asgardius/virtualx-engine
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Update libwebp to 0.6.0

Browse source
This commit is contained in:
volzhs 2017-02-17 23:49:40 +09:00
parent d5c2a6b76b
commit f7ef78c998
139 changed files with 10209 additions and 3709 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

194
modules/webp/SCsub
View file

@ -9,108 +9,118 @@ env_webp = env_modules.Clone()
if (env['builtin_libwebp'] != 'no'):
thirdparty_dir = "#thirdparty/libwebp/"
thirdparty_sources = [
"enc/webpenc.c",
"enc/near_lossless.c",
"enc/frame.c",
"enc/alpha.c",
"enc/picture_csp.c",
"enc/vp8l.c",
"enc/picture_psnr.c",
"enc/delta_palettization.c",
"enc/syntax.c",
"enc/backward_references.c",
"enc/token.c",
"enc/analysis.c",
"enc/iterator.c",
"enc/picture_tools.c",
"enc/picture_rescale.c",
"enc/config.c",
"enc/tree.c",
"enc/cost.c",
"enc/picture.c",
"enc/quant.c",
"enc/filter.c",
"enc/histogram.c",
"utils/rescaler.c",
"utils/filters.c",
"utils/quant_levels_dec.c",
"utils/huffman.c",
"utils/thread.c",
"utils/quant_levels.c",
"utils/bit_writer.c",
"utils/bit_reader.c",
"utils/random.c",
"utils/utils.c",
"utils/huffman_encode.c",
"utils/color_cache.c",
"mux/muxinternal.c",
"mux/muxread.c",
"mux/anim_encode.c",
"mux/muxedit.c",
"dec/webp.c",
"dec/frame.c",
"dec/alpha.c",
"dec/vp8l.c",
"dec/io.c",
"dec/vp8.c",
"dec/idec.c",
"dec/tree.c",
"dec/buffer.c",
"dec/quant.c",
"demux/demux.c",
"dec/alpha_dec.c",
"dec/buffer_dec.c",
"dec/frame_dec.c",
"dec/idec_dec.c",
"dec/io_dec.c",
"dec/quant_dec.c",
"dec/tree_dec.c",
"dec/vp8_dec.c",
"dec/vp8l_dec.c",
"dec/webp_dec.c",
"demux/anim_decode.c",
"dsp/yuv.c",
"dsp/filters_sse2.c",
"dsp/dec_sse41.c",
"dsp/rescaler.c",
"dsp/lossless_sse2.c",
"dsp/alpha_processing_sse41.c",
"dsp/alpha_processing_sse2.c",
"dsp/filters.c",
"dsp/upsampling_mips_dsp_r2.c",
"dsp/dec_neon.c",
"dsp/enc_neon.c",
"dsp/lossless_enc_mips32.c",
"dsp/lossless_enc_sse2.c",
"dsp/upsampling.c",
"dsp/lossless_enc_neon.c",
"demux/demux.c",
"dsp/alpha_processing.c",
"dsp/cost_sse2.c",
"dsp/dec_mips32.c",
"dsp/enc_avx2.c",
"dsp/rescaler_mips32.c",
"dsp/enc.c",
"dsp/lossless_enc_sse41.c",
"dsp/cost_mips32.c",
"dsp/lossless_mips_dsp_r2.c",
"dsp/filters_mips_dsp_r2.c",
"dsp/upsampling_neon.c",
"dsp/alpha_processing_mips_dsp_r2.c",
"dsp/enc_mips_dsp_r2.c",
"dsp/lossless.c",
"dsp/yuv_mips_dsp_r2.c",
"dsp/cost_mips_dsp_r2.c",
"dsp/alpha_processing_neon.c",
"dsp/alpha_processing_sse2.c",
"dsp/alpha_processing_sse41.c",
"dsp/argb.c",
"dsp/dec_sse2.c",
"dsp/rescaler_sse2.c",
"dsp/enc_sse41.c",
"dsp/argb_mips_dsp_r2.c",
"dsp/lossless_enc_mips_dsp_r2.c",
"dsp/dec_clip_tables.c",
"dsp/yuv_mips32.c",
"dsp/argb_sse2.c",
"dsp/cost.c",
"dsp/cost_mips32.c",
"dsp/cost_mips_dsp_r2.c",
"dsp/cost_sse2.c",
"dsp/cpu.c",
"dsp/dec.c",
"dsp/argb_sse2.c",
"dsp/lossless_neon.c",
"dsp/lossless_enc.c",
"dsp/enc_mips32.c",
"dsp/cost.c",
"dsp/rescaler_mips_dsp_r2.c",
"dsp/dec_clip_tables.c",
"dsp/dec_mips32.c",
"dsp/dec_mips_dsp_r2.c",
"dsp/rescaler_neon.c",
"dsp/yuv_sse2.c",
"dsp/dec_msa.c",
"dsp/dec_neon.c",
"dsp/dec_sse2.c",
"dsp/dec_sse41.c",
"dsp/enc_avx2.c",
"dsp/enc.c",
"dsp/enc_mips32.c",
"dsp/enc_mips_dsp_r2.c",
"dsp/enc_msa.c",
"dsp/enc_neon.c",
"dsp/enc_sse2.c",
"dsp/enc_sse41.c",
"dsp/filters.c",
"dsp/filters_mips_dsp_r2.c",
"dsp/filters_msa.c",
"dsp/filters_neon.c",
"dsp/filters_sse2.c",
"dsp/lossless.c",
"dsp/lossless_enc.c",
"dsp/lossless_enc_mips32.c",
"dsp/lossless_enc_mips_dsp_r2.c",
"dsp/lossless_enc_msa.c",
"dsp/lossless_enc_neon.c",
"dsp/lossless_enc_sse2.c",
"dsp/lossless_enc_sse41.c",
"dsp/lossless_mips_dsp_r2.c",
"dsp/lossless_msa.c",
"dsp/lossless_neon.c",
"dsp/lossless_sse2.c",
"dsp/rescaler.c",
"dsp/rescaler_mips32.c",
"dsp/rescaler_mips_dsp_r2.c",
"dsp/rescaler_msa.c",
"dsp/rescaler_neon.c",
"dsp/rescaler_sse2.c",
"dsp/upsampling.c",
"dsp/upsampling_mips_dsp_r2.c",
"dsp/upsampling_msa.c",
"dsp/upsampling_neon.c",
"dsp/upsampling_sse2.c",
"dsp/yuv.c",
"dsp/yuv_mips32.c",
"dsp/yuv_mips_dsp_r2.c",
"dsp/yuv_sse2.c",
"enc/alpha_enc.c",
"enc/analysis_enc.c",
"enc/backward_references_enc.c",
"enc/config_enc.c",
"enc/cost_enc.c",
"enc/delta_palettization_enc.c",
"enc/filter_enc.c",
"enc/frame_enc.c",
"enc/histogram_enc.c",
"enc/iterator_enc.c",
"enc/near_lossless_enc.c",
"enc/picture_csp_enc.c",
"enc/picture_enc.c",
"enc/picture_psnr_enc.c",
"enc/picture_rescale_enc.c",
"enc/picture_tools_enc.c",
"enc/predictor_enc.c",
"enc/quant_enc.c",
"enc/syntax_enc.c",
"enc/token_enc.c",
"enc/tree_enc.c",
"enc/vp8l_enc.c",
"enc/webp_enc.c",
"mux/anim_encode.c",
"mux/muxedit.c",
"mux/muxinternal.c",
"mux/muxread.c",
"utils/bit_reader_utils.c",
"utils/bit_writer_utils.c",
"utils/color_cache_utils.c",
"utils/filters_utils.c",
"utils/huffman_encode_utils.c",
"utils/huffman_utils.c",
"utils/quant_levels_dec_utils.c",
"utils/quant_levels_utils.c",
"utils/random_utils.c",
"utils/rescaler_utils.c",
"utils/thread_utils.c",
"utils/utils.c",
]
thirdparty_sources = [thirdparty_dir + file for file in thirdparty_sources]

6
thirdparty/README.md vendored
View file

@ -94,15 +94,15 @@ Files extracted from upstream source:
## libwebp
- Upstream: https://chromium.googlesource.com/webm/libwebp/
- Version: 0.5.2
- Version: 0.6.0
- License: BSD-3-Clause
Files extracted from upstream source:
- `src/*` except from: .am and .in, files, extras/, `webp/extras.h`
- `src/*` except from: .am, .rc and .in files
- AUTHORS, COPYING, PATENTS
Important: The files `utils/bit_reader.{c,h}` have Godot-made
Important: The files `utils/bit_reader_utils.{c,h}` have Godot-made
changes to ensure they build for Javascript/HTML5. Those
changes are marked with `// -- GODOT --` comments.

8
thirdparty/libwebp/dec/alpha.c → thirdparty/libwebp/dec/alpha_dec.c vendored
View file

@ -12,11 +12,11 @@
// Author: Skal (pascal.massimino@gmail.com)
#include <stdlib.h>
#include "./alphai.h"
#include "./vp8i.h"
#include "./vp8li.h"
#include "./alphai_dec.h"
#include "./vp8i_dec.h"
#include "./vp8li_dec.h"
#include "../dsp/dsp.h"
#include "../utils/quant_levels_dec.h"
#include "../utils/quant_levels_dec_utils.h"
#include "../utils/utils.h"
#include "../webp/format_constants.h"

4
thirdparty/libwebp/dec/alphai.h → thirdparty/libwebp/dec/alphai_dec.h vendored
View file

@ -14,8 +14,8 @@
#ifndef WEBP_DEC_ALPHAI_H_
#define WEBP_DEC_ALPHAI_H_
#include "./webpi.h"
#include "../utils/filters.h"
#include "./webpi_dec.h"
#include "../utils/filters_utils.h"
#ifdef __cplusplus
extern "C" {

4
thirdparty/libwebp/dec/buffer.c → thirdparty/libwebp/dec/buffer_dec.c vendored
View file

@ -13,8 +13,8 @@
#include <stdlib.h>
#include "./vp8i.h"
#include "./webpi.h"
#include "./vp8i_dec.h"
#include "./webpi_dec.h"
#include "../utils/utils.h"
//------------------------------------------------------------------------------

0
thirdparty/libwebp/dec/common.h → thirdparty/libwebp/dec/common_dec.h vendored
View file

4
thirdparty/libwebp/dec/frame.c → thirdparty/libwebp/dec/frame_dec.c vendored
View file

@ -12,7 +12,7 @@
// Author: Skal (pascal.massimino@gmail.com)
#include <stdlib.h>
#include "./vp8i.h"
#include "./vp8i_dec.h"
#include "../utils/utils.h"
//------------------------------------------------------------------------------
@ -723,7 +723,7 @@ static int AllocateMemory(VP8Decoder* const dec) {
return VP8SetError(dec, VP8_STATUS_OUT_OF_MEMORY,
"no memory during frame initialization.");
}
// down-cast is ok, thanks to WebPSafeAlloc() above.
// down-cast is ok, thanks to WebPSafeMalloc() above.
dec->mem_size_ = (size_t)needed;
}

6
thirdparty/libwebp/dec/idec.c → thirdparty/libwebp/dec/idec_dec.c vendored
View file

@ -15,9 +15,9 @@
#include <string.h>
#include <stdlib.h>
#include "./alphai.h"
#include "./webpi.h"
#include "./vp8i.h"
#include "./alphai_dec.h"
#include "./webpi_dec.h"
#include "./vp8i_dec.h"
#include "../utils/utils.h"
// In append mode, buffer allocations increase as multiples of this value.

113
thirdparty/libwebp/dec/io.c → thirdparty/libwebp/dec/io_dec.c vendored
View file

@ -13,8 +13,8 @@
#include <assert.h>
#include <stdlib.h>
#include "../dec/vp8i.h"
#include "./webpi.h"
#include "../dec/vp8i_dec.h"
#include "./webpi_dec.h"
#include "../dsp/dsp.h"
#include "../dsp/yuv.h"
#include "../utils/utils.h"
@ -256,7 +256,7 @@ static int Rescale(const uint8_t* src, int src_stride,
static int EmitRescaledYUV(const VP8Io* const io, WebPDecParams* const p) {
const int mb_h = io->mb_h;
const int uv_mb_h = (mb_h + 1) >> 1;
WebPRescaler* const scaler = &p->scaler_y;
WebPRescaler* const scaler = p->scaler_y;
int num_lines_out = 0;
if (WebPIsAlphaMode(p->output->colorspace) && io->a != NULL) {
// Before rescaling, we premultiply the luma directly into the io->y
@ -267,29 +267,28 @@ static int EmitRescaledYUV(const VP8Io* const io, WebPDecParams* const p) {
io->a, io->width, io->mb_w, mb_h, 0);
}
num_lines_out = Rescale(io->y, io->y_stride, mb_h, scaler);
Rescale(io->u, io->uv_stride, uv_mb_h, &p->scaler_u);
Rescale(io->v, io->uv_stride, uv_mb_h, &p->scaler_v);
Rescale(io->u, io->uv_stride, uv_mb_h, p->scaler_u);
Rescale(io->v, io->uv_stride, uv_mb_h, p->scaler_v);
return num_lines_out;
}
static int EmitRescaledAlphaYUV(const VP8Io* const io, WebPDecParams* const p,
int expected_num_lines_out) {
const WebPYUVABuffer* const buf = &p->output->u.YUVA;
uint8_t* const dst_a = buf->a + p->last_y * buf->a_stride;
if (io->a != NULL) {
uint8_t* dst_y = buf->y + p->last_y * buf->y_stride;
const uint8_t* src_a = buf->a + p->last_y * buf->a_stride;
const int num_lines_out = Rescale(io->a, io->width, io->mb_h, &p->scaler_a);
(void)expected_num_lines_out;
uint8_t* const dst_y = buf->y + p->last_y * buf->y_stride;
const int num_lines_out = Rescale(io->a, io->width, io->mb_h, p->scaler_a);
assert(expected_num_lines_out == num_lines_out);
if (num_lines_out > 0) { // unmultiply the Y
WebPMultRows(dst_y, buf->y_stride, src_a, buf->a_stride,
p->scaler_a.dst_width, num_lines_out, 1);
WebPMultRows(dst_y, buf->y_stride, dst_a, buf->a_stride,
p->scaler_a->dst_width, num_lines_out, 1);
}
} else if (buf->a != NULL) {
// the user requested alpha, but there is none, set it to opaque.
assert(p->last_y + expected_num_lines_out <= io->scaled_height);
FillAlphaPlane(buf->a + p->last_y * buf->a_stride,
io->scaled_width, expected_num_lines_out, buf->a_stride);
FillAlphaPlane(dst_a, io->scaled_width, expected_num_lines_out,
buf->a_stride);
}
return 0;
}
@ -305,31 +304,42 @@ static int InitYUVRescaler(const VP8Io* const io, WebPDecParams* const p) {
const int uv_in_height = (io->mb_h + 1) >> 1;
const size_t work_size = 2 * out_width; // scratch memory for luma rescaler
const size_t uv_work_size = 2 * uv_out_width; // and for each u/v ones
size_t tmp_size;
size_t tmp_size, rescaler_size;
rescaler_t* work;
WebPRescaler* scalers;
const int num_rescalers = has_alpha ? 4 : 3;
tmp_size = (work_size + 2 * uv_work_size) * sizeof(*work);
if (has_alpha) {
tmp_size += work_size * sizeof(*work);
}
p->memory = WebPSafeMalloc(1ULL, tmp_size);
rescaler_size = num_rescalers * sizeof(*p->scaler_y) + WEBP_ALIGN_CST;
p->memory = WebPSafeMalloc(1ULL, tmp_size + rescaler_size);
if (p->memory == NULL) {
return 0; // memory error
}
work = (rescaler_t*)p->memory;
WebPRescalerInit(&p->scaler_y, io->mb_w, io->mb_h,
scalers = (WebPRescaler*)WEBP_ALIGN((const uint8_t*)work + tmp_size);
p->scaler_y = &scalers[0];
p->scaler_u = &scalers[1];
p->scaler_v = &scalers[2];
p->scaler_a = has_alpha ? &scalers[3] : NULL;
WebPRescalerInit(p->scaler_y, io->mb_w, io->mb_h,
buf->y, out_width, out_height, buf->y_stride, 1,
work);
WebPRescalerInit(&p->scaler_u, uv_in_width, uv_in_height,
WebPRescalerInit(p->scaler_u, uv_in_width, uv_in_height,
buf->u, uv_out_width, uv_out_height, buf->u_stride, 1,
work + work_size);
WebPRescalerInit(&p->scaler_v, uv_in_width, uv_in_height,
WebPRescalerInit(p->scaler_v, uv_in_width, uv_in_height,
buf->v, uv_out_width, uv_out_height, buf->v_stride, 1,
work + work_size + uv_work_size);
p->emit = EmitRescaledYUV;
if (has_alpha) {
WebPRescalerInit(&p->scaler_a, io->mb_w, io->mb_h,
WebPRescalerInit(p->scaler_a, io->mb_w, io->mb_h,
buf->a, out_width, out_height, buf->a_stride, 1,
work + work_size + 2 * uv_work_size);
p->emit_alpha = EmitRescaledAlphaYUV;
@ -349,15 +359,15 @@ static int ExportRGB(WebPDecParams* const p, int y_pos) {
int num_lines_out = 0;
// For RGB rescaling, because of the YUV420, current scan position
// U/V can be +1/-1 line from the Y one. Hence the double test.
while (WebPRescalerHasPendingOutput(&p->scaler_y) &&
WebPRescalerHasPendingOutput(&p->scaler_u)) {
while (WebPRescalerHasPendingOutput(p->scaler_y) &&
WebPRescalerHasPendingOutput(p->scaler_u)) {
assert(y_pos + num_lines_out < p->output->height);
assert(p->scaler_u.y_accum == p->scaler_v.y_accum);
WebPRescalerExportRow(&p->scaler_y);
WebPRescalerExportRow(&p->scaler_u);
WebPRescalerExportRow(&p->scaler_v);
convert(p->scaler_y.dst, p->scaler_u.dst, p->scaler_v.dst,
dst, p->scaler_y.dst_width);
assert(p->scaler_u->y_accum == p->scaler_v->y_accum);
WebPRescalerExportRow(p->scaler_y);
WebPRescalerExportRow(p->scaler_u);
WebPRescalerExportRow(p->scaler_v);
convert(p->scaler_y->dst, p->scaler_u->dst, p->scaler_v->dst,
dst, p->scaler_y->dst_width);
dst += buf->stride;
++num_lines_out;
}
@ -371,15 +381,15 @@ static int EmitRescaledRGB(const VP8Io* const io, WebPDecParams* const p) {
int num_lines_out = 0;
while (j < mb_h) {
const int y_lines_in =
WebPRescalerImport(&p->scaler_y, mb_h - j,
WebPRescalerImport(p->scaler_y, mb_h - j,
io->y + j * io->y_stride, io->y_stride);
j += y_lines_in;
if (WebPRescaleNeededLines(&p->scaler_u, uv_mb_h - uv_j)) {
if (WebPRescaleNeededLines(p->scaler_u, uv_mb_h - uv_j)) {
const int u_lines_in =
WebPRescalerImport(&p->scaler_u, uv_mb_h - uv_j,
WebPRescalerImport(p->scaler_u, uv_mb_h - uv_j,
io->u + uv_j * io->uv_stride, io->uv_stride);
const int v_lines_in =
WebPRescalerImport(&p->scaler_v, uv_mb_h - uv_j,
WebPRescalerImport(p->scaler_v, uv_mb_h - uv_j,
io->v + uv_j * io->uv_stride, io->uv_stride);
(void)v_lines_in; // remove a gcc warning
assert(u_lines_in == v_lines_in);
@ -400,13 +410,13 @@ static int ExportAlpha(WebPDecParams* const p, int y_pos, int max_lines_out) {
int num_lines_out = 0;
const int is_premult_alpha = WebPIsPremultipliedMode(colorspace);
uint32_t non_opaque = 0;
const int width = p->scaler_a.dst_width;
const int width = p->scaler_a->dst_width;
while (WebPRescalerHasPendingOutput(&p->scaler_a) &&
while (WebPRescalerHasPendingOutput(p->scaler_a) &&
num_lines_out < max_lines_out) {
assert(y_pos + num_lines_out < p->output->height);
WebPRescalerExportRow(&p->scaler_a);
non_opaque |= WebPDispatchAlpha(p->scaler_a.dst, 0, width, 1, dst, 0);
WebPRescalerExportRow(p->scaler_a);
non_opaque |= WebPDispatchAlpha(p->scaler_a->dst, 0, width, 1, dst, 0);
dst += buf->stride;
++num_lines_out;
}
@ -428,18 +438,18 @@ static int ExportAlphaRGBA4444(WebPDecParams* const p, int y_pos,
#endif
int num_lines_out = 0;
const WEBP_CSP_MODE colorspace = p->output->colorspace;
const int width = p->scaler_a.dst_width;
const int width = p->scaler_a->dst_width;
const int is_premult_alpha = WebPIsPremultipliedMode(colorspace);
uint32_t alpha_mask = 0x0f;
while (WebPRescalerHasPendingOutput(&p->scaler_a) &&
while (WebPRescalerHasPendingOutput(p->scaler_a) &&
num_lines_out < max_lines_out) {
int i;
assert(y_pos + num_lines_out < p->output->height);
WebPRescalerExportRow(&p->scaler_a);
WebPRescalerExportRow(p->scaler_a);
for (i = 0; i < width; ++i) {
// Fill in the alpha value (converted to 4 bits).
const uint32_t alpha_value = p->scaler_a.dst[i] >> 4;
const uint32_t alpha_value = p->scaler_a->dst[i] >> 4;
alpha_dst[2 * i] = (alpha_dst[2 * i] & 0xf0) | alpha_value;
alpha_mask &= alpha_value;
}
@ -455,7 +465,7 @@ static int ExportAlphaRGBA4444(WebPDecParams* const p, int y_pos,
static int EmitRescaledAlphaRGB(const VP8Io* const io, WebPDecParams* const p,
int expected_num_out_lines) {
if (io->a != NULL) {
WebPRescaler* const scaler = &p->scaler_a;
WebPRescaler* const scaler = p->scaler_a;
int lines_left = expected_num_out_lines;
const int y_end = p->last_y + lines_left;
while (lines_left > 0) {
@ -477,7 +487,9 @@ static int InitRGBRescaler(const VP8Io* const io, WebPDecParams* const p) {
const size_t work_size = 2 * out_width; // scratch memory for one rescaler
rescaler_t* work; // rescalers work area
uint8_t* tmp; // tmp storage for scaled YUV444 samples before RGB conversion
size_t tmp_size1, tmp_size2, total_size;
size_t tmp_size1, tmp_size2, total_size, rescaler_size;
WebPRescaler* scalers;
const int num_rescalers = has_alpha ? 4 : 3;
tmp_size1 = 3 * work_size;
tmp_size2 = 3 * out_width;
@ -486,26 +498,35 @@ static int InitRGBRescaler(const VP8Io* const io, WebPDecParams* const p) {
tmp_size2 += out_width;
}
total_size = tmp_size1 * sizeof(*work) + tmp_size2 * sizeof(*tmp);
p->memory = WebPSafeMalloc(1ULL, total_size);
rescaler_size = num_rescalers * sizeof(*p->scaler_y) + WEBP_ALIGN_CST;
p->memory = WebPSafeMalloc(1ULL, total_size + rescaler_size);
if (p->memory == NULL) {
return 0; // memory error
}
work = (rescaler_t*)p->memory;
tmp = (uint8_t*)(work + tmp_size1);
WebPRescalerInit(&p->scaler_y, io->mb_w, io->mb_h,
scalers = (WebPRescaler*)WEBP_ALIGN((const uint8_t*)work + total_size);
p->scaler_y = &scalers[0];
p->scaler_u = &scalers[1];
p->scaler_v = &scalers[2];
p->scaler_a = has_alpha ? &scalers[3] : NULL;
WebPRescalerInit(p->scaler_y, io->mb_w, io->mb_h,
tmp + 0 * out_width, out_width, out_height, 0, 1,
work + 0 * work_size);
WebPRescalerInit(&p->scaler_u, uv_in_width, uv_in_height,
WebPRescalerInit(p->scaler_u, uv_in_width, uv_in_height,
tmp + 1 * out_width, out_width, out_height, 0, 1,
work + 1 * work_size);
WebPRescalerInit(&p->scaler_v, uv_in_width, uv_in_height,
WebPRescalerInit(p->scaler_v, uv_in_width, uv_in_height,
tmp + 2 * out_width, out_width, out_height, 0, 1,
work + 2 * work_size);
p->emit = EmitRescaledRGB;
WebPInitYUV444Converters();
if (has_alpha) {
WebPRescalerInit(&p->scaler_a, io->mb_w, io->mb_h,
WebPRescalerInit(p->scaler_a, io->mb_w, io->mb_h,
tmp + 3 * out_width, out_width, out_height, 0, 1,
work + 3 * work_size);
p->emit_alpha = EmitRescaledAlphaRGB;

2
thirdparty/libwebp/dec/quant.c → thirdparty/libwebp/dec/quant_dec.c vendored
View file

@ -11,7 +11,7 @@
//
// Author: Skal (pascal.massimino@gmail.com)
#include "./vp8i.h"
#include "./vp8i_dec.h"
static WEBP_INLINE int clip(int v, int M) {
return v < 0 ? 0 : v > M ? M : v;

7
thirdparty/libwebp/dec/tree.c → thirdparty/libwebp/dec/tree_dec.c vendored
View file

@ -11,10 +11,13 @@
//
// Author: Skal (pascal.massimino@gmail.com)
#include "./vp8i.h"
#include "../utils/bit_reader_inl.h"
#include "./vp8i_dec.h"
#include "../utils/bit_reader_inl_utils.h"
#if !defined(__arm__) && !defined(_M_ARM) && !defined(__aarch64__)
// using a table is ~1-2% slower on ARM. Prefer the coded-tree approach then.
#define USE_GENERIC_TREE
#endif
#ifdef USE_GENERIC_TREE
static const int8_t kYModesIntra4[18] = {

72
thirdparty/libwebp/dec/vp8.c → thirdparty/libwebp/dec/vp8_dec.c vendored
View file

@ -13,11 +13,11 @@
#include <stdlib.h>
#include "./alphai.h"
#include "./vp8i.h"
#include "./vp8li.h"
#include "./webpi.h"
#include "../utils/bit_reader_inl.h"
#include "./alphai_dec.h"
#include "./vp8i_dec.h"
#include "./vp8li_dec.h"
#include "./webpi_dec.h"
#include "../utils/bit_reader_inl_utils.h"
#include "../utils/utils.h"
//------------------------------------------------------------------------------
@ -26,6 +26,16 @@ int WebPGetDecoderVersion(void) {
return (DEC_MAJ_VERSION << 16) | (DEC_MIN_VERSION << 8) | DEC_REV_VERSION;
}
//------------------------------------------------------------------------------
// Signature and pointer-to-function for GetCoeffs() variants below.
typedef int (*GetCoeffsFunc)(VP8BitReader* const br,
const VP8BandProbas* const prob[],
int ctx, const quant_t dq, int n, int16_t* out);
static volatile GetCoeffsFunc GetCoeffs = NULL;
static void InitGetCoeffs(void);
//------------------------------------------------------------------------------
// VP8Decoder
@ -51,6 +61,7 @@ VP8Decoder* VP8New(void) {
WebPGetWorkerInterface()->Init(&dec->worker_);
dec->ready_ = 0;
dec->num_parts_minus_one_ = 0;
InitGetCoeffs();
}
return dec;
}
@ -273,12 +284,14 @@ int VP8GetHeaders(VP8Decoder* const dec, VP8Io* const io) {
frm_hdr->profile_ = (bits >> 1) & 7;
frm_hdr->show_ = (bits >> 4) & 1;
frm_hdr->partition_length_ = (bits >> 5);
if (frm_hdr->profile_ > 3)
if (frm_hdr->profile_ > 3) {
return VP8SetError(dec, VP8_STATUS_BITSTREAM_ERROR,
"Incorrect keyframe parameters.");
if (!frm_hdr->show_)
}
if (!frm_hdr->show_) {
return VP8SetError(dec, VP8_STATUS_UNSUPPORTED_FEATURE,
"Frame not displayable.");
}
buf += 3;
buf_size -= 3;
}
@ -420,8 +433,9 @@ static int GetLargeValue(VP8BitReader* const br, const uint8_t* const p) {
}
// Returns the position of the last non-zero coeff plus one
static int GetCoeffs(VP8BitReader* const br, const VP8BandProbas* const prob[],
int ctx, const quant_t dq, int n, int16_t* out) {
static int GetCoeffsFast(VP8BitReader* const br,
const VP8BandProbas* const prob[],
int ctx, const quant_t dq, int n, int16_t* out) {
const uint8_t* p = prob[n]->probas_[ctx];
for (; n < 16; ++n) {
if (!VP8GetBit(br, p[0])) {
@ -447,6 +461,46 @@ static int GetCoeffs(VP8BitReader* const br, const VP8BandProbas* const prob[],
return 16;
}
// This version of GetCoeffs() uses VP8GetBitAlt() which is an alternate version
// of VP8GetBitAlt() targeting specific platforms.
static int GetCoeffsAlt(VP8BitReader* const br,
const VP8BandProbas* const prob[],
int ctx, const quant_t dq, int n, int16_t* out) {
const uint8_t* p = prob[n]->probas_[ctx];
for (; n < 16; ++n) {
if (!VP8GetBitAlt(br, p[0])) {
return n; // previous coeff was last non-zero coeff
}
while (!VP8GetBitAlt(br, p[1])) { // sequence of zero coeffs
p = prob[++n]->probas_[0];
if (n == 16) return 16;
}
{ // non zero coeff
const VP8ProbaArray* const p_ctx = &prob[n + 1]->probas_[0];
int v;
if (!VP8GetBitAlt(br, p[2])) {
v = 1;
p = p_ctx[1];
} else {
v = GetLargeValue(br, p);
p = p_ctx[2];
}
out[kZigzag[n]] = VP8GetSigned(br, v) * dq[n > 0];
}
}
return 16;
}
WEBP_TSAN_IGNORE_FUNCTION static void InitGetCoeffs(void) {
if (GetCoeffs == NULL) {
if (VP8GetCPUInfo != NULL && VP8GetCPUInfo(kSlowSSSE3)) {
GetCoeffs = GetCoeffsAlt;
} else {
GetCoeffs = GetCoeffsFast;
}
}
}
static WEBP_INLINE uint32_t NzCodeBits(uint32_t nz_coeffs, int nz, int dc_nz) {
nz_coeffs <<= 2;
nz_coeffs |= (nz > 3) ? 3 : (nz > 1) ? 2 : dc_nz;

0
thirdparty/libwebp/dec/decode_vp8.h → thirdparty/libwebp/dec/vp8_dec.h vendored
View file

14
thirdparty/libwebp/dec/vp8i.h → thirdparty/libwebp/dec/vp8i_dec.h vendored
View file

@ -15,11 +15,11 @@
#define WEBP_DEC_VP8I_H_
#include <string.h> // for memcpy()
#include "./common.h"
#include "./vp8li.h"
#include "../utils/bit_reader.h"
#include "../utils/random.h"
#include "../utils/thread.h"
#include "./common_dec.h"
#include "./vp8li_dec.h"
#include "../utils/bit_reader_utils.h"
#include "../utils/random_utils.h"
#include "../utils/thread_utils.h"
#include "../dsp/dsp.h"
#ifdef __cplusplus
@ -31,8 +31,8 @@ extern "C" {
// version numbers
#define DEC_MAJ_VERSION 0
#define DEC_MIN_VERSION 5
#define DEC_REV_VERSION 2
#define DEC_MIN_VERSION 6
#define DEC_REV_VERSION 0
// YUV-cache parameters. Cache is 32-bytes wide (= one cacheline).
// Constraints are: We need to store one 16x16 block of luma samples (y),

37
thirdparty/libwebp/dec/vp8l.c → thirdparty/libwebp/dec/vp8l_dec.c vendored
View file

@ -14,13 +14,14 @@
#include <stdlib.h>
#include "./alphai.h"
#include "./vp8li.h"
#include "./alphai_dec.h"
#include "./vp8li_dec.h"
#include "../dsp/dsp.h"
#include "../dsp/lossless.h"
#include "../dsp/lossless_common.h"
#include "../dsp/yuv.h"
#include "../utils/endian_inl.h"
#include "../utils/huffman.h"
#include "../utils/endian_inl_utils.h"
#include "../utils/huffman_utils.h"
#include "../utils/utils.h"
#define NUM_ARGB_CACHE_ROWS 16
@ -547,11 +548,14 @@ static int EmitRescaledRowsRGBA(const VP8LDecoder* const dec,
uint8_t* const row_out = out + num_lines_out * out_stride;
const int lines_left = mb_h - num_lines_in;
const int needed_lines = WebPRescaleNeededLines(dec->rescaler, lines_left);
int lines_imported;
assert(needed_lines > 0 && needed_lines <= lines_left);
WebPMultARGBRows(row_in, in_stride,
dec->rescaler->src_width, needed_lines, 0);
WebPRescalerImport(dec->rescaler, lines_left, row_in, in_stride);
num_lines_in += needed_lines;
lines_imported =
WebPRescalerImport(dec->rescaler, lines_left, row_in, in_stride);
assert(lines_imported == needed_lines);
num_lines_in += lines_imported;
num_lines_out += Export(dec->rescaler, colorspace, out_stride, row_out);
}
return num_lines_out;
@ -623,9 +627,12 @@ static int EmitRescaledRowsYUVA(const VP8LDecoder* const dec,
while (num_lines_in < mb_h) {
const int lines_left = mb_h - num_lines_in;
const int needed_lines = WebPRescaleNeededLines(dec->rescaler, lines_left);
int lines_imported;
WebPMultARGBRows(in, in_stride, dec->rescaler->src_width, needed_lines, 0);
WebPRescalerImport(dec->rescaler, lines_left, in, in_stride);
num_lines_in += needed_lines;
lines_imported =
WebPRescalerImport(dec->rescaler, lines_left, in, in_stride);
assert(lines_imported == needed_lines);
num_lines_in += lines_imported;
in += needed_lines * in_stride;
y_pos += ExportYUVA(dec, y_pos);
}
@ -705,13 +712,15 @@ static void ApplyInverseTransforms(VP8LDecoder* const dec, int num_rows,
uint32_t* const rows_out = dec->argb_cache_;
// Inverse transforms.
// TODO: most transforms only need to operate on the cropped region only.
memcpy(rows_out, rows_in, cache_pixs * sizeof(*rows_out));
while (n-- > 0) {
VP8LTransform* const transform = &dec->transforms_[n];
VP8LInverseTransform(transform, start_row, end_row, rows_in, rows_out);
rows_in = rows_out;
}
if (rows_in != rows_out) {
// No transform called, hence just copy.
memcpy(rows_out, rows_in, cache_pixs * sizeof(*rows_out));
}
}
// Processes (transforms, scales & color-converts) the rows decoded after the
@ -1210,8 +1219,9 @@ static int ExpandColorMap(int num_colors, VP8LTransform* const transform) {
// Equivalent to AddPixelEq(), on a byte-basis.
new_data[i] = (data[i] + new_data[i - 4]) & 0xff;
}
for (; i < 4 * final_num_colors; ++i)
for (; i < 4 * final_num_colors; ++i) {
new_data[i] = 0; // black tail.
}
WebPSafeFree(transform->data_);
transform->data_ = new_color_map;
}
@ -1482,9 +1492,8 @@ static void ExtractAlphaRows(VP8LDecoder* const dec, int last_row) {
const int cache_pixs = width * num_rows_to_process;
uint8_t* const dst = output + width * cur_row;
const uint32_t* const src = dec->argb_cache_;
int i;
ApplyInverseTransforms(dec, num_rows_to_process, in);
for (i = 0; i < cache_pixs; ++i) dst[i] = (src[i] >> 8) & 0xff;
WebPExtractGreen(src, dst, cache_pixs);
AlphaApplyFilter(alph_dec,
cur_row, cur_row + num_rows_to_process, dst, width);
num_rows -= num_rows_to_process;
@ -1552,6 +1561,8 @@ int VP8LDecodeAlphaImageStream(ALPHDecoder* const alph_dec, int last_row) {
return 1; // done
}
if (!alph_dec->use_8b_decode_) WebPInitAlphaProcessing();
// Decode (with special row processing).
return alph_dec->use_8b_decode_ ?
DecodeAlphaData(dec, (uint8_t*)dec->pixels_, dec->width_, dec->height_,

8
thirdparty/libwebp/dec/vp8li.h → thirdparty/libwebp/dec/vp8li_dec.h vendored
View file

@ -16,10 +16,10 @@
#define WEBP_DEC_VP8LI_H_
#include <string.h> // for memcpy()
#include "./webpi.h"
#include "../utils/bit_reader.h"
#include "../utils/color_cache.h"
#include "../utils/huffman.h"
#include "./webpi_dec.h"
#include "../utils/bit_reader_utils.h"
#include "../utils/color_cache_utils.h"
#include "../utils/huffman_utils.h"
#ifdef __cplusplus
extern "C" {

17
thirdparty/libwebp/dec/webp.c → thirdparty/libwebp/dec/webp_dec.c vendored
View file

@ -13,9 +13,9 @@
#include <stdlib.h>
#include "./vp8i.h"
#include "./vp8li.h"
#include "./webpi.h"
#include "./vp8i_dec.h"
#include "./vp8li_dec.h"
#include "./webpi_dec.h"
#include "../utils/utils.h"
#include "../webp/mux_types.h" // ALPHA_FLAG
@ -39,8 +39,8 @@
// 20..23 VP8X flags bit-map corresponding to the chunk-types present.
// 24..26 Width of the Canvas Image.
// 27..29 Height of the Canvas Image.
// There can be extra chunks after the "VP8X" chunk (ICCP, FRGM, ANMF, VP8,
// VP8L, XMP, EXIF ...)
// There can be extra chunks after the "VP8X" chunk (ICCP, ANMF, VP8, VP8L,
// XMP, EXIF ...)
// All sizes are in little-endian order.
// Note: chunk data size must be padded to multiple of 2 when written.
@ -289,7 +289,6 @@ static VP8StatusCode ParseHeadersInternal(const uint8_t* data,
int found_riff = 0;
int found_vp8x = 0;
int animation_present = 0;
int fragments_present = 0;
const int have_all_data = (headers != NULL) ? headers->have_all_data : 0;
VP8StatusCode status;
@ -318,7 +317,6 @@ static VP8StatusCode ParseHeadersInternal(const uint8_t* data,
return status; // Wrong VP8X / insufficient data.
}
animation_present = !!(flags & ANIMATION_FLAG);
fragments_present = !!(flags & FRAGMENTS_FLAG);
if (!found_riff && found_vp8x) {
// Note: This restriction may be removed in the future, if it becomes
// necessary to send VP8X chunk to the decoder.
@ -330,8 +328,7 @@ static VP8StatusCode ParseHeadersInternal(const uint8_t* data,
image_width = canvas_width;
image_height = canvas_height;
if (found_vp8x && (animation_present || fragments_present) &&
headers == NULL) {
if (found_vp8x && animation_present && headers == NULL) {
status = VP8_STATUS_OK;
goto ReturnWidthHeight; // Just return features from VP8X header.
}
@ -362,7 +359,7 @@ static VP8StatusCode ParseHeadersInternal(const uint8_t* data,
return VP8_STATUS_BITSTREAM_ERROR;
}
if (format != NULL && !(animation_present || fragments_present)) {
if (format != NULL && !animation_present) {
*format = hdrs.is_lossless ? 2 : 1;
}

17
thirdparty/libwebp/dec/webpi.h → thirdparty/libwebp/dec/webpi_dec.h vendored
View file

@ -18,8 +18,8 @@
extern "C" {
#endif
#include "../utils/rescaler.h"
#include "./decode_vp8.h"
#include "../utils/rescaler_utils.h"
#include "./vp8_dec.h"
//------------------------------------------------------------------------------
// WebPDecParams: Decoding output parameters. Transient internal object.
@ -38,27 +38,18 @@ struct WebPDecParams {
int last_y; // coordinate of the line that was last output
const WebPDecoderOptions* options; // if not NULL, use alt decoding features
// rescalers
WebPRescaler scaler_y, scaler_u, scaler_v, scaler_a;
WebPRescaler* scaler_y, *scaler_u, *scaler_v, *scaler_a; // rescalers
void* memory; // overall scratch memory for the output work.
OutputFunc emit; // output RGB or YUV samples
OutputAlphaFunc emit_alpha; // output alpha channel
OutputRowFunc emit_alpha_row; // output one line of rescaled alpha values
WebPDecBuffer* final_output; // In case the user supplied a slow-memory
// output, we decode image in temporary buffer
// (this::output) and copy it here.
WebPDecBuffer tmp_buffer; // this::output will point to this one in case
// of slow memory.
};
// Should be called first, before any use of the WebPDecParams object.
void WebPResetDecParams(WebPDecParams* const params);
// Delete all memory (after an error occurred, for instance)
void WebPFreeDecParams(WebPDecParams* const params);
//------------------------------------------------------------------------------
// Header parsing helpers

5
thirdparty/libwebp/demux/demux.c vendored
View file

@ -25,7 +25,7 @@
#define DMUX_MAJ_VERSION 0
#define DMUX_MIN_VERSION 3
#define DMUX_REV_VERSION 1
#define DMUX_REV_VERSION 2
typedef struct {
size_t start_; // start location of the data
@ -590,7 +590,6 @@ static int CheckFrameBounds(const Frame* const frame, int exact,
static int IsValidExtendedFormat(const WebPDemuxer* const dmux) {
const int is_animation = !!(dmux->feature_flags_ & ANIMATION_FLAG);
const int is_fragmented = !!(dmux->feature_flags_ & FRAGMENTS_FLAG);
const Frame* f = dmux->frames_;
if (dmux->state_ == WEBP_DEMUX_PARSING_HEADER) return 1;
@ -598,7 +597,7 @@ static int IsValidExtendedFormat(const WebPDemuxer* const dmux) {
if (dmux->canvas_width_ <= 0 || dmux->canvas_height_ <= 0) return 0;
if (dmux->loop_count_ < 0) return 0;
if (dmux->state_ == WEBP_DEMUX_DONE && dmux->frames_ == NULL) return 0;
if (is_fragmented) return 0;
if (dmux->feature_flags_ & ~ALL_VALID_FLAGS) return 0; // invalid bitstream
while (f != NULL) {
const int cur_frame_set = f->frame_num_;

38
thirdparty/libwebp/dsp/alpha_processing.c vendored
View file

@ -284,9 +284,9 @@ static void ApplyAlphaMultiply_16b(uint8_t* rgba4444,
#endif
}
static int DispatchAlpha(const uint8_t* alpha, int alpha_stride,
int width, int height,
uint8_t* dst, int dst_stride) {
static int DispatchAlpha_C(const uint8_t* alpha, int alpha_stride,
int width, int height,
uint8_t* dst, int dst_stride) {
uint32_t alpha_mask = 0xff;
int i, j;
@ -303,9 +303,9 @@ static int DispatchAlpha(const uint8_t* alpha, int alpha_stride,
return (alpha_mask != 0xff);
}
static void DispatchAlphaToGreen(const uint8_t* alpha, int alpha_stride,
int width, int height,
uint32_t* dst, int dst_stride) {
static void DispatchAlphaToGreen_C(const uint8_t* alpha, int alpha_stride,
int width, int height,
uint32_t* dst, int dst_stride) {
int i, j;
for (j = 0; j < height; ++j) {
for (i = 0; i < width; ++i) {
@ -316,9 +316,9 @@ static void DispatchAlphaToGreen(const uint8_t* alpha, int alpha_stride,
}
}
static int ExtractAlpha(const uint8_t* argb, int argb_stride,
int width, int height,
uint8_t* alpha, int alpha_stride) {
static int ExtractAlpha_C(const uint8_t* argb, int argb_stride,
int width, int height,
uint8_t* alpha, int alpha_stride) {
uint8_t alpha_mask = 0xff;
int i, j;
@ -334,11 +334,17 @@ static int ExtractAlpha(const uint8_t* argb, int argb_stride,
return (alpha_mask == 0xff);
}
static void ExtractGreen_C(const uint32_t* argb, uint8_t* alpha, int size) {
int i;
for (i = 0; i < size; ++i) alpha[i] = argb[i] >> 8;
}
void (*WebPApplyAlphaMultiply)(uint8_t*, int, int, int, int);
void (*WebPApplyAlphaMultiply4444)(uint8_t*, int, int, int);
int (*WebPDispatchAlpha)(const uint8_t*, int, int, int, uint8_t*, int);
void (*WebPDispatchAlphaToGreen)(const uint8_t*, int, int, int, uint32_t*, int);
int (*WebPExtractAlpha)(const uint8_t*, int, int, int, uint8_t*, int);
void (*WebPExtractGreen)(const uint32_t* argb, uint8_t* alpha, int size);
//------------------------------------------------------------------------------
// Init function
@ -346,6 +352,7 @@ int (*WebPExtractAlpha)(const uint8_t*, int, int, int, uint8_t*, int);
extern void WebPInitAlphaProcessingMIPSdspR2(void);
extern void WebPInitAlphaProcessingSSE2(void);
extern void WebPInitAlphaProcessingSSE41(void);
extern void WebPInitAlphaProcessingNEON(void);
static volatile VP8CPUInfo alpha_processing_last_cpuinfo_used =
(VP8CPUInfo)&alpha_processing_last_cpuinfo_used;
@ -357,9 +364,11 @@ WEBP_TSAN_IGNORE_FUNCTION void WebPInitAlphaProcessing(void) {
WebPMultRow = WebPMultRowC;
WebPApplyAlphaMultiply = ApplyAlphaMultiply;
WebPApplyAlphaMultiply4444 = ApplyAlphaMultiply_16b;
WebPDispatchAlpha = DispatchAlpha;
WebPDispatchAlphaToGreen = DispatchAlphaToGreen;
WebPExtractAlpha = ExtractAlpha;
WebPDispatchAlpha = DispatchAlpha_C;
WebPDispatchAlphaToGreen = DispatchAlphaToGreen_C;
WebPExtractAlpha = ExtractAlpha_C;
WebPExtractGreen = ExtractGreen_C;
// If defined, use CPUInfo() to overwrite some pointers with faster versions.
if (VP8GetCPUInfo != NULL) {
@ -373,6 +382,11 @@ WEBP_TSAN_IGNORE_FUNCTION void WebPInitAlphaProcessing(void) {
#endif
}
#endif
#if defined(WEBP_USE_NEON)
if (VP8GetCPUInfo(kNEON)) {
WebPInitAlphaProcessingNEON();
}
#endif
#if defined(WEBP_USE_MIPS_DSP_R2)
if (VP8GetCPUInfo(kMIPSdspR2)) {
WebPInitAlphaProcessingMIPSdspR2();

191
thirdparty/libwebp/dsp/alpha_processing_neon.c vendored Normal file
View file

@ -0,0 +1,191 @@
// Copyright 2017 Google Inc. All Rights Reserved.
//
// Use of this source code is governed by a BSD-style license
// that can be found in the COPYING file in the root of the source
// tree. An additional intellectual property rights grant can be found
// in the file PATENTS. All contributing project authors may
// be found in the AUTHORS file in the root of the source tree.
// -----------------------------------------------------------------------------
//
// Utilities for processing transparent channel, NEON version.
//
// Author: Skal (pascal.massimino@gmail.com)
#include "./dsp.h"
#if defined(WEBP_USE_NEON)
#include "./neon.h"
//------------------------------------------------------------------------------
#define MULTIPLIER(a) ((a) * 0x8081)
#define PREMULTIPLY(x, m) (((x) * (m)) >> 23)
#define MULTIPLY_BY_ALPHA(V, ALPHA, OTHER) do { \
const uint8x8_t alpha = (V).val[(ALPHA)]; \
const uint16x8_t r1 = vmull_u8((V).val[1], alpha); \
const uint16x8_t g1 = vmull_u8((V).val[2], alpha); \
const uint16x8_t b1 = vmull_u8((V).val[(OTHER)], alpha); \
/* we use: v / 255 = (v + 1 + (v >> 8)) >> 8 */ \
const uint16x8_t r2 = vsraq_n_u16(r1, r1, 8); \
const uint16x8_t g2 = vsraq_n_u16(g1, g1, 8); \
const uint16x8_t b2 = vsraq_n_u16(b1, b1, 8); \
const uint16x8_t r3 = vaddq_u16(r2, kOne); \
const uint16x8_t g3 = vaddq_u16(g2, kOne); \
const uint16x8_t b3 = vaddq_u16(b2, kOne); \
(V).val[1] = vshrn_n_u16(r3, 8); \
(V).val[2] = vshrn_n_u16(g3, 8); \
(V).val[(OTHER)] = vshrn_n_u16(b3, 8); \
} while (0)
static void ApplyAlphaMultiply_NEON(uint8_t* rgba, int alpha_first,
int w, int h, int stride) {
const uint16x8_t kOne = vdupq_n_u16(1u);
while (h-- > 0) {
uint32_t* const rgbx = (uint32_t*)rgba;
int i = 0;
if (alpha_first) {
for (; i + 8 <= w; i += 8) {
// load aaaa...|rrrr...|gggg...|bbbb...
uint8x8x4_t RGBX = vld4_u8((const uint8_t*)(rgbx + i));
MULTIPLY_BY_ALPHA(RGBX, 0, 3);
vst4_u8((uint8_t*)(rgbx + i), RGBX);
}
} else {
for (; i + 8 <= w; i += 8) {
uint8x8x4_t RGBX = vld4_u8((const uint8_t*)(rgbx + i));
MULTIPLY_BY_ALPHA(RGBX, 3, 0);
vst4_u8((uint8_t*)(rgbx + i), RGBX);
}
}
// Finish with left-overs.
for (; i < w; ++i) {
uint8_t* const rgb = rgba + (alpha_first ? 1 : 0);
const uint8_t* const alpha = rgba + (alpha_first ? 0 : 3);
const uint32_t a = alpha[4 * i];
if (a != 0xff) {
const uint32_t mult = MULTIPLIER(a);
rgb[4 * i + 0] = PREMULTIPLY(rgb[4 * i + 0], mult);
rgb[4 * i + 1] = PREMULTIPLY(rgb[4 * i + 1], mult);
rgb[4 * i + 2] = PREMULTIPLY(rgb[4 * i + 2], mult);
}
}
rgba += stride;
}
}
#undef MULTIPLY_BY_ALPHA
#undef MULTIPLIER
#undef PREMULTIPLY
//------------------------------------------------------------------------------
static int DispatchAlpha_NEON(const uint8_t* alpha, int alpha_stride,
int width, int height,
uint8_t* dst, int dst_stride) {
uint32_t alpha_mask = 0xffffffffu;
uint8x8_t mask8 = vdup_n_u8(0xff);
uint32_t tmp[2];
int i, j;
for (j = 0; j < height; ++j) {
// We don't know if alpha is first or last in dst[] (depending on rgbA/Argb
// mode). So we must be sure dst[4*i + 8 - 1] is writable for the store.
// Hence the test with 'width - 1' instead of just 'width'.
for (i = 0; i + 8 <= width - 1; i += 8) {
uint8x8x4_t rgbX = vld4_u8((const uint8_t*)(dst + 4 * i));
const uint8x8_t alphas = vld1_u8(alpha + i);
rgbX.val[0] = alphas;
vst4_u8((uint8_t*)(dst + 4 * i), rgbX);
mask8 = vand_u8(mask8, alphas);
}
for (; i < width; ++i) {
const uint32_t alpha_value = alpha[i];
dst[4 * i] = alpha_value;
alpha_mask &= alpha_value;
}
alpha += alpha_stride;
dst += dst_stride;
}
vst1_u8((uint8_t*)tmp, mask8);
alpha_mask &= tmp[0];
alpha_mask &= tmp[1];
return (alpha_mask != 0xffffffffu);
}
static void DispatchAlphaToGreen_NEON(const uint8_t* alpha, int alpha_stride,
int width, int height,
uint32_t* dst, int dst_stride) {
int i, j;
uint8x8x4_t greens; // leave A/R/B channels zero'd.
greens.val[0] = vdup_n_u8(0);
greens.val[2] = vdup_n_u8(0);
greens.val[3] = vdup_n_u8(0);
for (j = 0; j < height; ++j) {
for (i = 0; i + 8 <= width; i += 8) {
greens.val[1] = vld1_u8(alpha + i);
vst4_u8((uint8_t*)(dst + i), greens);
}
for (; i < width; ++i) dst[i] = alpha[i] << 8;
alpha += alpha_stride;
dst += dst_stride;
}
}
static int ExtractAlpha_NEON(const uint8_t* argb, int argb_stride,
int width, int height,
uint8_t* alpha, int alpha_stride) {
uint32_t alpha_mask = 0xffffffffu;
uint8x8_t mask8 = vdup_n_u8(0xff);
uint32_t tmp[2];
int i, j;
for (j = 0; j < height; ++j) {
// We don't know if alpha is first or last in dst[] (depending on rgbA/Argb
// mode). So we must be sure dst[4*i + 8 - 1] is writable for the store.
// Hence the test with 'width - 1' instead of just 'width'.
for (i = 0; i + 8 <= width - 1; i += 8) {
const uint8x8x4_t rgbX = vld4_u8((const uint8_t*)(argb + 4 * i));
const uint8x8_t alphas = rgbX.val[0];
vst1_u8((uint8_t*)(alpha + i), alphas);
mask8 = vand_u8(mask8, alphas);
}
for (; i < width; ++i) {
alpha[i] = argb[4 * i];
alpha_mask &= alpha[i];
}
argb += argb_stride;
alpha += alpha_stride;
}
vst1_u8((uint8_t*)tmp, mask8);
alpha_mask &= tmp[0];
alpha_mask &= tmp[1];
return (alpha_mask == 0xffffffffu);
}
static void ExtractGreen_NEON(const uint32_t* argb,
uint8_t* alpha, int size) {
int i;
for (i = 0; i + 16 <= size; i += 16) {
const uint8x16x4_t rgbX = vld4q_u8((const uint8_t*)(argb + i));
const uint8x16_t greens = rgbX.val[1];
vst1q_u8(alpha + i, greens);
}
for (; i < size; ++i) alpha[i] = (argb[i] >> 8) & 0xff;
}
//------------------------------------------------------------------------------
extern void WebPInitAlphaProcessingNEON(void);
WEBP_TSAN_IGNORE_FUNCTION void WebPInitAlphaProcessingNEON(void) {
WebPApplyAlphaMultiply = ApplyAlphaMultiply_NEON;
WebPDispatchAlpha = DispatchAlpha_NEON;
WebPDispatchAlphaToGreen = DispatchAlphaToGreen_NEON;
WebPExtractAlpha = ExtractAlpha_NEON;
WebPExtractGreen = ExtractGreen_NEON;
}
#else // !WEBP_USE_NEON
WEBP_DSP_INIT_STUB(WebPInitAlphaProcessingNEON)
#endif // WEBP_USE_NEON

141
thirdparty/libwebp/dsp/alpha_processing_sse2.c vendored
View file

@ -150,46 +150,46 @@ static int ExtractAlpha(const uint8_t* argb, int argb_stride,
#define PREMULTIPLY(x, m) (((x) * (m)) >> 23)
// We can't use a 'const int' for the SHUFFLE value, because it has to be an
// immediate in the _mm_shufflexx_epi16() instruction. We really a macro here.
#define APPLY_ALPHA(RGBX, SHUFFLE, MASK, MULT) do { \
const __m128i argb0 = _mm_loadl_epi64((__m128i*)&(RGBX)); \
const __m128i argb1 = _mm_unpacklo_epi8(argb0, zero); \
const __m128i alpha0 = _mm_and_si128(argb1, MASK); \
const __m128i alpha1 = _mm_shufflelo_epi16(alpha0, SHUFFLE); \
const __m128i alpha2 = _mm_shufflehi_epi16(alpha1, SHUFFLE); \
/* alpha2 = [0 a0 a0 a0][0 a1 a1 a1] */ \
const __m128i scale0 = _mm_mullo_epi16(alpha2, MULT); \
const __m128i scale1 = _mm_mulhi_epu16(alpha2, MULT); \
const __m128i argb2 = _mm_mulhi_epu16(argb1, scale0); \
const __m128i argb3 = _mm_mullo_epi16(argb1, scale1); \
const __m128i argb4 = _mm_adds_epu16(argb2, argb3); \
const __m128i argb5 = _mm_srli_epi16(argb4, 7); \
const __m128i argb6 = _mm_or_si128(argb5, alpha0); \
const __m128i argb7 = _mm_packus_epi16(argb6, zero); \
_mm_storel_epi64((__m128i*)&(RGBX), argb7); \
// immediate in the _mm_shufflexx_epi16() instruction. We really need a macro.
// We use: v / 255 = (v * 0x8081) >> 23, where v = alpha * {r,g,b} is a 16bit
// value.
#define APPLY_ALPHA(RGBX, SHUFFLE) do { \
const __m128i argb0 = _mm_loadu_si128((const __m128i*)&(RGBX)); \
const __m128i argb1_lo = _mm_unpacklo_epi8(argb0, zero); \
const __m128i argb1_hi = _mm_unpackhi_epi8(argb0, zero); \
const __m128i alpha0_lo = _mm_or_si128(argb1_lo, kMask); \
const __m128i alpha0_hi = _mm_or_si128(argb1_hi, kMask); \
const __m128i alpha1_lo = _mm_shufflelo_epi16(alpha0_lo, SHUFFLE); \
const __m128i alpha1_hi = _mm_shufflelo_epi16(alpha0_hi, SHUFFLE); \
const __m128i alpha2_lo = _mm_shufflehi_epi16(alpha1_lo, SHUFFLE); \
const __m128i alpha2_hi = _mm_shufflehi_epi16(alpha1_hi, SHUFFLE); \
/* alpha2 = [ff a0 a0 a0][ff a1 a1 a1] */ \
const __m128i A0_lo = _mm_mullo_epi16(alpha2_lo, argb1_lo); \
const __m128i A0_hi = _mm_mullo_epi16(alpha2_hi, argb1_hi); \
const __m128i A1_lo = _mm_mulhi_epu16(A0_lo, kMult); \
const __m128i A1_hi = _mm_mulhi_epu16(A0_hi, kMult); \
const __m128i A2_lo = _mm_srli_epi16(A1_lo, 7); \
const __m128i A2_hi = _mm_srli_epi16(A1_hi, 7); \
const __m128i A3 = _mm_packus_epi16(A2_lo, A2_hi); \
_mm_storeu_si128((__m128i*)&(RGBX), A3); \
} while (0)
static void ApplyAlphaMultiply(uint8_t* rgba, int alpha_first,
int w, int h, int stride) {
static void ApplyAlphaMultiply_SSE2(uint8_t* rgba, int alpha_first,
int w, int h, int stride) {
const __m128i zero = _mm_setzero_si128();
const int kSpan = 2;
const int w2 = w & ~(kSpan - 1);
const __m128i kMult = _mm_set1_epi16(0x8081u);
const __m128i kMask = _mm_set_epi16(0, 0xff, 0xff, 0, 0, 0xff, 0xff, 0);
const int kSpan = 4;
while (h-- > 0) {
uint32_t* const rgbx = (uint32_t*)rgba;
int i;
if (!alpha_first) {
const __m128i kMask = _mm_set_epi16(0xff, 0, 0, 0, 0xff, 0, 0, 0);
const __m128i kMult =
_mm_set_epi16(0, 0x8081, 0x8081, 0x8081, 0, 0x8081, 0x8081, 0x8081);
for (i = 0; i < w2; i += kSpan) {
APPLY_ALPHA(rgbx[i], _MM_SHUFFLE(0, 3, 3, 3), kMask, kMult);
for (i = 0; i + kSpan <= w; i += kSpan) {
APPLY_ALPHA(rgbx[i], _MM_SHUFFLE(2, 3, 3, 3));
}
} else {
const __m128i kMask = _mm_set_epi16(0, 0, 0, 0xff, 0, 0, 0, 0xff);
const __m128i kMult =
_mm_set_epi16(0x8081, 0x8081, 0x8081, 0, 0x8081, 0x8081, 0x8081, 0);
for (i = 0; i < w2; i += kSpan) {
APPLY_ALPHA(rgbx[i], _MM_SHUFFLE(0, 0, 0, 3), kMask, kMult);
for (i = 0; i + kSpan <= w; i += kSpan) {
APPLY_ALPHA(rgbx[i], _MM_SHUFFLE(0, 0, 0, 1));
}
}
// Finish with left-overs.
@ -213,64 +213,51 @@ static void ApplyAlphaMultiply(uint8_t* rgba, int alpha_first,
// -----------------------------------------------------------------------------
// Apply alpha value to rows
// We use: kINV255 = (1 << 24) / 255 = 0x010101
// So: a * kINV255 = (a << 16) | [(a << 8) | a]
// -> _mm_mulhi_epu16() takes care of the (a<<16) part,
// and _mm_mullo_epu16(a * 0x0101,...) takes care of the "(a << 8) | a" one.
static void MultARGBRow(uint32_t* const ptr, int width, int inverse) {
static void MultARGBRow_SSE2(uint32_t* const ptr, int width, int inverse) {
int x = 0;
if (!inverse) {
const int kSpan = 2;
const __m128i zero = _mm_setzero_si128();
const __m128i kRound =
_mm_set_epi16(0, 1 << 7, 1 << 7, 1 << 7, 0, 1 << 7, 1 << 7, 1 << 7);
const __m128i kMult =
_mm_set_epi16(0, 0x0101, 0x0101, 0x0101, 0, 0x0101, 0x0101, 0x0101);
const __m128i kOne64 = _mm_set_epi16(1u << 8, 0, 0, 0, 1u << 8, 0, 0, 0);
const int w2 = width & ~(kSpan - 1);
for (x = 0; x < w2; x += kSpan) {
const __m128i argb0 = _mm_loadl_epi64((__m128i*)&ptr[x]);
const __m128i argb1 = _mm_unpacklo_epi8(argb0, zero);
const __m128i tmp0 = _mm_shufflelo_epi16(argb1, _MM_SHUFFLE(3, 3, 3, 3));
const __m128i tmp1 = _mm_shufflehi_epi16(tmp0, _MM_SHUFFLE(3, 3, 3, 3));
const __m128i tmp2 = _mm_srli_epi64(tmp1, 16);
const __m128i scale0 = _mm_mullo_epi16(tmp1, kMult);
const __m128i scale1 = _mm_or_si128(tmp2, kOne64);
const __m128i argb2 = _mm_mulhi_epu16(argb1, scale0);
const __m128i argb3 = _mm_mullo_epi16(argb1, scale1);
const __m128i argb4 = _mm_adds_epu16(argb2, argb3);
const __m128i argb5 = _mm_adds_epu16(argb4, kRound);
const __m128i argb6 = _mm_srli_epi16(argb5, 8);
const __m128i argb7 = _mm_packus_epi16(argb6, zero);
_mm_storel_epi64((__m128i*)&ptr[x], argb7);
const __m128i k128 = _mm_set1_epi16(128);
const __m128i kMult = _mm_set1_epi16(0x0101);
const __m128i kMask = _mm_set_epi16(0, 0xff, 0, 0, 0, 0xff, 0, 0);
for (x = 0; x + kSpan <= width; x += kSpan) {
// To compute 'result = (int)(a * x / 255. + .5)', we use:
// tmp = a * v + 128, result = (tmp * 0x0101u) >> 16
const __m128i A0 = _mm_loadl_epi64((const __m128i*)&ptr[x]);
const __m128i A1 = _mm_unpacklo_epi8(A0, zero);
const __m128i A2 = _mm_or_si128(A1, kMask);
const __m128i A3 = _mm_shufflelo_epi16(A2, _MM_SHUFFLE(2, 3, 3, 3));
const __m128i A4 = _mm_shufflehi_epi16(A3, _MM_SHUFFLE(2, 3, 3, 3));
// here, A4 = [ff a0 a0 a0][ff a1 a1 a1]
const __m128i A5 = _mm_mullo_epi16(A4, A1);
const __m128i A6 = _mm_add_epi16(A5, k128);
const __m128i A7 = _mm_mulhi_epu16(A6, kMult);
const __m128i A10 = _mm_packus_epi16(A7, zero);
_mm_storel_epi64((__m128i*)&ptr[x], A10);
}
}
width -= x;
if (width > 0) WebPMultARGBRowC(ptr + x, width, inverse);
}
static void MultRow(uint8_t* const ptr, const uint8_t* const alpha,
int width, int inverse) {
static void MultRow_SSE2(uint8_t* const ptr, const uint8_t* const alpha,
int width, int inverse) {
int x = 0;
if (!inverse) {
const int kSpan = 8;
const __m128i zero = _mm_setzero_si128();
const __m128i kRound = _mm_set1_epi16(1 << 7);
const int w2 = width & ~(kSpan - 1);
for (x = 0; x < w2; x += kSpan) {
const __m128i k128 = _mm_set1_epi16(128);
const __m128i kMult = _mm_set1_epi16(0x0101);
for (x = 0; x + 8 <= width; x += 8) {
const __m128i v0 = _mm_loadl_epi64((__m128i*)&ptr[x]);
const __m128i a0 = _mm_loadl_epi64((const __m128i*)&alpha[x]);
const __m128i v1 = _mm_unpacklo_epi8(v0, zero);
const __m128i alpha0 = _mm_loadl_epi64((const __m128i*)&alpha[x]);
const __m128i alpha1 = _mm_unpacklo_epi8(alpha0, zero);
const __m128i alpha2 = _mm_unpacklo_epi8(alpha0, alpha0);
const __m128i v2 = _mm_mulhi_epu16(v1, alpha2);
const __m128i v3 = _mm_mullo_epi16(v1, alpha1);
const __m128i v4 = _mm_adds_epu16(v2, v3);
const __m128i v5 = _mm_adds_epu16(v4, kRound);
const __m128i v6 = _mm_srli_epi16(v5, 8);
const __m128i v7 = _mm_packus_epi16(v6, zero);
_mm_storel_epi64((__m128i*)&ptr[x], v7);
const __m128i a1 = _mm_unpacklo_epi8(a0, zero);
const __m128i v2 = _mm_mullo_epi16(v1, a1);
const __m128i v3 = _mm_add_epi16(v2, k128);
const __m128i v4 = _mm_mulhi_epu16(v3, kMult);
const __m128i v5 = _mm_packus_epi16(v4, zero);
_mm_storel_epi64((__m128i*)&ptr[x], v5);
}
}
width -= x;
@ -283,9 +270,9 @@ static void MultRow(uint8_t* const ptr, const uint8_t* const alpha,
extern void WebPInitAlphaProcessingSSE2(void);
WEBP_TSAN_IGNORE_FUNCTION void WebPInitAlphaProcessingSSE2(void) {
WebPMultARGBRow = MultARGBRow;
WebPMultRow = MultRow;
WebPApplyAlphaMultiply = ApplyAlphaMultiply;
WebPMultARGBRow = MultARGBRow_SSE2;
WebPMultRow = MultRow_SSE2;
WebPApplyAlphaMultiply = ApplyAlphaMultiply_SSE2;
WebPDispatchAlpha = DispatchAlpha;
WebPDispatchAlphaToGreen = DispatchAlphaToGreen;
WebPExtractAlpha = ExtractAlpha;

85
thirdparty/libwebp/dsp/common_sse2.h vendored
View file

@ -100,6 +100,91 @@ static WEBP_INLINE void VP8Transpose_2_4x4_16b(
// a03 a13 a23 a33 b03 b13 b23 b33
}
//------------------------------------------------------------------------------
// Channel mixing.
// Function used several times in VP8PlanarTo24b.
// It samples the in buffer as follows: one every two unsigned char is stored
// at the beginning of the buffer, while the other half is stored at the end.
#define VP8PlanarTo24bHelper(IN, OUT) \
do { \
const __m128i v_mask = _mm_set1_epi16(0x00ff); \
/* Take one every two upper 8b values.*/ \
(OUT##0) = _mm_packus_epi16(_mm_and_si128((IN##0), v_mask), \
_mm_and_si128((IN##1), v_mask)); \
(OUT##1) = _mm_packus_epi16(_mm_and_si128((IN##2), v_mask), \
_mm_and_si128((IN##3), v_mask)); \
(OUT##2) = _mm_packus_epi16(_mm_and_si128((IN##4), v_mask), \
_mm_and_si128((IN##5), v_mask)); \
/* Take one every two lower 8b values.*/ \
(OUT##3) = _mm_packus_epi16(_mm_srli_epi16((IN##0), 8), \
_mm_srli_epi16((IN##1), 8)); \
(OUT##4) = _mm_packus_epi16(_mm_srli_epi16((IN##2), 8), \
_mm_srli_epi16((IN##3), 8)); \
(OUT##5) = _mm_packus_epi16(_mm_srli_epi16((IN##4), 8), \
_mm_srli_epi16((IN##5), 8)); \
} while (0)
// Pack the planar buffers
// rrrr... rrrr... gggg... gggg... bbbb... bbbb....
// triplet by triplet in the output buffer rgb as rgbrgbrgbrgb ...
static WEBP_INLINE void VP8PlanarTo24b(__m128i* const in0, __m128i* const in1,
__m128i* const in2, __m128i* const in3,
__m128i* const in4, __m128i* const in5) {
// The input is 6 registers of sixteen 8b but for the sake of explanation,
// let's take 6 registers of four 8b values.
// To pack, we will keep taking one every two 8b integer and move it
// around as follows:
// Input:
// r0r1r2r3 | r4r5r6r7 | g0g1g2g3 | g4g5g6g7 | b0b1b2b3 | b4b5b6b7
// Split the 6 registers in two sets of 3 registers: the first set as the even
// 8b bytes, the second the odd ones:
// r0r2r4r6 | g0g2g4g6 | b0b2b4b6 | r1r3r5r7 | g1g3g5g7 | b1b3b5b7
// Repeat the same permutations twice more:
// r0r4g0g4 | b0b4r1r5 | g1g5b1b5 | r2r6g2g6 | b2b6r3r7 | g3g7b3b7
// r0g0b0r1 | g1b1r2g2 | b2r3g3b3 | r4g4b4r5 | g5b5r6g6 | b6r7g7b7
__m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5;
VP8PlanarTo24bHelper(*in, tmp);
VP8PlanarTo24bHelper(tmp, *in);
VP8PlanarTo24bHelper(*in, tmp);
// We need to do it two more times than the example as we have sixteen bytes.
{
__m128i out0, out1, out2, out3, out4, out5;
VP8PlanarTo24bHelper(tmp, out);
VP8PlanarTo24bHelper(out, *in);
}
}
#undef VP8PlanarTo24bHelper
// Convert four packed four-channel buffers like argbargbargbargb... into the
// split channels aaaaa ... rrrr ... gggg .... bbbbb ......
static WEBP_INLINE void VP8L32bToPlanar(__m128i* const in0,
__m128i* const in1,
__m128i* const in2,
__m128i* const in3) {
// Column-wise transpose.
const __m128i A0 = _mm_unpacklo_epi8(*in0, *in1);
const __m128i A1 = _mm_unpackhi_epi8(*in0, *in1);
const __m128i A2 = _mm_unpacklo_epi8(*in2, *in3);
const __m128i A3 = _mm_unpackhi_epi8(*in2, *in3);
const __m128i B0 = _mm_unpacklo_epi8(A0, A1);
const __m128i B1 = _mm_unpackhi_epi8(A0, A1);
const __m128i B2 = _mm_unpacklo_epi8(A2, A3);
const __m128i B3 = _mm_unpackhi_epi8(A2, A3);
// C0 = g7 g6 ... g1 g0 | b7 b6 ... b1 b0
// C1 = a7 a6 ... a1 a0 | r7 r6 ... r1 r0
const __m128i C0 = _mm_unpacklo_epi8(B0, B1);
const __m128i C1 = _mm_unpackhi_epi8(B0, B1);
const __m128i C2 = _mm_unpacklo_epi8(B2, B3);
const __m128i C3 = _mm_unpackhi_epi8(B2, B3);
// Gather the channels.
*in0 = _mm_unpackhi_epi64(C1, C3);
*in1 = _mm_unpacklo_epi64(C1, C3);
*in2 = _mm_unpackhi_epi64(C0, C2);
*in3 = _mm_unpacklo_epi64(C0, C2);
}
#endif // WEBP_USE_SSE2
#ifdef __cplusplus

2
thirdparty/libwebp/dsp/cost.c vendored
View file

@ -10,7 +10,7 @@
// Author: Skal (pascal.massimino@gmail.com)
#include "./dsp.h"
#include "../enc/cost.h"
#include "../enc/cost_enc.h"
//------------------------------------------------------------------------------
// Boolean-cost cost table

2
thirdparty/libwebp/dsp/cost_mips32.c vendored
View file

@ -13,7 +13,7 @@
#if defined(WEBP_USE_MIPS32)
#include "../enc/cost.h"
#include "../enc/cost_enc.h"
static int GetResidualCost(int ctx0, const VP8Residual* const res) {
int temp0, temp1;

2
thirdparty/libwebp/dsp/cost_mips_dsp_r2.c vendored
View file

@ -13,7 +13,7 @@
#if defined(WEBP_USE_MIPS_DSP_R2)
#include "../enc/cost.h"
#include "../enc/cost_enc.h"
static int GetResidualCost(int ctx0, const VP8Residual* const res) {
int temp0, temp1;

4
thirdparty/libwebp/dsp/cost_sse2.c vendored
View file

@ -16,8 +16,8 @@
#if defined(WEBP_USE_SSE2)
#include <emmintrin.h>
#include "../enc/cost.h"
#include "../enc/vp8enci.h"
#include "../enc/cost_enc.h"
#include "../enc/vp8i_enc.h"
#include "../utils/utils.h"
//------------------------------------------------------------------------------

45
thirdparty/libwebp/dsp/cpu.c vendored
View file

@ -95,26 +95,62 @@ static WEBP_INLINE uint64_t xgetbv(void) {
#endif
#if defined(__i386__) || defined(__x86_64__) || defined(WEBP_MSC_SSE2)
// helper function for run-time detection of slow SSSE3 platforms
static int CheckSlowModel(int info) {
// Table listing display models with longer latencies for the bsr instruction
// (ie 2 cycles vs 10/16 cycles) and some SSSE3 instructions like pshufb.
// Refer to Intel 64 and IA-32 Architectures Optimization Reference Manual.
static const uint8_t kSlowModels[] = {
0x37, 0x4a, 0x4d, // Silvermont Microarchitecture
0x1c, 0x26, 0x27 // Atom Microarchitecture
};
const uint32_t model = ((info & 0xf0000) >> 12) | ((info >> 4) & 0xf);
const uint32_t family = (info >> 8) & 0xf;
if (family == 0x06) {
size_t i;
for (i = 0; i < sizeof(kSlowModels) / sizeof(kSlowModels[0]); ++i) {
if (model == kSlowModels[i]) return 1;
}
}
return 0;
}
static int x86CPUInfo(CPUFeature feature) {
int max_cpuid_value;
int cpu_info[4];
int is_intel = 0;
// get the highest feature value cpuid supports
GetCPUInfo(cpu_info, 0);
max_cpuid_value = cpu_info[0];
if (max_cpuid_value < 1) {
return 0;
} else {
const int VENDOR_ID_INTEL_EBX = 0x756e6547; // uneG
const int VENDOR_ID_INTEL_EDX = 0x49656e69; // Ieni
const int VENDOR_ID_INTEL_ECX = 0x6c65746e; // letn
is_intel = (cpu_info[1] == VENDOR_ID_INTEL_EBX &&
cpu_info[2] == VENDOR_ID_INTEL_ECX &&
cpu_info[3] == VENDOR_ID_INTEL_EDX); // genuine Intel?
}
GetCPUInfo(cpu_info, 1);
if (feature == kSSE2) {
return 0 != (cpu_info[3] & 0x04000000);
return !!(cpu_info[3] & (1 << 26));
}
if (feature == kSSE3) {
return 0 != (cpu_info[2] & 0x00000001);
return !!(cpu_info[2] & (1 << 0));
}
if (feature == kSlowSSSE3) {
if (is_intel && (cpu_info[2] & (1 << 0))) { // SSSE3?
return CheckSlowModel(cpu_info[0]);
}
return 0;
}
if (feature == kSSE4_1) {
return 0 != (cpu_info[2] & 0x00080000);
return !!(cpu_info[2] & (1 << 19));
}
if (feature == kAVX) {
// bits 27 (OSXSAVE) & 28 (256-bit AVX)
@ -126,7 +162,7 @@ static int x86CPUInfo(CPUFeature feature) {
if (feature == kAVX2) {
if (x86CPUInfo(kAVX) && max_cpuid_value >= 7) {
GetCPUInfo(cpu_info, 7);
return ((cpu_info[1] & 0x00000020) == 0x00000020);
return !!(cpu_info[1] & (1 << 5));
}
}
return 0;
@ -184,4 +220,3 @@ VP8CPUInfo VP8GetCPUInfo = mipsCPUInfo;
#else
VP8CPUInfo VP8GetCPUInfo = NULL;
#endif

2
thirdparty/libwebp/dsp/dec.c vendored
View file

@ -12,7 +12,7 @@
// Author: Skal (pascal.massimino@gmail.com)
#include "./dsp.h"
#include "../dec/vp8i.h"
#include "../dec/vp8i_dec.h"
#include "../utils/utils.h"
//------------------------------------------------------------------------------

8
thirdparty/libwebp/dsp/dec_clip_tables.c vendored
View file

@ -63,7 +63,7 @@ static const uint8_t abs0[255 + 255 + 1] = {
0xf9, 0xfa, 0xfb, 0xfc, 0xfd, 0xfe, 0xff
};
static const int8_t sclip1[1020 + 1020 + 1] = {
static const uint8_t sclip1[1020 + 1020 + 1] = {
0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
@ -236,7 +236,7 @@ static const int8_t sclip1[1020 + 1020 + 1] = {
0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f
};
static const int8_t sclip2[112 + 112 + 1] = {
static const uint8_t sclip2[112 + 112 + 1] = {
0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0,
0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0,
0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0,
@ -339,8 +339,8 @@ static volatile int tables_ok = 0;
#endif
const int8_t* const VP8ksclip1 = &sclip1[1020];
const int8_t* const VP8ksclip2 = &sclip2[112];
const int8_t* const VP8ksclip1 = (const int8_t*)&sclip1[1020];
const int8_t* const VP8ksclip2 = (const int8_t*)&sclip2[112];
const uint8_t* const VP8kclip1 = &clip1[255];
const uint8_t* const VP8kabs0 = &abs0[255];

