Upgrade Embree to the latest official release.

Since Embree v3.13.0 supports AARCH64, switch back to the
official repo instead of using Embree-aarch64.

`thirdparty/embree/patches/godot-changes.patch` should now contain
an accurate diff of the changes done to the library.

COPYRIGHT.txt

@@ -131,8 +131,8 @@ Comment: doctest
 Copyright: 2016-2020, Viktor Kirilov
 License: Expat
 
-Files: ./thirdparty/embree-aarch64/
+Files: ./thirdparty/embree/
-Comment: Embree-aarch64
+Comment: Embree
 Copyright: 2009-2021 Intel Corporation
 License: Apache-2.0
 

modules/raycast/SCsub

@@ -10,7 +10,7 @@ env_raycast = env_modules.Clone()
 thirdparty_obj = []
 
 if env["builtin_embree"]:
-    thirdparty_dir = "#thirdparty/embree-aarch64/"
+    thirdparty_dir = "#thirdparty/embree/"
 
     embree_src = [
         "common/sys/sysinfo.cpp",
@@ -28,16 +28,6 @@ if env["builtin_embree"]:
         "common/lexers/stringstream.cpp",
         "common/lexers/tokenstream.cpp",
         "common/tasking/taskschedulerinternal.cpp",
-        "common/algorithms/parallel_for.cpp",
-        "common/algorithms/parallel_reduce.cpp",
-        "common/algorithms/parallel_prefix_sum.cpp",
-        "common/algorithms/parallel_for_for.cpp",
-        "common/algorithms/parallel_for_for_prefix_sum.cpp",
-        "common/algorithms/parallel_partition.cpp",
-        "common/algorithms/parallel_sort.cpp",
-        "common/algorithms/parallel_set.cpp",
-        "common/algorithms/parallel_map.cpp",
-        "common/algorithms/parallel_filter.cpp",
         "kernels/common/device.cpp",
         "kernels/common/stat.cpp",
         "kernels/common/acceln.cpp",
@@ -82,13 +72,17 @@ if env["builtin_embree"]:
     if env["platform"] == "windows":
         if env.msvc:
             env.Append(LINKFLAGS=["psapi.lib"])
-            env_raycast.Append(CPPDEFINES=["__SSE2__", "__SSE__"])
         else:
             env.Append(LIBS=["psapi"])
 
     env_thirdparty = env_raycast.Clone()
     env_thirdparty.disable_warnings()
     env_thirdparty.add_source_files(thirdparty_obj, thirdparty_sources)
+
+    if not env["arch"] in ["x86", "x86_64"] or env.msvc:
+        # Embree needs those, it will automatically use SSE2NEON in ARM
+        env_thirdparty.Append(CPPDEFINES=["__SSE2__", "__SSE__"])
+
     env.modules_sources += thirdparty_obj
 
 

modules/raycast/config.py

@@ -1,6 +1,5 @@
 def can_build(env, platform):
-    # Depends on Embree library, which supports only x86_64 (originally)
+    # Depends on Embree library, which only supports x86_64 and aarch64.
-    # and aarch64 (thanks to the embree-aarch64 fork).
 
     if platform == "android":
         return env["android_arch"] in ["arm64v8", "x86_64"]

modules/raycast/godot_update_embree.py

@@ -11,6 +11,7 @@ include_dirs = [
     "common/algorithms",
     "common/lexers",
     "common/simd",
+    "common/simd/arm",
     "include/embree3",
     "kernels/subdiv",
     "kernels/geometry",
@@ -32,16 +33,6 @@ cpp_files = [
     "common/lexers/stringstream.cpp",
     "common/lexers/tokenstream.cpp",
     "common/tasking/taskschedulerinternal.cpp",
-    "common/algorithms/parallel_for.cpp",
-    "common/algorithms/parallel_reduce.cpp",
-    "common/algorithms/parallel_prefix_sum.cpp",
-    "common/algorithms/parallel_for_for.cpp",
-    "common/algorithms/parallel_for_for_prefix_sum.cpp",
-    "common/algorithms/parallel_partition.cpp",
-    "common/algorithms/parallel_sort.cpp",
-    "common/algorithms/parallel_set.cpp",
-    "common/algorithms/parallel_map.cpp",
-    "common/algorithms/parallel_filter.cpp",
     "kernels/common/device.cpp",
     "kernels/common/stat.cpp",
     "kernels/common/acceln.cpp",
@@ -74,11 +65,11 @@ cpp_files = [
 
 os.chdir("../../thirdparty")
 
-dir_name = "embree-aarch64"
+dir_name = "embree"
 if os.path.exists(dir_name):
     shutil.rmtree(dir_name)
 
-subprocess.run(["git", "clone", "https://github.com/lighttransport/embree-aarch64.git", "embree-tmp"])
+subprocess.run(["git", "clone", "https://github.com/embree/embree.git", "embree-tmp"])
 os.chdir("embree-tmp")
 
 commit_hash = str(subprocess.check_output(["git", "rev-parse", "HEAD"], universal_newlines=True)).strip()
@@ -197,7 +188,7 @@ with open("CMakeLists.txt", "r") as cmake_file:
 with open(os.path.join(dest_dir, "include/embree3/rtcore_config.h"), "w") as config_file:
     config_file.write(
         f"""
-// Copyright 2009-2020 Intel Corporation
+// Copyright 2009-2021 Intel Corporation
 // SPDX-License-Identifier: Apache-2.0
 
 #pragma once

thirdparty/README.md

@@ -61,10 +61,10 @@ Files extracted from upstream source:
 Extracted from .zip provided. Extracted license and header only.
 
 
-## embree-aarch64
+## embree
 
-- Upstream: https://github.com/lighttransport/embree-aarch64
+- Upstream: https://github.com/embree/embree
-- Version: 3.12.1 (6ef362f99af80c9dfe8dd2bfc582d9067897edc6, 2020)
+- Version: 3.13.0 (7c53133eb21424f7f0ae1e25bf357e358feaf6ab, 2021)
 - License: Apache 2.0
 
 Files extracted from upstream:
@@ -73,7 +73,7 @@ Files extracted from upstream:
 - All header files in the directories listed in `modules/raycast/godot_update_embree.py`
 
 The `modules/raycast/godot_update_embree.py` script can be used to pull the
-relevant files from the latest Embree-aarch64 release and apply some automatic changes.
+relevant files from the latest Embree release and apply some automatic changes.
 
 Some changes have been made in order to remove exceptions and fix minor build errors.
 They are marked with `// -- GODOT start --` and `// -- GODOT end --`

thirdparty/embree-aarch64/common/algorithms/parallel_filter.cpp

@@ -1,56 +0,0 @@
-// Copyright 2009-2020 Intel Corporation
-// SPDX-License-Identifier: Apache-2.0
-
-#include "parallel_filter.h"
-#include "../sys/regression.h"
-#include <map>
-
-namespace embree
-{
-  struct parallel_filter_regression_test : public RegressionTest
-  {
-    parallel_filter_regression_test(const char* name) : RegressionTest(name) {
-      registerRegressionTest(this);
-    }
-    
-    bool run ()
-    {
-      bool passed = true;
-      auto pred = [&]( uint32_t v ) { return (v & 0x3) == 0; };
-      
-      for (size_t N=10; N<1000000; N=size_t(2.1*N))
-      {
-        size_t N0 = rand() % N;
-        
-	/* initialize array with random numbers */
-	std::vector<uint32_t> src(N);
-        std::map<uint32_t,int> m;
-	for (size_t i=0; i<N; i++) src[i] = rand();
-
-        /* count elements up */
-	for (size_t i=N0; i<N; i++)
-          if (pred(src[i]))
-            m[src[i]] = 0;
-        for (size_t i=N0; i<N; i++)
-          if (pred(src[i]))
-            m[src[i]]++;
-
-        /* filter array */
-        //size_t M = sequential_filter(src.data(),N0,N,pred);
-        size_t M = parallel_filter(src.data(),N0,N,size_t(1024),pred);
-        
-	/* check if filtered data is correct */
-	for (size_t i=N0; i<M; i++) {
-          passed &= pred(src[i]);
-          m[src[i]]--;
-        }
-	for (size_t i=N0; i<M; i++)
-          passed &= (m[src[i]] == 0);
-      }
-
-      return passed;
-    }
-  };
-
-  parallel_filter_regression_test parallel_filter_regression("parallel_filter_regression");
-}

thirdparty/embree-aarch64/common/algorithms/parallel_for.cpp

@@ -1,48 +0,0 @@
-// Copyright 2009-2020 Intel Corporation
-// SPDX-License-Identifier: Apache-2.0
-
-#include "parallel_for.h"
-#include "../sys/regression.h"
-
-namespace embree
-{
-  struct parallel_for_regression_test : public RegressionTest
-  {
-    parallel_for_regression_test(const char* name) : RegressionTest(name) {
-      registerRegressionTest(this);
-    }
-    
-    bool run ()
-    {
-      bool passed = true;
-
-      const size_t M = 10;
-      for (size_t N=10; N<10000000; N=size_t(2.1*N))
-      {
-        /* sequentially calculate sum of squares */
-        size_t sum0 = 0;
-        for (size_t i=0; i<N; i++) {
-          sum0 += i*i;
-        }
-
-        /* parallel calculation of sum of squares */
-        for (size_t m=0; m<M; m++)
-        {
-          std::atomic<size_t> sum1(0);
-          parallel_for( size_t(0), size_t(N), size_t(1024), [&](const range<size_t>& r) 
-          {
-            size_t s = 0;
-            for (size_t i=r.begin(); i<r.end(); i++) 
-              s += i*i;
-            sum1 += s;
-          });
-          passed = sum0 == sum1;
-        }
-      }
-      
-      return passed;
-    }
-  };
-
-  parallel_for_regression_test parallel_for_regression("parallel_for_regression_test");
-}

thirdparty/embree-aarch64/common/algorithms/parallel_for_for.cpp

@@ -1,63 +0,0 @@
-// Copyright 2009-2020 Intel Corporation
-// SPDX-License-Identifier: Apache-2.0
-
-#include "parallel_for_for.h"
-#include "../sys/regression.h"
-
-namespace embree
-{
-  struct parallel_for_for_regression_test : public RegressionTest
-  {
-    parallel_for_for_regression_test(const char* name) : RegressionTest(name) {
-      registerRegressionTest(this);
-    }
-    
-    bool run ()
-    {
-      bool passed = true;
-
-      /* create vector with random numbers */
-      size_t sum0 = 0;
-      size_t K = 0;
-      const size_t M = 1000;
-      std::vector<std::vector<size_t>* > array2(M);
-      for (size_t i=0; i<M; i++) {
-        const size_t N = rand() % 1024;
-        K+=N;
-        array2[i] = new std::vector<size_t>(N);
-        for (size_t j=0; j<N; j++) 
-          sum0 += (*array2[i])[j] = rand();
-      }
-
-      /* array to test global index */
-      std::vector<atomic<size_t>> verify_k(K);
-      for (size_t i=0; i<K; i++) verify_k[i].store(0);
-
-      /* add all numbers using parallel_for_for */
-      std::atomic<size_t> sum1(0);
-      parallel_for_for( array2, size_t(1), [&](std::vector<size_t>* v, const range<size_t>& r, size_t k) -> size_t
-      {
-        size_t s = 0;
-	for (size_t i=r.begin(); i<r.end(); i++) {
-	  s += (*v)[i];
-          verify_k[k++]++;
-        }
-        sum1 += s;
-	return sum1;
-      });
-      passed &= (sum0 == sum1);
-
-      /* check global index */
-      for (size_t i=0; i<K; i++) 
-        passed &= (verify_k[i] == 1);
-
-      /* delete vectors again */
-      for (size_t i=0; i<array2.size(); i++)
-	delete array2[i];
-      
-      return passed;
-    }
-  };
-
-  parallel_for_for_regression_test parallel_for_for_regression("parallel_for_for_regression_test");
-}

thirdparty/embree-aarch64/common/algorithms/parallel_for_for_prefix_sum.cpp

@@ -1,85 +0,0 @@
-// Copyright 2009-2020 Intel Corporation
-// SPDX-License-Identifier: Apache-2.0
-
-#include "parallel_for_for_prefix_sum.h"
-#include "../sys/regression.h"
-
-namespace embree
-{
-  struct parallel_for_for_prefix_sum_regression_test : public RegressionTest
-  {
-    parallel_for_for_prefix_sum_regression_test(const char* name) : RegressionTest(name) {
-      registerRegressionTest(this);
-    }
-    
-    bool run ()
-    {
-      bool passed = true;
-
-      /* create vector with random numbers */
-      const size_t M = 10;
-      std::vector<atomic<size_t>> flattened;
-      typedef std::vector<std::vector<size_t>* > ArrayArray;
-      ArrayArray array2(M);
-      size_t K = 0;
-      for (size_t i=0; i<M; i++) {
-        const size_t N = rand() % 10;
-        K += N;
-        array2[i] = new std::vector<size_t>(N);
-        for (size_t j=0; j<N; j++) 
-          (*array2[i])[j] = rand() % 10;
-      }
-  
-      /* array to test global index */
-      std::vector<atomic<size_t>> verify_k(K);
-      for (size_t i=0; i<K; i++) verify_k[i].store(0);
-
-      ParallelForForPrefixSumState<size_t> state(array2,size_t(1));
-  
-      /* dry run only counts */
-      size_t S = parallel_for_for_prefix_sum0( state, array2, size_t(0), [&](std::vector<size_t>* v, const range<size_t>& r, size_t k, size_t i) -> size_t
-      {
-        size_t s = 0;
-	for (size_t i=r.begin(); i<r.end(); i++) {
-          s += (*v)[i];
-          verify_k[k++]++;
-        }
-        return s;
-      }, [](size_t v0, size_t v1) { return v0+v1; });
-      
-      /* create properly sized output array */
-      flattened.resize(S);
-      for (auto& a : flattened) a.store(0);
-
-      /* now we actually fill the flattened array */
-      parallel_for_for_prefix_sum1( state, array2, size_t(0), [&](std::vector<size_t>* v, const range<size_t>& r, size_t k, size_t i, const size_t base) -> size_t
-      {
-        size_t s = 0;
-	for (size_t i=r.begin(); i<r.end(); i++) {
-          for (size_t j=0; j<(*v)[i]; j++) {
-            flattened[base+s+j]++;
-          }
-          s += (*v)[i];
-          verify_k[k++]++;
-        }
-        return s;
-      }, [](size_t v0, size_t v1) { return v0+v1; });
-
-      /* check global index */
-      for (size_t i=0; i<K; i++) 
-        passed &= (verify_k[i] == 2);
-
-      /* check if each element was assigned exactly once */
-      for (size_t i=0; i<flattened.size(); i++)
-        passed &= (flattened[i] == 1);
-      
-      /* delete arrays again */
-      for (size_t i=0; i<array2.size(); i++)
-	delete array2[i];
-
-      return passed;
-    }
-  };
-
-  parallel_for_for_prefix_sum_regression_test parallel_for_for_prefix_sum_regression("parallel_for_for_prefix_sum_regression_test");
-}

thirdparty/embree-aarch64/common/algorithms/parallel_map.cpp

@@ -1,47 +0,0 @@
-// Copyright 2009-2020 Intel Corporation
-// SPDX-License-Identifier: Apache-2.0
-
-#include "parallel_map.h"
-#include "../sys/regression.h"
-
-namespace embree
-{
-  struct parallel_map_regression_test : public RegressionTest
-  {
-    parallel_map_regression_test(const char* name) : RegressionTest(name) {
-      registerRegressionTest(this);
-    }
-    
-    bool run ()
-    {
-      bool passed = true;
-
-      /* create key/value vectors with random numbers */
-      const size_t N = 10000;
-      std::vector<uint32_t> keys(N);
-      std::vector<uint32_t> vals(N);
-      for (size_t i=0; i<N; i++) keys[i] = 2*unsigned(i)*647382649;
-      for (size_t i=0; i<N; i++) std::swap(keys[i],keys[rand()%N]);
-      for (size_t i=0; i<N; i++) vals[i] = 2*rand();
-      
-      /* create map */
-      parallel_map<uint32_t,uint32_t> map;
-      map.init(keys,vals);
-
-      /* check that all keys are properly mapped */
-      for (size_t i=0; i<N; i++) {
-        const uint32_t* val = map.lookup(keys[i]);
-        passed &= val && (*val == vals[i]);
-      }
-
-      /* check that these keys are not in the map */
-      for (size_t i=0; i<N; i++) {
-        passed &= !map.lookup(keys[i]+1);
-      }
-
-      return passed;
-    }
-  };
-
-  parallel_map_regression_test parallel_map_regression("parallel_map_regression_test");
-}

thirdparty/embree-aarch64/common/algorithms/parallel_partition.cpp

@@ -1,53 +0,0 @@
-// Copyright 2009-2020 Intel Corporation
-// SPDX-License-Identifier: Apache-2.0
-
-#include "parallel_partition.h"
-#include "../sys/regression.h"
-
-namespace embree
-{
-  struct parallel_partition_regression_test : public RegressionTest
-  {
-    parallel_partition_regression_test(const char* name) : RegressionTest(name) {
-      registerRegressionTest(this);
-    }
-    
-    bool run ()
-    {
-      bool passed = true;
-
-      for (size_t i=0; i<100; i++)
-      {
-        /* create random permutation */
-        size_t N = std::rand() % 1000000;
-        std::vector<unsigned> array(N);
-        for (unsigned i=0; i<N; i++) array[i] = i;
-        for (auto& v : array) std::swap(v,array[std::rand()%array.size()]);
-        size_t split = std::rand() % (N+1);
-
-        /* perform parallel partitioning */
-        size_t left_sum = 0, right_sum = 0;
-        size_t mid = parallel_partitioning(array.data(),0,array.size(),0,left_sum,right_sum,
-                                           [&] ( size_t i ) { return i < split; },
-                                           []  ( size_t& sum, unsigned v) { sum += v; },
-                                           []  ( size_t& sum, size_t v) { sum += v; },
-                                           128);
-        
-        /*serial_partitioning(array.data(),0,array.size(),left_sum,right_sum,
-                            [&] ( size_t i ) { return i < split; },
-                            []  ( size_t& left_sum, int v) { left_sum += v; });*/
-
-        /* verify result */
-        passed &= mid == split;
-        passed &= left_sum == split*(split-1)/2;
-        passed &= right_sum == N*(N-1)/2-left_sum;
-        for (size_t i=0; i<split; i++) passed &= array[i] < split;
-        for (size_t i=split; i<N; i++) passed &= array[i] >= split;
-      }
-      
-      return passed;
-    }
-  };
-
-  parallel_partition_regression_test parallel_partition_regression("parallel_partition_regression_test");
-}

thirdparty/embree-aarch64/common/algorithms/parallel_prefix_sum.cpp

@@ -1,48 +0,0 @@
-// Copyright 2009-2020 Intel Corporation
-// SPDX-License-Identifier: Apache-2.0
-
-#include "parallel_prefix_sum.h"
-#include "../sys/regression.h"
-
-namespace embree
-{
-  struct parallel_prefix_sum_regression_test : public RegressionTest
-  {
-    parallel_prefix_sum_regression_test(const char* name) : RegressionTest(name) {
-      registerRegressionTest(this);
-    }
-    
-    bool run ()
-    {
-      bool passed = true;
-      const size_t M = 10;
-      
-      for (size_t N=10; N<10000000; N=size_t(2.1*N))
-      {
-	/* initialize array with random numbers */
-        uint32_t sum0 = 0;
-	std::vector<uint32_t> src(N);
-	for (size_t i=0; i<N; i++) {
-	  sum0 += src[i] = rand();
-        }
-        
-	/* calculate parallel prefix sum */
-	std::vector<uint32_t> dst(N);
-	for (auto& v : dst) v = 0;
-	
-	for (size_t i=0; i<M; i++) {
-	  uint32_t sum1 = parallel_prefix_sum(src,dst,N,0,std::plus<uint32_t>());
-          passed &= (sum0 == sum1);
-        }
-        
-	/* check if prefix sum is correct */
-	for (size_t i=0, sum=0; i<N; sum+=src[i++])
-	  passed &= ((uint32_t)sum == dst[i]);
-      }
-      
-      return passed;
-    }
-  };
-
-  parallel_prefix_sum_regression_test parallel_prefix_sum_regression("parallel_prefix_sum_regression");
-}

thirdparty/embree-aarch64/common/algorithms/parallel_reduce.cpp

@@ -1,49 +0,0 @@
-// Copyright 2009-2020 Intel Corporation
-// SPDX-License-Identifier: Apache-2.0
-
-#include "parallel_reduce.h"
-#include "../sys/regression.h"
-
-namespace embree
-{
-  struct parallel_reduce_regression_test : public RegressionTest
-  {
-    parallel_reduce_regression_test(const char* name) : RegressionTest(name) {
-      registerRegressionTest(this);
-    }
-    
-    bool run ()
-    {
-      bool passed = true;
-
-      const size_t M = 10;
-      for (size_t N=10; N<10000000; N=size_t(2.1*N))
-      {
-        /* sequentially calculate sum of squares */
-        size_t sum0 = 0;
-        for (size_t i=0; i<N; i++) {
-          sum0 += i*i;
-        }
-
-        /* parallel calculation of sum of squares */
-        for (size_t m=0; m<M; m++)
-        {
-          size_t sum1 = parallel_reduce( size_t(0), size_t(N), size_t(1024), size_t(0), [&](const range<size_t>& r) -> size_t
-          {
-            size_t s = 0;
-            for (size_t i=r.begin(); i<r.end(); i++) 
-              s += i*i;
-            return s;
-          }, 
-          [](const size_t v0, const size_t v1) {
-            return v0+v1;
-          });
-          passed = sum0 == sum1;
-        }
-      }
-      return passed;
-    }
-  };
-
-  parallel_reduce_regression_test parallel_reduce_regression("parallel_reduce_regression_test");
-}

thirdparty/embree-aarch64/common/algorithms/parallel_set.cpp

@@ -1,43 +0,0 @@
-// Copyright 2009-2020 Intel Corporation
-// SPDX-License-Identifier: Apache-2.0
-
-#include "parallel_set.h"
-#include "../sys/regression.h"
-
-namespace embree
-{
-  struct parallel_set_regression_test : public RegressionTest
-  {
-    parallel_set_regression_test(const char* name) : RegressionTest(name) {
-      registerRegressionTest(this);
-    }
-    
-    bool run ()
-    {
-      bool passed = true;
-
-      /* create vector with random numbers */
-      const size_t N = 10000;
-      std::vector<uint32_t> unsorted(N);
-      for (size_t i=0; i<N; i++) unsorted[i] = 2*rand();
-      
-      /* created set from numbers */
-      parallel_set<uint32_t> sorted;
-      sorted.init(unsorted);
-
-      /* check that all elements are in the set */
-      for (size_t i=0; i<N; i++) {
-	passed &= sorted.lookup(unsorted[i]);
-      }
-
-      /* check that these elements are not in the set */
-      for (size_t i=0; i<N; i++) {
-	passed &= !sorted.lookup(unsorted[i]+1);
-      }
-
-      return passed;
-    }
-  };
-
-  parallel_set_regression_test parallel_set_regression("parallel_set_regression_test");
-}

thirdparty/embree-aarch64/common/algorithms/parallel_sort.cpp

@@ -1,50 +0,0 @@
-// Copyright 2009-2020 Intel Corporation
-// SPDX-License-Identifier: Apache-2.0
-
-#include "parallel_sort.h"
-#include "../sys/regression.h"
-
-namespace embree
-{
-  template<typename Key>
-  struct RadixSortRegressionTest : public RegressionTest
-  {
-    RadixSortRegressionTest(const char* name) : RegressionTest(name) {
-      registerRegressionTest(this);
-    }
-    
-    bool run ()
-    {
-      bool passed = true;
-      const size_t M = 10;
-
-      for (size_t N=10; N<1000000; N=size_t(2.1*N))
-      {
-	std::vector<Key> src(N); memset(src.data(),0,N*sizeof(Key));
-	std::vector<Key> tmp(N); memset(tmp.data(),0,N*sizeof(Key));
-	for (size_t i=0; i<N; i++) src[i] = uint64_t(rand())*uint64_t(rand());
-	
-	/* calculate checksum */
-	Key sum0 = 0; for (size_t i=0; i<N; i++) sum0 += src[i];
-        
-	/* sort numbers */
-	for (size_t i=0; i<M; i++) {
-          radix_sort<Key>(src.data(),tmp.data(),N);
-        }
-	
-	/* calculate checksum */
-	Key sum1 = 0; for (size_t i=0; i<N; i++) sum1 += src[i];
-	if (sum0 != sum1) passed = false;
-        
-	/* check if numbers are sorted */
-	for (size_t i=1; i<N; i++)
-	  passed &= src[i-1] <= src[i];
-      }
-      
-      return passed;
-    }
-  };
-
-  RadixSortRegressionTest<uint32_t> test_u32("RadixSortRegressionTestU32");
-  RadixSortRegressionTest<uint64_t> test_u64("RadixSortRegressionTestU64");
-}

thirdparty/embree-aarch64/common/math/AVX2NEON.h

@@ -1,986 +0,0 @@
-#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-aarch64/common/math/constants.cpp

@@ -1,61 +0,0 @@
-// Copyright 2009-2020 Intel Corporation
-// SPDX-License-Identifier: Apache-2.0
-
-#if defined(__aarch64__)
-#include <arm_neon.h>
-#endif
-
-#include "constants.h"
-
-namespace embree
-{
-  TrueTy True;
-  FalseTy False;
-  ZeroTy zero;
-  OneTy one;
-  NegInfTy neg_inf;
-  PosInfTy inf;
-  PosInfTy pos_inf;
-  NaNTy nan;
-  UlpTy ulp;
-  PiTy pi;
-  OneOverPiTy one_over_pi;
-  TwoPiTy two_pi;
-  OneOverTwoPiTy one_over_two_pi;
-  FourPiTy four_pi;
-  OneOverFourPiTy one_over_four_pi;
-  StepTy step;
-  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-aarch64/common/tasking/taskschedulergcd.h

@@ -1,49 +0,0 @@
-#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-aarch64/kernels/builders/primrefgen.h

@@ -1,28 +0,0 @@
-// Copyright 2009-2020 Intel Corporation
-// SPDX-License-Identifier: Apache-2.0
-
-#pragma once
-
-#include "../common/scene.h"
-#include "../common/primref.h"
-#include "../common/primref_mb.h"
-#include "priminfo.h"
-#include "bvh_builder_morton.h"
-
-namespace embree
-{
-  namespace isa
-  {
-    PrimInfo createPrimRefArray(Geometry* geometry, unsigned int geomID, mvector<PrimRef>& prims, BuildProgressMonitor& progressMonitor);
-   
-    PrimInfo createPrimRefArray(Scene* scene, Geometry::GTypeMask types, bool mblur, mvector<PrimRef>& prims, BuildProgressMonitor& progressMonitor);
-   
-    PrimInfo createPrimRefArrayMBlur(Scene* scene, Geometry::GTypeMask types, mvector<PrimRef>& prims, BuildProgressMonitor& progressMonitor, size_t itime = 0);
-
-    PrimInfoMB createPrimRefArrayMSMBlur(Scene* scene, Geometry::GTypeMask types, mvector<PrimRefMB>& prims, BuildProgressMonitor& progressMonitor, BBox1f t0t1 = BBox1f(0.0f,1.0f));
-
-    template<typename Mesh>
-      size_t createMortonCodeArray(Mesh* mesh, mvector<BVHBuilderMorton::BuildPrim>& morton, BuildProgressMonitor& progressMonitor);
-  }
-}
-

