Speedtest now uses Crytek Sponza.

tiny_bvh.h

@@ -73,9 +73,9 @@ THE SOFTWARE.
 #define BVHBINS 8
 
 // SAH BVH building: Heuristic parameters
-// CPU builds: C_INT = 1, C_TRAV = 0.001 seems optimal.
+// CPU builds: C_INT = 1, C_TRAV = 0.01 seems optimal.
 #define C_INT	1
-#define C_TRAV	0.001f
+#define C_TRAV	0.01f
 
 // include fast AVX BVH builder
 #if defined(__x86_64__) || defined(_M_X64) || defined(__wasm_simd128__) || defined(__wasm_relaxed_simd__)
@@ -1080,7 +1080,7 @@ void BVH::Convert( BVHLayout from, BVHLayout to, const bool deleteOriginal )
 	else if (from == WALD_32BYTE && to == VERBOSE)
 	{
 		// allocate space
-		unsigned spaceNeeded = triCount * 2; // this one needs space to grow to 2N
+		unsigned spaceNeeded = triCount * (refittable ? 2 : 3); // this one needs space to grow to 2N
 		if (allocatedVerbose < spaceNeeded)
 		{
 			ALIGNED_FREE( verbose );
@@ -2495,14 +2495,13 @@ void BVH::BuildAVX( const bvhvec4* vertices, const unsigned primCount )
 			for (unsigned i = 0; i < node.triCount - 1; i++)
 			{
 				unsigned fid = *ti++;
+			#if defined __GNUC__ || _MSC_VER < 1920
+				if (fid > triCount) fid = triCount - 1; // never happens but g++ *and* vs2017 need this to not crash...
+			#endif
 				const __m256 b0 = binbox[i0], b1 = binbox[BVHBINS + i1], b2 = binbox[2 * BVHBINS + i2];
 				const __m128 fmin = frag4[fid].bmin4, fmax = frag4[fid].bmax4;
 				r0 = _mm256_max_ps( b0, f ), r1 = _mm256_max_ps( b1, f ), r2 = _mm256_max_ps( b2, f );
 				const __m128i b4 = _mm_cvtps_epi32( _mm_sub_ps( _mm_mul_ps( _mm_sub_ps( _mm_sub_ps( fmax, fmin ), nmin4 ), rpd4 ), half4 ) );
-			#if _MSC_VER < 1920 
-				// VS2017 needs a random check on fid to produce correct code... The condition is never true. 
-				if (fid > triCount) fid = triCount - 1;
-			#endif
 				f = frag8[fid], count[0][i0]++, count[1][i1]++, count[2][i2]++;
 				binbox[i0] = r0, i0 = ILANE( b4, 0 );
 				binbox[BVHBINS + i1] = r1, i1 = ILANE( b4, 1 );

tiny_bvh_fenster.cpp

@@ -20,9 +20,10 @@ BVH bvh;
 #endif
 
 #ifdef LOADSPONZA
-bvhvec4* triangles;
+bvhvec4* triangles = 0;
+#include <fstream>
 #else
-bvhvec4 triangles[259 /* level 3 */ * 6 * 2 * 49 * 3]{};
+ALIGNED( 16 ) bvhvec4 triangles[259 /* level 3 */ * 6 * 2 * 49 * 3]{};
 #endif
 int verts = 0;
 
@@ -52,12 +53,20 @@ void Init()
 {
 #ifdef LOADSPONZA
 	// load raw vertex data for Crytek's Sponza
-	FILE* f;
-	fopen_s( &f, "../testdata/cryteksponza.bin", "rb" );
-	fread( &verts, 1, 4, f );
-	verts *= 3, triangles = new bvhvec4[verts];
-	fread( triangles, sizeof( bvhvec4 ), verts, f );
-	fclose( f );
+	std::string filename{ "../testdata/cryteksponza.bin" };
+	std::fstream s{ filename, s.binary | s.in };
+	if (!s.is_open())
+	{
+		// try again, look in .\testdata
+		std::string filename{ "./testdata/cryteksponza.bin" };
+		s = std::fstream{ filename, s.binary | s.in };
+		assert( s.is_open() );
+	}
+	s.seekp( 0 );
+	s.read( (char*)&verts, 4 );
+	printf( "Loading triangle data (%i tris).\n", verts );
+	verts *= 3, triangles = (bvhvec4*)ALIGNED_MALLOC( verts * 16 );
+	s.read( (char*)triangles, verts * 16 );
 #else
 	// generate a sphere flake scene
 	sphere_flake( 0, 0, 0, 1.5f );
@@ -87,7 +96,7 @@ void Init()
 #if defined(BVH_USEAVX)
 	bvh.BuildAVX( triangles, verts / 3 );
 #elif defined(BVH_USENEON)
-    bvh.BuildNEON( triangles, verts / 3 );
+	bvh.BuildNEON( triangles, verts / 3 );
 #else
 	// bvh.Build( triangles, verts / 3 );
 #endif

tiny_bvh_speedtest.cpp

@@ -18,24 +18,32 @@
 #define SCRWIDTH	800
 #define SCRHEIGHT	600
 
+// scene selection
+#define LOADSPONZA
+
 // tests to perform
-// #define BUILD_REFERENCE
+#define BUILD_REFERENCE
 #define BUILD_AVX
 #define BUILD_NEON
-// #define BUILD_SBVH
+#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 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.
 
