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

opus: Packaging fixups after #33311

Browse source 
Removes the custom code for VS2012 support in config.h,
and reduces diff with upstream config.h.

We still have many custom defines for ARM optimizations
which we probably don't use properly, since we don't
compile the included asm code, so a thorough review and
cleanup would be welcome.
This commit is contained in:
Rémi Verschelde 2019-11-12 12:57:27 +01:00
parent 7d836a7cc3
commit 0387657fa4
6 changed files with 568 additions and 478 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

51
thirdparty/opus/celt/fixed_c6x.h → thirdparty/opus/celt/arm/armopts.s vendored
View file

@ -1,8 +1,4 @@
/* Copyright (C) 2008 CSIRO */
/**
@file fixed_c6x.h
@brief Fixed-point operations for the TI C6x DSP family
*/
/* Copyright (C) 2013 Mozilla Corporation */
/*
Redistribution and use in source and binary forms, with or without
modification, are permitted provided that the following conditions
@ -28,43 +24,14 @@
SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*/
#ifndef FIXED_C6X_H
#define FIXED_C6X_H
; Set the following to 1 if we have EDSP instructions
; (LDRD/STRD, etc., ARMv5E and later).
OPUS_ARM_MAY_HAVE_EDSP *
#undef MULT16_16SU
#define MULT16_16SU(a,b) _mpysu(a,b)
; Set the following to 1 if we have ARMv6 media instructions.
OPUS_ARM_MAY_HAVE_MEDIA *
#undef MULT_16_16
#define MULT_16_16(a,b) _mpy(a,b)
; Set the following to 1 if we have NEON (some ARMv7)
OPUS_ARM_MAY_HAVE_NEON *
#define celt_ilog2(x) (30 - _norm(x))
#define OVERRIDE_CELT_ILOG2
#undef MULT16_32_Q15
#define MULT16_32_Q15(a,b) (_mpylill(a, b) >> 15)
#if 0
#include "dsplib.h"
#undef MAX16
#define MAX16(a,b) _max(a,b)
#undef MIN16
#define MIN16(a,b) _min(a,b)
#undef MAX32
#define MAX32(a,b) _lmax(a,b)
#undef MIN32
#define MIN32(a,b) _lmin(a,b)
#undef VSHR32
#define VSHR32(a, shift) _lshl(a,-(shift))
#undef MULT16_16_Q15
#define MULT16_16_Q15(a,b) (_smpy(a,b))
#define celt_maxabs16(x, len) MAX32(EXTEND32(maxval((DATA *)x, len)),-EXTEND32(minval((DATA *)x, len)))
#define OVERRIDE_CELT_MAXABS16
#endif /* FIXED_C6X_H */
END

555
thirdparty/opus/celt/arm/celt_pitch_xcorr_arm-gnu.S vendored Normal file
View file