thirdparty/embree-aarch64/kernels/common/instance_stack.h

@@ -1,199 +0,0 @@
-// Copyright 2009-2020 Intel Corporation
-// SPDX-License-Identifier: Apache-2.0
-
-#pragma once
-
-#include "rtcore.h"
-
-namespace embree {
-namespace instance_id_stack {
-
-static_assert(RTC_MAX_INSTANCE_LEVEL_COUNT > 0, 
-              "RTC_MAX_INSTANCE_LEVEL_COUNT must be greater than 0.");
-
-/*******************************************************************************
- * Instance ID stack manipulation.
- * This is used from the instance intersector.
- ******************************************************************************/
-
-/* 
- * Push an instance to the stack. 
- */
-RTC_FORCEINLINE bool push(RTCIntersectContext* context, 
-                          unsigned instanceId)
-{
-#if RTC_MAX_INSTANCE_LEVEL_COUNT > 1
-  const bool spaceAvailable = context->instStackSize < RTC_MAX_INSTANCE_LEVEL_COUNT;
-  /* We assert here because instances are silently dropped when the stack is full. 
-     This might be quite hard to find in production. */
-  assert(spaceAvailable); 
-  if (likely(spaceAvailable))
-    context->instID[context->instStackSize++] = instanceId;
-  return spaceAvailable;
-#else
-  const bool spaceAvailable = (context->instID[0] == RTC_INVALID_GEOMETRY_ID);
-  assert(spaceAvailable); 
-  if (likely(spaceAvailable))
-    context->instID[0] = instanceId;
-  return spaceAvailable;
-#endif
-}
-
-
-/* 
- * Pop the last instance pushed to the stack. 
- * Do not call on an empty stack. 
- */
-RTC_FORCEINLINE void pop(RTCIntersectContext* context)
-{
-  assert(context);
-#if RTC_MAX_INSTANCE_LEVEL_COUNT > 1
-  assert(context->instStackSize > 0);
-  context->instID[--context->instStackSize] = RTC_INVALID_GEOMETRY_ID;
-#else
-  assert(context->instID[0] != RTC_INVALID_GEOMETRY_ID);
-  context->instID[0] = RTC_INVALID_GEOMETRY_ID;
-#endif
-}
-
-/*******************************************************************************
- * Optimized instance id stack copy.
- * The copy() function at the bottom of this block will either copy full
- * stacks or copy only until the last valid element has been copied, depending
- * on RTC_MAX_INSTANCE_LEVEL_COUNT.
- ******************************************************************************/
-
-/*
- * Plain array assignment. This works for scalar->scalar,
- * scalar->vector, and vector->vector.
- */
-template <class Src, class Tgt>
-RTC_FORCEINLINE void level_copy(unsigned level, Src* src, Tgt* tgt)
-{
-  tgt[level] = src[level];
-}
-
-/*
- * Masked SIMD vector->vector store.
- */
-template <int K>
-RTC_FORCEINLINE void level_copy(unsigned level, const vuint<K>* src, vuint<K>* tgt, const vbool<K>& mask)
-{
-  vuint<K>::storeu(mask, tgt + level, src[level]);
-}
-
-/*
- * Masked scalar->SIMD vector store.
- */
-template <int K>
-RTC_FORCEINLINE void level_copy(unsigned level, const unsigned* src, vuint<K>* tgt, const vbool<K>& mask)
-{
-  vuint<K>::store(mask, tgt + level, src[level]);
-}
-
-/*
- * Indexed assign from vector to scalar.
- */
-template <int K>
-RTC_FORCEINLINE void level_copy(unsigned level, const vuint<K>* src, unsigned* tgt, const size_t& idx)
-{
-  tgt[level] = src[level][idx];
-}
-
-/*
- * Indexed assign from scalar to vector.
- */
-template <int K>
-RTC_FORCEINLINE void level_copy(unsigned level, const unsigned* src, vuint<K>* tgt, const size_t& idx)
-{
-  tgt[level][idx] = src[level];
-}
-
-/*
- * Indexed assign from vector to vector.
- */
-template <int K>
-RTC_FORCEINLINE void level_copy(unsigned level, const vuint<K>* src, vuint<K>* tgt, const size_t& i, const size_t& j)
-{
-  tgt[level][j] = src[level][i];
-}
-
-/*
- * Check if the given stack level is valid.
- * These are only used for large max stack sizes.
- */
-RTC_FORCEINLINE bool level_valid(unsigned level, const unsigned* stack)
-{
-  return stack[level] != RTC_INVALID_GEOMETRY_ID;
-}
-RTC_FORCEINLINE bool level_valid(unsigned level, const unsigned* stack, const size_t& /*i*/)
-{
-  return stack[level] != RTC_INVALID_GEOMETRY_ID;
-}
-template <int K>
-RTC_FORCEINLINE bool level_valid(unsigned level, const unsigned* stack, const vbool<K>& /*mask*/)
-{
-  return stack[level] != RTC_INVALID_GEOMETRY_ID;
-}
-
-template <int K>
-RTC_FORCEINLINE bool level_valid(unsigned level, const vuint<K>* stack)
-{
-  return any(stack[level] != RTC_INVALID_GEOMETRY_ID);
-}
-template <int K>
-RTC_FORCEINLINE bool level_valid(unsigned level, const vuint<K>* stack, const vbool<K>& mask)
-{
-  return any(mask & (stack[level] != RTC_INVALID_GEOMETRY_ID));
-}
-
-template <int K>
-RTC_FORCEINLINE bool level_valid(unsigned level, const vuint<K>* stack, const size_t& i)
-{
-  return stack[level][i] != RTC_INVALID_GEOMETRY_ID;
-}
-template <int K>
-RTC_FORCEINLINE bool level_valid(unsigned level, const vuint<K>* stack, const size_t& i, const size_t& /*j*/)
-{
-  return stack[level][i] != RTC_INVALID_GEOMETRY_ID;
-}
-
-/*
- * Copy an instance ID stack.
- *
- * This function automatically selects a LevelFunctor from the above Assign 
- * structs.
- */
-template <class Src, class Tgt, class... Args>
-RTC_FORCEINLINE void copy(Src src, Tgt tgt, Args&&... args)
-{
-#if (RTC_MAX_INSTANCE_LEVEL_COUNT == 1)
-  /* 
-   * Avoid all loops for only one level. 
-   */
-  level_copy(0, src, tgt, std::forward<Args>(args)...);
-
-#elif (RTC_MAX_INSTANCE_LEVEL_COUNT <= 4)
-  /* 
-   * It is faster to avoid the valid test for low level counts.
-   * Just copy the whole stack.
-   */
-  for (unsigned l = 0; l < RTC_MAX_INSTANCE_LEVEL_COUNT; ++l)
-    level_copy(l, src, tgt, std::forward<Args>(args)...);
-
-#else
-  /* 
-   * For general stack sizes, it pays off to test for validity.
-   */
-  bool valid = true;
-  for (unsigned l = 0; l < RTC_MAX_INSTANCE_LEVEL_COUNT && valid; ++l)
-  {
-    level_copy(l, src, tgt, std::forward<Args>(args)...);
-    valid = level_valid(l, src, std::forward<Args>(args)...);
-  }
-#endif
-}
-
-} // namespace instance_id_stack
-} // namespace embree
-

thirdparty/embree-aarch64/kernels/common/scene_curves.h

@@ -1,341 +0,0 @@
-// Copyright 2009-2020 Intel Corporation
-// SPDX-License-Identifier: Apache-2.0
-
-#pragma once
-
-#include "default.h"
-#include "geometry.h"
-#include "buffer.h"
-
-namespace embree
-{
-  /*! represents an array of bicubic bezier curves */
-  struct CurveGeometry : public Geometry
-  {
-    /*! type of this geometry */
-    static const Geometry::GTypeMask geom_type = Geometry::MTY_CURVE4;
-
-  public:
-    
-    /*! bezier curve construction */
-    CurveGeometry (Device* device, Geometry::GType gtype);
-    
-  public:
-    void setMask(unsigned mask);
-    void setNumTimeSteps (unsigned int numTimeSteps);
-    void setVertexAttributeCount (unsigned int N);
-    void setBuffer(RTCBufferType type, unsigned int slot, RTCFormat format, const Ref<Buffer>& buffer, size_t offset, size_t stride, unsigned int num);
-    void* getBuffer(RTCBufferType type, unsigned int slot);
-    void updateBuffer(RTCBufferType type, unsigned int slot);
-    void commit();
-    bool verify();
-    void setTessellationRate(float N);
-    void setMaxRadiusScale(float s);
-    void addElementsToCount (GeometryCounts & counts) const;
-
-  public:
-    
-    /*! returns the number of vertices */
-    __forceinline size_t numVertices() const {
-      return vertices[0].size();
-    }
-
-    /*! returns the i'th curve */
-    __forceinline const unsigned int& curve(size_t i) const {
-      return curves[i];
-    }
-
-    /*! returns i'th vertex of the first time step */
-    __forceinline Vec3ff vertex(size_t i) const {
-      return vertices0[i];
-    }
-
-    /*! returns i'th normal of the first time step */
-    __forceinline Vec3fa normal(size_t i) const {
-      return normals0[i];
-    }
-
-    /*! returns i'th tangent of the first time step */
-    __forceinline Vec3ff tangent(size_t i) const {
-      return tangents0[i];
-    }
-
-    /*! returns i'th normal derivative of the first time step */
-    __forceinline Vec3fa dnormal(size_t i) const {
-      return dnormals0[i];
-    }
-
-    /*! returns i'th radius of the first time step */
-    __forceinline float radius(size_t i) const {
-      return vertices0[i].w;
-    }
-
-    /*! returns i'th vertex of itime'th timestep */
-    __forceinline Vec3ff vertex(size_t i, size_t itime) const {
-      return vertices[itime][i];
-    }
-
-    /*! returns i'th normal of itime'th timestep */
-    __forceinline Vec3fa normal(size_t i, size_t itime) const {
-      return normals[itime][i];
-    }
-
-    /*! returns i'th tangent of itime'th timestep */
-    __forceinline Vec3ff tangent(size_t i, size_t itime) const {
-      return tangents[itime][i];
-    }
-
-    /*! returns i'th normal derivative of itime'th timestep */
-    __forceinline Vec3fa dnormal(size_t i, size_t itime) const {
-      return dnormals[itime][i];
-    }
-
-    /*! returns i'th radius of itime'th timestep */
-    __forceinline float radius(size_t i, size_t itime) const {
-      return vertices[itime][i].w;
-    }
-
-    /*! gathers the curve starting with i'th vertex */
-    __forceinline void gather(Vec3ff& p0, Vec3ff& p1, Vec3ff& p2, Vec3ff& p3, size_t i) const
-    {
-      p0 = vertex(i+0);
-      p1 = vertex(i+1);
-      p2 = vertex(i+2);
-      p3 = vertex(i+3);
-    }
-
-    /*! gathers the curve starting with i'th vertex of itime'th timestep */
-    __forceinline void gather(Vec3ff& p0, Vec3ff& p1, Vec3ff& p2, Vec3ff& p3, size_t i, size_t itime) const
-    {
-      p0 = vertex(i+0,itime);
-      p1 = vertex(i+1,itime);
-      p2 = vertex(i+2,itime);
-      p3 = vertex(i+3,itime);
-    }
-
-    /*! gathers the curve starting with i'th vertex */
-    __forceinline void gather(Vec3ff& p0, Vec3ff& p1, Vec3ff& p2, Vec3ff& p3, Vec3fa& n0, Vec3fa& n1, Vec3fa& n2, Vec3fa& n3, size_t i) const
-    {
-      p0 = vertex(i+0);
-      p1 = vertex(i+1);
-      p2 = vertex(i+2);
-      p3 = vertex(i+3);
-      n0 = normal(i+0);
-      n1 = normal(i+1);
-      n2 = normal(i+2);
-      n3 = normal(i+3);
-    }
-
-    /*! gathers the curve starting with i'th vertex of itime'th timestep */
-    __forceinline void gather(Vec3ff& p0, Vec3ff& p1, Vec3ff& p2, Vec3ff& p3, Vec3fa& n0, Vec3fa& n1, Vec3fa& n2, Vec3fa& n3, size_t i, size_t itime) const
-    {
-      p0 = vertex(i+0,itime);
-      p1 = vertex(i+1,itime);
-      p2 = vertex(i+2,itime);
-      p3 = vertex(i+3,itime);
-      n0 = normal(i+0,itime);
-      n1 = normal(i+1,itime);
-      n2 = normal(i+2,itime);
-      n3 = normal(i+3,itime);
-    }
-
-    /*! prefetches the curve starting with i'th vertex of itime'th timestep */
-    __forceinline void prefetchL1_vertices(size_t i) const
-    {
-      prefetchL1(vertices0.getPtr(i)+0);
-      prefetchL1(vertices0.getPtr(i)+64);
-    }
-
-    /*! prefetches the curve starting with i'th vertex of itime'th timestep */
-    __forceinline void prefetchL2_vertices(size_t i) const
-    {
-      prefetchL2(vertices0.getPtr(i)+0);
-      prefetchL2(vertices0.getPtr(i)+64);
-    }  
-
-    /*! loads curve vertices for specified time */
-    __forceinline void gather(Vec3ff& p0, Vec3ff& p1, Vec3ff& p2, Vec3ff& p3, size_t i, float time) const
-    {
-      float ftime;
-      const size_t itime = timeSegment(time, ftime);
-
-      const float t0 = 1.0f - ftime;
-      const float t1 = ftime;
-      Vec3ff a0,a1,a2,a3;
-      gather(a0,a1,a2,a3,i,itime);
-      Vec3ff b0,b1,b2,b3;
-      gather(b0,b1,b2,b3,i,itime+1);
-      p0 = madd(Vec3ff(t0),a0,t1*b0);
-      p1 = madd(Vec3ff(t0),a1,t1*b1);
-      p2 = madd(Vec3ff(t0),a2,t1*b2);
-      p3 = madd(Vec3ff(t0),a3,t1*b3);
-    }
-
-    /*! loads curve vertices for specified time */
-    __forceinline void gather(Vec3ff& p0, Vec3ff& p1, Vec3ff& p2, Vec3ff& p3, Vec3fa& n0, Vec3fa& n1, Vec3fa& n2, Vec3fa& n3, size_t i, float time) const
-    {
-      float ftime;
-      const size_t itime = timeSegment(time, ftime);
-
-      const float t0 = 1.0f - ftime;
-      const float t1 = ftime;
-      Vec3ff a0,a1,a2,a3; Vec3fa an0,an1,an2,an3;
-      gather(a0,a1,a2,a3,an0,an1,an2,an3,i,itime);
-      Vec3ff b0,b1,b2,b3; Vec3fa bn0,bn1,bn2,bn3;
-      gather(b0,b1,b2,b3,bn0,bn1,bn2,bn3,i,itime+1);
-      p0 = madd(Vec3ff(t0),a0,t1*b0);
-      p1 = madd(Vec3ff(t0),a1,t1*b1);
-      p2 = madd(Vec3ff(t0),a2,t1*b2);
-      p3 = madd(Vec3ff(t0),a3,t1*b3);
-      n0 = madd(Vec3ff(t0),an0,t1*bn0);
-      n1 = madd(Vec3ff(t0),an1,t1*bn1);
-      n2 = madd(Vec3ff(t0),an2,t1*bn2);
-      n3 = madd(Vec3ff(t0),an3,t1*bn3);
-    }
-
-    template<typename SourceCurve3ff, typename SourceCurve3fa, typename TensorLinearCubicBezierSurface3fa>
-    __forceinline TensorLinearCubicBezierSurface3fa getNormalOrientedCurve(IntersectContext* context, const Vec3fa& ray_org, const unsigned int primID, const size_t itime) const
-    {
-      Vec3ff v0,v1,v2,v3; Vec3fa n0,n1,n2,n3;
-      unsigned int vertexID = curve(primID);
-      gather(v0,v1,v2,v3,n0,n1,n2,n3,vertexID,itime);
-      SourceCurve3ff ccurve(v0,v1,v2,v3);
-      SourceCurve3fa ncurve(n0,n1,n2,n3);
-      ccurve = enlargeRadiusToMinWidth(context,this,ray_org,ccurve);
-      return TensorLinearCubicBezierSurface3fa::fromCenterAndNormalCurve(ccurve,ncurve);
-    }
-
-    template<typename SourceCurve3ff, typename SourceCurve3fa, typename TensorLinearCubicBezierSurface3fa>
-    __forceinline TensorLinearCubicBezierSurface3fa getNormalOrientedCurve(IntersectContext* context, const Vec3fa& ray_org, const unsigned int primID, const float time) const
-    {
-      float ftime;
-      const size_t itime = timeSegment(time, ftime);
-      const TensorLinearCubicBezierSurface3fa curve0 = getNormalOrientedCurve<SourceCurve3ff, SourceCurve3fa, TensorLinearCubicBezierSurface3fa>(context,ray_org,primID,itime+0);
-      const TensorLinearCubicBezierSurface3fa curve1 = getNormalOrientedCurve<SourceCurve3ff, SourceCurve3fa, TensorLinearCubicBezierSurface3fa>(context,ray_org,primID,itime+1);
-      return clerp(curve0,curve1,ftime);
-    }
-
-    /*! gathers the hermite curve starting with i'th vertex */
-    __forceinline void gather_hermite(Vec3ff& p0, Vec3ff& t0, Vec3ff& p1, Vec3ff& t1, size_t i) const
-    {
-      p0 = vertex (i+0);
-      p1 = vertex (i+1);
-      t0 = tangent(i+0);
-      t1 = tangent(i+1);
-    }
-
-    /*! gathers the hermite curve starting with i'th vertex of itime'th timestep */
-    __forceinline void gather_hermite(Vec3ff& p0, Vec3ff& t0, Vec3ff& p1, Vec3ff& t1, size_t i, size_t itime) const
-    {
-      p0 = vertex (i+0,itime);
-      p1 = vertex (i+1,itime);
-      t0 = tangent(i+0,itime);
-      t1 = tangent(i+1,itime);
-    }
-
-    /*! loads curve vertices for specified time */
-    __forceinline void gather_hermite(Vec3ff& p0, Vec3ff& t0, Vec3ff& p1, Vec3ff& t1, size_t i, float time) const
-    {
-      float ftime;
-      const size_t itime = timeSegment(time, ftime);
-      const float f0 = 1.0f - ftime, f1 = ftime;
-      Vec3ff ap0,at0,ap1,at1;
-      gather_hermite(ap0,at0,ap1,at1,i,itime);
-      Vec3ff bp0,bt0,bp1,bt1;
-      gather_hermite(bp0,bt0,bp1,bt1,i,itime+1);
-      p0 = madd(Vec3ff(f0),ap0,f1*bp0);
-      t0 = madd(Vec3ff(f0),at0,f1*bt0);
-      p1 = madd(Vec3ff(f0),ap1,f1*bp1);
-      t1 = madd(Vec3ff(f0),at1,f1*bt1);
-    }
-
-    /*! gathers the hermite curve starting with i'th vertex */
-    __forceinline void gather_hermite(Vec3ff& p0, Vec3ff& t0, Vec3fa& n0, Vec3fa& dn0, Vec3ff& p1, Vec3ff& t1, Vec3fa& n1, Vec3fa& dn1, size_t i) const
-    {
-      p0 = vertex (i+0);
-      p1 = vertex (i+1);
-      t0 = tangent(i+0);
-      t1 = tangent(i+1);
-      n0 = normal(i+0);
-      n1 = normal(i+1);
-      dn0 = dnormal(i+0);
-      dn1 = dnormal(i+1);
-    }
-
-    /*! gathers the hermite curve starting with i'th vertex of itime'th timestep */
-    __forceinline void gather_hermite(Vec3ff& p0, Vec3ff& t0, Vec3fa& n0, Vec3fa& dn0, Vec3ff& p1, Vec3ff& t1, Vec3fa& n1, Vec3fa& dn1, size_t i, size_t itime) const
-    {
-      p0 = vertex (i+0,itime);
-      p1 = vertex (i+1,itime);
-      t0 = tangent(i+0,itime);
-      t1 = tangent(i+1,itime);
-      n0 = normal(i+0,itime);
-      n1 = normal(i+1,itime);
-      dn0 = dnormal(i+0,itime);
-      dn1 = dnormal(i+1,itime);
-    }
-
-    /*! loads curve vertices for specified time */
-    __forceinline void gather_hermite(Vec3ff& p0, Vec3fa& t0, Vec3fa& n0, Vec3fa& dn0, Vec3ff& p1, Vec3fa& t1, Vec3fa& n1, Vec3fa& dn1, size_t i, float time) const
-    {
-      float ftime;
-      const size_t itime = timeSegment(time, ftime);
-      const float f0 = 1.0f - ftime, f1 = ftime;
-      Vec3ff ap0,at0,ap1,at1; Vec3fa an0,adn0,an1,adn1;
-      gather_hermite(ap0,at0,an0,adn0,ap1,at1,an1,adn1,i,itime);
-      Vec3ff bp0,bt0,bp1,bt1; Vec3fa bn0,bdn0,bn1,bdn1;
-      gather_hermite(bp0,bt0,bn0,bdn0,bp1,bt1,bn1,bdn1,i,itime+1);
-      p0 = madd(Vec3ff(f0),ap0,f1*bp0);
-      t0 = madd(Vec3ff(f0),at0,f1*bt0);
-      n0 = madd(Vec3ff(f0),an0,f1*bn0);
-      dn0= madd(Vec3ff(f0),adn0,f1*bdn0);
-      p1 = madd(Vec3ff(f0),ap1,f1*bp1);
-      t1 = madd(Vec3ff(f0),at1,f1*bt1);
-      n1 = madd(Vec3ff(f0),an1,f1*bn1);
-      dn1= madd(Vec3ff(f0),adn1,f1*bdn1);
-    }
-
-    template<typename SourceCurve3ff, typename SourceCurve3fa, typename TensorLinearCubicBezierSurface3fa>
-      __forceinline TensorLinearCubicBezierSurface3fa getNormalOrientedHermiteCurve(IntersectContext* context, const Vec3fa& ray_org, const unsigned int primID, const size_t itime) const
-    {
-      Vec3ff v0,t0,v1,t1; Vec3fa n0,dn0,n1,dn1;
-      unsigned int vertexID = curve(primID);
-      gather_hermite(v0,t0,n0,dn0,v1,t1,n1,dn1,vertexID,itime);
-
-      SourceCurve3ff ccurve(v0,t0,v1,t1);
-      SourceCurve3fa ncurve(n0,dn0,n1,dn1);
-      ccurve = enlargeRadiusToMinWidth(context,this,ray_org,ccurve);
-      return TensorLinearCubicBezierSurface3fa::fromCenterAndNormalCurve(ccurve,ncurve);
-    }
-
-    template<typename SourceCurve3ff, typename SourceCurve3fa, typename TensorLinearCubicBezierSurface3fa>
-    __forceinline TensorLinearCubicBezierSurface3fa getNormalOrientedHermiteCurve(IntersectContext* context, const Vec3fa& ray_org, const unsigned int primID, const float time) const
-    {
-      float ftime;
-      const size_t itime = timeSegment(time, ftime);
-      const TensorLinearCubicBezierSurface3fa curve0 = getNormalOrientedHermiteCurve<SourceCurve3ff, SourceCurve3fa, TensorLinearCubicBezierSurface3fa>(context, ray_org, primID,itime+0);
-      const TensorLinearCubicBezierSurface3fa curve1 = getNormalOrientedHermiteCurve<SourceCurve3ff, SourceCurve3fa, TensorLinearCubicBezierSurface3fa>(context, ray_org, primID,itime+1);
-      return clerp(curve0,curve1,ftime);
-    }
-
-  private:
-    void resizeBuffers(unsigned int numSteps);
-
-  public:
-    BufferView<unsigned int> curves;        //!< array of curve indices
-    BufferView<Vec3ff> vertices0;           //!< fast access to first vertex buffer
-    BufferView<Vec3fa> normals0;            //!< fast access to first normal buffer
-    BufferView<Vec3ff> tangents0;           //!< fast access to first tangent buffer
-    BufferView<Vec3fa> dnormals0;           //!< fast access to first normal derivative buffer
-    vector<BufferView<Vec3ff>> vertices;    //!< vertex array for each timestep
-    vector<BufferView<Vec3fa>> normals;     //!< normal array for each timestep
-    vector<BufferView<Vec3ff>> tangents;    //!< tangent array for each timestep
-    vector<BufferView<Vec3fa>> dnormals;    //!< normal derivative array for each timestep
-    BufferView<char> flags;                 //!< start, end flag per segment
-    vector<BufferView<char>> vertexAttribs; //!< user buffers
-    int tessellationRate;                   //!< tessellation rate for flat curve
-    float maxRadiusScale = 1.0;             //!< maximal min-width scaling of curve radii
-  };
-  
-  DECLARE_ISA_FUNCTION(CurveGeometry*, createCurves, Device* COMMA Geometry::GType);
-}

thirdparty/embree-aarch64/kernels/geometry/curve_intersector_virtual_bezier_curve.h

@@ -1,21 +0,0 @@
-// 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-aarch64/kernels/geometry/curve_intersector_virtual_bspline_curve.h

@@ -1,21 +0,0 @@
-// 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-aarch64/kernels/geometry/curve_intersector_virtual_catmullrom_curve.h

@@ -1,21 +0,0 @@
-// 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-aarch64/kernels/geometry/curve_intersector_virtual_hermite_curve.h

@@ -1,21 +0,0 @@
-// 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-aarch64/kernels/geometry/curve_intersector_virtual_linear_curve.h

@@ -1,21 +0,0 @@
-// 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-aarch64/kernels/geometry/curve_intersector_virtual_point.h

@@ -1,22 +0,0 @@
-// 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-aarch64/kernels/geometry/subgrid_intersector_moeller.h

