Merge pull request #48455 from JFonS/3.x_embree_aarch64

[3.x] Switch to embree-aarch64

COPYRIGHT.txt

@@ -135,6 +135,11 @@ Copyright: 2018, Eric Lasota
  2018, Microsoft Corp.
 License: Expat
 
+Files: ./thirdparty/embree/
+Comment: Embree
+Copyright: 2009-2021 Intel Corporation
+License: Apache-2.0
+
 Files: ./thirdparty/enet/
 Comment: ENet
 Copyright: 2002-2020, Lee Salzman

modules/lightmapper_cpu/config.py

@@ -6,23 +6,13 @@ def can_build(env, platform):
     # `can_build()` for that module, so we need to duplicate that code as a short-term
     # solution.
 
-    # Embree requires at least SSE2 to be available, so 32-bit and ARM64 builds are
+    if platform == "android":
-    # not supported.
+        return env["android_arch"] in ["arm64v8", "x86", "x86_64"]
-    # It's also only relevant for tools build and desktop platforms,
-    # as doing lightmap generation on Android or HTML5 would be a bit far-fetched.
-    supported_platform = platform in ["x11", "osx", "windows", "server"]
-    supported_bits = env["bits"] == "64"
-    supported_arch = env["arch"] != "arm64"
 
-    # Hack to disable on Linux arm64. This won't work well for cross-compilation (checks
+    if platform in ["javascript", "server"]:
-    # host, not target) and would need a more thorough fix by refactoring our arch and
+        return False
-    # bits-handling code.
-    from platform import machine
 
-    if platform == "x11" and machine() != "x86_64":
+    return True
-        supported_arch = False
-
-    return supported_platform and supported_bits and supported_arch
 
 
 def configure(env):

modules/raycast/SCsub

@@ -70,25 +70,19 @@ if env["builtin_embree"]:
     thirdparty_sources = [thirdparty_dir + file for file in embree_src]
 
     env_raycast.Prepend(CPPPATH=[thirdparty_dir, thirdparty_dir + "include"])
-    env_raycast.Append(
+    env_raycast.Append(CPPDEFINES=["EMBREE_TARGET_SSE2", "EMBREE_LOWEST_ISA", "TASKING_INTERNAL", "NDEBUG"])
-        CPPDEFINES=[
-            "EMBREE_TARGET_SSE2",
-            "EMBREE_LOWEST_ISA",
-            "TASKING_INTERNAL",
-            "NDEBUG",
-            "__SSE2__",
-            "__SSE__",
-        ]
-    )
 
     if not env.msvc:
-        env_raycast.Append(CPPFLAGS=["-msse2", "-mxsave"])
+        if env["arch"] in ["x86", "x86_64"]:
+            env_raycast.Append(CPPFLAGS=["-msse2", "-mxsave"])
+
         if env["platform"] == "windows":
             env_raycast.Append(CPPFLAGS=["-mstackrealign"])
 
     if env["platform"] == "windows":
         if env.msvc:
             env.Append(LINKFLAGS=["psapi.lib"])
+            env_raycast.Append(CPPDEFINES=["__SSE2__", "__SSE__"])
         else:
             env.Append(LIBS=["psapi"])
 

modules/raycast/config.py

@@ -1,21 +1,14 @@
 def can_build(env, platform):
-    # Embree requires at least SSE2 to be available, so 32-bit and ARM64 builds are
+    if not env["tools"]:
-    # not supported.
+        return False
-    # It's also only relevant for tools build and desktop platforms,
-    # as doing lightmap generation on Android or HTML5 would be a bit far-fetched.
-    supported_platform = platform in ["x11", "osx", "windows", "server"]
-    supported_bits = env["bits"] == "64"
-    supported_arch = env["arch"] != "arm64"
 
-    # Hack to disable on Linux arm64. This won't work well for cross-compilation (checks
+    if platform == "android":
-    # host, not target) and would need a more thorough fix by refactoring our arch and
+        return env["android_arch"] in ["arm64v8", "x86", "x86_64"]
-    # bits-handling code.
-    from platform import machine
 
-    if platform == "x11" and machine() != "x86_64":
+    if platform in ["javascript", "server"]:
-        supported_arch = False
+        return False
 
-    return env["tools"] and supported_platform and supported_bits and supported_arch
+    return True
 
 
 def configure(env):

modules/raycast/godot_update_embree.py

@@ -74,17 +74,18 @@ cpp_files = [
 
 os.chdir("../../thirdparty")
 
-if os.path.exists("embree"):
+dir_name = "embree"
-    shutil.rmtree("embree")
+if os.path.exists(dir_name):
+    shutil.rmtree(dir_name)
 
-subprocess.run(["git", "clone", "https://github.com/embree/embree.git", "embree-tmp"])
+subprocess.run(["git", "clone", "https://github.com/lighttransport/embree-aarch64.git", "embree-tmp"])
 os.chdir("embree-tmp")
 
 commit_hash = str(subprocess.check_output(["git", "rev-parse", "HEAD"], universal_newlines=True)).strip()
 
-dest_dir = "../embree"
 all_files = set(cpp_files)
 
+dest_dir = os.path.join("..", dir_name)
 for include_dir in include_dirs:
     headers = glob.iglob(os.path.join(include_dir, "*.h"))
     all_files.update(headers)

modules/raycast/lightmap_raycaster.cpp

@@ -190,8 +190,11 @@ LightmapRaycasterEmbree::~LightmapRaycasterEmbree() {
 	_MM_SET_FLUSH_ZERO_MODE(_MM_FLUSH_ZERO_OFF);
 	_MM_SET_DENORMALS_ZERO_MODE(_MM_DENORMALS_ZERO_OFF);
 
-	if (embree_scene != nullptr)
+	if (embree_scene != nullptr) {
 		rtcReleaseScene(embree_scene);
-	if (embree_device != nullptr)
+	}
+
+	if (embree_device != nullptr) {
 		rtcReleaseDevice(embree_device);
+	}
 }

thirdparty/README.md

@@ -41,19 +41,19 @@ Files extracted from upstream source:
 
 ## embree
 
-- Upstream: https://github.com/embree/embree
+- Upstream: https://github.com/lighttransport/embree-aarch64
-- Version: 3.12.1 (69bd4c272f1ed608494f233ecfff3feec516880b, 2020)
+- Version: 3.12.1 (6ef362f99af80c9dfe8dd2bfc582d9067897edc6, 2020)
 - License: Apache 2.0
 
 Files extracted from upstream:
 
-- All cpp files listed in `modules/raytrace/godot_update_embree.py`
+- All cpp files listed in `modules/raycast/godot_update_embree.py`
-- All header files in the directories listed in `modules/raytrace/godot_update_embree.py`
+- All header files in the directories listed in `modules/raycast/godot_update_embree.py`
 
-The `modules/raytrace/godot_update_embree.py`script can be used to pull the 
+The `modules/raycast/godot_update_embree.py`script can be used to pull the 
-relevant files from the latest Embree release and apply some automatic changes.
+relevant files from the latest Embree-aarch64 release and apply some automatic changes.
 
-Some minor changes have been made in order to fix build errors.
+Some changes have been made in order to remove exceptions and fix minor build errors.
 They are marked with `// -- GODOT start --` and `// -- GODOT end --`
 comments. Apply the patches in the `patches/` folder when syncing on newer upstream
 commits.

thirdparty/embree/common/algorithms/parallel_for.h

@@ -8,6 +8,12 @@
 #include "../math/math.h"
 #include "../math/range.h"
 
+#if defined(TASKING_GCD) && defined(BUILD_IOS)
+#include <dispatch/dispatch.h>
+#include <algorithm>
+#include <type_traits>
+#endif
+
 namespace embree
 {
   /* parallel_for without range */
@@ -21,9 +27,30 @@ namespace embree
           func(r.begin());
         });
       if (!TaskScheduler::wait())
-        throw std::runtime_error("task cancelled");
+        // -- GODOT start --
+        // throw std::runtime_error("task cancelled");
+        abort(); 
+        // -- GODOT end --
     }
-    
+#elif defined(TASKING_GCD) && defined(BUILD_IOS)
+      
+    const size_t baselineNumBlocks = (TaskScheduler::threadCount() > 1)? TaskScheduler::threadCount() : 1;
+    const size_t length = N;
+    const size_t blockSize = (length + baselineNumBlocks-1) / baselineNumBlocks;
+    const size_t numBlocks = (length + blockSize-1) / blockSize;
+      
+    dispatch_apply(numBlocks, DISPATCH_APPLY_AUTO, ^(size_t currentBlock) {
+          
+        const size_t start = (currentBlock * blockSize);
+        const size_t blockLength = std::min(length - start, blockSize);
+        const size_t end = start + blockLength;
+          
+        for(size_t i=start; i < end; i++)
+        {
+            func(i);
+        }
+    });
+      
 #elif defined(TASKING_TBB)
   #if TBB_INTERFACE_VERSION >= 12002
     tbb::task_group_context context;
@@ -31,13 +58,19 @@ namespace embree
         func(i);
       },context);
     if (context.is_group_execution_cancelled())
-      throw std::runtime_error("task cancelled");
+      // -- GODOT start --
+      // throw std::runtime_error("task cancelled");
+      abort(); 
+      // -- GODOT end --
   #else
     tbb::parallel_for(Index(0),N,Index(1),[&](Index i) {
         func(i);
       });
     if (tbb::task::self().is_cancelled())
-      throw std::runtime_error("task cancelled");
+      // -- GODOT start --
+      // throw std::runtime_error("task cancelled");
+      abort(); 
+      // -- GODOT end --
   #endif
 
 #elif defined(TASKING_PPL)
@@ -57,7 +90,29 @@ namespace embree
 #if defined(TASKING_INTERNAL)
     TaskScheduler::spawn(first,last,minStepSize,func);
     if (!TaskScheduler::wait())
-      throw std::runtime_error("task cancelled");
+      // -- GODOT start --
+      // throw std::runtime_error("task cancelled");
+      abort(); 
+      // -- GODOT end --
+
+#elif defined(TASKING_GCD) && defined(BUILD_IOS)
+      
+    const size_t baselineNumBlocks = (TaskScheduler::threadCount() > 1)? 4*TaskScheduler::threadCount() : 1;
+    const size_t length = last - first;
+    const size_t blockSizeByThreads = (length + baselineNumBlocks-1) / baselineNumBlocks;
+    size_t blockSize = std::max<size_t>(minStepSize,blockSizeByThreads);
+    blockSize += blockSize % 4;
+      
+    const size_t numBlocks = (length + blockSize-1) / blockSize;
+      
+    dispatch_apply(numBlocks, DISPATCH_APPLY_AUTO, ^(size_t currentBlock) {
+          
+        const size_t start = first + (currentBlock * blockSize);
+        const size_t end = std::min<size_t>(last, start + blockSize);
+          
+        func( embree::range<Index>(start,end) );
+    });
+      
 
 #elif defined(TASKING_TBB)
   #if TBB_INTERFACE_VERSION >= 12002
@@ -66,13 +121,19 @@ namespace embree
         func(range<Index>(r.begin(),r.end()));
       },context);
     if (context.is_group_execution_cancelled())
-      throw std::runtime_error("task cancelled");
+      // -- GODOT start --
+      // throw std::runtime_error("task cancelled");
+      abort(); 
+      // -- GODOT end --
   #else
     tbb::parallel_for(tbb::blocked_range<Index>(first,last,minStepSize),[&](const tbb::blocked_range<Index>& r) {
         func(range<Index>(r.begin(),r.end()));
       });
     if (tbb::task::self().is_cancelled())
-      throw std::runtime_error("task cancelled");
+      // -- GODOT start --
+      // throw std::runtime_error("task cancelled");
+      abort(); 
+      // -- GODOT end --
   #endif
 
 #elif defined(TASKING_PPL)
@@ -104,13 +165,19 @@ namespace embree
           func(i);
         },tbb::simple_partitioner(),context);
       if (context.is_group_execution_cancelled())
-        throw std::runtime_error("task cancelled");
+        // -- GODOT start --
+        // throw std::runtime_error("task cancelled");
+        abort(); 
+        // -- GODOT end --
     #else
       tbb::parallel_for(Index(0),N,Index(1),[&](Index i) {
           func(i);
         },tbb::simple_partitioner());
       if (tbb::task::self().is_cancelled())
-        throw std::runtime_error("task cancelled");
+        // -- GODOT start --
+        // throw std::runtime_error("task cancelled");
+        abort(); 
+        // -- GODOT end --
     #endif
   }
 
@@ -125,13 +192,19 @@ namespace embree
           func(i);
         },ap,context);
       if (context.is_group_execution_cancelled())
-        throw std::runtime_error("task cancelled");
+       // -- GODOT start --
+       // throw std::runtime_error("task cancelled");
+       abort(); 
+       // -- GODOT end --
     #else
       tbb::parallel_for(Index(0),N,Index(1),[&](Index i) {
           func(i);
         },ap);
       if (tbb::task::self().is_cancelled())
-        throw std::runtime_error("task cancelled");
+        // -- GODOT start --
+        // throw std::runtime_error("task cancelled");
+        abort(); 
+        // -- GODOT end --
     #endif
   }
 

thirdparty/embree/common/algorithms/parallel_reduce.h

@@ -43,7 +43,7 @@ namespace embree
   template<typename Index, typename Value, typename Func, typename Reduction>
     __forceinline Value parallel_reduce( const Index first, const Index last, const Index minStepSize, const Value& identity, const Func& func, const Reduction& reduction )
   {
-#if defined(TASKING_INTERNAL)
+#if defined(TASKING_INTERNAL) || (defined(TASKING_GCD) && defined(BUILD_IOS))
 
     /* fast path for small number of iterations */
     Index taskCount = (last-first+minStepSize-1)/minStepSize;
@@ -58,15 +58,19 @@ namespace embree
     const Value v = tbb::parallel_reduce(tbb::blocked_range<Index>(first,last,minStepSize),identity,
       [&](const tbb::blocked_range<Index>& r, const Value& start) { return reduction(start,func(range<Index>(r.begin(),r.end()))); },
       reduction,context);
-    if (context.is_group_execution_cancelled())
+    // -- GODOT start --
-      throw std::runtime_error("task cancelled");
+    // if (context.is_group_execution_cancelled())
+    //   throw std::runtime_error("task cancelled");
+    // -- GODOT end --
     return v;
   #else
     const Value v = tbb::parallel_reduce(tbb::blocked_range<Index>(first,last,minStepSize),identity,
       [&](const tbb::blocked_range<Index>& r, const Value& start) { return reduction(start,func(range<Index>(r.begin(),r.end()))); },
       reduction);
-    if (tbb::task::self().is_cancelled())
+    // -- GODOT start --
-      throw std::runtime_error("task cancelled");
+    // if (tbb::task::self().is_cancelled())
+    //   throw std::runtime_error("task cancelled");
+    // -- GODOT end --
     return v;
   #endif
 #else // TASKING_PPL

thirdparty/embree/common/algorithms/parallel_sort.h

@@ -5,6 +5,9 @@
 
 #include "../simd/simd.h"
 #include "parallel_for.h"
+#if defined(TASKING_GCD) && defined(BUILD_IOS)
+#include "../sys/alloc.h"
+#endif
 #include <algorithm>
 
 namespace embree
@@ -320,7 +323,7 @@ namespace embree
 #pragma nounroll      
 #endif
       for (size_t i=startID; i<endID; i++) {
-#if defined(__X86_64__)
+#if defined(__X86_64__) || defined(__aarch64__)
         const size_t index = ((size_t)(Key)src[i] >> (size_t)shift) & (size_t)mask;
 #else
         const Key index = ((Key)src[i] >> shift) & mask;
@@ -382,7 +385,7 @@ namespace embree
 #endif
       for (size_t i=startID; i<endID; i++) {
         const Ty elt = src[i];
-#if defined(__X86_64__)
+#if defined(__X86_64__) || defined(__aarch64__)
         const size_t index = ((size_t)(Key)src[i] >> (size_t)shift) & (size_t)mask;
 #else
         const size_t index = ((Key)src[i] >> shift) & mask;

thirdparty/embree/common/lexers/stringstream.cpp

@@ -39,7 +39,10 @@ namespace embree
     std::vector<char> str; str.reserve(64);
     while (cin->peek() != EOF && !isSeparator(cin->peek())) {
       int c = cin->get();
-      if (!isValidChar(c)) throw std::runtime_error("invalid character "+std::string(1,c)+" in input");
+      // -- GODOT start --
+      // if (!isValidChar(c)) throw std::runtime_error("invalid character "+std::string(1,c)+" in input");
+      if (!isValidChar(c)) abort();
+      // -- GODOT end --
       str.push_back((char)c);
     }
     str.push_back(0);

thirdparty/embree/common/math/AVX2NEON.h

@@ -0,0 +1,986 @@
+#pragma once
+
+#include "SSE2NEON.h"
+
+
+#define AVX2NEON_ABI static inline  __attribute__((always_inline))
+
+
+struct __m256d;
+
+struct __m256 {
+    __m128 lo,hi;
+    __m256() {}
+};
+
+
+
+
+struct __m256i {
+    __m128i lo,hi;
+    explicit __m256i(const __m256 a) : lo(__m128i(a.lo)),hi(__m128i(a.hi)) {}
+    operator __m256() const {__m256 res; res.lo = __m128(lo);res.hi = __m128(hi); return res;}
+    __m256i() {}
+};
+ 
+
+
+
+struct __m256d {
+    float64x2_t lo,hi;
+    __m256d() {}
+    __m256d(const __m256& a) : lo(float64x2_t(a.lo)),hi(float64x2_t(a.hi)) {}
+    __m256d(const __m256i& a) : lo(float64x2_t(a.lo)),hi(float64x2_t(a.hi)) {}
+};
+
+#define UNARY_AVX_OP(type,func,basic_func) AVX2NEON_ABI type func(const type& a) {type res;res.lo=basic_func(a.lo);res.hi=basic_func(a.hi);return res;}
+
+
+#define BINARY_AVX_OP(type,func,basic_func) AVX2NEON_ABI type func(const type& a,const type& b) {type res;res.lo=basic_func(a.lo,b.lo);res.hi=basic_func(a.hi,b.hi);return res;}
+#define BINARY_AVX_OP_CAST(type,func,basic_func,bdst,bsrc) AVX2NEON_ABI type func(const type& a,const type& b) {type res;res.lo=bdst(basic_func(bsrc(a.lo),bsrc(b.lo)));res.hi=bdst(basic_func(bsrc(a.hi),bsrc(b.hi)));return res;}
+
+#define TERNARY_AVX_OP(type,func,basic_func) AVX2NEON_ABI type func(const type& a,const type& b,const type& c) {type res;res.lo=basic_func(a.lo,b.lo,c.lo);res.hi=basic_func(a.hi,b.hi,c.hi);return res;}
+
+
+#define CAST_SIMD_TYPE(to,name,from,basic_dst) AVX2NEON_ABI to name(const from& a) { to res; res.lo = basic_dst(a.lo); res.hi=basic_dst(a.hi); return res;}
+
+
+
+#define _mm_stream_load_si128 _mm_load_si128
+#define _mm256_stream_load_si256 _mm256_load_si256
+
+
+AVX2NEON_ABI
+__m128 _mm_blend_ps (__m128 a, __m128 b, const int imm8)
+{
+    __m128 res;
+    for (int i=0;i<4;i++)
+    {
+        if (imm8 & (1<<i))
+        {
+            res[i] = b[i];
+        }
+        else{
+            res[i] = a[i];
+        }
+    }
+    
+    return res;
+}
+
+AVX2NEON_ABI
+__m128i _mm_blend_epi32 (__m128i a, __m128i b, const int imm8)
+{
+    __m128i res;
+    for (int i=0;i<4;i++)
+    {
+        if (imm8 & (1<<i))
+        {
+            res[i] = b[i];
+        }
+        else{
+            res[i] = a[i];
+        }
+    }
+    return res;
+}
+
+AVX2NEON_ABI
+__m128 _mm_cmpngt_ps (__m128 a, __m128 b)
+{
+    return __m128(vmvnq_s32(__m128i(_mm_cmpgt_ps(a,b))));
+}
+
+
+AVX2NEON_ABI
+__m128i _mm_loadl_epi64 (__m128i const* mem_addr)
+{
+    int64x2_t y;
+    y[0] = *(int64_t *)mem_addr;
+    y[1] = 0;
+    return __m128i(y);
+}
+
+AVX2NEON_ABI
+int _mm_movemask_popcnt(__m128 a)
+{
+    return __builtin_popcount(_mm_movemask_ps(a));
+}
+
+AVX2NEON_ABI
+__m128 _mm_maskload_ps (float const * mem_addr, __m128i mask)
+{
+    __m128 res;
+    for (int i=0;i<4;i++) {
+        if (mask[i] & 0x80000000) res[i] = mem_addr[i]; else res[i] = 0;
+    }
+    return res;
+}
+
+AVX2NEON_ABI
+void _mm_maskstore_ps (float * mem_addr, __m128i mask, __m128 a)
+{
+    for (int i=0;i<4;i++) {
+        if (mask[i] & 0x80000000) mem_addr[i] = a[i];
+    }
+}
+
+AVX2NEON_ABI
+void _mm_maskstore_epi32 (int * mem_addr, __m128i mask, __m128i a)
+{
+    for (int i=0;i<4;i++) {
+        if (mask[i] & 0x80000000) mem_addr[i] = a[i];
+    }
+}
+
+AVX2NEON_ABI
+__m128 _mm_fnmsub_ps (__m128 a, __m128 b, __m128 c)
+{
+    return vnegq_f32(vfmaq_f32(c,a,b));
+}
+
+#define _mm_fnmsub_ss _mm_fnmsub_ps
+
+AVX2NEON_ABI
+__m128 _mm_fnmadd_ps (__m128 a, __m128 b, __m128 c)
+{
+    return vfmsq_f32(c,a,b);
+}
+
+#define _mm_fnmadd_ss _mm_fnmadd_ps
+
+
+AVX2NEON_ABI
+__m128 _mm_broadcast_ss (float const * mem_addr)
+{
+    return vdupq_n_f32(*mem_addr);
+}
+
+
+AVX2NEON_ABI
+__m128 _mm_fmsub_ps (__m128 a, __m128 b, __m128 c)
+{
+    return vfmaq_f32(vnegq_f32(c),a,b);
+}
+
+#define _mm_fmsub_ss _mm_fmsub_ps
+#define _mm_fmadd_ps _mm_madd_ps
+#define _mm_fmadd_ss _mm_madd_ps
+
+
+
+template<int code>
+AVX2NEON_ABI float32x4_t dpps_neon(const float32x4_t& a,const float32x4_t& b)
+{
+    float v;
+    v = 0;
+    v += (code & 0x10) ? a[0]*b[0] : 0;
+    v += (code & 0x20) ? a[1]*b[1] : 0;
+    v += (code & 0x40) ? a[2]*b[2] : 0;
+    v += (code & 0x80) ? a[3]*b[3] : 0;
+    float32x4_t res;
+    res[0] = (code & 0x1) ? v : 0;
+    res[1] = (code & 0x2) ? v : 0;
+    res[2] = (code & 0x4) ? v : 0;
+    res[3] = (code & 0x8) ? v : 0;
+    return res;
+}
+
+template<>
+inline float32x4_t dpps_neon<0x7f>(const float32x4_t& a,const float32x4_t& b)
+{
+    float v;
+    float32x4_t m = _mm_mul_ps(a,b);
+    m[3] = 0;
+    v = vaddvq_f32(m);
+    return _mm_set1_ps(v);
+}
+
+template<>
+inline float32x4_t dpps_neon<0xff>(const float32x4_t& a,const float32x4_t& b)
+{
+    float v;
+    float32x4_t m = _mm_mul_ps(a,b);
+    v = vaddvq_f32(m);
+    return _mm_set1_ps(v);
+}
+
+#define _mm_dp_ps(a,b,c) dpps_neon<c>((a),(b))
+
+
+
+AVX2NEON_ABI
+__m128 _mm_cmpnge_ps (__m128 a, __m128 b)
+{
+    return __m128(vmvnq_s32(__m128i(_mm_cmpge_ps(a,b))));
+}
+
+
+AVX2NEON_ABI
+__m128 _mm_permutevar_ps (__m128 a, __m128i b)
+{
+    __m128 x;
+    for (int i=0;i<4;i++)
+    {
+        x[i] = a[b[i&3]];
+    }
+    return x;
+}
+
+AVX2NEON_ABI
+__m256i _mm256_setzero_si256()
+{
+    __m256i res;
+    res.lo = res.hi = vdupq_n_s32(0);
+    return res;
+}
+
+AVX2NEON_ABI
+__m256 _mm256_setzero_ps()
+{
+    __m256 res;
+    res.lo = res.hi = vdupq_n_f32(0.0f);
+    return res;
+}
+
+AVX2NEON_ABI
+__m256i _mm256_undefined_si256()
+{
+    return _mm256_setzero_si256();
+}
+
+AVX2NEON_ABI
+__m256 _mm256_undefined_ps()
+{
+    return _mm256_setzero_ps();
+}
+
+CAST_SIMD_TYPE(__m256d,_mm256_castps_pd,__m256,float64x2_t)
+CAST_SIMD_TYPE(__m256i,_mm256_castps_si256,__m256,__m128i)
+CAST_SIMD_TYPE(__m256, _mm256_castsi256_ps, __m256i,__m128)
+CAST_SIMD_TYPE(__m256, _mm256_castpd_ps ,__m256d,__m128)
+CAST_SIMD_TYPE(__m256d, _mm256_castsi256_pd, __m256i,float64x2_t)
+CAST_SIMD_TYPE(__m256i, _mm256_castpd_si256, __m256d,__m128i)
+
+
+
+
+AVX2NEON_ABI
+__m128 _mm256_castps256_ps128 (__m256 a)
+{
+    return a.lo;
+}
+
+AVX2NEON_ABI
+__m256i _mm256_castsi128_si256 (__m128i a)
+{
+    __m256i res;
+    res.lo = a ;
+    res.hi = vdupq_n_s32(0);
+    return res;
+}
+
+AVX2NEON_ABI
+__m128i _mm256_castsi256_si128 (__m256i a)
+{
+    return a.lo;
+}
+
+AVX2NEON_ABI
+__m256 _mm256_castps128_ps256 (__m128 a)
+{
+    __m256 res;
+    res.lo = a;
+    res.hi = vdupq_n_f32(0);
+    return res;
+}
+
+
+AVX2NEON_ABI
+__m256 _mm256_broadcast_ss (float const * mem_addr)
+{
+    __m256 res;
+    res.lo = res.hi = vdupq_n_f32(*mem_addr);
+    return res;
+}
+
+
+
+AVX2NEON_ABI
+__m256i _mm256_set_epi32 (int e7, int e6, int e5, int e4, int e3, int e2, int e1, int e0)
+{
+    __m128i lo = {e0,e1,e2,e3}, hi = {e4,e5,e6,e7};
+    __m256i res;
+    res.lo = lo; res.hi = hi;
+    return res;
+    
+}
+
+AVX2NEON_ABI
+__m256i _mm256_set1_epi32 (int a)
+{
+    __m256i res;
+    res.lo = res.hi = vdupq_n_s32(a);
+    return res;
+}
+
+
+
+
+AVX2NEON_ABI
+int _mm256_movemask_ps(const __m256& v)
+{
+    return (_mm_movemask_ps(v.hi) << 4) | _mm_movemask_ps(v.lo);
+}
+
+template<int imm8>
+AVX2NEON_ABI
+__m256 __mm256_permute_ps (const __m256& a)
+{
+    __m256 res;
+    res.lo = _mm_shuffle_ps(a.lo,a.lo,imm8);
+    res.hi = _mm_shuffle_ps(a.hi,a.hi,imm8);
+    return res;
+
+}
+
+#define _mm256_permute_ps(a,c) __mm256_permute_ps<c>(a)
+
+
+template<int imm8>
+AVX2NEON_ABI
+__m256 __mm256_shuffle_ps (const __m256 a,const __m256& b)
+{
+    __m256 res;
+    res.lo = _mm_shuffle_ps(a.lo,b.lo,imm8);
+    res.hi = _mm_shuffle_ps(a.hi,b.hi,imm8);
+    return res;
+
+}
+
+#define _mm256_shuffle_ps(a,b,c) __mm256_shuffle_ps<c>(a,b)
+
+AVX2NEON_ABI
+__m256i _mm256_set1_epi64x (long long a)
+{
+    __m256i res;
+    int64x2_t t = vdupq_n_s64(a);
+    res.lo = res.hi = __m128i(t);
+    return res;
+}
+
+
+AVX2NEON_ABI
+__m256 _mm256_permute2f128_ps (__m256 a, __m256 b, int imm8)
+{
+    __m256 res;
+    __m128 tmp;
+    switch (imm8 & 0x7)
+    {
+        case 0: tmp = a.lo; break;
+        case 1: tmp = a.hi; break;
+        case 2: tmp = b.lo; break;
+        case 3: tmp = b.hi; break;
+    }
+    if (imm8 & 0x8)
+        tmp = _mm_setzero_ps();
+
+    
+    
+    res.lo = tmp;
+    imm8 >>= 4;
+    
+    switch (imm8 & 0x7)
+    {
+        case 0: tmp = a.lo; break;
+        case 1: tmp = a.hi; break;
+        case 2: tmp = b.lo; break;
+        case 3: tmp = b.hi; break;
+    }
+    if (imm8 & 0x8)
+        tmp = _mm_setzero_ps();
+    
+    res.hi = tmp;
+    
+    return res;
+}
+
+AVX2NEON_ABI
+__m256 _mm256_moveldup_ps (__m256 a)
+{
+    __m256 res;
+    res.lo[0] = res.lo[1] = a.lo[0];
+    res.lo[2] = res.lo[3] = a.lo[2];
+    res.hi[0] = res.hi[1] = a.hi[0];
+    res.hi[2] = res.hi[3] = a.hi[2];
+    return res;
+
+}
+
+AVX2NEON_ABI
+__m256 _mm256_movehdup_ps (__m256 a)
+{
+    __m256 res;
+    res.lo[0] = res.lo[1] = a.lo[1];
+    res.lo[2] = res.lo[3] = a.lo[3];
+    res.hi[0] = res.hi[1] = a.hi[1];
+    res.hi[2] = res.hi[3] = a.hi[3];
+    return res;
+}
+
+AVX2NEON_ABI
+__m256 _mm256_insertf128_ps (__m256 a, __m128 b, int imm8)
+{
+    __m256 res = a;
+    if (imm8 & 1) res.hi = b;
+    else res.lo = b;
+    return res;
+}
+
+
+AVX2NEON_ABI
+__m128 _mm256_extractf128_ps (__m256 a, const int imm8)
+{
+    if (imm8 & 1) return a.hi;
+    return a.lo;
+}
+
+
+AVX2NEON_ABI
+__m256d _mm256_movedup_pd (__m256d a)
+{
+    __m256d res;
+    res.hi = a.hi;
+    res.lo[0] = res.lo[1] = a.lo[0];
+    return res;
+}
+
+AVX2NEON_ABI
+__m256i _mm256_abs_epi32(__m256i a)
+{
+   __m256i res;
+   res.lo = vabsq_s32(a.lo);
+   res.hi = vabsq_s32(a.hi);
+   return res;
+}
+
+UNARY_AVX_OP(__m256,_mm256_sqrt_ps,_mm_sqrt_ps)
+UNARY_AVX_OP(__m256,_mm256_rsqrt_ps,_mm_rsqrt_ps)
+UNARY_AVX_OP(__m256,_mm256_rcp_ps,_mm_rcp_ps)
+UNARY_AVX_OP(__m256,_mm256_floor_ps,vrndmq_f32)
+UNARY_AVX_OP(__m256,_mm256_ceil_ps,vrndpq_f32)
+
+
+BINARY_AVX_OP(__m256i,_mm256_add_epi32,_mm_add_epi32)
+BINARY_AVX_OP(__m256i,_mm256_sub_epi32,_mm_sub_epi32)
+BINARY_AVX_OP(__m256i,_mm256_mullo_epi32,_mm_mullo_epi32)
+
+BINARY_AVX_OP(__m256i,_mm256_min_epi32,_mm_min_epi32)
+BINARY_AVX_OP(__m256i,_mm256_max_epi32,_mm_max_epi32)
+BINARY_AVX_OP_CAST(__m256i,_mm256_min_epu32,vminq_u32,__m128i,uint32x4_t)
+BINARY_AVX_OP_CAST(__m256i,_mm256_max_epu32,vmaxq_u32,__m128i,uint32x4_t)
+
+BINARY_AVX_OP(__m256,_mm256_min_ps,_mm_min_ps)
+BINARY_AVX_OP(__m256,_mm256_max_ps,_mm_max_ps)
+
+BINARY_AVX_OP(__m256,_mm256_add_ps,_mm_add_ps)
+BINARY_AVX_OP(__m256,_mm256_mul_ps,_mm_mul_ps)
+BINARY_AVX_OP(__m256,_mm256_sub_ps,_mm_sub_ps)
+BINARY_AVX_OP(__m256,_mm256_div_ps,_mm_div_ps)
+
+BINARY_AVX_OP(__m256,_mm256_and_ps,_mm_and_ps)
+BINARY_AVX_OP(__m256,_mm256_andnot_ps,_mm_andnot_ps)
+BINARY_AVX_OP(__m256,_mm256_or_ps,_mm_or_ps)
+BINARY_AVX_OP(__m256,_mm256_xor_ps,_mm_xor_ps)
+
+BINARY_AVX_OP_CAST(__m256d,_mm256_and_pd,vandq_s64,float64x2_t,int64x2_t)
+BINARY_AVX_OP_CAST(__m256d,_mm256_or_pd,vorrq_s64,float64x2_t,int64x2_t)
+BINARY_AVX_OP_CAST(__m256d,_mm256_xor_pd,veorq_s64,float64x2_t,int64x2_t)
+
+
+
+BINARY_AVX_OP(__m256i,_mm256_and_si256,_mm_and_si128)
+BINARY_AVX_OP(__m256i,_mm256_or_si256,_mm_or_si128)
+BINARY_AVX_OP(__m256i,_mm256_xor_si256,_mm_xor_si128)
+
+
+BINARY_AVX_OP(__m256,_mm256_unpackhi_ps,_mm_unpackhi_ps)
+BINARY_AVX_OP(__m256,_mm256_unpacklo_ps,_mm_unpacklo_ps)
+TERNARY_AVX_OP(__m256,_mm256_blendv_ps,_mm_blendv_ps)
+
+
+TERNARY_AVX_OP(__m256,_mm256_fmadd_ps,_mm_fmadd_ps)
+TERNARY_AVX_OP(__m256,_mm256_fnmadd_ps,_mm_fnmadd_ps)
+TERNARY_AVX_OP(__m256,_mm256_fmsub_ps,_mm_fmsub_ps)
+TERNARY_AVX_OP(__m256,_mm256_fnmsub_ps,_mm_fnmsub_ps)
+
+
+BINARY_AVX_OP(__m256i,_mm256_unpackhi_epi32,_mm_unpackhi_epi32)
+BINARY_AVX_OP(__m256i,_mm256_unpacklo_epi32,_mm_unpacklo_epi32)
+
+
+BINARY_AVX_OP(__m256i,_mm256_cmpeq_epi32,_mm_cmpeq_epi32)
+BINARY_AVX_OP(__m256i,_mm256_cmpgt_epi32,_mm_cmpgt_epi32)
+BINARY_AVX_OP(__m256,_mm256_cmpeq_ps,_mm_cmpeq_ps)
+BINARY_AVX_OP(__m256,_mm256_cmpneq_ps,_mm_cmpneq_ps)
+BINARY_AVX_OP(__m256,_mm256_cmpnlt_ps,_mm_cmpnlt_ps)
+BINARY_AVX_OP(__m256,_mm256_cmpngt_ps,_mm_cmpngt_ps)
+BINARY_AVX_OP(__m256,_mm256_cmpge_ps,_mm_cmpge_ps)
+BINARY_AVX_OP(__m256,_mm256_cmpnge_ps,_mm_cmpnge_ps)
+BINARY_AVX_OP(__m256,_mm256_cmplt_ps,_mm_cmplt_ps)
+BINARY_AVX_OP(__m256,_mm256_cmple_ps,_mm_cmple_ps)
+BINARY_AVX_OP(__m256,_mm256_cmpgt_ps,_mm_cmpgt_ps)
+BINARY_AVX_OP(__m256,_mm256_cmpnle_ps,_mm_cmpnle_ps)
+
+
+AVX2NEON_ABI
+__m256i _mm256_cvtps_epi32 (__m256 a)
+{
+    __m256i res;
+    res.lo = _mm_cvtps_epi32(a.lo);
+    res.hi = _mm_cvtps_epi32(a.hi);
+    return res;
+    
+}
+
+AVX2NEON_ABI
+__m256i _mm256_cvttps_epi32 (__m256 a)
+{
+    __m256i res;
+    res.lo = _mm_cvttps_epi32(a.lo);
+    res.hi = _mm_cvttps_epi32(a.hi);
+    return res;
+    
+}
+
+AVX2NEON_ABI
+__m256 _mm256_loadu_ps (float const * mem_addr)
+{
+    __m256 res;
+    res.lo = *(__m128 *)(mem_addr + 0);
+    res.hi = *(__m128 *)(mem_addr + 4);
+    return res;
+}
+#define _mm256_load_ps _mm256_loadu_ps
+
+
+AVX2NEON_ABI
+int _mm256_testz_ps (const __m256& a, const __m256& b)
+{
+    __m256 t = a;
+    if (&a != &b)
+        t = _mm256_and_ps(a,b);
+
+    __m128i l  = vshrq_n_s32(__m128i(t.lo),31);
+    __m128i h  = vshrq_n_s32(__m128i(t.hi),31);
+    return vaddvq_s32(vaddq_s32(l,h)) == 0;
+}
+
+
+AVX2NEON_ABI
+__m256i _mm256_set_epi64x (int64_t e3, int64_t e2, int64_t e1, int64_t e0)
+{
+    __m256i res;
+    int64x2_t t0 = {e0,e1};
+    int64x2_t t1 = {e2,e3};
+    res.lo = __m128i(t0);
+    res.hi = __m128i(t1);
+    return res;
+}
+
+AVX2NEON_ABI
+__m256d _mm256_setzero_pd ()
+{
+    __m256d res;
+    res.lo = res.hi = vdupq_n_f64(0);
+    return res;
+}
+
+AVX2NEON_ABI
+int _mm256_movemask_pd (__m256d a)
+{
+    int res = 0;
+    uint64x2_t x;
+    x = uint64x2_t(a.lo);
+    res |= (x[0] >> 63) ? 1 : 0;
+    res |= (x[0] >> 63) ? 2 : 0;
+    x = uint64x2_t(a.hi);
+    res |= (x[0] >> 63) ? 4 : 0;
+    res |= (x[0] >> 63) ? 8 : 0;
+    return res;
+}
+
+AVX2NEON_ABI
+__m256i _mm256_cmpeq_epi64 (__m256i a, __m256i b)
+{
+    __m256i res;
+    res.lo = __m128i(vceqq_s64(int64x2_t(a.lo),int64x2_t(b.lo)));
+    res.hi = __m128i(vceqq_s64(int64x2_t(a.hi),int64x2_t(b.hi)));
+    return res;
+}
+
+AVX2NEON_ABI
+__m256i _mm256_cmpeq_pd (__m256d a, __m256d b)
+{
+    __m256i res;
+    res.lo = __m128i(vceqq_f64(a.lo,b.lo));
+    res.hi = __m128i(vceqq_f64(a.hi,b.hi));
+    return res;
+}
+
+
+AVX2NEON_ABI
+int _mm256_testz_pd (const __m256d& a, const __m256d& b)
+{
+    __m256d t = a;
+
+    if (&a != &b)
+        t = _mm256_and_pd(a,b);
+
+    return _mm256_movemask_pd(t) == 0;
+}
+
+AVX2NEON_ABI
+__m256d _mm256_blendv_pd (__m256d a, __m256d b, __m256d mask)
+{
+    __m256d res;
+    uint64x2_t t = uint64x2_t(mask.lo);
+    res.lo[0] = (t[0] >> 63) ? b.lo[0] : a.lo[0];
+    res.lo[1] = (t[1] >> 63) ? b.lo[1] : a.lo[1];
+    t = uint64x2_t(mask.hi);
+    res.hi[0] = (t[0] >> 63) ? b.hi[0] : a.hi[0];
+    res.hi[1] = (t[1] >> 63) ? b.hi[1] : a.hi[1];
+    return res;
+}
+
+template<int imm8>
+__m256 __mm256_dp_ps (__m256 a, __m256 b)
+{
+    __m256 res;
+    res.lo = _mm_dp_ps(a.lo,b.lo,imm8);
+    res.hi = _mm_dp_ps(a.hi,b.hi,imm8);
+    return res;
+}
+
+#define _mm256_dp_ps(a,b,c) __mm256_dp_ps<c>(a,b)
+
+AVX2NEON_ABI
+double _mm256_permute4x64_pd_select(__m256d a, const int imm8)
+{
+    switch (imm8 & 3) {
+        case 0:
+            return a.lo[0];
+        case 1:
+            return a.lo[1];
+        case 2:
+            return a.hi[0];
+        case 3:
+            return a.hi[1];
+    }
+    __builtin_unreachable();
+    return 0;
+}
+
+AVX2NEON_ABI
+__m256d _mm256_permute4x64_pd (__m256d a, const int imm8)
+{
+    __m256d res;
+    res.lo[0] = _mm256_permute4x64_pd_select(a,imm8 >> 0);
+    res.lo[1] = _mm256_permute4x64_pd_select(a,imm8 >> 2);
+    res.hi[0] = _mm256_permute4x64_pd_select(a,imm8 >> 4);
+    res.hi[1] = _mm256_permute4x64_pd_select(a,imm8 >> 6);
+    
+    return res;
+}
+
+AVX2NEON_ABI
+__m256i _mm256_insertf128_si256 (__m256i a, __m128i b, int imm8)
+{
+    return __m256i(_mm256_insertf128_ps((__m256)a,(__m128)b,imm8));
+}
+
+
+AVX2NEON_ABI
+__m256i _mm256_loadu_si256 (__m256i const * mem_addr)
+{
+    __m256i res;
+    res.lo = *(__m128i *)((int32_t *)mem_addr + 0);
+    res.hi = *(__m128i *)((int32_t *)mem_addr + 4);
+    return res;
+}
+
+#define _mm256_load_si256 _mm256_loadu_si256
+
+AVX2NEON_ABI
+void _mm256_storeu_ps (float * mem_addr, __m256 a)
+{
+    *(__m128 *)(mem_addr + 0) = a.lo;
+    *(__m128 *)(mem_addr + 4) = a.hi;
+
+}
+
+#define _mm256_store_ps _mm256_storeu_ps
+#define _mm256_stream_ps _mm256_storeu_ps
+
+
+AVX2NEON_ABI
+void _mm256_storeu_si256 (__m256i * mem_addr, __m256i a)
+{
+    *(__m128i *)((int *)mem_addr + 0) = a.lo;
+    *(__m128i *)((int *)mem_addr + 4) = a.hi;
+
+}
+
+#define _mm256_store_si256 _mm256_storeu_si256
+
+
+
+AVX2NEON_ABI
+__m256 _mm256_maskload_ps (float const * mem_addr, __m256i mask)
+{
+    __m256 res;
+    res.lo = _mm_maskload_ps(mem_addr,mask.lo);
+    res.hi = _mm_maskload_ps(mem_addr + 4,mask.hi);
+    return res;
+
+}
+
+
+AVX2NEON_ABI
+__m256i _mm256_cvtepu8_epi32 (__m128i a)
+{
+    __m256i res;
+    uint8x16_t x = uint8x16_t(a);
+    for (int i=0;i<4;i++)
+    {
+        res.lo[i] = x[i];
+        res.hi[i] = x[i+4];
+    }
+    return res;
+}
+
+
+AVX2NEON_ABI
+__m256i _mm256_cvtepi8_epi32 (__m128i a)
+{
+    __m256i res;
+    int8x16_t x = int8x16_t(a);
+    for (int i=0;i<4;i++)
+    {
+        res.lo[i] = x[i];
+        res.hi[i] = x[i+4];
+    }
+    return res;
+}
+
+
+AVX2NEON_ABI
+__m256i _mm256_cvtepu16_epi32 (__m128i a)
+{
+    __m256i res;
+    uint16x8_t x = uint16x8_t(a);
+    for (int i=0;i<4;i++)
+    {
+        res.lo[i] = x[i];
+        res.hi[i] = x[i+4];
+    }
+    return res;
+}
+
+AVX2NEON_ABI
+__m256i _mm256_cvtepi16_epi32 (__m128i a)
+{
+    __m256i res;
+    int16x8_t x = int16x8_t(a);
+    for (int i=0;i<4;i++)
+    {
+        res.lo[i] = x[i];
+        res.hi[i] = x[i+4];
+    }
+    return res;
+}
+
+
+
+AVX2NEON_ABI
+void _mm256_maskstore_epi32 (int* mem_addr, __m256i mask, __m256i a)
+{
+    _mm_maskstore_epi32(mem_addr,mask.lo,a.lo);
+    _mm_maskstore_epi32(mem_addr + 4,mask.hi,a.hi);
+}
+
+AVX2NEON_ABI
+__m256i _mm256_slli_epi32 (__m256i a, int imm8)
+{
+    __m256i res;
+    res.lo = _mm_slli_epi32(a.lo,imm8);
+    res.hi = _mm_slli_epi32(a.hi,imm8);
+    return res;
+}
+
+
+AVX2NEON_ABI
+__m256i _mm256_srli_epi32 (__m256i a, int imm8)
+{
+    __m256i res;
+    res.lo = _mm_srli_epi32(a.lo,imm8);
+    res.hi = _mm_srli_epi32(a.hi,imm8);
+    return res;
+}
+
+AVX2NEON_ABI
+__m256i _mm256_srai_epi32 (__m256i a, int imm8)
+{
+    __m256i res;
+    res.lo = _mm_srai_epi32(a.lo,imm8);
+    res.hi = _mm_srai_epi32(a.hi,imm8);
+    return res;
+}
+
+
+AVX2NEON_ABI
+__m256i _mm256_sllv_epi32 (__m256i a, __m256i count)
+{
+    __m256i res;
+    res.lo = vshlq_s32(a.lo,count.lo);
+    res.hi = vshlq_s32(a.hi,count.hi);
+    return res;
+
+}
+
+
+AVX2NEON_ABI
+__m256i _mm256_srav_epi32 (__m256i a, __m256i count)
+{
+    __m256i res;
+    res.lo = vshlq_s32(a.lo,vnegq_s32(count.lo));
+    res.hi = vshlq_s32(a.hi,vnegq_s32(count.hi));
+    return res;
+
+}
+
+AVX2NEON_ABI
+__m256i _mm256_srlv_epi32 (__m256i a, __m256i count)
+{
+    __m256i res;
+    res.lo = __m128i(vshlq_u32(uint32x4_t(a.lo),vnegq_s32(count.lo)));
+    res.hi = __m128i(vshlq_u32(uint32x4_t(a.hi),vnegq_s32(count.hi)));
+    return res;
+
+}
+
+
+AVX2NEON_ABI
+__m256i _mm256_permute2f128_si256 (__m256i a, __m256i b, int imm8)
+{
+    return __m256i(_mm256_permute2f128_ps(__m256(a),__m256(b),imm8));
+}
+
+
+AVX2NEON_ABI
+__m128i _mm256_extractf128_si256 (__m256i a, const int imm8)
+{
+    if (imm8 & 1) return a.hi;
+    return a.lo;
+}
+
+AVX2NEON_ABI
+__m256 _mm256_set1_ps(float x)
+{
+    __m256 res;
+    res.lo = res.hi = vdupq_n_f32(x);
+    return res;
+}
+
+AVX2NEON_ABI
+__m256 _mm256_set_ps (float e7, float e6, float e5, float e4, float e3, float e2, float e1, float e0)
+{
+    __m256 res;
+    res.lo = _mm_set_ps(e3,e2,e1,e0);
+    res.hi = _mm_set_ps(e7,e6,e5,e4);
+    return res;
+}
+
+AVX2NEON_ABI
+__m256 _mm256_broadcast_ps (__m128 const * mem_addr)
+{
+    __m256 res;
+    res.lo = res.hi = *mem_addr;
+    return res;
+}
+
+AVX2NEON_ABI
+__m256 _mm256_cvtepi32_ps (__m256i a)
+{
+    __m256 res;
+    res.lo = _mm_cvtepi32_ps(a.lo);
+    res.hi = _mm_cvtepi32_ps(a.hi);
+    return res;
+}
+AVX2NEON_ABI
+void _mm256_maskstore_ps (float * mem_addr, __m256i mask, __m256 a)
+{
+    for (int i=0;i<4;i++) {
+        if (mask.lo[i] & 0x80000000) mem_addr[i] = a.lo[i];
+        if (mask.hi[i] & 0x80000000) mem_addr[i+4] = a.hi[i];
+    }
+}
+
+AVX2NEON_ABI
+__m256d _mm256_andnot_pd (__m256d a, __m256d b)
+{
+    __m256d res;
+    res.lo = float64x2_t(_mm_andnot_ps(__m128(a.lo),__m128(b.lo)));
+    res.hi = float64x2_t(_mm_andnot_ps(__m128(a.hi),__m128(b.hi)));
+    return res;
+}
+
+AVX2NEON_ABI
+__m256 _mm256_blend_ps (__m256 a, __m256 b, const int imm8)
+{
+    __m256 res;
+    res.lo = _mm_blend_ps(a.lo,b.lo,imm8 & 0xf);
+    res.hi = _mm_blend_ps(a.hi,b.hi,imm8 >> 4);
+    return res;
+
+}
+
+
+AVX2NEON_ABI
+__m256i _mm256_blend_epi32 (__m256i a, __m256i b, const int imm8)
+{
+    __m256i res;
+    res.lo = _mm_blend_epi32(a.lo,b.lo,imm8 & 0xf);
+    res.hi = _mm_blend_epi32(a.hi,b.hi,imm8 >> 4);
+    return res;
+
+}
+
+AVX2NEON_ABI
+__m256i _mm256_i32gather_epi32 (int const* base_addr, __m256i vindex, const int scale)
+{
+    __m256i res;
+    for (int i=0;i<4;i++)
+    {
+        res.lo[i] = *(int *)((char *) base_addr + (vindex.lo[i]*scale));
+        res.hi[i] = *(int *)((char *) base_addr + (vindex.hi[i]*scale));
+    }
+    return res;
+}
+
+
+AVX2NEON_ABI
+__m256i _mm256_mask_i32gather_epi32 (__m256i src, int const* base_addr, __m256i vindex, __m256i mask, const int scale)
+{
+    __m256i res = _mm256_setzero_si256();
+    for (int i=0;i<4;i++)
+    {
+        if (mask.lo[i] >> 31) res.lo[i] = *(int *)((char *) base_addr + (vindex.lo[i]*scale));
+        if (mask.hi[i] >> 31) res.hi[i] = *(int *)((char *) base_addr + (vindex.hi[i]*scale));
+    }
+    
+    return res;
+
+}
+
+

thirdparty/embree/common/math/bbox.h

@@ -77,7 +77,7 @@ namespace embree
     return lower > upper;
   }
 
-#if defined(__SSE__)
+#if defined(__SSE__) || defined(__ARM_NEON)
   template<> __forceinline bool BBox<Vec3fa>::empty() const {
     return !all(le_mask(lower,upper));
   }
@@ -228,11 +228,11 @@ namespace embree
 /// SSE / AVX / MIC specializations
 ////////////////////////////////////////////////////////////////////////////////
 
-#if defined __SSE__
+#if defined (__SSE__) || defined(__ARM_NEON)
 #include "../simd/sse.h"
 #endif
 
-#if defined __AVX__
+#if defined (__AVX__)
 #include "../simd/avx.h"
 #endif
 

thirdparty/embree/common/math/col3.h

@@ -42,6 +42,6 @@ namespace embree
   }
 
   /*! default template instantiations */
-  typedef Col3<unsigned char> Col3uc;
+  typedef Col3<uint8_t      > Col3uc;
   typedef Col3<float        > Col3f;
 }

thirdparty/embree/common/math/col4.h

@@ -42,6 +42,6 @@ namespace embree
   }
 
   /*! default template instantiations */
-  typedef Col4<unsigned char> Col4uc;
+  typedef Col4<uint8_t      > Col4uc;
   typedef Col4<float        > Col4f;
 }

thirdparty/embree/common/math/color.h

@@ -52,17 +52,17 @@ namespace embree
     __forceinline void set(Col3uc& d) const 
     {
       vfloat4 s = clamp(vfloat4(m128))*255.0f;
-      d.r = (unsigned char)(s[0]); 
+      d.r = (uint8_t)(s[0]); 
-      d.g = (unsigned char)(s[1]); 
+      d.g = (uint8_t)(s[1]); 
-      d.b = (unsigned char)(s[2]); 
+      d.b = (uint8_t)(s[2]); 
     }
     __forceinline void set(Col4uc& d) const 
     {
       vfloat4 s = clamp(vfloat4(m128))*255.0f;
-      d.r = (unsigned char)(s[0]); 
+      d.r = (uint8_t)(s[0]); 
-      d.g = (unsigned char)(s[1]); 
+      d.g = (uint8_t)(s[1]); 
-      d.b = (unsigned char)(s[2]); 
+      d.b = (uint8_t)(s[2]); 
-      d.a = (unsigned char)(s[3]); 
+      d.a = (uint8_t)(s[3]); 
     }
 
     ////////////////////////////////////////////////////////////////////////////////
@@ -114,16 +114,16 @@ namespace embree
     __forceinline void set(Col3uc& d) const 
     { 
       vfloat4 s = clamp(vfloat4(m128))*255.0f;
-      d.r = (unsigned char)(s[0]); 
+      d.r = (uint8_t)(s[0]); 
-      d.g = (unsigned char)(s[1]); 
+      d.g = (uint8_t)(s[1]); 
-      d.b = (unsigned char)(s[2]); 
+      d.b = (uint8_t)(s[2]); 
     }
     __forceinline void set(Col4uc& d) const 
     { 
       vfloat4 s = clamp(vfloat4(m128))*255.0f;
-      d.r = (unsigned char)(s[0]); 
+      d.r = (uint8_t)(s[0]); 
-      d.g = (unsigned char)(s[1]); 
+      d.g = (uint8_t)(s[1]); 
-      d.b = (unsigned char)(s[2]); 
+      d.b = (uint8_t)(s[2]); 
       d.a = 255; 
     }
 
@@ -152,21 +152,37 @@ namespace embree
   }
   __forceinline const Color rcp  ( const Color& a )
   {
+#if defined(__aarch64__) && defined(BUILD_IOS)
+    __m128 reciprocal = _mm_rcp_ps(a.m128);
+    reciprocal = vmulq_f32(vrecpsq_f32(a.m128, reciprocal), reciprocal);
+    reciprocal = vmulq_f32(vrecpsq_f32(a.m128, reciprocal), reciprocal);
+    return (const Color)reciprocal;
+#else
 #if defined(__AVX512VL__)
     const Color r = _mm_rcp14_ps(a.m128);
 #else
     const Color r = _mm_rcp_ps(a.m128);
 #endif
     return _mm_sub_ps(_mm_add_ps(r, r), _mm_mul_ps(_mm_mul_ps(r, r), a));
+#endif  //defined(__aarch64__) && defined(BUILD_IOS)
   }
   __forceinline const Color rsqrt( const Color& a )
   {
+#if defined(__aarch64__) && defined(BUILD_IOS)
+    __m128 r = _mm_rsqrt_ps(a.m128);
+    r = vmulq_f32(r, vrsqrtsq_f32(vmulq_f32(a.m128, r), r));
+    r = vmulq_f32(r, vrsqrtsq_f32(vmulq_f32(a.m128, r), r));
+    return r;
+#else
+      
 #if defined(__AVX512VL__)
     __m128 r = _mm_rsqrt14_ps(a.m128);
 #else
     __m128 r = _mm_rsqrt_ps(a.m128);
 #endif
     return _mm_add_ps(_mm_mul_ps(_mm_set1_ps(1.5f),r), _mm_mul_ps(_mm_mul_ps(_mm_mul_ps(a, _mm_set1_ps(-0.5f)), r), _mm_mul_ps(r, r)));
+      
+#endif  //defined(__aarch64__) && defined(BUILD_IOS)
   }
   __forceinline const Color sqrt ( const Color& a ) { return _mm_sqrt_ps(a.m128); }
 

thirdparty/embree/common/math/constants.cpp

@@ -1,6 +1,10 @@
 // Copyright 2009-2020 Intel Corporation
 // SPDX-License-Identifier: Apache-2.0
 
+#if defined(__aarch64__)
+#include <arm_neon.h>
+#endif
+
 #include "constants.h"
 
 namespace embree
@@ -24,4 +28,34 @@ namespace embree
   ReverseStepTy reverse_step;
   EmptyTy empty;
   UndefinedTy undefined;
+
+#if defined(__aarch64__)
+const uint32x4_t movemask_mask = { 1, 2, 4, 8 };
+const uint32x4_t vzero = { 0, 0, 0, 0 };
+const uint32x4_t v0x80000000 = { 0x80000000, 0x80000000, 0x80000000, 0x80000000 };
+const uint32x4_t v0x7fffffff = { 0x7fffffff, 0x7fffffff, 0x7fffffff, 0x7fffffff };
+const uint32x4_t v000F = { 0x00000000, 0x00000000, 0x00000000, 0xFFFFFFFF };
+const uint32x4_t v00F0 = { 0x00000000, 0x00000000, 0xFFFFFFFF, 0x00000000 };
+const uint32x4_t v00FF = { 0x00000000, 0x00000000, 0xFFFFFFFF, 0xFFFFFFFF };
+const uint32x4_t v0F00 = { 0x00000000, 0xFFFFFFFF, 0x00000000, 0x00000000 };
+const uint32x4_t v0F0F = { 0x00000000, 0xFFFFFFFF, 0x00000000, 0xFFFFFFFF };
+const uint32x4_t v0FF0 = { 0x00000000, 0xFFFFFFFF, 0xFFFFFFFF, 0x00000000 };
+const uint32x4_t v0FFF = { 0x00000000, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF };
+const uint32x4_t vF000 = { 0xFFFFFFFF, 0x00000000, 0x00000000, 0x00000000 };
+const uint32x4_t vF00F = { 0xFFFFFFFF, 0x00000000, 0x00000000, 0xFFFFFFFF };
+const uint32x4_t vF0F0 = { 0xFFFFFFFF, 0x00000000, 0xFFFFFFFF, 0x00000000 };
+const uint32x4_t vF0FF = { 0xFFFFFFFF, 0x00000000, 0xFFFFFFFF, 0xFFFFFFFF };
+const uint32x4_t vFF00 = { 0xFFFFFFFF, 0xFFFFFFFF, 0x00000000, 0x00000000 };
+const uint32x4_t vFF0F = { 0xFFFFFFFF, 0xFFFFFFFF, 0x00000000, 0xFFFFFFFF };
+const uint32x4_t vFFF0 = { 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0x00000000 };
+const uint32x4_t vFFFF = { 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF };
+const uint8x16_t v0022 = {0,1,2,3, 0,1,2,3, 8,9,10,11, 8,9,10,11};
+const uint8x16_t v1133 = {4,5,6,7, 4,5,6,7, 12,13,14,15, 12,13,14,15};
+const uint8x16_t v0101 = {0,1,2,3, 4,5,6,7, 0,1,2,3, 4,5,6,7};
+const float32x4_t vOne = { 1.0f, 1.0f, 1.0f, 1.0f };
+const float32x4_t vmOne = { -1.0f, -1.0f, -1.0f, -1.0f };
+const float32x4_t vInf = { INFINITY, INFINITY, INFINITY, INFINITY };
+const float32x4_t vmInf = { -INFINITY, -INFINITY, -INFINITY, -INFINITY };
+#endif
+
 }

thirdparty/embree/common/math/constants.h

@@ -12,6 +12,19 @@
 #include <cfloat>
 #include <climits>
 
+// Math constants may not be defined in libcxx + mingw + strict C++ standard
+#if defined(__MINGW32__)
+
+// TODO(LTE): use constexpr
+#ifndef M_PI
+#define M_PI 3.14159265358979323846
+#endif
+#ifndef M_1_PI
+#define M_1_PI 0.31830988618379067154
+#endif
+
+#endif // __MINGW32__
+
 namespace embree
 {
   static MAYBE_UNUSED const float one_over_255 = 1.0f/255.0f;
@@ -44,8 +57,8 @@ namespace embree
     __forceinline operator unsigned int      ( ) const { return 0; }
     __forceinline operator          short    ( ) const { return 0; }
     __forceinline operator unsigned short    ( ) const { return 0; }
-    __forceinline operator          char     ( ) const { return 0; }
+    __forceinline operator          int8_t     ( ) const { return 0; }
-    __forceinline operator unsigned char     ( ) const { return 0; }
+    __forceinline operator uint8_t     ( ) const { return 0; }
   }; 
 
   extern MAYBE_UNUSED ZeroTy zero;
@@ -62,8 +75,8 @@ namespace embree
     __forceinline operator unsigned int      ( ) const { return 1; }
     __forceinline operator          short    ( ) const { return 1; }
     __forceinline operator unsigned short    ( ) const { return 1; }
-    __forceinline operator          char     ( ) const { return 1; }
+    __forceinline operator          int8_t     ( ) const { return 1; }
-    __forceinline operator unsigned char     ( ) const { return 1; }
+    __forceinline operator uint8_t     ( ) const { return 1; }
   };
 
   extern MAYBE_UNUSED OneTy one;
@@ -80,8 +93,8 @@ namespace embree
     __forceinline operator unsigned int      ( ) const { return std::numeric_limits<unsigned int>::min(); }
     __forceinline operator          short    ( ) const { return std::numeric_limits<short>::min(); }
     __forceinline operator unsigned short    ( ) const { return std::numeric_limits<unsigned short>::min(); }
-    __forceinline operator          char     ( ) const { return std::numeric_limits<char>::min(); }
+    __forceinline operator          int8_t     ( ) const { return std::numeric_limits<int8_t>::min(); }
-    __forceinline operator unsigned char     ( ) const { return std::numeric_limits<unsigned char>::min(); }
+    __forceinline operator uint8_t     ( ) const { return std::numeric_limits<uint8_t>::min(); }
 