@ -0,0 +1,555 @@
.syntax unified
@ Copyright (c) 2007-2008 CSIRO
@ Copyright (c) 2007-2009 Xiph.Org Foundation
@ Copyright (c) 2013 Parrot
@ Written by Aurélien Zanelli
@
@ Redistribution and use in source and binary forms, with or without
@ modification, are permitted provided that the following conditions
@ are met:
@
@ - Redistributions of source code must retain the above copyright
@ notice, this list of conditions and the following disclaimer.
@
@ - Redistributions in binary form must reproduce the above copyright
@ notice, this list of conditions and the following disclaimer in the
@ documentation and/or other materials provided with the distribution.
@
@ THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
@ ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
@ LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
@ A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
@ OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
@ EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
@ PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
@ PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
@ LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
@ NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
@ SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
.text; .p2align 2; .arch armv7-a
.fpu neon
.object_arch armv4t
.include "celt/arm/armopts-gnu.S"
.if OPUS_ARM_MAY_HAVE_EDSP
.global celt_pitch_xcorr_edsp
.endif
.if OPUS_ARM_MAY_HAVE_NEON
.global celt_pitch_xcorr_neon
.endif
.if OPUS_ARM_MAY_HAVE_NEON
@ Compute sum[k]=sum(x[j]*y[j+k],j=0...len-1), k=0...3
.type xcorr_kernel_neon, %function; xcorr_kernel_neon: @ PROC
xcorr_kernel_neon_start:
@ input:
@ r3 = int len
@ r4 = opus_val16 *x
@ r5 = opus_val16 *y
@ q0 = opus_val32 sum[4]
@ output:
@ q0 = opus_val32 sum[4]
@ preserved: r0-r3, r6-r11, d2, q4-q7, q9-q15
@ internal usage:
@ r12 = int j
@ d3 = y_3|y_2|y_1|y_0
@ q2 = y_B|y_A|y_9|y_8|y_7|y_6|y_5|y_4
@ q3 = x_7|x_6|x_5|x_4|x_3|x_2|x_1|x_0
@ q8 = scratch
@
@ Load y[0...3]
@ This requires len>0 to always be valid (which we assert in the C code).
VLD1.16 {d5}, [r5]!
SUBS r12, r3, #8
BLE xcorr_kernel_neon_process4
@ Process 8 samples at a time.
@ This loop loads one y value more than we actually need. Therefore we have to
@ stop as soon as there are 8 or fewer samples left (instead of 7), to avoid
@ reading past the end of the array.
xcorr_kernel_neon_process8:
@ This loop has 19 total instructions (10 cycles to issue, minimum), with
@ - 2 cycles of ARM insrtuctions,
@ - 10 cycles of load/store/byte permute instructions, and
@ - 9 cycles of data processing instructions.
@ On a Cortex A8, we dual-issue the maximum amount (9 cycles) between the
@ latter two categories, meaning the whole loop should run in 10 cycles per
@ iteration, barring cache misses.
@
@ Load x[0...7]
VLD1.16 {d6, d7}, [r4]!
@ Unlike VMOV, VAND is a data processsing instruction (and doesn't get
@ assembled to VMOV, like VORR would), so it dual-issues with the prior VLD1.
VAND d3, d5, d5
SUBS r12, r12, #8
@ Load y[4...11]
VLD1.16 {d4, d5}, [r5]!
VMLAL.S16 q0, d3, d6[0]
VEXT.16 d16, d3, d4, #1
VMLAL.S16 q0, d4, d7[0]
VEXT.16 d17, d4, d5, #1
VMLAL.S16 q0, d16, d6[1]
VEXT.16 d16, d3, d4, #2
VMLAL.S16 q0, d17, d7[1]
VEXT.16 d17, d4, d5, #2
VMLAL.S16 q0, d16, d6[2]
VEXT.16 d16, d3, d4, #3
VMLAL.S16 q0, d17, d7[2]
VEXT.16 d17, d4, d5, #3
VMLAL.S16 q0, d16, d6[3]
VMLAL.S16 q0, d17, d7[3]
BGT xcorr_kernel_neon_process8
@ Process 4 samples here if we have > 4 left (still reading one extra y value).
xcorr_kernel_neon_process4:
ADDS r12, r12, #4
BLE xcorr_kernel_neon_process2
@ Load x[0...3]
VLD1.16 d6, [r4]!
@ Use VAND since it's a data processing instruction again.
VAND d4, d5, d5
SUB r12, r12, #4
@ Load y[4...7]
VLD1.16 d5, [r5]!
VMLAL.S16 q0, d4, d6[0]
VEXT.16 d16, d4, d5, #1
VMLAL.S16 q0, d16, d6[1]
VEXT.16 d16, d4, d5, #2
VMLAL.S16 q0, d16, d6[2]
VEXT.16 d16, d4, d5, #3
VMLAL.S16 q0, d16, d6[3]
@ Process 2 samples here if we have > 2 left (still reading one extra y value).
xcorr_kernel_neon_process2:
ADDS r12, r12, #2
BLE xcorr_kernel_neon_process1
@ Load x[0...1]
VLD2.16 {d6[],d7[]}, [r4]!
@ Use VAND since it's a data processing instruction again.
VAND d4, d5, d5
SUB r12, r12, #2
@ Load y[4...5]
VLD1.32 {d5[]}, [r5]!
VMLAL.S16 q0, d4, d6
VEXT.16 d16, d4, d5, #1
@ Replace bottom copy of {y5,y4} in d5 with {y3,y2} from d4, using VSRI
@ instead of VEXT, since it's a data-processing instruction.
VSRI.64 d5, d4, #32
VMLAL.S16 q0, d16, d7
@ Process 1 sample using the extra y value we loaded above.
xcorr_kernel_neon_process1:
@ Load next *x
VLD1.16 {d6[]}, [r4]!
ADDS r12, r12, #1
@ y[0...3] are left in d5 from prior iteration(s) (if any)
VMLAL.S16 q0, d5, d6
MOVLE pc, lr
@ Now process 1 last sample, not reading ahead.
@ Load last *y
VLD1.16 {d4[]}, [r5]!
VSRI.64 d4, d5, #16
@ Load last *x
VLD1.16 {d6[]}, [r4]!
VMLAL.S16 q0, d4, d6
MOV pc, lr
.size xcorr_kernel_neon, .-xcorr_kernel_neon @ ENDP
@ opus_val32 celt_pitch_xcorr_neon(opus_val16 *_x, opus_val16 *_y,
@ opus_val32 *xcorr, int len, int max_pitch, int arch)
.type celt_pitch_xcorr_neon, %function; celt_pitch_xcorr_neon: @ PROC
@ input:
@ r0 = opus_val16 *_x
@ r1 = opus_val16 *_y
@ r2 = opus_val32 *xcorr
@ r3 = int len
@ output:
@ r0 = int maxcorr
@ internal usage:
@ r4 = opus_val16 *x (for xcorr_kernel_neon())
@ r5 = opus_val16 *y (for xcorr_kernel_neon())
@ r6 = int max_pitch
@ r12 = int j
@ q15 = int maxcorr[4] (q15 is not used by xcorr_kernel_neon())
@ ignored:
@ int arch
STMFD sp!, {r4-r6, lr}
LDR r6, [sp, #16]
VMOV.S32 q15, #1
@ if (max_pitch < 4) goto celt_pitch_xcorr_neon_process4_done
SUBS r6, r6, #4
BLT celt_pitch_xcorr_neon_process4_done
celt_pitch_xcorr_neon_process4:
@ xcorr_kernel_neon parameters:
@ r3 = len, r4 = _x, r5 = _y, q0 = {0, 0, 0, 0}
MOV r4, r0
MOV r5, r1
VEOR q0, q0, q0
@ xcorr_kernel_neon only modifies r4, r5, r12, and q0...q3.
@ So we don't save/restore any other registers.
BL xcorr_kernel_neon_start
SUBS r6, r6, #4
VST1.32 {q0}, [r2]!
@ _y += 4
ADD r1, r1, #8
VMAX.S32 q15, q15, q0
@ if (max_pitch < 4) goto celt_pitch_xcorr_neon_process4_done
BGE celt_pitch_xcorr_neon_process4
@ We have less than 4 sums left to compute.
celt_pitch_xcorr_neon_process4_done:
ADDS r6, r6, #4
@ Reduce maxcorr to a single value
VMAX.S32 d30, d30, d31
VPMAX.S32 d30, d30, d30
@ if (max_pitch <= 0) goto celt_pitch_xcorr_neon_done
BLE celt_pitch_xcorr_neon_done
@ Now compute each remaining sum one at a time.
celt_pitch_xcorr_neon_process_remaining:
MOV r4, r0
MOV r5, r1
VMOV.I32 q0, #0
SUBS r12, r3, #8
BLT celt_pitch_xcorr_neon_process_remaining4
@ Sum terms 8 at a time.
celt_pitch_xcorr_neon_process_remaining_loop8:
@ Load x[0...7]
VLD1.16 {q1}, [r4]!
@ Load y[0...7]
VLD1.16 {q2}, [r5]!
SUBS r12, r12, #8
VMLAL.S16 q0, d4, d2
VMLAL.S16 q0, d5, d3
BGE celt_pitch_xcorr_neon_process_remaining_loop8
@ Sum terms 4 at a time.
celt_pitch_xcorr_neon_process_remaining4:
ADDS r12, r12, #4
BLT celt_pitch_xcorr_neon_process_remaining4_done
@ Load x[0...3]
VLD1.16 {d2}, [r4]!
@ Load y[0...3]
VLD1.16 {d3}, [r5]!
SUB r12, r12, #4
VMLAL.S16 q0, d3, d2
celt_pitch_xcorr_neon_process_remaining4_done:
@ Reduce the sum to a single value.
VADD.S32 d0, d0, d1
VPADDL.S32 d0, d0
ADDS r12, r12, #4
BLE celt_pitch_xcorr_neon_process_remaining_loop_done
@ Sum terms 1 at a time.
celt_pitch_xcorr_neon_process_remaining_loop1:
VLD1.16 {d2[]}, [r4]!
VLD1.16 {d3[]}, [r5]!
SUBS r12, r12, #1
VMLAL.S16 q0, d2, d3
BGT celt_pitch_xcorr_neon_process_remaining_loop1
celt_pitch_xcorr_neon_process_remaining_loop_done:
VST1.32 {d0[0]}, [r2]!
VMAX.S32 d30, d30, d0
SUBS r6, r6, #1
@ _y++
ADD r1, r1, #2
@ if (--max_pitch > 0) goto celt_pitch_xcorr_neon_process_remaining
BGT celt_pitch_xcorr_neon_process_remaining
celt_pitch_xcorr_neon_done:
VMOV.32 r0, d30[0]
LDMFD sp!, {r4-r6, pc}
.size celt_pitch_xcorr_neon, .-celt_pitch_xcorr_neon @ ENDP
.endif
.if OPUS_ARM_MAY_HAVE_EDSP
@ This will get used on ARMv7 devices without NEON, so it has been optimized
@ to take advantage of dual-issuing where possible.
.type xcorr_kernel_edsp, %function; xcorr_kernel_edsp: @ PROC
xcorr_kernel_edsp_start:
@ input:
@ r3 = int len
@ r4 = opus_val16 *_x (must be 32-bit aligned)
@ r5 = opus_val16 *_y (must be 32-bit aligned)
@ r6...r9 = opus_val32 sum[4]
@ output:
@ r6...r9 = opus_val32 sum[4]
@ preserved: r0-r5
@ internal usage
@ r2 = int j
@ r12,r14 = opus_val16 x[4]
@ r10,r11 = opus_val16 y[4]
STMFD sp!, {r2,r4,r5,lr}
LDR r10, [r5], #4 @ Load y[0...1]
SUBS r2, r3, #4 @ j = len-4
LDR r11, [r5], #4 @ Load y[2...3]
BLE xcorr_kernel_edsp_process4_done
LDR r12, [r4], #4 @ Load x[0...1]
@ Stall
xcorr_kernel_edsp_process4:
@ The multiplies must issue from pipeline 0, and can't dual-issue with each
@ other. Every other instruction here dual-issues with a multiply, and is
@ thus "free". There should be no stalls in the body of the loop.
SMLABB r6, r12, r10, r6 @ sum[0] = MAC16_16(sum[0],x_0,y_0)
LDR r14, [r4], #4 @ Load x[2...3]
SMLABT r7, r12, r10, r7 @ sum[1] = MAC16_16(sum[1],x_0,y_1)
SUBS r2, r2, #4 @ j-=4
SMLABB r8, r12, r11, r8 @ sum[2] = MAC16_16(sum[2],x_0,y_2)
SMLABT r9, r12, r11, r9 @ sum[3] = MAC16_16(sum[3],x_0,y_3)
SMLATT r6, r12, r10, r6 @ sum[0] = MAC16_16(sum[0],x_1,y_1)
LDR r10, [r5], #4 @ Load y[4...5]
SMLATB r7, r12, r11, r7 @ sum[1] = MAC16_16(sum[1],x_1,y_2)
SMLATT r8, r12, r11, r8 @ sum[2] = MAC16_16(sum[2],x_1,y_3)
SMLATB r9, r12, r10, r9 @ sum[3] = MAC16_16(sum[3],x_1,y_4)
LDRGT r12, [r4], #4 @ Load x[0...1]
SMLABB r6, r14, r11, r6 @ sum[0] = MAC16_16(sum[0],x_2,y_2)
SMLABT r7, r14, r11, r7 @ sum[1] = MAC16_16(sum[1],x_2,y_3)
SMLABB r8, r14, r10, r8 @ sum[2] = MAC16_16(sum[2],x_2,y_4)
SMLABT r9, r14, r10, r9 @ sum[3] = MAC16_16(sum[3],x_2,y_5)
SMLATT r6, r14, r11, r6 @ sum[0] = MAC16_16(sum[0],x_3,y_3)
LDR r11, [r5], #4 @ Load y[6...7]
SMLATB r7, r14, r10, r7 @ sum[1] = MAC16_16(sum[1],x_3,y_4)
SMLATT r8, r14, r10, r8 @ sum[2] = MAC16_16(sum[2],x_3,y_5)
SMLATB r9, r14, r11, r9 @ sum[3] = MAC16_16(sum[3],x_3,y_6)
BGT xcorr_kernel_edsp_process4
xcorr_kernel_edsp_process4_done:
ADDS r2, r2, #4
BLE xcorr_kernel_edsp_done
LDRH r12, [r4], #2 @ r12 = *x++
SUBS r2, r2, #1 @ j--
@ Stall
SMLABB r6, r12, r10, r6 @ sum[0] = MAC16_16(sum[0],x,y_0)
LDRHGT r14, [r4], #2 @ r14 = *x++
SMLABT r7, r12, r10, r7 @ sum[1] = MAC16_16(sum[1],x,y_1)
SMLABB r8, r12, r11, r8 @ sum[2] = MAC16_16(sum[2],x,y_2)
SMLABT r9, r12, r11, r9 @ sum[3] = MAC16_16(sum[3],x,y_3)
BLE xcorr_kernel_edsp_done
SMLABT r6, r14, r10, r6 @ sum[0] = MAC16_16(sum[0],x,y_1)
SUBS r2, r2, #1 @ j--
SMLABB r7, r14, r11, r7 @ sum[1] = MAC16_16(sum[1],x,y_2)
LDRH r10, [r5], #2 @ r10 = y_4 = *y++
SMLABT r8, r14, r11, r8 @ sum[2] = MAC16_16(sum[2],x,y_3)
LDRHGT r12, [r4], #2 @ r12 = *x++
SMLABB r9, r14, r10, r9 @ sum[3] = MAC16_16(sum[3],x,y_4)
BLE xcorr_kernel_edsp_done
SMLABB r6, r12, r11, r6 @ sum[0] = MAC16_16(sum[0],tmp,y_2)
CMP r2, #1 @ j--
SMLABT r7, r12, r11, r7 @ sum[1] = MAC16_16(sum[1],tmp,y_3)
LDRH r2, [r5], #2 @ r2 = y_5 = *y++
SMLABB r8, r12, r10, r8 @ sum[2] = MAC16_16(sum[2],tmp,y_4)
LDRHGT r14, [r4] @ r14 = *x
SMLABB r9, r12, r2, r9 @ sum[3] = MAC16_16(sum[3],tmp,y_5)
BLE xcorr_kernel_edsp_done
SMLABT r6, r14, r11, r6 @ sum[0] = MAC16_16(sum[0],tmp,y_3)
LDRH r11, [r5] @ r11 = y_6 = *y
SMLABB r7, r14, r10, r7 @ sum[1] = MAC16_16(sum[1],tmp,y_4)
SMLABB r8, r14, r2, r8 @ sum[2] = MAC16_16(sum[2],tmp,y_5)
SMLABB r9, r14, r11, r9 @ sum[3] = MAC16_16(sum[3],tmp,y_6)