@@ -1,493 +0,0 @@
-// Copyright 2009-2020 Intel Corporation
-// SPDX-License-Identifier: Apache-2.0
-
-#pragma once
-
-#include "subgrid.h"
-#include "quad_intersector_moeller.h"
-
-namespace embree
-{
-  namespace isa
-  {
-
-    /* ----------------------------- */
-    /* -- single ray intersectors -- */
-    /* ----------------------------- */
-
-    template<int M>
-      __forceinline void interpolateUV(MoellerTrumboreHitM<M> &hit,const GridMesh::Grid &g, const SubGrid& subgrid) 
-    {
-      /* correct U,V interpolation across the entire grid */
-      const vint<M> sx((int)subgrid.x());
-      const vint<M> sy((int)subgrid.y());
-      const vint<M> sxM(sx + vint<M>(0,1,1,0));
-      const vint<M> syM(sy + vint<M>(0,0,1,1));
-      const float inv_resX = rcp((float)((int)g.resX-1));
-      const float inv_resY = rcp((float)((int)g.resY-1));          
-      hit.U = (hit.U + (vfloat<M>)sxM * hit.absDen) * inv_resX;
-      hit.V = (hit.V + (vfloat<M>)syM * hit.absDen) * inv_resY;
-    }
-    
-    template<int M, bool filter>
-      struct SubGridQuadMIntersector1MoellerTrumbore;
-
-    template<int M, bool filter>
-      struct SubGridQuadMIntersector1MoellerTrumbore
-      {
-        __forceinline SubGridQuadMIntersector1MoellerTrumbore() {}
-
-        __forceinline SubGridQuadMIntersector1MoellerTrumbore(const Ray& ray, const void* ptr) {}
-
-        __forceinline void intersect(RayHit& ray, IntersectContext* context,
-                                     const Vec3vf<M>& v0, const Vec3vf<M>& v1, const Vec3vf<M>& v2, const Vec3vf<M>& v3,
-                                     const GridMesh::Grid &g, const SubGrid& subgrid) const
-        {
-          MoellerTrumboreHitM<M> hit;
-          MoellerTrumboreIntersector1<M> intersector(ray,nullptr);
-          Intersect1EpilogMU<M,filter> epilog(ray,context,subgrid.geomID(),subgrid.primID());
-
-          /* intersect first triangle */
-          if (intersector.intersect(ray,v0,v1,v3,hit)) 
-          {
-            interpolateUV<M>(hit,g,subgrid);
-            epilog(hit.valid,hit);
-          }
-
-          /* intersect second triangle */
-          if (intersector.intersect(ray,v2,v3,v1,hit)) 
-          {
-            hit.U = hit.absDen - hit.U;
-            hit.V = hit.absDen - hit.V;
-            interpolateUV<M>(hit,g,subgrid);
-            epilog(hit.valid,hit);
-          }
-        }
-      
-        __forceinline bool occluded(Ray& ray, IntersectContext* context,
-                                    const Vec3vf<M>& v0, const Vec3vf<M>& v1, const Vec3vf<M>& v2, const Vec3vf<M>& v3,
-                                    const GridMesh::Grid &g, const SubGrid& subgrid) const
-        {
-          MoellerTrumboreHitM<M> hit;
-          MoellerTrumboreIntersector1<M> intersector(ray,nullptr);
-          Occluded1EpilogMU<M,filter> epilog(ray,context,subgrid.geomID(),subgrid.primID());
-          
-          /* intersect first triangle */
-          if (intersector.intersect(ray,v0,v1,v3,hit)) 
-          {
-            interpolateUV<M>(hit,g,subgrid);
-            if (epilog(hit.valid,hit))
-              return true;
-          }
-
-          /* intersect second triangle */
-          if (intersector.intersect(ray,v2,v3,v1,hit)) 
-          {
-            hit.U = hit.absDen - hit.U;
-            hit.V = hit.absDen - hit.V;
-            interpolateUV<M>(hit,g,subgrid);
-            if (epilog(hit.valid,hit))
-              return true;
-          }
-          return false;
-        }
-      };
-
-#if defined (__AVX__)
-
-    /*! Intersects 4 quads with 1 ray using AVX */
-    template<bool filter>
-      struct SubGridQuadMIntersector1MoellerTrumbore<4,filter>
-    {
-      __forceinline SubGridQuadMIntersector1MoellerTrumbore() {}
-
-      __forceinline SubGridQuadMIntersector1MoellerTrumbore(const Ray& ray, const void* ptr) {}
-      
-      template<typename Epilog>
-        __forceinline bool intersect(Ray& ray, const Vec3vf4& v0, const Vec3vf4& v1, const Vec3vf4& v2, const Vec3vf4& v3, const GridMesh::Grid &g, const SubGrid& subgrid, const Epilog& epilog) const
-      {
-        const Vec3vf8 vtx0(vfloat8(v0.x,v2.x),vfloat8(v0.y,v2.y),vfloat8(v0.z,v2.z));
-#if !defined(EMBREE_BACKFACE_CULLING)
-        const Vec3vf8 vtx1(vfloat8(v1.x),vfloat8(v1.y),vfloat8(v1.z));
-        const Vec3vf8 vtx2(vfloat8(v3.x),vfloat8(v3.y),vfloat8(v3.z));        
-#else
-        const Vec3vf8 vtx1(vfloat8(v1.x,v3.x),vfloat8(v1.y,v3.y),vfloat8(v1.z,v3.z));
-        const Vec3vf8 vtx2(vfloat8(v3.x,v1.x),vfloat8(v3.y,v1.y),vfloat8(v3.z,v1.z));
-#endif
-        MoellerTrumboreHitM<8> hit;
-        MoellerTrumboreIntersector1<8> intersector(ray,nullptr);
-        const vbool8 flags(0,0,0,0,1,1,1,1);
-        if (unlikely(intersector.intersect(ray,vtx0,vtx1,vtx2,hit)))
-        {
-          vfloat8 U = hit.U, V = hit.V, absDen = hit.absDen;
-
-#if !defined(EMBREE_BACKFACE_CULLING)
-          hit.U = select(flags,absDen-V,U);
-          hit.V = select(flags,absDen-U,V);
-          hit.vNg *= select(flags,vfloat8(-1.0f),vfloat8(1.0f)); 
-#else
-          hit.U = select(flags,absDen-U,U);
-          hit.V = select(flags,absDen-V,V);
-#endif
-          /* correct U,V interpolation across the entire grid */
-          const vint8 sx((int)subgrid.x());
-          const vint8 sy((int)subgrid.y());
-          const vint8 sx8(sx + vint8(0,1,1,0,0,1,1,0));
-          const vint8 sy8(sy + vint8(0,0,1,1,0,0,1,1));
-          const float inv_resX = rcp((float)((int)g.resX-1));
-          const float inv_resY = rcp((float)((int)g.resY-1));          
-          hit.U = (hit.U + (vfloat8)sx8 * absDen) * inv_resX;
-          hit.V = (hit.V + (vfloat8)sy8 * absDen) * inv_resY;          
-
-          if (unlikely(epilog(hit.valid,hit)))
-            return true;
-        }
-        return false;
-      }
-      
-      __forceinline bool intersect(RayHit& ray, IntersectContext* context,
-                                   const Vec3vf4& v0, const Vec3vf4& v1, const Vec3vf4& v2, const Vec3vf4& v3, 
-                                   const GridMesh::Grid &g, const SubGrid& subgrid) const
-      {
-          return intersect(ray,v0,v1,v2,v3,g,subgrid,Intersect1EpilogMU<8,filter>(ray,context,subgrid.geomID(),subgrid.primID()));
-      }
-      
-      __forceinline bool occluded(Ray& ray, IntersectContext* context,
-                                  const Vec3vf4& v0, const Vec3vf4& v1, const Vec3vf4& v2, const Vec3vf4& v3, 
-                                  const GridMesh::Grid &g, const SubGrid& subgrid) const
-      {
-          return intersect(ray,v0,v1,v2,v3,g,subgrid,Occluded1EpilogMU<8,filter>(ray,context,subgrid.geomID(),subgrid.primID()));
-      }
-    };
-
-#endif
-
-    // ============================================================================================================================
-    // ============================================================================================================================
-    // ============================================================================================================================
-
-
-    /* ----------------------------- */
-    /* -- ray packet intersectors -- */
-    /* ----------------------------- */
-
-    template<int K>
-      struct SubGridQuadHitK
-      {
-        __forceinline SubGridQuadHitK(const vfloat<K>& U,
-                                      const vfloat<K>& V,
-                                      const vfloat<K>& T,
-                                      const vfloat<K>& absDen,
-                                      const Vec3vf<K>& Ng,
-                                      const vbool<K>& flags,
-                                      const GridMesh::Grid &g, 
-                                      const SubGrid& subgrid,
-                                      const unsigned int i)
-        : U(U), V(V), T(T), absDen(absDen), flags(flags), tri_Ng(Ng), g(g), subgrid(subgrid), i(i) {}
-
-        __forceinline std::tuple<vfloat<K>,vfloat<K>,vfloat<K>,Vec3vf<K>> operator() () const
-        {
-          const vfloat<K> rcpAbsDen = rcp(absDen);
-          const vfloat<K> t = T * rcpAbsDen;
-          const vfloat<K> u0 = min(U * rcpAbsDen,1.0f);
-          const vfloat<K> v0 = min(V * rcpAbsDen,1.0f);
-          const vfloat<K> u1 = vfloat<K>(1.0f) - u0;
-          const vfloat<K> v1 = vfloat<K>(1.0f) - v0;
-          const vfloat<K> uu = select(flags,u1,u0);
-          const vfloat<K> vv = select(flags,v1,v0);
-          const unsigned int sx = subgrid.x() + (unsigned int)(i % 2);
-          const unsigned int sy = subgrid.y() + (unsigned int)(i >>1);
-          const float inv_resX = rcp((float)(int)(g.resX-1));
-          const float inv_resY = rcp((float)(int)(g.resY-1));
-          const vfloat<K> u = (uu + (float)(int)sx) * inv_resX;
-          const vfloat<K> v = (vv + (float)(int)sy) * inv_resY;
-          const Vec3vf<K> Ng(tri_Ng.x,tri_Ng.y,tri_Ng.z);
-          return std::make_tuple(u,v,t,Ng);
-        }
-
-      private:
-        const vfloat<K> U;
-        const vfloat<K> V;
-        const vfloat<K> T;
-        const vfloat<K> absDen;
-        const vbool<K> flags;
-        const Vec3vf<K> tri_Ng;
-
-        const GridMesh::Grid &g;
-        const SubGrid& subgrid;
-        const size_t i;
-      };
-
-    template<int M, int K, bool filter>
-      struct SubGridQuadMIntersectorKMoellerTrumboreBase
-      {
-        __forceinline SubGridQuadMIntersectorKMoellerTrumboreBase(const vbool<K>& valid, const RayK<K>& ray) {}
-            
-        template<typename Epilog>
-        __forceinline vbool<K> intersectK(const vbool<K>& valid0,
-                                          RayK<K>& ray,
-                                          const Vec3vf<K>& tri_v0,
-                                          const Vec3vf<K>& tri_e1,
-                                          const Vec3vf<K>& tri_e2,
-                                          const Vec3vf<K>& tri_Ng,
-                                          const vbool<K>& flags,
-                                          const GridMesh::Grid &g, 
-                                          const SubGrid &subgrid,
-                                          const unsigned int i,
-                                          const Epilog& epilog) const
-        { 
-          /* calculate denominator */
-          vbool<K> valid = valid0;
-          const Vec3vf<K> C = tri_v0 - ray.org;
-          const Vec3vf<K> R = cross(C,ray.dir);
-          const vfloat<K> den = dot(tri_Ng,ray.dir);
-          const vfloat<K> absDen = abs(den);
-          const vfloat<K> sgnDen = signmsk(den);
-        
-          /* test against edge p2 p0 */
-          const vfloat<K> U = dot(R,tri_e2) ^ sgnDen;
-          valid &= U >= 0.0f;
-          if (likely(none(valid))) return false;
-        
-          /* test against edge p0 p1 */
-          const vfloat<K> V = dot(R,tri_e1) ^ sgnDen;
-          valid &= V >= 0.0f;
-          if (likely(none(valid))) return false;
-        
-          /* test against edge p1 p2 */
-          const vfloat<K> W = absDen-U-V;
-          valid &= W >= 0.0f;
-          if (likely(none(valid))) return false;
-        
-          /* perform depth test */
-          const vfloat<K> T = dot(tri_Ng,C) ^ sgnDen;
-          valid &= (absDen*ray.tnear() < T) & (T <= absDen*ray.tfar);
-          if (unlikely(none(valid))) return false;
-        
-          /* perform backface culling */
-#if defined(EMBREE_BACKFACE_CULLING)
-          valid &= den < vfloat<K>(zero);
-          if (unlikely(none(valid))) return false;
-#else
-          valid &= den != vfloat<K>(zero);
-          if (unlikely(none(valid))) return false;
-#endif
-        
-          /* calculate hit information */
-          SubGridQuadHitK<K> hit(U,V,T,absDen,tri_Ng,flags,g,subgrid,i);
-          return epilog(valid,hit);
-        }
-      
-        template<typename Epilog>
-        __forceinline vbool<K> intersectK(const vbool<K>& valid0, 
-                                          RayK<K>& ray,
-                                          const Vec3vf<K>& tri_v0,
-                                          const Vec3vf<K>& tri_v1,
-                                          const Vec3vf<K>& tri_v2,
-                                          const vbool<K>& flags,
-                                          const GridMesh::Grid &g, 
-                                          const SubGrid &subgrid,
-                                          const unsigned int i,
-                                          const Epilog& epilog) const
-        {
-          const Vec3vf<K> e1 = tri_v0-tri_v1;
-          const Vec3vf<K> e2 = tri_v2-tri_v0;
-          const Vec3vf<K> Ng = cross(e2,e1);
-          return intersectK(valid0,ray,tri_v0,e1,e2,Ng,flags,g,subgrid,i,epilog);
-        }
-
-        template<typename Epilog>
-        __forceinline bool intersectK(const vbool<K>& valid0, 
-                                      RayK<K>& ray,
-                                      const Vec3vf<K>& v0,
-                                      const Vec3vf<K>& v1,
-                                      const Vec3vf<K>& v2,
-                                      const Vec3vf<K>& v3,
-                                      const GridMesh::Grid &g, 
-                                      const SubGrid &subgrid,
-                                      const unsigned int i,
-                                      const Epilog& epilog) const
-        {
-          intersectK(valid0,ray,v0,v1,v3,vbool<K>(false),g,subgrid,i,epilog);
-          if (none(valid0)) return true;
-          intersectK(valid0,ray,v2,v3,v1,vbool<K>(true ),g,subgrid,i,epilog);
-          return none(valid0);
-        }
-
-        static  __forceinline bool intersect1(RayK<K>& ray,
-                                              size_t k,
-                                              const Vec3vf<M>& tri_v0,
-                                              const Vec3vf<M>& tri_e1,
-                                              const Vec3vf<M>& tri_e2,
-                                              const Vec3vf<M>& tri_Ng,
-                                              MoellerTrumboreHitM<M> &hit)
-        {
-          /* calculate denominator */
-          const Vec3vf<M> O = broadcast<vfloat<M>>(ray.org,k);
-          const Vec3vf<M> D = broadcast<vfloat<M>>(ray.dir,k);
-          const Vec3vf<M> C = Vec3vf<M>(tri_v0) - O;
-          const Vec3vf<M> R = cross(C,D);
-          const vfloat<M> den = dot(Vec3vf<M>(tri_Ng),D);
-          const vfloat<M> absDen = abs(den);
-          const vfloat<M> sgnDen = signmsk(den);
-        
-          /* perform edge tests */
-          const vfloat<M> U = dot(R,Vec3vf<M>(tri_e2)) ^ sgnDen;
-          const vfloat<M> V = dot(R,Vec3vf<M>(tri_e1)) ^ sgnDen;
-        
-          /* perform backface culling */
-#if defined(EMBREE_BACKFACE_CULLING)
-          vbool<M> valid = (den < vfloat<M>(zero)) & (U >= 0.0f) & (V >= 0.0f) & (U+V<=absDen);
-#else
-          vbool<M> valid = (den != vfloat<M>(zero)) & (U >= 0.0f) & (V >= 0.0f) & (U+V<=absDen);
-#endif
-          if (likely(none(valid))) return false;
-        
-          /* perform depth test */
-          const vfloat<M> T = dot(Vec3vf<M>(tri_Ng),C) ^ sgnDen;
-          valid &= (absDen*vfloat<M>(ray.tnear()[k]) < T) & (T <= absDen*vfloat<M>(ray.tfar[k]));
-          if (likely(none(valid))) return false;
-        
-          /* calculate hit information */
-          new (&hit) MoellerTrumboreHitM<M>(valid,U,V,T,absDen,tri_Ng);
-          return true;
-        }
-
-        static __forceinline bool intersect1(RayK<K>& ray,
-                                             size_t k,
-                                             const Vec3vf<M>& v0,
-                                             const Vec3vf<M>& v1,
-                                             const Vec3vf<M>& v2,
-                                             MoellerTrumboreHitM<M> &hit)
-        {
-          const Vec3vf<M> e1 = v0-v1;
-          const Vec3vf<M> e2 = v2-v0;
-          const Vec3vf<M> Ng = cross(e2,e1);
-          return intersect1(ray,k,v0,e1,e2,Ng,hit);
-        }
-
-      };
-
-    template<int M, int K, bool filter>
-      struct SubGridQuadMIntersectorKMoellerTrumbore : public SubGridQuadMIntersectorKMoellerTrumboreBase<M,K,filter>
-    {
-      __forceinline SubGridQuadMIntersectorKMoellerTrumbore(const vbool<K>& valid, const RayK<K>& ray)
-        : SubGridQuadMIntersectorKMoellerTrumboreBase<M,K,filter>(valid,ray) {}
-
-      __forceinline void intersect1(RayHitK<K>& ray, size_t k, IntersectContext* context,
-                                    const Vec3vf<M>& v0, const Vec3vf<M>& v1, const Vec3vf<M>& v2, const Vec3vf<M>& v3, const GridMesh::Grid &g, const SubGrid &subgrid) const
-      {
-        Intersect1KEpilogMU<M,K,filter> epilog(ray,k,context,subgrid.geomID(),subgrid.primID());
-
-        MoellerTrumboreHitM<4> hit;
-        if (SubGridQuadMIntersectorKMoellerTrumboreBase<4,K,filter>::intersect1(ray,k,v0,v1,v3,hit))
-        {
-          interpolateUV<M>(hit,g,subgrid);
-          epilog(hit.valid,hit);
-        }
-
-        if (SubGridQuadMIntersectorKMoellerTrumboreBase<4,K,filter>::intersect1(ray,k,v2,v3,v1,hit))
-        {
-          hit.U = hit.absDen - hit.U;
-          hit.V = hit.absDen - hit.V;
-          interpolateUV<M>(hit,g,subgrid);
-          epilog(hit.valid,hit);
-        }
-
-      }
-      
-      __forceinline bool occluded1(RayK<K>& ray, size_t k, IntersectContext* context,
-                                   const Vec3vf<M>& v0, const Vec3vf<M>& v1, const Vec3vf<M>& v2, const Vec3vf<M>& v3, const GridMesh::Grid &g, const SubGrid &subgrid) const
-      {
-        Occluded1KEpilogMU<M,K,filter> epilog(ray,k,context,subgrid.geomID(),subgrid.primID());
-
-        MoellerTrumboreHitM<4> hit;
-        if (SubGridQuadMIntersectorKMoellerTrumboreBase<4,K,filter>::intersect1(ray,k,v0,v1,v3,hit))
-        {
-          interpolateUV<M>(hit,g,subgrid);
-          if (epilog(hit.valid,hit)) return true;
-        }
-
-        if (SubGridQuadMIntersectorKMoellerTrumboreBase<4,K,filter>::intersect1(ray,k,v2,v3,v1,hit))
-        {
-          hit.U = hit.absDen - hit.U;
-          hit.V = hit.absDen - hit.V;
-          interpolateUV<M>(hit,g,subgrid);
-          if (epilog(hit.valid,hit)) return true;
-        }
-        return false;
-      }
-    };
-
-
-#if defined (__AVX__)
-
-    /*! Intersects 4 quads with 1 ray using AVX */
-    template<int K, bool filter>
-      struct SubGridQuadMIntersectorKMoellerTrumbore<4,K,filter> : public SubGridQuadMIntersectorKMoellerTrumboreBase<4,K,filter>
-    {
-      __forceinline SubGridQuadMIntersectorKMoellerTrumbore(const vbool<K>& valid, const RayK<K>& ray)
-        : SubGridQuadMIntersectorKMoellerTrumboreBase<4,K,filter>(valid,ray) {}
-      
-      template<typename Epilog>
-        __forceinline bool intersect1(RayK<K>& ray, size_t k,const Vec3vf4& v0, const Vec3vf4& v1, const Vec3vf4& v2, const Vec3vf4& v3, 
-                                      const GridMesh::Grid &g, const SubGrid &subgrid, const Epilog& epilog) const
-      {
-        const Vec3vf8 vtx0(vfloat8(v0.x,v2.x),vfloat8(v0.y,v2.y),vfloat8(v0.z,v2.z));
-#if !defined(EMBREE_BACKFACE_CULLING)
-        const Vec3vf8 vtx1(vfloat8(v1.x),vfloat8(v1.y),vfloat8(v1.z));
-        const Vec3vf8 vtx2(vfloat8(v3.x),vfloat8(v3.y),vfloat8(v3.z));
-#else
-        const Vec3vf8 vtx1(vfloat8(v1.x,v3.x),vfloat8(v1.y,v3.y),vfloat8(v1.z,v3.z));
-        const Vec3vf8 vtx2(vfloat8(v3.x,v1.x),vfloat8(v3.y,v1.y),vfloat8(v3.z,v1.z));
-#endif
-        const vbool8 flags(0,0,0,0,1,1,1,1);
-
-        MoellerTrumboreHitM<8> hit;
-        if (SubGridQuadMIntersectorKMoellerTrumboreBase<8,K,filter>::intersect1(ray,k,vtx0,vtx1,vtx2,hit))
-        {
-          vfloat8 U = hit.U, V = hit.V, absDen = hit.absDen;
-#if !defined(EMBREE_BACKFACE_CULLING)
-          hit.U = select(flags,absDen-V,U);
-          hit.V = select(flags,absDen-U,V);
-          hit.vNg *= select(flags,vfloat8(-1.0f),vfloat8(1.0f)); 
-#else
-          hit.U = select(flags,absDen-U,U);
-          hit.V = select(flags,absDen-V,V);
-#endif
-
-          /* correct U,V interpolation across the entire grid */
-          const vint8 sx((int)subgrid.x());
-          const vint8 sy((int)subgrid.y());
-          const vint8 sx8(sx + vint8(0,1,1,0,0,1,1,0));
-          const vint8 sy8(sy + vint8(0,0,1,1,0,0,1,1));
-          const float inv_resX = rcp((float)((int)g.resX-1));
-          const float inv_resY = rcp((float)((int)g.resY-1));          
-          hit.U = (hit.U + (vfloat8)sx8 * absDen) * inv_resX;
-          hit.V = (hit.V + (vfloat8)sy8 * absDen) * inv_resY;          
-          if (unlikely(epilog(hit.valid,hit)))
-            return true;
-
-        }
-        return false;
-      }
-      
-      __forceinline bool intersect1(RayHitK<K>& ray, size_t k, IntersectContext* context,
-                                    const Vec3vf4& v0, const Vec3vf4& v1, const Vec3vf4& v2, const Vec3vf4& v3, const GridMesh::Grid &g, const SubGrid &subgrid) const
-      {
-        return intersect1(ray,k,v0,v1,v2,v3,g,subgrid,Intersect1KEpilogMU<8,K,filter>(ray,k,context,subgrid.geomID(),subgrid.primID()));
-      }
-      
-      __forceinline bool occluded1(RayK<K>& ray, size_t k, IntersectContext* context,
-                                   const Vec3vf4& v0, const Vec3vf4& v1, const Vec3vf4& v2, const Vec3vf4& v3, const GridMesh::Grid &g, const SubGrid &subgrid) const
-      {
-        return intersect1(ray,k,v0,v1,v2,v3,g,subgrid,Occluded1KEpilogMU<8,K,filter>(ray,k,context,subgrid.geomID(),subgrid.primID()));
-      }
-    };
-
-#endif
-
-
-
-  }
-}

thirdparty/embree-aarch64/kernels/geometry/subgrid_intersector_pluecker.h

@@ -1,508 +0,0 @@
-// Copyright 2009-2020 Intel Corporation
-// SPDX-License-Identifier: Apache-2.0
-
-#pragma once
-
-#include "subgrid.h"
-#include "quad_intersector_moeller.h"
-#include "quad_intersector_pluecker.h"
-
-namespace embree
-{
-  namespace isa
-  {
-
-    template<int M>
-    struct SubGridQuadHitPlueckerM
-    {
-      __forceinline SubGridQuadHitPlueckerM() {}
-
-      __forceinline SubGridQuadHitPlueckerM(const vbool<M>& valid,
-                                            const vfloat<M>& U,
-                                            const vfloat<M>& V,
-                                            const vfloat<M>& UVW,
-                                            const vfloat<M>& t,
-                                            const Vec3vf<M>& Ng,
-                                            const vbool<M>& flags) : valid(valid), vt(t)
-      {
-        const vbool<M> invalid = abs(UVW) < min_rcp_input;
-        const vfloat<M> rcpUVW = select(invalid,vfloat<M>(0.0f),rcp(UVW));
-        const vfloat<M> u = min(U * rcpUVW,1.0f);
-        const vfloat<M> v = min(V * rcpUVW,1.0f);
-        const vfloat<M> u1 = vfloat<M>(1.0f) - u;
-        const vfloat<M> v1 = vfloat<M>(1.0f) - v;
-#if !defined(__AVX__) || defined(EMBREE_BACKFACE_CULLING)
-        vu = select(flags,u1,u);
-        vv = select(flags,v1,v);
-        vNg = Vec3vf<M>(Ng.x,Ng.y,Ng.z);
-#else
-        const vfloat<M> flip = select(flags,vfloat<M>(-1.0f),vfloat<M>(1.0f));
-        vv = select(flags,u1,v);
-        vu = select(flags,v1,u);
-        vNg = Vec3vf<M>(flip*Ng.x,flip*Ng.y,flip*Ng.z);
-#endif
-      }
-
-      __forceinline void finalize()
-      {
-      }
-
-      __forceinline Vec2f uv(const size_t i)
-      {
-        const float u = vu[i];
-        const float v = vv[i];
-        return Vec2f(u,v);
-      }
-
-      __forceinline float   t(const size_t i) { return vt[i]; }
-      __forceinline Vec3fa Ng(const size_t i) { return Vec3fa(vNg.x[i],vNg.y[i],vNg.z[i]); }
-
-    public:
-      vbool<M> valid;
-      vfloat<M> vu;
-      vfloat<M> vv;
-      vfloat<M> vt;
-      Vec3vf<M> vNg;
-    };
-
-    template<int M>
-      __forceinline void interpolateUV(SubGridQuadHitPlueckerM<M> &hit,const GridMesh::Grid &g, const SubGrid& subgrid, const vint<M> &stepX, const vint<M> &stepY) 
-    {
-      /* correct U,V interpolation across the entire grid */
-      const vint<M> sx((int)subgrid.x());
-      const vint<M> sy((int)subgrid.y());
-      const vint<M> sxM(sx + stepX);
-      const vint<M> syM(sy + stepY);
-      const float inv_resX = rcp((float)((int)g.resX-1));
-      const float inv_resY = rcp((float)((int)g.resY-1));          
-      hit.vu = (hit.vu + vfloat<M>(sxM)) * inv_resX;
-      hit.vv = (hit.vv + vfloat<M>(syM)) * inv_resY;
-    }
-
-    template<int M>
-    __forceinline static bool intersectPluecker(Ray& ray,
-                                                const Vec3vf<M>& tri_v0,
-                                                const Vec3vf<M>& tri_v1,
-                                                const Vec3vf<M>& tri_v2,
-                                                const vbool<M>& flags,
-                                                SubGridQuadHitPlueckerM<M> &hit)
-    {
-        /* calculate vertices relative to ray origin */
-      const Vec3vf<M> O = Vec3vf<M>((Vec3fa)ray.org);
-      const Vec3vf<M> D = Vec3vf<M>((Vec3fa)ray.dir);
-        const Vec3vf<M> v0 = tri_v0-O;
-        const Vec3vf<M> v1 = tri_v1-O;
-        const Vec3vf<M> v2 = tri_v2-O;
-
-        /* calculate triangle edges */
-        const Vec3vf<M> e0 = v2-v0;
-        const Vec3vf<M> e1 = v0-v1;
-        const Vec3vf<M> e2 = v1-v2;
-
-        /* perform edge tests */
-        const vfloat<M> U = dot(cross(e0,v2+v0),D);
-        const vfloat<M> V = dot(cross(e1,v0+v1),D);
-        const vfloat<M> W = dot(cross(e2,v1+v2),D);
-        const vfloat<M> UVW = U+V+W;
-        const vfloat<M> eps = float(ulp)*abs(UVW);
-#if defined(EMBREE_BACKFACE_CULLING)
-        vbool<M> valid = max(U,V,W) <= eps;
-#else
-        vbool<M> valid = (min(U,V,W) >= -eps) | (max(U,V,W) <= eps);
-#endif
-        if (unlikely(none(valid))) return false;
-
-        /* calculate geometry normal and denominator */
-        const Vec3vf<M> Ng = stable_triangle_normal(e0,e1,e2);
-        const vfloat<M> den = twice(dot(Ng,D));
-
-        /* perform depth test */
-        const vfloat<M> T = twice(dot(v0,Ng));
-        const vfloat<M> t = rcp(den)*T;
-        valid &= vfloat<M>(ray.tnear()) <= t & t <= vfloat<M>(ray.tfar);
-        valid &= den != vfloat<M>(zero);
-        if (unlikely(none(valid))) return false;
-
-        /* update hit information */
-        new (&hit) SubGridQuadHitPlueckerM<M>(valid,U,V,UVW,t,Ng,flags);
-        return true;
-      }
-
-    template<int M, bool filter>
-      struct SubGridQuadMIntersector1Pluecker;
-
-    template<int M, bool filter>
-      struct SubGridQuadMIntersector1Pluecker
-      {
-        __forceinline SubGridQuadMIntersector1Pluecker() {}
-
-        __forceinline SubGridQuadMIntersector1Pluecker(const Ray& ray, const void* ptr) {}
-
-        __forceinline void intersect(RayHit& ray, IntersectContext* context,
-                                     const Vec3vf<M>& v0, const Vec3vf<M>& v1, const Vec3vf<M>& v2, const Vec3vf<M>& v3,
-                                     const GridMesh::Grid &g, const SubGrid& subgrid) const
-        {
-          SubGridQuadHitPlueckerM<M> hit;
-          Intersect1EpilogMU<M,filter> epilog(ray,context,subgrid.geomID(),subgrid.primID());
-
-          /* intersect first triangle */
-          if (intersectPluecker(ray,v0,v1,v3,vbool<M>(false),hit)) 
-          {
-            interpolateUV<M>(hit,g,subgrid,vint<M>(0,1,1,0),vint<M>(0,0,1,1));
-            epilog(hit.valid,hit);
-          }
-
-          /* intersect second triangle */
-          if (intersectPluecker(ray,v2,v3,v1,vbool<M>(true),hit)) 
-          {
-            interpolateUV<M>(hit,g,subgrid,vint<M>(0,1,1,0),vint<M>(0,0,1,1));
-            epilog(hit.valid,hit);
-          }
-        }
-      
-        __forceinline bool occluded(Ray& ray, IntersectContext* context,
-                                    const Vec3vf<M>& v0, const Vec3vf<M>& v1, const Vec3vf<M>& v2, const Vec3vf<M>& v3,
-                                    const GridMesh::Grid &g, const SubGrid& subgrid) const
-        {
-          SubGridQuadHitPlueckerM<M> hit;
-          Occluded1EpilogMU<M,filter> epilog(ray,context,subgrid.geomID(),subgrid.primID());
-          
-          /* intersect first triangle */
-          if (intersectPluecker(ray,v0,v1,v3,vbool<M>(false),hit)) 
-          {
-            interpolateUV<M>(hit,g,subgrid,vint<M>(0,1,1,0),vint<M>(0,0,1,1));
-            if (epilog(hit.valid,hit))
-              return true;
-          }
-
-          /* intersect second triangle */
-          if (intersectPluecker(ray,v2,v3,v1,vbool<M>(true),hit)) 
-          {
-            interpolateUV<M>(hit,g,subgrid,vint<M>(0,1,1,0),vint<M>(0,0,1,1));
-            if (epilog(hit.valid,hit))
-              return true;
-          }
-
-          return false;
-        }
-      };
-
-#if defined (__AVX__)
-
-    /*! Intersects 4 quads with 1 ray using AVX */
-    template<bool filter>
-      struct SubGridQuadMIntersector1Pluecker<4,filter>
-    {
-      __forceinline SubGridQuadMIntersector1Pluecker() {}
-
-      __forceinline SubGridQuadMIntersector1Pluecker(const Ray& ray, const void* ptr) {}
-      
-      template<typename Epilog>
-        __forceinline bool intersect(Ray& ray, const Vec3vf4& v0, const Vec3vf4& v1, const Vec3vf4& v2, const Vec3vf4& v3, const GridMesh::Grid &g, const SubGrid& subgrid, const Epilog& epilog) const
-      {
-        const Vec3vf8 vtx0(vfloat8(v0.x,v2.x),vfloat8(v0.y,v2.y),vfloat8(v0.z,v2.z));
-#if !defined(EMBREE_BACKFACE_CULLING)
-        const Vec3vf8 vtx1(vfloat8(v1.x),vfloat8(v1.y),vfloat8(v1.z));
-        const Vec3vf8 vtx2(vfloat8(v3.x),vfloat8(v3.y),vfloat8(v3.z));        
-#else
-        const Vec3vf8 vtx1(vfloat8(v1.x,v3.x),vfloat8(v1.y,v3.y),vfloat8(v1.z,v3.z));
-        const Vec3vf8 vtx2(vfloat8(v3.x,v1.x),vfloat8(v3.y,v1.y),vfloat8(v3.z,v1.z));
-#endif
-        SubGridQuadHitPlueckerM<8> hit;
-        const vbool8 flags(0,0,0,0,1,1,1,1);
-        if (unlikely(intersectPluecker(ray,vtx0,vtx1,vtx2,flags,hit)))
-        {
-          /* correct U,V interpolation across the entire grid */
-          interpolateUV<8>(hit,g,subgrid,vint<8>(0,1,1,0,0,1,1,0),vint<8>(0,0,1,1,0,0,1,1));            
-          if (unlikely(epilog(hit.valid,hit)))
-            return true;
-        }
-        return false;
-      }
-      
-      __forceinline bool intersect(RayHit& ray, IntersectContext* context,
-                                   const Vec3vf4& v0, const Vec3vf4& v1, const Vec3vf4& v2, const Vec3vf4& v3, 
-                                   const GridMesh::Grid &g, const SubGrid& subgrid) const
-      {
-          return intersect(ray,v0,v1,v2,v3,g,subgrid,Intersect1EpilogMU<8,filter>(ray,context,subgrid.geomID(),subgrid.primID()));
-      }
-      
-      __forceinline bool occluded(Ray& ray, IntersectContext* context,
-                                  const Vec3vf4& v0, const Vec3vf4& v1, const Vec3vf4& v2, const Vec3vf4& v3, 
-                                  const GridMesh::Grid &g, const SubGrid& subgrid) const
-      {
-          return intersect(ray,v0,v1,v2,v3,g,subgrid,Occluded1EpilogMU<8,filter>(ray,context,subgrid.geomID(),subgrid.primID()));
-      }
-    };
-
-#endif
-
-
-    /* ----------------------------- */
-    /* -- ray packet intersectors -- */
-    /* ----------------------------- */
-
-    template<int K>
-      struct SubGridQuadHitPlueckerK
-      {
-         __forceinline SubGridQuadHitPlueckerK(const vfloat<K>& U,
-                                               const vfloat<K>& V,
-                                               const vfloat<K>& UVW,
-                                               const vfloat<K>& t,
-                                               const Vec3vf<K>& Ng,
-                                               const vbool<K>& flags,
-                                               const GridMesh::Grid &g, 
-                                               const SubGrid& subgrid,
-                                               const unsigned int i)
-         : U(U), V(V), UVW(UVW), t(t), flags(flags), tri_Ng(Ng), g(g), subgrid(subgrid), i(i) {}
-
-        __forceinline std::tuple<vfloat<K>,vfloat<K>,vfloat<K>,Vec3vf<K>> operator() () const
-        {
-          const vbool<K> invalid = abs(UVW) < min_rcp_input;
-          const vfloat<K> rcpUVW = select(invalid,vfloat<K>(0.0f),rcp(UVW));
-          const vfloat<K> u0 = min(U * rcpUVW,1.0f);
-          const vfloat<K> v0 = min(V * rcpUVW,1.0f);
-          const vfloat<K> u1 = vfloat<K>(1.0f) - u0;
-          const vfloat<K> v1 = vfloat<K>(1.0f) - v0;
-          const vfloat<K> uu = select(flags,u1,u0);
-          const vfloat<K> vv = select(flags,v1,v0);
-          const unsigned int sx = subgrid.x() + (unsigned int)(i % 2);
-          const unsigned int sy = subgrid.y() + (unsigned int)(i >>1);
-          const float inv_resX = rcp((float)(int)(g.resX-1));
-          const float inv_resY = rcp((float)(int)(g.resY-1));
-          const vfloat<K> u = (uu + (float)(int)sx) * inv_resX;
-          const vfloat<K> v = (vv + (float)(int)sy) * inv_resY;
-          const Vec3vf<K> Ng(tri_Ng.x,tri_Ng.y,tri_Ng.z);
-          return std::make_tuple(u,v,t,Ng);
-        }
-
-      private:
-        const vfloat<K> U;
-        const vfloat<K> V;
-        const vfloat<K> UVW;
-        const vfloat<K> t;
-        const vfloat<K> absDen;
-        const vbool<K> flags;
-        const Vec3vf<K> tri_Ng;
-
-        const GridMesh::Grid &g;
-        const SubGrid& subgrid;
-        const size_t i;
-      };
-
-
-    template<int M, int K, bool filter>
-      struct SubGridQuadMIntersectorKPlueckerBase
-      {
-        __forceinline SubGridQuadMIntersectorKPlueckerBase(const vbool<K>& valid, const RayK<K>& ray) {}
-            
-        template<typename Epilog>
-        __forceinline vbool<K> intersectK(const vbool<K>& valid0,
-                                          RayK<K>& ray,
-                                          const Vec3vf<K>& tri_v0,
-                                          const Vec3vf<K>& tri_v1,
-                                          const Vec3vf<K>& tri_v2,
-                                          const Vec3vf<K>& tri_Ng,
-                                          const vbool<K>& flags,
-                                          const GridMesh::Grid &g, 
-                                          const SubGrid &subgrid,
-                                          const unsigned int i,
-                                          const Epilog& epilog) const
-        { 
-          /* calculate denominator */
-        /* calculate vertices relative to ray origin */
-          vbool<K> valid = valid0;
-          const Vec3vf<K> O = ray.org;
-          const Vec3vf<K> D = ray.dir;
-          const Vec3vf<K> v0 = tri_v0-O;
-          const Vec3vf<K> v1 = tri_v1-O;
-          const Vec3vf<K> v2 = tri_v2-O;
-          
-          /* calculate triangle edges */
-          const Vec3vf<K> e0 = v2-v0;
-          const Vec3vf<K> e1 = v0-v1;
-          const Vec3vf<K> e2 = v1-v2;
-           
-          /* perform edge tests */
-          const vfloat<K> U = dot(Vec3vf<K>(cross(e0,v2+v0)),D);
-          const vfloat<K> V = dot(Vec3vf<K>(cross(e1,v0+v1)),D);
-          const vfloat<K> W = dot(Vec3vf<K>(cross(e2,v1+v2)),D);
-          const vfloat<K> UVW = U+V+W;
-          const vfloat<K> eps = float(ulp)*abs(UVW);
-#if defined(EMBREE_BACKFACE_CULLING)
-          valid &= max(U,V,W) <= eps;
-#else
-          valid &= (min(U,V,W) >= -eps) | (max(U,V,W) <= eps);
-#endif
-          if (unlikely(none(valid))) return false;
-          
-          /* calculate geometry normal and denominator */
-          const Vec3vf<K> Ng = stable_triangle_normal(e0,e1,e2);
-          const vfloat<K> den = twice(dot(Vec3vf<K>(Ng),D));
-
-          /* perform depth test */
-          const vfloat<K> T = twice(dot(v0,Vec3vf<K>(Ng)));
-          const vfloat<K> t = rcp(den)*T;
-          valid &= ray.tnear() <= t & t <= ray.tfar;
-          valid &= den != vfloat<K>(zero);
-          if (unlikely(none(valid))) return false;
-          
-          /* calculate hit information */
-          SubGridQuadHitPlueckerK<K> hit(U,V,UVW,t,tri_Ng,flags,g,subgrid,i);
-          return epilog(valid,hit);
-        }
-      
-        template<typename Epilog>
-        __forceinline vbool<K> intersectK(const vbool<K>& valid0, 
-                                          RayK<K>& ray,
-                                          const Vec3vf<K>& v0,
-                                          const Vec3vf<K>& v1,
-                                          const Vec3vf<K>& v2,
-                                          const vbool<K>& flags,
-                                          const GridMesh::Grid &g, 
-                                          const SubGrid &subgrid,
-                                          const unsigned int i,
-                                          const Epilog& epilog) const
-        {
-          const Vec3vf<K> e1 = v0-v1;
-          const Vec3vf<K> e2 = v2-v0;
-          const Vec3vf<K> Ng = cross(e2,e1);
-          return intersectK(valid0,ray,v0,v1,v2,Ng,flags,g,subgrid,i,epilog);
-        }
-
-        template<typename Epilog>
-        __forceinline bool intersectK(const vbool<K>& valid0, 
-                                      RayK<K>& ray,
-                                      const Vec3vf<K>& v0,
-                                      const Vec3vf<K>& v1,
-                                      const Vec3vf<K>& v2,
-                                      const Vec3vf<K>& v3,
-                                      const GridMesh::Grid &g, 
-                                      const SubGrid &subgrid,
-                                      const unsigned int i,
-                                      const Epilog& epilog) const
-        {
-          intersectK(valid0,ray,v0,v1,v3,vbool<K>(false),g,subgrid,i,epilog);
-          if (none(valid0)) return true;
-          intersectK(valid0,ray,v2,v3,v1,vbool<K>(true ),g,subgrid,i,epilog);
-          return none(valid0);
-        }
-
-        static  __forceinline bool intersect1(RayK<K>& ray,
-                                              size_t k,
-                                              const Vec3vf<M>& tri_v0,
-                                              const Vec3vf<M>& tri_v1,
-                                              const Vec3vf<M>& tri_v2,
-                                              const Vec3vf<M>& tri_Ng,
-                                              const vbool<M>& flags,
-                                              SubGridQuadHitPlueckerM<M> &hit)
-        {
-          /* calculate vertices relative to ray origin */
-          const Vec3vf<M> O = broadcast<vfloat<M>>(ray.org,k);
-          const Vec3vf<M> D = broadcast<vfloat<M>>(ray.dir,k);
-          const Vec3vf<M> v0 = tri_v0-O;
-          const Vec3vf<M> v1 = tri_v1-O;
-          const Vec3vf<M> v2 = tri_v2-O;
-          
-          /* calculate triangle edges */
-          const Vec3vf<M> e0 = v2-v0;
-          const Vec3vf<M> e1 = v0-v1;
-          const Vec3vf<M> e2 = v1-v2;
-          
-          /* perform edge tests */
-          const vfloat<M> U = dot(cross(e0,v2+v0),D);
-          const vfloat<M> V = dot(cross(e1,v0+v1),D);
-          const vfloat<M> W = dot(cross(e2,v1+v2),D);
-          const vfloat<M> UVW = U+V+W;
-          const vfloat<M> eps = float(ulp)*abs(UVW);
-#if defined(EMBREE_BACKFACE_CULLING)
-          vbool<M> valid = max(U,V,W) <= eps ;
-#else
-          vbool<M> valid = (min(U,V,W) >= -eps) | (max(U,V,W) <= eps);
-#endif
-          if (unlikely(none(valid))) return false;
-          
-          /* calculate geometry normal and denominator */
-          const Vec3vf<M> Ng = stable_triangle_normal(e0,e1,e2);
-          const vfloat<M> den = twice(dot(Ng,D));
-
-          /* perform depth test */
-          const vfloat<M> T = twice(dot(v0,Ng));
-          const vfloat<M> t = rcp(den)*T;
-          valid &= vfloat<M>(ray.tnear()[k]) <= t & t <= vfloat<M>(ray.tfar[k]);
-          if (unlikely(none(valid))) return false;
-          
-          /* avoid division by 0 */
-          valid &= den != vfloat<M>(zero);
-          if (unlikely(none(valid))) return false;
-          
-          /* update hit information */
-          new (&hit) SubGridQuadHitPlueckerM<M>(valid,U,V,UVW,t,tri_Ng,flags);
-          return true;
-        }
-
-        static __forceinline bool intersect1(RayK<K>& ray,
-                                             size_t k,
-                                             const Vec3vf<M>& v0,
-                                             const Vec3vf<M>& v1,
-                                             const Vec3vf<M>& v2,
-                                             const vbool<M>& flags,
-                                             SubGridQuadHitPlueckerM<M> &hit)
-        {
-          const Vec3vf<M> e1 = v0-v1;
-          const Vec3vf<M> e2 = v2-v0;
-          const Vec3vf<M> Ng = cross(e2,e1); // FIXME: optimize!!!
-          return intersect1(ray,k,v0,v1,v2,Ng,flags,hit);
-        }
-
-      };
-
-    template<int M, int K, bool filter>
-      struct SubGridQuadMIntersectorKPluecker : public SubGridQuadMIntersectorKPlueckerBase<M,K,filter>
-    {
-      __forceinline SubGridQuadMIntersectorKPluecker(const vbool<K>& valid, const RayK<K>& ray)
-        : SubGridQuadMIntersectorKPlueckerBase<M,K,filter>(valid,ray) {}
-
-      __forceinline void intersect1(RayHitK<K>& ray, size_t k, IntersectContext* context,
-                                    const Vec3vf<M>& v0, const Vec3vf<M>& v1, const Vec3vf<M>& v2, const Vec3vf<M>& v3, const GridMesh::Grid &g, const SubGrid &subgrid) const
-      {
-        Intersect1KEpilogMU<M,K,filter> epilog(ray,k,context,subgrid.geomID(),subgrid.primID());
-
-        SubGridQuadHitPlueckerM<4> hit;
-        if (SubGridQuadMIntersectorKPlueckerBase<4,K,filter>::intersect1(ray,k,v0,v1,v3,vboolf4(false),hit))
-        {
-          interpolateUV<M>(hit,g,subgrid,vint<M>(0,1,1,0),vint<M>(0,0,1,1));
-          epilog(hit.valid,hit);
-        }
-
-        if (SubGridQuadMIntersectorKPlueckerBase<4,K,filter>::intersect1(ray,k,v2,v3,v1,vboolf4(true),hit))
-        {
-          interpolateUV<M>(hit,g,subgrid,vint<M>(0,1,1,0),vint<M>(0,0,1,1));
-          epilog(hit.valid,hit);
-        }
-
-      }
-      
-      __forceinline bool occluded1(RayK<K>& ray, size_t k, IntersectContext* context,
-                                   const Vec3vf<M>& v0, const Vec3vf<M>& v1, const Vec3vf<M>& v2, const Vec3vf<M>& v3, const GridMesh::Grid &g, const SubGrid &subgrid) const
-      {
-        Occluded1KEpilogMU<M,K,filter> epilog(ray,k,context,subgrid.geomID(),subgrid.primID());
-
-        SubGridQuadHitPlueckerM<4> hit;
-        if (SubGridQuadMIntersectorKPlueckerBase<4,K,filter>::intersect1(ray,k,v0,v1,v3,vboolf4(false),hit))
-        {
-          interpolateUV<M>(hit,g,subgrid,vint<M>(0,1,1,0),vint<M>(0,0,1,1));
-          if (epilog(hit.valid,hit)) return true;
-        }
-
-        if (SubGridQuadMIntersectorKPlueckerBase<4,K,filter>::intersect1(ray,k,v2,v3,v1,vboolf4(true),hit))
-        {
-          interpolateUV<M>(hit,g,subgrid,vint<M>(0,1,1,0),vint<M>(0,0,1,1));
-          if (epilog(hit.valid,hit)) return true;
-        }
-        return false;
-      }
-    };
-
-  }
-}