   };
 
@@ -99,8 +112,8 @@ namespace embree
     __forceinline operator unsigned int      ( ) const { return std::numeric_limits<unsigned int>::max(); }
     __forceinline operator          short    ( ) const { return std::numeric_limits<short>::max(); }
     __forceinline operator unsigned short    ( ) const { return std::numeric_limits<unsigned short>::max(); }
-    __forceinline operator          char     ( ) const { return std::numeric_limits<char>::max(); }
+    __forceinline operator          int8_t     ( ) const { return std::numeric_limits<int8_t>::max(); }
-    __forceinline operator unsigned char     ( ) const { return std::numeric_limits<unsigned char>::max(); }
+    __forceinline operator uint8_t     ( ) const { return std::numeric_limits<uint8_t>::max(); }
   };
 
   extern MAYBE_UNUSED PosInfTy inf;
@@ -194,4 +207,33 @@ namespace embree
   };
 
   extern MAYBE_UNUSED UndefinedTy undefined;
+    
+#if defined(__aarch64__)
+  extern const uint32x4_t movemask_mask;
+  extern const uint32x4_t vzero;
+  extern const uint32x4_t v0x80000000;
+  extern const uint32x4_t v0x7fffffff;
+  extern const uint32x4_t v000F;
+  extern const uint32x4_t v00F0;
+  extern const uint32x4_t v00FF;
+  extern const uint32x4_t v0F00;
+  extern const uint32x4_t v0F0F;
+  extern const uint32x4_t v0FF0;
+  extern const uint32x4_t v0FFF;
+  extern const uint32x4_t vF000;
+  extern const uint32x4_t vF00F;
+  extern const uint32x4_t vF0F0;
+  extern const uint32x4_t vF0FF;
+  extern const uint32x4_t vFF00;
+  extern const uint32x4_t vFF0F;
+  extern const uint32x4_t vFFF0;
+  extern const uint32x4_t vFFFF;
+  extern const uint8x16_t v0022;
+  extern const uint8x16_t v1133;
+  extern const uint8x16_t v0101;
+  extern const float32x4_t vOne;
+  extern const float32x4_t vmOne;
+  extern const float32x4_t vInf;
+  extern const float32x4_t vmInf;
+#endif
 }

thirdparty/embree/common/math/math.h

@@ -8,12 +8,19 @@
 #include "constants.h"
 #include <cmath>
 
+#if defined(__ARM_NEON)
+#include "SSE2NEON.h"
+#if defined(NEON_AVX2_EMULATION)
+#include "AVX2NEON.h"
+#endif
+#else
 #include <emmintrin.h>
 #include <xmmintrin.h>
 #include <immintrin.h>
+#endif
 
-#if defined(__WIN32__)
+#if defined(__WIN32__) && !defined(__MINGW32__)
-#if defined(_MSC_VER) && (_MSC_VER <= 1700)
+#if (__MSV_VER <= 1700)
 namespace std
 {
   __forceinline bool isinf ( const float x ) { return _finite(x) == 0; }
@@ -40,7 +47,7 @@ namespace embree
   __forceinline int   toInt  (const float& a) { return int(a); }
   __forceinline float toFloat(const int&   a) { return float(a); }
 
-#if defined(__WIN32__)
+#if defined(__WIN32__) && !defined(__MINGW32__)
   __forceinline bool finite ( const float x ) { return _finite(x) != 0; }
 #endif
 
@@ -49,6 +56,16 @@ namespace embree
 
   __forceinline float rcp  ( const float x )
   {
+#if defined(__aarch64__)
+      // Move scalar to vector register and do rcp.
+      __m128 a;
+      a[0] = x;
+      float32x4_t reciprocal = vrecpeq_f32(a);
+      reciprocal = vmulq_f32(vrecpsq_f32(a, reciprocal), reciprocal);
+      reciprocal = vmulq_f32(vrecpsq_f32(a, reciprocal), reciprocal);
+      return reciprocal[0];
+#else
+
     const __m128 a = _mm_set_ss(x);
 
 #if defined(__AVX512VL__)
@@ -62,19 +79,61 @@ namespace embree
 #else
     return _mm_cvtss_f32(_mm_mul_ss(r,_mm_sub_ss(_mm_set_ss(2.0f), _mm_mul_ss(r, a))));
 #endif
+
+#endif  //defined(__aarch64__)
   }
 
   __forceinline float signmsk ( const float x ) {
+#if defined(__aarch64__)
+      // FP and Neon shares same vector register in arm64
+      __m128 a;
+      __m128i b;
+      a[0] = x;
+      b[0] = 0x80000000;
+      a = _mm_and_ps(a, vreinterpretq_f32_s32(b));
+      return a[0];
+#else
     return _mm_cvtss_f32(_mm_and_ps(_mm_set_ss(x),_mm_castsi128_ps(_mm_set1_epi32(0x80000000))));
+#endif
   }
   __forceinline float xorf( const float x, const float y ) {
+#if defined(__aarch64__)
+      // FP and Neon shares same vector register in arm64
+      __m128 a;
+      __m128 b;
+      a[0] = x;
+      b[0] = y;
+      a = _mm_xor_ps(a, b);
+      return a[0];
+#else
     return _mm_cvtss_f32(_mm_xor_ps(_mm_set_ss(x),_mm_set_ss(y)));
+#endif
   }
   __forceinline float andf( const float x, const unsigned y ) {
+#if defined(__aarch64__) 
+      // FP and Neon shares same vector register in arm64
+      __m128 a;
+      __m128i b;
+      a[0] = x;
+      b[0] = y;
+      a = _mm_and_ps(a, vreinterpretq_f32_s32(b));
+      return a[0];
+#else
     return _mm_cvtss_f32(_mm_and_ps(_mm_set_ss(x),_mm_castsi128_ps(_mm_set1_epi32(y))));
+#endif
   }
   __forceinline float rsqrt( const float x )
   {
+#if defined(__aarch64__)
+      // FP and Neon shares same vector register in arm64
+      __m128 a;
+      a[0] = x;
+      __m128 value = _mm_rsqrt_ps(a);
+      value = vmulq_f32(value, vrsqrtsq_f32(vmulq_f32(a, value), value));
+      value = vmulq_f32(value, vrsqrtsq_f32(vmulq_f32(a, value), value));
+      return value[0];
+#else
+
     const __m128 a = _mm_set_ss(x);
 #if defined(__AVX512VL__)
     const __m128 r = _mm_rsqrt14_ss(_mm_set_ss(0.0f),a);
@@ -84,9 +143,10 @@ namespace embree
     const __m128 c = _mm_add_ss(_mm_mul_ss(_mm_set_ss(1.5f), r),
                                 _mm_mul_ss(_mm_mul_ss(_mm_mul_ss(a, _mm_set_ss(-0.5f)), r), _mm_mul_ss(r, r)));
     return _mm_cvtss_f32(c);
+#endif
   }
 
-#if defined(__WIN32__) && defined(_MSC_VER) && (_MSC_VER <= 1700)
+#if defined(__WIN32__) && (__MSC_VER <= 1700) && !defined(__MINGW32__)
   __forceinline float nextafter(float x, float y) { if ((x<y) == (x>0)) return x*(1.1f+float(ulp)); else return x*(0.9f-float(ulp)); }
   __forceinline double nextafter(double x, double y) { return _nextafter(x, y); }
   __forceinline int roundf(float f) { return (int)(f + 0.5f); }
@@ -140,7 +200,17 @@ namespace embree
   __forceinline double floor( const double x ) { return ::floor (x); }
   __forceinline double ceil ( const double x ) { return ::ceil (x); }
 
-#if defined(__SSE4_1__)
+#if defined(__aarch64__) 
+    __forceinline float mini(float a, float b) {
+        // FP and Neon shares same vector register in arm64
+        __m128 x;
+        __m128 y;
+        x[0] = a;
+        y[0] = b;
+        x = _mm_min_ps(x, y);
+        return x[0];
+    }
+#elif defined(__SSE4_1__)
   __forceinline float mini(float a, float b) {
     const __m128i ai = _mm_castps_si128(_mm_set_ss(a));
     const __m128i bi = _mm_castps_si128(_mm_set_ss(b));
@@ -149,7 +219,17 @@ namespace embree
   }
 #endif
 
-#if defined(__SSE4_1__)
+#if defined(__aarch64__) 
+    __forceinline float maxi(float a, float b) {
+        // FP and Neon shares same vector register in arm64
+        __m128 x;
+        __m128 y;
+        x[0] = a;
+        y[0] = b;
+        x = _mm_max_ps(x, y);
+        return x[0];
+    }
+#elif defined(__SSE4_1__)
   __forceinline float maxi(float a, float b) {
     const __m128i ai = _mm_castps_si128(_mm_set_ss(a));
     const __m128i bi = _mm_castps_si128(_mm_set_ss(b));
@@ -166,7 +246,7 @@ namespace embree
   __forceinline  int64_t min(int64_t  a, int64_t  b) { return a<b ? a:b; }
   __forceinline    float min(float    a, float    b) { return a<b ? a:b; }
   __forceinline   double min(double   a, double   b) { return a<b ? a:b; }
-#if defined(__X86_64__)
+#if defined(__X86_64__) || defined(__aarch64__)
   __forceinline   size_t min(size_t   a, size_t   b) { return a<b ? a:b; }
 #endif
 
@@ -183,7 +263,7 @@ namespace embree
   __forceinline  int64_t max(int64_t  a, int64_t  b) { return a<b ? b:a; }
   __forceinline    float max(float    a, float    b) { return a<b ? b:a; }
   __forceinline   double max(double   a, double   b) { return a<b ? b:a; }
-#if defined(__X86_64__)
+#if defined(__X86_64__) || defined(__aarch64__)
   __forceinline   size_t max(size_t   a, size_t   b) { return a<b ? b:a; }
 #endif
 
@@ -225,6 +305,16 @@ namespace embree
   __forceinline float msub  ( const float a, const float b, const float c) { return _mm_cvtss_f32(_mm_fmsub_ss(_mm_set_ss(a),_mm_set_ss(b),_mm_set_ss(c))); }
   __forceinline float nmadd ( const float a, const float b, const float c) { return _mm_cvtss_f32(_mm_fnmadd_ss(_mm_set_ss(a),_mm_set_ss(b),_mm_set_ss(c))); }
   __forceinline float nmsub ( const float a, const float b, const float c) { return _mm_cvtss_f32(_mm_fnmsub_ss(_mm_set_ss(a),_mm_set_ss(b),_mm_set_ss(c))); }
+#elif defined (__aarch64__) && defined(__clang__)
+#pragma clang fp contract(fast)
+
+
+__forceinline float madd  ( const float a, const float b, const float c) { return a*b + c; }
+__forceinline float msub  ( const float a, const float b, const float c) { return a*b - c; }
+__forceinline float nmadd ( const float a, const float b, const float c) { return c - a*b; }
+__forceinline float nmsub ( const float a, const float b, const float c) { return -(c + a*b); }
+
+#pragma clang fp contract(on)
 #else
   __forceinline float madd  ( const float a, const float b, const float c) { return a*b+c; }
   __forceinline float msub  ( const float a, const float b, const float c) { return a*b-c; }
@@ -273,6 +363,15 @@ namespace embree
   /*! exchange */
   template<typename T> __forceinline void xchg ( T& a, T& b ) { const T tmp = a; a = b; b = tmp; }
 
+
+  template<typename T> __forceinline T prod_diff(const T& a,const T& b,const T& c,const T& d) {
+#if 1//!defined(__aarch64__)
+      return msub(a,b,c*d);
+#else
+      return nmadd(c,d,a*b);
+#endif
+  }
+
   /*! bit reverse operation */
   template<class T>
     __forceinline T bitReverse(const T& vin)
@@ -290,7 +389,7 @@ namespace embree
   template<class T>
     __forceinline T bitInterleave(const T& xin, const T& yin, const T& zin)
   {
-	T x = xin, y = yin, z = zin;
+	  T x = xin, y = yin, z = zin;
     x = (x | (x << 16)) & 0x030000FF;
     x = (x | (x <<  8)) & 0x0300F00F;
     x = (x | (x <<  4)) & 0x030C30C3;
@@ -309,7 +408,7 @@ namespace embree
     return x | (y << 1) | (z << 2);
   }
 
-#if defined(__AVX2__)
+#if defined(__AVX2__) && !defined(__aarch64__)
 
   template<>
     __forceinline unsigned int bitInterleave(const unsigned int &xi, const unsigned int& yi, const unsigned int& zi)

thirdparty/embree/common/math/vec2.h

@@ -205,11 +205,11 @@ namespace embree
 
 #include "vec2fa.h"
 
-#if defined __SSE__
+#if defined(__SSE__) || defined(__ARM_NEON)
 #include "../simd/sse.h"
 #endif
 
-#if defined __AVX__
+#if defined(__AVX__)
 #include "../simd/avx.h"
 #endif
 
@@ -221,7 +221,7 @@ namespace embree
 {
   template<> __forceinline Vec2<float>::Vec2(const Vec2fa& a) : x(a.x), y(a.y) {}
 
-#if defined(__SSE__)
+#if defined(__SSE__) || defined(__ARM_NEON)
   template<> __forceinline Vec2<vfloat4>::Vec2(const Vec2fa& a) : x(a.x), y(a.y) {}
 #endif
 

thirdparty/embree/common/math/vec2fa.h

@@ -97,6 +97,12 @@ namespace embree
 
   __forceinline Vec2fa rcp  ( const Vec2fa& a )
   {
+#if defined(__aarch64__)
+        __m128 reciprocal = _mm_rcp_ps(a.m128);
+        reciprocal = vmulq_f32(vrecpsq_f32(a.m128, reciprocal), reciprocal);
+        reciprocal = vmulq_f32(vrecpsq_f32(a.m128, reciprocal), reciprocal);
+        return (const Vec2fa)reciprocal;
+#else
 #if defined(__AVX512VL__)
     const Vec2fa r = _mm_rcp14_ps(a.m128);
 #else
@@ -111,6 +117,7 @@ namespace embree
 #endif
 
     return res;
+#endif  //defined(__aarch64__) 
   }
 
   __forceinline Vec2fa sqrt ( const Vec2fa& a ) { return _mm_sqrt_ps(a.m128); }
@@ -118,12 +125,21 @@ namespace embree
 
   __forceinline Vec2fa rsqrt( const Vec2fa& a )
   {
+#if defined(__aarch64__)
+        __m128 r = _mm_rsqrt_ps(a.m128);
+        r = vmulq_f32(r, vrsqrtsq_f32(vmulq_f32(a.m128, r), r));
+        r = vmulq_f32(r, vrsqrtsq_f32(vmulq_f32(a.m128, r), r));
+        return r;
+#else
+        
 #if defined(__AVX512VL__)
     __m128 r = _mm_rsqrt14_ps(a.m128);
 #else
     __m128 r = _mm_rsqrt_ps(a.m128);
 #endif
     return _mm_add_ps(_mm_mul_ps(_mm_set1_ps(1.5f),r), _mm_mul_ps(_mm_mul_ps(_mm_mul_ps(a, _mm_set1_ps(-0.5f)), r), _mm_mul_ps(r, r)));
+        
+#endif
   }
 
   __forceinline Vec2fa zero_fix(const Vec2fa& a) {
@@ -156,7 +172,7 @@ namespace embree
   __forceinline Vec2fa min( const Vec2fa& a, const Vec2fa& b ) { return _mm_min_ps(a.m128,b.m128); }
   __forceinline Vec2fa max( const Vec2fa& a, const Vec2fa& b ) { return _mm_max_ps(a.m128,b.m128); }
 
-#if defined(__SSE4_1__)
+#if defined(__aarch64__) || defined(__SSE4_1__)
     __forceinline Vec2fa mini(const Vec2fa& a, const Vec2fa& b) {
       const vint4 ai = _mm_castps_si128(a);
       const vint4 bi = _mm_castps_si128(b);
@@ -165,7 +181,7 @@ namespace embree
     }
 #endif
 
-#if defined(__SSE4_1__)
+#if defined(__aarch64__) || defined(__SSE4_1__)
     __forceinline Vec2fa maxi(const Vec2fa& a, const Vec2fa& b) {
       const vint4 ai = _mm_castps_si128(a);
       const vint4 bi = _mm_castps_si128(b);
@@ -275,7 +291,11 @@ namespace embree
   /// Rounding Functions
   ////////////////////////////////////////////////////////////////////////////////
 
-#if defined (__SSE4_1__)
+#if defined(__aarch64__)
+__forceinline Vec2fa floor(const Vec2fa& a) { return vrndmq_f32(a); }
+__forceinline Vec2fa ceil (const Vec2fa& a) { return vrndpq_f32(a); }
+//__forceinline Vec2fa trunc(const Vec2fa& a) { return vrndq_f32(a); }
+#elif defined (__SSE4_1__)
   //__forceinline Vec2fa trunc( const Vec2fa& a ) { return _mm_round_ps(a, _MM_FROUND_TO_NEAREST_INT); }
   __forceinline Vec2fa floor( const Vec2fa& a ) { return _mm_round_ps(a, _MM_FROUND_TO_NEG_INF    ); }
   __forceinline Vec2fa ceil ( const Vec2fa& a ) { return _mm_round_ps(a, _MM_FROUND_TO_POS_INF    ); }

thirdparty/embree/common/math/vec3.h

@@ -206,8 +206,7 @@ namespace embree
   template<typename T> __forceinline T       rcp_length( const Vec3<T>& a )                  { return rsqrt(sqr(a)); }
   template<typename T> __forceinline Vec3<T> normalize( const Vec3<T>& a )                   { return a*rsqrt(sqr(a)); }
   template<typename T> __forceinline T       distance ( const Vec3<T>& a, const Vec3<T>& b ) { return length(a-b); }
-  template<typename T> __forceinline Vec3<T> cross    ( const Vec3<T>& a, const Vec3<T>& b ) { return Vec3<T>(msub(a.y,b.z,a.z*b.y), msub(a.z,b.x,a.x*b.z), msub(a.x,b.y,a.y*b.x)); }
+  template<typename T> __forceinline Vec3<T> cross    ( const Vec3<T>& a, const Vec3<T>& b ) { return Vec3<T>(prod_diff(a.y,b.z,a.z,b.y), prod_diff(a.z,b.x,a.x,b.z), prod_diff(a.x,b.y,a.y,b.x)); }
-
   template<typename T> __forceinline Vec3<T> stable_triangle_normal( const Vec3<T>& a, const Vec3<T>& b, const Vec3<T>& c )
   {
     const T ab_x = a.z*b.y, ab_y = a.x*b.z, ab_z = a.y*b.x;
@@ -266,11 +265,11 @@ namespace embree
 /// SSE / AVX / MIC specializations
 ////////////////////////////////////////////////////////////////////////////////
 
-#if defined __SSE__
+#if defined(__SSE__) || defined(__ARM_NEON)
 #include "../simd/sse.h"
 #endif
 
-#if defined __AVX__
+#if defined(__AVX__)
 #include "../simd/avx.h"
 #endif
 
@@ -291,14 +290,14 @@ namespace embree
   template<> __forceinline Vec3<vfloat4>::Vec3(const Vec3fa& a) {
     x = a.x; y = a.y; z = a.z;
   }
-#elif defined(__SSE__)
+#elif defined(__SSE__) || defined(__ARM_NEON)
   template<>
   __forceinline Vec3<vfloat4>::Vec3(const Vec3fa& a) {
     const vfloat4 v = vfloat4(a.m128); x = shuffle<0,0,0,0>(v); y = shuffle<1,1,1,1>(v); z = shuffle<2,2,2,2>(v);
   }
 #endif
 
-#if defined(__SSE__)
+#if defined(__SSE__) || defined(__ARM_NEON)
   __forceinline Vec3<vfloat4> broadcast4f(const Vec3<vfloat4>& a, const size_t k) {
     return Vec3<vfloat4>(vfloat4::broadcast(&a.x[k]), vfloat4::broadcast(&a.y[k]), vfloat4::broadcast(&a.z[k]));
   }

thirdparty/embree/common/math/vec3fa.h

@@ -55,7 +55,13 @@ namespace embree
     ////////////////////////////////////////////////////////////////////////////////
 
     static __forceinline Vec3fa load( const void* const a ) {
+#if defined(__aarch64__)
+        __m128 t = _mm_load_ps((float*)a);
+        t[3] = 0.0f;
+        return Vec3fa(t);
+#else
       return Vec3fa(_mm_and_ps(_mm_load_ps((float*)a),_mm_castsi128_ps(_mm_set_epi32(0, -1, -1, -1))));
+#endif
     }
 
     static __forceinline Vec3fa loadu( const void* const a ) {
@@ -89,19 +95,42 @@ namespace embree
 
   __forceinline Vec3fa operator +( const Vec3fa& a ) { return a; }
   __forceinline Vec3fa operator -( const Vec3fa& a ) {
+#if defined(__aarch64__)
+    return vnegq_f32(a.m128);
+#else
     const __m128 mask = _mm_castsi128_ps(_mm_set1_epi32(0x80000000));
+
     return _mm_xor_ps(a.m128, mask);
+#endif
   }
   __forceinline Vec3fa abs  ( const Vec3fa& a ) {
+#if defined(__aarch64__)
+    return _mm_abs_ps(a.m128);
+#else
     const __m128 mask = _mm_castsi128_ps(_mm_set1_epi32(0x7fffffff));
     return _mm_and_ps(a.m128, mask);
+#endif
   }
   __forceinline Vec3fa sign ( const Vec3fa& a ) {
+#if defined(__aarch64__)
+    Vec3fa r = blendv_ps(vOne, vmOne, _mm_cmplt_ps (a.m128,vdupq_n_f32(0.0f)));
+    return r;
+#else
     return blendv_ps(Vec3fa(one).m128, (-Vec3fa(one)).m128, _mm_cmplt_ps (a.m128,Vec3fa(zero).m128));
+#endif
   }
 
   __forceinline Vec3fa rcp  ( const Vec3fa& a )
   {
+#if defined(__aarch64__) && defined(BUILD_IOS)
+  return vdivq_f32(vdupq_n_f32(1.0f),a.m128);
+#elif defined(__aarch64__)
+  __m128 reciprocal = _mm_rcp_ps(a.m128);
+  reciprocal = vmulq_f32(vrecpsq_f32(a.m128, reciprocal), reciprocal);
+  reciprocal = vmulq_f32(vrecpsq_f32(a.m128, reciprocal), reciprocal);
+  return (const Vec3fa)reciprocal;
+#else
+        
 #if defined(__AVX512VL__)
     const Vec3fa r = _mm_rcp14_ps(a.m128);
 #else
@@ -116,6 +145,7 @@ namespace embree
 #endif
 
     return res;
+#endif  //defined(__aarch64__)
   }
 
   __forceinline Vec3fa sqrt ( const Vec3fa& a ) { return _mm_sqrt_ps(a.m128); }
@@ -123,12 +153,20 @@ namespace embree
 
   __forceinline Vec3fa rsqrt( const Vec3fa& a )
   {
+#if defined(__aarch64__)
+        __m128 r = _mm_rsqrt_ps(a.m128);
+        r = vmulq_f32(r, vrsqrtsq_f32(vmulq_f32(a.m128, r), r));
+        r = vmulq_f32(r, vrsqrtsq_f32(vmulq_f32(a.m128, r), r));
+        return r;
+#else
+        
 #if defined(__AVX512VL__)
     __m128 r = _mm_rsqrt14_ps(a.m128);
 #else
     __m128 r = _mm_rsqrt_ps(a.m128);
 #endif
     return _mm_add_ps(_mm_mul_ps(_mm_set1_ps(1.5f),r), _mm_mul_ps(_mm_mul_ps(_mm_mul_ps(a.m128, _mm_set1_ps(-0.5f)), r), _mm_mul_ps(r, r)));
+#endif
   }
 
   __forceinline Vec3fa zero_fix(const Vec3fa& a) {
@@ -161,7 +199,7 @@ namespace embree
   __forceinline Vec3fa min( const Vec3fa& a, const Vec3fa& b ) { return _mm_min_ps(a.m128,b.m128); }
   __forceinline Vec3fa max( const Vec3fa& a, const Vec3fa& b ) { return _mm_max_ps(a.m128,b.m128); }
 
-#if defined(__SSE4_1__)
+#if defined(__aarch64__) || defined(__SSE4_1__)
     __forceinline Vec3fa mini(const Vec3fa& a, const Vec3fa& b) {
       const vint4 ai = _mm_castps_si128(a.m128);
       const vint4 bi = _mm_castps_si128(b.m128);
@@ -170,7 +208,7 @@ namespace embree
     }
 #endif
 
-#if defined(__SSE4_1__)
+#if defined(__aarch64__) || defined(__SSE4_1__)
     __forceinline Vec3fa maxi(const Vec3fa& a, const Vec3fa& b) {
       const vint4 ai = _mm_castps_si128(a.m128);
       const vint4 bi = _mm_castps_si128(b.m128);
@@ -192,11 +230,30 @@ namespace embree
   __forceinline Vec3fa msub  ( const Vec3fa& a, const Vec3fa& b, const Vec3fa& c) { return _mm_fmsub_ps(a.m128,b.m128,c.m128); }
   __forceinline Vec3fa nmadd ( const Vec3fa& a, const Vec3fa& b, const Vec3fa& c) { return _mm_fnmadd_ps(a.m128,b.m128,c.m128); }
   __forceinline Vec3fa nmsub ( const Vec3fa& a, const Vec3fa& b, const Vec3fa& c) { return _mm_fnmsub_ps(a.m128,b.m128,c.m128); }
+#else
+                                                                                
+#if defined(__aarch64__)
+  __forceinline Vec3fa madd  ( const Vec3fa& a, const Vec3fa& b, const Vec3fa& c) {
+        return _mm_madd_ps(a.m128, b.m128, c.m128);  //a*b+c;
+    }
+  __forceinline Vec3fa nmadd ( const Vec3fa& a, const Vec3fa& b, const Vec3fa& c) {
+        return _mm_msub_ps(a.m128, b.m128, c.m128);  //-a*b+c;
+    }
+  __forceinline Vec3fa nmsub( const Vec3fa& a, const Vec3fa& b, const Vec3fa& c) {
+        Vec3fa t = _mm_madd_ps(a.m128, b.m128, c.m128);
+        return -t;
+    }
+  __forceinline Vec3fa msub( const Vec3fa& a, const Vec3fa& b, const Vec3fa& c) {
+        return _mm_madd_ps(a.m128,b.m128,vnegq_f32(c.m128)); //a*b-c
+    }
+
 #else
   __forceinline Vec3fa madd  ( const Vec3fa& a, const Vec3fa& b, const Vec3fa& c) { return a*b+c; }
-  __forceinline Vec3fa msub  ( const Vec3fa& a, const Vec3fa& b, const Vec3fa& c) { return a*b-c; }
   __forceinline Vec3fa nmadd ( const Vec3fa& a, const Vec3fa& b, const Vec3fa& c) { return -a*b+c;}
   __forceinline Vec3fa nmsub ( const Vec3fa& a, const Vec3fa& b, const Vec3fa& c) { return -a*b-c; }
+  __forceinline Vec3fa msub  ( const Vec3fa& a, const Vec3fa& b, const Vec3fa& c) { return a*b-c; }
+#endif
+
 #endif
 
   __forceinline Vec3fa madd  ( const float a, const Vec3fa& b, const Vec3fa& c) { return madd(Vec3fa(a),b,c); }
@@ -218,18 +275,37 @@ namespace embree
   ////////////////////////////////////////////////////////////////////////////////
   /// Reductions
   ////////////////////////////////////////////////////////////////////////////////
-
+#if defined(__aarch64__) && defined(BUILD_IOS)
-  __forceinline float reduce_add(const Vec3fa& v) { 
+  __forceinline float reduce_add(const Vec3fa& v) {
+    float32x4_t t = v.m128;
+    t[3] = 0.0f;
+    return vaddvq_f32(t);
+  }
+                                                                                
+  __forceinline float reduce_mul(const Vec3fa& v) { return v.x*v.y*v.z; }
+  __forceinline float reduce_min(const Vec3fa& v) {
+    float32x4_t t = v.m128;
+      t[3] = t[2];
+    return vminvq_f32(t);
+  }
+  __forceinline float reduce_max(const Vec3fa& v) {
+    float32x4_t t = v.m128;
+      t[3] = t[2];
+    return vmaxvq_f32(t);
+  }
+#else
+  __forceinline float reduce_add(const Vec3fa& v) {
     const vfloat4 a(v.m128);
     const vfloat4 b = shuffle<1>(a);
     const vfloat4 c = shuffle<2>(a);
-    return _mm_cvtss_f32(a+b+c); 
+    return _mm_cvtss_f32(a+b+c);
   }
 
   __forceinline float reduce_mul(const Vec3fa& v) { return v.x*v.y*v.z; }
   __forceinline float reduce_min(const Vec3fa& v) { return min(v.x,v.y,v.z); }
   __forceinline float reduce_max(const Vec3fa& v) { return max(v.x,v.y,v.z); }
-
+#endif
+                                                                                
   ////////////////////////////////////////////////////////////////////////////////
   /// Comparison Operators
   ////////////////////////////////////////////////////////////////////////////////
@@ -241,8 +317,13 @@ namespace embree
   __forceinline Vec3ba neq_mask(const Vec3fa& a, const Vec3fa& b ) { return _mm_cmpneq_ps(a.m128, b.m128); }
   __forceinline Vec3ba lt_mask( const Vec3fa& a, const Vec3fa& b ) { return _mm_cmplt_ps (a.m128, b.m128); }
   __forceinline Vec3ba le_mask( const Vec3fa& a, const Vec3fa& b ) { return _mm_cmple_ps (a.m128, b.m128); }
-  __forceinline Vec3ba gt_mask( const Vec3fa& a, const Vec3fa& b ) { return _mm_cmpnle_ps(a.m128, b.m128); }
+ #if defined(__aarch64__)
-  __forceinline Vec3ba ge_mask( const Vec3fa& a, const Vec3fa& b ) { return _mm_cmpnlt_ps(a.m128, b.m128); }
+  __forceinline Vec3ba gt_mask( const Vec3fa& a, const Vec3fa& b ) { return _mm_cmpgt_ps (a.m128, b.m128); }
+  __forceinline Vec3ba ge_mask( const Vec3fa& a, const Vec3fa& b ) { return _mm_cmpge_ps (a.m128, b.m128); }
+#else
+  __forceinline Vec3ba gt_mask(const Vec3fa& a, const Vec3fa& b) { return _mm_cmpnle_ps(a.m128, b.m128); }
+  __forceinline Vec3ba ge_mask(const Vec3fa& a, const Vec3fa& b) { return _mm_cmpnlt_ps(a.m128, b.m128); }
+#endif
 
   __forceinline bool isvalid ( const Vec3fa& v ) {
     return all(gt_mask(v,Vec3fa(-FLT_LARGE)) & lt_mask(v,Vec3fa(+FLT_LARGE)));
@@ -280,7 +361,7 @@ namespace embree
     vfloat4 b0 = shuffle<1,2,0,3>(vfloat4(b.m128));
     vfloat4 a1 = shuffle<1,2,0,3>(vfloat4(a.m128));
     vfloat4 b1 = vfloat4(b.m128);
-    return Vec3fa(shuffle<1,2,0,3>(msub(a0,b0,a1*b1)));
+    return Vec3fa(shuffle<1,2,0,3>(prod_diff(a0,b0,a1,b1)));
   }
 
   __forceinline float  sqr_length ( const Vec3fa& a )                { return dot(a,a); }
@@ -335,7 +416,11 @@ namespace embree
   /// Rounding Functions
   ////////////////////////////////////////////////////////////////////////////////
 
-#if defined (__SSE4_1__)
+#if defined(__aarch64__)
+  __forceinline Vec3fa floor(const Vec3fa& a) { return vrndmq_f32(a.m128); }
+  __forceinline Vec3fa ceil (const Vec3fa& a) { return vrndpq_f32(a.m128); }
+  __forceinline Vec3fa trunc(const Vec3fa& a) { return vrndq_f32(a.m128); }
+#elif defined (__SSE4_1__)
   __forceinline Vec3fa trunc( const Vec3fa& a ) { return _mm_round_ps(a.m128, _MM_FROUND_TO_NEAREST_INT); }
   __forceinline Vec3fa floor( const Vec3fa& a ) { return _mm_round_ps(a.m128, _MM_FROUND_TO_NEG_INF    ); }
   __forceinline Vec3fa ceil ( const Vec3fa& a ) { return _mm_round_ps(a.m128, _MM_FROUND_TO_POS_INF    ); }
@@ -393,8 +478,10 @@ namespace embree
 
     __forceinline Vec3fx( const Vec3fa& other, const int      a1) { m128 = other.m128; a = a1; }
     __forceinline Vec3fx( const Vec3fa& other, const unsigned a1) { m128 = other.m128; u = a1; }
-    __forceinline Vec3fx( const Vec3fa& other, const float    w1) {      
+    __forceinline Vec3fx( const Vec3fa& other, const float    w1) {
-#if defined (__SSE4_1__)
+#if defined (__aarch64__)
+      m128 = other.m128; m128[3] = w1;
+#elif defined (__SSE4_1__)
       m128 = _mm_insert_ps(other.m128, _mm_set_ss(w1),3 << 4);
 #else
       const vint4 mask(-1,-1,-1,0);
@@ -526,7 +613,7 @@ namespace embree
   __forceinline Vec3fx min( const Vec3fx& a, const Vec3fx& b ) { return _mm_min_ps(a.m128,b.m128); }
   __forceinline Vec3fx max( const Vec3fx& a, const Vec3fx& b ) { return _mm_max_ps(a.m128,b.m128); }
 
-#if defined(__SSE4_1__)
+#if defined(__SSE4_1__) || defined(__aarch64__)
     __forceinline Vec3fx mini(const Vec3fx& a, const Vec3fx& b) {
       const vint4 ai = _mm_castps_si128(a.m128);
       const vint4 bi = _mm_castps_si128(b.m128);
@@ -535,7 +622,7 @@ namespace embree
     }
 #endif
 
-#if defined(__SSE4_1__)
+#if defined(__SSE4_1__) || defined(__aarch64__)
     __forceinline Vec3fx maxi(const Vec3fx& a, const Vec3fx& b) {
       const vint4 ai = _mm_castps_si128(a.m128);
       const vint4 bi = _mm_castps_si128(b.m128);
@@ -584,11 +671,11 @@ namespace embree
   /// Reductions
   ////////////////////////////////////////////////////////////////////////////////
 
-  __forceinline float reduce_add(const Vec3fx& v) { 
+  __forceinline float reduce_add(const Vec3fx& v) {
     const vfloat4 a(v.m128);
     const vfloat4 b = shuffle<1>(a);
     const vfloat4 c = shuffle<2>(a);
-    return _mm_cvtss_f32(a+b+c); 
+    return _mm_cvtss_f32(a+b+c);
   }
 
   __forceinline float reduce_mul(const Vec3fx& v) { return v.x*v.y*v.z; }
@@ -700,7 +787,7 @@ namespace embree
   /// Rounding Functions
   ////////////////////////////////////////////////////////////////////////////////
 
-#if defined (__SSE4_1__)
+#if defined (__SSE4_1__) && !defined(__aarch64__)
   __forceinline Vec3fx trunc( const Vec3fx& a ) { return _mm_round_ps(a.m128, _MM_FROUND_TO_NEAREST_INT); }
   __forceinline Vec3fx floor( const Vec3fx& a ) { return _mm_round_ps(a.m128, _MM_FROUND_TO_NEG_INF    ); }
   __forceinline Vec3fx ceil ( const Vec3fx& a ) { return _mm_round_ps(a.m128, _MM_FROUND_TO_POS_INF    ); }

thirdparty/embree/common/math/vec3ia.h

@@ -65,7 +65,9 @@ namespace embree
 
   __forceinline Vec3ia operator +( const Vec3ia& a ) { return a; }
   __forceinline Vec3ia operator -( const Vec3ia& a ) { return _mm_sub_epi32(_mm_setzero_si128(), a.m128); }
-#if defined(__SSSE3__)
+#if (defined(__aarch64__)) 
+  __forceinline Vec3ia abs       ( const Vec3ia& a ) { return vabsq_s32(a.m128); }
+#elif defined(__SSSE3__)
   __forceinline Vec3ia abs       ( const Vec3ia& a ) { return _mm_abs_epi32(a.m128); }
 #endif
 
@@ -81,7 +83,7 @@ namespace embree
   __forceinline Vec3ia operator -( const Vec3ia& a, const int     b ) { return a-Vec3ia(b); }
   __forceinline Vec3ia operator -( const int     a, const Vec3ia& b ) { return Vec3ia(a)-b; }
 
-#if defined(__SSE4_1__)
+#if defined(__aarch64__) || defined(__SSE4_1__)
   __forceinline Vec3ia operator *( const Vec3ia& a, const Vec3ia& b ) { return _mm_mullo_epi32(a.m128, b.m128); }
   __forceinline Vec3ia operator *( const Vec3ia& a, const int     b ) { return a * Vec3ia(b); }
   __forceinline Vec3ia operator *( const int     a, const Vec3ia& b ) { return Vec3ia(a) * b; }
@@ -99,12 +101,14 @@ namespace embree
   __forceinline Vec3ia operator ^( const Vec3ia& a, const int     b ) { return a ^ Vec3ia(b); }
   __forceinline Vec3ia operator ^( const int     a, const Vec3ia& b ) { return Vec3ia(a) ^ b; }
 
+#if !defined(__ARM_NEON)
   __forceinline Vec3ia operator <<( const Vec3ia& a, const int n ) { return _mm_slli_epi32(a.m128, n); }
   __forceinline Vec3ia operator >>( const Vec3ia& a, const int n ) { return _mm_srai_epi32(a.m128, n); }
 
   __forceinline Vec3ia sll ( const Vec3ia& a, const int b ) { return _mm_slli_epi32(a.m128, b); }
   __forceinline Vec3ia sra ( const Vec3ia& a, const int b ) { return _mm_srai_epi32(a.m128, b); }
   __forceinline Vec3ia srl ( const Vec3ia& a, const int b ) { return _mm_srli_epi32(a.m128, b); }
+#endif
 
   ////////////////////////////////////////////////////////////////////////////////
   /// Assignment Operators
@@ -116,7 +120,7 @@ namespace embree
   __forceinline Vec3ia& operator -=( Vec3ia& a, const Vec3ia& b ) { return a = a - b; }
   __forceinline Vec3ia& operator -=( Vec3ia& a, const int&   b ) { return a = a - b; }
   
-#if defined(__SSE4_1__)
+#if defined(__aarch64__) || defined(__SSE4_1__)
   __forceinline Vec3ia& operator *=( Vec3ia& a, const Vec3ia& b ) { return a = a * b; }
   __forceinline Vec3ia& operator *=( Vec3ia& a, const int&    b ) { return a = a * b; }
 #endif
@@ -127,18 +131,38 @@ namespace embree
   __forceinline Vec3ia& operator |=( Vec3ia& a, const Vec3ia& b ) { return a = a | b; }
   __forceinline Vec3ia& operator |=( Vec3ia& a, const int&    b ) { return a = a | b; }
   
+#if !defined(__ARM_NEON)
   __forceinline Vec3ia& operator <<=( Vec3ia& a, const int& b ) { return a = a << b; }
   __forceinline Vec3ia& operator >>=( Vec3ia& a, const int& b ) { return a = a >> b; }
+#endif
 
   ////////////////////////////////////////////////////////////////////////////////
   /// Reductions
   ////////////////////////////////////////////////////////////////////////////////
-
+#if defined(__aarch64__)
+  __forceinline int reduce_add(const Vec3ia& v) {
+    int32x4_t t = v.m128;
+    t[3] = 0;
+    return vaddvq_s32(t);
+        
+  }
+  __forceinline int reduce_mul(const Vec3ia& v) { return v.x*v.y*v.z; }
+  __forceinline int reduce_min(const Vec3ia& v) {
+    int32x4_t t = (__m128i)blendv_ps((__m128)v0x7fffffff, (__m128)v.m128, (__m128)vFFF0);
+    return vminvq_s32(t);
+        
+  }
+  __forceinline int reduce_max(const Vec3ia& v) {
+    int32x4_t t = (__m128i)blendv_ps((__m128)v0x80000000, (__m128)v.m128, (__m128)vFFF0);
+    return vmaxvq_s32(t);
+        
+  }
+#else
   __forceinline int reduce_add(const Vec3ia& v) { return v.x+v.y+v.z; }
   __forceinline int reduce_mul(const Vec3ia& v) { return v.x*v.y*v.z; }
   __forceinline int reduce_min(const Vec3ia& v) { return min(v.x,v.y,v.z); }
   __forceinline int reduce_max(const Vec3ia& v) { return max(v.x,v.y,v.z); }
-
+#endif
   ////////////////////////////////////////////////////////////////////////////////
   /// Comparison Operators
   ////////////////////////////////////////////////////////////////////////////////
@@ -161,14 +185,14 @@ namespace embree
   ////////////////////////////////////////////////////////////////////////////////
 
   __forceinline Vec3ia select( const Vec3ba& m, const Vec3ia& t, const Vec3ia& f ) {
-#if defined(__SSE4_1__)
+#if defined(__aarch64__) || defined(__SSE4_1__)
     return _mm_castps_si128(_mm_blendv_ps(_mm_castsi128_ps(f), _mm_castsi128_ps(t), m));
 #else
     return _mm_or_si128(_mm_and_si128(_mm_castps_si128(m), t), _mm_andnot_si128(_mm_castps_si128(m), f)); 
 #endif
   }
 
-#if defined(__SSE4_1__)
+#if defined(__aarch64__) || defined(__SSE4_1__)
   __forceinline Vec3ia min( const Vec3ia& a, const Vec3ia& b ) { return _mm_min_epi32(a.m128,b.m128); }
   __forceinline Vec3ia max( const Vec3ia& a, const Vec3ia& b ) { return _mm_max_epi32(a.m128,b.m128); }
 #else

thirdparty/embree/common/math/vec4.h

@@ -192,7 +192,7 @@ namespace embree
   ////////////////////////////////////////////////////////////////////////////////
 
   typedef Vec4<bool         > Vec4b;
-  typedef Vec4<unsigned char> Vec4uc;
+  typedef Vec4<uint8_t      > Vec4uc;
   typedef Vec4<int          > Vec4i;
   typedef Vec4<float        > Vec4f;
 }
@@ -205,7 +205,7 @@ namespace embree
 /// SSE / AVX / MIC specializations
 ////////////////////////////////////////////////////////////////////////////////
 
-#if defined __SSE__
+#if defined(__SSE__) || defined(__ARM_NEON)
 #include "../simd/sse.h"
 #endif
 
@@ -225,13 +225,13 @@ namespace embree
   template<> __forceinline Vec4<vfloat4>::Vec4( const Vec3fx& a ) {
     x = a.x; y = a.y; z = a.z; w = a.w;
   }
-#elif defined(__SSE__)
+#elif defined(__SSE__) || defined(__ARM_NEON)
   template<> __forceinline Vec4<vfloat4>::Vec4( const Vec3fx& a ) {
     const vfloat4 v = vfloat4(a.m128); x = shuffle<0,0,0,0>(v); y = shuffle<1,1,1,1>(v); z = shuffle<2,2,2,2>(v); w = shuffle<3,3,3,3>(v);
   }
 #endif
 
-#if defined(__SSE__)
+#if defined(__SSE__) || defined(__ARM_NEON)
   __forceinline Vec4<vfloat4> broadcast4f( const Vec4<vfloat4>& a, const size_t k ) {
     return Vec4<vfloat4>(vfloat4::broadcast(&a.x[k]), vfloat4::broadcast(&a.y[k]), vfloat4::broadcast(&a.z[k]), vfloat4::broadcast(&a.w[k]));
   }

thirdparty/embree/common/simd/simd.h

@@ -6,7 +6,7 @@
 #include "../math/math.h"
 
 /* include SSE wrapper classes */
-#if defined(__SSE__)
+#if defined(__SSE__) || defined(__ARM_NEON)
 #  include "sse.h"
 #endif
 

thirdparty/embree/common/simd/sse.h

@@ -11,7 +11,7 @@
 
 namespace embree 
 {
-#if defined(__SSE4_1__)
+#if (defined(__aarch64__) && defined(BUILD_IOS)) || defined(__SSE4_1__)
   __forceinline __m128 blendv_ps(__m128 f, __m128 t, __m128 mask) { 
     return _mm_blendv_ps(f,t,mask);
   }

thirdparty/embree/common/simd/vboold4_avx.h

@@ -56,8 +56,12 @@ namespace embree
     ////////////////////////////////////////////////////////////////////////////////
 
     __forceinline vboold(FalseTy) : v(_mm256_setzero_pd()) {}
+#if !defined(__aarch64__)
     __forceinline vboold(TrueTy)  : v(_mm256_cmp_pd(_mm256_setzero_pd(), _mm256_setzero_pd(), _CMP_EQ_OQ)) {}
-
+#else
+    __forceinline vboold(TrueTy)  : v(_mm256_cmpeq_pd(_mm256_setzero_pd(), _mm256_setzero_pd())) {}
+#endif
+      
     ////////////////////////////////////////////////////////////////////////////////
     /// Array Access
     ////////////////////////////////////////////////////////////////////////////////
@@ -101,9 +105,10 @@ namespace embree
   /// Movement/Shifting/Shuffling Functions
   ////////////////////////////////////////////////////////////////////////////////
 
+#if !defined(__aarch64__)
   __forceinline vboold4 unpacklo(const vboold4& a, const vboold4& b) { return _mm256_unpacklo_pd(a, b); }
   __forceinline vboold4 unpackhi(const vboold4& a, const vboold4& b) { return _mm256_unpackhi_pd(a, b); }
-
+#endif
 
 #if defined(__AVX2__)
   template<int i0, int i1, int i2, int i3>

thirdparty/embree/common/simd/vboolf4_sse2.h

@@ -37,9 +37,13 @@ namespace embree
       : v(mm_lookupmask_ps[(size_t(b) << 3) | (size_t(a) << 2) | (size_t(b) << 1) | size_t(a)]) {}
     __forceinline vboolf(bool a, bool b, bool c, bool d)
       : v(mm_lookupmask_ps[(size_t(d) << 3) | (size_t(c) << 2) | (size_t(b) << 1) | size_t(a)]) {}
+#if defined(__aarch64__) && defined(BUILD_IOS)
+      __forceinline vboolf(int mask) { v = mm_lookupmask_ps[mask]; }
+      __forceinline vboolf(unsigned int mask) { v = mm_lookupmask_ps[mask]; }
+#else
     __forceinline vboolf(int mask) { assert(mask >= 0 && mask < 16); v = mm_lookupmask_ps[mask]; }
     __forceinline vboolf(unsigned int mask) { assert(mask < 16); v = mm_lookupmask_ps[mask]; }
-
+#endif
     /* return int32 mask */
     __forceinline __m128i mask32() const { 
       return _mm_castps_si128(v);
@@ -56,8 +60,13 @@ namespace embree
     /// Array Access
     ////////////////////////////////////////////////////////////////////////////////
 
+#if defined(__aarch64__) && defined(BUILD_IOS)
+      __forceinline bool operator [](size_t index) const { return (_mm_movemask_ps(v) >> index) & 1; }
+      __forceinline int& operator [](size_t index)       { return i[index]; }
+#else
     __forceinline bool operator [](size_t index) const { assert(index < 4); return (_mm_movemask_ps(v) >> index) & 1; }
     __forceinline int& operator [](size_t index)       { assert(index < 4); return i[index]; }
+#endif
   };
 
   ////////////////////////////////////////////////////////////////////////////////
@@ -92,7 +101,7 @@ namespace embree
   __forceinline vboolf4 operator ==(const vboolf4& a, const vboolf4& b) { return _mm_castsi128_ps(_mm_cmpeq_epi32(a, b)); }
   
   __forceinline vboolf4 select(const vboolf4& m, const vboolf4& t, const vboolf4& f) {
-#if defined(__SSE4_1__)
+#if (defined(__aarch64__) && defined(BUILD_IOS)) || defined(__SSE4_1__)
     return _mm_blendv_ps(f, t, m); 
 #else
     return _mm_or_ps(_mm_and_ps(m, t), _mm_andnot_ps(m, f)); 
@@ -106,6 +115,17 @@ namespace embree
   __forceinline vboolf4 unpacklo(const vboolf4& a, const vboolf4& b) { return _mm_unpacklo_ps(a, b); }
   __forceinline vboolf4 unpackhi(const vboolf4& a, const vboolf4& b) { return _mm_unpackhi_ps(a, b); }
 
+#if defined(__aarch64__)
+  template<int i0, int i1, int i2, int i3>
+  __forceinline vboolf4 shuffle(const vboolf4& v) {
+    return vreinterpretq_f32_u8(vqtbl1q_u8( vreinterpretq_u8_s32(v), _MN_SHUFFLE(i0, i1, i2, i3)));
+  }
+
+  template<int i0, int i1, int i2, int i3>
+  __forceinline vboolf4 shuffle(const vboolf4& a, const vboolf4& b) {
+    return vreinterpretq_f32_u8(vqtbl2q_u8( (uint8x16x2_t){(uint8x16_t)a.v, (uint8x16_t)b.v}, _MF_SHUFFLE(i0, i1, i2, i3)));
+  }
+#else
   template<int i0, int i1, int i2, int i3>
   __forceinline vboolf4 shuffle(const vboolf4& v) {
     return _mm_castsi128_ps(_mm_shuffle_epi32(v, _MM_SHUFFLE(i3, i2, i1, i0)));
@@ -115,7 +135,8 @@ namespace embree
   __forceinline vboolf4 shuffle(const vboolf4& a, const vboolf4& b) {
     return _mm_shuffle_ps(a, b, _MM_SHUFFLE(i3, i2, i1, i0));
   }
-
+#endif
+                                                                
   template<int i0>
   __forceinline vboolf4 shuffle(const vboolf4& v) {
     return shuffle<i0,i0,i0,i0>(v);
@@ -127,7 +148,7 @@ namespace embree
   template<> __forceinline vboolf4 shuffle<0, 1, 0, 1>(const vboolf4& v) { return _mm_castpd_ps(_mm_movedup_pd(v)); }
 #endif
 
-#if defined(__SSE4_1__)
+#if defined(__SSE4_1__) && !defined(__aarch64__)
   template<int dst, int src, int clr> __forceinline vboolf4 insert(const vboolf4& a, const vboolf4& b) { return _mm_insert_ps(a, b, (dst << 4) | (src << 6) | clr); }
   template<int dst, int src> __forceinline vboolf4 insert(const vboolf4& a, const vboolf4& b) { return insert<dst, src, 0>(a, b); }
   template<int dst> __forceinline vboolf4 insert(const vboolf4& a, const bool b) { return insert<dst, 0>(a, vboolf4(b)); }
@@ -149,10 +170,14 @@ namespace embree
   __forceinline bool none(const vboolf4& valid, const vboolf4& b) { return none(valid & b); }
   
   __forceinline size_t movemask(const vboolf4& a) { return _mm_movemask_ps(a); }
+#if defined(__aarch64__) && defined(BUILD_IOS)
+__forceinline size_t popcnt(const vboolf4& a) { return _mm_movemask_popcnt_ps(a); }
+#else
 #if defined(__SSE4_2__)
   __forceinline size_t popcnt(const vboolf4& a) { return popcnt((size_t)_mm_movemask_ps(a)); }
 #else
   __forceinline size_t popcnt(const vboolf4& a) { return bool(a[0])+bool(a[1])+bool(a[2])+bool(a[3]); }
+#endif
 #endif
 
   ////////////////////////////////////////////////////////////////////////////////

thirdparty/embree/common/simd/vboolf8_avx.h

@@ -68,8 +68,11 @@ namespace embree
     ////////////////////////////////////////////////////////////////////////////////
 
     __forceinline vboolf(FalseTy) : v(_mm256_setzero_ps()) {}
+#if !defined(__aarch64__)
     __forceinline vboolf(TrueTy)  : v(_mm256_cmp_ps(_mm256_setzero_ps(), _mm256_setzero_ps(), _CMP_EQ_OQ)) {}
-
+#else
+    __forceinline vboolf(TrueTy)  : v(_mm256_cmpeq_ps(_mm256_setzero_ps(), _mm256_setzero_ps())) {}
+#endif
     ////////////////////////////////////////////////////////////////////////////////
     /// Array Access
     ////////////////////////////////////////////////////////////////////////////////

thirdparty/embree/common/simd/vdouble4_avx.h

@@ -181,13 +181,20 @@ namespace embree
   __forceinline vboold4 operator >=(const vdouble4& a, const vdouble4& b) { return _mm256_cmp_pd_mask(a, b, _MM_CMPINT_GE); }
   __forceinline vboold4 operator > (const vdouble4& a, const vdouble4& b) { return _mm256_cmp_pd_mask(a, b, _MM_CMPINT_GT); }
   __forceinline vboold4 operator <=(const vdouble4& a, const vdouble4& b) { return _mm256_cmp_pd_mask(a, b, _MM_CMPINT_LE); }
-#else
+#elif !defined(__aarch64__)
   __forceinline vboold4 operator ==(const vdouble4& a, const vdouble4& b) { return _mm256_cmp_pd(a, b, _CMP_EQ_OQ);  }
   __forceinline vboold4 operator !=(const vdouble4& a, const vdouble4& b) { return _mm256_cmp_pd(a, b, _CMP_NEQ_UQ); }
   __forceinline vboold4 operator < (const vdouble4& a, const vdouble4& b) { return _mm256_cmp_pd(a, b, _CMP_LT_OS);  }
   __forceinline vboold4 operator >=(const vdouble4& a, const vdouble4& b) { return _mm256_cmp_pd(a, b, _CMP_NLT_US); }
   __forceinline vboold4 operator > (const vdouble4& a, const vdouble4& b) { return _mm256_cmp_pd(a, b, _CMP_NLE_US); }
   __forceinline vboold4 operator <=(const vdouble4& a, const vdouble4& b) { return _mm256_cmp_pd(a, b, _CMP_LE_OS);  }
+#else
+  __forceinline vboold4 operator ==(const vdouble4& a, const vdouble4& b) { return _mm256_cmpeq_pd(a, b);  }
+  __forceinline vboold4 operator !=(const vdouble4& a, const vdouble4& b) { return _mm256_cmpneq_pd(a, b); }
+  __forceinline vboold4 operator < (const vdouble4& a, const vdouble4& b) { return _mm256_cmplt_pd(a, b);  }
+  __forceinline vboold4 operator >=(const vdouble4& a, const vdouble4& b) { return _mm256_cmpnlt_pd(a, b); }
+  __forceinline vboold4 operator > (const vdouble4& a, const vdouble4& b) { return _mm256_cmpnle_pd(a, b); }
+  __forceinline vboold4 operator <=(const vdouble4& a, const vdouble4& b) { return _mm256_cmple_pd(a, b);  }
 #endif
 
   __forceinline vboold4 operator ==(const vdouble4& a, double          b) { return a == vdouble4(b); }

thirdparty/embree/common/simd/vfloat4_sse2.h

@@ -10,18 +10,18 @@ namespace embree
   struct vfloat<4>
   {
     ALIGNED_STRUCT_(16);
-    
+
     typedef vboolf4 Bool;
     typedef vint4   Int;
     typedef vfloat4 Float;
-    
+
     enum  { size = 4 };                        // number of SIMD elements
     union { __m128 v; float f[4]; int i[4]; }; // data
 
     ////////////////////////////////////////////////////////////////////////////////
     /// Constructors, Assignment & Cast Operators
     ////////////////////////////////////////////////////////////////////////////////
-    
+
     __forceinline vfloat() {}
     __forceinline vfloat(const vfloat4& other) { v = other.v; }
     __forceinline vfloat4& operator =(const vfloat4& other) { v = other.v; return *this; }
@@ -34,14 +34,19 @@ namespace embree
     __forceinline vfloat(float a, float b, float c, float d) : v(_mm_set_ps(d, c, b, a)) {}
 
     __forceinline explicit vfloat(const vint4& a) : v(_mm_cvtepi32_ps(a)) {}
+#if defined(__aarch64__)
+    __forceinline explicit vfloat(const vuint4& x) {
+        v = vcvtq_f32_u32(vreinterpretq_u32_s32(x.v));
+    }
+#else
     __forceinline explicit vfloat(const vuint4& x) {
       const __m128i a   = _mm_and_si128(x,_mm_set1_epi32(0x7FFFFFFF));
-      const __m128i b   = _mm_and_si128(_mm_srai_epi32(x,31),_mm_set1_epi32(0x4F000000)); //0x4F000000 = 2^31 
+      const __m128i b   = _mm_and_si128(_mm_srai_epi32(x,31),_mm_set1_epi32(0x4F000000)); //0x4F000000 = 2^31
       const __m128  af  = _mm_cvtepi32_ps(a);
-      const __m128  bf  = _mm_castsi128_ps(b);  
+      const __m128  bf  = _mm_castsi128_ps(b);
       v  = _mm_add_ps(af,bf);
     }
-
+#endif
     ////////////////////////////////////////////////////////////////////////////////
     /// Constants
     ////////////////////////////////////////////////////////////////////////////////
@@ -102,32 +107,44 @@ namespace embree
 #if defined (__SSE4_1__)
     return _mm_castsi128_ps(_mm_stream_load_si128((__m128i*)ptr));
 #else
-    return _mm_load_ps(ptr); 
+    return _mm_load_ps(ptr);
 #endif
   }
 
-#if defined(__SSE4_1__)
+#if defined(__aarch64__)
-    static __forceinline vfloat4 load(const char* ptr) {
+    static __forceinline vfloat4 load(const int8_t* ptr) {
+        return __m128(_mm_load4epi8_f32(((__m128i*)ptr)));
+    }
+#elif defined(__SSE4_1__)
+    static __forceinline vfloat4 load(const int8_t* ptr) {
       return _mm_cvtepi32_ps(_mm_cvtepi8_epi32(_mm_loadu_si128((__m128i*)ptr)));
     }
 #else
-    static __forceinline vfloat4 load(const char* ptr) {
+    static __forceinline vfloat4 load(const int8_t* ptr) {
       return vfloat4(ptr[0],ptr[1],ptr[2],ptr[3]);
     }
 #endif
 
-#if defined(__SSE4_1__)
+#if defined(__aarch64__)
-    static __forceinline vfloat4 load(const unsigned char* ptr) {
+    static __forceinline vfloat4 load(const uint8_t* ptr) {
+        return __m128(_mm_load4epu8_f32(((__m128i*)ptr)));
+    }
+#elif defined(__SSE4_1__)
+    static __forceinline vfloat4 load(const uint8_t* ptr) {
       return _mm_cvtepi32_ps(_mm_cvtepu8_epi32(_mm_loadu_si128((__m128i*)ptr)));
     }
 #else
-    static __forceinline vfloat4 load(const unsigned char* ptr) {
+    static __forceinline vfloat4 load(const uint8_t* ptr) {
       //return _mm_cvtpu8_ps(*(__m64*)ptr); // don't enable, will use MMX instructions
       return vfloat4(ptr[0],ptr[1],ptr[2],ptr[3]);
     }
 #endif
 
-#if defined(__SSE4_1__)
+#if defined(__aarch64__)
+    static __forceinline vfloat4 load(const short* ptr) {
+        return __m128(_mm_load4epi16_f32(((__m128i*)ptr)));
+    }
+#elif defined(__SSE4_1__)
     static __forceinline vfloat4 load(const short* ptr) {
       return _mm_cvtepi32_ps(_mm_cvtepi16_epi32(_mm_loadu_si128((__m128i*)ptr)));
     }
@@ -140,11 +157,15 @@ namespace embree
     static __forceinline vfloat4 load(const unsigned short* ptr) {
       return _mm_mul_ps(vfloat4(vint4::load(ptr)),vfloat4(1.0f/65535.0f));
     }
-    
+
     static __forceinline void store_nt(void* ptr, const vfloat4& v)
     {
 #if defined (__SSE4_1__)
+#if defined(__aarch64__)
+      _mm_stream_ps((float*)ptr,vreinterpretq_s32_f32(v.v));
+#else
       _mm_stream_ps((float*)ptr,v);
+#endif
 #else
       _mm_store_ps((float*)ptr,v);
 #endif
@@ -152,14 +173,14 @@ namespace embree
 
     template<int scale = 4>
     static __forceinline vfloat4 gather(const float* ptr, const vint4& index) {
-#if defined(__AVX2__)
+#if defined(__AVX2__) && !defined(__aarch64__)
       return _mm_i32gather_ps(ptr, index, scale);
 #else
       return vfloat4(
-        *(float*)(((char*)ptr)+scale*index[0]),
+        *(float*)(((int8_t*)ptr)+scale*index[0]),
-        *(float*)(((char*)ptr)+scale*index[1]),
+        *(float*)(((int8_t*)ptr)+scale*index[1]),
-        *(float*)(((char*)ptr)+scale*index[2]),
+        *(float*)(((int8_t*)ptr)+scale*index[2]),
-        *(float*)(((char*)ptr)+scale*index[3]));
+        *(float*)(((int8_t*)ptr)+scale*index[3]));
 #endif
     }
 