xcorr_kernel_edsp_done:
LDMFD sp!, {r2,r4,r5,pc}
.size xcorr_kernel_edsp, .-xcorr_kernel_edsp @ ENDP
.type celt_pitch_xcorr_edsp, %function; celt_pitch_xcorr_edsp: @ PROC
@ input:
@ r0 = opus_val16 *_x (must be 32-bit aligned)
@ r1 = opus_val16 *_y (only needs to be 16-bit aligned)
@ r2 = opus_val32 *xcorr
@ r3 = int len
@ output:
@ r0 = maxcorr
@ internal usage
@ r4 = opus_val16 *x
@ r5 = opus_val16 *y
@ r6 = opus_val32 sum0
@ r7 = opus_val32 sum1
@ r8 = opus_val32 sum2
@ r9 = opus_val32 sum3
@ r1 = int max_pitch
@ r12 = int j
@ ignored:
@ int arch
STMFD sp!, {r4-r11, lr}
MOV r5, r1
LDR r1, [sp, #36]
MOV r4, r0
TST r5, #3
@ maxcorr = 1
MOV r0, #1
BEQ celt_pitch_xcorr_edsp_process1u_done
@ Compute one sum at the start to make y 32-bit aligned.
SUBS r12, r3, #4
@ r14 = sum = 0
MOV r14, #0
LDRH r8, [r5], #2
BLE celt_pitch_xcorr_edsp_process1u_loop4_done
LDR r6, [r4], #4
MOV r8, r8, LSL #16
celt_pitch_xcorr_edsp_process1u_loop4:
LDR r9, [r5], #4
SMLABT r14, r6, r8, r14 @ sum = MAC16_16(sum, x_0, y_0)
LDR r7, [r4], #4
SMLATB r14, r6, r9, r14 @ sum = MAC16_16(sum, x_1, y_1)
LDR r8, [r5], #4
SMLABT r14, r7, r9, r14 @ sum = MAC16_16(sum, x_2, y_2)
SUBS r12, r12, #4 @ j-=4
SMLATB r14, r7, r8, r14 @ sum = MAC16_16(sum, x_3, y_3)
LDRGT r6, [r4], #4
BGT celt_pitch_xcorr_edsp_process1u_loop4
MOV r8, r8, LSR #16
celt_pitch_xcorr_edsp_process1u_loop4_done:
ADDS r12, r12, #4
celt_pitch_xcorr_edsp_process1u_loop1:
LDRHGE r6, [r4], #2
@ Stall
SMLABBGE r14, r6, r8, r14 @ sum = MAC16_16(sum, *x, *y)
SUBSGE r12, r12, #1
LDRHGT r8, [r5], #2
BGT celt_pitch_xcorr_edsp_process1u_loop1
@ Restore _x
SUB r4, r4, r3, LSL #1
@ Restore and advance _y
SUB r5, r5, r3, LSL #1
@ maxcorr = max(maxcorr, sum)
CMP r0, r14
ADD r5, r5, #2
MOVLT r0, r14
SUBS r1, r1, #1
@ xcorr[i] = sum
STR r14, [r2], #4
BLE celt_pitch_xcorr_edsp_done
celt_pitch_xcorr_edsp_process1u_done:
@ if (max_pitch < 4) goto celt_pitch_xcorr_edsp_process2
SUBS r1, r1, #4
BLT celt_pitch_xcorr_edsp_process2
celt_pitch_xcorr_edsp_process4:
@ xcorr_kernel_edsp parameters:
@ r3 = len, r4 = _x, r5 = _y, r6...r9 = sum[4] = {0, 0, 0, 0}
MOV r6, #0
MOV r7, #0
MOV r8, #0
MOV r9, #0
BL xcorr_kernel_edsp_start @ xcorr_kernel_edsp(_x, _y+i, xcorr+i, len)
@ maxcorr = max(maxcorr, sum0, sum1, sum2, sum3)
CMP r0, r6
@ _y+=4
ADD r5, r5, #8
MOVLT r0, r6
CMP r0, r7
MOVLT r0, r7
CMP r0, r8
MOVLT r0, r8
CMP r0, r9
MOVLT r0, r9
STMIA r2!, {r6-r9}
SUBS r1, r1, #4
BGE celt_pitch_xcorr_edsp_process4
celt_pitch_xcorr_edsp_process2:
ADDS r1, r1, #2
BLT celt_pitch_xcorr_edsp_process1a
SUBS r12, r3, #4
@ {r10, r11} = {sum0, sum1} = {0, 0}
MOV r10, #0
MOV r11, #0
LDR r8, [r5], #4
BLE celt_pitch_xcorr_edsp_process2_loop_done
LDR r6, [r4], #4
LDR r9, [r5], #4
celt_pitch_xcorr_edsp_process2_loop4:
SMLABB r10, r6, r8, r10 @ sum0 = MAC16_16(sum0, x_0, y_0)
LDR r7, [r4], #4
SMLABT r11, r6, r8, r11 @ sum1 = MAC16_16(sum1, x_0, y_1)
SUBS r12, r12, #4 @ j-=4
SMLATT r10, r6, r8, r10 @ sum0 = MAC16_16(sum0, x_1, y_1)
LDR r8, [r5], #4
SMLATB r11, r6, r9, r11 @ sum1 = MAC16_16(sum1, x_1, y_2)
LDRGT r6, [r4], #4
SMLABB r10, r7, r9, r10 @ sum0 = MAC16_16(sum0, x_2, y_2)
SMLABT r11, r7, r9, r11 @ sum1 = MAC16_16(sum1, x_2, y_3)
SMLATT r10, r7, r9, r10 @ sum0 = MAC16_16(sum0, x_3, y_3)
LDRGT r9, [r5], #4
SMLATB r11, r7, r8, r11 @ sum1 = MAC16_16(sum1, x_3, y_4)
BGT celt_pitch_xcorr_edsp_process2_loop4
celt_pitch_xcorr_edsp_process2_loop_done:
ADDS r12, r12, #2
BLE celt_pitch_xcorr_edsp_process2_1
LDR r6, [r4], #4
@ Stall
SMLABB r10, r6, r8, r10 @ sum0 = MAC16_16(sum0, x_0, y_0)
LDR r9, [r5], #4
SMLABT r11, r6, r8, r11 @ sum1 = MAC16_16(sum1, x_0, y_1)
SUB r12, r12, #2
SMLATT r10, r6, r8, r10 @ sum0 = MAC16_16(sum0, x_1, y_1)
MOV r8, r9
SMLATB r11, r6, r9, r11 @ sum1 = MAC16_16(sum1, x_1, y_2)
celt_pitch_xcorr_edsp_process2_1:
LDRH r6, [r4], #2
ADDS r12, r12, #1
@ Stall
SMLABB r10, r6, r8, r10 @ sum0 = MAC16_16(sum0, x_0, y_0)
LDRHGT r7, [r4], #2
SMLABT r11, r6, r8, r11 @ sum1 = MAC16_16(sum1, x_0, y_1)
BLE celt_pitch_xcorr_edsp_process2_done
LDRH r9, [r5], #2
SMLABT r10, r7, r8, r10 @ sum0 = MAC16_16(sum0, x_0, y_1)
SMLABB r11, r7, r9, r11 @ sum1 = MAC16_16(sum1, x_0, y_2)
celt_pitch_xcorr_edsp_process2_done:
@ Restore _x
SUB r4, r4, r3, LSL #1
@ Restore and advance _y
SUB r5, r5, r3, LSL #1
@ maxcorr = max(maxcorr, sum0)
CMP r0, r10
ADD r5, r5, #2
MOVLT r0, r10
SUB r1, r1, #2
@ maxcorr = max(maxcorr, sum1)
CMP r0, r11
@ xcorr[i] = sum
STR r10, [r2], #4
MOVLT r0, r11
STR r11, [r2], #4
celt_pitch_xcorr_edsp_process1a:
ADDS r1, r1, #1
BLT celt_pitch_xcorr_edsp_done
SUBS r12, r3, #4
@ r14 = sum = 0
MOV r14, #0
BLT celt_pitch_xcorr_edsp_process1a_loop_done
LDR r6, [r4], #4
LDR r8, [r5], #4
LDR r7, [r4], #4
LDR r9, [r5], #4
celt_pitch_xcorr_edsp_process1a_loop4:
SMLABB r14, r6, r8, r14 @ sum = MAC16_16(sum, x_0, y_0)
SUBS r12, r12, #4 @ j-=4
SMLATT r14, r6, r8, r14 @ sum = MAC16_16(sum, x_1, y_1)
LDRGE r6, [r4], #4
SMLABB r14, r7, r9, r14 @ sum = MAC16_16(sum, x_2, y_2)
LDRGE r8, [r5], #4
SMLATT r14, r7, r9, r14 @ sum = MAC16_16(sum, x_3, y_3)
LDRGE r7, [r4], #4
LDRGE r9, [r5], #4
BGE celt_pitch_xcorr_edsp_process1a_loop4
celt_pitch_xcorr_edsp_process1a_loop_done:
ADDS r12, r12, #2
LDRGE r6, [r4], #4
LDRGE r8, [r5], #4
@ Stall
SMLABBGE r14, r6, r8, r14 @ sum = MAC16_16(sum, x_0, y_0)
SUBGE r12, r12, #2
SMLATTGE r14, r6, r8, r14 @ sum = MAC16_16(sum, x_1, y_1)
ADDS r12, r12, #1
LDRHGE r6, [r4], #2
LDRHGE r8, [r5], #2
@ Stall
SMLABBGE r14, r6, r8, r14 @ sum = MAC16_16(sum, *x, *y)
@ maxcorr = max(maxcorr, sum)
CMP r0, r14
@ xcorr[i] = sum
STR r14, [r2], #4
MOVLT r0, r14
celt_pitch_xcorr_edsp_done:
LDMFD sp!, {r4-r11, pc}
.size celt_pitch_xcorr_edsp, .-celt_pitch_xcorr_edsp @ ENDP
.endif
@ END:
.section .note.GNU-stack,"",%progbits