thirdparty/embree/common/algorithms/parallel_any_of.h

@@ -1,4 +1,4 @@
-// Copyright 2009-2020 Intel Corporation
+// Copyright 2009-2021 Intel Corporation
 // SPDX-License-Identifier: Apache-2.0
 
 #pragma once

thirdparty/embree/common/algorithms/parallel_filter.h

@@ -1,4 +1,4 @@
-// Copyright 2009-2020 Intel Corporation
+// Copyright 2009-2021 Intel Corporation
 // SPDX-License-Identifier: Apache-2.0
 
 #pragma once

thirdparty/embree/common/algorithms/parallel_for.h

@@ -1,4 +1,4 @@
-// Copyright 2009-2020 Intel Corporation
+// Copyright 2009-2021 Intel Corporation
 // SPDX-License-Identifier: Apache-2.0
 
 #pragma once
@@ -8,12 +8,6 @@
 #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 */
@@ -32,24 +26,6 @@ namespace embree
         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
@@ -95,25 +71,6 @@ namespace embree
       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
     tbb::task_group_context context;

thirdparty/embree/common/algorithms/parallel_for_for.h

@@ -1,4 +1,4 @@
-// Copyright 2009-2020 Intel Corporation
+// Copyright 2009-2021 Intel Corporation
 // SPDX-License-Identifier: Apache-2.0
 
 #pragma once

thirdparty/embree/common/algorithms/parallel_for_for_prefix_sum.h

@@ -1,4 +1,4 @@
-// Copyright 2009-2020 Intel Corporation
+// Copyright 2009-2021 Intel Corporation
 // SPDX-License-Identifier: Apache-2.0
 
 #pragma once

thirdparty/embree/common/algorithms/parallel_map.h

@@ -1,4 +1,4 @@
-// Copyright 2009-2020 Intel Corporation
+// Copyright 2009-2021 Intel Corporation
 // SPDX-License-Identifier: Apache-2.0
 
 #pragma once

thirdparty/embree/common/algorithms/parallel_partition.h

@@ -1,4 +1,4 @@
-// Copyright 2009-2020 Intel Corporation
+// Copyright 2009-2021 Intel Corporation
 // SPDX-License-Identifier: Apache-2.0
 
 #pragma once

thirdparty/embree/common/algorithms/parallel_prefix_sum.h

@@ -1,4 +1,4 @@
-// Copyright 2009-2020 Intel Corporation
+// Copyright 2009-2021 Intel Corporation
 // SPDX-License-Identifier: Apache-2.0
 
 #pragma once

thirdparty/embree/common/algorithms/parallel_reduce.h

@@ -1,4 +1,4 @@
-// Copyright 2009-2020 Intel Corporation
+// Copyright 2009-2021 Intel Corporation
 // SPDX-License-Identifier: Apache-2.0
 
 #pragma once
@@ -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) || (defined(TASKING_GCD) && defined(BUILD_IOS))
+#if defined(TASKING_INTERNAL)
 
     /* fast path for small number of iterations */
     Index taskCount = (last-first+minStepSize-1)/minStepSize;

thirdparty/embree/common/algorithms/parallel_set.h

@@ -1,4 +1,4 @@
-// Copyright 2009-2020 Intel Corporation
+// Copyright 2009-2021 Intel Corporation
 // SPDX-License-Identifier: Apache-2.0
 
 #pragma once

thirdparty/embree/common/algorithms/parallel_sort.h

@@ -1,13 +1,10 @@
-// Copyright 2009-2020 Intel Corporation
+// Copyright 2009-2021 Intel Corporation
 // SPDX-License-Identifier: Apache-2.0
 
 #pragma once
 
 #include "../simd/simd.h"
 #include "parallel_for.h"
-#if defined(TASKING_GCD) && defined(BUILD_IOS)
-#include "../sys/alloc.h"
-#endif
 #include <algorithm>
 
 namespace embree
@@ -323,7 +320,7 @@ namespace embree
 #pragma nounroll      
 #endif
       for (size_t i=startID; i<endID; i++) {
-#if defined(__X86_64__) || defined(__aarch64__)
+#if defined(__64BIT__)
         const size_t index = ((size_t)(Key)src[i] >> (size_t)shift) & (size_t)mask;
 #else
         const Key index = ((Key)src[i] >> shift) & mask;
@@ -385,7 +382,7 @@ namespace embree
 #endif
       for (size_t i=startID; i<endID; i++) {
         const Ty elt = src[i];
-#if defined(__X86_64__) || defined(__aarch64__)
+#if defined(__64BIT__)
         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/parsestream.h

@@ -1,4 +1,4 @@
-// Copyright 2009-2020 Intel Corporation
+// Copyright 2009-2021 Intel Corporation
 // SPDX-License-Identifier: Apache-2.0
 
 #pragma once

thirdparty/embree/common/lexers/stream.h

@@ -1,4 +1,4 @@
-// Copyright 2009-2020 Intel Corporation
+// Copyright 2009-2021 Intel Corporation
 // SPDX-License-Identifier: Apache-2.0
 
 #pragma once

thirdparty/embree/common/lexers/streamfilters.h

@@ -1,4 +1,4 @@
-// Copyright 2009-2020 Intel Corporation
+// Copyright 2009-2021 Intel Corporation
 // SPDX-License-Identifier: Apache-2.0
 
 #pragma once

thirdparty/embree/common/lexers/stringstream.cpp

@@ -1,4 +1,4 @@
-// Copyright 2009-2020 Intel Corporation
+// Copyright 2009-2021 Intel Corporation
 // SPDX-License-Identifier: Apache-2.0
 
 #include "stringstream.h"

thirdparty/embree/common/lexers/stringstream.h

@@ -1,4 +1,4 @@
-// Copyright 2009-2020 Intel Corporation
+// Copyright 2009-2021 Intel Corporation
 // SPDX-License-Identifier: Apache-2.0
 
 #pragma once

thirdparty/embree/common/lexers/tokenstream.cpp

@@ -1,4 +1,4 @@
-// Copyright 2009-2020 Intel Corporation
+// Copyright 2009-2021 Intel Corporation
 // SPDX-License-Identifier: Apache-2.0
 
 #include "tokenstream.h"

thirdparty/embree/common/lexers/tokenstream.h

@@ -1,4 +1,4 @@
-// Copyright 2009-2020 Intel Corporation
+// Copyright 2009-2021 Intel Corporation
 // SPDX-License-Identifier: Apache-2.0
 
 #pragma once

thirdparty/embree/common/math/affinespace.h

@@ -1,4 +1,4 @@
-// Copyright 2009-2020 Intel Corporation
+// Copyright 2009-2021 Intel Corporation
 // SPDX-License-Identifier: Apache-2.0
 
 #pragma once

thirdparty/embree/common/math/bbox.h

@@ -1,4 +1,4 @@
-// Copyright 2009-2020 Intel Corporation
+// Copyright 2009-2021 Intel Corporation
 // SPDX-License-Identifier: Apache-2.0
 
 #pragma once
@@ -77,7 +77,7 @@ namespace embree
     return lower > upper;
   }
 
-#if defined(__SSE__) || defined(__ARM_NEON)
+#if defined(__SSE__)
   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__) || defined(__ARM_NEON)
+#if defined __SSE__
 #include "../simd/sse.h"
 #endif
 
-#if defined (__AVX__)
+#if defined __AVX__
 #include "../simd/avx.h"
 #endif
 

thirdparty/embree/common/math/col3.h

@@ -1,4 +1,4 @@
-// Copyright 2009-2020 Intel Corporation
+// Copyright 2009-2021 Intel Corporation
 // SPDX-License-Identifier: Apache-2.0
 
 #pragma once
@@ -42,6 +42,6 @@ namespace embree
   }
 
   /*! default template instantiations */
-  typedef Col3<uint8_t      > Col3uc;
+  typedef Col3<unsigned char> Col3uc;
   typedef Col3<float        > Col3f;
 }

thirdparty/embree/common/math/col4.h

@@ -1,4 +1,4 @@
-// Copyright 2009-2020 Intel Corporation
+// Copyright 2009-2021 Intel Corporation
 // SPDX-License-Identifier: Apache-2.0
 
 #pragma once
@@ -42,6 +42,6 @@ namespace embree
   }
 
   /*! default template instantiations */
-  typedef Col4<uint8_t      > Col4uc;
+  typedef Col4<unsigned char> Col4uc;
   typedef Col4<float        > Col4f;
 }

thirdparty/embree/common/math/color.h

@@ -1,4 +1,4 @@
-// Copyright 2009-2020 Intel Corporation
+// Copyright 2009-2021 Intel Corporation
 // SPDX-License-Identifier: Apache-2.0
 
 #pragma once
@@ -52,17 +52,17 @@ namespace embree
     __forceinline void set(Col3uc& d) const 
     {
       vfloat4 s = clamp(vfloat4(m128))*255.0f;
-      d.r = (uint8_t)(s[0]); 
+      d.r = (unsigned char)(s[0]); 
-      d.g = (uint8_t)(s[1]); 
+      d.g = (unsigned char)(s[1]); 
-      d.b = (uint8_t)(s[2]); 
+      d.b = (unsigned char)(s[2]); 
     }
     __forceinline void set(Col4uc& d) const 
     {
       vfloat4 s = clamp(vfloat4(m128))*255.0f;
-      d.r = (uint8_t)(s[0]); 
+      d.r = (unsigned char)(s[0]); 
-      d.g = (uint8_t)(s[1]); 
+      d.g = (unsigned char)(s[1]); 
-      d.b = (uint8_t)(s[2]); 
+      d.b = (unsigned char)(s[2]); 
-      d.a = (uint8_t)(s[3]); 
+      d.a = (unsigned char)(s[3]); 
     }
 
     ////////////////////////////////////////////////////////////////////////////////
@@ -114,16 +114,16 @@ namespace embree
     __forceinline void set(Col3uc& d) const 
     { 
       vfloat4 s = clamp(vfloat4(m128))*255.0f;
-      d.r = (uint8_t)(s[0]); 
+      d.r = (unsigned char)(s[0]); 
-      d.g = (uint8_t)(s[1]); 
+      d.g = (unsigned char)(s[1]); 
-      d.b = (uint8_t)(s[2]); 
+      d.b = (unsigned char)(s[2]); 
     }
     __forceinline void set(Col4uc& d) const 
     { 
       vfloat4 s = clamp(vfloat4(m128))*255.0f;
-      d.r = (uint8_t)(s[0]); 
+      d.r = (unsigned char)(s[0]); 
-      d.g = (uint8_t)(s[1]); 
+      d.g = (unsigned char)(s[1]); 
-      d.b = (uint8_t)(s[2]); 
+      d.b = (unsigned char)(s[2]); 
       d.a = 255; 
     }
 
@@ -152,37 +152,21 @@ 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

@@ -0,0 +1,27 @@
+// Copyright 2009-2021 Intel Corporation
+// SPDX-License-Identifier: Apache-2.0
+
+#include "constants.h"
+
+namespace embree
+{
+  TrueTy True;
+  FalseTy False;
+  ZeroTy zero;
+  OneTy one;
+  NegInfTy neg_inf;
+  PosInfTy inf;
+  PosInfTy pos_inf;
+  NaNTy nan;
+  UlpTy ulp;
+  PiTy pi;
+  OneOverPiTy one_over_pi;
+  TwoPiTy two_pi;
+  OneOverTwoPiTy one_over_two_pi;
+  FourPiTy four_pi;
+  OneOverFourPiTy one_over_four_pi;
+  StepTy step;
+  ReverseStepTy reverse_step;
+  EmptyTy empty;
+  UndefinedTy undefined;
+}

thirdparty/embree/common/math/constants.h

@@ -1,4 +1,4 @@
-// Copyright 2009-2020 Intel Corporation
+// Copyright 2009-2021 Intel Corporation
 // SPDX-License-Identifier: Apache-2.0
 
 #pragma once
@@ -12,19 +12,6 @@
 #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;
@@ -57,8 +44,8 @@ namespace embree
     __forceinline operator unsigned int      ( ) const { return 0; }
     __forceinline operator          short    ( ) const { return 0; }
     __forceinline operator unsigned short    ( ) const { return 0; }
-    __forceinline operator          int8_t     ( ) const { return 0; }
+    __forceinline operator          char     ( ) const { return 0; }
-    __forceinline operator uint8_t     ( ) const { return 0; }
+    __forceinline operator unsigned char     ( ) const { return 0; }
   }; 
 
   extern MAYBE_UNUSED ZeroTy zero;
@@ -75,8 +62,8 @@ namespace embree
     __forceinline operator unsigned int      ( ) const { return 1; }
     __forceinline operator          short    ( ) const { return 1; }
     __forceinline operator unsigned short    ( ) const { return 1; }
-    __forceinline operator          int8_t     ( ) const { return 1; }
+    __forceinline operator          char     ( ) const { return 1; }
-    __forceinline operator uint8_t     ( ) const { return 1; }
+    __forceinline operator unsigned char     ( ) const { return 1; }
   };
 
   extern MAYBE_UNUSED OneTy one;
@@ -93,8 +80,8 @@ 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          int8_t     ( ) const { return std::numeric_limits<int8_t>::min(); }
+    __forceinline operator          char     ( ) const { return std::numeric_limits<char>::min(); }
-    __forceinline operator uint8_t     ( ) const { return std::numeric_limits<uint8_t>::min(); }
+    __forceinline operator unsigned char     ( ) const { return std::numeric_limits<unsigned char>::min(); }
 
   };
 
@@ -112,8 +99,8 @@ 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          int8_t     ( ) const { return std::numeric_limits<int8_t>::max(); }
+    __forceinline operator          char     ( ) const { return std::numeric_limits<char>::max(); }
-    __forceinline operator uint8_t     ( ) const { return std::numeric_limits<uint8_t>::max(); }
+    __forceinline operator unsigned char     ( ) const { return std::numeric_limits<unsigned char>::max(); }
   };
 
   extern MAYBE_UNUSED PosInfTy inf;
@@ -207,33 +194,4 @@ 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/interval.h

@@ -1,4 +1,4 @@
-// Copyright 2009-2020 Intel Corporation
+// Copyright 2009-2021 Intel Corporation
 // SPDX-License-Identifier: Apache-2.0
 
 #pragma once

thirdparty/embree/common/math/lbbox.h

@@ -1,4 +1,4 @@
-// Copyright 2009-2020 Intel Corporation
+// Copyright 2009-2021 Intel Corporation
 // SPDX-License-Identifier: Apache-2.0
 
 #pragma once

thirdparty/embree/common/math/linearspace2.h

@@ -1,4 +1,4 @@
-// Copyright 2009-2020 Intel Corporation
+// Copyright 2009-2021 Intel Corporation
 // SPDX-License-Identifier: Apache-2.0
 
 #pragma once

thirdparty/embree/common/math/linearspace3.h

@@ -1,4 +1,4 @@
-// Copyright 2009-2020 Intel Corporation
+// Copyright 2009-2021 Intel Corporation
 // SPDX-License-Identifier: Apache-2.0
 
 #pragma once

thirdparty/embree/common/math/math.h

@@ -1,4 +1,4 @@
-// Copyright 2009-2020 Intel Corporation
+// Copyright 2009-2021 Intel Corporation
 // SPDX-License-Identifier: Apache-2.0
 
 #pragma once
@@ -9,18 +9,15 @@
 #include <cmath>
 
 #if defined(__ARM_NEON)
-#include "SSE2NEON.h"
+#include "../simd/arm/emulation.h"
-#if defined(NEON_AVX2_EMULATION)
-#include "AVX2NEON.h"
-#endif
 #else
 #include <emmintrin.h>
 #include <xmmintrin.h>
 #include <immintrin.h>
 #endif
 
-#if defined(__WIN32__) && !defined(__MINGW32__)
+#if defined(__WIN32__)
-#if (__MSV_VER <= 1700)
+#if defined(_MSC_VER) && (_MSC_VER <= 1700)
 namespace std
 {
   __forceinline bool isinf ( const float x ) { return _finite(x) == 0; }
@@ -47,7 +44,7 @@ namespace embree
   __forceinline int   toInt  (const float& a) { return int(a); }
   __forceinline float toFloat(const int&   a) { return float(a); }
 
-#if defined(__WIN32__) && !defined(__MINGW32__)
+#if defined(__WIN32__)
   __forceinline bool finite ( const float x ) { return _finite(x) != 0; }
 #endif
 
@@ -56,16 +53,6 @@ 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__)
@@ -79,74 +66,33 @@ 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);
+    __m128 r = _mm_rsqrt14_ss(_mm_set_ss(0.0f),a);
 #else
-    const __m128 r = _mm_rsqrt_ss(a);
+    __m128 r = _mm_rsqrt_ss(a);
 #endif
-    const __m128 c = _mm_add_ss(_mm_mul_ss(_mm_set_ss(1.5f), r),
+    r = _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)));
-                                _mm_mul_ss(_mm_mul_ss(_mm_mul_ss(a, _mm_set_ss(-0.5f)), r), _mm_mul_ss(r, r)));
+#if defined(__ARM_NEON)
-    return _mm_cvtss_f32(c);
+    r = _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)));
 #endif
+    return _mm_cvtss_f32(r);
   }
 
-#if defined(__WIN32__) && (__MSC_VER <= 1700) && !defined(__MINGW32__)
+#if defined(__WIN32__) && defined(_MSC_VER) && (_MSC_VER <= 1700)
   __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); }
@@ -200,17 +146,7 @@ namespace embree
   __forceinline double floor( const double x ) { return ::floor (x); }
   __forceinline double ceil ( const double x ) { return ::ceil (x); }
 
-#if defined(__aarch64__) 
+#if defined(__SSE4_1__)
-    __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));
@@ -219,17 +155,7 @@ namespace embree
   }
 #endif
 
-#if defined(__aarch64__) 
+#if defined(__SSE4_1__)
-    __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));
@@ -246,7 +172,7 @@ 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__) || defined(__aarch64__)
+#if defined(__64BIT__)
   __forceinline   size_t min(size_t   a, size_t   b) { return a<b ? a:b; }
 #endif
 
@@ -263,7 +189,7 @@ 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__) || defined(__aarch64__)
+#if defined(__64BIT__)
   __forceinline   size_t max(size_t   a, size_t   b) { return a<b ? b:a; }
 #endif
 
@@ -305,16 +231,6 @@ 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; }
@@ -363,14 +279,16 @@ __forceinline float nmsub ( const float a, const float b, const float c) { retur
   /*! exchange */
   template<typename T> __forceinline void xchg ( T& a, T& b ) { const T tmp = a; a = b; b = tmp; }
 
+  /*  load/store */
+  template<typename Ty> struct mem;
  
-  template<typename T> __forceinline T prod_diff(const T& a,const T& b,const T& c,const T& d) {
+  template<> struct mem<float> {
-#if 1//!defined(__aarch64__)
+    static __forceinline float load (bool mask, const void* ptr) { return mask ? *(float*)ptr : 0.0f; }
-      return msub(a,b,c*d);
+    static __forceinline float loadu(bool mask, const void* ptr) { return mask ? *(float*)ptr : 0.0f; }
-#else
+  
-      return nmadd(c,d,a*b);
+    static __forceinline void store (bool mask, void* ptr, const float v) { if (mask) *(float*)ptr = v; }
-#endif
+    static __forceinline void storeu(bool mask, void* ptr, const float v) { if (mask) *(float*)ptr = v; }
-  }
+  };
   
   /*! bit reverse operation */
   template<class T>
@@ -408,7 +326,7 @@ __forceinline float nmsub ( const float a, const float b, const float c) { retur
     return x | (y << 1) | (z << 2);
   }
 
-#if defined(__AVX2__) && !defined(__aarch64__)
+#if defined(__AVX2__)
 
   template<>
     __forceinline unsigned int bitInterleave(const unsigned int &xi, const unsigned int& yi, const unsigned int& zi)

thirdparty/embree/common/math/obbox.h

@@ -1,4 +1,4 @@
-// Copyright 2009-2020 Intel Corporation
+// Copyright 2009-2021 Intel Corporation
 // SPDX-License-Identifier: Apache-2.0
 
 #pragma once

thirdparty/embree/common/math/quaternion.h

@@ -1,4 +1,4 @@
-// Copyright 2009-2020 Intel Corporation
+// Copyright 2009-2021 Intel Corporation
 // SPDX-License-Identifier: Apache-2.0
 
 #pragma once

thirdparty/embree/common/math/range.h

@@ -1,4 +1,4 @@
-// Copyright 2009-2020 Intel Corporation
+// Copyright 2009-2021 Intel Corporation
 // SPDX-License-Identifier: Apache-2.0
 
 #pragma once

thirdparty/embree/common/math/transcendental.h

@@ -1,4 +1,4 @@
-// Copyright 2009-2020 Intel Corporation
+// Copyright 2009-2021 Intel Corporation
 // SPDX-License-Identifier: Apache-2.0
 