@@ -168,13 +189,13 @@ namespace embree
       vfloat4 r = zero;
 #if defined(__AVX512VL__)
       return _mm_mmask_i32gather_ps(r, mask, index, ptr, scale);
-#elif defined(__AVX2__)
+#elif defined(__AVX2__)  && !defined(__aarch64__)
       return _mm_mask_i32gather_ps(r, ptr, index, mask, scale);
 #else
-      if (likely(mask[0])) r[0] = *(float*)(((char*)ptr)+scale*index[0]);
+      if (likely(mask[0])) r[0] = *(float*)(((int8_t*)ptr)+scale*index[0]);
-      if (likely(mask[1])) r[1] = *(float*)(((char*)ptr)+scale*index[1]);
+      if (likely(mask[1])) r[1] = *(float*)(((int8_t*)ptr)+scale*index[1]);
-      if (likely(mask[2])) r[2] = *(float*)(((char*)ptr)+scale*index[2]);
+      if (likely(mask[2])) r[2] = *(float*)(((int8_t*)ptr)+scale*index[2]);
-      if (likely(mask[3])) r[3] = *(float*)(((char*)ptr)+scale*index[3]);
+      if (likely(mask[3])) r[3] = *(float*)(((int8_t*)ptr)+scale*index[3]);
       return r;
 #endif
     }
@@ -185,10 +206,10 @@ namespace embree
 #if defined(__AVX512VL__)
       _mm_i32scatter_ps((float*)ptr, index, v, scale);
 #else
-      *(float*)(((char*)ptr)+scale*index[0]) = v[0];
+      *(float*)(((int8_t*)ptr)+scale*index[0]) = v[0];
-      *(float*)(((char*)ptr)+scale*index[1]) = v[1];
+      *(float*)(((int8_t*)ptr)+scale*index[1]) = v[1];
-      *(float*)(((char*)ptr)+scale*index[2]) = v[2];
+      *(float*)(((int8_t*)ptr)+scale*index[2]) = v[2];
-      *(float*)(((char*)ptr)+scale*index[3]) = v[3];
+      *(float*)(((int8_t*)ptr)+scale*index[3]) = v[3];
 #endif
     }
 
@@ -198,20 +219,20 @@ namespace embree
 #if defined(__AVX512VL__)
       _mm_mask_i32scatter_ps((float*)ptr ,mask, index, v, scale);
 #else
-      if (likely(mask[0])) *(float*)(((char*)ptr)+scale*index[0]) = v[0];
+      if (likely(mask[0])) *(float*)(((int8_t*)ptr)+scale*index[0]) = v[0];
-      if (likely(mask[1])) *(float*)(((char*)ptr)+scale*index[1]) = v[1];
+      if (likely(mask[1])) *(float*)(((int8_t*)ptr)+scale*index[1]) = v[1];
-      if (likely(mask[2])) *(float*)(((char*)ptr)+scale*index[2]) = v[2];
+      if (likely(mask[2])) *(float*)(((int8_t*)ptr)+scale*index[2]) = v[2];
-      if (likely(mask[3])) *(float*)(((char*)ptr)+scale*index[3]) = v[3];
+      if (likely(mask[3])) *(float*)(((int8_t*)ptr)+scale*index[3]) = v[3];
 #endif
     }
 
-    static __forceinline void store(const vboolf4& mask, char* ptr, const vint4& ofs, const vfloat4& v) {
+    static __forceinline void store(const vboolf4& mask, int8_t* ptr, const vint4& ofs, const vfloat4& v) {
       scatter<1>(mask,ptr,ofs,v);
     }
     static __forceinline void store(const vboolf4& mask, float* ptr, const vint4& ofs, const vfloat4& v) {
       scatter<4>(mask,ptr,ofs,v);
     }
-    
+
     ////////////////////////////////////////////////////////////////////////////////
     /// Array Access
     ////////////////////////////////////////////////////////////////////////////////
@@ -222,10 +243,10 @@ namespace embree
     friend __forceinline vfloat4 select(const vboolf4& m, const vfloat4& t, const vfloat4& f) {
 #if defined(__AVX512VL__)
       return _mm_mask_blend_ps(m, f, t);
-#elif defined(__SSE4_1__)
+#elif defined(__SSE4_1__) || (defined(__aarch64__))
-      return _mm_blendv_ps(f, t, m); 
+      return _mm_blendv_ps(f, t, m);
 #else
-      return _mm_or_ps(_mm_and_ps(m, t), _mm_andnot_ps(m, f)); 
+      return _mm_or_ps(_mm_and_ps(m, t), _mm_andnot_ps(m, f));
 #endif
     }
   };
@@ -243,18 +264,47 @@ namespace embree
   __forceinline vfloat4 toFloat(const vint4&   a) { return vfloat4(a); }
 
   __forceinline vfloat4 operator +(const vfloat4& a) { return a; }
+#if defined(__aarch64__)
+  __forceinline vfloat4 operator -(const vfloat4& a) {
+    return vnegq_f32(a);
+  }
+#else
   __forceinline vfloat4 operator -(const vfloat4& a) { return _mm_xor_ps(a, _mm_castsi128_ps(_mm_set1_epi32(0x80000000))); }
+#endif
 
+#if defined(__aarch64__)
+  __forceinline vfloat4 abs(const vfloat4& a) { return _mm_abs_ps(a); }
+#else
   __forceinline vfloat4 abs(const vfloat4& a) { return _mm_and_ps(a, _mm_castsi128_ps(_mm_set1_epi32(0x7fffffff))); }
+#endif
+
 #if defined(__AVX512VL__)
   __forceinline vfloat4 sign(const vfloat4& a) { return _mm_mask_blend_ps(_mm_cmp_ps_mask(a, vfloat4(zero), _CMP_LT_OQ), vfloat4(one), -vfloat4(one)); }
 #else
   __forceinline vfloat4 sign(const vfloat4& a) { return blendv_ps(vfloat4(one), -vfloat4(one), _mm_cmplt_ps(a, vfloat4(zero))); }
 #endif
+
+#if defined(__aarch64__)
+  __forceinline vfloat4 signmsk(const vfloat4& a) { return _mm_and_ps(a, vreinterpretq_f32_u32(v0x80000000)); }
+#else
   __forceinline vfloat4 signmsk(const vfloat4& a) { return _mm_and_ps(a,_mm_castsi128_ps(_mm_set1_epi32(0x80000000))); }
-  
+#endif
+
   __forceinline vfloat4 rcp(const vfloat4& a)
   {
+#if defined(__aarch64__)
+#if defined(BUILD_IOS)
+    return vfloat4(vdivq_f32(vdupq_n_f32(1.0f),a.v));
+#else //BUILD_IOS
+    __m128 reciprocal = _mm_rcp_ps(a);
+    reciprocal = vmulq_f32(vrecpsq_f32(a, reciprocal), reciprocal);
+    reciprocal = vmulq_f32(vrecpsq_f32(a, reciprocal), reciprocal);
+    // +1 round since NEON's reciprocal estimate instruction has less accuracy than SSE2's rcp.
+    reciprocal = vmulq_f32(vrecpsq_f32(a, reciprocal), reciprocal);
+    return (const vfloat4)reciprocal;
+#endif // BUILD_IOS
+#else
+
 #if defined(__AVX512VL__)
     const vfloat4 r = _mm_rcp14_ps(a);
 #else
@@ -266,12 +316,22 @@ namespace embree
 #else
     return _mm_mul_ps(r,_mm_sub_ps(vfloat4(2.0f), _mm_mul_ps(r, a)));
 #endif
+
+#endif  //defined(__aarch64__)
   }
   __forceinline vfloat4 sqr (const vfloat4& a) { return _mm_mul_ps(a,a); }
   __forceinline vfloat4 sqrt(const vfloat4& a) { return _mm_sqrt_ps(a); }
 
   __forceinline vfloat4 rsqrt(const vfloat4& a)
   {
+#if defined(__aarch64__)
+    vfloat4 r = _mm_rsqrt_ps(a);
+    r = vmulq_f32(r, vrsqrtsq_f32(vmulq_f32(a, r), r));
+    r = vmulq_f32(r, vrsqrtsq_f32(vmulq_f32(a, r), r));
+    r = vmulq_f32(r, vrsqrtsq_f32(vmulq_f32(a, r), r));
+    return r;
+#else
+
 #if defined(__AVX512VL__)
     const vfloat4 r = _mm_rsqrt14_ps(a);
 #else
@@ -284,11 +344,17 @@ namespace embree
 #else
     return _mm_add_ps(_mm_mul_ps(_mm_set1_ps(1.5f), r),
                       _mm_mul_ps(_mm_mul_ps(_mm_mul_ps(a, _mm_set1_ps(-0.5f)), r), _mm_mul_ps(r, r)));
+#endif
+
 #endif
   }
 
   __forceinline vboolf4 isnan(const vfloat4& a) {
+#if defined(__aarch64__)
+    const vfloat4 b = _mm_and_ps(a, vreinterpretq_f32_u32(v0x7fffffff));
+#else
     const vfloat4 b = _mm_and_ps(a, _mm_castsi128_ps(_mm_set1_epi32(0x7fffffff)));
+#endif
 #if defined(__AVX512VL__)
     return _mm_cmp_epi32_mask(_mm_castps_si128(b), _mm_set1_epi32(0x7f800000), _MM_CMPINT_GT);
 #else
@@ -329,7 +395,8 @@ namespace embree
   __forceinline vfloat4 max(const vfloat4& a, float          b) { return _mm_max_ps(a,vfloat4(b)); }
   __forceinline vfloat4 max(float          a, const vfloat4& b) { return _mm_max_ps(vfloat4(a),b); }
 
-#if defined(__SSE4_1__)
+#if defined(__SSE4_1__) || defined(__aarch64__)
+
     __forceinline vfloat4 mini(const vfloat4& a, const vfloat4& b) {
       const vint4 ai = _mm_castps_si128(a);
       const vint4 bi = _mm_castps_si128(b);
@@ -377,10 +444,24 @@ namespace embree
   __forceinline vfloat4 nmadd(const vfloat4& a, const vfloat4& b, const vfloat4& c) { return _mm_fnmadd_ps(a,b,c); }
   __forceinline vfloat4 nmsub(const vfloat4& a, const vfloat4& b, const vfloat4& c) { return _mm_fnmsub_ps(a,b,c); }
 #else
+
+#if defined(__aarch64__)
+  __forceinline vfloat4 madd (const vfloat4& a, const vfloat4& b, const vfloat4& c) {
+    return _mm_madd_ps(a, b, c);  //a*b+c;
+  }
+  __forceinline vfloat4 nmadd(const vfloat4& a, const vfloat4& b, const vfloat4& c) {
+    return _mm_msub_ps(a, b, c);  //-a*b+c;
+  }
+  __forceinline vfloat4 nmsub(const vfloat4& a, const vfloat4& b, const vfloat4& c) {
+    return vnegq_f32(vfmaq_f32(c,a, b));
+  }
+#else
   __forceinline vfloat4 madd (const vfloat4& a, const vfloat4& b, const vfloat4& c) { return a*b+c; }
-  __forceinline vfloat4 msub (const vfloat4& a, const vfloat4& b, const vfloat4& c) { return a*b-c; }
   __forceinline vfloat4 nmadd(const vfloat4& a, const vfloat4& b, const vfloat4& c) { return -a*b+c;}
   __forceinline vfloat4 nmsub(const vfloat4& a, const vfloat4& b, const vfloat4& c) { return -a*b-c; }
+#endif
+  __forceinline vfloat4 msub (const vfloat4& a, const vfloat4& b, const vfloat4& c) { return a*b-c; }
+
 #endif
 
   ////////////////////////////////////////////////////////////////////////////////
@@ -414,8 +495,13 @@ namespace embree
   __forceinline vboolf4 operator ==(const vfloat4& a, const vfloat4& b) { return _mm_cmpeq_ps (a, b); }
   __forceinline vboolf4 operator !=(const vfloat4& a, const vfloat4& b) { return _mm_cmpneq_ps(a, b); }
   __forceinline vboolf4 operator < (const vfloat4& a, const vfloat4& b) { return _mm_cmplt_ps (a, b); }
+#if defined(__aarch64__)
+  __forceinline vboolf4 operator >=(const vfloat4& a, const vfloat4& b) { return _mm_cmpge_ps (a, b); }
+  __forceinline vboolf4 operator > (const vfloat4& a, const vfloat4& b) { return _mm_cmpgt_ps (a, b); }
+#else
   __forceinline vboolf4 operator >=(const vfloat4& a, const vfloat4& b) { return _mm_cmpnlt_ps(a, b); }
   __forceinline vboolf4 operator > (const vfloat4& a, const vfloat4& b) { return _mm_cmpnle_ps(a, b); }
+#endif
   __forceinline vboolf4 operator <=(const vfloat4& a, const vfloat4& b) { return _mm_cmple_ps (a, b); }
 #endif
 
@@ -427,7 +513,7 @@ namespace embree
 
   __forceinline vboolf4 operator < (const vfloat4& a, float          b) { return a <  vfloat4(b); }
   __forceinline vboolf4 operator < (float          a, const vfloat4& b) { return vfloat4(a) <  b; }
-  
+
   __forceinline vboolf4 operator >=(const vfloat4& a, float          b) { return a >= vfloat4(b); }
   __forceinline vboolf4 operator >=(float          a, const vfloat4& b) { return vfloat4(a) >= b; }
 
@@ -463,17 +549,68 @@ namespace embree
   template<int mask>
     __forceinline vfloat4 select(const vfloat4& t, const vfloat4& f)
   {
-#if defined(__SSE4_1__) 
+#if defined(__SSE4_1__)
     return _mm_blend_ps(f, t, mask);
 #else
     return select(vboolf4(mask), t, f);
 #endif
   }
-  
+
+#if defined(__aarch64__)
+    template<> __forceinline vfloat4 select<0>(const vfloat4& t, const vfloat4& f) {
+        return _mm_blendv_ps(f, t, vreinterpretq_f32_u32(vzero));
+    }
+    template<> __forceinline vfloat4 select<1>(const vfloat4& t, const vfloat4& f) {
+        return _mm_blendv_ps(f, t, vreinterpretq_f32_u32(v000F));
+    }
+    template<> __forceinline vfloat4 select<2>(const vfloat4& t, const vfloat4& f) {
+        return _mm_blendv_ps(f, t, vreinterpretq_f32_u32(v00F0));
+    }
+    template<> __forceinline vfloat4 select<3>(const vfloat4& t, const vfloat4& f) {
+        return _mm_blendv_ps(f, t, vreinterpretq_f32_u32(v00FF));
+    }
+    template<> __forceinline vfloat4 select<4>(const vfloat4& t, const vfloat4& f) {
+        return _mm_blendv_ps(f, t, vreinterpretq_f32_u32(v0F00));
+    }
+    template<> __forceinline vfloat4 select<5>(const vfloat4& t, const vfloat4& f) {
+        return _mm_blendv_ps(f, t, vreinterpretq_f32_u32(v0F0F));
+    }
+    template<> __forceinline vfloat4 select<6>(const vfloat4& t, const vfloat4& f) {
+        return _mm_blendv_ps(f, t, vreinterpretq_f32_u32(v0FF0));
+    }
+    template<> __forceinline vfloat4 select<7>(const vfloat4& t, const vfloat4& f) {
+        return _mm_blendv_ps(f, t, vreinterpretq_f32_u32(v0FFF));
+    }
+    template<> __forceinline vfloat4 select<8>(const vfloat4& t, const vfloat4& f) {
+        return _mm_blendv_ps(f, t, vreinterpretq_f32_u32(vF000));
+    }
+    template<> __forceinline vfloat4 select<9>(const vfloat4& t, const vfloat4& f) {
+        return _mm_blendv_ps(f, t, vreinterpretq_f32_u32(vF00F));
+    }
+    template<> __forceinline vfloat4 select<10>(const vfloat4& t, const vfloat4& f) {
+        return _mm_blendv_ps(f, t, vreinterpretq_f32_u32(vF0F0));
+    }
+    template<> __forceinline vfloat4 select<11>(const vfloat4& t, const vfloat4& f) {
+        return _mm_blendv_ps(f, t, vreinterpretq_f32_u32(vF0FF));
+    }
+    template<> __forceinline vfloat4 select<12>(const vfloat4& t, const vfloat4& f) {
+        return _mm_blendv_ps(f, t, vreinterpretq_f32_u32(vFF00));
+    }
+    template<> __forceinline vfloat4 select<13>(const vfloat4& t, const vfloat4& f) {
+        return _mm_blendv_ps(f, t, vreinterpretq_f32_u32(vFF0F));
+    }
+    template<> __forceinline vfloat4 select<14>(const vfloat4& t, const vfloat4& f) {
+        return _mm_blendv_ps(f, t, vreinterpretq_f32_u32(vFFF0));
+    }
+    template<> __forceinline vfloat4 select<15>(const vfloat4& t, const vfloat4& f) {
+        return _mm_blendv_ps(f, t, vreinterpretq_f32_u32(vFFFF));
+    }
+#endif
+
   __forceinline vfloat4 lerp(const vfloat4& a, const vfloat4& b, const vfloat4& t) {
     return madd(t,b-a,a);
   }
-  
+
   __forceinline bool isvalid(const vfloat4& v) {
     return all((v > vfloat4(-FLT_LARGE)) & (v < vfloat4(+FLT_LARGE)));
   }
@@ -485,16 +622,21 @@ namespace embree
   __forceinline bool is_finite(const vboolf4& valid, const vfloat4& a) {
     return all(valid, (a >= vfloat4(-FLT_MAX)) & (a <= vfloat4(+FLT_MAX)));
   }
-  
+
   ////////////////////////////////////////////////////////////////////////////////
   /// Rounding Functions
   ////////////////////////////////////////////////////////////////////////////////
 
-#if defined (__SSE4_1__)
+#if defined(__aarch64__)
-  __forceinline vfloat4 floor(const vfloat4& a) { return _mm_round_ps(a, _MM_FROUND_TO_NEG_INF    ); }
+  __forceinline vfloat4 floor(const vfloat4& a) { return vrndmq_f32(a.v); } // towards -inf
-  __forceinline vfloat4 ceil (const vfloat4& a) { return _mm_round_ps(a, _MM_FROUND_TO_POS_INF    ); }
+  __forceinline vfloat4 ceil (const vfloat4& a) { return vrndpq_f32(a.v); } // toward +inf
-  __forceinline vfloat4 trunc(const vfloat4& a) { return _mm_round_ps(a, _MM_FROUND_TO_ZERO       ); }
+  __forceinline vfloat4 trunc(const vfloat4& a) { return vrndq_f32(a.v); } // towards 0
-  __forceinline vfloat4 round(const vfloat4& a) { return _mm_round_ps(a, _MM_FROUND_TO_NEAREST_INT); }
+  __forceinline vfloat4 round(const vfloat4& a) { return vrndnq_f32(a.v); } // to nearest, ties to even. NOTE(LTE): arm clang uses vrndnq, old gcc uses vrndqn?
+#elif defined (__SSE4_1__)
+  __forceinline vfloat4 floor(const vfloat4& a) { return _mm_round_ps(a, _MM_FROUND_TO_NEG_INF   ); }
+  __forceinline vfloat4 ceil (const vfloat4& a) { return _mm_round_ps(a, _MM_FROUND_TO_POS_INF   ); }
+  __forceinline vfloat4 trunc(const vfloat4& a) { return _mm_round_ps(a, _MM_FROUND_TO_ZERO      ); }
+  __forceinline vfloat4 round(const vfloat4& a) { return _mm_round_ps(a, _MM_FROUND_TO_NEAREST_INT); } // (even) https://www.felixcloutier.com/x86/roundpd
 #else
   __forceinline vfloat4 floor(const vfloat4& a) { return vfloat4(floorf(a[0]),floorf(a[1]),floorf(a[2]),floorf(a[3])); }
   __forceinline vfloat4 ceil (const vfloat4& a) { return vfloat4(ceilf (a[0]),ceilf (a[1]),ceilf (a[2]),ceilf (a[3])); }
@@ -504,7 +646,9 @@ namespace embree
   __forceinline vfloat4 frac(const vfloat4& a) { return a-floor(a); }
 
   __forceinline vint4 floori(const vfloat4& a) {
-#if defined(__SSE4_1__)
+#if defined(__aarch64__)
+    return vcvtq_s32_f32(floor(a));
+#elif defined(__SSE4_1__)
     return vint4(floor(a));
 #else
     return vint4(a-vfloat4(0.5f));
@@ -518,6 +662,16 @@ namespace embree
   __forceinline vfloat4 unpacklo(const vfloat4& a, const vfloat4& b) { return _mm_unpacklo_ps(a, b); }
   __forceinline vfloat4 unpackhi(const vfloat4& a, const vfloat4& b) { return _mm_unpackhi_ps(a, b); }
 
+#if defined(__aarch64__)
+      template<int i0, int i1, int i2, int i3>
+      __forceinline vfloat4 shuffle(const vfloat4& v) {
+          return vreinterpretq_f32_u8(vqtbl1q_u8( (uint8x16_t)v.v, _MN_SHUFFLE(i0, i1, i2, i3)));
+      }
+      template<int i0, int i1, int i2, int i3>
+      __forceinline vfloat4 shuffle(const vfloat4& a, const vfloat4& b) {
+          return vreinterpretq_f32_u8(vqtbl2q_u8( (uint8x16x2_t){(uint8x16_t)a.v, (uint8x16_t)b.v}, _MF_SHUFFLE(i0, i1, i2, i3)));
+      }
+#else
   template<int i0, int i1, int i2, int i3>
   __forceinline vfloat4 shuffle(const vfloat4& v) {
     return _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(v), _MM_SHUFFLE(i3, i2, i1, i0)));
@@ -527,14 +681,19 @@ namespace embree
   __forceinline vfloat4 shuffle(const vfloat4& a, const vfloat4& b) {
     return _mm_shuffle_ps(a, b, _MM_SHUFFLE(i3, i2, i1, i0));
   }
+#endif
 
 #if defined (__SSSE3__)
   __forceinline vfloat4 shuffle8(const vfloat4& a, const vint4& shuf) {
-    return _mm_castsi128_ps(_mm_shuffle_epi8(_mm_castps_si128(a), shuf)); 
+    return _mm_castsi128_ps(_mm_shuffle_epi8(_mm_castps_si128(a), shuf));
   }
 #endif
 
-#if defined(__SSE3__)
+#if defined(__aarch64__) 
+  template<> __forceinline vfloat4 shuffle<0, 0, 2, 2>(const vfloat4& v) { return __m128(vqtbl1q_u8( uint8x16_t(v.v), v0022 )); }
+  template<> __forceinline vfloat4 shuffle<1, 1, 3, 3>(const vfloat4& v) { return __m128(vqtbl1q_u8( uint8x16_t(v.v), v1133)); }
+  template<> __forceinline vfloat4 shuffle<0, 1, 0, 1>(const vfloat4& v) { return __m128(vqtbl1q_u8( uint8x16_t(v.v), v0101)); }
+#elif defined(__SSE3__)
   template<> __forceinline vfloat4 shuffle<0, 0, 2, 2>(const vfloat4& v) { return _mm_moveldup_ps(v); }
   template<> __forceinline vfloat4 shuffle<1, 1, 3, 3>(const vfloat4& v) { return _mm_movehdup_ps(v); }
   template<> __forceinline vfloat4 shuffle<0, 1, 0, 1>(const vfloat4& v) { return _mm_castpd_ps(_mm_movedup_pd(_mm_castps_pd(v))); }
@@ -545,14 +704,56 @@ namespace embree
     return shuffle<i,i,i,i>(v);
   }
 
-#if defined (__SSE4_1__) && !defined(__GNUC__)
+#if defined(__aarch64__)
+  template<int i> __forceinline float extract(const vfloat4& a);
+  template<> __forceinline float extract<0>(const vfloat4& b) {
+      return b[0];
+  }
+  template<> __forceinline float extract<1>(const vfloat4& b) {
+      return b[1];
+  }
+  template<> __forceinline float extract<2>(const vfloat4& b) {
+      return b[2];
+  }
+  template<> __forceinline float extract<3>(const vfloat4& b) {
+      return b[3];
+  }
+#elif defined (__SSE4_1__) && !defined(__GNUC__)
   template<int i> __forceinline float extract(const vfloat4& a) { return _mm_cvtss_f32(_mm_extract_ps(a,i)); }
+  template<> __forceinline float extract<0>(const vfloat4& a) { return _mm_cvtss_f32(a); }
 #else
   template<int i> __forceinline float extract(const vfloat4& a) { return _mm_cvtss_f32(shuffle<i,i,i,i>(a)); }
-#endif
   template<> __forceinline float extract<0>(const vfloat4& a) { return _mm_cvtss_f32(a); }
+#endif
 
-#if defined (__SSE4_1__)
+
+#if defined(__aarch64__)
+  template<int dst>  __forceinline vfloat4 insert(const vfloat4& a, float b);
+  template<> __forceinline vfloat4 insert<0>(const vfloat4& a, float b)
+  {
+        vfloat4 c = a;
+        c[0] = b;
+        return c;
+  }
+  template<> __forceinline vfloat4 insert<1>(const vfloat4& a, float b)
+  {
+        vfloat4 c = a;
+        c[1] = b;
+        return c;
+  }
+  template<> __forceinline vfloat4 insert<2>(const vfloat4& a, float b)
+  {
+        vfloat4 c = a;
+        c[2] = b;
+        return c;
+  }
+  template<> __forceinline vfloat4 insert<3>(const vfloat4& a, float b)
+  {
+        vfloat4 c = a;
+        c[3] = b;
+        return c;
+  }
+#elif defined (__SSE4_1__)
   template<int dst, int src, int clr> __forceinline vfloat4 insert(const vfloat4& a, const vfloat4& b) { return _mm_insert_ps(a, b, (dst << 4) | (src << 6) | clr); }
   template<int dst, int src> __forceinline vfloat4 insert(const vfloat4& a, const vfloat4& b) { return insert<dst, src, 0>(a, b); }
   template<int dst> __forceinline vfloat4 insert(const vfloat4& a, const float b) { return insert<dst, 0>(a, _mm_set_ss(b)); }
@@ -561,14 +762,19 @@ namespace embree
   template<int dst>  __forceinline vfloat4 insert(const vfloat4& a, float b) { vfloat4 c = a; c[dst&3] = b; return c; }
 #endif
 
+#if defined(__aarch64__)
+  __forceinline float toScalar(const vfloat4& v) {
+    return v[0];
+  }
+#else
   __forceinline float toScalar(const vfloat4& v) { return _mm_cvtss_f32(v); }
-
+#endif
   __forceinline vfloat4 broadcast4f(const vfloat4& a, size_t k) {
     return vfloat4::broadcast(&a[k]);
   }
 
   __forceinline vfloat4 shift_right_1(const vfloat4& x) {
-    return _mm_castsi128_ps(_mm_srli_si128(_mm_castps_si128(x), 4)); 
+    return _mm_castsi128_ps(_mm_srli_si128(_mm_castps_si128(x), 4));
   }
 
 #if defined (__AVX2__)
@@ -584,7 +790,7 @@ namespace embree
   template<int i>
   __forceinline vfloat4 align_shift_right(const vfloat4& a, const vfloat4& b) {
     return _mm_castsi128_ps(_mm_alignr_epi32(_mm_castps_si128(a), _mm_castps_si128(b), i));
-  }  
+  }
 #endif
 
 
@@ -658,28 +864,39 @@ namespace embree
   ////////////////////////////////////////////////////////////////////////////////
   /// Reductions
   ////////////////////////////////////////////////////////////////////////////////
-
+#if defined(__aarch64__)
+      __forceinline vfloat4 vreduce_min(const vfloat4& v) { float h = vminvq_f32(v); return vdupq_n_f32(h); }
+      __forceinline vfloat4 vreduce_max(const vfloat4& v) { float h = vmaxvq_f32(v); return vdupq_n_f32(h); }
+      __forceinline vfloat4 vreduce_add(const vfloat4& v) { float h = vaddvq_f32(v); return vdupq_n_f32(h); }
+#else
   __forceinline vfloat4 vreduce_min(const vfloat4& v) { vfloat4 h = min(shuffle<1,0,3,2>(v),v); return min(shuffle<2,3,0,1>(h),h); }
   __forceinline vfloat4 vreduce_max(const vfloat4& v) { vfloat4 h = max(shuffle<1,0,3,2>(v),v); return max(shuffle<2,3,0,1>(h),h); }
   __forceinline vfloat4 vreduce_add(const vfloat4& v) { vfloat4 h = shuffle<1,0,3,2>(v)   + v ; return shuffle<2,3,0,1>(h)   + h ; }
+#endif
 
+#if defined(__aarch64__)
+  __forceinline float reduce_min(const vfloat4& v) { return vminvq_f32(v); }
+  __forceinline float reduce_max(const vfloat4& v) { return vmaxvq_f32(v); }
+  __forceinline float reduce_add(const vfloat4& v) { return vaddvq_f32(v); }
+#else
   __forceinline float reduce_min(const vfloat4& v) { return _mm_cvtss_f32(vreduce_min(v)); }
   __forceinline float reduce_max(const vfloat4& v) { return _mm_cvtss_f32(vreduce_max(v)); }
   __forceinline float reduce_add(const vfloat4& v) { return _mm_cvtss_f32(vreduce_add(v)); }
+#endif
 
-  __forceinline size_t select_min(const vboolf4& valid, const vfloat4& v) 
+  __forceinline size_t select_min(const vboolf4& valid, const vfloat4& v)
-  { 
+  {
-    const vfloat4 a = select(valid,v,vfloat4(pos_inf)); 
+    const vfloat4 a = select(valid,v,vfloat4(pos_inf));
     const vbool4 valid_min = valid & (a == vreduce_min(a));
-    return bsf(movemask(any(valid_min) ? valid_min : valid)); 
+    return bsf(movemask(any(valid_min) ? valid_min : valid));
   }
-  __forceinline size_t select_max(const vboolf4& valid, const vfloat4& v) 
+  __forceinline size_t select_max(const vboolf4& valid, const vfloat4& v)
-  { 
+  {
-    const vfloat4 a = select(valid,v,vfloat4(neg_inf)); 
+    const vfloat4 a = select(valid,v,vfloat4(neg_inf));
     const vbool4 valid_max = valid & (a == vreduce_max(a));
-    return bsf(movemask(any(valid_max) ? valid_max : valid)); 
+    return bsf(movemask(any(valid_max) ? valid_max : valid));
   }
-  
+
   ////////////////////////////////////////////////////////////////////////////////
   /// Euclidian Space Operators
   ////////////////////////////////////////////////////////////////////////////////
@@ -694,7 +911,7 @@ namespace embree
     const vfloat4 b0 = shuffle<1,2,0,3>(b);
     const vfloat4 a1 = shuffle<1,2,0,3>(a);
     const vfloat4 b1 = b;
-    return shuffle<1,2,0,3>(msub(a0,b0,a1*b1));
+    return shuffle<1,2,0,3>(prod_diff(a0,b0,a1,b1));
   }
 
   ////////////////////////////////////////////////////////////////////////////////

thirdparty/embree/common/simd/vfloat8_avx.h

@@ -33,7 +33,7 @@ namespace embree
     __forceinline explicit vfloat(const vfloat4& a) : v(_mm256_insertf128_ps(_mm256_castps128_ps256(a),a,1)) {}
     __forceinline vfloat(const vfloat4& a, const vfloat4& b) : v(_mm256_insertf128_ps(_mm256_castps128_ps256(a),b,1)) {}
 
-    __forceinline explicit vfloat(const char* a) : v(_mm256_loadu_ps((const float*)a)) {}
+    __forceinline explicit vfloat(const int8_t* a) : v(_mm256_loadu_ps((const float*)a)) {}
     __forceinline vfloat(float a) : v(_mm256_set1_ps(a)) {}
     __forceinline vfloat(float a, float b) : v(_mm256_set_ps(b, a, b, a, b, a, b, a)) {}
     __forceinline vfloat(float a, float b, float c, float d) : v(_mm256_set_ps(d, c, b, a, d, c, b, a)) {}
@@ -75,7 +75,7 @@ namespace embree
       return _mm256_broadcast_ps((__m128*)ptr); 
     }
 
-    static __forceinline vfloat8 load(const char* ptr) {
+    static __forceinline vfloat8 load(const int8_t* ptr) {
 #if defined(__AVX2__)
       return _mm256_cvtepi32_ps(_mm256_cvtepi8_epi32(_mm_loadu_si128((__m128i*)ptr)));
 #else
@@ -83,7 +83,7 @@ namespace embree
 #endif
     }
 
-    static __forceinline vfloat8 load(const unsigned char* ptr) {
+    static __forceinline vfloat8 load(const uint8_t* ptr) {
 #if defined(__AVX2__)
       return _mm256_cvtepi32_ps(_mm256_cvtepu8_epi32(_mm_loadu_si128((__m128i*)ptr)));
 #else
@@ -119,6 +119,12 @@ namespace embree
 
     static __forceinline void store (const vboolf8& mask, void* ptr, const vfloat8& v) { _mm256_mask_store_ps ((float*)ptr,mask,v); }
     static __forceinline void storeu(const vboolf8& mask, void* ptr, const vfloat8& v) { _mm256_mask_storeu_ps((float*)ptr,mask,v); }
+#elif defined(__aarch64__)
+    static __forceinline vfloat8 load (const vboolf8& mask, const void* ptr) { return _mm256_maskload_ps((float*)ptr,(__m256i)mask.v); }
+    static __forceinline vfloat8 loadu(const vboolf8& mask, const void* ptr) { return _mm256_maskload_ps((float*)ptr,(__m256i)mask.v); }
+
+    static __forceinline void store (const vboolf8& mask, void* ptr, const vfloat8& v) { _mm256_maskstore_ps((float*)ptr,(__m256i)mask.v,v); }
+    static __forceinline void storeu(const vboolf8& mask, void* ptr, const vfloat8& v) { _mm256_maskstore_ps((float*)ptr,(__m256i)mask.v,v); }
 #else
     static __forceinline vfloat8 load (const vboolf8& mask, const void* ptr) { return _mm256_maskload_ps((float*)ptr,(__m256i)mask); }
     static __forceinline vfloat8 loadu(const vboolf8& mask, const void* ptr) { return _mm256_maskload_ps((float*)ptr,(__m256i)mask); }
@@ -139,18 +145,18 @@ namespace embree
 
     template<int scale = 4>
     static __forceinline vfloat8 gather(const float* ptr, const vint8& index) {
-#if defined(__AVX2__)
+#if defined(__AVX2__) && !defined(__aarch64__)
       return _mm256_i32gather_ps(ptr, index ,scale);
 #else
       return vfloat8(
-          *(float*)(((char*)ptr)+scale*index[0]),
+          *(float*)(((int8_t*)ptr)+scale*index[0]),
-          *(float*)(((char*)ptr)+scale*index[1]),
+          *(float*)(((int8_t*)ptr)+scale*index[1]),
-          *(float*)(((char*)ptr)+scale*index[2]),
+          *(float*)(((int8_t*)ptr)+scale*index[2]),
-          *(float*)(((char*)ptr)+scale*index[3]),
+          *(float*)(((int8_t*)ptr)+scale*index[3]),
-          *(float*)(((char*)ptr)+scale*index[4]),
+          *(float*)(((int8_t*)ptr)+scale*index[4]),
-          *(float*)(((char*)ptr)+scale*index[5]),
+          *(float*)(((int8_t*)ptr)+scale*index[5]),
-          *(float*)(((char*)ptr)+scale*index[6]),
+          *(float*)(((int8_t*)ptr)+scale*index[6]),
-          *(float*)(((char*)ptr)+scale*index[7]));
+          *(float*)(((int8_t*)ptr)+scale*index[7]));
 #endif
     }
 
@@ -159,17 +165,17 @@ namespace embree
       vfloat8 r = zero;
 #if defined(__AVX512VL__)
       return _mm256_mmask_i32gather_ps(r, mask, index, ptr, scale);
-#elif defined(__AVX2__)
+#elif defined(__AVX2__) && !defined(__aarch64__)
       return _mm256_mask_i32gather_ps(r, ptr, index, mask, scale);
 #else
-      if (likely(mask[0])) r[0] = *(float*)(((char*)ptr)+scale*index[0]);
+      if (likely(mask[0])) r[0] = *(float*)(((int8_t*)ptr)+scale*index[0]);
-      if (likely(mask[1])) r[1] = *(float*)(((char*)ptr)+scale*index[1]);
+      if (likely(mask[1])) r[1] = *(float*)(((int8_t*)ptr)+scale*index[1]);
-      if (likely(mask[2])) r[2] = *(float*)(((char*)ptr)+scale*index[2]);
+      if (likely(mask[2])) r[2] = *(float*)(((int8_t*)ptr)+scale*index[2]);
-      if (likely(mask[3])) r[3] = *(float*)(((char*)ptr)+scale*index[3]);
+      if (likely(mask[3])) r[3] = *(float*)(((int8_t*)ptr)+scale*index[3]);
-      if (likely(mask[4])) r[4] = *(float*)(((char*)ptr)+scale*index[4]);
+      if (likely(mask[4])) r[4] = *(float*)(((int8_t*)ptr)+scale*index[4]);
-      if (likely(mask[5])) r[5] = *(float*)(((char*)ptr)+scale*index[5]);
+      if (likely(mask[5])) r[5] = *(float*)(((int8_t*)ptr)+scale*index[5]);
-      if (likely(mask[6])) r[6] = *(float*)(((char*)ptr)+scale*index[6]);
+      if (likely(mask[6])) r[6] = *(float*)(((int8_t*)ptr)+scale*index[6]);
-      if (likely(mask[7])) r[7] = *(float*)(((char*)ptr)+scale*index[7]);
+      if (likely(mask[7])) r[7] = *(float*)(((int8_t*)ptr)+scale*index[7]);
       return r;
     #endif
     }
@@ -180,14 +186,14 @@ namespace embree
 #if defined(__AVX512VL__)
       _mm256_i32scatter_ps((float*)ptr, ofs, v, scale);
 #else
-      *(float*)(((char*)ptr)+scale*ofs[0]) = v[0];
+      *(float*)(((int8_t*)ptr)+scale*ofs[0]) = v[0];
-      *(float*)(((char*)ptr)+scale*ofs[1]) = v[1];
+      *(float*)(((int8_t*)ptr)+scale*ofs[1]) = v[1];
-      *(float*)(((char*)ptr)+scale*ofs[2]) = v[2];
+      *(float*)(((int8_t*)ptr)+scale*ofs[2]) = v[2];
-      *(float*)(((char*)ptr)+scale*ofs[3]) = v[3];
+      *(float*)(((int8_t*)ptr)+scale*ofs[3]) = v[3];
-      *(float*)(((char*)ptr)+scale*ofs[4]) = v[4];
+      *(float*)(((int8_t*)ptr)+scale*ofs[4]) = v[4];
-      *(float*)(((char*)ptr)+scale*ofs[5]) = v[5];
+      *(float*)(((int8_t*)ptr)+scale*ofs[5]) = v[5];
-      *(float*)(((char*)ptr)+scale*ofs[6]) = v[6];
+      *(float*)(((int8_t*)ptr)+scale*ofs[6]) = v[6];
-      *(float*)(((char*)ptr)+scale*ofs[7]) = v[7];
+      *(float*)(((int8_t*)ptr)+scale*ofs[7]) = v[7];
 #endif
     }
 
@@ -197,18 +203,18 @@ namespace embree
 #if defined(__AVX512VL__)
       _mm256_mask_i32scatter_ps((float*)ptr, mask, ofs, v, scale);
 #else
-      if (likely(mask[0])) *(float*)(((char*)ptr)+scale*ofs[0]) = v[0];
+      if (likely(mask[0])) *(float*)(((int8_t*)ptr)+scale*ofs[0]) = v[0];
-      if (likely(mask[1])) *(float*)(((char*)ptr)+scale*ofs[1]) = v[1];
+      if (likely(mask[1])) *(float*)(((int8_t*)ptr)+scale*ofs[1]) = v[1];
-      if (likely(mask[2])) *(float*)(((char*)ptr)+scale*ofs[2]) = v[2];
+      if (likely(mask[2])) *(float*)(((int8_t*)ptr)+scale*ofs[2]) = v[2];
-      if (likely(mask[3])) *(float*)(((char*)ptr)+scale*ofs[3]) = v[3];
+      if (likely(mask[3])) *(float*)(((int8_t*)ptr)+scale*ofs[3]) = v[3];
-      if (likely(mask[4])) *(float*)(((char*)ptr)+scale*ofs[4]) = v[4];
+      if (likely(mask[4])) *(float*)(((int8_t*)ptr)+scale*ofs[4]) = v[4];
-      if (likely(mask[5])) *(float*)(((char*)ptr)+scale*ofs[5]) = v[5];
+      if (likely(mask[5])) *(float*)(((int8_t*)ptr)+scale*ofs[5]) = v[5];
-      if (likely(mask[6])) *(float*)(((char*)ptr)+scale*ofs[6]) = v[6];
+      if (likely(mask[6])) *(float*)(((int8_t*)ptr)+scale*ofs[6]) = v[6];
-      if (likely(mask[7])) *(float*)(((char*)ptr)+scale*ofs[7]) = v[7];
+      if (likely(mask[7])) *(float*)(((int8_t*)ptr)+scale*ofs[7]) = v[7];
 #endif
     }
 
-    static __forceinline void store(const vboolf8& mask, char* ptr, const vint8& ofs, const vfloat8& v) {
+    static __forceinline void store(const vboolf8& mask, int8_t* ptr, const vint8& ofs, const vfloat8& v) {
       scatter<1>(mask,ptr,ofs,v);
     }
     static __forceinline void store(const vboolf8& mask, float* ptr, const vint8& ofs, const vfloat8& v) {
@@ -235,27 +241,60 @@ namespace embree
   __forceinline vfloat8 toFloat(const vint8&   a) { return vfloat8(a); }
 
   __forceinline vfloat8 operator +(const vfloat8& a) { return a; }
+#if !defined(__aarch64__)
   __forceinline vfloat8 operator -(const vfloat8& a) {
     const __m256 mask = _mm256_castsi256_ps(_mm256_set1_epi32(0x80000000)); 
     return _mm256_xor_ps(a, mask);
   }
-  __forceinline vfloat8 abs(const vfloat8& a) {
+#else
-    const __m256 mask = _mm256_castsi256_ps(_mm256_set1_epi32(0x7fffffff));
+  __forceinline vfloat8 operator -(const vfloat8& a) {
-    return _mm256_and_ps(a, mask);
+      __m256 res;
-  }
+      res.lo = vnegq_f32(a.v.lo);
+      res.hi = vnegq_f32(a.v.hi);
+      return res;
+}
+#endif
+
+#if !defined(__aarch64__)
+__forceinline vfloat8 abs(const vfloat8& a) {
+  const __m256 mask = _mm256_castsi256_ps(_mm256_set1_epi32(0x7fffffff));
+  return _mm256_and_ps(a, mask);
+}
+#else
+__forceinline vfloat8 abs(const vfloat8& a) {
+    __m256 res;
+    res.lo = vabsq_f32(a.v.lo);
+    res.hi = vabsq_f32(a.v.hi);
+    return res;
+}
+#endif
+
+#if !defined(__aarch64__)
   __forceinline vfloat8 sign   (const vfloat8& a) { return _mm256_blendv_ps(vfloat8(one), -vfloat8(one), _mm256_cmp_ps(a, vfloat8(zero), _CMP_NGE_UQ)); }
+#else
+  __forceinline vfloat8 sign   (const vfloat8& a) { return _mm256_blendv_ps(vfloat8(one), -vfloat8(one), _mm256_cmplt_ps(a, vfloat8(zero))); }
+#endif
   __forceinline vfloat8 signmsk(const vfloat8& a) { return _mm256_and_ps(a,_mm256_castsi256_ps(_mm256_set1_epi32(0x80000000))); }
 
 
   static __forceinline vfloat8 rcp(const vfloat8& a)
   {
+#if defined(BUILD_IOS) && defined(__aarch64__)
+    // ios devices are faster doing full divide, no need for NR fixup
+    vfloat8 ret;
+    const float32x4_t one = vdupq_n_f32(1.0f);
+    ret.v.lo = vdivq_f32(one, a.v.lo);
+    ret.v.hi = vdivq_f32(one, a.v.hi);
+    return ret;
+#endif
+
 #if defined(__AVX512VL__)
     const vfloat8 r = _mm256_rcp14_ps(a);
 #else
     const vfloat8 r = _mm256_rcp_ps(a);
 #endif
-
+      
-#if defined(__AVX2__)
+#if defined(__AVX2__) //&& !defined(aarch64)
     return _mm256_mul_ps(r, _mm256_fnmadd_ps(r, a, vfloat8(2.0f)));
 #else
     return _mm256_mul_ps(r, _mm256_sub_ps(vfloat8(2.0f), _mm256_mul_ps(r, a)));
@@ -404,17 +443,29 @@ namespace embree
   static __forceinline vfloat8 select(const vboolf8& m, const vfloat8& t, const vfloat8& f) {
     return _mm256_mask_blend_ps(m, f, t);
   }
-#else
+#elif !defined(__aarch64__)
-  static __forceinline vboolf8 operator ==(const vfloat8& a, const vfloat8& b) { return _mm256_cmp_ps(a, b, _CMP_EQ_OQ);  }
+  __forceinline vboolf8 operator ==(const vfloat8& a, const vfloat8& b) { return _mm256_cmp_ps(a, b, _CMP_EQ_OQ);  }
-  static __forceinline vboolf8 operator !=(const vfloat8& a, const vfloat8& b) { return _mm256_cmp_ps(a, b, _CMP_NEQ_UQ); }
+  __forceinline vboolf8 operator !=(const vfloat8& a, const vfloat8& b) { return _mm256_cmp_ps(a, b, _CMP_NEQ_UQ); }
-  static __forceinline vboolf8 operator < (const vfloat8& a, const vfloat8& b) { return _mm256_cmp_ps(a, b, _CMP_LT_OS);  }
+  __forceinline vboolf8 operator < (const vfloat8& a, const vfloat8& b) { return _mm256_cmp_ps(a, b, _CMP_LT_OS);  }
-  static __forceinline vboolf8 operator >=(const vfloat8& a, const vfloat8& b) { return _mm256_cmp_ps(a, b, _CMP_NLT_US); }
+  __forceinline vboolf8 operator >=(const vfloat8& a, const vfloat8& b) { return _mm256_cmp_ps(a, b, _CMP_NLT_US); }
-  static __forceinline vboolf8 operator > (const vfloat8& a, const vfloat8& b) { return _mm256_cmp_ps(a, b, _CMP_NLE_US); }
+  __forceinline vboolf8 operator > (const vfloat8& a, const vfloat8& b) { return _mm256_cmp_ps(a, b, _CMP_NLE_US); }
-  static __forceinline vboolf8 operator <=(const vfloat8& a, const vfloat8& b) { return _mm256_cmp_ps(a, b, _CMP_LE_OS);  }
+  __forceinline vboolf8 operator <=(const vfloat8& a, const vfloat8& b) { return _mm256_cmp_ps(a, b, _CMP_LE_OS);  }
 
-  static __forceinline vfloat8 select(const vboolf8& m, const vfloat8& t, const vfloat8& f) {
+  __forceinline vfloat8 select(const vboolf8& m, const vfloat8& t, const vfloat8& f) {
     return _mm256_blendv_ps(f, t, m); 
   }
+#else
+  __forceinline vboolf8 operator ==(const vfloat8& a, const vfloat8& b) { return _mm256_cmpeq_ps(a, b);  }
+  __forceinline vboolf8 operator !=(const vfloat8& a, const vfloat8& b) { return _mm256_cmpneq_ps(a, b); }
+  __forceinline vboolf8 operator < (const vfloat8& a, const vfloat8& b) { return _mm256_cmplt_ps(a, b);  }
+  __forceinline vboolf8 operator >=(const vfloat8& a, const vfloat8& b) { return _mm256_cmpge_ps(a, b);  }
+  __forceinline vboolf8 operator > (const vfloat8& a, const vfloat8& b) { return _mm256_cmpgt_ps(a, b);  }
+  __forceinline vboolf8 operator <=(const vfloat8& a, const vfloat8& b) { return _mm256_cmple_ps(a, b);  }
+
+  __forceinline vfloat8 select(const vboolf8& m, const vfloat8& t, const vfloat8& f) {
+    return _mm256_blendv_ps(f, t, m);
+  }
+
 #endif
 
   template<int mask>
@@ -483,10 +534,17 @@ namespace embree
   /// Rounding Functions
   ////////////////////////////////////////////////////////////////////////////////
 
+#if !defined(__aarch64__)
   __forceinline vfloat8 floor(const vfloat8& a) { return _mm256_round_ps(a, _MM_FROUND_TO_NEG_INF    ); }
   __forceinline vfloat8 ceil (const vfloat8& a) { return _mm256_round_ps(a, _MM_FROUND_TO_POS_INF    ); }
   __forceinline vfloat8 trunc(const vfloat8& a) { return _mm256_round_ps(a, _MM_FROUND_TO_ZERO       ); }
   __forceinline vfloat8 round(const vfloat8& a) { return _mm256_round_ps(a, _MM_FROUND_TO_NEAREST_INT); }
+#else
+  __forceinline vfloat8 floor(const vfloat8& a) { return _mm256_floor_ps(a); }
+  __forceinline vfloat8 ceil (const vfloat8& a) { return _mm256_ceil_ps(a); }
+#endif
+
+
   __forceinline vfloat8 frac (const vfloat8& a) { return a-floor(a); }
 
   ////////////////////////////////////////////////////////////////////////////////
@@ -521,9 +579,11 @@ namespace embree
     return _mm256_shuffle_ps(a, b, _MM_SHUFFLE(i3, i2, i1, i0));
   }
 
+#if !defined(__aarch64__)
   template<> __forceinline vfloat8 shuffle<0, 0, 2, 2>(const vfloat8& v) { return _mm256_moveldup_ps(v); }
   template<> __forceinline vfloat8 shuffle<1, 1, 3, 3>(const vfloat8& v) { return _mm256_movehdup_ps(v); }
   template<> __forceinline vfloat8 shuffle<0, 1, 0, 1>(const vfloat8& v) { return _mm256_castpd_ps(_mm256_movedup_pd(_mm256_castps_pd(v))); }
+#endif
 
   __forceinline vfloat8 broadcast(const float* ptr) { return _mm256_broadcast_ss(ptr); }
   template<size_t i> __forceinline vfloat8 insert4(const vfloat8& a, const vfloat4& b) { return _mm256_insertf128_ps(a, b, i); }
@@ -534,8 +594,8 @@ namespace embree
 
   __forceinline vfloat8 assign(const vfloat4& a) { return _mm256_castps128_ps256(a); }
 
-#if defined (__AVX2__)
+#if defined (__AVX2__) && !defined(__aarch64__)
-  static __forceinline vfloat8 permute(const vfloat8& a, const __m256i& index) {
+  __forceinline vfloat8 permute(const vfloat8& a, const __m256i& index) {
     return _mm256_permutevar8x32_ps(a, index);
   }
 #endif
@@ -639,7 +699,7 @@ namespace embree
   ////////////////////////////////////////////////////////////////////////////////
   /// Reductions
   ////////////////////////////////////////////////////////////////////////////////
-
+#if !defined(__aarch64__)
   __forceinline vfloat8 vreduce_min2(const vfloat8& v) { return min(v,shuffle<1,0,3,2>(v)); }
   __forceinline vfloat8 vreduce_min4(const vfloat8& v) { vfloat8 v1 = vreduce_min2(v); return min(v1,shuffle<2,3,0,1>(v1)); }
   __forceinline vfloat8 vreduce_min (const vfloat8& v) { vfloat8 v1 = vreduce_min4(v); return min(v1,shuffle4<1,0>(v1)); }
@@ -655,7 +715,14 @@ namespace embree
   __forceinline float reduce_min(const vfloat8& v) { return toScalar(vreduce_min(v)); }
   __forceinline float reduce_max(const vfloat8& v) { return toScalar(vreduce_max(v)); }
   __forceinline float reduce_add(const vfloat8& v) { return toScalar(vreduce_add(v)); }
+#else
+  __forceinline float reduce_min(const vfloat8& v) { return vminvq_f32(_mm_min_ps(v.v.lo,v.v.hi)); }
+  __forceinline float reduce_max(const vfloat8& v) { return vmaxvq_f32(_mm_max_ps(v.v.lo,v.v.hi)); }
+  __forceinline vfloat8 vreduce_min(const vfloat8& v) { return vfloat8(reduce_min(v)); }
+  __forceinline vfloat8 vreduce_max(const vfloat8& v) { return vfloat8(reduce_max(v)); }
+  __forceinline float reduce_add(const vfloat8& v) { return vaddvq_f32(_mm_add_ps(v.v.lo,v.v.hi)); }
 
+#endif
   __forceinline size_t select_min(const vboolf8& valid, const vfloat8& v) 
   { 
     const vfloat8 a = select(valid,v,vfloat8(pos_inf)); 

thirdparty/embree/common/simd/vint16_avx512.h

@@ -90,10 +90,10 @@ namespace embree
 
     static __forceinline vint16 load (const void* addr) { return _mm512_load_si512((int*)addr); }
 
-    static __forceinline vint16 load(const unsigned char* ptr) { return _mm512_cvtepu8_epi32(_mm_load_si128((__m128i*)ptr)); }
+    static __forceinline vint16 load(const uint8_t* ptr) { return _mm512_cvtepu8_epi32(_mm_load_si128((__m128i*)ptr)); }
     static __forceinline vint16 load(const unsigned short* ptr) { return _mm512_cvtepu16_epi32(_mm256_load_si256((__m256i*)ptr)); }
 
-    static __forceinline vint16 loadu(const unsigned char* ptr) { return _mm512_cvtepu8_epi32(_mm_loadu_si128((__m128i*)ptr)); }
+    static __forceinline vint16 loadu(const uint8_t* ptr) { return _mm512_cvtepu8_epi32(_mm_loadu_si128((__m128i*)ptr)); }
     static __forceinline vint16 loadu(const unsigned short* ptr) { return _mm512_cvtepu16_epi32(_mm256_loadu_si256((__m256i*)ptr)); }
 
     static __forceinline vint16 loadu(const void* addr) { return _mm512_loadu_si512(addr); }

thirdparty/embree/common/simd/vint4_sse2.h

@@ -23,7 +23,7 @@ namespace embree
     ////////////////////////////////////////////////////////////////////////////////
     /// Constructors, Assignment & Cast Operators
     ////////////////////////////////////////////////////////////////////////////////
-    
+
     __forceinline vint() {}
     __forceinline vint(const vint4& a) { v = a.v; }
     __forceinline vint4& operator =(const vint4& a) { v = a.v; return *this; }
@@ -68,7 +68,7 @@ namespace embree
 
     static __forceinline void store (void* ptr, const vint4& v) { _mm_store_si128((__m128i*)ptr,v); }
     static __forceinline void storeu(void* ptr, const vint4& v) { _mm_storeu_si128((__m128i*)ptr,v); }
-    
+
 #if defined(__AVX512VL__)
 
     static __forceinline vint4 compact(const vboolf4& mask, vint4 &v) {
@@ -98,61 +98,81 @@ namespace embree
 #endif
 
 
-#if defined(__SSE4_1__)
+#if defined(__aarch64__)
-    static __forceinline vint4 load(const unsigned char* ptr) {
+    static __forceinline vint4 load(const uint8_t* ptr) {
+        return _mm_load4epu8_epi32(((__m128i*)ptr));
+    }
+    static __forceinline vint4 loadu(const uint8_t* ptr) {
+        return  _mm_load4epu8_epi32(((__m128i*)ptr));
+    }
+#elif defined(__SSE4_1__)
+    static __forceinline vint4 load(const uint8_t* ptr) {
       return _mm_cvtepu8_epi32(_mm_loadl_epi64((__m128i*)ptr));
     }
 
-    static __forceinline vint4 loadu(const unsigned char* ptr) {
+    static __forceinline vint4 loadu(const uint8_t* ptr) {
       return  _mm_cvtepu8_epi32(_mm_loadl_epi64((__m128i*)ptr));
     }
 #else
 
-    static __forceinline vint4 load(const unsigned char* ptr) {
+    static __forceinline vint4 load(const uint8_t* ptr) {
       return vint4(ptr[0],ptr[1],ptr[2],ptr[3]);
-    } 
+    }
 
-    static __forceinline vint4 loadu(const unsigned char* ptr) {
+    static __forceinline vint4 loadu(const uint8_t* ptr) {
       return vint4(ptr[0],ptr[1],ptr[2],ptr[3]);
     }
 
 #endif
 
     static __forceinline vint4 load(const unsigned short* ptr) {
-#if defined (__SSE4_1__)
+#if defined(__aarch64__)
+      return __m128i(vmovl_u16(vld1_u16(ptr)));
+#elif defined (__SSE4_1__)
       return _mm_cvtepu16_epi32(_mm_loadu_si128((__m128i*)ptr));
 #else
       return vint4(ptr[0],ptr[1],ptr[2],ptr[3]);
 #endif
-    } 
+    }
 
-    static __forceinline void store(unsigned char* ptr, const vint4& v) {
+    static __forceinline void store(uint8_t* ptr, const vint4& v) {
-#if defined(__SSE4_1__)
+#if defined(__aarch64__)
+        int32x4_t x = v;
+        uint16x4_t y = vqmovn_u32(uint32x4_t(x));
+        uint8x8_t z = vqmovn_u16(vcombine_u16(y, y));
+        vst1_lane_u32((uint32_t *)ptr,uint32x2_t(z), 0);
+#elif defined(__SSE4_1__)
       __m128i x = v;
       x = _mm_packus_epi32(x, x);
       x = _mm_packus_epi16(x, x);
       *(int*)ptr = _mm_cvtsi128_si32(x);
 #else
       for (size_t i=0;i<4;i++)
-        ptr[i] = (unsigned char)v[i];
+        ptr[i] = (uint8_t)v[i];
 #endif
     }
 
     static __forceinline void store(unsigned short* ptr, const vint4& v) {
+#if defined(__aarch64__)
+      uint32x4_t x = uint32x4_t(v.v);
+      uint16x4_t y = vqmovn_u32(x);
+      vst1_u16(ptr, y);
+#else
       for (size_t i=0;i<4;i++)
         ptr[i] = (unsigned short)v[i];
+#endif
     }
 
     static __forceinline vint4 load_nt(void* ptr) {
-#if defined(__SSE4_1__)
+#if defined(__aarch64__) || defined(__SSE4_1__)
-      return _mm_stream_load_si128((__m128i*)ptr); 
+      return _mm_stream_load_si128((__m128i*)ptr);
 #else
-      return _mm_load_si128((__m128i*)ptr); 
+      return _mm_load_si128((__m128i*)ptr);
 #endif
     }
-    
+
     static __forceinline void store_nt(void* ptr, const vint4& v) {
-#if defined(__SSE4_1__)
+#if !defined(__aarch64__) && defined(__SSE4_1__)
       _mm_stream_ps((float*)ptr, _mm_castsi128_ps(v));
 #else
       _mm_store_si128((__m128i*)ptr,v);
@@ -161,14 +181,14 @@ namespace embree
 
     template<int scale = 4>
     static __forceinline vint4 gather(const int* ptr, const vint4& index) {
-#if defined(__AVX2__)
+#if defined(__AVX2__) && !defined(__aarch64__)
       return _mm_i32gather_epi32(ptr, index, scale);
 #else
       return vint4(
-          *(int*)(((char*)ptr)+scale*index[0]),
+          *(int*)(((int8_t*)ptr)+scale*index[0]),
-          *(int*)(((char*)ptr)+scale*index[1]),
+          *(int*)(((int8_t*)ptr)+scale*index[1]),
-          *(int*)(((char*)ptr)+scale*index[2]),
+          *(int*)(((int8_t*)ptr)+scale*index[2]),
-          *(int*)(((char*)ptr)+scale*index[3]));
+          *(int*)(((int8_t*)ptr)+scale*index[3]));
 #endif
     }
 