847
thirdparty/libwebp/dsp/dec_msa.c vendored
View file

@ -153,6 +153,820 @@ static void TransformAC3(const int16_t* in, uint8_t* dst) {
ST4x4_UB(res0, res0, 3, 2, 1, 0, dst, BPS);
}
//------------------------------------------------------------------------------
// Edge filtering functions
#define FLIP_SIGN2(in0, in1, out0, out1) { \
out0 = (v16i8)__msa_xori_b(in0, 0x80); \
out1 = (v16i8)__msa_xori_b(in1, 0x80); \
}
#define FLIP_SIGN4(in0, in1, in2, in3, out0, out1, out2, out3) { \
FLIP_SIGN2(in0, in1, out0, out1); \
FLIP_SIGN2(in2, in3, out2, out3); \
}
#define FILT_VAL(q0_m, p0_m, mask, filt) do { \
v16i8 q0_sub_p0; \
q0_sub_p0 = __msa_subs_s_b(q0_m, p0_m); \
filt = __msa_adds_s_b(filt, q0_sub_p0); \
filt = __msa_adds_s_b(filt, q0_sub_p0); \
filt = __msa_adds_s_b(filt, q0_sub_p0); \
filt = filt & mask; \
} while (0)
#define FILT2(q_m, p_m, q, p) do { \
u_r = SRAI_H(temp1, 7); \
u_r = __msa_sat_s_h(u_r, 7); \
u_l = SRAI_H(temp3, 7); \
u_l = __msa_sat_s_h(u_l, 7); \
u = __msa_pckev_b((v16i8)u_l, (v16i8)u_r); \
q_m = __msa_subs_s_b(q_m, u); \
p_m = __msa_adds_s_b(p_m, u); \
q = __msa_xori_b((v16u8)q_m, 0x80); \
p = __msa_xori_b((v16u8)p_m, 0x80); \
} while (0)
#define LPF_FILTER4_4W(p1, p0, q0, q1, mask, hev) do { \
v16i8 p1_m, p0_m, q0_m, q1_m; \
v16i8 filt, t1, t2; \
const v16i8 cnst4b = __msa_ldi_b(4); \
const v16i8 cnst3b = __msa_ldi_b(3); \
\
FLIP_SIGN4(p1, p0, q0, q1, p1_m, p0_m, q0_m, q1_m); \
filt = __msa_subs_s_b(p1_m, q1_m); \
filt = filt & hev; \
FILT_VAL(q0_m, p0_m, mask, filt); \
t1 = __msa_adds_s_b(filt, cnst4b); \
t1 = SRAI_B(t1, 3); \
t2 = __msa_adds_s_b(filt, cnst3b); \
t2 = SRAI_B(t2, 3); \
q0_m = __msa_subs_s_b(q0_m, t1); \
q0 = __msa_xori_b((v16u8)q0_m, 0x80); \
p0_m = __msa_adds_s_b(p0_m, t2); \
p0 = __msa_xori_b((v16u8)p0_m, 0x80); \
filt = __msa_srari_b(t1, 1); \
hev = __msa_xori_b(hev, 0xff); \
filt = filt & hev; \
q1_m = __msa_subs_s_b(q1_m, filt); \
q1 = __msa_xori_b((v16u8)q1_m, 0x80); \
p1_m = __msa_adds_s_b(p1_m, filt); \
p1 = __msa_xori_b((v16u8)p1_m, 0x80); \
} while (0)
#define LPF_MBFILTER(p2, p1, p0, q0, q1, q2, mask, hev) do { \
v16i8 p2_m, p1_m, p0_m, q2_m, q1_m, q0_m; \
v16i8 u, filt, t1, t2, filt_sign; \
v8i16 filt_r, filt_l, u_r, u_l; \
v8i16 temp0, temp1, temp2, temp3; \
const v16i8 cnst4b = __msa_ldi_b(4); \
const v16i8 cnst3b = __msa_ldi_b(3); \
const v8i16 cnst9h = __msa_ldi_h(9); \
\
FLIP_SIGN4(p1, p0, q0, q1, p1_m, p0_m, q0_m, q1_m); \
filt = __msa_subs_s_b(p1_m, q1_m); \
FILT_VAL(q0_m, p0_m, mask, filt); \
FLIP_SIGN2(p2, q2, p2_m, q2_m); \
t2 = filt & hev; \
/* filt_val &= ~hev */ \
hev = __msa_xori_b(hev, 0xff); \
filt = filt & hev; \
t1 = __msa_adds_s_b(t2, cnst4b); \
t1 = SRAI_B(t1, 3); \
t2 = __msa_adds_s_b(t2, cnst3b); \
t2 = SRAI_B(t2, 3); \
q0_m = __msa_subs_s_b(q0_m, t1); \
p0_m = __msa_adds_s_b(p0_m, t2); \
filt_sign = __msa_clti_s_b(filt, 0); \
ILVRL_B2_SH(filt_sign, filt, filt_r, filt_l); \
/* update q2/p2 */ \
temp0 = filt_r * cnst9h; \
temp1 = ADDVI_H(temp0, 63); \
temp2 = filt_l * cnst9h; \
temp3 = ADDVI_H(temp2, 63); \
FILT2(q2_m, p2_m, q2, p2); \
/* update q1/p1 */ \
temp1 = temp1 + temp0; \
temp3 = temp3 + temp2; \
FILT2(q1_m, p1_m, q1, p1); \
/* update q0/p0 */ \
temp1 = temp1 + temp0; \
temp3 = temp3 + temp2; \
FILT2(q0_m, p0_m, q0, p0); \
} while (0)
#define LPF_MASK_HEV(p3_in, p2_in, p1_in, p0_in, \
q0_in, q1_in, q2_in, q3_in, \
limit_in, b_limit_in, thresh_in, \
hev_out, mask_out) do { \
v16u8 p3_asub_p2_m, p2_asub_p1_m, p1_asub_p0_m, q1_asub_q0_m; \
v16u8 p1_asub_q1_m, p0_asub_q0_m, q3_asub_q2_m, q2_asub_q1_m; \
v16u8 flat_out; \
\
/* absolute subtraction of pixel values */ \
p3_asub_p2_m = __msa_asub_u_b(p3_in, p2_in); \
p2_asub_p1_m = __msa_asub_u_b(p2_in, p1_in); \
p1_asub_p0_m = __msa_asub_u_b(p1_in, p0_in); \
q1_asub_q0_m = __msa_asub_u_b(q1_in, q0_in); \
q2_asub_q1_m = __msa_asub_u_b(q2_in, q1_in); \
q3_asub_q2_m = __msa_asub_u_b(q3_in, q2_in); \
p0_asub_q0_m = __msa_asub_u_b(p0_in, q0_in); \
p1_asub_q1_m = __msa_asub_u_b(p1_in, q1_in); \
/* calculation of hev */ \
flat_out = __msa_max_u_b(p1_asub_p0_m, q1_asub_q0_m); \
hev_out = (thresh_in < flat_out); \
/* calculation of mask */ \
p0_asub_q0_m = __msa_adds_u_b(p0_asub_q0_m, p0_asub_q0_m); \
p1_asub_q1_m = SRAI_B(p1_asub_q1_m, 1); \
p0_asub_q0_m = __msa_adds_u_b(p0_asub_q0_m, p1_asub_q1_m); \
mask_out = (b_limit_in < p0_asub_q0_m); \
mask_out = __msa_max_u_b(flat_out, mask_out); \
p3_asub_p2_m = __msa_max_u_b(p3_asub_p2_m, p2_asub_p1_m); \
mask_out = __msa_max_u_b(p3_asub_p2_m, mask_out); \
q2_asub_q1_m = __msa_max_u_b(q2_asub_q1_m, q3_asub_q2_m); \
mask_out = __msa_max_u_b(q2_asub_q1_m, mask_out); \
mask_out = (limit_in < mask_out); \
mask_out = __msa_xori_b(mask_out, 0xff); \
} while (0)
#define ST6x1_UB(in0, in0_idx, in1, in1_idx, pdst, stride) do { \
const uint16_t tmp0_h = __msa_copy_s_h((v8i16)in1, in1_idx); \
const uint32_t tmp0_w = __msa_copy_s_w((v4i32)in0, in0_idx); \
SW(tmp0_w, pdst); \
SH(tmp0_h, pdst + stride); \
} while (0)
#define ST6x4_UB(in0, start_in0_idx, in1, start_in1_idx, pdst, stride) do { \
uint8_t* ptmp1 = (uint8_t*)pdst; \
ST6x1_UB(in0, start_in0_idx, in1, start_in1_idx, ptmp1, 4); \
ptmp1 += stride; \
ST6x1_UB(in0, start_in0_idx + 1, in1, start_in1_idx + 1, ptmp1, 4); \
ptmp1 += stride; \
ST6x1_UB(in0, start_in0_idx + 2, in1, start_in1_idx + 2, ptmp1, 4); \
ptmp1 += stride; \
ST6x1_UB(in0, start_in0_idx + 3, in1, start_in1_idx + 3, ptmp1, 4); \
} while (0)
#define LPF_SIMPLE_FILT(p1_in, p0_in, q0_in, q1_in, mask) do { \
v16i8 p1_m, p0_m, q0_m, q1_m, filt, filt1, filt2; \
const v16i8 cnst4b = __msa_ldi_b(4); \
const v16i8 cnst3b = __msa_ldi_b(3); \
\
FLIP_SIGN4(p1_in, p0_in, q0_in, q1_in, p1_m, p0_m, q0_m, q1_m); \
filt = __msa_subs_s_b(p1_m, q1_m); \
FILT_VAL(q0_m, p0_m, mask, filt); \
filt1 = __msa_adds_s_b(filt, cnst4b); \
filt1 = SRAI_B(filt1, 3); \
filt2 = __msa_adds_s_b(filt, cnst3b); \
filt2 = SRAI_B(filt2, 3); \
q0_m = __msa_subs_s_b(q0_m, filt1); \
p0_m = __msa_adds_s_b(p0_m, filt2); \
q0_in = __msa_xori_b((v16u8)q0_m, 0x80); \
p0_in = __msa_xori_b((v16u8)p0_m, 0x80); \
} while (0)
#define LPF_SIMPLE_MASK(p1, p0, q0, q1, b_limit, mask) do { \
v16u8 p1_a_sub_q1, p0_a_sub_q0; \
\
p0_a_sub_q0 = __msa_asub_u_b(p0, q0); \
p1_a_sub_q1 = __msa_asub_u_b(p1, q1); \
p1_a_sub_q1 = (v16u8)__msa_srli_b((v16i8)p1_a_sub_q1, 1); \
p0_a_sub_q0 = __msa_adds_u_b(p0_a_sub_q0, p0_a_sub_q0); \
mask = __msa_adds_u_b(p0_a_sub_q0, p1_a_sub_q1); \
mask = (mask <= b_limit); \
} while (0)
static void VFilter16(uint8_t* src, int stride,
int b_limit_in, int limit_in, int thresh_in) {
uint8_t* ptemp = src - 4 * stride;
v16u8 p3, p2, p1, p0, q3, q2, q1, q0;
v16u8 mask, hev;
const v16u8 thresh = (v16u8)__msa_fill_b(thresh_in);
const v16u8 limit = (v16u8)__msa_fill_b(limit_in);
const v16u8 b_limit = (v16u8)__msa_fill_b(b_limit_in);
LD_UB8(ptemp, stride, p3, p2, p1, p0, q0, q1, q2, q3);
LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit, b_limit, thresh,
hev, mask);
LPF_MBFILTER(p2, p1, p0, q0, q1, q2, mask, hev);
ptemp = src - 3 * stride;
ST_UB4(p2, p1, p0, q0, ptemp, stride);
ptemp += (4 * stride);
ST_UB2(q1, q2, ptemp, stride);
}
static void HFilter16(uint8_t* src, int stride,
int b_limit_in, int limit_in, int thresh_in) {
uint8_t* ptmp = src - 4;
v16u8 p3, p2, p1, p0, q3, q2, q1, q0;
v16u8 mask, hev;
v16u8 row0, row1, row2, row3, row4, row5, row6, row7, row8;
v16u8 row9, row10, row11, row12, row13, row14, row15;
v8i16 tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
const v16u8 b_limit = (v16u8)__msa_fill_b(b_limit_in);
const v16u8 limit = (v16u8)__msa_fill_b(limit_in);
const v16u8 thresh = (v16u8)__msa_fill_b(thresh_in);
LD_UB8(ptmp, stride, row0, row1, row2, row3, row4, row5, row6, row7);
ptmp += (8 * stride);
LD_UB8(ptmp, stride, row8, row9, row10, row11, row12, row13, row14, row15);
TRANSPOSE16x8_UB_UB(row0, row1, row2, row3, row4, row5, row6, row7,
row8, row9, row10, row11, row12, row13, row14, row15,
p3, p2, p1, p0, q0, q1, q2, q3);
LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit, b_limit, thresh,
hev, mask);
LPF_MBFILTER(p2, p1, p0, q0, q1, q2, mask, hev);
ILVR_B2_SH(p1, p2, q0, p0, tmp0, tmp1);
ILVRL_H2_SH(tmp1, tmp0, tmp3, tmp4);
ILVL_B2_SH(p1, p2, q0, p0, tmp0, tmp1);
ILVRL_H2_SH(tmp1, tmp0, tmp6, tmp7);
ILVRL_B2_SH(q2, q1, tmp2, tmp5);
ptmp = src - 3;
ST6x1_UB(tmp3, 0, tmp2, 0, ptmp, 4);
ptmp += stride;
ST6x1_UB(tmp3, 1, tmp2, 1, ptmp, 4);
ptmp += stride;
ST6x1_UB(tmp3, 2, tmp2, 2, ptmp, 4);
ptmp += stride;
ST6x1_UB(tmp3, 3, tmp2, 3, ptmp, 4);
ptmp += stride;
ST6x1_UB(tmp4, 0, tmp2, 4, ptmp, 4);
ptmp += stride;
ST6x1_UB(tmp4, 1, tmp2, 5, ptmp, 4);
ptmp += stride;
ST6x1_UB(tmp4, 2, tmp2, 6, ptmp, 4);
ptmp += stride;
ST6x1_UB(tmp4, 3, tmp2, 7, ptmp, 4);
ptmp += stride;
ST6x1_UB(tmp6, 0, tmp5, 0, ptmp, 4);
ptmp += stride;
ST6x1_UB(tmp6, 1, tmp5, 1, ptmp, 4);
ptmp += stride;
ST6x1_UB(tmp6, 2, tmp5, 2, ptmp, 4);
ptmp += stride;
ST6x1_UB(tmp6, 3, tmp5, 3, ptmp, 4);
ptmp += stride;
ST6x1_UB(tmp7, 0, tmp5, 4, ptmp, 4);
ptmp += stride;
ST6x1_UB(tmp7, 1, tmp5, 5, ptmp, 4);
ptmp += stride;
ST6x1_UB(tmp7, 2, tmp5, 6, ptmp, 4);
ptmp += stride;
ST6x1_UB(tmp7, 3, tmp5, 7, ptmp, 4);
}
// on three inner edges
static void VFilterHorEdge16i(uint8_t* src, int stride,
int b_limit, int limit, int thresh) {
v16u8 mask, hev;
v16u8 p3, p2, p1, p0, q3, q2, q1, q0;
const v16u8 thresh0 = (v16u8)__msa_fill_b(thresh);
const v16u8 b_limit0 = (v16u8)__msa_fill_b(b_limit);
const v16u8 limit0 = (v16u8)__msa_fill_b(limit);
LD_UB8((src - 4 * stride), stride, p3, p2, p1, p0, q0, q1, q2, q3);
LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit0, b_limit0, thresh0,
hev, mask);
LPF_FILTER4_4W(p1, p0, q0, q1, mask, hev);
ST_UB4(p1, p0, q0, q1, (src - 2 * stride), stride);
}
static void VFilter16i(uint8_t* src_y, int stride,
int b_limit, int limit, int thresh) {
VFilterHorEdge16i(src_y + 4 * stride, stride, b_limit, limit, thresh);
VFilterHorEdge16i(src_y + 8 * stride, stride, b_limit, limit, thresh);
VFilterHorEdge16i(src_y + 12 * stride, stride, b_limit, limit, thresh);
}
static void HFilterVertEdge16i(uint8_t* src, int stride,
int b_limit, int limit, int thresh) {
v16u8 mask, hev;
v16u8 p3, p2, p1, p0, q3, q2, q1, q0;
v16u8 row0, row1, row2, row3, row4, row5, row6, row7;
v16u8 row8, row9, row10, row11, row12, row13, row14, row15;
v8i16 tmp0, tmp1, tmp2, tmp3, tmp4, tmp5;
const v16u8 thresh0 = (v16u8)__msa_fill_b(thresh);
const v16u8 b_limit0 = (v16u8)__msa_fill_b(b_limit);
const v16u8 limit0 = (v16u8)__msa_fill_b(limit);
LD_UB8(src - 4, stride, row0, row1, row2, row3, row4, row5, row6, row7);
LD_UB8(src - 4 + (8 * stride), stride,
row8, row9, row10, row11, row12, row13, row14, row15);
TRANSPOSE16x8_UB_UB(row0, row1, row2, row3, row4, row5, row6, row7,
row8, row9, row10, row11, row12, row13, row14, row15,
p3, p2, p1, p0, q0, q1, q2, q3);
LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit0, b_limit0, thresh0,
hev, mask);
LPF_FILTER4_4W(p1, p0, q0, q1, mask, hev);
ILVR_B2_SH(p0, p1, q1, q0, tmp0, tmp1);
ILVRL_H2_SH(tmp1, tmp0, tmp2, tmp3);
ILVL_B2_SH(p0, p1, q1, q0, tmp0, tmp1);
ILVRL_H2_SH(tmp1, tmp0, tmp4, tmp5);
src -= 2;
ST4x8_UB(tmp2, tmp3, src, stride);
src += (8 * stride);
ST4x8_UB(tmp4, tmp5, src, stride);
}
static void HFilter16i(uint8_t* src_y, int stride,
int b_limit, int limit, int thresh) {
HFilterVertEdge16i(src_y + 4, stride, b_limit, limit, thresh);
HFilterVertEdge16i(src_y + 8, stride, b_limit, limit, thresh);
HFilterVertEdge16i(src_y + 12, stride, b_limit, limit, thresh);
}
// 8-pixels wide variants, for chroma filtering
static void VFilter8(uint8_t* src_u, uint8_t* src_v, int stride,
int b_limit_in, int limit_in, int thresh_in) {
uint8_t* ptmp_src_u = src_u - 4 * stride;
uint8_t* ptmp_src_v = src_v - 4 * stride;
uint64_t p2_d, p1_d, p0_d, q0_d, q1_d, q2_d;
v16u8 p3, p2, p1, p0, q3, q2, q1, q0, mask, hev;
v16u8 p3_u, p2_u, p1_u, p0_u, q3_u, q2_u, q1_u, q0_u;
v16u8 p3_v, p2_v, p1_v, p0_v, q3_v, q2_v, q1_v, q0_v;
const v16u8 b_limit = (v16u8)__msa_fill_b(b_limit_in);
const v16u8 limit = (v16u8)__msa_fill_b(limit_in);
const v16u8 thresh = (v16u8)__msa_fill_b(thresh_in);
LD_UB8(ptmp_src_u, stride, p3_u, p2_u, p1_u, p0_u, q0_u, q1_u, q2_u, q3_u);
LD_UB8(ptmp_src_v, stride, p3_v, p2_v, p1_v, p0_v, q0_v, q1_v, q2_v, q3_v);
ILVR_D4_UB(p3_v, p3_u, p2_v, p2_u, p1_v, p1_u, p0_v, p0_u, p3, p2, p1, p0);
ILVR_D4_UB(q0_v, q0_u, q1_v, q1_u, q2_v, q2_u, q3_v, q3_u, q0, q1, q2, q3);
LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit, b_limit, thresh,
hev, mask);
LPF_MBFILTER(p2, p1, p0, q0, q1, q2, mask, hev);
p2_d = __msa_copy_s_d((v2i64)p2, 0);
p1_d = __msa_copy_s_d((v2i64)p1, 0);
p0_d = __msa_copy_s_d((v2i64)p0, 0);
q0_d = __msa_copy_s_d((v2i64)q0, 0);
q1_d = __msa_copy_s_d((v2i64)q1, 0);
q2_d = __msa_copy_s_d((v2i64)q2, 0);
ptmp_src_u += stride;
SD4(p2_d, p1_d, p0_d, q0_d, ptmp_src_u, stride);
ptmp_src_u += (4 * stride);
SD(q1_d, ptmp_src_u);
ptmp_src_u += stride;
SD(q2_d, ptmp_src_u);
p2_d = __msa_copy_s_d((v2i64)p2, 1);
p1_d = __msa_copy_s_d((v2i64)p1, 1);
p0_d = __msa_copy_s_d((v2i64)p0, 1);
q0_d = __msa_copy_s_d((v2i64)q0, 1);
q1_d = __msa_copy_s_d((v2i64)q1, 1);
q2_d = __msa_copy_s_d((v2i64)q2, 1);
ptmp_src_v += stride;
SD4(p2_d, p1_d, p0_d, q0_d, ptmp_src_v, stride);
ptmp_src_v += (4 * stride);
SD(q1_d, ptmp_src_v);
ptmp_src_v += stride;
SD(q2_d, ptmp_src_v);
}
static void HFilter8(uint8_t* src_u, uint8_t* src_v, int stride,
int b_limit_in, int limit_in, int thresh_in) {
uint8_t* ptmp_src_u = src_u - 4;
uint8_t* ptmp_src_v = src_v - 4;
v16u8 p3, p2, p1, p0, q3, q2, q1, q0, mask, hev;
v16u8 row0, row1, row2, row3, row4, row5, row6, row7, row8;
v16u8 row9, row10, row11, row12, row13, row14, row15;
v8i16 tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
const v16u8 b_limit = (v16u8)__msa_fill_b(b_limit_in);
const v16u8 limit = (v16u8)__msa_fill_b(limit_in);
const v16u8 thresh = (v16u8)__msa_fill_b(thresh_in);
LD_UB8(ptmp_src_u, stride, row0, row1, row2, row3, row4, row5, row6, row7);
LD_UB8(ptmp_src_v, stride,
row8, row9, row10, row11, row12, row13, row14, row15);
TRANSPOSE16x8_UB_UB(row0, row1, row2, row3, row4, row5, row6, row7,
row8, row9, row10, row11, row12, row13, row14, row15,
p3, p2, p1, p0, q0, q1, q2, q3);
LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit, b_limit, thresh,
hev, mask);
LPF_MBFILTER(p2, p1, p0, q0, q1, q2, mask, hev);
ILVR_B2_SH(p1, p2, q0, p0, tmp0, tmp1);
ILVRL_H2_SH(tmp1, tmp0, tmp3, tmp4);
ILVL_B2_SH(p1, p2, q0, p0, tmp0, tmp1);
ILVRL_H2_SH(tmp1, tmp0, tmp6, tmp7);
ILVRL_B2_SH(q2, q1, tmp2, tmp5);
ptmp_src_u += 1;
ST6x4_UB(tmp3, 0, tmp2, 0, ptmp_src_u, stride);
ptmp_src_u += 4 * stride;
ST6x4_UB(tmp4, 0, tmp2, 4, ptmp_src_u, stride);
ptmp_src_v += 1;
ST6x4_UB(tmp6, 0, tmp5, 0, ptmp_src_v, stride);
ptmp_src_v += 4 * stride;
ST6x4_UB(tmp7, 0, tmp5, 4, ptmp_src_v, stride);
}
static void VFilter8i(uint8_t* src_u, uint8_t* src_v, int stride,
int b_limit_in, int limit_in, int thresh_in) {
uint64_t p1_d, p0_d, q0_d, q1_d;
v16u8 p3, p2, p1, p0, q3, q2, q1, q0, mask, hev;
v16u8 p3_u, p2_u, p1_u, p0_u, q3_u, q2_u, q1_u, q0_u;
v16u8 p3_v, p2_v, p1_v, p0_v, q3_v, q2_v, q1_v, q0_v;
const v16u8 thresh = (v16u8)__msa_fill_b(thresh_in);
const v16u8 limit = (v16u8)__msa_fill_b(limit_in);
const v16u8 b_limit = (v16u8)__msa_fill_b(b_limit_in);
LD_UB8(src_u, stride, p3_u, p2_u, p1_u, p0_u, q0_u, q1_u, q2_u, q3_u);
src_u += (5 * stride);
LD_UB8(src_v, stride, p3_v, p2_v, p1_v, p0_v, q0_v, q1_v, q2_v, q3_v);
src_v += (5 * stride);
ILVR_D4_UB(p3_v, p3_u, p2_v, p2_u, p1_v, p1_u, p0_v, p0_u, p3, p2, p1, p0);
ILVR_D4_UB(q0_v, q0_u, q1_v, q1_u, q2_v, q2_u, q3_v, q3_u, q0, q1, q2, q3);
LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit, b_limit, thresh,
hev, mask);
LPF_FILTER4_4W(p1, p0, q0, q1, mask, hev);
p1_d = __msa_copy_s_d((v2i64)p1, 0);
p0_d = __msa_copy_s_d((v2i64)p0, 0);
q0_d = __msa_copy_s_d((v2i64)q0, 0);
q1_d = __msa_copy_s_d((v2i64)q1, 0);
SD4(q1_d, q0_d, p0_d, p1_d, src_u, -stride);
p1_d = __msa_copy_s_d((v2i64)p1, 1);
p0_d = __msa_copy_s_d((v2i64)p0, 1);
q0_d = __msa_copy_s_d((v2i64)q0, 1);
q1_d = __msa_copy_s_d((v2i64)q1, 1);
SD4(q1_d, q0_d, p0_d, p1_d, src_v, -stride);
}
static void HFilter8i(uint8_t* src_u, uint8_t* src_v, int stride,
int b_limit_in, int limit_in, int thresh_in) {
v16u8 p3, p2, p1, p0, q3, q2, q1, q0, mask, hev;
v16u8 row0, row1, row2, row3, row4, row5, row6, row7, row8;
v16u8 row9, row10, row11, row12, row13, row14, row15;
v4i32 tmp0, tmp1, tmp2, tmp3, tmp4, tmp5;
const v16u8 thresh = (v16u8)__msa_fill_b(thresh_in);
const v16u8 limit = (v16u8)__msa_fill_b(limit_in);
const v16u8 b_limit = (v16u8)__msa_fill_b(b_limit_in);
LD_UB8(src_u, stride, row0, row1, row2, row3, row4, row5, row6, row7);
LD_UB8(src_v, stride,
row8, row9, row10, row11, row12, row13, row14, row15);
TRANSPOSE16x8_UB_UB(row0, row1, row2, row3, row4, row5, row6, row7,
row8, row9, row10, row11, row12, row13, row14, row15,
p3, p2, p1, p0, q0, q1, q2, q3);
LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit, b_limit, thresh,
hev, mask);
LPF_FILTER4_4W(p1, p0, q0, q1, mask, hev);
ILVR_B2_SW(p0, p1, q1, q0, tmp0, tmp1);
ILVRL_H2_SW(tmp1, tmp0, tmp2, tmp3);
ILVL_B2_SW(p0, p1, q1, q0, tmp0, tmp1);
ILVRL_H2_SW(tmp1, tmp0, tmp4, tmp5);
src_u += 2;
ST4x4_UB(tmp2, tmp2, 0, 1, 2, 3, src_u, stride);
src_u += 4 * stride;
ST4x4_UB(tmp3, tmp3, 0, 1, 2, 3, src_u, stride);
src_v += 2;
ST4x4_UB(tmp4, tmp4, 0, 1, 2, 3, src_v, stride);
src_v += 4 * stride;
ST4x4_UB(tmp5, tmp5, 0, 1, 2, 3, src_v, stride);
}
static void SimpleVFilter16(uint8_t* src, int stride, int b_limit_in) {
v16u8 p1, p0, q1, q0, mask;
const v16u8 b_limit = (v16u8)__msa_fill_b(b_limit_in);
LD_UB4(src - 2 * stride, stride, p1, p0, q0, q1);
LPF_SIMPLE_MASK(p1, p0, q0, q1, b_limit, mask);
LPF_SIMPLE_FILT(p1, p0, q0, q1, mask);
ST_UB2(p0, q0, src - stride, stride);
}
static void SimpleHFilter16(uint8_t* src, int stride, int b_limit_in) {
v16u8 p1, p0, q1, q0, mask, row0, row1, row2, row3, row4, row5, row6, row7;
v16u8 row8, row9, row10, row11, row12, row13, row14, row15;
v8i16 tmp0, tmp1;
const v16u8 b_limit = (v16u8)__msa_fill_b(b_limit_in);
uint8_t* ptemp_src = src - 2;
LD_UB8(ptemp_src, stride, row0, row1, row2, row3, row4, row5, row6, row7);
LD_UB8(ptemp_src + 8 * stride, stride,
row8, row9, row10, row11, row12, row13, row14, row15);
TRANSPOSE16x4_UB_UB(row0, row1, row2, row3, row4, row5, row6, row7,
row8, row9, row10, row11, row12, row13, row14, row15,
p1, p0, q0, q1);
LPF_SIMPLE_MASK(p1, p0, q0, q1, b_limit, mask);
LPF_SIMPLE_FILT(p1, p0, q0, q1, mask);
ILVRL_B2_SH(q0, p0, tmp1, tmp0);
ptemp_src += 1;
ST2x4_UB(tmp1, 0, ptemp_src, stride);
ptemp_src += 4 * stride;
ST2x4_UB(tmp1, 4, ptemp_src, stride);
ptemp_src += 4 * stride;
ST2x4_UB(tmp0, 0, ptemp_src, stride);
ptemp_src += 4 * stride;
ST2x4_UB(tmp0, 4, ptemp_src, stride);
ptemp_src += 4 * stride;
}
static void SimpleVFilter16i(uint8_t* src_y, int stride, int b_limit_in) {
SimpleVFilter16(src_y + 4 * stride, stride, b_limit_in);
SimpleVFilter16(src_y + 8 * stride, stride, b_limit_in);
SimpleVFilter16(src_y + 12 * stride, stride, b_limit_in);
}
static void SimpleHFilter16i(uint8_t* src_y, int stride, int b_limit_in) {
SimpleHFilter16(src_y + 4, stride, b_limit_in);
SimpleHFilter16(src_y + 8, stride, b_limit_in);
SimpleHFilter16(src_y + 12, stride, b_limit_in);
}
//------------------------------------------------------------------------------
// Intra predictions
//------------------------------------------------------------------------------
// 4x4
static void DC4(uint8_t* dst) { // DC
uint32_t dc = 4;
int i;
for (i = 0; i < 4; ++i) dc += dst[i - BPS] + dst[-1 + i * BPS];
dc >>= 3;
dc = dc | (dc << 8) | (dc << 16) | (dc << 24);
SW4(dc, dc, dc, dc, dst, BPS);
}
static void TM4(uint8_t* dst) {
const uint8_t* const ptemp = dst - BPS - 1;
v8i16 T, d, r0, r1, r2, r3;
const v16i8 zero = { 0 };
const v8i16 TL = (v8i16)__msa_fill_h(ptemp[0 * BPS]);
const v8i16 L0 = (v8i16)__msa_fill_h(ptemp[1 * BPS]);
const v8i16 L1 = (v8i16)__msa_fill_h(ptemp[2 * BPS]);
const v8i16 L2 = (v8i16)__msa_fill_h(ptemp[3 * BPS]);
const v8i16 L3 = (v8i16)__msa_fill_h(ptemp[4 * BPS]);
const v16u8 T1 = LD_UB(ptemp + 1);
T = (v8i16)__msa_ilvr_b(zero, (v16i8)T1);
d = T - TL;
ADD4(d, L0, d, L1, d, L2, d, L3, r0, r1, r2, r3);
CLIP_SH4_0_255(r0, r1, r2, r3);
PCKEV_ST4x4_UB(r0, r1, r2, r3, dst, BPS);
}
static void VE4(uint8_t* dst) { // vertical
const uint8_t* const ptop = dst - BPS - 1;
const uint32_t val0 = LW(ptop + 0);
const uint32_t val1 = LW(ptop + 4);
uint32_t out;
v16u8 A, B, C, AC, B2, R;
INSERT_W2_UB(val0, val1, A);
B = SLDI_UB(A, A, 1);
C = SLDI_UB(A, A, 2);
AC = __msa_ave_u_b(A, C);
B2 = __msa_ave_u_b(B, B);
R = __msa_aver_u_b(AC, B2);
out = __msa_copy_s_w((v4i32)R, 0);
SW4(out, out, out, out, dst, BPS);
}
static void RD4(uint8_t* dst) { // Down-right
const uint8_t* const ptop = dst - 1 - BPS;
uint32_t val0 = LW(ptop + 0);
uint32_t val1 = LW(ptop + 4);
uint32_t val2, val3;
v16u8 A, B, C, AC, B2, R, A1;
INSERT_W2_UB(val0, val1, A1);
A = SLDI_UB(A1, A1, 12);
A = (v16u8)__msa_insert_b((v16i8)A, 3, ptop[1 * BPS]);
A = (v16u8)__msa_insert_b((v16i8)A, 2, ptop[2 * BPS]);
A = (v16u8)__msa_insert_b((v16i8)A, 1, ptop[3 * BPS]);
A = (v16u8)__msa_insert_b((v16i8)A, 0, ptop[4 * BPS]);
B = SLDI_UB(A, A, 1);
C = SLDI_UB(A, A, 2);
AC = __msa_ave_u_b(A, C);
B2 = __msa_ave_u_b(B, B);
R = __msa_aver_u_b(AC, B2);
val3 = __msa_copy_s_w((v4i32)R, 0);
R = SLDI_UB(R, R, 1);
val2 = __msa_copy_s_w((v4i32)R, 0);
R = SLDI_UB(R, R, 1);
val1 = __msa_copy_s_w((v4i32)R, 0);
R = SLDI_UB(R, R, 1);
val0 = __msa_copy_s_w((v4i32)R, 0);
SW4(val0, val1, val2, val3, dst, BPS);
}
static void LD4(uint8_t* dst) { // Down-Left
const uint8_t* const ptop = dst - BPS;
uint32_t val0 = LW(ptop + 0);
uint32_t val1 = LW(ptop + 4);
uint32_t val2, val3;
v16u8 A, B, C, AC, B2, R;
INSERT_W2_UB(val0, val1, A);
B = SLDI_UB(A, A, 1);
C = SLDI_UB(A, A, 2);
C = (v16u8)__msa_insert_b((v16i8)C, 6, ptop[7]);
AC = __msa_ave_u_b(A, C);
B2 = __msa_ave_u_b(B, B);
R = __msa_aver_u_b(AC, B2);
val0 = __msa_copy_s_w((v4i32)R, 0);
R = SLDI_UB(R, R, 1);
val1 = __msa_copy_s_w((v4i32)R, 0);
R = SLDI_UB(R, R, 1);
val2 = __msa_copy_s_w((v4i32)R, 0);
R = SLDI_UB(R, R, 1);
val3 = __msa_copy_s_w((v4i32)R, 0);
SW4(val0, val1, val2, val3, dst, BPS);
}
// 16x16
static void DC16(uint8_t* dst) { // DC
uint32_t dc = 16;
int i;
const v16u8 rtop = LD_UB(dst - BPS);
const v8u16 dctop = __msa_hadd_u_h(rtop, rtop);
v16u8 out;
for (i = 0; i < 16; ++i) {
dc += dst[-1 + i * BPS];
}
dc += HADD_UH_U32(dctop);
out = (v16u8)__msa_fill_b(dc >> 5);
ST_UB8(out, out, out, out, out, out, out, out, dst, BPS);
ST_UB8(out, out, out, out, out, out, out, out, dst + 8 * BPS, BPS);
}
static void TM16(uint8_t* dst) {
int j;
v8i16 d1, d2;
const v16i8 zero = { 0 };
const v8i16 TL = (v8i16)__msa_fill_h(dst[-1 - BPS]);
const v16i8 T = LD_SB(dst - BPS);
ILVRL_B2_SH(zero, T, d1, d2);
SUB2(d1, TL, d2, TL, d1, d2);
for (j = 0; j < 16; j += 4) {
v16i8 t0, t1, t2, t3;
v8i16 r0, r1, r2, r3, r4, r5, r6, r7;
const v8i16 L0 = (v8i16)__msa_fill_h(dst[-1 + 0 * BPS]);
const v8i16 L1 = (v8i16)__msa_fill_h(dst[-1 + 1 * BPS]);
const v8i16 L2 = (v8i16)__msa_fill_h(dst[-1 + 2 * BPS]);
const v8i16 L3 = (v8i16)__msa_fill_h(dst[-1 + 3 * BPS]);
ADD4(d1, L0, d1, L1, d1, L2, d1, L3, r0, r1, r2, r3);
ADD4(d2, L0, d2, L1, d2, L2, d2, L3, r4, r5, r6, r7);
CLIP_SH4_0_255(r0, r1, r2, r3);
CLIP_SH4_0_255(r4, r5, r6, r7);
PCKEV_B4_SB(r4, r0, r5, r1, r6, r2, r7, r3, t0, t1, t2, t3);
ST_SB4(t0, t1, t2, t3, dst, BPS);
dst += 4 * BPS;
}
}
static void VE16(uint8_t* dst) { // vertical
const v16u8 rtop = LD_UB(dst - BPS);
ST_UB8(rtop, rtop, rtop, rtop, rtop, rtop, rtop, rtop, dst, BPS);
ST_UB8(rtop, rtop, rtop, rtop, rtop, rtop, rtop, rtop, dst + 8 * BPS, BPS);
}
static void HE16(uint8_t* dst) { // horizontal
int j;
for (j = 16; j > 0; j -= 4) {
const v16u8 L0 = (v16u8)__msa_fill_b(dst[-1 + 0 * BPS]);
const v16u8 L1 = (v16u8)__msa_fill_b(dst[-1 + 1 * BPS]);
const v16u8 L2 = (v16u8)__msa_fill_b(dst[-1 + 2 * BPS]);
const v16u8 L3 = (v16u8)__msa_fill_b(dst[-1 + 3 * BPS]);
ST_UB4(L0, L1, L2, L3, dst, BPS);
dst += 4 * BPS;
}
}
static void DC16NoTop(uint8_t* dst) { // DC with top samples not available
int j;
uint32_t dc = 8;
v16u8 out;
for (j = 0; j < 16; ++j) {
dc += dst[-1 + j * BPS];
}
out = (v16u8)__msa_fill_b(dc >> 4);
ST_UB8(out, out, out, out, out, out, out, out, dst, BPS);
ST_UB8(out, out, out, out, out, out, out, out, dst + 8 * BPS, BPS);
}
static void DC16NoLeft(uint8_t* dst) { // DC with left samples not available
uint32_t dc = 8;
const v16u8 rtop = LD_UB(dst - BPS);
const v8u16 dctop = __msa_hadd_u_h(rtop, rtop);
v16u8 out;
dc += HADD_UH_U32(dctop);
out = (v16u8)__msa_fill_b(dc >> 4);
ST_UB8(out, out, out, out, out, out, out, out, dst, BPS);
ST_UB8(out, out, out, out, out, out, out, out, dst + 8 * BPS, BPS);
}
static void DC16NoTopLeft(uint8_t* dst) { // DC with nothing
const v16u8 out = (v16u8)__msa_fill_b(0x80);
ST_UB8(out, out, out, out, out, out, out, out, dst, BPS);
ST_UB8(out, out, out, out, out, out, out, out, dst + 8 * BPS, BPS);
}
// Chroma
#define STORE8x8(out, dst) do { \
SD4(out, out, out, out, dst + 0 * BPS, BPS); \
SD4(out, out, out, out, dst + 4 * BPS, BPS); \
} while (0)
static void DC8uv(uint8_t* dst) { // DC
uint32_t dc = 8;
int i;
uint64_t out;
const v16u8 rtop = LD_UB(dst - BPS);
const v8u16 temp0 = __msa_hadd_u_h(rtop, rtop);
const v4u32 temp1 = __msa_hadd_u_w(temp0, temp0);
const v2u64 temp2 = __msa_hadd_u_d(temp1, temp1);
v16u8 dctemp;
for (i = 0; i < 8; ++i) {
dc += dst[-1 + i * BPS];
}
dc += __msa_copy_s_w((v4i32)temp2, 0);
dctemp = (v16u8)__msa_fill_b(dc >> 4);
out = __msa_copy_s_d((v2i64)dctemp, 0);
STORE8x8(out, dst);
}
static void TM8uv(uint8_t* dst) {
int j;
const v16i8 T1 = LD_SB(dst - BPS);
const v16i8 zero = { 0 };
const v8i16 T = (v8i16)__msa_ilvr_b(zero, T1);
const v8i16 TL = (v8i16)__msa_fill_h(dst[-1 - BPS]);
const v8i16 d = T - TL;
for (j = 0; j < 8; j += 4) {
v16i8 t0, t1;
v8i16 r0 = (v8i16)__msa_fill_h(dst[-1 + 0 * BPS]);
v8i16 r1 = (v8i16)__msa_fill_h(dst[-1 + 1 * BPS]);
v8i16 r2 = (v8i16)__msa_fill_h(dst[-1 + 2 * BPS]);
v8i16 r3 = (v8i16)__msa_fill_h(dst[-1 + 3 * BPS]);
ADD4(d, r0, d, r1, d, r2, d, r3, r0, r1, r2, r3);
CLIP_SH4_0_255(r0, r1, r2, r3);
PCKEV_B2_SB(r1, r0, r3, r2, t0, t1);
ST4x4_UB(t0, t1, 0, 2, 0, 2, dst, BPS);
ST4x4_UB(t0, t1, 1, 3, 1, 3, dst + 4, BPS);
dst += 4 * BPS;
}
}
static void VE8uv(uint8_t* dst) { // vertical
const v16u8 rtop = LD_UB(dst - BPS);
const uint64_t out = __msa_copy_s_d((v2i64)rtop, 0);
STORE8x8(out, dst);
}
static void HE8uv(uint8_t* dst) { // horizontal
int j;
for (j = 0; j < 8; j += 4) {
const v16u8 L0 = (v16u8)__msa_fill_b(dst[-1 + 0 * BPS]);
const v16u8 L1 = (v16u8)__msa_fill_b(dst[-1 + 1 * BPS]);
const v16u8 L2 = (v16u8)__msa_fill_b(dst[-1 + 2 * BPS]);
const v16u8 L3 = (v16u8)__msa_fill_b(dst[-1 + 3 * BPS]);
const uint64_t out0 = __msa_copy_s_d((v2i64)L0, 0);
const uint64_t out1 = __msa_copy_s_d((v2i64)L1, 0);
const uint64_t out2 = __msa_copy_s_d((v2i64)L2, 0);
const uint64_t out3 = __msa_copy_s_d((v2i64)L3, 0);
SD4(out0, out1, out2, out3, dst, BPS);
dst += 4 * BPS;
}
}
static void DC8uvNoLeft(uint8_t* dst) { // DC with no left samples
const uint32_t dc = 4;
const v16u8 rtop = LD_UB(dst - BPS);
const v8u16 temp0 = __msa_hadd_u_h(rtop, rtop);
const v4u32 temp1 = __msa_hadd_u_w(temp0, temp0);
const v2u64 temp2 = __msa_hadd_u_d(temp1, temp1);
const uint32_t sum_m = __msa_copy_s_w((v4i32)temp2, 0);
const v16u8 dcval = (v16u8)__msa_fill_b((dc + sum_m) >> 3);
const uint64_t out = __msa_copy_s_d((v2i64)dcval, 0);
STORE8x8(out, dst);
}
static void DC8uvNoTop(uint8_t* dst) { // DC with no top samples
uint32_t dc = 4;
int i;
uint64_t out;
v16u8 dctemp;
for (i = 0; i < 8; ++i) {
dc += dst[-1 + i * BPS];
}
dctemp = (v16u8)__msa_fill_b(dc >> 3);
out = __msa_copy_s_d((v2i64)dctemp, 0);
STORE8x8(out, dst);
}
static void DC8uvNoTopLeft(uint8_t* dst) { // DC with nothing
const uint64_t out = 0x8080808080808080ULL;
STORE8x8(out, dst);
}
//------------------------------------------------------------------------------
// Entry point
@ -163,6 +977,39 @@ WEBP_TSAN_IGNORE_FUNCTION void VP8DspInitMSA(void) {
VP8Transform = TransformTwo;
VP8TransformDC = TransformDC;
VP8TransformAC3 = TransformAC3;
VP8VFilter16 = VFilter16;
VP8HFilter16 = HFilter16;
VP8VFilter16i = VFilter16i;
VP8HFilter16i = HFilter16i;
VP8VFilter8 = VFilter8;
VP8HFilter8 = HFilter8;
VP8VFilter8i = VFilter8i;
VP8HFilter8i = HFilter8i;
VP8SimpleVFilter16 = SimpleVFilter16;
VP8SimpleHFilter16 = SimpleHFilter16;
VP8SimpleVFilter16i = SimpleVFilter16i;
VP8SimpleHFilter16i = SimpleHFilter16i;
VP8PredLuma4[0] = DC4;
VP8PredLuma4[1] = TM4;
VP8PredLuma4[2] = VE4;
VP8PredLuma4[4] = RD4;
VP8PredLuma4[6] = LD4;
VP8PredLuma16[0] = DC16;
VP8PredLuma16[1] = TM16;
VP8PredLuma16[2] = VE16;
VP8PredLuma16[3] = HE16;
VP8PredLuma16[4] = DC16NoTop;
VP8PredLuma16[5] = DC16NoLeft;
VP8PredLuma16[6] = DC16NoTopLeft;
VP8PredChroma8[0] = DC8uv;
VP8PredChroma8[1] = TM8uv;
VP8PredChroma8[2] = VE8uv;
VP8PredChroma8[3] = HE8uv;
VP8PredChroma8[4] = DC8uvNoTop;
VP8PredChroma8[5] = DC8uvNoLeft;
VP8PredChroma8[6] = DC8uvNoTopLeft;
}
#else // !WEBP_USE_MSA

40
thirdparty/libwebp/dsp/dec_neon.c vendored
View file

@ -17,7 +17,7 @@
#if defined(WEBP_USE_NEON)
#include "./neon.h"
#include "../dec/vp8i.h"
#include "../dec/vp8i_dec.h"
//------------------------------------------------------------------------------
// NxM Loading functions
@ -666,9 +666,8 @@ static uint8x16_t NeedsHev(const uint8x16_t p1, const uint8x16_t p0,
const uint8x16_t hev_thresh_v = vdupq_n_u8((uint8_t)hev_thresh);
const uint8x16_t a_p1_p0 = vabdq_u8(p1, p0); // abs(p1 - p0)
const uint8x16_t a_q1_q0 = vabdq_u8(q1, q0); // abs(q1 - q0)
const uint8x16_t mask1 = vcgtq_u8(a_p1_p0, hev_thresh_v);
const uint8x16_t mask2 = vcgtq_u8(a_q1_q0, hev_thresh_v);
const uint8x16_t mask = vorrq_u8(mask1, mask2);
const uint8x16_t a_max = vmaxq_u8(a_p1_p0, a_q1_q0);
const uint8x16_t mask = vcgtq_u8(a_max, hev_thresh_v);
return mask;
}
@ -756,24 +755,25 @@ static void ApplyFilter6(
const int8x16_t delta,
uint8x16_t* const op2, uint8x16_t* const op1, uint8x16_t* const op0,
uint8x16_t* const oq0, uint8x16_t* const oq1, uint8x16_t* const oq2) {
const int16x8_t kCst63 = vdupq_n_s16(63);
const int8x8_t kCst27 = vdup_n_s8(27);
const int8x8_t kCst18 = vdup_n_s8(18);
const int8x8_t kCst9 = vdup_n_s8(9);
// We have to compute: X = (9*a+63) >> 7, Y = (18*a+63)>>7, Z = (27*a+63) >> 7
// Turns out, there's a common sub-expression S=9 * a - 1 that can be used
// with the special vqrshrn_n_s16 rounding-shift-and-narrow instruction:
// X = (S + 64) >> 7, Y = (S + 32) >> 6, Z = (18 * a + S + 64) >> 7
const int8x8_t delta_lo = vget_low_s8(delta);
const int8x8_t delta_hi = vget_high_s8(delta);
const int16x8_t s1_lo = vmlal_s8(kCst63, kCst27, delta_lo); // 63 + 27 * a
const int16x8_t s1_hi = vmlal_s8(kCst63, kCst27, delta_hi); // 63 + 27 * a
const int16x8_t s2_lo = vmlal_s8(kCst63, kCst18, delta_lo); // 63 + 18 * a
const int16x8_t s2_hi = vmlal_s8(kCst63, kCst18, delta_hi); // 63 + 18 * a
const int16x8_t s3_lo = vmlal_s8(kCst63, kCst9, delta_lo); // 63 + 9 * a
const int16x8_t s3_hi = vmlal_s8(kCst63, kCst9, delta_hi); // 63 + 9 * a
const int8x8_t a1_lo = vqshrn_n_s16(s1_lo, 7);
const int8x8_t a1_hi = vqshrn_n_s16(s1_hi, 7);
const int8x8_t a2_lo = vqshrn_n_s16(s2_lo, 7);
const int8x8_t a2_hi = vqshrn_n_s16(s2_hi, 7);
const int8x8_t a3_lo = vqshrn_n_s16(s3_lo, 7);
const int8x8_t a3_hi = vqshrn_n_s16(s3_hi, 7);
const int8x8_t kCst9 = vdup_n_s8(9);
const int16x8_t kCstm1 = vdupq_n_s16(-1);
const int8x8_t kCst18 = vdup_n_s8(18);
const int16x8_t S_lo = vmlal_s8(kCstm1, kCst9, delta_lo); // S = 9 * a - 1
const int16x8_t S_hi = vmlal_s8(kCstm1, kCst9, delta_hi);
const int16x8_t Z_lo = vmlal_s8(S_lo, kCst18, delta_lo); // S + 18 * a
const int16x8_t Z_hi = vmlal_s8(S_hi, kCst18, delta_hi);
const int8x8_t a3_lo = vqrshrn_n_s16(S_lo, 7); // (9 * a + 63) >> 7
const int8x8_t a3_hi = vqrshrn_n_s16(S_hi, 7);
const int8x8_t a2_lo = vqrshrn_n_s16(S_lo, 6); // (9 * a + 31) >> 6
const int8x8_t a2_hi = vqrshrn_n_s16(S_hi, 6);
const int8x8_t a1_lo = vqrshrn_n_s16(Z_lo, 7); // (27 * a + 63) >> 7
const int8x8_t a1_hi = vqrshrn_n_s16(Z_hi, 7);
const int8x16_t a1 = vcombine_s8(a1_lo, a1_hi);
const int8x16_t a2 = vcombine_s8(a2_lo, a2_hi);
const int8x16_t a3 = vcombine_s8(a3_lo, a3_hi);

4
thirdparty/libwebp/dsp/dec_sse2.c vendored
View file

@ -22,7 +22,7 @@
#include <emmintrin.h>
#include "./common_sse2.h"
#include "../dec/vp8i.h"
#include "../dec/vp8i_dec.h"
#include "../utils/utils.h"
//------------------------------------------------------------------------------
@ -140,7 +140,7 @@ static void Transform(const int16_t* in, uint8_t* dst, int do_two) {
// Transpose the two 4x4.
VP8Transpose_2_4x4_16b(&shifted0, &shifted1, &shifted2, &shifted3, &T0, &T1,
&T2, &T3);
&T2, &T3);
}
// Add inverse transform to 'dst' and store.

2
thirdparty/libwebp/dsp/dec_sse41.c vendored
View file

@ -16,7 +16,7 @@
#if defined(WEBP_USE_SSE41)
#include <smmintrin.h>
#include "../dec/vp8i.h"
#include "../dec/vp8i_dec.h"
#include "../utils/utils.h"
static void HE16(uint8_t* dst) { // horizontal

57
thirdparty/libwebp/dsp/dsp.h vendored
View file

@ -111,8 +111,7 @@ extern "C" {
#define WEBP_UBSAN_IGNORE_UNDEF
#define WEBP_UBSAN_IGNORE_UNSIGNED_OVERFLOW
#if !defined(WEBP_FORCE_ALIGNED) && defined(__clang__) && \
defined(__has_attribute)
#if defined(__clang__) && defined(__has_attribute)
#if __has_attribute(no_sanitize)
// This macro prevents the undefined behavior sanitizer from reporting
// failures. This is only meant to silence unaligned loads on platforms that
@ -133,6 +132,7 @@ extern "C" {
typedef enum {
kSSE2,
kSSE3,
kSlowSSSE3, // special feature for slow SSSE3 architectures
kSSE4_1,
kAVX,
kAVX2,
@ -185,6 +185,11 @@ typedef int (*VP8WMetric)(const uint8_t* pix, const uint8_t* ref,
// 4 by 4 symmetric matrix.
extern VP8WMetric VP8TDisto4x4, VP8TDisto16x16;
// Compute the average (DC) of four 4x4 blocks.
// Each sub-4x4 block #i sum is stored in dc[i].
typedef void (*VP8MeanMetric)(const uint8_t* ref, uint32_t dc[4]);
extern VP8MeanMetric VP8Mean16x4;
typedef void (*VP8BlockCopy)(const uint8_t* src, uint8_t* dst);
extern VP8BlockCopy VP8Copy4x4;
extern VP8BlockCopy VP8Copy16x8;
@ -246,30 +251,37 @@ extern VP8GetResidualCostFunc VP8GetResidualCost;
void VP8EncDspCostInit(void);
//------------------------------------------------------------------------------
// SSIM utils
// SSIM / PSNR utils
// struct for accumulating statistical moments
typedef struct {
double w; // sum(w_i) : sum of weights
double xm, ym; // sum(w_i * x_i), sum(w_i * y_i)
double xxm, xym, yym; // sum(w_i * x_i * x_i), etc.
uint32_t w; // sum(w_i) : sum of weights
uint32_t xm, ym; // sum(w_i * x_i), sum(w_i * y_i)
uint32_t xxm, xym, yym; // sum(w_i * x_i * x_i), etc.
} VP8DistoStats;
// Compute the final SSIM value
// The non-clipped version assumes stats->w = (2 * VP8_SSIM_KERNEL + 1)^2.
double VP8SSIMFromStats(const VP8DistoStats* const stats);
double VP8SSIMFromStatsClipped(const VP8DistoStats* const stats);
#define VP8_SSIM_KERNEL 3 // total size of the kernel: 2 * VP8_SSIM_KERNEL + 1
typedef void (*VP8SSIMAccumulateClippedFunc)(const uint8_t* src1, int stride1,
const uint8_t* src2, int stride2,
int xo, int yo, // center position
int W, int H, // plane dimension
VP8DistoStats* const stats);
typedef double (*VP8SSIMGetClippedFunc)(const uint8_t* src1, int stride1,
const uint8_t* src2, int stride2,
int xo, int yo, // center position
int W, int H); // plane dimension
// This version is called with the guarantee that you can load 8 bytes and
// 8 rows at offset src1 and src2
typedef void (*VP8SSIMAccumulateFunc)(const uint8_t* src1, int stride1,
const uint8_t* src2, int stride2,
VP8DistoStats* const stats);
typedef double (*VP8SSIMGetFunc)(const uint8_t* src1, int stride1,
const uint8_t* src2, int stride2);
extern VP8SSIMAccumulateFunc VP8SSIMAccumulate; // unclipped / unchecked
extern VP8SSIMAccumulateClippedFunc VP8SSIMAccumulateClipped; // with clipping
extern VP8SSIMGetFunc VP8SSIMGet; // unclipped / unchecked
extern VP8SSIMGetClippedFunc VP8SSIMGetClipped; // with clipping
typedef uint32_t (*VP8AccumulateSSEFunc)(const uint8_t* src1,
const uint8_t* src2, int len);
extern VP8AccumulateSSEFunc VP8AccumulateSSE;
// must be called before using any of the above directly
void VP8SSIMDspInit(void);
@ -416,6 +428,15 @@ extern void WebPConvertARGBToUV_C(const uint32_t* argb, uint8_t* u, uint8_t* v,
extern void WebPConvertRGBA32ToUV_C(const uint16_t* rgb,
uint8_t* u, uint8_t* v, int width);
// utilities for accurate RGB->YUV conversion
extern uint64_t (*WebPSharpYUVUpdateY)(const uint16_t* src, const uint16_t* ref,
uint16_t* dst, int len);
extern void (*WebPSharpYUVUpdateRGB)(const int16_t* src, const int16_t* ref,
int16_t* dst, int len);
extern void (*WebPSharpYUVFilterRow)(const int16_t* A, const int16_t* B,
int len,
const uint16_t* best_y, uint16_t* out);
// Must be called before using the above.
void WebPInitConvertARGBToYUV(void);
@ -488,6 +509,10 @@ extern int (*WebPExtractAlpha)(const uint8_t* argb, int argb_stride,
int width, int height,
uint8_t* alpha, int alpha_stride);
// Extract the green values from 32b values in argb[] and pack them into alpha[]
// (this is the opposite of WebPDispatchAlphaToGreen).
extern void (*WebPExtractGreen)(const uint32_t* argb, uint8_t* alpha, int size);
// Pre-Multiply operation transforms x into x * A / 255 (where x=Y,R,G or B).
// Un-Multiply operation transforms x into x * 255 / A.

178
thirdparty/libwebp/dsp/enc.c vendored
View file

@ -15,7 +15,7 @@
#include <stdlib.h> // for abs()
#include "./dsp.h"
#include "../enc/vp8enci.h"
#include "../enc/vp8i_enc.h"
static WEBP_INLINE uint8_t clip_8b(int v) {
return (!(v & ~0xff)) ? v : (v < 0) ? 0 : 255;
@ -551,6 +551,20 @@ static int SSE4x4(const uint8_t* a, const uint8_t* b) {
return GetSSE(a, b, 4, 4);
}
static void Mean16x4(const uint8_t* ref, uint32_t dc[4]) {
int k, x, y;
for (k = 0; k < 4; ++k) {
uint32_t avg = 0;
for (y = 0; y < 4; ++y) {
for (x = 0; x < 4; ++x) {
avg += ref[x + y * BPS];
}
}
dc[k] = avg;
ref += 4; // go to next 4x4 block.
}
}
//------------------------------------------------------------------------------
// Texture distortion
//
@ -656,32 +670,6 @@ static int Quantize2Blocks(int16_t in[32], int16_t out[32],
return nz;
}
static int QuantizeBlockWHT(int16_t in[16], int16_t out[16],
const VP8Matrix* const mtx) {
int n, last = -1;
for (n = 0; n < 16; ++n) {
const int j = kZigzag[n];
const int sign = (in[j] < 0);
const uint32_t coeff = sign ? -in[j] : in[j];
assert(mtx->sharpen_[j] == 0);
if (coeff > mtx->zthresh_[j]) {
const uint32_t Q = mtx->q_[j];
const uint32_t iQ = mtx->iq_[j];
const uint32_t B = mtx->bias_[j];
int level = QUANTDIV(coeff, iQ, B);
if (level > MAX_LEVEL) level = MAX_LEVEL;
if (sign) level = -level;
in[j] = level * (int)Q;
out[n] = level;
if (level) last = n;
} else {
out[n] = 0;
in[j] = 0;
}
}
return (last >= 0);
}
//------------------------------------------------------------------------------
// Block copy
@ -703,11 +691,51 @@ static void Copy16x8(const uint8_t* src, uint8_t* dst) {
}
//------------------------------------------------------------------------------
// SSIM / PSNR
static void SSIMAccumulateClipped(const uint8_t* src1, int stride1,
const uint8_t* src2, int stride2,
int xo, int yo, int W, int H,
VP8DistoStats* const stats) {
// hat-shaped filter. Sum of coefficients is equal to 16.
static const uint32_t kWeight[2 * VP8_SSIM_KERNEL + 1] = {
1, 2, 3, 4, 3, 2, 1
};
static const uint32_t kWeightSum = 16 * 16; // sum{kWeight}^2
static WEBP_INLINE double SSIMCalculation(
const VP8DistoStats* const stats, uint32_t N /*num samples*/) {
const uint32_t w2 = N * N;
const uint32_t C1 = 20 * w2;
const uint32_t C2 = 60 * w2;
const uint32_t C3 = 8 * 8 * w2; // 'dark' limit ~= 6
const uint64_t xmxm = (uint64_t)stats->xm * stats->xm;
const uint64_t ymym = (uint64_t)stats->ym * stats->ym;
if (xmxm + ymym >= C3) {
const int64_t xmym = (int64_t)stats->xm * stats->ym;
const int64_t sxy = (int64_t)stats->xym * N - xmym; // can be negative
const uint64_t sxx = (uint64_t)stats->xxm * N - xmxm;
const uint64_t syy = (uint64_t)stats->yym * N - ymym;
// we descale by 8 to prevent overflow during the fnum/fden multiply.
const uint64_t num_S = (2 * (uint64_t)(sxy < 0 ? 0 : sxy) + C2) >> 8;
const uint64_t den_S = (sxx + syy + C2) >> 8;
const uint64_t fnum = (2 * xmym + C1) * num_S;
const uint64_t fden = (xmxm + ymym + C1) * den_S;
const double r = (double)fnum / fden;
assert(r >= 0. && r <= 1.0);
return r;
}
return 1.; // area is too dark to contribute meaningfully
}
double VP8SSIMFromStats(const VP8DistoStats* const stats) {
return SSIMCalculation(stats, kWeightSum);
}
double VP8SSIMFromStatsClipped(const VP8DistoStats* const stats) {
return SSIMCalculation(stats, stats->w);
}
static double SSIMGetClipped_C(const uint8_t* src1, int stride1,
const uint8_t* src2, int stride2,
int xo, int yo, int W, int H) {
VP8DistoStats stats = { 0, 0, 0, 0, 0, 0 };
const int ymin = (yo - VP8_SSIM_KERNEL < 0) ? 0 : yo - VP8_SSIM_KERNEL;
const int ymax = (yo + VP8_SSIM_KERNEL > H - 1) ? H - 1
: yo + VP8_SSIM_KERNEL;
@ -719,38 +747,61 @@ static void SSIMAccumulateClipped(const uint8_t* src1, int stride1,
src2 += ymin * stride2;
for (y = ymin; y <= ymax; ++y, src1 += stride1, src2 += stride2) {
for (x = xmin; x <= xmax; ++x) {
const int s1 = src1[x];
const int s2 = src2[x];
stats->w += 1;
stats->xm += s1;
stats->ym += s2;
stats->xxm += s1 * s1;
stats->xym += s1 * s2;
stats->yym += s2 * s2;
const uint32_t w = kWeight[VP8_SSIM_KERNEL + x - xo]
* kWeight[VP8_SSIM_KERNEL + y - yo];
const uint32_t s1 = src1[x];
const uint32_t s2 = src2[x];
stats.w += w;
stats.xm += w * s1;
stats.ym += w * s2;
stats.xxm += w * s1 * s1;
stats.xym += w * s1 * s2;
stats.yym += w * s2 * s2;
}
}
return VP8SSIMFromStatsClipped(&stats);
}
static void SSIMAccumulate(const uint8_t* src1, int stride1,
const uint8_t* src2, int stride2,
VP8DistoStats* const stats) {
static double SSIMGet_C(const uint8_t* src1, int stride1,
const uint8_t* src2, int stride2) {
VP8DistoStats stats = { 0, 0, 0, 0, 0, 0 };
int x, y;
for (y = 0; y <= 2 * VP8_SSIM_KERNEL; ++y, src1 += stride1, src2 += stride2) {
for (x = 0; x <= 2 * VP8_SSIM_KERNEL; ++x) {
const int s1 = src1[x];
const int s2 = src2[x];
stats->w += 1;
stats->xm += s1;
stats->ym += s2;
stats->xxm += s1 * s1;
stats->xym += s1 * s2;
stats->yym += s2 * s2;
const uint32_t w = kWeight[x] * kWeight[y];
const uint32_t s1 = src1[x];
const uint32_t s2 = src2[x];
stats.xm += w * s1;
stats.ym += w * s2;
stats.xxm += w * s1 * s1;
stats.xym += w * s1 * s2;
stats.yym += w * s2 * s2;
}
}
return VP8SSIMFromStats(&stats);
}
VP8SSIMAccumulateFunc VP8SSIMAccumulate;
VP8SSIMAccumulateClippedFunc VP8SSIMAccumulateClipped;
//------------------------------------------------------------------------------
static uint32_t AccumulateSSE(const uint8_t* src1,
const uint8_t* src2, int len) {
int i;
uint32_t sse2 = 0;
assert(len <= 65535); // to ensure that accumulation fits within uint32_t
for (i = 0; i < len; ++i) {
const int32_t diff = src1[i] - src2[i];
sse2 += diff * diff;
}
return sse2;
}
//------------------------------------------------------------------------------
VP8SSIMGetFunc VP8SSIMGet;
VP8SSIMGetClippedFunc VP8SSIMGetClipped;
VP8AccumulateSSEFunc VP8AccumulateSSE;
extern void VP8SSIMDspInitSSE2(void);
static volatile VP8CPUInfo ssim_last_cpuinfo_used =
(VP8CPUInfo)&ssim_last_cpuinfo_used;
@ -758,8 +809,17 @@ static volatile VP8CPUInfo ssim_last_cpuinfo_used =
WEBP_TSAN_IGNORE_FUNCTION void VP8SSIMDspInit(void) {
if (ssim_last_cpuinfo_used == VP8GetCPUInfo) return;
VP8SSIMAccumulate = SSIMAccumulate;
VP8SSIMAccumulateClipped = SSIMAccumulateClipped;
VP8SSIMGetClipped = SSIMGetClipped_C;
VP8SSIMGet = SSIMGet_C;
VP8AccumulateSSE = AccumulateSSE;
if (VP8GetCPUInfo != NULL) {
#if defined(WEBP_USE_SSE2)
if (VP8GetCPUInfo(kSSE2)) {
VP8SSIMDspInitSSE2();
}
#endif
}
ssim_last_cpuinfo_used = VP8GetCPUInfo;
}
@ -783,6 +843,7 @@ VP8Metric VP8SSE16x8;
VP8Metric VP8SSE4x4;
VP8WMetric VP8TDisto4x4;
VP8WMetric VP8TDisto16x16;
VP8MeanMetric VP8Mean16x4;
VP8QuantizeBlock VP8EncQuantizeBlock;
VP8Quantize2Blocks VP8EncQuantize2Blocks;
VP8QuantizeBlockWHT VP8EncQuantizeBlockWHT;
@ -795,6 +856,7 @@ extern void VP8EncDspInitAVX2(void);
extern void VP8EncDspInitNEON(void);
extern void VP8EncDspInitMIPS32(void);
extern void VP8EncDspInitMIPSdspR2(void);
extern void VP8EncDspInitMSA(void);
static volatile VP8CPUInfo enc_last_cpuinfo_used =
(VP8CPUInfo)&enc_last_cpuinfo_used;
@ -820,9 +882,10 @@ WEBP_TSAN_IGNORE_FUNCTION void VP8EncDspInit(void) {
VP8SSE4x4 = SSE4x4;
VP8TDisto4x4 = Disto4x4;
VP8TDisto16x16 = Disto16x16;
VP8Mean16x4 = Mean16x4;
VP8EncQuantizeBlock = QuantizeBlock;
VP8EncQuantize2Blocks = Quantize2Blocks;
VP8EncQuantizeBlockWHT = QuantizeBlockWHT;
VP8EncQuantizeBlockWHT = QuantizeBlock;
VP8Copy4x4 = Copy4x4;
VP8Copy16x8 = Copy16x8;
@ -857,6 +920,11 @@ WEBP_TSAN_IGNORE_FUNCTION void VP8EncDspInit(void) {
if (VP8GetCPUInfo(kMIPSdspR2)) {
VP8EncDspInitMIPSdspR2();
}
#endif
#if defined(WEBP_USE_MSA)
if (VP8GetCPUInfo(kMSA)) {
VP8EncDspInitMSA();
}
#endif
}
enc_last_cpuinfo_used = VP8GetCPUInfo;

4
thirdparty/libwebp/dsp/enc_mips32.c vendored
View file

@ -18,8 +18,8 @@
#if defined(WEBP_USE_MIPS32)
#include "./mips_macro.h"
#include "../enc/vp8enci.h"
#include "../enc/cost.h"
#include "../enc/vp8i_enc.h"
#include "../enc/cost_enc.h"
static const int kC1 = 20091 + (1 << 16);
static const int kC2 = 35468;

4
thirdparty/libwebp/dsp/enc_mips_dsp_r2.c vendored
View file

@ -17,8 +17,8 @@
#if defined(WEBP_USE_MIPS_DSP_R2)
#include "./mips_macro.h"
#include "../enc/cost.h"
#include "../enc/vp8enci.h"
#include "../enc/cost_enc.h"
#include "../enc/vp8i_enc.h"
static const int kC1 = 20091 + (1 << 16);
static const int kC2 = 35468;