79
thirdparty/opus/celt/fixed_c5x.h vendored
View file

@ -1,79 +0,0 @@
/* Copyright (C) 2003 Jean-Marc Valin */
/**
@file fixed_c5x.h
@brief Fixed-point operations for the TI C5x DSP family
*/
/*
Redistribution and use in source and binary forms, with or without
modification, are permitted provided that the following conditions
are met:
- Redistributions of source code must retain the above copyright
notice, this list of conditions and the following disclaimer.
- Redistributions in binary form must reproduce the above copyright
notice, this list of conditions and the following disclaimer in the
documentation and/or other materials provided with the distribution.
THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*/
#ifndef FIXED_C5X_H
#define FIXED_C5X_H
#include "dsplib.h"
#undef IMUL32
static OPUS_INLINE long IMUL32(long i, long j)
{
long ac0, ac1;
ac0 = _lmpy(i>>16,j);
ac1 = ac0 + _lmpy(i,j>>16);
return _lmpyu(i,j) + (ac1<<16);
}
#undef MAX16
#define MAX16(a,b) _max(a,b)
#undef MIN16
#define MIN16(a,b) _min(a,b)
#undef MAX32
#define MAX32(a,b) _lmax(a,b)
#undef MIN32
#define MIN32(a,b) _lmin(a,b)
#undef VSHR32
#define VSHR32(a, shift) _lshl(a,-(shift))
#undef MULT16_16_Q15
#define MULT16_16_Q15(a,b) (_smpy(a,b))
#undef MULT16_16SU
#define MULT16_16SU(a,b) _lmpysu(a,b)
#undef MULT_16_16
#define MULT_16_16(a,b) _lmpy(a,b)
/* FIXME: This is technically incorrect and is bound to cause problems. Is there any cleaner solution? */
#undef MULT16_32_Q15
#define MULT16_32_Q15(a,b) ADD32(SHL(MULT16_16((a),SHR((b),16)),1), SHR(MULT16_16SU((a),(b)),15))
#define celt_ilog2(x) (30 - _lnorm(x))
#define OVERRIDE_CELT_ILOG2
#define celt_maxabs16(x, len) MAX32(EXTEND32(maxval((DATA *)x, len)),-EXTEND32(minval((DATA *)x, len)))
#define OVERRIDE_CELT_MAXABS16
#endif /* FIXED_C5X_H */