@@ -177,13 +197,13 @@ namespace embree
       vint4 r = zero;
 #if defined(__AVX512VL__)
       return _mm_mmask_i32gather_epi32(r, mask, index, ptr, scale);
-#elif defined(__AVX2__)
+#elif defined(__AVX2__) && !defined(__aarch64__)
       return _mm_mask_i32gather_epi32(r, ptr, index, mask, scale);
 #else
-      if (likely(mask[0])) r[0] = *(int*)(((char*)ptr)+scale*index[0]);
+      if (likely(mask[0])) r[0] = *(int*)(((int8_t*)ptr)+scale*index[0]);
-      if (likely(mask[1])) r[1] = *(int*)(((char*)ptr)+scale*index[1]);
+      if (likely(mask[1])) r[1] = *(int*)(((int8_t*)ptr)+scale*index[1]);
-      if (likely(mask[2])) r[2] = *(int*)(((char*)ptr)+scale*index[2]);
+      if (likely(mask[2])) r[2] = *(int*)(((int8_t*)ptr)+scale*index[2]);
-      if (likely(mask[3])) r[3] = *(int*)(((char*)ptr)+scale*index[3]);
+      if (likely(mask[3])) r[3] = *(int*)(((int8_t*)ptr)+scale*index[3]);
       return r;
 #endif
     }
@@ -194,10 +214,10 @@ namespace embree
 #if defined(__AVX512VL__)
       _mm_i32scatter_epi32((int*)ptr, index, v, scale);
 #else
-      *(int*)(((char*)ptr)+scale*index[0]) = v[0];
+      *(int*)(((int8_t*)ptr)+scale*index[0]) = v[0];
-      *(int*)(((char*)ptr)+scale*index[1]) = v[1];
+      *(int*)(((int8_t*)ptr)+scale*index[1]) = v[1];
-      *(int*)(((char*)ptr)+scale*index[2]) = v[2];
+      *(int*)(((int8_t*)ptr)+scale*index[2]) = v[2];
-      *(int*)(((char*)ptr)+scale*index[3]) = v[3];
+      *(int*)(((int8_t*)ptr)+scale*index[3]) = v[3];
 #endif
     }
 
@@ -207,14 +227,14 @@ namespace embree
 #if defined(__AVX512VL__)
       _mm_mask_i32scatter_epi32((int*)ptr, mask, index, v, scale);
 #else
-      if (likely(mask[0])) *(int*)(((char*)ptr)+scale*index[0]) = v[0];
+      if (likely(mask[0])) *(int*)(((int8_t*)ptr)+scale*index[0]) = v[0];
-      if (likely(mask[1])) *(int*)(((char*)ptr)+scale*index[1]) = v[1];
+      if (likely(mask[1])) *(int*)(((int8_t*)ptr)+scale*index[1]) = v[1];
-      if (likely(mask[2])) *(int*)(((char*)ptr)+scale*index[2]) = v[2];
+      if (likely(mask[2])) *(int*)(((int8_t*)ptr)+scale*index[2]) = v[2];
-      if (likely(mask[3])) *(int*)(((char*)ptr)+scale*index[3]) = v[3];
+      if (likely(mask[3])) *(int*)(((int8_t*)ptr)+scale*index[3]) = v[3];
 #endif
     }
 
-#if defined(__x86_64__)
+#if defined(__x86_64__) || defined(__aarch64__)
     static __forceinline vint4 broadcast64(long long a) { return _mm_set1_epi64x(a); }
 #endif
 
@@ -228,10 +248,12 @@ namespace embree
     friend __forceinline vint4 select(const vboolf4& m, const vint4& t, const vint4& f) {
 #if defined(__AVX512VL__)
       return _mm_mask_blend_epi32(m, (__m128i)f, (__m128i)t);
+#elif defined(__aarch64__)
+      return _mm_castps_si128(_mm_blendv_ps((__m128)f.v,(__m128) t.v, (__m128)m.v));
 #elif defined(__SSE4_1__)
-      return _mm_castps_si128(_mm_blendv_ps(_mm_castsi128_ps(f), _mm_castsi128_ps(t), m)); 
+      return _mm_castps_si128(_mm_blendv_ps(_mm_castsi128_ps(f), _mm_castsi128_ps(t), m));
 #else
-      return _mm_or_si128(_mm_and_si128(m, t), _mm_andnot_si128(m, f)); 
+      return _mm_or_si128(_mm_and_si128(m, t), _mm_andnot_si128(m, f));
 #endif
     }
   };
@@ -248,7 +270,9 @@ namespace embree
 
   __forceinline vint4 operator +(const vint4& a) { return a; }
   __forceinline vint4 operator -(const vint4& a) { return _mm_sub_epi32(_mm_setzero_si128(), a); }
-#if defined(__SSSE3__)
+#if defined(__aarch64__)
+  __forceinline vint4 abs(const vint4& a) { return vabsq_s32(a.v); }
+#elif defined(__SSSE3__)
   __forceinline vint4 abs(const vint4& a) { return _mm_abs_epi32(a); }
 #endif
 
@@ -264,7 +288,7 @@ namespace embree
   __forceinline vint4 operator -(const vint4& a, int          b) { return a - vint4(b); }
   __forceinline vint4 operator -(int          a, const vint4& b) { return vint4(a) - b; }
 
-#if defined(__SSE4_1__)
+#if (defined(__aarch64__)) || defined(__SSE4_1__)
   __forceinline vint4 operator *(const vint4& a, const vint4& b) { return _mm_mullo_epi32(a, b); }
 #else
   __forceinline vint4 operator *(const vint4& a, const vint4& b) { return vint4(a[0]*b[0],a[1]*b[1],a[2]*b[2],a[3]*b[3]); }
@@ -284,34 +308,34 @@ namespace embree
   __forceinline vint4 operator ^(const vint4& a, int          b) { return a ^ vint4(b); }
   __forceinline vint4 operator ^(int          a, const vint4& b) { return vint4(a) ^ b; }
 
-  __forceinline vint4 operator <<(const vint4& a, int n) { return _mm_slli_epi32(a, n); }
+  __forceinline vint4 operator <<(const vint4& a, const int n) { return _mm_slli_epi32(a, n); }
-  __forceinline vint4 operator >>(const vint4& a, int n) { return _mm_srai_epi32(a, n); }
+  __forceinline vint4 operator >>(const vint4& a, const int n) { return _mm_srai_epi32(a, n); }
 
   __forceinline vint4 sll (const vint4& a, int b) { return _mm_slli_epi32(a, b); }
   __forceinline vint4 sra (const vint4& a, int b) { return _mm_srai_epi32(a, b); }
   __forceinline vint4 srl (const vint4& a, int b) { return _mm_srli_epi32(a, b); }
-  
+
   ////////////////////////////////////////////////////////////////////////////////
   /// Assignment Operators
   ////////////////////////////////////////////////////////////////////////////////
 
   __forceinline vint4& operator +=(vint4& a, const vint4& b) { return a = a + b; }
   __forceinline vint4& operator +=(vint4& a, int          b) { return a = a + b; }
-  
+
   __forceinline vint4& operator -=(vint4& a, const vint4& b) { return a = a - b; }
   __forceinline vint4& operator -=(vint4& a, int          b) { return a = a - b; }
 
-#if defined(__SSE4_1__)
+#if (defined(__aarch64__)) || defined(__SSE4_1__)
   __forceinline vint4& operator *=(vint4& a, const vint4& b) { return a = a * b; }
   __forceinline vint4& operator *=(vint4& a, int          b) { return a = a * b; }
 #endif
-  
+
   __forceinline vint4& operator &=(vint4& a, const vint4& b) { return a = a & b; }
   __forceinline vint4& operator &=(vint4& a, int          b) { return a = a & b; }
-  
+
   __forceinline vint4& operator |=(vint4& a, const vint4& b) { return a = a | b; }
   __forceinline vint4& operator |=(vint4& a, int          b) { return a = a | b; }
-  
+
   __forceinline vint4& operator <<=(vint4& a, int b) { return a = a << b; }
   __forceinline vint4& operator >>=(vint4& a, int b) { return a = a >> b; }
 
@@ -378,14 +402,15 @@ namespace embree
 
   template<int mask>
   __forceinline vint4 select(const vint4& t, const vint4& f) {
-#if defined(__SSE4_1__) 
+#if defined(__SSE4_1__)
     return _mm_castps_si128(_mm_blend_ps(_mm_castsi128_ps(f), _mm_castsi128_ps(t), mask));
 #else
     return select(vboolf4(mask), t, f);
-#endif    
+#endif
   }
 
-#if defined(__SSE4_1__)
+      
+#if defined(__aarch64__) || defined(__SSE4_1__)
   __forceinline vint4 min(const vint4& a, const vint4& b) { return _mm_min_epi32(a, b); }
   __forceinline vint4 max(const vint4& a, const vint4& b) { return _mm_max_epi32(a, b); }
 
@@ -409,16 +434,25 @@ namespace embree
   __forceinline vint4 unpacklo(const vint4& a, const vint4& b) { return _mm_castps_si128(_mm_unpacklo_ps(_mm_castsi128_ps(a), _mm_castsi128_ps(b))); }
   __forceinline vint4 unpackhi(const vint4& a, const vint4& b) { return _mm_castps_si128(_mm_unpackhi_ps(_mm_castsi128_ps(a), _mm_castsi128_ps(b))); }
 
+#if defined(__aarch64__)
+    template<int i0, int i1, int i2, int i3>
+    __forceinline vint4 shuffle(const vint4& v) {
+        return vreinterpretq_s32_u8(vqtbl1q_u8( (uint8x16_t)v.v, _MN_SHUFFLE(i0, i1, i2, i3)));
+    }
+    template<int i0, int i1, int i2, int i3>
+    __forceinline vint4 shuffle(const vint4& a, const vint4& b) {
+        return vreinterpretq_s32_u8(vqtbl2q_u8( (uint8x16x2_t){(uint8x16_t)a.v, (uint8x16_t)b.v}, _MF_SHUFFLE(i0, i1, i2, i3)));
+    }
+#else
   template<int i0, int i1, int i2, int i3>
   __forceinline vint4 shuffle(const vint4& v) {
     return _mm_shuffle_epi32(v, _MM_SHUFFLE(i3, i2, i1, i0));
   }
-
   template<int i0, int i1, int i2, int i3>
   __forceinline vint4 shuffle(const vint4& a, const vint4& b) {
     return _mm_castps_si128(_mm_shuffle_ps(_mm_castsi128_ps(a), _mm_castsi128_ps(b), _MM_SHUFFLE(i3, i2, i1, i0)));
   }
-
+#endif
 #if defined(__SSE3__)
   template<> __forceinline vint4 shuffle<0, 0, 2, 2>(const vint4& v) { return _mm_castps_si128(_mm_moveldup_ps(_mm_castsi128_ps(v))); }
   template<> __forceinline vint4 shuffle<1, 1, 3, 3>(const vint4& v) { return _mm_castps_si128(_mm_movehdup_ps(_mm_castsi128_ps(v))); }
@@ -430,7 +464,10 @@ namespace embree
     return shuffle<i,i,i,i>(v);
   }
 
-#if defined(__SSE4_1__)
+#if defined(__aarch64__)
+    template<int src> __forceinline int extract(const vint4& b);
+    template<int dst> __forceinline vint4 insert(const vint4& a, const int b);
+#elif defined(__SSE4_1__)
   template<int src> __forceinline int extract(const vint4& b) { return _mm_extract_epi32(b, src); }
   template<int dst> __forceinline vint4 insert(const vint4& a, const int b) { return _mm_insert_epi32(a, b, dst); }
 #else
@@ -438,19 +475,69 @@ namespace embree
   template<int dst> __forceinline vint4 insert(const vint4& a, int b) { vint4 c = a; c[dst&3] = b; return c; }
 #endif
 
-
+#if defined(__aarch64__)
+    template<> __forceinline int extract<0>(const vint4& b) {
+        return b.v[0];
+    }
+    template<> __forceinline int extract<1>(const vint4& b) {
+        return b.v[1];
+    }
+    template<> __forceinline int extract<2>(const vint4& b) {
+        return b.v[2];
+    }
+    template<> __forceinline int extract<3>(const vint4& b) {
+        return b.v[3];
+    }
+    template<> __forceinline vint4 insert<0>(const vint4& a, int b)
+    {
+        vint4 c = a;
+        c[0] = b;
+        return c;
+    }
+    template<> __forceinline vint4 insert<1>(const vint4& a, int b)
+    {
+        vint4 c = a;
+        c[1] = b;
+        return c;
+    }
+    template<> __forceinline vint4 insert<2>(const vint4& a, int b)
+    {
+        vint4 c = a;
+        c[2] = b;
+        return c;
+    }
+    template<> __forceinline vint4 insert<3>(const vint4& a, int b)
+    {
+        vint4 c = a;
+        c[3] = b;
+        return c;
+    }
+      
+    __forceinline int toScalar(const vint4& v) {
+        return v[0];
+    }
+      
+    __forceinline size_t toSizeT(const vint4& v) {
+        uint64x2_t x = uint64x2_t(v.v);
+        return x[0];
+    }
+#else
   template<> __forceinline int extract<0>(const vint4& b) { return _mm_cvtsi128_si32(b); }
 
   __forceinline int toScalar(const vint4& v) { return _mm_cvtsi128_si32(v); }
 
-  __forceinline size_t toSizeT(const vint4& v) { 
+  __forceinline size_t toSizeT(const vint4& v) {
 #if defined(__WIN32__) && !defined(__X86_64__) // win32 workaround
     return toScalar(v);
+#elif defined(__ARM_NEON)
+    // FIXME(LTE): Do we need a swap(i.e. use lane 1)?
+    return vgetq_lane_u64(*(reinterpret_cast<const uint64x2_t *>(&v)), 0);
 #else
-    return _mm_cvtsi128_si64(v); 
+    return _mm_cvtsi128_si64(v);
 #endif
   }
-
+#endif
+      
 #if defined(__AVX512VL__)
 
   __forceinline vint4 permute(const vint4 &a, const vint4 &index) {
@@ -459,15 +546,25 @@ namespace embree
 
   template<int i>
   __forceinline vint4 align_shift_right(const vint4& a, const vint4& b) {
-    return _mm_alignr_epi32(a, b, i);    
+    return _mm_alignr_epi32(a, b, i);
-  }  
+  }
 #endif
 
   ////////////////////////////////////////////////////////////////////////////////
   /// Reductions
   ////////////////////////////////////////////////////////////////////////////////
 
-#if defined(__SSE4_1__)
+#if defined(__aarch64__) || defined(__SSE4_1__)
+      
+#if defined(__aarch64__)
+    __forceinline vint4 vreduce_min(const vint4& v) { int h = vminvq_s32(v); return vdupq_n_s32(h); }
+    __forceinline vint4 vreduce_max(const vint4& v) { int h = vmaxvq_s32(v); return vdupq_n_s32(h); }
+    __forceinline vint4 vreduce_add(const vint4& v) { int h = vaddvq_s32(v); return vdupq_n_s32(h); }
+      
+    __forceinline int reduce_min(const vint4& v) { return vminvq_s32(v); }
+    __forceinline int reduce_max(const vint4& v) { return vmaxvq_s32(v); }
+    __forceinline int reduce_add(const vint4& v) { return vaddvq_s32(v); }
+#else
   __forceinline vint4 vreduce_min(const vint4& v) { vint4 h = min(shuffle<1,0,3,2>(v),v); return min(shuffle<2,3,0,1>(h),h); }
   __forceinline vint4 vreduce_max(const vint4& v) { vint4 h = max(shuffle<1,0,3,2>(v),v); return max(shuffle<2,3,0,1>(h),h); }
   __forceinline vint4 vreduce_add(const vint4& v) { vint4 h = shuffle<1,0,3,2>(v)   + v ; return shuffle<2,3,0,1>(h)   + h ; }
@@ -475,7 +572,8 @@ namespace embree
   __forceinline int reduce_min(const vint4& v) { return toScalar(vreduce_min(v)); }
   __forceinline int reduce_max(const vint4& v) { return toScalar(vreduce_max(v)); }
   __forceinline int reduce_add(const vint4& v) { return toScalar(vreduce_add(v)); }
-
+#endif
+      
   __forceinline size_t select_min(const vint4& v) { return bsf(movemask(v == vreduce_min(v))); }
   __forceinline size_t select_max(const vint4& v) { return bsf(movemask(v == vreduce_max(v))); }
 
@@ -494,7 +592,7 @@ namespace embree
   /// Sorting networks
   ////////////////////////////////////////////////////////////////////////////////
 
-#if defined(__SSE4_1__)
+#if (defined(__aarch64__)) || defined(__SSE4_1__)
 
   __forceinline vint4 usort_ascending(const vint4& v)
   {

thirdparty/embree/common/simd/vint8_avx.h

@@ -71,20 +71,25 @@ namespace embree
     static __forceinline void store (void* ptr, const vint8& f) { _mm256_store_ps((float*)ptr,_mm256_castsi256_ps(f)); }
     static __forceinline void storeu(void* ptr, const vint8& f) { _mm256_storeu_ps((float*)ptr,_mm256_castsi256_ps(f)); }
     
+#if !defined(__aarch64__)
     static __forceinline void store (const vboolf8& mask, void* ptr, const vint8& f) { _mm256_maskstore_ps((float*)ptr,(__m256i)mask,_mm256_castsi256_ps(f)); }
     static __forceinline void storeu(const vboolf8& mask, void* ptr, const vint8& f) { _mm256_maskstore_ps((float*)ptr,(__m256i)mask,_mm256_castsi256_ps(f)); }
+#else
+    static __forceinline void store (const vboolf8& mask, void* ptr, const vint8& f) { _mm256_maskstore_ps((float*)ptr,(__m256i)mask.v,_mm256_castsi256_ps(f)); }
+    static __forceinline void storeu(const vboolf8& mask, void* ptr, const vint8& f) { _mm256_maskstore_ps((float*)ptr,(__m256i)mask.v,_mm256_castsi256_ps(f)); }
+#endif
 
     static __forceinline void store_nt(void* ptr, const vint8& v) {
       _mm256_stream_ps((float*)ptr,_mm256_castsi256_ps(v));
     }
 
-    static __forceinline vint8 load(const unsigned char* ptr) {
+    static __forceinline vint8 load(const uint8_t* ptr) {
       vint4 il = vint4::load(ptr+0);
       vint4 ih = vint4::load(ptr+4);
       return vint8(il,ih);
     }
 
-    static __forceinline vint8 loadu(const unsigned char* ptr) {
+    static __forceinline vint8 loadu(const uint8_t* ptr) {
       vint4 il = vint4::loadu(ptr+0);
       vint4 ih = vint4::loadu(ptr+4);
       return vint8(il,ih);
@@ -102,7 +107,7 @@ namespace embree
       return vint8(il,ih);
     }
 
-    static __forceinline void store(unsigned char* ptr, const vint8& i) {
+    static __forceinline void store(uint8_t* ptr, const vint8& i) {
       vint4 il(i.vl);
       vint4 ih(i.vh);
       vint4::store(ptr + 0,il);
@@ -117,54 +122,54 @@ namespace embree
     template<int scale = 4>
     static __forceinline vint8 gather(const int* ptr, const vint8& index) {
       return vint8(
-          *(int*)(((char*)ptr)+scale*index[0]),
+          *(int*)(((int8_t*)ptr)+scale*index[0]),
-          *(int*)(((char*)ptr)+scale*index[1]),
+          *(int*)(((int8_t*)ptr)+scale*index[1]),
-          *(int*)(((char*)ptr)+scale*index[2]),
+          *(int*)(((int8_t*)ptr)+scale*index[2]),
-          *(int*)(((char*)ptr)+scale*index[3]),
+          *(int*)(((int8_t*)ptr)+scale*index[3]),
-          *(int*)(((char*)ptr)+scale*index[4]),
+          *(int*)(((int8_t*)ptr)+scale*index[4]),
-          *(int*)(((char*)ptr)+scale*index[5]),
+          *(int*)(((int8_t*)ptr)+scale*index[5]),
-          *(int*)(((char*)ptr)+scale*index[6]),
+          *(int*)(((int8_t*)ptr)+scale*index[6]),
-          *(int*)(((char*)ptr)+scale*index[7]));
+          *(int*)(((int8_t*)ptr)+scale*index[7]));
     }
 
     template<int scale = 4>
     static __forceinline vint8 gather(const vboolf8& mask, const int* ptr, const vint8& index) {
       vint8 r = zero;
-      if (likely(mask[0])) r[0] = *(int*)(((char*)ptr)+scale*index[0]);
+      if (likely(mask[0])) r[0] = *(int*)(((int8_t*)ptr)+scale*index[0]);
-      if (likely(mask[1])) r[1] = *(int*)(((char*)ptr)+scale*index[1]);
+      if (likely(mask[1])) r[1] = *(int*)(((int8_t*)ptr)+scale*index[1]);
-      if (likely(mask[2])) r[2] = *(int*)(((char*)ptr)+scale*index[2]);
+      if (likely(mask[2])) r[2] = *(int*)(((int8_t*)ptr)+scale*index[2]);
-      if (likely(mask[3])) r[3] = *(int*)(((char*)ptr)+scale*index[3]);
+      if (likely(mask[3])) r[3] = *(int*)(((int8_t*)ptr)+scale*index[3]);
-      if (likely(mask[4])) r[4] = *(int*)(((char*)ptr)+scale*index[4]);
+      if (likely(mask[4])) r[4] = *(int*)(((int8_t*)ptr)+scale*index[4]);
-      if (likely(mask[5])) r[5] = *(int*)(((char*)ptr)+scale*index[5]);
+      if (likely(mask[5])) r[5] = *(int*)(((int8_t*)ptr)+scale*index[5]);
-      if (likely(mask[6])) r[6] = *(int*)(((char*)ptr)+scale*index[6]);
+      if (likely(mask[6])) r[6] = *(int*)(((int8_t*)ptr)+scale*index[6]);
-      if (likely(mask[7])) r[7] = *(int*)(((char*)ptr)+scale*index[7]);
+      if (likely(mask[7])) r[7] = *(int*)(((int8_t*)ptr)+scale*index[7]);
       return r;
     }
 
     template<int scale = 4>
     static __forceinline void scatter(void* ptr, const vint8& ofs, const vint8& v)
     {
-      *(int*)(((char*)ptr)+scale*ofs[0]) = v[0];
+      *(int*)(((int8_t*)ptr)+scale*ofs[0]) = v[0];
-      *(int*)(((char*)ptr)+scale*ofs[1]) = v[1];
+      *(int*)(((int8_t*)ptr)+scale*ofs[1]) = v[1];
-      *(int*)(((char*)ptr)+scale*ofs[2]) = v[2];
+      *(int*)(((int8_t*)ptr)+scale*ofs[2]) = v[2];
-      *(int*)(((char*)ptr)+scale*ofs[3]) = v[3];
+      *(int*)(((int8_t*)ptr)+scale*ofs[3]) = v[3];
-      *(int*)(((char*)ptr)+scale*ofs[4]) = v[4];
+      *(int*)(((int8_t*)ptr)+scale*ofs[4]) = v[4];
-      *(int*)(((char*)ptr)+scale*ofs[5]) = v[5];
+      *(int*)(((int8_t*)ptr)+scale*ofs[5]) = v[5];
-      *(int*)(((char*)ptr)+scale*ofs[6]) = v[6];
+      *(int*)(((int8_t*)ptr)+scale*ofs[6]) = v[6];
-      *(int*)(((char*)ptr)+scale*ofs[7]) = v[7];
+      *(int*)(((int8_t*)ptr)+scale*ofs[7]) = v[7];
     }
 
     template<int scale = 4>
     static __forceinline void scatter(const vboolf8& mask, void* ptr, const vint8& ofs, const vint8& v)
     {
-      if (likely(mask[0])) *(int*)(((char*)ptr)+scale*ofs[0]) = v[0];
+      if (likely(mask[0])) *(int*)(((int8_t*)ptr)+scale*ofs[0]) = v[0];
-      if (likely(mask[1])) *(int*)(((char*)ptr)+scale*ofs[1]) = v[1];
+      if (likely(mask[1])) *(int*)(((int8_t*)ptr)+scale*ofs[1]) = v[1];
-      if (likely(mask[2])) *(int*)(((char*)ptr)+scale*ofs[2]) = v[2];
+      if (likely(mask[2])) *(int*)(((int8_t*)ptr)+scale*ofs[2]) = v[2];
-      if (likely(mask[3])) *(int*)(((char*)ptr)+scale*ofs[3]) = v[3];
+      if (likely(mask[3])) *(int*)(((int8_t*)ptr)+scale*ofs[3]) = v[3];
-      if (likely(mask[4])) *(int*)(((char*)ptr)+scale*ofs[4]) = v[4];
+      if (likely(mask[4])) *(int*)(((int8_t*)ptr)+scale*ofs[4]) = v[4];
-      if (likely(mask[5])) *(int*)(((char*)ptr)+scale*ofs[5]) = v[5];
+      if (likely(mask[5])) *(int*)(((int8_t*)ptr)+scale*ofs[5]) = v[5];
-      if (likely(mask[6])) *(int*)(((char*)ptr)+scale*ofs[6]) = v[6];
+      if (likely(mask[6])) *(int*)(((int8_t*)ptr)+scale*ofs[6]) = v[6];
-      if (likely(mask[7])) *(int*)(((char*)ptr)+scale*ofs[7]) = v[7];
+      if (likely(mask[7])) *(int*)(((int8_t*)ptr)+scale*ofs[7]) = v[7];
     }
 
 

thirdparty/embree/common/simd/vint8_avx2.h

@@ -67,8 +67,8 @@ namespace embree
     /// Loads and Stores
     ////////////////////////////////////////////////////////////////////////////////
 
-    static __forceinline vint8 load(const unsigned char* ptr)  { return _mm256_cvtepu8_epi32(_mm_loadl_epi64((__m128i*)ptr)); }
+    static __forceinline vint8 load(const uint8_t* ptr)  { return _mm256_cvtepu8_epi32(_mm_loadl_epi64((__m128i*)ptr)); }
-    static __forceinline vint8 loadu(const unsigned char* ptr) { return _mm256_cvtepu8_epi32(_mm_loadl_epi64((__m128i*)ptr)); }
+    static __forceinline vint8 loadu(const uint8_t* ptr) { return _mm256_cvtepu8_epi32(_mm_loadl_epi64((__m128i*)ptr)); }
     static __forceinline vint8 load(const unsigned short* ptr)  { return _mm256_cvtepu16_epi32(_mm_load_si128((__m128i*)ptr)); }
     static __forceinline vint8 loadu(const unsigned short* ptr) { return _mm256_cvtepu16_epi32(_mm_loadu_si128((__m128i*)ptr)); }
 
@@ -108,7 +108,7 @@ namespace embree
       _mm256_stream_ps((float*)ptr,_mm256_castsi256_ps(v));
     }
 
-    static __forceinline void store(unsigned char* ptr, const vint8& i)
+    static __forceinline void store(uint8_t* ptr, const vint8& i)
     {
       for (size_t j=0; j<8; j++)
         ptr[j] = i[j];
@@ -140,14 +140,14 @@ namespace embree
 #if defined(__AVX512VL__)
       _mm256_i32scatter_epi32((int*)ptr, ofs, v, scale);
 #else
-      *(int*)(((char*)ptr)+scale*ofs[0]) = v[0];
+      *(int*)(((int8_t*)ptr)+scale*ofs[0]) = v[0];
-      *(int*)(((char*)ptr)+scale*ofs[1]) = v[1];
+      *(int*)(((int8_t*)ptr)+scale*ofs[1]) = v[1];
-      *(int*)(((char*)ptr)+scale*ofs[2]) = v[2];
+      *(int*)(((int8_t*)ptr)+scale*ofs[2]) = v[2];
-      *(int*)(((char*)ptr)+scale*ofs[3]) = v[3];
+      *(int*)(((int8_t*)ptr)+scale*ofs[3]) = v[3];
-      *(int*)(((char*)ptr)+scale*ofs[4]) = v[4];
+      *(int*)(((int8_t*)ptr)+scale*ofs[4]) = v[4];
-      *(int*)(((char*)ptr)+scale*ofs[5]) = v[5];
+      *(int*)(((int8_t*)ptr)+scale*ofs[5]) = v[5];
-      *(int*)(((char*)ptr)+scale*ofs[6]) = v[6];
+      *(int*)(((int8_t*)ptr)+scale*ofs[6]) = v[6];
-      *(int*)(((char*)ptr)+scale*ofs[7]) = v[7];
+      *(int*)(((int8_t*)ptr)+scale*ofs[7]) = v[7];
 #endif
     }
 
@@ -157,14 +157,14 @@ namespace embree
 #if defined(__AVX512VL__)
       _mm256_mask_i32scatter_epi32((int*)ptr, mask, ofs, v, scale);
 #else
-      if (likely(mask[0])) *(int*)(((char*)ptr)+scale*ofs[0]) = v[0];
+      if (likely(mask[0])) *(int*)(((int8_t*)ptr)+scale*ofs[0]) = v[0];
-      if (likely(mask[1])) *(int*)(((char*)ptr)+scale*ofs[1]) = v[1];
+      if (likely(mask[1])) *(int*)(((int8_t*)ptr)+scale*ofs[1]) = v[1];
-      if (likely(mask[2])) *(int*)(((char*)ptr)+scale*ofs[2]) = v[2];
+      if (likely(mask[2])) *(int*)(((int8_t*)ptr)+scale*ofs[2]) = v[2];
-      if (likely(mask[3])) *(int*)(((char*)ptr)+scale*ofs[3]) = v[3];
+      if (likely(mask[3])) *(int*)(((int8_t*)ptr)+scale*ofs[3]) = v[3];
-      if (likely(mask[4])) *(int*)(((char*)ptr)+scale*ofs[4]) = v[4];
+      if (likely(mask[4])) *(int*)(((int8_t*)ptr)+scale*ofs[4]) = v[4];
-      if (likely(mask[5])) *(int*)(((char*)ptr)+scale*ofs[5]) = v[5];
+      if (likely(mask[5])) *(int*)(((int8_t*)ptr)+scale*ofs[5]) = v[5];
-      if (likely(mask[6])) *(int*)(((char*)ptr)+scale*ofs[6]) = v[6];
+      if (likely(mask[6])) *(int*)(((int8_t*)ptr)+scale*ofs[6]) = v[6];
-      if (likely(mask[7])) *(int*)(((char*)ptr)+scale*ofs[7]) = v[7];
+      if (likely(mask[7])) *(int*)(((int8_t*)ptr)+scale*ofs[7]) = v[7];
 #endif
     }
 
@@ -385,7 +385,9 @@ namespace embree
 
   __forceinline int toScalar(const vint8& v) { return _mm_cvtsi128_si32(_mm256_castsi256_si128(v)); }
 
-  __forceinline vint8 permute(const vint8& v, const __m256i& index) {
+#if !defined(__aarch64__)
+
+__forceinline vint8 permute(const vint8& v, const __m256i& index) {
     return _mm256_permutevar8x32_epi32(v, index);
   }
 
@@ -393,6 +395,8 @@ namespace embree
     return _mm256_castps_si256(_mm256_permutevar_ps(_mm256_castsi256_ps(v), index));
   }
 
+
+
   template<int i>
   static __forceinline vint8 align_shift_right(const vint8& a, const vint8& b) {
 #if defined(__AVX512VL__)
@@ -402,6 +406,9 @@ namespace embree
 #endif
   }  
 
+#endif
+
+
   ////////////////////////////////////////////////////////////////////////////////
   /// Reductions
   ////////////////////////////////////////////////////////////////////////////////

thirdparty/embree/common/simd/vllong8_avx512.h

@@ -78,7 +78,7 @@ namespace embree
       return _mm512_load_si512(addr);
     }
 
-    static __forceinline vllong8 load(const unsigned char* ptr) {
+    static __forceinline vllong8 load(const uint8_t* ptr) {
       return _mm512_cvtepu8_epi64(*(__m128i*)ptr); 
     }
 

thirdparty/embree/common/simd/vuint16_avx512.h

@@ -83,7 +83,7 @@ namespace embree
       return _mm512_loadu_si512(addr);
     }
 
-    static __forceinline vuint16 loadu(const unsigned char* ptr) { return _mm512_cvtepu8_epi32(_mm_loadu_si128((__m128i*)ptr)); }
+    static __forceinline vuint16 loadu(const uint8_t* ptr) { return _mm512_cvtepu8_epi32(_mm_loadu_si128((__m128i*)ptr)); }
     static __forceinline vuint16 loadu(const unsigned short* ptr) { return _mm512_cvtepu16_epi32(_mm256_loadu_si256((__m256i*)ptr)); }
 
     static __forceinline vuint16 load(const vuint16* addr) {

thirdparty/embree/common/simd/vuint4_sse2.h

@@ -87,44 +87,64 @@ namespace embree
     static __forceinline void storeu(const vboolf4& mask, void* ptr, const vuint4& i) { storeu(ptr,select(mask,i,loadu(ptr))); }
 #endif
 
-#if defined(__SSE4_1__)
+#if defined(__aarch64__)
-    static __forceinline vuint4 load(const unsigned char* ptr) {
+    static __forceinline vuint4 load(const uint8_t* ptr) {
+        return _mm_load4epu8_epi32(((__m128i*)ptr));
+    }
+    static __forceinline vuint4 loadu(const uint8_t* ptr) {
+        return _mm_load4epu8_epi32(((__m128i*)ptr));
+    }
+#elif defined(__SSE4_1__)
+    static __forceinline vuint4 load(const uint8_t* ptr) {
       return _mm_cvtepu8_epi32(_mm_loadl_epi64((__m128i*)ptr));
     }
 
-    static __forceinline vuint4 loadu(const unsigned char* ptr) {
+    static __forceinline vuint4 loadu(const uint8_t* ptr) {
       return  _mm_cvtepu8_epi32(_mm_loadl_epi64((__m128i*)ptr));
     }
 
 #endif
 
     static __forceinline vuint4 load(const unsigned short* ptr) {
-#if defined (__SSE4_1__)
+#if defined(__aarch64__)
+      return _mm_load4epu16_epi32(((__m128i*)ptr));
+#elif defined (__SSE4_1__)
       return _mm_cvtepu16_epi32(_mm_loadu_si128((__m128i*)ptr));
 #else
       return vuint4(ptr[0],ptr[1],ptr[2],ptr[3]);
 #endif
     } 
 
-    static __forceinline void store_uchar(unsigned char* ptr, const vuint4& v) {
+    static __forceinline void store_uint8(uint8_t* ptr, const vuint4& v) {
-#if defined(__SSE4_1__)
+#if defined(__aarch64__) 
+        uint32x4_t x = uint32x4_t(v.v);
+        uint16x4_t y = vqmovn_u32(x);
+        uint8x8_t z = vqmovn_u16(vcombine_u16(y, y));
+        vst1_lane_u32((uint32_t *)ptr, uint32x2_t(z), 0);
+#elif defined(__SSE4_1__)
       __m128i x = v;
       x = _mm_packus_epi32(x, x);
       x = _mm_packus_epi16(x, x);
       *(unsigned*)ptr = _mm_cvtsi128_si32(x);
 #else
       for (size_t i=0;i<4;i++)
-        ptr[i] = (unsigned char)v[i];
+        ptr[i] = (uint8_t)v[i];
 #endif
     }
 
-    static __forceinline void store_uchar(unsigned short* ptr, const vuint4& v) {
+    static __forceinline void store_uint8(unsigned short* ptr, const vuint4& v) {
+#if defined(__aarch64__)
+        uint32x4_t x = (uint32x4_t)v.v;
+        uint16x4_t y = vqmovn_u32(x);
+        vst1_u16(ptr, y);
+#else
       for (size_t i=0;i<4;i++)
         ptr[i] = (unsigned short)v[i];
+#endif
     }
 
     static __forceinline vuint4 load_nt(void* ptr) {
-#if defined(__SSE4_1__)
+#if (defined(__aarch64__)) || defined(__SSE4_1__)
       return _mm_stream_load_si128((__m128i*)ptr); 
 #else
       return _mm_load_si128((__m128i*)ptr); 
@@ -132,8 +152,8 @@ namespace embree
     }
     
     static __forceinline void store_nt(void* ptr, const vuint4& v) {
-#if defined(__SSE4_1__)
+#if !defined(__aarch64__) && defined(__SSE4_1__)
-      _mm_stream_ps((float*)ptr,_mm_castsi128_ps(v)); 
+      _mm_stream_ps((float*)ptr, _mm_castsi128_ps(v)); 
 #else
       _mm_store_si128((__m128i*)ptr,v);
 #endif
@@ -141,14 +161,14 @@ namespace embree
 
     template<int scale = 4>
     static __forceinline vuint4 gather(const unsigned int* ptr, const vint4& index) {
-#if defined(__AVX2__)
+#if defined(__AVX2__) && !defined(__aarch64__)
       return _mm_i32gather_epi32((const int*)ptr, index, scale);
 #else
       return vuint4(
-          *(unsigned int*)(((char*)ptr)+scale*index[0]),
+          *(unsigned int*)(((int8_t*)ptr)+scale*index[0]),
-          *(unsigned int*)(((char*)ptr)+scale*index[1]),
+          *(unsigned int*)(((int8_t*)ptr)+scale*index[1]),
-          *(unsigned int*)(((char*)ptr)+scale*index[2]),
+          *(unsigned int*)(((int8_t*)ptr)+scale*index[2]),
-          *(unsigned int*)(((char*)ptr)+scale*index[3]));
+          *(unsigned int*)(((int8_t*)ptr)+scale*index[3]));
 #endif
     }
 
@@ -157,13 +177,13 @@ namespace embree
       vuint4 r = zero;
 #if defined(__AVX512VL__)
       return _mm_mmask_i32gather_epi32(r, mask, index, ptr, scale);
-#elif defined(__AVX2__)
+#elif defined(__AVX2__) && !defined(__aarch64__)
       return _mm_mask_i32gather_epi32(r, (const int*)ptr, index, mask, scale);
 #else
-      if (likely(mask[0])) r[0] = *(unsigned int*)(((char*)ptr)+scale*index[0]);
+      if (likely(mask[0])) r[0] = *(unsigned int*)(((int8_t*)ptr)+scale*index[0]);
-      if (likely(mask[1])) r[1] = *(unsigned int*)(((char*)ptr)+scale*index[1]);
+      if (likely(mask[1])) r[1] = *(unsigned int*)(((int8_t*)ptr)+scale*index[1]);
-      if (likely(mask[2])) r[2] = *(unsigned int*)(((char*)ptr)+scale*index[2]);
+      if (likely(mask[2])) r[2] = *(unsigned int*)(((int8_t*)ptr)+scale*index[2]);
-      if (likely(mask[3])) r[3] = *(unsigned int*)(((char*)ptr)+scale*index[3]);
+      if (likely(mask[3])) r[3] = *(unsigned int*)(((int8_t*)ptr)+scale*index[3]);
       return r;
 #endif
     }
@@ -353,16 +373,25 @@ namespace embree
   __forceinline vuint4 unpacklo(const vuint4& a, const vuint4& b) { return _mm_castps_si128(_mm_unpacklo_ps(_mm_castsi128_ps(a), _mm_castsi128_ps(b))); }
   __forceinline vuint4 unpackhi(const vuint4& a, const vuint4& b) { return _mm_castps_si128(_mm_unpackhi_ps(_mm_castsi128_ps(a), _mm_castsi128_ps(b))); }
 
+#if defined(__aarch64__)
+  template<int i0, int i1, int i2, int i3>
+  __forceinline vuint4 shuffle(const vuint4& v) {
+    return vreinterpretq_s32_u8(vqtbl1q_u8( (uint8x16_t)v.v, _MN_SHUFFLE(i0, i1, i2, i3)));
+  }
+  template<int i0, int i1, int i2, int i3>
+  __forceinline vuint4 shuffle(const vuint4& a, const vuint4& b) {
+    return vreinterpretq_s32_u8(vqtbl2q_u8( (uint8x16x2_t){(uint8x16_t)a.v, (uint8x16_t)b.v}, _MF_SHUFFLE(i0, i1, i2, i3)));
+  }
+#else
   template<int i0, int i1, int i2, int i3>
   __forceinline vuint4 shuffle(const vuint4& v) {
     return _mm_shuffle_epi32(v, _MM_SHUFFLE(i3, i2, i1, i0));
   }
-
   template<int i0, int i1, int i2, int i3>
   __forceinline vuint4 shuffle(const vuint4& a, const vuint4& b) {
     return _mm_castps_si128(_mm_shuffle_ps(_mm_castsi128_ps(a), _mm_castsi128_ps(b), _MM_SHUFFLE(i3, i2, i1, i0)));
   }
-
+#endif
 #if defined(__SSE3__)
   template<> __forceinline vuint4 shuffle<0, 0, 2, 2>(const vuint4& v) { return _mm_castps_si128(_mm_moveldup_ps(_mm_castsi128_ps(v))); }
   template<> __forceinline vuint4 shuffle<1, 1, 3, 3>(const vuint4& v) { return _mm_castps_si128(_mm_movehdup_ps(_mm_castsi128_ps(v))); }
@@ -374,7 +403,10 @@ namespace embree
     return shuffle<i,i,i,i>(v);
   }
 
-#if defined(__SSE4_1__)
+#if defined(__aarch64__)
+  template<int src> __forceinline unsigned int extract(const vuint4& b);
+  template<int dst> __forceinline vuint4 insert(const vuint4& a, const unsigned b);
+#elif defined(__SSE4_1__)
   template<int src> __forceinline unsigned int extract(const vuint4& b) { return _mm_extract_epi32(b, src); }
   template<int dst> __forceinline vuint4 insert(const vuint4& a, const unsigned b) { return _mm_insert_epi32(a, b, dst); }
 #else
@@ -382,11 +414,50 @@ namespace embree
   template<int dst> __forceinline vuint4 insert(const vuint4& a, const unsigned b) { vuint4 c = a; c[dst&3] = b; return c; }
 #endif
 
-
+#if defined(__aarch64__)
+  template<> __forceinline unsigned int extract<0>(const vuint4& b) {
+    return b[0];
+  }
+  template<> __forceinline unsigned int extract<1>(const vuint4& b) {
+    return b[1];
+  }
+  template<> __forceinline unsigned int extract<2>(const vuint4& b) {
+    return b[2];
+  }
+  template<> __forceinline unsigned int extract<3>(const vuint4& b) {
+    return b[3];
+  }
+                                                                               
+  template<> __forceinline vuint4 insert<0>(const vuint4& a, unsigned b){
+    vuint4 c = a;
+    c[0] = b;
+    return c;
+  }
+  template<> __forceinline vuint4 insert<1>(const vuint4& a, unsigned b){
+    vuint4 c = a;
+    c[1] = b;
+    return c;
+  }
+  template<> __forceinline vuint4 insert<2>(const vuint4& a, unsigned b){
+    vuint4 c = a;
+    c[2] = b;
+    return c;
+  }
+  template<> __forceinline vuint4 insert<3>(const vuint4& a, unsigned b){
+    vuint4 c = a;
+    c[3] = b;
+    return c;
+  }
+                                                                               
+  __forceinline unsigned int toScalar(const vuint4& v) {
+    return v[0];
+  }
+#else
   template<> __forceinline unsigned int extract<0>(const vuint4& b) { return _mm_cvtsi128_si32(b); }
 
   __forceinline unsigned int toScalar(const vuint4& v) { return _mm_cvtsi128_si32(v); }
-
+#endif
+                                                                               
   ////////////////////////////////////////////////////////////////////////////////
   /// Reductions
   ////////////////////////////////////////////////////////////////////////////////

thirdparty/embree/common/simd/vuint8_avx.h

@@ -69,20 +69,24 @@ namespace embree
     static __forceinline void store (void* ptr, const vuint8& f) { _mm256_store_ps((float*)ptr,_mm256_castsi256_ps(f)); }
     static __forceinline void storeu(void* ptr, const vuint8& f) { _mm256_storeu_ps((float*)ptr,_mm256_castsi256_ps(f)); }
     
+#if !defined(__aarch64__)
     static __forceinline void store (const vboolf8& mask, void* ptr, const vuint8& f) { _mm256_maskstore_ps((float*)ptr,(__m256i)mask,_mm256_castsi256_ps(f)); }
     static __forceinline void storeu(const vboolf8& mask, void* ptr, const vuint8& f) { _mm256_maskstore_ps((float*)ptr,(__m256i)mask,_mm256_castsi256_ps(f)); }
-
+#else
+    static __forceinline void store (const vboolf8& mask, void* ptr, const vuint8& f) { _mm256_maskstore_ps((float*)ptr,(__m256i)mask.v,_mm256_castsi256_ps(f)); }
+    static __forceinline void storeu(const vboolf8& mask, void* ptr, const vuint8& f) { _mm256_maskstore_ps((float*)ptr,(__m256i)mask.v,_mm256_castsi256_ps(f)); }
+#endif
     static __forceinline void store_nt(void* ptr, const vuint8& v) {
       _mm256_stream_ps((float*)ptr,_mm256_castsi256_ps(v));
     }
 
-    static __forceinline vuint8 load(const unsigned char* ptr) {
+    static __forceinline vuint8 load(const uint8_t* ptr) {
       vuint4 il = vuint4::load(ptr+0);
       vuint4 ih = vuint4::load(ptr+4);
       return vuint8(il,ih);
     }
 
-    static __forceinline vuint8 loadu(const unsigned char* ptr) {
+    static __forceinline vuint8 loadu(const uint8_t* ptr) {
       vuint4 il = vuint4::loadu(ptr+0);
       vuint4 ih = vuint4::loadu(ptr+4);
       return vuint8(il,ih);
@@ -100,7 +104,7 @@ namespace embree
       return vuint8(il,ih);
     }
 
-    static __forceinline void store(unsigned char* ptr, const vuint8& i) {
+    static __forceinline void store(uint8_t* ptr, const vuint8& i) {
       vuint4 il(i.vl);
       vuint4 ih(i.vh);
       vuint4::store(ptr + 0,il);
@@ -115,54 +119,54 @@ namespace embree
     template<int scale = 4>
     static __forceinline vuint8 gather(const unsigned int* ptr, const vint8& index) {
       return vuint8(
-          *(unsigned int*)(((char*)ptr)+scale*index[0]),
+          *(unsigned int*)(((int8_t*)ptr)+scale*index[0]),
-          *(unsigned int*)(((char*)ptr)+scale*index[1]),
+          *(unsigned int*)(((int8_t*)ptr)+scale*index[1]),
-          *(unsigned int*)(((char*)ptr)+scale*index[2]),
+          *(unsigned int*)(((int8_t*)ptr)+scale*index[2]),
-          *(unsigned int*)(((char*)ptr)+scale*index[3]),
+          *(unsigned int*)(((int8_t*)ptr)+scale*index[3]),
-          *(unsigned int*)(((char*)ptr)+scale*index[4]),
+          *(unsigned int*)(((int8_t*)ptr)+scale*index[4]),
-          *(unsigned int*)(((char*)ptr)+scale*index[5]),
+          *(unsigned int*)(((int8_t*)ptr)+scale*index[5]),
-          *(unsigned int*)(((char*)ptr)+scale*index[6]),
+          *(unsigned int*)(((int8_t*)ptr)+scale*index[6]),
-          *(unsigned int*)(((char*)ptr)+scale*index[7]));
+          *(unsigned int*)(((int8_t*)ptr)+scale*index[7]));
     }
 
     template<int scale = 4>
     static __forceinline vuint8 gather(const vboolf8& mask, const unsigned int* ptr, const vint8& index) {
       vuint8 r = zero;
-      if (likely(mask[0])) r[0] = *(unsigned int*)(((char*)ptr)+scale*index[0]);
+      if (likely(mask[0])) r[0] = *(unsigned int*)(((int8_t*)ptr)+scale*index[0]);
-      if (likely(mask[1])) r[1] = *(unsigned int*)(((char*)ptr)+scale*index[1]);
+      if (likely(mask[1])) r[1] = *(unsigned int*)(((int8_t*)ptr)+scale*index[1]);
-      if (likely(mask[2])) r[2] = *(unsigned int*)(((char*)ptr)+scale*index[2]);
+      if (likely(mask[2])) r[2] = *(unsigned int*)(((int8_t*)ptr)+scale*index[2]);
-      if (likely(mask[3])) r[3] = *(unsigned int*)(((char*)ptr)+scale*index[3]);
+      if (likely(mask[3])) r[3] = *(unsigned int*)(((int8_t*)ptr)+scale*index[3]);
-      if (likely(mask[4])) r[4] = *(unsigned int*)(((char*)ptr)+scale*index[4]);
+      if (likely(mask[4])) r[4] = *(unsigned int*)(((int8_t*)ptr)+scale*index[4]);
-      if (likely(mask[5])) r[5] = *(unsigned int*)(((char*)ptr)+scale*index[5]);
+      if (likely(mask[5])) r[5] = *(unsigned int*)(((int8_t*)ptr)+scale*index[5]);
-      if (likely(mask[6])) r[6] = *(unsigned int*)(((char*)ptr)+scale*index[6]);
+      if (likely(mask[6])) r[6] = *(unsigned int*)(((int8_t*)ptr)+scale*index[6]);
-      if (likely(mask[7])) r[7] = *(unsigned int*)(((char*)ptr)+scale*index[7]);
+      if (likely(mask[7])) r[7] = *(unsigned int*)(((int8_t*)ptr)+scale*index[7]);
       return r;
     }
 
     template<int scale = 4>
     static __forceinline void scatter(void* ptr, const vint8& ofs, const vuint8& v)
     {
-      *(unsigned int*)(((char*)ptr)+scale*ofs[0]) = v[0];
+      *(unsigned int*)(((int8_t*)ptr)+scale*ofs[0]) = v[0];
-      *(unsigned int*)(((char*)ptr)+scale*ofs[1]) = v[1];
+      *(unsigned int*)(((int8_t*)ptr)+scale*ofs[1]) = v[1];
-      *(unsigned int*)(((char*)ptr)+scale*ofs[2]) = v[2];
+      *(unsigned int*)(((int8_t*)ptr)+scale*ofs[2]) = v[2];
-      *(unsigned int*)(((char*)ptr)+scale*ofs[3]) = v[3];
+      *(unsigned int*)(((int8_t*)ptr)+scale*ofs[3]) = v[3];
-      *(unsigned int*)(((char*)ptr)+scale*ofs[4]) = v[4];
+      *(unsigned int*)(((int8_t*)ptr)+scale*ofs[4]) = v[4];
-      *(unsigned int*)(((char*)ptr)+scale*ofs[5]) = v[5];
+      *(unsigned int*)(((int8_t*)ptr)+scale*ofs[5]) = v[5];
-      *(unsigned int*)(((char*)ptr)+scale*ofs[6]) = v[6];
+      *(unsigned int*)(((int8_t*)ptr)+scale*ofs[6]) = v[6];
-      *(unsigned int*)(((char*)ptr)+scale*ofs[7]) = v[7];
+      *(unsigned int*)(((int8_t*)ptr)+scale*ofs[7]) = v[7];
     }
 
     template<int scale = 4>
     static __forceinline void scatter(const vboolf8& mask, void* ptr, const vint8& ofs, const vuint8& v)
     {
-      if (likely(mask[0])) *(unsigned int*)(((char*)ptr)+scale*ofs[0]) = v[0];
+      if (likely(mask[0])) *(unsigned int*)(((int8_t*)ptr)+scale*ofs[0]) = v[0];
-      if (likely(mask[1])) *(unsigned int*)(((char*)ptr)+scale*ofs[1]) = v[1];
+      if (likely(mask[1])) *(unsigned int*)(((int8_t*)ptr)+scale*ofs[1]) = v[1];
-      if (likely(mask[2])) *(unsigned int*)(((char*)ptr)+scale*ofs[2]) = v[2];
+      if (likely(mask[2])) *(unsigned int*)(((int8_t*)ptr)+scale*ofs[2]) = v[2];
-      if (likely(mask[3])) *(unsigned int*)(((char*)ptr)+scale*ofs[3]) = v[3];
+      if (likely(mask[3])) *(unsigned int*)(((int8_t*)ptr)+scale*ofs[3]) = v[3];
-      if (likely(mask[4])) *(unsigned int*)(((char*)ptr)+scale*ofs[4]) = v[4];
+      if (likely(mask[4])) *(unsigned int*)(((int8_t*)ptr)+scale*ofs[4]) = v[4];
-      if (likely(mask[5])) *(unsigned int*)(((char*)ptr)+scale*ofs[5]) = v[5];
+      if (likely(mask[5])) *(unsigned int*)(((int8_t*)ptr)+scale*ofs[5]) = v[5];
-      if (likely(mask[6])) *(unsigned int*)(((char*)ptr)+scale*ofs[6]) = v[6];
+      if (likely(mask[6])) *(unsigned int*)(((int8_t*)ptr)+scale*ofs[6]) = v[6];
-      if (likely(mask[7])) *(unsigned int*)(((char*)ptr)+scale*ofs[7]) = v[7];
+      if (likely(mask[7])) *(unsigned int*)(((int8_t*)ptr)+scale*ofs[7]) = v[7];
     }
 
 

thirdparty/embree/common/simd/vuint8_avx2.h

@@ -66,8 +66,8 @@ namespace embree
     /// Loads and Stores
     ////////////////////////////////////////////////////////////////////////////////
 
-    static __forceinline vuint8 load(const unsigned char* ptr)  { return _mm256_cvtepu8_epi32(_mm_loadl_epi64((__m128i*)ptr)); }
+    static __forceinline vuint8 load(const uint8_t* ptr)  { return _mm256_cvtepu8_epi32(_mm_loadl_epi64((__m128i*)ptr)); }
-    static __forceinline vuint8 loadu(const unsigned char* ptr) { return _mm256_cvtepu8_epi32(_mm_loadl_epi64((__m128i*)ptr)); }
+    static __forceinline vuint8 loadu(const uint8_t* ptr) { return _mm256_cvtepu8_epi32(_mm_loadl_epi64((__m128i*)ptr)); }
     static __forceinline vuint8 load(const unsigned short* ptr)  { return _mm256_cvtepu16_epi32(_mm_load_si128((__m128i*)ptr)); }
     static __forceinline vuint8 loadu(const unsigned short* ptr) { return _mm256_cvtepu16_epi32(_mm_loadu_si128((__m128i*)ptr)); }
 
@@ -107,7 +107,7 @@ namespace embree
       _mm256_stream_ps((float*)ptr,_mm256_castsi256_ps(v));
     }
 
-    static __forceinline void store(unsigned char* ptr, const vuint8& i)
+    static __forceinline void store(uint8_t* ptr, const vuint8& i)
     {
       for (size_t j=0; j<8; j++)
         ptr[j] = i[j];
@@ -139,14 +139,14 @@ namespace embree
 #if defined(__AVX512VL__)
       _mm256_i32scatter_epi32((int*)ptr, ofs, v, scale);
 #else
-      *(unsigned int*)(((char*)ptr)+scale*ofs[0]) = v[0];
+      *(unsigned int*)(((int8_t*)ptr) + scale * ofs[0]) = v[0];
-      *(unsigned int*)(((char*)ptr)+scale*ofs[1]) = v[1];
+      *(unsigned int*)(((int8_t*)ptr) + scale * ofs[1]) = v[1];
-      *(unsigned int*)(((char*)ptr)+scale*ofs[2]) = v[2];
+      *(unsigned int*)(((int8_t*)ptr) + scale * ofs[2]) = v[2];
-      *(unsigned int*)(((char*)ptr)+scale*ofs[3]) = v[3];
+      *(unsigned int*)(((int8_t*)ptr) + scale * ofs[3]) = v[3];
-      *(unsigned int*)(((char*)ptr)+scale*ofs[4]) = v[4];
+      *(unsigned int*)(((int8_t*)ptr) + scale * ofs[4]) = v[4];
-      *(unsigned int*)(((char*)ptr)+scale*ofs[5]) = v[5];
+      *(unsigned int*)(((int8_t*)ptr) + scale * ofs[5]) = v[5];
-      *(unsigned int*)(((char*)ptr)+scale*ofs[6]) = v[6];
+      *(unsigned int*)(((int8_t*)ptr) + scale * ofs[6]) = v[6];
-      *(unsigned int*)(((char*)ptr)+scale*ofs[7]) = v[7];
+      *(unsigned int*)(((int8_t*)ptr) + scale * ofs[7]) = v[7];
 #endif
     }
 
@@ -156,14 +156,14 @@ namespace embree
 #if defined(__AVX512VL__)
       _mm256_mask_i32scatter_epi32((int*)ptr, mask, ofs, v, scale);
 #else
-      if (likely(mask[0])) *(unsigned int*)(((char*)ptr)+scale*ofs[0]) = v[0];
+      if (likely(mask[0])) *(unsigned int*)(((int8_t*)ptr)+scale*ofs[0]) = v[0];
-      if (likely(mask[1])) *(unsigned int*)(((char*)ptr)+scale*ofs[1]) = v[1];
+      if (likely(mask[1])) *(unsigned int*)(((int8_t*)ptr)+scale*ofs[1]) = v[1];
-      if (likely(mask[2])) *(unsigned int*)(((char*)ptr)+scale*ofs[2]) = v[2];
+      if (likely(mask[2])) *(unsigned int*)(((int8_t*)ptr)+scale*ofs[2]) = v[2];
-      if (likely(mask[3])) *(unsigned int*)(((char*)ptr)+scale*ofs[3]) = v[3];
+      if (likely(mask[3])) *(unsigned int*)(((int8_t*)ptr)+scale*ofs[3]) = v[3];
-      if (likely(mask[4])) *(unsigned int*)(((char*)ptr)+scale*ofs[4]) = v[4];
+      if (likely(mask[4])) *(unsigned int*)(((int8_t*)ptr)+scale*ofs[4]) = v[4];
-      if (likely(mask[5])) *(unsigned int*)(((char*)ptr)+scale*ofs[5]) = v[5];
+      if (likely(mask[5])) *(unsigned int*)(((int8_t*)ptr)+scale*ofs[5]) = v[5];
-      if (likely(mask[6])) *(unsigned int*)(((char*)ptr)+scale*ofs[6]) = v[6];
+      if (likely(mask[6])) *(unsigned int*)(((int8_t*)ptr)+scale*ofs[6]) = v[6];
-      if (likely(mask[7])) *(unsigned int*)(((char*)ptr)+scale*ofs[7]) = v[7];
+      if (likely(mask[7])) *(unsigned int*)(((int8_t*)ptr)+scale*ofs[7]) = v[7];
 #endif
     }
 
@@ -379,6 +379,8 @@ namespace embree
 
   __forceinline int toScalar(const vuint8& v) { return _mm_cvtsi128_si32(_mm256_castsi256_si128(v)); }
 
+#if !defined(__aarch64__)
+
   __forceinline vuint8 permute(const vuint8& v, const __m256i& index) {
     return _mm256_permutevar8x32_epi32(v, index);
   }
@@ -394,7 +396,10 @@ namespace embree
 #else
     return _mm256_alignr_epi8(a, b, 4*i);
 #endif
-  }  
+  }
+
+#endif
+
 
   ////////////////////////////////////////////////////////////////////////////////
   /// Reductions

thirdparty/embree/common/sys/alloc.cpp

@@ -21,7 +21,10 @@ namespace embree
     void* ptr = _mm_malloc(size,align);
 
     if (size != 0 && ptr == nullptr)
-      throw std::bad_alloc();
+      // -- GODOT start --
+      // throw std::bad_alloc();
+      abort(); 
+      // -- GODOT end --
     
     return ptr;
   }
@@ -128,7 +131,10 @@ namespace embree
     /* fall back to 4k pages */
     int flags = MEM_COMMIT | MEM_RESERVE;
     char* ptr = (char*) VirtualAlloc(nullptr,bytes,flags,PAGE_READWRITE);
-    if (ptr == nullptr) throw std::bad_alloc();
+    // -- GODOT start --
+    // if (ptr == nullptr) throw std::bad_alloc();
+    if (ptr == nullptr) abort();
+    // -- GODOT end --
     hugepages = false;
     return ptr;
   }
@@ -145,7 +151,10 @@ namespace embree
       return bytesOld;
 
     if (!VirtualFree((char*)ptr+bytesNew,bytesOld-bytesNew,MEM_DECOMMIT))
-      throw std::bad_alloc();
+      // -- GODOT start --
+      // throw std::bad_alloc();
+      abort();
+      // -- GODOT end --
 
     return bytesNew;
   }
@@ -156,7 +165,10 @@ namespace embree
       return;
 
     if (!VirtualFree(ptr,0,MEM_RELEASE))
-      throw std::bad_alloc();
+      // -- GODOT start --
+      // throw std::bad_alloc();
+      abort();
+      // -- GODOT end --
   }
 
   void os_advise(void *ptr, size_t bytes)
@@ -260,7 +272,10 @@ namespace embree
 
     /* fallback to 4k pages */
     void* ptr = (char*) mmap(0, bytes, PROT_READ | PROT_WRITE, MAP_PRIVATE | MAP_ANON, -1, 0);
-    if (ptr == MAP_FAILED) throw std::bad_alloc();
+    // -- GODOT start --
+    // if (ptr == MAP_FAILED) throw std::bad_alloc();
+    if (ptr == MAP_FAILED) abort();
+    // -- GODOT end --
     hugepages = false;
 