892
thirdparty/libwebp/dsp/enc_msa.c vendored Normal file
View file

@ -0,0 +1,892 @@
// Copyright 2016 Google Inc. All Rights Reserved.
//
// Use of this source code is governed by a BSD-style license
// that can be found in the COPYING file in the root of the source
// tree. An additional intellectual property rights grant can be found
// in the file PATENTS. All contributing project authors may
// be found in the AUTHORS file in the root of the source tree.
// -----------------------------------------------------------------------------
//
// MSA version of encoder dsp functions.
//
// Author: Prashant Patil (prashant.patil@imgtec.com)
#include "./dsp.h"
#if defined(WEBP_USE_MSA)
#include <stdlib.h>
#include "./msa_macro.h"
#include "../enc/vp8i_enc.h"
//------------------------------------------------------------------------------
// Transforms
#define IDCT_1D_W(in0, in1, in2, in3, out0, out1, out2, out3) do { \
v4i32 a1_m, b1_m, c1_m, d1_m; \
const v4i32 cospi8sqrt2minus1 = __msa_fill_w(20091); \
const v4i32 sinpi8sqrt2 = __msa_fill_w(35468); \
v4i32 c_tmp1_m = in1 * sinpi8sqrt2; \
v4i32 c_tmp2_m = in3 * cospi8sqrt2minus1; \
v4i32 d_tmp1_m = in1 * cospi8sqrt2minus1; \
v4i32 d_tmp2_m = in3 * sinpi8sqrt2; \
\
ADDSUB2(in0, in2, a1_m, b1_m); \
SRAI_W2_SW(c_tmp1_m, c_tmp2_m, 16); \
c_tmp2_m = c_tmp2_m + in3; \
c1_m = c_tmp1_m - c_tmp2_m; \
SRAI_W2_SW(d_tmp1_m, d_tmp2_m, 16); \
d_tmp1_m = d_tmp1_m + in1; \
d1_m = d_tmp1_m + d_tmp2_m; \
BUTTERFLY_4(a1_m, b1_m, c1_m, d1_m, out0, out1, out2, out3); \
} while (0)
static WEBP_INLINE void ITransformOne(const uint8_t* ref, const int16_t* in,
uint8_t* dst) {
v8i16 input0, input1;
v4i32 in0, in1, in2, in3, hz0, hz1, hz2, hz3, vt0, vt1, vt2, vt3;
v4i32 res0, res1, res2, res3;
v16i8 dest0, dest1, dest2, dest3;
const v16i8 zero = { 0 };
LD_SH2(in, 8, input0, input1);
UNPCK_SH_SW(input0, in0, in1);
UNPCK_SH_SW(input1, in2, in3);
IDCT_1D_W(in0, in1, in2, in3, hz0, hz1, hz2, hz3);
TRANSPOSE4x4_SW_SW(hz0, hz1, hz2, hz3, hz0, hz1, hz2, hz3);
IDCT_1D_W(hz0, hz1, hz2, hz3, vt0, vt1, vt2, vt3);
SRARI_W4_SW(vt0, vt1, vt2, vt3, 3);
TRANSPOSE4x4_SW_SW(vt0, vt1, vt2, vt3, vt0, vt1, vt2, vt3);
LD_SB4(ref, BPS, dest0, dest1, dest2, dest3);
ILVR_B4_SW(zero, dest0, zero, dest1, zero, dest2, zero, dest3,
res0, res1, res2, res3);
ILVR_H4_SW(zero, res0, zero, res1, zero, res2, zero, res3,
res0, res1, res2, res3);
ADD4(res0, vt0, res1, vt1, res2, vt2, res3, vt3, res0, res1, res2, res3);
CLIP_SW4_0_255(res0, res1, res2, res3);
PCKEV_B2_SW(res0, res1, res2, res3, vt0, vt1);
res0 = (v4i32)__msa_pckev_b((v16i8)vt0, (v16i8)vt1);
ST4x4_UB(res0, res0, 3, 2, 1, 0, dst, BPS);
}
static void ITransform(const uint8_t* ref, const int16_t* in, uint8_t* dst,
int do_two) {
ITransformOne(ref, in, dst);
if (do_two) {
ITransformOne(ref + 4, in + 16, dst + 4);
}
}
static void FTransform(const uint8_t* src, const uint8_t* ref, int16_t* out) {
uint64_t out0, out1, out2, out3;
uint32_t in0, in1, in2, in3;
v4i32 tmp0, tmp1, tmp2, tmp3, tmp4, tmp5;
v8i16 t0, t1, t2, t3;
v16u8 srcl0, srcl1, src0, src1;
const v8i16 mask0 = { 0, 4, 8, 12, 1, 5, 9, 13 };
const v8i16 mask1 = { 3, 7, 11, 15, 2, 6, 10, 14 };
const v8i16 mask2 = { 4, 0, 5, 1, 6, 2, 7, 3 };
const v8i16 mask3 = { 0, 4, 1, 5, 2, 6, 3, 7 };
const v8i16 cnst0 = { 2217, -5352, 2217, -5352, 2217, -5352, 2217, -5352 };
const v8i16 cnst1 = { 5352, 2217, 5352, 2217, 5352, 2217, 5352, 2217 };
LW4(src, BPS, in0, in1, in2, in3);
INSERT_W4_UB(in0, in1, in2, in3, src0);
LW4(ref, BPS, in0, in1, in2, in3);
INSERT_W4_UB(in0, in1, in2, in3, src1);
ILVRL_B2_UB(src0, src1, srcl0, srcl1);
HSUB_UB2_SH(srcl0, srcl1, t0, t1);
VSHF_H2_SH(t0, t1, t0, t1, mask0, mask1, t2, t3);
ADDSUB2(t2, t3, t0, t1);
t0 = SRLI_H(t0, 3);
VSHF_H2_SH(t0, t0, t1, t1, mask2, mask3, t3, t2);
tmp0 = __msa_hadd_s_w(t3, t3);
tmp2 = __msa_hsub_s_w(t3, t3);
FILL_W2_SW(1812, 937, tmp1, tmp3);
DPADD_SH2_SW(t2, t2, cnst0, cnst1, tmp3, tmp1);
SRAI_W2_SW(tmp1, tmp3, 9);
PCKEV_H2_SH(tmp1, tmp0, tmp3, tmp2, t0, t1);
VSHF_H2_SH(t0, t1, t0, t1, mask0, mask1, t2, t3);
ADDSUB2(t2, t3, t0, t1);
VSHF_H2_SH(t0, t0, t1, t1, mask2, mask3, t3, t2);
tmp0 = __msa_hadd_s_w(t3, t3);
tmp2 = __msa_hsub_s_w(t3, t3);
ADDVI_W2_SW(tmp0, 7, tmp2, 7, tmp0, tmp2);
SRAI_W2_SW(tmp0, tmp2, 4);
FILL_W2_SW(12000, 51000, tmp1, tmp3);
DPADD_SH2_SW(t2, t2, cnst0, cnst1, tmp3, tmp1);
SRAI_W2_SW(tmp1, tmp3, 16);
UNPCK_R_SH_SW(t1, tmp4);
tmp5 = __msa_ceqi_w(tmp4, 0);
tmp4 = (v4i32)__msa_nor_v((v16u8)tmp5, (v16u8)tmp5);
tmp5 = __msa_fill_w(1);
tmp5 = (v4i32)__msa_and_v((v16u8)tmp5, (v16u8)tmp4);
tmp1 += tmp5;
PCKEV_H2_SH(tmp1, tmp0, tmp3, tmp2, t0, t1);
out0 = __msa_copy_s_d((v2i64)t0, 0);
out1 = __msa_copy_s_d((v2i64)t0, 1);
out2 = __msa_copy_s_d((v2i64)t1, 0);
out3 = __msa_copy_s_d((v2i64)t1, 1);
SD4(out0, out1, out2, out3, out, 8);
}
static void FTransformWHT(const int16_t* in, int16_t* out) {
v8i16 in0 = { 0 };
v8i16 in1 = { 0 };
v8i16 tmp0, tmp1, tmp2, tmp3;
v8i16 out0, out1;
const v8i16 mask0 = { 0, 1, 2, 3, 8, 9, 10, 11 };
const v8i16 mask1 = { 4, 5, 6, 7, 12, 13, 14, 15 };
const v8i16 mask2 = { 0, 4, 8, 12, 1, 5, 9, 13 };
const v8i16 mask3 = { 3, 7, 11, 15, 2, 6, 10, 14 };
in0 = __msa_insert_h(in0, 0, in[ 0]);
in0 = __msa_insert_h(in0, 1, in[ 64]);
in0 = __msa_insert_h(in0, 2, in[128]);
in0 = __msa_insert_h(in0, 3, in[192]);
in0 = __msa_insert_h(in0, 4, in[ 16]);
in0 = __msa_insert_h(in0, 5, in[ 80]);
in0 = __msa_insert_h(in0, 6, in[144]);
in0 = __msa_insert_h(in0, 7, in[208]);
in1 = __msa_insert_h(in1, 0, in[ 48]);
in1 = __msa_insert_h(in1, 1, in[112]);
in1 = __msa_insert_h(in1, 2, in[176]);
in1 = __msa_insert_h(in1, 3, in[240]);
in1 = __msa_insert_h(in1, 4, in[ 32]);
in1 = __msa_insert_h(in1, 5, in[ 96]);
in1 = __msa_insert_h(in1, 6, in[160]);
in1 = __msa_insert_h(in1, 7, in[224]);
ADDSUB2(in0, in1, tmp0, tmp1);
VSHF_H2_SH(tmp0, tmp1, tmp0, tmp1, mask0, mask1, tmp2, tmp3);
ADDSUB2(tmp2, tmp3, tmp0, tmp1);
VSHF_H2_SH(tmp0, tmp1, tmp0, tmp1, mask2, mask3, in0, in1);
ADDSUB2(in0, in1, tmp0, tmp1);
VSHF_H2_SH(tmp0, tmp1, tmp0, tmp1, mask0, mask1, tmp2, tmp3);
ADDSUB2(tmp2, tmp3, out0, out1);
SRAI_H2_SH(out0, out1, 1);
ST_SH2(out0, out1, out, 8);
}
static int TTransform(const uint8_t* in, const uint16_t* w) {
int sum;
uint32_t in0_m, in1_m, in2_m, in3_m;
v16i8 src0;
v8i16 in0, in1, tmp0, tmp1, tmp2, tmp3;
v4i32 dst0, dst1;
const v16i8 zero = { 0 };
const v8i16 mask0 = { 0, 1, 2, 3, 8, 9, 10, 11 };
const v8i16 mask1 = { 4, 5, 6, 7, 12, 13, 14, 15 };
const v8i16 mask2 = { 0, 4, 8, 12, 1, 5, 9, 13 };
const v8i16 mask3 = { 3, 7, 11, 15, 2, 6, 10, 14 };
LW4(in, BPS, in0_m, in1_m, in2_m, in3_m);
INSERT_W4_SB(in0_m, in1_m, in2_m, in3_m, src0);
ILVRL_B2_SH(zero, src0, tmp0, tmp1);
VSHF_H2_SH(tmp0, tmp1, tmp0, tmp1, mask2, mask3, in0, in1);
ADDSUB2(in0, in1, tmp0, tmp1);
VSHF_H2_SH(tmp0, tmp1, tmp0, tmp1, mask0, mask1, tmp2, tmp3);
ADDSUB2(tmp2, tmp3, tmp0, tmp1);
VSHF_H2_SH(tmp0, tmp1, tmp0, tmp1, mask2, mask3, in0, in1);
ADDSUB2(in0, in1, tmp0, tmp1);
VSHF_H2_SH(tmp0, tmp1, tmp0, tmp1, mask0, mask1, tmp2, tmp3);
ADDSUB2(tmp2, tmp3, tmp0, tmp1);
tmp0 = __msa_add_a_h(tmp0, (v8i16)zero);
tmp1 = __msa_add_a_h(tmp1, (v8i16)zero);
LD_SH2(w, 8, tmp2, tmp3);
DOTP_SH2_SW(tmp0, tmp1, tmp2, tmp3, dst0, dst1);
dst0 = dst0 + dst1;
sum = HADD_SW_S32(dst0);
return sum;
}
static int Disto4x4(const uint8_t* const a, const uint8_t* const b,
const uint16_t* const w) {
const int sum1 = TTransform(a, w);
const int sum2 = TTransform(b, w);
return abs(sum2 - sum1) >> 5;
}
static int Disto16x16(const uint8_t* const a, const uint8_t* const b,
const uint16_t* const w) {
int D = 0;
int x, y;
for (y = 0; y < 16 * BPS; y += 4 * BPS) {
for (x = 0; x < 16; x += 4) {
D += Disto4x4(a + x + y, b + x + y, w);
}
}
return D;
}
//------------------------------------------------------------------------------
// Histogram
static void CollectHistogram(const uint8_t* ref, const uint8_t* pred,
int start_block, int end_block,
VP8Histogram* const histo) {
int j;
int distribution[MAX_COEFF_THRESH + 1] = { 0 };
for (j = start_block; j < end_block; ++j) {
int16_t out[16];
VP8FTransform(ref + VP8DspScan[j], pred + VP8DspScan[j], out);
{
int k;
v8i16 coeff0, coeff1;
const v8i16 zero = { 0 };
const v8i16 max_coeff_thr = __msa_ldi_h(MAX_COEFF_THRESH);
LD_SH2(&out[0], 8, coeff0, coeff1);
coeff0 = __msa_add_a_h(coeff0, zero);
coeff1 = __msa_add_a_h(coeff1, zero);
SRAI_H2_SH(coeff0, coeff1, 3);
coeff0 = __msa_min_s_h(coeff0, max_coeff_thr);
coeff1 = __msa_min_s_h(coeff1, max_coeff_thr);
ST_SH2(coeff0, coeff1, &out[0], 8);
for (k = 0; k < 16; ++k) {
++distribution[out[k]];
}
}
}
VP8SetHistogramData(distribution, histo);
}
//------------------------------------------------------------------------------
// Intra predictions
// luma 4x4 prediction
#define DST(x, y) dst[(x) + (y) * BPS]
#define AVG3(a, b, c) (((a) + 2 * (b) + (c) + 2) >> 2)
#define AVG2(a, b) (((a) + (b) + 1) >> 1)
static WEBP_INLINE void VE4(uint8_t* dst, const uint8_t* top) { // vertical
const uint64_t val_m = LD(top - 1);
const v16u8 A = (v16u8)__msa_insert_d((v2i64)A, 0, val_m);
const v16u8 B = SLDI_UB(A, A, 1);
const v16u8 C = SLDI_UB(A, A, 2);
const v16u8 AC = __msa_ave_u_b(A, C);
const v16u8 B2 = __msa_ave_u_b(B, B);
const v16u8 R = __msa_aver_u_b(AC, B2);
const uint32_t out = __msa_copy_s_w((v4i32)R, 0);
SW4(out, out, out, out, dst, BPS);
}
static WEBP_INLINE void HE4(uint8_t* dst, const uint8_t* top) { // horizontal
const int X = top[-1];
const int I = top[-2];
const int J = top[-3];
const int K = top[-4];
const int L = top[-5];
WebPUint32ToMem(dst + 0 * BPS, 0x01010101U * AVG3(X, I, J));
WebPUint32ToMem(dst + 1 * BPS, 0x01010101U * AVG3(I, J, K));
WebPUint32ToMem(dst + 2 * BPS, 0x01010101U * AVG3(J, K, L));
WebPUint32ToMem(dst + 3 * BPS, 0x01010101U * AVG3(K, L, L));
}
static WEBP_INLINE void DC4(uint8_t* dst, const uint8_t* top) {
uint32_t dc = 4;
int i;
for (i = 0; i < 4; ++i) dc += top[i] + top[-5 + i];
dc >>= 3;
dc = dc | (dc << 8) | (dc << 16) | (dc << 24);
SW4(dc, dc, dc, dc, dst, BPS);
}
static WEBP_INLINE void RD4(uint8_t* dst, const uint8_t* top) {
const uint64_t val_m = LD(top - 5);
const v16u8 A1 = (v16u8)__msa_insert_d((v2i64)A1, 0, val_m);
const v16u8 A = (v16u8)__msa_insert_b((v16i8)A1, 8, top[3]);
const v16u8 B = SLDI_UB(A, A, 1);
const v16u8 C = SLDI_UB(A, A, 2);
const v16u8 AC = __msa_ave_u_b(A, C);
const v16u8 B2 = __msa_ave_u_b(B, B);
const v16u8 R0 = __msa_aver_u_b(AC, B2);
const v16u8 R1 = SLDI_UB(R0, R0, 1);
const v16u8 R2 = SLDI_UB(R1, R1, 1);
const v16u8 R3 = SLDI_UB(R2, R2, 1);
const uint32_t val0 = __msa_copy_s_w((v4i32)R0, 0);
const uint32_t val1 = __msa_copy_s_w((v4i32)R1, 0);
const uint32_t val2 = __msa_copy_s_w((v4i32)R2, 0);
const uint32_t val3 = __msa_copy_s_w((v4i32)R3, 0);
SW4(val3, val2, val1, val0, dst, BPS);
}
static WEBP_INLINE void LD4(uint8_t* dst, const uint8_t* top) {
const uint64_t val_m = LD(top);
const v16u8 A = (v16u8)__msa_insert_d((v2i64)A, 0, val_m);
const v16u8 B = SLDI_UB(A, A, 1);
const v16u8 C1 = SLDI_UB(A, A, 2);
const v16u8 C = (v16u8)__msa_insert_b((v16i8)C1, 6, top[7]);
const v16u8 AC = __msa_ave_u_b(A, C);
const v16u8 B2 = __msa_ave_u_b(B, B);
const v16u8 R0 = __msa_aver_u_b(AC, B2);
const v16u8 R1 = SLDI_UB(R0, R0, 1);
const v16u8 R2 = SLDI_UB(R1, R1, 1);
const v16u8 R3 = SLDI_UB(R2, R2, 1);
const uint32_t val0 = __msa_copy_s_w((v4i32)R0, 0);
const uint32_t val1 = __msa_copy_s_w((v4i32)R1, 0);
const uint32_t val2 = __msa_copy_s_w((v4i32)R2, 0);
const uint32_t val3 = __msa_copy_s_w((v4i32)R3, 0);
SW4(val0, val1, val2, val3, dst, BPS);
}
static WEBP_INLINE void VR4(uint8_t* dst, const uint8_t* top) {
const int X = top[-1];
const int I = top[-2];
const int J = top[-3];
const int K = top[-4];
const int A = top[0];
const int B = top[1];
const int C = top[2];
const int D = top[3];
DST(0, 0) = DST(1, 2) = AVG2(X, A);
DST(1, 0) = DST(2, 2) = AVG2(A, B);
DST(2, 0) = DST(3, 2) = AVG2(B, C);
DST(3, 0) = AVG2(C, D);
DST(0, 3) = AVG3(K, J, I);
DST(0, 2) = AVG3(J, I, X);
DST(0, 1) = DST(1, 3) = AVG3(I, X, A);
DST(1, 1) = DST(2, 3) = AVG3(X, A, B);
DST(2, 1) = DST(3, 3) = AVG3(A, B, C);
DST(3, 1) = AVG3(B, C, D);
}
static WEBP_INLINE void VL4(uint8_t* dst, const uint8_t* top) {
const int A = top[0];
const int B = top[1];
const int C = top[2];
const int D = top[3];
const int E = top[4];
const int F = top[5];
const int G = top[6];
const int H = top[7];
DST(0, 0) = AVG2(A, B);
DST(1, 0) = DST(0, 2) = AVG2(B, C);
DST(2, 0) = DST(1, 2) = AVG2(C, D);
DST(3, 0) = DST(2, 2) = AVG2(D, E);
DST(0, 1) = AVG3(A, B, C);
DST(1, 1) = DST(0, 3) = AVG3(B, C, D);
DST(2, 1) = DST(1, 3) = AVG3(C, D, E);
DST(3, 1) = DST(2, 3) = AVG3(D, E, F);
DST(3, 2) = AVG3(E, F, G);
DST(3, 3) = AVG3(F, G, H);
}
static WEBP_INLINE void HU4(uint8_t* dst, const uint8_t* top) {
const int I = top[-2];
const int J = top[-3];
const int K = top[-4];
const int L = top[-5];
DST(0, 0) = AVG2(I, J);
DST(2, 0) = DST(0, 1) = AVG2(J, K);
DST(2, 1) = DST(0, 2) = AVG2(K, L);
DST(1, 0) = AVG3(I, J, K);
DST(3, 0) = DST(1, 1) = AVG3(J, K, L);
DST(3, 1) = DST(1, 2) = AVG3(K, L, L);
DST(3, 2) = DST(2, 2) =
DST(0, 3) = DST(1, 3) = DST(2, 3) = DST(3, 3) = L;
}
static WEBP_INLINE void HD4(uint8_t* dst, const uint8_t* top) {
const int X = top[-1];
const int I = top[-2];
const int J = top[-3];
const int K = top[-4];
const int L = top[-5];
const int A = top[0];
const int B = top[1];
const int C = top[2];
DST(0, 0) = DST(2, 1) = AVG2(I, X);
DST(0, 1) = DST(2, 2) = AVG2(J, I);
DST(0, 2) = DST(2, 3) = AVG2(K, J);
DST(0, 3) = AVG2(L, K);
DST(3, 0) = AVG3(A, B, C);
DST(2, 0) = AVG3(X, A, B);
DST(1, 0) = DST(3, 1) = AVG3(I, X, A);
DST(1, 1) = DST(3, 2) = AVG3(J, I, X);
DST(1, 2) = DST(3, 3) = AVG3(K, J, I);
DST(1, 3) = AVG3(L, K, J);
}
static WEBP_INLINE void TM4(uint8_t* dst, const uint8_t* top) {
const v16i8 zero = { 0 };
const v8i16 TL = (v8i16)__msa_fill_h(top[-1]);
const v8i16 L0 = (v8i16)__msa_fill_h(top[-2]);
const v8i16 L1 = (v8i16)__msa_fill_h(top[-3]);
const v8i16 L2 = (v8i16)__msa_fill_h(top[-4]);
const v8i16 L3 = (v8i16)__msa_fill_h(top[-5]);
const v16u8 T1 = LD_UB(top);
const v8i16 T = (v8i16)__msa_ilvr_b(zero, (v16i8)T1);
const v8i16 d = T - TL;
v8i16 r0, r1, r2, r3;
ADD4(d, L0, d, L1, d, L2, d, L3, r0, r1, r2, r3);
CLIP_SH4_0_255(r0, r1, r2, r3);
PCKEV_ST4x4_UB(r0, r1, r2, r3, dst, BPS);
}
#undef DST
#undef AVG3
#undef AVG2
static void Intra4Preds(uint8_t* dst, const uint8_t* top) {
DC4(I4DC4 + dst, top);
TM4(I4TM4 + dst, top);
VE4(I4VE4 + dst, top);
HE4(I4HE4 + dst, top);
RD4(I4RD4 + dst, top);
VR4(I4VR4 + dst, top);
LD4(I4LD4 + dst, top);
VL4(I4VL4 + dst, top);
HD4(I4HD4 + dst, top);
HU4(I4HU4 + dst, top);
}
// luma 16x16 prediction
#define STORE16x16(out, dst) do { \
ST_UB8(out, out, out, out, out, out, out, out, dst + 0 * BPS, BPS); \
ST_UB8(out, out, out, out, out, out, out, out, dst + 8 * BPS, BPS); \
} while (0)
static WEBP_INLINE void VerticalPred16x16(uint8_t* dst, const uint8_t* top) {
if (top != NULL) {
const v16u8 out = LD_UB(top);
STORE16x16(out, dst);
} else {
const v16u8 out = (v16u8)__msa_fill_b(0x7f);
STORE16x16(out, dst);
}
}
static WEBP_INLINE void HorizontalPred16x16(uint8_t* dst,
const uint8_t* left) {
if (left != NULL) {
int j;
for (j = 0; j < 16; j += 4) {
const v16u8 L0 = (v16u8)__msa_fill_b(left[0]);
const v16u8 L1 = (v16u8)__msa_fill_b(left[1]);
const v16u8 L2 = (v16u8)__msa_fill_b(left[2]);
const v16u8 L3 = (v16u8)__msa_fill_b(left[3]);
ST_UB4(L0, L1, L2, L3, dst, BPS);
dst += 4 * BPS;
left += 4;
}
} else {
const v16u8 out = (v16u8)__msa_fill_b(0x81);
STORE16x16(out, dst);
}
}
static WEBP_INLINE void TrueMotion16x16(uint8_t* dst, const uint8_t* left,
const uint8_t* top) {
if (left != NULL) {
if (top != NULL) {
int j;
v8i16 d1, d2;
const v16i8 zero = { 0 };
const v8i16 TL = (v8i16)__msa_fill_h(left[-1]);
const v16u8 T = LD_UB(top);
ILVRL_B2_SH(zero, T, d1, d2);
SUB2(d1, TL, d2, TL, d1, d2);
for (j = 0; j < 16; j += 4) {
v16i8 t0, t1, t2, t3;
v8i16 r0, r1, r2, r3, r4, r5, r6, r7;
const v8i16 L0 = (v8i16)__msa_fill_h(left[j + 0]);
const v8i16 L1 = (v8i16)__msa_fill_h(left[j + 1]);
const v8i16 L2 = (v8i16)__msa_fill_h(left[j + 2]);
const v8i16 L3 = (v8i16)__msa_fill_h(left[j + 3]);
ADD4(d1, L0, d1, L1, d1, L2, d1, L3, r0, r1, r2, r3);
ADD4(d2, L0, d2, L1, d2, L2, d2, L3, r4, r5, r6, r7);
CLIP_SH4_0_255(r0, r1, r2, r3);
CLIP_SH4_0_255(r4, r5, r6, r7);
PCKEV_B4_SB(r4, r0, r5, r1, r6, r2, r7, r3, t0, t1, t2, t3);
ST_SB4(t0, t1, t2, t3, dst, BPS);
dst += 4 * BPS;
}
} else {
HorizontalPred16x16(dst, left);
}
} else {
if (top != NULL) {
VerticalPred16x16(dst, top);
} else {
const v16u8 out = (v16u8)__msa_fill_b(0x81);
STORE16x16(out, dst);
}
}
}
static WEBP_INLINE void DCMode16x16(uint8_t* dst, const uint8_t* left,
const uint8_t* top) {
int DC;
v16u8 out;
if (top != NULL && left != NULL) {
const v16u8 rtop = LD_UB(top);
const v8u16 dctop = __msa_hadd_u_h(rtop, rtop);
const v16u8 rleft = LD_UB(left);
const v8u16 dcleft = __msa_hadd_u_h(rleft, rleft);
const v8u16 dctemp = dctop + dcleft;
DC = HADD_UH_U32(dctemp);
DC = (DC + 16) >> 5;
} else if (left != NULL) { // left but no top
const v16u8 rleft = LD_UB(left);
const v8u16 dcleft = __msa_hadd_u_h(rleft, rleft);
DC = HADD_UH_U32(dcleft);
DC = (DC + DC + 16) >> 5;
} else if (top != NULL) { // top but no left
const v16u8 rtop = LD_UB(top);
const v8u16 dctop = __msa_hadd_u_h(rtop, rtop);
DC = HADD_UH_U32(dctop);
DC = (DC + DC + 16) >> 5;
} else { // no top, no left, nothing.
DC = 0x80;
}
out = (v16u8)__msa_fill_b(DC);
STORE16x16(out, dst);
}
static void Intra16Preds(uint8_t* dst,
const uint8_t* left, const uint8_t* top) {
DCMode16x16(I16DC16 + dst, left, top);
VerticalPred16x16(I16VE16 + dst, top);
HorizontalPred16x16(I16HE16 + dst, left);
TrueMotion16x16(I16TM16 + dst, left, top);
}
// Chroma 8x8 prediction
#define CALC_DC8(in, out) do { \
const v8u16 temp0 = __msa_hadd_u_h(in, in); \
const v4u32 temp1 = __msa_hadd_u_w(temp0, temp0); \
const v2i64 temp2 = (v2i64)__msa_hadd_u_d(temp1, temp1); \
const v2i64 temp3 = __msa_splati_d(temp2, 1); \
const v2i64 temp4 = temp3 + temp2; \
const v16i8 temp5 = (v16i8)__msa_srari_d(temp4, 4); \
const v2i64 temp6 = (v2i64)__msa_splati_b(temp5, 0); \
out = __msa_copy_s_d(temp6, 0); \
} while (0)
#define STORE8x8(out, dst) do { \
SD4(out, out, out, out, dst + 0 * BPS, BPS); \
SD4(out, out, out, out, dst + 4 * BPS, BPS); \
} while (0)
static WEBP_INLINE void VerticalPred8x8(uint8_t* dst, const uint8_t* top) {
if (top != NULL) {
const uint64_t out = LD(top);
STORE8x8(out, dst);
} else {
const uint64_t out = 0x7f7f7f7f7f7f7f7fULL;
STORE8x8(out, dst);
}
}
static WEBP_INLINE void HorizontalPred8x8(uint8_t* dst, const uint8_t* left) {
if (left != NULL) {
int j;
for (j = 0; j < 8; j += 4) {
const v16u8 L0 = (v16u8)__msa_fill_b(left[0]);
const v16u8 L1 = (v16u8)__msa_fill_b(left[1]);
const v16u8 L2 = (v16u8)__msa_fill_b(left[2]);
const v16u8 L3 = (v16u8)__msa_fill_b(left[3]);
const uint64_t out0 = __msa_copy_s_d((v2i64)L0, 0);
const uint64_t out1 = __msa_copy_s_d((v2i64)L1, 0);
const uint64_t out2 = __msa_copy_s_d((v2i64)L2, 0);
const uint64_t out3 = __msa_copy_s_d((v2i64)L3, 0);
SD4(out0, out1, out2, out3, dst, BPS);
dst += 4 * BPS;
left += 4;
}
} else {
const uint64_t out = 0x8181818181818181ULL;
STORE8x8(out, dst);
}
}
static WEBP_INLINE void TrueMotion8x8(uint8_t* dst, const uint8_t* left,
const uint8_t* top) {
if (left != NULL) {
if (top != NULL) {
int j;
const v8i16 TL = (v8i16)__msa_fill_h(left[-1]);
const v16u8 T1 = LD_UB(top);
const v16i8 zero = { 0 };
const v8i16 T = (v8i16)__msa_ilvr_b(zero, (v16i8)T1);
const v8i16 d = T - TL;
for (j = 0; j < 8; j += 4) {
uint64_t out0, out1, out2, out3;
v16i8 t0, t1;
v8i16 r0 = (v8i16)__msa_fill_h(left[j + 0]);
v8i16 r1 = (v8i16)__msa_fill_h(left[j + 1]);
v8i16 r2 = (v8i16)__msa_fill_h(left[j + 2]);
v8i16 r3 = (v8i16)__msa_fill_h(left[j + 3]);
ADD4(d, r0, d, r1, d, r2, d, r3, r0, r1, r2, r3);
CLIP_SH4_0_255(r0, r1, r2, r3);
PCKEV_B2_SB(r1, r0, r3, r2, t0, t1);
out0 = __msa_copy_s_d((v2i64)t0, 0);
out1 = __msa_copy_s_d((v2i64)t0, 1);
out2 = __msa_copy_s_d((v2i64)t1, 0);
out3 = __msa_copy_s_d((v2i64)t1, 1);
SD4(out0, out1, out2, out3, dst, BPS);
dst += 4 * BPS;
}
} else {
HorizontalPred8x8(dst, left);
}
} else {
if (top != NULL) {
VerticalPred8x8(dst, top);
} else {
const uint64_t out = 0x8181818181818181ULL;
STORE8x8(out, dst);
}
}
}
static WEBP_INLINE void DCMode8x8(uint8_t* dst, const uint8_t* left,
const uint8_t* top) {
uint64_t out;
v16u8 src;
if (top != NULL && left != NULL) {
const uint64_t left_m = LD(left);
const uint64_t top_m = LD(top);
INSERT_D2_UB(left_m, top_m, src);
CALC_DC8(src, out);
} else if (left != NULL) { // left but no top
const uint64_t left_m = LD(left);
INSERT_D2_UB(left_m, left_m, src);
CALC_DC8(src, out);
} else if (top != NULL) { // top but no left
const uint64_t top_m = LD(top);
INSERT_D2_UB(top_m, top_m, src);
CALC_DC8(src, out);
} else { // no top, no left, nothing.
src = (v16u8)__msa_fill_b(0x80);
out = __msa_copy_s_d((v2i64)src, 0);
}
STORE8x8(out, dst);
}
static void IntraChromaPreds(uint8_t* dst, const uint8_t* left,
const uint8_t* top) {
// U block
DCMode8x8(C8DC8 + dst, left, top);
VerticalPred8x8(C8VE8 + dst, top);
HorizontalPred8x8(C8HE8 + dst, left);
TrueMotion8x8(C8TM8 + dst, left, top);
// V block
dst += 8;
if (top != NULL) top += 8;
if (left != NULL) left += 16;
DCMode8x8(C8DC8 + dst, left, top);
VerticalPred8x8(C8VE8 + dst, top);
HorizontalPred8x8(C8HE8 + dst, left);
TrueMotion8x8(C8TM8 + dst, left, top);
}
//------------------------------------------------------------------------------
// Metric
#define PACK_DOTP_UB4_SW(in0, in1, in2, in3, out0, out1, out2, out3) do { \
v16u8 tmp0, tmp1; \
v8i16 tmp2, tmp3; \
ILVRL_B2_UB(in0, in1, tmp0, tmp1); \
HSUB_UB2_SH(tmp0, tmp1, tmp2, tmp3); \
DOTP_SH2_SW(tmp2, tmp3, tmp2, tmp3, out0, out1); \
ILVRL_B2_UB(in2, in3, tmp0, tmp1); \
HSUB_UB2_SH(tmp0, tmp1, tmp2, tmp3); \
DOTP_SH2_SW(tmp2, tmp3, tmp2, tmp3, out2, out3); \
} while (0)
#define PACK_DPADD_UB4_SW(in0, in1, in2, in3, out0, out1, out2, out3) do { \
v16u8 tmp0, tmp1; \
v8i16 tmp2, tmp3; \
ILVRL_B2_UB(in0, in1, tmp0, tmp1); \
HSUB_UB2_SH(tmp0, tmp1, tmp2, tmp3); \
DPADD_SH2_SW(tmp2, tmp3, tmp2, tmp3, out0, out1); \
ILVRL_B2_UB(in2, in3, tmp0, tmp1); \
HSUB_UB2_SH(tmp0, tmp1, tmp2, tmp3); \
DPADD_SH2_SW(tmp2, tmp3, tmp2, tmp3, out2, out3); \
} while (0)
static int SSE16x16(const uint8_t* a, const uint8_t* b) {
uint32_t sum;
v16u8 src0, src1, src2, src3, src4, src5, src6, src7;
v16u8 ref0, ref1, ref2, ref3, ref4, ref5, ref6, ref7;
v4i32 out0, out1, out2, out3;
LD_UB8(a, BPS, src0, src1, src2, src3, src4, src5, src6, src7);
LD_UB8(b, BPS, ref0, ref1, ref2, ref3, ref4, ref5, ref6, ref7);
PACK_DOTP_UB4_SW(src0, ref0, src1, ref1, out0, out1, out2, out3);
PACK_DPADD_UB4_SW(src2, ref2, src3, ref3, out0, out1, out2, out3);
PACK_DPADD_UB4_SW(src4, ref4, src5, ref5, out0, out1, out2, out3);
PACK_DPADD_UB4_SW(src6, ref6, src7, ref7, out0, out1, out2, out3);
a += 8 * BPS;
b += 8 * BPS;
LD_UB8(a, BPS, src0, src1, src2, src3, src4, src5, src6, src7);
LD_UB8(b, BPS, ref0, ref1, ref2, ref3, ref4, ref5, ref6, ref7);
PACK_DPADD_UB4_SW(src0, ref0, src1, ref1, out0, out1, out2, out3);
PACK_DPADD_UB4_SW(src2, ref2, src3, ref3, out0, out1, out2, out3);
PACK_DPADD_UB4_SW(src4, ref4, src5, ref5, out0, out1, out2, out3);
PACK_DPADD_UB4_SW(src6, ref6, src7, ref7, out0, out1, out2, out3);
out0 += out1;
out2 += out3;
out0 += out2;
sum = HADD_SW_S32(out0);
return sum;
}
static int SSE16x8(const uint8_t* a, const uint8_t* b) {
uint32_t sum;
v16u8 src0, src1, src2, src3, src4, src5, src6, src7;
v16u8 ref0, ref1, ref2, ref3, ref4, ref5, ref6, ref7;
v4i32 out0, out1, out2, out3;
LD_UB8(a, BPS, src0, src1, src2, src3, src4, src5, src6, src7);
LD_UB8(b, BPS, ref0, ref1, ref2, ref3, ref4, ref5, ref6, ref7);
PACK_DOTP_UB4_SW(src0, ref0, src1, ref1, out0, out1, out2, out3);
PACK_DPADD_UB4_SW(src2, ref2, src3, ref3, out0, out1, out2, out3);
PACK_DPADD_UB4_SW(src4, ref4, src5, ref5, out0, out1, out2, out3);
PACK_DPADD_UB4_SW(src6, ref6, src7, ref7, out0, out1, out2, out3);
out0 += out1;
out2 += out3;
out0 += out2;
sum = HADD_SW_S32(out0);
return sum;
}
static int SSE8x8(const uint8_t* a, const uint8_t* b) {
uint32_t sum;
v16u8 src0, src1, src2, src3, src4, src5, src6, src7;
v16u8 ref0, ref1, ref2, ref3, ref4, ref5, ref6, ref7;
v16u8 t0, t1, t2, t3;
v4i32 out0, out1, out2, out3;
LD_UB8(a, BPS, src0, src1, src2, src3, src4, src5, src6, src7);
LD_UB8(b, BPS, ref0, ref1, ref2, ref3, ref4, ref5, ref6, ref7);
ILVR_B4_UB(src0, src1, src2, src3, ref0, ref1, ref2, ref3, t0, t1, t2, t3);
PACK_DOTP_UB4_SW(t0, t2, t1, t3, out0, out1, out2, out3);
ILVR_B4_UB(src4, src5, src6, src7, ref4, ref5, ref6, ref7, t0, t1, t2, t3);
PACK_DPADD_UB4_SW(t0, t2, t1, t3, out0, out1, out2, out3);
out0 += out1;
out2 += out3;
out0 += out2;
sum = HADD_SW_S32(out0);
return sum;
}
static int SSE4x4(const uint8_t* a, const uint8_t* b) {
uint32_t sum = 0;
uint32_t src0, src1, src2, src3, ref0, ref1, ref2, ref3;
v16u8 src, ref, tmp0, tmp1;
v8i16 diff0, diff1;
v4i32 out0, out1;
LW4(a, BPS, src0, src1, src2, src3);
LW4(b, BPS, ref0, ref1, ref2, ref3);
INSERT_W4_UB(src0, src1, src2, src3, src);
INSERT_W4_UB(ref0, ref1, ref2, ref3, ref);
ILVRL_B2_UB(src, ref, tmp0, tmp1);
HSUB_UB2_SH(tmp0, tmp1, diff0, diff1);
DOTP_SH2_SW(diff0, diff1, diff0, diff1, out0, out1);
out0 += out1;
sum = HADD_SW_S32(out0);
return sum;
}
//------------------------------------------------------------------------------
// Quantization
static int QuantizeBlock(int16_t in[16], int16_t out[16],
const VP8Matrix* const mtx) {
int sum;
v8i16 in0, in1, sh0, sh1, out0, out1;
v8i16 tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, sign0, sign1;
v4i32 s0, s1, s2, s3, b0, b1, b2, b3, t0, t1, t2, t3;
const v8i16 zero = { 0 };
const v8i16 zigzag0 = { 0, 1, 4, 8, 5, 2, 3, 6 };
const v8i16 zigzag1 = { 9, 12, 13, 10, 7, 11, 14, 15 };
const v8i16 maxlevel = __msa_fill_h(MAX_LEVEL);
LD_SH2(&in[0], 8, in0, in1);
LD_SH2(&mtx->sharpen_[0], 8, sh0, sh1);
tmp4 = __msa_add_a_h(in0, zero);
tmp5 = __msa_add_a_h(in1, zero);
ILVRL_H2_SH(sh0, tmp4, tmp0, tmp1);
ILVRL_H2_SH(sh1, tmp5, tmp2, tmp3);
HADD_SH4_SW(tmp0, tmp1, tmp2, tmp3, s0, s1, s2, s3);
sign0 = (in0 < zero);
sign1 = (in1 < zero); // sign
LD_SH2(&mtx->iq_[0], 8, tmp0, tmp1); // iq
ILVRL_H2_SW(zero, tmp0, t0, t1);
ILVRL_H2_SW(zero, tmp1, t2, t3);
LD_SW4(&mtx->bias_[0], 4, b0, b1, b2, b3); // bias
MUL4(t0, s0, t1, s1, t2, s2, t3, s3, t0, t1, t2, t3);
ADD4(b0, t0, b1, t1, b2, t2, b3, t3, b0, b1, b2, b3);
SRAI_W4_SW(b0, b1, b2, b3, 17);
PCKEV_H2_SH(b1, b0, b3, b2, tmp2, tmp3);
tmp0 = (tmp2 > maxlevel);
tmp1 = (tmp3 > maxlevel);
tmp2 = (v8i16)__msa_bmnz_v((v16u8)tmp2, (v16u8)maxlevel, (v16u8)tmp0);
tmp3 = (v8i16)__msa_bmnz_v((v16u8)tmp3, (v16u8)maxlevel, (v16u8)tmp1);
SUB2(0, tmp2, 0, tmp3, tmp0, tmp1);
tmp2 = (v8i16)__msa_bmnz_v((v16u8)tmp2, (v16u8)tmp0, (v16u8)sign0);
tmp3 = (v8i16)__msa_bmnz_v((v16u8)tmp3, (v16u8)tmp1, (v16u8)sign1);
LD_SW4(&mtx->zthresh_[0], 4, t0, t1, t2, t3); // zthresh
t0 = (s0 > t0);
t1 = (s1 > t1);
t2 = (s2 > t2);
t3 = (s3 > t3);
PCKEV_H2_SH(t1, t0, t3, t2, tmp0, tmp1);
tmp4 = (v8i16)__msa_bmnz_v((v16u8)zero, (v16u8)tmp2, (v16u8)tmp0);
tmp5 = (v8i16)__msa_bmnz_v((v16u8)zero, (v16u8)tmp3, (v16u8)tmp1);
LD_SH2(&mtx->q_[0], 8, tmp0, tmp1);
MUL2(tmp4, tmp0, tmp5, tmp1, in0, in1);
VSHF_H2_SH(tmp4, tmp5, tmp4, tmp5, zigzag0, zigzag1, out0, out1);
ST_SH2(in0, in1, &in[0], 8);
ST_SH2(out0, out1, &out[0], 8);
out0 = __msa_add_a_h(out0, out1);
sum = HADD_SH_S32(out0);
return (sum > 0);
}
static int Quantize2Blocks(int16_t in[32], int16_t out[32],
const VP8Matrix* const mtx) {
int nz;
nz = VP8EncQuantizeBlock(in + 0 * 16, out + 0 * 16, mtx) << 0;
nz |= VP8EncQuantizeBlock(in + 1 * 16, out + 1 * 16, mtx) << 1;
return nz;
}
//------------------------------------------------------------------------------
// Entry point
extern void VP8EncDspInitMSA(void);
WEBP_TSAN_IGNORE_FUNCTION void VP8EncDspInitMSA(void) {
VP8ITransform = ITransform;
VP8FTransform = FTransform;
VP8FTransformWHT = FTransformWHT;
VP8TDisto4x4 = Disto4x4;
VP8TDisto16x16 = Disto16x16;
VP8CollectHistogram = CollectHistogram;
VP8EncPredLuma4 = Intra4Preds;
VP8EncPredLuma16 = Intra16Preds;
VP8EncPredChroma8 = IntraChromaPreds;
VP8SSE16x16 = SSE16x16;
VP8SSE16x8 = SSE16x8;
VP8SSE8x8 = SSE8x8;
VP8SSE4x4 = SSE4x4;
VP8EncQuantizeBlock = QuantizeBlock;
VP8EncQuantize2Blocks = Quantize2Blocks;
VP8EncQuantizeBlockWHT = QuantizeBlock;
}
#else // !WEBP_USE_MSA
WEBP_DSP_INIT_STUB(VP8EncDspInitMSA)
#endif // WEBP_USE_MSA

42
thirdparty/libwebp/dsp/enc_neon.c vendored
View file

@ -18,7 +18,7 @@
#include <assert.h>
#include "./neon.h"
#include "../enc/vp8enci.h"
#include "../enc/vp8i_enc.h"
//------------------------------------------------------------------------------
// Transforms (Paragraph 14.4)
@ -746,9 +746,14 @@ static WEBP_INLINE void AccumulateSSE16(const uint8_t* const a,
const uint8x16_t a0 = vld1q_u8(a);
const uint8x16_t b0 = vld1q_u8(b);
const uint8x16_t abs_diff = vabdq_u8(a0, b0);
uint16x8_t prod = vmull_u8(vget_low_u8(abs_diff), vget_low_u8(abs_diff));
prod = vmlal_u8(prod, vget_high_u8(abs_diff), vget_high_u8(abs_diff));
*sum = vpadalq_u16(*sum, prod); // pair-wise add and accumulate
const uint16x8_t prod1 = vmull_u8(vget_low_u8(abs_diff),
vget_low_u8(abs_diff));
const uint16x8_t prod2 = vmull_u8(vget_high_u8(abs_diff),
vget_high_u8(abs_diff));
/* pair-wise adds and widen */
const uint32x4_t sum1 = vpaddlq_u16(prod1);
const uint32x4_t sum2 = vpaddlq_u16(prod2);
*sum = vaddq_u32(*sum, vaddq_u32(sum1, sum2));
}
// Horizontal sum of all four uint32_t values in 'sum'.
@ -758,7 +763,7 @@ static int SumToInt(uint32x4_t sum) {
return (int)sum3;
}
static int SSE16x16(const uint8_t* a, const uint8_t* b) {
static int SSE16x16_NEON(const uint8_t* a, const uint8_t* b) {
uint32x4_t sum = vdupq_n_u32(0);
int y;
for (y = 0; y < 16; ++y) {
@ -767,7 +772,7 @@ static int SSE16x16(const uint8_t* a, const uint8_t* b) {
return SumToInt(sum);
}
static int SSE16x8(const uint8_t* a, const uint8_t* b) {
static int SSE16x8_NEON(const uint8_t* a, const uint8_t* b) {
uint32x4_t sum = vdupq_n_u32(0);
int y;
for (y = 0; y < 8; ++y) {
@ -776,7 +781,7 @@ static int SSE16x8(const uint8_t* a, const uint8_t* b) {
return SumToInt(sum);
}
static int SSE8x8(const uint8_t* a, const uint8_t* b) {
static int SSE8x8_NEON(const uint8_t* a, const uint8_t* b) {
uint32x4_t sum = vdupq_n_u32(0);
int y;
for (y = 0; y < 8; ++y) {
@ -789,13 +794,18 @@ static int SSE8x8(const uint8_t* a, const uint8_t* b) {
return SumToInt(sum);
}
static int SSE4x4(const uint8_t* a, const uint8_t* b) {
static int SSE4x4_NEON(const uint8_t* a, const uint8_t* b) {
const uint8x16_t a0 = Load4x4(a);
const uint8x16_t b0 = Load4x4(b);
const uint8x16_t abs_diff = vabdq_u8(a0, b0);
uint16x8_t prod = vmull_u8(vget_low_u8(abs_diff), vget_low_u8(abs_diff));
prod = vmlal_u8(prod, vget_high_u8(abs_diff), vget_high_u8(abs_diff));
return SumToInt(vpaddlq_u16(prod));
const uint16x8_t prod1 = vmull_u8(vget_low_u8(abs_diff),
vget_low_u8(abs_diff));
const uint16x8_t prod2 = vmull_u8(vget_high_u8(abs_diff),
vget_high_u8(abs_diff));
/* pair-wise adds and widen */
const uint32x4_t sum1 = vpaddlq_u16(prod1);
const uint32x4_t sum2 = vpaddlq_u16(prod2);
return SumToInt(vaddq_u32(sum1, sum2));
}
//------------------------------------------------------------------------------
@ -903,10 +913,12 @@ WEBP_TSAN_IGNORE_FUNCTION void VP8EncDspInitNEON(void) {
VP8TDisto4x4 = Disto4x4;
VP8TDisto16x16 = Disto16x16;
VP8CollectHistogram = CollectHistogram;
VP8SSE16x16 = SSE16x16;
VP8SSE16x8 = SSE16x8;
VP8SSE8x8 = SSE8x8;
VP8SSE4x4 = SSE4x4;
VP8SSE16x16 = SSE16x16_NEON;
VP8SSE16x8 = SSE16x8_NEON;
VP8SSE8x8 = SSE8x8_NEON;
VP8SSE4x4 = SSE4x4_NEON;
#if !defined(WORK_AROUND_GCC)
VP8EncQuantizeBlock = QuantizeBlock;
VP8EncQuantize2Blocks = Quantize2Blocks;

182
thirdparty/libwebp/dsp/enc_sse2.c vendored
View file

@ -14,12 +14,13 @@
#include "./dsp.h"
#if defined(WEBP_USE_SSE2)
#include <assert.h>
#include <stdlib.h> // for abs()
#include <emmintrin.h>
#include "./common_sse2.h"
#include "../enc/cost.h"
#include "../enc/vp8enci.h"
#include "../enc/cost_enc.h"
#include "../enc/vp8i_enc.h"
//------------------------------------------------------------------------------
// Transforms (Paragraph 14.4)
@ -139,7 +140,7 @@ static void ITransform(const uint8_t* ref, const int16_t* in, uint8_t* dst,
// Transpose the two 4x4.
VP8Transpose_2_4x4_16b(&shifted0, &shifted1, &shifted2, &shifted3, &T0, &T1,
&T2, &T3);
&T2, &T3);
}
// Add inverse transform to 'ref' and store.
@ -250,25 +251,11 @@ static void FTransformPass2(const __m128i* const v01, const __m128i* const v32,
const __m128i k51000 = _mm_set1_epi32(51000);
// Same operations are done on the (0,3) and (1,2) pairs.
// a0 = v0 + v3
// a1 = v1 + v2
// a3 = v0 - v3
// a2 = v1 - v2
const __m128i a01 = _mm_add_epi16(*v01, *v32);
const __m128i a32 = _mm_sub_epi16(*v01, *v32);
const __m128i a11 = _mm_unpackhi_epi64(a01, a01);
const __m128i a22 = _mm_unpackhi_epi64(a32, a32);
const __m128i a01_plus_7 = _mm_add_epi16(a01, seven);
// d0 = (a0 + a1 + 7) >> 4;
// d2 = (a0 - a1 + 7) >> 4;
const __m128i c0 = _mm_add_epi16(a01_plus_7, a11);
const __m128i c2 = _mm_sub_epi16(a01_plus_7, a11);
const __m128i d0 = _mm_srai_epi16(c0, 4);
const __m128i d2 = _mm_srai_epi16(c2, 4);
// f1 = ((b3 * 5352 + b2 * 2217 + 12000) >> 16)
// f3 = ((b3 * 2217 - b2 * 5352 + 51000) >> 16)
const __m128i b23 = _mm_unpacklo_epi16(a22, a32);
const __m128i c1 = _mm_madd_epi16(b23, k5352_2217);
const __m128i c3 = _mm_madd_epi16(b23, k2217_5352);
@ -276,14 +263,28 @@ static void FTransformPass2(const __m128i* const v01, const __m128i* const v32,
const __m128i d3 = _mm_add_epi32(c3, k51000);
const __m128i e1 = _mm_srai_epi32(d1, 16);
const __m128i e3 = _mm_srai_epi32(d3, 16);
// f1 = ((b3 * 5352 + b2 * 2217 + 12000) >> 16)
// f3 = ((b3 * 2217 - b2 * 5352 + 51000) >> 16)
const __m128i f1 = _mm_packs_epi32(e1, e1);
const __m128i f3 = _mm_packs_epi32(e3, e3);
// f1 = f1 + (a3 != 0);
// g1 = f1 + (a3 != 0);
// The compare will return (0xffff, 0) for (==0, !=0). To turn that into the
// desired (0, 1), we add one earlier through k12000_plus_one.
// -> f1 = f1 + 1 - (a3 == 0)
// -> g1 = f1 + 1 - (a3 == 0)
const __m128i g1 = _mm_add_epi16(f1, _mm_cmpeq_epi16(a32, zero));
// a0 = v0 + v3
// a1 = v1 + v2
const __m128i a01 = _mm_add_epi16(*v01, *v32);
const __m128i a01_plus_7 = _mm_add_epi16(a01, seven);
const __m128i a11 = _mm_unpackhi_epi64(a01, a01);
const __m128i c0 = _mm_add_epi16(a01_plus_7, a11);
const __m128i c2 = _mm_sub_epi16(a01_plus_7, a11);
// d0 = (a0 + a1 + 7) >> 4;
// d2 = (a0 - a1 + 7) >> 4;
const __m128i d0 = _mm_srai_epi16(c0, 4);
const __m128i d2 = _mm_srai_epi16(c2, 4);
const __m128i d0_g1 = _mm_unpacklo_epi64(d0, g1);
const __m128i d2_f3 = _mm_unpacklo_epi64(d2, f3);
_mm_storeu_si128((__m128i*)&out[0], d0_g1);
@ -1045,6 +1046,37 @@ static int SSE4x4(const uint8_t* a, const uint8_t* b) {
return (tmp[3] + tmp[2] + tmp[1] + tmp[0]);
}
//------------------------------------------------------------------------------
static void Mean16x4(const uint8_t* ref, uint32_t dc[4]) {
const __m128i mask = _mm_set1_epi16(0x00ff);
const __m128i a0 = _mm_loadu_si128((const __m128i*)&ref[BPS * 0]);
const __m128i a1 = _mm_loadu_si128((const __m128i*)&ref[BPS * 1]);
const __m128i a2 = _mm_loadu_si128((const __m128i*)&ref[BPS * 2]);
const __m128i a3 = _mm_loadu_si128((const __m128i*)&ref[BPS * 3]);
const __m128i b0 = _mm_srli_epi16(a0, 8); // hi byte
const __m128i b1 = _mm_srli_epi16(a1, 8);
const __m128i b2 = _mm_srli_epi16(a2, 8);
const __m128i b3 = _mm_srli_epi16(a3, 8);
const __m128i c0 = _mm_and_si128(a0, mask); // lo byte
const __m128i c1 = _mm_and_si128(a1, mask);
const __m128i c2 = _mm_and_si128(a2, mask);
const __m128i c3 = _mm_and_si128(a3, mask);
const __m128i d0 = _mm_add_epi32(b0, c0);
const __m128i d1 = _mm_add_epi32(b1, c1);
const __m128i d2 = _mm_add_epi32(b2, c2);
const __m128i d3 = _mm_add_epi32(b3, c3);
const __m128i e0 = _mm_add_epi32(d0, d1);
const __m128i e1 = _mm_add_epi32(d2, d3);
const __m128i f0 = _mm_add_epi32(e0, e1);
uint16_t tmp[8];
_mm_storeu_si128((__m128i*)tmp, f0);
dc[0] = tmp[0] + tmp[1];
dc[1] = tmp[2] + tmp[3];
dc[2] = tmp[4] + tmp[5];
dc[3] = tmp[6] + tmp[7];
}
//------------------------------------------------------------------------------
// Texture distortion
//
@ -1331,10 +1363,122 @@ WEBP_TSAN_IGNORE_FUNCTION void VP8EncDspInitSSE2(void) {
VP8SSE4x4 = SSE4x4;
VP8TDisto4x4 = Disto4x4;
VP8TDisto16x16 = Disto16x16;
VP8Mean16x4 = Mean16x4;
}
//------------------------------------------------------------------------------
// SSIM / PSNR entry point (TODO(skal): move to its own file later)
static uint32_t AccumulateSSE_SSE2(const uint8_t* src1,
const uint8_t* src2, int len) {
int i = 0;
uint32_t sse2 = 0;
if (len >= 16) {
const int limit = len - 32;
int32_t tmp[4];
__m128i sum1;
__m128i sum = _mm_setzero_si128();
__m128i a0 = _mm_loadu_si128((const __m128i*)&src1[i]);
__m128i b0 = _mm_loadu_si128((const __m128i*)&src2[i]);
i += 16;
while (i <= limit) {
const __m128i a1 = _mm_loadu_si128((const __m128i*)&src1[i]);
const __m128i b1 = _mm_loadu_si128((const __m128i*)&src2[i]);
__m128i sum2;
i += 16;
SubtractAndAccumulate(a0, b0, &sum1);
sum = _mm_add_epi32(sum, sum1);
a0 = _mm_loadu_si128((const __m128i*)&src1[i]);
b0 = _mm_loadu_si128((const __m128i*)&src2[i]);
i += 16;
SubtractAndAccumulate(a1, b1, &sum2);
sum = _mm_add_epi32(sum, sum2);
}
SubtractAndAccumulate(a0, b0, &sum1);
sum = _mm_add_epi32(sum, sum1);
_mm_storeu_si128((__m128i*)tmp, sum);
sse2 += (tmp[3] + tmp[2] + tmp[1] + tmp[0]);
}
for (; i < len; ++i) {
const int32_t diff = src1[i] - src2[i];
sse2 += diff * diff;
}
return sse2;
}
static uint32_t HorizontalAdd16b(const __m128i* const m) {
uint16_t tmp[8];
const __m128i a = _mm_srli_si128(*m, 8);
const __m128i b = _mm_add_epi16(*m, a);
_mm_storeu_si128((__m128i*)tmp, b);
return (uint32_t)tmp[3] + tmp[2] + tmp[1] + tmp[0];
}
static uint32_t HorizontalAdd32b(const __m128i* const m) {
const __m128i a = _mm_srli_si128(*m, 8);
const __m128i b = _mm_add_epi32(*m, a);
const __m128i c = _mm_add_epi32(b, _mm_srli_si128(b, 4));
return (uint32_t)_mm_cvtsi128_si32(c);
}
static const uint16_t kWeight[] = { 1, 2, 3, 4, 3, 2, 1, 0 };
#define ACCUMULATE_ROW(WEIGHT) do { \
/* compute row weight (Wx * Wy) */ \
const __m128i Wy = _mm_set1_epi16((WEIGHT)); \
const __m128i W = _mm_mullo_epi16(Wx, Wy); \
/* process 8 bytes at a time (7 bytes, actually) */ \
const __m128i a0 = _mm_loadl_epi64((const __m128i*)src1); \
const __m128i b0 = _mm_loadl_epi64((const __m128i*)src2); \
/* convert to 16b and multiply by weight */ \
const __m128i a1 = _mm_unpacklo_epi8(a0, zero); \
const __m128i b1 = _mm_unpacklo_epi8(b0, zero); \
const __m128i wa1 = _mm_mullo_epi16(a1, W); \
const __m128i wb1 = _mm_mullo_epi16(b1, W); \
/* accumulate */ \
xm = _mm_add_epi16(xm, wa1); \
ym = _mm_add_epi16(ym, wb1); \
xxm = _mm_add_epi32(xxm, _mm_madd_epi16(a1, wa1)); \
xym = _mm_add_epi32(xym, _mm_madd_epi16(a1, wb1)); \
yym = _mm_add_epi32(yym, _mm_madd_epi16(b1, wb1)); \
src1 += stride1; \
src2 += stride2; \
} while (0)
static double SSIMGet_SSE2(const uint8_t* src1, int stride1,
const uint8_t* src2, int stride2) {
VP8DistoStats stats;
const __m128i zero = _mm_setzero_si128();
__m128i xm = zero, ym = zero; // 16b accums
__m128i xxm = zero, yym = zero, xym = zero; // 32b accum
const __m128i Wx = _mm_loadu_si128((const __m128i*)kWeight);
assert(2 * VP8_SSIM_KERNEL + 1 == 7);
ACCUMULATE_ROW(1);
ACCUMULATE_ROW(2);
ACCUMULATE_ROW(3);
ACCUMULATE_ROW(4);
ACCUMULATE_ROW(3);
ACCUMULATE_ROW(2);
ACCUMULATE_ROW(1);
stats.xm = HorizontalAdd16b(&xm);
stats.ym = HorizontalAdd16b(&ym);
stats.xxm = HorizontalAdd32b(&xxm);
stats.xym = HorizontalAdd32b(&xym);
stats.yym = HorizontalAdd32b(&yym);
return VP8SSIMFromStats(&stats);
}
extern void VP8SSIMDspInitSSE2(void);
WEBP_TSAN_IGNORE_FUNCTION void VP8SSIMDspInitSSE2(void) {
VP8AccumulateSSE = AccumulateSSE_SSE2;
VP8SSIMGet = SSIMGet_SSE2;
}
#else // !WEBP_USE_SSE2
WEBP_DSP_INIT_STUB(VP8EncDspInitSSE2)
WEBP_DSP_INIT_STUB(VP8SSIMDspInitSSE2)
#endif // WEBP_USE_SSE2

2
thirdparty/libwebp/dsp/enc_sse41.c vendored
View file

@ -18,7 +18,7 @@
#include <stdlib.h> // for abs()
#include "./common_sse2.h"
#include "../enc/vp8enci.h"
#include "../enc/vp8i_enc.h"
//------------------------------------------------------------------------------
// Compute susceptibility based on DCT-coeff histograms.

12
thirdparty/libwebp/dsp/filters.c vendored
View file

@ -227,6 +227,8 @@ WebPFilterFunc WebPFilters[WEBP_FILTER_LAST];
WebPUnfilterFunc WebPUnfilters[WEBP_FILTER_LAST];
extern void VP8FiltersInitMIPSdspR2(void);
extern void VP8FiltersInitMSA(void);
extern void VP8FiltersInitNEON(void);
extern void VP8FiltersInitSSE2(void);
static volatile VP8CPUInfo filters_last_cpuinfo_used =
@ -251,10 +253,20 @@ WEBP_TSAN_IGNORE_FUNCTION void VP8FiltersInit(void) {
VP8FiltersInitSSE2();
}
#endif
#if defined(WEBP_USE_NEON)
if (VP8GetCPUInfo(kNEON)) {
VP8FiltersInitNEON();
}
#endif
#if defined(WEBP_USE_MIPS_DSP_R2)
if (VP8GetCPUInfo(kMIPSdspR2)) {
VP8FiltersInitMIPSdspR2();
}
#endif
#if defined(WEBP_USE_MSA)
if (VP8GetCPUInfo(kMSA)) {
VP8FiltersInitMSA();
}
#endif
}
filters_last_cpuinfo_used = VP8GetCPUInfo;

202
thirdparty/libwebp/dsp/filters_msa.c vendored Normal file
View file

@ -0,0 +1,202 @@
// Copyright 2016 Google Inc. All Rights Reserved.
//
// Use of this source code is governed by a BSD-style license
// that can be found in the COPYING file in the root of the source
// tree. An additional intellectual property rights grant can be found
// in the file PATENTS. All contributing project authors may
// be found in the AUTHORS file in the root of the source tree.
// -----------------------------------------------------------------------------
//
// MSA variant of alpha filters
//
// Author: Prashant Patil (prashant.patil@imgtec.com)
#include "./dsp.h"
#if defined(WEBP_USE_MSA)
#include "./msa_macro.h"
#include <assert.h>
static WEBP_INLINE void PredictLineInverse0(const uint8_t* src,
const uint8_t* pred,
uint8_t* dst, int length) {
v16u8 src0, pred0, dst0;
assert(length >= 0);
while (length >= 32) {
v16u8 src1, pred1, dst1;
LD_UB2(src, 16, src0, src1);
LD_UB2(pred, 16, pred0, pred1);
SUB2(src0, pred0, src1, pred1, dst0, dst1);
ST_UB2(dst0, dst1, dst, 16);
src += 32;
pred += 32;
dst += 32;
length -= 32;
}
if (length > 0) {
int i;
if (length >= 16) {
src0 = LD_UB(src);
pred0 = LD_UB(pred);
dst0 = src0 - pred0;
ST_UB(dst0, dst);
src += 16;
pred += 16;
dst += 16;
length -= 16;
}
for (i = 0; i < length; i++) {
dst[i] = src[i] - pred[i];
}
}
}
//------------------------------------------------------------------------------
// Helpful macro.
#define SANITY_CHECK(in, out) \
assert(in != NULL); \
assert(out != NULL); \
assert(width > 0); \
assert(height > 0); \
assert(stride >= width);
//------------------------------------------------------------------------------
// Horrizontal filter
static void HorizontalFilter(const uint8_t* data, int width, int height,
int stride, uint8_t* filtered_data) {
const uint8_t* preds = data;
const uint8_t* in = data;
uint8_t* out = filtered_data;
int row = 1;
SANITY_CHECK(in, out);
// Leftmost pixel is the same as input for topmost scanline.
out[0] = in[0];
PredictLineInverse0(in + 1, preds, out + 1, width - 1);
preds += stride;
in += stride;
out += stride;
// Filter line-by-line.
while (row < height) {
// Leftmost pixel is predicted from above.
PredictLineInverse0(in, preds - stride, out, 1);
PredictLineInverse0(in + 1, preds, out + 1, width - 1);
++row;
preds += stride;
in += stride;
out += stride;
}
}
//------------------------------------------------------------------------------
// Gradient filter
static WEBP_INLINE void PredictLineGradient(const uint8_t* pinput,
const uint8_t* ppred,
uint8_t* poutput, int stride,
int size) {
int w;
const v16i8 zero = { 0 };
while (size >= 16) {
v16u8 pred0, dst0;
v8i16 a0, a1, b0, b1, c0, c1;
const v16u8 tmp0 = LD_UB(ppred - 1);
const v16u8 tmp1 = LD_UB(ppred - stride);
const v16u8 tmp2 = LD_UB(ppred - stride - 1);
const v16u8 src0 = LD_UB(pinput);
ILVRL_B2_SH(zero, tmp0, a0, a1);
ILVRL_B2_SH(zero, tmp1, b0, b1);
ILVRL_B2_SH(zero, tmp2, c0, c1);
ADD2(a0, b0, a1, b1, a0, a1);
SUB2(a0, c0, a1, c1, a0, a1);
CLIP_SH2_0_255(a0, a1);
pred0 = (v16u8)__msa_pckev_b((v16i8)a1, (v16i8)a0);
dst0 = src0 - pred0;
ST_UB(dst0, poutput);
ppred += 16;
pinput += 16;
poutput += 16;
size -= 16;
}
for (w = 0; w < size; ++w) {
const int pred = ppred[w - 1] + ppred[w - stride] - ppred[w - stride - 1];
poutput[w] = pinput[w] - (pred < 0 ? 0 : pred > 255 ? 255 : pred);
}
}
static void GradientFilter(const uint8_t* data, int width, int height,
int stride, uint8_t* filtered_data) {
const uint8_t* in = data;
const uint8_t* preds = data;
uint8_t* out = filtered_data;
int row = 1;
SANITY_CHECK(in, out);
// left prediction for top scan-line
out[0] = in[0];
PredictLineInverse0(in + 1, preds, out + 1, width - 1);
preds += stride;
in += stride;
out += stride;
// Filter line-by-line.
while (row < height) {
out[0] = in[0] - preds[- stride];
PredictLineGradient(preds + 1, in + 1, out + 1, stride, width - 1);
++row;
preds += stride;
in += stride;
out += stride;
}
}
//------------------------------------------------------------------------------
// Vertical filter
static void VerticalFilter(const uint8_t* data, int width, int height,
int stride, uint8_t* filtered_data) {
const uint8_t* in = data;
const uint8_t* preds = data;
uint8_t* out = filtered_data;
int row = 1;
SANITY_CHECK(in, out);
// Very first top-left pixel is copied.
out[0] = in[0];
// Rest of top scan-line is left-predicted.
PredictLineInverse0(in + 1, preds, out + 1, width - 1);
in += stride;
out += stride;
// Filter line-by-line.
while (row < height) {
PredictLineInverse0(in, preds, out, width);
++row;
preds += stride;
in += stride;
out += stride;
}
}
#undef SANITY_CHECK
//------------------------------------------------------------------------------
// Entry point
extern void VP8FiltersInitMSA(void);
WEBP_TSAN_IGNORE_FUNCTION void VP8FiltersInitMSA(void) {
WebPFilters[WEBP_FILTER_HORIZONTAL] = HorizontalFilter;
WebPFilters[WEBP_FILTER_VERTICAL] = VerticalFilter;
WebPFilters[WEBP_FILTER_GRADIENT] = GradientFilter;
}
#else // !WEBP_USE_MSA
WEBP_DSP_INIT_STUB(VP8FiltersInitMSA)
#endif // WEBP_USE_MSA

327
thirdparty/libwebp/dsp/filters_neon.c vendored Normal file
View file

@ -0,0 +1,327 @@
// Copyright 2017 Google Inc. All Rights Reserved.
//
// Use of this source code is governed by a BSD-style license
// that can be found in the COPYING file in the root of the source
// tree. An additional intellectual property rights grant can be found
// in the file PATENTS. All contributing project authors may
// be found in the AUTHORS file in the root of the source tree.
// -----------------------------------------------------------------------------
//
// NEON variant of alpha filters
//
// Author: Skal (pascal.massimino@gmail.com)
#include "./dsp.h"
#if defined(WEBP_USE_NEON)
#include <assert.h>
#include "./neon.h"
//------------------------------------------------------------------------------
// Helpful macros.
# define SANITY_CHECK(in, out) \
assert(in != NULL); \
assert(out != NULL); \
assert(width > 0); \
assert(height > 0); \
assert(stride >= width); \
assert(row >= 0 && num_rows > 0 && row + num_rows <= height); \
(void)height; // Silence unused warning.
// load eight u8 and widen to s16
#define U8_TO_S16(A) vreinterpretq_s16_u16(vmovl_u8(A))
#define LOAD_U8_TO_S16(A) U8_TO_S16(vld1_u8(A))
// shift left or right by N byte, inserting zeros
#define SHIFT_RIGHT_N_Q(A, N) vextq_u8((A), zero, (N))
#define SHIFT_LEFT_N_Q(A, N) vextq_u8(zero, (A), (16 - (N)) % 16)
// rotate left by N bytes
#define ROTATE_LEFT_N(A, N) vext_u8((A), (A), (N))
// rotate right by N bytes
#define ROTATE_RIGHT_N(A, N) vext_u8((A), (A), (8 - (N)) % 8)
static void PredictLine_NEON(const uint8_t* src, const uint8_t* pred,
uint8_t* dst, int length) {
int i;
assert(length >= 0);
for (i = 0; i + 16 <= length; i += 16) {
const uint8x16_t A = vld1q_u8(&src[i]);
const uint8x16_t B = vld1q_u8(&pred[i]);
const uint8x16_t C = vsubq_u8(A, B);
vst1q_u8(&dst[i], C);
}
for (; i < length; ++i) dst[i] = src[i] - pred[i];
}
// Special case for left-based prediction (when preds==dst-1 or preds==src-1).
static void PredictLineLeft_NEON(const uint8_t* src, uint8_t* dst, int length) {
PredictLine_NEON(src, src - 1, dst, length);
}
//------------------------------------------------------------------------------
// Horizontal filter.
static WEBP_INLINE void DoHorizontalFilter_NEON(const uint8_t* in,
int width, int height,
int stride,
int row, int num_rows,
uint8_t* out) {
const size_t start_offset = row * stride;
const int last_row = row + num_rows;
SANITY_CHECK(in, out);
in += start_offset;
out += start_offset;
if (row == 0) {
// Leftmost pixel is the same as input for topmost scanline.
out[0] = in[0];
PredictLineLeft_NEON(in + 1, out + 1, width - 1);
row = 1;
in += stride;
out += stride;
}
// Filter line-by-line.
while (row < last_row) {
// Leftmost pixel is predicted from above.
out[0] = in[0] - in[-stride];
PredictLineLeft_NEON(in + 1, out + 1, width - 1);
++row;
in += stride;
out += stride;
}
}
static void HorizontalFilter_NEON(const uint8_t* data, int width, int height,
int stride, uint8_t* filtered_data) {
DoHorizontalFilter_NEON(data, width, height, stride, 0, height,
filtered_data);
}
//------------------------------------------------------------------------------
// Vertical filter.
static WEBP_INLINE void DoVerticalFilter_NEON(const uint8_t* in,
int width, int height, int stride,
int row, int num_rows,
uint8_t* out) {
const size_t start_offset = row * stride;
const int last_row = row + num_rows;
SANITY_CHECK(in, out);
in += start_offset;
out += start_offset;
if (row == 0) {
// Very first top-left pixel is copied.
out[0] = in[0];
// Rest of top scan-line is left-predicted.
PredictLineLeft_NEON(in + 1, out + 1, width - 1);
row = 1;
in += stride;
out += stride;
}
// Filter line-by-line.
while (row < last_row) {
PredictLine_NEON(in, in - stride, out, width);
++row;
in += stride;
out += stride;
}
}
static void VerticalFilter_NEON(const uint8_t* data, int width, int height,
int stride, uint8_t* filtered_data) {
DoVerticalFilter_NEON(data, width, height, stride, 0, height,
filtered_data);
}
//------------------------------------------------------------------------------
// Gradient filter.
static WEBP_INLINE int GradientPredictor_C(uint8_t a, uint8_t b, uint8_t c) {
const int g = a + b - c;
return ((g & ~0xff) == 0) ? g : (g < 0) ? 0 : 255; // clip to 8bit
}
static void GradientPredictDirect_NEON(const uint8_t* const row,
const uint8_t* const top,
uint8_t* const out, int length) {
int i;
for (i = 0; i + 8 <= length; i += 8) {
const uint8x8_t A = vld1_u8(&row[i - 1]);
const uint8x8_t B = vld1_u8(&top[i + 0]);
const int16x8_t C = vreinterpretq_s16_u16(vaddl_u8(A, B));
const int16x8_t D = LOAD_U8_TO_S16(&top[i - 1]);
const uint8x8_t E = vqmovun_s16(vsubq_s16(C, D));
const uint8x8_t F = vld1_u8(&row[i + 0]);
vst1_u8(&out[i], vsub_u8(F, E));
}
for (; i < length; ++i) {
out[i] = row[i] - GradientPredictor_C(row[i - 1], top[i], top[i - 1]);
}
}
static WEBP_INLINE void DoGradientFilter_NEON(const uint8_t* in,
int width, int height,
int stride,
int row, int num_rows,
uint8_t* out) {
const size_t start_offset = row * stride;
const int last_row = row + num_rows;
SANITY_CHECK(in, out);
in += start_offset;
out += start_offset;
// left prediction for top scan-line
if (row == 0) {
out[0] = in[0];
PredictLineLeft_NEON(in + 1, out + 1, width - 1);
row = 1;
in += stride;
out += stride;
}
// Filter line-by-line.
while (row < last_row) {
out[0] = in[0] - in[-stride];
GradientPredictDirect_NEON(in + 1, in + 1 - stride, out + 1, width - 1);
++row;
in += stride;
out += stride;
}
}
static void GradientFilter_NEON(const uint8_t* data, int width, int height,
int stride, uint8_t* filtered_data) {
DoGradientFilter_NEON(data, width, height, stride, 0, height,
filtered_data);
}
#undef SANITY_CHECK
//------------------------------------------------------------------------------
// Inverse transforms
static void HorizontalUnfilter_NEON(const uint8_t* prev, const uint8_t* in,
uint8_t* out, int width) {
int i;
const uint8x16_t zero = vdupq_n_u8(0);
uint8x16_t last;
out[0] = in[0] + (prev == NULL ? 0 : prev[0]);
if (width <= 1) return;
last = vsetq_lane_u8(out[0], zero, 0);
for (i = 1; i + 16 <= width; i += 16) {
const uint8x16_t A0 = vld1q_u8(&in[i]);
const uint8x16_t A1 = vaddq_u8(A0, last);
const uint8x16_t A2 = SHIFT_LEFT_N_Q(A1, 1);
const uint8x16_t A3 = vaddq_u8(A1, A2);
const uint8x16_t A4 = SHIFT_LEFT_N_Q(A3, 2);
const uint8x16_t A5 = vaddq_u8(A3, A4);
const uint8x16_t A6 = SHIFT_LEFT_N_Q(A5, 4);
const uint8x16_t A7 = vaddq_u8(A5, A6);
const uint8x16_t A8 = SHIFT_LEFT_N_Q(A7, 8);
const uint8x16_t A9 = vaddq_u8(A7, A8);
vst1q_u8(&out[i], A9);
last = SHIFT_RIGHT_N_Q(A9, 15);
}
for (; i < width; ++i) out[i] = in[i] + out[i - 1];
}
static void VerticalUnfilter_NEON(const uint8_t* prev, const uint8_t* in,
uint8_t* out, int width) {
if (prev == NULL) {
HorizontalUnfilter_NEON(NULL, in, out, width);
} else {
int i;
assert(width >= 0);
for (i = 0; i + 16 <= width; i += 16) {
const uint8x16_t A = vld1q_u8(&in[i]);
const uint8x16_t B = vld1q_u8(&prev[i]);
const uint8x16_t C = vaddq_u8(A, B);
vst1q_u8(&out[i], C);
}
for (; i < width; ++i) out[i] = in[i] + prev[i];
}
}
// GradientUnfilter_NEON is correct but slower than the C-version,
// at least on ARM64. For armv7, it's a wash.
// So best is to disable it for now, but keep the idea around...
// #define USE_GRADIENT_UNFILTER
#if defined(USE_GRADIENT_UNFILTER)
#define GRAD_PROCESS_LANE(L) do { \
const uint8x8_t tmp1 = ROTATE_RIGHT_N(pred, 1); /* rotate predictor in */ \
const int16x8_t tmp2 = vaddq_s16(BC, U8_TO_S16(tmp1)); \
const uint8x8_t delta = vqmovun_s16(tmp2); \
pred = vadd_u8(D, delta); \
out = vext_u8(out, ROTATE_LEFT_N(pred, (L)), 1); \
} while (0)
static void GradientPredictInverse_NEON(const uint8_t* const in,
const uint8_t* const top,
uint8_t* const row, int length) {
if (length > 0) {
int i;
uint8x8_t pred = vdup_n_u8(row[-1]); // left sample
uint8x8_t out = vdup_n_u8(0);
for (i = 0; i + 8 <= length; i += 8) {
const int16x8_t B = LOAD_U8_TO_S16(&top[i + 0]);
const int16x8_t C = LOAD_U8_TO_S16(&top[i - 1]);
const int16x8_t BC = vsubq_s16(B, C); // unclipped gradient basis B - C
const uint8x8_t D = vld1_u8(&in[i]); // base input
GRAD_PROCESS_LANE(0);
GRAD_PROCESS_LANE(1);
GRAD_PROCESS_LANE(2);
GRAD_PROCESS_LANE(3);
GRAD_PROCESS_LANE(4);
GRAD_PROCESS_LANE(5);
GRAD_PROCESS_LANE(6);
GRAD_PROCESS_LANE(7);
vst1_u8(&row[i], out);
}
for (; i < length; ++i) {
row[i] = in[i] + GradientPredictor_C(row[i - 1], top[i], top[i - 1]);
}
}
}
#undef GRAD_PROCESS_LANE
static void GradientUnfilter_NEON(const uint8_t* prev, const uint8_t* in,
uint8_t* out, int width) {
if (prev == NULL) {
HorizontalUnfilter_NEON(NULL, in, out, width);
} else {
out[0] = in[0] + prev[0]; // predict from above
GradientPredictInverse_NEON(in + 1, prev + 1, out + 1, width - 1);
}
}
#endif // USE_GRADIENT_UNFILTER
//------------------------------------------------------------------------------
// Entry point
extern void VP8FiltersInitNEON(void);
WEBP_TSAN_IGNORE_FUNCTION void VP8FiltersInitNEON(void) {
WebPUnfilters[WEBP_FILTER_HORIZONTAL] = HorizontalUnfilter_NEON;
WebPUnfilters[WEBP_FILTER_VERTICAL] = VerticalUnfilter_NEON;
#if defined(USE_GRADIENT_UNFILTER)
WebPUnfilters[WEBP_FILTER_GRADIENT] = GradientUnfilter_NEON;
#endif
WebPFilters[WEBP_FILTER_HORIZONTAL] = HorizontalFilter_NEON;
WebPFilters[WEBP_FILTER_VERTICAL] = VerticalFilter_NEON;
WebPFilters[WEBP_FILTER_GRADIENT] = GradientFilter_NEON;
}
#else // !WEBP_USE_NEON
WEBP_DSP_INIT_STUB(VP8FiltersInitNEON)
#endif // WEBP_USE_NEON

186
thirdparty/libwebp/dsp/lossless.c vendored
View file

@ -17,20 +17,16 @@
#include <math.h>
#include <stdlib.h>
#include "../dec/vp8li.h"
#include "../utils/endian_inl.h"
#include "../dec/vp8li_dec.h"
#include "../utils/endian_inl_utils.h"
#include "./lossless.h"
#include "./lossless_common.h"
#define MAX_DIFF_COST (1e30f)
//------------------------------------------------------------------------------
// Image transforms.
// In-place sum of each component with mod 256.
static WEBP_INLINE void AddPixelsEq(uint32_t* a, uint32_t b) {
*a = VP8LAddPixels(*a, b);
}
static WEBP_INLINE uint32_t Average2(uint32_t a0, uint32_t a1) {
return (((a0 ^ a1) & 0xfefefefeu) >> 1) + (a0 & a1);
}
@ -171,21 +167,41 @@ static uint32_t Predictor13(uint32_t left, const uint32_t* const top) {
return pred;
}
GENERATE_PREDICTOR_ADD(Predictor0, PredictorAdd0)
static void PredictorAdd1(const uint32_t* in, const uint32_t* upper,
int num_pixels, uint32_t* out) {
int i;
uint32_t left = out[-1];
for (i = 0; i < num_pixels; ++i) {
out[i] = left = VP8LAddPixels(in[i], left);
}
(void)upper;
}
GENERATE_PREDICTOR_ADD(Predictor2, PredictorAdd2)
GENERATE_PREDICTOR_ADD(Predictor3, PredictorAdd3)
GENERATE_PREDICTOR_ADD(Predictor4, PredictorAdd4)
GENERATE_PREDICTOR_ADD(Predictor5, PredictorAdd5)
GENERATE_PREDICTOR_ADD(Predictor6, PredictorAdd6)
GENERATE_PREDICTOR_ADD(Predictor7, PredictorAdd7)
GENERATE_PREDICTOR_ADD(Predictor8, PredictorAdd8)
GENERATE_PREDICTOR_ADD(Predictor9, PredictorAdd9)
GENERATE_PREDICTOR_ADD(Predictor10, PredictorAdd10)
GENERATE_PREDICTOR_ADD(Predictor11, PredictorAdd11)
GENERATE_PREDICTOR_ADD(Predictor12, PredictorAdd12)
GENERATE_PREDICTOR_ADD(Predictor13, PredictorAdd13)
//------------------------------------------------------------------------------
// Inverse prediction.
static void PredictorInverseTransform(const VP8LTransform* const transform,
int y_start, int y_end, uint32_t* data) {
int y_start, int y_end,
const uint32_t* in, uint32_t* out) {
const int width = transform->xsize_;
if (y_start == 0) { // First Row follows the L (mode=1) mode.
int x;
const uint32_t pred0 = Predictor0(data[-1], NULL);
AddPixelsEq(data, pred0);
for (x = 1; x < width; ++x) {
const uint32_t pred1 = Predictor1(data[x - 1], NULL);
AddPixelsEq(data + x, pred1);
}
data += width;
PredictorAdd0(in, NULL, 1, out);
PredictorAdd1(in + 1, NULL, width - 1, out + 1);
in += width;
out += width;
++y_start;
}
@ -193,36 +209,26 @@ static void PredictorInverseTransform(const VP8LTransform* const transform,
int y = y_start;
const int tile_width = 1 << transform->bits_;
const int mask = tile_width - 1;
const int safe_width = width & ~mask;
const int tiles_per_row = VP8LSubSampleSize(width, transform->bits_);
const uint32_t* pred_mode_base =
transform->data_ + (y >> transform->bits_) * tiles_per_row;
while (y < y_end) {
const uint32_t pred2 = Predictor2(data[-1], data - width);
const uint32_t* pred_mode_src = pred_mode_base;
VP8LPredictorFunc pred_func;
int x = 1;
int t = 1;
// First pixel follows the T (mode=2) mode.
AddPixelsEq(data, pred2);
PredictorAdd2(in, out - width, 1, out);
// .. the rest:
while (x < safe_width) {
pred_func = VP8LPredictors[((*pred_mode_src++) >> 8) & 0xf];
for (; t < tile_width; ++t, ++x) {
const uint32_t pred = pred_func(data[x - 1], data + x - width);
AddPixelsEq(data + x, pred);
}
t = 0;
while (x < width) {
const VP8LPredictorAddSubFunc pred_func =
VP8LPredictorsAdd[((*pred_mode_src++) >> 8) & 0xf];
int x_end = (x & ~mask) + tile_width;
if (x_end > width) x_end = width;
pred_func(in + x, out + x - width, x_end - x, out + x);
x = x_end;
}
if (x < width) {
pred_func = VP8LPredictors[((*pred_mode_src++) >> 8) & 0xf];
for (; x < width; ++x) {
const uint32_t pred = pred_func(data[x - 1], data + x - width);
AddPixelsEq(data + x, pred);
}
}
data += width;
in += width;
out += width;
++y;
if ((y & mask) == 0) { // Use the same mask, since tiles are squares.
pred_mode_base += tiles_per_row;
@ -233,21 +239,22 @@ static void PredictorInverseTransform(const VP8LTransform* const transform,
// Add green to blue and red channels (i.e. perform the inverse transform of
// 'subtract green').
void VP8LAddGreenToBlueAndRed_C(uint32_t* data, int num_pixels) {
void VP8LAddGreenToBlueAndRed_C(const uint32_t* src, int num_pixels,
uint32_t* dst) {
int i;
for (i = 0; i < num_pixels; ++i) {
const uint32_t argb = data[i];
const uint32_t argb = src[i];
const uint32_t green = ((argb >> 8) & 0xff);
uint32_t red_blue = (argb & 0x00ff00ffu);
red_blue += (green << 16) | green;
red_blue &= 0x00ff00ffu;
data[i] = (argb & 0xff00ff00u) | red_blue;
dst[i] = (argb & 0xff00ff00u) | red_blue;
}
}
static WEBP_INLINE uint32_t ColorTransformDelta(int8_t color_pred,
int8_t color) {
return (uint32_t)((int)(color_pred) * color) >> 5;
static WEBP_INLINE int ColorTransformDelta(int8_t color_pred,
int8_t color) {
return ((int)color_pred * color) >> 5;
}
static WEBP_INLINE void ColorCodeToMultipliers(uint32_t color_code,
@ -257,27 +264,29 @@ static WEBP_INLINE void ColorCodeToMultipliers(uint32_t color_code,
m->red_to_blue_ = (color_code >> 16) & 0xff;
}
void VP8LTransformColorInverse_C(const VP8LMultipliers* const m, uint32_t* data,
int num_pixels) {
void VP8LTransformColorInverse_C(const VP8LMultipliers* const m,
const uint32_t* src, int num_pixels,
uint32_t* dst) {
int i;
for (i = 0; i < num_pixels; ++i) {
const uint32_t argb = data[i];
const uint32_t argb = src[i];
const uint32_t green = argb >> 8;
const uint32_t red = argb >> 16;
uint32_t new_red = red;
uint32_t new_blue = argb;
int new_red = red;
int new_blue = argb;
new_red += ColorTransformDelta(m->green_to_red_, green);
new_red &= 0xff;
new_blue += ColorTransformDelta(m->green_to_blue_, green);
new_blue += ColorTransformDelta(m->red_to_blue_, new_red);
new_blue &= 0xff;
data[i] = (argb & 0xff00ff00u) | (new_red << 16) | (new_blue);
dst[i] = (argb & 0xff00ff00u) | (new_red << 16) | (new_blue);
}
}
// Color space inverse transform.
static void ColorSpaceInverseTransform(const VP8LTransform* const transform,
int y_start, int y_end, uint32_t* data) {
int y_start, int y_end,
const uint32_t* src, uint32_t* dst) {
const int width = transform->xsize_;
const int tile_width = 1 << transform->bits_;
const int mask = tile_width - 1;
@ -291,17 +300,19 @@ static void ColorSpaceInverseTransform(const VP8LTransform* const transform,
while (y < y_end) {
const uint32_t* pred = pred_row;
VP8LMultipliers m = { 0, 0, 0 };
const uint32_t* const data_safe_end = data + safe_width;
const uint32_t* const data_end = data + width;
while (data < data_safe_end) {
const uint32_t* const src_safe_end = src + safe_width;
const uint32_t* const src_end = src + width;
while (src < src_safe_end) {
ColorCodeToMultipliers(*pred++, &m);
VP8LTransformColorInverse(&m, data, tile_width);
data += tile_width;
VP8LTransformColorInverse(&m, src, tile_width, dst);
src += tile_width;
dst += tile_width;
}
if (data < data_end) { // Left-overs using C-version.
if (src < src_end) { // Left-overs using C-version.
ColorCodeToMultipliers(*pred++, &m);
VP8LTransformColorInverse(&m, data, remaining_width);
data += remaining_width;
VP8LTransformColorInverse(&m, src, remaining_width, dst);
src += remaining_width;
dst += remaining_width;
}
++y;
if ((y & mask) == 0) pred_row += tiles_per_row;
@ -366,10 +377,10 @@ void VP8LInverseTransform(const VP8LTransform* const transform,
assert(row_end <= transform->ysize_);
switch (transform->type_) {
case SUBTRACT_GREEN:
VP8LAddGreenToBlueAndRed(out, (row_end - row_start) * width);
VP8LAddGreenToBlueAndRed(in, (row_end - row_start) * width, out);
break;
case PREDICTOR_TRANSFORM:
PredictorInverseTransform(transform, row_start, row_end, out);
PredictorInverseTransform(transform, row_start, row_end, in, out);
if (row_end != transform->ysize_) {
// The last predicted row in this iteration will be the top-pred row
// for the first row in next iteration.
@ -378,7 +389,7 @@ void VP8LInverseTransform(const VP8LTransform* const transform,
}
break;
case CROSS_COLOR_TRANSFORM:
ColorSpaceInverseTransform(transform, row_start, row_end, out);
ColorSpaceInverseTransform(transform, row_start, row_end, in, out);
break;
case COLOR_INDEXING_TRANSFORM:
if (in == out && transform->bits_ > 0) {
@ -555,10 +566,15 @@ void VP8LConvertFromBGRA(const uint32_t* const in_data, int num_pixels,
//------------------------------------------------------------------------------
VP8LProcessBlueAndRedFunc VP8LAddGreenToBlueAndRed;
VP8LProcessDecBlueAndRedFunc VP8LAddGreenToBlueAndRed;
VP8LPredictorAddSubFunc VP8LPredictorsAdd[16];
VP8LPredictorFunc VP8LPredictors[16];
VP8LTransformColorFunc VP8LTransformColorInverse;
// exposed plain-C implementations
VP8LPredictorAddSubFunc VP8LPredictorsAdd_C[16];
VP8LPredictorFunc VP8LPredictors_C[16];
VP8LTransformColorInverseFunc VP8LTransformColorInverse;
VP8LConvertFunc VP8LConvertBGRAToRGB;
VP8LConvertFunc VP8LConvertBGRAToRGBA;
@ -572,29 +588,37 @@ VP8LMapAlphaFunc VP8LMapColor8b;
extern void VP8LDspInitSSE2(void);
extern void VP8LDspInitNEON(void);
extern void VP8LDspInitMIPSdspR2(void);
extern void VP8LDspInitMSA(void);
static volatile VP8CPUInfo lossless_last_cpuinfo_used =
(VP8CPUInfo)&lossless_last_cpuinfo_used;
#define COPY_PREDICTOR_ARRAY(IN, OUT) do { \
(OUT)[0] = IN##0; \
(OUT)[1] = IN##1; \
(OUT)[2] = IN##2; \
(OUT)[3] = IN##3; \
(OUT)[4] = IN##4; \
(OUT)[5] = IN##5; \
(OUT)[6] = IN##6; \
(OUT)[7] = IN##7; \
(OUT)[8] = IN##8; \
(OUT)[9] = IN##9; \
(OUT)[10] = IN##10; \
(OUT)[11] = IN##11; \
(OUT)[12] = IN##12; \
(OUT)[13] = IN##13; \
(OUT)[14] = IN##0; /* <- padding security sentinels*/ \
(OUT)[15] = IN##0; \
} while (0);
WEBP_TSAN_IGNORE_FUNCTION void VP8LDspInit(void) {
if (lossless_last_cpuinfo_used == VP8GetCPUInfo) return;
VP8LPredictors[0] = Predictor0;
VP8LPredictors[1] = Predictor1;
VP8LPredictors[2] = Predictor2;
VP8LPredictors[3] = Predictor3;
VP8LPredictors[4] = Predictor4;
VP8LPredictors[5] = Predictor5;
VP8LPredictors[6] = Predictor6;
VP8LPredictors[7] = Predictor7;
VP8LPredictors[8] = Predictor8;
VP8LPredictors[9] = Predictor9;
VP8LPredictors[10] = Predictor10;
VP8LPredictors[11] = Predictor11;
VP8LPredictors[12] = Predictor12;
VP8LPredictors[13] = Predictor13;
VP8LPredictors[14] = Predictor0; // <- padding security sentinels
VP8LPredictors[15] = Predictor0;
COPY_PREDICTOR_ARRAY(Predictor, VP8LPredictors)
COPY_PREDICTOR_ARRAY(Predictor, VP8LPredictors_C)
COPY_PREDICTOR_ARRAY(PredictorAdd, VP8LPredictorsAdd)
COPY_PREDICTOR_ARRAY(PredictorAdd, VP8LPredictorsAdd_C)
VP8LAddGreenToBlueAndRed = VP8LAddGreenToBlueAndRed_C;
@ -625,9 +649,15 @@ WEBP_TSAN_IGNORE_FUNCTION void VP8LDspInit(void) {
if (VP8GetCPUInfo(kMIPSdspR2)) {
VP8LDspInitMIPSdspR2();
}
#endif
#if defined(WEBP_USE_MSA)
if (VP8GetCPUInfo(kMSA)) {
VP8LDspInitMSA();
}
#endif
}
lossless_last_cpuinfo_used = VP8GetCPUInfo;
}
#undef COPY_PREDICTOR_ARRAY
//------------------------------------------------------------------------------

223
thirdparty/libwebp/dsp/lossless.h vendored
View file

@ -18,7 +18,7 @@
#include "../webp/types.h"
#include "../webp/decode.h"
#include "../enc/histogram.h"
#include "../enc/histogram_enc.h"
#include "../utils/utils.h"
#ifdef __cplusplus
@ -26,7 +26,7 @@ extern "C" {
#endif
#ifdef WEBP_EXPERIMENTAL_FEATURES
#include "../enc/delta_palettization.h"
#include "../enc/delta_palettization_enc.h"
#endif // WEBP_EXPERIMENTAL_FEATURES
//------------------------------------------------------------------------------
@ -34,9 +34,17 @@ extern "C" {
typedef uint32_t (*VP8LPredictorFunc)(uint32_t left, const uint32_t* const top);
extern VP8LPredictorFunc VP8LPredictors[16];
extern VP8LPredictorFunc VP8LPredictors_C[16];
// These Add/Sub function expects upper[-1] and out[-1] to be readable.
typedef void (*VP8LPredictorAddSubFunc)(const uint32_t* in,
const uint32_t* upper, int num_pixels,
uint32_t* out);
extern VP8LPredictorAddSubFunc VP8LPredictorsAdd[16];
extern VP8LPredictorAddSubFunc VP8LPredictorsAdd_C[16];
typedef void (*VP8LProcessBlueAndRedFunc)(uint32_t* argb_data, int num_pixels);
extern VP8LProcessBlueAndRedFunc VP8LAddGreenToBlueAndRed;
typedef void (*VP8LProcessDecBlueAndRedFunc)(const uint32_t* src,
int num_pixels, uint32_t* dst);
extern VP8LProcessDecBlueAndRedFunc VP8LAddGreenToBlueAndRed;
typedef struct {
// Note: the members are uint8_t, so that any negative values are
@ -45,9 +53,10 @@ typedef struct {
uint8_t green_to_blue_;
uint8_t red_to_blue_;
} VP8LMultipliers;
typedef void (*VP8LTransformColorFunc)(const VP8LMultipliers* const m,
uint32_t* argb_data, int num_pixels);
extern VP8LTransformColorFunc VP8LTransformColorInverse;
typedef void (*VP8LTransformColorInverseFunc)(const VP8LMultipliers* const m,
const uint32_t* src,
int num_pixels, uint32_t* dst);
extern VP8LTransformColorInverseFunc VP8LTransformColorInverse;
struct VP8LTransform; // Defined in dec/vp8li.h.
@ -72,23 +81,6 @@ extern VP8LConvertFunc VP8LConvertBGRAToBGR;
void VP8LConvertFromBGRA(const uint32_t* const in_data, int num_pixels,
WEBP_CSP_MODE out_colorspace, uint8_t* const rgba);
// color mapping related functions.
static WEBP_INLINE uint32_t VP8GetARGBIndex(uint32_t idx) {
return (idx >> 8) & 0xff;
}
static WEBP_INLINE uint8_t VP8GetAlphaIndex(uint8_t idx) {
return idx;
}
static WEBP_INLINE uint32_t VP8GetARGBValue(uint32_t val) {
return val;
}
static WEBP_INLINE uint8_t VP8GetAlphaValue(uint32_t val) {
return (val >> 8) & 0xff;
}
typedef void (*VP8LMapARGBFunc)(const uint32_t* src,
const uint32_t* const color_map,
uint32_t* dst, int y_start,
@ -110,7 +102,8 @@ void VP8LColorIndexInverseTransformAlpha(
// Expose some C-only fallback functions
void VP8LTransformColorInverse_C(const VP8LMultipliers* const m,
uint32_t* data, int num_pixels);
const uint32_t* src, int num_pixels,
uint32_t* dst);
void VP8LConvertBGRAToRGB_C(const uint32_t* src, int num_pixels, uint8_t* dst);
void VP8LConvertBGRAToRGBA_C(const uint32_t* src, int num_pixels, uint8_t* dst);
@ -119,7 +112,8 @@ void VP8LConvertBGRAToRGBA4444_C(const uint32_t* src,
void VP8LConvertBGRAToRGB565_C(const uint32_t* src,
int num_pixels, uint8_t* dst);
void VP8LConvertBGRAToBGR_C(const uint32_t* src, int num_pixels, uint8_t* dst);
void VP8LAddGreenToBlueAndRed_C(uint32_t* data, int num_pixels);
void VP8LAddGreenToBlueAndRed_C(const uint32_t* src, int num_pixels,
uint32_t* dst);
// Must be called before calling any of the above methods.
void VP8LDspInit(void);
@ -127,7 +121,10 @@ void VP8LDspInit(void);
//------------------------------------------------------------------------------
// Encoding
extern VP8LProcessBlueAndRedFunc VP8LSubtractGreenFromBlueAndRed;
typedef void (*VP8LProcessEncBlueAndRedFunc)(uint32_t* dst, int num_pixels);
extern VP8LProcessEncBlueAndRedFunc VP8LSubtractGreenFromBlueAndRed;
typedef void (*VP8LTransformColorFunc)(const VP8LMultipliers* const m,
uint32_t* const dst, int num_pixels);
extern VP8LTransformColorFunc VP8LTransformColor;
typedef void (*VP8LCollectColorBlueTransformsFunc)(
const uint32_t* argb, int stride,
@ -153,62 +150,8 @@ void VP8LCollectColorBlueTransforms_C(const uint32_t* argb, int stride,
int green_to_blue, int red_to_blue,
int histo[]);
//------------------------------------------------------------------------------
// Image transforms.
void VP8LResidualImage(int width, int height, int bits, int low_effort,
uint32_t* const argb, uint32_t* const argb_scratch,
uint32_t* const image, int near_lossless, int exact,
int used_subtract_green);
void VP8LColorSpaceTransform(int width, int height, int bits, int quality,
uint32_t* const argb, uint32_t* image);
//------------------------------------------------------------------------------
// Misc methods.
// Computes sampled size of 'size' when sampling using 'sampling bits'.
static WEBP_INLINE uint32_t VP8LSubSampleSize(uint32_t size,
uint32_t sampling_bits) {
return (size + (1 << sampling_bits) - 1) >> sampling_bits;
}
// Converts near lossless quality into max number of bits shaved off.
static WEBP_INLINE int VP8LNearLosslessBits(int near_lossless_quality) {
// 100 -> 0
// 80..99 -> 1
// 60..79 -> 2
// 40..59 -> 3
// 20..39 -> 4
// 0..19 -> 5
return 5 - near_lossless_quality / 20;
}
// -----------------------------------------------------------------------------
// Faster logarithm for integers. Small values use a look-up table.
// The threshold till approximate version of log_2 can be used.
// Practically, we can get rid of the call to log() as the two values match to
// very high degree (the ratio of these two is 0.99999x).
// Keeping a high threshold for now.
#define APPROX_LOG_WITH_CORRECTION_MAX 65536
#define APPROX_LOG_MAX 4096
#define LOG_2_RECIPROCAL 1.44269504088896338700465094007086
#define LOG_LOOKUP_IDX_MAX 256
extern const float kLog2Table[LOG_LOOKUP_IDX_MAX];
extern const float kSLog2Table[LOG_LOOKUP_IDX_MAX];
typedef float (*VP8LFastLog2SlowFunc)(uint32_t v);
extern VP8LFastLog2SlowFunc VP8LFastLog2Slow;
extern VP8LFastLog2SlowFunc VP8LFastSLog2Slow;
static WEBP_INLINE float VP8LFastLog2(uint32_t v) {
return (v < LOG_LOOKUP_IDX_MAX) ? kLog2Table[v] : VP8LFastLog2Slow(v);
}
// Fast calculation of v * log2(v) for integer input.
static WEBP_INLINE float VP8LFastSLog2(uint32_t v) {
return (v < LOG_LOOKUP_IDX_MAX) ? kSLog2Table[v] : VP8LFastSLog2Slow(v);
}
extern VP8LPredictorAddSubFunc VP8LPredictorsSub[16];
extern VP8LPredictorAddSubFunc VP8LPredictorsSub_C[16];
// -----------------------------------------------------------------------------
// Huffman-cost related functions.
@ -228,11 +171,6 @@ typedef struct { // small struct to hold counters
int streaks[2][2]; // [zero/non-zero][streak<3 / streak>=3]
} VP8LStreaks;
typedef VP8LStreaks (*VP8LCostCombinedCountFunc)(const uint32_t* X,
const uint32_t* Y, int length);
extern VP8LCostCombinedCountFunc VP8LHuffmanCostCombinedCount;
typedef struct { // small struct to hold bit entropy results
double entropy; // entropy
uint32_t sum; // sum of the population
@ -246,26 +184,20 @@ void VP8LBitEntropyInit(VP8LBitEntropy* const entropy);
// Get the combined symbol bit entropy and Huffman cost stats for the
// distributions 'X' and 'Y'. Those results can then be refined according to
// codec specific heuristics.
void VP8LGetCombinedEntropyUnrefined(const uint32_t* const X,
const uint32_t* const Y, int length,
VP8LBitEntropy* const bit_entropy,
VP8LStreaks* const stats);
typedef void (*VP8LGetCombinedEntropyUnrefinedFunc)(
const uint32_t X[], const uint32_t Y[], int length,
VP8LBitEntropy* const bit_entropy, VP8LStreaks* const stats);
extern VP8LGetCombinedEntropyUnrefinedFunc VP8LGetCombinedEntropyUnrefined;
// Get the entropy for the distribution 'X'.
void VP8LGetEntropyUnrefined(const uint32_t* const X, int length,
VP8LBitEntropy* const bit_entropy,
VP8LStreaks* const stats);
typedef void (*VP8LGetEntropyUnrefinedFunc)(const uint32_t X[], int length,
VP8LBitEntropy* const bit_entropy,
VP8LStreaks* const stats);
extern VP8LGetEntropyUnrefinedFunc VP8LGetEntropyUnrefined;
void VP8LBitsEntropyUnrefined(const uint32_t* const array, int n,
VP8LBitEntropy* const entropy);
typedef void (*GetEntropyUnrefinedHelperFunc)(uint32_t val, int i,
uint32_t* const val_prev,
int* const i_prev,
VP8LBitEntropy* const bit_entropy,
VP8LStreaks* const stats);
// Internal function used by VP8LGet*EntropyUnrefined.
extern GetEntropyUnrefinedHelperFunc VP8LGetEntropyUnrefinedHelper;
typedef void (*VP8LHistogramAddFunc)(const VP8LHistogram* const a,
const VP8LHistogram* const b,
VP8LHistogram* const out);
@ -279,86 +211,11 @@ typedef int (*VP8LVectorMismatchFunc)(const uint32_t* const array1,
// Returns the first index where array1 and array2 are different.
extern VP8LVectorMismatchFunc VP8LVectorMismatch;
static WEBP_INLINE int VP8LBitsLog2Ceiling(uint32_t n) {
const int log_floor = BitsLog2Floor(n);
if (n == (n & ~(n - 1))) // zero or a power of two.
return log_floor;
else
return log_floor + 1;
}
// Splitting of distance and length codes into prefixes and
// extra bits. The prefixes are encoded with an entropy code
// while the extra bits are stored just as normal bits.
static WEBP_INLINE void VP8LPrefixEncodeBitsNoLUT(int distance, int* const code,
int* const extra_bits) {
const int highest_bit = BitsLog2Floor(--distance);
const int second_highest_bit = (distance >> (highest_bit - 1)) & 1;
*extra_bits = highest_bit - 1;
*code = 2 * highest_bit + second_highest_bit;
}
static WEBP_INLINE void VP8LPrefixEncodeNoLUT(int distance, int* const code,
int* const extra_bits,
int* const extra_bits_value) {
const int highest_bit = BitsLog2Floor(--distance);
const int second_highest_bit = (distance >> (highest_bit - 1)) & 1;
*extra_bits = highest_bit - 1;
*extra_bits_value = distance & ((1 << *extra_bits) - 1);
*code = 2 * highest_bit + second_highest_bit;
}
#define PREFIX_LOOKUP_IDX_MAX 512
typedef struct {
int8_t code_;
int8_t extra_bits_;
} VP8LPrefixCode;
// These tables are derived using VP8LPrefixEncodeNoLUT.
extern const VP8LPrefixCode kPrefixEncodeCode[PREFIX_LOOKUP_IDX_MAX];
extern const uint8_t kPrefixEncodeExtraBitsValue[PREFIX_LOOKUP_IDX_MAX];
static WEBP_INLINE void VP8LPrefixEncodeBits(int distance, int* const code,
int* const extra_bits) {
if (distance < PREFIX_LOOKUP_IDX_MAX) {
const VP8LPrefixCode prefix_code = kPrefixEncodeCode[distance];
*code = prefix_code.code_;
*extra_bits = prefix_code.extra_bits_;
} else {
VP8LPrefixEncodeBitsNoLUT(distance, code, extra_bits);
}
}
static WEBP_INLINE void VP8LPrefixEncode(int distance, int* const code,
int* const extra_bits,
int* const extra_bits_value) {
if (distance < PREFIX_LOOKUP_IDX_MAX) {
const VP8LPrefixCode prefix_code = kPrefixEncodeCode[distance];
*code = prefix_code.code_;
*extra_bits = prefix_code.extra_bits_;
*extra_bits_value = kPrefixEncodeExtraBitsValue[distance];
} else {
VP8LPrefixEncodeNoLUT(distance, code, extra_bits, extra_bits_value);
}
}
// Sum of each component, mod 256.
static WEBP_INLINE uint32_t VP8LAddPixels(uint32_t a, uint32_t b) {
const uint32_t alpha_and_green = (a & 0xff00ff00u) + (b & 0xff00ff00u);
const uint32_t red_and_blue = (a & 0x00ff00ffu) + (b & 0x00ff00ffu);
return (alpha_and_green & 0xff00ff00u) | (red_and_blue & 0x00ff00ffu);
}
// Difference of each component, mod 256.
static WEBP_INLINE uint32_t VP8LSubPixels(uint32_t a, uint32_t b) {
const uint32_t alpha_and_green =
0x00ff00ffu + (a & 0xff00ff00u) - (b & 0xff00ff00u);
const uint32_t red_and_blue =
0xff00ff00u + (a & 0x00ff00ffu) - (b & 0x00ff00ffu);
return (alpha_and_green & 0xff00ff00u) | (red_and_blue & 0x00ff00ffu);
}
void VP8LBundleColorMap(const uint8_t* const row, int width,
int xbits, uint32_t* const dst);
typedef void (*VP8LBundleColorMapFunc)(const uint8_t* const row, int width,
int xbits, uint32_t* dst);
extern VP8LBundleColorMapFunc VP8LBundleColorMap;
void VP8LBundleColorMap_C(const uint8_t* const row, int width, int xbits,
uint32_t* dst);
// Must be called before calling any of the above methods.
void VP8LEncDspInit(void);