 #pragma once
@@ -418,7 +418,7 @@ __forceinline void __rangeReduceLog(const T &input,
 }
 
 template <typename T> struct ExponentType            { };
-template <int N>      struct ExponentType<vfloat<N>> { typedef vint<N> Ty; };
+template <int N>      struct ExponentType<vfloat_impl<N>> { typedef vint<N> Ty; };
 template <>           struct ExponentType<float>     { typedef int     Ty; };
 
 template <typename T>

thirdparty/embree/common/math/vec2.h

@@ -1,4 +1,4 @@
-// Copyright 2009-2020 Intel Corporation
+// Copyright 2009-2021 Intel Corporation
 // SPDX-License-Identifier: Apache-2.0
 
 #pragma once
@@ -205,11 +205,11 @@ namespace embree
 
 #include "vec2fa.h"
 
-#if defined(__SSE__) || defined(__ARM_NEON)
+#if defined __SSE__
 #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__) || defined(__ARM_NEON)
+#if defined(__SSE__)
   template<> __forceinline Vec2<vfloat4>::Vec2(const Vec2fa& a) : x(a.x), y(a.y) {}
 #endif
 

thirdparty/embree/common/math/vec2fa.h

@@ -1,4 +1,4 @@
-// Copyright 2009-2020 Intel Corporation
+// Copyright 2009-2021 Intel Corporation
 // SPDX-License-Identifier: Apache-2.0
 
 #pragma once
@@ -97,12 +97,6 @@ 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
@@ -117,7 +111,6 @@ namespace embree
 #endif
 
     return res;
-#endif  //defined(__aarch64__) 
   }
 
   __forceinline Vec2fa sqrt ( const Vec2fa& a ) { return _mm_sqrt_ps(a.m128); }
@@ -125,21 +118,12 @@ 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) {
@@ -172,7 +156,7 @@ 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(__aarch64__) || defined(__SSE4_1__)
+#if 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);
@@ -181,7 +165,7 @@ namespace embree
     }
 #endif
 
-#if defined(__aarch64__) || defined(__SSE4_1__)
+#if 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);
@@ -292,9 +276,9 @@ namespace embree
   ////////////////////////////////////////////////////////////////////////////////
 
 #if defined(__aarch64__)
+  //__forceinline Vec2fa trunc(const Vec2fa& a) { return vrndq_f32(a); }
   __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    ); }

thirdparty/embree/common/math/vec3.h

@@ -1,4 +1,4 @@
-// Copyright 2009-2020 Intel Corporation
+// Copyright 2009-2021 Intel Corporation
 // SPDX-License-Identifier: Apache-2.0
 
 #pragma once
@@ -206,7 +206,8 @@ 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>(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> 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> 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;
@@ -265,11 +266,11 @@ namespace embree
 /// SSE / AVX / MIC specializations
 ////////////////////////////////////////////////////////////////////////////////
 
-#if defined(__SSE__) || defined(__ARM_NEON)
+#if defined __SSE__
 #include "../simd/sse.h"
 #endif
 
-#if defined(__AVX__)
+#if defined __AVX__
 #include "../simd/avx.h"
 #endif
 
@@ -290,18 +291,14 @@ namespace embree
   template<> __forceinline Vec3<vfloat4>::Vec3(const Vec3fa& a) {
     x = a.x; y = a.y; z = a.z;
   }
-#elif defined(__SSE__) || defined(__ARM_NEON)
+#elif defined(__SSE__)
   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__) || defined(__ARM_NEON)
+#if defined(__SSE__)
-  __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]));
-  }
-
   template<>
   __forceinline Vec3<vfloat4> broadcast<vfloat4,vfloat4>(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]));
@@ -318,15 +315,6 @@ namespace embree
   __forceinline Vec3<vfloat8>::Vec3(const Vec3fa& a) {
     x = a.x; y = a.y; z = a.z;
   }
-  __forceinline Vec3<vfloat4> broadcast4f(const Vec3<vfloat8>& a, const size_t k) {
-    return Vec3<vfloat4>(vfloat4::broadcast(&a.x[k]), vfloat4::broadcast(&a.y[k]), vfloat4::broadcast(&a.z[k]));
-  }
-  __forceinline Vec3<vfloat8> broadcast8f(const Vec3<vfloat4>& a, const size_t k) {
-    return Vec3<vfloat8>(vfloat8::broadcast(&a.x[k]), vfloat8::broadcast(&a.y[k]), vfloat8::broadcast(&a.z[k]));
-  }
-  __forceinline Vec3<vfloat8> broadcast8f(const Vec3<vfloat8>& a, const size_t k) {
-    return Vec3<vfloat8>(vfloat8::broadcast(&a.x[k]), vfloat8::broadcast(&a.y[k]), vfloat8::broadcast(&a.z[k]));
-  }
 
   template<>
   __forceinline Vec3<vfloat8> broadcast<vfloat8,vfloat4>(const Vec3<vfloat4>& a, const size_t k) {

thirdparty/embree/common/math/vec3ba.h

@@ -1,4 +1,4 @@
-// Copyright 2009-2020 Intel Corporation
+// Copyright 2009-2021 Intel Corporation
 // SPDX-License-Identifier: Apache-2.0
 
 #pragma once

thirdparty/embree/common/math/vec3fa.h

@@ -1,4 +1,4 @@
-// Copyright 2009-2020 Intel Corporation
+// Copyright 2009-2021 Intel Corporation
 // SPDX-License-Identifier: Apache-2.0
 
 #pragma once
@@ -55,13 +55,7 @@ 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 ) {
@@ -95,42 +89,19 @@ 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
@@ -145,7 +116,6 @@ namespace embree
 #endif
 
     return res;
-#endif  //defined(__aarch64__)
   }
 
   __forceinline Vec3fa sqrt ( const Vec3fa& a ) { return _mm_sqrt_ps(a.m128); }
@@ -153,20 +123,12 @@ 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) {
@@ -199,7 +161,7 @@ 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(__aarch64__) || defined(__SSE4_1__)
+#if 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);
@@ -208,7 +170,7 @@ namespace embree
     }
 #endif
 
-#if defined(__aarch64__) || defined(__SSE4_1__)
+#if 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);
@@ -230,30 +192,11 @@ 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); }
@@ -275,25 +218,7 @@ namespace embree
   ////////////////////////////////////////////////////////////////////////////////
   /// Reductions
   ////////////////////////////////////////////////////////////////////////////////
-#if defined(__aarch64__) && defined(BUILD_IOS)
-  __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);
@@ -304,7 +229,6 @@ namespace embree
   __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
@@ -317,13 +241,8 @@ 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); }
- #if defined(__aarch64__)
-  __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)));
@@ -361,7 +280,7 @@ 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>(prod_diff(a0,b0,a1,b1)));
+    return Vec3fa(shuffle<1,2,0,3>(msub(a0,b0,a1*b1)));
   }
 
   __forceinline float  sqr_length ( const Vec3fa& a )                { return dot(a,a); }
@@ -416,11 +335,7 @@ namespace embree
   /// Rounding Functions
   ////////////////////////////////////////////////////////////////////////////////
 
-#if defined(__aarch64__)
+#if defined (__SSE4_1__)
-  __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    ); }
@@ -479,9 +394,7 @@ 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) {      
-#if defined (__aarch64__)
+#if defined (__SSE4_1__)
-      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);
@@ -613,7 +526,7 @@ 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__) || defined(__aarch64__)
+#if defined(__SSE4_1__)
     __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);
@@ -622,7 +535,7 @@ namespace embree
     }
 #endif
 
-#if defined(__SSE4_1__) || defined(__aarch64__)
+#if defined(__SSE4_1__)
     __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);
@@ -787,7 +700,11 @@ namespace embree
   /// Rounding Functions
   ////////////////////////////////////////////////////////////////////////////////
 
-#if defined (__SSE4_1__) && !defined(__aarch64__)
+#if defined(__aarch64__)
+  __forceinline Vec3fx trunc(const Vec3fx& a) { return vrndq_f32(a.m128); }
+  __forceinline Vec3fx floor(const Vec3fx& a) { return vrndmq_f32(a.m128); }
+  __forceinline Vec3fx ceil (const Vec3fx& a) { return vrndpq_f32(a.m128); }
+#elif defined (__SSE4_1__)
   __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

@@ -1,4 +1,4 @@
-// Copyright 2009-2020 Intel Corporation
+// Copyright 2009-2021 Intel Corporation
 // SPDX-License-Identifier: Apache-2.0
 
 #pragma once
@@ -65,9 +65,7 @@ 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(__aarch64__)) 
+#if defined(__SSSE3__)
-  __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
 
@@ -83,7 +81,7 @@ 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(__aarch64__) || defined(__SSE4_1__)
+#if 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; }
@@ -101,14 +99,12 @@ 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
@@ -120,7 +116,7 @@ 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(__aarch64__) || defined(__SSE4_1__)
+#if 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
@@ -131,38 +127,18 @@ 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
   ////////////////////////////////////////////////////////////////////////////////
@@ -185,14 +161,14 @@ namespace embree
   ////////////////////////////////////////////////////////////////////////////////
 
   __forceinline Vec3ia select( const Vec3ba& m, const Vec3ia& t, const Vec3ia& f ) {
-#if defined(__aarch64__) || defined(__SSE4_1__)
+#if 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(__aarch64__) || defined(__SSE4_1__)
+#if 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

@@ -1,4 +1,4 @@
-// Copyright 2009-2020 Intel Corporation
+// Copyright 2009-2021 Intel Corporation
 // SPDX-License-Identifier: Apache-2.0
 
 #pragma once
@@ -192,7 +192,7 @@ namespace embree
   ////////////////////////////////////////////////////////////////////////////////
 
   typedef Vec4<bool         > Vec4b;
-  typedef Vec4<uint8_t      > Vec4uc;
+  typedef Vec4<unsigned char> Vec4uc;
   typedef Vec4<int          > Vec4i;
   typedef Vec4<float        > Vec4f;
 }
@@ -205,7 +205,7 @@ namespace embree
 /// SSE / AVX / MIC specializations
 ////////////////////////////////////////////////////////////////////////////////
 
-#if defined(__SSE__) || defined(__ARM_NEON)
+#if defined __SSE__
 #include "../simd/sse.h"
 #endif
 
@@ -225,31 +225,16 @@ 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__) || defined(__ARM_NEON)
+#elif defined(__SSE__)
   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__) || 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]));
-  }
-#endif
-
 #if defined(__AVX__)
   template<> __forceinline Vec4<vfloat8>::Vec4( const Vec3fx& a ) {
     x = a.x; y = a.y; z = a.z; w = a.w;
   }
-  __forceinline Vec4<vfloat4> broadcast4f( const Vec4<vfloat8>& 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]));
-  }
-  __forceinline Vec4<vfloat8> broadcast8f( const Vec4<vfloat4>& a, const size_t k ) {
-    return Vec4<vfloat8>(vfloat8::broadcast(&a.x[k]), vfloat8::broadcast(&a.y[k]), vfloat8::broadcast(&a.z[k]), vfloat8::broadcast(&a.w[k]));
-  }
-  __forceinline Vec4<vfloat8> broadcast8f( const Vec4<vfloat8>& a, const size_t k ) {
-    return Vec4<vfloat8>(vfloat8::broadcast(&a.x[k]), vfloat8::broadcast(&a.y[k]), vfloat8::broadcast(&a.z[k]), vfloat8::broadcast(&a.w[k]));
-  }
 #endif
 
 #if defined(__AVX512F__)

thirdparty/embree/common/simd/arm/emulation.h

@@ -0,0 +1,50 @@
+// Copyright 2009-2020 Intel Corporation
+// SPDX-License-Identifier: Apache-2.0
+
+#pragma once
+
+/* Make precision match SSE, at the cost of some performance */
+#if !defined(__aarch64__)
+#  define SSE2NEON_PRECISE_DIV 1
+#  define SSE2NEON_PRECISE_SQRT 1
+#endif
+
+#include "sse2neon.h"
+
+__forceinline __m128 _mm_fmsub_ps(__m128 a, __m128 b, __m128 c) {
+   __m128 neg_c = vreinterpretq_m128_f32(vnegq_f32(vreinterpretq_f32_m128(c)));
+  return _mm_fmadd_ps(a, b, neg_c);
+}
+
+__forceinline __m128 _mm_fnmadd_ps(__m128 a, __m128 b, __m128 c) {
+#if defined(__aarch64__)
+    return vreinterpretq_m128_f32(vfmsq_f32(vreinterpretq_f32_m128(c),
+                                            vreinterpretq_f32_m128(b),
+                                            vreinterpretq_f32_m128(a)));
+#else
+    return _mm_sub_ps(c, _mm_mul_ps(a, b));
+#endif
+}
+
+__forceinline __m128 _mm_fnmsub_ps(__m128 a, __m128 b, __m128 c) {
+  return vreinterpretq_m128_f32(vnegq_f32(vreinterpretq_f32_m128(_mm_fmadd_ps(a,b,c))));
+}
+
+
+/* Dummy defines for floating point control */
+#define _MM_MASK_MASK 0x1f80
+#define _MM_MASK_DIV_ZERO 0x200
+#define _MM_FLUSH_ZERO_ON 0x8000
+#define _MM_MASK_DENORM 0x100
+#define _MM_SET_EXCEPTION_MASK(x)
+#define _MM_SET_FLUSH_ZERO_MODE(x)
+
+__forceinline int _mm_getcsr()
+{
+  return 0;
+}
+
+__forceinline void _mm_mfence()
+{
+  __sync_synchronize();
+}

thirdparty/embree/common/simd/avx.h

@@ -1,4 +1,4 @@
-// Copyright 2009-2020 Intel Corporation
+// Copyright 2009-2021 Intel Corporation
 // SPDX-License-Identifier: Apache-2.0
 
 #pragma once

thirdparty/embree/common/simd/avx512.h

@@ -1,4 +1,4 @@
-// Copyright 2009-2020 Intel Corporation
+// Copyright 2009-2021 Intel Corporation
 // SPDX-License-Identifier: Apache-2.0
 
 #pragma once

thirdparty/embree/common/simd/simd.h

@@ -1,4 +1,4 @@
-// Copyright 2009-2020 Intel Corporation
+// Copyright 2009-2021 Intel Corporation
 // SPDX-License-Identifier: Apache-2.0
 
 #pragma once
@@ -6,7 +6,7 @@
 #include "../math/math.h"
 
 /* include SSE wrapper classes */
-#if defined(__SSE__) || defined(__ARM_NEON)
+#if defined(__SSE__)
 #  include "sse.h"
 #endif
 

thirdparty/embree/common/simd/sse.cpp

@@ -1,4 +1,4 @@
-// Copyright 2009-2020 Intel Corporation
+// Copyright 2009-2021 Intel Corporation
 // SPDX-License-Identifier: Apache-2.0
 
 #include "sse.h"

thirdparty/embree/common/simd/sse.h

@@ -1,4 +1,4 @@
-// Copyright 2009-2020 Intel Corporation
+// Copyright 2009-2021 Intel Corporation
 // SPDX-License-Identifier: Apache-2.0
 
 #pragma once
@@ -11,7 +11,7 @@
 
 namespace embree 
 {
-#if (defined(__aarch64__) && defined(BUILD_IOS)) || defined(__SSE4_1__)
+#if defined(__SSE4_1__)
   __forceinline __m128 blendv_ps(__m128 f, __m128 t, __m128 mask) { 
     return _mm_blendv_ps(f,t,mask);
   }

thirdparty/embree/common/simd/varying.h

@@ -1,4 +1,4 @@
-// Copyright 2009-2020 Intel Corporation
+// Copyright 2009-2021 Intel Corporation
 // SPDX-License-Identifier: Apache-2.0
 
 #pragma once
@@ -9,7 +9,7 @@ namespace embree
 {
   /* Varying numeric types */
   template<int N>
-  struct vfloat
+  struct vfloat_impl
   {
     union { float f[N]; int i[N]; };
     __forceinline const float& operator [](size_t index) const { assert(index < N); return f[index]; }
@@ -17,7 +17,7 @@ namespace embree
   };
 
   template<int N>
-  struct vdouble
+  struct vdouble_impl
   {
     union { double f[N]; long long i[N]; };
     __forceinline const double& operator [](size_t index) const { assert(index < N); return f[index]; }
@@ -25,7 +25,7 @@ namespace embree
   };
 
   template<int N>
-  struct vint
+  struct vint_impl
   {
     int i[N];
     __forceinline const int& operator [](size_t index) const { assert(index < N); return i[index]; }
@@ -33,7 +33,7 @@ namespace embree
   };
 
   template<int N>
-  struct vuint
+  struct vuint_impl
   {
     unsigned int i[N];
     __forceinline const unsigned int& operator [](size_t index) const { assert(index < N); return i[index]; }
@@ -41,7 +41,7 @@ namespace embree
   };
 
   template<int N>
-  struct vllong
+  struct vllong_impl
   {
     long long i[N];
     __forceinline const long long& operator [](size_t index) const { assert(index < N); return i[index]; }
@@ -49,20 +49,13 @@ namespace embree
   };
 
   /* Varying bool types */
-  template<int N> struct vboolf { int       i[N]; }; // for float/int
+  template<int N> struct vboolf_impl { int       i[N]; }; // for float/int
-  template<int N> struct vboold { long long i[N]; }; // for double/long long
+  template<int N> struct vboold_impl { long long i[N]; }; // for double/long long
-
-  /* Aliases to default types */
-  template<int N> using vreal = vfloat<N>;
-  template<int N> using vbool = vboolf<N>;
  
   /* Varying size constants */
 #if defined(__AVX512VL__) // SKX
   const int VSIZEX = 8;  // default size
   const int VSIZEL = 16; // large size
-#elif defined(__AVX512F__) // KNL
-  const int VSIZEX = 16;
-  const int VSIZEL = 16;
 #elif defined(__AVX__)
   const int VSIZEX = 8;
   const int VSIZEL = 8;
@@ -71,21 +64,41 @@ namespace embree
   const int VSIZEL = 4;
 #endif
 
-  /* Extends varying size N to optimal or up to max(N, N2) */
+  template<int N>
-  template<int N, int N2 = VSIZEX>
+  struct vtypes {
-  struct vextend
+    using vbool = vboolf_impl<N>;
-  {
+    using vboolf = vboolf_impl<N>;
-#if defined(__AVX512F__) && !defined(__AVX512VL__) // KNL
+    using vboold = vboold_impl<N>;
-    /* use 16-wide SIMD calculations on KNL even for 4 and 8 wide SIMD */
+    using vint = vint_impl<N>;
-    static const int size = (N2 == VSIZEX) ? VSIZEX : N;
+    using vuint = vuint_impl<N>;
-    #define SIMD_MODE(N) N, 16
+    using vllong = vllong_impl<N>;
-#else
+    using vfloat = vfloat_impl<N>;
-    /* calculate with same SIMD width otherwise */
+    using vdouble = vdouble_impl<N>;
-    static const int size = N;
-    #define SIMD_MODE(N) N, N
-#endif
   };
 
+  template<>
+  struct vtypes<1> {
+    using vbool = bool;
+    using vboolf = bool;
+    using vboold = bool;
+    using vint = int;
+    using vuint = unsigned int;
+    using vllong = long long;
+    using vfloat = float;
+    using vdouble = double;
+  };
+
+  /* Aliases to default types */
+  template<int N> using vbool = typename vtypes<N>::vbool;
+  template<int N> using vboolf = typename vtypes<N>::vboolf;
+  template<int N> using vboold = typename vtypes<N>::vboold;
+  template<int N> using vint = typename vtypes<N>::vint;
+  template<int N> using vuint = typename vtypes<N>::vuint;
+  template<int N> using vllong = typename vtypes<N>::vllong;
+  template<int N> using vreal = typename vtypes<N>::vfloat;
+  template<int N> using vfloat = typename vtypes<N>::vfloat;
+  template<int N> using vdouble = typename vtypes<N>::vdouble;
+
   /* 4-wide shortcuts */
   typedef vfloat<4>  vfloat4;
   typedef vdouble<4> vdouble4;

thirdparty/embree/common/simd/vboold4_avx.h

@@ -1,8 +1,16 @@
-// Copyright 2009-2020 Intel Corporation
+// Copyright 2009-2021 Intel Corporation
 // SPDX-License-Identifier: Apache-2.0
 
 #pragma once
 
+#define vboolf vboolf_impl
+#define vboold vboold_impl
+#define vint vint_impl
+#define vuint vuint_impl
+#define vllong vllong_impl
+#define vfloat vfloat_impl
+#define vdouble vdouble_impl
+
 namespace embree
 {
   /* 4-wide AVX bool type for 64bit data types*/
@@ -49,18 +57,12 @@ namespace embree
 #endif
     }
     
-    __forceinline vboold(__m128d a, __m128d b) : vl(a), vh(b) {}
-
     ////////////////////////////////////////////////////////////////////////////////
     /// Constants
     ////////////////////////////////////////////////////////////////////////////////
 
     __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
@@ -105,10 +107,9 @@ 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>
@@ -158,3 +159,11 @@ namespace embree
                        << a[4] << ", " << a[5] << ", " << a[6] << ", " << a[7] << ">";
   }
 }
+
+#undef vboolf
+#undef vboold
+#undef vint
+#undef vuint
+#undef vllong
+#undef vfloat
+#undef vdouble

thirdparty/embree/common/simd/vboold4_avx512.h

@@ -1,8 +1,16 @@
-// Copyright 2009-2020 Intel Corporation
+// Copyright 2009-2021 Intel Corporation
 // SPDX-License-Identifier: Apache-2.0
 
 #pragma once
 
+#define vboolf vboolf_impl
+#define vboold vboold_impl
+#define vint vint_impl
+#define vuint vuint_impl
+#define vllong vllong_impl
+#define vfloat vfloat_impl
+#define vdouble vdouble_impl
+
 namespace embree
 {
   /* 4-wide AVX-512 bool type */
@@ -138,3 +146,11 @@ namespace embree
     return cout << ">";
   }
 }
+
+#undef vboolf
+#undef vboold
+#undef vint
+#undef vuint
+#undef vllong
+#undef vfloat
+#undef vdouble

thirdparty/embree/common/simd/vboold8_avx512.h

@@ -1,8 +1,16 @@
-// Copyright 2009-2020 Intel Corporation
+// Copyright 2009-2021 Intel Corporation
 // SPDX-License-Identifier: Apache-2.0
 
 #pragma once
 
+#define vboolf vboolf_impl
+#define vboold vboold_impl
+#define vint vint_impl
+#define vuint vuint_impl
+#define vllong vllong_impl
+#define vfloat vfloat_impl
+#define vdouble vdouble_impl
+
 namespace embree
 {
   /* 8-wide AVX-512 bool type */
@@ -32,25 +40,12 @@ namespace embree
 
     /* return int8 mask */
     __forceinline __m128i mask8() const {
-#if defined(__AVX512BW__)
       return _mm_movm_epi8(v);
-#else
-      const __m512i f = _mm512_set1_epi64(0);
-      const __m512i t = _mm512_set1_epi64(-1);
-      const __m512i m =  _mm512_mask_or_epi64(f,v,t,t); 
-      return _mm512_cvtepi64_epi8(m);
-#endif
     }
 
     /* return int64 mask */
     __forceinline __m512i mask64() const { 
-#if defined(__AVX512DQ__)
       return _mm512_movm_epi64(v);
-#else
-      const __m512i f = _mm512_set1_epi64(0);
-      const __m512i t = _mm512_set1_epi64(-1);
-      return _mm512_mask_or_epi64(f,v,t,t); 
-#endif
     }
 
     ////////////////////////////////////////////////////////////////////////////////
@@ -146,3 +141,11 @@ namespace embree
     return cout << ">";
   }
 }
+
+#undef vboolf
+#undef vboold
+#undef vint
+#undef vuint
+#undef vllong
+#undef vfloat
+#undef vdouble

thirdparty/embree/common/simd/vboolf16_avx512.h

@@ -1,8 +1,16 @@
-// Copyright 2009-2020 Intel Corporation
+// Copyright 2009-2021 Intel Corporation
 // SPDX-License-Identifier: Apache-2.0
 
 #pragma once
 
+#define vboolf vboolf_impl
+#define vboold vboold_impl
+#define vint vint_impl
+#define vuint vuint_impl
+#define vllong vllong_impl
+#define vfloat vfloat_impl
+#define vdouble vdouble_impl
+
 namespace embree
 {
   /* 16-wide AVX-512 bool type */
@@ -33,25 +41,12 @@ namespace embree
 
     /* return int8 mask */
     __forceinline __m128i mask8() const {
-#if defined(__AVX512BW__)
       return _mm_movm_epi8(v);
-#else
-      const __m512i f = _mm512_set1_epi32(0);
-      const __m512i t = _mm512_set1_epi32(-1);
-      const __m512i m =  _mm512_mask_or_epi32(f,v,t,t);
-      return _mm512_cvtepi32_epi8(m);
-#endif
     }
 
     /* return int32 mask */
     __forceinline __m512i mask32() const {
-#if defined(__AVX512DQ__)
       return _mm512_movm_epi32(v);
-#else
-      const __m512i f = _mm512_set1_epi32(0);
-      const __m512i t = _mm512_set1_epi32(-1);
-      return _mm512_mask_or_epi32(f,v,t,t);
-#endif
     }
 
     ////////////////////////////////////////////////////////////////////////////////
@@ -148,3 +143,11 @@ namespace embree
     return cout << ">";
   }
 }
+
+#undef vboolf
+#undef vboold
+#undef vint
+#undef vuint
+#undef vllong
+#undef vfloat
+#undef vdouble

thirdparty/embree/common/simd/vboolf4_avx512.h

@@ -1,8 +1,16 @@
-// Copyright 2009-2020 Intel Corporation
+// Copyright 2009-2021 Intel Corporation
 // SPDX-License-Identifier: Apache-2.0
 
 #pragma once
 
+#define vboolf vboolf_impl
+#define vboold vboold_impl
+#define vint vint_impl
+#define vuint vuint_impl
+#define vllong vllong_impl
+#define vfloat vfloat_impl
+#define vdouble vdouble_impl
+
 namespace embree
 {
   /* 4-wide AVX-512 bool type */
@@ -141,3 +149,11 @@ namespace embree
     return cout << ">";
   }
 }
+
+#undef vboolf
+#undef vboold
+#undef vint
+#undef vuint
+#undef vllong
+#undef vfloat
+#undef vdouble

thirdparty/embree/common/simd/vboolf4_sse2.h

@@ -1,8 +1,16 @@
-// Copyright 2009-2020 Intel Corporation
+// Copyright 2009-2021 Intel Corporation
 // SPDX-License-Identifier: Apache-2.0
 
 #pragma once
 
+#define vboolf vboolf_impl
+#define vboold vboold_impl
+#define vint vint_impl
+#define vuint vuint_impl
+#define vllong vllong_impl
+#define vfloat vfloat_impl
+#define vdouble vdouble_impl
+
 namespace embree
 {
   /* 4-wide SSE bool type */
@@ -37,13 +45,9 @@ 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);
@@ -60,13 +64,8 @@ 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
   };
 
   ////////////////////////////////////////////////////////////////////////////////
@@ -101,7 +100,7 @@ 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(__aarch64__) && defined(BUILD_IOS)) || defined(__SSE4_1__)
+#if 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)); 
@@ -115,17 +114,6 @@ 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)));
@@ -135,7 +123,6 @@ 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) {
@@ -148,7 +135,7 @@ 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__) && !defined(__aarch64__)
+#if defined(__SSE4_1__)
   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)); }
@@ -170,14 +157,10 @@ 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
 
   ////////////////////////////////////////////////////////////////////////////////
@@ -196,3 +179,11 @@ __forceinline size_t popcnt(const vboolf4& a) { return _mm_movemask_popcnt_ps(a)
     return cout << "<" << a[0] << ", " << a[1] << ", " << a[2] << ", " << a[3] << ">";
   }
 }
+
+#undef vboolf
+#undef vboold
+#undef vint
+#undef vuint
+#undef vllong
+#undef vfloat
+#undef vdouble

thirdparty/embree/common/simd/vboolf8_avx.h

@@ -1,8 +1,16 @@
-// Copyright 2009-2020 Intel Corporation
+// Copyright 2009-2021 Intel Corporation
 // SPDX-License-Identifier: Apache-2.0
 
 #pragma once
 
+#define vboolf vboolf_impl
+#define vboold vboold_impl
+#define vint vint_impl
+#define vuint vuint_impl
+#define vllong vllong_impl
+#define vfloat vfloat_impl
+#define vdouble vdouble_impl
+
 namespace embree
 {
   /* 8-wide AVX bool type */
@@ -68,11 +76,8 @@ 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
     ////////////////////////////////////////////////////////////////////////////////
@@ -187,3 +192,11 @@ namespace embree
                        << a[4] << ", " << a[5] << ", " << a[6] << ", " << a[7] << ">";
   }
 }
+
+#undef vboolf
+#undef vboold
+#undef vint
+#undef vuint
+#undef vllong
+#undef vfloat
+#undef vdouble

thirdparty/embree/common/simd/vboolf8_avx512.h

@@ -1,8 +1,16 @@
-// Copyright 2009-2020 Intel Corporation
+// Copyright 2009-2021 Intel Corporation
 // SPDX-License-Identifier: Apache-2.0
 
 #pragma once
 
+#define vboolf vboolf_impl
+#define vboold vboold_impl
+#define vint vint_impl
+#define vuint vuint_impl
+#define vllong vllong_impl
+#define vfloat vfloat_impl
+#define vdouble vdouble_impl
+
 namespace embree
 {
   /* 8-wide AVX-512 bool type */
@@ -141,3 +149,11 @@ namespace embree
     return cout << ">";
   }
 }
+
+#undef vboolf
+#undef vboold
+#undef vint
+#undef vuint
+#undef vllong
+#undef vfloat
+#undef vdouble

thirdparty/embree/common/simd/vdouble4_avx.h

@@ -1,8 +1,16 @@
-// Copyright 2009-2020 Intel Corporation
+// Copyright 2009-2021 Intel Corporation
 // SPDX-License-Identifier: Apache-2.0
 
 #pragma once
 
+#define vboolf vboolf_impl
+#define vboold vboold_impl
+#define vint vint_impl
+#define vuint vuint_impl
+#define vllong vllong_impl
+#define vfloat vfloat_impl
+#define vdouble vdouble_impl
+
 namespace embree
 { 
   /* 4-wide AVX 64-bit double type */
@@ -181,20 +189,13 @@ 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); }
-#elif !defined(__aarch64__)
+#else
   __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); }
@@ -246,18 +247,6 @@ namespace embree
 #endif
   }
 