     /* advise huge page hint for THP */
@@ -277,7 +292,10 @@ namespace embree
       return bytesOld;
 
     if (munmap((char*)ptr+bytesNew,bytesOld-bytesNew) == -1)
-      throw std::bad_alloc();
+      // -- GODOT start --
+      // throw std::bad_alloc();
+      abort();
+      // -- GODOT end --
 
     return bytesNew;
   }
@@ -291,7 +309,10 @@ namespace embree
     const size_t pageSize = hugepages ? PAGE_SIZE_2M : PAGE_SIZE_4K;
     bytes = (bytes+pageSize-1) & ~(pageSize-1);
     if (munmap(ptr,bytes) == -1)
-      throw std::bad_alloc();
+      // -- GODOT start --
+      // throw std::bad_alloc();
+      abort();
+      // -- GODOT end --
   }
 
   /* hint for transparent huge pages (THP) */

thirdparty/embree/common/sys/array.h

@@ -139,7 +139,7 @@ namespace embree
     __forceinline       Ty& operator[](const unsigned i)       { assert(i<N); return data[i]; }
     __forceinline const Ty& operator[](const unsigned i) const { assert(i<N); return data[i]; }
 
-#if defined(__X86_64__)
+#if defined(__X86_64__) || defined(__aarch64__)
     __forceinline       Ty& operator[](const size_t i)       { assert(i<N); return data[i]; }
     __forceinline const Ty& operator[](const size_t i) const { assert(i<N); return data[i]; }
 #endif
@@ -196,7 +196,7 @@ namespace embree
     __forceinline       Ty& operator[](const int i)      { assert(i>=0 && i<max_total_elements); resize(i+1); return data[i]; }
     __forceinline       Ty& operator[](const unsigned i) { assert(i<max_total_elements); resize(i+1); return data[i]; }
 
-#if defined(__X86_64__)
+#if defined(__X86_64__) || defined(__aarch64__)
     __forceinline       Ty& operator[](const size_t i) { assert(i<max_total_elements); resize(i+1); return data[i]; }
 #endif
 

thirdparty/embree/common/sys/intrinsics.h

@@ -9,7 +9,14 @@
 #include <intrin.h>
 #endif
 
+#if defined(__ARM_NEON)
+#include "../math/SSE2NEON.h"
+#if defined(NEON_AVX2_EMULATION)
+#include "../math/AVX2NEON.h"
+#endif
+#else
 #include <immintrin.h>
+#endif
 
 #if defined(__BMI__) && defined(__GNUC__) && !defined(__INTEL_COMPILER)
   #if !defined(_tzcnt_u32)
@@ -20,6 +27,14 @@
   #endif
 #endif
 
+#if defined(__aarch64__)
+#if !defined(_lzcnt_u32)
+  #define _lzcnt_u32 __builtin_clz
+#endif
+#if !defined(_lzcnt_u32)
+  #define _lzcnt_u32 __builtin_clzll
+#endif
+#else
 #if defined(__LZCNT__)
   #if !defined(_lzcnt_u32)
     #define _lzcnt_u32 __lzcnt32
@@ -28,16 +43,13 @@
     #define _lzcnt_u64 __lzcnt64
   #endif
 #endif
+#endif
 
 #if defined(__WIN32__)
-// -- GODOT start --
+#  ifndef NOMINMAX
-#if !defined(NOMINMAX)
+#  define NOMINMAX
-// -- GODOT end --
+#  endif
-#define NOMINMAX
+#  include <windows.h>
-// -- GODOT start --
-#endif
-#include "windows.h"
-// -- GODOT end --
 #endif
 
 /* normally defined in pmmintrin.h, but we always need this */
@@ -50,133 +62,133 @@
 
 namespace embree
 {
-  
+
 ////////////////////////////////////////////////////////////////////////////////
 /// Windows Platform
 ////////////////////////////////////////////////////////////////////////////////
-  
+
 #if defined(__WIN32__)
-  
+
-  __forceinline size_t read_tsc()  
+  __forceinline size_t read_tsc()
   {
     LARGE_INTEGER li;
     QueryPerformanceCounter(&li);
     return (size_t)li.QuadPart;
   }
-  
+
   __forceinline int bsf(int v) {
-#if defined(__AVX2__) 
+#if defined(__AVX2__) && !defined(__aarch64__)
     return _tzcnt_u32(v);
 #else
     unsigned long r = 0; _BitScanForward(&r,v); return r;
 #endif
   }
-  
+
   __forceinline unsigned bsf(unsigned v) {
-#if defined(__AVX2__) 
+#if defined(__AVX2__) && !defined(__aarch64__)
     return _tzcnt_u32(v);
 #else
     unsigned long r = 0; _BitScanForward(&r,v); return r;
 #endif
   }
-  
+
 #if defined(__X86_64__)
   __forceinline size_t bsf(size_t v) {
-#if defined(__AVX2__) 
+#if defined(__AVX2__)
     return _tzcnt_u64(v);
 #else
     unsigned long r = 0; _BitScanForward64(&r,v); return r;
 #endif
   }
 #endif
-  
+
-  __forceinline int bscf(int& v) 
+  __forceinline int bscf(int& v)
   {
     int i = bsf(v);
     v &= v-1;
     return i;
   }
-  
+
-  __forceinline unsigned bscf(unsigned& v) 
+  __forceinline unsigned bscf(unsigned& v)
   {
     unsigned i = bsf(v);
     v &= v-1;
     return i;
   }
-  
+
 #if defined(__X86_64__)
-  __forceinline size_t bscf(size_t& v) 
+  __forceinline size_t bscf(size_t& v)
   {
     size_t i = bsf(v);
     v &= v-1;
     return i;
   }
 #endif
-  
+
   __forceinline int bsr(int v) {
-#if defined(__AVX2__) 
+#if defined(__AVX2__)  && !defined(__aarch64__)
     return 31 - _lzcnt_u32(v);
 #else
     unsigned long r = 0; _BitScanReverse(&r,v); return r;
 #endif
   }
-  
+
   __forceinline unsigned bsr(unsigned v) {
-#if defined(__AVX2__) 
+#if defined(__AVX2__) && !defined(__aarch64__)
     return 31 - _lzcnt_u32(v);
 #else
     unsigned long r = 0; _BitScanReverse(&r,v); return r;
 #endif
   }
-  
+
 #if defined(__X86_64__)
   __forceinline size_t bsr(size_t v) {
-#if defined(__AVX2__) 
+#if defined(__AVX2__)
     return 63 -_lzcnt_u64(v);
 #else
     unsigned long r = 0; _BitScanReverse64(&r, v); return r;
 #endif
   }
 #endif
-  
+
   __forceinline int lzcnt(const int x)
   {
-#if defined(__AVX2__)
+#if defined(__AVX2__) && !defined(__aarch64__)
     return _lzcnt_u32(x);
 #else
     if (unlikely(x == 0)) return 32;
-    return 31 - bsr(x);    
+    return 31 - bsr(x);
 #endif
   }
-  
+
   __forceinline int btc(int v, int i) {
     long r = v; _bittestandcomplement(&r,i); return r;
   }
-  
+
   __forceinline int bts(int v, int i) {
     long r = v; _bittestandset(&r,i); return r;
   }
-  
+
   __forceinline int btr(int v, int i) {
     long r = v; _bittestandreset(&r,i); return r;
   }
-  
+
 #if defined(__X86_64__)
-  
+
   __forceinline size_t btc(size_t v, size_t i) {
     size_t r = v; _bittestandcomplement64((__int64*)&r,i); return r;
   }
-  
+
   __forceinline size_t bts(size_t v, size_t i) {
     __int64 r = v; _bittestandset64(&r,i); return r;
   }
-  
+
   __forceinline size_t btr(size_t v, size_t i) {
     __int64 r = v; _bittestandreset64(&r,i); return r;
   }
-  
+
 #endif
-  
+
   __forceinline int32_t atomic_cmpxchg(volatile int32_t* p, const int32_t c, const int32_t v) {
     return _InterlockedCompareExchange((volatile long*)p,v,c);
   }
@@ -184,143 +196,174 @@ namespace embree
 ////////////////////////////////////////////////////////////////////////////////
 /// Unix Platform
 ////////////////////////////////////////////////////////////////////////////////
-  
+
 #else
-  
+
 #if defined(__i386__) && defined(__PIC__)
-  
+
-  __forceinline void __cpuid(int out[4], int op) 
+  __forceinline void __cpuid(int out[4], int op)
   {
     asm volatile ("xchg{l}\t{%%}ebx, %1\n\t"
                   "cpuid\n\t"
                   "xchg{l}\t{%%}ebx, %1\n\t"
-                  : "=a"(out[0]), "=r"(out[1]), "=c"(out[2]), "=d"(out[3]) 
+                  : "=a"(out[0]), "=r"(out[1]), "=c"(out[2]), "=d"(out[3])
-                  : "0"(op)); 
+                  : "0"(op));
   }
-  
+
-  __forceinline void __cpuid_count(int out[4], int op1, int op2) 
+  __forceinline void __cpuid_count(int out[4], int op1, int op2)
   {
     asm volatile ("xchg{l}\t{%%}ebx, %1\n\t"
                   "cpuid\n\t"
                   "xchg{l}\t{%%}ebx, %1\n\t"
                   : "=a" (out[0]), "=r" (out[1]), "=c" (out[2]), "=d" (out[3])
-                  : "0" (op1), "2" (op2)); 
+                  : "0" (op1), "2" (op2));
   }
-  
+
 #else
-  
+
   __forceinline void __cpuid(int out[4], int op) {
-    asm volatile ("cpuid" : "=a"(out[0]), "=b"(out[1]), "=c"(out[2]), "=d"(out[3]) : "a"(op)); 
+#if defined(__ARM_NEON)
-  }
+    if (op == 0) { // Get CPU name
-  
+      out[0] = 0x41524d20;
-  __forceinline void __cpuid_count(int out[4], int op1, int op2) {
+      out[1] = 0x41524d20;
-    asm volatile ("cpuid" : "=a"(out[0]), "=b"(out[1]), "=c"(out[2]), "=d"(out[3]) : "a"(op1), "c"(op2)); 
+      out[2] = 0x41524d20;
-  }
+      out[3] = 0x41524d20;
-  
+    }
+#else
+    asm volatile ("cpuid" : "=a"(out[0]), "=b"(out[1]), "=c"(out[2]), "=d"(out[3]) : "a"(op));
 #endif
-  
+  }
+
+#if !defined(__ARM_NEON)
+  __forceinline void __cpuid_count(int out[4], int op1, int op2) {
+    asm volatile ("cpuid" : "=a"(out[0]), "=b"(out[1]), "=c"(out[2]), "=d"(out[3]) : "a"(op1), "c"(op2));
+  }
+#endif
+
+#endif
+
   __forceinline uint64_t read_tsc()  {
+#if defined(__ARM_NEON)
+    return 0; // FIXME(LTE): mimic rdtsc
+#else
     uint32_t high,low;
     asm volatile ("rdtsc" : "=d"(high), "=a"(low));
     return (((uint64_t)high) << 32) + (uint64_t)low;
+#endif
   }
-  
+
   __forceinline int bsf(int v) {
-#if defined(__AVX2__) 
+#if defined(__ARM_NEON)
+    return __builtin_ctz(v);
+#else
+#if defined(__AVX2__)
     return _tzcnt_u32(v);
 #else
     int r = 0; asm ("bsf %1,%0" : "=r"(r) : "r"(v)); return r;
+#endif
 #endif
   }
-  
+
-#if defined(__X86_64__)
+#if defined(__X86_64__) || defined(__aarch64__)
-  __forceinline unsigned bsf(unsigned v) 
+  __forceinline unsigned bsf(unsigned v)
   {
-#if defined(__AVX2__) 
+#if defined(__ARM_NEON)
+    return __builtin_ctz(v);
+#else
+#if defined(__AVX2__)
     return _tzcnt_u32(v);
 #else
     unsigned r = 0; asm ("bsf %1,%0" : "=r"(r) : "r"(v)); return r;
+#endif
 #endif
   }
 #endif
-  
+
   __forceinline size_t bsf(size_t v) {
-#if defined(__AVX2__)
+#if defined(__AVX2__) && !defined(__aarch64__)
 #if defined(__X86_64__)
     return _tzcnt_u64(v);
 #else
     return _tzcnt_u32(v);
 #endif
+#elif defined(__ARM_NEON)
+    return __builtin_ctzl(v);
 #else
     size_t r = 0; asm ("bsf %1,%0" : "=r"(r) : "r"(v)); return r;
 #endif
   }
 
-  __forceinline int bscf(int& v) 
+  __forceinline int bscf(int& v)
   {
     int i = bsf(v);
     v &= v-1;
     return i;
   }
-  
+
-#if defined(__X86_64__)
+#if defined(__X86_64__) || defined(__aarch64__)
-  __forceinline unsigned int bscf(unsigned int& v) 
+  __forceinline unsigned int bscf(unsigned int& v)
   {
     unsigned int i = bsf(v);
     v &= v-1;
     return i;
   }
 #endif
-  
+
-  __forceinline size_t bscf(size_t& v) 
+  __forceinline size_t bscf(size_t& v)
   {
     size_t i = bsf(v);
     v &= v-1;
     return i;
   }
-  
+
   __forceinline int bsr(int v) {
-#if defined(__AVX2__) 
+#if defined(__AVX2__) && !defined(__aarch64__)
     return 31 - _lzcnt_u32(v);
+#elif defined(__ARM_NEON)
+    return __builtin_clz(v)^31;
 #else
     int r = 0; asm ("bsr %1,%0" : "=r"(r) : "r"(v)); return r;
 #endif
   }
-  
+
-#if defined(__X86_64__)
+#if defined(__X86_64__) || defined(__aarch64__)
   __forceinline unsigned bsr(unsigned v) {
-#if defined(__AVX2__) 
+#if defined(__AVX2__)
     return 31 - _lzcnt_u32(v);
+#elif defined(__ARM_NEON)
+    return __builtin_clz(v)^31;
 #else
     unsigned r = 0; asm ("bsr %1,%0" : "=r"(r) : "r"(v)); return r;
 #endif
   }
 #endif
-  
+
   __forceinline size_t bsr(size_t v) {
-#if defined(__AVX2__)
+#if defined(__AVX2__) && !defined(__aarch64__)
 #if defined(__X86_64__)
     return 63 - _lzcnt_u64(v);
 #else
     return 31 - _lzcnt_u32(v);
 #endif
+#elif defined(__aarch64__)
+    return (sizeof(v) * 8 - 1) - __builtin_clzl(v);
 #else
     size_t r = 0; asm ("bsr %1,%0" : "=r"(r) : "r"(v)); return r;
 #endif
   }
-  
+
   __forceinline int lzcnt(const int x)
   {
-#if defined(__AVX2__)
+#if defined(__AVX2__) && !defined(__aarch64__)
     return _lzcnt_u32(x);
 #else
     if (unlikely(x == 0)) return 32;
-    return 31 - bsr(x);    
+    return 31 - bsr(x);
 #endif
   }
 
   __forceinline size_t blsr(size_t v) {
-#if defined(__AVX2__) 
+#if defined(__AVX2__) && !defined(__aarch64__)
 #if defined(__INTEL_COMPILER)
     return _blsr_u64(v);
 #else
@@ -334,41 +377,79 @@ namespace embree
     return v & (v-1);
 #endif
   }
-  
+
   __forceinline int btc(int v, int i) {
+#if defined(__aarch64__)
+    // _bittestandcomplement(long *a, long b) {
+    // unsigned char x = (*a >> b) & 1;
+    // *a = *a ^ (1 << b);
+    // return x;
+
+    // We only need `*a`
+    return (v ^ (1 << i));
+#else
     int r = 0; asm ("btc %1,%0" : "=r"(r) : "r"(i), "0"(v) : "flags" ); return r;
+#endif
   }
-  
+
   __forceinline int bts(int v, int i) {
+#if defined(__aarch64__)
+    // _bittestandset(long *a, long b) {
+    // unsigned char x = (*a >> b) & 1;
+    //  *a = *a | (1 << b);
+    //  return x;
+    return (v | (v << i));
+#else
     int r = 0; asm ("bts %1,%0" : "=r"(r) : "r"(i), "0"(v) : "flags"); return r;
+#endif
   }
-  
+
   __forceinline int btr(int v, int i) {
+#if defined(__aarch64__)
+    // _bittestandreset(long *a, long b) {
+    // unsigned char x = (*a >> b) & 1;
+    //  *a = *a & ~(1 << b);
+    //  return x;
+    return (v & ~(v << i));
+#else
     int r = 0; asm ("btr %1,%0" : "=r"(r) : "r"(i), "0"(v) : "flags"); return r;
+#endif
   }
-  
+
   __forceinline size_t btc(size_t v, size_t i) {
+#if defined(__aarch64__)
+    return (v ^ (1 << i));
+#else
     size_t r = 0; asm ("btc %1,%0" : "=r"(r) : "r"(i), "0"(v) : "flags" ); return r;
+#endif
   }
-  
+
   __forceinline size_t bts(size_t v, size_t i) {
+#if defined(__aarch64__)
+    return (v | (v << i));
+#else
     size_t r = 0; asm ("bts %1,%0" : "=r"(r) : "r"(i), "0"(v) : "flags"); return r;
+#endif
   }
-  
+
   __forceinline size_t btr(size_t v, size_t i) {
+#if defined(__ARM_NEON)
+    return (v & ~(v << i));
+#else
     size_t r = 0; asm ("btr %1,%0" : "=r"(r) : "r"(i), "0"(v) : "flags"); return r;
+#endif
   }
 
   __forceinline int32_t atomic_cmpxchg(int32_t volatile* value, int32_t comparand, const int32_t input) {
     return __sync_val_compare_and_swap(value, comparand, input);
   }
-  
+
 #endif
-  
+
 ////////////////////////////////////////////////////////////////////////////////
 /// All Platforms
 ////////////////////////////////////////////////////////////////////////////////
-  
+
 #if defined(__clang__) || defined(__GNUC__)
 #if !defined(_mm_undefined_ps)
   __forceinline __m128 _mm_undefined_ps() { return _mm_setzero_ps(); }
@@ -390,39 +471,39 @@ namespace embree
 #endif
 #endif
 
-#if defined(__SSE4_2__)
+#if defined(__SSE4_2__) || defined(__ARM_NEON)
-  
+
   __forceinline int popcnt(int in) {
     return _mm_popcnt_u32(in);
   }
-  
+
   __forceinline unsigned popcnt(unsigned in) {
     return _mm_popcnt_u32(in);
   }
-  
+
-#if defined(__X86_64__)
+#if defined(__X86_64__) || defined(__ARM_NEON)
   __forceinline size_t popcnt(size_t in) {
     return _mm_popcnt_u64(in);
   }
 #endif
-  
+
 #endif
 
   __forceinline uint64_t rdtsc()
   {
-    int dummy[4]; 
+    int dummy[4];
-    __cpuid(dummy,0); 
+    __cpuid(dummy,0);
-    uint64_t clock = read_tsc(); 
+    uint64_t clock = read_tsc();
-    __cpuid(dummy,0); 
+    __cpuid(dummy,0);
     return clock;
   }
-  
+
   __forceinline void pause_cpu(const size_t N = 8)
   {
     for (size_t i=0; i<N; i++)
       _mm_pause();
   }
-  
+
   /* prefetches */
   __forceinline void prefetchL1 (const void* ptr) { _mm_prefetch((const char*)ptr,_MM_HINT_T0); }
   __forceinline void prefetchL2 (const void* ptr) { _mm_prefetch((const char*)ptr,_MM_HINT_T1); }
@@ -432,18 +513,18 @@ namespace embree
 #if defined(__INTEL_COMPILER)
     _mm_prefetch((const char*)ptr,_MM_HINT_ET0);
 #else
-    _mm_prefetch((const char*)ptr,_MM_HINT_T0);    
+    _mm_prefetch((const char*)ptr,_MM_HINT_T0);
 #endif
   }
 
-  __forceinline void prefetchL1EX(const void* ptr) { 
+  __forceinline void prefetchL1EX(const void* ptr) {
-    prefetchEX(ptr); 
+    prefetchEX(ptr);
   }
-  
+
-  __forceinline void prefetchL2EX(const void* ptr) { 
+  __forceinline void prefetchL2EX(const void* ptr) {
-    prefetchEX(ptr); 
+    prefetchEX(ptr);
   }
-#if defined(__AVX2__)
+#if defined(__AVX2__) && !defined(__aarch64__)
    __forceinline unsigned int pext(unsigned int a, unsigned int b) { return _pext_u32(a, b); }
    __forceinline unsigned int pdep(unsigned int a, unsigned int b) { return _pdep_u32(a, b); }
 #if defined(__X86_64__)

thirdparty/embree/common/sys/library.cpp

@@ -27,9 +27,7 @@ namespace embree
 
   /* returns address of a symbol from the library */
   void* getSymbol(lib_t lib, const std::string& sym) {
-    // -- GODOT start --
+    return reinterpret_cast<void *>(GetProcAddress(HMODULE(lib),sym.c_str()));
-    return (void*) GetProcAddress(HMODULE(lib),sym.c_str());
-    // -- GODOT end --
   }
 
   /* closes the shared library */
@@ -63,7 +61,7 @@ namespace embree
     lib = dlopen((executable.path() + fullName).c_str(),RTLD_NOW);
     if (lib == nullptr) {
       const char* error = dlerror();
-      if (error) { 
+      if (error) {
         THROW_RUNTIME_ERROR(error);
       } else {
         THROW_RUNTIME_ERROR("could not load library "+executable.str());

thirdparty/embree/common/sys/mutex.cpp

@@ -36,6 +36,7 @@ namespace embree
     MAYBE_UNUSED bool ok = pthread_mutex_destroy((pthread_mutex_t*)mutex) == 0;
     assert(ok);
     delete (pthread_mutex_t*)mutex; 
+    mutex = nullptr;
   }
   
   void MutexSys::lock() 

thirdparty/embree/common/sys/mutex.h

@@ -47,7 +47,7 @@ namespace embree
       {
         while (flag.load()) 
         {
-          _mm_pause();
+          _mm_pause(); 
           _mm_pause();
         }
         
@@ -74,7 +74,7 @@ namespace embree
     {
       while(flag.load())
       {
-        _mm_pause();
+        _mm_pause(); 
         _mm_pause();
       }
     }

thirdparty/embree/common/sys/platform.h

@@ -88,10 +88,10 @@
 #define dll_import __declspec(dllimport)
 #else
 #define dll_export __attribute__ ((visibility ("default")))
-#define dll_import 
+#define dll_import
 #endif
 
-#if defined(__WIN32__) && !defined(__MINGW32__)
+#ifdef __WIN32__
 #if !defined(__noinline)
 #define __noinline             __declspec(noinline)
 #endif
@@ -103,11 +103,16 @@
 #define __restrict__           //__restrict // causes issues with MSVC
 #endif
 #if !defined(__thread)
+// NOTE: Require `-fms-extensions` for clang
 #define __thread               __declspec(thread)
 #endif
 #if !defined(__aligned)
+#if defined(__MINGW32__)
+#define __aligned(...)           __attribute__((aligned(__VA_ARGS__)))
+#else
 #define __aligned(...)           __declspec(align(__VA_ARGS__))
 #endif
+#endif
 //#define __FUNCTION__           __FUNCTION__
 #define debugbreak()           __debugbreak()
 
@@ -142,7 +147,7 @@
 #endif
 
 // -- GODOT start --
-#if !defined(likely)
+#ifndef likely
 // -- GODOT end --
 #if defined(_MSC_VER) && !defined(__INTEL_COMPILER)
 #define   likely(expr) (expr)
@@ -169,11 +174,19 @@
 #define PRINT4(x,y,z,w) embree_cout << STRING(x) << " = " << (x) << ", " << STRING(y) << " = " << (y) << ", " << STRING(z) << " = " << (z) << ", " << STRING(w) << " = " << (w) << embree_endl
 
 #if defined(DEBUG) // only report file and line in debug mode
+  // -- GODOT start --
+  // #define THROW_RUNTIME_ERROR(str)
+  //   throw std::runtime_error(std::string(__FILE__) + " (" + toString(__LINE__) + "): " + std::string(str));
   #define THROW_RUNTIME_ERROR(str) \
-    throw std::runtime_error(std::string(__FILE__) + " (" + toString(__LINE__) + "): " + std::string(str));
+    printf(std::string(__FILE__) + " (" + toString(__LINE__) + "): " + std::string(str)), abort();
+  // -- GODOT end --
 #else
+  // -- GODOT start --
+  // #define THROW_RUNTIME_ERROR(str)
+  //   throw std::runtime_error(str);
   #define THROW_RUNTIME_ERROR(str) \
-    throw std::runtime_error(str);
+    abort();
+  // -- GODOT end --
 #endif
 
 #define FATAL(x)   THROW_RUNTIME_ERROR(x)
@@ -192,7 +205,7 @@ namespace embree {
 
 /* windows does not have ssize_t */
 #if defined(__WIN32__)
-#if defined(__X86_64__)
+#if defined(__X86_64__) || defined(__aarch64__)
 typedef int64_t ssize_t;
 #else
 typedef int32_t ssize_t;
@@ -316,7 +329,7 @@ __forceinline std::string toString(long long value) {
 /// Some macros for static profiling
 ////////////////////////////////////////////////////////////////////////////////
 
-#if defined (__GNUC__) 
+#if defined (__GNUC__)
 #define IACA_SSC_MARK( MARK_ID )						\
 __asm__ __volatile__ (									\
 					  "\n\t  movl $"#MARK_ID", %%ebx"	\
@@ -355,7 +368,7 @@ namespace embree
     bool active;
     const Closure f;
   };
-  
+
   template <typename Closure>
     OnScopeExitHelper<Closure> OnScopeExit(const Closure f) {
     return OnScopeExitHelper<Closure>(f);

thirdparty/embree/common/sys/sysinfo.cpp

@@ -18,10 +18,16 @@ typedef cpuset_t cpu_set_t;
 namespace embree
 {
   NullTy null;
-  
+
-  std::string getPlatformName() 
+  std::string getPlatformName()
   {
-#if defined(__LINUX__) && !defined(__X86_64__)
+#if defined(__LINUX__) && defined(__ANDROID__) && defined(__aarch64__) && defined(__ARM_NEON)
+    return "Android Linux (aarch64 / arm64)";
+#elif defined(__LINUX__) && defined(__ANDROID__) && defined(__X86_64__)
+    return "Android Linux (x64)";
+#elif defined(__LINUX__) && defined(__ANDROID__) && (defined(_X86_) || defined(__X86__) || defined(_M_IX86))
+    return "Android Linux (x86)";
+#elif defined(__LINUX__) && !defined(__X86_64__)
     return "Linux (32bit)";
 #elif defined(__LINUX__) && defined(__X86_64__)
     return "Linux (64bit)";
@@ -37,10 +43,16 @@ namespace embree
     return "Windows (32bit)";
 #elif defined(__WIN32__) && defined(__X86_64__)
     return "Windows (64bit)";
+#elif defined(TARGET_IPHONE_SIMULATOR) && defined(__X86_64__)
+    return "iOS Simulator (x64)";
+#elif defined(TARGET_OS_IPHONE) && defined(__aarch64__) && defined(__ARM_NEON)
+    return "iOS (aarch64 / arm64)";
 #elif defined(__MACOSX__) && !defined(__X86_64__)
     return "Mac OS X (32bit)";
 #elif defined(__MACOSX__) && defined(__X86_64__)
     return "Mac OS X (64bit)";
+#elif defined(__UNIX__) && defined(__aarch64__)
+    return "Unix (aarch64)";
 #elif defined(__UNIX__) && !defined(__X86_64__)
     return "Unix (32bit)";
 #elif defined(__UNIX__) && defined(__X86_64__)
@@ -79,8 +91,8 @@ namespace embree
 
   std::string getCPUVendor()
   {
-    int cpuinfo[4]; 
+    int cpuinfo[4];
-    __cpuid (cpuinfo, 0); 
+    __cpuid (cpuinfo, 0);
     int name[4];
     name[0] = cpuinfo[1];
     name[1] = cpuinfo[3];
@@ -89,11 +101,11 @@ namespace embree
     return (char*)name;
   }
 
-  CPU getCPUModel() 
+  CPU getCPUModel()
   {
     if (getCPUVendor() != "GenuineIntel")
       return CPU::UNKNOWN;
-    
+
     int out[4];
     __cpuid(out, 0);
     if (out[0] < 1) return CPU::UNKNOWN;
@@ -183,11 +195,13 @@ namespace embree
     case CPU::NEHALEM                 : return "Nehalem";
     case CPU::CORE2                   : return "Core2";
     case CPU::CORE1                   : return "Core";
+    case CPU::ARM                     : return "Arm";
     case CPU::UNKNOWN                 : return "Unknown CPU";
     }
     return "Unknown CPU (error)";
   }
 
+#if !defined(__ARM_NEON)
   /* constants to access destination registers of CPUID instruction */
   static const int EAX = 0;
   static const int EBX = 1;
@@ -227,13 +241,16 @@ namespace embree
   static const int CPU_FEATURE_BIT_AVX512BW = 1 << 30;    // AVX512BW (byte and word instructions)
   static const int CPU_FEATURE_BIT_AVX512VL = 1 << 31;    // AVX512VL (vector length extensions)
   static const int CPU_FEATURE_BIT_AVX512IFMA = 1 << 21;  // AVX512IFMA (integer fused multiple-add instructions)
-  
+
   /* cpuid[eax=7,ecx=0].ecx */
   static const int CPU_FEATURE_BIT_AVX512VBMI = 1 << 1;   // AVX512VBMI (vector bit manipulation instructions)
+#endif
 
-  __noinline int64_t get_xcr0() 
+#if !defined(__ARM_NEON)
+  __noinline int64_t get_xcr0()
   {
-#if defined (__WIN32__) /* -- GODOT start -- */ && !defined (__MINGW32__) /* -- GODOT end -- */
+    // https://github.com/opencv/opencv/blob/master/modules/core/src/system.cpp#L466
+#if defined (__WIN32__) && defined(_XCR_XFEATURE_ENABLED_MASK)
     int64_t xcr0 = 0; // int64_t is workaround for compiler bug under VS2013, Win32
     xcr0 = _xgetbv(0);
     return xcr0;
@@ -243,21 +260,44 @@ namespace embree
     return xcr0;
 #endif
   }
+#endif
 
   int getCPUFeatures()
   {
+#if defined(__ARM_NEON)
+      int cpu_features = CPU_FEATURE_NEON|CPU_FEATURE_SSE|CPU_FEATURE_SSE2;
+#if defined(NEON_AVX2_EMULATION)
+      cpu_features |= CPU_FEATURE_SSE3|CPU_FEATURE_SSSE3|CPU_FEATURE_SSE42;
+      cpu_features |= CPU_FEATURE_XMM_ENABLED;
+      cpu_features |= CPU_FEATURE_YMM_ENABLED;
+      cpu_features |= CPU_FEATURE_SSE41 | CPU_FEATURE_RDRAND | CPU_FEATURE_F16C;
+      cpu_features |= CPU_FEATURE_POPCNT;
+      cpu_features |= CPU_FEATURE_AVX;
+      cpu_features |= CPU_FEATURE_AVX2;
+      cpu_features |= CPU_FEATURE_FMA3;
+      cpu_features |= CPU_FEATURE_LZCNT;
+      cpu_features |= CPU_FEATURE_BMI1;
+      cpu_features |= CPU_FEATURE_BMI2;
+      cpu_features |= CPU_FEATURE_NEON_2X;
+
+
+ 
+#endif
+     return cpu_features;
+      
+#else
     /* cache CPU features access */
     static int cpu_features = 0;
-    if (cpu_features) 
+    if (cpu_features)
       return cpu_features;
 
     /* get number of CPUID leaves */
-    int cpuid_leaf0[4]; 
+    int cpuid_leaf0[4];
     __cpuid(cpuid_leaf0, 0x00000000);
-    unsigned nIds = cpuid_leaf0[EAX];  
+    unsigned nIds = cpuid_leaf0[EAX];
 
     /* get number of extended CPUID leaves */
-    int cpuid_leafe[4]; 
+    int cpuid_leafe[4];
     __cpuid(cpuid_leafe, 0x80000000);
     unsigned nExIds = cpuid_leafe[EAX];
 
@@ -289,7 +329,7 @@ namespace embree
     if (xmm_enabled) cpu_features |= CPU_FEATURE_XMM_ENABLED;
     if (ymm_enabled) cpu_features |= CPU_FEATURE_YMM_ENABLED;
     if (zmm_enabled) cpu_features |= CPU_FEATURE_ZMM_ENABLED;
-    
+
     if (cpuid_leaf_1[EDX] & CPU_FEATURE_BIT_SSE   ) cpu_features |= CPU_FEATURE_SSE;
     if (cpuid_leaf_1[EDX] & CPU_FEATURE_BIT_SSE2  ) cpu_features |= CPU_FEATURE_SSE2;
     if (cpuid_leaf_1[ECX] & CPU_FEATURE_BIT_SSE3  ) cpu_features |= CPU_FEATURE_SSE3;
@@ -297,8 +337,8 @@ namespace embree
     if (cpuid_leaf_1[ECX] & CPU_FEATURE_BIT_SSE4_1) cpu_features |= CPU_FEATURE_SSE41;
     if (cpuid_leaf_1[ECX] & CPU_FEATURE_BIT_SSE4_2) cpu_features |= CPU_FEATURE_SSE42;
     if (cpuid_leaf_1[ECX] & CPU_FEATURE_BIT_POPCNT) cpu_features |= CPU_FEATURE_POPCNT;
-    
     if (cpuid_leaf_1[ECX] & CPU_FEATURE_BIT_AVX   ) cpu_features |= CPU_FEATURE_AVX;
+
     if (cpuid_leaf_1[ECX] & CPU_FEATURE_BIT_F16C  ) cpu_features |= CPU_FEATURE_F16C;
     if (cpuid_leaf_1[ECX] & CPU_FEATURE_BIT_RDRAND) cpu_features |= CPU_FEATURE_RDRAND;
     if (cpuid_leaf_7[EBX] & CPU_FEATURE_BIT_AVX2  ) cpu_features |= CPU_FEATURE_AVX2;
@@ -310,7 +350,7 @@ namespace embree
     if (cpuid_leaf_7[EBX] & CPU_FEATURE_BIT_AVX512F   ) cpu_features |= CPU_FEATURE_AVX512F;
     if (cpuid_leaf_7[EBX] & CPU_FEATURE_BIT_AVX512DQ  ) cpu_features |= CPU_FEATURE_AVX512DQ;
     if (cpuid_leaf_7[EBX] & CPU_FEATURE_BIT_AVX512PF  ) cpu_features |= CPU_FEATURE_AVX512PF;
-    if (cpuid_leaf_7[EBX] & CPU_FEATURE_BIT_AVX512ER  ) cpu_features |= CPU_FEATURE_AVX512ER; 
+    if (cpuid_leaf_7[EBX] & CPU_FEATURE_BIT_AVX512ER  ) cpu_features |= CPU_FEATURE_AVX512ER;
     if (cpuid_leaf_7[EBX] & CPU_FEATURE_BIT_AVX512CD  ) cpu_features |= CPU_FEATURE_AVX512CD;
     if (cpuid_leaf_7[EBX] & CPU_FEATURE_BIT_AVX512BW  ) cpu_features |= CPU_FEATURE_AVX512BW;
     if (cpuid_leaf_7[EBX] & CPU_FEATURE_BIT_AVX512IFMA) cpu_features |= CPU_FEATURE_AVX512IFMA;
@@ -318,6 +358,7 @@ namespace embree
     if (cpuid_leaf_7[ECX] & CPU_FEATURE_BIT_AVX512VBMI) cpu_features |= CPU_FEATURE_AVX512VBMI;
 
     return cpu_features;
+#endif
   }
 
   std::string stringOfCPUFeatures(int features)
@@ -350,9 +391,11 @@ namespace embree
     if (features & CPU_FEATURE_AVX512VL) str += "AVX512VL ";
     if (features & CPU_FEATURE_AVX512IFMA) str += "AVX512IFMA ";
     if (features & CPU_FEATURE_AVX512VBMI) str += "AVX512VBMI ";
+    if (features & CPU_FEATURE_NEON) str += "NEON ";
+    if (features & CPU_FEATURE_NEON_2X) str += "2xNEON ";
     return str;
   }
-  
+
   std::string stringOfISA (int isa)
   {
     if (isa == SSE) return "SSE";
@@ -365,13 +408,15 @@ namespace embree
     if (isa == AVX2) return "AVX2";
     if (isa == AVX512KNL) return "AVX512KNL";
     if (isa == AVX512SKX) return "AVX512SKX";
+    if (isa == NEON) return "NEON";    
+    if (isa == NEON_2X) return "2xNEON";
     return "UNKNOWN";
   }
 
   bool hasISA(int features, int isa) {
     return (features & isa) == isa;
   }
-  
+
   std::string supportedTargetList (int features)
   {
     std::string v;
@@ -386,6 +431,8 @@ namespace embree
     if (hasISA(features,AVX2)) v += "AVX2 ";
     if (hasISA(features,AVX512KNL)) v += "AVX512KNL ";
     if (hasISA(features,AVX512SKX)) v += "AVX512SKX ";
+    if (hasISA(features,NEON)) v += "NEON ";
+    if (hasISA(features,NEON_2X)) v += "2xNEON ";
     return v;
   }
 }
@@ -409,7 +456,7 @@ namespace embree
     return std::string(filename);
   }
 
-  unsigned int getNumberOfLogicalThreads() 
+  unsigned int getNumberOfLogicalThreads()
   {
     static int nThreads = -1;
     if (nThreads != -1) return nThreads;
@@ -420,11 +467,11 @@ namespace embree
     GetActiveProcessorGroupCountFunc pGetActiveProcessorGroupCount = (GetActiveProcessorGroupCountFunc)GetProcAddress(hlib, "GetActiveProcessorGroupCount");
     GetActiveProcessorCountFunc      pGetActiveProcessorCount      = (GetActiveProcessorCountFunc)     GetProcAddress(hlib, "GetActiveProcessorCount");
 
-    if (pGetActiveProcessorGroupCount && pGetActiveProcessorCount) 
+    if (pGetActiveProcessorGroupCount && pGetActiveProcessorCount)
     {
       int groups = pGetActiveProcessorGroupCount();
       int totalProcessors = 0;
-      for (int i = 0; i < groups; i++) 
+      for (int i = 0; i < groups; i++)
         totalProcessors += pGetActiveProcessorCount(i);
       nThreads = totalProcessors;
     }
@@ -438,7 +485,7 @@ namespace embree
     return nThreads;
   }
 
-  int getTerminalWidth() 
+  int getTerminalWidth()
   {
     HANDLE handle = GetStdHandle(STD_OUTPUT_HANDLE);
     if (handle == INVALID_HANDLE_VALUE) return 80;
@@ -448,7 +495,7 @@ namespace embree
     return info.dwSize.X;
   }
 
-  double getSeconds() 
+  double getSeconds()
   {
     LARGE_INTEGER freq, val;
     QueryPerformanceFrequency(&freq);
@@ -487,7 +534,7 @@ namespace embree
 
 namespace embree
 {
-  std::string getExecutableFileName() 
+  std::string getExecutableFileName()
   {
     std::string pid = "/proc/" + toString(getpid()) + "/exe";
     char buf[4096];
@@ -540,7 +587,7 @@ namespace embree
   size_t getVirtualMemoryBytes() {
     return 0;
   }
-   
+
   size_t getResidentMemoryBytes() {
     return 0;
   }
@@ -570,7 +617,7 @@ namespace embree
   size_t getVirtualMemoryBytes() {
     return 0;
   }
-   
+
   size_t getResidentMemoryBytes() {
     return 0;
   }
@@ -591,12 +638,12 @@ namespace embree
 
 namespace embree
 {
-  unsigned int getNumberOfLogicalThreads() 
+  unsigned int getNumberOfLogicalThreads()
   {
     static int nThreads = -1;
     if (nThreads != -1) return nThreads;
 
-#if defined(__MACOSX__)
+#if defined(__MACOSX__) || defined(__ANDROID__)
     nThreads = sysconf(_SC_NPROCESSORS_ONLN); // does not work in Linux LXC container
     assert(nThreads);
 #else
@@ -604,12 +651,12 @@ namespace embree
     if (pthread_getaffinity_np(pthread_self(), sizeof(set), &set) == 0)
       nThreads = CPU_COUNT(&set);
 #endif
-    
+
     assert(nThreads);
     return nThreads;
   }
 
-  int getTerminalWidth() 
+  int getTerminalWidth()
   {
     struct winsize info;
     if (ioctl(STDOUT_FILENO, TIOCGWINSZ, &info) < 0) return 80;

thirdparty/embree/common/sys/sysinfo.h

@@ -59,7 +59,12 @@
 #  define isa sse
 #  define ISA SSE
 #  define ISA_STR "SSE"
-#else 
+#elif defined(__ARM_NEON)
+// NOTE(LTE): Use sse2 for `isa` for the compatibility at the moment.
+#define isa sse2
+#define ISA NEON
+#define ISA_STR "NEON"
+#else
 #error Unknown ISA
 #endif
 
@@ -87,6 +92,7 @@ namespace embree
     NEHALEM,
     CORE2,
     CORE1,
+    ARM,
     UNKNOWN,
   };
   
@@ -114,7 +120,7 @@ namespace embree
   static const int CPU_FEATURE_SSE3   = 1 << 2;
   static const int CPU_FEATURE_SSSE3  = 1 << 3;
   static const int CPU_FEATURE_SSE41  = 1 << 4;
-  static const int CPU_FEATURE_SSE42  = 1 << 5; 
+  static const int CPU_FEATURE_SSE42  = 1 << 5;
   static const int CPU_FEATURE_POPCNT = 1 << 6;
   static const int CPU_FEATURE_AVX    = 1 << 7;
   static const int CPU_FEATURE_F16C   = 1 << 8;
@@ -125,7 +131,7 @@ namespace embree
   static const int CPU_FEATURE_BMI1   = 1 << 13;
   static const int CPU_FEATURE_BMI2   = 1 << 14;
   static const int CPU_FEATURE_AVX512F = 1 << 16;
-  static const int CPU_FEATURE_AVX512DQ = 1 << 17;    
+  static const int CPU_FEATURE_AVX512DQ = 1 << 17;
   static const int CPU_FEATURE_AVX512PF = 1 << 18;
   static const int CPU_FEATURE_AVX512ER = 1 << 19;
   static const int CPU_FEATURE_AVX512CD = 1 << 20;
@@ -136,7 +142,9 @@ namespace embree
   static const int CPU_FEATURE_XMM_ENABLED = 1 << 25;
   static const int CPU_FEATURE_YMM_ENABLED = 1 << 26;
   static const int CPU_FEATURE_ZMM_ENABLED = 1 << 27;
- 
+  static const int CPU_FEATURE_NEON = 1 << 28;
+  static const int CPU_FEATURE_NEON_2X = 1 << 29;
+
   /*! get CPU features */
   int getCPUFeatures();
 
@@ -147,7 +155,7 @@ namespace embree
   std::string supportedTargetList (int isa);
 
   /*! ISAs */
-  static const int SSE    = CPU_FEATURE_SSE | CPU_FEATURE_XMM_ENABLED; 
+  static const int SSE    = CPU_FEATURE_SSE | CPU_FEATURE_XMM_ENABLED;
   static const int SSE2   = SSE | CPU_FEATURE_SSE2;
   static const int SSE3   = SSE2 | CPU_FEATURE_SSE3;
   static const int SSSE3  = SSE3 | CPU_FEATURE_SSSE3;
@@ -158,6 +166,8 @@ namespace embree
   static const int AVX2   = AVXI | CPU_FEATURE_AVX2 | CPU_FEATURE_FMA3 | CPU_FEATURE_BMI1 | CPU_FEATURE_BMI2 | CPU_FEATURE_LZCNT;
   static const int AVX512KNL = AVX2 | CPU_FEATURE_AVX512F | CPU_FEATURE_AVX512PF | CPU_FEATURE_AVX512ER | CPU_FEATURE_AVX512CD | CPU_FEATURE_ZMM_ENABLED;
   static const int AVX512SKX = AVX2 | CPU_FEATURE_AVX512F | CPU_FEATURE_AVX512DQ | CPU_FEATURE_AVX512CD | CPU_FEATURE_AVX512BW | CPU_FEATURE_AVX512VL | CPU_FEATURE_ZMM_ENABLED;
+  static const int NEON = CPU_FEATURE_NEON | CPU_FEATURE_SSE | CPU_FEATURE_SSE2;
+  static const int NEON_2X = CPU_FEATURE_NEON_2X | AVX2;
 
   /*! converts ISA bitvector into a string */
   std::string stringOfISA(int features);

thirdparty/embree/common/sys/thread.cpp

@@ -6,7 +6,11 @@
 #include "string.h"
 
 #include <iostream>
+#if defined(__ARM_NEON)
+#include "../math/SSE2NEON.h"
+#else
 #include <xmmintrin.h>
+#endif
 
 #if defined(PTHREADS_WIN32)
 #pragma comment (lib, "pthreadVC.lib")
@@ -35,7 +39,7 @@ namespace embree
     GetActiveProcessorCountFunc pGetActiveProcessorCount = (GetActiveProcessorCountFunc)GetProcAddress(hlib, "GetActiveProcessorCount");
     SetThreadGroupAffinityFunc pSetThreadGroupAffinity = (SetThreadGroupAffinityFunc)GetProcAddress(hlib, "SetThreadGroupAffinity");
     SetThreadIdealProcessorExFunc pSetThreadIdealProcessorEx = (SetThreadIdealProcessorExFunc)GetProcAddress(hlib, "SetThreadIdealProcessorEx");
-    if (pGetActiveProcessorGroupCount && pGetActiveProcessorCount && pSetThreadGroupAffinity && pSetThreadIdealProcessorEx) 
+    if (pGetActiveProcessorGroupCount && pGetActiveProcessorCount && pSetThreadGroupAffinity && pSetThreadIdealProcessorEx)
     {
       int groups = pGetActiveProcessorGroupCount();
       int totalProcessors = 0, group = 0, number = 0;
@@ -48,7 +52,7 @@ namespace embree
         }
         totalProcessors += processors;
       }
-  
+
       GROUP_AFFINITY groupAffinity;
       groupAffinity.Group = (WORD)group;
       groupAffinity.Mask = (KAFFINITY)(uint64_t(1) << number);
@@ -57,15 +61,15 @@ namespace embree
       groupAffinity.Reserved[2] = 0;
       if (!pSetThreadGroupAffinity(thread, &groupAffinity, nullptr))
         WARNING("SetThreadGroupAffinity failed"); // on purpose only a warning
-  
+
       PROCESSOR_NUMBER processorNumber;
       processorNumber.Group = group;
       processorNumber.Number = number;
       processorNumber.Reserved = 0;
       if (!pSetThreadIdealProcessorEx(thread, &processorNumber, nullptr))
         WARNING("SetThreadIdealProcessorEx failed"); // on purpose only a warning
-    } 
+    }
-    else 
+    else
     {
       if (!SetThreadAffinityMask(thread, DWORD_PTR(uint64_t(1) << affinity)))
         WARNING("SetThreadAffinityMask failed"); // on purpose only a warning
@@ -79,10 +83,10 @@ namespace embree
     setAffinity(GetCurrentThread(), affinity);
   }
 
-  struct ThreadStartupData 
+  struct ThreadStartupData
   {
   public:
-    ThreadStartupData (thread_func f, void* arg) 
+    ThreadStartupData (thread_func f, void* arg)
       : f(f), arg(arg) {}
   public:
     thread_func f;
@@ -95,6 +99,7 @@ namespace embree
     _mm_setcsr(_mm_getcsr() | /*FTZ:*/ (1<<15) | /*DAZ:*/ (1<<6));
     parg->f(parg->arg);
     delete parg;
+    parg = nullptr;
     return 0;
   }
 
@@ -120,12 +125,6 @@ namespace embree
     CloseHandle(HANDLE(tid));
   }
 
-  /*! destroy a hardware thread by its handle */
-  void destroyThread(thread_t tid) {
-    TerminateThread(HANDLE(tid),0);
-    CloseHandle(HANDLE(tid));
-  }
-
   /*! creates thread local storage */
   tls_t createTls() {
     return tls_t(size_t(TlsAlloc()));
@@ -160,16 +159,21 @@ namespace embree
 #include <sstream>
 #include <algorithm>
 
+#if defined(__ANDROID__)
+#include <pthread.h>
+#endif
+
 namespace embree
 {
   static MutexSys mutex;
   static std::vector<size_t> threadIDs;
-  
+
+#if !defined(__ANDROID__) // TODO(LTE): Implement for Android target
   /* changes thread ID mapping such that we first fill up all thread on one core */
   size_t mapThreadID(size_t threadID)
   {
     Lock<MutexSys> lock(mutex);
-    
+
     if (threadIDs.size() == 0)
     {
       /* parse thread/CPU topology */
@@ -181,11 +185,11 @@ namespace embree
         if (fs.fail()) break;
 
         int i;
-        while (fs >> i) 
+        while (fs >> i)
         {
           if (std::none_of(threadIDs.begin(),threadIDs.end(),[&] (int id) { return id == i; }))
             threadIDs.push_back(i);
-          if (fs.peek() == ',') 
+          if (fs.peek() == ',')
             fs.ignore();
         }
         fs.close();
@@ -229,16 +233,21 @@ namespace embree
 
     return ID;
   }
+#endif
 
   /*! set affinity of the calling thread */
   void setAffinity(ssize_t affinity)
   {
+#if defined(__ANDROID__)
+    // TODO(LTE): Implement
+#else
     cpu_set_t cset;
     CPU_ZERO(&cset);
     size_t threadID = mapThreadID(affinity);
     CPU_SET(threadID, &cset);
 
     pthread_setaffinity_np(pthread_self(), sizeof(cset), &cset);
+#endif
   }
 }
 #endif
@@ -303,21 +312,21 @@ namespace embree
 
 namespace embree
 {
-  struct ThreadStartupData 
+  struct ThreadStartupData
   {
   public:
-    ThreadStartupData (thread_func f, void* arg, int affinity) 
+    ThreadStartupData (thread_func f, void* arg, int affinity)
       : f(f), arg(arg), affinity(affinity) {}
-  public: 
+  public:
     thread_func f;
     void* arg;
     ssize_t affinity;
   };
-  
+
   static void* threadStartup(ThreadStartupData* parg)
   {
     _mm_setcsr(_mm_getcsr() | /*FTZ:*/ (1<<15) | /*DAZ:*/ (1<<6));
-    
+
     /*! Mac OS X does not support setting affinity at thread creation time */
 #if defined(__MACOSX__)
     if (parg->affinity >= 0)
@@ -326,6 +335,7 @@ namespace embree
 
     parg->f(parg->arg);
     delete parg;
+    parg = nullptr;
     return nullptr;
   }
 
@@ -341,13 +351,13 @@ namespace embree
     pthread_t* tid = new pthread_t;
     if (pthread_create(tid,&attr,(void*(*)(void*))threadStartup,new ThreadStartupData(f,arg,threadID)) != 0) {
       pthread_attr_destroy(&attr);
-      delete tid; 
+      delete tid;
       FATAL("pthread_create failed");
     }
     pthread_attr_destroy(&attr);
 
     /* set affinity */
-#if defined(__LINUX__)
+#if defined(__LINUX__) && !defined(__ANDROID__)
     if (threadID >= 0) {
       cpu_set_t cset;
       CPU_ZERO(&cset);
@@ -379,14 +389,8 @@ namespace embree
     delete (pthread_t*)tid;
   }
 
-  /*! destroy a hardware thread by its handle */
-  void destroyThread(thread_t tid) {
-    pthread_cancel(*(pthread_t*)tid);
-    delete (pthread_t*)tid;
-  }
-
   /*! creates thread local storage */
-  tls_t createTls() 
+  tls_t createTls()
   {
     pthread_key_t* key = new pthread_key_t;
     if (pthread_key_create(key,nullptr) != 0) {
@@ -398,14 +402,14 @@ namespace embree
   }
 
   /*! return the thread local storage pointer */
-  void* getTls(tls_t tls) 
+  void* getTls(tls_t tls)
   {
     assert(tls);
     return pthread_getspecific(*(pthread_key_t*)tls);
   }
 
   /*! set the thread local storage pointer */
-  void setTls(tls_t tls, void* const ptr) 
+  void setTls(tls_t tls, void* const ptr)
   {
     assert(tls);
     if (pthread_setspecific(*(pthread_key_t*)tls, ptr) != 0)
@@ -413,7 +417,7 @@ namespace embree
   }
 
   /*! destroys thread local storage identifier */
-  void destroyTls(tls_t tls) 
+  void destroyTls(tls_t tls)
   {
     assert(tls);
     if (pthread_key_delete(*(pthread_key_t*)tls) != 0)

thirdparty/embree/common/sys/thread.h

@@ -29,9 +29,6 @@ namespace embree
   /*! waits until the given thread has terminated */
   void join(thread_t tid);
 
-  /*! destroy handle of a thread */
-  void destroyThread(thread_t tid);
-
   /*! type for handle to thread local storage */
   typedef struct opaque_tls_t* tls_t;
 

thirdparty/embree/common/tasking/taskscheduler.h

@@ -5,6 +5,8 @@
 
 #if defined(TASKING_INTERNAL)
 #  include "taskschedulerinternal.h"
+#elif defined(TASKING_GCD) && defined(BUILD_IOS)
+#  include "taskschedulergcd.h"
 #elif defined(TASKING_TBB)
 #  include "taskschedulertbb.h"
 #elif defined(TASKING_PPL)

thirdparty/embree/common/tasking/taskschedulergcd.h

@@ -0,0 +1,49 @@
+#pragma once
+
+#include "../sys/platform.h"
+#include "../sys/alloc.h"
+#include "../sys/barrier.h"
+#include "../sys/thread.h"
+#include "../sys/mutex.h"
+#include "../sys/condition.h"
+#include "../sys/ref.h"
+
+#include <dispatch/dispatch.h>
+
+namespace embree
+{
+  struct TaskScheduler
+  {
+    /*! initializes the task scheduler */
+    static void create(size_t numThreads, bool set_affinity, bool start_threads);
+
+    /*! destroys the task scheduler again */
+    static void destroy() {}
+
+    /* returns the ID of the current thread */
+    static __forceinline size_t threadID()
+    {
+      return threadIndex();
+    }
+
+    /* returns the index (0..threadCount-1) of the current thread */
+    static __forceinline size_t threadIndex()
+    {
+        currentThreadIndex = (currentThreadIndex + 1) % GCDNumThreads;
+        return currentThreadIndex;
+    }
+
+    /* returns the total number of threads */
+    static __forceinline size_t threadCount()
+    {
+        return GCDNumThreads;
+    }
+
+    private:
+      static size_t GCDNumThreads;
+      static size_t currentThreadIndex;
+
+  };
+
+};
+

thirdparty/embree/common/tasking/taskschedulerinternal.cpp

@@ -48,13 +48,15 @@ namespace embree
     {
       Task* prevTask = thread.task;
       thread.task = this;
-      try {
+      // -- GODOT start --
-        if (thread.scheduler->cancellingException == nullptr)
+      // try {
+      // if (thread.scheduler->cancellingException == nullptr)
           closure->execute();
-      } catch (...) {
+      // } catch (...) {
-        if (thread.scheduler->cancellingException == nullptr)
+      //   if (thread.scheduler->cancellingException == nullptr)
-          thread.scheduler->cancellingException = std::current_exception();
+      //     thread.scheduler->cancellingException = std::current_exception();
-      }
+      // }
+      // -- GODOT end --
       thread.task = prevTask;
       add_dependencies(-1);
     }
@@ -152,6 +154,12 @@ namespace embree
     assert(newNumThreads);
     newNumThreads = min(newNumThreads, (size_t) getNumberOfLogicalThreads());
 
+    // We are observing a few % gain by increasing number threads by 2 on aarch64.
+#if defined(__aarch64__) && defined(BUILD_IOS)
+    numThreads = newNumThreads*2;
+#else
+    numThreads = newNumThreads;
+#endif
     numThreads = newNumThreads;
     if (!startThreads && !running) return;
     running = true;
@@ -291,8 +299,11 @@ namespace embree
     size_t threadIndex = allocThreadIndex();
     condition.wait(mutex, [&] () { return hasRootTask.load(); });
     mutex.unlock();
-    std::exception_ptr except = thread_loop(threadIndex);
+    // -- GODOT start --
-    if (except != nullptr) std::rethrow_exception(except);
+    // std::exception_ptr except = thread_loop(threadIndex);
+    // if (except != nullptr) std::rethrow_exception(except);
+    thread_loop(threadIndex);
+    // -- GODOT end --
   }
 
   void TaskScheduler::reset() {
@@ -324,7 +335,10 @@ namespace embree
     return thread->scheduler->cancellingException == nullptr;
   }
 
-  std::exception_ptr TaskScheduler::thread_loop(size_t threadIndex)
+// -- GODOT start --
+//   std::exception_ptr TaskScheduler::thread_loop(size_t threadIndex)
+  void TaskScheduler::thread_loop(size_t threadIndex)
+// -- GODOT end --
   {
     /* allocate thread structure */
     std::unique_ptr<Thread> mthread(new Thread(threadIndex,this)); // too large for stack allocation
@@ -347,9 +361,10 @@ namespace embree
     swapThread(oldThread);
 
     /* remember exception to throw */
-    std::exception_ptr except = nullptr;
+    // -- GODOT start --
-    if (cancellingException != nullptr) except = cancellingException;
+    // std::exception_ptr except = nullptr;
-
+    // if (cancellingException != nullptr) except = cancellingException;
+    // -- GODOT end --
     /* wait for all threads to terminate */
     threadCounter--;
 #if defined(__WIN32__)
@@ -367,7 +382,10 @@ namespace embree
           yield();
 #endif
 	}
-    return except;
+    // -- GODOT start --
+    // return except;
+    return;
+    // -- GODOT end --
   }
 
   bool TaskScheduler::steal_from_other_threads(Thread& thread)

thirdparty/embree/common/tasking/taskschedulerinternal.h

@@ -123,7 +123,10 @@ namespace embree
       {
         size_t ofs = bytes + ((align - stackPtr) & (align-1));
         if (stackPtr + ofs > CLOSURE_STACK_SIZE)
-          throw std::runtime_error("closure stack overflow");
+          // -- GODOT start --
+          // throw std::runtime_error("closure stack overflow");
+          abort();
+          // -- GODOT end --
         stackPtr += ofs;
         return &stack[stackPtr-bytes];
       }
@@ -132,12 +135,16 @@ namespace embree
       __forceinline void push_right(Thread& thread, const size_t size, const Closure& closure)
       {
         if (right >= TASK_STACK_SIZE)
-          throw std::runtime_error("task stack overflow");
+          // -- GODOT start --
+          // throw std::runtime_error("task stack overflow");
+          abort();
+          // -- GODOT end --
 
 	/* allocate new task on right side of stack */
         size_t oldStackPtr = stackPtr;
         TaskFunction* func = new (alloc(sizeof(ClosureTaskFunction<Closure>))) ClosureTaskFunction<Closure>(closure);
-        new (&tasks[right]) Task(func,thread.task,oldStackPtr,size);
+        /* gcc 8 or later fails to compile without explicit .load() */
+        new (&(tasks[right.load()])) Task(func,thread.task,oldStackPtr,size);
         right++;
 
 	/* also move left pointer */
@@ -238,7 +245,10 @@ namespace embree
     void wait_for_threads(size_t threadCount);
 
     /*! thread loop for all worker threads */
-    std::exception_ptr thread_loop(size_t threadIndex);
+    // -- GODOT start --
+    // std::exception_ptr thread_loop(size_t threadIndex);
+    void thread_loop(size_t threadIndex);
+    // -- GODOT end --
 
     /*! steals a task from a different thread */
     bool steal_from_other_threads(Thread& thread);

thirdparty/embree/common/tasking/taskschedulertbb.h

@@ -12,13 +12,7 @@
 #include "../sys/ref.h"
 
 #if defined(__WIN32__)
-// -- GODOT start --
-#if !defined(NOMINMAX)
-// -- GODOT end --
 #  define NOMINMAX
-// -- GODOT start --
-#endif
-// -- GODOT end --
 #endif
 
 // We need to define these to avoid implicit linkage against

thirdparty/embree/include/embree3/rtcore_common.h

@@ -19,7 +19,7 @@ typedef int ssize_t;
 #endif
 #endif
 
-#if defined(_WIN32) && defined(_MSC_VER)
+#if defined(_WIN32) && !defined(__MINGW32__)
 #  define RTC_ALIGN(...) __declspec(align(__VA_ARGS__))
 #else
 #  define RTC_ALIGN(...) __attribute__((aligned(__VA_ARGS__)))
@@ -35,7 +35,7 @@ typedef int ssize_t;
 #endif
 #endif
 
-#if defined(_WIN32) 
+#if defined(_WIN32)
 #  define RTC_FORCEINLINE __forceinline
 #else
 #  define RTC_FORCEINLINE inline __attribute__((always_inline))
@@ -224,13 +224,13 @@ RTC_FORCEINLINE void rtcInitIntersectContext(struct RTCIntersectContext* context
 }
 
 /* Point query structure for closest point query */
-struct RTC_ALIGN(16) RTCPointQuery 
+struct RTC_ALIGN(16) RTCPointQuery
 {
   float x;                // x coordinate of the query point
   float y;                // y coordinate of the query point
   float z;                // z coordinate of the query point
   float time;             // time of the point query
-  float radius;           // radius of the point query 
+  float radius;           // radius of the point query
 };
 