210
thirdparty/libwebp/dsp/lossless_common.h vendored Normal file
View file

@ -0,0 +1,210 @@
// Copyright 2012 Google Inc. All Rights Reserved.
//
// Use of this source code is governed by a BSD-style license
// that can be found in the COPYING file in the root of the source
// tree. An additional intellectual property rights grant can be found
// in the file PATENTS. All contributing project authors may
// be found in the AUTHORS file in the root of the source tree.
// -----------------------------------------------------------------------------
//
// Image transforms and color space conversion methods for lossless decoder.
//
// Authors: Vikas Arora (vikaas.arora@gmail.com)
// Jyrki Alakuijala (jyrki@google.com)
// Vincent Rabaud (vrabaud@google.com)
#ifndef WEBP_DSP_LOSSLESS_COMMON_H_
#define WEBP_DSP_LOSSLESS_COMMON_H_
#include "../webp/types.h"
#include "../utils/utils.h"
#ifdef __cplusplus
extern "C" {
#endif
//------------------------------------------------------------------------------
// Decoding
// color mapping related functions.
static WEBP_INLINE uint32_t VP8GetARGBIndex(uint32_t idx) {
return (idx >> 8) & 0xff;
}
static WEBP_INLINE uint8_t VP8GetAlphaIndex(uint8_t idx) {
return idx;
}
static WEBP_INLINE uint32_t VP8GetARGBValue(uint32_t val) {
return val;
}
static WEBP_INLINE uint8_t VP8GetAlphaValue(uint32_t val) {
return (val >> 8) & 0xff;
}
//------------------------------------------------------------------------------
// Misc methods.
// Computes sampled size of 'size' when sampling using 'sampling bits'.
static WEBP_INLINE uint32_t VP8LSubSampleSize(uint32_t size,
uint32_t sampling_bits) {
return (size + (1 << sampling_bits) - 1) >> sampling_bits;
}
// Converts near lossless quality into max number of bits shaved off.
static WEBP_INLINE int VP8LNearLosslessBits(int near_lossless_quality) {
// 100 -> 0
// 80..99 -> 1
// 60..79 -> 2
// 40..59 -> 3
// 20..39 -> 4
// 0..19 -> 5
return 5 - near_lossless_quality / 20;
}
// -----------------------------------------------------------------------------
// Faster logarithm for integers. Small values use a look-up table.
// The threshold till approximate version of log_2 can be used.
// Practically, we can get rid of the call to log() as the two values match to
// very high degree (the ratio of these two is 0.99999x).
// Keeping a high threshold for now.
#define APPROX_LOG_WITH_CORRECTION_MAX 65536
#define APPROX_LOG_MAX 4096
#define LOG_2_RECIPROCAL 1.44269504088896338700465094007086
#define LOG_LOOKUP_IDX_MAX 256
extern const float kLog2Table[LOG_LOOKUP_IDX_MAX];
extern const float kSLog2Table[LOG_LOOKUP_IDX_MAX];
typedef float (*VP8LFastLog2SlowFunc)(uint32_t v);
extern VP8LFastLog2SlowFunc VP8LFastLog2Slow;
extern VP8LFastLog2SlowFunc VP8LFastSLog2Slow;
static WEBP_INLINE float VP8LFastLog2(uint32_t v) {
return (v < LOG_LOOKUP_IDX_MAX) ? kLog2Table[v] : VP8LFastLog2Slow(v);
}
// Fast calculation of v * log2(v) for integer input.
static WEBP_INLINE float VP8LFastSLog2(uint32_t v) {
return (v < LOG_LOOKUP_IDX_MAX) ? kSLog2Table[v] : VP8LFastSLog2Slow(v);
}
// -----------------------------------------------------------------------------
// PrefixEncode()
static WEBP_INLINE int VP8LBitsLog2Ceiling(uint32_t n) {
const int log_floor = BitsLog2Floor(n);
if (n == (n & ~(n - 1))) { // zero or a power of two.
return log_floor;
}
return log_floor + 1;
}
// Splitting of distance and length codes into prefixes and
// extra bits. The prefixes are encoded with an entropy code
// while the extra bits are stored just as normal bits.
static WEBP_INLINE void VP8LPrefixEncodeBitsNoLUT(int distance, int* const code,
int* const extra_bits) {
const int highest_bit = BitsLog2Floor(--distance);
const int second_highest_bit = (distance >> (highest_bit - 1)) & 1;
*extra_bits = highest_bit - 1;
*code = 2 * highest_bit + second_highest_bit;
}
static WEBP_INLINE void VP8LPrefixEncodeNoLUT(int distance, int* const code,
int* const extra_bits,
int* const extra_bits_value) {
const int highest_bit = BitsLog2Floor(--distance);
const int second_highest_bit = (distance >> (highest_bit - 1)) & 1;
*extra_bits = highest_bit - 1;
*extra_bits_value = distance & ((1 << *extra_bits) - 1);
*code = 2 * highest_bit + second_highest_bit;
}
#define PREFIX_LOOKUP_IDX_MAX 512
typedef struct {
int8_t code_;
int8_t extra_bits_;
} VP8LPrefixCode;
// These tables are derived using VP8LPrefixEncodeNoLUT.
extern const VP8LPrefixCode kPrefixEncodeCode[PREFIX_LOOKUP_IDX_MAX];
extern const uint8_t kPrefixEncodeExtraBitsValue[PREFIX_LOOKUP_IDX_MAX];
static WEBP_INLINE void VP8LPrefixEncodeBits(int distance, int* const code,
int* const extra_bits) {
if (distance < PREFIX_LOOKUP_IDX_MAX) {
const VP8LPrefixCode prefix_code = kPrefixEncodeCode[distance];
*code = prefix_code.code_;
*extra_bits = prefix_code.extra_bits_;
} else {
VP8LPrefixEncodeBitsNoLUT(distance, code, extra_bits);
}
}
static WEBP_INLINE void VP8LPrefixEncode(int distance, int* const code,
int* const extra_bits,
int* const extra_bits_value) {
if (distance < PREFIX_LOOKUP_IDX_MAX) {
const VP8LPrefixCode prefix_code = kPrefixEncodeCode[distance];
*code = prefix_code.code_;
*extra_bits = prefix_code.extra_bits_;
*extra_bits_value = kPrefixEncodeExtraBitsValue[distance];
} else {
VP8LPrefixEncodeNoLUT(distance, code, extra_bits, extra_bits_value);
}
}
// Sum of each component, mod 256.
static WEBP_UBSAN_IGNORE_UNSIGNED_OVERFLOW WEBP_INLINE
uint32_t VP8LAddPixels(uint32_t a, uint32_t b) {
const uint32_t alpha_and_green = (a & 0xff00ff00u) + (b & 0xff00ff00u);
const uint32_t red_and_blue = (a & 0x00ff00ffu) + (b & 0x00ff00ffu);
return (alpha_and_green & 0xff00ff00u) | (red_and_blue & 0x00ff00ffu);
}
// Difference of each component, mod 256.
static WEBP_UBSAN_IGNORE_UNSIGNED_OVERFLOW WEBP_INLINE
uint32_t VP8LSubPixels(uint32_t a, uint32_t b) {
const uint32_t alpha_and_green =
0x00ff00ffu + (a & 0xff00ff00u) - (b & 0xff00ff00u);
const uint32_t red_and_blue =
0xff00ff00u + (a & 0x00ff00ffu) - (b & 0x00ff00ffu);
return (alpha_and_green & 0xff00ff00u) | (red_and_blue & 0x00ff00ffu);
}
//------------------------------------------------------------------------------
// Transform-related functions use din both encoding and decoding.
// Macros used to create a batch predictor that iteratively uses a
// one-pixel predictor.
// The predictor is added to the output pixel (which
// is therefore considered as a residual) to get the final prediction.
#define GENERATE_PREDICTOR_ADD(PREDICTOR, PREDICTOR_ADD) \
static void PREDICTOR_ADD(const uint32_t* in, const uint32_t* upper, \
int num_pixels, uint32_t* out) { \
int x; \
for (x = 0; x < num_pixels; ++x) { \
const uint32_t pred = (PREDICTOR)(out[x - 1], upper + x); \
out[x] = VP8LAddPixels(in[x], pred); \
} \
}
// It subtracts the prediction from the input pixel and stores the residual
// in the output pixel.
#define GENERATE_PREDICTOR_SUB(PREDICTOR, PREDICTOR_SUB) \
static void PREDICTOR_SUB(const uint32_t* in, const uint32_t* upper, \
int num_pixels, uint32_t* out) { \
int x; \
for (x = 0; x < num_pixels; ++x) { \
const uint32_t pred = (PREDICTOR)(in[x - 1], upper + x); \
out[x] = VP8LSubPixels(in[x], pred); \
} \
}
#ifdef __cplusplus
} // extern "C"
#endif
#endif // WEBP_DSP_LOSSLESS_COMMON_H_

953
thirdparty/libwebp/dsp/lossless_enc.c vendored
View file

File diff suppressed because it is too large Load diff

47
thirdparty/libwebp/dsp/lossless_enc_mips32.c vendored
View file

@ -14,6 +14,7 @@
#include "./dsp.h"
#include "./lossless.h"
#include "./lossless_common.h"
#if defined(WEBP_USE_MIPS32)
@ -240,6 +241,49 @@ static WEBP_INLINE void GetEntropyUnrefinedHelper(
*i_prev = i;
}
static void GetEntropyUnrefined(const uint32_t X[], int length,
VP8LBitEntropy* const bit_entropy,
VP8LStreaks* const stats) {
int i;
int i_prev = 0;
uint32_t x_prev = X[0];
memset(stats, 0, sizeof(*stats));
VP8LBitEntropyInit(bit_entropy);
for (i = 1; i < length; ++i) {
const uint32_t x = X[i];
if (x != x_prev) {
GetEntropyUnrefinedHelper(x, i, &x_prev, &i_prev, bit_entropy, stats);
}
}
GetEntropyUnrefinedHelper(0, i, &x_prev, &i_prev, bit_entropy, stats);
bit_entropy->entropy += VP8LFastSLog2(bit_entropy->sum);
}
static void GetCombinedEntropyUnrefined(const uint32_t X[], const uint32_t Y[],
int length,
VP8LBitEntropy* const bit_entropy,
VP8LStreaks* const stats) {
int i = 1;
int i_prev = 0;
uint32_t xy_prev = X[0] + Y[0];
memset(stats, 0, sizeof(*stats));
VP8LBitEntropyInit(bit_entropy);
for (i = 1; i < length; ++i) {
const uint32_t xy = X[i] + Y[i];
if (xy != xy_prev) {
GetEntropyUnrefinedHelper(xy, i, &xy_prev, &i_prev, bit_entropy, stats);
}
}
GetEntropyUnrefinedHelper(0, i, &xy_prev, &i_prev, bit_entropy, stats);
bit_entropy->entropy += VP8LFastSLog2(bit_entropy->sum);
}
#define ASM_START \
__asm__ volatile( \
".set push \n\t" \
@ -375,7 +419,8 @@ WEBP_TSAN_IGNORE_FUNCTION void VP8LEncDspInitMIPS32(void) {
VP8LFastLog2Slow = FastLog2Slow;
VP8LExtraCost = ExtraCost;
VP8LExtraCostCombined = ExtraCostCombined;
VP8LGetEntropyUnrefinedHelper = GetEntropyUnrefinedHelper;
VP8LGetEntropyUnrefined = GetEntropyUnrefined;
VP8LGetCombinedEntropyUnrefined = GetCombinedEntropyUnrefined;
VP8LHistogramAdd = HistogramAdd;
}

147
thirdparty/libwebp/dsp/lossless_enc_msa.c vendored Normal file
View file

@ -0,0 +1,147 @@
// Copyright 2016 Google Inc. All Rights Reserved.
//
// Use of this source code is governed by a BSD-style license
// that can be found in the COPYING file in the root of the source
// tree. An additional intellectual property rights grant can be found
// in the file PATENTS. All contributing project authors may
// be found in the AUTHORS file in the root of the source tree.
// -----------------------------------------------------------------------------
//
// MSA variant of Image transform methods for lossless encoder.
//
// Authors: Prashant Patil (Prashant.Patil@imgtec.com)
#include "./dsp.h"
#if defined(WEBP_USE_MSA)
#include "./lossless.h"
#include "./msa_macro.h"
#define TRANSFORM_COLOR_8(src0, src1, dst0, dst1, c0, c1, mask0, mask1) do { \
v8i16 g0, g1, t0, t1, t2, t3; \
v4i32 t4, t5; \
VSHF_B2_SH(src0, src0, src1, src1, mask0, mask0, g0, g1); \
DOTP_SB2_SH(g0, g1, c0, c0, t0, t1); \
SRAI_H2_SH(t0, t1, 5); \
t0 = __msa_subv_h((v8i16)src0, t0); \
t1 = __msa_subv_h((v8i16)src1, t1); \
t4 = __msa_srli_w((v4i32)src0, 16); \
t5 = __msa_srli_w((v4i32)src1, 16); \
DOTP_SB2_SH(t4, t5, c1, c1, t2, t3); \
SRAI_H2_SH(t2, t3, 5); \
SUB2(t0, t2, t1, t3, t0, t1); \
VSHF_B2_UB(src0, t0, src1, t1, mask1, mask1, dst0, dst1); \
} while (0)
#define TRANSFORM_COLOR_4(src, dst, c0, c1, mask0, mask1) do { \
const v16i8 g0 = VSHF_SB(src, src, mask0); \
v8i16 t0 = __msa_dotp_s_h(c0, g0); \
v8i16 t1; \
v4i32 t2; \
t0 = SRAI_H(t0, 5); \
t0 = __msa_subv_h((v8i16)src, t0); \
t2 = __msa_srli_w((v4i32)src, 16); \
t1 = __msa_dotp_s_h(c1, (v16i8)t2); \
t1 = SRAI_H(t1, 5); \
t0 = t0 - t1; \
dst = VSHF_UB(src, t0, mask1); \
} while (0)
static void TransformColor(const VP8LMultipliers* const m, uint32_t* data,
int num_pixels) {
v16u8 src0, dst0;
const v16i8 g2br = (v16i8)__msa_fill_w(m->green_to_blue_ |
(m->green_to_red_ << 16));
const v16i8 r2b = (v16i8)__msa_fill_w(m->red_to_blue_);
const v16u8 mask0 = { 1, 255, 1, 255, 5, 255, 5, 255, 9, 255, 9, 255,
13, 255, 13, 255 };
const v16u8 mask1 = { 16, 1, 18, 3, 20, 5, 22, 7, 24, 9, 26, 11,
28, 13, 30, 15 };
while (num_pixels >= 8) {
v16u8 src1, dst1;
LD_UB2(data, 4, src0, src1);
TRANSFORM_COLOR_8(src0, src1, dst0, dst1, g2br, r2b, mask0, mask1);
ST_UB2(dst0, dst1, data, 4);
data += 8;
num_pixels -= 8;
}
if (num_pixels > 0) {
if (num_pixels >= 4) {
src0 = LD_UB(data);
TRANSFORM_COLOR_4(src0, dst0, g2br, r2b, mask0, mask1);
ST_UB(dst0, data);
data += 4;
num_pixels -= 4;
}
if (num_pixels > 0) {
src0 = LD_UB(data);
TRANSFORM_COLOR_4(src0, dst0, g2br, r2b, mask0, mask1);
if (num_pixels == 3) {
const uint64_t pix_d = __msa_copy_s_d((v2i64)dst0, 0);
const uint32_t pix_w = __msa_copy_s_w((v4i32)dst0, 2);
SD(pix_d, data + 0);
SW(pix_w, data + 2);
} else if (num_pixels == 2) {
const uint64_t pix_d = __msa_copy_s_d((v2i64)dst0, 0);
SD(pix_d, data);
} else {
const uint32_t pix_w = __msa_copy_s_w((v4i32)dst0, 0);
SW(pix_w, data);
}
}
}
}
static void SubtractGreenFromBlueAndRed(uint32_t* argb_data, int num_pixels) {
int i;
uint8_t* ptemp_data = (uint8_t*)argb_data;
v16u8 src0, dst0, tmp0;
const v16u8 mask = { 1, 255, 1, 255, 5, 255, 5, 255, 9, 255, 9, 255,
13, 255, 13, 255 };
while (num_pixels >= 8) {
v16u8 src1, dst1, tmp1;
LD_UB2(ptemp_data, 16, src0, src1);
VSHF_B2_UB(src0, src1, src1, src0, mask, mask, tmp0, tmp1);
SUB2(src0, tmp0, src1, tmp1, dst0, dst1);
ST_UB2(dst0, dst1, ptemp_data, 16);
ptemp_data += 8 * 4;
num_pixels -= 8;
}
if (num_pixels > 0) {
if (num_pixels >= 4) {
src0 = LD_UB(ptemp_data);
tmp0 = VSHF_UB(src0, src0, mask);
dst0 = src0 - tmp0;
ST_UB(dst0, ptemp_data);
ptemp_data += 4 * 4;
num_pixels -= 4;
}
for (i = 0; i < num_pixels; i++) {
const uint8_t b = ptemp_data[0];
const uint8_t g = ptemp_data[1];
const uint8_t r = ptemp_data[2];
ptemp_data[0] = (b - g) & 0xff;
ptemp_data[2] = (r - g) & 0xff;
ptemp_data += 4;
}
}
}
//------------------------------------------------------------------------------
// Entry point
extern void VP8LEncDspInitMSA(void);
WEBP_TSAN_IGNORE_FUNCTION void VP8LEncDspInitMSA(void) {
VP8LSubtractGreenFromBlueAndRed = SubtractGreenFromBlueAndRed;
VP8LTransformColor = TransformColor;
}
#else // !WEBP_USE_MSA
WEBP_DSP_INIT_STUB(VP8LEncDspInitMSA)
#endif // WEBP_USE_MSA

320
thirdparty/libwebp/dsp/lossless_enc_sse2.c vendored
View file

@ -17,6 +17,8 @@
#include <assert.h>
#include <emmintrin.h>
#include "./lossless.h"
#include "./common_sse2.h"
#include "./lossless_common.h"
// For sign-extended multiplying constants, pre-shifted by 5:
#define CST_5b(X) (((int16_t)((uint16_t)X << 8)) >> 5)
@ -35,7 +37,9 @@ static void SubtractGreenFromBlueAndRed(uint32_t* argb_data, int num_pixels) {
_mm_storeu_si128((__m128i*)&argb_data[i], out);
}
// fallthrough and finish off with plain-C
VP8LSubtractGreenFromBlueAndRed_C(argb_data + i, num_pixels - i);
if (i != num_pixels) {
VP8LSubtractGreenFromBlueAndRed_C(argb_data + i, num_pixels - i);
}
}
//------------------------------------------------------------------------------
@ -69,7 +73,9 @@ static void TransformColor(const VP8LMultipliers* const m,
_mm_storeu_si128((__m128i*)&argb_data[i], out);
}
// fallthrough and finish off with plain-C
VP8LTransformColor_C(m, argb_data + i, num_pixels - i);
if (i != num_pixels) {
VP8LTransformColor_C(m, argb_data + i, num_pixels - i);
}
}
//------------------------------------------------------------------------------
@ -364,8 +370,9 @@ static int VectorMismatch(const uint32_t* const array1,
if (length >= 8 &&
_mm_movemask_epi8(_mm_cmpeq_epi32(
_mm_loadu_si128((const __m128i*)&array1[4]),
_mm_loadu_si128((const __m128i*)&array2[4]))) == 0xffff)
_mm_loadu_si128((const __m128i*)&array2[4]))) == 0xffff) {
match_len = 8;
}
}
}
@ -375,6 +382,295 @@ static int VectorMismatch(const uint32_t* const array1,
return match_len;
}
// Bundles multiple (1, 2, 4 or 8) pixels into a single pixel.
static void BundleColorMap_SSE2(const uint8_t* const row, int width, int xbits,
uint32_t* dst) {
int x;
assert(xbits >= 0);
assert(xbits <= 3);
switch (xbits) {
case 0: {
const __m128i ff = _mm_set1_epi16(0xff00);
const __m128i zero = _mm_setzero_si128();
// Store 0xff000000 | (row[x] << 8).
for (x = 0; x + 16 <= width; x += 16, dst += 16) {
const __m128i in = _mm_loadu_si128((const __m128i*)&row[x]);
const __m128i in_lo = _mm_unpacklo_epi8(zero, in);
const __m128i dst0 = _mm_unpacklo_epi16(in_lo, ff);
const __m128i dst1 = _mm_unpackhi_epi16(in_lo, ff);
const __m128i in_hi = _mm_unpackhi_epi8(zero, in);
const __m128i dst2 = _mm_unpacklo_epi16(in_hi, ff);
const __m128i dst3 = _mm_unpackhi_epi16(in_hi, ff);
_mm_storeu_si128((__m128i*)&dst[0], dst0);
_mm_storeu_si128((__m128i*)&dst[4], dst1);
_mm_storeu_si128((__m128i*)&dst[8], dst2);
_mm_storeu_si128((__m128i*)&dst[12], dst3);
}
break;
}
case 1: {
const __m128i ff = _mm_set1_epi16(0xff00);
const __m128i mul = _mm_set1_epi16(0x110);
for (x = 0; x + 16 <= width; x += 16, dst += 8) {
// 0a0b | (where a/b are 4 bits).
const __m128i in = _mm_loadu_si128((const __m128i*)&row[x]);
const __m128i tmp = _mm_mullo_epi16(in, mul); // aba0
const __m128i pack = _mm_and_si128(tmp, ff); // ab00
const __m128i dst0 = _mm_unpacklo_epi16(pack, ff);
const __m128i dst1 = _mm_unpackhi_epi16(pack, ff);
_mm_storeu_si128((__m128i*)&dst[0], dst0);
_mm_storeu_si128((__m128i*)&dst[4], dst1);
}
break;
}
case 2: {
const __m128i mask_or = _mm_set1_epi32(0xff000000);
const __m128i mul_cst = _mm_set1_epi16(0x0104);
const __m128i mask_mul = _mm_set1_epi16(0x0f00);
for (x = 0; x + 16 <= width; x += 16, dst += 4) {
// 000a000b000c000d | (where a/b/c/d are 2 bits).
const __m128i in = _mm_loadu_si128((const __m128i*)&row[x]);
const __m128i mul = _mm_mullo_epi16(in, mul_cst); // 00ab00b000cd00d0
const __m128i tmp = _mm_and_si128(mul, mask_mul); // 00ab000000cd0000
const __m128i shift = _mm_srli_epi32(tmp, 12); // 00000000ab000000
const __m128i pack = _mm_or_si128(shift, tmp); // 00000000abcd0000
// Convert to 0xff00**00.
const __m128i res = _mm_or_si128(pack, mask_or);
_mm_storeu_si128((__m128i*)dst, res);
}
break;
}
default: {
assert(xbits == 3);
for (x = 0; x + 16 <= width; x += 16, dst += 2) {
// 0000000a00000000b... | (where a/b are 1 bit).
const __m128i in = _mm_loadu_si128((const __m128i*)&row[x]);
const __m128i shift = _mm_slli_epi64(in, 7);
const uint32_t move = _mm_movemask_epi8(shift);
dst[0] = 0xff000000 | ((move & 0xff) << 8);
dst[1] = 0xff000000 | (move & 0xff00);
}
break;
}
}
if (x != width) {
VP8LBundleColorMap_C(row + x, width - x, xbits, dst);
}
}
//------------------------------------------------------------------------------
// Batch version of Predictor Transform subtraction
static WEBP_INLINE void Average2_m128i(const __m128i* const a0,
const __m128i* const a1,
__m128i* const avg) {
// (a + b) >> 1 = ((a + b + 1) >> 1) - ((a ^ b) & 1)
const __m128i ones = _mm_set1_epi8(1);
const __m128i avg1 = _mm_avg_epu8(*a0, *a1);
const __m128i one = _mm_and_si128(_mm_xor_si128(*a0, *a1), ones);
*avg = _mm_sub_epi8(avg1, one);
}
// Predictor0: ARGB_BLACK.
static void PredictorSub0_SSE2(const uint32_t* in, const uint32_t* upper,
int num_pixels, uint32_t* out) {
int i;
const __m128i black = _mm_set1_epi32(ARGB_BLACK);
for (i = 0; i + 4 <= num_pixels; i += 4) {
const __m128i src = _mm_loadu_si128((const __m128i*)&in[i]);
const __m128i res = _mm_sub_epi8(src, black);
_mm_storeu_si128((__m128i*)&out[i], res);
}
if (i != num_pixels) {
VP8LPredictorsSub_C[0](in + i, upper + i, num_pixels - i, out + i);
}
}
#define GENERATE_PREDICTOR_1(X, IN) \
static void PredictorSub##X##_SSE2(const uint32_t* in, const uint32_t* upper, \
int num_pixels, uint32_t* out) { \
int i; \
for (i = 0; i + 4 <= num_pixels; i += 4) { \
const __m128i src = _mm_loadu_si128((const __m128i*)&in[i]); \
const __m128i pred = _mm_loadu_si128((const __m128i*)&(IN)); \
const __m128i res = _mm_sub_epi8(src, pred); \
_mm_storeu_si128((__m128i*)&out[i], res); \
} \
if (i != num_pixels) { \
VP8LPredictorsSub_C[(X)](in + i, upper + i, num_pixels - i, out + i); \
} \
}
GENERATE_PREDICTOR_1(1, in[i - 1]) // Predictor1: L
GENERATE_PREDICTOR_1(2, upper[i]) // Predictor2: T
GENERATE_PREDICTOR_1(3, upper[i + 1]) // Predictor3: TR
GENERATE_PREDICTOR_1(4, upper[i - 1]) // Predictor4: TL
#undef GENERATE_PREDICTOR_1
// Predictor5: avg2(avg2(L, TR), T)
static void PredictorSub5_SSE2(const uint32_t* in, const uint32_t* upper,
int num_pixels, uint32_t* out) {
int i;
for (i = 0; i + 4 <= num_pixels; i += 4) {
const __m128i L = _mm_loadu_si128((const __m128i*)&in[i - 1]);
const __m128i T = _mm_loadu_si128((const __m128i*)&upper[i]);
const __m128i TR = _mm_loadu_si128((const __m128i*)&upper[i + 1]);
const __m128i src = _mm_loadu_si128((const __m128i*)&in[i]);
__m128i avg, pred, res;
Average2_m128i(&L, &TR, &avg);
Average2_m128i(&avg, &T, &pred);
res = _mm_sub_epi8(src, pred);
_mm_storeu_si128((__m128i*)&out[i], res);
}
if (i != num_pixels) {
VP8LPredictorsSub_C[5](in + i, upper + i, num_pixels - i, out + i);
}
}
#define GENERATE_PREDICTOR_2(X, A, B) \
static void PredictorSub##X##_SSE2(const uint32_t* in, const uint32_t* upper, \
int num_pixels, uint32_t* out) { \
int i; \
for (i = 0; i + 4 <= num_pixels; i += 4) { \
const __m128i tA = _mm_loadu_si128((const __m128i*)&(A)); \
const __m128i tB = _mm_loadu_si128((const __m128i*)&(B)); \
const __m128i src = _mm_loadu_si128((const __m128i*)&in[i]); \
__m128i pred, res; \
Average2_m128i(&tA, &tB, &pred); \
res = _mm_sub_epi8(src, pred); \
_mm_storeu_si128((__m128i*)&out[i], res); \
} \
if (i != num_pixels) { \
VP8LPredictorsSub_C[(X)](in + i, upper + i, num_pixels - i, out + i); \
} \
}
GENERATE_PREDICTOR_2(6, in[i - 1], upper[i - 1]) // Predictor6: avg(L, TL)
GENERATE_PREDICTOR_2(7, in[i - 1], upper[i]) // Predictor7: avg(L, T)
GENERATE_PREDICTOR_2(8, upper[i - 1], upper[i]) // Predictor8: avg(TL, T)
GENERATE_PREDICTOR_2(9, upper[i], upper[i + 1]) // Predictor9: average(T, TR)
#undef GENERATE_PREDICTOR_2
// Predictor10: avg(avg(L,TL), avg(T, TR)).
static void PredictorSub10_SSE2(const uint32_t* in, const uint32_t* upper,
int num_pixels, uint32_t* out) {
int i;
for (i = 0; i + 4 <= num_pixels; i += 4) {
const __m128i L = _mm_loadu_si128((const __m128i*)&in[i - 1]);
const __m128i src = _mm_loadu_si128((const __m128i*)&in[i]);
const __m128i TL = _mm_loadu_si128((const __m128i*)&upper[i - 1]);
const __m128i T = _mm_loadu_si128((const __m128i*)&upper[i]);
const __m128i TR = _mm_loadu_si128((const __m128i*)&upper[i + 1]);
__m128i avgTTR, avgLTL, avg, res;
Average2_m128i(&T, &TR, &avgTTR);
Average2_m128i(&L, &TL, &avgLTL);
Average2_m128i(&avgTTR, &avgLTL, &avg);
res = _mm_sub_epi8(src, avg);
_mm_storeu_si128((__m128i*)&out[i], res);
}
if (i != num_pixels) {
VP8LPredictorsSub_C[10](in + i, upper + i, num_pixels - i, out + i);
}
}
// Predictor11: select.
static void GetSumAbsDiff32(const __m128i* const A, const __m128i* const B,
__m128i* const out) {
// We can unpack with any value on the upper 32 bits, provided it's the same
// on both operands (to that their sum of abs diff is zero). Here we use *A.
const __m128i A_lo = _mm_unpacklo_epi32(*A, *A);
const __m128i B_lo = _mm_unpacklo_epi32(*B, *A);
const __m128i A_hi = _mm_unpackhi_epi32(*A, *A);
const __m128i B_hi = _mm_unpackhi_epi32(*B, *A);
const __m128i s_lo = _mm_sad_epu8(A_lo, B_lo);
const __m128i s_hi = _mm_sad_epu8(A_hi, B_hi);
*out = _mm_packs_epi32(s_lo, s_hi);
}
static void PredictorSub11_SSE2(const uint32_t* in, const uint32_t* upper,
int num_pixels, uint32_t* out) {
int i;
for (i = 0; i + 4 <= num_pixels; i += 4) {
const __m128i L = _mm_loadu_si128((const __m128i*)&in[i - 1]);
const __m128i T = _mm_loadu_si128((const __m128i*)&upper[i]);
const __m128i TL = _mm_loadu_si128((const __m128i*)&upper[i - 1]);
const __m128i src = _mm_loadu_si128((const __m128i*)&in[i]);
__m128i pa, pb;
GetSumAbsDiff32(&T, &TL, &pa); // pa = sum |T-TL|
GetSumAbsDiff32(&L, &TL, &pb); // pb = sum |L-TL|
{
const __m128i mask = _mm_cmpgt_epi32(pb, pa);
const __m128i A = _mm_and_si128(mask, L);
const __m128i B = _mm_andnot_si128(mask, T);
const __m128i pred = _mm_or_si128(A, B); // pred = (L > T)? L : T
const __m128i res = _mm_sub_epi8(src, pred);
_mm_storeu_si128((__m128i*)&out[i], res);
}
}
if (i != num_pixels) {
VP8LPredictorsSub_C[11](in + i, upper + i, num_pixels - i, out + i);
}
}
// Predictor12: ClampedSubSubtractFull.
static void PredictorSub12_SSE2(const uint32_t* in, const uint32_t* upper,
int num_pixels, uint32_t* out) {
int i;
const __m128i zero = _mm_setzero_si128();
for (i = 0; i + 4 <= num_pixels; i += 4) {
const __m128i src = _mm_loadu_si128((const __m128i*)&in[i]);
const __m128i L = _mm_loadu_si128((const __m128i*)&in[i - 1]);
const __m128i L_lo = _mm_unpacklo_epi8(L, zero);
const __m128i L_hi = _mm_unpackhi_epi8(L, zero);
const __m128i T = _mm_loadu_si128((const __m128i*)&upper[i]);
const __m128i T_lo = _mm_unpacklo_epi8(T, zero);
const __m128i T_hi = _mm_unpackhi_epi8(T, zero);
const __m128i TL = _mm_loadu_si128((const __m128i*)&upper[i - 1]);
const __m128i TL_lo = _mm_unpacklo_epi8(TL, zero);
const __m128i TL_hi = _mm_unpackhi_epi8(TL, zero);
const __m128i diff_lo = _mm_sub_epi16(T_lo, TL_lo);
const __m128i diff_hi = _mm_sub_epi16(T_hi, TL_hi);
const __m128i pred_lo = _mm_add_epi16(L_lo, diff_lo);
const __m128i pred_hi = _mm_add_epi16(L_hi, diff_hi);
const __m128i pred = _mm_packus_epi16(pred_lo, pred_hi);
const __m128i res = _mm_sub_epi8(src, pred);
_mm_storeu_si128((__m128i*)&out[i], res);
}
if (i != num_pixels) {
VP8LPredictorsSub_C[12](in + i, upper + i, num_pixels - i, out + i);
}
}
// Predictors13: ClampedAddSubtractHalf
static void PredictorSub13_SSE2(const uint32_t* in, const uint32_t* upper,
int num_pixels, uint32_t* out) {
int i;
const __m128i zero = _mm_setzero_si128();
for (i = 0; i + 2 <= num_pixels; i += 2) {
// we can only process two pixels at a time
const __m128i L = _mm_loadl_epi64((const __m128i*)&in[i - 1]);
const __m128i src = _mm_loadl_epi64((const __m128i*)&in[i]);
const __m128i T = _mm_loadl_epi64((const __m128i*)&upper[i]);
const __m128i TL = _mm_loadl_epi64((const __m128i*)&upper[i - 1]);
const __m128i L_lo = _mm_unpacklo_epi8(L, zero);
const __m128i T_lo = _mm_unpacklo_epi8(T, zero);
const __m128i TL_lo = _mm_unpacklo_epi8(TL, zero);
const __m128i sum = _mm_add_epi16(T_lo, L_lo);
const __m128i avg = _mm_srli_epi16(sum, 1);
const __m128i A1 = _mm_sub_epi16(avg, TL_lo);
const __m128i bit_fix = _mm_cmpgt_epi16(TL_lo, avg);
const __m128i A2 = _mm_sub_epi16(A1, bit_fix);
const __m128i A3 = _mm_srai_epi16(A2, 1);
const __m128i A4 = _mm_add_epi16(avg, A3);
const __m128i pred = _mm_packus_epi16(A4, A4);
const __m128i res = _mm_sub_epi8(src, pred);
_mm_storel_epi64((__m128i*)&out[i], res);
}
if (i != num_pixels) {
VP8LPredictorsSub_C[13](in + i, upper + i, num_pixels - i, out + i);
}
}
//------------------------------------------------------------------------------
// Entry point
@ -388,6 +684,24 @@ WEBP_TSAN_IGNORE_FUNCTION void VP8LEncDspInitSSE2(void) {
VP8LHistogramAdd = HistogramAdd;
VP8LCombinedShannonEntropy = CombinedShannonEntropy;
VP8LVectorMismatch = VectorMismatch;
VP8LBundleColorMap = BundleColorMap_SSE2;
VP8LPredictorsSub[0] = PredictorSub0_SSE2;
VP8LPredictorsSub[1] = PredictorSub1_SSE2;
VP8LPredictorsSub[2] = PredictorSub2_SSE2;
VP8LPredictorsSub[3] = PredictorSub3_SSE2;
VP8LPredictorsSub[4] = PredictorSub4_SSE2;
VP8LPredictorsSub[5] = PredictorSub5_SSE2;
VP8LPredictorsSub[6] = PredictorSub6_SSE2;
VP8LPredictorsSub[7] = PredictorSub7_SSE2;
VP8LPredictorsSub[8] = PredictorSub8_SSE2;
VP8LPredictorsSub[9] = PredictorSub9_SSE2;
VP8LPredictorsSub[10] = PredictorSub10_SSE2;
VP8LPredictorsSub[11] = PredictorSub11_SSE2;
VP8LPredictorsSub[12] = PredictorSub12_SSE2;
VP8LPredictorsSub[13] = PredictorSub13_SSE2;
VP8LPredictorsSub[14] = PredictorSub0_SSE2; // <- padding security sentinels
VP8LPredictorsSub[15] = PredictorSub0_SSE2;
}
#else // !WEBP_USE_SSE2

4
thirdparty/libwebp/dsp/lossless_enc_sse41.c vendored
View file

@ -32,7 +32,9 @@ static void SubtractGreenFromBlueAndRed(uint32_t* argb_data, int num_pixels) {
_mm_storeu_si128((__m128i*)&argb_data[i], out);
}
// fallthrough and finish off with plain-C
VP8LSubtractGreenFromBlueAndRed_C(argb_data + i, num_pixels - i);
if (i != num_pixels) {
VP8LSubtractGreenFromBlueAndRed_C(argb_data + i, num_pixels - i);
}
}
//------------------------------------------------------------------------------

79
thirdparty/libwebp/dsp/lossless_mips_dsp_r2.c vendored
View file

@ -17,6 +17,7 @@
#if defined(WEBP_USE_MIPS_DSP_R2)
#include "./lossless.h"
#include "./lossless_common.h"
#define MAP_COLOR_FUNCS(FUNC_NAME, TYPE, GET_INDEX, GET_VALUE) \
static void FUNC_NAME(const TYPE* src, \
@ -227,25 +228,27 @@ static uint32_t Predictor13(uint32_t left, const uint32_t* const top) {
// Add green to blue and red channels (i.e. perform the inverse transform of
// 'subtract green').
static void AddGreenToBlueAndRed(uint32_t* data, int num_pixels) {
static void AddGreenToBlueAndRed(const uint32_t* src, int num_pixels,
uint32_t* dst) {
uint32_t temp0, temp1, temp2, temp3, temp4, temp5, temp6, temp7;
uint32_t* const p_loop1_end = data + (num_pixels & ~3);
uint32_t* const p_loop2_end = data + num_pixels;
const uint32_t* const p_loop1_end = src + (num_pixels & ~3);
const uint32_t* const p_loop2_end = src + num_pixels;
__asm__ volatile (
".set push \n\t"
".set noreorder \n\t"
"beq %[data], %[p_loop1_end], 3f \n\t"
"beq %[src], %[p_loop1_end], 3f \n\t"
" nop \n\t"
"0: \n\t"
"lw %[temp0], 0(%[data]) \n\t"
"lw %[temp1], 4(%[data]) \n\t"
"lw %[temp2], 8(%[data]) \n\t"
"lw %[temp3], 12(%[data]) \n\t"
"lw %[temp0], 0(%[src]) \n\t"
"lw %[temp1], 4(%[src]) \n\t"
"lw %[temp2], 8(%[src]) \n\t"
"lw %[temp3], 12(%[src]) \n\t"
"ext %[temp4], %[temp0], 8, 8 \n\t"
"ext %[temp5], %[temp1], 8, 8 \n\t"
"ext %[temp6], %[temp2], 8, 8 \n\t"
"ext %[temp7], %[temp3], 8, 8 \n\t"
"addiu %[data], %[data], 16 \n\t"
"addiu %[src], %[src], 16 \n\t"
"addiu %[dst], %[dst], 16 \n\t"
"replv.ph %[temp4], %[temp4] \n\t"
"replv.ph %[temp5], %[temp5] \n\t"
"replv.ph %[temp6], %[temp6] \n\t"
@ -254,44 +257,47 @@ static void AddGreenToBlueAndRed(uint32_t* data, int num_pixels) {
"addu.qb %[temp1], %[temp1], %[temp5] \n\t"
"addu.qb %[temp2], %[temp2], %[temp6] \n\t"
"addu.qb %[temp3], %[temp3], %[temp7] \n\t"
"sw %[temp0], -16(%[data]) \n\t"
"sw %[temp1], -12(%[data]) \n\t"
"sw %[temp2], -8(%[data]) \n\t"
"bne %[data], %[p_loop1_end], 0b \n\t"
" sw %[temp3], -4(%[data]) \n\t"
"sw %[temp0], -16(%[dst]) \n\t"
"sw %[temp1], -12(%[dst]) \n\t"
"sw %[temp2], -8(%[dst]) \n\t"
"bne %[src], %[p_loop1_end], 0b \n\t"
" sw %[temp3], -4(%[dst]) \n\t"
"3: \n\t"
"beq %[data], %[p_loop2_end], 2f \n\t"
"beq %[src], %[p_loop2_end], 2f \n\t"
" nop \n\t"
"1: \n\t"
"lw %[temp0], 0(%[data]) \n\t"
"addiu %[data], %[data], 4 \n\t"
"lw %[temp0], 0(%[src]) \n\t"
"addiu %[src], %[src], 4 \n\t"
"addiu %[dst], %[dst], 4 \n\t"
"ext %[temp4], %[temp0], 8, 8 \n\t"
"replv.ph %[temp4], %[temp4] \n\t"
"addu.qb %[temp0], %[temp0], %[temp4] \n\t"
"bne %[data], %[p_loop2_end], 1b \n\t"
" sw %[temp0], -4(%[data]) \n\t"
"bne %[src], %[p_loop2_end], 1b \n\t"
" sw %[temp0], -4(%[dst]) \n\t"
"2: \n\t"
".set pop \n\t"
: [data]"+&r"(data), [temp0]"=&r"(temp0), [temp1]"=&r"(temp1),
[temp2]"=&r"(temp2), [temp3]"=&r"(temp3), [temp4]"=&r"(temp4),
[temp5]"=&r"(temp5), [temp6]"=&r"(temp6), [temp7]"=&r"(temp7)
: [dst]"+&r"(dst), [src]"+&r"(src), [temp0]"=&r"(temp0),
[temp1]"=&r"(temp1), [temp2]"=&r"(temp2), [temp3]"=&r"(temp3),
[temp4]"=&r"(temp4), [temp5]"=&r"(temp5), [temp6]"=&r"(temp6),
[temp7]"=&r"(temp7)
: [p_loop1_end]"r"(p_loop1_end), [p_loop2_end]"r"(p_loop2_end)
: "memory"
);
}
static void TransformColorInverse(const VP8LMultipliers* const m,
uint32_t* data, int num_pixels) {
const uint32_t* src, int num_pixels,
uint32_t* dst) {
int temp0, temp1, temp2, temp3, temp4, temp5;
uint32_t argb, argb1, new_red;
const uint32_t G_to_R = m->green_to_red_;
const uint32_t G_to_B = m->green_to_blue_;
const uint32_t R_to_B = m->red_to_blue_;
uint32_t* const p_loop_end = data + (num_pixels & ~1);
const uint32_t* const p_loop_end = src + (num_pixels & ~1);
__asm__ volatile (
".set push \n\t"
".set noreorder \n\t"
"beq %[data], %[p_loop_end], 1f \n\t"
"beq %[src], %[p_loop_end], 1f \n\t"
" nop \n\t"
"replv.ph %[temp0], %[G_to_R] \n\t"
"replv.ph %[temp1], %[G_to_B] \n\t"
@ -303,9 +309,12 @@ static void TransformColorInverse(const VP8LMultipliers* const m,
"shra.ph %[temp1], %[temp1], 8 \n\t"
"shra.ph %[temp2], %[temp2], 8 \n\t"
"0: \n\t"
"lw %[argb], 0(%[data]) \n\t"
"lw %[argb1], 4(%[data]) \n\t"
"addiu %[data], %[data], 8 \n\t"
"lw %[argb], 0(%[src]) \n\t"
"lw %[argb1], 4(%[src]) \n\t"
"sw %[argb], 0(%[dst]) \n\t"
"sw %[argb1], 4(%[dst]) \n\t"
"addiu %[src], %[src], 8 \n\t"
"addiu %[dst], %[dst], 8 \n\t"
"precrq.qb.ph %[temp3], %[argb], %[argb1] \n\t"
"preceu.ph.qbra %[temp3], %[temp3] \n\t"
"shll.ph %[temp3], %[temp3], 8 \n\t"
@ -322,29 +331,29 @@ static void TransformColorInverse(const VP8LMultipliers* const m,
"shll.ph %[temp4], %[temp5], 8 \n\t"
"shra.ph %[temp4], %[temp4], 8 \n\t"
"mul.ph %[temp4], %[temp4], %[temp2] \n\t"
"sb %[temp5], -2(%[data]) \n\t"
"sb %[temp5], -2(%[dst]) \n\t"
"sra %[temp5], %[temp5], 16 \n\t"
"shra.ph %[temp4], %[temp4], 5 \n\t"
"addu.ph %[argb1], %[argb1], %[temp4] \n\t"
"preceu.ph.qbra %[temp3], %[argb1] \n\t"
"sb %[temp5], -6(%[data]) \n\t"
"sb %[temp3], -4(%[data]) \n\t"
"sb %[temp5], -6(%[dst]) \n\t"
"sb %[temp3], -4(%[dst]) \n\t"
"sra %[temp3], %[temp3], 16 \n\t"
"bne %[data], %[p_loop_end], 0b \n\t"
" sb %[temp3], -8(%[data]) \n\t"
"bne %[src], %[p_loop_end], 0b \n\t"
" sb %[temp3], -8(%[dst]) \n\t"
"1: \n\t"
".set pop \n\t"
: [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2),
[temp3]"=&r"(temp3), [temp4]"=&r"(temp4), [temp5]"=&r"(temp5),
[new_red]"=&r"(new_red), [argb]"=&r"(argb),
[argb1]"=&r"(argb1), [data]"+&r"(data)
[argb1]"=&r"(argb1), [dst]"+&r"(dst), [src]"+&r"(src)
: [G_to_R]"r"(G_to_R), [R_to_B]"r"(R_to_B),
[G_to_B]"r"(G_to_B), [p_loop_end]"r"(p_loop_end)
: "memory", "hi", "lo"
);
// Fall-back to C-version for left-overs.
if (num_pixels & 1) VP8LTransformColorInverse_C(m, data, 1);
if (num_pixels & 1) VP8LTransformColorInverse_C(m, src, 1, dst);
}
static void ConvertBGRAToRGB(const uint32_t* src,

355
thirdparty/libwebp/dsp/lossless_msa.c vendored Normal file
View file

@ -0,0 +1,355 @@
// Copyright 2016 Google Inc. All Rights Reserved.
//
// Use of this source code is governed by a BSD-style license
// that can be found in the COPYING file in the root of the source
// tree. An additional intellectual property rights grant can be found
// in the file PATENTS. All contributing project authors may
// be found in the AUTHORS file in the root of the source tree.
// -----------------------------------------------------------------------------
//
// MSA variant of methods for lossless decoder
//
// Author: Prashant Patil (prashant.patil@imgtec.com)
#include "./dsp.h"
#if defined(WEBP_USE_MSA)
#include "./lossless.h"
#include "./msa_macro.h"
//------------------------------------------------------------------------------
// Colorspace conversion functions
#define CONVERT16_BGRA_XXX(psrc, pdst, m0, m1, m2) do { \
v16u8 src0, src1, src2, src3, dst0, dst1, dst2; \
LD_UB4(psrc, 16, src0, src1, src2, src3); \
VSHF_B2_UB(src0, src1, src1, src2, m0, m1, dst0, dst1); \
dst2 = VSHF_UB(src2, src3, m2); \
ST_UB2(dst0, dst1, pdst, 16); \
ST_UB(dst2, pdst + 32); \
} while (0)
#define CONVERT12_BGRA_XXX(psrc, pdst, m0, m1, m2) do { \
uint32_t pix_w; \
v16u8 src0, src1, src2, dst0, dst1, dst2; \
LD_UB3(psrc, 16, src0, src1, src2); \
VSHF_B2_UB(src0, src1, src1, src2, m0, m1, dst0, dst1); \
dst2 = VSHF_UB(src2, src2, m2); \
ST_UB2(dst0, dst1, pdst, 16); \
pix_w = __msa_copy_s_w((v4i32)dst2, 0); \
SW(pix_w, pdst + 32); \
} while (0)
#define CONVERT8_BGRA_XXX(psrc, pdst, m0, m1) do { \
uint64_t pix_d; \
v16u8 src0, src1, src2, dst0, dst1; \
LD_UB2(psrc, 16, src0, src1); \
VSHF_B2_UB(src0, src1, src1, src2, m0, m1, dst0, dst1); \
ST_UB(dst0, pdst); \
pix_d = __msa_copy_s_d((v2i64)dst1, 0); \
SD(pix_d, pdst + 16); \
} while (0)
#define CONVERT4_BGRA_XXX(psrc, pdst, m) do { \
const v16u8 src0 = LD_UB(psrc); \
const v16u8 dst0 = VSHF_UB(src0, src0, m); \
uint64_t pix_d = __msa_copy_s_d((v2i64)dst0, 0); \
uint32_t pix_w = __msa_copy_s_w((v4i32)dst0, 2); \
SD(pix_d, pdst + 0); \
SW(pix_w, pdst + 8); \
} while (0)
#define CONVERT1_BGRA_BGR(psrc, pdst) do { \
const int32_t b = (psrc)[0]; \
const int32_t g = (psrc)[1]; \
const int32_t r = (psrc)[2]; \
(pdst)[0] = b; \
(pdst)[1] = g; \
(pdst)[2] = r; \
} while (0)
#define CONVERT1_BGRA_RGB(psrc, pdst) do { \
const int32_t b = (psrc)[0]; \
const int32_t g = (psrc)[1]; \
const int32_t r = (psrc)[2]; \
(pdst)[0] = r; \
(pdst)[1] = g; \
(pdst)[2] = b; \
} while (0)
#define TRANSFORM_COLOR_INVERSE_8(src0, src1, dst0, dst1, \
c0, c1, mask0, mask1) do { \
v8i16 g0, g1, t0, t1, t2, t3; \
v4i32 t4, t5; \
VSHF_B2_SH(src0, src0, src1, src1, mask0, mask0, g0, g1); \
DOTP_SB2_SH(g0, g1, c0, c0, t0, t1); \
SRAI_H2_SH(t0, t1, 5); \
t0 = __msa_addv_h(t0, (v8i16)src0); \
t1 = __msa_addv_h(t1, (v8i16)src1); \
t4 = __msa_srli_w((v4i32)t0, 16); \
t5 = __msa_srli_w((v4i32)t1, 16); \
DOTP_SB2_SH(t4, t5, c1, c1, t2, t3); \
SRAI_H2_SH(t2, t3, 5); \
ADD2(t0, t2, t1, t3, t0, t1); \
VSHF_B2_UB(src0, t0, src1, t1, mask1, mask1, dst0, dst1); \
} while (0)
#define TRANSFORM_COLOR_INVERSE_4(src, dst, c0, c1, mask0, mask1) do { \
const v16i8 g0 = VSHF_SB(src, src, mask0); \
v8i16 t0 = __msa_dotp_s_h(c0, g0); \
v8i16 t1; \
v4i32 t2; \
t0 = SRAI_H(t0, 5); \
t0 = __msa_addv_h(t0, (v8i16)src); \
t2 = __msa_srli_w((v4i32)t0, 16); \
t1 = __msa_dotp_s_h(c1, (v16i8)t2); \
t1 = SRAI_H(t1, 5); \
t0 = t0 + t1; \
dst = VSHF_UB(src, t0, mask1); \
} while (0)
static void ConvertBGRAToRGBA(const uint32_t* src,
int num_pixels, uint8_t* dst) {
int i;
const uint8_t* ptemp_src = (const uint8_t*)src;
uint8_t* ptemp_dst = (uint8_t*)dst;
v16u8 src0, dst0;
const v16u8 mask = { 2, 1, 0, 3, 6, 5, 4, 7, 10, 9, 8, 11, 14, 13, 12, 15 };
while (num_pixels >= 8) {
v16u8 src1, dst1;
LD_UB2(ptemp_src, 16, src0, src1);
VSHF_B2_UB(src0, src0, src1, src1, mask, mask, dst0, dst1);
ST_UB2(dst0, dst1, ptemp_dst, 16);
ptemp_src += 32;
ptemp_dst += 32;
num_pixels -= 8;
}
if (num_pixels > 0) {
if (num_pixels >= 4) {
src0 = LD_UB(ptemp_src);
dst0 = VSHF_UB(src0, src0, mask);
ST_UB(dst0, ptemp_dst);
ptemp_src += 16;
ptemp_dst += 16;
num_pixels -= 4;
}
for (i = 0; i < num_pixels; i++) {
const uint8_t b = ptemp_src[2];
const uint8_t g = ptemp_src[1];
const uint8_t r = ptemp_src[0];
const uint8_t a = ptemp_src[3];
ptemp_dst[0] = b;
ptemp_dst[1] = g;
ptemp_dst[2] = r;
ptemp_dst[3] = a;
ptemp_src += 4;
ptemp_dst += 4;
}
}
}
static void ConvertBGRAToBGR(const uint32_t* src,
int num_pixels, uint8_t* dst) {
const uint8_t* ptemp_src = (const uint8_t*)src;
uint8_t* ptemp_dst = (uint8_t*)dst;
const v16u8 mask0 = { 0, 1, 2, 4, 5, 6, 8, 9, 10, 12, 13, 14,
16, 17, 18, 20 };
const v16u8 mask1 = { 5, 6, 8, 9, 10, 12, 13, 14, 16, 17, 18, 20,
21, 22, 24, 25 };
const v16u8 mask2 = { 10, 12, 13, 14, 16, 17, 18, 20, 21, 22, 24, 25,
26, 28, 29, 30 };
while (num_pixels >= 16) {
CONVERT16_BGRA_XXX(ptemp_src, ptemp_dst, mask0, mask1, mask2);
ptemp_src += 64;
ptemp_dst += 48;
num_pixels -= 16;
}
if (num_pixels > 0) {
if (num_pixels >= 12) {
CONVERT12_BGRA_XXX(ptemp_src, ptemp_dst, mask0, mask1, mask2);
ptemp_src += 48;
ptemp_dst += 36;
num_pixels -= 12;
} else if (num_pixels >= 8) {
CONVERT8_BGRA_XXX(ptemp_src, ptemp_dst, mask0, mask1);
ptemp_src += 32;
ptemp_dst += 24;
num_pixels -= 8;
} else if (num_pixels >= 4) {
CONVERT4_BGRA_XXX(ptemp_src, ptemp_dst, mask0);
ptemp_src += 16;
ptemp_dst += 12;
num_pixels -= 4;
}
if (num_pixels == 3) {
CONVERT1_BGRA_BGR(ptemp_src + 0, ptemp_dst + 0);
CONVERT1_BGRA_BGR(ptemp_src + 4, ptemp_dst + 3);
CONVERT1_BGRA_BGR(ptemp_src + 8, ptemp_dst + 6);
} else if (num_pixels == 2) {
CONVERT1_BGRA_BGR(ptemp_src + 0, ptemp_dst + 0);
CONVERT1_BGRA_BGR(ptemp_src + 4, ptemp_dst + 3);
} else if (num_pixels == 1) {
CONVERT1_BGRA_BGR(ptemp_src, ptemp_dst);
}
}
}
static void ConvertBGRAToRGB(const uint32_t* src,
int num_pixels, uint8_t* dst) {
const uint8_t* ptemp_src = (const uint8_t*)src;
uint8_t* ptemp_dst = (uint8_t*)dst;
const v16u8 mask0 = { 2, 1, 0, 6, 5, 4, 10, 9, 8, 14, 13, 12,
18, 17, 16, 22 };
const v16u8 mask1 = { 5, 4, 10, 9, 8, 14, 13, 12, 18, 17, 16, 22,
21, 20, 26, 25 };
const v16u8 mask2 = { 8, 14, 13, 12, 18, 17, 16, 22, 21, 20, 26, 25,
24, 30, 29, 28 };
while (num_pixels >= 16) {
CONVERT16_BGRA_XXX(ptemp_src, ptemp_dst, mask0, mask1, mask2);
ptemp_src += 64;
ptemp_dst += 48;
num_pixels -= 16;
}
if (num_pixels) {
if (num_pixels >= 12) {
CONVERT12_BGRA_XXX(ptemp_src, ptemp_dst, mask0, mask1, mask2);
ptemp_src += 48;
ptemp_dst += 36;
num_pixels -= 12;
} else if (num_pixels >= 8) {
CONVERT8_BGRA_XXX(ptemp_src, ptemp_dst, mask0, mask1);
ptemp_src += 32;
ptemp_dst += 24;
num_pixels -= 8;
} else if (num_pixels >= 4) {
CONVERT4_BGRA_XXX(ptemp_src, ptemp_dst, mask0);
ptemp_src += 16;
ptemp_dst += 12;
num_pixels -= 4;
}
if (num_pixels == 3) {
CONVERT1_BGRA_RGB(ptemp_src + 0, ptemp_dst + 0);
CONVERT1_BGRA_RGB(ptemp_src + 4, ptemp_dst + 3);
CONVERT1_BGRA_RGB(ptemp_src + 8, ptemp_dst + 6);
} else if (num_pixels == 2) {
CONVERT1_BGRA_RGB(ptemp_src + 0, ptemp_dst + 0);
CONVERT1_BGRA_RGB(ptemp_src + 4, ptemp_dst + 3);
} else if (num_pixels == 1) {
CONVERT1_BGRA_RGB(ptemp_src, ptemp_dst);
}
}
}
static void AddGreenToBlueAndRed(const uint32_t* const src, int num_pixels,
uint32_t* dst) {
int i;
const uint8_t* in = (const uint8_t*)src;
uint8_t* out = (uint8_t*)dst;
v16u8 src0, dst0, tmp0;
const v16u8 mask = { 1, 255, 1, 255, 5, 255, 5, 255, 9, 255, 9, 255,
13, 255, 13, 255 };
while (num_pixels >= 8) {
v16u8 src1, dst1, tmp1;
LD_UB2(in, 16, src0, src1);
VSHF_B2_UB(src0, src1, src1, src0, mask, mask, tmp0, tmp1);
ADD2(src0, tmp0, src1, tmp1, dst0, dst1);
ST_UB2(dst0, dst1, out, 16);
in += 32;
out += 32;
num_pixels -= 8;
}
if (num_pixels > 0) {
if (num_pixels >= 4) {
src0 = LD_UB(in);
tmp0 = VSHF_UB(src0, src0, mask);
dst0 = src0 + tmp0;
ST_UB(dst0, out);
in += 16;
out += 16;
num_pixels -= 4;
}
for (i = 0; i < num_pixels; i++) {
const uint8_t b = in[0];
const uint8_t g = in[1];
const uint8_t r = in[2];
out[0] = (b + g) & 0xff;
out[1] = g;
out[2] = (r + g) & 0xff;
out[4] = in[4];
out += 4;
}
}
}
static void TransformColorInverse(const VP8LMultipliers* const m,
const uint32_t* src, int num_pixels,
uint32_t* dst) {
v16u8 src0, dst0;
const v16i8 g2br = (v16i8)__msa_fill_w(m->green_to_blue_ |
(m->green_to_red_ << 16));
const v16i8 r2b = (v16i8)__msa_fill_w(m->red_to_blue_);
const v16u8 mask0 = { 1, 255, 1, 255, 5, 255, 5, 255, 9, 255, 9, 255,
13, 255, 13, 255 };
const v16u8 mask1 = { 16, 1, 18, 3, 20, 5, 22, 7, 24, 9, 26, 11,
28, 13, 30, 15 };
while (num_pixels >= 8) {
v16u8 src1, dst1;
LD_UB2(src, 4, src0, src1);
TRANSFORM_COLOR_INVERSE_8(src0, src1, dst0, dst1, g2br, r2b, mask0, mask1);
ST_UB2(dst0, dst1, dst, 4);
src += 8;
dst += 8;
num_pixels -= 8;
}
if (num_pixels > 0) {
if (num_pixels >= 4) {
src0 = LD_UB(src);
TRANSFORM_COLOR_INVERSE_4(src0, dst0, g2br, r2b, mask0, mask1);
ST_UB(dst0, dst);
src += 4;
dst += 4;
num_pixels -= 4;
}
if (num_pixels > 0) {
src0 = LD_UB(src);
TRANSFORM_COLOR_INVERSE_4(src0, dst0, g2br, r2b, mask0, mask1);
if (num_pixels == 3) {
const uint64_t pix_d = __msa_copy_s_d((v2i64)dst0, 0);
const uint32_t pix_w = __msa_copy_s_w((v4i32)dst0, 2);
SD(pix_d, dst + 0);
SW(pix_w, dst + 2);
} else if (num_pixels == 2) {
const uint64_t pix_d = __msa_copy_s_d((v2i64)dst0, 0);
SD(pix_d, dst);
} else {
const uint32_t pix_w = __msa_copy_s_w((v4i32)dst0, 0);
SW(pix_w, dst);
}
}
}
}
//------------------------------------------------------------------------------
// Entry point
extern void VP8LDspInitMSA(void);
WEBP_TSAN_IGNORE_FUNCTION void VP8LDspInitMSA(void) {
VP8LConvertBGRAToRGBA = ConvertBGRAToRGBA;
VP8LConvertBGRAToBGR = ConvertBGRAToBGR;
VP8LConvertBGRAToRGB = ConvertBGRAToRGB;
VP8LAddGreenToBlueAndRed = AddGreenToBlueAndRed;
VP8LTransformColorInverse = TransformColorInverse;
}
#else // !WEBP_USE_MSA
WEBP_DSP_INIT_STUB(VP8LDspInitMSA)
#endif // WEBP_USE_MSA

395
thirdparty/libwebp/dsp/lossless_neon.c vendored
View file

@ -139,6 +139,357 @@ static void ConvertBGRAToRGB(const uint32_t* src,
#endif // !WORK_AROUND_GCC
//------------------------------------------------------------------------------
// Predictor Transform
#define LOAD_U32_AS_U8(IN) vreinterpret_u8_u32(vdup_n_u32((IN)))
#define LOAD_U32P_AS_U8(IN) vreinterpret_u8_u32(vld1_u32((IN)))
#define LOADQ_U32_AS_U8(IN) vreinterpretq_u8_u32(vdupq_n_u32((IN)))
#define LOADQ_U32P_AS_U8(IN) vreinterpretq_u8_u32(vld1q_u32((IN)))
#define GET_U8_AS_U32(IN) vget_lane_u32(vreinterpret_u32_u8((IN)), 0);
#define GETQ_U8_AS_U32(IN) vgetq_lane_u32(vreinterpretq_u32_u8((IN)), 0);
#define STOREQ_U8_AS_U32P(OUT, IN) vst1q_u32((OUT), vreinterpretq_u32_u8((IN)));
#define ROTATE32_LEFT(L) vextq_u8((L), (L), 12) // D|C|B|A -> C|B|A|D
static WEBP_INLINE uint8x8_t Average2_u8_NEON(uint32_t a0, uint32_t a1) {
const uint8x8_t A0 = LOAD_U32_AS_U8(a0);
const uint8x8_t A1 = LOAD_U32_AS_U8(a1);
return vhadd_u8(A0, A1);
}
static WEBP_INLINE uint32_t ClampedAddSubtractHalf_NEON(uint32_t c0,
uint32_t c1,
uint32_t c2) {
const uint8x8_t avg = Average2_u8_NEON(c0, c1);
// Remove one to c2 when bigger than avg.
const uint8x8_t C2 = LOAD_U32_AS_U8(c2);
const uint8x8_t cmp = vcgt_u8(C2, avg);
const uint8x8_t C2_1 = vadd_u8(C2, cmp);
// Compute half of the difference between avg and c2.
const int8x8_t diff_avg = vreinterpret_s8_u8(vhsub_u8(avg, C2_1));
// Compute the sum with avg and saturate.
const int16x8_t avg_16 = vreinterpretq_s16_u16(vmovl_u8(avg));
const uint8x8_t res = vqmovun_s16(vaddw_s8(avg_16, diff_avg));
const uint32_t output = GET_U8_AS_U32(res);
return output;
}
static WEBP_INLINE uint32_t Average2_NEON(uint32_t a0, uint32_t a1) {
const uint8x8_t avg_u8x8 = Average2_u8_NEON(a0, a1);
const uint32_t avg = GET_U8_AS_U32(avg_u8x8);
return avg;
}
static WEBP_INLINE uint32_t Average3_NEON(uint32_t a0, uint32_t a1,
uint32_t a2) {
const uint8x8_t avg0 = Average2_u8_NEON(a0, a2);
const uint8x8_t A1 = LOAD_U32_AS_U8(a1);
const uint32_t avg = GET_U8_AS_U32(vhadd_u8(avg0, A1));
return avg;
}
static uint32_t Predictor5_NEON(uint32_t left, const uint32_t* const top) {
return Average3_NEON(left, top[0], top[1]);
}
static uint32_t Predictor6_NEON(uint32_t left, const uint32_t* const top) {
return Average2_NEON(left, top[-1]);
}
static uint32_t Predictor7_NEON(uint32_t left, const uint32_t* const top) {
return Average2_NEON(left, top[0]);
}
static uint32_t Predictor13_NEON(uint32_t left, const uint32_t* const top) {
return ClampedAddSubtractHalf_NEON(left, top[0], top[-1]);
}
// Batch versions of those functions.
// Predictor0: ARGB_BLACK.
static void PredictorAdd0_NEON(const uint32_t* in, const uint32_t* upper,
int num_pixels, uint32_t* out) {
int i;
const uint8x16_t black = vreinterpretq_u8_u32(vdupq_n_u32(ARGB_BLACK));
for (i = 0; i + 4 <= num_pixels; i += 4) {
const uint8x16_t src = LOADQ_U32P_AS_U8(&in[i]);
const uint8x16_t res = vaddq_u8(src, black);
STOREQ_U8_AS_U32P(&out[i], res);
}
VP8LPredictorsAdd_C[0](in + i, upper + i, num_pixels - i, out + i);
}
// Predictor1: left.
static void PredictorAdd1_NEON(const uint32_t* in, const uint32_t* upper,
int num_pixels, uint32_t* out) {
int i;
const uint8x16_t zero = LOADQ_U32_AS_U8(0);
for (i = 0; i + 4 <= num_pixels; i += 4) {
// a | b | c | d
const uint8x16_t src = LOADQ_U32P_AS_U8(&in[i]);
// 0 | a | b | c
const uint8x16_t shift0 = vextq_u8(zero, src, 12);
// a | a + b | b + c | c + d
const uint8x16_t sum0 = vaddq_u8(src, shift0);
// 0 | 0 | a | a + b
const uint8x16_t shift1 = vextq_u8(zero, sum0, 8);
// a | a + b | a + b + c | a + b + c + d
const uint8x16_t sum1 = vaddq_u8(sum0, shift1);
const uint8x16_t prev = LOADQ_U32_AS_U8(out[i - 1]);
const uint8x16_t res = vaddq_u8(sum1, prev);
STOREQ_U8_AS_U32P(&out[i], res);
}
VP8LPredictorsAdd_C[1](in + i, upper + i, num_pixels - i, out + i);
}
// Macro that adds 32-bit integers from IN using mod 256 arithmetic
// per 8 bit channel.
#define GENERATE_PREDICTOR_1(X, IN) \
static void PredictorAdd##X##_NEON(const uint32_t* in, \
const uint32_t* upper, int num_pixels, \
uint32_t* out) { \
int i; \
for (i = 0; i + 4 <= num_pixels; i += 4) { \
const uint8x16_t src = LOADQ_U32P_AS_U8(&in[i]); \
const uint8x16_t other = LOADQ_U32P_AS_U8(&(IN)); \
const uint8x16_t res = vaddq_u8(src, other); \
STOREQ_U8_AS_U32P(&out[i], res); \
} \
VP8LPredictorsAdd_C[(X)](in + i, upper + i, num_pixels - i, out + i); \
}
// Predictor2: Top.
GENERATE_PREDICTOR_1(2, upper[i])
// Predictor3: Top-right.
GENERATE_PREDICTOR_1(3, upper[i + 1])
// Predictor4: Top-left.
GENERATE_PREDICTOR_1(4, upper[i - 1])
#undef GENERATE_PREDICTOR_1
// Predictor5: average(average(left, TR), T)
#define DO_PRED5(LANE) do { \
const uint8x16_t avgLTR = vhaddq_u8(L, TR); \
const uint8x16_t avg = vhaddq_u8(avgLTR, T); \
const uint8x16_t res = vaddq_u8(avg, src); \
vst1q_lane_u32(&out[i + (LANE)], vreinterpretq_u32_u8(res), (LANE)); \
L = ROTATE32_LEFT(res); \
} while (0)
static void PredictorAdd5_NEON(const uint32_t* in, const uint32_t* upper,
int num_pixels, uint32_t* out) {
int i;
uint8x16_t L = LOADQ_U32_AS_U8(out[-1]);
for (i = 0; i + 4 <= num_pixels; i += 4) {
const uint8x16_t src = LOADQ_U32P_AS_U8(&in[i]);
const uint8x16_t T = LOADQ_U32P_AS_U8(&upper[i + 0]);
const uint8x16_t TR = LOADQ_U32P_AS_U8(&upper[i + 1]);
DO_PRED5(0);
DO_PRED5(1);
DO_PRED5(2);
DO_PRED5(3);
}
VP8LPredictorsAdd_C[5](in + i, upper + i, num_pixels - i, out + i);
}
#undef DO_PRED5
#define DO_PRED67(LANE) do { \
const uint8x16_t avg = vhaddq_u8(L, top); \
const uint8x16_t res = vaddq_u8(avg, src); \
vst1q_lane_u32(&out[i + (LANE)], vreinterpretq_u32_u8(res), (LANE)); \
L = ROTATE32_LEFT(res); \
} while (0)
// Predictor6: average(left, TL)
static void PredictorAdd6_NEON(const uint32_t* in, const uint32_t* upper,
int num_pixels, uint32_t* out) {
int i;
uint8x16_t L = LOADQ_U32_AS_U8(out[-1]);
for (i = 0; i + 4 <= num_pixels; i += 4) {
const uint8x16_t src = LOADQ_U32P_AS_U8(&in[i]);
const uint8x16_t top = LOADQ_U32P_AS_U8(&upper[i - 1]);
DO_PRED67(0);
DO_PRED67(1);
DO_PRED67(2);
DO_PRED67(3);
}
VP8LPredictorsAdd_C[6](in + i, upper + i, num_pixels - i, out + i);
}
// Predictor7: average(left, T)
static void PredictorAdd7_NEON(const uint32_t* in, const uint32_t* upper,
int num_pixels, uint32_t* out) {
int i;
uint8x16_t L = LOADQ_U32_AS_U8(out[-1]);
for (i = 0; i + 4 <= num_pixels; i += 4) {
const uint8x16_t src = LOADQ_U32P_AS_U8(&in[i]);
const uint8x16_t top = LOADQ_U32P_AS_U8(&upper[i]);
DO_PRED67(0);
DO_PRED67(1);
DO_PRED67(2);
DO_PRED67(3);
}
VP8LPredictorsAdd_C[7](in + i, upper + i, num_pixels - i, out + i);
}
#undef DO_PRED67
#define GENERATE_PREDICTOR_2(X, IN) \
static void PredictorAdd##X##_NEON(const uint32_t* in, \
const uint32_t* upper, int num_pixels, \
uint32_t* out) { \
int i; \
for (i = 0; i + 4 <= num_pixels; i += 4) { \
const uint8x16_t src = LOADQ_U32P_AS_U8(&in[i]); \
const uint8x16_t Tother = LOADQ_U32P_AS_U8(&(IN)); \
const uint8x16_t T = LOADQ_U32P_AS_U8(&upper[i]); \
const uint8x16_t avg = vhaddq_u8(T, Tother); \
const uint8x16_t res = vaddq_u8(avg, src); \
STOREQ_U8_AS_U32P(&out[i], res); \
} \
VP8LPredictorsAdd_C[(X)](in + i, upper + i, num_pixels - i, out + i); \
}
// Predictor8: average TL T.
GENERATE_PREDICTOR_2(8, upper[i - 1])
// Predictor9: average T TR.
GENERATE_PREDICTOR_2(9, upper[i + 1])
#undef GENERATE_PREDICTOR_2
// Predictor10: average of (average of (L,TL), average of (T, TR)).
#define DO_PRED10(LANE) do { \
const uint8x16_t avgLTL = vhaddq_u8(L, TL); \
const uint8x16_t avg = vhaddq_u8(avgTTR, avgLTL); \
const uint8x16_t res = vaddq_u8(avg, src); \
vst1q_lane_u32(&out[i + (LANE)], vreinterpretq_u32_u8(res), (LANE)); \
L = ROTATE32_LEFT(res); \
} while (0)
static void PredictorAdd10_NEON(const uint32_t* in, const uint32_t* upper,
int num_pixels, uint32_t* out) {
int i;
uint8x16_t L = LOADQ_U32_AS_U8(out[-1]);
for (i = 0; i + 4 <= num_pixels; i += 4) {
const uint8x16_t src = LOADQ_U32P_AS_U8(&in[i]);
const uint8x16_t TL = LOADQ_U32P_AS_U8(&upper[i - 1]);
const uint8x16_t T = LOADQ_U32P_AS_U8(&upper[i]);
const uint8x16_t TR = LOADQ_U32P_AS_U8(&upper[i + 1]);
const uint8x16_t avgTTR = vhaddq_u8(T, TR);
DO_PRED10(0);
DO_PRED10(1);
DO_PRED10(2);
DO_PRED10(3);
}
VP8LPredictorsAdd_C[10](in + i, upper + i, num_pixels - i, out + i);
}
#undef DO_PRED10
// Predictor11: select.
#define DO_PRED11(LANE) do { \
const uint8x16_t sumLin = vaddq_u8(L, src); /* in + L */ \
const uint8x16_t pLTL = vabdq_u8(L, TL); /* |L - TL| */ \
const uint16x8_t sum_LTL = vpaddlq_u8(pLTL); \
const uint32x4_t pa = vpaddlq_u16(sum_LTL); \
const uint32x4_t mask = vcleq_u32(pa, pb); \
const uint8x16_t res = vbslq_u8(vreinterpretq_u8_u32(mask), sumTin, sumLin); \
vst1q_lane_u32(&out[i + (LANE)], vreinterpretq_u32_u8(res), (LANE)); \
L = ROTATE32_LEFT(res); \
} while (0)
static void PredictorAdd11_NEON(const uint32_t* in, const uint32_t* upper,
int num_pixels, uint32_t* out) {
int i;
uint8x16_t L = LOADQ_U32_AS_U8(out[-1]);
for (i = 0; i + 4 <= num_pixels; i += 4) {
const uint8x16_t T = LOADQ_U32P_AS_U8(&upper[i]);
const uint8x16_t TL = LOADQ_U32P_AS_U8(&upper[i - 1]);
const uint8x16_t pTTL = vabdq_u8(T, TL); // |T - TL|
const uint16x8_t sum_TTL = vpaddlq_u8(pTTL);
const uint32x4_t pb = vpaddlq_u16(sum_TTL);
const uint8x16_t src = LOADQ_U32P_AS_U8(&in[i]);
const uint8x16_t sumTin = vaddq_u8(T, src); // in + T
DO_PRED11(0);
DO_PRED11(1);
DO_PRED11(2);
DO_PRED11(3);
}
VP8LPredictorsAdd_C[11](in + i, upper + i, num_pixels - i, out + i);
}
#undef DO_PRED11
// Predictor12: ClampedAddSubtractFull.
#define DO_PRED12(DIFF, LANE) do { \
const uint8x8_t pred = \
vqmovun_s16(vaddq_s16(vreinterpretq_s16_u16(L), (DIFF))); \
const uint8x8_t res = \
vadd_u8(pred, (LANE <= 1) ? vget_low_u8(src) : vget_high_u8(src)); \
const uint16x8_t res16 = vmovl_u8(res); \
vst1_lane_u32(&out[i + (LANE)], vreinterpret_u32_u8(res), (LANE) & 1); \
/* rotate in the left predictor for next iteration */ \
L = vextq_u16(res16, res16, 4); \
} while (0)
static void PredictorAdd12_NEON(const uint32_t* in, const uint32_t* upper,
int num_pixels, uint32_t* out) {
int i;
uint16x8_t L = vmovl_u8(LOAD_U32_AS_U8(out[-1]));
for (i = 0; i + 4 <= num_pixels; i += 4) {
// load four pixels of source
const uint8x16_t src = LOADQ_U32P_AS_U8(&in[i]);
// precompute the difference T - TL once for all, stored as s16
const uint8x16_t TL = LOADQ_U32P_AS_U8(&upper[i - 1]);
const uint8x16_t T = LOADQ_U32P_AS_U8(&upper[i]);
const int16x8_t diff_lo =
vreinterpretq_s16_u16(vsubl_u8(vget_low_u8(T), vget_low_u8(TL)));
const int16x8_t diff_hi =
vreinterpretq_s16_u16(vsubl_u8(vget_high_u8(T), vget_high_u8(TL)));
// loop over the four reconstructed pixels
DO_PRED12(diff_lo, 0);
DO_PRED12(diff_lo, 1);
DO_PRED12(diff_hi, 2);
DO_PRED12(diff_hi, 3);
}
VP8LPredictorsAdd_C[12](in + i, upper + i, num_pixels - i, out + i);
}
#undef DO_PRED12
// Predictor13: ClampedAddSubtractHalf
#define DO_PRED13(LANE, LOW_OR_HI) do { \
const uint8x16_t avg = vhaddq_u8(L, T); \
const uint8x16_t cmp = vcgtq_u8(TL, avg); \
const uint8x16_t TL_1 = vaddq_u8(TL, cmp); \
/* Compute half of the difference between avg and TL'. */ \
const int8x8_t diff_avg = \
vreinterpret_s8_u8(LOW_OR_HI(vhsubq_u8(avg, TL_1))); \
/* Compute the sum with avg and saturate. */ \
const int16x8_t avg_16 = vreinterpretq_s16_u16(vmovl_u8(LOW_OR_HI(avg))); \
const uint8x8_t delta = vqmovun_s16(vaddw_s8(avg_16, diff_avg)); \
const uint8x8_t res = vadd_u8(LOW_OR_HI(src), delta); \
const uint8x16_t res2 = vcombine_u8(res, res); \
vst1_lane_u32(&out[i + (LANE)], vreinterpret_u32_u8(res), (LANE) & 1); \
L = ROTATE32_LEFT(res2); \
} while (0)
static void PredictorAdd13_NEON(const uint32_t* in, const uint32_t* upper,
int num_pixels, uint32_t* out) {
int i;
uint8x16_t L = LOADQ_U32_AS_U8(out[-1]);
for (i = 0; i + 4 <= num_pixels; i += 4) {
const uint8x16_t src = LOADQ_U32P_AS_U8(&in[i]);
const uint8x16_t T = LOADQ_U32P_AS_U8(&upper[i]);
const uint8x16_t TL = LOADQ_U32P_AS_U8(&upper[i - 1]);
DO_PRED13(0, vget_low_u8);
DO_PRED13(1, vget_low_u8);
DO_PRED13(2, vget_high_u8);
DO_PRED13(3, vget_high_u8);
}
VP8LPredictorsAdd_C[13](in + i, upper + i, num_pixels - i, out + i);
}
#undef DO_PRED13
#undef LOAD_U32_AS_U8
#undef LOAD_U32P_AS_U8
#undef LOADQ_U32_AS_U8
#undef LOADQ_U32P_AS_U8
#undef GET_U8_AS_U32
#undef GETQ_U8_AS_U32
#undef STOREQ_U8_AS_U32P
#undef ROTATE32_LEFT
//------------------------------------------------------------------------------
// Subtract-Green Transform
@ -171,28 +522,30 @@ static WEBP_INLINE uint8x16_t DoGreenShuffle(const uint8x16_t argb,
}
#endif // USE_VTBLQ
static void AddGreenToBlueAndRed(uint32_t* argb_data, int num_pixels) {
const uint32_t* const end = argb_data + (num_pixels & ~3);
static void AddGreenToBlueAndRed(const uint32_t* src, int num_pixels,
uint32_t* dst) {
const uint32_t* const end = src + (num_pixels & ~3);
#ifdef USE_VTBLQ
const uint8x16_t shuffle = vld1q_u8(kGreenShuffle);
#else
const uint8x8_t shuffle = vld1_u8(kGreenShuffle);
#endif
for (; argb_data < end; argb_data += 4) {
const uint8x16_t argb = vld1q_u8((uint8_t*)argb_data);
for (; src < end; src += 4, dst += 4) {
const uint8x16_t argb = vld1q_u8((const uint8_t*)src);
const uint8x16_t greens = DoGreenShuffle(argb, shuffle);
vst1q_u8((uint8_t*)argb_data, vaddq_u8(argb, greens));
vst1q_u8((uint8_t*)dst, vaddq_u8(argb, greens));
}
// fallthrough and finish off with plain-C
VP8LAddGreenToBlueAndRed_C(argb_data, num_pixels & 3);
VP8LAddGreenToBlueAndRed_C(src, num_pixels & 3, dst);
}
//------------------------------------------------------------------------------
// Color Transform
static void TransformColorInverse(const VP8LMultipliers* const m,
uint32_t* argb_data, int num_pixels) {
// sign-extended multiplying constants, pre-shifted by 6.
const uint32_t* const src, int num_pixels,
uint32_t* dst) {
// sign-extended multiplying constants, pre-shifted by 6.
#define CST(X) (((int16_t)(m->X << 8)) >> 6)
const int16_t rb[8] = {
CST(green_to_blue_), CST(green_to_red_),
@ -219,7 +572,7 @@ static void TransformColorInverse(const VP8LMultipliers* const m,
const uint32x4_t mask_ag = vdupq_n_u32(0xff00ff00u);
int i;
for (i = 0; i + 4 <= num_pixels; i += 4) {
const uint8x16_t in = vld1q_u8((uint8_t*)(argb_data + i));
const uint8x16_t in = vld1q_u8((const uint8_t*)(src + i));
const uint32x4_t a0g0 = vandq_u32(vreinterpretq_u32_u8(in), mask_ag);
// 0 g 0 g
const uint8x16_t greens = DoGreenShuffle(in, shuffle);
@ -240,10 +593,10 @@ static void TransformColorInverse(const VP8LMultipliers* const m,
// 0 r' 0 b''
const uint16x8_t G = vshrq_n_u16(vreinterpretq_u16_s8(F), 8);
const uint32x4_t out = vorrq_u32(vreinterpretq_u32_u16(G), a0g0);
vst1q_u32(argb_data + i, out);
vst1q_u32(dst + i, out);
}
// Fall-back to C-version for left-overs.
VP8LTransformColorInverse_C(m, argb_data + i, num_pixels - i);
VP8LTransformColorInverse_C(m, src + i, num_pixels - i, dst + i);
}
#undef USE_VTBLQ
@ -254,6 +607,26 @@ static void TransformColorInverse(const VP8LMultipliers* const m,
extern void VP8LDspInitNEON(void);
WEBP_TSAN_IGNORE_FUNCTION void VP8LDspInitNEON(void) {
VP8LPredictors[5] = Predictor5_NEON;
VP8LPredictors[6] = Predictor6_NEON;
VP8LPredictors[7] = Predictor7_NEON;
VP8LPredictors[13] = Predictor13_NEON;
VP8LPredictorsAdd[0] = PredictorAdd0_NEON;
VP8LPredictorsAdd[1] = PredictorAdd1_NEON;
VP8LPredictorsAdd[2] = PredictorAdd2_NEON;
VP8LPredictorsAdd[3] = PredictorAdd3_NEON;
VP8LPredictorsAdd[4] = PredictorAdd4_NEON;
VP8LPredictorsAdd[5] = PredictorAdd5_NEON;
VP8LPredictorsAdd[6] = PredictorAdd6_NEON;
VP8LPredictorsAdd[7] = PredictorAdd7_NEON;
VP8LPredictorsAdd[8] = PredictorAdd8_NEON;
VP8LPredictorsAdd[9] = PredictorAdd9_NEON;
VP8LPredictorsAdd[10] = PredictorAdd10_NEON;
VP8LPredictorsAdd[11] = PredictorAdd11_NEON;
VP8LPredictorsAdd[12] = PredictorAdd12_NEON;
VP8LPredictorsAdd[13] = PredictorAdd13_NEON;
VP8LConvertBGRAToRGBA = ConvertBGRAToRGBA;
VP8LConvertBGRAToBGR = ConvertBGRAToBGR;
VP8LConvertBGRAToRGB = ConvertBGRAToRGB;