-  __forceinline void xchg(const vboold4& m, vdouble4& a, vdouble4& b) {
-    const vdouble4 c = a; a = select(m,b,a); b = select(m,c,b);
-  }
-
-  __forceinline vboold4 test(const vdouble4& a, const vdouble4& b) {
-#if defined(__AVX512VL__)
-    return _mm256_test_epi64_mask(_mm256_castpd_si256(a),_mm256_castpd_si256(b));
-#else
-    return _mm256_testz_si256(_mm256_castpd_si256(a),_mm256_castpd_si256(b));
-#endif
-  }
-
   ////////////////////////////////////////////////////////////////////////////////
   // Movement/Shifting/Shuffling Functions
   ////////////////////////////////////////////////////////////////////////////////
@@ -322,3 +311,11 @@ namespace embree
     return cout;
   }
 }
+
+#undef vboolf
+#undef vboold
+#undef vint
+#undef vuint
+#undef vllong
+#undef vfloat
+#undef vdouble

thirdparty/embree/common/simd/vdouble8_avx512.h

@@ -1,8 +1,16 @@
-// Copyright 2009-2020 Intel Corporation
+// Copyright 2009-2021 Intel Corporation
 // SPDX-License-Identifier: Apache-2.0
 
 #pragma once
 
+#define vboolf vboolf_impl
+#define vboold vboold_impl
+#define vint vint_impl
+#define vuint vuint_impl
+#define vllong vllong_impl
+#define vfloat vfloat_impl
+#define vdouble vdouble_impl
+
 namespace embree
 {
   /* 8-wide AVX-512 64-bit double type */
@@ -91,15 +99,6 @@ namespace embree
       _mm512_mask_store_pd(addr, mask, v2);
     }
 
-    /* pass by value to avoid compiler generating inefficient code */
-    static __forceinline void storeu_compact(const vboold8 mask,void * addr, const vdouble8& reg) {
-      _mm512_mask_compressstoreu_pd(addr, mask, reg);
-    }
-
-    static __forceinline vdouble8 compact64bit(const vboold8& mask, vdouble8& v) {
-      return _mm512_mask_compress_pd(v, mask, v);
-    }
-
     static __forceinline vdouble8 compact(const vboold8& mask, vdouble8& v) {
       return _mm512_mask_compress_pd(v, mask, v);
     }
@@ -260,18 +259,6 @@ namespace embree
     return _mm512_mask_or_pd(f,m,t,t);
   }
 
-  __forceinline void xchg(const vboold8& m, vdouble8& a, vdouble8& b) {
-    const vdouble8 c = a; a = select(m,b,a); b = select(m,c,b);
-  }
-
-  __forceinline vboold8 test(const vboold8& m, const vdouble8& a, const vdouble8& b) {
-    return _mm512_mask_test_epi64_mask(m,_mm512_castpd_si512(a),_mm512_castpd_si512(b));
-  }
-
-  __forceinline vboold8 test(const vdouble8& a, const vdouble8& b) {
-    return _mm512_test_epi64_mask(_mm512_castpd_si512(a),_mm512_castpd_si512(b));
-  }
-
   ////////////////////////////////////////////////////////////////////////////////
   // Movement/Shifting/Shuffling Functions
   ////////////////////////////////////////////////////////////////////////////////
@@ -354,3 +341,11 @@ namespace embree
     return cout;
   }
 }
+
+#undef vboolf
+#undef vboold
+#undef vint
+#undef vuint
+#undef vllong
+#undef vfloat
+#undef vdouble

thirdparty/embree/common/simd/vfloat16_avx512.h

@@ -1,8 +1,16 @@
-// Copyright 2009-2020 Intel Corporation
+// Copyright 2009-2021 Intel Corporation
 // SPDX-License-Identifier: Apache-2.0
 
 #pragma once
 
+#define vboolf vboolf_impl
+#define vboold vboold_impl
+#define vint vint_impl
+#define vuint vuint_impl
+#define vllong vllong_impl
+#define vfloat vfloat_impl
+#define vdouble vdouble_impl
+
 namespace embree
 {
   /* 16-wide AVX-512 float type */
@@ -73,11 +81,11 @@ namespace embree
     }
 
     /* WARNING: due to f64x4 the mask is considered as an 8bit mask */
-    __forceinline vfloat(const vboolf16& mask, const vfloat8& a, const vfloat8& b) {
+    /*__forceinline vfloat(const vboolf16& mask, const vfloat8& a, const vfloat8& b) {
       __m512d aa = _mm512_broadcast_f64x4(_mm256_castps_pd(a));
       aa = _mm512_mask_broadcast_f64x4(aa,mask,_mm256_castps_pd(b));
       v = _mm512_castpd_ps(aa);
-    }
+      }*/
     
     __forceinline explicit vfloat(const vint16& a) {
       v = _mm512_cvtepi32_ps(a);
@@ -123,30 +131,6 @@ namespace embree
       return _mm512_set1_ps(*f);
     }
 
-    static __forceinline vfloat16 compact(const vboolf16& mask, vfloat16 &v) {
-      return _mm512_mask_compress_ps(v, mask, v);
-    }
-    static __forceinline vfloat16 compact(const vboolf16& mask, vfloat16 &a, const vfloat16& b) {
-      return _mm512_mask_compress_ps(a, mask, b);
-    }
-
-    static __forceinline vfloat16 expand(const vboolf16& mask, const vfloat16& a, vfloat16& b) {
-      return _mm512_mask_expand_ps(b, mask, a);
-    }
-
-    static __forceinline vfloat16 loadu_compact(const vboolf16& mask, const void* ptr) {
-      return _mm512_mask_expandloadu_ps(_mm512_setzero_ps(), mask, (float*)ptr);
-    }
-
-    static __forceinline void storeu_compact(const vboolf16& mask, float *addr, const vfloat16 reg) {
-      _mm512_mask_compressstoreu_ps(addr, mask, reg);
-    }
-    
-    static __forceinline void storeu_compact_single(const vboolf16& mask, float * addr, const vfloat16& reg) {
-      //_mm512_mask_compressstoreu_ps(addr,mask,reg);
-      *addr = mm512_cvtss_f32(_mm512_mask_compress_ps(reg, mask, reg));
-    }
-
     template<int scale = 4>
     static __forceinline vfloat16 gather(const float* ptr, const vint16& index) {
       return _mm512_i32gather_ps(index, ptr, scale);
@@ -194,12 +178,8 @@ namespace embree
   __forceinline vfloat16 signmsk(const vfloat16& a) { return _mm512_castsi512_ps(_mm512_and_epi32(_mm512_castps_si512(a),_mm512_set1_epi32(0x80000000))); }
 
   __forceinline vfloat16 rcp(const vfloat16& a) {
-#if defined(__AVX512ER__)
-    return _mm512_rcp28_ps(a);
-#else
     const vfloat16 r = _mm512_rcp14_ps(a);
     return _mm512_mul_ps(r, _mm512_fnmadd_ps(r, a, vfloat16(2.0f)));
-#endif
   }
 
   __forceinline vfloat16 sqr (const vfloat16& a) { return _mm512_mul_ps(a,a); }
@@ -207,13 +187,9 @@ namespace embree
 
   __forceinline vfloat16 rsqrt(const vfloat16& a)
   {
-#if defined(__AVX512VL__)
     const vfloat16 r = _mm512_rsqrt14_ps(a);
     return _mm512_fmadd_ps(_mm512_set1_ps(1.5f), r,
                            _mm512_mul_ps(_mm512_mul_ps(_mm512_mul_ps(a, _mm512_set1_ps(-0.5f)), r), _mm512_mul_ps(r, r))); 
-#else
-    return _mm512_rsqrt28_ps(a);
-#endif
   }
 
   ////////////////////////////////////////////////////////////////////////////////
@@ -242,54 +218,26 @@ namespace embree
     return  _mm512_castsi512_ps(_mm512_xor_epi32(_mm512_castps_si512(a),_mm512_castps_si512(b))); 
   }
   
-  __forceinline vfloat16 min(const vfloat16& a, const vfloat16& b) {
+  __forceinline vfloat16 min(const vfloat16& a, const vfloat16& b) { return _mm512_min_ps(a,b);  }
-    return _mm512_min_ps(a,b); 
+  __forceinline vfloat16 min(const vfloat16& a, float           b) { return _mm512_min_ps(a,vfloat16(b)); }
-  }
+  __forceinline vfloat16 min(const float&    a, const vfloat16& b) { return _mm512_min_ps(vfloat16(a),b); }
-  __forceinline vfloat16 min(const vfloat16& a, float b) {
-    return _mm512_min_ps(a,vfloat16(b));
-  }
-  __forceinline vfloat16 min(const float& a, const vfloat16& b) {
-    return _mm512_min_ps(vfloat16(a),b);
-  }
 
-  __forceinline vfloat16 max(const vfloat16& a, const vfloat16& b) {
+  __forceinline vfloat16 max(const vfloat16& a, const vfloat16& b) { return _mm512_max_ps(a,b); }
-    return _mm512_max_ps(a,b); 
+  __forceinline vfloat16 max(const vfloat16& a, float           b) { return _mm512_max_ps(a,vfloat16(b)); }
-  }
+  __forceinline vfloat16 max(const float&    a, const vfloat16& b) { return _mm512_max_ps(vfloat16(a),b); }
-  __forceinline vfloat16 max(const vfloat16& a, float b) {
-    return _mm512_max_ps(a,vfloat16(b));
-  }
-  __forceinline vfloat16 max(const float& a, const vfloat16& b) {
-    return _mm512_max_ps(vfloat16(a),b);
-  }
-
-  __forceinline vfloat16 mask_add(const vboolf16& mask, const vfloat16& c, const vfloat16& a, const vfloat16& b) { return _mm512_mask_add_ps (c,mask,a,b); }
-  __forceinline vfloat16 mask_min(const vboolf16& mask, const vfloat16& c, const vfloat16& a, const vfloat16& b) {
-    return _mm512_mask_min_ps(c,mask,a,b); 
-  }; 
-  __forceinline vfloat16 mask_max(const vboolf16& mask, const vfloat16& c, const vfloat16& a, const vfloat16& b) {
-    return _mm512_mask_max_ps(c,mask,a,b); 
-  }; 
 
   __forceinline vfloat16 mini(const vfloat16& a, const vfloat16& b) {
-#if !defined(__AVX512ER__) // SKX
     const vint16 ai = _mm512_castps_si512(a);
     const vint16 bi = _mm512_castps_si512(b);
     const vint16 ci = _mm512_min_epi32(ai,bi);
     return _mm512_castsi512_ps(ci);
-#else // KNL
-    return min(a,b);
-#endif
   }
 
   __forceinline vfloat16 maxi(const vfloat16& a, const vfloat16& b) {
-#if !defined(__AVX512ER__) // SKX
     const vint16 ai = _mm512_castps_si512(a);
     const vint16 bi = _mm512_castps_si512(b);
     const vint16 ci = _mm512_max_epi32(ai,bi);
     return _mm512_castsi512_ps(ci);
-#else // KNL
-    return max(a,b);
-#endif
   }
 
   ////////////////////////////////////////////////////////////////////////////////
@@ -301,43 +249,6 @@ namespace embree
   __forceinline vfloat16 nmadd(const vfloat16& a, const vfloat16& b, const vfloat16& c) { return _mm512_fnmadd_ps(a,b,c); }
   __forceinline vfloat16 nmsub(const vfloat16& a, const vfloat16& b, const vfloat16& c) { return _mm512_fnmsub_ps(a,b,c); }
   
-  __forceinline vfloat16 mask_msub(const vboolf16& mask,const vfloat16& a, const vfloat16& b, const vfloat16& c) { return _mm512_mask_fmsub_ps(a,mask,b,c); }
-  
-  __forceinline vfloat16 madd231 (const vfloat16& a, const vfloat16& b, const vfloat16& c) { return _mm512_fmadd_ps(c,b,a); }
-  __forceinline vfloat16 msub213 (const vfloat16& a, const vfloat16& b, const vfloat16& c) { return _mm512_fmsub_ps(a,b,c); }
-  __forceinline vfloat16 msub231 (const vfloat16& a, const vfloat16& b, const vfloat16& c) { return _mm512_fmsub_ps(c,b,a); }
-  __forceinline vfloat16 msubr231(const vfloat16& a, const vfloat16& b, const vfloat16& c) { return _mm512_fnmadd_ps(c,b,a); }
-
-
-  ////////////////////////////////////////////////////////////////////////////////
-  /// Operators with rounding
-  ////////////////////////////////////////////////////////////////////////////////
-  
-  __forceinline vfloat16 madd_round_down(const vfloat16& a, const vfloat16& b, const vfloat16& c) { return _mm512_fmadd_round_ps(a,b,c,_MM_FROUND_TO_NEG_INF | _MM_FROUND_NO_EXC); }
-  __forceinline vfloat16 madd_round_up  (const vfloat16& a, const vfloat16& b, const vfloat16& c) { return _mm512_fmadd_round_ps(a,b,c,_MM_FROUND_TO_POS_INF | _MM_FROUND_NO_EXC); }
-
-  __forceinline vfloat16 mul_round_down(const vfloat16& a, const vfloat16& b) { return _mm512_mul_round_ps(a,b,_MM_FROUND_TO_NEG_INF | _MM_FROUND_NO_EXC); }
-  __forceinline vfloat16 mul_round_up  (const vfloat16& a, const vfloat16& b) { return _mm512_mul_round_ps(a,b,_MM_FROUND_TO_POS_INF | _MM_FROUND_NO_EXC); }
-
-  __forceinline vfloat16 add_round_down(const vfloat16& a, const vfloat16& b) { return _mm512_add_round_ps(a,b,_MM_FROUND_TO_NEG_INF | _MM_FROUND_NO_EXC); }
-  __forceinline vfloat16 add_round_up  (const vfloat16& a, const vfloat16& b) { return _mm512_add_round_ps(a,b,_MM_FROUND_TO_POS_INF | _MM_FROUND_NO_EXC); }
-
-  __forceinline vfloat16 sub_round_down(const vfloat16& a, const vfloat16& b) { return _mm512_sub_round_ps(a,b,_MM_FROUND_TO_NEG_INF | _MM_FROUND_NO_EXC); }
-  __forceinline vfloat16 sub_round_up  (const vfloat16& a, const vfloat16& b) { return _mm512_sub_round_ps(a,b,_MM_FROUND_TO_POS_INF | _MM_FROUND_NO_EXC); }
-
-  __forceinline vfloat16 div_round_down(const vfloat16& a, const vfloat16& b) { return _mm512_div_round_ps(a,b,_MM_FROUND_TO_NEG_INF | _MM_FROUND_NO_EXC); }
-  __forceinline vfloat16 div_round_up  (const vfloat16& a, const vfloat16& b) { return _mm512_div_round_ps(a,b,_MM_FROUND_TO_POS_INF | _MM_FROUND_NO_EXC); }
-
-  __forceinline vfloat16 mask_msub_round_down(const vboolf16& mask,const vfloat16& a, const vfloat16& b, const vfloat16& c) { return _mm512_mask_fmsub_round_ps(a,mask,b,c,_MM_FROUND_TO_NEG_INF | _MM_FROUND_NO_EXC); }
-  __forceinline vfloat16 mask_msub_round_up  (const vboolf16& mask,const vfloat16& a, const vfloat16& b, const vfloat16& c) { return _mm512_mask_fmsub_round_ps(a,mask,b,c,_MM_FROUND_TO_POS_INF | _MM_FROUND_NO_EXC); }
-  
-  __forceinline vfloat16 mask_mul_round_down(const vboolf16& mask,const vfloat16& a, const vfloat16& b, const vfloat16& c) { return _mm512_mask_mul_round_ps(a,mask,b,c,_MM_FROUND_TO_NEG_INF | _MM_FROUND_NO_EXC); }
-  __forceinline vfloat16 mask_mul_round_up  (const vboolf16& mask,const vfloat16& a, const vfloat16& b, const vfloat16& c) { return _mm512_mask_mul_round_ps(a,mask,b,c,_MM_FROUND_TO_POS_INF | _MM_FROUND_NO_EXC); }
-
-  __forceinline vfloat16 mask_sub_round_down(const vboolf16& mask,const vfloat16& a, const vfloat16& b, const vfloat16& c) { return _mm512_mask_sub_round_ps(a,mask,b,c,_MM_FROUND_TO_NEG_INF | _MM_FROUND_NO_EXC); }
-  __forceinline vfloat16 mask_sub_round_up  (const vboolf16& mask,const vfloat16& a, const vfloat16& b, const vfloat16& c) { return _mm512_mask_sub_round_ps(a,mask,b,c,_MM_FROUND_TO_POS_INF | _MM_FROUND_NO_EXC); }
-
-  
   ////////////////////////////////////////////////////////////////////////////////
   /// Assignment Operators
   ////////////////////////////////////////////////////////////////////////////////
@@ -404,13 +315,6 @@ namespace embree
     return madd(t,b-a,a);
   }
 
-  __forceinline void xchg(vboolf16 m, vfloat16& a, vfloat16& b)
-  {
-    vfloat16 c = a;
-    a = select(m,b,a);
-    b = select(m,c,b); 
-  }
-
   ////////////////////////////////////////////////////////////////////////////////
   /// Rounding Functions
   ////////////////////////////////////////////////////////////////////////////////
@@ -455,24 +359,6 @@ namespace embree
     return _mm512_shuffle_f32x4(v, v, _MM_SHUFFLE(i3, i2, i1, i0));
   }
 
-  __forceinline vfloat16 interleave_even(const vfloat16& a, const vfloat16& b) {
-    return _mm512_castsi512_ps(_mm512_mask_shuffle_epi32(_mm512_castps_si512(a), mm512_int2mask(0xaaaa), _mm512_castps_si512(b), (_MM_PERM_ENUM)0xb1));
-  }
-
-  __forceinline vfloat16 interleave_odd(const vfloat16& a, const vfloat16& b) {
-    return _mm512_castsi512_ps(_mm512_mask_shuffle_epi32(_mm512_castps_si512(b), mm512_int2mask(0x5555), _mm512_castps_si512(a), (_MM_PERM_ENUM)0xb1));
-  }
-
-  __forceinline vfloat16 interleave2_even(const vfloat16& a, const vfloat16& b) {
-    /* mask should be 8-bit but is 16-bit to reuse for interleave_even */
-    return _mm512_castsi512_ps(_mm512_mask_permutex_epi64(_mm512_castps_si512(a), mm512_int2mask(0xaaaa), _mm512_castps_si512(b), (_MM_PERM_ENUM)0xb1));
-  }
-
-  __forceinline vfloat16 interleave2_odd(const vfloat16& a, const vfloat16& b) {
-    /* mask should be 8-bit but is 16-bit to reuse for interleave_odd */
-    return _mm512_castsi512_ps(_mm512_mask_permutex_epi64(_mm512_castps_si512(b), mm512_int2mask(0x5555), _mm512_castps_si512(a), (_MM_PERM_ENUM)0xb1));
-  }
-
   __forceinline vfloat16 interleave4_even(const vfloat16& a, const vfloat16& b) {
     return _mm512_castsi512_ps(_mm512_mask_permutex_epi64(_mm512_castps_si512(a), mm512_int2mask(0xcc), _mm512_castps_si512(b), (_MM_PERM_ENUM)0x4e));
   }
@@ -537,17 +423,6 @@ namespace embree
   __forceinline void transpose(const vfloat16& r0, const vfloat16& r1, const vfloat16& r2, const vfloat16& r3,
                                vfloat16& c0, vfloat16& c1, vfloat16& c2, vfloat16& c3)
   {
-#if defined(__AVX512F__) && !defined(__AVX512VL__) // KNL
-    vfloat16 a0a1_c0c1 = interleave_even(r0, r1);
-    vfloat16 a2a3_c2c3 = interleave_even(r2, r3);
-    vfloat16 b0b1_d0d1 = interleave_odd (r0, r1);
-    vfloat16 b2b3_d2d3 = interleave_odd (r2, r3);
-
-    c0 = interleave2_even(a0a1_c0c1, a2a3_c2c3);
-    c1 = interleave2_even(b0b1_d0d1, b2b3_d2d3);
-    c2 = interleave2_odd (a0a1_c0c1, a2a3_c2c3);
-    c3 = interleave2_odd (b0b1_d0d1, b2b3_d2d3);
-#else
     vfloat16 a0a2_b0b2 = unpacklo(r0, r2);
     vfloat16 c0c2_d0d2 = unpackhi(r0, r2);
     vfloat16 a1a3_b1b3 = unpacklo(r1, r3);
@@ -557,7 +432,6 @@ namespace embree
     c1 = unpackhi(a0a2_b0b2, a1a3_b1b3);
     c2 = unpacklo(c0c2_d0d2, c1c3_d1d3);
     c3 = unpackhi(c0c2_d0d2, c1c3_d1d3);
-#endif
   }
 
   __forceinline void transpose(const vfloat4& r0,  const vfloat4& r1,  const vfloat4& r2,  const vfloat4& r3,
@@ -715,44 +589,6 @@ namespace embree
     return v;  
   }
 
-  ////////////////////////////////////////////////////////////////////////////////
-  /// Memory load and store operations
-  ////////////////////////////////////////////////////////////////////////////////
-
-  __forceinline vfloat16 loadAOS4to16f(const float& x, const float& y, const float& z)
-  {
-    vfloat16 f = zero;
-    f = select(0x1111,vfloat16::broadcast(&x),f);
-    f = select(0x2222,vfloat16::broadcast(&y),f);
-    f = select(0x4444,vfloat16::broadcast(&z),f);
-    return f;
-  }
-
-  __forceinline vfloat16 loadAOS4to16f(unsigned int index,
-                                       const vfloat16& x,
-                                       const vfloat16& y,
-                                       const vfloat16& z)
-  {
-    vfloat16 f = zero;
-    f = select(0x1111,vfloat16::broadcast((float*)&x + index),f);
-    f = select(0x2222,vfloat16::broadcast((float*)&y + index),f);
-    f = select(0x4444,vfloat16::broadcast((float*)&z + index),f);
-    return f;
-  }
-
-  __forceinline vfloat16 loadAOS4to16f(unsigned int index,
-                                       const vfloat16& x,
-                                       const vfloat16& y,
-                                       const vfloat16& z,
-                                       const vfloat16& fill)
-  {
-    vfloat16 f = fill;
-    f = select(0x1111,vfloat16::broadcast((float*)&x + index),f);
-    f = select(0x2222,vfloat16::broadcast((float*)&y + index),f);
-    f = select(0x4444,vfloat16::broadcast((float*)&z + index),f);
-    return f;
-  }
-
   __forceinline vfloat16 rcp_safe(const vfloat16& a) {
     return rcp(select(a != vfloat16(zero), a, vfloat16(min_rcp_input)));
   }
@@ -769,3 +605,11 @@ namespace embree
     return cout;
   }
 }
+
+#undef vboolf
+#undef vboold
+#undef vint
+#undef vuint
+#undef vllong
+#undef vfloat
+#undef vdouble

thirdparty/embree/common/simd/vfloat4_sse2.h

@@ -1,8 +1,16 @@
-// Copyright 2009-2020 Intel Corporation
+// Copyright 2009-2021 Intel Corporation
 // SPDX-License-Identifier: Apache-2.0
 
 #pragma once
 
+#define vboolf vboolf_impl
+#define vboold vboold_impl
+#define vint vint_impl
+#define vuint vuint_impl
+#define vllong vllong_impl
+#define vfloat vfloat_impl
+#define vdouble vdouble_impl
+
 namespace embree
 {
   /* 4-wide SSE float type */
@@ -34,11 +42,6 @@ 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 
@@ -46,7 +49,7 @@ namespace embree
       const __m128  bf  = _mm_castsi128_ps(b);  
       v  = _mm_add_ps(af,bf);
     }
-#endif
+
     ////////////////////////////////////////////////////////////////////////////////
     /// Constants
     ////////////////////////////////////////////////////////////////////////////////
@@ -71,13 +74,6 @@ namespace embree
 
 #if defined(__AVX512VL__)
 
-    static __forceinline vfloat4 compact(const vboolf4& mask, vfloat4 &v) {
-      return _mm_mask_compress_ps(v, mask, v);
-    }
-    static __forceinline vfloat4 compact(const vboolf4& mask, vfloat4 &a, const vfloat4& b) {
-      return _mm_mask_compress_ps(a, mask, b);
-    }
-
     static __forceinline vfloat4 load (const vboolf4& mask, const void* ptr) { return _mm_mask_load_ps (_mm_setzero_ps(),mask,(float*)ptr); }
     static __forceinline vfloat4 loadu(const vboolf4& mask, const void* ptr) { return _mm_mask_loadu_ps(_mm_setzero_ps(),mask,(float*)ptr); }
 
@@ -111,40 +107,28 @@ namespace embree
 #endif
   }
 
-#if defined(__aarch64__)
+#if defined(__SSE4_1__)
-    static __forceinline vfloat4 load(const int8_t* ptr) {
+    static __forceinline vfloat4 load(const char* 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 int8_t* ptr) {
+    static __forceinline vfloat4 load(const char* ptr) {
       return vfloat4(ptr[0],ptr[1],ptr[2],ptr[3]);
     }
 #endif
 
-#if defined(__aarch64__)
+#if defined(__SSE4_1__)
-    static __forceinline vfloat4 load(const uint8_t* ptr) {
+    static __forceinline vfloat4 load(const unsigned char* 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 uint8_t* ptr) {
+    static __forceinline vfloat4 load(const unsigned char* 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(__aarch64__)
+#if defined(__SSE4_1__)
-    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)));
     }
@@ -161,11 +145,7 @@ namespace embree
     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
@@ -173,14 +153,14 @@ namespace embree
 
     template<int scale = 4>
     static __forceinline vfloat4 gather(const float* ptr, const vint4& index) {
-#if defined(__AVX2__) && !defined(__aarch64__)
+#if defined(__AVX2__)
       return _mm_i32gather_ps(ptr, index, scale);
 #else
       return vfloat4(
-        *(float*)(((int8_t*)ptr)+scale*index[0]),
+        *(float*)(((char*)ptr)+scale*index[0]),
-        *(float*)(((int8_t*)ptr)+scale*index[1]),
+        *(float*)(((char*)ptr)+scale*index[1]),
-        *(float*)(((int8_t*)ptr)+scale*index[2]),
+        *(float*)(((char*)ptr)+scale*index[2]),
-        *(float*)(((int8_t*)ptr)+scale*index[3]));
+        *(float*)(((char*)ptr)+scale*index[3]));
 #endif
     }
 
@@ -189,13 +169,13 @@ namespace embree
       vfloat4 r = zero;
 #if defined(__AVX512VL__)
       return _mm_mmask_i32gather_ps(r, mask, index, ptr, scale);
-#elif defined(__AVX2__)  && !defined(__aarch64__)
+#elif defined(__AVX2__)
       return _mm_mask_i32gather_ps(r, ptr, index, mask, scale);
 #else
-      if (likely(mask[0])) r[0] = *(float*)(((int8_t*)ptr)+scale*index[0]);
+      if (likely(mask[0])) r[0] = *(float*)(((char*)ptr)+scale*index[0]);
-      if (likely(mask[1])) r[1] = *(float*)(((int8_t*)ptr)+scale*index[1]);
+      if (likely(mask[1])) r[1] = *(float*)(((char*)ptr)+scale*index[1]);
-      if (likely(mask[2])) r[2] = *(float*)(((int8_t*)ptr)+scale*index[2]);
+      if (likely(mask[2])) r[2] = *(float*)(((char*)ptr)+scale*index[2]);
-      if (likely(mask[3])) r[3] = *(float*)(((int8_t*)ptr)+scale*index[3]);
+      if (likely(mask[3])) r[3] = *(float*)(((char*)ptr)+scale*index[3]);
       return r;
 #endif
     }
@@ -206,10 +186,10 @@ namespace embree
 #if defined(__AVX512VL__)
       _mm_i32scatter_ps((float*)ptr, index, v, scale);
 #else
-      *(float*)(((int8_t*)ptr)+scale*index[0]) = v[0];
+      *(float*)(((char*)ptr)+scale*index[0]) = v[0];
-      *(float*)(((int8_t*)ptr)+scale*index[1]) = v[1];
+      *(float*)(((char*)ptr)+scale*index[1]) = v[1];
-      *(float*)(((int8_t*)ptr)+scale*index[2]) = v[2];
+      *(float*)(((char*)ptr)+scale*index[2]) = v[2];
-      *(float*)(((int8_t*)ptr)+scale*index[3]) = v[3];
+      *(float*)(((char*)ptr)+scale*index[3]) = v[3];
 #endif
     }
 
@@ -219,14 +199,14 @@ namespace embree
 #if defined(__AVX512VL__)
       _mm_mask_i32scatter_ps((float*)ptr ,mask, index, v, scale);
 #else
-      if (likely(mask[0])) *(float*)(((int8_t*)ptr)+scale*index[0]) = v[0];
+      if (likely(mask[0])) *(float*)(((char*)ptr)+scale*index[0]) = v[0];
-      if (likely(mask[1])) *(float*)(((int8_t*)ptr)+scale*index[1]) = v[1];
+      if (likely(mask[1])) *(float*)(((char*)ptr)+scale*index[1]) = v[1];
-      if (likely(mask[2])) *(float*)(((int8_t*)ptr)+scale*index[2]) = v[2];
+      if (likely(mask[2])) *(float*)(((char*)ptr)+scale*index[2]) = v[2];
-      if (likely(mask[3])) *(float*)(((int8_t*)ptr)+scale*index[3]) = v[3];
+      if (likely(mask[3])) *(float*)(((char*)ptr)+scale*index[3]) = v[3];
 #endif
     }
 
-    static __forceinline void store(const vboolf4& mask, int8_t* ptr, const vint4& ofs, const vfloat4& v) {
+    static __forceinline void store(const vboolf4& mask, char* 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) {
@@ -243,7 +223,7 @@ 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__) || (defined(__aarch64__))
+#elif 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)); 
@@ -251,6 +231,18 @@ namespace embree
     }
   };
 
+  ////////////////////////////////////////////////////////////////////////////////
+  /// Load/Store
+  ////////////////////////////////////////////////////////////////////////////////
+
+  template<> struct mem<vfloat4>
+  {
+    static __forceinline vfloat4 load (const vboolf4& mask, const void* ptr) { return vfloat4::load (mask,ptr); }
+    static __forceinline vfloat4 loadu(const vboolf4& mask, const void* ptr) { return vfloat4::loadu(mask,ptr); }
+    
+    static __forceinline void store (const vboolf4& mask, void* ptr, const vfloat4& v) { vfloat4::store (mask,ptr,v); }
+    static __forceinline void storeu(const vboolf4& mask, void* ptr, const vfloat4& v) { vfloat4::storeu(mask,ptr,v); }
+  };
     
   ////////////////////////////////////////////////////////////////////////////////
   /// Unary Operators
@@ -264,47 +256,18 @@ 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
@@ -316,45 +279,31 @@ 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);
+    vfloat4 r = _mm_rsqrt14_ps(a);
 #else
-    const vfloat4 r = _mm_rsqrt_ps(a);
+    vfloat4 r = _mm_rsqrt_ps(a);
 #endif
 
