Cleaned up new memory interface.

jbikker · Nov 18, 2024 · 2eb739a · 2eb739a
1 parent c8c3d3c
commit 2eb739a
Show file tree

Hide file tree

Showing 3 changed files with 38 additions and 34 deletions.
diff --git a/tiny_bvh.h b/tiny_bvh.h
@@ -124,51 +124,51 @@ inline size_t make_multiple_64( size_t x ) { return (x + 63) & ~0x3f; }
 #ifdef _MSC_VER // Visual Studio / C11
 #define ALIGNED( x ) __declspec( align( x ) )
 namespace tinybvh {
-inline void* default_aligned_malloc( size_t size, void* = nullptr )
+inline void* malloc64( size_t size, void* = nullptr )
 {
 	return size == 0 ? 0 : _aligned_malloc( make_multiple_64( size ), 64 );
 }
-inline void default_aligned_free( void* ptr, void* = nullptr ) { _aligned_free( ptr ); }
+inline void free64( void* ptr, void* = nullptr ) { _aligned_free( ptr ); }
 }
 #elif defined(__EMSCRIPTEN__) // EMSCRIPTEN - needs to be before gcc and clang to avoid misdetection
 #define ALIGNED( x ) __attribute__( ( aligned( x ) ) )
 #if defined(__wasm_simd128__) || defined(__wasm_relaxed_simd__)
 // https://emscripten.org/docs/porting/simd.html
 #include <xmmintrin.h>
 namespace tinybvh {
-inline void* default_aligned_malloc( size_t size, void* = nullptr )
+inline void* malloc64( size_t size, void* = nullptr )
 {
 	return size == 0 ? 0 : _mm_malloc( size, 64 );
 }
-inline void default_aligned_free( void* ptr, void* = nullptr ) { _mm_free( ptr ); }
+inline void free64( void* ptr, void* = nullptr ) { _mm_free( ptr ); }
 }
 #else
 namespace tinybvh {
-inline void* default_aligned_malloc( size_t size, void* = nullptr )
+inline void* malloc64( size_t size, void* = nullptr )
 {
 	return size == 0 ? 0 : aligned_alloc( 64, make_multiple_64( size ) );
 }
-inline void default_aligned_free( void* ptr, void* = nullptr ) { free( ptr ); }
+inline void free64( void* ptr, void* = nullptr ) { free( ptr ); }
 }
 #endif
 #else // gcc / clang
 #define ALIGNED( x ) __attribute__( ( aligned( x ) ) )
 #if defined(__x86_64__) || defined(_M_X64)
 #include <xmmintrin.h>
 namespace tinybvh {
-inline void* default_aligned_malloc( size_t size, void* = nullptr )
+inline void* malloc64( size_t size, void* = nullptr )
 {
 	return size == 0 ? 0 : _mm_malloc( make_multiple_64( size ), 64 );
 }
-inline void default_aligned_free( void* ptr, void* = nullptr ) { _mm_free( ptr ); }
+inline void free64( void* ptr, void* = nullptr ) { _mm_free( ptr ); }
 }
 #else
 namespace tinybvh {
-inline void* default_aligned_malloc( size_t size, void* = nullptr )
+inline void* malloc64( size_t size, void* = nullptr )
 {
 	return size == 0 ? 0 : aligned_alloc( 64, make_multiple_64( size ) );
 }
-inline void default_aligned_free( void* ptr, void* = nullptr ) { free( ptr ); }
+inline void free64( void* ptr, void* = nullptr ) { free( ptr ); }
 }
 #endif
 #endif
@@ -375,8 +375,8 @@ struct Ray
 
 struct BVHContext
 {
-	void* (*malloc)(size_t size, void* userdata) = default_aligned_malloc;
-	void (*free)(void* ptr, void* userdata) = default_aligned_free;
+	void* (*malloc)(size_t size, void* userdata) = malloc64;
+	void (*free)(void* ptr, void* userdata) = free64;
 	void* userdata = nullptr;
 };
 