 /* Structure of a packet of 4 query points */
@@ -250,7 +250,7 @@ struct RTC_ALIGN(32) RTCPointQuery8
   float y[8];                // y coordinate of the query point
   float z[8];                // z coordinate of the query point
   float time[8];             // time of the point query
-  float radius[8];           // radius ofr the point query 
+  float radius[8];           // radius ofr the point query
 };
 
 /* Structure of a packet of 16 query points */
@@ -269,11 +269,11 @@ struct RTC_ALIGN(16) RTCPointQueryContext
 {
   // accumulated 4x4 column major matrices from world space to instance space.
   // undefined if size == 0.
-  float world2inst[RTC_MAX_INSTANCE_LEVEL_COUNT][16]; 
+  float world2inst[RTC_MAX_INSTANCE_LEVEL_COUNT][16];
 
   // accumulated 4x4 column major matrices from instance space to world space.
   // undefined if size == 0.
-  float inst2world[RTC_MAX_INSTANCE_LEVEL_COUNT][16]; 
+  float inst2world[RTC_MAX_INSTANCE_LEVEL_COUNT][16];
 
   // instance ids.
   unsigned int instID[RTC_MAX_INSTANCE_LEVEL_COUNT];
@@ -301,13 +301,13 @@ struct RTC_ALIGN(16) RTCPointQueryFunctionArguments
   void* userPtr;
 
   // primitive and geometry ID of primitive
-  unsigned int  primID;        
+  unsigned int  primID;
-  unsigned int  geomID;    
+  unsigned int  geomID;
 
   // the context with transformation and instance ID stack
   struct RTCPointQueryContext* context;
 
-  // If the current instance transform M (= context->world2inst[context->instStackSize]) 
+  // If the current instance transform M (= context->world2inst[context->instStackSize])
   // is a similarity matrix, i.e there is a constant factor similarityScale such that,
   //    for all x,y: dist(Mx, My) = similarityScale * dist(x, y),
   // The similarity scale is 0, if the current instance transform is not a
@@ -322,5 +322,5 @@ struct RTC_ALIGN(16) RTCPointQueryFunctionArguments
 };
 
 typedef bool (*RTCPointQueryFunction)(struct RTCPointQueryFunctionArguments* args);
-  
+
 RTC_NAMESPACE_END

thirdparty/embree/kernels/builders/bvh_builder_sah.h

@@ -43,7 +43,7 @@ namespace embree
         {
           if (RTC_BUILD_ARGUMENTS_HAS(settings,maxBranchingFactor)) branchingFactor = settings.maxBranchingFactor;
           if (RTC_BUILD_ARGUMENTS_HAS(settings,maxDepth          )) maxDepth        = settings.maxDepth;
-          if (RTC_BUILD_ARGUMENTS_HAS(settings,sahBlockSize      )) logBlockSize    = bsr(settings.sahBlockSize);
+          if (RTC_BUILD_ARGUMENTS_HAS(settings,sahBlockSize      )) logBlockSize    = bsr(static_cast<size_t>(settings.sahBlockSize));
           if (RTC_BUILD_ARGUMENTS_HAS(settings,minLeafSize       )) minLeafSize     = settings.minLeafSize;
           if (RTC_BUILD_ARGUMENTS_HAS(settings,maxLeafSize       )) maxLeafSize     = settings.maxLeafSize;
           if (RTC_BUILD_ARGUMENTS_HAS(settings,traversalCost     )) travCost        = settings.traversalCost;

thirdparty/embree/kernels/bvh/bvh.cpp

@@ -51,7 +51,7 @@ namespace embree
   template<int N>
   void BVHN<N>::layoutLargeNodes(size_t num)
   {
-#if defined(__X86_64__) // do not use tree rotations on 32 bit platforms, barrier bit in NodeRef will cause issues
+#if defined(__X86_64__) || defined(__aarch64__) // do not use tree rotations on 32 bit platforms, barrier bit in NodeRef will cause issues
     struct NodeArea 
     {
       __forceinline NodeArea() {}
@@ -183,7 +183,7 @@ namespace embree
   template class BVHN<8>;
 #endif
 
-#if !defined(__AVX__) || !defined(EMBREE_TARGET_SSE2) && !defined(EMBREE_TARGET_SSE42)
+#if !defined(__AVX__) || !defined(EMBREE_TARGET_SSE2) && !defined(EMBREE_TARGET_SSE42) || defined(__aarch64__)
   template class BVHN<4>;
 #endif
 }

thirdparty/embree/kernels/bvh/bvh.h

@@ -81,7 +81,7 @@ namespace embree
     struct CreateAlloc : public FastAllocator::Create {
       __forceinline CreateAlloc (BVHN* bvh) : FastAllocator::Create(&bvh->alloc) {}
     };
-    
+
     typedef BVHNodeRecord<NodeRef>     NodeRecord;
     typedef BVHNodeRecordMB<NodeRef>   NodeRecordMB;
     typedef BVHNodeRecordMB4D<NodeRef> NodeRecordMB4D;

thirdparty/embree/kernels/bvh/bvh_builder_morton.cpp

@@ -18,7 +18,7 @@
 #include "../geometry/object.h"
 #include "../geometry/instance.h"
 
-#if defined(__X86_64__)
+#if defined(__X86_64__) || defined(__aarch64__)
 #  define ROTATE_TREE 1 // specifies number of tree rotation rounds to perform
 #else
 #  define ROTATE_TREE 0 // do not use tree rotations on 32 bit platforms, barrier bit in NodeRef will cause issues

thirdparty/embree/kernels/bvh/bvh_intersector_stream.h

@@ -172,12 +172,23 @@ namespace embree
           TravRayKStream<K,robust> &p = packets[rayID / K];
           const size_t i = rayID % K;
           const vint<Nx> bitmask(shiftTable[rayID]);
+
+#if defined (__aarch64__)
+          const vfloat<Nx> tNearX = madd(bminX, p.rdir.x[i], p.neg_org_rdir.x[i]);
+          const vfloat<Nx> tNearY = madd(bminY, p.rdir.y[i], p.neg_org_rdir.y[i]);
+          const vfloat<Nx> tNearZ = madd(bminZ, p.rdir.z[i], p.neg_org_rdir.z[i]);
+          const vfloat<Nx> tFarX  = madd(bmaxX, p.rdir.x[i], p.neg_org_rdir.x[i]);
+          const vfloat<Nx> tFarY  = madd(bmaxY, p.rdir.y[i], p.neg_org_rdir.y[i]);
+          const vfloat<Nx> tFarZ  = madd(bmaxZ, p.rdir.z[i], p.neg_org_rdir.z[i]); 
+#else
           const vfloat<Nx> tNearX = msub(bminX, p.rdir.x[i], p.org_rdir.x[i]);
           const vfloat<Nx> tNearY = msub(bminY, p.rdir.y[i], p.org_rdir.y[i]);
           const vfloat<Nx> tNearZ = msub(bminZ, p.rdir.z[i], p.org_rdir.z[i]);
           const vfloat<Nx> tFarX  = msub(bmaxX, p.rdir.x[i], p.org_rdir.x[i]);
           const vfloat<Nx> tFarY  = msub(bmaxY, p.rdir.y[i], p.org_rdir.y[i]);
           const vfloat<Nx> tFarZ  = msub(bmaxZ, p.rdir.z[i], p.org_rdir.z[i]); 
+#endif
+
           const vfloat<Nx> tNear  = maxi(tNearX, tNearY, tNearZ, vfloat<Nx>(p.tnear[i]));
           const vfloat<Nx> tFar   = mini(tFarX , tFarY , tFarZ,  vfloat<Nx>(p.tfar[i]));      
 

thirdparty/embree/kernels/bvh/bvh_node_ref.h

@@ -102,7 +102,7 @@ namespace embree
     
     /*! Sets the barrier bit. */
     __forceinline void setBarrier() {
-#if defined(__X86_64__)
+#if defined(__X86_64__) || defined(__aarch64__)
       assert(!isBarrier());
       ptr |= barrier_mask;
 #else
@@ -112,7 +112,7 @@ namespace embree
     
     /*! Clears the barrier bit. */
     __forceinline void clearBarrier() {
-#if defined(__X86_64__)
+#if defined(__X86_64__) || defined(__aarch64__)
       ptr &= ~barrier_mask;
 #else
       assert(false);

thirdparty/embree/kernels/bvh/bvh_statistics.cpp

@@ -150,7 +150,10 @@ namespace embree
       }
     }
     else {
-      throw std::runtime_error("not supported node type in bvh_statistics");
+      // -- GODOT start --
+      // throw std::runtime_error("not supported node type in bvh_statistics");
+      abort();
+      // -- GODOT end --
     }
     return s;
   } 
@@ -159,7 +162,7 @@ namespace embree
   template class BVHNStatistics<8>;
 #endif
 
-#if !defined(__AVX__) || !defined(EMBREE_TARGET_SSE2) && !defined(EMBREE_TARGET_SSE42)
+#if !defined(__AVX__) || (!defined(EMBREE_TARGET_SSE2) && !defined(EMBREE_TARGET_SSE42)) || defined(__aarch64__)
   template class BVHNStatistics<4>;
 #endif
 }

thirdparty/embree/kernels/bvh/node_intersector1.h

@@ -5,6 +5,15 @@
 
 #include "node_intersector.h"
 
+#if defined(__AVX2__)
+#define __FMA_X4__
+#endif
+
+#if defined(__aarch64__)
+#define __FMA_X4__
+#endif
+
+
 namespace embree
 {
   namespace isa
@@ -29,9 +38,15 @@ namespace embree
         org = Vec3vf<N>(ray_org.x,ray_org.y,ray_org.z);
         dir = Vec3vf<N>(ray_dir.x,ray_dir.y,ray_dir.z);
         rdir = Vec3vf<N>(ray_rdir.x,ray_rdir.y,ray_rdir.z);
-#if defined(__AVX2__)
+#if defined(__FMA_X4__)
         const Vec3fa ray_org_rdir = ray_org*ray_rdir;
+#if !defined(__aarch64__)
         org_rdir = Vec3vf<N>(ray_org_rdir.x,ray_org_rdir.y,ray_org_rdir.z);
+#else
+          //for aarch64, we do not have msub equal instruction, so we negeate orig and use madd
+          //x86 will use msub
+        neg_org_rdir = Vec3vf<N>(-ray_org_rdir.x,-ray_org_rdir.y,-ray_org_rdir.z);
+#endif
 #endif
         nearX = ray_rdir.x >= 0.0f ? 0*sizeof(vfloat<N>) : 1*sizeof(vfloat<N>);
         nearY = ray_rdir.y >= 0.0f ? 2*sizeof(vfloat<N>) : 3*sizeof(vfloat<N>);
@@ -59,8 +74,12 @@ namespace embree
         org  = Vec3vf<Nx>(ray_org.x[k], ray_org.y[k], ray_org.z[k]);
         dir  = Vec3vf<Nx>(ray_dir.x[k], ray_dir.y[k], ray_dir.z[k]);
         rdir = Vec3vf<Nx>(ray_rdir.x[k], ray_rdir.y[k], ray_rdir.z[k]);
-#if defined(__AVX2__)
+#if defined(__FMA_X4__)
-	org_rdir = org*rdir;
+#if !defined(__aarch64__)
+        org_rdir = org*rdir;
+#else
+        neg_org_rdir = -(org*rdir);
+#endif
 #endif
 	nearX = nearXYZ.x[k];
 	nearY = nearXYZ.y[k];
@@ -81,8 +100,14 @@ namespace embree
 
       Vec3fa org_xyz, dir_xyz;
       Vec3vf<Nx> org, dir, rdir;
-#if defined(__AVX2__)
+#if defined(__FMA_X4__)
+#if !defined(__aarch64__)
       Vec3vf<Nx> org_rdir;
+#else
+        //aarch64 version are keeping negation of the org_rdir and use madd
+        //x86 uses msub
+      Vec3vf<Nx> neg_org_rdir;
+#endif
 #endif
 #if defined(__AVX512ER__) // KNL+
       vint16 permX, permY, permZ;
@@ -110,7 +135,6 @@ namespace embree
         dir = Vec3vf<N>(ray_dir.x,ray_dir.y,ray_dir.z);
         rdir_near = Vec3vf<N>(ray_rdir_near.x,ray_rdir_near.y,ray_rdir_near.z);
         rdir_far  = Vec3vf<N>(ray_rdir_far .x,ray_rdir_far .y,ray_rdir_far .z);
-
         nearX = ray_rdir_near.x >= 0.0f ? 0*sizeof(vfloat<N>) : 1*sizeof(vfloat<N>);
         nearY = ray_rdir_near.y >= 0.0f ? 2*sizeof(vfloat<N>) : 3*sizeof(vfloat<N>);
         nearZ = ray_rdir_near.z >= 0.0f ? 4*sizeof(vfloat<N>) : 5*sizeof(vfloat<N>);
@@ -447,13 +471,22 @@ namespace embree
     template<>
       __forceinline size_t intersectNode<4,4>(const typename BVH4::AABBNode* node, const TravRay<4,4,false>& ray, vfloat4& dist)
     {
-#if defined(__AVX2__)
+#if defined(__FMA_X4__)
+#if defined(__aarch64__)
+      const vfloat4 tNearX = madd(vfloat4::load((float*)((const char*)&node->lower_x+ray.nearX)), ray.rdir.x, ray.neg_org_rdir.x);
+      const vfloat4 tNearY = madd(vfloat4::load((float*)((const char*)&node->lower_x+ray.nearY)), ray.rdir.y, ray.neg_org_rdir.y);
+      const vfloat4 tNearZ = madd(vfloat4::load((float*)((const char*)&node->lower_x+ray.nearZ)), ray.rdir.z, ray.neg_org_rdir.z);
+      const vfloat4 tFarX  = madd(vfloat4::load((float*)((const char*)&node->lower_x+ray.farX )), ray.rdir.x, ray.neg_org_rdir.x);
+      const vfloat4 tFarY  = madd(vfloat4::load((float*)((const char*)&node->lower_x+ray.farY )), ray.rdir.y, ray.neg_org_rdir.y);
+      const vfloat4 tFarZ  = madd(vfloat4::load((float*)((const char*)&node->lower_x+ray.farZ )), ray.rdir.z, ray.neg_org_rdir.z);
+#else
       const vfloat4 tNearX = msub(vfloat4::load((float*)((const char*)&node->lower_x+ray.nearX)), ray.rdir.x, ray.org_rdir.x);
       const vfloat4 tNearY = msub(vfloat4::load((float*)((const char*)&node->lower_x+ray.nearY)), ray.rdir.y, ray.org_rdir.y);
       const vfloat4 tNearZ = msub(vfloat4::load((float*)((const char*)&node->lower_x+ray.nearZ)), ray.rdir.z, ray.org_rdir.z);
       const vfloat4 tFarX  = msub(vfloat4::load((float*)((const char*)&node->lower_x+ray.farX )), ray.rdir.x, ray.org_rdir.x);
       const vfloat4 tFarY  = msub(vfloat4::load((float*)((const char*)&node->lower_x+ray.farY )), ray.rdir.y, ray.org_rdir.y);
       const vfloat4 tFarZ  = msub(vfloat4::load((float*)((const char*)&node->lower_x+ray.farZ )), ray.rdir.z, ray.org_rdir.z);
+#endif
 #else
       const vfloat4 tNearX = (vfloat4::load((float*)((const char*)&node->lower_x+ray.nearX)) - ray.org.x) * ray.rdir.x;
       const vfloat4 tNearY = (vfloat4::load((float*)((const char*)&node->lower_x+ray.nearY)) - ray.org.y) * ray.rdir.y;
@@ -462,8 +495,13 @@ namespace embree
       const vfloat4 tFarY  = (vfloat4::load((float*)((const char*)&node->lower_x+ray.farY )) - ray.org.y) * ray.rdir.y;
       const vfloat4 tFarZ  = (vfloat4::load((float*)((const char*)&node->lower_x+ray.farZ )) - ray.org.z) * ray.rdir.z;
 #endif
-      
+
-#if defined(__SSE4_1__) && !defined(__AVX512F__) // up to HSW
+#if defined(__aarch64__)
+      const vfloat4 tNear = maxi(tNearX, tNearY, tNearZ, ray.tnear);
+      const vfloat4 tFar = mini(tFarX, tFarY, tFarZ, ray.tfar);
+      const vbool4 vmask = asInt(tNear) <= asInt(tFar);
+      const size_t mask = movemask(vmask);
+#elif defined(__SSE4_1__) && !defined(__AVX512F__) // up to HSW
       const vfloat4 tNear = maxi(tNearX,tNearY,tNearZ,ray.tnear);
       const vfloat4 tFar  = mini(tFarX ,tFarY ,tFarZ ,ray.tfar);
       const vbool4 vmask = asInt(tNear) > asInt(tFar);
@@ -489,12 +527,22 @@ namespace embree
       __forceinline size_t intersectNode<8,8>(const typename BVH8::AABBNode* node, const TravRay<8,8,false>& ray, vfloat8& dist)
     {
 #if defined(__AVX2__)
+#if defined(__aarch64__)
+      const vfloat8 tNearX = madd(vfloat8::load((float*)((const char*)&node->lower_x+ray.nearX)), ray.rdir.x, ray.neg_org_rdir.x);
+      const vfloat8 tNearY = madd(vfloat8::load((float*)((const char*)&node->lower_x+ray.nearY)), ray.rdir.y, ray.neg_org_rdir.y);
+      const vfloat8 tNearZ = madd(vfloat8::load((float*)((const char*)&node->lower_x+ray.nearZ)), ray.rdir.z, ray.neg_org_rdir.z);
+      const vfloat8 tFarX  = madd(vfloat8::load((float*)((const char*)&node->lower_x+ray.farX )), ray.rdir.x, ray.neg_org_rdir.x);
+      const vfloat8 tFarY  = madd(vfloat8::load((float*)((const char*)&node->lower_x+ray.farY )), ray.rdir.y, ray.neg_org_rdir.y);
+      const vfloat8 tFarZ  = madd(vfloat8::load((float*)((const char*)&node->lower_x+ray.farZ )), ray.rdir.z, ray.neg_org_rdir.z);
+#else
       const vfloat8 tNearX = msub(vfloat8::load((float*)((const char*)&node->lower_x+ray.nearX)), ray.rdir.x, ray.org_rdir.x);
       const vfloat8 tNearY = msub(vfloat8::load((float*)((const char*)&node->lower_x+ray.nearY)), ray.rdir.y, ray.org_rdir.y);
       const vfloat8 tNearZ = msub(vfloat8::load((float*)((const char*)&node->lower_x+ray.nearZ)), ray.rdir.z, ray.org_rdir.z);
       const vfloat8 tFarX  = msub(vfloat8::load((float*)((const char*)&node->lower_x+ray.farX )), ray.rdir.x, ray.org_rdir.x);
       const vfloat8 tFarY  = msub(vfloat8::load((float*)((const char*)&node->lower_x+ray.farY )), ray.rdir.y, ray.org_rdir.y);
       const vfloat8 tFarZ  = msub(vfloat8::load((float*)((const char*)&node->lower_x+ray.farZ )), ray.rdir.z, ray.org_rdir.z);
+#endif
+        
 #else
       const vfloat8 tNearX = (vfloat8::load((float*)((const char*)&node->lower_x+ray.nearX)) - ray.org.x) * ray.rdir.x;
       const vfloat8 tNearY = (vfloat8::load((float*)((const char*)&node->lower_x+ray.nearY)) - ray.org.y) * ray.rdir.y;
@@ -638,13 +686,22 @@ namespace embree
       const vfloat<N>* pFarX  = (const vfloat<N>*)((const char*)&node->lower_x+ray.farX);
       const vfloat<N>* pFarY  = (const vfloat<N>*)((const char*)&node->lower_x+ray.farY);
       const vfloat<N>* pFarZ  = (const vfloat<N>*)((const char*)&node->lower_x+ray.farZ);
-#if defined(__AVX2__)
+#if defined(__FMA_X4__)
+#if defined(__aarch64__)
+      const vfloat<N> tNearX = madd(madd(time,pNearX[6],vfloat<N>(pNearX[0])), ray.rdir.x, ray.neg_org_rdir.x);
+      const vfloat<N> tNearY = madd(madd(time,pNearY[6],vfloat<N>(pNearY[0])), ray.rdir.y, ray.neg_org_rdir.y);
+      const vfloat<N> tNearZ = madd(madd(time,pNearZ[6],vfloat<N>(pNearZ[0])), ray.rdir.z, ray.neg_org_rdir.z);
+      const vfloat<N> tFarX  = madd(madd(time,pFarX [6],vfloat<N>(pFarX [0])), ray.rdir.x, ray.neg_org_rdir.x);
+      const vfloat<N> tFarY  = madd(madd(time,pFarY [6],vfloat<N>(pFarY [0])), ray.rdir.y, ray.neg_org_rdir.y);
+      const vfloat<N> tFarZ  = madd(madd(time,pFarZ [6],vfloat<N>(pFarZ [0])), ray.rdir.z, ray.neg_org_rdir.z);
+#else
       const vfloat<N> tNearX = msub(madd(time,pNearX[6],vfloat<N>(pNearX[0])), ray.rdir.x, ray.org_rdir.x);
       const vfloat<N> tNearY = msub(madd(time,pNearY[6],vfloat<N>(pNearY[0])), ray.rdir.y, ray.org_rdir.y);
       const vfloat<N> tNearZ = msub(madd(time,pNearZ[6],vfloat<N>(pNearZ[0])), ray.rdir.z, ray.org_rdir.z);
       const vfloat<N> tFarX  = msub(madd(time,pFarX [6],vfloat<N>(pFarX [0])), ray.rdir.x, ray.org_rdir.x);
       const vfloat<N> tFarY  = msub(madd(time,pFarY [6],vfloat<N>(pFarY [0])), ray.rdir.y, ray.org_rdir.y);
       const vfloat<N> tFarZ  = msub(madd(time,pFarZ [6],vfloat<N>(pFarZ [0])), ray.rdir.z, ray.org_rdir.z);
+#endif
 #else
       const vfloat<N> tNearX = (madd(time,pNearX[6],vfloat<N>(pNearX[0])) - ray.org.x) * ray.rdir.x;
       const vfloat<N> tNearY = (madd(time,pNearY[6],vfloat<N>(pNearY[0])) - ray.org.y) * ray.rdir.y;
@@ -653,7 +710,7 @@ namespace embree
       const vfloat<N> tFarY  = (madd(time,pFarY [6],vfloat<N>(pFarY [0])) - ray.org.y) * ray.rdir.y;
       const vfloat<N> tFarZ  = (madd(time,pFarZ [6],vfloat<N>(pFarZ [0])) - ray.org.z) * ray.rdir.z;
 #endif
-#if defined(__AVX2__) && !defined(__AVX512F__) // HSW
+#if defined(__FMA_X4__) && !defined(__AVX512F__) // HSW
       const vfloat<N> tNear = maxi(tNearX,tNearY,tNearZ,ray.tnear);
       const vfloat<N> tFar  = mini(tFarX ,tFarY ,tFarZ ,ray.tfar);
       const vbool<N> vmask = asInt(tNear) > asInt(tFar);
@@ -714,13 +771,22 @@ namespace embree
       const vfloat<N>* pFarX  = (const vfloat<N>*)((const char*)&node->lower_x+ray.farX);
       const vfloat<N>* pFarY  = (const vfloat<N>*)((const char*)&node->lower_x+ray.farY);
       const vfloat<N>* pFarZ  = (const vfloat<N>*)((const char*)&node->lower_x+ray.farZ);
-#if defined (__AVX2__)
+#if defined (__FMA_X4__)
+#if defined(__aarch64__)
+      const vfloat<N> tNearX = madd(madd(time,pNearX[6],vfloat<N>(pNearX[0])), ray.rdir.x, ray.neg_org_rdir.x);
+      const vfloat<N> tNearY = madd(madd(time,pNearY[6],vfloat<N>(pNearY[0])), ray.rdir.y, ray.neg_org_rdir.y);
+      const vfloat<N> tNearZ = madd(madd(time,pNearZ[6],vfloat<N>(pNearZ[0])), ray.rdir.z, ray.neg_org_rdir.z);
+      const vfloat<N> tFarX  = madd(madd(time,pFarX [6],vfloat<N>(pFarX [0])), ray.rdir.x, ray.neg_org_rdir.x);
+      const vfloat<N> tFarY  = madd(madd(time,pFarY [6],vfloat<N>(pFarY [0])), ray.rdir.y, ray.neg_org_rdir.y);
+      const vfloat<N> tFarZ  = madd(madd(time,pFarZ [6],vfloat<N>(pFarZ [0])), ray.rdir.z, ray.neg_org_rdir.z);
+#else
       const vfloat<N> tNearX = msub(madd(time,pNearX[6],vfloat<N>(pNearX[0])), ray.rdir.x, ray.org_rdir.x);
       const vfloat<N> tNearY = msub(madd(time,pNearY[6],vfloat<N>(pNearY[0])), ray.rdir.y, ray.org_rdir.y);
       const vfloat<N> tNearZ = msub(madd(time,pNearZ[6],vfloat<N>(pNearZ[0])), ray.rdir.z, ray.org_rdir.z);
       const vfloat<N> tFarX  = msub(madd(time,pFarX [6],vfloat<N>(pFarX [0])), ray.rdir.x, ray.org_rdir.x);
       const vfloat<N> tFarY  = msub(madd(time,pFarY [6],vfloat<N>(pFarY [0])), ray.rdir.y, ray.org_rdir.y);
       const vfloat<N> tFarZ  = msub(madd(time,pFarZ [6],vfloat<N>(pFarZ [0])), ray.rdir.z, ray.org_rdir.z);
+#endif
 #else
       const vfloat<N> tNearX = (madd(time,pNearX[6],vfloat<N>(pNearX[0])) - ray.org.x) * ray.rdir.x;
       const vfloat<N> tNearY = (madd(time,pNearY[6],vfloat<N>(pNearY[0])) - ray.org.y) * ray.rdir.y;
@@ -729,7 +795,7 @@ namespace embree
       const vfloat<N> tFarY  = (madd(time,pFarY [6],vfloat<N>(pFarY [0])) - ray.org.y) * ray.rdir.y;
       const vfloat<N> tFarZ  = (madd(time,pFarZ [6],vfloat<N>(pFarZ [0])) - ray.org.z) * ray.rdir.z;
 #endif
-#if defined(__AVX2__) && !defined(__AVX512F__)
+#if defined(__FMA_X4__) && !defined(__AVX512F__)
       const vfloat<N> tNear = maxi(maxi(tNearX,tNearY),maxi(tNearZ,ray.tnear));
       const vfloat<N> tFar  = mini(mini(tFarX ,tFarY ),mini(tFarZ ,ray.tfar ));
 #else
@@ -803,13 +869,22 @@ namespace embree
       const vfloat4 lower_z = madd(node->dequantize<4>(ray.nearZ >> 2),scale_z,start_z);
       const vfloat4 upper_z = madd(node->dequantize<4>(ray.farZ  >> 2),scale_z,start_z);
 
-#if defined(__AVX2__)
+#if defined(__FMA_X4__)
+#if defined(__aarch64__)
+      const vfloat4 tNearX = madd(lower_x, ray.rdir.x, ray.neg_org_rdir.x);
+      const vfloat4 tNearY = madd(lower_y, ray.rdir.y, ray.neg_org_rdir.y);
+      const vfloat4 tNearZ = madd(lower_z, ray.rdir.z, ray.neg_org_rdir.z);
+      const vfloat4 tFarX  = madd(upper_x, ray.rdir.x, ray.neg_org_rdir.x);
+      const vfloat4 tFarY  = madd(upper_y, ray.rdir.y, ray.neg_org_rdir.y);
+      const vfloat4 tFarZ  = madd(upper_z, ray.rdir.z, ray.neg_org_rdir.z);
+#else
       const vfloat4 tNearX = msub(lower_x, ray.rdir.x, ray.org_rdir.x);
       const vfloat4 tNearY = msub(lower_y, ray.rdir.y, ray.org_rdir.y);
       const vfloat4 tNearZ = msub(lower_z, ray.rdir.z, ray.org_rdir.z);
       const vfloat4 tFarX  = msub(upper_x, ray.rdir.x, ray.org_rdir.x);
       const vfloat4 tFarY  = msub(upper_y, ray.rdir.y, ray.org_rdir.y);
       const vfloat4 tFarZ  = msub(upper_z, ray.rdir.z, ray.org_rdir.z);
+#endif
 #else
       const vfloat4 tNearX = (lower_x - ray.org.x) * ray.rdir.x;
       const vfloat4 tNearY = (lower_y - ray.org.y) * ray.rdir.y;
@@ -819,7 +894,7 @@ namespace embree
       const vfloat4 tFarZ  = (upper_z - ray.org.z) * ray.rdir.z;
 #endif
       
-#if defined(__SSE4_1__) && !defined(__AVX512F__) // up to HSW
+#if (defined(__aarch64__) && defined(BUILD_IOS)) || defined(__SSE4_1__) && !defined(__AVX512F__) // up to HSW
       const vfloat4 tNear = maxi(tNearX,tNearY,tNearZ,ray.tnear);
       const vfloat4 tFar  = mini(tFarX ,tFarY ,tFarZ ,ray.tfar);
       const vbool4 vmask = asInt(tNear) > asInt(tFar);
@@ -892,12 +967,21 @@ namespace embree
       const vfloat8 upper_z = madd(node->dequantize<8>(ray.farZ  >> 2),scale_z,start_z);
 
 #if defined(__AVX2__)
+#if defined(__aarch64__)
+      const vfloat8 tNearX = madd(lower_x, ray.rdir.x, ray.neg_org_rdir.x);
+      const vfloat8 tNearY = madd(lower_y, ray.rdir.y, ray.neg_org_rdir.y);
+      const vfloat8 tNearZ = madd(lower_z, ray.rdir.z, ray.neg_org_rdir.z);
+      const vfloat8 tFarX  = madd(upper_x, ray.rdir.x, ray.neg_org_rdir.x);
+      const vfloat8 tFarY  = madd(upper_y, ray.rdir.y, ray.neg_org_rdir.y);
+      const vfloat8 tFarZ  = madd(upper_z, ray.rdir.z, ray.neg_org_rdir.z);
+#else
       const vfloat8 tNearX = msub(lower_x, ray.rdir.x, ray.org_rdir.x);
       const vfloat8 tNearY = msub(lower_y, ray.rdir.y, ray.org_rdir.y);
       const vfloat8 tNearZ = msub(lower_z, ray.rdir.z, ray.org_rdir.z);
       const vfloat8 tFarX  = msub(upper_x, ray.rdir.x, ray.org_rdir.x);
       const vfloat8 tFarY  = msub(upper_y, ray.rdir.y, ray.org_rdir.y);
       const vfloat8 tFarZ  = msub(upper_z, ray.rdir.z, ray.org_rdir.z);
+#endif
 #else
       const vfloat8 tNearX = (lower_x - ray.org.x) * ray.rdir.x;
       const vfloat8 tNearY = (lower_y - ray.org.y) * ray.rdir.y;
@@ -1078,13 +1162,22 @@ namespace embree
       const vfloat<N> upper_y   = node->dequantizeUpperY(time);
       const vfloat<N> lower_z   = node->dequantizeLowerZ(time);
       const vfloat<N> upper_z   = node->dequantizeUpperZ(time);     
-#if defined(__AVX2__)
+#if defined(__FMA_X4__)
+#if defined(__aarch64__)
+      const vfloat<N> tNearX = madd(lower_x, ray.rdir.x, ray.neg_org_rdir.x);
+      const vfloat<N> tNearY = madd(lower_y, ray.rdir.y, ray.neg_org_rdir.y);
+      const vfloat<N> tNearZ = madd(lower_z, ray.rdir.z, ray.neg_org_rdir.z);
+      const vfloat<N> tFarX  = madd(upper_x, ray.rdir.x, ray.neg_org_rdir.x);
+      const vfloat<N> tFarY  = madd(upper_y, ray.rdir.y, ray.neg_org_rdir.y);
+      const vfloat<N> tFarZ  = madd(upper_z, ray.rdir.z, ray.neg_org_rdir.z);
+#else
       const vfloat<N> tNearX = msub(lower_x, ray.rdir.x, ray.org_rdir.x);
       const vfloat<N> tNearY = msub(lower_y, ray.rdir.y, ray.org_rdir.y);
       const vfloat<N> tNearZ = msub(lower_z, ray.rdir.z, ray.org_rdir.z);
       const vfloat<N> tFarX  = msub(upper_x, ray.rdir.x, ray.org_rdir.x);
       const vfloat<N> tFarY  = msub(upper_y, ray.rdir.y, ray.org_rdir.y);
       const vfloat<N> tFarZ  = msub(upper_z, ray.rdir.z, ray.org_rdir.z);
+#endif
 #else
       const vfloat<N> tNearX = (lower_x - ray.org.x) * ray.rdir.x;
       const vfloat<N> tNearY = (lower_y - ray.org.y) * ray.rdir.y;

thirdparty/embree/kernels/bvh/node_intersector_frustum.h

@@ -81,9 +81,13 @@ namespace embree
         min_rdir = select(pos_rdir, reduced_min_rdir, reduced_max_rdir);
         max_rdir = select(pos_rdir, reduced_max_rdir, reduced_min_rdir);
 
+#if defined (__aarch64__)
+        neg_min_org_rdir = -(min_rdir * select(pos_rdir, reduced_max_org, reduced_min_org));
+        neg_max_org_rdir = -(max_rdir * select(pos_rdir, reduced_min_org, reduced_max_org));
+#else
         min_org_rdir = min_rdir * select(pos_rdir, reduced_max_org, reduced_min_org);
         max_org_rdir = max_rdir * select(pos_rdir, reduced_min_org, reduced_max_org);
-
+#endif
         min_dist = reduced_min_dist;
         max_dist = reduced_max_dist;
 
@@ -101,9 +105,13 @@ namespace embree
       Vec3fa min_rdir;
       Vec3fa max_rdir;
 
+#if defined (__aarch64__)
+      Vec3fa neg_min_org_rdir;
+      Vec3fa neg_max_org_rdir;
+#else
       Vec3fa min_org_rdir;
       Vec3fa max_org_rdir;
-
+#endif
       float min_dist;
       float max_dist;
     };
@@ -203,13 +211,21 @@ namespace embree
       const vfloat<Nx> bmaxY = *(const vfloat<N>*)((const char*)&node->lower_x + frustum.nf.farY);
       const vfloat<Nx> bmaxZ = *(const vfloat<N>*)((const char*)&node->lower_x + frustum.nf.farZ);
 
+#if defined (__aarch64__)
+      const vfloat<Nx> fminX = madd(bminX, vfloat<Nx>(frustum.min_rdir.x), vfloat<Nx>(frustum.neg_min_org_rdir.x));
+      const vfloat<Nx> fminY = madd(bminY, vfloat<Nx>(frustum.min_rdir.y), vfloat<Nx>(frustum.neg_min_org_rdir.y));
+      const vfloat<Nx> fminZ = madd(bminZ, vfloat<Nx>(frustum.min_rdir.z), vfloat<Nx>(frustum.neg_min_org_rdir.z));
+      const vfloat<Nx> fmaxX = madd(bmaxX, vfloat<Nx>(frustum.max_rdir.x), vfloat<Nx>(frustum.neg_max_org_rdir.x));
+      const vfloat<Nx> fmaxY = madd(bmaxY, vfloat<Nx>(frustum.max_rdir.y), vfloat<Nx>(frustum.neg_max_org_rdir.y));
+      const vfloat<Nx> fmaxZ = madd(bmaxZ, vfloat<Nx>(frustum.max_rdir.z), vfloat<Nx>(frustum.neg_max_org_rdir.z));
+#else
       const vfloat<Nx> fminX = msub(bminX, vfloat<Nx>(frustum.min_rdir.x), vfloat<Nx>(frustum.min_org_rdir.x));
       const vfloat<Nx> fminY = msub(bminY, vfloat<Nx>(frustum.min_rdir.y), vfloat<Nx>(frustum.min_org_rdir.y));
       const vfloat<Nx> fminZ = msub(bminZ, vfloat<Nx>(frustum.min_rdir.z), vfloat<Nx>(frustum.min_org_rdir.z));
       const vfloat<Nx> fmaxX = msub(bmaxX, vfloat<Nx>(frustum.max_rdir.x), vfloat<Nx>(frustum.max_org_rdir.x));
       const vfloat<Nx> fmaxY = msub(bmaxY, vfloat<Nx>(frustum.max_rdir.y), vfloat<Nx>(frustum.max_org_rdir.y));
       const vfloat<Nx> fmaxZ = msub(bmaxZ, vfloat<Nx>(frustum.max_rdir.z), vfloat<Nx>(frustum.max_org_rdir.z));
-
+#endif
       const vfloat<Nx> fmin  = maxi(fminX, fminY, fminZ, vfloat<Nx>(frustum.min_dist));
       dist = fmin;
       const vfloat<Nx> fmax  = mini(fmaxX, fmaxY, fmaxZ, vfloat<Nx>(frustum.max_dist));

thirdparty/embree/kernels/bvh/node_intersector_packet.h

@@ -39,10 +39,11 @@ namespace embree
         org = ray_org;
         dir = ray_dir;
         rdir = rcp_safe(ray_dir);
-#if defined(__AVX2__)
+#if defined(__aarch64__)
+        neg_org_rdir = -(org * rdir);
+#elif defined(__AVX2__)
         org_rdir = org * rdir;
 #endif
-
         if (N)
         {
           const int size = sizeof(float)*N;
@@ -55,7 +56,9 @@ namespace embree
       Vec3vf<K> org;
       Vec3vf<K> dir;
       Vec3vf<K> rdir;
-#if defined(__AVX2__)
+#if defined(__aarch64__)
+      Vec3vf<K> neg_org_rdir;
+#elif defined(__AVX2__)
       Vec3vf<K> org_rdir;
 #endif
       Vec3vi<K> nearXYZ;
@@ -119,7 +122,14 @@ namespace embree
                                          const TravRayKFast<K>& ray, vfloat<K>& dist)
 
     {
-  #if defined(__AVX2__)
+#if defined(__aarch64__)
+      const vfloat<K> lclipMinX = madd(node->lower_x[i], ray.rdir.x, ray.neg_org_rdir.x);
+      const vfloat<K> lclipMinY = madd(node->lower_y[i], ray.rdir.y, ray.neg_org_rdir.y);
+      const vfloat<K> lclipMinZ = madd(node->lower_z[i], ray.rdir.z, ray.neg_org_rdir.z);
+      const vfloat<K> lclipMaxX = madd(node->upper_x[i], ray.rdir.x, ray.neg_org_rdir.x);
+      const vfloat<K> lclipMaxY = madd(node->upper_y[i], ray.rdir.y, ray.neg_org_rdir.y);
+      const vfloat<K> lclipMaxZ = madd(node->upper_z[i], ray.rdir.z, ray.neg_org_rdir.z);
+#elif defined(__AVX2__)
       const vfloat<K> lclipMinX = msub(node->lower_x[i], ray.rdir.x, ray.org_rdir.x);
       const vfloat<K> lclipMinY = msub(node->lower_y[i], ray.rdir.y, ray.org_rdir.y);
       const vfloat<K> lclipMinZ = msub(node->lower_z[i], ray.rdir.z, ray.org_rdir.z);
@@ -199,7 +209,14 @@ namespace embree
       const vfloat<K> vupper_y = madd(time, vfloat<K>(node->upper_dy[i]), vfloat<K>(node->upper_y[i]));
       const vfloat<K> vupper_z = madd(time, vfloat<K>(node->upper_dz[i]), vfloat<K>(node->upper_z[i]));
 
-#if defined(__AVX2__)
+#if defined(__aarch64__)
+      const vfloat<K> lclipMinX = madd(vlower_x, ray.rdir.x, ray.neg_org_rdir.x);
+      const vfloat<K> lclipMinY = madd(vlower_y, ray.rdir.y, ray.neg_org_rdir.y);
+      const vfloat<K> lclipMinZ = madd(vlower_z, ray.rdir.z, ray.neg_org_rdir.z);
+      const vfloat<K> lclipMaxX = madd(vupper_x, ray.rdir.x, ray.neg_org_rdir.x);
+      const vfloat<K> lclipMaxY = madd(vupper_y, ray.rdir.y, ray.neg_org_rdir.y);
+      const vfloat<K> lclipMaxZ = madd(vupper_z, ray.rdir.z, ray.neg_org_rdir.z);
+#elif defined(__AVX2__)
       const vfloat<K> lclipMinX = msub(vlower_x, ray.rdir.x, ray.org_rdir.x);
       const vfloat<K> lclipMinY = msub(vlower_y, ray.rdir.y, ray.org_rdir.y);
       const vfloat<K> lclipMinZ = msub(vlower_z, ray.rdir.z, ray.org_rdir.z);
@@ -302,7 +319,14 @@ namespace embree
       const vfloat<K> vupper_y = madd(time, vfloat<K>(node->upper_dy[i]), vfloat<K>(node->upper_y[i]));
       const vfloat<K> vupper_z = madd(time, vfloat<K>(node->upper_dz[i]), vfloat<K>(node->upper_z[i]));
 
-#if defined(__AVX2__)
+#if defined(__aarch64__)
+      const vfloat<K> lclipMinX = madd(vlower_x, ray.rdir.x, ray.neg_org_rdir.x);
+      const vfloat<K> lclipMinY = madd(vlower_y, ray.rdir.y, ray.neg_org_rdir.y);
+      const vfloat<K> lclipMinZ = madd(vlower_z, ray.rdir.z, ray.neg_org_rdir.z);
+      const vfloat<K> lclipMaxX = madd(vupper_x, ray.rdir.x, ray.neg_org_rdir.x);
+      const vfloat<K> lclipMaxY = madd(vupper_y, ray.rdir.y, ray.neg_org_rdir.y);
+      const vfloat<K> lclipMaxZ = madd(vupper_z, ray.rdir.z, ray.neg_org_rdir.z);
+#elif defined(__AVX2__)
       const vfloat<K> lclipMinX = msub(vlower_x, ray.rdir.x, ray.org_rdir.x);
       const vfloat<K> lclipMinY = msub(vlower_y, ray.rdir.y, ray.org_rdir.y);
       const vfloat<K> lclipMinZ = msub(vlower_z, ray.rdir.z, ray.org_rdir.z);
@@ -464,7 +488,14 @@ namespace embree
       const vfloat<N> lower_z = node->dequantizeLowerZ();
       const vfloat<N> upper_z = node->dequantizeUpperZ();
 
-  #if defined(__AVX2__)
+  #if defined(__aarch64__)
+      const vfloat<K> lclipMinX = madd(lower_x[i], ray.rdir.x, ray.neg_org_rdir.x);
+      const vfloat<K> lclipMinY = madd(lower_y[i], ray.rdir.y, ray.neg_org_rdir.y);
+      const vfloat<K> lclipMinZ = madd(lower_z[i], ray.rdir.z, ray.neg_org_rdir.z);
+      const vfloat<K> lclipMaxX = madd(upper_x[i], ray.rdir.x, ray.neg_org_rdir.x);
+      const vfloat<K> lclipMaxY = madd(upper_y[i], ray.rdir.y, ray.neg_org_rdir.y);
+      const vfloat<K> lclipMaxZ = madd(upper_z[i], ray.rdir.z, ray.neg_org_rdir.z);
+  #elif defined(__AVX2__)
       const vfloat<K> lclipMinX = msub(lower_x[i], ray.rdir.x, ray.org_rdir.x);
       const vfloat<K> lclipMinY = msub(lower_y[i], ray.rdir.y, ray.org_rdir.y);
       const vfloat<K> lclipMinZ = msub(lower_z[i], ray.rdir.z, ray.org_rdir.z);
@@ -549,7 +580,14 @@ namespace embree
         const vfloat<K> lower_z = node->dequantizeLowerZ(i,time);
         const vfloat<K> upper_z = node->dequantizeUpperZ(i,time);
         
-#if defined(__AVX2__)
+#if defined(__aarch64__)
+        const vfloat<K> lclipMinX = madd(lower_x, ray.rdir.x, ray.neg_org_rdir.x);
+        const vfloat<K> lclipMinY = madd(lower_y, ray.rdir.y, ray.neg_org_rdir.y);
+        const vfloat<K> lclipMinZ = madd(lower_z, ray.rdir.z, ray.neg_org_rdir.z);
+        const vfloat<K> lclipMaxX = madd(upper_x, ray.rdir.x, ray.neg_org_rdir.x);
+        const vfloat<K> lclipMaxY = madd(upper_y, ray.rdir.y, ray.neg_org_rdir.y);
+        const vfloat<K> lclipMaxZ = madd(upper_z, ray.rdir.z, ray.neg_org_rdir.z);
+#elif defined(__AVX2__)
         const vfloat<K> lclipMinX = msub(lower_x, ray.rdir.x, ray.org_rdir.x);
         const vfloat<K> lclipMinY = msub(lower_y, ray.rdir.y, ray.org_rdir.y);
         const vfloat<K> lclipMinZ = msub(lower_z, ray.rdir.z, ray.org_rdir.z);

thirdparty/embree/kernels/bvh/node_intersector_packet_stream.h

@@ -32,11 +32,19 @@ namespace embree
       __forceinline void init(const Vec3vf<K>& ray_org, const Vec3vf<K>& ray_dir)
       {
         rdir = rcp_safe(ray_dir);
+#if defined(__aarch64__)
+        neg_org_rdir = -(ray_org * rdir);
+#else
         org_rdir = ray_org * rdir;
+#endif
       }
 
       Vec3vf<K> rdir;
+#if defined(__aarch64__)
+      Vec3vf<K> neg_org_rdir;
+#else
       Vec3vf<K> org_rdir;
+#endif
       vfloat<K> tnear;
       vfloat<K> tfar;
     };
@@ -87,12 +95,21 @@ namespace embree
       const vfloat<Nx> bmaxY = vfloat<Nx>(*(const vfloat<N>*)((const char*)&node->lower_x + nf.farY));
       const vfloat<Nx> bmaxZ = vfloat<Nx>(*(const vfloat<N>*)((const char*)&node->lower_x + nf.farZ));
 
+#if defined (__aarch64__)
+      const vfloat<Nx> rminX = madd(bminX, vfloat<Nx>(ray.rdir.x[k]), vfloat<Nx>(ray.neg_org_rdir.x[k]));
+      const vfloat<Nx> rminY = madd(bminY, vfloat<Nx>(ray.rdir.y[k]), vfloat<Nx>(ray.neg_org_rdir.y[k]));
+      const vfloat<Nx> rminZ = madd(bminZ, vfloat<Nx>(ray.rdir.z[k]), vfloat<Nx>(ray.neg_org_rdir.z[k]));
+      const vfloat<Nx> rmaxX = madd(bmaxX, vfloat<Nx>(ray.rdir.x[k]), vfloat<Nx>(ray.neg_org_rdir.x[k]));
+      const vfloat<Nx> rmaxY = madd(bmaxY, vfloat<Nx>(ray.rdir.y[k]), vfloat<Nx>(ray.neg_org_rdir.y[k]));
+      const vfloat<Nx> rmaxZ = madd(bmaxZ, vfloat<Nx>(ray.rdir.z[k]), vfloat<Nx>(ray.neg_org_rdir.z[k]));
+#else
       const vfloat<Nx> rminX = msub(bminX, vfloat<Nx>(ray.rdir.x[k]), vfloat<Nx>(ray.org_rdir.x[k]));
       const vfloat<Nx> rminY = msub(bminY, vfloat<Nx>(ray.rdir.y[k]), vfloat<Nx>(ray.org_rdir.y[k]));
       const vfloat<Nx> rminZ = msub(bminZ, vfloat<Nx>(ray.rdir.z[k]), vfloat<Nx>(ray.org_rdir.z[k]));
       const vfloat<Nx> rmaxX = msub(bmaxX, vfloat<Nx>(ray.rdir.x[k]), vfloat<Nx>(ray.org_rdir.x[k]));
       const vfloat<Nx> rmaxY = msub(bmaxY, vfloat<Nx>(ray.rdir.y[k]), vfloat<Nx>(ray.org_rdir.y[k]));
       const vfloat<Nx> rmaxZ = msub(bmaxZ, vfloat<Nx>(ray.rdir.z[k]), vfloat<Nx>(ray.org_rdir.z[k]));
+#endif
       const vfloat<Nx> rmin  = maxi(rminX, rminY, rminZ, vfloat<Nx>(ray.tnear[k]));
       const vfloat<Nx> rmax  = mini(rmaxX, rmaxY, rmaxZ, vfloat<Nx>(ray.tfar[k]));
 
@@ -113,12 +130,21 @@ namespace embree
       const vfloat<K> bmaxY = *(const float*)(ptr + nf.farY);
       const vfloat<K> bmaxZ = *(const float*)(ptr + nf.farZ);
 
+#if defined (__aarch64__)
+      const vfloat<K> rminX = madd(bminX, ray.rdir.x, ray.neg_org_rdir.x);
+      const vfloat<K> rminY = madd(bminY, ray.rdir.y, ray.neg_org_rdir.y);
+      const vfloat<K> rminZ = madd(bminZ, ray.rdir.z, ray.neg_org_rdir.z);
+      const vfloat<K> rmaxX = madd(bmaxX, ray.rdir.x, ray.neg_org_rdir.x);
+      const vfloat<K> rmaxY = madd(bmaxY, ray.rdir.y, ray.neg_org_rdir.y);
+      const vfloat<K> rmaxZ = madd(bmaxZ, ray.rdir.z, ray.neg_org_rdir.z);
+#else
       const vfloat<K> rminX = msub(bminX, ray.rdir.x, ray.org_rdir.x);
       const vfloat<K> rminY = msub(bminY, ray.rdir.y, ray.org_rdir.y);
       const vfloat<K> rminZ = msub(bminZ, ray.rdir.z, ray.org_rdir.z);
       const vfloat<K> rmaxX = msub(bmaxX, ray.rdir.x, ray.org_rdir.x);
       const vfloat<K> rmaxY = msub(bmaxY, ray.rdir.y, ray.org_rdir.y);
       const vfloat<K> rmaxZ = msub(bmaxZ, ray.rdir.z, ray.org_rdir.z);
+#endif
 
       const vfloat<K> rmin  = maxi(rminX, rminY, rminZ, ray.tnear);
       const vfloat<K> rmax  = mini(rmaxX, rmaxY, rmaxZ, ray.tfar);

thirdparty/embree/kernels/common/accel.h

@@ -332,7 +332,7 @@ namespace embree
         intersectorN.intersect(this,rayN,N,context);
       }
       
-#if defined(__SSE__)
+#if defined(__SSE__) || defined(__ARM_NEON)
       __forceinline void intersect(const vbool4& valid, RayHitK<4>& ray, IntersectContext* context) {
         const vint<4> mask = valid.mask32();
         intersect4(&mask,(RTCRayHit4&)ray,context);
@@ -388,7 +388,7 @@ namespace embree
         intersectorN.occluded(this,rayN,N,context);
       }
       
-#if defined(__SSE__)
+#if defined(__SSE__) || defined(__ARM_NEON)
       __forceinline void occluded(const vbool4& valid, RayK<4>& ray, IntersectContext* context) {
         const vint<4> mask = valid.mask32();
         occluded4(&mask,(RTCRay4&)ray,context);

thirdparty/embree/kernels/common/acceln.cpp

@@ -97,7 +97,7 @@ namespace embree
     for (size_t i=0; i<This->accels.size(); i++) {
       if (This->accels[i]->isEmpty()) continue;
       This->accels[i]->intersectors.occluded4(valid,ray,context);
-#if defined(__SSE2__)
+#if defined(__SSE2__) || defined(__ARM_NEON)
       vbool4 valid0 = asBool(((vint4*)valid)[0]);
       vbool4 hit0   = ((vfloat4*)ray.tfar)[0] >= vfloat4(zero);
       if (unlikely(none(valid0 & hit0))) break;
@@ -111,7 +111,7 @@ namespace embree
     for (size_t i=0; i<This->accels.size(); i++) {
       if (This->accels[i]->isEmpty()) continue;
       This->accels[i]->intersectors.occluded8(valid,ray,context);
-#if defined(__SSE2__) // FIXME: use higher ISA
+#if defined(__SSE2__) || defined(__ARM_NEON) // FIXME: use higher ISA
       vbool4 valid0 = asBool(((vint4*)valid)[0]);
       vbool4 hit0   = ((vfloat4*)ray.tfar)[0] >= vfloat4(zero);
       vbool4 valid1 = asBool(((vint4*)valid)[1]);
@@ -127,7 +127,7 @@ namespace embree
     for (size_t i=0; i<This->accels.size(); i++) {
       if (This->accels[i]->isEmpty()) continue;
       This->accels[i]->intersectors.occluded16(valid,ray,context);
-#if defined(__SSE2__) // FIXME: use higher ISA
+#if defined(__SSE2__) || defined(__ARM_NEON) // FIXME: use higher ISA
       vbool4 valid0 = asBool(((vint4*)valid)[0]);
       vbool4 hit0   = ((vfloat4*)ray.tfar)[0] >= vfloat4(zero);
       vbool4 valid1 = asBool(((vint4*)valid)[1]);

thirdparty/embree/kernels/common/alloc.cpp

@@ -3,6 +3,9 @@
 
 #include "alloc.h"
 #include "../../common/sys/thread.h"
+#if defined(__aarch64__) && defined(BUILD_IOS)
+#include "../../common/sys/barrier.h"
+#endif
 
 namespace embree
 {

thirdparty/embree/kernels/common/alloc.h

@@ -8,6 +8,10 @@
 #include "scene.h"
 #include "primref.h"
 
+#if defined(__aarch64__) && defined(BUILD_IOS)
+#include <mutex>
+#endif
+
 namespace embree
 {
   class FastAllocator
@@ -26,7 +30,7 @@ namespace embree
   public:
 
     struct ThreadLocal2;
-    enum AllocationType { ALIGNED_MALLOC, OS_MALLOC, SHARED, ANY_TYPE };
+    enum AllocationType { ALIGNED_MALLOC, EMBREE_OS_MALLOC, SHARED, ANY_TYPE };
 
     /*! Per thread structure holding the current memory block. */
     struct __aligned(64) ThreadLocal
@@ -132,7 +136,11 @@ namespace embree
       {
         assert(alloc_i);
         if (alloc.load() == alloc_i) return;
+#if defined(__aarch64__) && defined(BUILD_IOS)
+        std::scoped_lock lock(mutex);
+#else
         Lock<SpinLock> lock(mutex);
+#endif
         //if (alloc.load() == alloc_i) return; // not required as only one thread calls bind
         if (alloc.load()) {
           alloc.load()->bytesUsed   += alloc0.getUsedBytes()   + alloc1.getUsedBytes();
@@ -150,7 +158,11 @@ namespace embree
       {
         assert(alloc_i);
         if (alloc.load() != alloc_i) return;
+#if defined(__aarch64__) && defined(BUILD_IOS)
+        std::scoped_lock lock(mutex);
+#else
         Lock<SpinLock> lock(mutex);
+#endif
         if (alloc.load() != alloc_i) return; // required as a different thread calls unbind
         alloc.load()->bytesUsed   += alloc0.getUsedBytes()   + alloc1.getUsedBytes();
         alloc.load()->bytesFree   += alloc0.getFreeBytes()   + alloc1.getFreeBytes();
@@ -161,7 +173,11 @@ namespace embree
       }
 
     public:
+#if defined(__aarch64__) && defined(BUILD_IOS)
+      std::mutex mutex;
+#else
       SpinLock mutex;        //!< required as unbind is called from other threads
+#endif
       std::atomic<FastAllocator*> alloc;  //!< parent allocator
       ThreadLocal alloc0;
       ThreadLocal alloc1;
@@ -169,7 +185,7 @@ namespace embree
 
     FastAllocator (Device* device, bool osAllocation) 
       : device(device), slotMask(0), usedBlocks(nullptr), freeBlocks(nullptr), use_single_mode(false), defaultBlockSize(PAGE_SIZE), estimatedSize(0),
-        growSize(PAGE_SIZE), maxGrowSize(maxAllocationSize), log2_grow_size_scale(0), bytesUsed(0), bytesFree(0), bytesWasted(0), atype(osAllocation ? OS_MALLOC : ALIGNED_MALLOC),
+        growSize(PAGE_SIZE), maxGrowSize(maxAllocationSize), log2_grow_size_scale(0), bytesUsed(0), bytesFree(0), bytesWasted(0), atype(osAllocation ? EMBREE_OS_MALLOC : ALIGNED_MALLOC),
         primrefarray(device,0)
     {
       for (size_t i=0; i<MAX_THREAD_USED_BLOCK_SLOTS; i++)
@@ -206,7 +222,7 @@ namespace embree
 
     void setOSallocation(bool flag)
     {
-      atype = flag ? OS_MALLOC : ALIGNED_MALLOC;
+      atype = flag ? EMBREE_OS_MALLOC : ALIGNED_MALLOC;
     }
 
   private:
@@ -217,7 +233,11 @@ namespace embree
       ThreadLocal2* alloc = thread_local_allocator2;
       if (alloc == nullptr) {
         thread_local_allocator2 = alloc = new ThreadLocal2;
+#if defined(__aarch64__) && defined(BUILD_IOS)
+        std::scoped_lock lock(s_thread_local_allocators_lock);
+#else
         Lock<SpinLock> lock(s_thread_local_allocators_lock);
+#endif
         s_thread_local_allocators.push_back(make_unique(alloc));
       }
       return alloc;
@@ -227,7 +247,11 @@ namespace embree
 
     __forceinline void join(ThreadLocal2* alloc)
     {
+#if defined(__aarch64__) && defined(BUILD_IOS)
+      std::scoped_lock lock(s_thread_local_allocators_lock);
+#else
       Lock<SpinLock> lock(thread_local_allocators_lock);
+#endif
       thread_local_allocators.push_back(alloc);
     }
 
@@ -496,7 +520,11 @@ namespace embree
         /* parallel block creation in case of no freeBlocks, avoids single global mutex */
         if (likely(freeBlocks.load() == nullptr))
         {
+#if defined(__aarch64__) && defined(BUILD_IOS)
+          std::scoped_lock lock(slotMutex[slot]);
+#else
           Lock<SpinLock> lock(slotMutex[slot]);
+#endif
           if (myUsedBlocks == threadUsedBlocks[slot]) {
             const size_t alignedBytes = (bytes+(align-1)) & ~(align-1);
             const size_t allocSize = max(min(growSize,maxGrowSize),alignedBytes);
@@ -509,7 +537,11 @@ namespace embree
 
         /* if this fails allocate new block */
         {
-          Lock<SpinLock> lock(mutex);
+#if defined(__aarch64__) && defined(BUILD_IOS)
+            std::scoped_lock lock(mutex);
+#else
+            Lock<SpinLock> lock(mutex);
+#endif
 	  if (myUsedBlocks == threadUsedBlocks[slot])
 	  {
             if (freeBlocks.load() != nullptr) {
@@ -531,7 +563,11 @@ namespace embree
     /*! add new block */
     void addBlock(void* ptr, ssize_t bytes)
     {
+#if defined(__aarch64__) && defined(BUILD_IOS)
+      std::scoped_lock lock(mutex);
+#else
       Lock<SpinLock> lock(mutex);
+#endif
       const size_t sizeof_Header = offsetof(Block,data[0]);
       void* aptr = (void*) ((((size_t)ptr)+maxAlignment-1) & ~(maxAlignment-1));
       size_t ofs = (size_t) aptr - (size_t) ptr;
@@ -617,8 +653,8 @@ namespace embree
         bytesWasted(alloc->bytesWasted),
         stat_all(alloc,ANY_TYPE),
         stat_malloc(alloc,ALIGNED_MALLOC),
-        stat_4K(alloc,OS_MALLOC,false),
+        stat_4K(alloc,EMBREE_OS_MALLOC,false),
-        stat_2M(alloc,OS_MALLOC,true),
+        stat_2M(alloc,EMBREE_OS_MALLOC,true),
         stat_shared(alloc,SHARED) {}
 
       AllStatistics (size_t bytesUsed,
@@ -711,7 +747,7 @@ namespace embree
         /* We avoid using os_malloc for small blocks as this could
          * cause a risk of fragmenting the virtual address space and
          * reach the limit of vm.max_map_count = 65k under Linux. */
-        if (atype == OS_MALLOC && bytesAllocate < maxAllocationSize)
+        if (atype == EMBREE_OS_MALLOC && bytesAllocate < maxAllocationSize)
           atype = ALIGNED_MALLOC;
 
         /* we need to additionally allocate some header */
@@ -720,7 +756,7 @@ namespace embree
         bytesReserve  = sizeof_Header+bytesReserve;
 