-#if defined(__AVX2__)
+#if defined(__ARM_NEON)
-    return _mm_fmadd_ps(_mm_set1_ps(1.5f), r,
+    r = _mm_fmadd_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)));
-                        _mm_mul_ps(_mm_mul_ps(_mm_mul_ps(a, _mm_set1_ps(-0.5f)), r), _mm_mul_ps(r, r)));
+    r = _mm_fmadd_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)));
+#elif defined(__AVX2__)
+    r = _mm_fmadd_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)));
 #else
-    return _mm_add_ps(_mm_mul_ps(_mm_set1_ps(1.5f), r),
+    r = _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)));
-                      _mm_mul_ps(_mm_mul_ps(_mm_mul_ps(a, _mm_set1_ps(-0.5f)), r), _mm_mul_ps(r, r)));
-#endif
-
 #endif
+    return r;
   }
 
   __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
@@ -395,8 +344,7 @@ 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__) || defined(__aarch64__)
+#if defined(__SSE4_1__)
-
     __forceinline vfloat4 mini(const vfloat4& a, const vfloat4& b) {
       const vint4 ai = _mm_castps_si128(a);
       const vint4 bi = _mm_castps_si128(b);
@@ -438,30 +386,16 @@ namespace embree
   /// Ternary Operators
   ////////////////////////////////////////////////////////////////////////////////
 
-#if defined(__AVX2__)
+#if defined(__AVX2__) || defined(__ARM_NEON)
   __forceinline vfloat4 madd (const vfloat4& a, const vfloat4& b, const vfloat4& c) { return _mm_fmadd_ps(a,b,c); }
   __forceinline vfloat4 msub (const vfloat4& a, const vfloat4& b, const vfloat4& c) { return _mm_fmsub_ps(a,b,c); }
   __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
 
   ////////////////////////////////////////////////////////////////////////////////
@@ -495,13 +429,8 @@ 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
 
@@ -556,57 +485,6 @@ namespace embree
 #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);
   }
@@ -628,15 +506,15 @@ namespace embree
   ////////////////////////////////////////////////////////////////////////////////
 
 #if defined(__aarch64__)
-  __forceinline vfloat4 floor(const vfloat4& a) { return vrndmq_f32(a.v); } // towards -inf
+  __forceinline vfloat4 floor(const vfloat4& a) { return vrndmq_f32(a.v); }
-  __forceinline vfloat4 ceil (const vfloat4& a) { return vrndpq_f32(a.v); } // toward +inf
+  __forceinline vfloat4 ceil (const vfloat4& a) { return vrndpq_f32(a.v); }
-  __forceinline vfloat4 trunc(const vfloat4& a) { return vrndq_f32(a.v); } // towards 0
+  __forceinline vfloat4 trunc(const vfloat4& a) { return vrndq_f32(a.v); }
-  __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?
+  __forceinline vfloat4 round(const vfloat4& a) { return vrndnq_f32(a.v); }
 #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
+  __forceinline vfloat4 round(const vfloat4& a) { return _mm_round_ps(a, _MM_FROUND_TO_NEAREST_INT); }
 #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])); }
@@ -646,9 +524,7 @@ namespace embree
   __forceinline vfloat4 frac(const vfloat4& a) { return a-floor(a); }
 
   __forceinline vint4 floori(const vfloat4& a) {
-#if defined(__aarch64__)
+#if defined(__SSE4_1__)
-    return vcvtq_s32_f32(floor(a));
-#elif defined(__SSE4_1__)
     return vint4(floor(a));
 #else
     return vint4(a-vfloat4(0.5f));
@@ -662,16 +538,6 @@ 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)));
@@ -681,19 +547,8 @@ 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__)
+#if defined(__SSE3__)
-  __forceinline vfloat4 shuffle8(const vfloat4& a, const vint4& shuf) {
-    return _mm_castsi128_ps(_mm_shuffle_epi8(_mm_castps_si128(a), shuf));
-  }
-#endif
-
-#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))); }
@@ -704,56 +559,10 @@ namespace embree
     return shuffle<i,i,i,i>(v);
   }
 
-#if defined(__aarch64__)
+  template<int i> __forceinline float extract   (const vfloat4& a) { return _mm_cvtss_f32(shuffle<i>(a)); }
-  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)); }
-  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)); }
@@ -762,16 +571,7 @@ 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)); 
@@ -864,25 +664,14 @@ 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) 
   { 
@@ -911,7 +700,7 @@ 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>(prod_diff(a0,b0,a1,b1));
+    return shuffle<1,2,0,3>(msub(a0,b0,a1*b1));
   }
 
   ////////////////////////////////////////////////////////////////////////////////
@@ -923,3 +712,11 @@ namespace embree
   }
 
 }
+
+#undef vboolf
+#undef vboold
+#undef vint
+#undef vuint
+#undef vllong
+#undef vfloat
+#undef vdouble

thirdparty/embree/common/simd/vfloat8_avx.h

@@ -1,8 +1,16 @@
-// Copyright 2009-2020 Intel Corporation
+// Copyright 2009-2021 Intel Corporation
 // SPDX-License-Identifier: Apache-2.0
 
 #pragma once
 
+#define vboolf vboolf_impl
+#define vboold vboold_impl
+#define vint vint_impl
+#define vuint vuint_impl
+#define vllong vllong_impl
+#define vfloat vfloat_impl
+#define vdouble vdouble_impl
+
 namespace embree
 {
   /* 8-wide AVX float type */
@@ -33,7 +41,7 @@ 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 int8_t* a) : v(_mm256_loadu_ps((const float*)a)) {}
+    __forceinline explicit vfloat(const char* 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)) {}
@@ -61,21 +69,7 @@ namespace embree
       return _mm256_broadcast_ss((float*)a); 
     }
 
-    static __forceinline vfloat8 broadcast2(const float* a, const float* b) {
+    static __forceinline vfloat8 load(const char* ptr) {
-#if defined(__INTEL_COMPILER)
-      const vfloat8 v0 = _mm256_broadcast_ss(a); 
-      const vfloat8 v1 = _mm256_broadcast_ss(b); 
-      return _mm256_blend_ps(v1, v0, 0xf);
-#else
-      return _mm256_set_ps(*b,*b,*b,*b,*a,*a,*a,*a);
-#endif
-    }
-
-    static __forceinline vfloat8 broadcast4f(const vfloat4* ptr) {
-      return _mm256_broadcast_ps((__m128*)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 +77,7 @@ namespace embree
 #endif
     }
 
-    static __forceinline vfloat8 load(const uint8_t* ptr) {
+    static __forceinline vfloat8 load(const unsigned char* ptr) {
 #if defined(__AVX2__)
       return _mm256_cvtepi32_ps(_mm256_cvtepu8_epi32(_mm_loadu_si128((__m128i*)ptr)));
 #else
@@ -107,24 +101,11 @@ namespace embree
 
 #if defined(__AVX512VL__)
 
-    static __forceinline vfloat8 compact(const vboolf8& mask, vfloat8 &v) {
-      return _mm256_mask_compress_ps(v, mask, v);
-    }
-    static __forceinline vfloat8 compact(const vboolf8& mask, vfloat8 &a, const vfloat8& b) {
-      return _mm256_mask_compress_ps(a, mask, b);
-    }
-
     static __forceinline vfloat8 load (const vboolf8& mask, const void* ptr) { return _mm256_mask_load_ps (_mm256_setzero_ps(),mask,(float*)ptr); }
     static __forceinline vfloat8 loadu(const vboolf8& mask, const void* ptr) { return _mm256_mask_loadu_ps(_mm256_setzero_ps(),mask,(float*)ptr); }
 
     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); }
@@ -145,18 +126,18 @@ namespace embree
 
     template<int scale = 4>
     static __forceinline vfloat8 gather(const float* ptr, const vint8& index) {
-#if defined(__AVX2__) && !defined(__aarch64__)
+#if defined(__AVX2__)
       return _mm256_i32gather_ps(ptr, index ,scale);
 #else
       return vfloat8(
-          *(float*)(((int8_t*)ptr)+scale*index[0]),
+          *(float*)(((char*)ptr)+scale*index[0]),
-          *(float*)(((int8_t*)ptr)+scale*index[1]),
+          *(float*)(((char*)ptr)+scale*index[1]),
-          *(float*)(((int8_t*)ptr)+scale*index[2]),
+          *(float*)(((char*)ptr)+scale*index[2]),
-          *(float*)(((int8_t*)ptr)+scale*index[3]),
+          *(float*)(((char*)ptr)+scale*index[3]),
-          *(float*)(((int8_t*)ptr)+scale*index[4]),
+          *(float*)(((char*)ptr)+scale*index[4]),
-          *(float*)(((int8_t*)ptr)+scale*index[5]),
+          *(float*)(((char*)ptr)+scale*index[5]),
-          *(float*)(((int8_t*)ptr)+scale*index[6]),
+          *(float*)(((char*)ptr)+scale*index[6]),
-          *(float*)(((int8_t*)ptr)+scale*index[7]));
+          *(float*)(((char*)ptr)+scale*index[7]));
 #endif
     }
 
@@ -165,17 +146,17 @@ namespace embree
       vfloat8 r = zero;
 #if defined(__AVX512VL__)
       return _mm256_mmask_i32gather_ps(r, mask, index, ptr, scale);
-#elif defined(__AVX2__) && !defined(__aarch64__)
+#elif defined(__AVX2__)
       return _mm256_mask_i32gather_ps(r, ptr, index, mask, scale);
 #else
-      if (likely(mask[0])) r[0] = *(float*)(((int8_t*)ptr)+scale*index[0]);
+      if (likely(mask[0])) r[0] = *(float*)(((char*)ptr)+scale*index[0]);
-      if (likely(mask[1])) r[1] = *(float*)(((int8_t*)ptr)+scale*index[1]);
+      if (likely(mask[1])) r[1] = *(float*)(((char*)ptr)+scale*index[1]);
-      if (likely(mask[2])) r[2] = *(float*)(((int8_t*)ptr)+scale*index[2]);
+      if (likely(mask[2])) r[2] = *(float*)(((char*)ptr)+scale*index[2]);
-      if (likely(mask[3])) r[3] = *(float*)(((int8_t*)ptr)+scale*index[3]);
+      if (likely(mask[3])) r[3] = *(float*)(((char*)ptr)+scale*index[3]);
-      if (likely(mask[4])) r[4] = *(float*)(((int8_t*)ptr)+scale*index[4]);
+      if (likely(mask[4])) r[4] = *(float*)(((char*)ptr)+scale*index[4]);
-      if (likely(mask[5])) r[5] = *(float*)(((int8_t*)ptr)+scale*index[5]);
+      if (likely(mask[5])) r[5] = *(float*)(((char*)ptr)+scale*index[5]);
-      if (likely(mask[6])) r[6] = *(float*)(((int8_t*)ptr)+scale*index[6]);
+      if (likely(mask[6])) r[6] = *(float*)(((char*)ptr)+scale*index[6]);
-      if (likely(mask[7])) r[7] = *(float*)(((int8_t*)ptr)+scale*index[7]);
+      if (likely(mask[7])) r[7] = *(float*)(((char*)ptr)+scale*index[7]);
       return r;
     #endif
     }
@@ -186,14 +167,14 @@ namespace embree
 #if defined(__AVX512VL__)
       _mm256_i32scatter_ps((float*)ptr, ofs, v, scale);
 #else
-      *(float*)(((int8_t*)ptr)+scale*ofs[0]) = v[0];
+      *(float*)(((char*)ptr)+scale*ofs[0]) = v[0];
-      *(float*)(((int8_t*)ptr)+scale*ofs[1]) = v[1];
+      *(float*)(((char*)ptr)+scale*ofs[1]) = v[1];
-      *(float*)(((int8_t*)ptr)+scale*ofs[2]) = v[2];
+      *(float*)(((char*)ptr)+scale*ofs[2]) = v[2];
-      *(float*)(((int8_t*)ptr)+scale*ofs[3]) = v[3];
+      *(float*)(((char*)ptr)+scale*ofs[3]) = v[3];
-      *(float*)(((int8_t*)ptr)+scale*ofs[4]) = v[4];
+      *(float*)(((char*)ptr)+scale*ofs[4]) = v[4];
-      *(float*)(((int8_t*)ptr)+scale*ofs[5]) = v[5];
+      *(float*)(((char*)ptr)+scale*ofs[5]) = v[5];
-      *(float*)(((int8_t*)ptr)+scale*ofs[6]) = v[6];
+      *(float*)(((char*)ptr)+scale*ofs[6]) = v[6];
-      *(float*)(((int8_t*)ptr)+scale*ofs[7]) = v[7];
+      *(float*)(((char*)ptr)+scale*ofs[7]) = v[7];
 #endif
     }
 
@@ -203,24 +184,17 @@ namespace embree
 #if defined(__AVX512VL__)
       _mm256_mask_i32scatter_ps((float*)ptr, mask, ofs, v, scale);
 #else
-      if (likely(mask[0])) *(float*)(((int8_t*)ptr)+scale*ofs[0]) = v[0];
+      if (likely(mask[0])) *(float*)(((char*)ptr)+scale*ofs[0]) = v[0];
-      if (likely(mask[1])) *(float*)(((int8_t*)ptr)+scale*ofs[1]) = v[1];
+      if (likely(mask[1])) *(float*)(((char*)ptr)+scale*ofs[1]) = v[1];
-      if (likely(mask[2])) *(float*)(((int8_t*)ptr)+scale*ofs[2]) = v[2];
+      if (likely(mask[2])) *(float*)(((char*)ptr)+scale*ofs[2]) = v[2];
-      if (likely(mask[3])) *(float*)(((int8_t*)ptr)+scale*ofs[3]) = v[3];
+      if (likely(mask[3])) *(float*)(((char*)ptr)+scale*ofs[3]) = v[3];
-      if (likely(mask[4])) *(float*)(((int8_t*)ptr)+scale*ofs[4]) = v[4];
+      if (likely(mask[4])) *(float*)(((char*)ptr)+scale*ofs[4]) = v[4];
-      if (likely(mask[5])) *(float*)(((int8_t*)ptr)+scale*ofs[5]) = v[5];
+      if (likely(mask[5])) *(float*)(((char*)ptr)+scale*ofs[5]) = v[5];
-      if (likely(mask[6])) *(float*)(((int8_t*)ptr)+scale*ofs[6]) = v[6];
+      if (likely(mask[6])) *(float*)(((char*)ptr)+scale*ofs[6]) = v[6];
-      if (likely(mask[7])) *(float*)(((int8_t*)ptr)+scale*ofs[7]) = v[7];
+      if (likely(mask[7])) *(float*)(((char*)ptr)+scale*ofs[7]) = v[7];
 #endif
     }
 
-    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) {
-      scatter<4>(mask,ptr,ofs,v);
-    }
-
     ////////////////////////////////////////////////////////////////////////////////
     /// Array Access
     ////////////////////////////////////////////////////////////////////////////////
@@ -241,60 +215,27 @@ 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);
   }
-#else
-  __forceinline vfloat8 operator -(const vfloat8& a) {
-      __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__) //&& !defined(aarch64)
+#if defined(__AVX2__)
     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)));
@@ -443,29 +384,17 @@ __forceinline vfloat8 abs(const vfloat8& a) {
   static __forceinline vfloat8 select(const vboolf8& m, const vfloat8& t, const vfloat8& f) {
     return _mm256_mask_blend_ps(m, f, t);
   }
-#elif !defined(__aarch64__)
-  __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_NEQ_UQ); }
-  __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_NLT_US); }
-  __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_LE_OS);  }
-
-  __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);  }
+  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_cmpneq_ps(a, b); }
+  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_cmplt_ps(a, b);  }
+  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_cmpge_ps(a, b);  }
+  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_cmpgt_ps(a, b);  }
+  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_cmple_ps(a, b);  }
+  static __forceinline vboolf8 operator <=(const vfloat8& a, const vfloat8& b) { return _mm256_cmp_ps(a, b, _CMP_LE_OS);  }
 
-  __forceinline vfloat8 select(const vboolf8& m, const vfloat8& t, const vfloat8& f) {
+  static __forceinline vfloat8 select(const vboolf8& m, const vfloat8& t, const vfloat8& f) {
     return _mm256_blendv_ps(f, t, m); 
   }
-
 #endif
 
   template<int mask>
@@ -534,17 +463,10 @@ __forceinline vfloat8 abs(const vfloat8& a) {
   /// 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); }
 
   ////////////////////////////////////////////////////////////////////////////////
@@ -579,11 +501,9 @@ __forceinline vfloat8 abs(const vfloat8& a) {
     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); }
@@ -592,10 +512,8 @@ __forceinline vfloat8 abs(const vfloat8& a) {
 
   __forceinline float toScalar(const vfloat8& v) { return _mm_cvtss_f32(_mm256_castps256_ps128(v)); }
 
-  __forceinline vfloat8 assign(const vfloat4& a) { return _mm256_castps128_ps256(a); }
+#if defined (__AVX2__)
-
+  static __forceinline vfloat8 permute(const vfloat8& a, const __m256i& index) {
-#if defined (__AVX2__) && !defined(__aarch64__)
-  __forceinline vfloat8 permute(const vfloat8& a, const __m256i& index) {
     return _mm256_permutevar8x32_ps(a, index);
   }
 #endif
@@ -618,14 +536,6 @@ __forceinline vfloat8 abs(const vfloat8& a) {
   }
 #endif
 
-  __forceinline vfloat4 broadcast4f(const vfloat8& a, const size_t k) {
-    return vfloat4::broadcast(&a[k]);
-  }
-
-  __forceinline vfloat8 broadcast8f(const vfloat8& a, const size_t k) {
-    return vfloat8::broadcast(&a[k]);
-  }
-
 #if defined(__AVX512VL__)
   static __forceinline vfloat8 shift_right_1(const vfloat8& x) {
     return align_shift_right<1>(zero,x);
@@ -699,7 +609,7 @@ __forceinline vfloat8 abs(const vfloat8& a) {
   ////////////////////////////////////////////////////////////////////////////////
   /// 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)); }
@@ -715,14 +625,7 @@ __forceinline vfloat8 abs(const vfloat8& a) {
   __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)); 
@@ -845,3 +748,11 @@ __forceinline vfloat8 abs(const vfloat8& a) {
     return cout << "<" << a[0] << ", " << a[1] << ", " << a[2] << ", " << a[3] << ", " << a[4] << ", " << a[5] << ", " << a[6] << ", " << a[7] << ">";
   }
 }
+
+#undef vboolf
+#undef vboold
+#undef vint
+#undef vuint
+#undef vllong
+#undef vfloat
+#undef vdouble

thirdparty/embree/common/simd/vint16_avx512.h

@@ -1,8 +1,16 @@
-// Copyright 2009-2020 Intel Corporation
+// Copyright 2009-2021 Intel Corporation
 // SPDX-License-Identifier: Apache-2.0
 
 #pragma once
 
+#define vboolf vboolf_impl
+#define vboold vboold_impl
+#define vint vint_impl
+#define vuint vuint_impl
+#define vllong vllong_impl
+#define vfloat vfloat_impl
+#define vdouble vdouble_impl
+
 namespace embree
 { 
   /* 16-wide AVX-512 integer type */
@@ -90,10 +98,10 @@ namespace embree
 
     static __forceinline vint16 load (const void* addr) { return _mm512_load_si512((int*)addr); }
 
-    static __forceinline vint16 load(const uint8_t* ptr) { return _mm512_cvtepu8_epi32(_mm_load_si128((__m128i*)ptr)); }
+    static __forceinline vint16 load(const unsigned char* 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 uint8_t* ptr) { return _mm512_cvtepu8_epi32(_mm_loadu_si128((__m128i*)ptr)); }
+    static __forceinline vint16 loadu(const unsigned char* 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); }
@@ -109,20 +117,6 @@ namespace embree
 
     static __forceinline void store_nt(void* __restrict__ ptr, const vint16& a) { _mm512_stream_si512((__m512i*)ptr,a); }
 
-    /* pass by value to avoid compiler generating inefficient code */
-    static __forceinline void storeu_compact(const vboolf16 mask, void* addr, vint16 reg) {
-      _mm512_mask_compressstoreu_epi32(addr,mask,reg);
-    }
-
-    static __forceinline void storeu_compact_single(const vboolf16 mask, void* addr, vint16 reg) {
-      //_mm512_mask_compressstoreu_epi32(addr,mask,reg);
-      *(float*)addr = mm512_cvtss_f32(_mm512_mask_compress_ps(_mm512_castsi512_ps(reg),mask,_mm512_castsi512_ps(reg)));
-    }
-
-    static __forceinline vint16 compact64bit(const vboolf16& mask, vint16 &v) {
-      return _mm512_mask_compress_epi64(v,mask,v);
-    }
-
     static __forceinline vint16 compact(const vboolf16& mask, vint16 &v) {
       return _mm512_mask_compress_epi32(v,mask,v);
     }
@@ -160,10 +154,6 @@ namespace embree
       _mm512_mask_i32scatter_epi32((int*)ptr,mask,index,v,scale);
     }
 
-    static __forceinline vint16 broadcast64bit(size_t v) {
-      return _mm512_set1_epi64(v);
-    }
-
     ////////////////////////////////////////////////////////////////////////////////
     /// Array Access
     ////////////////////////////////////////////////////////////////////////////////
@@ -313,18 +303,6 @@ namespace embree
     return _mm512_mask_or_epi32(f,m,t,t); 
   }
 
-  __forceinline void xchg(const vboolf16& m, vint16& a, vint16& b) {
-    const vint16 c = a; a = select(m,b,a); b = select(m,c,b);
-  }
-
-  __forceinline vboolf16 test(const vboolf16& m, const vint16& a, const vint16& b) {
-    return _mm512_mask_test_epi32_mask(m,a,b);
-  }
-
-  __forceinline vboolf16 test(const vint16& a, const vint16& b) {
-    return _mm512_test_epi32_mask(a,b);
-  }
-
   ////////////////////////////////////////////////////////////////////////////////
   // Movement/Shifting/Shuffling Functions
   ////////////////////////////////////////////////////////////////////////////////
@@ -363,10 +341,6 @@ namespace embree
 
   template<int i> __forceinline vint16 insert4(const vint16& a, const vint4& b) { return _mm512_inserti32x4(a, b, i); }
 
-  __forceinline size_t extract64bit(const vint16& v) {
-    return _mm_cvtsi128_si64(_mm512_castsi512_si128(v));
-  }
-
   template<int N, int i>
   vint<N> extractN(const vint16& v);
 
@@ -488,3 +462,11 @@ namespace embree
     return cout;
   }
 }
+
+#undef vboolf
+#undef vboold
+#undef vint
+#undef vuint
+#undef vllong
+#undef vfloat
+#undef vdouble

thirdparty/embree/common/simd/vint4_sse2.h

@@ -1,10 +1,18 @@
-// Copyright 2009-2020 Intel Corporation
+// Copyright 2009-2021 Intel Corporation
 // SPDX-License-Identifier: Apache-2.0
 
 #pragma once
 
 #include "../math/math.h"
 
+#define vboolf vboolf_impl
+#define vboold vboold_impl
+#define vint vint_impl
+#define vuint vuint_impl
+#define vllong vllong_impl
+#define vfloat vfloat_impl
+#define vdouble vdouble_impl
+
 namespace embree
 {
   /* 4-wide SSE integer type */
@@ -98,73 +106,53 @@ namespace embree
 #endif
 
 
-#if defined(__aarch64__)
+#if defined(__SSE4_1__)
-    static __forceinline vint4 load(const uint8_t* ptr) {
+    static __forceinline vint4 load(const unsigned char* 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 uint8_t* ptr) {
+    static __forceinline vint4 loadu(const unsigned char* ptr) {
       return  _mm_cvtepu8_epi32(_mm_loadl_epi64((__m128i*)ptr));
     }
 #else
 
-    static __forceinline vint4 load(const uint8_t* ptr) {
+    static __forceinline vint4 load(const unsigned char* ptr) {
       return vint4(ptr[0],ptr[1],ptr[2],ptr[3]);
     } 
 
-    static __forceinline vint4 loadu(const uint8_t* ptr) {
+    static __forceinline vint4 loadu(const unsigned char* ptr) {
       return vint4(ptr[0],ptr[1],ptr[2],ptr[3]);
     }
 
 #endif
 
     static __forceinline vint4 load(const unsigned short* ptr) {
-#if defined(__aarch64__)
+#if defined (__SSE4_1__)
-      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(uint8_t* ptr, const vint4& v) {
+    static __forceinline void store(unsigned char* ptr, const vint4& v) {
-#if defined(__aarch64__)
+#if defined(__SSE4_1__)
-        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] = (uint8_t)v[i];
+        ptr[i] = (unsigned char)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(__aarch64__) || defined(__SSE4_1__)
+#if defined(__SSE4_1__)
       return _mm_stream_load_si128((__m128i*)ptr); 
 #else
       return _mm_load_si128((__m128i*)ptr); 
@@ -172,7 +160,7 @@ namespace embree
     }
     
     static __forceinline void store_nt(void* ptr, const vint4& v) {
-#if !defined(__aarch64__) && defined(__SSE4_1__)
+#if defined(__SSE4_1__)
       _mm_stream_ps((float*)ptr, _mm_castsi128_ps(v));
 #else
       _mm_store_si128((__m128i*)ptr,v);
@@ -181,14 +169,14 @@ namespace embree
 
     template<int scale = 4>
     static __forceinline vint4 gather(const int* ptr, const vint4& index) {
-#if defined(__AVX2__) && !defined(__aarch64__)
+#if defined(__AVX2__)
       return _mm_i32gather_epi32(ptr, index, scale);
 #else
       return vint4(
-          *(int*)(((int8_t*)ptr)+scale*index[0]),
+          *(int*)(((char*)ptr)+scale*index[0]),
-          *(int*)(((int8_t*)ptr)+scale*index[1]),
+          *(int*)(((char*)ptr)+scale*index[1]),
-          *(int*)(((int8_t*)ptr)+scale*index[2]),
+          *(int*)(((char*)ptr)+scale*index[2]),
-          *(int*)(((int8_t*)ptr)+scale*index[3]));
+          *(int*)(((char*)ptr)+scale*index[3]));
 #endif
     }
 
@@ -197,13 +185,13 @@ namespace embree
       vint4 r = zero;
 #if defined(__AVX512VL__)
       return _mm_mmask_i32gather_epi32(r, mask, index, ptr, scale);
-#elif defined(__AVX2__) && !defined(__aarch64__)
+#elif defined(__AVX2__)
       return _mm_mask_i32gather_epi32(r, ptr, index, mask, scale);
 #else
-      if (likely(mask[0])) r[0] = *(int*)(((int8_t*)ptr)+scale*index[0]);
+      if (likely(mask[0])) r[0] = *(int*)(((char*)ptr)+scale*index[0]);
-      if (likely(mask[1])) r[1] = *(int*)(((int8_t*)ptr)+scale*index[1]);
+      if (likely(mask[1])) r[1] = *(int*)(((char*)ptr)+scale*index[1]);
-      if (likely(mask[2])) r[2] = *(int*)(((int8_t*)ptr)+scale*index[2]);
+      if (likely(mask[2])) r[2] = *(int*)(((char*)ptr)+scale*index[2]);
-      if (likely(mask[3])) r[3] = *(int*)(((int8_t*)ptr)+scale*index[3]);
+      if (likely(mask[3])) r[3] = *(int*)(((char*)ptr)+scale*index[3]);
       return r;
 #endif
     }
@@ -214,10 +202,10 @@ namespace embree
 #if defined(__AVX512VL__)
       _mm_i32scatter_epi32((int*)ptr, index, v, scale);
 #else
-      *(int*)(((int8_t*)ptr)+scale*index[0]) = v[0];
+      *(int*)(((char*)ptr)+scale*index[0]) = v[0];
-      *(int*)(((int8_t*)ptr)+scale*index[1]) = v[1];
+      *(int*)(((char*)ptr)+scale*index[1]) = v[1];
-      *(int*)(((int8_t*)ptr)+scale*index[2]) = v[2];
+      *(int*)(((char*)ptr)+scale*index[2]) = v[2];
-      *(int*)(((int8_t*)ptr)+scale*index[3]) = v[3];
+      *(int*)(((char*)ptr)+scale*index[3]) = v[3];
 #endif
     }
 
@@ -227,14 +215,14 @@ namespace embree
 #if defined(__AVX512VL__)
       _mm_mask_i32scatter_epi32((int*)ptr, mask, index, v, scale);
 #else
-      if (likely(mask[0])) *(int*)(((int8_t*)ptr)+scale*index[0]) = v[0];
+      if (likely(mask[0])) *(int*)(((char*)ptr)+scale*index[0]) = v[0];
-      if (likely(mask[1])) *(int*)(((int8_t*)ptr)+scale*index[1]) = v[1];
+      if (likely(mask[1])) *(int*)(((char*)ptr)+scale*index[1]) = v[1];
-      if (likely(mask[2])) *(int*)(((int8_t*)ptr)+scale*index[2]) = v[2];
+      if (likely(mask[2])) *(int*)(((char*)ptr)+scale*index[2]) = v[2];
-      if (likely(mask[3])) *(int*)(((int8_t*)ptr)+scale*index[3]) = v[3];
+      if (likely(mask[3])) *(int*)(((char*)ptr)+scale*index[3]) = v[3];
 #endif
     }
 
-#if defined(__x86_64__) || defined(__aarch64__)
+#if defined(__x86_64__)
     static __forceinline vint4 broadcast64(long long a) { return _mm_set1_epi64x(a); }
 #endif
 
@@ -248,8 +236,6 @@ 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)); 
 #else
@@ -270,9 +256,7 @@ 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(__aarch64__)
+#if defined(__SSSE3__)
-  __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
 
@@ -288,7 +272,7 @@ 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(__aarch64__)) || defined(__SSE4_1__)
+#if 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]); }
@@ -308,8 +292,8 @@ 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, const int n) { return _mm_slli_epi32(a, n); }
+  __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_srai_epi32(a, n); }
+  __forceinline vint4 operator >>(const vint4& a, 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); }
@@ -325,7 +309,7 @@ namespace embree
   __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(__aarch64__)) || defined(__SSE4_1__)
+#if 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
@@ -409,8 +393,7 @@ namespace embree
 #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); }
 
@@ -434,25 +417,16 @@ 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))); }
@@ -464,10 +438,7 @@ namespace embree
     return shuffle<i,i,i,i>(v);
   }
 
-#if defined(__aarch64__)
+#if defined(__SSE4_1__)
-    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
@@ -475,53 +446,7 @@ 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); }
@@ -529,14 +454,10 @@ namespace embree
   __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); 
 #endif
   }
-#endif
 
 #if defined(__AVX512VL__)
 