389
thirdparty/libwebp/dsp/lossless_sse2.c vendored
View file

@ -14,9 +14,12 @@
#include "./dsp.h"
#if defined(WEBP_USE_SSE2)
#include "./common_sse2.h"
#include "./lossless.h"
#include "./lossless_common.h"
#include <assert.h>
#include <emmintrin.h>
#include "./lossless.h"
//------------------------------------------------------------------------------
// Predictor Transform
@ -75,25 +78,44 @@ static WEBP_INLINE uint32_t Select(uint32_t a, uint32_t b, uint32_t c) {
return (pa_minus_pb <= 0) ? a : b;
}
static WEBP_INLINE __m128i Average2_128i(uint32_t a0, uint32_t a1) {
static WEBP_INLINE void Average2_m128i(const __m128i* const a0,
const __m128i* const a1,
__m128i* const avg) {
// (a + b) >> 1 = ((a + b + 1) >> 1) - ((a ^ b) & 1)
const __m128i ones = _mm_set1_epi8(1);
const __m128i avg1 = _mm_avg_epu8(*a0, *a1);
const __m128i one = _mm_and_si128(_mm_xor_si128(*a0, *a1), ones);
*avg = _mm_sub_epi8(avg1, one);
}
static WEBP_INLINE void Average2_uint32(const uint32_t a0, const uint32_t a1,
__m128i* const avg) {
// (a + b) >> 1 = ((a + b + 1) >> 1) - ((a ^ b) & 1)
const __m128i ones = _mm_set1_epi8(1);
const __m128i A0 = _mm_cvtsi32_si128(a0);
const __m128i A1 = _mm_cvtsi32_si128(a1);
const __m128i avg1 = _mm_avg_epu8(A0, A1);
const __m128i one = _mm_and_si128(_mm_xor_si128(A0, A1), ones);
*avg = _mm_sub_epi8(avg1, one);
}
static WEBP_INLINE __m128i Average2_uint32_16(uint32_t a0, uint32_t a1) {
const __m128i zero = _mm_setzero_si128();
const __m128i A0 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(a0), zero);
const __m128i A1 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(a1), zero);
const __m128i sum = _mm_add_epi16(A1, A0);
const __m128i avg = _mm_srli_epi16(sum, 1);
return avg;
return _mm_srli_epi16(sum, 1);
}
static WEBP_INLINE uint32_t Average2(uint32_t a0, uint32_t a1) {
const __m128i avg = Average2_128i(a0, a1);
const __m128i A2 = _mm_packus_epi16(avg, avg);
const uint32_t output = _mm_cvtsi128_si32(A2);
return output;
__m128i output;
Average2_uint32(a0, a1, &output);
return _mm_cvtsi128_si32(output);
}
static WEBP_INLINE uint32_t Average3(uint32_t a0, uint32_t a1, uint32_t a2) {
const __m128i zero = _mm_setzero_si128();
const __m128i avg1 = Average2_128i(a0, a2);
const __m128i avg1 = Average2_uint32_16(a0, a2);
const __m128i A1 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(a1), zero);
const __m128i sum = _mm_add_epi16(avg1, A1);
const __m128i avg2 = _mm_srli_epi16(sum, 1);
@ -104,8 +126,8 @@ static WEBP_INLINE uint32_t Average3(uint32_t a0, uint32_t a1, uint32_t a2) {
static WEBP_INLINE uint32_t Average4(uint32_t a0, uint32_t a1,
uint32_t a2, uint32_t a3) {
const __m128i avg1 = Average2_128i(a0, a1);
const __m128i avg2 = Average2_128i(a2, a3);
const __m128i avg1 = Average2_uint32_16(a0, a1);
const __m128i avg2 = Average2_uint32_16(a2, a3);
const __m128i sum = _mm_add_epi16(avg2, avg1);
const __m128i avg3 = _mm_srli_epi16(sum, 1);
const __m128i A0 = _mm_packus_epi16(avg3, avg3);
@ -113,68 +135,289 @@ static WEBP_INLINE uint32_t Average4(uint32_t a0, uint32_t a1,
return output;
}
static uint32_t Predictor5(uint32_t left, const uint32_t* const top) {
static uint32_t Predictor5_SSE2(uint32_t left, const uint32_t* const top) {
const uint32_t pred = Average3(left, top[0], top[1]);
return pred;
}
static uint32_t Predictor6(uint32_t left, const uint32_t* const top) {
static uint32_t Predictor6_SSE2(uint32_t left, const uint32_t* const top) {
const uint32_t pred = Average2(left, top[-1]);
return pred;
}
static uint32_t Predictor7(uint32_t left, const uint32_t* const top) {
static uint32_t Predictor7_SSE2(uint32_t left, const uint32_t* const top) {
const uint32_t pred = Average2(left, top[0]);
return pred;
}
static uint32_t Predictor8(uint32_t left, const uint32_t* const top) {
static uint32_t Predictor8_SSE2(uint32_t left, const uint32_t* const top) {
const uint32_t pred = Average2(top[-1], top[0]);
(void)left;
return pred;
}
static uint32_t Predictor9(uint32_t left, const uint32_t* const top) {
static uint32_t Predictor9_SSE2(uint32_t left, const uint32_t* const top) {
const uint32_t pred = Average2(top[0], top[1]);
(void)left;
return pred;
}
static uint32_t Predictor10(uint32_t left, const uint32_t* const top) {
static uint32_t Predictor10_SSE2(uint32_t left, const uint32_t* const top) {
const uint32_t pred = Average4(left, top[-1], top[0], top[1]);
return pred;
}
static uint32_t Predictor11(uint32_t left, const uint32_t* const top) {
static uint32_t Predictor11_SSE2(uint32_t left, const uint32_t* const top) {
const uint32_t pred = Select(top[0], left, top[-1]);
return pred;
}
static uint32_t Predictor12(uint32_t left, const uint32_t* const top) {
static uint32_t Predictor12_SSE2(uint32_t left, const uint32_t* const top) {
const uint32_t pred = ClampedAddSubtractFull(left, top[0], top[-1]);
return pred;
}
static uint32_t Predictor13(uint32_t left, const uint32_t* const top) {
static uint32_t Predictor13_SSE2(uint32_t left, const uint32_t* const top) {
const uint32_t pred = ClampedAddSubtractHalf(left, top[0], top[-1]);
return pred;
}
// Batch versions of those functions.
// Predictor0: ARGB_BLACK.
static void PredictorAdd0_SSE2(const uint32_t* in, const uint32_t* upper,
int num_pixels, uint32_t* out) {
int i;
const __m128i black = _mm_set1_epi32(ARGB_BLACK);
for (i = 0; i + 4 <= num_pixels; i += 4) {
const __m128i src = _mm_loadu_si128((const __m128i*)&in[i]);
const __m128i res = _mm_add_epi8(src, black);
_mm_storeu_si128((__m128i*)&out[i], res);
}
if (i != num_pixels) {
VP8LPredictorsAdd_C[0](in + i, upper + i, num_pixels - i, out + i);
}
}
// Predictor1: left.
static void PredictorAdd1_SSE2(const uint32_t* in, const uint32_t* upper,
int num_pixels, uint32_t* out) {
int i;
__m128i prev = _mm_set1_epi32(out[-1]);
for (i = 0; i + 4 <= num_pixels; i += 4) {
// a | b | c | d
const __m128i src = _mm_loadu_si128((const __m128i*)&in[i]);
// 0 | a | b | c
const __m128i shift0 = _mm_slli_si128(src, 4);
// a | a + b | b + c | c + d
const __m128i sum0 = _mm_add_epi8(src, shift0);
// 0 | 0 | a | a + b
const __m128i shift1 = _mm_slli_si128(sum0, 8);
// a | a + b | a + b + c | a + b + c + d
const __m128i sum1 = _mm_add_epi8(sum0, shift1);
const __m128i res = _mm_add_epi8(sum1, prev);
_mm_storeu_si128((__m128i*)&out[i], res);
// replicate prev output on the four lanes
prev = _mm_shuffle_epi32(res, (3 << 0) | (3 << 2) | (3 << 4) | (3 << 6));
}
if (i != num_pixels) {
VP8LPredictorsAdd_C[1](in + i, upper + i, num_pixels - i, out + i);
}
}
// Macro that adds 32-bit integers from IN using mod 256 arithmetic
// per 8 bit channel.
#define GENERATE_PREDICTOR_1(X, IN) \
static void PredictorAdd##X##_SSE2(const uint32_t* in, const uint32_t* upper, \
int num_pixels, uint32_t* out) { \
int i; \
for (i = 0; i + 4 <= num_pixels; i += 4) { \
const __m128i src = _mm_loadu_si128((const __m128i*)&in[i]); \
const __m128i other = _mm_loadu_si128((const __m128i*)&(IN)); \
const __m128i res = _mm_add_epi8(src, other); \
_mm_storeu_si128((__m128i*)&out[i], res); \
} \
if (i != num_pixels) { \
VP8LPredictorsAdd_C[(X)](in + i, upper + i, num_pixels - i, out + i); \
} \
}
// Predictor2: Top.
GENERATE_PREDICTOR_1(2, upper[i])
// Predictor3: Top-right.
GENERATE_PREDICTOR_1(3, upper[i + 1])
// Predictor4: Top-left.
GENERATE_PREDICTOR_1(4, upper[i - 1])
#undef GENERATE_PREDICTOR_1
// Due to averages with integers, values cannot be accumulated in parallel for
// predictors 5 to 7.
GENERATE_PREDICTOR_ADD(Predictor5_SSE2, PredictorAdd5_SSE2)
GENERATE_PREDICTOR_ADD(Predictor6_SSE2, PredictorAdd6_SSE2)
GENERATE_PREDICTOR_ADD(Predictor7_SSE2, PredictorAdd7_SSE2)
#define GENERATE_PREDICTOR_2(X, IN) \
static void PredictorAdd##X##_SSE2(const uint32_t* in, const uint32_t* upper, \
int num_pixels, uint32_t* out) { \
int i; \
for (i = 0; i + 4 <= num_pixels; i += 4) { \
const __m128i Tother = _mm_loadu_si128((const __m128i*)&(IN)); \
const __m128i T = _mm_loadu_si128((const __m128i*)&upper[i]); \
const __m128i src = _mm_loadu_si128((const __m128i*)&in[i]); \
__m128i avg, res; \
Average2_m128i(&T, &Tother, &avg); \
res = _mm_add_epi8(avg, src); \
_mm_storeu_si128((__m128i*)&out[i], res); \
} \
if (i != num_pixels) { \
VP8LPredictorsAdd_C[(X)](in + i, upper + i, num_pixels - i, out + i); \
} \
}
// Predictor8: average TL T.
GENERATE_PREDICTOR_2(8, upper[i - 1])
// Predictor9: average T TR.
GENERATE_PREDICTOR_2(9, upper[i + 1])
#undef GENERATE_PREDICTOR_2
// Predictor10: average of (average of (L,TL), average of (T, TR)).
static void PredictorAdd10_SSE2(const uint32_t* in, const uint32_t* upper,
int num_pixels, uint32_t* out) {
int i, j;
__m128i L = _mm_cvtsi32_si128(out[-1]);
for (i = 0; i + 4 <= num_pixels; i += 4) {
__m128i src = _mm_loadu_si128((const __m128i*)&in[i]);
__m128i TL = _mm_loadu_si128((const __m128i*)&upper[i - 1]);
const __m128i T = _mm_loadu_si128((const __m128i*)&upper[i]);
const __m128i TR = _mm_loadu_si128((const __m128i*)&upper[i + 1]);
__m128i avgTTR;
Average2_m128i(&T, &TR, &avgTTR);
for (j = 0; j < 4; ++j) {
__m128i avgLTL, avg;
Average2_m128i(&L, &TL, &avgLTL);
Average2_m128i(&avgTTR, &avgLTL, &avg);
L = _mm_add_epi8(avg, src);
out[i + j] = _mm_cvtsi128_si32(L);
// Rotate the pre-computed values for the next iteration.
avgTTR = _mm_srli_si128(avgTTR, 4);
TL = _mm_srli_si128(TL, 4);
src = _mm_srli_si128(src, 4);
}
}
if (i != num_pixels) {
VP8LPredictorsAdd_C[10](in + i, upper + i, num_pixels - i, out + i);
}
}
// Predictor11: select.
static void GetSumAbsDiff32(const __m128i* const A, const __m128i* const B,
__m128i* const out) {
// We can unpack with any value on the upper 32 bits, provided it's the same
// on both operands (to that their sum of abs diff is zero). Here we use *A.
const __m128i A_lo = _mm_unpacklo_epi32(*A, *A);
const __m128i B_lo = _mm_unpacklo_epi32(*B, *A);
const __m128i A_hi = _mm_unpackhi_epi32(*A, *A);
const __m128i B_hi = _mm_unpackhi_epi32(*B, *A);
const __m128i s_lo = _mm_sad_epu8(A_lo, B_lo);
const __m128i s_hi = _mm_sad_epu8(A_hi, B_hi);
*out = _mm_packs_epi32(s_lo, s_hi);
}
static void PredictorAdd11_SSE2(const uint32_t* in, const uint32_t* upper,
int num_pixels, uint32_t* out) {
int i, j;
__m128i L = _mm_cvtsi32_si128(out[-1]);
for (i = 0; i + 4 <= num_pixels; i += 4) {
__m128i T = _mm_loadu_si128((const __m128i*)&upper[i]);
__m128i TL = _mm_loadu_si128((const __m128i*)&upper[i - 1]);
__m128i src = _mm_loadu_si128((const __m128i*)&in[i]);
__m128i pa;
GetSumAbsDiff32(&T, &TL, &pa); // pa = sum |T-TL|
for (j = 0; j < 4; ++j) {
const __m128i L_lo = _mm_unpacklo_epi32(L, L);
const __m128i TL_lo = _mm_unpacklo_epi32(TL, L);
const __m128i pb = _mm_sad_epu8(L_lo, TL_lo); // pb = sum |L-TL|
const __m128i mask = _mm_cmpgt_epi32(pb, pa);
const __m128i A = _mm_and_si128(mask, L);
const __m128i B = _mm_andnot_si128(mask, T);
const __m128i pred = _mm_or_si128(A, B); // pred = (L > T)? L : T
L = _mm_add_epi8(src, pred);
out[i + j] = _mm_cvtsi128_si32(L);
// Shift the pre-computed value for the next iteration.
T = _mm_srli_si128(T, 4);
TL = _mm_srli_si128(TL, 4);
src = _mm_srli_si128(src, 4);
pa = _mm_srli_si128(pa, 4);
}
}
if (i != num_pixels) {
VP8LPredictorsAdd_C[11](in + i, upper + i, num_pixels - i, out + i);
}
}
// Predictor12: ClampedAddSubtractFull.
#define DO_PRED12(DIFF, LANE, OUT) \
do { \
const __m128i all = _mm_add_epi16(L, (DIFF)); \
const __m128i alls = _mm_packus_epi16(all, all); \
const __m128i res = _mm_add_epi8(src, alls); \
out[i + (OUT)] = _mm_cvtsi128_si32(res); \
L = _mm_unpacklo_epi8(res, zero); \
/* Shift the pre-computed value for the next iteration.*/ \
if (LANE == 0) (DIFF) = _mm_srli_si128((DIFF), 8); \
src = _mm_srli_si128(src, 4); \
} while (0)
static void PredictorAdd12_SSE2(const uint32_t* in, const uint32_t* upper,
int num_pixels, uint32_t* out) {
int i;
const __m128i zero = _mm_setzero_si128();
const __m128i L8 = _mm_cvtsi32_si128(out[-1]);
__m128i L = _mm_unpacklo_epi8(L8, zero);
for (i = 0; i + 4 <= num_pixels; i += 4) {
// Load 4 pixels at a time.
__m128i src = _mm_loadu_si128((const __m128i*)&in[i]);
const __m128i T = _mm_loadu_si128((const __m128i*)&upper[i]);
const __m128i T_lo = _mm_unpacklo_epi8(T, zero);
const __m128i T_hi = _mm_unpackhi_epi8(T, zero);
const __m128i TL = _mm_loadu_si128((const __m128i*)&upper[i - 1]);
const __m128i TL_lo = _mm_unpacklo_epi8(TL, zero);
const __m128i TL_hi = _mm_unpackhi_epi8(TL, zero);
__m128i diff_lo = _mm_sub_epi16(T_lo, TL_lo);
__m128i diff_hi = _mm_sub_epi16(T_hi, TL_hi);
DO_PRED12(diff_lo, 0, 0);
DO_PRED12(diff_lo, 1, 1);
DO_PRED12(diff_hi, 0, 2);
DO_PRED12(diff_hi, 1, 3);
}
if (i != num_pixels) {
VP8LPredictorsAdd_C[12](in + i, upper + i, num_pixels - i, out + i);
}
}
#undef DO_PRED12
// Due to averages with integers, values cannot be accumulated in parallel for
// predictors 13.
GENERATE_PREDICTOR_ADD(Predictor13_SSE2, PredictorAdd13_SSE2)
//------------------------------------------------------------------------------
// Subtract-Green Transform
static void AddGreenToBlueAndRed(uint32_t* argb_data, int num_pixels) {
static void AddGreenToBlueAndRed(const uint32_t* const src, int num_pixels,
uint32_t* dst) {
int i;
for (i = 0; i + 4 <= num_pixels; i += 4) {
const __m128i in = _mm_loadu_si128((__m128i*)&argb_data[i]); // argb
const __m128i in = _mm_loadu_si128((const __m128i*)&src[i]); // argb
const __m128i A = _mm_srli_epi16(in, 8); // 0 a 0 g
const __m128i B = _mm_shufflelo_epi16(A, _MM_SHUFFLE(2, 2, 0, 0));
const __m128i C = _mm_shufflehi_epi16(B, _MM_SHUFFLE(2, 2, 0, 0)); // 0g0g
const __m128i out = _mm_add_epi8(in, C);
_mm_storeu_si128((__m128i*)&argb_data[i], out);
_mm_storeu_si128((__m128i*)&dst[i], out);
}
// fallthrough and finish off with plain-C
VP8LAddGreenToBlueAndRed_C(argb_data + i, num_pixels - i);
if (i != num_pixels) {
VP8LAddGreenToBlueAndRed_C(src + i, num_pixels - i, dst + i);
}
}
//------------------------------------------------------------------------------
// Color Transform
static void TransformColorInverse(const VP8LMultipliers* const m,
uint32_t* argb_data, int num_pixels) {
// sign-extended multiplying constants, pre-shifted by 5.
const uint32_t* const src, int num_pixels,
uint32_t* dst) {
// sign-extended multiplying constants, pre-shifted by 5.
#define CST(X) (((int16_t)(m->X << 8)) >> 5) // sign-extend
const __m128i mults_rb = _mm_set_epi16(
CST(green_to_red_), CST(green_to_blue_),
@ -188,7 +431,7 @@ static void TransformColorInverse(const VP8LMultipliers* const m,
const __m128i mask_ag = _mm_set1_epi32(0xff00ff00); // alpha-green masks
int i;
for (i = 0; i + 4 <= num_pixels; i += 4) {
const __m128i in = _mm_loadu_si128((__m128i*)&argb_data[i]); // argb
const __m128i in = _mm_loadu_si128((const __m128i*)&src[i]); // argb
const __m128i A = _mm_and_si128(in, mask_ag); // a 0 g 0
const __m128i B = _mm_shufflelo_epi16(A, _MM_SHUFFLE(2, 2, 0, 0));
const __m128i C = _mm_shufflehi_epi16(B, _MM_SHUFFLE(2, 2, 0, 0)); // g0g0
@ -200,15 +443,53 @@ static void TransformColorInverse(const VP8LMultipliers* const m,
const __m128i I = _mm_add_epi8(H, F); // r' x b'' 0
const __m128i J = _mm_srli_epi16(I, 8); // 0 r' 0 b''
const __m128i out = _mm_or_si128(J, A);
_mm_storeu_si128((__m128i*)&argb_data[i], out);
_mm_storeu_si128((__m128i*)&dst[i], out);
}
// Fall-back to C-version for left-overs.
VP8LTransformColorInverse_C(m, argb_data + i, num_pixels - i);
if (i != num_pixels) {
VP8LTransformColorInverse_C(m, src + i, num_pixels - i, dst + i);
}
}
//------------------------------------------------------------------------------
// Color-space conversion functions
static void ConvertBGRAToRGB(const uint32_t* src, int num_pixels,
uint8_t* dst) {
const __m128i* in = (const __m128i*)src;
__m128i* out = (__m128i*)dst;
while (num_pixels >= 32) {
// Load the BGRA buffers.
__m128i in0 = _mm_loadu_si128(in + 0);
__m128i in1 = _mm_loadu_si128(in + 1);
__m128i in2 = _mm_loadu_si128(in + 2);
__m128i in3 = _mm_loadu_si128(in + 3);
__m128i in4 = _mm_loadu_si128(in + 4);
__m128i in5 = _mm_loadu_si128(in + 5);
__m128i in6 = _mm_loadu_si128(in + 6);
__m128i in7 = _mm_loadu_si128(in + 7);
VP8L32bToPlanar(&in0, &in1, &in2, &in3);
VP8L32bToPlanar(&in4, &in5, &in6, &in7);
// At this points, in1/in5 contains red only, in2/in6 green only ...
// Pack the colors in 24b RGB.
VP8PlanarTo24b(&in1, &in5, &in2, &in6, &in3, &in7);
_mm_storeu_si128(out + 0, in1);
_mm_storeu_si128(out + 1, in5);
_mm_storeu_si128(out + 2, in2);
_mm_storeu_si128(out + 3, in6);
_mm_storeu_si128(out + 4, in3);
_mm_storeu_si128(out + 5, in7);
in += 8;
out += 6;
num_pixels -= 32;
}
// left-overs
if (num_pixels > 0) {
VP8LConvertBGRAToRGB_C((const uint32_t*)in, num_pixels, (uint8_t*)out);
}
}
static void ConvertBGRAToRGBA(const uint32_t* src,
int num_pixels, uint8_t* dst) {
const __m128i* in = (const __m128i*)src;
@ -233,7 +514,9 @@ static void ConvertBGRAToRGBA(const uint32_t* src,
num_pixels -= 8;
}
// left-overs
VP8LConvertBGRAToRGBA_C((const uint32_t*)in, num_pixels, (uint8_t*)out);
if (num_pixels > 0) {
VP8LConvertBGRAToRGBA_C((const uint32_t*)in, num_pixels, (uint8_t*)out);
}
}
static void ConvertBGRAToRGBA4444(const uint32_t* src,
@ -267,7 +550,9 @@ static void ConvertBGRAToRGBA4444(const uint32_t* src,
num_pixels -= 8;
}
// left-overs
VP8LConvertBGRAToRGBA4444_C((const uint32_t*)in, num_pixels, (uint8_t*)out);
if (num_pixels > 0) {
VP8LConvertBGRAToRGBA4444_C((const uint32_t*)in, num_pixels, (uint8_t*)out);
}
}
static void ConvertBGRAToRGB565(const uint32_t* src,
@ -306,7 +591,9 @@ static void ConvertBGRAToRGB565(const uint32_t* src,
num_pixels -= 8;
}
// left-overs
VP8LConvertBGRAToRGB565_C((const uint32_t*)in, num_pixels, (uint8_t*)out);
if (num_pixels > 0) {
VP8LConvertBGRAToRGB565_C((const uint32_t*)in, num_pixels, (uint8_t*)out);
}
}
static void ConvertBGRAToBGR(const uint32_t* src,
@ -337,7 +624,9 @@ static void ConvertBGRAToBGR(const uint32_t* src,
num_pixels -= 8;
}
// left-overs
VP8LConvertBGRAToBGR_C((const uint32_t*)in, num_pixels, dst);
if (num_pixels > 0) {
VP8LConvertBGRAToBGR_C((const uint32_t*)in, num_pixels, dst);
}
}
//------------------------------------------------------------------------------
@ -346,19 +635,35 @@ static void ConvertBGRAToBGR(const uint32_t* src,
extern void VP8LDspInitSSE2(void);
WEBP_TSAN_IGNORE_FUNCTION void VP8LDspInitSSE2(void) {
VP8LPredictors[5] = Predictor5;
VP8LPredictors[6] = Predictor6;
VP8LPredictors[7] = Predictor7;
VP8LPredictors[8] = Predictor8;
VP8LPredictors[9] = Predictor9;
VP8LPredictors[10] = Predictor10;
VP8LPredictors[11] = Predictor11;
VP8LPredictors[12] = Predictor12;
VP8LPredictors[13] = Predictor13;
VP8LPredictors[5] = Predictor5_SSE2;
VP8LPredictors[6] = Predictor6_SSE2;
VP8LPredictors[7] = Predictor7_SSE2;
VP8LPredictors[8] = Predictor8_SSE2;
VP8LPredictors[9] = Predictor9_SSE2;
VP8LPredictors[10] = Predictor10_SSE2;
VP8LPredictors[11] = Predictor11_SSE2;
VP8LPredictors[12] = Predictor12_SSE2;
VP8LPredictors[13] = Predictor13_SSE2;
VP8LPredictorsAdd[0] = PredictorAdd0_SSE2;
VP8LPredictorsAdd[1] = PredictorAdd1_SSE2;
VP8LPredictorsAdd[2] = PredictorAdd2_SSE2;
VP8LPredictorsAdd[3] = PredictorAdd3_SSE2;
VP8LPredictorsAdd[4] = PredictorAdd4_SSE2;
VP8LPredictorsAdd[5] = PredictorAdd5_SSE2;
VP8LPredictorsAdd[6] = PredictorAdd6_SSE2;
VP8LPredictorsAdd[7] = PredictorAdd7_SSE2;
VP8LPredictorsAdd[8] = PredictorAdd8_SSE2;
VP8LPredictorsAdd[9] = PredictorAdd9_SSE2;
VP8LPredictorsAdd[10] = PredictorAdd10_SSE2;
VP8LPredictorsAdd[11] = PredictorAdd11_SSE2;
VP8LPredictorsAdd[12] = PredictorAdd12_SSE2;
VP8LPredictorsAdd[13] = PredictorAdd13_SSE2;
VP8LAddGreenToBlueAndRed = AddGreenToBlueAndRed;
VP8LTransformColorInverse = TransformColorInverse;
VP8LConvertBGRAToRGB = ConvertBGRAToRGB;
VP8LConvertBGRAToRGBA = ConvertBGRAToRGBA;
VP8LConvertBGRAToRGBA4444 = ConvertBGRAToRGBA4444;
VP8LConvertBGRAToRGB565 = ConvertBGRAToRGB565;

1097
thirdparty/libwebp/dsp/msa_macro.h vendored
View file

File diff suppressed because it is too large Load diff

18
thirdparty/libwebp/dsp/neon.h vendored
View file

@ -79,4 +79,22 @@ static WEBP_INLINE int32x4x4_t Transpose4x4(const int32x4x4_t rows) {
}
}
#if 0 // Useful debug macro.
#include <stdio.h>
#define PRINT_REG(REG, SIZE) do { \
int i; \
printf("%s \t[%d]: 0x", #REG, SIZE); \
if (SIZE == 8) { \
uint8_t _tmp[8]; \
vst1_u8(_tmp, (REG)); \
for (i = 0; i < 8; ++i) printf("%.2x ", _tmp[i]); \
} else if (SIZE == 16) { \
uint16_t _tmp[4]; \
vst1_u16(_tmp, (REG)); \
for (i = 0; i < 4; ++i) printf("%.4x ", _tmp[i]); \
} \
printf("\n"); \
} while (0)
#endif
#endif // WEBP_DSP_NEON_H_

8
thirdparty/libwebp/dsp/rescaler.c vendored
View file

@ -14,7 +14,7 @@
#include <assert.h>
#include "./dsp.h"
#include "../utils/rescaler.h"
#include "../utils/rescaler_utils.h"
//------------------------------------------------------------------------------
// Implementations of critical functions ImportRow / ExportRow
@ -199,6 +199,7 @@ WebPRescalerExportRowFunc WebPRescalerExportRowShrink;
extern void WebPRescalerDspInitSSE2(void);
extern void WebPRescalerDspInitMIPS32(void);
extern void WebPRescalerDspInitMIPSdspR2(void);
extern void WebPRescalerDspInitMSA(void);
extern void WebPRescalerDspInitNEON(void);
static volatile VP8CPUInfo rescaler_last_cpuinfo_used =
@ -232,6 +233,11 @@ WEBP_TSAN_IGNORE_FUNCTION void WebPRescalerDspInit(void) {
if (VP8GetCPUInfo(kMIPSdspR2)) {
WebPRescalerDspInitMIPSdspR2();
}
#endif
#if defined(WEBP_USE_MSA)
if (VP8GetCPUInfo(kMSA)) {
WebPRescalerDspInitMSA();
}
#endif
}
rescaler_last_cpuinfo_used = VP8GetCPUInfo;

2
thirdparty/libwebp/dsp/rescaler_mips32.c vendored
View file

@ -16,7 +16,7 @@
#if defined(WEBP_USE_MIPS32)
#include <assert.h>
#include "../utils/rescaler.h"
#include "../utils/rescaler_utils.h"
//------------------------------------------------------------------------------
// Row import

2
thirdparty/libwebp/dsp/rescaler_mips_dsp_r2.c vendored
View file

@ -16,7 +16,7 @@
#if defined(WEBP_USE_MIPS_DSP_R2)
#include <assert.h>
#include "../utils/rescaler.h"
#include "../utils/rescaler_utils.h"
#define ROUNDER (WEBP_RESCALER_ONE >> 1)
#define MULT_FIX(x, y) (((uint64_t)(x) * (y) + ROUNDER) >> WEBP_RESCALER_RFIX)

444
thirdparty/libwebp/dsp/rescaler_msa.c vendored Normal file
View file

@ -0,0 +1,444 @@
// Copyright 2016 Google Inc. All Rights Reserved.
//
// Use of this source code is governed by a BSD-style license
// that can be found in the COPYING file in the root of the source
// tree. An additional intellectual property rights grant can be found
// in the file PATENTS. All contributing project authors may
// be found in the AUTHORS file in the root of the source tree.
// -----------------------------------------------------------------------------
//
// MSA version of rescaling functions
//
// Author: Prashant Patil (prashant.patil@imgtec.com)
#include "./dsp.h"
#if defined(WEBP_USE_MSA)
#include <assert.h>
#include "../utils/rescaler_utils.h"
#include "./msa_macro.h"
#define ROUNDER (WEBP_RESCALER_ONE >> 1)
#define MULT_FIX(x, y) (((uint64_t)(x) * (y) + ROUNDER) >> WEBP_RESCALER_RFIX)
#define CALC_MULT_FIX_16(in0, in1, in2, in3, scale, shift, dst) do { \
v4u32 tmp0, tmp1, tmp2, tmp3; \
v16u8 t0, t1, t2, t3, t4, t5; \
v2u64 out0, out1, out2, out3; \
ILVRL_W2_UW(zero, in0, tmp0, tmp1); \
ILVRL_W2_UW(zero, in1, tmp2, tmp3); \
DOTP_UW2_UD(tmp0, tmp1, scale, scale, out0, out1); \
DOTP_UW2_UD(tmp2, tmp3, scale, scale, out2, out3); \
SRAR_D4_UD(out0, out1, out2, out3, shift); \
PCKEV_B2_UB(out1, out0, out3, out2, t0, t1); \
ILVRL_W2_UW(zero, in2, tmp0, tmp1); \
ILVRL_W2_UW(zero, in3, tmp2, tmp3); \
DOTP_UW2_UD(tmp0, tmp1, scale, scale, out0, out1); \
DOTP_UW2_UD(tmp2, tmp3, scale, scale, out2, out3); \
SRAR_D4_UD(out0, out1, out2, out3, shift); \
PCKEV_B2_UB(out1, out0, out3, out2, t2, t3); \
PCKEV_B2_UB(t1, t0, t3, t2, t4, t5); \
dst = (v16u8)__msa_pckev_b((v16i8)t5, (v16i8)t4); \
} while (0)
#define CALC_MULT_FIX_4(in0, scale, shift, dst) do { \
v4u32 tmp0, tmp1; \
v16i8 t0, t1; \
v2u64 out0, out1; \
ILVRL_W2_UW(zero, in0, tmp0, tmp1); \
DOTP_UW2_UD(tmp0, tmp1, scale, scale, out0, out1); \
SRAR_D2_UD(out0, out1, shift); \
t0 = __msa_pckev_b((v16i8)out1, (v16i8)out0); \
t1 = __msa_pckev_b(t0, t0); \
t0 = __msa_pckev_b(t1, t1); \
dst = __msa_copy_s_w((v4i32)t0, 0); \
} while (0)
#define CALC_MULT_FIX1_16(in0, in1, in2, in3, fyscale, shift, \
dst0, dst1, dst2, dst3) do { \
v4u32 tmp0, tmp1, tmp2, tmp3; \
v2u64 out0, out1, out2, out3; \
ILVRL_W2_UW(zero, in0, tmp0, tmp1); \
ILVRL_W2_UW(zero, in1, tmp2, tmp3); \
DOTP_UW2_UD(tmp0, tmp1, fyscale, fyscale, out0, out1); \
DOTP_UW2_UD(tmp2, tmp3, fyscale, fyscale, out2, out3); \
SRAR_D4_UD(out0, out1, out2, out3, shift); \
PCKEV_W2_UW(out1, out0, out3, out2, dst0, dst1); \
ILVRL_W2_UW(zero, in2, tmp0, tmp1); \
ILVRL_W2_UW(zero, in3, tmp2, tmp3); \
DOTP_UW2_UD(tmp0, tmp1, fyscale, fyscale, out0, out1); \
DOTP_UW2_UD(tmp2, tmp3, fyscale, fyscale, out2, out3); \
SRAR_D4_UD(out0, out1, out2, out3, shift); \
PCKEV_W2_UW(out1, out0, out3, out2, dst2, dst3); \
} while (0)
#define CALC_MULT_FIX1_4(in0, scale, shift, dst) do { \
v4u32 tmp0, tmp1; \
v2u64 out0, out1; \
ILVRL_W2_UW(zero, in0, tmp0, tmp1); \
DOTP_UW2_UD(tmp0, tmp1, scale, scale, out0, out1); \
SRAR_D2_UD(out0, out1, shift); \
dst = (v4u32)__msa_pckev_w((v4i32)out1, (v4i32)out0); \
} while (0)
#define CALC_MULT_FIX2_16(in0, in1, in2, in3, mult, scale, shift, \
dst0, dst1) do { \
v4u32 tmp0, tmp1, tmp2, tmp3; \
v2u64 out0, out1, out2, out3; \
ILVRL_W2_UW(in0, in2, tmp0, tmp1); \
ILVRL_W2_UW(in1, in3, tmp2, tmp3); \
DOTP_UW2_UD(tmp0, tmp1, mult, mult, out0, out1); \
DOTP_UW2_UD(tmp2, tmp3, mult, mult, out2, out3); \
SRAR_D4_UD(out0, out1, out2, out3, shift); \
DOTP_UW2_UD(out0, out1, scale, scale, out0, out1); \
DOTP_UW2_UD(out2, out3, scale, scale, out2, out3); \
SRAR_D4_UD(out0, out1, out2, out3, shift); \
PCKEV_B2_UB(out1, out0, out3, out2, dst0, dst1); \
} while (0)
#define CALC_MULT_FIX2_4(in0, in1, mult, scale, shift, dst) do { \
v4u32 tmp0, tmp1; \
v2u64 out0, out1; \
v16i8 t0, t1; \
ILVRL_W2_UW(in0, in1, tmp0, tmp1); \
DOTP_UW2_UD(tmp0, tmp1, mult, mult, out0, out1); \
SRAR_D2_UD(out0, out1, shift); \
DOTP_UW2_UD(out0, out1, scale, scale, out0, out1); \
SRAR_D2_UD(out0, out1, shift); \
t0 = __msa_pckev_b((v16i8)out1, (v16i8)out0); \
t1 = __msa_pckev_b(t0, t0); \
t0 = __msa_pckev_b(t1, t1); \
dst = __msa_copy_s_w((v4i32)t0, 0); \
} while (0)
static WEBP_INLINE void ExportRowExpand_0(const uint32_t* frow, uint8_t* dst,
int length,
WebPRescaler* const wrk) {
const v4u32 scale = (v4u32)__msa_fill_w(wrk->fy_scale);
const v4u32 shift = (v4u32)__msa_fill_w(WEBP_RESCALER_RFIX);
const v4i32 zero = { 0 };
while (length >= 16) {
v4u32 src0, src1, src2, src3;
v16u8 out;
LD_UW4(frow, 4, src0, src1, src2, src3);
CALC_MULT_FIX_16(src0, src1, src2, src3, scale, shift, out);
ST_UB(out, dst);
length -= 16;
frow += 16;
dst += 16;
}
if (length > 0) {
int x_out;
if (length >= 12) {
uint32_t val0_m, val1_m, val2_m;
v4u32 src0, src1, src2;
LD_UW3(frow, 4, src0, src1, src2);
CALC_MULT_FIX_4(src0, scale, shift, val0_m);
CALC_MULT_FIX_4(src1, scale, shift, val1_m);
CALC_MULT_FIX_4(src2, scale, shift, val2_m);
SW3(val0_m, val1_m, val2_m, dst, 4);
length -= 12;
frow += 12;
dst += 12;
} else if (length >= 8) {
uint32_t val0_m, val1_m;
v4u32 src0, src1;
LD_UW2(frow, 4, src0, src1);
CALC_MULT_FIX_4(src0, scale, shift, val0_m);
CALC_MULT_FIX_4(src1, scale, shift, val1_m);
SW2(val0_m, val1_m, dst, 4);
length -= 8;
frow += 8;
dst += 8;
} else if (length >= 4) {
uint32_t val0_m;
const v4u32 src0 = LD_UW(frow);
CALC_MULT_FIX_4(src0, scale, shift, val0_m);
SW(val0_m, dst);
length -= 4;
frow += 4;
dst += 4;
}
for (x_out = 0; x_out < length; ++x_out) {
const uint32_t J = frow[x_out];
const int v = (int)MULT_FIX(J, wrk->fy_scale);
assert(v >= 0 && v <= 255);
dst[x_out] = v;
}
}
}
static WEBP_INLINE void ExportRowExpand_1(const uint32_t* frow, uint32_t* irow,
uint8_t* dst, int length,
WebPRescaler* const wrk) {
const uint32_t B = WEBP_RESCALER_FRAC(-wrk->y_accum, wrk->y_sub);
const uint32_t A = (uint32_t)(WEBP_RESCALER_ONE - B);
const v4i32 B1 = __msa_fill_w(B);
const v4i32 A1 = __msa_fill_w(A);
const v4i32 AB = __msa_ilvr_w(A1, B1);
const v4u32 scale = (v4u32)__msa_fill_w(wrk->fy_scale);
const v4u32 shift = (v4u32)__msa_fill_w(WEBP_RESCALER_RFIX);
while (length >= 16) {
v4u32 frow0, frow1, frow2, frow3, irow0, irow1, irow2, irow3;
v16u8 t0, t1, t2, t3, t4, t5;
LD_UW4(frow, 4, frow0, frow1, frow2, frow3);
LD_UW4(irow, 4, irow0, irow1, irow2, irow3);
CALC_MULT_FIX2_16(frow0, frow1, irow0, irow1, AB, scale, shift, t0, t1);
CALC_MULT_FIX2_16(frow2, frow3, irow2, irow3, AB, scale, shift, t2, t3);
PCKEV_B2_UB(t1, t0, t3, t2, t4, t5);
t0 = (v16u8)__msa_pckev_b((v16i8)t5, (v16i8)t4);
ST_UB(t0, dst);
frow += 16;
irow += 16;
dst += 16;
length -= 16;
}
if (length > 0) {
int x_out;
if (length >= 12) {
uint32_t val0_m, val1_m, val2_m;
v4u32 frow0, frow1, frow2, irow0, irow1, irow2;
LD_UW3(frow, 4, frow0, frow1, frow2);
LD_UW3(irow, 4, irow0, irow1, irow2);
CALC_MULT_FIX2_4(frow0, irow0, AB, scale, shift, val0_m);
CALC_MULT_FIX2_4(frow1, irow1, AB, scale, shift, val1_m);
CALC_MULT_FIX2_4(frow2, irow2, AB, scale, shift, val2_m);
SW3(val0_m, val1_m, val2_m, dst, 4);
frow += 12;
irow += 12;
dst += 12;
length -= 12;
} else if (length >= 8) {
uint32_t val0_m, val1_m;
v4u32 frow0, frow1, irow0, irow1;
LD_UW2(frow, 4, frow0, frow1);
LD_UW2(irow, 4, irow0, irow1);
CALC_MULT_FIX2_4(frow0, irow0, AB, scale, shift, val0_m);
CALC_MULT_FIX2_4(frow1, irow1, AB, scale, shift, val1_m);
SW2(val0_m, val1_m, dst, 4);
frow += 4;
irow += 4;
dst += 4;
length -= 4;
} else if (length >= 4) {
uint32_t val0_m;
const v4u32 frow0 = LD_UW(frow + 0);
const v4u32 irow0 = LD_UW(irow + 0);
CALC_MULT_FIX2_4(frow0, irow0, AB, scale, shift, val0_m);
SW(val0_m, dst);
frow += 4;
irow += 4;
dst += 4;
length -= 4;
}
for (x_out = 0; x_out < length; ++x_out) {
const uint64_t I = (uint64_t)A * frow[x_out]
+ (uint64_t)B * irow[x_out];
const uint32_t J = (uint32_t)((I + ROUNDER) >> WEBP_RESCALER_RFIX);
const int v = (int)MULT_FIX(J, wrk->fy_scale);
assert(v >= 0 && v <= 255);
dst[x_out] = v;
}
}
}
static void RescalerExportRowExpand(WebPRescaler* const wrk) {
uint8_t* dst = wrk->dst;
rescaler_t* irow = wrk->irow;
const int x_out_max = wrk->dst_width * wrk->num_channels;
const rescaler_t* frow = wrk->frow;
assert(!WebPRescalerOutputDone(wrk));
assert(wrk->y_accum <= 0);
assert(wrk->y_expand);
assert(wrk->y_sub != 0);
if (wrk->y_accum == 0) {
ExportRowExpand_0(frow, dst, x_out_max, wrk);
} else {
ExportRowExpand_1(frow, irow, dst, x_out_max, wrk);
}
}
static WEBP_INLINE void ExportRowShrink_0(const uint32_t* frow, uint32_t* irow,
uint8_t* dst, int length,
const uint32_t yscale,
WebPRescaler* const wrk) {
const v4u32 y_scale = (v4u32)__msa_fill_w(yscale);
const v4u32 fxyscale = (v4u32)__msa_fill_w(wrk->fxy_scale);
const v4u32 shiftval = (v4u32)__msa_fill_w(WEBP_RESCALER_RFIX);
const v4i32 zero = { 0 };
while (length >= 16) {
v4u32 src0, src1, src2, src3, frac0, frac1, frac2, frac3;
v16u8 out;
LD_UW4(frow, 4, src0, src1, src2, src3);
CALC_MULT_FIX1_16(src0, src1, src2, src3, y_scale, shiftval,
frac0, frac1, frac2, frac3);
LD_UW4(irow, 4, src0, src1, src2, src3);
SUB4(src0, frac0, src1, frac1, src2, frac2, src3, frac3,
src0, src1, src2, src3);
CALC_MULT_FIX_16(src0, src1, src2, src3, fxyscale, shiftval, out);
ST_UB(out, dst);
ST_UW4(frac0, frac1, frac2, frac3, irow, 4);
frow += 16;
irow += 16;
dst += 16;
length -= 16;
}
if (length > 0) {
int x_out;
if (length >= 12) {
uint32_t val0_m, val1_m, val2_m;
v4u32 src0, src1, src2, frac0, frac1, frac2;
LD_UW3(frow, 4, src0, src1, src2);
CALC_MULT_FIX1_4(src0, y_scale, shiftval, frac0);
CALC_MULT_FIX1_4(src1, y_scale, shiftval, frac1);
CALC_MULT_FIX1_4(src2, y_scale, shiftval, frac2);
LD_UW3(irow, 4, src0, src1, src2);
SUB3(src0, frac0, src1, frac1, src2, frac2, src0, src1, src2);
CALC_MULT_FIX_4(src0, fxyscale, shiftval, val0_m);
CALC_MULT_FIX_4(src1, fxyscale, shiftval, val1_m);
CALC_MULT_FIX_4(src2, fxyscale, shiftval, val2_m);
SW3(val0_m, val1_m, val2_m, dst, 4);
ST_UW3(frac0, frac1, frac2, irow, 4);
frow += 12;
irow += 12;
dst += 12;
length -= 12;
} else if (length >= 8) {
uint32_t val0_m, val1_m;
v4u32 src0, src1, frac0, frac1;
LD_UW2(frow, 4, src0, src1);
CALC_MULT_FIX1_4(src0, y_scale, shiftval, frac0);
CALC_MULT_FIX1_4(src1, y_scale, shiftval, frac1);
LD_UW2(irow, 4, src0, src1);
SUB2(src0, frac0, src1, frac1, src0, src1);
CALC_MULT_FIX_4(src0, fxyscale, shiftval, val0_m);
CALC_MULT_FIX_4(src1, fxyscale, shiftval, val1_m);
SW2(val0_m, val1_m, dst, 4);
ST_UW2(frac0, frac1, irow, 4);
frow += 8;
irow += 8;
dst += 8;
length -= 8;
} else if (length >= 4) {
uint32_t val0_m;
v4u32 frac0;
v4u32 src0 = LD_UW(frow);
CALC_MULT_FIX1_4(src0, y_scale, shiftval, frac0);
src0 = LD_UW(irow);
src0 = src0 - frac0;
CALC_MULT_FIX_4(src0, fxyscale, shiftval, val0_m);
SW(val0_m, dst);
ST_UW(frac0, irow);
frow += 4;
irow += 4;
dst += 4;
length -= 4;
}
for (x_out = 0; x_out < length; ++x_out) {
const uint32_t frac = (uint32_t)MULT_FIX(frow[x_out], yscale);
const int v = (int)MULT_FIX(irow[x_out] - frac, wrk->fxy_scale);
assert(v >= 0 && v <= 255);
dst[x_out] = v;
irow[x_out] = frac;
}
}
}
static WEBP_INLINE void ExportRowShrink_1(uint32_t* irow, uint8_t* dst,
int length,
WebPRescaler* const wrk) {
const v4u32 scale = (v4u32)__msa_fill_w(wrk->fxy_scale);
const v4u32 shift = (v4u32)__msa_fill_w(WEBP_RESCALER_RFIX);
const v4i32 zero = { 0 };
while (length >= 16) {
v4u32 src0, src1, src2, src3;
v16u8 dst0;
LD_UW4(irow, 4, src0, src1, src2, src3);
CALC_MULT_FIX_16(src0, src1, src2, src3, scale, shift, dst0);
ST_UB(dst0, dst);
ST_SW4(zero, zero, zero, zero, irow, 4);
length -= 16;
irow += 16;
dst += 16;
}
if (length > 0) {
int x_out;
if (length >= 12) {
uint32_t val0_m, val1_m, val2_m;
v4u32 src0, src1, src2;
LD_UW3(irow, 4, src0, src1, src2);
CALC_MULT_FIX_4(src0, scale, shift, val0_m);
CALC_MULT_FIX_4(src1, scale, shift, val1_m);
CALC_MULT_FIX_4(src2, scale, shift, val2_m);
SW3(val0_m, val1_m, val2_m, dst, 4);
ST_SW3(zero, zero, zero, irow, 4);
length -= 12;
irow += 12;
dst += 12;
} else if (length >= 8) {
uint32_t val0_m, val1_m;
v4u32 src0, src1;
LD_UW2(irow, 4, src0, src1);
CALC_MULT_FIX_4(src0, scale, shift, val0_m);
CALC_MULT_FIX_4(src1, scale, shift, val1_m);
SW2(val0_m, val1_m, dst, 4);
ST_SW2(zero, zero, irow, 4);
length -= 8;
irow += 8;
dst += 8;
} else if (length >= 4) {
uint32_t val0_m;
const v4u32 src0 = LD_UW(irow + 0);
CALC_MULT_FIX_4(src0, scale, shift, val0_m);
SW(val0_m, dst);
ST_SW(zero, irow);
length -= 4;
irow += 4;
dst += 4;
}
for (x_out = 0; x_out < length; ++x_out) {
const int v = (int)MULT_FIX(irow[x_out], wrk->fxy_scale);
assert(v >= 0 && v <= 255);
dst[x_out] = v;
irow[x_out] = 0;
}
}
}
static void RescalerExportRowShrink(WebPRescaler* const wrk) {
uint8_t* dst = wrk->dst;
rescaler_t* irow = wrk->irow;
const int x_out_max = wrk->dst_width * wrk->num_channels;
const rescaler_t* frow = wrk->frow;
const uint32_t yscale = wrk->fy_scale * (-wrk->y_accum);
assert(!WebPRescalerOutputDone(wrk));
assert(wrk->y_accum <= 0);
assert(!wrk->y_expand);
if (yscale) {
ExportRowShrink_0(frow, irow, dst, x_out_max, yscale, wrk);
} else {
ExportRowShrink_1(irow, dst, x_out_max, wrk);
}
}
//------------------------------------------------------------------------------
// Entry point
extern void WebPRescalerDspInitMSA(void);
WEBP_TSAN_IGNORE_FUNCTION void WebPRescalerDspInitMSA(void) {
WebPRescalerExportRowExpand = RescalerExportRowExpand;
WebPRescalerExportRowShrink = RescalerExportRowShrink;
}
#else // !WEBP_USE_MSA
WEBP_DSP_INIT_STUB(WebPRescalerDspInitMSA)
#endif // WEBP_USE_MSA

2
thirdparty/libwebp/dsp/rescaler_neon.c vendored
View file

@ -18,7 +18,7 @@
#include <arm_neon.h>
#include <assert.h>
#include "./neon.h"
#include "../utils/rescaler.h"
#include "../utils/rescaler_utils.h"
#define ROUNDER (WEBP_RESCALER_ONE >> 1)
#define MULT_FIX_C(x, y) (((uint64_t)(x) * (y) + ROUNDER) >> WEBP_RESCALER_RFIX)

2
thirdparty/libwebp/dsp/rescaler_sse2.c vendored
View file

@ -17,7 +17,7 @@
#include <emmintrin.h>
#include <assert.h>
#include "../utils/rescaler.h"
#include "../utils/rescaler_utils.h"
#include "../utils/utils.h"
//------------------------------------------------------------------------------

6
thirdparty/libwebp/dsp/upsampling.c vendored
View file

@ -215,6 +215,7 @@ WEBP_TSAN_IGNORE_FUNCTION void WebPInitYUV444Converters(void) {
extern void WebPInitUpsamplersSSE2(void);
extern void WebPInitUpsamplersNEON(void);
extern void WebPInitUpsamplersMIPSdspR2(void);
extern void WebPInitUpsamplersMSA(void);
static volatile VP8CPUInfo upsampling_last_cpuinfo_used2 =
(VP8CPUInfo)&upsampling_last_cpuinfo_used2;
@ -251,6 +252,11 @@ WEBP_TSAN_IGNORE_FUNCTION void WebPInitUpsamplers(void) {
if (VP8GetCPUInfo(kMIPSdspR2)) {
WebPInitUpsamplersMIPSdspR2();
}
#endif
#if defined(WEBP_USE_MSA)
if (VP8GetCPUInfo(kMSA)) {
WebPInitUpsamplersMSA();
}
#endif
}
#endif // FANCY_UPSAMPLING

678
thirdparty/libwebp/dsp/upsampling_msa.c vendored Normal file
View file

@ -0,0 +1,678 @@
// Copyright 2016 Google Inc. All Rights Reserved.
//
// Use of this source code is governed by a BSD-style license
// that can be found in the COPYING file in the root of the source
// tree. An additional intellectual property rights grant can be found
// in the file PATENTS. All contributing project authors may
// be found in the AUTHORS file in the root of the source tree.
// -----------------------------------------------------------------------------
//
// MSA version of YUV to RGB upsampling functions.
//
// Author: Prashant Patil (prashant.patil@imgtec.com)
#include <string.h>
#include "./dsp.h"
#if defined(WEBP_USE_MSA)
#include "./msa_macro.h"
#include "./yuv.h"
#ifdef FANCY_UPSAMPLING
#define ILVR_UW2(in, out0, out1) do { \
const v8i16 t0 = (v8i16)__msa_ilvr_b((v16i8)zero, (v16i8)in); \
out0 = (v4u32)__msa_ilvr_h((v8i16)zero, t0); \
out1 = (v4u32)__msa_ilvl_h((v8i16)zero, t0); \
} while (0)
#define ILVRL_UW4(in, out0, out1, out2, out3) do { \
v16u8 t0, t1; \
ILVRL_B2_UB(zero, in, t0, t1); \
ILVRL_H2_UW(zero, t0, out0, out1); \
ILVRL_H2_UW(zero, t1, out2, out3); \
} while (0)
#define MULTHI_16(in0, in1, in2, in3, cnst, out0, out1) do { \
const v4i32 const0 = (v4i32)__msa_fill_w(cnst * 256); \
v4u32 temp0, temp1, temp2, temp3; \
MUL4(in0, const0, in1, const0, in2, const0, in3, const0, \
temp0, temp1, temp2, temp3); \
PCKOD_H2_UH(temp1, temp0, temp3, temp2, out0, out1); \
} while (0)
#define MULTHI_8(in0, in1, cnst, out0) do { \
const v4i32 const0 = (v4i32)__msa_fill_w(cnst * 256); \
v4u32 temp0, temp1; \
MUL2(in0, const0, in1, const0, temp0, temp1); \
out0 = (v8u16)__msa_pckod_h((v8i16)temp1, (v8i16)temp0); \
} while (0)
#define CALC_R16(y0, y1, v0, v1, dst) do { \
const v8i16 const_a = (v8i16)__msa_fill_h(14234); \
const v8i16 a0 = __msa_adds_s_h((v8i16)y0, (v8i16)v0); \
const v8i16 a1 = __msa_adds_s_h((v8i16)y1, (v8i16)v1); \
v8i16 b0 = __msa_subs_s_h(a0, const_a); \
v8i16 b1 = __msa_subs_s_h(a1, const_a); \
SRAI_H2_SH(b0, b1, 6); \
CLIP_SH2_0_255(b0, b1); \
dst = (v16u8)__msa_pckev_b((v16i8)b1, (v16i8)b0); \
} while (0)
#define CALC_R8(y0, v0, dst) do { \
const v8i16 const_a = (v8i16)__msa_fill_h(14234); \
const v8i16 a0 = __msa_adds_s_h((v8i16)y0, (v8i16)v0); \
v8i16 b0 = __msa_subs_s_h(a0, const_a); \
b0 = SRAI_H(b0, 6); \
CLIP_SH_0_255(b0); \
dst = (v16u8)__msa_pckev_b((v16i8)b0, (v16i8)b0); \
} while (0)
#define CALC_G16(y0, y1, u0, u1, v0, v1, dst) do { \
const v8i16 const_a = (v8i16)__msa_fill_h(8708); \
v8i16 a0 = __msa_subs_s_h((v8i16)y0, (v8i16)u0); \
v8i16 a1 = __msa_subs_s_h((v8i16)y1, (v8i16)u1); \
const v8i16 b0 = __msa_subs_s_h(a0, (v8i16)v0); \
const v8i16 b1 = __msa_subs_s_h(a1, (v8i16)v1); \
a0 = __msa_adds_s_h(b0, const_a); \
a1 = __msa_adds_s_h(b1, const_a); \
SRAI_H2_SH(a0, a1, 6); \
CLIP_SH2_0_255(a0, a1); \
dst = (v16u8)__msa_pckev_b((v16i8)a1, (v16i8)a0); \
} while (0)
#define CALC_G8(y0, u0, v0, dst) do { \
const v8i16 const_a = (v8i16)__msa_fill_h(8708); \
v8i16 a0 = __msa_subs_s_h((v8i16)y0, (v8i16)u0); \
const v8i16 b0 = __msa_subs_s_h(a0, (v8i16)v0); \
a0 = __msa_adds_s_h(b0, const_a); \
a0 = SRAI_H(a0, 6); \
CLIP_SH_0_255(a0); \
dst = (v16u8)__msa_pckev_b((v16i8)a0, (v16i8)a0); \
} while (0)
#define CALC_B16(y0, y1, u0, u1, dst) do { \
const v8u16 const_a = (v8u16)__msa_fill_h(17685); \
const v8u16 a0 = __msa_adds_u_h((v8u16)y0, u0); \
const v8u16 a1 = __msa_adds_u_h((v8u16)y1, u1); \
v8u16 b0 = __msa_subs_u_h(a0, const_a); \
v8u16 b1 = __msa_subs_u_h(a1, const_a); \
SRAI_H2_UH(b0, b1, 6); \
CLIP_UH2_0_255(b0, b1); \
dst = (v16u8)__msa_pckev_b((v16i8)b1, (v16i8)b0); \
} while (0)
#define CALC_B8(y0, u0, dst) do { \
const v8u16 const_a = (v8u16)__msa_fill_h(17685); \
const v8u16 a0 = __msa_adds_u_h((v8u16)y0, u0); \
v8u16 b0 = __msa_subs_u_h(a0, const_a); \
b0 = SRAI_H(b0, 6); \
CLIP_UH_0_255(b0); \
dst = (v16u8)__msa_pckev_b((v16i8)b0, (v16i8)b0); \
} while (0)
#define CALC_RGB16(y, u, v, R, G, B) do { \
const v16u8 zero = { 0 }; \
v8u16 y0, y1, u0, u1, v0, v1; \
v4u32 p0, p1, p2, p3; \
const v16u8 in_y = LD_UB(y); \
const v16u8 in_u = LD_UB(u); \
const v16u8 in_v = LD_UB(v); \
ILVRL_UW4(in_y, p0, p1, p2, p3); \
MULTHI_16(p0, p1, p2, p3, 19077, y0, y1); \
ILVRL_UW4(in_v, p0, p1, p2, p3); \
MULTHI_16(p0, p1, p2, p3, 26149, v0, v1); \
CALC_R16(y0, y1, v0, v1, R); \
MULTHI_16(p0, p1, p2, p3, 13320, v0, v1); \
ILVRL_UW4(in_u, p0, p1, p2, p3); \
MULTHI_16(p0, p1, p2, p3, 6419, u0, u1); \
CALC_G16(y0, y1, u0, u1, v0, v1, G); \
MULTHI_16(p0, p1, p2, p3, 33050, u0, u1); \
CALC_B16(y0, y1, u0, u1, B); \
} while (0)
#define CALC_RGB8(y, u, v, R, G, B) do { \
const v16u8 zero = { 0 }; \
v8u16 y0, u0, v0; \
v4u32 p0, p1; \
const v16u8 in_y = LD_UB(y); \
const v16u8 in_u = LD_UB(u); \
const v16u8 in_v = LD_UB(v); \
ILVR_UW2(in_y, p0, p1); \
MULTHI_8(p0, p1, 19077, y0); \
ILVR_UW2(in_v, p0, p1); \
MULTHI_8(p0, p1, 26149, v0); \
CALC_R8(y0, v0, R); \
MULTHI_8(p0, p1, 13320, v0); \
ILVR_UW2(in_u, p0, p1); \
MULTHI_8(p0, p1, 6419, u0); \
CALC_G8(y0, u0, v0, G); \
MULTHI_8(p0, p1, 33050, u0); \
CALC_B8(y0, u0, B); \
} while (0)
#define STORE16_3(a0, a1, a2, dst) do { \
const v16u8 mask0 = { 0, 1, 16, 2, 3, 17, 4, 5, 18, 6, 7, 19, \
8, 9, 20, 10 }; \
const v16u8 mask1 = { 0, 21, 1, 2, 22, 3, 4, 23, 5, 6, 24, 7, \
8, 25, 9, 10 }; \
const v16u8 mask2 = { 26, 0, 1, 27, 2, 3, 28, 4, 5, 29, 6, 7, \
30, 8, 9, 31 }; \
v16u8 out0, out1, out2, tmp0, tmp1, tmp2; \
ILVRL_B2_UB(a1, a0, tmp0, tmp1); \
out0 = VSHF_UB(tmp0, a2, mask0); \
tmp2 = SLDI_UB(tmp1, tmp0, 11); \
out1 = VSHF_UB(tmp2, a2, mask1); \
tmp2 = SLDI_UB(tmp1, tmp1, 6); \
out2 = VSHF_UB(tmp2, a2, mask2); \
ST_UB(out0, dst + 0); \
ST_UB(out1, dst + 16); \
ST_UB(out2, dst + 32); \
} while (0)
#define STORE8_3(a0, a1, a2, dst) do { \
int64_t out_m; \
const v16u8 mask0 = { 0, 1, 16, 2, 3, 17, 4, 5, 18, 6, 7, 19, \
8, 9, 20, 10 }; \
const v16u8 mask1 = { 11, 21, 12, 13, 22, 14, 15, 23, \
255, 255, 255, 255, 255, 255, 255, 255 }; \
const v16u8 tmp0 = (v16u8)__msa_ilvr_b((v16i8)a1, (v16i8)a0); \
v16u8 out0, out1; \
VSHF_B2_UB(tmp0, a2, tmp0, a2, mask0, mask1, out0, out1); \
ST_UB(out0, dst); \
out_m = __msa_copy_s_d((v2i64)out1, 0); \
SD(out_m, dst + 16); \
} while (0)
#define STORE16_4(a0, a1, a2, a3, dst) do { \
v16u8 tmp0, tmp1, tmp2, tmp3; \
v16u8 out0, out1, out2, out3; \
ILVRL_B2_UB(a1, a0, tmp0, tmp1); \
ILVRL_B2_UB(a3, a2, tmp2, tmp3); \
ILVRL_H2_UB(tmp2, tmp0, out0, out1); \
ILVRL_H2_UB(tmp3, tmp1, out2, out3); \
ST_UB(out0, dst + 0); \
ST_UB(out1, dst + 16); \
ST_UB(out2, dst + 32); \
ST_UB(out3, dst + 48); \
} while (0)
#define STORE8_4(a0, a1, a2, a3, dst) do { \
v16u8 tmp0, tmp1, tmp2, tmp3; \
ILVR_B2_UB(a1, a0, a3, a2, tmp0, tmp1); \
ILVRL_H2_UB(tmp1, tmp0, tmp2, tmp3); \
ST_UB(tmp2, dst + 0); \
ST_UB(tmp3, dst + 16); \
} while (0)
#define STORE2_16(a0, a1, dst) do { \
v16u8 out0, out1; \
ILVRL_B2_UB(a1, a0, out0, out1); \
ST_UB(out0, dst + 0); \
ST_UB(out1, dst + 16); \
} while (0)
#define STORE2_8(a0, a1, dst) do { \
const v16u8 out0 = (v16u8)__msa_ilvr_b((v16i8)a1, (v16i8)a0); \
ST_UB(out0, dst); \
} while (0)
#define CALC_RGBA4444(y, u, v, out0, out1, N, dst) do { \
CALC_RGB##N(y, u, v, R, G, B); \
tmp0 = ANDI_B(R, 0xf0); \
tmp1 = SRAI_B(G, 4); \
RG = tmp0 | tmp1; \
tmp0 = ANDI_B(B, 0xf0); \
BA = ORI_B(tmp0, 0x0f); \
STORE2_##N(out0, out1, dst); \
} while (0)
#define CALC_RGB565(y, u, v, out0, out1, N, dst) do { \
CALC_RGB##N(y, u, v, R, G, B); \
tmp0 = ANDI_B(R, 0xf8); \
tmp1 = SRAI_B(G, 5); \
RG = tmp0 | tmp1; \
tmp0 = SLLI_B(G, 3); \
tmp1 = ANDI_B(tmp0, 0xe0); \
tmp0 = SRAI_B(B, 3); \
GB = tmp0 | tmp1; \
STORE2_##N(out0, out1, dst); \
} while (0)
static WEBP_INLINE int Clip8(int v) {
return v < 0 ? 0 : v > 255 ? 255 : v;
}
static void YuvToRgb(int y, int u, int v, uint8_t* const rgb) {
const int y1 = MultHi(y, 19077);
const int r1 = y1 + MultHi(v, 26149) - 14234;
const int g1 = y1 - MultHi(u, 6419) - MultHi(v, 13320) + 8708;
const int b1 = y1 + MultHi(u, 33050) - 17685;
rgb[0] = Clip8(r1 >> 6);
rgb[1] = Clip8(g1 >> 6);
rgb[2] = Clip8(b1 >> 6);
}
static void YuvToBgr(int y, int u, int v, uint8_t* const bgr) {
const int y1 = MultHi(y, 19077);
const int r1 = y1 + MultHi(v, 26149) - 14234;
const int g1 = y1 - MultHi(u, 6419) - MultHi(v, 13320) + 8708;
const int b1 = y1 + MultHi(u, 33050) - 17685;
bgr[0] = Clip8(b1 >> 6);
bgr[1] = Clip8(g1 >> 6);
bgr[2] = Clip8(r1 >> 6);
}
static void YuvToRgb565(int y, int u, int v, uint8_t* const rgb) {
const int y1 = MultHi(y, 19077);
const int r1 = y1 + MultHi(v, 26149) - 14234;
const int g1 = y1 - MultHi(u, 6419) - MultHi(v, 13320) + 8708;
const int b1 = y1 + MultHi(u, 33050) - 17685;
const int r = Clip8(r1 >> 6);
const int g = Clip8(g1 >> 6);
const int b = Clip8(b1 >> 6);
const int rg = (r & 0xf8) | (g >> 5);
const int gb = ((g << 3) & 0xe0) | (b >> 3);
#ifdef WEBP_SWAP_16BIT_CSP
rgb[0] = gb;
rgb[1] = rg;
#else
rgb[0] = rg;
rgb[1] = gb;
#endif
}
static void YuvToRgba4444(int y, int u, int v, uint8_t* const argb) {
const int y1 = MultHi(y, 19077);
const int r1 = y1 + MultHi(v, 26149) - 14234;
const int g1 = y1 - MultHi(u, 6419) - MultHi(v, 13320) + 8708;
const int b1 = y1 + MultHi(u, 33050) - 17685;
const int r = Clip8(r1 >> 6);
const int g = Clip8(g1 >> 6);
const int b = Clip8(b1 >> 6);
const int rg = (r & 0xf0) | (g >> 4);
const int ba = (b & 0xf0) | 0x0f; // overwrite the lower 4 bits
#ifdef WEBP_SWAP_16BIT_CSP
argb[0] = ba;
argb[1] = rg;
#else
argb[0] = rg;
argb[1] = ba;
#endif
}
static void YuvToArgb(uint8_t y, uint8_t u, uint8_t v, uint8_t* const argb) {
argb[0] = 0xff;
YuvToRgb(y, u, v, argb + 1);
}
static void YuvToBgra(uint8_t y, uint8_t u, uint8_t v, uint8_t* const bgra) {
YuvToBgr(y, u, v, bgra);
bgra[3] = 0xff;
}
static void YuvToRgba(uint8_t y, uint8_t u, uint8_t v, uint8_t* const rgba) {
YuvToRgb(y, u, v, rgba);
rgba[3] = 0xff;
}
static void YuvToRgbLine(const uint8_t* y, const uint8_t* u,
const uint8_t* v, uint8_t* dst, int length) {
v16u8 R, G, B;
while (length >= 16) {
CALC_RGB16(y, u, v, R, G, B);
STORE16_3(R, G, B, dst);
y += 16;
u += 16;
v += 16;
dst += 16 * 3;
length -= 16;
}
if (length > 8) {
uint8_t temp[3 * 16] = { 0 };
memcpy(temp, y, length * sizeof(*temp));
CALC_RGB16(temp, u, v, R, G, B);
STORE16_3(R, G, B, temp);
memcpy(dst, temp, length * 3 * sizeof(*dst));
} else if (length > 0) {
uint8_t temp[3 * 8] = { 0 };
memcpy(temp, y, length * sizeof(*temp));
CALC_RGB8(temp, u, v, R, G, B);
STORE8_3(R, G, B, temp);
memcpy(dst, temp, length * 3 * sizeof(*dst));
}
}
static void YuvToBgrLine(const uint8_t* y, const uint8_t* u,
const uint8_t* v, uint8_t* dst, int length) {
v16u8 R, G, B;
while (length >= 16) {
CALC_RGB16(y, u, v, R, G, B);
STORE16_3(B, G, R, dst);
y += 16;
u += 16;
v += 16;
dst += 16 * 3;
length -= 16;
}
if (length > 8) {
uint8_t temp[3 * 16] = { 0 };
memcpy(temp, y, length * sizeof(*temp));
CALC_RGB16(temp, u, v, R, G, B);
STORE16_3(B, G, R, temp);
memcpy(dst, temp, length * 3 * sizeof(*dst));
} else if (length > 0) {
uint8_t temp[3 * 8] = { 0 };
memcpy(temp, y, length * sizeof(*temp));
CALC_RGB8(temp, u, v, R, G, B);
STORE8_3(B, G, R, temp);
memcpy(dst, temp, length * 3 * sizeof(*dst));
}
}
static void YuvToRgbaLine(const uint8_t* y, const uint8_t* u,
const uint8_t* v, uint8_t* dst, int length) {
v16u8 R, G, B;
const v16u8 A = (v16u8)__msa_ldi_b(0xff);
while (length >= 16) {
CALC_RGB16(y, u, v, R, G, B);
STORE16_4(R, G, B, A, dst);
y += 16;
u += 16;
v += 16;
dst += 16 * 4;
length -= 16;
}
if (length > 8) {
uint8_t temp[4 * 16] = { 0 };
memcpy(temp, y, length * sizeof(*temp));
CALC_RGB16(&temp[0], u, v, R, G, B);
STORE16_4(R, G, B, A, temp);
memcpy(dst, temp, length * 4 * sizeof(*dst));
} else if (length > 0) {
uint8_t temp[4 * 8] = { 0 };
memcpy(temp, y, length * sizeof(*temp));
CALC_RGB8(temp, u, v, R, G, B);
STORE8_4(R, G, B, A, temp);
memcpy(dst, temp, length * 4 * sizeof(*dst));
}
}
static void YuvToBgraLine(const uint8_t* y, const uint8_t* u,
const uint8_t* v, uint8_t* dst, int length) {
v16u8 R, G, B;
const v16u8 A = (v16u8)__msa_ldi_b(0xff);
while (length >= 16) {
CALC_RGB16(y, u, v, R, G, B);
STORE16_4(B, G, R, A, dst);
y += 16;
u += 16;
v += 16;
dst += 16 * 4;
length -= 16;
}
if (length > 8) {
uint8_t temp[4 * 16] = { 0 };
memcpy(temp, y, length * sizeof(*temp));
CALC_RGB16(temp, u, v, R, G, B);
STORE16_4(B, G, R, A, temp);
memcpy(dst, temp, length * 4 * sizeof(*dst));
} else if (length > 0) {
uint8_t temp[4 * 8] = { 0 };
memcpy(temp, y, length * sizeof(*temp));
CALC_RGB8(temp, u, v, R, G, B);
STORE8_4(B, G, R, A, temp);
memcpy(dst, temp, length * 4 * sizeof(*dst));
}
}
static void YuvToArgbLine(const uint8_t* y, const uint8_t* u,
const uint8_t* v, uint8_t* dst, int length) {
v16u8 R, G, B;
const v16u8 A = (v16u8)__msa_ldi_b(0xff);
while (length >= 16) {
CALC_RGB16(y, u, v, R, G, B);
STORE16_4(A, R, G, B, dst);
y += 16;
u += 16;
v += 16;
dst += 16 * 4;
length -= 16;
}
if (length > 8) {
uint8_t temp[4 * 16] = { 0 };
memcpy(temp, y, length * sizeof(*temp));
CALC_RGB16(temp, u, v, R, G, B);
STORE16_4(A, R, G, B, temp);
memcpy(dst, temp, length * 4 * sizeof(*dst));
} else if (length > 0) {
uint8_t temp[4 * 8] = { 0 };
memcpy(temp, y, length * sizeof(*temp));
CALC_RGB8(temp, u, v, R, G, B);
STORE8_4(A, R, G, B, temp);
memcpy(dst, temp, length * 4 * sizeof(*dst));
}
}
static void YuvToRgba4444Line(const uint8_t* y, const uint8_t* u,
const uint8_t* v, uint8_t* dst, int length) {
v16u8 R, G, B, RG, BA, tmp0, tmp1;
while (length >= 16) {
#ifdef WEBP_SWAP_16BIT_CSP
CALC_RGBA4444(y, u, v, BA, RG, 16, dst);
#else
CALC_RGBA4444(y, u, v, RG, BA, 16, dst);
#endif
y += 16;
u += 16;
v += 16;
dst += 16 * 2;
length -= 16;
}
if (length > 8) {
uint8_t temp[2 * 16] = { 0 };
memcpy(temp, y, length * sizeof(*temp));
#ifdef WEBP_SWAP_16BIT_CSP
CALC_RGBA4444(temp, u, v, BA, RG, 16, temp);
#else
CALC_RGBA4444(temp, u, v, RG, BA, 16, temp);
#endif
memcpy(dst, temp, length * 2 * sizeof(*dst));
} else if (length > 0) {
uint8_t temp[2 * 8] = { 0 };
memcpy(temp, y, length * sizeof(*temp));
#ifdef WEBP_SWAP_16BIT_CSP
CALC_RGBA4444(temp, u, v, BA, RG, 8, temp);
#else
CALC_RGBA4444(temp, u, v, RG, BA, 8, temp);
#endif
memcpy(dst, temp, length * 2 * sizeof(*dst));
}
}
static void YuvToRgb565Line(const uint8_t* y, const uint8_t* u,
const uint8_t* v, uint8_t* dst, int length) {
v16u8 R, G, B, RG, GB, tmp0, tmp1;
while (length >= 16) {
#ifdef WEBP_SWAP_16BIT_CSP
CALC_RGB565(y, u, v, GB, RG, 16, dst);
#else
CALC_RGB565(y, u, v, RG, GB, 16, dst);
#endif
y += 16;
u += 16;
v += 16;
dst += 16 * 2;
length -= 16;
}
if (length > 8) {
uint8_t temp[2 * 16] = { 0 };
memcpy(temp, y, length * sizeof(*temp));
#ifdef WEBP_SWAP_16BIT_CSP
CALC_RGB565(temp, u, v, GB, RG, 16, temp);
#else
CALC_RGB565(temp, u, v, RG, GB, 16, temp);
#endif
memcpy(dst, temp, length * 2 * sizeof(*dst));
} else if (length > 0) {
uint8_t temp[2 * 8] = { 0 };
memcpy(temp, y, length * sizeof(*temp));
#ifdef WEBP_SWAP_16BIT_CSP
CALC_RGB565(temp, u, v, GB, RG, 8, temp);
#else
CALC_RGB565(temp, u, v, RG, GB, 8, temp);
#endif
memcpy(dst, temp, length * 2 * sizeof(*dst));
}
}
#define UPSAMPLE_32PIXELS(a, b, c, d) do { \
v16u8 s = __msa_aver_u_b(a, d); \
v16u8 t = __msa_aver_u_b(b, c); \
const v16u8 st = s ^ t; \
v16u8 ad = a ^ d; \
v16u8 bc = b ^ c; \
v16u8 t0 = ad | bc; \
v16u8 t1 = t0 | st; \
v16u8 t2 = ANDI_B(t1, 1); \
v16u8 t3 = __msa_aver_u_b(s, t); \
const v16u8 k = t3 - t2; \
v16u8 diag1, diag2; \
AVER_UB2_UB(t, k, s, k, t0, t1); \
bc = bc & st; \
ad = ad & st; \
t = t ^ k; \
s = s ^ k; \
t2 = bc | t; \
t3 = ad | s; \
t2 = ANDI_B(t2, 1); \
t3 = ANDI_B(t3, 1); \
SUB2(t0, t2, t1, t3, diag1, diag2); \
AVER_UB2_UB(a, diag1, b, diag2, t0, t1); \
ILVRL_B2_UB(t1, t0, a, b); \
if (pbot_y != NULL) { \
AVER_UB2_UB(c, diag2, d, diag1, t0, t1); \
ILVRL_B2_UB(t1, t0, c, d); \
} \
} while (0)
#define UPSAMPLE_FUNC(FUNC_NAME, FUNC, XSTEP) \
static void FUNC_NAME(const uint8_t* top_y, const uint8_t* bot_y, \
const uint8_t* top_u, const uint8_t* top_v, \
const uint8_t* cur_u, const uint8_t* cur_v, \
uint8_t* top_dst, uint8_t* bot_dst, int len) \
{ \
int size = (len - 1) >> 1; \
uint8_t temp_u[64]; \
uint8_t temp_v[64]; \
const uint32_t tl_uv = ((top_u[0]) | ((top_v[0]) << 16)); \
const uint32_t l_uv = ((cur_u[0]) | ((cur_v[0]) << 16)); \
const uint32_t uv0 = (3 * tl_uv + l_uv + 0x00020002u) >> 2; \
const uint8_t* ptop_y = &top_y[1]; \
uint8_t *ptop_dst = top_dst + XSTEP; \
const uint8_t* pbot_y = &bot_y[1]; \
uint8_t *pbot_dst = bot_dst + XSTEP; \
\
FUNC(top_y[0], uv0 & 0xff, (uv0 >> 16), top_dst); \
if (bot_y != NULL) { \
const uint32_t uv1 = (3 * l_uv + tl_uv + 0x00020002u) >> 2; \
FUNC(bot_y[0], uv1 & 0xff, (uv1 >> 16), bot_dst); \
} \
while (size >= 16) { \
v16u8 tu0, tu1, tv0, tv1, cu0, cu1, cv0, cv1; \
LD_UB2(top_u, 1, tu0, tu1); \
LD_UB2(cur_u, 1, cu0, cu1); \
LD_UB2(top_v, 1, tv0, tv1); \
LD_UB2(cur_v, 1, cv0, cv1); \
UPSAMPLE_32PIXELS(tu0, tu1, cu0, cu1); \
UPSAMPLE_32PIXELS(tv0, tv1, cv0, cv1); \
ST_UB4(tu0, tu1, cu0, cu1, &temp_u[0], 16); \
ST_UB4(tv0, tv1, cv0, cv1, &temp_v[0], 16); \
FUNC##Line(ptop_y, &temp_u[ 0], &temp_v[0], ptop_dst, 32); \
if (bot_y != NULL) { \
FUNC##Line(pbot_y, &temp_u[32], &temp_v[32], pbot_dst, 32); \
} \
ptop_y += 32; \
pbot_y += 32; \
ptop_dst += XSTEP * 32; \
pbot_dst += XSTEP * 32; \
top_u += 16; \
top_v += 16; \
cur_u += 16; \
cur_v += 16; \
size -= 16; \
} \
if (size > 0) { \
v16u8 tu0, tu1, tv0, tv1, cu0, cu1, cv0, cv1; \
memcpy(&temp_u[ 0], top_u, 17 * sizeof(uint8_t)); \
memcpy(&temp_u[32], cur_u, 17 * sizeof(uint8_t)); \
memcpy(&temp_v[ 0], top_v, 17 * sizeof(uint8_t)); \
memcpy(&temp_v[32], cur_v, 17 * sizeof(uint8_t)); \
LD_UB2(&temp_u[ 0], 1, tu0, tu1); \
LD_UB2(&temp_u[32], 1, cu0, cu1); \
LD_UB2(&temp_v[ 0], 1, tv0, tv1); \
LD_UB2(&temp_v[32], 1, cv0, cv1); \
UPSAMPLE_32PIXELS(tu0, tu1, cu0, cu1); \
UPSAMPLE_32PIXELS(tv0, tv1, cv0, cv1); \
ST_UB4(tu0, tu1, cu0, cu1, &temp_u[0], 16); \
ST_UB4(tv0, tv1, cv0, cv1, &temp_v[0], 16); \
FUNC##Line(ptop_y, &temp_u[ 0], &temp_v[0], ptop_dst, size * 2); \
if (bot_y != NULL) { \
FUNC##Line(pbot_y, &temp_u[32], &temp_v[32], pbot_dst, size * 2); \
} \
top_u += size; \
top_v += size; \
cur_u += size; \
cur_v += size; \
} \
if (!(len & 1)) { \
const uint32_t t0 = ((top_u[0]) | ((top_v[0]) << 16)); \
const uint32_t c0 = ((cur_u[0]) | ((cur_v[0]) << 16)); \
const uint32_t tmp0 = (3 * t0 + c0 + 0x00020002u) >> 2; \
FUNC(top_y[len - 1], tmp0 & 0xff, (tmp0 >> 16), \
top_dst + (len - 1) * XSTEP); \
if (bot_y != NULL) { \
const uint32_t tmp1 = (3 * c0 + t0 + 0x00020002u) >> 2; \
FUNC(bot_y[len - 1], tmp1 & 0xff, (tmp1 >> 16), \
bot_dst + (len - 1) * XSTEP); \
} \
} \
}
UPSAMPLE_FUNC(UpsampleRgbLinePair, YuvToRgb, 3)
UPSAMPLE_FUNC(UpsampleBgrLinePair, YuvToBgr, 3)
UPSAMPLE_FUNC(UpsampleRgbaLinePair, YuvToRgba, 4)
UPSAMPLE_FUNC(UpsampleBgraLinePair, YuvToBgra, 4)
UPSAMPLE_FUNC(UpsampleArgbLinePair, YuvToArgb, 4)
UPSAMPLE_FUNC(UpsampleRgba4444LinePair, YuvToRgba4444, 2)
UPSAMPLE_FUNC(UpsampleRgb565LinePair, YuvToRgb565, 2)
//------------------------------------------------------------------------------
// Entry point
extern WebPUpsampleLinePairFunc WebPUpsamplers[/* MODE_LAST */];
extern void WebPInitUpsamplersMSA(void);
WEBP_TSAN_IGNORE_FUNCTION void WebPInitUpsamplersMSA(void) {
WebPUpsamplers[MODE_RGB] = UpsampleRgbLinePair;
WebPUpsamplers[MODE_RGBA] = UpsampleRgbaLinePair;
WebPUpsamplers[MODE_BGR] = UpsampleBgrLinePair;
WebPUpsamplers[MODE_BGRA] = UpsampleBgraLinePair;
WebPUpsamplers[MODE_ARGB] = UpsampleArgbLinePair;
WebPUpsamplers[MODE_rgbA] = UpsampleRgbaLinePair;
WebPUpsamplers[MODE_bgrA] = UpsampleBgraLinePair;
WebPUpsamplers[MODE_Argb] = UpsampleArgbLinePair;
WebPUpsamplers[MODE_RGB_565] = UpsampleRgb565LinePair;
WebPUpsamplers[MODE_RGBA_4444] = UpsampleRgba4444LinePair;
WebPUpsamplers[MODE_rgbA_4444] = UpsampleRgba4444LinePair;
}
#endif // FANCY_UPSAMPLING
#endif // WEBP_USE_MSA
#if !(defined(FANCY_UPSAMPLING) && defined(WEBP_USE_MSA))
WEBP_DSP_INIT_STUB(WebPInitUpsamplersMSA)
#endif