         /* consume full 4k pages with using os_malloc */
-        if (atype == OS_MALLOC) {
+        if (atype == EMBREE_OS_MALLOC) {
           bytesAllocate = ((bytesAllocate+PAGE_SIZE-1) & ~(PAGE_SIZE-1));
           bytesReserve  = ((bytesReserve +PAGE_SIZE-1) & ~(PAGE_SIZE-1));
         }
@@ -752,11 +788,11 @@ namespace embree
             return new (ptr) Block(ALIGNED_MALLOC,bytesAllocate-sizeof_Header,bytesAllocate-sizeof_Header,next,alignment);
           }
         }
-        else if (atype == OS_MALLOC)
+        else if (atype == EMBREE_OS_MALLOC)
         {
           if (device) device->memoryMonitor(bytesAllocate,false);
           bool huge_pages; ptr = os_malloc(bytesReserve,huge_pages);
-          return new (ptr) Block(OS_MALLOC,bytesAllocate-sizeof_Header,bytesReserve-sizeof_Header,next,0,huge_pages);
+          return new (ptr) Block(EMBREE_OS_MALLOC,bytesAllocate-sizeof_Header,bytesReserve-sizeof_Header,next,0,huge_pages);
         }
         else
           assert(false);
@@ -800,7 +836,7 @@ namespace embree
           if (device) device->memoryMonitor(-sizeof_Alloced,true);
         }
 
-        else if (atype == OS_MALLOC) {
+        else if (atype == EMBREE_OS_MALLOC) {
          size_t sizeof_This = sizeof_Header+reserveEnd;
          os_free(this,sizeof_This,huge_pages);
          if (device) device->memoryMonitor(-sizeof_Alloced,true);
@@ -861,7 +897,7 @@ namespace embree
       bool hasType(AllocationType atype_i, bool huge_pages_i) const
       {
         if      (atype_i == ANY_TYPE ) return true;
-        else if (atype   == OS_MALLOC) return atype_i == atype && huge_pages_i == huge_pages;
+        else if (atype   == EMBREE_OS_MALLOC) return atype_i == atype && huge_pages_i == huge_pages;
         else                           return atype_i == atype;
       }
 
@@ -910,7 +946,7 @@ namespace embree
       void print_block() const
       {
         if (atype == ALIGNED_MALLOC) std::cout << "A";
-        else if (atype == OS_MALLOC) std::cout << "O";
+        else if (atype == EMBREE_OS_MALLOC) std::cout << "O";
         else if (atype == SHARED) std::cout << "S";
         if (huge_pages) std::cout << "H";
         size_t bytesUsed = getBlockUsedBytes();
@@ -940,7 +976,11 @@ namespace embree
     std::atomic<Block*> freeBlocks;
 
     std::atomic<Block*> threadBlocks[MAX_THREAD_USED_BLOCK_SLOTS];
+#if defined(__aarch64__) && defined(BUILD_IOS)
+    std::mutex slotMutex[MAX_THREAD_USED_BLOCK_SLOTS];
+#else
     SpinLock slotMutex[MAX_THREAD_USED_BLOCK_SLOTS];
+#endif
 
     bool use_single_mode;
     size_t defaultBlockSize;
@@ -954,7 +994,11 @@ namespace embree
     static __thread ThreadLocal2* thread_local_allocator2;
     static SpinLock s_thread_local_allocators_lock;
     static std::vector<std::unique_ptr<ThreadLocal2>> s_thread_local_allocators;
+#if defined(__aarch64__) && defined(BUILD_IOS)
+    std::mutex thread_local_allocators_lock;
+#else
     SpinLock thread_local_allocators_lock;
+#endif
     std::vector<ThreadLocal2*> thread_local_allocators;
     AllocationType atype;
     mvector<PrimRef> primrefarray;     //!< primrefarray used to allocate nodes

thirdparty/embree/kernels/common/default.h

@@ -55,6 +55,11 @@
 #include <utility>
 #include <sstream>
 
+#if !defined(_DEBUG) && defined(BUILD_IOS)
+#undef assert
+#define assert(_EXPR)
+#endif
+
 namespace embree
 {
   ////////////////////////////////////////////////////////////////////////////////

thirdparty/embree/kernels/common/device.cpp

@@ -221,6 +221,9 @@ namespace embree
 #if defined(TASKING_INTERNAL)
     std::cout << "internal_tasking_system ";
 #endif
+#if defined(TASKING_GCD) && defined(BUILD_IOS)
+    std::cout << "GCD tasking system ";
+#endif
 #if defined(TASKING_PPL)
 	std::cout << "PPL ";
 #endif
@@ -503,6 +506,10 @@ namespace embree
 #if defined(TASKING_PPL)
     case RTC_DEVICE_PROPERTY_TASKING_SYSTEM: return 2;
 #endif
+            
+#if defined(TASKING_GCD) && defined(BUILD_IOS)
+    case RTC_DEVICE_PROPERTY_TASKING_SYSTEM: return 3;
+#endif
 
 #if defined(EMBREE_GEOMETRY_TRIANGLE)
     case RTC_DEVICE_PROPERTY_TRIANGLE_GEOMETRY_SUPPORTED: return 1;

thirdparty/embree/kernels/common/isa.h

@@ -46,7 +46,7 @@ namespace embree
 #define SELECT_SYMBOL_DEFAULT(features,intersector) \
   intersector = isa::intersector;
 
-#if defined(__SSE__)
+#if defined(__SSE__) || defined(__ARM_NEON)
 #if !defined(EMBREE_TARGET_SIMD4)
 #define EMBREE_TARGET_SIMD4
 #endif

thirdparty/embree/kernels/common/primref.h

@@ -29,7 +29,7 @@ namespace embree
 
     __forceinline PrimRef (const BBox3fa& bounds, size_t id) 
     {
-#if defined(__X86_64__)
+#if defined(__X86_64__) || defined(__aarch64__)
       lower = Vec3fx(bounds.lower, (unsigned)(id & 0xFFFFFFFF));
       upper = Vec3fx(bounds.upper, (unsigned)((id >> 32) & 0xFFFFFFFF));
 #else
@@ -79,7 +79,7 @@ namespace embree
 
     /*! returns an size_t sized ID */
     __forceinline size_t ID() const { 
-#if defined(__X86_64__)
+#if defined(__X86_64__) || defined(__aarch64__)
       return size_t(lower.u) + (size_t(upper.u) << 32);
 #else
       return size_t(lower.u);

thirdparty/embree/kernels/common/primref_mb.h

@@ -32,7 +32,7 @@ namespace embree
       : lbounds((LBBox3fx)lbounds_i), time_range(time_range)
     {
       assert(activeTimeSegments > 0);
-#if defined(__X86_64__)
+#if defined(__X86_64__) || defined(__aarch64__)
       lbounds.bounds0.lower.a = id & 0xFFFFFFFF;
       lbounds.bounds0.upper.a = (id >> 32) & 0xFFFFFFFF;
 #else
@@ -47,7 +47,7 @@ namespace embree
       : lbounds((LBBox3fx)lbounds_i), time_range(time_range)
     {
       assert(activeTimeSegments > 0);
-#if defined(__X86_64__)
+#if defined(__X86_64__) || defined(__aarch64__)
       lbounds.bounds0.lower.u = id & 0xFFFFFFFF;
       lbounds.bounds0.upper.u = (id >> 32) & 0xFFFFFFFF;
 #else
@@ -115,7 +115,7 @@ namespace embree
 
     /*! returns an size_t sized ID */
     __forceinline size_t ID() const {
-#if defined(__X86_64__)
+#if defined(__X86_64__) || defined(__aarch64__)
       return size_t(lbounds.bounds0.lower.u) + (size_t(lbounds.bounds0.upper.u) << 32);
 #else
       return size_t(lbounds.bounds0.lower.u);
@@ -163,7 +163,7 @@ namespace embree
       : bbox(bounds.interpolate(0.5f)), _activeTimeSegments(activeTimeSegments), _totalTimeSegments(totalTimeSegments), time_range(time_range)
     {
       assert(activeTimeSegments > 0);
-#if defined(__X86_64__)
+#if defined(__X86_64__) || defined(__aarch64__)
       bbox.lower.u = id & 0xFFFFFFFF;
       bbox.upper.u = (id >> 32) & 0xFFFFFFFF;
 #else
@@ -229,7 +229,7 @@ namespace embree
 
     /*! returns an size_t sized ID */
     __forceinline size_t ID() const { 
-#if defined(__X86_64__)
+#if defined(__X86_64__) || defined(__aarch64__)
       return size_t(bbox.lower.u) + (size_t(bbox.upper.u) << 32);
 #else
       return size_t(bbox.lower.u);

thirdparty/embree/kernels/common/rtcore.cpp

@@ -8,18 +8,31 @@
 #include "scene.h"
 #include "context.h"
 #include "../../include/embree3/rtcore_ray.h"
+
+#if defined(__aarch64__) && defined(BUILD_IOS)
+#include <mutex>
+#endif
+
 using namespace embree;
 
 RTC_NAMESPACE_BEGIN;
 
   /* mutex to make API thread safe */
-  static MutexSys g_mutex;
+#if defined(__aarch64__) && defined(BUILD_IOS)
+    static std::mutex g_mutex;
+#else
+    static MutexSys g_mutex;
+#endif
 
   RTC_API RTCDevice rtcNewDevice(const char* config)
   {
     RTC_CATCH_BEGIN;
     RTC_TRACE(rtcNewDevice);
+#if defined(__aarch64__) && defined(BUILD_IOS)
+    std::scoped_lock lock(g_mutex);
+#else
     Lock<MutexSys> lock(g_mutex);
+#endif
     Device* device = new Device(config);
     return (RTCDevice) device->refInc();
     RTC_CATCH_END(nullptr);
@@ -32,7 +45,11 @@ RTC_NAMESPACE_BEGIN;
     RTC_CATCH_BEGIN;
     RTC_TRACE(rtcRetainDevice);
     RTC_VERIFY_HANDLE(hdevice);
+#if defined(__aarch64__) && defined(BUILD_IOS)
+    std::scoped_lock lock(g_mutex);
+#else
     Lock<MutexSys> lock(g_mutex);
+#endif
     device->refInc();
     RTC_CATCH_END(nullptr);
   }
@@ -43,7 +60,11 @@ RTC_NAMESPACE_BEGIN;
     RTC_CATCH_BEGIN;
     RTC_TRACE(rtcReleaseDevice);
     RTC_VERIFY_HANDLE(hdevice);
+#if defined(__aarch64__) && defined(BUILD_IOS)
+    std::scoped_lock lock(g_mutex);
+#else
     Lock<MutexSys> lock(g_mutex);
+#endif
     device->refDec();
     RTC_CATCH_END(nullptr);
   }
@@ -54,7 +75,11 @@ RTC_NAMESPACE_BEGIN;
     RTC_CATCH_BEGIN;
     RTC_TRACE(rtcGetDeviceProperty);
     RTC_VERIFY_HANDLE(hdevice);
+#if defined(__aarch64__) && defined(BUILD_IOS)
+    std::scoped_lock lock(g_mutex);
+#else
     Lock<MutexSys> lock(g_mutex);
+#endif
     return device->getProperty(prop);
     RTC_CATCH_END(device);
     return 0;
@@ -67,7 +92,11 @@ RTC_NAMESPACE_BEGIN;
     RTC_TRACE(rtcSetDeviceProperty);
     const bool internal_prop = (size_t)prop >= 1000000 && (size_t)prop < 1000004;
     if (!internal_prop) RTC_VERIFY_HANDLE(hdevice); // allow NULL device for special internal settings
+#if defined(__aarch64__) && defined(BUILD_IOS)
+    std::scoped_lock lock(g_mutex);
+#else
     Lock<MutexSys> lock(g_mutex);
+#endif
     device->setProperty(prop,val);
     RTC_CATCH_END(device);
   }
@@ -183,7 +212,11 @@ RTC_NAMESPACE_BEGIN;
     RTC_CATCH_BEGIN;
     RTC_TRACE(rtcSetSceneProgressMonitorFunction);
     RTC_VERIFY_HANDLE(hscene);
+#if defined(__aarch64__) && defined(BUILD_IOS)
+    std::scoped_lock lock(g_mutex);
+#else
     Lock<MutexSys> lock(g_mutex);
+#endif
     scene->setProgressMonitorFunction(progress,ptr);
     RTC_CATCH_END2(scene);
   }
@@ -197,7 +230,10 @@ RTC_NAMESPACE_BEGIN;
     if (quality != RTC_BUILD_QUALITY_LOW &&
         quality != RTC_BUILD_QUALITY_MEDIUM &&
         quality != RTC_BUILD_QUALITY_HIGH)
-      throw std::runtime_error("invalid build quality");
+      // -- GODOT start --
+      // throw std::runtime_error("invalid build quality");
+      abort();
+      // -- GODOT end --
     scene->setBuildQuality(quality);
     RTC_CATCH_END2(scene);
   }
@@ -479,12 +515,12 @@ RTC_NAMESPACE_BEGIN;
 
     IntersectContext context(scene,user_context);
 #if !defined(EMBREE_RAY_PACKETS)
-    Ray4* ray4 = (Ray4*) rayhit;
+    RayHit4* rayhit4 = (RayHit4*)rayhit;
     for (size_t i=0; i<4; i++) {
       if (!valid[i]) continue;
-      RayHit ray1; ray4->get(i,ray1);
+      RayHit ray1; rayhit4->get(i,ray1);
       scene->intersectors.intersect((RTCRayHit&)ray1,&context);
-      ray4->set(i,ray1);
+      rayhit4->set(i,ray1);
     }
 #else
     scene->intersectors.intersect4(valid,*rayhit,&context);
@@ -510,12 +546,12 @@ RTC_NAMESPACE_BEGIN;
 
     IntersectContext context(scene,user_context);
 #if !defined(EMBREE_RAY_PACKETS)
-    Ray8* ray8 = (Ray8*) rayhit;
+    RayHit8* rayhit8 = (RayHit8*) rayhit;
     for (size_t i=0; i<8; i++) {
       if (!valid[i]) continue;
-      RayHit ray1; ray8->get(i,ray1);
+      RayHit ray1; rayhit8->get(i,ray1);
       scene->intersectors.intersect((RTCRayHit&)ray1,&context);
-      ray8->set(i,ray1);
+      rayhit8->set(i,ray1);
     }
 #else
     if (likely(scene->intersectors.intersector8))
@@ -543,12 +579,12 @@ RTC_NAMESPACE_BEGIN;
 
     IntersectContext context(scene,user_context);
 #if !defined(EMBREE_RAY_PACKETS)
-    Ray16* ray16 = (Ray16*) rayhit;
+    RayHit16* rayhit16 = (RayHit16*) rayhit;
     for (size_t i=0; i<16; i++) {
       if (!valid[i]) continue;
-      RayHit ray1; ray16->get(i,ray1);
+      RayHit ray1; rayhit16->get(i,ray1);
       scene->intersectors.intersect((RTCRayHit&)ray1,&context);
-      ray16->set(i,ray1);
+      rayhit16->set(i,ray1);
     }
 #else
     if (likely(scene->intersectors.intersector16))
@@ -730,12 +766,12 @@ RTC_NAMESPACE_BEGIN;
 
     IntersectContext context(scene,user_context);
 #if !defined(EMBREE_RAY_PACKETS)
-    RayHit4* ray4 = (RayHit4*) ray;
+    Ray4* ray4 = (Ray4*) ray;
     for (size_t i=0; i<4; i++) {
       if (!valid[i]) continue;
-      RayHit ray1; ray4->get(i,ray1);
+      Ray ray1; ray4->get(i,ray1);
       scene->intersectors.occluded((RTCRay&)ray1,&context);
-      ray4->geomID[i] = ray1.geomID; 
+      ray4->set(i,ray1);
     }
 #else
     scene->intersectors.occluded4(valid,*ray,&context);
@@ -761,10 +797,10 @@ RTC_NAMESPACE_BEGIN;
 
     IntersectContext context(scene,user_context);
 #if !defined(EMBREE_RAY_PACKETS)
-    RayHit8* ray8 = (RayHit8*) ray;
+    Ray8* ray8 = (Ray8*) ray;
     for (size_t i=0; i<8; i++) {
       if (!valid[i]) continue;
-      RayHit ray1; ray8->get(i,ray1);
+      Ray ray1; ray8->get(i,ray1);
       scene->intersectors.occluded((RTCRay&)ray1,&context);
       ray8->set(i,ray1);
     }
@@ -795,10 +831,10 @@ RTC_NAMESPACE_BEGIN;
 
     IntersectContext context(scene,user_context);
 #if !defined(EMBREE_RAY_PACKETS)
-    RayHit16* ray16 = (RayHit16*) ray;
+    Ray16* ray16 = (Ray16*) ray;
     for (size_t i=0; i<16; i++) {
       if (!valid[i]) continue;
-      RayHit ray1; ray16->get(i,ray1);
+      Ray ray1; ray16->get(i,ray1);
       scene->intersectors.occluded((RTCRay&)ray1,&context);
       ray16->set(i,ray1);
     }
@@ -1350,7 +1386,10 @@ RTC_NAMESPACE_BEGIN;
         quality != RTC_BUILD_QUALITY_MEDIUM &&
         quality != RTC_BUILD_QUALITY_HIGH &&
         quality != RTC_BUILD_QUALITY_REFIT)
-      throw std::runtime_error("invalid build quality");
+      // -- GODOT start --
+      // throw std::runtime_error("invalid build quality");
+      abort();
+      // -- GODOT end --
     geometry->setBuildQuality(quality);
     RTC_CATCH_END2(geometry);
   }

thirdparty/embree/kernels/common/rtcore.h

@@ -25,52 +25,58 @@ namespace embree
 #endif
 
 /*! Macros used in the rtcore API implementation */
-#define RTC_CATCH_BEGIN try {
+// -- GODOT start --
+// #define RTC_CATCH_BEGIN try {
+#define RTC_CATCH_BEGIN
   
-#define RTC_CATCH_END(device)                                                \
+// #define RTC_CATCH_END(device)                                                \
-  } catch (std::bad_alloc&) {                                                   \
+//   } catch (std::bad_alloc&) {                                                   \
-    Device::process_error(device,RTC_ERROR_OUT_OF_MEMORY,"out of memory");      \
+//     Device::process_error(device,RTC_ERROR_OUT_OF_MEMORY,"out of memory");      \
-  } catch (rtcore_error& e) {                                                   \
+//   } catch (rtcore_error& e) {                                                   \
-    Device::process_error(device,e.error,e.what());                             \
+//     Device::process_error(device,e.error,e.what());                             \
-  } catch (std::exception& e) {                                                 \
+//   } catch (std::exception& e) {                                                 \
-    Device::process_error(device,RTC_ERROR_UNKNOWN,e.what());                   \
+//     Device::process_error(device,RTC_ERROR_UNKNOWN,e.what());                   \
-  } catch (...) {                                                               \
+//   } catch (...) {                                                               \
-    Device::process_error(device,RTC_ERROR_UNKNOWN,"unknown exception caught"); \
+//     Device::process_error(device,RTC_ERROR_UNKNOWN,"unknown exception caught"); \
-  }
+//   }
+#define RTC_CATCH_END(device)
   
-#define RTC_CATCH_END2(scene)                                                \
+// #define RTC_CATCH_END2(scene)                                                \
-  } catch (std::bad_alloc&) {                                                   \
+//   } catch (std::bad_alloc&) {                                                   \
-    Device* device = scene ? scene->device : nullptr;                           \
+//     Device* device = scene ? scene->device : nullptr;                           \
-    Device::process_error(device,RTC_ERROR_OUT_OF_MEMORY,"out of memory");      \
+//     Device::process_error(device,RTC_ERROR_OUT_OF_MEMORY,"out of memory");      \
-  } catch (rtcore_error& e) {                                                   \
+//   } catch (rtcore_error& e) {                                                   \
-    Device* device = scene ? scene->device : nullptr;                           \
+//     Device* device = scene ? scene->device : nullptr;                           \
-    Device::process_error(device,e.error,e.what());                             \
+//     Device::process_error(device,e.error,e.what());                             \
-  } catch (std::exception& e) {                                                 \
+//   } catch (std::exception& e) {                                                 \
-    Device* device = scene ? scene->device : nullptr;                           \
+//     Device* device = scene ? scene->device : nullptr;                           \
-    Device::process_error(device,RTC_ERROR_UNKNOWN,e.what());                   \
+//     Device::process_error(device,RTC_ERROR_UNKNOWN,e.what());                   \
-  } catch (...) {                                                               \
+//   } catch (...) {                                                               \
-    Device* device = scene ? scene->device : nullptr;                           \
+//     Device* device = scene ? scene->device : nullptr;                           \
-    Device::process_error(device,RTC_ERROR_UNKNOWN,"unknown exception caught"); \
+//     Device::process_error(device,RTC_ERROR_UNKNOWN,"unknown exception caught"); \
-  }
+//   }
+#define RTC_CATCH_END2(scene)
 
-#define RTC_CATCH_END2_FALSE(scene)                                             \
+// #define RTC_CATCH_END2_FALSE(scene)                                             \
-  } catch (std::bad_alloc&) {                                                   \
+//   } catch (std::bad_alloc&) {                                                   \
-    Device* device = scene ? scene->device : nullptr;                           \
+//     Device* device = scene ? scene->device : nullptr;                           \
-    Device::process_error(device,RTC_ERROR_OUT_OF_MEMORY,"out of memory");      \
+//     Device::process_error(device,RTC_ERROR_OUT_OF_MEMORY,"out of memory");      \
-    return false;                                                               \
+//     return false;                                                               \
-  } catch (rtcore_error& e) {                                                   \
+//   } catch (rtcore_error& e) {                                                   \
-    Device* device = scene ? scene->device : nullptr;                           \
+//     Device* device = scene ? scene->device : nullptr;                           \
-    Device::process_error(device,e.error,e.what());                             \
+//     Device::process_error(device,e.error,e.what());                             \
-    return false;                                                               \
+//     return false;                                                               \
-  } catch (std::exception& e) {                                                 \
+//   } catch (std::exception& e) {                                                 \
-    Device* device = scene ? scene->device : nullptr;                           \
+//     Device* device = scene ? scene->device : nullptr;                           \
-    Device::process_error(device,RTC_ERROR_UNKNOWN,e.what());                   \
+//     Device::process_error(device,RTC_ERROR_UNKNOWN,e.what());                   \
-    return false;                                                               \
+//     return false;                                                               \
-  } catch (...) {                                                               \
+//   } catch (...) {                                                               \
-    Device* device = scene ? scene->device : nullptr;                           \
+//     Device* device = scene ? scene->device : nullptr;                           \
-    Device::process_error(device,RTC_ERROR_UNKNOWN,"unknown exception caught"); \
+//     Device::process_error(device,RTC_ERROR_UNKNOWN,"unknown exception caught"); \
-    return false;                                                               \
+//     return false;                                                               \
-  }
+//   }
+#define RTC_CATCH_END2_FALSE(scene) return false;
+// -- GODOT end --
 
 #define RTC_VERIFY_HANDLE(handle)                               \
   if (handle == nullptr) {                                         \
@@ -97,28 +103,38 @@ namespace embree
 #define RTC_TRACE(x) 
 #endif
 
-  /*! used to throw embree API errors */
+// -- GODOT begin --
-  struct rtcore_error : public std::exception
+//   /*! used to throw embree API errors */
-  {
+//   struct rtcore_error : public std::exception
-    __forceinline rtcore_error(RTCError error, const std::string& str)
+//   {
-      : error(error), str(str) {}
+//     __forceinline rtcore_error(RTCError error, const std::string& str)
-    
+//       : error(error), str(str) {}
-    ~rtcore_error() throw() {}
+//     
-    
+//     ~rtcore_error() throw() {}
-    const char* what () const throw () {
+//     
-      return str.c_str();
+//     const char* what () const throw () {
-    }
+//       return str.c_str();
-    
+//     }
-    RTCError error;
+//     
-    std::string str;
+//     RTCError error;
-  };
+//     std::string str;
+//   };
+// -- GODOT end --
 
 #if defined(DEBUG) // only report file and line in debug mode
+  // -- GODOT begin --
+  // #define throw_RTCError(error,str) \
+  //   throw rtcore_error(error,std::string(__FILE__) + " (" + toString(__LINE__) + "): " + std::string(str));
   #define throw_RTCError(error,str) \
-    throw rtcore_error(error,std::string(__FILE__) + " (" + toString(__LINE__) + "): " + std::string(str));
+    printf(std::string(__FILE__) + " (" + toString(__LINE__) + "): " + std::string(str)), abort();
+  // -- GODOT end --
 #else
+  // -- GODOT begin --
+  // #define throw_RTCError(error,str) \
+  //   throw rtcore_error(error,str);
   #define throw_RTCError(error,str) \
-    throw rtcore_error(error,str);
+    abort();
+  // -- GODOT end --
 #endif
 
 #define RTC_BUILD_ARGUMENTS_HAS(settings,member) \

thirdparty/embree/kernels/common/scene.cpp

@@ -6,7 +6,7 @@
 #include "../bvh/bvh4_factory.h"
 #include "../bvh/bvh8_factory.h"
 #include "../../common/algorithms/parallel_reduce.h"
- 
+
 namespace embree
 {
   /* error raising rtcIntersect and rtcOccluded functions */
@@ -40,7 +40,7 @@ namespace embree
   {
     device->refDec();
   }
-  
+
   void Scene::printStatistics()
   {
     /* calculate maximum number of time segments */
@@ -56,12 +56,12 @@ namespace embree
       statistics[i].resize(max_time_steps);
 
     /* gather statistics */
-    for (size_t i=0; i<size(); i++) 
+    for (size_t i=0; i<size(); i++)
     {
       if (!get(i)) continue;
-      int ty = get(i)->getType(); 
+      int ty = get(i)->getType();
       assert(ty<Geometry::GTY_END);
-      int timesegments = get(i)->numTimeSegments(); 
+      int timesegments = get(i)->numTimeSegments();
       assert((unsigned int)timesegments < max_time_steps);
       statistics[ty][timesegments] += get(i)->size();
     }
@@ -76,7 +76,7 @@ namespace embree
     for (size_t t=0; t<max_time_steps; t++)
       std::cout << "----------";
     std::cout << std::endl;
-    
+
     for (size_t p=0; p<Geometry::GTY_END; p++)
     {
       if (std::string(Geometry::gtype_names[p]) == "") continue;
@@ -90,34 +90,34 @@ namespace embree
   void Scene::createTriangleAccel()
   {
 #if defined(EMBREE_GEOMETRY_TRIANGLE)
-    if (device->tri_accel == "default") 
+    if (device->tri_accel == "default")
     {
       if (quality_flags != RTC_BUILD_QUALITY_LOW)
       {
-        int mode =  2*(int)isCompactAccel() + 1*(int)isRobustAccel(); 
+        int mode =  2*(int)isCompactAccel() + 1*(int)isRobustAccel();
         switch (mode) {
-        case /*0b00*/ 0: 
+        case /*0b00*/ 0:
 #if defined (EMBREE_TARGET_SIMD8)
           if (device->canUseAVX())
 	  {
-            if (quality_flags == RTC_BUILD_QUALITY_HIGH) 
+            if (quality_flags == RTC_BUILD_QUALITY_HIGH)
               accels_add(device->bvh8_factory->BVH8Triangle4(this,BVHFactory::BuildVariant::HIGH_QUALITY,BVHFactory::IntersectVariant::FAST));
             else
               accels_add(device->bvh8_factory->BVH8Triangle4(this,BVHFactory::BuildVariant::STATIC,BVHFactory::IntersectVariant::FAST));
           }
-          else 
+          else
 #endif
-          { 
+          {
-            if (quality_flags == RTC_BUILD_QUALITY_HIGH) 
+            if (quality_flags == RTC_BUILD_QUALITY_HIGH)
               accels_add(device->bvh4_factory->BVH4Triangle4(this,BVHFactory::BuildVariant::HIGH_QUALITY,BVHFactory::IntersectVariant::FAST));
-            else 
+            else
               accels_add(device->bvh4_factory->BVH4Triangle4(this,BVHFactory::BuildVariant::STATIC,BVHFactory::IntersectVariant::FAST));
           }
           break;
 
-        case /*0b01*/ 1: 
+        case /*0b01*/ 1:
 #if defined (EMBREE_TARGET_SIMD8)
-          if (device->canUseAVX()) 
+          if (device->canUseAVX())
             accels_add(device->bvh8_factory->BVH8Triangle4v(this,BVHFactory::BuildVariant::STATIC,BVHFactory::IntersectVariant::ROBUST));
           else
 #endif
@@ -175,8 +175,8 @@ namespace embree
 #if defined(EMBREE_GEOMETRY_TRIANGLE)
     if (device->tri_accel_mb == "default")
     {
-      int mode =  2*(int)isCompactAccel() + 1*(int)isRobustAccel(); 
+      int mode =  2*(int)isCompactAccel() + 1*(int)isRobustAccel();
-      
+
 #if defined (EMBREE_TARGET_SIMD8)
       if (device->canUseAVX2()) // BVH8 reduces performance on AVX only-machines
       {
@@ -211,18 +211,18 @@ namespace embree
   void Scene::createQuadAccel()
   {
 #if defined(EMBREE_GEOMETRY_QUAD)
-    if (device->quad_accel == "default") 
+    if (device->quad_accel == "default")
     {
       if (quality_flags != RTC_BUILD_QUALITY_LOW)
       {
         /* static */
-        int mode =  2*(int)isCompactAccel() + 1*(int)isRobustAccel(); 
+        int mode =  2*(int)isCompactAccel() + 1*(int)isRobustAccel();
         switch (mode) {
         case /*0b00*/ 0:
 #if defined (EMBREE_TARGET_SIMD8)
           if (device->canUseAVX())
           {
-            if (quality_flags == RTC_BUILD_QUALITY_HIGH) 
+            if (quality_flags == RTC_BUILD_QUALITY_HIGH)
               accels_add(device->bvh8_factory->BVH8Quad4v(this,BVHFactory::BuildVariant::HIGH_QUALITY,BVHFactory::IntersectVariant::FAST));
             else
               accels_add(device->bvh8_factory->BVH8Quad4v(this,BVHFactory::BuildVariant::STATIC,BVHFactory::IntersectVariant::FAST));
@@ -230,7 +230,7 @@ namespace embree
           else
 #endif
           {
-            if (quality_flags == RTC_BUILD_QUALITY_HIGH) 
+            if (quality_flags == RTC_BUILD_QUALITY_HIGH)
               accels_add(device->bvh4_factory->BVH4Quad4v(this,BVHFactory::BuildVariant::HIGH_QUALITY,BVHFactory::IntersectVariant::FAST));
             else
               accels_add(device->bvh4_factory->BVH4Quad4v(this,BVHFactory::BuildVariant::STATIC,BVHFactory::IntersectVariant::FAST));
@@ -292,9 +292,9 @@ namespace embree
   void Scene::createQuadMBAccel()
   {
 #if defined(EMBREE_GEOMETRY_QUAD)
-    if (device->quad_accel_mb == "default") 
+    if (device->quad_accel_mb == "default")
     {
-      int mode =  2*(int)isCompactAccel() + 1*(int)isRobustAccel(); 
+      int mode =  2*(int)isCompactAccel() + 1*(int)isRobustAccel();
       switch (mode) {
       case /*0b00*/ 0:
 #if defined (EMBREE_TARGET_SIMD8)
@@ -416,7 +416,7 @@ namespace embree
   void Scene::createUserGeometryAccel()
   {
 #if defined(EMBREE_GEOMETRY_USER)
-    if (device->object_accel == "default") 
+    if (device->object_accel == "default")
     {
 #if defined (EMBREE_TARGET_SIMD8)
       if (device->canUseAVX() && !isCompactAccel())
@@ -554,7 +554,7 @@ namespace embree
   {
     BVHFactory::IntersectVariant ivariant = isRobustAccel() ? BVHFactory::IntersectVariant::ROBUST : BVHFactory::IntersectVariant::FAST;
 #if defined(EMBREE_GEOMETRY_GRID)
-    if (device->grid_accel == "default") 
+    if (device->grid_accel == "default")
     {
 #if defined (EMBREE_TARGET_SIMD8)
       if (device->canUseAVX() && !isCompactAccel())
@@ -579,7 +579,7 @@ namespace embree
   void Scene::createGridMBAccel()
   {
 #if defined(EMBREE_GEOMETRY_GRID)
-    if (device->grid_accel_mb == "default") 
+    if (device->grid_accel_mb == "default")
     {
       accels_add(device->bvh4_factory->BVH4GridMB(this,BVHFactory::BuildVariant::STATIC));
     }
@@ -588,13 +588,17 @@ namespace embree
 #endif
 
   }
-  
+
   void Scene::clear() {
   }
 
-  unsigned Scene::bind(unsigned geomID, Ref<Geometry> geometry) 
+  unsigned Scene::bind(unsigned geomID, Ref<Geometry> geometry)
   {
+#if defined(__aarch64__) && defined(BUILD_IOS)
+    std::scoped_lock lock(geometriesMutex);
+#else
     Lock<SpinLock> lock(geometriesMutex);
+#endif
     if (geomID == RTC_INVALID_GEOMETRY_ID) {
       geomID = id_pool.allocate();
       if (geomID == RTC_INVALID_GEOMETRY_ID)
@@ -620,15 +624,19 @@ namespace embree
 
   void Scene::detachGeometry(size_t geomID)
   {
+#if defined(__aarch64__) && defined(BUILD_IOS)
+    std::scoped_lock lock(geometriesMutex);
+#else
     Lock<SpinLock> lock(geometriesMutex);
-    
+#endif
+
     if (geomID >= geometries.size())
       throw_RTCError(RTC_ERROR_INVALID_OPERATION,"invalid geometry ID");
 
     Ref<Geometry>& geometry = geometries[geomID];
     if (geometry == null)
       throw_RTCError(RTC_ERROR_INVALID_OPERATION,"invalid geometry");
-    
+
     if (geometry->isEnabled()) {
       setModified ();
     }
@@ -650,21 +658,21 @@ namespace embree
     if (!isModified()) {
       return;
     }
-    
+
     /* print scene statistics */
     if (device->verbosity(2))
       printStatistics();
 
     progress_monitor_counter = 0;
-    
+
     /* gather scene stats and call preCommit function of each geometry */
-    this->world = parallel_reduce (size_t(0), geometries.size(), GeometryCounts (), 
+    this->world = parallel_reduce (size_t(0), geometries.size(), GeometryCounts (),
       [this](const range<size_t>& r)->GeometryCounts
       {
         GeometryCounts c;
-        for (auto i=r.begin(); i<r.end(); ++i) 
+        for (auto i=r.begin(); i<r.end(); ++i)
         {
-          if (geometries[i] && geometries[i]->isEnabled()) 
+          if (geometries[i] && geometries[i]->isEnabled())
           {
             geometries[i]->preCommit();
             geometries[i]->addElementsToCount (c);
@@ -675,19 +683,19 @@ namespace embree
       },
       std::plus<GeometryCounts>()
     );
-    
+
     /* select acceleration structures to build */
     unsigned int new_enabled_geometry_types = world.enabledGeometryTypesMask();
     if (flags_modified || new_enabled_geometry_types != enabled_geometry_types)
     {
       accels_init();
 
-      /* we need to make all geometries modified, otherwise two level builder will 
+      /* we need to make all geometries modified, otherwise two level builder will
         not rebuild currently not modified geometries */
       parallel_for(geometryModCounters_.size(), [&] ( const size_t i ) {
           geometryModCounters_[i] = 0;
         });
-      
+
       if (getNumPrimitives(TriangleMesh::geom_type,false)) createTriangleAccel();
       if (getNumPrimitives(TriangleMesh::geom_type,true)) createTriangleMBAccel();
       if (getNumPrimitives(QuadMesh::geom_type,false)) createQuadAccel();
@@ -704,14 +712,14 @@ namespace embree
       if (getNumPrimitives(Geometry::MTY_INSTANCE_CHEAP,true)) createInstanceMBAccel();
       if (getNumPrimitives(Geometry::MTY_INSTANCE_EXPENSIVE,false)) createInstanceExpensiveAccel();
       if (getNumPrimitives(Geometry::MTY_INSTANCE_EXPENSIVE,true)) createInstanceExpensiveMBAccel();
-      
+
       flags_modified = false;
       enabled_geometry_types = new_enabled_geometry_types;
     }
-    
+
     /* select fast code path if no filter function is present */
     accels_select(hasFilterFunction());
-  
+
     /* build all hierarchies of this scene */
     accels_build();
 
@@ -729,7 +737,7 @@ namespace embree
           geometryModCounters_[i] = geometries[i]->getModCounter();
         }
       });
-      
+
     updateInterface();
 
     if (device->verbosity(2)) {
@@ -738,7 +746,7 @@ namespace embree
       std::cout << "selected scene intersector" << std::endl;
       intersectors.print(2);
     }
-    
+
     setModified(false);
   }
 
@@ -763,16 +771,16 @@ namespace embree
   RTCSceneFlags Scene::getSceneFlags() const {
     return scene_flags;
   }
-                   
+
 #if defined(TASKING_INTERNAL)
 
-  void Scene::commit (bool join) 
+  void Scene::commit (bool join)
   {
     Lock<MutexSys> buildLock(buildMutex,false);
 
     /* allocates own taskscheduler for each build */
     Ref<TaskScheduler> scheduler = nullptr;
-    { 
+    {
       Lock<MutexSys> lock(schedulerMutex);
       scheduler = this->scheduler;
       if (scheduler == null) {
@@ -784,31 +792,33 @@ namespace embree
     /* worker threads join build */
     if (!buildLock.isLocked())
     {
-      if (!join) 
+      if (!join)
         throw_RTCError(RTC_ERROR_INVALID_OPERATION,"use rtcJoinCommitScene to join a build operation");
-      
+
       scheduler->join();
       return;
     }
 
     /* initiate build */
-    try {
+    // -- GODOT start --
+    // try {
       scheduler->spawn_root([&]() { commit_task(); Lock<MutexSys> lock(schedulerMutex); this->scheduler = nullptr; }, 1, !join);
-    }
+    // }
-    catch (...) {
+    // catch (...) {
-      accels_clear();
+    //   accels_clear();
-      updateInterface();
+    //   updateInterface();
-      Lock<MutexSys> lock(schedulerMutex);
+    //   Lock<MutexSys> lock(schedulerMutex);
-      this->scheduler = nullptr;
+    //   this->scheduler = nullptr;
-      throw;
+    //   throw;
-    }
+    // }
+    // -- GODOT end --
   }
 
 #endif
 
-#if defined(TASKING_TBB)
+#if defined(TASKING_TBB) || defined(TASKING_GCD)
 
-  void Scene::commit (bool join) 
+  void Scene::commit (bool join)
   {
 #if defined(TASKING_TBB) && (TBB_INTERFACE_VERSION_MAJOR < 8)
     if (join)
@@ -822,12 +832,15 @@ namespace embree
     if (!lock.isLocked())
     {
 #if !TASKING_TBB_USE_TASK_ISOLATION
-      if (!join) 
+      if (!join)
         throw_RTCError(RTC_ERROR_INVALID_OPERATION,"invoking rtcCommitScene from multiple threads is not supported with this TBB version");
 #endif
-      
+
       do {
 
+#if defined(TASKING_GCD)
+      // Do Nothing
+#else
 #if USE_TASK_ARENA
         if (join) {
           device->arena->execute([&]{ group.wait(); });
@@ -837,21 +850,24 @@ namespace embree
         {
           group.wait();
         }
+#endif
 
         pause_cpu();
         yield();
+
       } while (!buildMutex.try_lock());
-      
+
       buildMutex.unlock();
       return;
-    }   
+    }
 
     /* for best performance set FTZ and DAZ flags in the MXCSR control and status register */
     const unsigned int mxcsr = _mm_getcsr();
     _mm_setcsr(mxcsr | /* FTZ */ (1<<15) | /* DAZ */ (1<<6));
-    
+
     try {
-#if TBB_INTERFACE_VERSION_MAJOR < 8    
+#if defined(TASKING_TBB)
+#if TBB_INTERFACE_VERSION_MAJOR < 8
       tbb::task_group_context ctx( tbb::task_group_context::isolated, tbb::task_group_context::default_traits);
 #else
       tbb::task_group_context ctx( tbb::task_group_context::isolated, tbb::task_group_context::default_traits | tbb::task_group_context::fp_settings );
@@ -876,15 +892,22 @@ namespace embree
           });
         group.wait();
       }
-     
+
       /* reset MXCSR register again */
       _mm_setcsr(mxcsr);
-    } 
+
+#elif defined(TASKING_GCD)
+
+      commit_task();
+
+#endif  // #if defined(TASKING_TBB)
+
+    }
     catch (...)
     {
       /* reset MXCSR register again */
       _mm_setcsr(mxcsr);
-      
+
       accels_clear();
       updateInterface();
       throw;
@@ -894,7 +917,7 @@ namespace embree
 
 #if defined(TASKING_PPL)
 
-  void Scene::commit (bool join) 
+  void Scene::commit (bool join)
   {
 #if defined(TASKING_PPL)
     if (join)
@@ -912,7 +935,7 @@ namespace embree
     /* for best performance set FTZ and DAZ flags in the MXCSR control and status register */
     const unsigned int mxcsr = _mm_getcsr();
     _mm_setcsr(mxcsr | /* FTZ */ (1<<15) | /* DAZ */ (1<<6));
-    
+
     try {
 
       group.run([&]{
@@ -922,12 +945,12 @@ namespace embree
 
        /* reset MXCSR register again */
       _mm_setcsr(mxcsr);
-    } 
+    }
     catch (...)
     {
       /* reset MXCSR register again */
       _mm_setcsr(mxcsr);
-      
+
       accels_clear();
       updateInterface();
       throw;
@@ -935,7 +958,7 @@ namespace embree
   }
 #endif
 
-  void Scene::setProgressMonitorFunction(RTCProgressMonitorFunction func, void* ptr) 
+  void Scene::setProgressMonitorFunction(RTCProgressMonitorFunction func, void* ptr)
   {
     progress_monitor_function = func;
     progress_monitor_ptr      = ptr;

thirdparty/embree/kernels/common/scene_subdiv_mesh.h

@@ -275,11 +275,11 @@ namespace embree
     parallel_set<uint32_t> holeSet;
 
     /*! fast lookup table to detect invalid faces */
-    mvector<char> invalid_face;
+    mvector<int8_t> invalid_face;
 
     /*! test if face i is invalid in timestep j */
-    __forceinline       char& invalidFace(size_t i, size_t j = 0)       { return invalid_face[i*numTimeSteps+j]; }
+    __forceinline       int8_t& invalidFace(size_t i, size_t j = 0)       { return invalid_face[i*numTimeSteps+j]; }
-    __forceinline const char& invalidFace(size_t i, size_t j = 0) const { return invalid_face[i*numTimeSteps+j]; }
+    __forceinline const int8_t& invalidFace(size_t i, size_t j = 0) const { return invalid_face[i*numTimeSteps+j]; }
 
     /*! interpolation cache */
   public:

thirdparty/embree/kernels/common/state.cpp

@@ -147,7 +147,20 @@ namespace embree
   }
 
   bool State::checkISASupport() {
+#if defined(__ARM_NEON)
+    /*
+     * NEON CPU type is a mixture of NEON and SSE2
+     */
+
+    bool hasSSE2 = (getCPUFeatures() & enabled_cpu_features) & CPU_FEATURE_SSE2;
+
+    /* this will be true when explicitly initialize Device with `isa=neon` config */
+    bool hasNEON = (getCPUFeatures() & enabled_cpu_features) & CPU_FEATURE_NEON;
+
+    return hasSSE2 || hasNEON;
+#else
     return (getCPUFeatures() & enabled_cpu_features) == enabled_cpu_features;
+#endif
   }
   
   void State::verify()
@@ -160,8 +173,10 @@ namespace embree
      * functions */
 #if defined(DEBUG)
 #if defined(EMBREE_TARGET_SSE2)
+#if !defined(__ARM_NEON)
     assert(sse2::getISA() <= SSE2);
 #endif
+#endif
 #if defined(EMBREE_TARGET_SSE42)
     assert(sse42::getISA() <= SSE42);
 #endif

thirdparty/embree/kernels/geometry/curveNi.h

@@ -43,10 +43,10 @@ namespace embree
     __forceinline void fill(const PrimRef* prims, size_t& begin, size_t _end, Scene* scene)
     {  
       size_t end = min(begin+M,_end);
-      N = (unsigned char)(end-begin);
+      N = (uint8_t)(end-begin);
       const unsigned int geomID0 = prims[begin].geomID();
       this->geomID(N) = geomID0;
-      ty = (unsigned char) scene->get(geomID0)->getType();
+      ty = (uint8_t) scene->get(geomID0)->getType();
 
       /* encode all primitives */
       BBox3fa bounds = empty;
@@ -76,25 +76,25 @@ namespace embree
         const LinearSpace3fa space3(trunc(126.0f*space2.vx),trunc(126.0f*space2.vy),trunc(126.0f*space2.vz));
         const BBox3fa bounds = scene->get(geomID)->vbounds(loffset,lscale,max(length(space3.vx),length(space3.vy),length(space3.vz)),space3.transposed(),primID);
         
-        bounds_vx_x(N)[i] = (char) space3.vx.x;
+        bounds_vx_x(N)[i] = (int8_t) space3.vx.x;
-        bounds_vx_y(N)[i] = (char) space3.vx.y;
+        bounds_vx_y(N)[i] = (int8_t) space3.vx.y;
-        bounds_vx_z(N)[i] = (char) space3.vx.z;
+        bounds_vx_z(N)[i] = (int8_t) space3.vx.z;
         bounds_vx_lower(N)[i] = (short) clamp(floor(bounds.lower.x),-32767.0f,32767.0f);
         bounds_vx_upper(N)[i] = (short) clamp(ceil (bounds.upper.x),-32767.0f,32767.0f);
         assert(-32767.0f <= floor(bounds.lower.x) && floor(bounds.lower.x) <= 32767.0f);
         assert(-32767.0f <= ceil (bounds.upper.x) && ceil (bounds.upper.x) <= 32767.0f);
 
-        bounds_vy_x(N)[i] = (char) space3.vy.x;
+        bounds_vy_x(N)[i] = (int8_t) space3.vy.x;
-        bounds_vy_y(N)[i] = (char) space3.vy.y;
+        bounds_vy_y(N)[i] = (int8_t) space3.vy.y;
-        bounds_vy_z(N)[i] = (char) space3.vy.z;
+        bounds_vy_z(N)[i] = (int8_t) space3.vy.z;
         bounds_vy_lower(N)[i] = (short) clamp(floor(bounds.lower.y),-32767.0f,32767.0f);
         bounds_vy_upper(N)[i] = (short) clamp(ceil (bounds.upper.y),-32767.0f,32767.0f);
         assert(-32767.0f <= floor(bounds.lower.y) && floor(bounds.lower.y) <= 32767.0f);
         assert(-32767.0f <= ceil (bounds.upper.y) && ceil (bounds.upper.y) <= 32767.0f);
 
-        bounds_vz_x(N)[i] = (char) space3.vz.x;
+        bounds_vz_x(N)[i] = (int8_t) space3.vz.x;
-        bounds_vz_y(N)[i] = (char) space3.vz.y;
+        bounds_vz_y(N)[i] = (int8_t) space3.vz.y;
-        bounds_vz_z(N)[i] = (char) space3.vz.z;
+        bounds_vz_z(N)[i] = (int8_t) space3.vz.z;
         bounds_vz_lower(N)[i] = (short) clamp(floor(bounds.lower.z),-32767.0f,32767.0f);
         bounds_vz_upper(N)[i] = (short) clamp(ceil (bounds.upper.z),-32767.0f,32767.0f);
         assert(-32767.0f <= floor(bounds.lower.z) && floor(bounds.lower.z) <= 32767.0f);
@@ -114,15 +114,15 @@ namespace embree
       for (size_t i=0; i<items; i++) {
         accel[i].fill(prims,start,set.end(),bvh->scene);
       }
-      return bvh->encodeLeaf((char*)accel,items);
+      return bvh->encodeLeaf((int8_t*)accel,items);
     };
     
   public:
     
     // 27.6 - 46 bytes per primitive
-    unsigned char ty;
+    uint8_t ty;
-    unsigned char N;
+    uint8_t N;
-    unsigned char data[4+25*M+16];
+    uint8_t data[4+25*M+16];
 
     /*
     struct Layout
@@ -130,21 +130,21 @@ namespace embree
       unsigned int geomID;
       unsigned int primID[N];
       
-      char bounds_vx_x[N];
+      int8_t bounds_vx_x[N];
-      char bounds_vx_y[N];
+      int8_t bounds_vx_y[N];
-      char bounds_vx_z[N];
+      int8_t bounds_vx_z[N];
       short bounds_vx_lower[N];
       short bounds_vx_upper[N];
       
-      char bounds_vy_x[N];
+      int8_t bounds_vy_x[N];
-      char bounds_vy_y[N];
+      int8_t bounds_vy_y[N];
-      char bounds_vy_z[N];
+      int8_t bounds_vy_z[N];
       short bounds_vy_lower[N];
       short bounds_vy_upper[N];
       
-      char bounds_vz_x[N];
+      int8_t bounds_vz_x[N];
-      char bounds_vz_y[N];
+      int8_t bounds_vz_y[N];
-      char bounds_vz_z[N];
+      int8_t bounds_vz_z[N];
       short bounds_vz_lower[N];
       short bounds_vz_upper[N];
       
@@ -153,65 +153,65 @@ namespace embree
     };
     */
     
-    __forceinline       unsigned int& geomID(size_t N)       { return *(unsigned int*)((char*)this+2); }
+    __forceinline       unsigned int& geomID(size_t N)       { return *(unsigned int*)((int8_t*)this+2); }
-    __forceinline const unsigned int& geomID(size_t N) const { return *(unsigned int*)((char*)this+2); }
+    __forceinline const unsigned int& geomID(size_t N) const { return *(unsigned int*)((int8_t*)this+2); }
     
-    __forceinline       unsigned int* primID(size_t N)       { return (unsigned int*)((char*)this+6); }
+    __forceinline       unsigned int* primID(size_t N)       { return (unsigned int*)((int8_t*)this+6); }
-    __forceinline const unsigned int* primID(size_t N) const { return (unsigned int*)((char*)this+6); }
+    __forceinline const unsigned int* primID(size_t N) const { return (unsigned int*)((int8_t*)this+6); }
     
-    __forceinline       char* bounds_vx_x(size_t N)       { return (char*)((char*)this+6+4*N); }
+    __forceinline       int8_t* bounds_vx_x(size_t N)       { return (int8_t*)((int8_t*)this+6+4*N); }
-    __forceinline const char* bounds_vx_x(size_t N) const { return (char*)((char*)this+6+4*N); }
+    __forceinline const int8_t* bounds_vx_x(size_t N) const { return (int8_t*)((int8_t*)this+6+4*N); }
     
-    __forceinline       char* bounds_vx_y(size_t N)       { return (char*)((char*)this+6+5*N); }
+    __forceinline       int8_t* bounds_vx_y(size_t N)       { return (int8_t*)((int8_t*)this+6+5*N); }
-    __forceinline const char* bounds_vx_y(size_t N) const { return (char*)((char*)this+6+5*N); }
+    __forceinline const int8_t* bounds_vx_y(size_t N) const { return (int8_t*)((int8_t*)this+6+5*N); }
     
-    __forceinline       char* bounds_vx_z(size_t N)       { return (char*)((char*)this+6+6*N); }
+    __forceinline       int8_t* bounds_vx_z(size_t N)       { return (int8_t*)((int8_t*)this+6+6*N); }
-    __forceinline const char* bounds_vx_z(size_t N) const { return (char*)((char*)this+6+6*N); }
+    __forceinline const int8_t* bounds_vx_z(size_t N) const { return (int8_t*)((int8_t*)this+6+6*N); }
     
-    __forceinline       short* bounds_vx_lower(size_t N)       { return (short*)((char*)this+6+7*N); }
+    __forceinline       short* bounds_vx_lower(size_t N)       { return (short*)((int8_t*)this+6+7*N); }
-    __forceinline const short* bounds_vx_lower(size_t N) const { return (short*)((char*)this+6+7*N); }
+    __forceinline const short* bounds_vx_lower(size_t N) const { return (short*)((int8_t*)this+6+7*N); }
     
-    __forceinline       short* bounds_vx_upper(size_t N)       { return (short*)((char*)this+6+9*N); }
+    __forceinline       short* bounds_vx_upper(size_t N)       { return (short*)((int8_t*)this+6+9*N); }
-    __forceinline const short* bounds_vx_upper(size_t N) const { return (short*)((char*)this+6+9*N); }
+    __forceinline const short* bounds_vx_upper(size_t N) const { return (short*)((int8_t*)this+6+9*N); }
     
-    __forceinline       char* bounds_vy_x(size_t N)       { return (char*)((char*)this+6+11*N); }
+    __forceinline       int8_t* bounds_vy_x(size_t N)       { return (int8_t*)((int8_t*)this+6+11*N); }
-    __forceinline const char* bounds_vy_x(size_t N) const { return (char*)((char*)this+6+11*N); }
+    __forceinline const int8_t* bounds_vy_x(size_t N) const { return (int8_t*)((int8_t*)this+6+11*N); }
     
-    __forceinline       char* bounds_vy_y(size_t N)       { return (char*)((char*)this+6+12*N); }
+    __forceinline       int8_t* bounds_vy_y(size_t N)       { return (int8_t*)((int8_t*)this+6+12*N); }
-    __forceinline const char* bounds_vy_y(size_t N) const { return (char*)((char*)this+6+12*N); }
+    __forceinline const int8_t* bounds_vy_y(size_t N) const { return (int8_t*)((int8_t*)this+6+12*N); }
     
-    __forceinline       char* bounds_vy_z(size_t N)       { return (char*)((char*)this+6+13*N); }
+    __forceinline       int8_t* bounds_vy_z(size_t N)       { return (int8_t*)((int8_t*)this+6+13*N); }
-    __forceinline const char* bounds_vy_z(size_t N) const { return (char*)((char*)this+6+13*N); }
+    __forceinline const int8_t* bounds_vy_z(size_t N) const { return (int8_t*)((int8_t*)this+6+13*N); }
     
-    __forceinline       short* bounds_vy_lower(size_t N)       { return (short*)((char*)this+6+14*N); }
+    __forceinline       short* bounds_vy_lower(size_t N)       { return (short*)((int8_t*)this+6+14*N); }
-    __forceinline const short* bounds_vy_lower(size_t N) const { return (short*)((char*)this+6+14*N); }
+    __forceinline const short* bounds_vy_lower(size_t N) const { return (short*)((int8_t*)this+6+14*N); }
     
-    __forceinline       short* bounds_vy_upper(size_t N)       { return (short*)((char*)this+6+16*N); }
+    __forceinline       short* bounds_vy_upper(size_t N)       { return (short*)((int8_t*)this+6+16*N); }
-    __forceinline const short* bounds_vy_upper(size_t N) const { return (short*)((char*)this+6+16*N); }
+    __forceinline const short* bounds_vy_upper(size_t N) const { return (short*)((int8_t*)this+6+16*N); }
     
-    __forceinline       char* bounds_vz_x(size_t N)       { return (char*)((char*)this+6+18*N); }
+    __forceinline       int8_t* bounds_vz_x(size_t N)       { return (int8_t*)((int8_t*)this+6+18*N); }
-    __forceinline const char* bounds_vz_x(size_t N) const { return (char*)((char*)this+6+18*N); }
+    __forceinline const int8_t* bounds_vz_x(size_t N) const { return (int8_t*)((int8_t*)this+6+18*N); }
     
-    __forceinline       char* bounds_vz_y(size_t N)       { return (char*)((char*)this+6+19*N); }
+    __forceinline       int8_t* bounds_vz_y(size_t N)       { return (int8_t*)((int8_t*)this+6+19*N); }
-    __forceinline const char* bounds_vz_y(size_t N) const { return (char*)((char*)this+6+19*N); }
+    __forceinline const int8_t* bounds_vz_y(size_t N) const { return (int8_t*)((int8_t*)this+6+19*N); }
     
-    __forceinline       char* bounds_vz_z(size_t N)       { return (char*)((char*)this+6+20*N); }
+    __forceinline       int8_t* bounds_vz_z(size_t N)       { return (int8_t*)((int8_t*)this+6+20*N); }
-    __forceinline const char* bounds_vz_z(size_t N) const { return (char*)((char*)this+6+20*N); }
+    __forceinline const int8_t* bounds_vz_z(size_t N) const { return (int8_t*)((int8_t*)this+6+20*N); }
     
-    __forceinline       short* bounds_vz_lower(size_t N)       { return (short*)((char*)this+6+21*N); }
+    __forceinline       short* bounds_vz_lower(size_t N)       { return (short*)((int8_t*)this+6+21*N); }
-    __forceinline const short* bounds_vz_lower(size_t N) const { return (short*)((char*)this+6+21*N); }
+    __forceinline const short* bounds_vz_lower(size_t N) const { return (short*)((int8_t*)this+6+21*N); }
     
-    __forceinline       short* bounds_vz_upper(size_t N)       { return (short*)((char*)this+6+23*N); }
+    __forceinline       short* bounds_vz_upper(size_t N)       { return (short*)((int8_t*)this+6+23*N); }
-    __forceinline const short* bounds_vz_upper(size_t N) const { return (short*)((char*)this+6+23*N); }
+    __forceinline const short* bounds_vz_upper(size_t N) const { return (short*)((int8_t*)this+6+23*N); }
     
-    __forceinline       Vec3f* offset(size_t N)       { return (Vec3f*)((char*)this+6+25*N); }
+    __forceinline       Vec3f* offset(size_t N)       { return (Vec3f*)((int8_t*)this+6+25*N); }
-    __forceinline const Vec3f* offset(size_t N) const { return (Vec3f*)((char*)this+6+25*N); }
+    __forceinline const Vec3f* offset(size_t N) const { return (Vec3f*)((int8_t*)this+6+25*N); }
     
-    __forceinline       float* scale(size_t N)       { return (float*)((char*)this+6+25*N+12); }
+    __forceinline       float* scale(size_t N)       { return (float*)((int8_t*)this+6+25*N+12); }
-    __forceinline const float* scale(size_t N) const { return (float*)((char*)this+6+25*N+12); }
+    __forceinline const float* scale(size_t N) const { return (float*)((int8_t*)this+6+25*N+12); }
 
-    __forceinline       char* end(size_t N)       { return (char*)this+6+25*N+16; }
+    __forceinline       int8_t* end(size_t N)       { return (int8_t*)this+6+25*N+16; }
-    __forceinline const char* end(size_t N) const { return (char*)this+6+25*N+16; }
+    __forceinline const int8_t* end(size_t N) const { return (int8_t*)this+6+25*N+16; }
   };
 
   template<int M>

thirdparty/embree/kernels/geometry/curveNi_mb.h

@@ -43,10 +43,10 @@ namespace embree
     __forceinline LBBox3fa fillMB(const PrimRefMB* prims, size_t& begin, size_t _end, Scene* scene, const BBox1f time_range)
     {
       size_t end = min(begin+M,_end);
-      N = (unsigned char)(end-begin);
+      N = (uint8_t)(end-begin);
       const unsigned int geomID0 = prims[begin].geomID();
       this->geomID(N) = geomID0;
-      ty = (unsigned char) scene->get(geomID0)->getType();
+      ty = (uint8_t) scene->get(geomID0)->getType();
 
       /* encode all primitives */
       LBBox3fa lbounds = empty;
@@ -79,10 +79,10 @@ namespace embree
         const LinearSpace3fa space3(trunc(126.0f*space2.vx),trunc(126.0f*space2.vy),trunc(126.0f*space2.vz));
         const LBBox3fa bounds = scene->get(geomID)->vlinearBounds(loffset,lscale,max(length(space3.vx),length(space3.vy),length(space3.vz)),space3.transposed(),primID,time_range);
         
-        // NOTE: this weird (char) (short) cast works around VS2015 Win32 compiler bug
+        // NOTE: this weird (int8_t) (short) cast works around VS2015 Win32 compiler bug
-        bounds_vx_x(N)[i] = (char) (short) space3.vx.x;
+        bounds_vx_x(N)[i] = (int8_t) (short) space3.vx.x;
-        bounds_vx_y(N)[i] = (char) (short) space3.vx.y;
+        bounds_vx_y(N)[i] = (int8_t) (short) space3.vx.y;
-        bounds_vx_z(N)[i] = (char) (short) space3.vx.z;
+        bounds_vx_z(N)[i] = (int8_t) (short) space3.vx.z;
         bounds_vx_lower0(N)[i] = (short) clamp(floor(bounds.bounds0.lower.x),-32767.0f,32767.0f);
         bounds_vx_upper0(N)[i] = (short) clamp(ceil (bounds.bounds0.upper.x),-32767.0f,32767.0f);
         bounds_vx_lower1(N)[i] = (short) clamp(floor(bounds.bounds1.lower.x),-32767.0f,32767.0f);
@@ -92,9 +92,9 @@ namespace embree
         assert(-32767.0f <= floor(bounds.bounds1.lower.x) && floor(bounds.bounds1.lower.x) <= 32767.0f);
         assert(-32767.0f <= ceil (bounds.bounds1.upper.x) && ceil (bounds.bounds1.upper.x) <= 32767.0f);
         