@@ -554,17 +475,7 @@ namespace embree
   /// Reductions
   ////////////////////////////////////////////////////////////////////////////////
 
-#if defined(__aarch64__) || defined(__SSE4_1__)
+#if 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 ; }
@@ -572,7 +483,6 @@ 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))); }
@@ -592,7 +502,7 @@ namespace embree
   /// Sorting networks
   ////////////////////////////////////////////////////////////////////////////////
 
-#if (defined(__aarch64__)) || defined(__SSE4_1__)
+#if defined(__SSE4_1__)
 
   __forceinline vint4 usort_ascending(const vint4& v)
   {
@@ -679,3 +589,10 @@ namespace embree
   }
 }
 
+#undef vboolf
+#undef vboold
+#undef vint
+#undef vuint
+#undef vllong
+#undef vfloat
+#undef vdouble

thirdparty/embree/common/simd/vint8_avx.h

@@ -1,8 +1,16 @@
-// Copyright 2009-2020 Intel Corporation
+// Copyright 2009-2021 Intel Corporation
 // SPDX-License-Identifier: Apache-2.0
 
 #pragma once
 
+#define vboolf vboolf_impl
+#define vboold vboold_impl
+#define vint vint_impl
+#define vuint vuint_impl
+#define vllong vllong_impl
+#define vfloat vfloat_impl
+#define vdouble vdouble_impl
+
 namespace embree
 {
   /* 8-wide AVX integer type */
@@ -71,25 +79,20 @@ 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 uint8_t* ptr) {
+    static __forceinline vint8 load(const unsigned char* ptr) {
       vint4 il = vint4::load(ptr+0);
       vint4 ih = vint4::load(ptr+4);
       return vint8(il,ih);
     }
 
-    static __forceinline vint8 loadu(const uint8_t* ptr) {
+    static __forceinline vint8 loadu(const unsigned char* ptr) {
       vint4 il = vint4::loadu(ptr+0);
       vint4 ih = vint4::loadu(ptr+4);
       return vint8(il,ih);
@@ -107,7 +110,7 @@ namespace embree
       return vint8(il,ih);
     }
 
-    static __forceinline void store(uint8_t* ptr, const vint8& i) {
+    static __forceinline void store(unsigned char* ptr, const vint8& i) {
       vint4 il(i.vl);
       vint4 ih(i.vh);
       vint4::store(ptr + 0,il);
@@ -122,54 +125,54 @@ namespace embree
     template<int scale = 4>
     static __forceinline vint8 gather(const int* ptr, const vint8& index) {
       return vint8(
-          *(int*)(((int8_t*)ptr)+scale*index[0]),
+          *(int*)(((char*)ptr)+scale*index[0]),
-          *(int*)(((int8_t*)ptr)+scale*index[1]),
+          *(int*)(((char*)ptr)+scale*index[1]),
-          *(int*)(((int8_t*)ptr)+scale*index[2]),
+          *(int*)(((char*)ptr)+scale*index[2]),
-          *(int*)(((int8_t*)ptr)+scale*index[3]),
+          *(int*)(((char*)ptr)+scale*index[3]),
-          *(int*)(((int8_t*)ptr)+scale*index[4]),
+          *(int*)(((char*)ptr)+scale*index[4]),
-          *(int*)(((int8_t*)ptr)+scale*index[5]),
+          *(int*)(((char*)ptr)+scale*index[5]),
-          *(int*)(((int8_t*)ptr)+scale*index[6]),
+          *(int*)(((char*)ptr)+scale*index[6]),
-          *(int*)(((int8_t*)ptr)+scale*index[7]));
+          *(int*)(((char*)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*)(((int8_t*)ptr)+scale*index[0]);
+      if (likely(mask[0])) r[0] = *(int*)(((char*)ptr)+scale*index[0]);
-      if (likely(mask[1])) r[1] = *(int*)(((int8_t*)ptr)+scale*index[1]);
+      if (likely(mask[1])) r[1] = *(int*)(((char*)ptr)+scale*index[1]);
-      if (likely(mask[2])) r[2] = *(int*)(((int8_t*)ptr)+scale*index[2]);
+      if (likely(mask[2])) r[2] = *(int*)(((char*)ptr)+scale*index[2]);
-      if (likely(mask[3])) r[3] = *(int*)(((int8_t*)ptr)+scale*index[3]);
+      if (likely(mask[3])) r[3] = *(int*)(((char*)ptr)+scale*index[3]);
-      if (likely(mask[4])) r[4] = *(int*)(((int8_t*)ptr)+scale*index[4]);
+      if (likely(mask[4])) r[4] = *(int*)(((char*)ptr)+scale*index[4]);
-      if (likely(mask[5])) r[5] = *(int*)(((int8_t*)ptr)+scale*index[5]);
+      if (likely(mask[5])) r[5] = *(int*)(((char*)ptr)+scale*index[5]);
-      if (likely(mask[6])) r[6] = *(int*)(((int8_t*)ptr)+scale*index[6]);
+      if (likely(mask[6])) r[6] = *(int*)(((char*)ptr)+scale*index[6]);
-      if (likely(mask[7])) r[7] = *(int*)(((int8_t*)ptr)+scale*index[7]);
+      if (likely(mask[7])) r[7] = *(int*)(((char*)ptr)+scale*index[7]);
       return r;
     }
 
     template<int scale = 4>
     static __forceinline void scatter(void* ptr, const vint8& ofs, const vint8& v)
     {
-      *(int*)(((int8_t*)ptr)+scale*ofs[0]) = v[0];
+      *(int*)(((char*)ptr)+scale*ofs[0]) = v[0];
-      *(int*)(((int8_t*)ptr)+scale*ofs[1]) = v[1];
+      *(int*)(((char*)ptr)+scale*ofs[1]) = v[1];
-      *(int*)(((int8_t*)ptr)+scale*ofs[2]) = v[2];
+      *(int*)(((char*)ptr)+scale*ofs[2]) = v[2];
-      *(int*)(((int8_t*)ptr)+scale*ofs[3]) = v[3];
+      *(int*)(((char*)ptr)+scale*ofs[3]) = v[3];
-      *(int*)(((int8_t*)ptr)+scale*ofs[4]) = v[4];
+      *(int*)(((char*)ptr)+scale*ofs[4]) = v[4];
-      *(int*)(((int8_t*)ptr)+scale*ofs[5]) = v[5];
+      *(int*)(((char*)ptr)+scale*ofs[5]) = v[5];
-      *(int*)(((int8_t*)ptr)+scale*ofs[6]) = v[6];
+      *(int*)(((char*)ptr)+scale*ofs[6]) = v[6];
-      *(int*)(((int8_t*)ptr)+scale*ofs[7]) = v[7];
+      *(int*)(((char*)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*)(((int8_t*)ptr)+scale*ofs[0]) = v[0];
+      if (likely(mask[0])) *(int*)(((char*)ptr)+scale*ofs[0]) = v[0];
-      if (likely(mask[1])) *(int*)(((int8_t*)ptr)+scale*ofs[1]) = v[1];
+      if (likely(mask[1])) *(int*)(((char*)ptr)+scale*ofs[1]) = v[1];
-      if (likely(mask[2])) *(int*)(((int8_t*)ptr)+scale*ofs[2]) = v[2];
+      if (likely(mask[2])) *(int*)(((char*)ptr)+scale*ofs[2]) = v[2];
-      if (likely(mask[3])) *(int*)(((int8_t*)ptr)+scale*ofs[3]) = v[3];
+      if (likely(mask[3])) *(int*)(((char*)ptr)+scale*ofs[3]) = v[3];
-      if (likely(mask[4])) *(int*)(((int8_t*)ptr)+scale*ofs[4]) = v[4];
+      if (likely(mask[4])) *(int*)(((char*)ptr)+scale*ofs[4]) = v[4];
-      if (likely(mask[5])) *(int*)(((int8_t*)ptr)+scale*ofs[5]) = v[5];
+      if (likely(mask[5])) *(int*)(((char*)ptr)+scale*ofs[5]) = v[5];
-      if (likely(mask[6])) *(int*)(((int8_t*)ptr)+scale*ofs[6]) = v[6];
+      if (likely(mask[6])) *(int*)(((char*)ptr)+scale*ofs[6]) = v[6];
-      if (likely(mask[7])) *(int*)(((int8_t*)ptr)+scale*ofs[7]) = v[7];
+      if (likely(mask[7])) *(int*)(((char*)ptr)+scale*ofs[7]) = v[7];
     }
 
 
@@ -315,11 +318,6 @@ namespace embree
     return _mm256_castps_si256(_mm256_blendv_ps(_mm256_castsi256_ps(f), _mm256_castsi256_ps(t), m)); 
   }
 
-  __forceinline vint8 notand(const vboolf8& m, const vint8& f) {
-    return _mm256_castps_si256(_mm256_andnot_ps(m, _mm256_castsi256_ps(f))); 
-  }
-
-
   ////////////////////////////////////////////////////////////////////////////////
   /// Movement/Shifting/Shuffling Functions
   ////////////////////////////////////////////////////////////////////////////////
@@ -462,3 +460,11 @@ namespace embree
     return cout << "<" << a[0] << ", " << a[1] << ", " << a[2] << ", " << a[3] << ", " << a[4] << ", " << a[5] << ", " << a[6] << ", " << a[7] << ">";
   }
 }
+
+#undef vboolf
+#undef vboold
+#undef vint
+#undef vuint
+#undef vllong
+#undef vfloat
+#undef vdouble

thirdparty/embree/common/simd/vint8_avx2.h

@@ -1,8 +1,16 @@
-// Copyright 2009-2020 Intel Corporation
+// Copyright 2009-2021 Intel Corporation
 // SPDX-License-Identifier: Apache-2.0
 
 #pragma once
 
+#define vboolf vboolf_impl
+#define vboold vboold_impl
+#define vint vint_impl
+#define vuint vuint_impl
+#define vllong vllong_impl
+#define vfloat vfloat_impl
+#define vdouble vdouble_impl
+
 namespace embree
 {
   /* 8-wide AVX integer type */
@@ -67,8 +75,8 @@ namespace embree
     /// Loads and Stores
     ////////////////////////////////////////////////////////////////////////////////
 
-    static __forceinline vint8 load(const uint8_t* ptr)  { return _mm256_cvtepu8_epi32(_mm_loadl_epi64((__m128i*)ptr)); }
+    static __forceinline vint8 load(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 loadu(const unsigned char* 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 +116,7 @@ namespace embree
       _mm256_stream_ps((float*)ptr,_mm256_castsi256_ps(v));
     }
 
-    static __forceinline void store(uint8_t* ptr, const vint8& i)
+    static __forceinline void store(unsigned char* ptr, const vint8& i)
     {
       for (size_t j=0; j<8; j++)
         ptr[j] = i[j];
@@ -140,14 +148,14 @@ namespace embree
 #if defined(__AVX512VL__)
       _mm256_i32scatter_epi32((int*)ptr, ofs, v, scale);
 #else
-      *(int*)(((int8_t*)ptr)+scale*ofs[0]) = v[0];
+      *(int*)(((char*)ptr)+scale*ofs[0]) = v[0];
-      *(int*)(((int8_t*)ptr)+scale*ofs[1]) = v[1];
+      *(int*)(((char*)ptr)+scale*ofs[1]) = v[1];
-      *(int*)(((int8_t*)ptr)+scale*ofs[2]) = v[2];
+      *(int*)(((char*)ptr)+scale*ofs[2]) = v[2];
-      *(int*)(((int8_t*)ptr)+scale*ofs[3]) = v[3];
+      *(int*)(((char*)ptr)+scale*ofs[3]) = v[3];
-      *(int*)(((int8_t*)ptr)+scale*ofs[4]) = v[4];
+      *(int*)(((char*)ptr)+scale*ofs[4]) = v[4];
-      *(int*)(((int8_t*)ptr)+scale*ofs[5]) = v[5];
+      *(int*)(((char*)ptr)+scale*ofs[5]) = v[5];
-      *(int*)(((int8_t*)ptr)+scale*ofs[6]) = v[6];
+      *(int*)(((char*)ptr)+scale*ofs[6]) = v[6];
-      *(int*)(((int8_t*)ptr)+scale*ofs[7]) = v[7];
+      *(int*)(((char*)ptr)+scale*ofs[7]) = v[7];
 #endif
     }
 
@@ -157,14 +165,14 @@ namespace embree
 #if defined(__AVX512VL__)
       _mm256_mask_i32scatter_epi32((int*)ptr, mask, ofs, v, scale);
 #else
-      if (likely(mask[0])) *(int*)(((int8_t*)ptr)+scale*ofs[0]) = v[0];
+      if (likely(mask[0])) *(int*)(((char*)ptr)+scale*ofs[0]) = v[0];
-      if (likely(mask[1])) *(int*)(((int8_t*)ptr)+scale*ofs[1]) = v[1];
+      if (likely(mask[1])) *(int*)(((char*)ptr)+scale*ofs[1]) = v[1];
-      if (likely(mask[2])) *(int*)(((int8_t*)ptr)+scale*ofs[2]) = v[2];
+      if (likely(mask[2])) *(int*)(((char*)ptr)+scale*ofs[2]) = v[2];
-      if (likely(mask[3])) *(int*)(((int8_t*)ptr)+scale*ofs[3]) = v[3];
+      if (likely(mask[3])) *(int*)(((char*)ptr)+scale*ofs[3]) = v[3];
-      if (likely(mask[4])) *(int*)(((int8_t*)ptr)+scale*ofs[4]) = v[4];
+      if (likely(mask[4])) *(int*)(((char*)ptr)+scale*ofs[4]) = v[4];
-      if (likely(mask[5])) *(int*)(((int8_t*)ptr)+scale*ofs[5]) = v[5];
+      if (likely(mask[5])) *(int*)(((char*)ptr)+scale*ofs[5]) = v[5];
-      if (likely(mask[6])) *(int*)(((int8_t*)ptr)+scale*ofs[6]) = v[6];
+      if (likely(mask[6])) *(int*)(((char*)ptr)+scale*ofs[6]) = v[6];
-      if (likely(mask[7])) *(int*)(((int8_t*)ptr)+scale*ofs[7]) = v[7];
+      if (likely(mask[7])) *(int*)(((char*)ptr)+scale*ofs[7]) = v[7];
 #endif
     }
 
@@ -385,8 +393,6 @@ namespace embree
 
   __forceinline int toScalar(const vint8& v) { return _mm_cvtsi128_si32(_mm256_castsi256_si128(v)); }
 
-#if !defined(__aarch64__)
-
   __forceinline vint8 permute(const vint8& v, const __m256i& index) {
     return _mm256_permutevar8x32_epi32(v, index);
   }
@@ -395,8 +401,6 @@ __forceinline vint8 permute(const vint8& v, const __m256i& index) {
     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__)
@@ -406,9 +410,6 @@ __forceinline vint8 permute(const vint8& v, const __m256i& index) {
 #endif
   }  
 
-#endif
-
-
   ////////////////////////////////////////////////////////////////////////////////
   /// Reductions
   ////////////////////////////////////////////////////////////////////////////////
@@ -435,9 +436,6 @@ __forceinline vint8 permute(const vint8& v, const __m256i& index) {
   __forceinline size_t select_min(const vboolf8& valid, const vint8& v) { const vint8 a = select(valid,v,vint8(pos_inf)); return bsf(movemask(valid & (a == vreduce_min(a)))); }
   __forceinline size_t select_max(const vboolf8& valid, const vint8& v) { const vint8 a = select(valid,v,vint8(neg_inf)); return bsf(movemask(valid & (a == vreduce_max(a)))); }
 
-
-  __forceinline vint8 assign(const vint4& a) { return _mm256_castsi128_si256(a); }
-
   ////////////////////////////////////////////////////////////////////////////////
   /// Sorting networks
   ////////////////////////////////////////////////////////////////////////////////
@@ -510,3 +508,11 @@ __forceinline vint8 permute(const vint8& v, const __m256i& index) {
     return cout << "<" << a[0] << ", " << a[1] << ", " << a[2] << ", " << a[3] << ", " << a[4] << ", " << a[5] << ", " << a[6] << ", " << a[7] << ">";
   }
 }
+
+#undef vboolf
+#undef vboold
+#undef vint
+#undef vuint
+#undef vllong
+#undef vfloat
+#undef vdouble

thirdparty/embree/common/simd/vllong4_avx2.h

@@ -1,8 +1,16 @@
-// Copyright 2009-2020 Intel Corporation
+// Copyright 2009-2021 Intel Corporation
 // SPDX-License-Identifier: Apache-2.0
 
 #pragma once
 
+#define vboolf vboolf_impl
+#define vboold vboold_impl
+#define vint vint_impl
+#define vuint vuint_impl
+#define vllong vllong_impl
+#define vfloat vfloat_impl
+#define vdouble vdouble_impl
+
 namespace embree
 { 
   /* 4-wide AVX2 64-bit long long type */
@@ -95,16 +103,6 @@ namespace embree
 #endif
     }
 
-    static __forceinline vllong4 broadcast64bit(size_t v) {
-      return _mm256_set1_epi64x(v);
-    }
-
-    static __forceinline size_t extract64bit(const vllong4& v)
-    {
-      return _mm_cvtsi128_si64(_mm256_castsi256_si128(v));
-    }
-
-
     ////////////////////////////////////////////////////////////////////////////////
     /// Array Access
     ////////////////////////////////////////////////////////////////////////////////
@@ -276,18 +274,6 @@ namespace embree
   __forceinline vboold4 le(const vboold4& mask, const vllong4& a, const vllong4& b) { return mask & (a <= b); }
 #endif
 
-  __forceinline void xchg(const vboold4& m, vllong4& a, vllong4& b) {
-    const vllong4 c = a; a = select(m,b,a); b = select(m,c,b);
-  }
-
-  __forceinline vboold4 test(const vllong4& a, const vllong4& b) {
-#if defined(__AVX512VL__)
-    return _mm256_test_epi64_mask(a,b);
-#else
-    return _mm256_testz_si256(a,b);
-#endif
-  }
-
   ////////////////////////////////////////////////////////////////////////////////
   // Movement/Shifting/Shuffling Functions
   ////////////////////////////////////////////////////////////////////////////////
@@ -356,3 +342,11 @@ namespace embree
     return cout;
   }
 }
+
+#undef vboolf
+#undef vboold
+#undef vint
+#undef vuint
+#undef vllong
+#undef vfloat
+#undef vdouble

thirdparty/embree/common/simd/vllong8_avx512.h

@@ -1,8 +1,16 @@
-// Copyright 2009-2020 Intel Corporation
+// Copyright 2009-2021 Intel Corporation
 // SPDX-License-Identifier: Apache-2.0
 
 #pragma once
 
+#define vboolf vboolf_impl
+#define vboold vboold_impl
+#define vint vint_impl
+#define vuint vuint_impl
+#define vllong vllong_impl
+#define vfloat vfloat_impl
+#define vdouble vdouble_impl
+
 namespace embree
 { 
   /* 8-wide AVX-512 64-bit long long type */
@@ -78,7 +86,7 @@ namespace embree
       return _mm512_load_si512(addr);
     }
 
-    static __forceinline vllong8 load(const uint8_t* ptr) {
+    static __forceinline vllong8 load(const unsigned char* ptr) {
       return _mm512_cvtepu8_epi64(*(__m128i*)ptr); 
     }
 
@@ -98,19 +106,6 @@ namespace embree
       _mm512_mask_store_epi64(addr,mask,v2);
     }
 
-    /* pass by value to avoid compiler generating inefficient code */
-    static __forceinline void storeu_compact(const vboold8 mask, void* addr, const vllong8& reg) {
-      _mm512_mask_compressstoreu_epi64(addr,mask,reg);
-    }
-
-    static __forceinline vllong8 compact64bit(const vboold8& mask, vllong8& v) {
-      return _mm512_mask_compress_epi64(v,mask,v);
-    }
-
-    static __forceinline vllong8 compact64bit(const vboold8& mask, vllong8& dest, const vllong8& source) {
-      return _mm512_mask_compress_epi64(dest,mask,source);
-    }
-
     static __forceinline vllong8 compact(const vboold8& mask, vllong8& v) {
       return _mm512_mask_compress_epi64(v,mask,v);
     }
@@ -123,16 +118,6 @@ namespace embree
       return _mm512_mask_expand_epi64(b,mask,a);
     }
 
-    static __forceinline vllong8 broadcast64bit(size_t v) {
-      return _mm512_set1_epi64(v);
-    }
-
-    static __forceinline size_t extract64bit(const vllong8& v)
-    {
-      return _mm_cvtsi128_si64(_mm512_castsi512_si128(v));
-    }
-
-
     ////////////////////////////////////////////////////////////////////////////////
     /// Array Access
     ////////////////////////////////////////////////////////////////////////////////
@@ -271,18 +256,6 @@ namespace embree
     return _mm512_mask_or_epi64(f,m,t,t); 
   }
 
-  __forceinline void xchg(const vboold8& m, vllong8& a, vllong8& b) {
-    const vllong8 c = a; a = select(m,b,a); b = select(m,c,b);
-  }
-
-  __forceinline vboold8 test(const vboold8& m, const vllong8& a, const vllong8& b) {
-    return _mm512_mask_test_epi64_mask(m,a,b);
-  }
-
-  __forceinline vboold8 test(const vllong8& a, const vllong8& b) {
-    return _mm512_test_epi64_mask(a,b);
-  }
-
   ////////////////////////////////////////////////////////////////////////////////
   // Movement/Shifting/Shuffling Functions
   ////////////////////////////////////////////////////////////////////////////////
@@ -321,10 +294,6 @@ namespace embree
     return _mm_cvtsi128_si64(_mm512_castsi512_si128(v));
   }
 
-  __forceinline vllong8 zeroExtend32Bit(const __m512i& a) {
-    return _mm512_cvtepu32_epi64(_mm512_castsi512_si256(a));
-  }
-
   ////////////////////////////////////////////////////////////////////////////////
   /// Reductions
   ////////////////////////////////////////////////////////////////////////////////
@@ -379,3 +348,11 @@ namespace embree
     return cout;
   }
 }
+
+#undef vboolf
+#undef vboold
+#undef vint
+#undef vuint
+#undef vllong
+#undef vfloat
+#undef vdouble

thirdparty/embree/common/simd/vuint16_avx512.h

@@ -1,8 +1,16 @@
-// Copyright 2009-2020 Intel Corporation
+// Copyright 2009-2021 Intel Corporation
 // SPDX-License-Identifier: Apache-2.0
 
 #pragma once
 
+#define vboolf vboolf_impl
+#define vboold vboold_impl
+#define vint vint_impl
+#define vuint vuint_impl
+#define vllong vllong_impl
+#define vfloat vfloat_impl
+#define vdouble vdouble_impl
+
 namespace embree
 { 
   /* 16-wide AVX-512 unsigned integer type */
@@ -83,7 +91,7 @@ namespace embree
       return _mm512_loadu_si512(addr);
     }
 
-    static __forceinline vuint16 loadu(const uint8_t* ptr) { return _mm512_cvtepu8_epi32(_mm_loadu_si128((__m128i*)ptr)); }
+    static __forceinline vuint16 loadu(const unsigned char* 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) {
@@ -113,20 +121,6 @@ namespace embree
       _mm512_mask_store_epi32(addr,mask,v2);
     }
 
-    /* pass by value to avoid compiler generating inefficient code */
-    static __forceinline void storeu_compact(const vboolf16 mask, void* addr, const vuint16 reg) {
-      _mm512_mask_compressstoreu_epi32(addr,mask,reg);
-    }
-
-    static __forceinline void storeu_compact_single(const vboolf16 mask, void* addr, vuint16 reg) {
-      //_mm512_mask_compressstoreu_epi32(addr,mask,reg);
-      *(float*)addr = mm512_cvtss_f32(_mm512_mask_compress_ps(_mm512_castsi512_ps(reg),mask,_mm512_castsi512_ps(reg)));
-    }
-
-    static __forceinline vuint16 compact64bit(const vboolf16& mask, vuint16& v) {
-      return _mm512_mask_compress_epi64(v,mask,v);
-    }
-
     static __forceinline vuint16 compact(const vboolf16& mask, vuint16& v) {
       return _mm512_mask_compress_epi32(v,mask,v);
     }
@@ -164,15 +158,6 @@ namespace embree
       _mm512_mask_i32scatter_epi32((int*)ptr,mask,index,v,scale);
     }
 
-    static __forceinline vuint16 broadcast64bit(size_t v) {
-      return _mm512_set1_epi64(v);
-    }
-
-    static __forceinline size_t extract64bit(const vuint16& v)
-    {
-      return _mm_cvtsi128_si64(_mm512_castsi512_si128(v));
-    }
-
     ////////////////////////////////////////////////////////////////////////////////
     /// Array Access
     ////////////////////////////////////////////////////////////////////////////////
@@ -315,18 +300,6 @@ namespace embree
     return _mm512_mask_or_epi32(f,m,t,t); 
   }
 
-  __forceinline void xchg(const vboolf16& m, vuint16& a, vuint16& b) {
-    const vuint16 c = a; a = select(m,b,a); b = select(m,c,b);
-  }
-
-  __forceinline vboolf16 test(const vboolf16& m, const vuint16& a, const vuint16& b) {
-    return _mm512_mask_test_epi32_mask(m,a,b);
-  }
-
-  __forceinline vboolf16 test(const vuint16& a, const vuint16& b) {
-    return _mm512_test_epi32_mask(a,b);
-  }
-
   ////////////////////////////////////////////////////////////////////////////////
   // Movement/Shifting/Shuffling Functions
   ////////////////////////////////////////////////////////////////////////////////
@@ -441,3 +414,11 @@ namespace embree
     return cout;
   }
 }
+
+#undef vboolf
+#undef vboold
+#undef vint
+#undef vuint
+#undef vllong
+#undef vfloat
+#undef vdouble

thirdparty/embree/common/simd/vuint4_sse2.h

@@ -1,10 +1,18 @@
-// Copyright 2009-2020 Intel Corporation
+// Copyright 2009-2021 Intel Corporation
 // SPDX-License-Identifier: Apache-2.0
 
 #pragma once
 
 #include "../math/math.h"
 
+#define vboolf vboolf_impl
+#define vboold vboold_impl
+#define vint vint_impl
+#define vuint vuint_impl
+#define vllong vllong_impl
+#define vfloat vfloat_impl
+#define vdouble vdouble_impl
+
 namespace embree
 {
   /* 4-wide SSE integer type */
@@ -87,64 +95,27 @@ namespace embree
     static __forceinline void storeu(const vboolf4& mask, void* ptr, const vuint4& i) { storeu(ptr,select(mask,i,loadu(ptr))); }
 #endif
 
-#if defined(__aarch64__)
+#if defined(__SSE4_1__)
-    static __forceinline vuint4 load(const uint8_t* ptr) {
+    static __forceinline vuint4 load(const unsigned char* 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 uint8_t* ptr) {
+    static __forceinline vuint4 loadu(const unsigned char* ptr) {
       return  _mm_cvtepu8_epi32(_mm_loadl_epi64((__m128i*)ptr));
     }
 
 #endif
 
     static __forceinline vuint4 load(const unsigned short* ptr) {
-#if defined(__aarch64__)
+#if defined (__SSE4_1__)
-      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_uint8(uint8_t* ptr, const vuint4& v) {
-#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] = (uint8_t)v[i];
-#endif
-    }
-
-    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(__aarch64__)) || defined(__SSE4_1__)
+#if defined(__SSE4_1__)
       return _mm_stream_load_si128((__m128i*)ptr); 
 #else
       return _mm_load_si128((__m128i*)ptr); 
@@ -152,7 +123,7 @@ namespace embree
     }
     
     static __forceinline void store_nt(void* ptr, const vuint4& v) {
-#if !defined(__aarch64__) && defined(__SSE4_1__)
+#if defined(__SSE4_1__)
       _mm_stream_ps((float*)ptr,_mm_castsi128_ps(v)); 
 #else
       _mm_store_si128((__m128i*)ptr,v);
@@ -161,14 +132,14 @@ namespace embree
 
     template<int scale = 4>
     static __forceinline vuint4 gather(const unsigned int* ptr, const vint4& index) {
-#if defined(__AVX2__) && !defined(__aarch64__)
+#if defined(__AVX2__)
       return _mm_i32gather_epi32((const int*)ptr, index, scale);
 #else
       return vuint4(
-          *(unsigned int*)(((int8_t*)ptr)+scale*index[0]),
+          *(unsigned int*)(((char*)ptr)+scale*index[0]),
-          *(unsigned int*)(((int8_t*)ptr)+scale*index[1]),
+          *(unsigned int*)(((char*)ptr)+scale*index[1]),
-          *(unsigned int*)(((int8_t*)ptr)+scale*index[2]),
+          *(unsigned int*)(((char*)ptr)+scale*index[2]),
-          *(unsigned int*)(((int8_t*)ptr)+scale*index[3]));
+          *(unsigned int*)(((char*)ptr)+scale*index[3]));
 #endif
     }
 
@@ -177,13 +148,13 @@ namespace embree
       vuint4 r = zero;
 #if defined(__AVX512VL__)
       return _mm_mmask_i32gather_epi32(r, mask, index, ptr, scale);
-#elif defined(__AVX2__) && !defined(__aarch64__)
+#elif defined(__AVX2__)
       return _mm_mask_i32gather_epi32(r, (const int*)ptr, index, mask, scale);
 #else
-      if (likely(mask[0])) r[0] = *(unsigned int*)(((int8_t*)ptr)+scale*index[0]);
+      if (likely(mask[0])) r[0] = *(unsigned int*)(((char*)ptr)+scale*index[0]);
-      if (likely(mask[1])) r[1] = *(unsigned int*)(((int8_t*)ptr)+scale*index[1]);
+      if (likely(mask[1])) r[1] = *(unsigned int*)(((char*)ptr)+scale*index[1]);
-      if (likely(mask[2])) r[2] = *(unsigned int*)(((int8_t*)ptr)+scale*index[2]);
+      if (likely(mask[2])) r[2] = *(unsigned int*)(((char*)ptr)+scale*index[2]);
-      if (likely(mask[3])) r[3] = *(unsigned int*)(((int8_t*)ptr)+scale*index[3]);
+      if (likely(mask[3])) r[3] = *(unsigned int*)(((char*)ptr)+scale*index[3]);
       return r;
 #endif
     }
@@ -373,25 +344,16 @@ 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))); }
@@ -403,10 +365,7 @@ namespace embree
     return shuffle<i,i,i,i>(v);
   }
 
-#if defined(__aarch64__)
+#if defined(__SSE4_1__)
-  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
@@ -414,49 +373,10 @@ 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
@@ -497,3 +417,10 @@ namespace embree
   }
 }
 
+#undef vboolf
+#undef vboold
+#undef vint
+#undef vuint
+#undef vllong
+#undef vfloat
+#undef vdouble