77
thirdparty/libwebp/dsp/upsampling_neon.c vendored
View file

@ -28,47 +28,34 @@
// U/V upsampling
// Loads 9 pixels each from rows r1 and r2 and generates 16 pixels.
#define UPSAMPLE_16PIXELS(r1, r2, out) { \
uint8x8_t a = vld1_u8(r1); \
uint8x8_t b = vld1_u8(r1 + 1); \
uint8x8_t c = vld1_u8(r2); \
uint8x8_t d = vld1_u8(r2 + 1); \
\
uint16x8_t al = vshll_n_u8(a, 1); \
uint16x8_t bl = vshll_n_u8(b, 1); \
uint16x8_t cl = vshll_n_u8(c, 1); \
uint16x8_t dl = vshll_n_u8(d, 1); \
\
uint8x8_t diag1, diag2; \
uint16x8_t sl; \
\
#define UPSAMPLE_16PIXELS(r1, r2, out) do { \
const uint8x8_t a = vld1_u8(r1 + 0); \
const uint8x8_t b = vld1_u8(r1 + 1); \
const uint8x8_t c = vld1_u8(r2 + 0); \
const uint8x8_t d = vld1_u8(r2 + 1); \
/* a + b + c + d */ \
sl = vaddl_u8(a, b); \
sl = vaddw_u8(sl, c); \
sl = vaddw_u8(sl, d); \
const uint16x8_t ad = vaddl_u8(a, d); \
const uint16x8_t bc = vaddl_u8(b, c); \
const uint16x8_t abcd = vaddq_u16(ad, bc); \
/* 3a + b + c + 3d */ \
const uint16x8_t al = vaddq_u16(abcd, vshlq_n_u16(ad, 1)); \
/* a + 3b + 3c + d */ \
const uint16x8_t bl = vaddq_u16(abcd, vshlq_n_u16(bc, 1)); \
\
al = vaddq_u16(sl, al); /* 3a + b + c + d */ \
bl = vaddq_u16(sl, bl); /* a + 3b + c + d */ \
const uint8x8_t diag2 = vshrn_n_u16(al, 3); \
const uint8x8_t diag1 = vshrn_n_u16(bl, 3); \
\
al = vaddq_u16(al, dl); /* 3a + b + c + 3d */ \
bl = vaddq_u16(bl, cl); /* a + 3b + 3c + d */ \
const uint8x8_t A = vrhadd_u8(a, diag1); \
const uint8x8_t B = vrhadd_u8(b, diag2); \
const uint8x8_t C = vrhadd_u8(c, diag2); \
const uint8x8_t D = vrhadd_u8(d, diag1); \
\
diag2 = vshrn_n_u16(al, 3); \
diag1 = vshrn_n_u16(bl, 3); \
\
a = vrhadd_u8(a, diag1); \
b = vrhadd_u8(b, diag2); \
c = vrhadd_u8(c, diag2); \
d = vrhadd_u8(d, diag1); \
\
{ \
uint8x8x2_t a_b, c_d; \
INIT_VECTOR2(a_b, a, b); \
INIT_VECTOR2(c_d, c, d); \
vst2_u8(out, a_b); \
vst2_u8(out + 32, c_d); \
} \
}
uint8x8x2_t A_B, C_D; \
INIT_VECTOR2(A_B, A, B); \
INIT_VECTOR2(C_D, C, D); \
vst2_u8(out + 0, A_B); \
vst2_u8(out + 32, C_D); \
} while (0)
// Turn the macro into a function for reducing code-size when non-critical
static void Upsample16Pixels(const uint8_t *r1, const uint8_t *r2,
@ -93,7 +80,6 @@ static void Upsample16Pixels(const uint8_t *r1, const uint8_t *r2,
static const int16_t kCoeffs1[4] = { 19077, 26149, 6419, 13320 };
#define v255 vdup_n_u8(255)
#define v_0x0f vdup_n_u8(15)
#define STORE_Rgb(out, r, g, b) do { \
uint8x8x3_t r_g_b; \
@ -132,21 +118,16 @@ static const int16_t kCoeffs1[4] = { 19077, 26149, 6419, 13320 };
#endif
#define STORE_Rgba4444(out, r, g, b) do { \
const uint8x8_t r1 = vshl_n_u8(vshr_n_u8(r, 4), 4); /* 4bits */ \
const uint8x8_t g1 = vshr_n_u8(g, 4); \
const uint8x8_t ba = vorr_u8(b, v_0x0f); \
const uint8x8_t rg = vorr_u8(r1, g1); \
const uint8x8_t rg = vsri_n_u8(r, g, 4); /* shift g, insert r */ \
const uint8x8_t ba = vsri_n_u8(b, v255, 4); /* shift a, insert b */ \
const uint8x8x2_t rgba4444 = ZIP_U8(rg, ba); \
vst1q_u8(out, vcombine_u8(rgba4444.val[0], rgba4444.val[1])); \
} while (0)
#define STORE_Rgb565(out, r, g, b) do { \
const uint8x8_t r1 = vshl_n_u8(vshr_n_u8(r, 3), 3); /* 5bits */ \
const uint8x8_t g1 = vshr_n_u8(g, 5); /* upper 3bits */\
const uint8x8_t g2 = vshl_n_u8(vshr_n_u8(g, 2), 5); /* lower 3bits */\
const uint8x8_t b1 = vshr_n_u8(b, 3); /* 5bits */ \
const uint8x8_t rg = vorr_u8(r1, g1); \
const uint8x8_t gb = vorr_u8(g2, b1); \
const uint8x8_t rg = vsri_n_u8(r, g, 5); /* shift g and insert r */ \
const uint8x8_t g1 = vshl_n_u8(g, 3); /* pre-shift g: 3bits */ \
const uint8x8_t gb = vsri_n_u8(g1, b, 3); /* shift b and insert g */ \
const uint8x8x2_t rgb565 = ZIP_U8(rg, gb); \
vst1q_u8(out, vcombine_u8(rgb565.val[0], rgb565.val[1])); \
} while (0)

57
thirdparty/libwebp/dsp/yuv.c vendored
View file

@ -13,6 +13,8 @@
#include "./yuv.h"
#include <stdlib.h>
#if defined(WEBP_YUV_USE_TABLE)
static int done = 0;
@ -244,6 +246,48 @@ void WebPConvertRGBA32ToUV_C(const uint16_t* rgb,
//-----------------------------------------------------------------------------
#define MAX_Y ((1 << 10) - 1) // 10b precision over 16b-arithmetic
static uint16_t clip_y(int v) {
return (v < 0) ? 0 : (v > MAX_Y) ? MAX_Y : (uint16_t)v;
}
static uint64_t SharpYUVUpdateY_C(const uint16_t* ref, const uint16_t* src,
uint16_t* dst, int len) {
uint64_t diff = 0;
int i;
for (i = 0; i < len; ++i) {
const int diff_y = ref[i] - src[i];
const int new_y = (int)dst[i] + diff_y;
dst[i] = clip_y(new_y);
diff += (uint64_t)abs(diff_y);
}
return diff;
}
static void SharpYUVUpdateRGB_C(const int16_t* ref, const int16_t* src,
int16_t* dst, int len) {
int i;
for (i = 0; i < len; ++i) {
const int diff_uv = ref[i] - src[i];
dst[i] += diff_uv;
}
}
static void SharpYUVFilterRow_C(const int16_t* A, const int16_t* B, int len,
const uint16_t* best_y, uint16_t* out) {
int i;
for (i = 0; i < len; ++i, ++A, ++B) {
const int v0 = (A[0] * 9 + A[1] * 3 + B[0] * 3 + B[1] + 8) >> 4;
const int v1 = (A[1] * 9 + A[0] * 3 + B[1] * 3 + B[0] + 8) >> 4;
out[2 * i + 0] = clip_y(best_y[2 * i + 0] + v0);
out[2 * i + 1] = clip_y(best_y[2 * i + 1] + v1);
}
}
#undef MAX_Y
//-----------------------------------------------------------------------------
void (*WebPConvertRGB24ToY)(const uint8_t* rgb, uint8_t* y, int width);
void (*WebPConvertBGR24ToY)(const uint8_t* bgr, uint8_t* y, int width);
void (*WebPConvertRGBA32ToUV)(const uint16_t* rgb,
@ -253,10 +297,18 @@ void (*WebPConvertARGBToY)(const uint32_t* argb, uint8_t* y, int width);
void (*WebPConvertARGBToUV)(const uint32_t* argb, uint8_t* u, uint8_t* v,
int src_width, int do_store);
uint64_t (*WebPSharpYUVUpdateY)(const uint16_t* ref, const uint16_t* src,
uint16_t* dst, int len);
void (*WebPSharpYUVUpdateRGB)(const int16_t* ref, const int16_t* src,
int16_t* dst, int len);
void (*WebPSharpYUVFilterRow)(const int16_t* A, const int16_t* B, int len,
const uint16_t* best_y, uint16_t* out);
static volatile VP8CPUInfo rgba_to_yuv_last_cpuinfo_used =
(VP8CPUInfo)&rgba_to_yuv_last_cpuinfo_used;
extern void WebPInitConvertARGBToYUVSSE2(void);
extern void WebPInitSharpYUVSSE2(void);
WEBP_TSAN_IGNORE_FUNCTION void WebPInitConvertARGBToYUV(void) {
if (rgba_to_yuv_last_cpuinfo_used == VP8GetCPUInfo) return;
@ -269,10 +321,15 @@ WEBP_TSAN_IGNORE_FUNCTION void WebPInitConvertARGBToYUV(void) {
WebPConvertRGBA32ToUV = WebPConvertRGBA32ToUV_C;
WebPSharpYUVUpdateY = SharpYUVUpdateY_C;
WebPSharpYUVUpdateRGB = SharpYUVUpdateRGB_C;
WebPSharpYUVFilterRow = SharpYUVFilterRow_C;
if (VP8GetCPUInfo != NULL) {
#if defined(WEBP_USE_SSE2)
if (VP8GetCPUInfo(kSSE2)) {
WebPInitConvertARGBToYUVSSE2();
WebPInitSharpYUVSSE2();
}
#endif // WEBP_USE_SSE2
}

2
thirdparty/libwebp/dsp/yuv.h vendored
View file

@ -36,7 +36,7 @@
#define WEBP_DSP_YUV_H_
#include "./dsp.h"
#include "../dec/decode_vp8.h"
#include "../dec/vp8_dec.h"
#if defined(WEBP_EXPERIMENTAL_FEATURES)
// Do NOT activate this feature for real compression. This is only experimental!

298
thirdparty/libwebp/dsp/yuv_sse2.c vendored
View file

@ -15,6 +15,8 @@
#if defined(WEBP_USE_SSE2)
#include "./common_sse2.h"
#include <stdlib.h>
#include <emmintrin.h>
//-----------------------------------------------------------------------------
@ -155,30 +157,13 @@ static WEBP_INLINE void PackAndStore565(const __m128i* const R,
_mm_storeu_si128((__m128i*)dst, rgb565);
}
// Function used several times in PlanarTo24b.
// It samples the in buffer as follows: one every two unsigned char is stored
// at the beginning of the buffer, while the other half is stored at the end.
static WEBP_INLINE void PlanarTo24bHelper(const __m128i* const in /*in[6]*/,
__m128i* const out /*out[6]*/) {
const __m128i v_mask = _mm_set1_epi16(0x00ff);
// Take one every two upper 8b values.
out[0] = _mm_packus_epi16(_mm_and_si128(in[0], v_mask),
_mm_and_si128(in[1], v_mask));
out[1] = _mm_packus_epi16(_mm_and_si128(in[2], v_mask),
_mm_and_si128(in[3], v_mask));
out[2] = _mm_packus_epi16(_mm_and_si128(in[4], v_mask),
_mm_and_si128(in[5], v_mask));
// Take one every two lower 8b values.
out[3] = _mm_packus_epi16(_mm_srli_epi16(in[0], 8), _mm_srli_epi16(in[1], 8));
out[4] = _mm_packus_epi16(_mm_srli_epi16(in[2], 8), _mm_srli_epi16(in[3], 8));
out[5] = _mm_packus_epi16(_mm_srli_epi16(in[4], 8), _mm_srli_epi16(in[5], 8));
}
// Pack the planar buffers
// rrrr... rrrr... gggg... gggg... bbbb... bbbb....
// triplet by triplet in the output buffer rgb as rgbrgbrgbrgb ...
static WEBP_INLINE void PlanarTo24b(__m128i* const in /*in[6]*/, uint8_t* rgb) {
static WEBP_INLINE void PlanarTo24b(__m128i* const in0, __m128i* const in1,
__m128i* const in2, __m128i* const in3,
__m128i* const in4, __m128i* const in5,
uint8_t* const rgb) {
// The input is 6 registers of sixteen 8b but for the sake of explanation,
// let's take 6 registers of four 8b values.
// To pack, we will keep taking one every two 8b integer and move it
@ -191,22 +176,15 @@ static WEBP_INLINE void PlanarTo24b(__m128i* const in /*in[6]*/, uint8_t* rgb) {
// Repeat the same permutations twice more:
// r0r4g0g4 | b0b4r1r5 | g1g5b1b5 | r2r6g2g6 | b2b6r3r7 | g3g7b3b7
// r0g0b0r1 | g1b1r2g2 | b2r3g3b3 | r4g4b4r5 | g5b5r6g6 | b6r7g7b7
__m128i tmp[6];
PlanarTo24bHelper(in, tmp);
PlanarTo24bHelper(tmp, in);
PlanarTo24bHelper(in, tmp);
// We need to do it two more times than the example as we have sixteen bytes.
PlanarTo24bHelper(tmp, in);
PlanarTo24bHelper(in, tmp);
VP8PlanarTo24b(in0, in1, in2, in3, in4, in5);
_mm_storeu_si128((__m128i*)(rgb + 0), tmp[0]);
_mm_storeu_si128((__m128i*)(rgb + 16), tmp[1]);
_mm_storeu_si128((__m128i*)(rgb + 32), tmp[2]);
_mm_storeu_si128((__m128i*)(rgb + 48), tmp[3]);
_mm_storeu_si128((__m128i*)(rgb + 64), tmp[4]);
_mm_storeu_si128((__m128i*)(rgb + 80), tmp[5]);
_mm_storeu_si128((__m128i*)(rgb + 0), *in0);
_mm_storeu_si128((__m128i*)(rgb + 16), *in1);
_mm_storeu_si128((__m128i*)(rgb + 32), *in2);
_mm_storeu_si128((__m128i*)(rgb + 48), *in3);
_mm_storeu_si128((__m128i*)(rgb + 64), *in4);
_mm_storeu_si128((__m128i*)(rgb + 80), *in5);
}
#undef MK_UINT32
void VP8YuvToRgba32(const uint8_t* y, const uint8_t* u, const uint8_t* v,
uint8_t* dst) {
@ -265,29 +243,29 @@ void VP8YuvToRgb56532(const uint8_t* y, const uint8_t* u, const uint8_t* v,
void VP8YuvToRgb32(const uint8_t* y, const uint8_t* u, const uint8_t* v,
uint8_t* dst) {
__m128i R0, R1, R2, R3, G0, G1, G2, G3, B0, B1, B2, B3;
__m128i rgb[6];
__m128i rgb0, rgb1, rgb2, rgb3, rgb4, rgb5;
YUV444ToRGB(y + 0, u + 0, v + 0, &R0, &G0, &B0);
YUV444ToRGB(y + 8, u + 8, v + 8, &R1, &G1, &B1);
YUV444ToRGB(y + 0, u + 0, v + 0, &R0, &G0, &B0);
YUV444ToRGB(y + 8, u + 8, v + 8, &R1, &G1, &B1);
YUV444ToRGB(y + 16, u + 16, v + 16, &R2, &G2, &B2);
YUV444ToRGB(y + 24, u + 24, v + 24, &R3, &G3, &B3);
// Cast to 8b and store as RRRRGGGGBBBB.
rgb[0] = _mm_packus_epi16(R0, R1);
rgb[1] = _mm_packus_epi16(R2, R3);
rgb[2] = _mm_packus_epi16(G0, G1);
rgb[3] = _mm_packus_epi16(G2, G3);
rgb[4] = _mm_packus_epi16(B0, B1);
rgb[5] = _mm_packus_epi16(B2, B3);
rgb0 = _mm_packus_epi16(R0, R1);
rgb1 = _mm_packus_epi16(R2, R3);
rgb2 = _mm_packus_epi16(G0, G1);
rgb3 = _mm_packus_epi16(G2, G3);
rgb4 = _mm_packus_epi16(B0, B1);
rgb5 = _mm_packus_epi16(B2, B3);
// Pack as RGBRGBRGBRGB.
PlanarTo24b(rgb, dst);
PlanarTo24b(&rgb0, &rgb1, &rgb2, &rgb3, &rgb4, &rgb5, dst);
}
void VP8YuvToBgr32(const uint8_t* y, const uint8_t* u, const uint8_t* v,
uint8_t* dst) {
__m128i R0, R1, R2, R3, G0, G1, G2, G3, B0, B1, B2, B3;
__m128i bgr[6];
__m128i bgr0, bgr1, bgr2, bgr3, bgr4, bgr5;
YUV444ToRGB(y + 0, u + 0, v + 0, &R0, &G0, &B0);
YUV444ToRGB(y + 8, u + 8, v + 8, &R1, &G1, &B1);
@ -295,15 +273,15 @@ void VP8YuvToBgr32(const uint8_t* y, const uint8_t* u, const uint8_t* v,
YUV444ToRGB(y + 24, u + 24, v + 24, &R3, &G3, &B3);
// Cast to 8b and store as BBBBGGGGRRRR.
bgr[0] = _mm_packus_epi16(B0, B1);
bgr[1] = _mm_packus_epi16(B2, B3);
bgr[2] = _mm_packus_epi16(G0, G1);
bgr[3] = _mm_packus_epi16(G2, G3);
bgr[4] = _mm_packus_epi16(R0, R1);
bgr[5] = _mm_packus_epi16(R2, R3);
bgr0 = _mm_packus_epi16(B0, B1);
bgr1 = _mm_packus_epi16(B2, B3);
bgr2 = _mm_packus_epi16(G0, G1);
bgr3 = _mm_packus_epi16(G2, G3);
bgr4 = _mm_packus_epi16(R0, R1);
bgr5= _mm_packus_epi16(R2, R3);
// Pack as BGRBGRBGRBGR.
PlanarTo24b(bgr, dst);
PlanarTo24b(&bgr0, &bgr1, &bgr2, &bgr3, &bgr4, &bgr5, dst);
}
//-----------------------------------------------------------------------------
@ -377,7 +355,7 @@ static void YuvToRgbRow(const uint8_t* y, const uint8_t* u, const uint8_t* v,
int n;
for (n = 0; n + 32 <= len; n += 32, dst += 32 * 3) {
__m128i R0, R1, R2, R3, G0, G1, G2, G3, B0, B1, B2, B3;
__m128i rgb[6];
__m128i rgb0, rgb1, rgb2, rgb3, rgb4, rgb5;
YUV420ToRGB(y + 0, u + 0, v + 0, &R0, &G0, &B0);
YUV420ToRGB(y + 8, u + 4, v + 4, &R1, &G1, &B1);
@ -385,15 +363,15 @@ static void YuvToRgbRow(const uint8_t* y, const uint8_t* u, const uint8_t* v,
YUV420ToRGB(y + 24, u + 12, v + 12, &R3, &G3, &B3);
// Cast to 8b and store as RRRRGGGGBBBB.
rgb[0] = _mm_packus_epi16(R0, R1);
rgb[1] = _mm_packus_epi16(R2, R3);
rgb[2] = _mm_packus_epi16(G0, G1);
rgb[3] = _mm_packus_epi16(G2, G3);
rgb[4] = _mm_packus_epi16(B0, B1);
rgb[5] = _mm_packus_epi16(B2, B3);
rgb0 = _mm_packus_epi16(R0, R1);
rgb1 = _mm_packus_epi16(R2, R3);
rgb2 = _mm_packus_epi16(G0, G1);
rgb3 = _mm_packus_epi16(G2, G3);
rgb4 = _mm_packus_epi16(B0, B1);
rgb5 = _mm_packus_epi16(B2, B3);
// Pack as RGBRGBRGBRGB.
PlanarTo24b(rgb, dst);
PlanarTo24b(&rgb0, &rgb1, &rgb2, &rgb3, &rgb4, &rgb5, dst);
y += 32;
u += 16;
@ -413,7 +391,7 @@ static void YuvToBgrRow(const uint8_t* y, const uint8_t* u, const uint8_t* v,
int n;
for (n = 0; n + 32 <= len; n += 32, dst += 32 * 3) {
__m128i R0, R1, R2, R3, G0, G1, G2, G3, B0, B1, B2, B3;
__m128i bgr[6];
__m128i bgr0, bgr1, bgr2, bgr3, bgr4, bgr5;
YUV420ToRGB(y + 0, u + 0, v + 0, &R0, &G0, &B0);
YUV420ToRGB(y + 8, u + 4, v + 4, &R1, &G1, &B1);
@ -421,15 +399,15 @@ static void YuvToBgrRow(const uint8_t* y, const uint8_t* u, const uint8_t* v,
YUV420ToRGB(y + 24, u + 12, v + 12, &R3, &G3, &B3);
// Cast to 8b and store as BBBBGGGGRRRR.
bgr[0] = _mm_packus_epi16(B0, B1);
bgr[1] = _mm_packus_epi16(B2, B3);
bgr[2] = _mm_packus_epi16(G0, G1);
bgr[3] = _mm_packus_epi16(G2, G3);
bgr[4] = _mm_packus_epi16(R0, R1);
bgr[5] = _mm_packus_epi16(R2, R3);
bgr0 = _mm_packus_epi16(B0, B1);
bgr1 = _mm_packus_epi16(B2, B3);
bgr2 = _mm_packus_epi16(G0, G1);
bgr3 = _mm_packus_epi16(G2, G3);
bgr4 = _mm_packus_epi16(R0, R1);
bgr5 = _mm_packus_epi16(R2, R3);
// Pack as BGRBGRBGRBGR.
PlanarTo24b(bgr, dst);
PlanarTo24b(&bgr0, &bgr1, &bgr2, &bgr3, &bgr4, &bgr5, dst);
y += 32;
u += 16;
@ -499,25 +477,19 @@ static WEBP_INLINE void RGB24PackedToPlanar(const uint8_t* const rgb,
// Convert 8 packed ARGB to r[], g[], b[]
static WEBP_INLINE void RGB32PackedToPlanar(const uint32_t* const argb,
__m128i* const r,
__m128i* const g,
__m128i* const b) {
__m128i* const rgb /*in[6]*/) {
const __m128i zero = _mm_setzero_si128();
const __m128i in0 = LOAD_16(argb + 0); // argb3 | argb2 | argb1 | argb0
const __m128i in1 = LOAD_16(argb + 4); // argb7 | argb6 | argb5 | argb4
// column-wise transpose
const __m128i A0 = _mm_unpacklo_epi8(in0, in1);
const __m128i A1 = _mm_unpackhi_epi8(in0, in1);
const __m128i B0 = _mm_unpacklo_epi8(A0, A1);
const __m128i B1 = _mm_unpackhi_epi8(A0, A1);
// C0 = g7 g6 ... g1 g0 | b7 b6 ... b1 b0
// C1 = a7 a6 ... a1 a0 | r7 r6 ... r1 r0
const __m128i C0 = _mm_unpacklo_epi8(B0, B1);
const __m128i C1 = _mm_unpackhi_epi8(B0, B1);
// store 16b
*r = _mm_unpacklo_epi8(C1, zero);
*g = _mm_unpackhi_epi8(C0, zero);
*b = _mm_unpacklo_epi8(C0, zero);
__m128i a0 = LOAD_16(argb + 0);
__m128i a1 = LOAD_16(argb + 4);
__m128i a2 = LOAD_16(argb + 8);
__m128i a3 = LOAD_16(argb + 12);
VP8L32bToPlanar(&a0, &a1, &a2, &a3);
rgb[0] = _mm_unpacklo_epi8(a1, zero);
rgb[1] = _mm_unpackhi_epi8(a1, zero);
rgb[2] = _mm_unpacklo_epi8(a2, zero);
rgb[3] = _mm_unpackhi_epi8(a2, zero);
rgb[4] = _mm_unpacklo_epi8(a3, zero);
rgb[5] = _mm_unpackhi_epi8(a3, zero);
}
// This macro computes (RG * MULT_RG + GB * MULT_GB + ROUNDER) >> DESCALE_FIX
@ -649,11 +621,10 @@ static void ConvertARGBToY(const uint32_t* argb, uint8_t* y, int width) {
const int max_width = width & ~15;
int i;
for (i = 0; i < max_width; i += 16) {
__m128i r, g, b, Y0, Y1;
RGB32PackedToPlanar(&argb[i + 0], &r, &g, &b);
ConvertRGBToY(&r, &g, &b, &Y0);
RGB32PackedToPlanar(&argb[i + 8], &r, &g, &b);
ConvertRGBToY(&r, &g, &b, &Y1);
__m128i Y0, Y1, rgb[6];
RGB32PackedToPlanar(&argb[i], rgb);
ConvertRGBToY(&rgb[0], &rgb[2], &rgb[4], &Y0);
ConvertRGBToY(&rgb[1], &rgb[3], &rgb[5], &Y1);
STORE_16(_mm_packus_epi16(Y0, Y1), y + i);
}
for (; i < width; ++i) { // left-over
@ -678,20 +649,18 @@ static void ConvertARGBToUV(const uint32_t* argb, uint8_t* u, uint8_t* v,
const int max_width = src_width & ~31;
int i;
for (i = 0; i < max_width; i += 32, u += 16, v += 16) {
__m128i r0, g0, b0, r1, g1, b1, U0, V0, U1, V1;
RGB32PackedToPlanar(&argb[i + 0], &r0, &g0, &b0);
RGB32PackedToPlanar(&argb[i + 8], &r1, &g1, &b1);
HorizontalAddPack(&r0, &r1, &r0);
HorizontalAddPack(&g0, &g1, &g0);
HorizontalAddPack(&b0, &b1, &b0);
ConvertRGBToUV(&r0, &g0, &b0, &U0, &V0);
__m128i rgb[6], U0, V0, U1, V1;
RGB32PackedToPlanar(&argb[i], rgb);
HorizontalAddPack(&rgb[0], &rgb[1], &rgb[0]);
HorizontalAddPack(&rgb[2], &rgb[3], &rgb[2]);
HorizontalAddPack(&rgb[4], &rgb[5], &rgb[4]);
ConvertRGBToUV(&rgb[0], &rgb[2], &rgb[4], &U0, &V0);
RGB32PackedToPlanar(&argb[i + 16], &r0, &g0, &b0);
RGB32PackedToPlanar(&argb[i + 24], &r1, &g1, &b1);
HorizontalAddPack(&r0, &r1, &r0);
HorizontalAddPack(&g0, &g1, &g0);
HorizontalAddPack(&b0, &b1, &b0);
ConvertRGBToUV(&r0, &g0, &b0, &U1, &V1);
RGB32PackedToPlanar(&argb[i + 16], rgb);
HorizontalAddPack(&rgb[0], &rgb[1], &rgb[0]);
HorizontalAddPack(&rgb[2], &rgb[3], &rgb[2]);
HorizontalAddPack(&rgb[4], &rgb[5], &rgb[4]);
ConvertRGBToUV(&rgb[0], &rgb[2], &rgb[4], &U1, &V1);
U0 = _mm_packus_epi16(U0, U1);
V0 = _mm_packus_epi16(V0, V1);
@ -767,9 +736,128 @@ WEBP_TSAN_IGNORE_FUNCTION void WebPInitConvertARGBToYUVSSE2(void) {
WebPConvertRGBA32ToUV = ConvertRGBA32ToUV;
}
//------------------------------------------------------------------------------
#define MAX_Y ((1 << 10) - 1) // 10b precision over 16b-arithmetic
static uint16_t clip_y(int v) {
return (v < 0) ? 0 : (v > MAX_Y) ? MAX_Y : (uint16_t)v;
}
static uint64_t SharpYUVUpdateY_SSE2(const uint16_t* ref, const uint16_t* src,
uint16_t* dst, int len) {
uint64_t diff = 0;
uint32_t tmp[4];
int i;
const __m128i zero = _mm_setzero_si128();
const __m128i max = _mm_set1_epi16(MAX_Y);
const __m128i one = _mm_set1_epi16(1);
__m128i sum = zero;
for (i = 0; i + 8 <= len; i += 8) {
const __m128i A = _mm_loadu_si128((const __m128i*)(ref + i));
const __m128i B = _mm_loadu_si128((const __m128i*)(src + i));
const __m128i C = _mm_loadu_si128((const __m128i*)(dst + i));
const __m128i D = _mm_sub_epi16(A, B); // diff_y
const __m128i E = _mm_cmpgt_epi16(zero, D); // sign (-1 or 0)
const __m128i F = _mm_add_epi16(C, D); // new_y
const __m128i G = _mm_or_si128(E, one); // -1 or 1
const __m128i H = _mm_max_epi16(_mm_min_epi16(F, max), zero);
const __m128i I = _mm_madd_epi16(D, G); // sum(abs(...))
_mm_storeu_si128((__m128i*)(dst + i), H);
sum = _mm_add_epi32(sum, I);
}
_mm_storeu_si128((__m128i*)tmp, sum);
diff = tmp[3] + tmp[2] + tmp[1] + tmp[0];
for (; i < len; ++i) {
const int diff_y = ref[i] - src[i];
const int new_y = (int)dst[i] + diff_y;
dst[i] = clip_y(new_y);
diff += (uint64_t)abs(diff_y);
}
return diff;
}
static void SharpYUVUpdateRGB_SSE2(const int16_t* ref, const int16_t* src,
int16_t* dst, int len) {
int i = 0;
for (i = 0; i + 8 <= len; i += 8) {
const __m128i A = _mm_loadu_si128((const __m128i*)(ref + i));
const __m128i B = _mm_loadu_si128((const __m128i*)(src + i));
const __m128i C = _mm_loadu_si128((const __m128i*)(dst + i));
const __m128i D = _mm_sub_epi16(A, B); // diff_uv
const __m128i E = _mm_add_epi16(C, D); // new_uv
_mm_storeu_si128((__m128i*)(dst + i), E);
}
for (; i < len; ++i) {
const int diff_uv = ref[i] - src[i];
dst[i] += diff_uv;
}
}
static void SharpYUVFilterRow_SSE2(const int16_t* A, const int16_t* B, int len,
const uint16_t* best_y, uint16_t* out) {
int i;
const __m128i kCst8 = _mm_set1_epi16(8);
const __m128i max = _mm_set1_epi16(MAX_Y);
const __m128i zero = _mm_setzero_si128();
for (i = 0; i + 8 <= len; i += 8) {
const __m128i a0 = _mm_loadu_si128((const __m128i*)(A + i + 0));
const __m128i a1 = _mm_loadu_si128((const __m128i*)(A + i + 1));
const __m128i b0 = _mm_loadu_si128((const __m128i*)(B + i + 0));
const __m128i b1 = _mm_loadu_si128((const __m128i*)(B + i + 1));
const __m128i a0b1 = _mm_add_epi16(a0, b1);
const __m128i a1b0 = _mm_add_epi16(a1, b0);
const __m128i a0a1b0b1 = _mm_add_epi16(a0b1, a1b0); // A0+A1+B0+B1
const __m128i a0a1b0b1_8 = _mm_add_epi16(a0a1b0b1, kCst8);
const __m128i a0b1_2 = _mm_add_epi16(a0b1, a0b1); // 2*(A0+B1)
const __m128i a1b0_2 = _mm_add_epi16(a1b0, a1b0); // 2*(A1+B0)
const __m128i c0 = _mm_srai_epi16(_mm_add_epi16(a0b1_2, a0a1b0b1_8), 3);
const __m128i c1 = _mm_srai_epi16(_mm_add_epi16(a1b0_2, a0a1b0b1_8), 3);
const __m128i d0 = _mm_add_epi16(c1, a0);
const __m128i d1 = _mm_add_epi16(c0, a1);
const __m128i e0 = _mm_srai_epi16(d0, 1);
const __m128i e1 = _mm_srai_epi16(d1, 1);
const __m128i f0 = _mm_unpacklo_epi16(e0, e1);
const __m128i f1 = _mm_unpackhi_epi16(e0, e1);
const __m128i g0 = _mm_loadu_si128((const __m128i*)(best_y + 2 * i + 0));
const __m128i g1 = _mm_loadu_si128((const __m128i*)(best_y + 2 * i + 8));
const __m128i h0 = _mm_add_epi16(g0, f0);
const __m128i h1 = _mm_add_epi16(g1, f1);
const __m128i i0 = _mm_max_epi16(_mm_min_epi16(h0, max), zero);
const __m128i i1 = _mm_max_epi16(_mm_min_epi16(h1, max), zero);
_mm_storeu_si128((__m128i*)(out + 2 * i + 0), i0);
_mm_storeu_si128((__m128i*)(out + 2 * i + 8), i1);
}
for (; i < len; ++i) {
// (9 * A0 + 3 * A1 + 3 * B0 + B1 + 8) >> 4 =
// = (8 * A0 + 2 * (A1 + B0) + (A0 + A1 + B0 + B1 + 8)) >> 4
// We reuse the common sub-expressions.
const int a0b1 = A[i + 0] + B[i + 1];
const int a1b0 = A[i + 1] + B[i + 0];
const int a0a1b0b1 = a0b1 + a1b0 + 8;
const int v0 = (8 * A[i + 0] + 2 * a1b0 + a0a1b0b1) >> 4;
const int v1 = (8 * A[i + 1] + 2 * a0b1 + a0a1b0b1) >> 4;
out[2 * i + 0] = clip_y(best_y[2 * i + 0] + v0);
out[2 * i + 1] = clip_y(best_y[2 * i + 1] + v1);
}
}
#undef MAX_Y
//------------------------------------------------------------------------------
extern void WebPInitSharpYUVSSE2(void);
WEBP_TSAN_IGNORE_FUNCTION void WebPInitSharpYUVSSE2(void) {
WebPSharpYUVUpdateY = SharpYUVUpdateY_SSE2;
WebPSharpYUVUpdateRGB = SharpYUVUpdateRGB_SSE2;
WebPSharpYUVFilterRow = SharpYUVFilterRow_SSE2;
}
#else // !WEBP_USE_SSE2
WEBP_DSP_INIT_STUB(WebPInitSamplersSSE2)
WEBP_DSP_INIT_STUB(WebPInitConvertARGBToYUVSSE2)
WEBP_DSP_INIT_STUB(WebPInitSharpYUVSSE2)
#endif // WEBP_USE_SSE2

536
thirdparty/libwebp/dsp/yuv_tables_sse2.h vendored
View file

@ -1,536 +0,0 @@
// Copyright 2014 Google Inc. All Rights Reserved.
//
// Use of this source code is governed by a BSD-style license
// that can be found in the COPYING file in the root of the source
// tree. An additional intellectual property rights grant can be found
// in the file PATENTS. All contributing project authors may
// be found in the AUTHORS file in the root of the source tree.
// -----------------------------------------------------------------------------
//
// SSE2 tables for YUV->RGB conversion (12kB overall)
//
// Author: Skal (pascal.massimino@gmail.com)
// This file is not compiled, but #include'd directly from yuv.c
// Only used if WEBP_YUV_USE_SSE2_TABLES is defined.
static const VP8kCstSSE2 VP8kYtoRGBA[256] = {
{{0xfffb77b0, 0xfffb77b0, 0xfffb77b0, 0x003fc000}},
{{0xfffbc235, 0xfffbc235, 0xfffbc235, 0x003fc000}},
{{0xfffc0cba, 0xfffc0cba, 0xfffc0cba, 0x003fc000}},
{{0xfffc573f, 0xfffc573f, 0xfffc573f, 0x003fc000}},
{{0xfffca1c4, 0xfffca1c4, 0xfffca1c4, 0x003fc000}},
{{0xfffcec49, 0xfffcec49, 0xfffcec49, 0x003fc000}},
{{0xfffd36ce, 0xfffd36ce, 0xfffd36ce, 0x003fc000}},
{{0xfffd8153, 0xfffd8153, 0xfffd8153, 0x003fc000}},
{{0xfffdcbd8, 0xfffdcbd8, 0xfffdcbd8, 0x003fc000}},
{{0xfffe165d, 0xfffe165d, 0xfffe165d, 0x003fc000}},
{{0xfffe60e2, 0xfffe60e2, 0xfffe60e2, 0x003fc000}},
{{0xfffeab67, 0xfffeab67, 0xfffeab67, 0x003fc000}},
{{0xfffef5ec, 0xfffef5ec, 0xfffef5ec, 0x003fc000}},
{{0xffff4071, 0xffff4071, 0xffff4071, 0x003fc000}},
{{0xffff8af6, 0xffff8af6, 0xffff8af6, 0x003fc000}},
{{0xffffd57b, 0xffffd57b, 0xffffd57b, 0x003fc000}},
{{0x00002000, 0x00002000, 0x00002000, 0x003fc000}},
{{0x00006a85, 0x00006a85, 0x00006a85, 0x003fc000}},
{{0x0000b50a, 0x0000b50a, 0x0000b50a, 0x003fc000}},
{{0x0000ff8f, 0x0000ff8f, 0x0000ff8f, 0x003fc000}},
{{0x00014a14, 0x00014a14, 0x00014a14, 0x003fc000}},
{{0x00019499, 0x00019499, 0x00019499, 0x003fc000}},
{{0x0001df1e, 0x0001df1e, 0x0001df1e, 0x003fc000}},
{{0x000229a3, 0x000229a3, 0x000229a3, 0x003fc000}},
{{0x00027428, 0x00027428, 0x00027428, 0x003fc000}},
{{0x0002bead, 0x0002bead, 0x0002bead, 0x003fc000}},
{{0x00030932, 0x00030932, 0x00030932, 0x003fc000}},
{{0x000353b7, 0x000353b7, 0x000353b7, 0x003fc000}},
{{0x00039e3c, 0x00039e3c, 0x00039e3c, 0x003fc000}},
{{0x0003e8c1, 0x0003e8c1, 0x0003e8c1, 0x003fc000}},
{{0x00043346, 0x00043346, 0x00043346, 0x003fc000}},
{{0x00047dcb, 0x00047dcb, 0x00047dcb, 0x003fc000}},
{{0x0004c850, 0x0004c850, 0x0004c850, 0x003fc000}},
{{0x000512d5, 0x000512d5, 0x000512d5, 0x003fc000}},
{{0x00055d5a, 0x00055d5a, 0x00055d5a, 0x003fc000}},
{{0x0005a7df, 0x0005a7df, 0x0005a7df, 0x003fc000}},
{{0x0005f264, 0x0005f264, 0x0005f264, 0x003fc000}},
{{0x00063ce9, 0x00063ce9, 0x00063ce9, 0x003fc000}},
{{0x0006876e, 0x0006876e, 0x0006876e, 0x003fc000}},
{{0x0006d1f3, 0x0006d1f3, 0x0006d1f3, 0x003fc000}},
{{0x00071c78, 0x00071c78, 0x00071c78, 0x003fc000}},
{{0x000766fd, 0x000766fd, 0x000766fd, 0x003fc000}},
{{0x0007b182, 0x0007b182, 0x0007b182, 0x003fc000}},
{{0x0007fc07, 0x0007fc07, 0x0007fc07, 0x003fc000}},
{{0x0008468c, 0x0008468c, 0x0008468c, 0x003fc000}},
{{0x00089111, 0x00089111, 0x00089111, 0x003fc000}},
{{0x0008db96, 0x0008db96, 0x0008db96, 0x003fc000}},
{{0x0009261b, 0x0009261b, 0x0009261b, 0x003fc000}},
{{0x000970a0, 0x000970a0, 0x000970a0, 0x003fc000}},
{{0x0009bb25, 0x0009bb25, 0x0009bb25, 0x003fc000}},
{{0x000a05aa, 0x000a05aa, 0x000a05aa, 0x003fc000}},
{{0x000a502f, 0x000a502f, 0x000a502f, 0x003fc000}},
{{0x000a9ab4, 0x000a9ab4, 0x000a9ab4, 0x003fc000}},
{{0x000ae539, 0x000ae539, 0x000ae539, 0x003fc000}},
{{0x000b2fbe, 0x000b2fbe, 0x000b2fbe, 0x003fc000}},
{{0x000b7a43, 0x000b7a43, 0x000b7a43, 0x003fc000}},
{{0x000bc4c8, 0x000bc4c8, 0x000bc4c8, 0x003fc000}},
{{0x000c0f4d, 0x000c0f4d, 0x000c0f4d, 0x003fc000}},
{{0x000c59d2, 0x000c59d2, 0x000c59d2, 0x003fc000}},
{{0x000ca457, 0x000ca457, 0x000ca457, 0x003fc000}},
{{0x000ceedc, 0x000ceedc, 0x000ceedc, 0x003fc000}},
{{0x000d3961, 0x000d3961, 0x000d3961, 0x003fc000}},
{{0x000d83e6, 0x000d83e6, 0x000d83e6, 0x003fc000}},
{{0x000dce6b, 0x000dce6b, 0x000dce6b, 0x003fc000}},
{{0x000e18f0, 0x000e18f0, 0x000e18f0, 0x003fc000}},
{{0x000e6375, 0x000e6375, 0x000e6375, 0x003fc000}},
{{0x000eadfa, 0x000eadfa, 0x000eadfa, 0x003fc000}},
{{0x000ef87f, 0x000ef87f, 0x000ef87f, 0x003fc000}},
{{0x000f4304, 0x000f4304, 0x000f4304, 0x003fc000}},
{{0x000f8d89, 0x000f8d89, 0x000f8d89, 0x003fc000}},
{{0x000fd80e, 0x000fd80e, 0x000fd80e, 0x003fc000}},
{{0x00102293, 0x00102293, 0x00102293, 0x003fc000}},
{{0x00106d18, 0x00106d18, 0x00106d18, 0x003fc000}},
{{0x0010b79d, 0x0010b79d, 0x0010b79d, 0x003fc000}},
{{0x00110222, 0x00110222, 0x00110222, 0x003fc000}},
{{0x00114ca7, 0x00114ca7, 0x00114ca7, 0x003fc000}},
{{0x0011972c, 0x0011972c, 0x0011972c, 0x003fc000}},
{{0x0011e1b1, 0x0011e1b1, 0x0011e1b1, 0x003fc000}},
{{0x00122c36, 0x00122c36, 0x00122c36, 0x003fc000}},
{{0x001276bb, 0x001276bb, 0x001276bb, 0x003fc000}},
{{0x0012c140, 0x0012c140, 0x0012c140, 0x003fc000}},
{{0x00130bc5, 0x00130bc5, 0x00130bc5, 0x003fc000}},
{{0x0013564a, 0x0013564a, 0x0013564a, 0x003fc000}},
{{0x0013a0cf, 0x0013a0cf, 0x0013a0cf, 0x003fc000}},
{{0x0013eb54, 0x0013eb54, 0x0013eb54, 0x003fc000}},
{{0x001435d9, 0x001435d9, 0x001435d9, 0x003fc000}},
{{0x0014805e, 0x0014805e, 0x0014805e, 0x003fc000}},
{{0x0014cae3, 0x0014cae3, 0x0014cae3, 0x003fc000}},
{{0x00151568, 0x00151568, 0x00151568, 0x003fc000}},
{{0x00155fed, 0x00155fed, 0x00155fed, 0x003fc000}},
{{0x0015aa72, 0x0015aa72, 0x0015aa72, 0x003fc000}},
{{0x0015f4f7, 0x0015f4f7, 0x0015f4f7, 0x003fc000}},
{{0x00163f7c, 0x00163f7c, 0x00163f7c, 0x003fc000}},
{{0x00168a01, 0x00168a01, 0x00168a01, 0x003fc000}},
{{0x0016d486, 0x0016d486, 0x0016d486, 0x003fc000}},
{{0x00171f0b, 0x00171f0b, 0x00171f0b, 0x003fc000}},
{{0x00176990, 0x00176990, 0x00176990, 0x003fc000}},
{{0x0017b415, 0x0017b415, 0x0017b415, 0x003fc000}},
{{0x0017fe9a, 0x0017fe9a, 0x0017fe9a, 0x003fc000}},
{{0x0018491f, 0x0018491f, 0x0018491f, 0x003fc000}},
{{0x001893a4, 0x001893a4, 0x001893a4, 0x003fc000}},
{{0x0018de29, 0x0018de29, 0x0018de29, 0x003fc000}},
{{0x001928ae, 0x001928ae, 0x001928ae, 0x003fc000}},
{{0x00197333, 0x00197333, 0x00197333, 0x003fc000}},
{{0x0019bdb8, 0x0019bdb8, 0x0019bdb8, 0x003fc000}},
{{0x001a083d, 0x001a083d, 0x001a083d, 0x003fc000}},
{{0x001a52c2, 0x001a52c2, 0x001a52c2, 0x003fc000}},
{{0x001a9d47, 0x001a9d47, 0x001a9d47, 0x003fc000}},
{{0x001ae7cc, 0x001ae7cc, 0x001ae7cc, 0x003fc000}},
{{0x001b3251, 0x001b3251, 0x001b3251, 0x003fc000}},
{{0x001b7cd6, 0x001b7cd6, 0x001b7cd6, 0x003fc000}},
{{0x001bc75b, 0x001bc75b, 0x001bc75b, 0x003fc000}},
{{0x001c11e0, 0x001c11e0, 0x001c11e0, 0x003fc000}},
{{0x001c5c65, 0x001c5c65, 0x001c5c65, 0x003fc000}},
{{0x001ca6ea, 0x001ca6ea, 0x001ca6ea, 0x003fc000}},
{{0x001cf16f, 0x001cf16f, 0x001cf16f, 0x003fc000}},
{{0x001d3bf4, 0x001d3bf4, 0x001d3bf4, 0x003fc000}},
{{0x001d8679, 0x001d8679, 0x001d8679, 0x003fc000}},
{{0x001dd0fe, 0x001dd0fe, 0x001dd0fe, 0x003fc000}},
{{0x001e1b83, 0x001e1b83, 0x001e1b83, 0x003fc000}},
{{0x001e6608, 0x001e6608, 0x001e6608, 0x003fc000}},
{{0x001eb08d, 0x001eb08d, 0x001eb08d, 0x003fc000}},
{{0x001efb12, 0x001efb12, 0x001efb12, 0x003fc000}},
{{0x001f4597, 0x001f4597, 0x001f4597, 0x003fc000}},
{{0x001f901c, 0x001f901c, 0x001f901c, 0x003fc000}},
{{0x001fdaa1, 0x001fdaa1, 0x001fdaa1, 0x003fc000}},
{{0x00202526, 0x00202526, 0x00202526, 0x003fc000}},
{{0x00206fab, 0x00206fab, 0x00206fab, 0x003fc000}},
{{0x0020ba30, 0x0020ba30, 0x0020ba30, 0x003fc000}},
{{0x002104b5, 0x002104b5, 0x002104b5, 0x003fc000}},
{{0x00214f3a, 0x00214f3a, 0x00214f3a, 0x003fc000}},
{{0x002199bf, 0x002199bf, 0x002199bf, 0x003fc000}},
{{0x0021e444, 0x0021e444, 0x0021e444, 0x003fc000}},
{{0x00222ec9, 0x00222ec9, 0x00222ec9, 0x003fc000}},
{{0x0022794e, 0x0022794e, 0x0022794e, 0x003fc000}},
{{0x0022c3d3, 0x0022c3d3, 0x0022c3d3, 0x003fc000}},
{{0x00230e58, 0x00230e58, 0x00230e58, 0x003fc000}},
{{0x002358dd, 0x002358dd, 0x002358dd, 0x003fc000}},
{{0x0023a362, 0x0023a362, 0x0023a362, 0x003fc000}},
{{0x0023ede7, 0x0023ede7, 0x0023ede7, 0x003fc000}},
{{0x0024386c, 0x0024386c, 0x0024386c, 0x003fc000}},
{{0x002482f1, 0x002482f1, 0x002482f1, 0x003fc000}},
{{0x0024cd76, 0x0024cd76, 0x0024cd76, 0x003fc000}},
{{0x002517fb, 0x002517fb, 0x002517fb, 0x003fc000}},
{{0x00256280, 0x00256280, 0x00256280, 0x003fc000}},
{{0x0025ad05, 0x0025ad05, 0x0025ad05, 0x003fc000}},
{{0x0025f78a, 0x0025f78a, 0x0025f78a, 0x003fc000}},
{{0x0026420f, 0x0026420f, 0x0026420f, 0x003fc000}},
{{0x00268c94, 0x00268c94, 0x00268c94, 0x003fc000}},
{{0x0026d719, 0x0026d719, 0x0026d719, 0x003fc000}},
{{0x0027219e, 0x0027219e, 0x0027219e, 0x003fc000}},
{{0x00276c23, 0x00276c23, 0x00276c23, 0x003fc000}},
{{0x0027b6a8, 0x0027b6a8, 0x0027b6a8, 0x003fc000}},
{{0x0028012d, 0x0028012d, 0x0028012d, 0x003fc000}},
{{0x00284bb2, 0x00284bb2, 0x00284bb2, 0x003fc000}},
{{0x00289637, 0x00289637, 0x00289637, 0x003fc000}},
{{0x0028e0bc, 0x0028e0bc, 0x0028e0bc, 0x003fc000}},
{{0x00292b41, 0x00292b41, 0x00292b41, 0x003fc000}},
{{0x002975c6, 0x002975c6, 0x002975c6, 0x003fc000}},
{{0x0029c04b, 0x0029c04b, 0x0029c04b, 0x003fc000}},
{{0x002a0ad0, 0x002a0ad0, 0x002a0ad0, 0x003fc000}},
{{0x002a5555, 0x002a5555, 0x002a5555, 0x003fc000}},
{{0x002a9fda, 0x002a9fda, 0x002a9fda, 0x003fc000}},
{{0x002aea5f, 0x002aea5f, 0x002aea5f, 0x003fc000}},
{{0x002b34e4, 0x002b34e4, 0x002b34e4, 0x003fc000}},
{{0x002b7f69, 0x002b7f69, 0x002b7f69, 0x003fc000}},
{{0x002bc9ee, 0x002bc9ee, 0x002bc9ee, 0x003fc000}},
{{0x002c1473, 0x002c1473, 0x002c1473, 0x003fc000}},
{{0x002c5ef8, 0x002c5ef8, 0x002c5ef8, 0x003fc000}},
{{0x002ca97d, 0x002ca97d, 0x002ca97d, 0x003fc000}},
{{0x002cf402, 0x002cf402, 0x002cf402, 0x003fc000}},
{{0x002d3e87, 0x002d3e87, 0x002d3e87, 0x003fc000}},
{{0x002d890c, 0x002d890c, 0x002d890c, 0x003fc000}},
{{0x002dd391, 0x002dd391, 0x002dd391, 0x003fc000}},
{{0x002e1e16, 0x002e1e16, 0x002e1e16, 0x003fc000}},
{{0x002e689b, 0x002e689b, 0x002e689b, 0x003fc000}},
{{0x002eb320, 0x002eb320, 0x002eb320, 0x003fc000}},
{{0x002efda5, 0x002efda5, 0x002efda5, 0x003fc000}},
{{0x002f482a, 0x002f482a, 0x002f482a, 0x003fc000}},
{{0x002f92af, 0x002f92af, 0x002f92af, 0x003fc000}},
{{0x002fdd34, 0x002fdd34, 0x002fdd34, 0x003fc000}},
{{0x003027b9, 0x003027b9, 0x003027b9, 0x003fc000}},
{{0x0030723e, 0x0030723e, 0x0030723e, 0x003fc000}},
{{0x0030bcc3, 0x0030bcc3, 0x0030bcc3, 0x003fc000}},
{{0x00310748, 0x00310748, 0x00310748, 0x003fc000}},
{{0x003151cd, 0x003151cd, 0x003151cd, 0x003fc000}},
{{0x00319c52, 0x00319c52, 0x00319c52, 0x003fc000}},
{{0x0031e6d7, 0x0031e6d7, 0x0031e6d7, 0x003fc000}},
{{0x0032315c, 0x0032315c, 0x0032315c, 0x003fc000}},
{{0x00327be1, 0x00327be1, 0x00327be1, 0x003fc000}},
{{0x0032c666, 0x0032c666, 0x0032c666, 0x003fc000}},
{{0x003310eb, 0x003310eb, 0x003310eb, 0x003fc000}},
{{0x00335b70, 0x00335b70, 0x00335b70, 0x003fc000}},
{{0x0033a5f5, 0x0033a5f5, 0x0033a5f5, 0x003fc000}},
{{0x0033f07a, 0x0033f07a, 0x0033f07a, 0x003fc000}},
{{0x00343aff, 0x00343aff, 0x00343aff, 0x003fc000}},
{{0x00348584, 0x00348584, 0x00348584, 0x003fc000}},
{{0x0034d009, 0x0034d009, 0x0034d009, 0x003fc000}},
{{0x00351a8e, 0x00351a8e, 0x00351a8e, 0x003fc000}},
{{0x00356513, 0x00356513, 0x00356513, 0x003fc000}},
{{0x0035af98, 0x0035af98, 0x0035af98, 0x003fc000}},
{{0x0035fa1d, 0x0035fa1d, 0x0035fa1d, 0x003fc000}},
{{0x003644a2, 0x003644a2, 0x003644a2, 0x003fc000}},
{{0x00368f27, 0x00368f27, 0x00368f27, 0x003fc000}},
{{0x0036d9ac, 0x0036d9ac, 0x0036d9ac, 0x003fc000}},
{{0x00372431, 0x00372431, 0x00372431, 0x003fc000}},
{{0x00376eb6, 0x00376eb6, 0x00376eb6, 0x003fc000}},
{{0x0037b93b, 0x0037b93b, 0x0037b93b, 0x003fc000}},
{{0x003803c0, 0x003803c0, 0x003803c0, 0x003fc000}},
{{0x00384e45, 0x00384e45, 0x00384e45, 0x003fc000}},
{{0x003898ca, 0x003898ca, 0x003898ca, 0x003fc000}},
{{0x0038e34f, 0x0038e34f, 0x0038e34f, 0x003fc000}},
{{0x00392dd4, 0x00392dd4, 0x00392dd4, 0x003fc000}},
{{0x00397859, 0x00397859, 0x00397859, 0x003fc000}},
{{0x0039c2de, 0x0039c2de, 0x0039c2de, 0x003fc000}},
{{0x003a0d63, 0x003a0d63, 0x003a0d63, 0x003fc000}},
{{0x003a57e8, 0x003a57e8, 0x003a57e8, 0x003fc000}},
{{0x003aa26d, 0x003aa26d, 0x003aa26d, 0x003fc000}},
{{0x003aecf2, 0x003aecf2, 0x003aecf2, 0x003fc000}},
{{0x003b3777, 0x003b3777, 0x003b3777, 0x003fc000}},
{{0x003b81fc, 0x003b81fc, 0x003b81fc, 0x003fc000}},
{{0x003bcc81, 0x003bcc81, 0x003bcc81, 0x003fc000}},
{{0x003c1706, 0x003c1706, 0x003c1706, 0x003fc000}},
{{0x003c618b, 0x003c618b, 0x003c618b, 0x003fc000}},
{{0x003cac10, 0x003cac10, 0x003cac10, 0x003fc000}},
{{0x003cf695, 0x003cf695, 0x003cf695, 0x003fc000}},
{{0x003d411a, 0x003d411a, 0x003d411a, 0x003fc000}},
{{0x003d8b9f, 0x003d8b9f, 0x003d8b9f, 0x003fc000}},
{{0x003dd624, 0x003dd624, 0x003dd624, 0x003fc000}},
{{0x003e20a9, 0x003e20a9, 0x003e20a9, 0x003fc000}},
{{0x003e6b2e, 0x003e6b2e, 0x003e6b2e, 0x003fc000}},
{{0x003eb5b3, 0x003eb5b3, 0x003eb5b3, 0x003fc000}},
{{0x003f0038, 0x003f0038, 0x003f0038, 0x003fc000}},
{{0x003f4abd, 0x003f4abd, 0x003f4abd, 0x003fc000}},
{{0x003f9542, 0x003f9542, 0x003f9542, 0x003fc000}},
{{0x003fdfc7, 0x003fdfc7, 0x003fdfc7, 0x003fc000}},
{{0x00402a4c, 0x00402a4c, 0x00402a4c, 0x003fc000}},
{{0x004074d1, 0x004074d1, 0x004074d1, 0x003fc000}},
{{0x0040bf56, 0x0040bf56, 0x0040bf56, 0x003fc000}},
{{0x004109db, 0x004109db, 0x004109db, 0x003fc000}},
{{0x00415460, 0x00415460, 0x00415460, 0x003fc000}},
{{0x00419ee5, 0x00419ee5, 0x00419ee5, 0x003fc000}},
{{0x0041e96a, 0x0041e96a, 0x0041e96a, 0x003fc000}},
{{0x004233ef, 0x004233ef, 0x004233ef, 0x003fc000}},
{{0x00427e74, 0x00427e74, 0x00427e74, 0x003fc000}},
{{0x0042c8f9, 0x0042c8f9, 0x0042c8f9, 0x003fc000}},
{{0x0043137e, 0x0043137e, 0x0043137e, 0x003fc000}},
{{0x00435e03, 0x00435e03, 0x00435e03, 0x003fc000}},
{{0x0043a888, 0x0043a888, 0x0043a888, 0x003fc000}},
{{0x0043f30d, 0x0043f30d, 0x0043f30d, 0x003fc000}},
{{0x00443d92, 0x00443d92, 0x00443d92, 0x003fc000}},
{{0x00448817, 0x00448817, 0x00448817, 0x003fc000}},
{{0x0044d29c, 0x0044d29c, 0x0044d29c, 0x003fc000}},
{{0x00451d21, 0x00451d21, 0x00451d21, 0x003fc000}},
{{0x004567a6, 0x004567a6, 0x004567a6, 0x003fc000}},
{{0x0045b22b, 0x0045b22b, 0x0045b22b, 0x003fc000}}
};
static const VP8kCstSSE2 VP8kUtoRGBA[256] = {
{{0, 0x000c8980, 0xffbf7300, 0}}, {{0, 0x000c706d, 0xffbff41a, 0}},
{{0, 0x000c575a, 0xffc07534, 0}}, {{0, 0x000c3e47, 0xffc0f64e, 0}},
{{0, 0x000c2534, 0xffc17768, 0}}, {{0, 0x000c0c21, 0xffc1f882, 0}},
{{0, 0x000bf30e, 0xffc2799c, 0}}, {{0, 0x000bd9fb, 0xffc2fab6, 0}},
{{0, 0x000bc0e8, 0xffc37bd0, 0}}, {{0, 0x000ba7d5, 0xffc3fcea, 0}},
{{0, 0x000b8ec2, 0xffc47e04, 0}}, {{0, 0x000b75af, 0xffc4ff1e, 0}},
{{0, 0x000b5c9c, 0xffc58038, 0}}, {{0, 0x000b4389, 0xffc60152, 0}},
{{0, 0x000b2a76, 0xffc6826c, 0}}, {{0, 0x000b1163, 0xffc70386, 0}},
{{0, 0x000af850, 0xffc784a0, 0}}, {{0, 0x000adf3d, 0xffc805ba, 0}},
{{0, 0x000ac62a, 0xffc886d4, 0}}, {{0, 0x000aad17, 0xffc907ee, 0}},
{{0, 0x000a9404, 0xffc98908, 0}}, {{0, 0x000a7af1, 0xffca0a22, 0}},
{{0, 0x000a61de, 0xffca8b3c, 0}}, {{0, 0x000a48cb, 0xffcb0c56, 0}},
{{0, 0x000a2fb8, 0xffcb8d70, 0}}, {{0, 0x000a16a5, 0xffcc0e8a, 0}},
{{0, 0x0009fd92, 0xffcc8fa4, 0}}, {{0, 0x0009e47f, 0xffcd10be, 0}},
{{0, 0x0009cb6c, 0xffcd91d8, 0}}, {{0, 0x0009b259, 0xffce12f2, 0}},
{{0, 0x00099946, 0xffce940c, 0}}, {{0, 0x00098033, 0xffcf1526, 0}},
{{0, 0x00096720, 0xffcf9640, 0}}, {{0, 0x00094e0d, 0xffd0175a, 0}},
{{0, 0x000934fa, 0xffd09874, 0}}, {{0, 0x00091be7, 0xffd1198e, 0}},
{{0, 0x000902d4, 0xffd19aa8, 0}}, {{0, 0x0008e9c1, 0xffd21bc2, 0}},
{{0, 0x0008d0ae, 0xffd29cdc, 0}}, {{0, 0x0008b79b, 0xffd31df6, 0}},
{{0, 0x00089e88, 0xffd39f10, 0}}, {{0, 0x00088575, 0xffd4202a, 0}},
{{0, 0x00086c62, 0xffd4a144, 0}}, {{0, 0x0008534f, 0xffd5225e, 0}},
{{0, 0x00083a3c, 0xffd5a378, 0}}, {{0, 0x00082129, 0xffd62492, 0}},
{{0, 0x00080816, 0xffd6a5ac, 0}}, {{0, 0x0007ef03, 0xffd726c6, 0}},
{{0, 0x0007d5f0, 0xffd7a7e0, 0}}, {{0, 0x0007bcdd, 0xffd828fa, 0}},
{{0, 0x0007a3ca, 0xffd8aa14, 0}}, {{0, 0x00078ab7, 0xffd92b2e, 0}},
{{0, 0x000771a4, 0xffd9ac48, 0}}, {{0, 0x00075891, 0xffda2d62, 0}},
{{0, 0x00073f7e, 0xffdaae7c, 0}}, {{0, 0x0007266b, 0xffdb2f96, 0}},
{{0, 0x00070d58, 0xffdbb0b0, 0}}, {{0, 0x0006f445, 0xffdc31ca, 0}},
{{0, 0x0006db32, 0xffdcb2e4, 0}}, {{0, 0x0006c21f, 0xffdd33fe, 0}},
{{0, 0x0006a90c, 0xffddb518, 0}}, {{0, 0x00068ff9, 0xffde3632, 0}},
{{0, 0x000676e6, 0xffdeb74c, 0}}, {{0, 0x00065dd3, 0xffdf3866, 0}},
{{0, 0x000644c0, 0xffdfb980, 0}}, {{0, 0x00062bad, 0xffe03a9a, 0}},
{{0, 0x0006129a, 0xffe0bbb4, 0}}, {{0, 0x0005f987, 0xffe13cce, 0}},
{{0, 0x0005e074, 0xffe1bde8, 0}}, {{0, 0x0005c761, 0xffe23f02, 0}},
{{0, 0x0005ae4e, 0xffe2c01c, 0}}, {{0, 0x0005953b, 0xffe34136, 0}},
{{0, 0x00057c28, 0xffe3c250, 0}}, {{0, 0x00056315, 0xffe4436a, 0}},
{{0, 0x00054a02, 0xffe4c484, 0}}, {{0, 0x000530ef, 0xffe5459e, 0}},
{{0, 0x000517dc, 0xffe5c6b8, 0}}, {{0, 0x0004fec9, 0xffe647d2, 0}},
{{0, 0x0004e5b6, 0xffe6c8ec, 0}}, {{0, 0x0004cca3, 0xffe74a06, 0}},
{{0, 0x0004b390, 0xffe7cb20, 0}}, {{0, 0x00049a7d, 0xffe84c3a, 0}},
{{0, 0x0004816a, 0xffe8cd54, 0}}, {{0, 0x00046857, 0xffe94e6e, 0}},
{{0, 0x00044f44, 0xffe9cf88, 0}}, {{0, 0x00043631, 0xffea50a2, 0}},
{{0, 0x00041d1e, 0xffead1bc, 0}}, {{0, 0x0004040b, 0xffeb52d6, 0}},
{{0, 0x0003eaf8, 0xffebd3f0, 0}}, {{0, 0x0003d1e5, 0xffec550a, 0}},
{{0, 0x0003b8d2, 0xffecd624, 0}}, {{0, 0x00039fbf, 0xffed573e, 0}},
{{0, 0x000386ac, 0xffedd858, 0}}, {{0, 0x00036d99, 0xffee5972, 0}},
{{0, 0x00035486, 0xffeeda8c, 0}}, {{0, 0x00033b73, 0xffef5ba6, 0}},
{{0, 0x00032260, 0xffefdcc0, 0}}, {{0, 0x0003094d, 0xfff05dda, 0}},
{{0, 0x0002f03a, 0xfff0def4, 0}}, {{0, 0x0002d727, 0xfff1600e, 0}},
{{0, 0x0002be14, 0xfff1e128, 0}}, {{0, 0x0002a501, 0xfff26242, 0}},
{{0, 0x00028bee, 0xfff2e35c, 0}}, {{0, 0x000272db, 0xfff36476, 0}},
{{0, 0x000259c8, 0xfff3e590, 0}}, {{0, 0x000240b5, 0xfff466aa, 0}},
{{0, 0x000227a2, 0xfff4e7c4, 0}}, {{0, 0x00020e8f, 0xfff568de, 0}},
{{0, 0x0001f57c, 0xfff5e9f8, 0}}, {{0, 0x0001dc69, 0xfff66b12, 0}},
{{0, 0x0001c356, 0xfff6ec2c, 0}}, {{0, 0x0001aa43, 0xfff76d46, 0}},
{{0, 0x00019130, 0xfff7ee60, 0}}, {{0, 0x0001781d, 0xfff86f7a, 0}},
{{0, 0x00015f0a, 0xfff8f094, 0}}, {{0, 0x000145f7, 0xfff971ae, 0}},
{{0, 0x00012ce4, 0xfff9f2c8, 0}}, {{0, 0x000113d1, 0xfffa73e2, 0}},
{{0, 0x0000fabe, 0xfffaf4fc, 0}}, {{0, 0x0000e1ab, 0xfffb7616, 0}},
{{0, 0x0000c898, 0xfffbf730, 0}}, {{0, 0x0000af85, 0xfffc784a, 0}},
{{0, 0x00009672, 0xfffcf964, 0}}, {{0, 0x00007d5f, 0xfffd7a7e, 0}},
{{0, 0x0000644c, 0xfffdfb98, 0}}, {{0, 0x00004b39, 0xfffe7cb2, 0}},
{{0, 0x00003226, 0xfffefdcc, 0}}, {{0, 0x00001913, 0xffff7ee6, 0}},
{{0, 0x00000000, 0x00000000, 0}}, {{0, 0xffffe6ed, 0x0000811a, 0}},
{{0, 0xffffcdda, 0x00010234, 0}}, {{0, 0xffffb4c7, 0x0001834e, 0}},
{{0, 0xffff9bb4, 0x00020468, 0}}, {{0, 0xffff82a1, 0x00028582, 0}},
{{0, 0xffff698e, 0x0003069c, 0}}, {{0, 0xffff507b, 0x000387b6, 0}},
{{0, 0xffff3768, 0x000408d0, 0}}, {{0, 0xffff1e55, 0x000489ea, 0}},
{{0, 0xffff0542, 0x00050b04, 0}}, {{0, 0xfffeec2f, 0x00058c1e, 0}},
{{0, 0xfffed31c, 0x00060d38, 0}}, {{0, 0xfffeba09, 0x00068e52, 0}},
{{0, 0xfffea0f6, 0x00070f6c, 0}}, {{0, 0xfffe87e3, 0x00079086, 0}},
{{0, 0xfffe6ed0, 0x000811a0, 0}}, {{0, 0xfffe55bd, 0x000892ba, 0}},
{{0, 0xfffe3caa, 0x000913d4, 0}}, {{0, 0xfffe2397, 0x000994ee, 0}},
{{0, 0xfffe0a84, 0x000a1608, 0}}, {{0, 0xfffdf171, 0x000a9722, 0}},
{{0, 0xfffdd85e, 0x000b183c, 0}}, {{0, 0xfffdbf4b, 0x000b9956, 0}},
{{0, 0xfffda638, 0x000c1a70, 0}}, {{0, 0xfffd8d25, 0x000c9b8a, 0}},
{{0, 0xfffd7412, 0x000d1ca4, 0}}, {{0, 0xfffd5aff, 0x000d9dbe, 0}},
{{0, 0xfffd41ec, 0x000e1ed8, 0}}, {{0, 0xfffd28d9, 0x000e9ff2, 0}},
{{0, 0xfffd0fc6, 0x000f210c, 0}}, {{0, 0xfffcf6b3, 0x000fa226, 0}},
{{0, 0xfffcdda0, 0x00102340, 0}}, {{0, 0xfffcc48d, 0x0010a45a, 0}},
{{0, 0xfffcab7a, 0x00112574, 0}}, {{0, 0xfffc9267, 0x0011a68e, 0}},
{{0, 0xfffc7954, 0x001227a8, 0}}, {{0, 0xfffc6041, 0x0012a8c2, 0}},
{{0, 0xfffc472e, 0x001329dc, 0}}, {{0, 0xfffc2e1b, 0x0013aaf6, 0}},
{{0, 0xfffc1508, 0x00142c10, 0}}, {{0, 0xfffbfbf5, 0x0014ad2a, 0}},
{{0, 0xfffbe2e2, 0x00152e44, 0}}, {{0, 0xfffbc9cf, 0x0015af5e, 0}},
{{0, 0xfffbb0bc, 0x00163078, 0}}, {{0, 0xfffb97a9, 0x0016b192, 0}},
{{0, 0xfffb7e96, 0x001732ac, 0}}, {{0, 0xfffb6583, 0x0017b3c6, 0}},
{{0, 0xfffb4c70, 0x001834e0, 0}}, {{0, 0xfffb335d, 0x0018b5fa, 0}},
{{0, 0xfffb1a4a, 0x00193714, 0}}, {{0, 0xfffb0137, 0x0019b82e, 0}},
{{0, 0xfffae824, 0x001a3948, 0}}, {{0, 0xfffacf11, 0x001aba62, 0}},
{{0, 0xfffab5fe, 0x001b3b7c, 0}}, {{0, 0xfffa9ceb, 0x001bbc96, 0}},
{{0, 0xfffa83d8, 0x001c3db0, 0}}, {{0, 0xfffa6ac5, 0x001cbeca, 0}},
{{0, 0xfffa51b2, 0x001d3fe4, 0}}, {{0, 0xfffa389f, 0x001dc0fe, 0}},
{{0, 0xfffa1f8c, 0x001e4218, 0}}, {{0, 0xfffa0679, 0x001ec332, 0}},
{{0, 0xfff9ed66, 0x001f444c, 0}}, {{0, 0xfff9d453, 0x001fc566, 0}},
{{0, 0xfff9bb40, 0x00204680, 0}}, {{0, 0xfff9a22d, 0x0020c79a, 0}},
{{0, 0xfff9891a, 0x002148b4, 0}}, {{0, 0xfff97007, 0x0021c9ce, 0}},
{{0, 0xfff956f4, 0x00224ae8, 0}}, {{0, 0xfff93de1, 0x0022cc02, 0}},
{{0, 0xfff924ce, 0x00234d1c, 0}}, {{0, 0xfff90bbb, 0x0023ce36, 0}},
{{0, 0xfff8f2a8, 0x00244f50, 0}}, {{0, 0xfff8d995, 0x0024d06a, 0}},
{{0, 0xfff8c082, 0x00255184, 0}}, {{0, 0xfff8a76f, 0x0025d29e, 0}},
{{0, 0xfff88e5c, 0x002653b8, 0}}, {{0, 0xfff87549, 0x0026d4d2, 0}},
{{0, 0xfff85c36, 0x002755ec, 0}}, {{0, 0xfff84323, 0x0027d706, 0}},
{{0, 0xfff82a10, 0x00285820, 0}}, {{0, 0xfff810fd, 0x0028d93a, 0}},
{{0, 0xfff7f7ea, 0x00295a54, 0}}, {{0, 0xfff7ded7, 0x0029db6e, 0}},
{{0, 0xfff7c5c4, 0x002a5c88, 0}}, {{0, 0xfff7acb1, 0x002adda2, 0}},
{{0, 0xfff7939e, 0x002b5ebc, 0}}, {{0, 0xfff77a8b, 0x002bdfd6, 0}},
{{0, 0xfff76178, 0x002c60f0, 0}}, {{0, 0xfff74865, 0x002ce20a, 0}},
{{0, 0xfff72f52, 0x002d6324, 0}}, {{0, 0xfff7163f, 0x002de43e, 0}},
{{0, 0xfff6fd2c, 0x002e6558, 0}}, {{0, 0xfff6e419, 0x002ee672, 0}},
{{0, 0xfff6cb06, 0x002f678c, 0}}, {{0, 0xfff6b1f3, 0x002fe8a6, 0}},
{{0, 0xfff698e0, 0x003069c0, 0}}, {{0, 0xfff67fcd, 0x0030eada, 0}},
{{0, 0xfff666ba, 0x00316bf4, 0}}, {{0, 0xfff64da7, 0x0031ed0e, 0}},
{{0, 0xfff63494, 0x00326e28, 0}}, {{0, 0xfff61b81, 0x0032ef42, 0}},
{{0, 0xfff6026e, 0x0033705c, 0}}, {{0, 0xfff5e95b, 0x0033f176, 0}},
{{0, 0xfff5d048, 0x00347290, 0}}, {{0, 0xfff5b735, 0x0034f3aa, 0}},
{{0, 0xfff59e22, 0x003574c4, 0}}, {{0, 0xfff5850f, 0x0035f5de, 0}},
{{0, 0xfff56bfc, 0x003676f8, 0}}, {{0, 0xfff552e9, 0x0036f812, 0}},
{{0, 0xfff539d6, 0x0037792c, 0}}, {{0, 0xfff520c3, 0x0037fa46, 0}},
{{0, 0xfff507b0, 0x00387b60, 0}}, {{0, 0xfff4ee9d, 0x0038fc7a, 0}},
{{0, 0xfff4d58a, 0x00397d94, 0}}, {{0, 0xfff4bc77, 0x0039feae, 0}},
{{0, 0xfff4a364, 0x003a7fc8, 0}}, {{0, 0xfff48a51, 0x003b00e2, 0}},
{{0, 0xfff4713e, 0x003b81fc, 0}}, {{0, 0xfff4582b, 0x003c0316, 0}},
{{0, 0xfff43f18, 0x003c8430, 0}}, {{0, 0xfff42605, 0x003d054a, 0}},
{{0, 0xfff40cf2, 0x003d8664, 0}}, {{0, 0xfff3f3df, 0x003e077e, 0}},
{{0, 0xfff3dacc, 0x003e8898, 0}}, {{0, 0xfff3c1b9, 0x003f09b2, 0}},
{{0, 0xfff3a8a6, 0x003f8acc, 0}}, {{0, 0xfff38f93, 0x00400be6, 0}}
};
static VP8kCstSSE2 VP8kVtoRGBA[256] = {
{{0xffcced80, 0x001a0400, 0, 0}}, {{0xffcd53a5, 0x0019cff8, 0, 0}},
{{0xffcdb9ca, 0x00199bf0, 0, 0}}, {{0xffce1fef, 0x001967e8, 0, 0}},
{{0xffce8614, 0x001933e0, 0, 0}}, {{0xffceec39, 0x0018ffd8, 0, 0}},
{{0xffcf525e, 0x0018cbd0, 0, 0}}, {{0xffcfb883, 0x001897c8, 0, 0}},
{{0xffd01ea8, 0x001863c0, 0, 0}}, {{0xffd084cd, 0x00182fb8, 0, 0}},
{{0xffd0eaf2, 0x0017fbb0, 0, 0}}, {{0xffd15117, 0x0017c7a8, 0, 0}},
{{0xffd1b73c, 0x001793a0, 0, 0}}, {{0xffd21d61, 0x00175f98, 0, 0}},
{{0xffd28386, 0x00172b90, 0, 0}}, {{0xffd2e9ab, 0x0016f788, 0, 0}},
{{0xffd34fd0, 0x0016c380, 0, 0}}, {{0xffd3b5f5, 0x00168f78, 0, 0}},
{{0xffd41c1a, 0x00165b70, 0, 0}}, {{0xffd4823f, 0x00162768, 0, 0}},
{{0xffd4e864, 0x0015f360, 0, 0}}, {{0xffd54e89, 0x0015bf58, 0, 0}},
{{0xffd5b4ae, 0x00158b50, 0, 0}}, {{0xffd61ad3, 0x00155748, 0, 0}},
{{0xffd680f8, 0x00152340, 0, 0}}, {{0xffd6e71d, 0x0014ef38, 0, 0}},
{{0xffd74d42, 0x0014bb30, 0, 0}}, {{0xffd7b367, 0x00148728, 0, 0}},
{{0xffd8198c, 0x00145320, 0, 0}}, {{0xffd87fb1, 0x00141f18, 0, 0}},
{{0xffd8e5d6, 0x0013eb10, 0, 0}}, {{0xffd94bfb, 0x0013b708, 0, 0}},
{{0xffd9b220, 0x00138300, 0, 0}}, {{0xffda1845, 0x00134ef8, 0, 0}},
{{0xffda7e6a, 0x00131af0, 0, 0}}, {{0xffdae48f, 0x0012e6e8, 0, 0}},
{{0xffdb4ab4, 0x0012b2e0, 0, 0}}, {{0xffdbb0d9, 0x00127ed8, 0, 0}},
{{0xffdc16fe, 0x00124ad0, 0, 0}}, {{0xffdc7d23, 0x001216c8, 0, 0}},
{{0xffdce348, 0x0011e2c0, 0, 0}}, {{0xffdd496d, 0x0011aeb8, 0, 0}},
{{0xffddaf92, 0x00117ab0, 0, 0}}, {{0xffde15b7, 0x001146a8, 0, 0}},
{{0xffde7bdc, 0x001112a0, 0, 0}}, {{0xffdee201, 0x0010de98, 0, 0}},
{{0xffdf4826, 0x0010aa90, 0, 0}}, {{0xffdfae4b, 0x00107688, 0, 0}},
{{0xffe01470, 0x00104280, 0, 0}}, {{0xffe07a95, 0x00100e78, 0, 0}},
{{0xffe0e0ba, 0x000fda70, 0, 0}}, {{0xffe146df, 0x000fa668, 0, 0}},
{{0xffe1ad04, 0x000f7260, 0, 0}}, {{0xffe21329, 0x000f3e58, 0, 0}},
{{0xffe2794e, 0x000f0a50, 0, 0}}, {{0xffe2df73, 0x000ed648, 0, 0}},
{{0xffe34598, 0x000ea240, 0, 0}}, {{0xffe3abbd, 0x000e6e38, 0, 0}},
{{0xffe411e2, 0x000e3a30, 0, 0}}, {{0xffe47807, 0x000e0628, 0, 0}},
{{0xffe4de2c, 0x000dd220, 0, 0}}, {{0xffe54451, 0x000d9e18, 0, 0}},
{{0xffe5aa76, 0x000d6a10, 0, 0}}, {{0xffe6109b, 0x000d3608, 0, 0}},
{{0xffe676c0, 0x000d0200, 0, 0}}, {{0xffe6dce5, 0x000ccdf8, 0, 0}},
{{0xffe7430a, 0x000c99f0, 0, 0}}, {{0xffe7a92f, 0x000c65e8, 0, 0}},
{{0xffe80f54, 0x000c31e0, 0, 0}}, {{0xffe87579, 0x000bfdd8, 0, 0}},
{{0xffe8db9e, 0x000bc9d0, 0, 0}}, {{0xffe941c3, 0x000b95c8, 0, 0}},
{{0xffe9a7e8, 0x000b61c0, 0, 0}}, {{0xffea0e0d, 0x000b2db8, 0, 0}},
{{0xffea7432, 0x000af9b0, 0, 0}}, {{0xffeada57, 0x000ac5a8, 0, 0}},
{{0xffeb407c, 0x000a91a0, 0, 0}}, {{0xffeba6a1, 0x000a5d98, 0, 0}},
{{0xffec0cc6, 0x000a2990, 0, 0}}, {{0xffec72eb, 0x0009f588, 0, 0}},
{{0xffecd910, 0x0009c180, 0, 0}}, {{0xffed3f35, 0x00098d78, 0, 0}},
{{0xffeda55a, 0x00095970, 0, 0}}, {{0xffee0b7f, 0x00092568, 0, 0}},
{{0xffee71a4, 0x0008f160, 0, 0}}, {{0xffeed7c9, 0x0008bd58, 0, 0}},
{{0xffef3dee, 0x00088950, 0, 0}}, {{0xffefa413, 0x00085548, 0, 0}},
{{0xfff00a38, 0x00082140, 0, 0}}, {{0xfff0705d, 0x0007ed38, 0, 0}},
{{0xfff0d682, 0x0007b930, 0, 0}}, {{0xfff13ca7, 0x00078528, 0, 0}},
{{0xfff1a2cc, 0x00075120, 0, 0}}, {{0xfff208f1, 0x00071d18, 0, 0}},
{{0xfff26f16, 0x0006e910, 0, 0}}, {{0xfff2d53b, 0x0006b508, 0, 0}},
{{0xfff33b60, 0x00068100, 0, 0}}, {{0xfff3a185, 0x00064cf8, 0, 0}},
{{0xfff407aa, 0x000618f0, 0, 0}}, {{0xfff46dcf, 0x0005e4e8, 0, 0}},
{{0xfff4d3f4, 0x0005b0e0, 0, 0}}, {{0xfff53a19, 0x00057cd8, 0, 0}},
{{0xfff5a03e, 0x000548d0, 0, 0}}, {{0xfff60663, 0x000514c8, 0, 0}},
{{0xfff66c88, 0x0004e0c0, 0, 0}}, {{0xfff6d2ad, 0x0004acb8, 0, 0}},
{{0xfff738d2, 0x000478b0, 0, 0}}, {{0xfff79ef7, 0x000444a8, 0, 0}},
{{0xfff8051c, 0x000410a0, 0, 0}}, {{0xfff86b41, 0x0003dc98, 0, 0}},
{{0xfff8d166, 0x0003a890, 0, 0}}, {{0xfff9378b, 0x00037488, 0, 0}},
{{0xfff99db0, 0x00034080, 0, 0}}, {{0xfffa03d5, 0x00030c78, 0, 0}},
{{0xfffa69fa, 0x0002d870, 0, 0}}, {{0xfffad01f, 0x0002a468, 0, 0}},
{{0xfffb3644, 0x00027060, 0, 0}}, {{0xfffb9c69, 0x00023c58, 0, 0}},
{{0xfffc028e, 0x00020850, 0, 0}}, {{0xfffc68b3, 0x0001d448, 0, 0}},
{{0xfffcced8, 0x0001a040, 0, 0}}, {{0xfffd34fd, 0x00016c38, 0, 0}},
{{0xfffd9b22, 0x00013830, 0, 0}}, {{0xfffe0147, 0x00010428, 0, 0}},
{{0xfffe676c, 0x0000d020, 0, 0}}, {{0xfffecd91, 0x00009c18, 0, 0}},
{{0xffff33b6, 0x00006810, 0, 0}}, {{0xffff99db, 0x00003408, 0, 0}},
{{0x00000000, 0x00000000, 0, 0}}, {{0x00006625, 0xffffcbf8, 0, 0}},
{{0x0000cc4a, 0xffff97f0, 0, 0}}, {{0x0001326f, 0xffff63e8, 0, 0}},
{{0x00019894, 0xffff2fe0, 0, 0}}, {{0x0001feb9, 0xfffefbd8, 0, 0}},
{{0x000264de, 0xfffec7d0, 0, 0}}, {{0x0002cb03, 0xfffe93c8, 0, 0}},
{{0x00033128, 0xfffe5fc0, 0, 0}}, {{0x0003974d, 0xfffe2bb8, 0, 0}},
{{0x0003fd72, 0xfffdf7b0, 0, 0}}, {{0x00046397, 0xfffdc3a8, 0, 0}},
{{0x0004c9bc, 0xfffd8fa0, 0, 0}}, {{0x00052fe1, 0xfffd5b98, 0, 0}},
{{0x00059606, 0xfffd2790, 0, 0}}, {{0x0005fc2b, 0xfffcf388, 0, 0}},
{{0x00066250, 0xfffcbf80, 0, 0}}, {{0x0006c875, 0xfffc8b78, 0, 0}},
{{0x00072e9a, 0xfffc5770, 0, 0}}, {{0x000794bf, 0xfffc2368, 0, 0}},
{{0x0007fae4, 0xfffbef60, 0, 0}}, {{0x00086109, 0xfffbbb58, 0, 0}},
{{0x0008c72e, 0xfffb8750, 0, 0}}, {{0x00092d53, 0xfffb5348, 0, 0}},
{{0x00099378, 0xfffb1f40, 0, 0}}, {{0x0009f99d, 0xfffaeb38, 0, 0}},
{{0x000a5fc2, 0xfffab730, 0, 0}}, {{0x000ac5e7, 0xfffa8328, 0, 0}},
{{0x000b2c0c, 0xfffa4f20, 0, 0}}, {{0x000b9231, 0xfffa1b18, 0, 0}},
{{0x000bf856, 0xfff9e710, 0, 0}}, {{0x000c5e7b, 0xfff9b308, 0, 0}},
{{0x000cc4a0, 0xfff97f00, 0, 0}}, {{0x000d2ac5, 0xfff94af8, 0, 0}},
{{0x000d90ea, 0xfff916f0, 0, 0}}, {{0x000df70f, 0xfff8e2e8, 0, 0}},
{{0x000e5d34, 0xfff8aee0, 0, 0}}, {{0x000ec359, 0xfff87ad8, 0, 0}},
{{0x000f297e, 0xfff846d0, 0, 0}}, {{0x000f8fa3, 0xfff812c8, 0, 0}},
{{0x000ff5c8, 0xfff7dec0, 0, 0}}, {{0x00105bed, 0xfff7aab8, 0, 0}},
{{0x0010c212, 0xfff776b0, 0, 0}}, {{0x00112837, 0xfff742a8, 0, 0}},
{{0x00118e5c, 0xfff70ea0, 0, 0}}, {{0x0011f481, 0xfff6da98, 0, 0}},
{{0x00125aa6, 0xfff6a690, 0, 0}}, {{0x0012c0cb, 0xfff67288, 0, 0}},
{{0x001326f0, 0xfff63e80, 0, 0}}, {{0x00138d15, 0xfff60a78, 0, 0}},
{{0x0013f33a, 0xfff5d670, 0, 0}}, {{0x0014595f, 0xfff5a268, 0, 0}},
{{0x0014bf84, 0xfff56e60, 0, 0}}, {{0x001525a9, 0xfff53a58, 0, 0}},
{{0x00158bce, 0xfff50650, 0, 0}}, {{0x0015f1f3, 0xfff4d248, 0, 0}},
{{0x00165818, 0xfff49e40, 0, 0}}, {{0x0016be3d, 0xfff46a38, 0, 0}},
{{0x00172462, 0xfff43630, 0, 0}}, {{0x00178a87, 0xfff40228, 0, 0}},
{{0x0017f0ac, 0xfff3ce20, 0, 0}}, {{0x001856d1, 0xfff39a18, 0, 0}},
{{0x0018bcf6, 0xfff36610, 0, 0}}, {{0x0019231b, 0xfff33208, 0, 0}},
{{0x00198940, 0xfff2fe00, 0, 0}}, {{0x0019ef65, 0xfff2c9f8, 0, 0}},
{{0x001a558a, 0xfff295f0, 0, 0}}, {{0x001abbaf, 0xfff261e8, 0, 0}},
{{0x001b21d4, 0xfff22de0, 0, 0}}, {{0x001b87f9, 0xfff1f9d8, 0, 0}},
{{0x001bee1e, 0xfff1c5d0, 0, 0}}, {{0x001c5443, 0xfff191c8, 0, 0}},
{{0x001cba68, 0xfff15dc0, 0, 0}}, {{0x001d208d, 0xfff129b8, 0, 0}},
{{0x001d86b2, 0xfff0f5b0, 0, 0}}, {{0x001decd7, 0xfff0c1a8, 0, 0}},
{{0x001e52fc, 0xfff08da0, 0, 0}}, {{0x001eb921, 0xfff05998, 0, 0}},
{{0x001f1f46, 0xfff02590, 0, 0}}, {{0x001f856b, 0xffeff188, 0, 0}},
{{0x001feb90, 0xffefbd80, 0, 0}}, {{0x002051b5, 0xffef8978, 0, 0}},
{{0x0020b7da, 0xffef5570, 0, 0}}, {{0x00211dff, 0xffef2168, 0, 0}},
{{0x00218424, 0xffeeed60, 0, 0}}, {{0x0021ea49, 0xffeeb958, 0, 0}},
{{0x0022506e, 0xffee8550, 0, 0}}, {{0x0022b693, 0xffee5148, 0, 0}},
{{0x00231cb8, 0xffee1d40, 0, 0}}, {{0x002382dd, 0xffede938, 0, 0}},
{{0x0023e902, 0xffedb530, 0, 0}}, {{0x00244f27, 0xffed8128, 0, 0}},
{{0x0024b54c, 0xffed4d20, 0, 0}}, {{0x00251b71, 0xffed1918, 0, 0}},
{{0x00258196, 0xffece510, 0, 0}}, {{0x0025e7bb, 0xffecb108, 0, 0}},
{{0x00264de0, 0xffec7d00, 0, 0}}, {{0x0026b405, 0xffec48f8, 0, 0}},
{{0x00271a2a, 0xffec14f0, 0, 0}}, {{0x0027804f, 0xffebe0e8, 0, 0}},
{{0x0027e674, 0xffebace0, 0, 0}}, {{0x00284c99, 0xffeb78d8, 0, 0}},
{{0x0028b2be, 0xffeb44d0, 0, 0}}, {{0x002918e3, 0xffeb10c8, 0, 0}},
{{0x00297f08, 0xffeadcc0, 0, 0}}, {{0x0029e52d, 0xffeaa8b8, 0, 0}},
{{0x002a4b52, 0xffea74b0, 0, 0}}, {{0x002ab177, 0xffea40a8, 0, 0}},
{{0x002b179c, 0xffea0ca0, 0, 0}}, {{0x002b7dc1, 0xffe9d898, 0, 0}},
{{0x002be3e6, 0xffe9a490, 0, 0}}, {{0x002c4a0b, 0xffe97088, 0, 0}},
{{0x002cb030, 0xffe93c80, 0, 0}}, {{0x002d1655, 0xffe90878, 0, 0}},
{{0x002d7c7a, 0xffe8d470, 0, 0}}, {{0x002de29f, 0xffe8a068, 0, 0}},
{{0x002e48c4, 0xffe86c60, 0, 0}}, {{0x002eaee9, 0xffe83858, 0, 0}},
{{0x002f150e, 0xffe80450, 0, 0}}, {{0x002f7b33, 0xffe7d048, 0, 0}},
{{0x002fe158, 0xffe79c40, 0, 0}}, {{0x0030477d, 0xffe76838, 0, 0}},
{{0x0030ada2, 0xffe73430, 0, 0}}, {{0x003113c7, 0xffe70028, 0, 0}},
{{0x003179ec, 0xffe6cc20, 0, 0}}, {{0x0031e011, 0xffe69818, 0, 0}},
{{0x00324636, 0xffe66410, 0, 0}}, {{0x0032ac5b, 0xffe63008, 0, 0}}
};