-        bounds_vy_x(N)[i] = (char) (short) space3.vy.x;
+        bounds_vy_x(N)[i] = (int8_t) (short) space3.vy.x;
-        bounds_vy_y(N)[i] = (char) (short) space3.vy.y;
+        bounds_vy_y(N)[i] = (int8_t) (short) space3.vy.y;
-        bounds_vy_z(N)[i] = (char) (short) space3.vy.z;
+        bounds_vy_z(N)[i] = (int8_t) (short) space3.vy.z;
         bounds_vy_lower0(N)[i] = (short) clamp(floor(bounds.bounds0.lower.y),-32767.0f,32767.0f);
         bounds_vy_upper0(N)[i] = (short) clamp(ceil (bounds.bounds0.upper.y),-32767.0f,32767.0f);
         bounds_vy_lower1(N)[i] = (short) clamp(floor(bounds.bounds1.lower.y),-32767.0f,32767.0f);
@@ -104,9 +104,9 @@ namespace embree
         assert(-32767.0f <= floor(bounds.bounds1.lower.y) && floor(bounds.bounds1.lower.y) <= 32767.0f);
         assert(-32767.0f <= ceil (bounds.bounds1.upper.y) && ceil (bounds.bounds1.upper.y) <= 32767.0f);
 
-        bounds_vz_x(N)[i] = (char) (short) space3.vz.x;
+        bounds_vz_x(N)[i] = (int8_t) (short) space3.vz.x;
-        bounds_vz_y(N)[i] = (char) (short) space3.vz.y;
+        bounds_vz_y(N)[i] = (int8_t) (short) space3.vz.y;
-        bounds_vz_z(N)[i] = (char) (short) space3.vz.z;
+        bounds_vz_z(N)[i] = (int8_t) (short) space3.vz.z;
         bounds_vz_lower0(N)[i] = (short) clamp(floor(bounds.bounds0.lower.z),-32767.0f,32767.0f);
         bounds_vz_upper0(N)[i] = (short) clamp(ceil (bounds.bounds0.upper.z),-32767.0f,32767.0f);
         bounds_vz_lower1(N)[i] = (short) clamp(floor(bounds.bounds1.lower.z),-32767.0f,32767.0f);
@@ -130,7 +130,7 @@ namespace embree
       size_t items = CurveNiMB::blocks(prims.size());
       size_t numbytes = CurveNiMB::bytes(prims.size());
       CurveNiMB* accel = (CurveNiMB*) alloc.malloc1(numbytes,BVH::byteAlignment);
-      const typename BVH::NodeRef node = bvh->encodeLeaf((char*)accel,items);
+      const typename BVH::NodeRef node = bvh->encodeLeaf((int8_t*)accel,items);
       
       LBBox3fa bounds = empty;
       for (size_t i=0; i<items; i++)
@@ -143,9 +143,9 @@ namespace embree
   public:
     
     // 27.6 - 46 bytes per primitive
-    unsigned char ty;
+    uint8_t ty;
-    unsigned char N;
+    uint8_t N;
-    unsigned char data[4+37*M+24];
+    uint8_t data[4+37*M+24];
 
     /*
     struct Layout
@@ -153,25 +153,25 @@ namespace embree
       unsigned int geomID;
       unsigned int primID[N];
       
-      char bounds_vx_x[N];
+      int8_t bounds_vx_x[N];
-      char bounds_vx_y[N];
+      int8_t bounds_vx_y[N];
-      char bounds_vx_z[N];
+      int8_t bounds_vx_z[N];
       short bounds_vx_lower0[N];
       short bounds_vx_upper0[N];
       short bounds_vx_lower1[N];
       short bounds_vx_upper1[N];
       
-      char bounds_vy_x[N];
+      int8_t bounds_vy_x[N];
-      char bounds_vy_y[N];
+      int8_t bounds_vy_y[N];
-      char bounds_vy_z[N];
+      int8_t bounds_vy_z[N];
       short bounds_vy_lower0[N];
       short bounds_vy_upper0[N];
       short bounds_vy_lower1[N];
       short bounds_vy_upper1[N];
       
-      char bounds_vz_x[N];
+      int8_t bounds_vz_x[N];
-      char bounds_vz_y[N];
+      int8_t bounds_vz_y[N];
-      char bounds_vz_z[N];
+      int8_t bounds_vz_z[N];
       short bounds_vz_lower0[N];
       short bounds_vz_upper0[N];
       short bounds_vz_lower1[N];
@@ -185,89 +185,89 @@ namespace embree
     };
     */
     
-    __forceinline       unsigned int& geomID(size_t N)       { return *(unsigned int*)((char*)this+2); }
+    __forceinline       unsigned int& geomID(size_t N)       { return *(unsigned int*)((int8_t*)this+2); }
-    __forceinline const unsigned int& geomID(size_t N) const { return *(unsigned int*)((char*)this+2); }
+    __forceinline const unsigned int& geomID(size_t N) const { return *(unsigned int*)((int8_t*)this+2); }
     
-    __forceinline       unsigned int* primID(size_t N)       { return (unsigned int*)((char*)this+6); }
+    __forceinline       unsigned int* primID(size_t N)       { return (unsigned int*)((int8_t*)this+6); }
-    __forceinline const unsigned int* primID(size_t N) const { return (unsigned int*)((char*)this+6); }
+    __forceinline const unsigned int* primID(size_t N) const { return (unsigned int*)((int8_t*)this+6); }
     
-    __forceinline       char* bounds_vx_x(size_t N)       { return (char*)((char*)this+6+4*N); }
+    __forceinline       int8_t* bounds_vx_x(size_t N)       { return (int8_t*)((int8_t*)this+6+4*N); }
-    __forceinline const char* bounds_vx_x(size_t N) const { return (char*)((char*)this+6+4*N); }
+    __forceinline const int8_t* bounds_vx_x(size_t N) const { return (int8_t*)((int8_t*)this+6+4*N); }
     
-    __forceinline       char* bounds_vx_y(size_t N)       { return (char*)((char*)this+6+5*N); }
+    __forceinline       int8_t* bounds_vx_y(size_t N)       { return (int8_t*)((int8_t*)this+6+5*N); }
-    __forceinline const char* bounds_vx_y(size_t N) const { return (char*)((char*)this+6+5*N); }
+    __forceinline const int8_t* bounds_vx_y(size_t N) const { return (int8_t*)((int8_t*)this+6+5*N); }
     
-    __forceinline       char* bounds_vx_z(size_t N)       { return (char*)((char*)this+6+6*N); }
+    __forceinline       int8_t* bounds_vx_z(size_t N)       { return (int8_t*)((int8_t*)this+6+6*N); }
-    __forceinline const char* bounds_vx_z(size_t N) const { return (char*)((char*)this+6+6*N); }
+    __forceinline const int8_t* bounds_vx_z(size_t N) const { return (int8_t*)((int8_t*)this+6+6*N); }
     
-    __forceinline       short* bounds_vx_lower0(size_t N)       { return (short*)((char*)this+6+7*N); }
+    __forceinline       short* bounds_vx_lower0(size_t N)       { return (short*)((int8_t*)this+6+7*N); }
-    __forceinline const short* bounds_vx_lower0(size_t N) const { return (short*)((char*)this+6+7*N); }
+    __forceinline const short* bounds_vx_lower0(size_t N) const { return (short*)((int8_t*)this+6+7*N); }
     
-    __forceinline       short* bounds_vx_upper0(size_t N)       { return (short*)((char*)this+6+9*N); }
+    __forceinline       short* bounds_vx_upper0(size_t N)       { return (short*)((int8_t*)this+6+9*N); }
-    __forceinline const short* bounds_vx_upper0(size_t N) const { return (short*)((char*)this+6+9*N); }
+    __forceinline const short* bounds_vx_upper0(size_t N) const { return (short*)((int8_t*)this+6+9*N); }
 
-    __forceinline       short* bounds_vx_lower1(size_t N)       { return (short*)((char*)this+6+11*N); }
+    __forceinline       short* bounds_vx_lower1(size_t N)       { return (short*)((int8_t*)this+6+11*N); }
-    __forceinline const short* bounds_vx_lower1(size_t N) const { return (short*)((char*)this+6+11*N); }
+    __forceinline const short* bounds_vx_lower1(size_t N) const { return (short*)((int8_t*)this+6+11*N); }
     
-    __forceinline       short* bounds_vx_upper1(size_t N)       { return (short*)((char*)this+6+13*N); }
+    __forceinline       short* bounds_vx_upper1(size_t N)       { return (short*)((int8_t*)this+6+13*N); }
-    __forceinline const short* bounds_vx_upper1(size_t N) const { return (short*)((char*)this+6+13*N); }
+    __forceinline const short* bounds_vx_upper1(size_t N) const { return (short*)((int8_t*)this+6+13*N); }
 
-    __forceinline       char* bounds_vy_x(size_t N)       { return (char*)((char*)this+6+15*N); }
+    __forceinline       int8_t* bounds_vy_x(size_t N)       { return (int8_t*)((int8_t*)this+6+15*N); }
-    __forceinline const char* bounds_vy_x(size_t N) const { return (char*)((char*)this+6+15*N); }
+    __forceinline const int8_t* bounds_vy_x(size_t N) const { return (int8_t*)((int8_t*)this+6+15*N); }
     
-    __forceinline       char* bounds_vy_y(size_t N)       { return (char*)((char*)this+6+16*N); }
+    __forceinline       int8_t* bounds_vy_y(size_t N)       { return (int8_t*)((int8_t*)this+6+16*N); }
-    __forceinline const char* bounds_vy_y(size_t N) const { return (char*)((char*)this+6+16*N); }
+    __forceinline const int8_t* bounds_vy_y(size_t N) const { return (int8_t*)((int8_t*)this+6+16*N); }
     
-    __forceinline       char* bounds_vy_z(size_t N)       { return (char*)((char*)this+6+17*N); }
+    __forceinline       int8_t* bounds_vy_z(size_t N)       { return (int8_t*)((int8_t*)this+6+17*N); }
-    __forceinline const char* bounds_vy_z(size_t N) const { return (char*)((char*)this+6+17*N); }
+    __forceinline const int8_t* bounds_vy_z(size_t N) const { return (int8_t*)((int8_t*)this+6+17*N); }
     
-    __forceinline       short* bounds_vy_lower0(size_t N)       { return (short*)((char*)this+6+18*N); }
+    __forceinline       short* bounds_vy_lower0(size_t N)       { return (short*)((int8_t*)this+6+18*N); }
-    __forceinline const short* bounds_vy_lower0(size_t N) const { return (short*)((char*)this+6+18*N); }
+    __forceinline const short* bounds_vy_lower0(size_t N) const { return (short*)((int8_t*)this+6+18*N); }
     
-    __forceinline       short* bounds_vy_upper0(size_t N)       { return (short*)((char*)this+6+20*N); }
+    __forceinline       short* bounds_vy_upper0(size_t N)       { return (short*)((int8_t*)this+6+20*N); }
-    __forceinline const short* bounds_vy_upper0(size_t N) const { return (short*)((char*)this+6+20*N); }
+    __forceinline const short* bounds_vy_upper0(size_t N) const { return (short*)((int8_t*)this+6+20*N); }
 
-    __forceinline       short* bounds_vy_lower1(size_t N)       { return (short*)((char*)this+6+22*N); }
+    __forceinline       short* bounds_vy_lower1(size_t N)       { return (short*)((int8_t*)this+6+22*N); }
-    __forceinline const short* bounds_vy_lower1(size_t N) const { return (short*)((char*)this+6+22*N); }
+    __forceinline const short* bounds_vy_lower1(size_t N) const { return (short*)((int8_t*)this+6+22*N); }
     
-    __forceinline       short* bounds_vy_upper1(size_t N)       { return (short*)((char*)this+6+24*N); }
+    __forceinline       short* bounds_vy_upper1(size_t N)       { return (short*)((int8_t*)this+6+24*N); }
-    __forceinline const short* bounds_vy_upper1(size_t N) const { return (short*)((char*)this+6+24*N); }
+    __forceinline const short* bounds_vy_upper1(size_t N) const { return (short*)((int8_t*)this+6+24*N); }
     
-    __forceinline       char* bounds_vz_x(size_t N)       { return (char*)((char*)this+6+26*N); }
+    __forceinline       int8_t* bounds_vz_x(size_t N)       { return (int8_t*)((int8_t*)this+6+26*N); }
-    __forceinline const char* bounds_vz_x(size_t N) const { return (char*)((char*)this+6+26*N); }
+    __forceinline const int8_t* bounds_vz_x(size_t N) const { return (int8_t*)((int8_t*)this+6+26*N); }
     
-    __forceinline       char* bounds_vz_y(size_t N)       { return (char*)((char*)this+6+27*N); }
+    __forceinline       int8_t* bounds_vz_y(size_t N)       { return (int8_t*)((int8_t*)this+6+27*N); }
-    __forceinline const char* bounds_vz_y(size_t N) const { return (char*)((char*)this+6+27*N); }
+    __forceinline const int8_t* bounds_vz_y(size_t N) const { return (int8_t*)((int8_t*)this+6+27*N); }
     
-    __forceinline       char* bounds_vz_z(size_t N)       { return (char*)((char*)this+6+28*N); }
+    __forceinline       int8_t* bounds_vz_z(size_t N)       { return (int8_t*)((int8_t*)this+6+28*N); }
-    __forceinline const char* bounds_vz_z(size_t N) const { return (char*)((char*)this+6+28*N); }
+    __forceinline const int8_t* bounds_vz_z(size_t N) const { return (int8_t*)((int8_t*)this+6+28*N); }
     
-    __forceinline       short* bounds_vz_lower0(size_t N)       { return (short*)((char*)this+6+29*N); }
+    __forceinline       short* bounds_vz_lower0(size_t N)       { return (short*)((int8_t*)this+6+29*N); }
-    __forceinline const short* bounds_vz_lower0(size_t N) const { return (short*)((char*)this+6+29*N); }
+    __forceinline const short* bounds_vz_lower0(size_t N) const { return (short*)((int8_t*)this+6+29*N); }
     
-    __forceinline       short* bounds_vz_upper0(size_t N)       { return (short*)((char*)this+6+31*N); }
+    __forceinline       short* bounds_vz_upper0(size_t N)       { return (short*)((int8_t*)this+6+31*N); }
-    __forceinline const short* bounds_vz_upper0(size_t N) const { return (short*)((char*)this+6+31*N); }
+    __forceinline const short* bounds_vz_upper0(size_t N) const { return (short*)((int8_t*)this+6+31*N); }
 
-    __forceinline       short* bounds_vz_lower1(size_t N)       { return (short*)((char*)this+6+33*N); }
+    __forceinline       short* bounds_vz_lower1(size_t N)       { return (short*)((int8_t*)this+6+33*N); }
-    __forceinline const short* bounds_vz_lower1(size_t N) const { return (short*)((char*)this+6+33*N); }
+    __forceinline const short* bounds_vz_lower1(size_t N) const { return (short*)((int8_t*)this+6+33*N); }
     
-    __forceinline       short* bounds_vz_upper1(size_t N)       { return (short*)((char*)this+6+35*N); }
+    __forceinline       short* bounds_vz_upper1(size_t N)       { return (short*)((int8_t*)this+6+35*N); }
-    __forceinline const short* bounds_vz_upper1(size_t N) const { return (short*)((char*)this+6+35*N); }
+    __forceinline const short* bounds_vz_upper1(size_t N) const { return (short*)((int8_t*)this+6+35*N); }
 
-    __forceinline       Vec3f* offset(size_t N)       { return (Vec3f*)((char*)this+6+37*N); }
+    __forceinline       Vec3f* offset(size_t N)       { return (Vec3f*)((int8_t*)this+6+37*N); }
-    __forceinline const Vec3f* offset(size_t N) const { return (Vec3f*)((char*)this+6+37*N); }
+    __forceinline const Vec3f* offset(size_t N) const { return (Vec3f*)((int8_t*)this+6+37*N); }
     
-    __forceinline       float* scale(size_t N)       { return (float*)((char*)this+6+37*N+12); }
+    __forceinline       float* scale(size_t N)       { return (float*)((int8_t*)this+6+37*N+12); }
-    __forceinline const float* scale(size_t N) const { return (float*)((char*)this+6+37*N+12); }
+    __forceinline const float* scale(size_t N) const { return (float*)((int8_t*)this+6+37*N+12); }
 
-    __forceinline       float& time_offset(size_t N)       { return *(float*)((char*)this+6+37*N+16); }
+    __forceinline       float& time_offset(size_t N)       { return *(float*)((int8_t*)this+6+37*N+16); }
-    __forceinline const float& time_offset(size_t N) const { return *(float*)((char*)this+6+37*N+16); }
+    __forceinline const float& time_offset(size_t N) const { return *(float*)((int8_t*)this+6+37*N+16); }
     
-    __forceinline       float& time_scale(size_t N)       { return *(float*)((char*)this+6+37*N+20); }
+    __forceinline       float& time_scale(size_t N)       { return *(float*)((int8_t*)this+6+37*N+20); }
-    __forceinline const float& time_scale(size_t N) const { return *(float*)((char*)this+6+37*N+20); }
+    __forceinline const float& time_scale(size_t N) const { return *(float*)((int8_t*)this+6+37*N+20); }
 
-    __forceinline       char* end(size_t N)       { return (char*)this+6+37*N+24; }
+    __forceinline       int8_t* end(size_t N)       { return (int8_t*)this+6+37*N+24; }
-    __forceinline const char* end(size_t N) const { return (char*)this+6+37*N+24; }
+    __forceinline const int8_t* end(size_t N) const { return (int8_t*)this+6+37*N+24; }
   };
 
   template<int M>

thirdparty/embree/kernels/geometry/curve_intersector_virtual_bezier_curve.h

@@ -0,0 +1,21 @@
+// Copyright 2020 Light Transport Entertainment Inc.
+// SPDX-License-Identifier: Apache-2.0
+
+#pragma once
+
+#include "curve_intersector_virtual.h"
+
+namespace embree
+{
+  namespace isa
+  {
+    void AddVirtualCurveBezierCurveInterector4i(VirtualCurveIntersector &prim);
+    void AddVirtualCurveBezierCurveInterector4v(VirtualCurveIntersector &prim);
+    void AddVirtualCurveBezierCurveInterector4iMB(VirtualCurveIntersector &prim);
+#if defined(__AVX__)
+    void AddVirtualCurveBezierCurveInterector8i(VirtualCurveIntersector &prim);
+    void AddVirtualCurveBezierCurveInterector8v(VirtualCurveIntersector &prim);
+    void AddVirtualCurveBezierCurveInterector8iMB(VirtualCurveIntersector &prim);
+#endif
+  }
+}

thirdparty/embree/kernels/geometry/curve_intersector_virtual_bspline_curve.h

@@ -0,0 +1,21 @@
+// Copyright 2020 Light Transport Entertainment Inc.
+// SPDX-License-Identifier: Apache-2.0
+
+#pragma once
+
+#include "curve_intersector_virtual.h"
+
+namespace embree
+{
+  namespace isa
+  {
+    void AddVirtualCurveBSplineCurveInterector4i(VirtualCurveIntersector &prim);
+    void AddVirtualCurveBSplineCurveInterector4v(VirtualCurveIntersector &prim);
+    void AddVirtualCurveBSplineCurveInterector4iMB(VirtualCurveIntersector &prim);
+#if defined(__AVX__)
+    void AddVirtualCurveBSplineCurveInterector8i(VirtualCurveIntersector &prim);
+    void AddVirtualCurveBSplineCurveInterector8v(VirtualCurveIntersector &prim);
+    void AddVirtualCurveBSplineCurveInterector8iMB(VirtualCurveIntersector &prim);
+#endif
+  }
+}

thirdparty/embree/kernels/geometry/curve_intersector_virtual_catmullrom_curve.h

@@ -0,0 +1,21 @@
+// Copyright 2020 Light Transport Entertainment Inc.
+// SPDX-License-Identifier: Apache-2.0
+
+#pragma once
+
+#include "curve_intersector_virtual.h"
+
+namespace embree
+{
+  namespace isa
+  {
+    void AddVirtualCurveCatmullRomCurveInterector4i(VirtualCurveIntersector &prim);
+    void AddVirtualCurveCatmullRomCurveInterector4v(VirtualCurveIntersector &prim);
+    void AddVirtualCurveCatmullRomCurveInterector4iMB(VirtualCurveIntersector &prim);
+#if defined(__AVX__)
+    void AddVirtualCurveCatmullRomCurveInterector8i(VirtualCurveIntersector &prim);
+    void AddVirtualCurveCatmullRomCurveInterector8v(VirtualCurveIntersector &prim);
+    void AddVirtualCurveCatmullRomCurveInterector8iMB(VirtualCurveIntersector &prim);
+#endif
+  }
+}

thirdparty/embree/kernels/geometry/curve_intersector_virtual_hermite_curve.h

@@ -0,0 +1,21 @@
+// Copyright 2020 Light Transport Entertainment Inc.
+// SPDX-License-Identifier: Apache-2.0
+
+#pragma once
+
+#include "curve_intersector_virtual.h"
+
+namespace embree
+{
+  namespace isa
+  {
+    void AddVirtualCurveHermiteCurveInterector4i(VirtualCurveIntersector &prim);
+    void AddVirtualCurveHermiteCurveInterector4v(VirtualCurveIntersector &prim);
+    void AddVirtualCurveHermiteCurveInterector4iMB(VirtualCurveIntersector &prim);
+#if defined(__AVX__)
+    void AddVirtualCurveHermiteCurveInterector8i(VirtualCurveIntersector &prim);
+    void AddVirtualCurveHermiteCurveInterector8v(VirtualCurveIntersector &prim);
+    void AddVirtualCurveHermiteCurveInterector8iMB(VirtualCurveIntersector &prim);
+#endif
+  }
+}

thirdparty/embree/kernels/geometry/curve_intersector_virtual_linear_curve.h

@@ -0,0 +1,21 @@
+// Copyright 2020 Light Transport Entertainment Inc.
+// SPDX-License-Identifier: Apache-2.0
+
+#pragma once
+
+#include "curve_intersector_virtual.h"
+
+namespace embree
+{
+  namespace isa
+  {
+    void AddVirtualCurveLinearCurveInterector4i(VirtualCurveIntersector &prim);
+    void AddVirtualCurveLinearCurveInterector4v(VirtualCurveIntersector &prim);
+    void AddVirtualCurveLinearCurveInterector4iMB(VirtualCurveIntersector &prim);
+#if defined(__AVX__)
+    void AddVirtualCurveLinearCurveInterector8i(VirtualCurveIntersector &prim);
+    void AddVirtualCurveLinearCurveInterector8v(VirtualCurveIntersector &prim);
+    void AddVirtualCurveLinearCurveInterector8iMB(VirtualCurveIntersector &prim);
+#endif
+  }
+}

thirdparty/embree/kernels/geometry/curve_intersector_virtual_point.h

@@ -0,0 +1,22 @@
+// Copyright 2020 Light Transport Entertainment Inc.
+// SPDX-License-Identifier: Apache-2.0
+
+#pragma once
+
+#include "curve_intersector_virtual.h"
+
+namespace embree
+{
+  namespace isa
+  {
+    void AddVirtualCurvePointInterector4i(VirtualCurveIntersector &prim);
+    void AddVirtualCurvePointInterector4v(VirtualCurveIntersector &prim);
+    void AddVirtualCurvePointInterector4iMB(VirtualCurveIntersector &prim);
+
+#if defined (__AVX__)
+    void AddVirtualCurvePointInterector8i(VirtualCurveIntersector &prim);
+    void AddVirtualCurvePointInterector8v(VirtualCurveIntersector &prim);
+    void AddVirtualCurvePointInterector8iMB(VirtualCurveIntersector &prim);
+#endif
+  }
+}

thirdparty/embree/kernels/geometry/grid_soa.h

@@ -41,7 +41,7 @@ namespace embree
         }
         const size_t gridBytes = 4*size_t(width)*size_t(height)*sizeof(float);  
         size_t rootBytes = time_steps*sizeof(BVH4::NodeRef);
-#if !defined(__X86_64__)
+#if !defined(__X86_64__) && !defined(__aarch64__)
         rootBytes += 4; // We read 2 elements behind the grid. As we store at least 8 root bytes after the grid we are fine in 64 bit mode. But in 32 bit mode we have to do additional padding.
 #endif
         void* data = alloc(offsetof(GridSOA,data)+bvhBytes+time_steps*gridBytes+rootBytes);
@@ -62,8 +62,8 @@ namespace embree
       __forceinline const BVH4::NodeRef& root(size_t t = 0) const { return (BVH4::NodeRef&)data[rootOffset + t*sizeof(BVH4::NodeRef)]; }
 
       /*! returns pointer to BVH array */
-      __forceinline       char* bvhData()       { return &data[0]; }
+      __forceinline       int8_t* bvhData()       { return &data[0]; }
-      __forceinline const char* bvhData() const { return &data[0]; }
+      __forceinline const int8_t* bvhData() const { return &data[0]; }
 
       /*! returns pointer to Grid array */
       __forceinline       float* gridData(size_t t = 0)       { return (float*) &data[gridOffset + t*gridBytes]; }
@@ -253,7 +253,7 @@ namespace embree
 
     public:
       BVH4::NodeRef troot;
-#if !defined(__X86_64__)
+#if !defined(__X86_64__) && !defined(__aarch64__)
       unsigned align1;
 #endif
       unsigned time_steps;
@@ -269,7 +269,7 @@ namespace embree
       unsigned gridBytes;
       unsigned rootOffset;
 
-      char data[1];        //!< after the struct we first store the BVH, then the grid, and finally the roots
+      int8_t data[1];      //!< after the struct we first store the BVH, then the grid, and finally the roots
     };
   }
 }

thirdparty/embree/kernels/hash.h

@@ -2,4 +2,4 @@
 // Copyright 2009-2020 Intel Corporation
 // SPDX-License-Identifier: Apache-2.0
 
-#define RTC_HASH "69bd4c272f1ed608494f233ecfff3feec516880b"
+#define RTC_HASH "6ef362f99af80c9dfe8dd2bfc582d9067897edc6"

thirdparty/embree/kernels/subdiv/tessellation_cache.h

@@ -63,7 +63,7 @@ namespace embree
    static const size_t NUM_CACHE_SEGMENTS              = 8;
    static const size_t NUM_PREALLOC_THREAD_WORK_STATES = 512;
    static const size_t COMMIT_INDEX_SHIFT              = 32+8;
-#if defined(__X86_64__)
+#if defined(__X86_64__) || defined(__aarch64__)
    static const size_t REF_TAG_MASK                    = 0xffffffffff;
 #else
    static const size_t REF_TAG_MASK                    = 0x7FFFFFFF;

thirdparty/embree/patches/godot-changes.patch

@@ -1,215 +1,630 @@
-diff --git a/common/math/math.h b/common/math/math.h
+diff --git a/thirdparty/embree/common/algorithms/parallel_for.h b/thirdparty/embree/common/algorithms/parallel_for.h
-index 5af0691a2..1982c27c1 100644
+index 76c6b740aa..51d296fb16 100644
---- a/common/math/math.h
+--- a/thirdparty/embree/common/algorithms/parallel_for.h
-+++ b/common/math/math.h
++++ b/thirdparty/embree/common/algorithms/parallel_for.h
-@@ -13,7 +13,7 @@
+@@ -27,7 +27,10 @@ namespace embree
- #include <immintrin.h>
+           func(r.begin());
+         });
+       if (!TaskScheduler::wait())
+-        throw std::runtime_error("task cancelled");
++        // -- GODOT start --
++        // throw std::runtime_error("task cancelled");
++        abort(); 
++        // -- GODOT end --
+     }
+ #elif defined(TASKING_GCD) && defined(BUILD_IOS)
+       
+@@ -55,13 +58,19 @@ namespace embree
+         func(i);
+       },context);
+     if (context.is_group_execution_cancelled())
+-      throw std::runtime_error("task cancelled");
++      // -- GODOT start --
++      // throw std::runtime_error("task cancelled");
++      abort(); 
++      // -- GODOT end --
+   #else
+     tbb::parallel_for(Index(0),N,Index(1),[&](Index i) {
+         func(i);
+       });
+     if (tbb::task::self().is_cancelled())
+-      throw std::runtime_error("task cancelled");
++      // -- GODOT start --
++      // throw std::runtime_error("task cancelled");
++      abort(); 
++      // -- GODOT end --
+   #endif
  
- #if defined(__WIN32__)
+ #elif defined(TASKING_PPL)
--#if (__MSV_VER <= 1700)
+@@ -81,7 +90,10 @@ namespace embree
-+#if defined(_MSC_VER) && (_MSC_VER <= 1700)
+ #if defined(TASKING_INTERNAL)
- namespace std
+     TaskScheduler::spawn(first,last,minStepSize,func);
- {
+     if (!TaskScheduler::wait())
-   __forceinline bool isinf ( const float x ) { return _finite(x) == 0; }
+-      throw std::runtime_error("task cancelled");
-@@ -86,7 +86,7 @@
++      // -- GODOT start --
-     return _mm_cvtss_f32(c);
++      // throw std::runtime_error("task cancelled");
++      abort(); 
++      // -- GODOT end --
+ 
+ #elif defined(TASKING_GCD) && defined(BUILD_IOS)
+       
+@@ -109,13 +121,19 @@ namespace embree
+         func(range<Index>(r.begin(),r.end()));
+       },context);
+     if (context.is_group_execution_cancelled())
+-      throw std::runtime_error("task cancelled");
++      // -- GODOT start --
++      // throw std::runtime_error("task cancelled");
++      abort(); 
++      // -- GODOT end --
+   #else
+     tbb::parallel_for(tbb::blocked_range<Index>(first,last,minStepSize),[&](const tbb::blocked_range<Index>& r) {
+         func(range<Index>(r.begin(),r.end()));
+       });
+     if (tbb::task::self().is_cancelled())
+-      throw std::runtime_error("task cancelled");
++      // -- GODOT start --
++      // throw std::runtime_error("task cancelled");
++      abort(); 
++      // -- GODOT end --
+   #endif
+ 
+ #elif defined(TASKING_PPL)
+@@ -147,13 +165,19 @@ namespace embree
+           func(i);
+         },tbb::simple_partitioner(),context);
+       if (context.is_group_execution_cancelled())
+-        throw std::runtime_error("task cancelled");
++        // -- GODOT start --
++        // throw std::runtime_error("task cancelled");
++        abort(); 
++        // -- GODOT end --
+     #else
+       tbb::parallel_for(Index(0),N,Index(1),[&](Index i) {
+           func(i);
+         },tbb::simple_partitioner());
+       if (tbb::task::self().is_cancelled())
+-        throw std::runtime_error("task cancelled");
++        // -- GODOT start --
++        // throw std::runtime_error("task cancelled");
++        abort(); 
++        // -- GODOT end --
+     #endif
    }
  
--#if defined(__WIN32__) && (__MSC_VER <= 1700)
+@@ -168,13 +192,19 @@ namespace embree
-+#if defined(__WIN32__) && defined(_MSC_VER) && (_MSC_VER <= 1700)
+           func(i);
-   __forceinline float nextafter(float x, float y) { if ((x<y) == (x>0)) return x*(1.1f+float(ulp)); else return x*(0.9f-float(ulp)); }
+         },ap,context);
-   __forceinline double nextafter(double x, double y) { return _nextafter(x, y); }
+       if (context.is_group_execution_cancelled())
-   __forceinline int roundf(float f) { return (int)(f + 0.5f); }
+-        throw std::runtime_error("task cancelled");
-diff --git a/common/sys/intrinsics.h b/common/sys/intrinsics.h
++       // -- GODOT start --
-index 3f0619cac..58f5c3bb4 100644
++       // throw std::runtime_error("task cancelled");
---- a/common/sys/intrinsics.h
++       abort(); 
-+++ b/common/sys/intrinsics.h
++       // -- GODOT end --
-@@ -11,6 +11,12 @@
+     #else
- 
+       tbb::parallel_for(Index(0),N,Index(1),[&](Index i) {
- #include <immintrin.h>
+           func(i);
- 
+         },ap);
-+// -- GODOT start --
+       if (tbb::task::self().is_cancelled())
-+#if defined(__WIN32__) && defined(__MINGW32__)
+-        throw std::runtime_error("task cancelled");
-+#include <unistd.h>
++        // -- GODOT start --
-+#endif
++        // throw std::runtime_error("task cancelled");
-+// -- GODOT end --
++        abort(); 
-+
++        // -- GODOT end --
- #if defined(__BMI__) && defined(__GNUC__) && !defined(__INTEL_COMPILER)
+     #endif
-   #if !defined(_tzcnt_u32)
-     #define _tzcnt_u32 __tzcnt_u32
-@@ -30,8 +36,14 @@
- #endif
- 
- #if defined(__WIN32__)
--#  define NOMINMAX
--#  include <windows.h>
-+// -- GODOT start --
-+#if !defined(NOMINMAX)
-+// -- GODOT end --
-+#define NOMINMAX
-+// -- GODOT start --
-+#endif
-+#include "windows.h"
-+// -- GODOT end --
- #endif
- 
- /* normally defined in pmmintrin.h, but we always need this */
-@@ -413,8 +425,16 @@ namespace embree
-   
-   __forceinline void pause_cpu(const size_t N = 8)
-   {
-+// -- GODOT start --
-     for (size_t i=0; i<N; i++)
-+#if !(defined(__WIN32__) && defined(__MINGW32__))
-+// -- GODOT end --
-       _mm_pause();    
-+// -- GODOT start --
-+#else
-+      usleep(1);
-+#endif
-+// -- GODOT end --
    }
-   
-   /* prefetches */
-diff --git a/common/sys/library.cpp b/common/sys/library.cpp
-index e448b195d..8ec918660 100644
---- a/common/sys/library.cpp
-+++ b/common/sys/library.cpp
-@@ -27,7 +27,9 @@ namespace embree
  
-   /* returns address of a symbol from the library */
+diff --git a/thirdparty/embree/common/algorithms/parallel_reduce.h b/thirdparty/embree/common/algorithms/parallel_reduce.h
-   void* getSymbol(lib_t lib, const std::string& sym) {
+index d444b6a2e4..0daf94e50e 100644
--    return GetProcAddress(HMODULE(lib),sym.c_str());
+--- a/thirdparty/embree/common/algorithms/parallel_reduce.h
++++ b/thirdparty/embree/common/algorithms/parallel_reduce.h
+@@ -58,15 +58,19 @@ namespace embree
+     const Value v = tbb::parallel_reduce(tbb::blocked_range<Index>(first,last,minStepSize),identity,
+       [&](const tbb::blocked_range<Index>& r, const Value& start) { return reduction(start,func(range<Index>(r.begin(),r.end()))); },
+       reduction,context);
+-    if (context.is_group_execution_cancelled())
+-      throw std::runtime_error("task cancelled");
 +    // -- GODOT start --
-+    return (void*) GetProcAddress(HMODULE(lib),sym.c_str());
++    // if (context.is_group_execution_cancelled())
++    //   throw std::runtime_error("task cancelled");
++    // -- GODOT end --
+     return v;
+   #else
+     const Value v = tbb::parallel_reduce(tbb::blocked_range<Index>(first,last,minStepSize),identity,
+       [&](const tbb::blocked_range<Index>& r, const Value& start) { return reduction(start,func(range<Index>(r.begin(),r.end()))); },
+       reduction);
+-    if (tbb::task::self().is_cancelled())
+-      throw std::runtime_error("task cancelled");
++    // -- GODOT start --
++    // if (tbb::task::self().is_cancelled())
++    //   throw std::runtime_error("task cancelled");
++    // -- GODOT end --
+     return v;
+   #endif
+ #else // TASKING_PPL
+diff --git a/thirdparty/embree/common/lexers/stringstream.cpp b/thirdparty/embree/common/lexers/stringstream.cpp
+index 7e7b9faef8..98dc80ad59 100644
+--- a/thirdparty/embree/common/lexers/stringstream.cpp
++++ b/thirdparty/embree/common/lexers/stringstream.cpp
+@@ -39,7 +39,10 @@ namespace embree
+     std::vector<char> str; str.reserve(64);
+     while (cin->peek() != EOF && !isSeparator(cin->peek())) {
+       int c = cin->get();
+-      if (!isValidChar(c)) throw std::runtime_error("invalid character "+std::string(1,c)+" in input");
++      // -- GODOT start --
++      // if (!isValidChar(c)) throw std::runtime_error("invalid character "+std::string(1,c)+" in input");
++      if (!isValidChar(c)) abort();
++      // -- GODOT end --
+       str.push_back((char)c);
+     }
+     str.push_back(0);
+diff --git a/thirdparty/embree/common/sys/alloc.cpp b/thirdparty/embree/common/sys/alloc.cpp
+index 4e8928242e..12f143f131 100644
+--- a/thirdparty/embree/common/sys/alloc.cpp
++++ b/thirdparty/embree/common/sys/alloc.cpp
+@@ -21,7 +21,10 @@ namespace embree
+     void* ptr = _mm_malloc(size,align);
+ 
+     if (size != 0 && ptr == nullptr)
+-      throw std::bad_alloc();
++      // -- GODOT start --
++      // throw std::bad_alloc();
++      abort(); 
++      // -- GODOT end --
+     
+     return ptr;
+   }
+@@ -128,7 +131,10 @@ namespace embree
+     /* fall back to 4k pages */
+     int flags = MEM_COMMIT | MEM_RESERVE;
+     char* ptr = (char*) VirtualAlloc(nullptr,bytes,flags,PAGE_READWRITE);
+-    if (ptr == nullptr) throw std::bad_alloc();
++    // -- GODOT start --
++    // if (ptr == nullptr) throw std::bad_alloc();
++    if (ptr == nullptr) abort();
++    // -- GODOT end --
+     hugepages = false;
+     return ptr;
+   }
+@@ -145,7 +151,10 @@ namespace embree
+       return bytesOld;
+ 
+     if (!VirtualFree((char*)ptr+bytesNew,bytesOld-bytesNew,MEM_DECOMMIT))
+-      throw std::bad_alloc();
++      // -- GODOT start --
++      // throw std::bad_alloc();
++      abort();
++      // -- GODOT end --
+ 
+     return bytesNew;
+   }
+@@ -156,7 +165,10 @@ namespace embree
+       return;
+ 
+     if (!VirtualFree(ptr,0,MEM_RELEASE))
+-      throw std::bad_alloc();
++      // -- GODOT start --
++      // throw std::bad_alloc();
++      abort();
++      // -- GODOT end --
+   }
+ 
+   void os_advise(void *ptr, size_t bytes)
+@@ -260,7 +272,10 @@ namespace embree
+ 
+     /* fallback to 4k pages */
+     void* ptr = (char*) mmap(0, bytes, PROT_READ | PROT_WRITE, MAP_PRIVATE | MAP_ANON, -1, 0);
+-    if (ptr == MAP_FAILED) throw std::bad_alloc();
++    // -- GODOT start --
++    // if (ptr == MAP_FAILED) throw std::bad_alloc();
++    if (ptr == MAP_FAILED) abort();
++    // -- GODOT end --
+     hugepages = false;
+ 
+     /* advise huge page hint for THP */
+@@ -277,7 +292,10 @@ namespace embree
+       return bytesOld;
+ 
+     if (munmap((char*)ptr+bytesNew,bytesOld-bytesNew) == -1)
+-      throw std::bad_alloc();
++      // -- GODOT start --
++      // throw std::bad_alloc();
++      abort();
++      // -- GODOT end --
+ 
+     return bytesNew;
+   }
+@@ -291,7 +309,10 @@ namespace embree
+     const size_t pageSize = hugepages ? PAGE_SIZE_2M : PAGE_SIZE_4K;
+     bytes = (bytes+pageSize-1) & ~(pageSize-1);
+     if (munmap(ptr,bytes) == -1)
+-      throw std::bad_alloc();
++      // -- GODOT start --
++      // throw std::bad_alloc();
++      abort();
++      // -- GODOT end --
+   }
+ 
+   /* hint for transparent huge pages (THP) */
+diff --git a/thirdparty/embree/common/sys/platform.h b/thirdparty/embree/common/sys/platform.h
+index 7914eb7a52..737f14aa6e 100644
+--- a/thirdparty/embree/common/sys/platform.h
++++ b/thirdparty/embree/common/sys/platform.h
+@@ -174,11 +174,19 @@
+ #define PRINT4(x,y,z,w) embree_cout << STRING(x) << " = " << (x) << ", " << STRING(y) << " = " << (y) << ", " << STRING(z) << " = " << (z) << ", " << STRING(w) << " = " << (w) << embree_endl
+ 
+ #if defined(DEBUG) // only report file and line in debug mode
++  // -- GODOT start --
++  // #define THROW_RUNTIME_ERROR(str)
++  //   throw std::runtime_error(std::string(__FILE__) + " (" + toString(__LINE__) + "): " + std::string(str));
+   #define THROW_RUNTIME_ERROR(str) \
+-    throw std::runtime_error(std::string(__FILE__) + " (" + toString(__LINE__) + "): " + std::string(str));
++    printf(std::string(__FILE__) + " (" + toString(__LINE__) + "): " + std::string(str)), abort();
++  // -- GODOT end --
+ #else
++  // -- GODOT start --
++  // #define THROW_RUNTIME_ERROR(str)
++  //   throw std::runtime_error(str);
+   #define THROW_RUNTIME_ERROR(str) \
+-    throw std::runtime_error(str);
++    abort();
++  // -- GODOT end --
+ #endif
+ 
+ #define FATAL(x)   THROW_RUNTIME_ERROR(x)
+diff --git a/thirdparty/embree/common/tasking/taskschedulerinternal.cpp b/thirdparty/embree/common/tasking/taskschedulerinternal.cpp
+index 98d7fb9249..ebf656d1a0 100644
+--- a/thirdparty/embree/common/tasking/taskschedulerinternal.cpp
++++ b/thirdparty/embree/common/tasking/taskschedulerinternal.cpp
+@@ -48,13 +48,15 @@ namespace embree
+     {
+       Task* prevTask = thread.task;
+       thread.task = this;
+-      try {
+-        if (thread.scheduler->cancellingException == nullptr)
++      // -- GODOT start --
++      // try {
++      // if (thread.scheduler->cancellingException == nullptr)
+           closure->execute();
+-      } catch (...) {
+-        if (thread.scheduler->cancellingException == nullptr)
+-          thread.scheduler->cancellingException = std::current_exception();
+-      }
++      // } catch (...) {
++      //   if (thread.scheduler->cancellingException == nullptr)
++      //     thread.scheduler->cancellingException = std::current_exception();
++      // }
++      // -- GODOT end --
+       thread.task = prevTask;
+       add_dependencies(-1);
+     }
+@@ -297,8 +299,11 @@ namespace embree
+     size_t threadIndex = allocThreadIndex();
+     condition.wait(mutex, [&] () { return hasRootTask.load(); });
+     mutex.unlock();
+-    std::exception_ptr except = thread_loop(threadIndex);
+-    if (except != nullptr) std::rethrow_exception(except);
++    // -- GODOT start --
++    // std::exception_ptr except = thread_loop(threadIndex);
++    // if (except != nullptr) std::rethrow_exception(except);
++    thread_loop(threadIndex);
 +    // -- GODOT end --
    }
  
-   /* closes the shared library */
+   void TaskScheduler::reset() {
-diff --git a/common/sys/mutex.h b/common/sys/mutex.h
+@@ -330,7 +335,10 @@ namespace embree
-index 1164210f2..f0f55340a 100644
+     return thread->scheduler->cancellingException == nullptr;
---- a/common/sys/mutex.h
+   }
-+++ b/common/sys/mutex.h
+ 
-@@ -47,8 +47,17 @@ namespace embree
+-  std::exception_ptr TaskScheduler::thread_loop(size_t threadIndex)
++// -- GODOT start --
++//   std::exception_ptr TaskScheduler::thread_loop(size_t threadIndex)
++  void TaskScheduler::thread_loop(size_t threadIndex)
++// -- GODOT end --
+   {
+     /* allocate thread structure */
+     std::unique_ptr<Thread> mthread(new Thread(threadIndex,this)); // too large for stack allocation
+@@ -353,9 +361,10 @@ namespace embree
+     swapThread(oldThread);
+ 
+     /* remember exception to throw */
+-    std::exception_ptr except = nullptr;
+-    if (cancellingException != nullptr) except = cancellingException;
+-
++    // -- GODOT start --
++    // std::exception_ptr except = nullptr;
++    // if (cancellingException != nullptr) except = cancellingException;
++    // -- GODOT end --
+     /* wait for all threads to terminate */
+     threadCounter--;
+ #if defined(__WIN32__)
+@@ -373,7 +382,10 @@ namespace embree
+           yield();
+ #endif
+ 	}
+-    return except;
++    // -- GODOT start --
++    // return except;
++    return;
++    // -- GODOT end --
+   }
+ 
+   bool TaskScheduler::steal_from_other_threads(Thread& thread)
+diff --git a/thirdparty/embree/common/tasking/taskschedulerinternal.h b/thirdparty/embree/common/tasking/taskschedulerinternal.h
+index c2a9391aea..8bd70b2b8c 100644
+--- a/thirdparty/embree/common/tasking/taskschedulerinternal.h
++++ b/thirdparty/embree/common/tasking/taskschedulerinternal.h
+@@ -123,7 +123,10 @@ namespace embree
        {
-         while (flag.load()) 
+         size_t ofs = bytes + ((align - stackPtr) & (align-1));
-         {
+         if (stackPtr + ofs > CLOSURE_STACK_SIZE)
-+// -- GODOT start --
+-          throw std::runtime_error("closure stack overflow");
-+#if !(defined (__WIN32__) && defined (__MINGW32__))
++          // -- GODOT start --
-+// -- GODOT end --
++          // throw std::runtime_error("closure stack overflow");
-           _mm_pause(); 
++          abort();
-           _mm_pause();
++          // -- GODOT end --
-+// -- GODOT start --
+         stackPtr += ofs;
-+#else
+         return &stack[stackPtr-bytes];
-+          __builtin_ia32_pause();
+       }
-+          __builtin_ia32_pause();
+@@ -132,7 +135,10 @@ namespace embree
-+#endif
+       __forceinline void push_right(Thread& thread, const size_t size, const Closure& closure)
-+// -- GODOT end --
-         }
-         
-         bool expected = false;
-@@ -74,8 +82,17 @@ namespace embree
-     {
-       while(flag.load())
        {
-+// -- GODOT start --
+         if (right >= TASK_STACK_SIZE)
-+#if !(defined (__WIN32__) && defined(__MINGW32__))
+-          throw std::runtime_error("task stack overflow");
-+// -- GODOT end --
++          // -- GODOT start --
-         _mm_pause(); 
++          // throw std::runtime_error("task stack overflow");
-         _mm_pause();
++          abort();
-+// -- GODOT start --
++          // -- GODOT end --
-+#else
+ 
-+        __builtin_ia32_pause();
+ 	/* allocate new task on right side of stack */
-+        __builtin_ia32_pause();
+         size_t oldStackPtr = stackPtr;
-+#endif
+@@ -239,7 +245,10 @@ namespace embree
-+// -- GODOT end --
+     void wait_for_threads(size_t threadCount);
+ 
+     /*! thread loop for all worker threads */
+-    std::exception_ptr thread_loop(size_t threadIndex);
++    // -- GODOT start --
++    // std::exception_ptr thread_loop(size_t threadIndex);
++    void thread_loop(size_t threadIndex);
++    // -- GODOT end --
+ 
+     /*! steals a task from a different thread */
+     bool steal_from_other_threads(Thread& thread);
+diff --git a/thirdparty/embree/kernels/bvh/bvh_statistics.cpp b/thirdparty/embree/kernels/bvh/bvh_statistics.cpp
+index 20cdd2d320..aa56035026 100644
+--- a/thirdparty/embree/kernels/bvh/bvh_statistics.cpp
++++ b/thirdparty/embree/kernels/bvh/bvh_statistics.cpp
+@@ -150,7 +150,10 @@ namespace embree
        }
      }
- 
+     else {
-diff --git a/common/sys/platform.h b/common/sys/platform.h
+-      throw std::runtime_error("not supported node type in bvh_statistics");
-index 96f9aab01..08617452f 100644
++      // -- GODOT start --
---- a/common/sys/platform.h
++      // throw std::runtime_error("not supported node type in bvh_statistics");
-+++ b/common/sys/platform.h
++      abort();
-@@ -141,6 +141,9 @@
++      // -- GODOT end --
-   #define DELETED  = delete
+     }
+     return s;
+   } 
+diff --git a/thirdparty/embree/kernels/common/rtcore.cpp b/thirdparty/embree/kernels/common/rtcore.cpp
+index ee5c37b238..625fbf6d4f 100644
+--- a/thirdparty/embree/kernels/common/rtcore.cpp
++++ b/thirdparty/embree/kernels/common/rtcore.cpp
+@@ -230,7 +230,10 @@ RTC_NAMESPACE_BEGIN;
+     if (quality != RTC_BUILD_QUALITY_LOW &&
+         quality != RTC_BUILD_QUALITY_MEDIUM &&
+         quality != RTC_BUILD_QUALITY_HIGH)
+-      throw std::runtime_error("invalid build quality");
++      // -- GODOT start --
++      // throw std::runtime_error("invalid build quality");
++      abort();
++      // -- GODOT end --
+     scene->setBuildQuality(quality);
+     RTC_CATCH_END2(scene);
+   }
+@@ -1383,7 +1386,10 @@ RTC_NAMESPACE_BEGIN;
+         quality != RTC_BUILD_QUALITY_MEDIUM &&
+         quality != RTC_BUILD_QUALITY_HIGH &&
+         quality != RTC_BUILD_QUALITY_REFIT)
+-      throw std::runtime_error("invalid build quality");
++      // -- GODOT start --
++      // throw std::runtime_error("invalid build quality");
++      abort();
++      // -- GODOT end --
+     geometry->setBuildQuality(quality);
+     RTC_CATCH_END2(geometry);
+   }
+diff --git a/thirdparty/embree/kernels/common/rtcore.h b/thirdparty/embree/kernels/common/rtcore.h
+index 6583d12d57..4b070e122b 100644
+--- a/thirdparty/embree/kernels/common/rtcore.h
++++ b/thirdparty/embree/kernels/common/rtcore.h
+@@ -25,52 +25,58 @@ namespace embree
  #endif
  
+ /*! Macros used in the rtcore API implementation */
+-#define RTC_CATCH_BEGIN try {
 +// -- GODOT start --
-+#if !defined(likely)
++// #define RTC_CATCH_BEGIN try {
++#define RTC_CATCH_BEGIN
+   
+-#define RTC_CATCH_END(device)                                                \
+-  } catch (std::bad_alloc&) {                                                   \
+-    Device::process_error(device,RTC_ERROR_OUT_OF_MEMORY,"out of memory");      \
+-  } catch (rtcore_error& e) {                                                   \
+-    Device::process_error(device,e.error,e.what());                             \
+-  } catch (std::exception& e) {                                                 \
+-    Device::process_error(device,RTC_ERROR_UNKNOWN,e.what());                   \
+-  } catch (...) {                                                               \
+-    Device::process_error(device,RTC_ERROR_UNKNOWN,"unknown exception caught"); \
+-  }
++// #define RTC_CATCH_END(device)                                                \
++//   } catch (std::bad_alloc&) {                                                   \
++//     Device::process_error(device,RTC_ERROR_OUT_OF_MEMORY,"out of memory");      \
++//   } catch (rtcore_error& e) {                                                   \
++//     Device::process_error(device,e.error,e.what());                             \
++//   } catch (std::exception& e) {                                                 \
++//     Device::process_error(device,RTC_ERROR_UNKNOWN,e.what());                   \
++//   } catch (...) {                                                               \
++//     Device::process_error(device,RTC_ERROR_UNKNOWN,"unknown exception caught"); \
++//   }
++#define RTC_CATCH_END(device)
+   
+-#define RTC_CATCH_END2(scene)                                                \
+-  } catch (std::bad_alloc&) {                                                   \
+-    Device* device = scene ? scene->device : nullptr;                           \
+-    Device::process_error(device,RTC_ERROR_OUT_OF_MEMORY,"out of memory");      \
+-  } catch (rtcore_error& e) {                                                   \
+-    Device* device = scene ? scene->device : nullptr;                           \
+-    Device::process_error(device,e.error,e.what());                             \
+-  } catch (std::exception& e) {                                                 \
+-    Device* device = scene ? scene->device : nullptr;                           \
+-    Device::process_error(device,RTC_ERROR_UNKNOWN,e.what());                   \
+-  } catch (...) {                                                               \
+-    Device* device = scene ? scene->device : nullptr;                           \
+-    Device::process_error(device,RTC_ERROR_UNKNOWN,"unknown exception caught"); \
+-  }
++// #define RTC_CATCH_END2(scene)                                                \
++//   } catch (std::bad_alloc&) {                                                   \
++//     Device* device = scene ? scene->device : nullptr;                           \
++//     Device::process_error(device,RTC_ERROR_OUT_OF_MEMORY,"out of memory");      \
++//   } catch (rtcore_error& e) {                                                   \
++//     Device* device = scene ? scene->device : nullptr;                           \
++//     Device::process_error(device,e.error,e.what());                             \
++//   } catch (std::exception& e) {                                                 \
++//     Device* device = scene ? scene->device : nullptr;                           \
++//     Device::process_error(device,RTC_ERROR_UNKNOWN,e.what());                   \
++//   } catch (...) {                                                               \
++//     Device* device = scene ? scene->device : nullptr;                           \
++//     Device::process_error(device,RTC_ERROR_UNKNOWN,"unknown exception caught"); \
++//   }
++#define RTC_CATCH_END2(scene)
+ 
+-#define RTC_CATCH_END2_FALSE(scene)                                             \
+-  } catch (std::bad_alloc&) {                                                   \
+-    Device* device = scene ? scene->device : nullptr;                           \
+-    Device::process_error(device,RTC_ERROR_OUT_OF_MEMORY,"out of memory");      \
+-    return false;                                                               \
+-  } catch (rtcore_error& e) {                                                   \
+-    Device* device = scene ? scene->device : nullptr;                           \
+-    Device::process_error(device,e.error,e.what());                             \
+-    return false;                                                               \
+-  } catch (std::exception& e) {                                                 \
+-    Device* device = scene ? scene->device : nullptr;                           \
+-    Device::process_error(device,RTC_ERROR_UNKNOWN,e.what());                   \
+-    return false;                                                               \
+-  } catch (...) {                                                               \
+-    Device* device = scene ? scene->device : nullptr;                           \
+-    Device::process_error(device,RTC_ERROR_UNKNOWN,"unknown exception caught"); \
+-    return false;                                                               \
+-  }
++// #define RTC_CATCH_END2_FALSE(scene)                                             \
++//   } catch (std::bad_alloc&) {                                                   \
++//     Device* device = scene ? scene->device : nullptr;                           \
++//     Device::process_error(device,RTC_ERROR_OUT_OF_MEMORY,"out of memory");      \
++//     return false;                                                               \
++//   } catch (rtcore_error& e) {                                                   \
++//     Device* device = scene ? scene->device : nullptr;                           \
++//     Device::process_error(device,e.error,e.what());                             \
++//     return false;                                                               \
++//   } catch (std::exception& e) {                                                 \
++//     Device* device = scene ? scene->device : nullptr;                           \
++//     Device::process_error(device,RTC_ERROR_UNKNOWN,e.what());                   \
++//     return false;                                                               \
++//   } catch (...) {                                                               \
++//     Device* device = scene ? scene->device : nullptr;                           \
++//     Device::process_error(device,RTC_ERROR_UNKNOWN,"unknown exception caught"); \
++//     return false;                                                               \
++//   }
++#define RTC_CATCH_END2_FALSE(scene) return false;
 +// -- GODOT end --
- #if defined(_MSC_VER) && !defined(__INTEL_COMPILER)
+ 
- #define   likely(expr) (expr)
+ #define RTC_VERIFY_HANDLE(handle)                               \
- #define unlikely(expr) (expr)
+   if (handle == nullptr) {                                         \
-@@ -148,6 +151,9 @@
+@@ -97,28 +103,38 @@ namespace embree
- #define   likely(expr) __builtin_expect((bool)(expr),true )
+ #define RTC_TRACE(x) 
- #define unlikely(expr) __builtin_expect((bool)(expr),false)
  #endif
-+// -- GODOT start --
+ 
-+#endif
+-  /*! used to throw embree API errors */
+-  struct rtcore_error : public std::exception
+-  {
+-    __forceinline rtcore_error(RTCError error, const std::string& str)
+-      : error(error), str(str) {}
+-    
+-    ~rtcore_error() throw() {}
+-    
+-    const char* what () const throw () {
+-      return str.c_str();
+-    }
+-    
+-    RTCError error;
+-    std::string str;
+-  };
++// -- GODOT begin --
++//   /*! used to throw embree API errors */
++//   struct rtcore_error : public std::exception
++//   {
++//     __forceinline rtcore_error(RTCError error, const std::string& str)
++//       : error(error), str(str) {}
++//     
++//     ~rtcore_error() throw() {}
++//     
++//     const char* what () const throw () {
++//       return str.c_str();
++//     }
++//     
++//     RTCError error;
++//     std::string str;
++//   };
 +// -- GODOT end --
  
- ////////////////////////////////////////////////////////////////////////////////
+ #if defined(DEBUG) // only report file and line in debug mode
- /// Error handling and debugging
++  // -- GODOT begin --
-diff --git a/common/sys/sysinfo.cpp b/common/sys/sysinfo.cpp
++  // #define throw_RTCError(error,str) \
-index eb0a10eaf..74438260d 100644
++  //   throw rtcore_error(error,std::string(__FILE__) + " (" + toString(__LINE__) + "): " + std::string(str));
---- a/common/sys/sysinfo.cpp
+   #define throw_RTCError(error,str) \
-+++ b/common/sys/sysinfo.cpp
+-    throw rtcore_error(error,std::string(__FILE__) + " (" + toString(__LINE__) + "): " + std::string(str));
-@@ -233,7 +233,7 @@ namespace embree
++    printf(std::string(__FILE__) + " (" + toString(__LINE__) + "): " + std::string(str)), abort();
- 
++  // -- GODOT end --
-   __noinline int64_t get_xcr0() 
-   {
--#if defined (__WIN32__)
-+#if defined (__WIN32__) /* -- GODOT start -- */ && !defined (__MINGW32__) /* -- GODOT end -- */
-     int64_t xcr0 = 0; // int64_t is workaround for compiler bug under VS2013, Win32
-     xcr0 = _xgetbv(0);
-     return xcr0;
-diff --git a/common/tasking/taskschedulerinternal.cpp b/common/tasking/taskschedulerinternal.cpp
-index 2152e92f4..923d62f83 100644
---- a/common/tasking/taskschedulerinternal.cpp
-+++ b/common/tasking/taskschedulerinternal.cpp
-@@ -361,7 +361,15 @@ namespace embree
-           if ((loopIndex % LOOP_YIELD_THRESHOLD) == 0)
-             yield();
-           else
-+// -- GODOT start --
-+#if !defined(__MINGW32__)
-+// -- GODOT end --
-             _mm_pause();
-+// -- GODOT start --
-+#else
-+            usleep(1);
-+#endif
-+// -- GODOT end --
- 	  loopIndex++;
  #else
-           yield();
++  // -- GODOT begin --
-diff --git a/common/tasking/taskschedulertbb.h b/common/tasking/taskschedulertbb.h
++  // #define throw_RTCError(error,str) \
-index 98dba2687..369e5edf0 100644
++  //   throw rtcore_error(error,str);
---- a/common/tasking/taskschedulertbb.h
+   #define throw_RTCError(error,str) \
-+++ b/common/tasking/taskschedulertbb.h
+-    throw rtcore_error(error,str);
-@@ -12,7 +12,13 @@
++    abort();
- #include "../sys/ref.h"
++  // -- GODOT end --
- 
- #if defined(__WIN32__)
-+// -- GODOT start --
-+#if !defined(NOMINMAX)
-+// -- GODOT end --
- #  define NOMINMAX
-+// -- GODOT start --
-+#endif
-+// -- GODOT end --
  #endif
  
- // We need to define these to avoid implicit linkage against
+ #define RTC_BUILD_ARGUMENTS_HAS(settings,member) \
-diff a/include/embree3/rtcore_common.h b/include/embree3/rtcore_common.h
+diff --git a/thirdparty/embree/kernels/common/scene.cpp b/thirdparty/embree/kernels/common/scene.cpp
---- a/include/embree3/rtcore_common.h
+index e75aa968f9..1e23aeb415 100644
-+++ b/include/embree3/rtcore_common.h
+--- a/thirdparty/embree/kernels/common/scene.cpp
-@@ -19,7 +19,7 @@
++++ b/thirdparty/embree/kernels/common/scene.cpp
- #endif
+@@ -800,16 +800,18 @@ namespace embree
- #endif
+     }
  
--#ifdef _WIN32
+     /* initiate build */
-+#if defined(_WIN32) && defined(_MSC_VER)
+-    try {
- #  define RTC_ALIGN(...) __declspec(align(__VA_ARGS__))
++    // -- GODOT start --
- #else
++    // try {
- #  define RTC_ALIGN(...) __attribute__((aligned(__VA_ARGS__)))
+       scheduler->spawn_root([&]() { commit_task(); Lock<MutexSys> lock(schedulerMutex); this->scheduler = nullptr; }, 1, !join);
+-    }
+-    catch (...) {
+-      accels_clear();
+-      updateInterface();
+-      Lock<MutexSys> lock(schedulerMutex);
+-      this->scheduler = nullptr;
+-      throw;
+-    }
++    // }
++    // catch (...) {
++    //   accels_clear();
++    //   updateInterface();
++    //   Lock<MutexSys> lock(schedulerMutex);
++    //   this->scheduler = nullptr;
++    //   throw;
++    // }
++    // -- GODOT end --
+   }
+ 
+ #endif