@@ -2480,11 +2480,12 @@ void BVH::BuildAVX( const bvhvec4* vertices, const unsigned primCount )
 {
 	int test = BVHBINS;
 	if (test != 8) assert( false ); // AVX builders require BVHBINS == 8.
+	assert( ((long long)vertices & 63) == 0 ); // buffer must be cacheline-aligned
 	// aligned data
-	ALIGNED( 64 ) __m256 binbox[3 * BVHBINS];				// 768 bytes
-	ALIGNED( 64 ) __m256 binboxOrig[3 * BVHBINS];			// 768 bytes
+	ALIGNED( 64 ) __m256 binbox[3 * BVHBINS];			// 768 bytes
+	ALIGNED( 64 ) __m256 binboxOrig[3 * BVHBINS];		// 768 bytes
 	ALIGNED( 64 ) unsigned count[3][BVHBINS]{};			// 96 bytes
-	ALIGNED( 64 ) __m256 bestLBox, bestRBox;				// 64 bytes
+	ALIGNED( 64 ) __m256 bestLBox, bestRBox;			// 64 bytes
 	// some constants
 	static const __m128 max4 = _mm_set1_ps( -1e30f ), half4 = _mm_set1_ps( 0.5f );
 	static const __m128 two4 = _mm_set1_ps( 2.0f ), min1 = _mm_set1_ps( -1 );
@@ -3141,11 +3142,12 @@ void BVH::BuildNEON( const bvhvec4* vertices, const unsigned primCount )
 {
 	int test = BVHBINS;
 	if (test != 8) assert( false ); // AVX builders require BVHBINS == 8.
+	assert( ((long long)vertices & 63) == 0 ); // buffer must be cacheline-aligned
 	// aligned data
-	ALIGNED( 64 ) float32x4x2_t binbox[3 * BVHBINS];                // 768 bytes
-	ALIGNED( 64 ) float32x4x2_t binboxOrig[3 * BVHBINS];            // 768 bytes
-	ALIGNED( 64 ) unsigned count[3][BVHBINS]{};            // 96 bytes
-	ALIGNED( 64 ) float32x4x2_t bestLBox, bestRBox;                // 64 bytes
+	ALIGNED( 64 ) float32x4x2_t binbox[3 * BVHBINS];		// 768 bytes
+	ALIGNED( 64 ) float32x4x2_t binboxOrig[3 * BVHBINS];	// 768 bytes
+	ALIGNED( 64 ) unsigned count[3][BVHBINS]{};				// 96 bytes
+	ALIGNED( 64 ) float32x4x2_t bestLBox, bestRBox;			// 64 bytes
 	// some constants
 	static const float32x4_t max4 = vdupq_n_f32( -1e30f ), half4 = vdupq_n_f32( 0.5f );
 	static const float32x4_t two4 = vdupq_n_f32( 2.0f ), min1 = vdupq_n_f32( -1 );

diff --git a/tiny_bvh_speedtest.cpp b/tiny_bvh_speedtest.cpp
@@ -14,16 +14,16 @@
 
 // tests to perform
 #define BUILD_REFERENCE
-// #define BUILD_AVX
+#define BUILD_AVX
 #define BUILD_NEON
-//#define BUILD_SBVH
-//#define TRAVERSE_2WAY_ST
-//#define TRAVERSE_ALT2WAY_ST
-//#define TRAVERSE_SOA2WAY_ST
-//#define TRAVERSE_2WAY_MT
-//#define TRAVERSE_2WAY_MT_PACKET
-//#define TRAVERSE_2WAY_MT_DIVERGENT
-//#define TRAVERSE_OPTIMIZED_ST
+#define BUILD_SBVH
+#define TRAVERSE_2WAY_ST
+#define TRAVERSE_ALT2WAY_ST
+#define TRAVERSE_SOA2WAY_ST
+#define TRAVERSE_2WAY_MT
+#define TRAVERSE_2WAY_MT_PACKET
+#define TRAVERSE_2WAY_MT_DIVERGENT
+#define TRAVERSE_OPTIMIZED_ST
 // #define EMBREE_BUILD // win64-only for now.
 // #define EMBREE_TRAVERSE // win64-only for now.
 
@@ -164,7 +164,7 @@ int main()
 	s.seekp( 0 );
 	s.read( (char*)&verts, 4 );
 	printf( "Loading triangle data (%i tris).\n", verts );
-	verts *= 3, triangles = (bvhvec4*)tinybvh::default_aligned_malloc( verts * 16 );
+	verts *= 3, triangles = (bvhvec4*)tinybvh::malloc64( verts * 16 );
 	s.read( (char*)triangles, verts * 16 );
 #else
 	// generate a sphere flake scene
@@ -186,7 +186,7 @@ int main()
 	// generate primary rays in a cacheline-aligned buffer - and, for data locality:
 	// organized in 4x4 pixel tiles, 16 samples per pixel, so 256 rays per tile.
 	int N = 0;
-	Ray* rays = (Ray*)tinybvh::default_aligned_malloc( SCRWIDTH * SCRHEIGHT * 16 * sizeof( Ray ) );
+	Ray* rays = (Ray*)tinybvh::malloc64( SCRWIDTH * SCRHEIGHT * 16 * sizeof( Ray ) );
 	for (int ty = 0; ty < SCRHEIGHT / 4; ty++) for (int tx = 0; tx < SCRWIDTH / 4; tx++)
 	{
 		for (int y = 0; y < 4; y++) for (int x = 0; x < 4; x++)

diff --git a/tiny_ocl.h b/tiny_ocl.h
@@ -115,10 +115,11 @@ struct oclvec3 { float x, y, z; };
 // ============================================================================
 
 // OpenCL context
-// You only need to explicitly instantiate this if you want to override the
-// default memory allocator used by tinyocl::Buffer. In all other cases, the
-// first use of tinyocl (creating a buffer, loading a kernel) will take care
-// of this for you transparently.
+// *Only!* in case you want to override the default memory allocator used by 
+// tinyocl::Buffer: call OpenCL::CreateInstance with OpenCLContext fields describing
+// your allocator and deallocator, before using any tinyocl functionality.
+// In all other cases, the first use of tinyocl (creating a buffer, loading a 
+// kernel) will take care of this for you transparently.
 struct OpenCLContext
 {
 	void* (*malloc)(size_t size, void* userdata) = default_aligned_malloc;
@@ -132,6 +133,7 @@ class OpenCL
 	OpenCLContext context;
 	void* AlignedAlloc( size_t size );
 	void AlignedFree( void* ptr );
+	static void CreateInstance( OpenCLContext ctx ) { ocl = new OpenCL( ctx ); }
 	static OpenCL* GetInstance() { return ocl; }
 	inline static OpenCL* ocl = 0;
 };