17
thirdparty/opus/config.h vendored
View file

@ -35,7 +35,7 @@
/* #undef FUZZING */
/* Define to 1 if you have the <alloca.h> header file. */
/* #undef HAVE_ALLOCA_H */
/* #undef HAVE_ALLOCA_H */
/* NE10 library is installed on host. Make sure it is on target! */
/* #undef HAVE_ARM_NE10 */
@ -46,16 +46,12 @@
/* Define to 1 if you have the <inttypes.h> header file. */
#define HAVE_INTTYPES_H 1
#if (!defined( _MSC_VER ) || ( _MSC_VER >= 1800 ))
/* Define to 1 if you have the `lrint' function. */
#define HAVE_LRINT 1
/* Define to 1 if you have the `lrintf' function. */
#define HAVE_LRINTF 1
#endif
/* Define to 1 if you have the <memory.h> header file. */
#define HAVE_MEMORY_H 1
@ -83,8 +79,7 @@
/* Define to 1 if you have the `__malloc_hook' function. */
#define HAVE___MALLOC_HOOK 1
/* Define to the sub-directory in which libtool stores uninstalled libraries.
*/
/* Define to the sub-directory where libtool stores uninstalled libraries. */
#define LT_OBJDIR ".libs/"
#ifdef OPUS_ARM_OPT
@ -191,7 +186,7 @@
#define PACKAGE_NAME "opus"
/* Define to the full name and version of this package. */
#define PACKAGE_STRING "opus unknown"
#define PACKAGE_STRING "opus 1.3.1"
/* Define to the one symbol short name of this package. */
#define PACKAGE_TARNAME "opus"
@ -200,7 +195,7 @@
#define PACKAGE_URL ""
/* Define to the version of this package. */
#define PACKAGE_VERSION "unknown"
#define PACKAGE_VERSION "1.3.1"
/* Define to 1 if you have the ANSI C header files. */
#define STDC_HEADERS 1
@ -232,11 +227,7 @@
/* Define to the equivalent of the C99 'restrict' keyword, or to
nothing if this is not supported. Do not define if restrict is
supported directly. */
#if (!defined( _MSC_VER ) || ( _MSC_VER >= 1800 ))
#define restrict __restrict
#else
#undef restrict
#endif
/* Work around a bug in Sun C++: it does not support _Restrict or
__restrict__, even though the corresponding Sun C compiler ends up with
"#define restrict _Restrict" or "#define restrict __restrict__" in the