8
thirdparty/libwebp/enc/alpha.c → thirdparty/libwebp/enc/alpha_enc.c vendored
View file

@ -14,10 +14,10 @@
#include <assert.h>
#include <stdlib.h>
#include "./vp8enci.h"
#include "./vp8i_enc.h"
#include "../dsp/dsp.h"
#include "../utils/filters.h"
#include "../utils/quant_levels.h"
#include "../utils/filters_utils.h"
#include "../utils/quant_levels_utils.h"
#include "../utils/utils.h"
#include "../webp/format_constants.h"
@ -44,7 +44,7 @@
// invalid quality or method, or
// memory allocation for the compressed data fails.
#include "../enc/vp8li.h"
#include "../enc/vp8li_enc.h"
static int EncodeLossless(const uint8_t* const data, int width, int height,
int effort_level, // in [0..6] range

47
thirdparty/libwebp/enc/analysis.c → thirdparty/libwebp/enc/analysis_enc.c vendored
View file

@ -15,8 +15,8 @@
#include <string.h>
#include <assert.h>
#include "./vp8enci.h"
#include "./cost.h"
#include "./vp8i_enc.h"
#include "./cost_enc.h"
#include "../utils/utils.h"
#define MAX_ITERS_K_MEANS 6
@ -262,6 +262,29 @@ static int MBAnalyzeBestIntra16Mode(VP8EncIterator* const it) {
return best_alpha;
}
static int FastMBAnalyze(VP8EncIterator* const it) {
// Empirical cut-off value, should be around 16 (~=block size). We use the
// [8-17] range and favor intra4 at high quality, intra16 for low quality.
const int q = (int)it->enc_->config_->quality;
const uint32_t kThreshold = 8 + (17 - 8) * q / 100;
int k;
uint32_t dc[16], m, m2;
for (k = 0; k < 16; k += 4) {
VP8Mean16x4(it->yuv_in_ + Y_OFF_ENC + k * BPS, &dc[k]);
}
for (m = 0, m2 = 0, k = 0; k < 16; ++k) {
m += dc[k];
m2 += dc[k] * dc[k];
}
if (kThreshold * m2 < m * m) {
VP8SetIntra16Mode(it, 0); // DC16
} else {
const uint8_t modes[16] = { 0 }; // DC4
VP8SetIntra4Mode(it, modes);
}
return 0;
}
static int MBAnalyzeBestIntra4Mode(VP8EncIterator* const it,
int best_alpha) {
uint8_t modes[16];
@ -344,13 +367,17 @@ static void MBAnalyze(VP8EncIterator* const it,
VP8SetSkip(it, 0); // not skipped
VP8SetSegment(it, 0); // default segment, spec-wise.
best_alpha = MBAnalyzeBestIntra16Mode(it);
if (enc->method_ >= 5) {
// We go and make a fast decision for intra4/intra16.
// It's usually not a good and definitive pick, but helps seeding the stats
// about level bit-cost.
// TODO(skal): improve criterion.
best_alpha = MBAnalyzeBestIntra4Mode(it, best_alpha);
if (enc->method_ <= 1) {
best_alpha = FastMBAnalyze(it);
} else {
best_alpha = MBAnalyzeBestIntra16Mode(it);
if (enc->method_ >= 5) {
// We go and make a fast decision for intra4/intra16.
// It's usually not a good and definitive pick, but helps seeding the
// stats about level bit-cost.
// TODO(skal): improve criterion.
best_alpha = MBAnalyzeBestIntra4Mode(it, best_alpha);
}
}
best_uv_alpha = MBAnalyzeBestUVMode(it);
@ -453,7 +480,7 @@ int VP8EncAnalyze(VP8Encoder* const enc) {
const int do_segments =
enc->config_->emulate_jpeg_size || // We need the complexity evaluation.
(enc->segment_hdr_.num_segments_ > 1) ||
(enc->method_ == 0); // for method 0, we need preds_[] to be filled.
(enc->method_ <= 1); // for method 0 - 1, we need preds_[] to be filled.
if (do_segments) {
const int last_row = enc->mb_h_;
// We give a little more than a half work to the main thread.

327
thirdparty/libwebp/enc/backward_references.c → thirdparty/libwebp/enc/backward_references_enc.c vendored
View file

@ -13,11 +13,12 @@
#include <assert.h>
#include <math.h>
#include "./backward_references.h"
#include "./histogram.h"
#include "./backward_references_enc.h"
#include "./histogram_enc.h"
#include "../dsp/lossless.h"
#include "../dsp/lossless_common.h"
#include "../dsp/dsp.h"
#include "../utils/color_cache.h"
#include "../utils/color_cache_utils.h"
#include "../utils/utils.h"
#define VALUES_IN_BYTE 256
@ -30,8 +31,9 @@
#define WINDOW_SIZE_BITS 20
#define WINDOW_SIZE ((1 << WINDOW_SIZE_BITS) - 120)
// Bounds for the match length.
#define MIN_LENGTH 2
// Minimum number of pixels for which it is cheaper to encode a
// distance + length instead of each pixel as a literal.
#define MIN_LENGTH 4
// If you change this, you need MAX_LENGTH_BITS + WINDOW_SIZE_BITS <= 32 as it
// is used in VP8LHashChain.
#define MAX_LENGTH_BITS 12
@ -211,13 +213,13 @@ void VP8LHashChainClear(VP8LHashChain* const p) {
// -----------------------------------------------------------------------------
#define HASH_MULTIPLIER_HI (0xc6a4a793U)
#define HASH_MULTIPLIER_LO (0x5bd1e996U)
#define HASH_MULTIPLIER_HI (0xc6a4a793ULL)
#define HASH_MULTIPLIER_LO (0x5bd1e996ULL)
static WEBP_INLINE uint32_t GetPixPairHash64(const uint32_t* const argb) {
uint32_t key;
key = argb[1] * HASH_MULTIPLIER_HI;
key += argb[0] * HASH_MULTIPLIER_LO;
key = (argb[1] * HASH_MULTIPLIER_HI) & 0xffffffffu;
key += (argb[0] * HASH_MULTIPLIER_LO) & 0xffffffffu;
key = key >> (32 - HASH_BITS);
return key;
}
@ -242,19 +244,26 @@ static WEBP_INLINE int MaxFindCopyLength(int len) {
}
int VP8LHashChainFill(VP8LHashChain* const p, int quality,
const uint32_t* const argb, int xsize, int ysize) {
const uint32_t* const argb, int xsize, int ysize,
int low_effort) {
const int size = xsize * ysize;
const int iter_max = GetMaxItersForQuality(quality);
const int iter_min = iter_max - quality / 10;
const uint32_t window_size = GetWindowSizeForHashChain(quality, xsize);
int pos;
int argb_comp;
uint32_t base_position;
int32_t* hash_to_first_index;
// Temporarily use the p->offset_length_ as a hash chain.
int32_t* chain = (int32_t*)p->offset_length_;
assert(size > 0);
assert(p->size_ != 0);
assert(p->offset_length_ != NULL);
if (size <= 2) {
p->offset_length_[0] = p->offset_length_[size - 1] = 0;
return 1;
}
hash_to_first_index =
(int32_t*)WebPSafeMalloc(HASH_SIZE, sizeof(*hash_to_first_index));
if (hash_to_first_index == NULL) return 0;
@ -262,48 +271,111 @@ int VP8LHashChainFill(VP8LHashChain* const p, int quality,
// Set the int32_t array to -1.
memset(hash_to_first_index, 0xff, HASH_SIZE * sizeof(*hash_to_first_index));
// Fill the chain linking pixels with the same hash.
for (pos = 0; pos < size - 1; ++pos) {
const uint32_t hash_code = GetPixPairHash64(argb + pos);
chain[pos] = hash_to_first_index[hash_code];
hash_to_first_index[hash_code] = pos;
argb_comp = (argb[0] == argb[1]);
for (pos = 0; pos < size - 2;) {
uint32_t hash_code;
const int argb_comp_next = (argb[pos + 1] == argb[pos + 2]);
if (argb_comp && argb_comp_next) {
// Consecutive pixels with the same color will share the same hash.
// We therefore use a different hash: the color and its repetition
// length.
uint32_t tmp[2];
uint32_t len = 1;
tmp[0] = argb[pos];
// Figure out how far the pixels are the same.
// The last pixel has a different 64 bit hash, as its next pixel does
// not have the same color, so we just need to get to the last pixel equal
// to its follower.
while (pos + (int)len + 2 < size && argb[pos + len + 2] == argb[pos]) {
++len;
}
if (len > MAX_LENGTH) {
// Skip the pixels that match for distance=1 and length>MAX_LENGTH
// because they are linked to their predecessor and we automatically
// check that in the main for loop below. Skipping means setting no
// predecessor in the chain, hence -1.
memset(chain + pos, 0xff, (len - MAX_LENGTH) * sizeof(*chain));
pos += len - MAX_LENGTH;
len = MAX_LENGTH;
}
// Process the rest of the hash chain.
while (len) {
tmp[1] = len--;
hash_code = GetPixPairHash64(tmp);
chain[pos] = hash_to_first_index[hash_code];
hash_to_first_index[hash_code] = pos++;
}
argb_comp = 0;
} else {
// Just move one pixel forward.
hash_code = GetPixPairHash64(argb + pos);
chain[pos] = hash_to_first_index[hash_code];
hash_to_first_index[hash_code] = pos++;
argb_comp = argb_comp_next;
}
}
// Process the penultimate pixel.
chain[pos] = hash_to_first_index[GetPixPairHash64(argb + pos)];
WebPSafeFree(hash_to_first_index);
// Find the best match interval at each pixel, defined by an offset to the
// pixel and a length. The right-most pixel cannot match anything to the right
// (hence a best length of 0) and the left-most pixel nothing to the left
// (hence an offset of 0).
assert(size > 2);
p->offset_length_[0] = p->offset_length_[size - 1] = 0;
for (base_position = size - 2 < 0 ? 0 : size - 2; base_position > 0;) {
for (base_position = size - 2; base_position > 0;) {
const int max_len = MaxFindCopyLength(size - 1 - base_position);
const uint32_t* const argb_start = argb + base_position;
int iter = iter_max;
int best_length = 0;
uint32_t best_distance = 0;
uint32_t best_argb;
const int min_pos =
(base_position > window_size) ? base_position - window_size : 0;
const int length_max = (max_len < 256) ? max_len : 256;
uint32_t max_base_position;
for (pos = chain[base_position]; pos >= min_pos; pos = chain[pos]) {
pos = chain[base_position];
if (!low_effort) {
int curr_length;
if (--iter < 0) {
break;
// Heuristic: use the comparison with the above line as an initialization.
if (base_position >= (uint32_t)xsize) {
curr_length = FindMatchLength(argb_start - xsize, argb_start,
best_length, max_len);
if (curr_length > best_length) {
best_length = curr_length;
best_distance = xsize;
}
--iter;
}
// Heuristic: compare to the previous pixel.
curr_length =
FindMatchLength(argb_start - 1, argb_start, best_length, max_len);
if (curr_length > best_length) {
best_length = curr_length;
best_distance = 1;
}
--iter;
// Skip the for loop if we already have the maximum.
if (best_length == MAX_LENGTH) pos = min_pos - 1;
}
best_argb = argb_start[best_length];
for (; pos >= min_pos && --iter; pos = chain[pos]) {
int curr_length;
assert(base_position > (uint32_t)pos);
curr_length =
FindMatchLength(argb + pos, argb_start, best_length, max_len);
if (argb[pos + best_length] != best_argb) continue;
curr_length = VP8LVectorMismatch(argb + pos, argb_start, max_len);
if (best_length < curr_length) {
best_length = curr_length;
best_distance = base_position - pos;
// Stop if we have reached the maximum length. Otherwise, make sure
// we have executed a minimum number of iterations depending on the
// quality.
if ((best_length == MAX_LENGTH) ||
(curr_length >= length_max && iter < iter_min)) {
break;
}
best_argb = argb_start[best_length];
// Stop if we have reached a good enough length.
if (best_length >= length_max) break;
}
}
// We have the best match but in case the two intervals continue matching
@ -392,17 +464,16 @@ static int BackwardReferencesRle(int xsize, int ysize,
i = 1;
while (i < pix_count) {
const int max_len = MaxFindCopyLength(pix_count - i);
const int kMinLength = 4;
const int rle_len = FindMatchLength(argb + i, argb + i - 1, 0, max_len);
const int prev_row_len = (i < xsize) ? 0 :
FindMatchLength(argb + i, argb + i - xsize, 0, max_len);
if (rle_len >= prev_row_len && rle_len >= kMinLength) {
if (rle_len >= prev_row_len && rle_len >= MIN_LENGTH) {
BackwardRefsCursorAdd(refs, PixOrCopyCreateCopy(1, rle_len));
// We don't need to update the color cache here since it is always the
// same pixel being copied, and that does not change the color cache
// state.
i += rle_len;
} else if (prev_row_len >= kMinLength) {
} else if (prev_row_len >= MIN_LENGTH) {
BackwardRefsCursorAdd(refs, PixOrCopyCreateCopy(xsize, prev_row_len));
if (use_color_cache) {
for (k = 0; k < prev_row_len; ++k) {
@ -442,7 +513,7 @@ static int BackwardReferencesLz77(int xsize, int ysize,
int len = 0;
int j;
HashChainFindCopy(hash_chain, i, &offset, &len);
if (len > MIN_LENGTH + 1) {
if (len >= MIN_LENGTH) {
const int len_ini = len;
int max_reach = 0;
assert(i + len < pix_count);
@ -457,7 +528,7 @@ static int BackwardReferencesLz77(int xsize, int ysize,
for (j = i_last_check + 1; j <= i + len_ini; ++j) {
const int len_j = HashChainFindLength(hash_chain, j);
const int reach =
j + (len_j > MIN_LENGTH + 1 ? len_j : 1); // 1 for single literal.
j + (len_j >= MIN_LENGTH ? len_j : 1); // 1 for single literal.
if (reach > max_reach) {
len = j - i;
max_reach = reach;
@ -581,9 +652,10 @@ static void AddSingleLiteralWithCostModel(const uint32_t* const argb,
uint16_t* const dist_array) {
double cost_val = prev_cost;
const uint32_t color = argb[0];
if (use_color_cache && VP8LColorCacheContains(hashers, color)) {
const int ix = use_color_cache ? VP8LColorCacheContains(hashers, color) : -1;
if (ix >= 0) {
// use_color_cache is true and hashers contains color
const double mul0 = 0.68;
const int ix = VP8LColorCacheGetIndex(hashers, color);
cost_val += GetCacheCost(cost_model, ix) * mul0;
} else {
const double mul1 = 0.82;
@ -1215,7 +1287,8 @@ static int BackwardReferencesHashChainDistanceOnly(
int offset = 0, len = 0;
double prev_cost = cost_manager->costs_[i - 1];
HashChainFindCopy(hash_chain, i, &offset, &len);
if (len >= MIN_LENGTH) {
if (len >= 2) {
// If we are dealing with a non-literal.
const int code = DistanceToPlaneCode(xsize, offset);
const double offset_cost = GetDistanceCost(cost_model, code);
const int first_i = i;
@ -1304,20 +1377,17 @@ static int BackwardReferencesHashChainDistanceOnly(
}
goto next_symbol;
}
if (len > MIN_LENGTH) {
int code_min_length;
double cost_total;
offset = HashChainFindOffset(hash_chain, i);
code_min_length = DistanceToPlaneCode(xsize, offset);
cost_total = prev_cost +
GetDistanceCost(cost_model, code_min_length) +
GetLengthCost(cost_model, 1);
if (len > 2) {
// Also try the smallest interval possible (size 2).
double cost_total =
prev_cost + offset_cost + GetLengthCost(cost_model, 1);
if (cost_manager->costs_[i + 1] > cost_total) {
cost_manager->costs_[i + 1] = (float)cost_total;
dist_array[i + 1] = 2;
}
}
} else { // len < MIN_LENGTH
} else {
// The pixel is added as a single literal so just update the costs.
UpdateCostPerIndex(cost_manager, i + 1);
}
@ -1393,9 +1463,11 @@ static int BackwardReferencesHashChainFollowChosenPath(
i += len;
} else {
PixOrCopy v;
if (use_color_cache && VP8LColorCacheContains(&hashers, argb[i])) {
const int idx =
use_color_cache ? VP8LColorCacheContains(&hashers, argb[i]) : -1;
if (idx >= 0) {
// use_color_cache is true and hashers contains argb[i]
// push pixel as a color cache index
const int idx = VP8LColorCacheGetIndex(&hashers, argb[i]);
v = PixOrCopyCreateCacheIdx(idx);
} else {
if (use_color_cache) VP8LColorCacheInsert(&hashers, argb[i]);
@ -1454,63 +1526,89 @@ static void BackwardReferences2DLocality(int xsize,
}
}
// Returns entropy for the given cache bits.
static double ComputeCacheEntropy(const uint32_t* argb,
const VP8LBackwardRefs* const refs,
int cache_bits) {
const int use_color_cache = (cache_bits > 0);
int cc_init = 0;
double entropy = MAX_ENTROPY;
const double kSmallPenaltyForLargeCache = 4.0;
VP8LColorCache hashers;
// Computes the entropies for a color cache size (in bits) between 0 (unused)
// and cache_bits_max (inclusive).
// Returns 1 on success, 0 in case of allocation error.
static int ComputeCacheEntropies(const uint32_t* argb,
const VP8LBackwardRefs* const refs,
int cache_bits_max, double entropies[]) {
int cc_init[MAX_COLOR_CACHE_BITS + 1] = { 0 };
VP8LColorCache hashers[MAX_COLOR_CACHE_BITS + 1];
VP8LRefsCursor c = VP8LRefsCursorInit(refs);
VP8LHistogram* histo = VP8LAllocateHistogram(cache_bits);
if (histo == NULL) goto Error;
VP8LHistogram* histos[MAX_COLOR_CACHE_BITS + 1] = { NULL };
int ok = 0;
int i;
if (use_color_cache) {
cc_init = VP8LColorCacheInit(&hashers, cache_bits);
if (!cc_init) goto Error;
for (i = 0; i <= cache_bits_max; ++i) {
histos[i] = VP8LAllocateHistogram(i);
if (histos[i] == NULL) goto Error;
if (i == 0) continue;
cc_init[i] = VP8LColorCacheInit(&hashers[i], i);
if (!cc_init[i]) goto Error;
}
if (!use_color_cache) {
while (VP8LRefsCursorOk(&c)) {
VP8LHistogramAddSinglePixOrCopy(histo, c.cur_pos);
VP8LRefsCursorNext(&c);
}
} else {
assert(cache_bits_max >= 0);
// Do not use the color cache for cache_bits=0.
while (VP8LRefsCursorOk(&c)) {
VP8LHistogramAddSinglePixOrCopy(histos[0], c.cur_pos);
VP8LRefsCursorNext(&c);
}
if (cache_bits_max > 0) {
c = VP8LRefsCursorInit(refs);
while (VP8LRefsCursorOk(&c)) {
const PixOrCopy* const v = c.cur_pos;
if (PixOrCopyIsLiteral(v)) {
const uint32_t pix = *argb++;
const uint32_t key = VP8LColorCacheGetIndex(&hashers, pix);
if (VP8LColorCacheLookup(&hashers, key) == pix) {
++histo->literal_[NUM_LITERAL_CODES + NUM_LENGTH_CODES + key];
} else {
VP8LColorCacheSet(&hashers, key, pix);
++histo->blue_[pix & 0xff];
++histo->literal_[(pix >> 8) & 0xff];
++histo->red_[(pix >> 16) & 0xff];
++histo->alpha_[pix >> 24];
// The keys of the caches can be derived from the longest one.
int key = HashPix(pix, 32 - cache_bits_max);
for (i = cache_bits_max; i >= 1; --i, key >>= 1) {
if (VP8LColorCacheLookup(&hashers[i], key) == pix) {
++histos[i]->literal_[NUM_LITERAL_CODES + NUM_LENGTH_CODES + key];
} else {
VP8LColorCacheSet(&hashers[i], key, pix);
++histos[i]->blue_[pix & 0xff];
++histos[i]->literal_[(pix >> 8) & 0xff];
++histos[i]->red_[(pix >> 16) & 0xff];
++histos[i]->alpha_[pix >> 24];
}
}
} else {
// Update the histograms for distance/length.
int len = PixOrCopyLength(v);
int code, extra_bits;
VP8LPrefixEncodeBits(len, &code, &extra_bits);
++histo->literal_[NUM_LITERAL_CODES + code];
VP8LPrefixEncodeBits(PixOrCopyDistance(v), &code, &extra_bits);
++histo->distance_[code];
int code_dist, code_len, extra_bits;
uint32_t argb_prev = *argb ^ 0xffffffffu;
VP8LPrefixEncodeBits(len, &code_len, &extra_bits);
VP8LPrefixEncodeBits(PixOrCopyDistance(v), &code_dist, &extra_bits);
for (i = 1; i <= cache_bits_max; ++i) {
++histos[i]->literal_[NUM_LITERAL_CODES + code_len];
++histos[i]->distance_[code_dist];
}
// Update the colors caches.
do {
VP8LColorCacheInsert(&hashers, *argb++);
} while(--len != 0);
if (*argb != argb_prev) {
// Efficiency: insert only if the color changes.
int key = HashPix(*argb, 32 - cache_bits_max);
for (i = cache_bits_max; i >= 1; --i, key >>= 1) {
hashers[i].colors_[key] = *argb;
}
argb_prev = *argb;
}
argb++;
} while (--len != 0);
}
VP8LRefsCursorNext(&c);
}
}
entropy = VP8LHistogramEstimateBits(histo) +
kSmallPenaltyForLargeCache * cache_bits;
Error:
if (cc_init) VP8LColorCacheClear(&hashers);
VP8LFreeHistogram(histo);
return entropy;
for (i = 0; i <= cache_bits_max; ++i) {
entropies[i] = VP8LHistogramEstimateBits(histos[i]);
}
ok = 1;
Error:
for (i = 0; i <= cache_bits_max; ++i) {
if (cc_init[i]) VP8LColorCacheClear(&hashers[i]);
VP8LFreeHistogram(histos[i]);
}
return ok;
}
// Evaluate optimal cache bits for the local color cache.
@ -1524,13 +1622,10 @@ static int CalculateBestCacheSize(const uint32_t* const argb,
VP8LBackwardRefs* const refs,
int* const lz77_computed,
int* const best_cache_bits) {
int eval_low = 1;
int eval_high = 1;
double entropy_low = MAX_ENTROPY;
double entropy_high = MAX_ENTROPY;
const double cost_mul = 5e-4;
int cache_bits_low = 0;
int i;
int cache_bits_high = (quality <= 25) ? 0 : *best_cache_bits;
double entropy_min = MAX_ENTROPY;
double entropies[MAX_COLOR_CACHE_BITS + 1];
assert(cache_bits_high <= MAX_COLOR_CACHE_BITS);
@ -1540,34 +1635,23 @@ static int CalculateBestCacheSize(const uint32_t* const argb,
// Local color cache is disabled.
return 1;
}
if (!BackwardReferencesLz77(xsize, ysize, argb, cache_bits_low, hash_chain,
refs)) {
// Compute LZ77 with no cache (0 bits), as the ideal LZ77 with a color cache
// is not that different in practice.
if (!BackwardReferencesLz77(xsize, ysize, argb, 0, hash_chain, refs)) {
return 0;
}
// Do a binary search to find the optimal entropy for cache_bits.
while (eval_low || eval_high) {
if (eval_low) {
entropy_low = ComputeCacheEntropy(argb, refs, cache_bits_low);
entropy_low += entropy_low * cache_bits_low * cost_mul;
eval_low = 0;
}
if (eval_high) {
entropy_high = ComputeCacheEntropy(argb, refs, cache_bits_high);
entropy_high += entropy_high * cache_bits_high * cost_mul;
eval_high = 0;
}
if (entropy_high < entropy_low) {
const int prev_cache_bits_low = cache_bits_low;
*best_cache_bits = cache_bits_high;
cache_bits_low = (cache_bits_low + cache_bits_high) / 2;
if (cache_bits_low != prev_cache_bits_low) eval_low = 1;
} else {
*best_cache_bits = cache_bits_low;
cache_bits_high = (cache_bits_low + cache_bits_high) / 2;
if (cache_bits_high != cache_bits_low) eval_high = 1;
// Find the cache_bits giving the lowest entropy. The search is done in a
// brute-force way as the function (entropy w.r.t cache_bits) can be
// anything in practice.
if (!ComputeCacheEntropies(argb, refs, cache_bits_high, entropies)) {
return 0;
}
for (i = 0; i <= cache_bits_high; ++i) {
if (i == 0 || entropies[i] < entropy_min) {
entropy_min = entropies[i];
*best_cache_bits = i;
}
}
*lz77_computed = 1;
return 1;
}
@ -1584,8 +1668,9 @@ static int BackwardRefsWithLocalCache(const uint32_t* const argb,
PixOrCopy* const v = c.cur_pos;
if (PixOrCopyIsLiteral(v)) {
const uint32_t argb_literal = v->argb_or_distance;
if (VP8LColorCacheContains(&hashers, argb_literal)) {
const int ix = VP8LColorCacheGetIndex(&hashers, argb_literal);
const int ix = VP8LColorCacheContains(&hashers, argb_literal);
if (ix >= 0) {
// hashers contains argb_literal
*v = PixOrCopyCreateCacheIdx(ix);
} else {
VP8LColorCacheInsert(&hashers, argb_literal);

3
thirdparty/libwebp/enc/backward_references.h → thirdparty/libwebp/enc/backward_references_enc.h vendored
View file

@ -130,7 +130,8 @@ struct VP8LHashChain {
int VP8LHashChainInit(VP8LHashChain* const p, int size);
// Pre-compute the best matches for argb.
int VP8LHashChainFill(VP8LHashChain* const p, int quality,
const uint32_t* const argb, int xsize, int ysize);
const uint32_t* const argb, int xsize, int ysize,
int low_effort);
void VP8LHashChainClear(VP8LHashChain* const p); // release memory
// -----------------------------------------------------------------------------

91
thirdparty/libwebp/enc/config.c → thirdparty/libwebp/enc/config_enc.c vendored
View file

@ -11,6 +11,10 @@
//
// Author: Skal (pascal.massimino@gmail.com)
#ifdef HAVE_CONFIG_H
#include "../webp/config.h"
#endif
#include "../webp/encode.h"
//------------------------------------------------------------------------------
@ -49,9 +53,8 @@ int WebPConfigInitInternal(WebPConfig* config,
config->thread_level = 0;
config->low_memory = 0;
config->near_lossless = 100;
#ifdef WEBP_EXPERIMENTAL_FEATURES
config->delta_palettization = 0;
#endif // WEBP_EXPERIMENTAL_FEATURES
config->use_delta_palette = 0;
config->use_sharp_yuv = 0;
// TODO(skal): tune.
switch (preset) {
@ -92,60 +95,36 @@ int WebPConfigInitInternal(WebPConfig* config,
int WebPValidateConfig(const WebPConfig* config) {
if (config == NULL) return 0;
if (config->quality < 0 || config->quality > 100)
if (config->quality < 0 || config->quality > 100) return 0;
if (config->target_size < 0) return 0;
if (config->target_PSNR < 0) return 0;
if (config->method < 0 || config->method > 6) return 0;
if (config->segments < 1 || config->segments > 4) return 0;
if (config->sns_strength < 0 || config->sns_strength > 100) return 0;
if (config->filter_strength < 0 || config->filter_strength > 100) return 0;
if (config->filter_sharpness < 0 || config->filter_sharpness > 7) return 0;
if (config->filter_type < 0 || config->filter_type > 1) return 0;
if (config->autofilter < 0 || config->autofilter > 1) return 0;
if (config->pass < 1 || config->pass > 10) return 0;
if (config->show_compressed < 0 || config->show_compressed > 1) return 0;
if (config->preprocessing < 0 || config->preprocessing > 7) return 0;
if (config->partitions < 0 || config->partitions > 3) return 0;
if (config->partition_limit < 0 || config->partition_limit > 100) return 0;
if (config->alpha_compression < 0) return 0;
if (config->alpha_filtering < 0) return 0;
if (config->alpha_quality < 0 || config->alpha_quality > 100) return 0;
if (config->lossless < 0 || config->lossless > 1) return 0;
if (config->near_lossless < 0 || config->near_lossless > 100) return 0;
if (config->image_hint >= WEBP_HINT_LAST) return 0;
if (config->emulate_jpeg_size < 0 || config->emulate_jpeg_size > 1) return 0;
if (config->thread_level < 0 || config->thread_level > 1) return 0;
if (config->low_memory < 0 || config->low_memory > 1) return 0;
if (config->exact < 0 || config->exact > 1) return 0;
if (config->use_delta_palette < 0 || config->use_delta_palette > 1) {
return 0;
if (config->target_size < 0)
return 0;
if (config->target_PSNR < 0)
return 0;
if (config->method < 0 || config->method > 6)
return 0;
if (config->segments < 1 || config->segments > 4)
return 0;
if (config->sns_strength < 0 || config->sns_strength > 100)
return 0;
if (config->filter_strength < 0 || config->filter_strength > 100)
return 0;
if (config->filter_sharpness < 0 || config->filter_sharpness > 7)
return 0;
if (config->filter_type < 0 || config->filter_type > 1)
return 0;
if (config->autofilter < 0 || config->autofilter > 1)
return 0;
if (config->pass < 1 || config->pass > 10)
return 0;
if (config->show_compressed < 0 || config->show_compressed > 1)
return 0;
if (config->preprocessing < 0 || config->preprocessing > 7)
return 0;
if (config->partitions < 0 || config->partitions > 3)
return 0;
if (config->partition_limit < 0 || config->partition_limit > 100)
return 0;
if (config->alpha_compression < 0)
return 0;
if (config->alpha_filtering < 0)
return 0;
if (config->alpha_quality < 0 || config->alpha_quality > 100)
return 0;
if (config->lossless < 0 || config->lossless > 1)
return 0;
if (config->near_lossless < 0 || config->near_lossless > 100)
return 0;
if (config->image_hint >= WEBP_HINT_LAST)
return 0;
if (config->emulate_jpeg_size < 0 || config->emulate_jpeg_size > 1)
return 0;
if (config->thread_level < 0 || config->thread_level > 1)
return 0;
if (config->low_memory < 0 || config->low_memory > 1)
return 0;
if (config->exact < 0 || config->exact > 1)
return 0;
#ifdef WEBP_EXPERIMENTAL_FEATURES
if (config->delta_palettization < 0 || config->delta_palettization > 1)
return 0;
#endif // WEBP_EXPERIMENTAL_FEATURES
}
if (config->use_sharp_yuv < 0 || config->use_sharp_yuv > 1) return 0;
return 1;
}

2
thirdparty/libwebp/enc/cost.c → thirdparty/libwebp/enc/cost_enc.c vendored
View file

@ -11,7 +11,7 @@
//
// Author: Skal (pascal.massimino@gmail.com)
#include "./cost.h"
#include "./cost_enc.h"
//------------------------------------------------------------------------------
// Level cost tables

2
thirdparty/libwebp/enc/cost.h → thirdparty/libwebp/enc/cost_enc.h vendored
View file

@ -16,7 +16,7 @@
#include <assert.h>
#include <stdlib.h>
#include "./vp8enci.h"
#include "./vp8i_enc.h"
#ifdef __cplusplus
extern "C" {

2
thirdparty/libwebp/enc/delta_palettization.c → thirdparty/libwebp/enc/delta_palettization_enc.c vendored
View file

@ -10,7 +10,7 @@
// Author: Mislav Bradac (mislavm@google.com)
//
#include "./delta_palettization.h"
#include "./delta_palettization_enc.h"
#ifdef WEBP_EXPERIMENTAL_FEATURES
#include "../webp/types.h"

2
thirdparty/libwebp/enc/delta_palettization.h → thirdparty/libwebp/enc/delta_palettization_enc.h vendored
View file

@ -14,7 +14,7 @@
#define WEBP_ENC_DELTA_PALETTIZATION_H_
#include "../webp/encode.h"
#include "../enc/vp8li.h"
#include "../enc/vp8li_enc.h"
// Replaces enc->argb_[] input by a palettizable approximation of it,
// and generates optimal enc->palette_[].

107
thirdparty/libwebp/enc/filter.c → thirdparty/libwebp/enc/filter_enc.c vendored
View file

@ -12,7 +12,7 @@
// Author: somnath@google.com (Somnath Banerjee)
#include <assert.h>
#include "./vp8enci.h"
#include "./vp8i_enc.h"
#include "../dsp/dsp.h"
// This table gives, for a given sharpness, the filtering strength to be
@ -105,115 +105,28 @@ static void DoFilter(const VP8EncIterator* const it, int level) {
}
//------------------------------------------------------------------------------
// SSIM metric
static const double kMinValue = 1.e-10; // minimal threshold
void VP8SSIMAddStats(const VP8DistoStats* const src, VP8DistoStats* const dst) {
dst->w += src->w;
dst->xm += src->xm;
dst->ym += src->ym;
dst->xxm += src->xxm;
dst->xym += src->xym;
dst->yym += src->yym;
}
double VP8SSIMGet(const VP8DistoStats* const stats) {
const double xmxm = stats->xm * stats->xm;
const double ymym = stats->ym * stats->ym;
const double xmym = stats->xm * stats->ym;
const double w2 = stats->w * stats->w;
double sxx = stats->xxm * stats->w - xmxm;
double syy = stats->yym * stats->w - ymym;
double sxy = stats->xym * stats->w - xmym;
double C1, C2;
double fnum;
double fden;
// small errors are possible, due to rounding. Clamp to zero.
if (sxx < 0.) sxx = 0.;
if (syy < 0.) syy = 0.;
C1 = 6.5025 * w2;
C2 = 58.5225 * w2;
fnum = (2 * xmym + C1) * (2 * sxy + C2);
fden = (xmxm + ymym + C1) * (sxx + syy + C2);
return (fden != 0.) ? fnum / fden : kMinValue;
}
double VP8SSIMGetSquaredError(const VP8DistoStats* const s) {
if (s->w > 0.) {
const double iw2 = 1. / (s->w * s->w);
const double sxx = s->xxm * s->w - s->xm * s->xm;
const double syy = s->yym * s->w - s->ym * s->ym;
const double sxy = s->xym * s->w - s->xm * s->ym;
const double SSE = iw2 * (sxx + syy - 2. * sxy);
if (SSE > kMinValue) return SSE;
}
return kMinValue;
}
#define LIMIT(A, M) ((A) > (M) ? (M) : (A))
static void VP8SSIMAccumulateRow(const uint8_t* src1, int stride1,
const uint8_t* src2, int stride2,
int y, int W, int H,
VP8DistoStats* const stats) {
int x = 0;
const int w0 = LIMIT(VP8_SSIM_KERNEL, W);
for (x = 0; x < w0; ++x) {
VP8SSIMAccumulateClipped(src1, stride1, src2, stride2, x, y, W, H, stats);
}
for (; x <= W - 8 + VP8_SSIM_KERNEL; ++x) {
VP8SSIMAccumulate(
src1 + (y - VP8_SSIM_KERNEL) * stride1 + (x - VP8_SSIM_KERNEL), stride1,
src2 + (y - VP8_SSIM_KERNEL) * stride2 + (x - VP8_SSIM_KERNEL), stride2,
stats);
}
for (; x < W; ++x) {
VP8SSIMAccumulateClipped(src1, stride1, src2, stride2, x, y, W, H, stats);
}
}
void VP8SSIMAccumulatePlane(const uint8_t* src1, int stride1,
const uint8_t* src2, int stride2,
int W, int H, VP8DistoStats* const stats) {
int x, y;
const int h0 = LIMIT(VP8_SSIM_KERNEL, H);
const int h1 = LIMIT(VP8_SSIM_KERNEL, H - VP8_SSIM_KERNEL);
for (y = 0; y < h0; ++y) {
for (x = 0; x < W; ++x) {
VP8SSIMAccumulateClipped(src1, stride1, src2, stride2, x, y, W, H, stats);
}
}
for (; y < h1; ++y) {
VP8SSIMAccumulateRow(src1, stride1, src2, stride2, y, W, H, stats);
}
for (; y < H; ++y) {
for (x = 0; x < W; ++x) {
VP8SSIMAccumulateClipped(src1, stride1, src2, stride2, x, y, W, H, stats);
}
}
}
#undef LIMIT
// SSIM metric for one macroblock
static double GetMBSSIM(const uint8_t* yuv1, const uint8_t* yuv2) {
int x, y;
VP8DistoStats s = { .0, .0, .0, .0, .0, .0 };
double sum = 0.;
// compute SSIM in a 10 x 10 window
for (y = VP8_SSIM_KERNEL; y < 16 - VP8_SSIM_KERNEL; y++) {
for (x = VP8_SSIM_KERNEL; x < 16 - VP8_SSIM_KERNEL; x++) {
VP8SSIMAccumulateClipped(yuv1 + Y_OFF_ENC, BPS, yuv2 + Y_OFF_ENC, BPS,
x, y, 16, 16, &s);
sum += VP8SSIMGetClipped(yuv1 + Y_OFF_ENC, BPS, yuv2 + Y_OFF_ENC, BPS,
x, y, 16, 16);
}
}
for (x = 1; x < 7; x++) {
for (y = 1; y < 7; y++) {
VP8SSIMAccumulateClipped(yuv1 + U_OFF_ENC, BPS, yuv2 + U_OFF_ENC, BPS,
x, y, 8, 8, &s);
VP8SSIMAccumulateClipped(yuv1 + V_OFF_ENC, BPS, yuv2 + V_OFF_ENC, BPS,
x, y, 8, 8, &s);
sum += VP8SSIMGetClipped(yuv1 + U_OFF_ENC, BPS, yuv2 + U_OFF_ENC, BPS,
x, y, 8, 8);
sum += VP8SSIMGetClipped(yuv1 + V_OFF_ENC, BPS, yuv2 + V_OFF_ENC, BPS,
x, y, 8, 8);
}
}
return VP8SSIMGet(&s);
return sum;
}
//------------------------------------------------------------------------------

10
thirdparty/libwebp/enc/frame.c → thirdparty/libwebp/enc/frame_enc.c vendored
View file

@ -14,8 +14,8 @@
#include <string.h>
#include <math.h>
#include "./cost.h"
#include "./vp8enci.h"
#include "./cost_enc.h"
#include "./vp8i_enc.h"
#include "../dsp/dsp.h"
#include "../webp/format_constants.h" // RIFF constants
@ -248,8 +248,9 @@ static int PutCoeffs(VP8BitWriter* const bw, int ctx, const VP8Residual* res) {
p = res->prob[VP8EncBands[n]][1];
} else {
if (!VP8PutBit(bw, v > 4, p[3])) {
if (VP8PutBit(bw, v != 2, p[4]))
if (VP8PutBit(bw, v != 2, p[4])) {
VP8PutBit(bw, v == 4, p[5]);
}
} else if (!VP8PutBit(bw, v > 10, p[6])) {
if (!VP8PutBit(bw, v > 6, p[7])) {
VP8PutBit(bw, v == 6, 159);
@ -557,8 +558,9 @@ static uint64_t OneStatPass(VP8Encoder* const enc, VP8RDLevel rd_opt,
size += info.R + info.H;
size_p0 += info.H;
distortion += info.D;
if (percent_delta && !VP8IteratorProgress(&it, percent_delta))
if (percent_delta && !VP8IteratorProgress(&it, percent_delta)) {
return 0;
}
VP8IteratorSaveBoundary(&it);
} while (VP8IteratorNext(&it) && --nb_mbs > 0);

302
thirdparty/libwebp/enc/histogram.c → thirdparty/libwebp/enc/histogram_enc.c vendored
View file

@ -15,9 +15,10 @@
#include <math.h>
#include "./backward_references.h"
#include "./histogram.h"
#include "./backward_references_enc.h"
#include "./histogram_enc.h"
#include "../dsp/lossless.h"
#include "../dsp/lossless_common.h"
#include "../utils/utils.h"
#define MAX_COST 1.e38
@ -213,10 +214,19 @@ static double InitialHuffmanCost(void) {
// Finalize the Huffman cost based on streak numbers and length type (<3 or >=3)
static double FinalHuffmanCost(const VP8LStreaks* const stats) {
// The constants in this function are experimental and got rounded from
// their original values in 1/8 when switched to 1/1024.
double retval = InitialHuffmanCost();
// Second coefficient: Many zeros in the histogram are covered efficiently
// by a run-length encode. Originally 2/8.
retval += stats->counts[0] * 1.5625 + 0.234375 * stats->streaks[0][1];
// Second coefficient: Constant values are encoded less efficiently, but still
// RLE'ed. Originally 6/8.
retval += stats->counts[1] * 2.578125 + 0.703125 * stats->streaks[1][1];
// 0s are usually encoded more efficiently than non-0s.
// Originally 15/8.
retval += 1.796875 * stats->streaks[0][0];
// Originally 26/8.
retval += 3.28125 * stats->streaks[1][0];
return retval;
}
@ -236,14 +246,30 @@ static double PopulationCost(const uint32_t* const population, int length,
return BitsEntropyRefine(&bit_entropy) + FinalHuffmanCost(&stats);
}
// trivial_at_end is 1 if the two histograms only have one element that is
// non-zero: both the zero-th one, or both the last one.
static WEBP_INLINE double GetCombinedEntropy(const uint32_t* const X,
const uint32_t* const Y,
int length) {
VP8LBitEntropy bit_entropy;
int length, int trivial_at_end) {
VP8LStreaks stats;
VP8LGetCombinedEntropyUnrefined(X, Y, length, &bit_entropy, &stats);
if (trivial_at_end) {
// This configuration is due to palettization that transforms an indexed
// pixel into 0xff000000 | (pixel << 8) in VP8LBundleColorMap.
// BitsEntropyRefine is 0 for histograms with only one non-zero value.
// Only FinalHuffmanCost needs to be evaluated.
memset(&stats, 0, sizeof(stats));
// Deal with the non-zero value at index 0 or length-1.
stats.streaks[1][0] += 1;
// Deal with the following/previous zero streak.
stats.counts[0] += 1;
stats.streaks[0][1] += length - 1;
return FinalHuffmanCost(&stats);
} else {
VP8LBitEntropy bit_entropy;
VP8LGetCombinedEntropyUnrefined(X, Y, length, &bit_entropy, &stats);
return BitsEntropyRefine(&bit_entropy) + FinalHuffmanCost(&stats);
return BitsEntropyRefine(&bit_entropy) + FinalHuffmanCost(&stats);
}
}
// Estimates the Entropy + Huffman + other block overhead size cost.
@ -267,24 +293,42 @@ static int GetCombinedHistogramEntropy(const VP8LHistogram* const a,
double cost_threshold,
double* cost) {
const int palette_code_bits = a->palette_code_bits_;
int trivial_at_end = 0;
assert(a->palette_code_bits_ == b->palette_code_bits_);
*cost += GetCombinedEntropy(a->literal_, b->literal_,
VP8LHistogramNumCodes(palette_code_bits));
VP8LHistogramNumCodes(palette_code_bits), 0);
*cost += VP8LExtraCostCombined(a->literal_ + NUM_LITERAL_CODES,
b->literal_ + NUM_LITERAL_CODES,
NUM_LENGTH_CODES);
if (*cost > cost_threshold) return 0;
*cost += GetCombinedEntropy(a->red_, b->red_, NUM_LITERAL_CODES);
if (a->trivial_symbol_ != VP8L_NON_TRIVIAL_SYM &&
a->trivial_symbol_ == b->trivial_symbol_) {
// A, R and B are all 0 or 0xff.
const uint32_t color_a = (a->trivial_symbol_ >> 24) & 0xff;
const uint32_t color_r = (a->trivial_symbol_ >> 16) & 0xff;
const uint32_t color_b = (a->trivial_symbol_ >> 0) & 0xff;
if ((color_a == 0 || color_a == 0xff) &&
(color_r == 0 || color_r == 0xff) &&
(color_b == 0 || color_b == 0xff)) {
trivial_at_end = 1;
}
}
*cost +=
GetCombinedEntropy(a->red_, b->red_, NUM_LITERAL_CODES, trivial_at_end);
if (*cost > cost_threshold) return 0;
*cost += GetCombinedEntropy(a->blue_, b->blue_, NUM_LITERAL_CODES);
*cost +=
GetCombinedEntropy(a->blue_, b->blue_, NUM_LITERAL_CODES, trivial_at_end);
if (*cost > cost_threshold) return 0;
*cost += GetCombinedEntropy(a->alpha_, b->alpha_, NUM_LITERAL_CODES);
*cost += GetCombinedEntropy(a->alpha_, b->alpha_, NUM_LITERAL_CODES,
trivial_at_end);
if (*cost > cost_threshold) return 0;
*cost += GetCombinedEntropy(a->distance_, b->distance_, NUM_DISTANCE_CODES);
*cost +=
GetCombinedEntropy(a->distance_, b->distance_, NUM_DISTANCE_CODES, 0);
*cost +=
VP8LExtraCostCombined(a->distance_, b->distance_, NUM_DISTANCE_CODES);
if (*cost > cost_threshold) return 0;
@ -292,6 +336,15 @@ static int GetCombinedHistogramEntropy(const VP8LHistogram* const a,
return 1;
}
static WEBP_INLINE void HistogramAdd(const VP8LHistogram* const a,
const VP8LHistogram* const b,
VP8LHistogram* const out) {
VP8LHistogramAdd(a, b, out);
out->trivial_symbol_ = (a->trivial_symbol_ == b->trivial_symbol_)
? a->trivial_symbol_
: VP8L_NON_TRIVIAL_SYM;
}
// Performs out = a + b, computing the cost C(a+b) - C(a) - C(b) while comparing
// to the threshold value 'cost_threshold'. The score returned is
// Score = C(a+b) - C(a) - C(b), where C(a) + C(b) is known and fixed.
@ -307,11 +360,9 @@ static double HistogramAddEval(const VP8LHistogram* const a,
cost_threshold += sum_cost;
if (GetCombinedHistogramEntropy(a, b, cost_threshold, &cost)) {
VP8LHistogramAdd(a, b, out);
HistogramAdd(a, b, out);
out->bit_cost_ = cost;
out->palette_code_bits_ = a->palette_code_bits_;
out->trivial_symbol_ = (a->trivial_symbol_ == b->trivial_symbol_) ?
a->trivial_symbol_ : VP8L_NON_TRIVIAL_SYM;
}
return cost - sum_cost;
@ -450,113 +501,103 @@ static void HistogramCopyAndAnalyze(
// Partition histograms to different entropy bins for three dominant (literal,
// red and blue) symbol costs and compute the histogram aggregate bit_cost.
static void HistogramAnalyzeEntropyBin(VP8LHistogramSet* const image_histo,
int16_t* const bin_map, int low_effort) {
uint16_t* const bin_map,
int low_effort) {
int i;
VP8LHistogram** const histograms = image_histo->histograms;
const int histo_size = image_histo->size;
const int bin_depth = histo_size + 1;
DominantCostRange cost_range;
DominantCostRangeInit(&cost_range);
// Analyze the dominant (literal, red and blue) entropy costs.
for (i = 0; i < histo_size; ++i) {
VP8LHistogram* const histo = histograms[i];
UpdateDominantCostRange(histo, &cost_range);
UpdateDominantCostRange(histograms[i], &cost_range);
}
// bin-hash histograms on three of the dominant (literal, red and blue)
// symbol costs.
// symbol costs and store the resulting bin_id for each histogram.
for (i = 0; i < histo_size; ++i) {
const VP8LHistogram* const histo = histograms[i];
const int bin_id = GetHistoBinIndex(histo, &cost_range, low_effort);
const int bin_offset = bin_id * bin_depth;
// bin_map[n][0] for every bin 'n' maintains the counter for the number of
// histograms in that bin.
// Get and increment the num_histos in that bin.
const int num_histos = ++bin_map[bin_offset];
assert(bin_offset + num_histos < bin_depth * BIN_SIZE);
// Add histogram i'th index at num_histos (last) position in the bin_map.
bin_map[bin_offset + num_histos] = i;
bin_map[i] = GetHistoBinIndex(histograms[i], &cost_range, low_effort);
}
}
// Compact the histogram set by removing unused entries.
static void HistogramCompactBins(VP8LHistogramSet* const image_histo) {
VP8LHistogram** const histograms = image_histo->histograms;
int i, j;
for (i = 0, j = 0; i < image_histo->size; ++i) {
if (histograms[i] != NULL && histograms[i]->bit_cost_ != 0.) {
if (j < i) {
histograms[j] = histograms[i];
histograms[i] = NULL;
}
++j;
}
}
image_histo->size = j;
}
// Compact image_histo[] by merging some histograms with same bin_id together if
// it's advantageous.
static VP8LHistogram* HistogramCombineEntropyBin(
VP8LHistogramSet* const image_histo,
VP8LHistogram* cur_combo,
int16_t* const bin_map, int bin_depth, int num_bins,
const uint16_t* const bin_map, int bin_map_size, int num_bins,
double combine_cost_factor, int low_effort) {
int bin_id;
VP8LHistogram** const histograms = image_histo->histograms;
int idx;
// Work in-place: processed histograms are put at the beginning of
// image_histo[]. At the end, we just have to truncate the array.
int size = 0;
struct {
int16_t first; // position of the histogram that accumulates all
// histograms with the same bin_id
uint16_t num_combine_failures; // number of combine failures per bin_id
} bin_info[BIN_SIZE];
for (bin_id = 0; bin_id < num_bins; ++bin_id) {
const int bin_offset = bin_id * bin_depth;
const int num_histos = bin_map[bin_offset];
const int idx1 = bin_map[bin_offset + 1];
int num_combine_failures = 0;
int n;
for (n = 2; n <= num_histos; ++n) {
const int idx2 = bin_map[bin_offset + n];
if (low_effort) {
// Merge all histograms with the same bin index, irrespective of cost of
// the merged histograms.
VP8LHistogramAdd(histograms[idx1], histograms[idx2], histograms[idx1]);
histograms[idx2]->bit_cost_ = 0.;
} else {
const double bit_cost_idx2 = histograms[idx2]->bit_cost_;
if (bit_cost_idx2 > 0.) {
const double bit_cost_thresh = -bit_cost_idx2 * combine_cost_factor;
const double curr_cost_diff =
HistogramAddEval(histograms[idx1], histograms[idx2],
cur_combo, bit_cost_thresh);
if (curr_cost_diff < bit_cost_thresh) {
// Try to merge two histograms only if the combo is a trivial one or
// the two candidate histograms are already non-trivial.
// For some images, 'try_combine' turns out to be false for a lot of
// histogram pairs. In that case, we fallback to combining
// histograms as usual to avoid increasing the header size.
const int try_combine =
(cur_combo->trivial_symbol_ != VP8L_NON_TRIVIAL_SYM) ||
((histograms[idx1]->trivial_symbol_ == VP8L_NON_TRIVIAL_SYM) &&
(histograms[idx2]->trivial_symbol_ == VP8L_NON_TRIVIAL_SYM));
const int max_combine_failures = 32;
if (try_combine || (num_combine_failures >= max_combine_failures)) {
HistogramSwap(&cur_combo, &histograms[idx1]);
histograms[idx2]->bit_cost_ = 0.;
} else {
++num_combine_failures;
}
}
assert(num_bins <= BIN_SIZE);
for (idx = 0; idx < num_bins; ++idx) {
bin_info[idx].first = -1;
bin_info[idx].num_combine_failures = 0;
}
for (idx = 0; idx < bin_map_size; ++idx) {
const int bin_id = bin_map[idx];
const int first = bin_info[bin_id].first;
assert(size <= idx);
if (first == -1) {
// just move histogram #idx to its final position
histograms[size] = histograms[idx];
bin_info[bin_id].first = size++;
} else if (low_effort) {
HistogramAdd(histograms[idx], histograms[first], histograms[first]);
} else {
// try to merge #idx into #first (both share the same bin_id)
const double bit_cost = histograms[idx]->bit_cost_;
const double bit_cost_thresh = -bit_cost * combine_cost_factor;
const double curr_cost_diff =
HistogramAddEval(histograms[first], histograms[idx],
cur_combo, bit_cost_thresh);
if (curr_cost_diff < bit_cost_thresh) {
// Try to merge two histograms only if the combo is a trivial one or
// the two candidate histograms are already non-trivial.
// For some images, 'try_combine' turns out to be false for a lot of
// histogram pairs. In that case, we fallback to combining
// histograms as usual to avoid increasing the header size.
const int try_combine =
(cur_combo->trivial_symbol_ != VP8L_NON_TRIVIAL_SYM) ||
((histograms[idx]->trivial_symbol_ == VP8L_NON_TRIVIAL_SYM) &&
(histograms[first]->trivial_symbol_ == VP8L_NON_TRIVIAL_SYM));
const int max_combine_failures = 32;
if (try_combine ||
bin_info[bin_id].num_combine_failures >= max_combine_failures) {
// move the (better) merged histogram to its final slot
HistogramSwap(&cur_combo, &histograms[first]);
} else {
histograms[size++] = histograms[idx];
++bin_info[bin_id].num_combine_failures;
}
} else {
histograms[size++] = histograms[idx];
}
}
if (low_effort) {
// Update the bit_cost for the merged histograms (per bin index).
UpdateHistogramCost(histograms[idx1]);
}
image_histo->size = size;
if (low_effort) {
// for low_effort case, update the final cost when everything is merged
for (idx = 0; idx < size; ++idx) {
UpdateHistogramCost(histograms[idx]);
}
}
HistogramCompactBins(image_histo);
return cur_combo;
}
static uint32_t MyRand(uint32_t *seed) {
*seed *= 16807U;
static uint32_t MyRand(uint32_t* const seed) {
*seed = (*seed * 16807ull) & 0xffffffffu;
if (*seed == 0) {
*seed = 1;
}
@ -682,7 +723,7 @@ static int HistogramCombineGreedy(VP8LHistogramSet* const image_histo) {
HistogramPair* copy_to;
const int idx1 = histo_queue.queue[0].idx1;
const int idx2 = histo_queue.queue[0].idx2;
VP8LHistogramAdd(histograms[idx2], histograms[idx1], histograms[idx1]);
HistogramAdd(histograms[idx2], histograms[idx1], histograms[idx1]);
histograms[idx1]->bit_cost_ = histo_queue.queue[0].cost_combo;
// Remove merged histogram.
for (i = 0; i + 1 < image_histo_size; ++i) {
@ -748,6 +789,8 @@ static void HistogramCombineStochastic(VP8LHistogramSet* const image_histo,
const int outer_iters = image_histo_size * iter_mult;
const int num_pairs = image_histo_size / 2;
const int num_tries_no_success = outer_iters / 2;
int idx2_max = image_histo_size - 1;
int do_brute_dorce = 0;
VP8LHistogram** const histograms = image_histo->histograms;
// Collapse similar histograms in 'image_histo'.
@ -758,43 +801,62 @@ static void HistogramCombineStochastic(VP8LHistogramSet* const image_histo,
double best_cost_diff = 0.;
int best_idx1 = -1, best_idx2 = 1;
int j;
const int num_tries =
int num_tries =
(num_pairs < image_histo_size) ? num_pairs : image_histo_size;
// Use a brute force approach if:
// - stochastic has not worked for a while and
// - if the number of iterations for brute force is less than the number of
// iterations if we never find a match ever again stochastically (hence
// num_tries times the number of remaining outer iterations).
do_brute_dorce =
(tries_with_no_success > 10) &&
(idx2_max * (idx2_max + 1) < 2 * num_tries * (outer_iters - iter));
if (do_brute_dorce) num_tries = idx2_max;
seed += iter;
for (j = 0; j < num_tries; ++j) {
double curr_cost_diff;
// Choose two histograms at random and try to combine them.
const uint32_t idx1 = MyRand(&seed) % image_histo_size;
const uint32_t tmp = (j & 7) + 1;
const uint32_t diff =
(tmp < 3) ? tmp : MyRand(&seed) % (image_histo_size - 1);
const uint32_t idx2 = (idx1 + diff + 1) % image_histo_size;
if (idx1 == idx2) {
continue;
uint32_t idx1, idx2;
if (do_brute_dorce) {
// Use a brute force approach.
idx1 = (uint32_t)j;
idx2 = (uint32_t)idx2_max;
} else {
const uint32_t tmp = (j & 7) + 1;
const uint32_t diff =
(tmp < 3) ? tmp : MyRand(&seed) % (image_histo_size - 1);
idx1 = MyRand(&seed) % image_histo_size;
idx2 = (idx1 + diff + 1) % image_histo_size;
if (idx1 == idx2) {
continue;
}
}
// Calculate cost reduction on combining.
curr_cost_diff = HistogramAddEval(histograms[idx1], histograms[idx2],
tmp_histo, best_cost_diff);
if (curr_cost_diff < best_cost_diff) { // found a better pair?
if (curr_cost_diff < best_cost_diff) { // found a better pair?
HistogramSwap(&best_combo, &tmp_histo);
best_cost_diff = curr_cost_diff;
best_idx1 = idx1;
best_idx2 = idx2;
}
}
if (do_brute_dorce) --idx2_max;
if (best_idx1 >= 0) {
HistogramSwap(&best_combo, &histograms[best_idx1]);
// swap best_idx2 slot with last one (which is now unused)
--image_histo_size;
if (idx2_max >= image_histo_size) idx2_max = image_histo_size - 1;
if (best_idx2 != image_histo_size) {
HistogramSwap(&histograms[image_histo_size], &histograms[best_idx2]);
histograms[image_histo_size] = NULL;
}
tries_with_no_success = 0;
}
if (++tries_with_no_success >= num_tries_no_success) {
if (++tries_with_no_success >= num_tries_no_success || idx2_max == 0) {
break;
}
}
@ -843,7 +905,7 @@ static void HistogramRemap(const VP8LHistogramSet* const in,
for (i = 0; i < in_size; ++i) {
const int idx = symbols[i];
VP8LHistogramAdd(in_histo[i], out_histo[idx], out_histo[idx]);
HistogramAdd(in_histo[i], out_histo[idx], out_histo[idx]);
}
}
@ -869,32 +931,18 @@ int VP8LGetHistoImageSymbols(int xsize, int ysize,
const int histo_xsize = histo_bits ? VP8LSubSampleSize(xsize, histo_bits) : 1;
const int histo_ysize = histo_bits ? VP8LSubSampleSize(ysize, histo_bits) : 1;
const int image_histo_raw_size = histo_xsize * histo_ysize;
const int entropy_combine_num_bins = low_effort ? NUM_PARTITIONS : BIN_SIZE;
// The bin_map for every bin follows following semantics:
// bin_map[n][0] = num_histo; // The number of histograms in that bin.
// bin_map[n][1] = index of first histogram in that bin;
// bin_map[n][num_histo] = index of last histogram in that bin;
// bin_map[n][num_histo + 1] ... bin_map[n][bin_depth - 1] = unused indices.
const int bin_depth = image_histo_raw_size + 1;
int16_t* bin_map = NULL;
VP8LHistogramSet* const orig_histo =
VP8LAllocateHistogramSet(image_histo_raw_size, cache_bits);
VP8LHistogram* cur_combo;
// Don't attempt linear bin-partition heuristic for
// histograms of small sizes (as bin_map will be very sparse) and
// maximum quality q==100 (to preserve the compression gains at that level).
const int entropy_combine_num_bins = low_effort ? NUM_PARTITIONS : BIN_SIZE;
const int entropy_combine =
(orig_histo->size > entropy_combine_num_bins * 2) && (quality < 100);
if (orig_histo == NULL) goto Error;
// Don't attempt linear bin-partition heuristic for:
// histograms of small sizes, as bin_map will be very sparse and;
// Maximum quality (q==100), to preserve the compression gains at that level.
if (entropy_combine) {
const int bin_map_size = bin_depth * entropy_combine_num_bins;
bin_map = (int16_t*)WebPSafeCalloc(bin_map_size, sizeof(*bin_map));
if (bin_map == NULL) goto Error;
}
// Construct the histograms from backward references.
HistogramBuild(xsize, histo_bits, refs, orig_histo);
// Copies the histograms and computes its bit_cost.
@ -902,12 +950,17 @@ int VP8LGetHistoImageSymbols(int xsize, int ysize,
cur_combo = tmp_histos->histograms[1]; // pick up working slot
if (entropy_combine) {
const int bin_map_size = orig_histo->size;
// Reuse histogram_symbols storage. By definition, it's guaranteed to be ok.
uint16_t* const bin_map = histogram_symbols;
const double combine_cost_factor =
GetCombineCostFactor(image_histo_raw_size, quality);
HistogramAnalyzeEntropyBin(orig_histo, bin_map, low_effort);
// Collapse histograms with similar entropy.
cur_combo = HistogramCombineEntropyBin(image_histo, cur_combo, bin_map,
bin_depth, entropy_combine_num_bins,
cur_combo = HistogramCombineEntropyBin(image_histo, cur_combo,
bin_map, bin_map_size,
entropy_combine_num_bins,
combine_cost_factor, low_effort);
}
@ -932,7 +985,6 @@ int VP8LGetHistoImageSymbols(int xsize, int ysize,
ok = 1;
Error:
WebPSafeFree(bin_map);
VP8LFreeHistogramSet(orig_histo);
return ok;
}

2
thirdparty/libwebp/enc/histogram.h → thirdparty/libwebp/enc/histogram_enc.h vendored
View file

@ -16,7 +16,7 @@
#include <string.h>
#include "./backward_references.h"
#include "./backward_references_enc.h"
#include "../webp/format_constants.h"
#include "../webp/types.h"

19
thirdparty/libwebp/enc/iterator.c → thirdparty/libwebp/enc/iterator_enc.c vendored
View file

@ -13,7 +13,7 @@
#include <string.h>
#include "./vp8enci.h"
#include "./vp8i_enc.h"
//------------------------------------------------------------------------------
// VP8Iterator
@ -53,7 +53,6 @@ void VP8IteratorReset(VP8EncIterator* const it) {
VP8IteratorSetRow(it, 0);
VP8IteratorSetCountDown(it, enc->mb_w_ * enc->mb_h_); // default
InitTop(it);
InitLeft(it);
memset(it->bit_count_, 0, sizeof(it->bit_count_));
it->do_trellis_ = 0;
}
@ -68,8 +67,6 @@ int VP8IteratorIsDone(const VP8EncIterator* const it) {
void VP8IteratorInit(VP8Encoder* const enc, VP8EncIterator* const it) {
it->enc_ = enc;
it->y_stride_ = enc->pic_->y_stride;
it->uv_stride_ = enc->pic_->uv_stride;
it->yuv_in_ = (uint8_t*)WEBP_ALIGN(it->yuv_mem_);
it->yuv_out_ = it->yuv_in_ + YUV_SIZE_ENC;
it->yuv_out2_ = it->yuv_out_ + YUV_SIZE_ENC;
@ -309,14 +306,14 @@ void VP8IteratorSaveBoundary(VP8EncIterator* const it) {
}
int VP8IteratorNext(VP8EncIterator* const it) {
it->preds_ += 4;
it->mb_ += 1;
it->nz_ += 1;
it->y_top_ += 16;
it->uv_top_ += 16;
it->x_ += 1;
if (it->x_ == it->enc_->mb_w_) {
if (++it->x_ == it->enc_->mb_w_) {
VP8IteratorSetRow(it, ++it->y_);
} else {
it->preds_ += 4;
it->mb_ += 1;
it->nz_ += 1;
it->y_top_ += 16;
it->uv_top_ += 16;
}
return (0 < --it->count_down_);
}

4
thirdparty/libwebp/enc/near_lossless.c → thirdparty/libwebp/enc/near_lossless_enc.c vendored
View file

@ -17,9 +17,9 @@
#include <assert.h>
#include <stdlib.h>
#include "../dsp/lossless.h"
#include "../dsp/lossless_common.h"
#include "../utils/utils.h"
#include "./vp8enci.h"
#include "./vp8i_enc.h"
#define MIN_DIM_FOR_NEAR_LOSSLESS 64
#define MAX_LIMIT_BITS 5

269
thirdparty/libwebp/enc/picture_csp.c → thirdparty/libwebp/enc/picture_csp_enc.c vendored
View file