184
thirdparty/opus/silk/fixed/mips/prefilter_FIX_mipsr1.h vendored
View file

@ -1,184 +0,0 @@
/***********************************************************************
Copyright (c) 2006-2011, Skype Limited. All rights reserved.
Redistribution and use in source and binary forms, with or without
modification, are permitted provided that the following conditions
are met:
- Redistributions of source code must retain the above copyright notice,
this list of conditions and the following disclaimer.
- Redistributions in binary form must reproduce the above copyright
notice, this list of conditions and the following disclaimer in the
documentation and/or other materials provided with the distribution.
- Neither the name of Internet Society, IETF or IETF Trust, nor the
names of specific contributors, may be used to endorse or promote
products derived from this software without specific prior written
permission.
THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
POSSIBILITY OF SUCH DAMAGE.
***********************************************************************/
#ifndef __PREFILTER_FIX_MIPSR1_H__
#define __PREFILTER_FIX_MIPSR1_H__
#ifdef HAVE_CONFIG_H
#include "config.h"
#endif
#include "main_FIX.h"
#include "stack_alloc.h"
#include "tuning_parameters.h"
#define OVERRIDE_silk_warped_LPC_analysis_filter_FIX
void silk_warped_LPC_analysis_filter_FIX(
opus_int32 state[], /* I/O State [order + 1] */
opus_int32 res_Q2[], /* O Residual signal [length] */
const opus_int16 coef_Q13[], /* I Coefficients [order] */
const opus_int16 input[], /* I Input signal [length] */
const opus_int16 lambda_Q16, /* I Warping factor */
const opus_int length, /* I Length of input signal */
const opus_int order, /* I Filter order (even) */
int arch
)
{
opus_int n, i;
opus_int32 acc_Q11, acc_Q22, tmp1, tmp2, tmp3, tmp4;
opus_int32 state_cur, state_next;
(void)arch;
/* Order must be even */
/* Length must be even */
silk_assert( ( order & 1 ) == 0 );
silk_assert( ( length & 1 ) == 0 );
for( n = 0; n < length; n+=2 ) {
/* Output of lowpass section */
tmp2 = silk_SMLAWB( state[ 0 ], state[ 1 ], lambda_Q16 );
state_cur = silk_LSHIFT( input[ n ], 14 );
/* Output of allpass section */
tmp1 = silk_SMLAWB( state[ 1 ], state[ 2 ] - tmp2, lambda_Q16 );
state_next = tmp2;
acc_Q11 = silk_RSHIFT( order, 1 );
acc_Q11 = silk_SMLAWB( acc_Q11, tmp2, coef_Q13[ 0 ] );
/* Output of lowpass section */
tmp4 = silk_SMLAWB( state_cur, state_next, lambda_Q16 );
state[ 0 ] = silk_LSHIFT( input[ n+1 ], 14 );
/* Output of allpass section */
tmp3 = silk_SMLAWB( state_next, tmp1 - tmp4, lambda_Q16 );
state[ 1 ] = tmp4;
acc_Q22 = silk_RSHIFT( order, 1 );
acc_Q22 = silk_SMLAWB( acc_Q22, tmp4, coef_Q13[ 0 ] );
/* Loop over allpass sections */
for( i = 2; i < order; i += 2 ) {
/* Output of allpass section */
tmp2 = silk_SMLAWB( state[ i ], state[ i + 1 ] - tmp1, lambda_Q16 );
state_cur = tmp1;
acc_Q11 = silk_SMLAWB( acc_Q11, tmp1, coef_Q13[ i - 1 ] );
/* Output of allpass section */
tmp1 = silk_SMLAWB( state[ i + 1 ], state[ i + 2 ] - tmp2, lambda_Q16 );
state_next = tmp2;
acc_Q11 = silk_SMLAWB( acc_Q11, tmp2, coef_Q13[ i ] );
/* Output of allpass section */
tmp4 = silk_SMLAWB( state_cur, state_next - tmp3, lambda_Q16 );
state[ i ] = tmp3;
acc_Q22 = silk_SMLAWB( acc_Q22, tmp3, coef_Q13[ i - 1 ] );
/* Output of allpass section */
tmp3 = silk_SMLAWB( state_next, tmp1 - tmp4, lambda_Q16 );
state[ i + 1 ] = tmp4;
acc_Q22 = silk_SMLAWB( acc_Q22, tmp4, coef_Q13[ i ] );
}
acc_Q11 = silk_SMLAWB( acc_Q11, tmp1, coef_Q13[ order - 1 ] );
res_Q2[ n ] = silk_LSHIFT( (opus_int32)input[ n ], 2 ) - silk_RSHIFT_ROUND( acc_Q11, 9 );
state[ order ] = tmp3;
acc_Q22 = silk_SMLAWB( acc_Q22, tmp3, coef_Q13[ order - 1 ] );
res_Q2[ n+1 ] = silk_LSHIFT( (opus_int32)input[ n+1 ], 2 ) - silk_RSHIFT_ROUND( acc_Q22, 9 );
}
}
/* Prefilter for finding Quantizer input signal */
#define OVERRIDE_silk_prefilt_FIX
static inline void silk_prefilt_FIX(
silk_prefilter_state_FIX *P, /* I/O state */
opus_int32 st_res_Q12[], /* I short term residual signal */
opus_int32 xw_Q3[], /* O prefiltered signal */
opus_int32 HarmShapeFIRPacked_Q12, /* I Harmonic shaping coeficients */
opus_int Tilt_Q14, /* I Tilt shaping coeficient */
opus_int32 LF_shp_Q14, /* I Low-frequancy shaping coeficients */
opus_int lag, /* I Lag for harmonic shaping */
opus_int length /* I Length of signals */
)
{
opus_int i, idx, LTP_shp_buf_idx;
opus_int32 n_LTP_Q12, n_Tilt_Q10, n_LF_Q10;
opus_int32 sLF_MA_shp_Q12, sLF_AR_shp_Q12;
opus_int16 *LTP_shp_buf;
/* To speed up use temp variables instead of using the struct */
LTP_shp_buf = P->sLTP_shp;
LTP_shp_buf_idx = P->sLTP_shp_buf_idx;
sLF_AR_shp_Q12 = P->sLF_AR_shp_Q12;
sLF_MA_shp_Q12 = P->sLF_MA_shp_Q12;
if( lag > 0 ) {
for( i = 0; i < length; i++ ) {
/* unrolled loop */
silk_assert( HARM_SHAPE_FIR_TAPS == 3 );
idx = lag + LTP_shp_buf_idx;
n_LTP_Q12 = silk_SMULBB( LTP_shp_buf[ ( idx - HARM_SHAPE_FIR_TAPS / 2 - 1) & LTP_MASK ], HarmShapeFIRPacked_Q12 );
n_LTP_Q12 = silk_SMLABT( n_LTP_Q12, LTP_shp_buf[ ( idx - HARM_SHAPE_FIR_TAPS / 2 ) & LTP_MASK ], HarmShapeFIRPacked_Q12 );
n_LTP_Q12 = silk_SMLABB( n_LTP_Q12, LTP_shp_buf[ ( idx - HARM_SHAPE_FIR_TAPS / 2 + 1) & LTP_MASK ], HarmShapeFIRPacked_Q12 );
n_Tilt_Q10 = silk_SMULWB( sLF_AR_shp_Q12, Tilt_Q14 );
n_LF_Q10 = silk_SMLAWB( silk_SMULWT( sLF_AR_shp_Q12, LF_shp_Q14 ), sLF_MA_shp_Q12, LF_shp_Q14 );
sLF_AR_shp_Q12 = silk_SUB32( st_res_Q12[ i ], silk_LSHIFT( n_Tilt_Q10, 2 ) );
sLF_MA_shp_Q12 = silk_SUB32( sLF_AR_shp_Q12, silk_LSHIFT( n_LF_Q10, 2 ) );
LTP_shp_buf_idx = ( LTP_shp_buf_idx - 1 ) & LTP_MASK;
LTP_shp_buf[ LTP_shp_buf_idx ] = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND( sLF_MA_shp_Q12, 12 ) );
xw_Q3[i] = silk_RSHIFT_ROUND( silk_SUB32( sLF_MA_shp_Q12, n_LTP_Q12 ), 9 );
}
}
else
{
for( i = 0; i < length; i++ ) {
n_LTP_Q12 = 0;
n_Tilt_Q10 = silk_SMULWB( sLF_AR_shp_Q12, Tilt_Q14 );
n_LF_Q10 = silk_SMLAWB( silk_SMULWT( sLF_AR_shp_Q12, LF_shp_Q14 ), sLF_MA_shp_Q12, LF_shp_Q14 );
sLF_AR_shp_Q12 = silk_SUB32( st_res_Q12[ i ], silk_LSHIFT( n_Tilt_Q10, 2 ) );
sLF_MA_shp_Q12 = silk_SUB32( sLF_AR_shp_Q12, silk_LSHIFT( n_LF_Q10, 2 ) );
LTP_shp_buf_idx = ( LTP_shp_buf_idx - 1 ) & LTP_MASK;
LTP_shp_buf[ LTP_shp_buf_idx ] = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND( sLF_MA_shp_Q12, 12 ) );
xw_Q3[i] = silk_RSHIFT_ROUND( sLF_MA_shp_Q12, 9 );
}
}
/* Copy temp variable back to state */
P->sLF_AR_shp_Q12 = sLF_AR_shp_Q12;
P->sLF_MA_shp_Q12 = sLF_MA_shp_Q12;
P->sLTP_shp_buf_idx = LTP_shp_buf_idx;
}
#endif /* __PREFILTER_FIX_MIPSR1_H__ */

160
thirdparty/opus/silk/fixed/x86/prefilter_FIX_sse.c vendored
View file

@ -1,160 +0,0 @@
/* Copyright (c) 2014, Cisco Systems, INC
Written by XiangMingZhu WeiZhou MinPeng YanWang
Redistribution and use in source and binary forms, with or without
modification, are permitted provided that the following conditions
are met:
- Redistributions of source code must retain the above copyright
notice, this list of conditions and the following disclaimer.
- Redistributions in binary form must reproduce the above copyright
notice, this list of conditions and the following disclaimer in the
documentation and/or other materials provided with the distribution.
THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*/
#ifdef HAVE_CONFIG_H
#include "config.h"
#endif
#include <xmmintrin.h>
#include <emmintrin.h>
#include <smmintrin.h>
#include "main.h"
#include "celt/x86/x86cpu.h"
void silk_warped_LPC_analysis_filter_FIX_sse4_1(
opus_int32 state[], /* I/O State [order + 1] */
opus_int32 res_Q2[], /* O Residual signal [length] */
const opus_int16 coef_Q13[], /* I Coefficients [order] */
const opus_int16 input[], /* I Input signal [length] */
const opus_int16 lambda_Q16, /* I Warping factor */
const opus_int length, /* I Length of input signal */
const opus_int order /* I Filter order (even) */
)
{
opus_int n, i;
opus_int32 acc_Q11, tmp1, tmp2;
/* Order must be even */
celt_assert( ( order & 1 ) == 0 );
if (order == 10)
{
if (0 == lambda_Q16)
{
__m128i coef_Q13_3210, coef_Q13_7654;
__m128i coef_Q13_0123, coef_Q13_4567;
__m128i state_0123, state_4567;
__m128i xmm_product1, xmm_product2;
__m128i xmm_tempa, xmm_tempb;
register opus_int32 sum;
register opus_int32 state_8, state_9, state_a;
register opus_int64 coef_Q13_8, coef_Q13_9;
celt_assert( length > 0 );
coef_Q13_3210 = OP_CVTEPI16_EPI32_M64( &coef_Q13[ 0 ] );
coef_Q13_7654 = OP_CVTEPI16_EPI32_M64( &coef_Q13[ 4 ] );
coef_Q13_0123 = _mm_shuffle_epi32( coef_Q13_3210, _MM_SHUFFLE( 0, 1, 2, 3 ) );
coef_Q13_4567 = _mm_shuffle_epi32( coef_Q13_7654, _MM_SHUFFLE( 0, 1, 2, 3 ) );
coef_Q13_8 = (opus_int64) coef_Q13[ 8 ];
coef_Q13_9 = (opus_int64) coef_Q13[ 9 ];
state_0123 = _mm_loadu_si128( (__m128i *)(&state[ 0 ] ) );
state_4567 = _mm_loadu_si128( (__m128i *)(&state[ 4 ] ) );
state_0123 = _mm_shuffle_epi32( state_0123, _MM_SHUFFLE( 0, 1, 2, 3 ) );
state_4567 = _mm_shuffle_epi32( state_4567, _MM_SHUFFLE( 0, 1, 2, 3 ) );
state_8 = state[ 8 ];
state_9 = state[ 9 ];
state_a = 0;
for( n = 0; n < length; n++ )
{
xmm_product1 = _mm_mul_epi32( coef_Q13_0123, state_0123 ); /* 64-bit multiply, only 2 pairs */
xmm_product2 = _mm_mul_epi32( coef_Q13_4567, state_4567 );
xmm_tempa = _mm_shuffle_epi32( state_0123, _MM_SHUFFLE( 0, 1, 2, 3 ) );
xmm_tempb = _mm_shuffle_epi32( state_4567, _MM_SHUFFLE( 0, 1, 2, 3 ) );
xmm_product1 = _mm_srli_epi64( xmm_product1, 16 ); /* >> 16, zero extending works */
xmm_product2 = _mm_srli_epi64( xmm_product2, 16 );
xmm_tempa = _mm_mul_epi32( coef_Q13_3210, xmm_tempa );
xmm_tempb = _mm_mul_epi32( coef_Q13_7654, xmm_tempb );
xmm_tempa = _mm_srli_epi64( xmm_tempa, 16 );
xmm_tempb = _mm_srli_epi64( xmm_tempb, 16 );
xmm_tempa = _mm_add_epi32( xmm_tempa, xmm_product1 );
xmm_tempb = _mm_add_epi32( xmm_tempb, xmm_product2 );
xmm_tempa = _mm_add_epi32( xmm_tempa, xmm_tempb );
sum = (opus_int32)((coef_Q13_8 * state_8) >> 16);
sum += (opus_int32)((coef_Q13_9 * state_9) >> 16);
xmm_tempa = _mm_add_epi32( xmm_tempa, _mm_shuffle_epi32( xmm_tempa, _MM_SHUFFLE( 0, 0, 0, 2 ) ) );
sum += _mm_cvtsi128_si32( xmm_tempa);
res_Q2[ n ] = silk_LSHIFT( (opus_int32)input[ n ], 2 ) - silk_RSHIFT_ROUND( ( 5 + sum ), 9);
/* move right */
state_a = state_9;
state_9 = state_8;
state_8 = _mm_cvtsi128_si32( state_4567 );
state_4567 = _mm_alignr_epi8( state_0123, state_4567, 4 );
state_0123 = _mm_alignr_epi8( _mm_cvtsi32_si128( silk_LSHIFT( input[ n ], 14 ) ), state_0123, 4 );
}
_mm_storeu_si128( (__m128i *)( &state[ 0 ] ), _mm_shuffle_epi32( state_0123, _MM_SHUFFLE( 0, 1, 2, 3 ) ) );
_mm_storeu_si128( (__m128i *)( &state[ 4 ] ), _mm_shuffle_epi32( state_4567, _MM_SHUFFLE( 0, 1, 2, 3 ) ) );
state[ 8 ] = state_8;
state[ 9 ] = state_9;
state[ 10 ] = state_a;
return;
}
}
for( n = 0; n < length; n++ ) {
/* Output of lowpass section */
tmp2 = silk_SMLAWB( state[ 0 ], state[ 1 ], lambda_Q16 );
state[ 0 ] = silk_LSHIFT( input[ n ], 14 );
/* Output of allpass section */
tmp1 = silk_SMLAWB( state[ 1 ], state[ 2 ] - tmp2, lambda_Q16 );
state[ 1 ] = tmp2;
acc_Q11 = silk_RSHIFT( order, 1 );
acc_Q11 = silk_SMLAWB( acc_Q11, tmp2, coef_Q13[ 0 ] );
/* Loop over allpass sections */
for( i = 2; i < order; i += 2 ) {
/* Output of allpass section */
tmp2 = silk_SMLAWB( state[ i ], state[ i + 1 ] - tmp1, lambda_Q16 );
state[ i ] = tmp1;
acc_Q11 = silk_SMLAWB( acc_Q11, tmp1, coef_Q13[ i - 1 ] );
/* Output of allpass section */
tmp1 = silk_SMLAWB( state[ i + 1 ], state[ i + 2 ] - tmp2, lambda_Q16 );
state[ i + 1 ] = tmp2;
acc_Q11 = silk_SMLAWB( acc_Q11, tmp2, coef_Q13[ i ] );
}
state[ order ] = tmp1;
acc_Q11 = silk_SMLAWB( acc_Q11, tmp1, coef_Q13[ order - 1 ] );
res_Q2[ n ] = silk_LSHIFT( (opus_int32)input[ n ], 2 ) - silk_RSHIFT_ROUND( acc_Q11, 9 );
}
}