@ -15,8 +15,8 @@
#include <stdlib.h>
#include <math.h>
#include "./vp8enci.h"
#include "../utils/random.h"
#include "./vp8i_enc.h"
#include "../utils/random_utils.h"
#include "../utils/utils.h"
#include "../dsp/yuv.h"
@ -153,9 +153,9 @@ static int RGBToV(int r, int g, int b, VP8Random* const rg) {
}
//------------------------------------------------------------------------------
// Smart RGB->YUV conversion
// Sharp RGB->YUV conversion
static const int kNumIterations = 6;
static const int kNumIterations = 4;
static const int kMinDimensionIterativeConversion = 4;
// We could use SFIX=0 and only uint8_t for fixed_y_t, but it produces some
@ -171,9 +171,9 @@ typedef uint16_t fixed_y_t; // unsigned type with extra SFIX precision for W
#if defined(USE_GAMMA_COMPRESSION)
// float variant of gamma-correction
// We use tables of different size and precision, along with a 'real-world'
// Gamma value close to ~2.
#define kGammaF 2.2
// We use tables of different size and precision for the Rec709
// transfer function.
#define kGammaF (1./0.45)
static float kGammaToLinearTabF[MAX_Y_T + 1]; // size scales with Y_FIX
static float kLinearToGammaTabF[kGammaTabSize + 2];
static volatile int kGammaTablesFOk = 0;
@ -183,11 +183,26 @@ static WEBP_TSAN_IGNORE_FUNCTION void InitGammaTablesF(void) {
int v;
const double norm = 1. / MAX_Y_T;
const double scale = 1. / kGammaTabSize;
const double a = 0.099;
const double thresh = 0.018;
for (v = 0; v <= MAX_Y_T; ++v) {
kGammaToLinearTabF[v] = (float)pow(norm * v, kGammaF);
const double g = norm * v;
if (g <= thresh * 4.5) {
kGammaToLinearTabF[v] = (float)(g / 4.5);
} else {
const double a_rec = 1. / (1. + a);
kGammaToLinearTabF[v] = (float)pow(a_rec * (g + a), kGammaF);
}
}
for (v = 0; v <= kGammaTabSize; ++v) {
kLinearToGammaTabF[v] = (float)(MAX_Y_T * pow(scale * v, 1. / kGammaF));
const double g = scale * v;
double value;
if (g <= thresh) {
value = 4.5 * g;
} else {
value = (1. + a) * pow(g, 1. / kGammaF) - a;
}
kLinearToGammaTabF[v] = (float)(MAX_Y_T * value);
}
// to prevent small rounding errors to cause read-overflow:
kLinearToGammaTabF[kGammaTabSize + 1] = kLinearToGammaTabF[kGammaTabSize];
@ -235,12 +250,12 @@ static fixed_y_t clip_y(int y) {
//------------------------------------------------------------------------------
static int RGBToGray(int r, int g, int b) {
const int luma = 19595 * r + 38470 * g + 7471 * b + YUV_HALF;
const int luma = 13933 * r + 46871 * g + 4732 * b + YUV_HALF;
return (luma >> YUV_FIX);
}
static float RGBToGrayF(float r, float g, float b) {
return 0.299f * r + 0.587f * g + 0.114f * b;
return (float)(0.2126 * r + 0.7152 * g + 0.0722 * b);
}
static int ScaleDown(int a, int b, int c, int d) {
@ -251,59 +266,51 @@ static int ScaleDown(int a, int b, int c, int d) {
return LinearToGammaF(0.25f * (A + B + C + D));
}
static WEBP_INLINE void UpdateW(const fixed_y_t* src, fixed_y_t* dst, int len) {
while (len-- > 0) {
const float R = GammaToLinearF(src[0]);
const float G = GammaToLinearF(src[1]);
const float B = GammaToLinearF(src[2]);
static WEBP_INLINE void UpdateW(const fixed_y_t* src, fixed_y_t* dst, int w) {
int i;
for (i = 0; i < w; ++i) {
const float R = GammaToLinearF(src[0 * w + i]);
const float G = GammaToLinearF(src[1 * w + i]);
const float B = GammaToLinearF(src[2 * w + i]);
const float Y = RGBToGrayF(R, G, B);
*dst++ = (fixed_y_t)LinearToGammaF(Y);
src += 3;
dst[i] = (fixed_y_t)LinearToGammaF(Y);
}
}
static int UpdateChroma(const fixed_y_t* src1,
const fixed_y_t* src2,
fixed_t* dst, fixed_y_t* tmp, int len) {
int diff = 0;
while (len--> 0) {
const int r = ScaleDown(src1[0], src1[3], src2[0], src2[3]);
const int g = ScaleDown(src1[1], src1[4], src2[1], src2[4]);
const int b = ScaleDown(src1[2], src1[5], src2[2], src2[5]);
static void UpdateChroma(const fixed_y_t* src1, const fixed_y_t* src2,
fixed_t* dst, int uv_w) {
int i;
for (i = 0; i < uv_w; ++i) {
const int r = ScaleDown(src1[0 * uv_w + 0], src1[0 * uv_w + 1],
src2[0 * uv_w + 0], src2[0 * uv_w + 1]);
const int g = ScaleDown(src1[2 * uv_w + 0], src1[2 * uv_w + 1],
src2[2 * uv_w + 0], src2[2 * uv_w + 1]);
const int b = ScaleDown(src1[4 * uv_w + 0], src1[4 * uv_w + 1],
src2[4 * uv_w + 0], src2[4 * uv_w + 1]);
const int W = RGBToGray(r, g, b);
const int r_avg = (src1[0] + src1[3] + src2[0] + src2[3] + 2) >> 2;
const int g_avg = (src1[1] + src1[4] + src2[1] + src2[4] + 2) >> 2;
const int b_avg = (src1[2] + src1[5] + src2[2] + src2[5] + 2) >> 2;
dst[0] = (fixed_t)(r - W);
dst[1] = (fixed_t)(g - W);
dst[2] = (fixed_t)(b - W);
dst += 3;
src1 += 6;
src2 += 6;
if (tmp != NULL) {
tmp[0] = tmp[1] = clip_y(W);
tmp += 2;
}
diff += abs(RGBToGray(r_avg, g_avg, b_avg) - W);
dst[0 * uv_w] = (fixed_t)(r - W);
dst[1 * uv_w] = (fixed_t)(g - W);
dst[2 * uv_w] = (fixed_t)(b - W);
dst += 1;
src1 += 2;
src2 += 2;
}
}
static void StoreGray(const fixed_y_t* rgb, fixed_y_t* y, int w) {
int i;
for (i = 0; i < w; ++i) {
y[i] = RGBToGray(rgb[0 * w + i], rgb[1 * w + i], rgb[2 * w + i]);
}
return diff;
}
//------------------------------------------------------------------------------
static WEBP_INLINE int Filter(const fixed_t* const A, const fixed_t* const B,
int rightwise) {
int v;
if (!rightwise) {
v = (A[0] * 9 + A[-3] * 3 + B[0] * 3 + B[-3]);
} else {
v = (A[0] * 9 + A[+3] * 3 + B[0] * 3 + B[+3]);
}
return (v + 8) >> 4;
static WEBP_INLINE fixed_y_t Filter2(int A, int B, int W0) {
const int v0 = (A * 3 + B + 2) >> 2;
return clip_y(v0 + W0);
}
static WEBP_INLINE int Filter2(int A, int B) { return (A * 3 + B + 2) >> 2; }
//------------------------------------------------------------------------------
static WEBP_INLINE fixed_y_t UpLift(uint8_t a) { // 8bit -> SFIX
@ -317,52 +324,50 @@ static void ImportOneRow(const uint8_t* const r_ptr,
int pic_width,
fixed_y_t* const dst) {
int i;
const int w = (pic_width + 1) & ~1;
for (i = 0; i < pic_width; ++i) {
const int off = i * step;
dst[3 * i + 0] = UpLift(r_ptr[off]);
dst[3 * i + 1] = UpLift(g_ptr[off]);
dst[3 * i + 2] = UpLift(b_ptr[off]);
dst[i + 0 * w] = UpLift(r_ptr[off]);
dst[i + 1 * w] = UpLift(g_ptr[off]);
dst[i + 2 * w] = UpLift(b_ptr[off]);
}
if (pic_width & 1) { // replicate rightmost pixel
memcpy(dst + 3 * pic_width, dst + 3 * (pic_width - 1), 3 * sizeof(*dst));
dst[pic_width + 0 * w] = dst[pic_width + 0 * w - 1];
dst[pic_width + 1 * w] = dst[pic_width + 1 * w - 1];
dst[pic_width + 2 * w] = dst[pic_width + 2 * w - 1];
}
}
static void InterpolateTwoRows(const fixed_y_t* const best_y,
const fixed_t* const prev_uv,
const fixed_t* const cur_uv,
const fixed_t* const next_uv,
const fixed_t* prev_uv,
const fixed_t* cur_uv,
const fixed_t* next_uv,
int w,
fixed_y_t* const out1,
fixed_y_t* const out2) {
int i, k;
{ // special boundary case for i==0
const int W0 = best_y[0];
const int W1 = best_y[w];
for (k = 0; k <= 2; ++k) {
out1[k] = clip_y(Filter2(cur_uv[k], prev_uv[k]) + W0);
out2[k] = clip_y(Filter2(cur_uv[k], next_uv[k]) + W1);
}
}
for (i = 1; i < w - 1; ++i) {
const int W0 = best_y[i + 0];
const int W1 = best_y[i + w];
const int off = 3 * (i >> 1);
for (k = 0; k <= 2; ++k) {
const int tmp0 = Filter(cur_uv + off + k, prev_uv + off + k, i & 1);
const int tmp1 = Filter(cur_uv + off + k, next_uv + off + k, i & 1);
out1[3 * i + k] = clip_y(tmp0 + W0);
out2[3 * i + k] = clip_y(tmp1 + W1);
}
}
{ // special boundary case for i == w - 1
const int W0 = best_y[i + 0];
const int W1 = best_y[i + w];
const int off = 3 * (i >> 1);
for (k = 0; k <= 2; ++k) {
out1[3 * i + k] = clip_y(Filter2(cur_uv[off + k], prev_uv[off + k]) + W0);
out2[3 * i + k] = clip_y(Filter2(cur_uv[off + k], next_uv[off + k]) + W1);
fixed_y_t* out1,
fixed_y_t* out2) {
const int uv_w = w >> 1;
const int len = (w - 1) >> 1; // length to filter
int k = 3;
while (k-- > 0) { // process each R/G/B segments in turn
// special boundary case for i==0
out1[0] = Filter2(cur_uv[0], prev_uv[0], best_y[0]);
out2[0] = Filter2(cur_uv[0], next_uv[0], best_y[w]);
WebPSharpYUVFilterRow(cur_uv, prev_uv, len, best_y + 0 + 1, out1 + 1);
WebPSharpYUVFilterRow(cur_uv, next_uv, len, best_y + w + 1, out2 + 1);
// special boundary case for i == w - 1 when w is even
if (!(w & 1)) {
out1[w - 1] = Filter2(cur_uv[uv_w - 1], prev_uv[uv_w - 1],
best_y[w - 1 + 0]);
out2[w - 1] = Filter2(cur_uv[uv_w - 1], next_uv[uv_w - 1],
best_y[w - 1 + w]);
}
out1 += w;
out2 += w;
prev_uv += uv_w;
cur_uv += uv_w;
next_uv += uv_w;
}
}
@ -394,11 +399,11 @@ static int ConvertWRGBToYUV(const fixed_y_t* best_y, const fixed_t* best_uv,
const int uv_h = h >> 1;
for (best_uv = best_uv_base, j = 0; j < picture->height; ++j) {
for (i = 0; i < picture->width; ++i) {
const int off = 3 * (i >> 1);
const int off = (i >> 1);
const int W = best_y[i];
const int r = best_uv[off + 0] + W;
const int g = best_uv[off + 1] + W;
const int b = best_uv[off + 2] + W;
const int r = best_uv[off + 0 * uv_w] + W;
const int g = best_uv[off + 1 * uv_w] + W;
const int b = best_uv[off + 2 * uv_w] + W;
dst_y[i] = ConvertRGBToY(r, g, b);
}
best_y += w;
@ -407,10 +412,10 @@ static int ConvertWRGBToYUV(const fixed_y_t* best_y, const fixed_t* best_uv,
}
for (best_uv = best_uv_base, j = 0; j < uv_h; ++j) {
for (i = 0; i < uv_w; ++i) {
const int off = 3 * i;
const int r = best_uv[off + 0];
const int g = best_uv[off + 1];
const int b = best_uv[off + 2];
const int off = i;
const int r = best_uv[off + 0 * uv_w];
const int g = best_uv[off + 1 * uv_w];
const int b = best_uv[off + 2 * uv_w];
dst_u[i] = ConvertRGBToU(r, g, b);
dst_v[i] = ConvertRGBToV(r, g, b);
}
@ -436,7 +441,8 @@ static int PreprocessARGB(const uint8_t* r_ptr,
const int h = (picture->height + 1) & ~1;
const int uv_w = w >> 1;
const int uv_h = h >> 1;
int i, j, iter;
uint64_t prev_diff_y_sum = ~0;
int j, iter;
// TODO(skal): allocate one big memory chunk. But for now, it's easier
// for valgrind debugging to have several chunks.
@ -451,11 +457,8 @@ static int PreprocessARGB(const uint8_t* r_ptr,
fixed_y_t* target_y = target_y_base;
fixed_t* best_uv = best_uv_base;
fixed_t* target_uv = target_uv_base;
const uint64_t diff_y_threshold = (uint64_t)(3.0 * w * h);
int ok;
int diff_sum = 0;
const int first_diff_threshold = (int)(2.5 * w * h);
const int min_improvement = 5; // stop if improvement is below this %
const int min_first_improvement = 80;
if (best_y_base == NULL || best_uv_base == NULL ||
target_y_base == NULL || target_uv_base == NULL ||
@ -467,10 +470,12 @@ static int PreprocessARGB(const uint8_t* r_ptr,
assert(picture->width >= kMinDimensionIterativeConversion);
assert(picture->height >= kMinDimensionIterativeConversion);
WebPInitConvertARGBToYUV();
// Import RGB samples to W/RGB representation.
for (j = 0; j < picture->height; j += 2) {
const int is_last_row = (j == picture->height - 1);
fixed_y_t* const src1 = tmp_buffer;
fixed_y_t* const src1 = tmp_buffer + 0 * w;
fixed_y_t* const src2 = tmp_buffer + 3 * w;
// prepare two rows of input
@ -481,11 +486,13 @@ static int PreprocessARGB(const uint8_t* r_ptr,
} else {
memcpy(src2, src1, 3 * w * sizeof(*src2));
}
StoreGray(src1, best_y + 0, w);
StoreGray(src2, best_y + w, w);
UpdateW(src1, target_y, w);
UpdateW(src2, target_y + w, w);
diff_sum += UpdateChroma(src1, src2, target_uv, best_y, uv_w);
UpdateChroma(src1, src2, target_uv, uv_w);
memcpy(best_uv, target_uv, 3 * uv_w * sizeof(*best_uv));
memcpy(best_y + w, best_y, w * sizeof(*best_y));
best_y += 2 * w;
best_uv += 3 * uv_w;
target_y += 2 * w;
@ -497,18 +504,16 @@ static int PreprocessARGB(const uint8_t* r_ptr,
// Iterate and resolve clipping conflicts.
for (iter = 0; iter < kNumIterations; ++iter) {
int k;
const fixed_t* cur_uv = best_uv_base;
const fixed_t* prev_uv = best_uv_base;
const int old_diff_sum = diff_sum;
diff_sum = 0;
uint64_t diff_y_sum = 0;
best_y = best_y_base;
best_uv = best_uv_base;
target_y = target_y_base;
target_uv = target_uv_base;
for (j = 0; j < h; j += 2) {
fixed_y_t* const src1 = tmp_buffer;
fixed_y_t* const src1 = tmp_buffer + 0 * w;
fixed_y_t* const src2 = tmp_buffer + 3 * w;
{
const fixed_t* const next_uv = cur_uv + ((j < h - 2) ? 3 * uv_w : 0);
@ -519,50 +524,24 @@ static int PreprocessARGB(const uint8_t* r_ptr,
UpdateW(src1, best_rgb_y + 0 * w, w);
UpdateW(src2, best_rgb_y + 1 * w, w);
diff_sum += UpdateChroma(src1, src2, best_rgb_uv, NULL, uv_w);
UpdateChroma(src1, src2, best_rgb_uv, uv_w);
// update two rows of Y and one row of RGB
for (i = 0; i < 2 * w; ++i) {
const int diff_y = target_y[i] - best_rgb_y[i];
const int new_y = (int)best_y[i] + diff_y;
best_y[i] = clip_y(new_y);
}
for (i = 0; i < uv_w; ++i) {
const int off = 3 * i;
int W;
for (k = 0; k <= 2; ++k) {
const int diff_uv = (int)target_uv[off + k] - best_rgb_uv[off + k];
best_uv[off + k] += diff_uv;
}
W = RGBToGray(best_uv[off + 0], best_uv[off + 1], best_uv[off + 2]);
for (k = 0; k <= 2; ++k) {
best_uv[off + k] -= W;
}
}
diff_y_sum += WebPSharpYUVUpdateY(target_y, best_rgb_y, best_y, 2 * w);
WebPSharpYUVUpdateRGB(target_uv, best_rgb_uv, best_uv, 3 * uv_w);
best_y += 2 * w;
best_uv += 3 * uv_w;
target_y += 2 * w;
target_uv += 3 * uv_w;
}
// test exit condition
if (diff_sum > 0) {
const int improvement = 100 * abs(diff_sum - old_diff_sum) / diff_sum;
// Check if first iteration gave good result already, without a large
// jump of improvement (otherwise it means we need to try few extra
// iterations, just to be sure).
if (iter == 0 && diff_sum < first_diff_threshold &&
improvement < min_first_improvement) {
break;
}
// then, check if improvement is stalling.
if (improvement < min_improvement) {
break;
}
} else {
break;
if (iter > 0) {
if (diff_y_sum < diff_y_threshold) break;
if (diff_y_sum > prev_diff_y_sum) break;
}
prev_diff_y_sum = diff_y_sum;
}
// final reconstruction
ok = ConvertWRGBToYUV(best_y_base, best_uv_base, picture);
@ -1032,9 +1011,13 @@ int WebPPictureARGBToYUVA(WebPPicture* picture, WebPEncCSP colorspace) {
return PictureARGBToYUVA(picture, colorspace, 0.f, 0);
}
int WebPPictureSmartARGBToYUVA(WebPPicture* picture) {
int WebPPictureSharpARGBToYUVA(WebPPicture* picture) {
return PictureARGBToYUVA(picture, WEBP_YUV420, 0.f, 1);
}
// for backward compatibility
int WebPPictureSmartARGBToYUVA(WebPPicture* picture) {
return WebPPictureSharpARGBToYUVA(picture);
}
//------------------------------------------------------------------------------
// call for YUVA -> ARGB conversion

2
thirdparty/libwebp/enc/picture.c → thirdparty/libwebp/enc/picture_enc.c vendored
View file

@ -14,7 +14,7 @@
#include <assert.h>
#include <stdlib.h>
#include "./vp8enci.h"
#include "./vp8i_enc.h"
#include "../dsp/dsp.h"
#include "../utils/utils.h"

177
thirdparty/libwebp/enc/picture_psnr.c vendored
View file

@ -1,177 +0,0 @@
// Copyright 2014 Google Inc. All Rights Reserved.
//
// Use of this source code is governed by a BSD-style license
// that can be found in the COPYING file in the root of the source
// tree. An additional intellectual property rights grant can be found
// in the file PATENTS. All contributing project authors may
// be found in the AUTHORS file in the root of the source tree.
// -----------------------------------------------------------------------------
//
// WebPPicture tools for measuring distortion
//
// Author: Skal (pascal.massimino@gmail.com)
#include <math.h>
#include <stdlib.h>
#include "./vp8enci.h"
#include "../utils/utils.h"
//------------------------------------------------------------------------------
// local-min distortion
//
// For every pixel in the *reference* picture, we search for the local best
// match in the compressed image. This is not a symmetrical measure.
#define RADIUS 2 // search radius. Shouldn't be too large.
static void AccumulateLSIM(const uint8_t* src, int src_stride,
const uint8_t* ref, int ref_stride,
int w, int h, VP8DistoStats* stats) {
int x, y;
double total_sse = 0.;
for (y = 0; y < h; ++y) {
const int y_0 = (y - RADIUS < 0) ? 0 : y - RADIUS;
const int y_1 = (y + RADIUS + 1 >= h) ? h : y + RADIUS + 1;
for (x = 0; x < w; ++x) {
const int x_0 = (x - RADIUS < 0) ? 0 : x - RADIUS;
const int x_1 = (x + RADIUS + 1 >= w) ? w : x + RADIUS + 1;
double best_sse = 255. * 255.;
const double value = (double)ref[y * ref_stride + x];
int i, j;
for (j = y_0; j < y_1; ++j) {
const uint8_t* const s = src + j * src_stride;
for (i = x_0; i < x_1; ++i) {
const double diff = s[i] - value;
const double sse = diff * diff;
if (sse < best_sse) best_sse = sse;
}
}
total_sse += best_sse;
}
}
stats->w = w * h;
stats->xm = 0;
stats->ym = 0;
stats->xxm = total_sse;
stats->yym = 0;
stats->xxm = 0;
}
#undef RADIUS
//------------------------------------------------------------------------------
// Distortion
// Max value returned in case of exact similarity.
static const double kMinDistortion_dB = 99.;
static float GetPSNR(const double v) {
return (float)((v > 0.) ? -4.3429448 * log(v / (255 * 255.))
: kMinDistortion_dB);
}
int WebPPictureDistortion(const WebPPicture* src, const WebPPicture* ref,
int type, float result[5]) {
VP8DistoStats stats[5];
int w, h;
memset(stats, 0, sizeof(stats));
VP8SSIMDspInit();
if (src == NULL || ref == NULL ||
src->width != ref->width || src->height != ref->height ||
src->use_argb != ref->use_argb || result == NULL) {
return 0;
}
w = src->width;
h = src->height;
if (src->use_argb == 1) {
if (src->argb == NULL || ref->argb == NULL) {
return 0;
} else {
int i, j, c;
uint8_t* tmp1, *tmp2;
uint8_t* const tmp_plane =
(uint8_t*)WebPSafeMalloc(2ULL * w * h, sizeof(*tmp_plane));
if (tmp_plane == NULL) return 0;
tmp1 = tmp_plane;
tmp2 = tmp_plane + w * h;
for (c = 0; c < 4; ++c) {
for (j = 0; j < h; ++j) {
for (i = 0; i < w; ++i) {
tmp1[j * w + i] = src->argb[i + j * src->argb_stride] >> (c * 8);
tmp2[j * w + i] = ref->argb[i + j * ref->argb_stride] >> (c * 8);
}
}
if (type >= 2) {
AccumulateLSIM(tmp1, w, tmp2, w, w, h, &stats[c]);
} else {
VP8SSIMAccumulatePlane(tmp1, w, tmp2, w, w, h, &stats[c]);
}
}
WebPSafeFree(tmp_plane);
}
} else {
int has_alpha, uv_w, uv_h;
if (src->y == NULL || ref->y == NULL ||
src->u == NULL || ref->u == NULL ||
src->v == NULL || ref->v == NULL) {
return 0;
}
has_alpha = !!(src->colorspace & WEBP_CSP_ALPHA_BIT);
if (has_alpha != !!(ref->colorspace & WEBP_CSP_ALPHA_BIT) ||
(has_alpha && (src->a == NULL || ref->a == NULL))) {
return 0;
}
uv_w = (src->width + 1) >> 1;
uv_h = (src->height + 1) >> 1;
if (type >= 2) {
AccumulateLSIM(src->y, src->y_stride, ref->y, ref->y_stride,
w, h, &stats[0]);
AccumulateLSIM(src->u, src->uv_stride, ref->u, ref->uv_stride,
uv_w, uv_h, &stats[1]);
AccumulateLSIM(src->v, src->uv_stride, ref->v, ref->uv_stride,
uv_w, uv_h, &stats[2]);
if (has_alpha) {
AccumulateLSIM(src->a, src->a_stride, ref->a, ref->a_stride,
w, h, &stats[3]);
}
} else {
VP8SSIMAccumulatePlane(src->y, src->y_stride,
ref->y, ref->y_stride,
w, h, &stats[0]);
VP8SSIMAccumulatePlane(src->u, src->uv_stride,
ref->u, ref->uv_stride,
uv_w, uv_h, &stats[1]);
VP8SSIMAccumulatePlane(src->v, src->uv_stride,
ref->v, ref->uv_stride,
uv_w, uv_h, &stats[2]);
if (has_alpha) {
VP8SSIMAccumulatePlane(src->a, src->a_stride,
ref->a, ref->a_stride,
w, h, &stats[3]);
}
}
}
// Final stat calculations.
{
int c;
for (c = 0; c <= 4; ++c) {
if (type == 1) {
const double v = VP8SSIMGet(&stats[c]);
result[c] = (float)((v < 1.) ? -10.0 * log10(1. - v)
: kMinDistortion_dB);
} else {
const double v = VP8SSIMGetSquaredError(&stats[c]);
result[c] = GetPSNR(v);
}
// Accumulate forward
if (c < 4) VP8SSIMAddStats(&stats[c], &stats[4]);
}
}
return 1;
}
//------------------------------------------------------------------------------

213
thirdparty/libwebp/enc/picture_psnr_enc.c vendored Normal file
View file

@ -0,0 +1,213 @@
// Copyright 2014 Google Inc. All Rights Reserved.
//
// Use of this source code is governed by a BSD-style license
// that can be found in the COPYING file in the root of the source
// tree. An additional intellectual property rights grant can be found
// in the file PATENTS. All contributing project authors may
// be found in the AUTHORS file in the root of the source tree.
// -----------------------------------------------------------------------------
//
// WebPPicture tools for measuring distortion
//
// Author: Skal (pascal.massimino@gmail.com)
#include <math.h>
#include <stdlib.h>
#include "./vp8i_enc.h"
#include "../utils/utils.h"
typedef double (*AccumulateFunc)(const uint8_t* src, int src_stride,
const uint8_t* ref, int ref_stride,
int w, int h);
//------------------------------------------------------------------------------
// local-min distortion
//
// For every pixel in the *reference* picture, we search for the local best
// match in the compressed image. This is not a symmetrical measure.
#define RADIUS 2 // search radius. Shouldn't be too large.
static double AccumulateLSIM(const uint8_t* src, int src_stride,
const uint8_t* ref, int ref_stride,
int w, int h) {
int x, y;
double total_sse = 0.;
for (y = 0; y < h; ++y) {
const int y_0 = (y - RADIUS < 0) ? 0 : y - RADIUS;
const int y_1 = (y + RADIUS + 1 >= h) ? h : y + RADIUS + 1;
for (x = 0; x < w; ++x) {
const int x_0 = (x - RADIUS < 0) ? 0 : x - RADIUS;
const int x_1 = (x + RADIUS + 1 >= w) ? w : x + RADIUS + 1;
double best_sse = 255. * 255.;
const double value = (double)ref[y * ref_stride + x];
int i, j;
for (j = y_0; j < y_1; ++j) {
const uint8_t* const s = src + j * src_stride;
for (i = x_0; i < x_1; ++i) {
const double diff = s[i] - value;
const double sse = diff * diff;
if (sse < best_sse) best_sse = sse;
}
}
total_sse += best_sse;
}
}
return total_sse;
}
#undef RADIUS
static double AccumulateSSE(const uint8_t* src, int src_stride,
const uint8_t* ref, int ref_stride,
int w, int h) {
int y;
double total_sse = 0.;
for (y = 0; y < h; ++y) {
total_sse += VP8AccumulateSSE(src, ref, w);
src += src_stride;
ref += ref_stride;
}
return total_sse;
}
//------------------------------------------------------------------------------
static double AccumulateSSIM(const uint8_t* src, int src_stride,
const uint8_t* ref, int ref_stride,
int w, int h) {
const int w0 = (w < VP8_SSIM_KERNEL) ? w : VP8_SSIM_KERNEL;
const int w1 = w - VP8_SSIM_KERNEL - 1;
const int h0 = (h < VP8_SSIM_KERNEL) ? h : VP8_SSIM_KERNEL;
const int h1 = h - VP8_SSIM_KERNEL - 1;
int x, y;
double sum = 0.;
for (y = 0; y < h0; ++y) {
for (x = 0; x < w; ++x) {
sum += VP8SSIMGetClipped(src, src_stride, ref, ref_stride, x, y, w, h);
}
}
for (; y < h1; ++y) {
for (x = 0; x < w0; ++x) {
sum += VP8SSIMGetClipped(src, src_stride, ref, ref_stride, x, y, w, h);
}
for (; x < w1; ++x) {
const int off1 = x - VP8_SSIM_KERNEL + (y - VP8_SSIM_KERNEL) * src_stride;
const int off2 = x - VP8_SSIM_KERNEL + (y - VP8_SSIM_KERNEL) * ref_stride;
sum += VP8SSIMGet(src + off1, src_stride, ref + off2, ref_stride);
}
for (; x < w; ++x) {
sum += VP8SSIMGetClipped(src, src_stride, ref, ref_stride, x, y, w, h);
}
}
for (; y < h; ++y) {
for (x = 0; x < w; ++x) {
sum += VP8SSIMGetClipped(src, src_stride, ref, ref_stride, x, y, w, h);
}
}
return sum;
}
//------------------------------------------------------------------------------
// Distortion
// Max value returned in case of exact similarity.
static const double kMinDistortion_dB = 99.;
static double GetPSNR(double v, double size) {
return (v > 0. && size > 0.) ? -4.3429448 * log(v / (size * 255 * 255.))
: kMinDistortion_dB;
}
static double GetLogSSIM(double v, double size) {
v = (size > 0.) ? v / size : 1.;
return (v < 1.) ? -10.0 * log10(1. - v) : kMinDistortion_dB;
}
int WebPPlaneDistortion(const uint8_t* src, size_t src_stride,
const uint8_t* ref, size_t ref_stride,
int width, int height, size_t x_step,
int type, float* distortion, float* result) {
uint8_t* allocated = NULL;
const AccumulateFunc metric = (type == 0) ? AccumulateSSE :
(type == 1) ? AccumulateSSIM :
AccumulateLSIM;
if (src == NULL || ref == NULL ||
src_stride < x_step * width || ref_stride < x_step * width ||
result == NULL || distortion == NULL) {
return 0;
}
VP8SSIMDspInit();
if (x_step != 1) { // extract a packed plane if needed
int x, y;
uint8_t* tmp1;
uint8_t* tmp2;
allocated =
(uint8_t*)WebPSafeMalloc(2ULL * width * height, sizeof(*allocated));
if (allocated == NULL) return 0;
tmp1 = allocated;
tmp2 = tmp1 + (size_t)width * height;
for (y = 0; y < height; ++y) {
for (x = 0; x < width; ++x) {
tmp1[x + y * width] = src[x * x_step + y * src_stride];
tmp2[x + y * width] = ref[x * x_step + y * ref_stride];
}
}
src = tmp1;
ref = tmp2;
}
*distortion = (float)metric(src, width, ref, width, width, height);
WebPSafeFree(allocated);
*result = (type == 1) ? (float)GetLogSSIM(*distortion, (double)width * height)
: (float)GetPSNR(*distortion, (double)width * height);
return 1;
}
int WebPPictureDistortion(const WebPPicture* src, const WebPPicture* ref,
int type, float results[5]) {
int w, h, c;
int ok = 0;
WebPPicture p0, p1;
double total_size = 0., total_distortion = 0.;
if (src == NULL || ref == NULL ||
src->width != ref->width || src->height != ref->height ||
results == NULL) {
return 0;
}
VP8SSIMDspInit();
if (!WebPPictureInit(&p0) || !WebPPictureInit(&p1)) return 0;
w = src->width;
h = src->height;
if (!WebPPictureView(src, 0, 0, w, h, &p0)) goto Error;
if (!WebPPictureView(ref, 0, 0, w, h, &p1)) goto Error;
// We always measure distortion in ARGB space.
if (p0.use_argb == 0 && !WebPPictureYUVAToARGB(&p0)) goto Error;
if (p1.use_argb == 0 && !WebPPictureYUVAToARGB(&p1)) goto Error;
for (c = 0; c < 4; ++c) {
float distortion;
const size_t stride0 = 4 * (size_t)p0.argb_stride;
const size_t stride1 = 4 * (size_t)p1.argb_stride;
if (!WebPPlaneDistortion((const uint8_t*)p0.argb + c, stride0,
(const uint8_t*)p1.argb + c, stride1,
w, h, 4, type, &distortion, results + c)) {
goto Error;
}
total_distortion += distortion;
total_size += w * h;
}
results[4] = (type == 1) ? (float)GetLogSSIM(total_distortion, total_size)
: (float)GetPSNR(total_distortion, total_size);
ok = 1;
Error:
WebPPictureFree(&p0);
WebPPictureFree(&p1);
return ok;
}
//------------------------------------------------------------------------------

4
thirdparty/libwebp/enc/picture_rescale.c → thirdparty/libwebp/enc/picture_rescale_enc.c vendored
View file

@ -14,8 +14,8 @@
#include <assert.h>
#include <stdlib.h>
#include "./vp8enci.h"
#include "../utils/rescaler.h"
#include "./vp8i_enc.h"
#include "../utils/rescaler_utils.h"
#include "../utils/utils.h"
#define HALVE(x) (((x) + 1) >> 1)

2
thirdparty/libwebp/enc/picture_tools.c → thirdparty/libwebp/enc/picture_tools_enc.c vendored
View file

@ -13,7 +13,7 @@
#include <assert.h>
#include "./vp8enci.h"
#include "./vp8i_enc.h"
#include "../dsp/yuv.h"
static WEBP_INLINE uint32_t MakeARGB32(int r, int g, int b) {

750
thirdparty/libwebp/enc/predictor_enc.c vendored Normal file
View file

@ -0,0 +1,750 @@
// Copyright 2016 Google Inc. All Rights Reserved.
//
// Use of this source code is governed by a BSD-style license
// that can be found in the COPYING file in the root of the source
// tree. An additional intellectual property rights grant can be found
// in the file PATENTS. All contributing project authors may
// be found in the AUTHORS file in the root of the source tree.
// -----------------------------------------------------------------------------
//
// Image transform methods for lossless encoder.
//
// Authors: Vikas Arora (vikaas.arora@gmail.com)
// Jyrki Alakuijala (jyrki@google.com)
// Urvang Joshi (urvang@google.com)
// Vincent Rabaud (vrabaud@google.com)
#include "../dsp/lossless.h"
#include "../dsp/lossless_common.h"
#include "./vp8li_enc.h"
#define MAX_DIFF_COST (1e30f)
static const float kSpatialPredictorBias = 15.f;
static const int kPredLowEffort = 11;
static const uint32_t kMaskAlpha = 0xff000000;
// Mostly used to reduce code size + readability
static WEBP_INLINE int GetMin(int a, int b) { return (a > b) ? b : a; }
static WEBP_INLINE int GetMax(int a, int b) { return (a < b) ? b : a; }
//------------------------------------------------------------------------------
// Methods to calculate Entropy (Shannon).
static float PredictionCostSpatial(const int counts[256], int weight_0,
double exp_val) {
const int significant_symbols = 256 >> 4;
const double exp_decay_factor = 0.6;
double bits = weight_0 * counts[0];
int i;
for (i = 1; i < significant_symbols; ++i) {
bits += exp_val * (counts[i] + counts[256 - i]);
exp_val *= exp_decay_factor;
}
return (float)(-0.1 * bits);
}
static float PredictionCostSpatialHistogram(const int accumulated[4][256],
const int tile[4][256]) {
int i;
double retval = 0;
for (i = 0; i < 4; ++i) {
const double kExpValue = 0.94;
retval += PredictionCostSpatial(tile[i], 1, kExpValue);
retval += VP8LCombinedShannonEntropy(tile[i], accumulated[i]);
}
return (float)retval;
}
static WEBP_INLINE void UpdateHisto(int histo_argb[4][256], uint32_t argb) {
++histo_argb[0][argb >> 24];
++histo_argb[1][(argb >> 16) & 0xff];
++histo_argb[2][(argb >> 8) & 0xff];
++histo_argb[3][argb & 0xff];
}
//------------------------------------------------------------------------------
// Spatial transform functions.
static WEBP_INLINE void PredictBatch(int mode, int x_start, int y,
int num_pixels, const uint32_t* current,
const uint32_t* upper, uint32_t* out) {
if (x_start == 0) {
if (y == 0) {
// ARGB_BLACK.
VP8LPredictorsSub[0](current, NULL, 1, out);
} else {
// Top one.
VP8LPredictorsSub[2](current, upper, 1, out);
}
++x_start;
++out;
--num_pixels;
}
if (y == 0) {
// Left one.
VP8LPredictorsSub[1](current + x_start, NULL, num_pixels, out);
} else {
VP8LPredictorsSub[mode](current + x_start, upper + x_start, num_pixels,
out);
}
}
static int MaxDiffBetweenPixels(uint32_t p1, uint32_t p2) {
const int diff_a = abs((int)(p1 >> 24) - (int)(p2 >> 24));
const int diff_r = abs((int)((p1 >> 16) & 0xff) - (int)((p2 >> 16) & 0xff));
const int diff_g = abs((int)((p1 >> 8) & 0xff) - (int)((p2 >> 8) & 0xff));
const int diff_b = abs((int)(p1 & 0xff) - (int)(p2 & 0xff));
return GetMax(GetMax(diff_a, diff_r), GetMax(diff_g, diff_b));
}
static int MaxDiffAroundPixel(uint32_t current, uint32_t up, uint32_t down,
uint32_t left, uint32_t right) {
const int diff_up = MaxDiffBetweenPixels(current, up);
const int diff_down = MaxDiffBetweenPixels(current, down);
const int diff_left = MaxDiffBetweenPixels(current, left);
const int diff_right = MaxDiffBetweenPixels(current, right);
return GetMax(GetMax(diff_up, diff_down), GetMax(diff_left, diff_right));
}
static uint32_t AddGreenToBlueAndRed(uint32_t argb) {
const uint32_t green = (argb >> 8) & 0xff;
uint32_t red_blue = argb & 0x00ff00ffu;
red_blue += (green << 16) | green;
red_blue &= 0x00ff00ffu;
return (argb & 0xff00ff00u) | red_blue;
}
static void MaxDiffsForRow(int width, int stride, const uint32_t* const argb,
uint8_t* const max_diffs, int used_subtract_green) {
uint32_t current, up, down, left, right;
int x;
if (width <= 2) return;
current = argb[0];
right = argb[1];
if (used_subtract_green) {
current = AddGreenToBlueAndRed(current);
right = AddGreenToBlueAndRed(right);
}
// max_diffs[0] and max_diffs[width - 1] are never used.
for (x = 1; x < width - 1; ++x) {
up = argb[-stride + x];
down = argb[stride + x];
left = current;
current = right;
right = argb[x + 1];
if (used_subtract_green) {
up = AddGreenToBlueAndRed(up);
down = AddGreenToBlueAndRed(down);
right = AddGreenToBlueAndRed(right);
}
max_diffs[x] = MaxDiffAroundPixel(current, up, down, left, right);
}
}
// Quantize the difference between the actual component value and its prediction
// to a multiple of quantization, working modulo 256, taking care not to cross
// a boundary (inclusive upper limit).
static uint8_t NearLosslessComponent(uint8_t value, uint8_t predict,
uint8_t boundary, int quantization) {
const int residual = (value - predict) & 0xff;
const int boundary_residual = (boundary - predict) & 0xff;
const int lower = residual & ~(quantization - 1);
const int upper = lower + quantization;
// Resolve ties towards a value closer to the prediction (i.e. towards lower
// if value comes after prediction and towards upper otherwise).
const int bias = ((boundary - value) & 0xff) < boundary_residual;
if (residual - lower < upper - residual + bias) {
// lower is closer to residual than upper.
if (residual > boundary_residual && lower <= boundary_residual) {
// Halve quantization step to avoid crossing boundary. This midpoint is
// on the same side of boundary as residual because midpoint >= residual
// (since lower is closer than upper) and residual is above the boundary.
return lower + (quantization >> 1);
}
return lower;
} else {
// upper is closer to residual than lower.
if (residual <= boundary_residual && upper > boundary_residual) {
// Halve quantization step to avoid crossing boundary. This midpoint is
// on the same side of boundary as residual because midpoint <= residual
// (since upper is closer than lower) and residual is below the boundary.
return lower + (quantization >> 1);
}
return upper & 0xff;
}
}
// Quantize every component of the difference between the actual pixel value and
// its prediction to a multiple of a quantization (a power of 2, not larger than
// max_quantization which is a power of 2, smaller than max_diff). Take care if
// value and predict have undergone subtract green, which means that red and
// blue are represented as offsets from green.
static uint32_t NearLossless(uint32_t value, uint32_t predict,
int max_quantization, int max_diff,
int used_subtract_green) {
int quantization;
uint8_t new_green = 0;
uint8_t green_diff = 0;
uint8_t a, r, g, b;
if (max_diff <= 2) {
return VP8LSubPixels(value, predict);
}
quantization = max_quantization;
while (quantization >= max_diff) {
quantization >>= 1;
}
if ((value >> 24) == 0 || (value >> 24) == 0xff) {
// Preserve transparency of fully transparent or fully opaque pixels.
a = ((value >> 24) - (predict >> 24)) & 0xff;
} else {
a = NearLosslessComponent(value >> 24, predict >> 24, 0xff, quantization);
}
g = NearLosslessComponent((value >> 8) & 0xff, (predict >> 8) & 0xff, 0xff,
quantization);
if (used_subtract_green) {
// The green offset will be added to red and blue components during decoding
// to obtain the actual red and blue values.
new_green = ((predict >> 8) + g) & 0xff;
// The amount by which green has been adjusted during quantization. It is
// subtracted from red and blue for compensation, to avoid accumulating two
// quantization errors in them.
green_diff = (new_green - (value >> 8)) & 0xff;
}
r = NearLosslessComponent(((value >> 16) - green_diff) & 0xff,
(predict >> 16) & 0xff, 0xff - new_green,
quantization);
b = NearLosslessComponent((value - green_diff) & 0xff, predict & 0xff,
0xff - new_green, quantization);
return ((uint32_t)a << 24) | ((uint32_t)r << 16) | ((uint32_t)g << 8) | b;
}
// Stores the difference between the pixel and its prediction in "out".
// In case of a lossy encoding, updates the source image to avoid propagating
// the deviation further to pixels which depend on the current pixel for their
// predictions.
static WEBP_INLINE void GetResidual(
int width, int height, uint32_t* const upper_row,
uint32_t* const current_row, const uint8_t* const max_diffs, int mode,
int x_start, int x_end, int y, int max_quantization, int exact,
int used_subtract_green, uint32_t* const out) {
if (exact) {
PredictBatch(mode, x_start, y, x_end - x_start, current_row, upper_row,
out);
} else {
const VP8LPredictorFunc pred_func = VP8LPredictors[mode];
int x;
for (x = x_start; x < x_end; ++x) {
uint32_t predict;
uint32_t residual;
if (y == 0) {
predict = (x == 0) ? ARGB_BLACK : current_row[x - 1]; // Left.
} else if (x == 0) {
predict = upper_row[x]; // Top.
} else {
predict = pred_func(current_row[x - 1], upper_row + x);
}
if (max_quantization == 1 || mode == 0 || y == 0 || y == height - 1 ||
x == 0 || x == width - 1) {
residual = VP8LSubPixels(current_row[x], predict);
} else {
residual = NearLossless(current_row[x], predict, max_quantization,
max_diffs[x], used_subtract_green);
// Update the source image.
current_row[x] = VP8LAddPixels(predict, residual);
// x is never 0 here so we do not need to update upper_row like below.
}
if ((current_row[x] & kMaskAlpha) == 0) {
// If alpha is 0, cleanup RGB. We can choose the RGB values of the
// residual for best compression. The prediction of alpha itself can be
// non-zero and must be kept though. We choose RGB of the residual to be
// 0.
residual &= kMaskAlpha;
// Update the source image.
current_row[x] = predict & ~kMaskAlpha;
// The prediction for the rightmost pixel in a row uses the leftmost
// pixel
// in that row as its top-right context pixel. Hence if we change the
// leftmost pixel of current_row, the corresponding change must be
// applied
// to upper_row as well where top-right context is being read from.
if (x == 0 && y != 0) upper_row[width] = current_row[0];
}
out[x - x_start] = residual;
}
}
}
// Returns best predictor and updates the accumulated histogram.
// If max_quantization > 1, assumes that near lossless processing will be
// applied, quantizing residuals to multiples of quantization levels up to
// max_quantization (the actual quantization level depends on smoothness near
// the given pixel).
static int GetBestPredictorForTile(int width, int height,
int tile_x, int tile_y, int bits,
int accumulated[4][256],
uint32_t* const argb_scratch,
const uint32_t* const argb,
int max_quantization,
int exact, int used_subtract_green,
const uint32_t* const modes) {
const int kNumPredModes = 14;
const int start_x = tile_x << bits;
const int start_y = tile_y << bits;
const int tile_size = 1 << bits;
const int max_y = GetMin(tile_size, height - start_y);
const int max_x = GetMin(tile_size, width - start_x);
// Whether there exist columns just outside the tile.
const int have_left = (start_x > 0);
const int have_right = (max_x < width - start_x);
// Position and size of the strip covering the tile and adjacent columns if
// they exist.
const int context_start_x = start_x - have_left;
const int context_width = max_x + have_left + have_right;
const int tiles_per_row = VP8LSubSampleSize(width, bits);
// Prediction modes of the left and above neighbor tiles.
const int left_mode = (tile_x > 0) ?
(modes[tile_y * tiles_per_row + tile_x - 1] >> 8) & 0xff : 0xff;
const int above_mode = (tile_y > 0) ?
(modes[(tile_y - 1) * tiles_per_row + tile_x] >> 8) & 0xff : 0xff;
// The width of upper_row and current_row is one pixel larger than image width
// to allow the top right pixel to point to the leftmost pixel of the next row
// when at the right edge.
uint32_t* upper_row = argb_scratch;
uint32_t* current_row = upper_row + width + 1;
uint8_t* const max_diffs = (uint8_t*)(current_row + width + 1);
float best_diff = MAX_DIFF_COST;
int best_mode = 0;
int mode;
int histo_stack_1[4][256];
int histo_stack_2[4][256];
// Need pointers to be able to swap arrays.
int (*histo_argb)[256] = histo_stack_1;
int (*best_histo)[256] = histo_stack_2;
int i, j;
uint32_t residuals[1 << MAX_TRANSFORM_BITS];
assert(bits <= MAX_TRANSFORM_BITS);
assert(max_x <= (1 << MAX_TRANSFORM_BITS));
for (mode = 0; mode < kNumPredModes; ++mode) {
float cur_diff;
int relative_y;
memset(histo_argb, 0, sizeof(histo_stack_1));
if (start_y > 0) {
// Read the row above the tile which will become the first upper_row.
// Include a pixel to the left if it exists; include a pixel to the right
// in all cases (wrapping to the leftmost pixel of the next row if it does
// not exist).
memcpy(current_row + context_start_x,
argb + (start_y - 1) * width + context_start_x,
sizeof(*argb) * (max_x + have_left + 1));
}
for (relative_y = 0; relative_y < max_y; ++relative_y) {
const int y = start_y + relative_y;
int relative_x;
uint32_t* tmp = upper_row;
upper_row = current_row;
current_row = tmp;
// Read current_row. Include a pixel to the left if it exists; include a
// pixel to the right in all cases except at the bottom right corner of
// the image (wrapping to the leftmost pixel of the next row if it does
// not exist in the current row).
memcpy(current_row + context_start_x,
argb + y * width + context_start_x,
sizeof(*argb) * (max_x + have_left + (y + 1 < height)));
if (max_quantization > 1 && y >= 1 && y + 1 < height) {
MaxDiffsForRow(context_width, width, argb + y * width + context_start_x,
max_diffs + context_start_x, used_subtract_green);
}
GetResidual(width, height, upper_row, current_row, max_diffs, mode,
start_x, start_x + max_x, y, max_quantization, exact,
used_subtract_green, residuals);
for (relative_x = 0; relative_x < max_x; ++relative_x) {
UpdateHisto(histo_argb, residuals[relative_x]);
}
}
cur_diff = PredictionCostSpatialHistogram(
(const int (*)[256])accumulated, (const int (*)[256])histo_argb);
// Favor keeping the areas locally similar.
if (mode == left_mode) cur_diff -= kSpatialPredictorBias;
if (mode == above_mode) cur_diff -= kSpatialPredictorBias;
if (cur_diff < best_diff) {
int (*tmp)[256] = histo_argb;
histo_argb = best_histo;
best_histo = tmp;
best_diff = cur_diff;
best_mode = mode;
}
}
for (i = 0; i < 4; i++) {
for (j = 0; j < 256; j++) {
accumulated[i][j] += best_histo[i][j];
}
}
return best_mode;
}
// Converts pixels of the image to residuals with respect to predictions.
// If max_quantization > 1, applies near lossless processing, quantizing
// residuals to multiples of quantization levels up to max_quantization
// (the actual quantization level depends on smoothness near the given pixel).
static void CopyImageWithPrediction(int width, int height,
int bits, uint32_t* const modes,
uint32_t* const argb_scratch,
uint32_t* const argb,
int low_effort, int max_quantization,
int exact, int used_subtract_green) {
const int tiles_per_row = VP8LSubSampleSize(width, bits);
// The width of upper_row and current_row is one pixel larger than image width
// to allow the top right pixel to point to the leftmost pixel of the next row
// when at the right edge.
uint32_t* upper_row = argb_scratch;
uint32_t* current_row = upper_row + width + 1;
uint8_t* current_max_diffs = (uint8_t*)(current_row + width + 1);
uint8_t* lower_max_diffs = current_max_diffs + width;
int y;
for (y = 0; y < height; ++y) {
int x;
uint32_t* const tmp32 = upper_row;
upper_row = current_row;
current_row = tmp32;
memcpy(current_row, argb + y * width,
sizeof(*argb) * (width + (y + 1 < height)));
if (low_effort) {
PredictBatch(kPredLowEffort, 0, y, width, current_row, upper_row,
argb + y * width);
} else {
if (max_quantization > 1) {
// Compute max_diffs for the lower row now, because that needs the
// contents of argb for the current row, which we will overwrite with
// residuals before proceeding with the next row.
uint8_t* const tmp8 = current_max_diffs;
current_max_diffs = lower_max_diffs;
lower_max_diffs = tmp8;
if (y + 2 < height) {
MaxDiffsForRow(width, width, argb + (y + 1) * width, lower_max_diffs,
used_subtract_green);
}
}
for (x = 0; x < width;) {
const int mode =
(modes[(y >> bits) * tiles_per_row + (x >> bits)] >> 8) & 0xff;
int x_end = x + (1 << bits);
if (x_end > width) x_end = width;
GetResidual(width, height, upper_row, current_row, current_max_diffs,
mode, x, x_end, y, max_quantization, exact,
used_subtract_green, argb + y * width + x);
x = x_end;
}
}
}
}
// Finds the best predictor for each tile, and converts the image to residuals
// with respect to predictions. If near_lossless_quality < 100, applies
// near lossless processing, shaving off more bits of residuals for lower
// qualities.
void VP8LResidualImage(int width, int height, int bits, int low_effort,
uint32_t* const argb, uint32_t* const argb_scratch,
uint32_t* const image, int near_lossless_quality,
int exact, int used_subtract_green) {
const int tiles_per_row = VP8LSubSampleSize(width, bits);
const int tiles_per_col = VP8LSubSampleSize(height, bits);
int tile_y;
int histo[4][256];
const int max_quantization = 1 << VP8LNearLosslessBits(near_lossless_quality);
if (low_effort) {
int i;
for (i = 0; i < tiles_per_row * tiles_per_col; ++i) {
image[i] = ARGB_BLACK | (kPredLowEffort << 8);
}
} else {
memset(histo, 0, sizeof(histo));
for (tile_y = 0; tile_y < tiles_per_col; ++tile_y) {
int tile_x;
for (tile_x = 0; tile_x < tiles_per_row; ++tile_x) {
const int pred = GetBestPredictorForTile(width, height, tile_x, tile_y,
bits, histo, argb_scratch, argb, max_quantization, exact,
used_subtract_green, image);
image[tile_y * tiles_per_row + tile_x] = ARGB_BLACK | (pred << 8);
}
}
}
CopyImageWithPrediction(width, height, bits, image, argb_scratch, argb,
low_effort, max_quantization, exact,
used_subtract_green);
}
//------------------------------------------------------------------------------
// Color transform functions.
static WEBP_INLINE void MultipliersClear(VP8LMultipliers* const m) {
m->green_to_red_ = 0;
m->green_to_blue_ = 0;
m->red_to_blue_ = 0;
}
static WEBP_INLINE void ColorCodeToMultipliers(uint32_t color_code,
VP8LMultipliers* const m) {
m->green_to_red_ = (color_code >> 0) & 0xff;
m->green_to_blue_ = (color_code >> 8) & 0xff;
m->red_to_blue_ = (color_code >> 16) & 0xff;
}
static WEBP_INLINE uint32_t MultipliersToColorCode(
const VP8LMultipliers* const m) {
return 0xff000000u |
((uint32_t)(m->red_to_blue_) << 16) |
((uint32_t)(m->green_to_blue_) << 8) |
m->green_to_red_;
}
static float PredictionCostCrossColor(const int accumulated[256],
const int counts[256]) {
// Favor low entropy, locally and globally.
// Favor small absolute values for PredictionCostSpatial
static const double kExpValue = 2.4;
return VP8LCombinedShannonEntropy(counts, accumulated) +
PredictionCostSpatial(counts, 3, kExpValue);
}
static float GetPredictionCostCrossColorRed(
const uint32_t* argb, int stride, int tile_width, int tile_height,
VP8LMultipliers prev_x, VP8LMultipliers prev_y, int green_to_red,
const int accumulated_red_histo[256]) {
int histo[256] = { 0 };
float cur_diff;
VP8LCollectColorRedTransforms(argb, stride, tile_width, tile_height,
green_to_red, histo);
cur_diff = PredictionCostCrossColor(accumulated_red_histo, histo);
if ((uint8_t)green_to_red == prev_x.green_to_red_) {
cur_diff -= 3; // favor keeping the areas locally similar
}
if ((uint8_t)green_to_red == prev_y.green_to_red_) {
cur_diff -= 3; // favor keeping the areas locally similar
}
if (green_to_red == 0) {
cur_diff -= 3;
}
return cur_diff;
}
static void GetBestGreenToRed(
const uint32_t* argb, int stride, int tile_width, int tile_height,
VP8LMultipliers prev_x, VP8LMultipliers prev_y, int quality,
const int accumulated_red_histo[256], VP8LMultipliers* const best_tx) {
const int kMaxIters = 4 + ((7 * quality) >> 8); // in range [4..6]
int green_to_red_best = 0;
int iter, offset;
float best_diff = GetPredictionCostCrossColorRed(
argb, stride, tile_width, tile_height, prev_x, prev_y,
green_to_red_best, accumulated_red_histo);
for (iter = 0; iter < kMaxIters; ++iter) {
// ColorTransformDelta is a 3.5 bit fixed point, so 32 is equal to
// one in color computation. Having initial delta here as 1 is sufficient
// to explore the range of (-2, 2).
const int delta = 32 >> iter;
// Try a negative and a positive delta from the best known value.
for (offset = -delta; offset <= delta; offset += 2 * delta) {
const int green_to_red_cur = offset + green_to_red_best;
const float cur_diff = GetPredictionCostCrossColorRed(
argb, stride, tile_width, tile_height, prev_x, prev_y,
green_to_red_cur, accumulated_red_histo);
if (cur_diff < best_diff) {
best_diff = cur_diff;
green_to_red_best = green_to_red_cur;
}
}
}
best_tx->green_to_red_ = green_to_red_best;
}
static float GetPredictionCostCrossColorBlue(
const uint32_t* argb, int stride, int tile_width, int tile_height,
VP8LMultipliers prev_x, VP8LMultipliers prev_y,
int green_to_blue, int red_to_blue, const int accumulated_blue_histo[256]) {
int histo[256] = { 0 };
float cur_diff;
VP8LCollectColorBlueTransforms(argb, stride, tile_width, tile_height,
green_to_blue, red_to_blue, histo);
cur_diff = PredictionCostCrossColor(accumulated_blue_histo, histo);
if ((uint8_t)green_to_blue == prev_x.green_to_blue_) {
cur_diff -= 3; // favor keeping the areas locally similar
}
if ((uint8_t)green_to_blue == prev_y.green_to_blue_) {
cur_diff -= 3; // favor keeping the areas locally similar
}
if ((uint8_t)red_to_blue == prev_x.red_to_blue_) {
cur_diff -= 3; // favor keeping the areas locally similar
}
if ((uint8_t)red_to_blue == prev_y.red_to_blue_) {
cur_diff -= 3; // favor keeping the areas locally similar
}
if (green_to_blue == 0) {
cur_diff -= 3;
}
if (red_to_blue == 0) {
cur_diff -= 3;
}
return cur_diff;
}
#define kGreenRedToBlueNumAxis 8
#define kGreenRedToBlueMaxIters 7
static void GetBestGreenRedToBlue(
const uint32_t* argb, int stride, int tile_width, int tile_height,
VP8LMultipliers prev_x, VP8LMultipliers prev_y, int quality,
const int accumulated_blue_histo[256],
VP8LMultipliers* const best_tx) {
const int8_t offset[kGreenRedToBlueNumAxis][2] =
{{0, -1}, {0, 1}, {-1, 0}, {1, 0}, {-1, -1}, {-1, 1}, {1, -1}, {1, 1}};
const int8_t delta_lut[kGreenRedToBlueMaxIters] = { 16, 16, 8, 4, 2, 2, 2 };
const int iters =
(quality < 25) ? 1 : (quality > 50) ? kGreenRedToBlueMaxIters : 4;
int green_to_blue_best = 0;
int red_to_blue_best = 0;
int iter;
// Initial value at origin:
float best_diff = GetPredictionCostCrossColorBlue(
argb, stride, tile_width, tile_height, prev_x, prev_y,
green_to_blue_best, red_to_blue_best, accumulated_blue_histo);
for (iter = 0; iter < iters; ++iter) {
const int delta = delta_lut[iter];
int axis;
for (axis = 0; axis < kGreenRedToBlueNumAxis; ++axis) {
const int green_to_blue_cur =
offset[axis][0] * delta + green_to_blue_best;
const int red_to_blue_cur = offset[axis][1] * delta + red_to_blue_best;
const float cur_diff = GetPredictionCostCrossColorBlue(
argb, stride, tile_width, tile_height, prev_x, prev_y,
green_to_blue_cur, red_to_blue_cur, accumulated_blue_histo);
if (cur_diff < best_diff) {
best_diff = cur_diff;
green_to_blue_best = green_to_blue_cur;
red_to_blue_best = red_to_blue_cur;
}
if (quality < 25 && iter == 4) {
// Only axis aligned diffs for lower quality.
break; // next iter.
}
}
if (delta == 2 && green_to_blue_best == 0 && red_to_blue_best == 0) {
// Further iterations would not help.
break; // out of iter-loop.
}
}
best_tx->green_to_blue_ = green_to_blue_best;
best_tx->red_to_blue_ = red_to_blue_best;
}
#undef kGreenRedToBlueMaxIters
#undef kGreenRedToBlueNumAxis
static VP8LMultipliers GetBestColorTransformForTile(
int tile_x, int tile_y, int bits,
VP8LMultipliers prev_x,
VP8LMultipliers prev_y,
int quality, int xsize, int ysize,
const int accumulated_red_histo[256],
const int accumulated_blue_histo[256],
const uint32_t* const argb) {
const int max_tile_size = 1 << bits;
const int tile_y_offset = tile_y * max_tile_size;
const int tile_x_offset = tile_x * max_tile_size;
const int all_x_max = GetMin(tile_x_offset + max_tile_size, xsize);
const int all_y_max = GetMin(tile_y_offset + max_tile_size, ysize);
const int tile_width = all_x_max - tile_x_offset;
const int tile_height = all_y_max - tile_y_offset;
const uint32_t* const tile_argb = argb + tile_y_offset * xsize
+ tile_x_offset;
VP8LMultipliers best_tx;
MultipliersClear(&best_tx);
GetBestGreenToRed(tile_argb, xsize, tile_width, tile_height,
prev_x, prev_y, quality, accumulated_red_histo, &best_tx);
GetBestGreenRedToBlue(tile_argb, xsize, tile_width, tile_height,
prev_x, prev_y, quality, accumulated_blue_histo,
&best_tx);
return best_tx;
}
static void CopyTileWithColorTransform(int xsize, int ysize,
int tile_x, int tile_y,
int max_tile_size,
VP8LMultipliers color_transform,
uint32_t* argb) {
const int xscan = GetMin(max_tile_size, xsize - tile_x);
int yscan = GetMin(max_tile_size, ysize - tile_y);
argb += tile_y * xsize + tile_x;
while (yscan-- > 0) {
VP8LTransformColor(&color_transform, argb, xscan);
argb += xsize;
}
}
void VP8LColorSpaceTransform(int width, int height, int bits, int quality,
uint32_t* const argb, uint32_t* image) {
const int max_tile_size = 1 << bits;
const int tile_xsize = VP8LSubSampleSize(width, bits);
const int tile_ysize = VP8LSubSampleSize(height, bits);
int accumulated_red_histo[256] = { 0 };
int accumulated_blue_histo[256] = { 0 };
int tile_x, tile_y;
VP8LMultipliers prev_x, prev_y;
MultipliersClear(&prev_y);
MultipliersClear(&prev_x);
for (tile_y = 0; tile_y < tile_ysize; ++tile_y) {
for (tile_x = 0; tile_x < tile_xsize; ++tile_x) {
int y;
const int tile_x_offset = tile_x * max_tile_size;
const int tile_y_offset = tile_y * max_tile_size;
const int all_x_max = GetMin(tile_x_offset + max_tile_size, width);
const int all_y_max = GetMin(tile_y_offset + max_tile_size, height);
const int offset = tile_y * tile_xsize + tile_x;
if (tile_y != 0) {
ColorCodeToMultipliers(image[offset - tile_xsize], &prev_y);
}
prev_x = GetBestColorTransformForTile(tile_x, tile_y, bits,
prev_x, prev_y,
quality, width, height,
accumulated_red_histo,
accumulated_blue_histo,
argb);
image[offset] = MultipliersToColorCode(&prev_x);
CopyTileWithColorTransform(width, height, tile_x_offset, tile_y_offset,
max_tile_size, prev_x, argb);
// Gather accumulated histogram data.
for (y = tile_y_offset; y < all_y_max; ++y) {
int ix = y * width + tile_x_offset;
const int ix_end = ix + all_x_max - tile_x_offset;
for (; ix < ix_end; ++ix) {
const uint32_t pix = argb[ix];
if (ix >= 2 &&
pix == argb[ix - 2] &&
pix == argb[ix - 1]) {
continue; // repeated pixels are handled by backward references
}
if (ix >= width + 2 &&
argb[ix - 2] == argb[ix - width - 2] &&
argb[ix - 1] == argb[ix - width - 1] &&
pix == argb[ix - width]) {
continue; // repeated pixels are handled by backward references
}
++accumulated_red_histo[(pix >> 16) & 0xff];
++accumulated_blue_histo[(pix >> 0) & 0xff];
}
}
}
}
}

21
thirdparty/libwebp/enc/quant.c → thirdparty/libwebp/enc/quant_enc.c vendored
View file

@ -15,8 +15,8 @@
#include <math.h>
#include <stdlib.h> // for abs()
#include "./vp8enci.h"
#include "./cost.h"
#include "./vp8i_enc.h"
#include "./cost_enc.h"
#define DO_TRELLIS_I4 1
#define DO_TRELLIS_I16 1 // not a huge gain, but ok at low bitrate.
@ -643,6 +643,8 @@ static int TrellisQuantizeBlock(const VP8Encoder* const enc,
const int sign = (in[j] < 0);
const uint32_t coeff0 = (sign ? -in[j] : in[j]) + mtx->sharpen_[j];
int level0 = QUANTDIV(coeff0, iQ, B);
int thresh_level = QUANTDIV(coeff0, iQ, BIAS(0x80));
if (thresh_level > MAX_LEVEL) thresh_level = MAX_LEVEL;
if (level0 > MAX_LEVEL) level0 = MAX_LEVEL;
{ // Swap current and previous score states
@ -657,23 +659,17 @@ static int TrellisQuantizeBlock(const VP8Encoder* const enc,
int level = level0 + m;
const int ctx = (level > 2) ? 2 : level;
const int band = VP8EncBands[n + 1];
score_t base_score, last_pos_score;
score_t base_score;
score_t best_cur_score = MAX_COST;
int best_prev = 0; // default, in case
ss_cur[m].score = MAX_COST;
ss_cur[m].costs = costs[n + 1][ctx];
if (level > MAX_LEVEL || level < 0) { // node is dead?
if (level < 0 || level > thresh_level) {
// Node is dead.
continue;
}
// Compute extra rate cost if last coeff's position is < 15
{
const score_t last_pos_cost =
(n < 15) ? VP8BitCost(0, probas[band][ctx][0]) : 0;
last_pos_score = RDScoreTrellis(lambda, last_pos_cost, 0);
}
{
// Compute delta_error = how much coding this level will
// subtract to max_error as distortion.
@ -705,6 +701,9 @@ static int TrellisQuantizeBlock(const VP8Encoder* const enc,
// Now, record best terminal node (and thus best entry in the graph).
if (level != 0) {
const score_t last_pos_cost =
(n < 15) ? VP8BitCost(0, probas[band][ctx][0]) : 0;
const score_t last_pos_score = RDScoreTrellis(lambda, last_pos_cost, 0);
const score_t score = best_cur_score + last_pos_score;
if (score < best_score) {
best_score = score;

5
thirdparty/libwebp/enc/syntax.c → thirdparty/libwebp/enc/syntax_enc.c vendored
View file

@ -16,7 +16,7 @@
#include "../utils/utils.h"
#include "../webp/format_constants.h" // RIFF constants
#include "../webp/mux_types.h" // ALPHA_FLAG
#include "./vp8enci.h"
#include "./vp8i_enc.h"
//------------------------------------------------------------------------------
// Helper functions
@ -362,8 +362,7 @@ int VP8EncWrite(VP8Encoder* const enc) {
for (p = 0; p < enc->num_parts_; ++p) {
const uint8_t* const buf = VP8BitWriterBuf(enc->parts_ + p);
const size_t size = VP8BitWriterSize(enc->parts_ + p);
if (size)
ok = ok && pic->writer(buf, size, pic);
if (size) ok = ok && pic->writer(buf, size, pic);
VP8BitWriterWipeOut(enc->parts_ + p); // will free the internal buffer.
ok = ok && WebPReportProgress(pic, enc->percent_ + percent_per_part,
&enc->percent_);

7
thirdparty/libwebp/enc/token.c → thirdparty/libwebp/enc/token_enc.c vendored
View file

@ -20,8 +20,8 @@
#include <stdlib.h>
#include <string.h>
#include "./cost.h"
#include "./vp8enci.h"
#include "./cost_enc.h"
#include "./vp8i_enc.h"
#include "../utils/utils.h"
#if !defined(DISABLE_TOKEN_BUFFER)
@ -137,8 +137,9 @@ int VP8RecordCoeffTokens(int ctx, const struct VP8Residual* const res,
s = res->stats[VP8EncBands[n]][1];
} else {
if (!AddToken(tokens, v > 4, base_id + 3, s + 3)) {
if (AddToken(tokens, v != 2, base_id + 4, s + 4))
if (AddToken(tokens, v != 2, base_id + 4, s + 4)) {
AddToken(tokens, v == 4, base_id + 5, s + 5);
}
} else if (!AddToken(tokens, v > 10, base_id + 6, s + 6)) {
if (!AddToken(tokens, v > 6, base_id + 7, s + 7)) {
AddConstantToken(tokens, v == 6, 159);

2
thirdparty/libwebp/enc/tree.c → thirdparty/libwebp/enc/tree_enc.c vendored
View file

@ -11,7 +11,7 @@
//
// Author: Skal (pascal.massimino@gmail.com)
#include "./vp8enci.h"
#include "./vp8i_enc.h"
//------------------------------------------------------------------------------
// Default probabilities

19
thirdparty/libwebp/enc/vp8enci.h → thirdparty/libwebp/enc/vp8i_enc.h vendored
View file

@ -15,10 +15,10 @@
#define WEBP_ENC_VP8ENCI_H_
#include <string.h> // for memcpy()
#include "../dec/common.h"
#include "../dec/common_dec.h"
#include "../dsp/dsp.h"
#include "../utils/bit_writer.h"
#include "../utils/thread.h"
#include "../utils/bit_writer_utils.h"
#include "../utils/thread_utils.h"
#include "../utils/utils.h"
#include "../webp/encode.h"
@ -31,8 +31,8 @@ extern "C" {
// version numbers
#define ENC_MAJ_VERSION 0
#define ENC_MIN_VERSION 5
#define ENC_REV_VERSION 2
#define ENC_MIN_VERSION 6
#define ENC_REV_VERSION 0
enum { MAX_LF_LEVELS = 64, // Maximum loop filter level
MAX_VARIABLE_LEVEL = 67, // last (inclusive) level with variable cost
@ -219,7 +219,6 @@ typedef struct {
// right neighbouring data (samples, predictions, contexts, ...)
typedef struct {
int x_, y_; // current macroblock
int y_stride_, uv_stride_; // respective strides
uint8_t* yuv_in_; // input samples
uint8_t* yuv_out_; // output samples
uint8_t* yuv_out2_; // secondary buffer swapped with yuv_out_.
@ -474,14 +473,6 @@ int VP8EncStartAlpha(VP8Encoder* const enc); // start alpha coding process
int VP8EncFinishAlpha(VP8Encoder* const enc); // finalize compressed data
int VP8EncDeleteAlpha(VP8Encoder* const enc); // delete compressed data
// in filter.c
void VP8SSIMAddStats(const VP8DistoStats* const src, VP8DistoStats* const dst);
void VP8SSIMAccumulatePlane(const uint8_t* src1, int stride1,
const uint8_t* src2, int stride2,
int W, int H, VP8DistoStats* const stats);
double VP8SSIMGet(const VP8DistoStats* const stats);
double VP8SSIMGetSquaredError(const VP8DistoStats* const stats);
// autofilter
void VP8InitFilter(VP8EncIterator* const it);
void VP8StoreFilterStats(VP8EncIterator* const it);

255
thirdparty/libwebp/enc/vp8l.c → thirdparty/libwebp/enc/vp8l_enc.c vendored
View file

@ -15,17 +15,18 @@
#include <assert.h>
#include <stdlib.h>
#include "./backward_references.h"
#include "./histogram.h"
#include "./vp8enci.h"
#include "./vp8li.h"
#include "./backward_references_enc.h"
#include "./histogram_enc.h"
#include "./vp8i_enc.h"
#include "./vp8li_enc.h"
#include "../dsp/lossless.h"
#include "../utils/bit_writer.h"
#include "../utils/huffman_encode.h"
#include "../dsp/lossless_common.h"
#include "../utils/bit_writer_utils.h"
#include "../utils/huffman_encode_utils.h"
#include "../utils/utils.h"
#include "../webp/format_constants.h"
#include "./delta_palettization.h"
#include "./delta_palettization_enc.h"
#define PALETTE_KEY_RIGHT_SHIFT 22 // Key for 1K buffer.
// Maximum number of histogram images (sub-blocks).
@ -163,18 +164,25 @@ typedef enum {
kHistoTotal // Must be last.
} HistoIx;
static void AddSingleSubGreen(uint32_t p, uint32_t* r, uint32_t* b) {
const uint32_t green = p >> 8; // The upper bits are masked away later.
static void AddSingleSubGreen(int p, uint32_t* const r, uint32_t* const b) {
const int green = p >> 8; // The upper bits are masked away later.
++r[((p >> 16) - green) & 0xff];
++b[(p - green) & 0xff];
++b[((p >> 0) - green) & 0xff];
}
static void AddSingle(uint32_t p,
uint32_t* a, uint32_t* r, uint32_t* g, uint32_t* b) {
++a[p >> 24];
uint32_t* const a, uint32_t* const r,
uint32_t* const g, uint32_t* const b) {
++a[(p >> 24) & 0xff];
++r[(p >> 16) & 0xff];
++g[(p >> 8) & 0xff];
++b[(p & 0xff)];
++g[(p >> 8) & 0xff];
++b[(p >> 0) & 0xff];
}
static WEBP_INLINE uint32_t HashPix(uint32_t pix) {
// Note that masking with 0xffffffffu is for preventing an
// 'unsigned int overflow' warning. Doesn't impact the compiled code.
return ((((uint64_t)pix + (pix >> 19)) * 0x39c5fba7ull) & 0xffffffffu) >> 24;
}
static int AnalyzeEntropy(const uint32_t* argb,
@ -214,8 +222,8 @@ static int AnalyzeEntropy(const uint32_t* argb,
&histo[kHistoBluePredSubGreen * 256]);
{
// Approximate the palette by the entropy of the multiplicative hash.
const int hash = ((pix + (pix >> 19)) * 0x39c5fba7) >> 24;
++histo[kHistoPalette * 256 + (hash & 0xff)];
const uint32_t hash = HashPix(pix);
++histo[kHistoPalette * 256 + hash];
}
}
prev_row = curr_row;
@ -311,7 +319,10 @@ static int GetHistoBits(int method, int use_palette, int width, int height) {
static int GetTransformBits(int method, int histo_bits) {
const int max_transform_bits = (method < 4) ? 6 : (method > 4) ? 4 : 5;
return (histo_bits > max_transform_bits) ? max_transform_bits : histo_bits;
const int res =
(histo_bits > max_transform_bits) ? max_transform_bits : histo_bits;
assert(res <= MAX_TRANSFORM_BITS);
return res;
}
static int AnalyzeAndInit(VP8LEncoder* const enc) {
@ -696,7 +707,7 @@ static WebPEncodingError EncodeImageNoHuffman(VP8LBitWriter* const bw,
VP8LHashChain* const hash_chain,
VP8LBackwardRefs refs_array[2],
int width, int height,
int quality) {
int quality, int low_effort) {
int i;
int max_tokens = 0;
WebPEncodingError err = VP8_ENC_OK;
@ -714,7 +725,8 @@ static WebPEncodingError EncodeImageNoHuffman(VP8LBitWriter* const bw,
}
// Calculate backward references from ARGB image.
if (VP8LHashChainFill(hash_chain, quality, argb, width, height) == 0) {
if (!VP8LHashChainFill(hash_chain, quality, argb, width, height,
low_effort)) {
err = VP8_ENC_ERROR_OUT_OF_MEMORY;
goto Error;
}
@ -814,11 +826,18 @@ static WebPEncodingError EncodeImageInternal(VP8LBitWriter* const bw,
goto Error;
}
*cache_bits = use_cache ? MAX_COLOR_CACHE_BITS : 0;
if (use_cache) {
// If the value is different from zero, it has been set during the
// palette analysis.
if (*cache_bits == 0) *cache_bits = MAX_COLOR_CACHE_BITS;
} else {
*cache_bits = 0;
}
// 'best_refs' is the reference to the best backward refs and points to one
// of refs_array[0] or refs_array[1].
// Calculate backward references from ARGB image.
if (VP8LHashChainFill(hash_chain, quality, argb, width, height) == 0) {
if (!VP8LHashChainFill(hash_chain, quality, argb, width, height,
low_effort)) {
err = VP8_ENC_ERROR_OUT_OF_MEMORY;
goto Error;
}
@ -899,7 +918,7 @@ static WebPEncodingError EncodeImageInternal(VP8LBitWriter* const bw,
err = EncodeImageNoHuffman(bw, histogram_argb, hash_chain, refs_array,
VP8LSubSampleSize(width, histogram_bits),
VP8LSubSampleSize(height, histogram_bits),
quality);
quality, low_effort);
WebPSafeFree(histogram_argb);
if (err != VP8_ENC_OK) goto Error;
}
@ -990,12 +1009,12 @@ static WebPEncodingError ApplyPredictFilter(const VP8LEncoder* const enc,
(VP8LHashChain*)&enc->hash_chain_,
(VP8LBackwardRefs*)enc->refs_, // cast const away
transform_width, transform_height,
quality);
quality, low_effort);
}
static WebPEncodingError ApplyCrossColorFilter(const VP8LEncoder* const enc,
int width, int height,
int quality,
int quality, int low_effort,
VP8LBitWriter* const bw) {
const int ccolor_transform_bits = enc->transform_bits_;
const int transform_width = VP8LSubSampleSize(width, ccolor_transform_bits);
@ -1011,7 +1030,7 @@ static WebPEncodingError ApplyCrossColorFilter(const VP8LEncoder* const enc,
(VP8LHashChain*)&enc->hash_chain_,
(VP8LBackwardRefs*)enc->refs_, // cast const away
transform_width, transform_height,
quality);
quality, low_effort);
}
// -----------------------------------------------------------------------------
@ -1156,7 +1175,8 @@ static WebPEncodingError MakeInputImageCopy(VP8LEncoder* const enc) {
// -----------------------------------------------------------------------------
static int SearchColor(const uint32_t sorted[], uint32_t color, int hi) {
static WEBP_INLINE int SearchColorNoIdx(const uint32_t sorted[], uint32_t color,
int hi) {
int low = 0;
if (sorted[low] == color) return low; // loop invariant: sorted[low] != color
while (1) {
@ -1171,35 +1191,68 @@ static int SearchColor(const uint32_t sorted[], uint32_t color, int hi) {
}
}
#define APPLY_PALETTE_GREEDY_MAX 4
static WEBP_INLINE uint32_t SearchColorGreedy(const uint32_t palette[],
int palette_size,
uint32_t color) {
(void)palette_size;
assert(palette_size < APPLY_PALETTE_GREEDY_MAX);
assert(3 == APPLY_PALETTE_GREEDY_MAX - 1);
if (color == palette[0]) return 0;
if (color == palette[1]) return 1;
if (color == palette[2]) return 2;
return 3;
}
static WEBP_INLINE uint32_t ApplyPaletteHash0(uint32_t color) {
// Focus on the green color.
return (color >> 8) & 0xff;
}
#define PALETTE_INV_SIZE_BITS 11
#define PALETTE_INV_SIZE (1 << PALETTE_INV_SIZE_BITS)
static WEBP_INLINE uint32_t ApplyPaletteHash1(uint32_t color) {
// Forget about alpha.
return ((color & 0x00ffffffu) * 4222244071u) >> (32 - PALETTE_INV_SIZE_BITS);
}
static WEBP_INLINE uint32_t ApplyPaletteHash2(uint32_t color) {
// Forget about alpha.
return (color & 0x00ffffffu) * ((1u << 31) - 1) >>
(32 - PALETTE_INV_SIZE_BITS);
}
// Sort palette in increasing order and prepare an inverse mapping array.
static void PrepareMapToPalette(const uint32_t palette[], int num_colors,
uint32_t sorted[], int idx_map[]) {
uint32_t sorted[], uint32_t idx_map[]) {
int i;
memcpy(sorted, palette, num_colors * sizeof(*sorted));
qsort(sorted, num_colors, sizeof(*sorted), PaletteCompareColorsForQsort);
for (i = 0; i < num_colors; ++i) {
idx_map[SearchColor(sorted, palette[i], num_colors)] = i;
idx_map[SearchColorNoIdx(sorted, palette[i], num_colors)] = i;
}
}
static void MapToPalette(const uint32_t sorted_palette[], int num_colors,
uint32_t* const last_pix, int* const last_idx,
const int idx_map[],
const uint32_t* src, uint8_t* dst, int width) {
int x;
int prev_idx = *last_idx;
uint32_t prev_pix = *last_pix;
for (x = 0; x < width; ++x) {
const uint32_t pix = src[x];
if (pix != prev_pix) {
prev_idx = idx_map[SearchColor(sorted_palette, pix, num_colors)];
prev_pix = pix;
}
dst[x] = prev_idx;
}
*last_idx = prev_idx;
*last_pix = prev_pix;
}
// Use 1 pixel cache for ARGB pixels.
#define APPLY_PALETTE_FOR(COLOR_INDEX) do { \
uint32_t prev_pix = palette[0]; \
uint32_t prev_idx = 0; \
for (y = 0; y < height; ++y) { \
for (x = 0; x < width; ++x) { \
const uint32_t pix = src[x]; \
if (pix != prev_pix) { \
prev_idx = COLOR_INDEX; \
prev_pix = pix; \
} \
tmp_row[x] = prev_idx; \
} \
VP8LBundleColorMap(tmp_row, width, xbits, dst); \
src += src_stride; \
dst += dst_stride; \
} \
} while (0)
// Remap argb values in src[] to packed palettes entries in dst[]
// using 'row' as a temporary buffer of size 'width'.
@ -1212,52 +1265,59 @@ static WebPEncodingError ApplyPalette(const uint32_t* src, uint32_t src_stride,
// TODO(skal): this tmp buffer is not needed if VP8LBundleColorMap() can be
// made to work in-place.
uint8_t* const tmp_row = (uint8_t*)WebPSafeMalloc(width, sizeof(*tmp_row));
int i, x, y;
int use_LUT = 1;
int x, y;
if (tmp_row == NULL) return VP8_ENC_ERROR_OUT_OF_MEMORY;
for (i = 0; i < palette_size; ++i) {
if ((palette[i] & 0xffff00ffu) != 0) {
use_LUT = 0;
break;
}
}
if (use_LUT) {
uint8_t inv_palette[MAX_PALETTE_SIZE] = { 0 };
for (i = 0; i < palette_size; ++i) {
const int color = (palette[i] >> 8) & 0xff;
inv_palette[color] = i;
}
for (y = 0; y < height; ++y) {
for (x = 0; x < width; ++x) {
const int color = (src[x] >> 8) & 0xff;
tmp_row[x] = inv_palette[color];
}
VP8LBundleColorMap(tmp_row, width, xbits, dst);
src += src_stride;
dst += dst_stride;
}
if (palette_size < APPLY_PALETTE_GREEDY_MAX) {
APPLY_PALETTE_FOR(SearchColorGreedy(palette, palette_size, pix));
} else {
// Use 1 pixel cache for ARGB pixels.
uint32_t last_pix;
int last_idx;
uint32_t sorted[MAX_PALETTE_SIZE];
int idx_map[MAX_PALETTE_SIZE];
PrepareMapToPalette(palette, palette_size, sorted, idx_map);
last_pix = palette[0];
last_idx = 0;
for (y = 0; y < height; ++y) {
MapToPalette(sorted, palette_size, &last_pix, &last_idx,
idx_map, src, tmp_row, width);
VP8LBundleColorMap(tmp_row, width, xbits, dst);
src += src_stride;
dst += dst_stride;
int i, j;
uint16_t buffer[PALETTE_INV_SIZE];
uint32_t (*const hash_functions[])(uint32_t) = {
ApplyPaletteHash0, ApplyPaletteHash1, ApplyPaletteHash2
};
// Try to find a perfect hash function able to go from a color to an index
// within 1 << PALETTE_INV_SIZE_BITS in order to build a hash map to go
// from color to index in palette.
for (i = 0; i < 3; ++i) {
int use_LUT = 1;
// Set each element in buffer to max uint16_t.
memset(buffer, 0xff, sizeof(buffer));
for (j = 0; j < palette_size; ++j) {
const uint32_t ind = hash_functions[i](palette[j]);
if (buffer[ind] != 0xffffu) {
use_LUT = 0;
break;
} else {
buffer[ind] = j;
}
}
if (use_LUT) break;
}
if (i == 0) {
APPLY_PALETTE_FOR(buffer[ApplyPaletteHash0(pix)]);
} else if (i == 1) {
APPLY_PALETTE_FOR(buffer[ApplyPaletteHash1(pix)]);
} else if (i == 2) {
APPLY_PALETTE_FOR(buffer[ApplyPaletteHash2(pix)]);
} else {
uint32_t idx_map[MAX_PALETTE_SIZE];
uint32_t palette_sorted[MAX_PALETTE_SIZE];
PrepareMapToPalette(palette, palette_size, palette_sorted, idx_map);
APPLY_PALETTE_FOR(
idx_map[SearchColorNoIdx(palette_sorted, pix, palette_size)]);
}
}
WebPSafeFree(tmp_row);
return VP8_ENC_OK;
}
#undef APPLY_PALETTE_FOR
#undef PALETTE_INV_SIZE_BITS
#undef PALETTE_INV_SIZE
#undef APPLY_PALETTE_GREEDY_MAX
// Note: Expects "enc->palette_" to be set properly.
static WebPEncodingError MapImageFromPalette(VP8LEncoder* const enc,
@ -1290,7 +1350,7 @@ static WebPEncodingError MapImageFromPalette(VP8LEncoder* const enc,
}
// Save palette_[] to bitstream.
static WebPEncodingError EncodePalette(VP8LBitWriter* const bw,
static WebPEncodingError EncodePalette(VP8LBitWriter* const bw, int low_effort,
VP8LEncoder* const enc) {
int i;
uint32_t tmp_palette[MAX_PALETTE_SIZE];
@ -1305,13 +1365,14 @@ static WebPEncodingError EncodePalette(VP8LBitWriter* const bw,
}
tmp_palette[0] = palette[0];
return EncodeImageNoHuffman(bw, tmp_palette, &enc->hash_chain_, enc->refs_,
palette_size, 1, 20 /* quality */);
palette_size, 1, 20 /* quality */, low_effort);
}
#ifdef WEBP_EXPERIMENTAL_FEATURES
static WebPEncodingError EncodeDeltaPalettePredictorImage(
VP8LBitWriter* const bw, VP8LEncoder* const enc, int quality) {
VP8LBitWriter* const bw, VP8LEncoder* const enc, int quality,
int low_effort) {
const WebPPicture* const pic = enc->pic_;
const int width = pic->width;
const int height = pic->height;
@ -1342,7 +1403,7 @@ static WebPEncodingError EncodeDeltaPalettePredictorImage(
err = EncodeImageNoHuffman(bw, predictors, &enc->hash_chain_,
(VP8LBackwardRefs*)enc->refs_, // cast const away
transform_width, transform_height,
quality);
quality, low_effort);
WebPSafeFree(predictors);
return err;
}
@ -1393,7 +1454,7 @@ WebPEncodingError VP8LEncodeStream(const WebPConfig* const config,
int use_near_lossless = 0;
int hdr_size = 0;
int data_size = 0;
int use_delta_palettization = 0;
int use_delta_palette = 0;
if (enc == NULL) {
err = VP8_ENC_ERROR_OUT_OF_MEMORY;
@ -1420,7 +1481,7 @@ WebPEncodingError VP8LEncodeStream(const WebPConfig* const config,
}
#ifdef WEBP_EXPERIMENTAL_FEATURES
if (config->delta_palettization) {
if (config->use_delta_palette) {
enc->use_predict_ = 1;
enc->use_cross_color_ = 0;
enc->use_subtract_green_ = 0;
@ -1432,21 +1493,25 @@ WebPEncodingError VP8LEncodeStream(const WebPConfig* const config,
if (enc->use_palette_) {
err = AllocateTransformBuffer(enc, width, height);
if (err != VP8_ENC_OK) goto Error;
err = EncodeDeltaPalettePredictorImage(bw, enc, quality);
err = EncodeDeltaPalettePredictorImage(bw, enc, quality, low_effort);
if (err != VP8_ENC_OK) goto Error;
use_delta_palettization = 1;
use_delta_palette = 1;
}
}
#endif // WEBP_EXPERIMENTAL_FEATURES
// Encode palette
if (enc->use_palette_) {
err = EncodePalette(bw, enc);
err = EncodePalette(bw, low_effort, enc);
if (err != VP8_ENC_OK) goto Error;
err = MapImageFromPalette(enc, use_delta_palettization);
err = MapImageFromPalette(enc, use_delta_palette);
if (err != VP8_ENC_OK) goto Error;
// If using a color cache, do not have it bigger than the number of colors.
if (use_cache && enc->palette_size_ < (1 << MAX_COLOR_CACHE_BITS)) {
enc->cache_bits_ = BitsLog2Floor(enc->palette_size_) + 1;
}
}
if (!use_delta_palettization) {
if (!use_delta_palette) {
// In case image is not packed.
if (enc->argb_ == NULL) {
err = MakeInputImageCopy(enc);
@ -1468,7 +1533,7 @@ WebPEncodingError VP8LEncodeStream(const WebPConfig* const config,
if (enc->use_cross_color_) {
err = ApplyCrossColorFilter(enc, enc->current_width_,
height, quality, bw);
height, quality, low_effort, bw);
if (err != VP8_ENC_OK) goto Error;
}
}

Some files were not shown because too many files have changed in this diff Show more