 using namespace tinybvh;
 
-bvhvec4 triangles[259 /* level 3 */ * 6 * 2 * 49 * 3]{};
+#ifdef LOADSPONZA
+bvhvec4* triangles = 0;
+#include <fstream>
+#else
+ALIGNED( 16 ) bvhvec4 triangles[259 /* level 3 */ * 6 * 2 * 49 * 3]{};
+#endif
 int verts = 0;
 BVH bvh;
 
@@ -87,12 +95,62 @@ void sphere_flake( float x, float y, float z, float s, int d = 0 )
 
 int main()
 {
+	int minor = TINY_BVH_VERSION_MINOR;
+	int major = TINY_BVH_VERSION_MAJOR;
+	int sub = TINY_BVH_VERSION_SUB;
+	printf( "tiny_bvh version %i.%i.%i performance statistics ", major, minor, sub );
+
+	// determine compiler
+#ifdef _MSC_VER
+	printf( "(MSVC %i build)\n", _MSC_VER );
+#elif defined __EMSCRIPTEN__
+	// EMSCRIPTEN needs to be before clang or gcc
+	printf( "(emcc %i.%i build)\n", __EMSCRIPTEN_major__, __EMSCRIPTEN_minor__ );
+#elif defined __clang__
+	printf( "(clang %i.%i build)\n", __clang_major__, __clang_minor__ );
+#elif defined __GNUC__
+	printf( "(gcc %i.%i build)\n", __GNUC__, __GNUC_MINOR__ );
+#else
+	printf( "\n" );
+#endif
+
+	// determine what CPU is running the tests.
+#ifdef _WIN32
+	char model[256]{};
+	for (unsigned i = 0; i < 3; ++i) __cpuidex( (int*)(model + i * 16), i + 0x80000002, 0 );
+	printf( "running on %s\n", model );
+#endif
+	printf( "----------------------------------------------------------------\n" );
+
+	#ifdef LOADSPONZA
+	// load raw vertex data for Crytek's Sponza
+	std::string filename{ "../testdata/cryteksponza.bin" };
+	std::fstream s{ filename, s.binary | s.in };
+	if (!s.is_open())
+	{
+		// try again, look in .\testdata
+		filename = std::string{ "./testdata/cryteksponza.bin" };
+		s = std::fstream{ filename, s.binary | s.in };
+		assert( s.is_open() );
+	}
+	s.seekp( 0 );
+	s.read( (char*)&verts, 4 );
+	printf( "Loading triangle data (%i tris).\n", verts );
+	verts *= 3, triangles = (bvhvec4*)ALIGNED_MALLOC( verts * 16 );
+	s.read( (char*)triangles, verts * 16 );
+#else
 	// generate a sphere flake scene
+	printf( "Creating sphere flake (%i tris).\n", verts / 3 );
 	sphere_flake( 0, 0, 0, 1.5f );
+#endif
 
 	// setup view pyramid for a pinhole camera: 
 	// eye, p1 (top-left), p2 (top-right) and p3 (bottom-left)
+#ifdef LOADSPONZA
+	bvhvec3 eye( 0, 30, 0 ), view = normalize( bvhvec3( -8, 2, -1.7f ) );
+#else
 	bvhvec3 eye( -3.5f, -1.5f, -6.5f ), view = normalize( bvhvec3( 3, 1.5f, 5 ) );
+#endif
 	bvhvec3 right = normalize( cross( bvhvec3( 0, 1, 0 ), view ) );
 	bvhvec3 up = 0.8f * cross( view, right ), C = eye + 2 * view;
 	bvhvec3 p1 = C - right + up, p2 = C + right + up, p3 = C - right - up;
@@ -119,40 +177,13 @@ int main()
 
 	//  T I N Y _ B V H   P E R F O R M A N C E   M E A S U R E M E N T S
 
-	int minor = TINY_BVH_VERSION_MINOR;
-	int major = TINY_BVH_VERSION_MAJOR;
-	int sub = TINY_BVH_VERSION_SUB;
-	printf( "tiny_bvh version %i.%i.%i performance statistics ", major, minor, sub );
-
-	// determine compiler
-#ifdef _MSC_VER
-	printf( "(MSVC %i build)\n", _MSC_VER );
-#elif defined __EMSCRIPTEN__
-	// EMSCRIPTEN needs to be before clang or gcc
-	printf( "(emcc %i.%i build)\n", __EMSCRIPTEN_major__, __EMSCRIPTEN_minor__ );
-#elif defined __clang__
-	printf( "(clang %i.%i build)\n", __clang_major__, __clang_minor__ );
-#elif defined __GNUC__
-	printf( "(gcc %i.%i build)\n", __GNUC__, __GNUC_MINOR__ );
-#else
-	printf( "\n" );
-#endif
-
-	// determine what CPU is running the tests.
-#ifdef _WIN32
-	char model[256]{};
-	for (unsigned i = 0; i < 3; ++i) __cpuidex( (int*)(model + i * 16), i + 0x80000002, 0 );
-	printf( "running on %s\n", model );
-#endif
-	printf( "----------------------------------------------------------------\n" );
-
 	Timer t;
 	float mrays;
 
 	// measure single-core bvh construction time - warming caches
 	printf( "BVH construction speed\n" );
 	printf( "warming caches...\n" );
-	bvh.Build( (bvhvec4*)triangles, verts / 3 );
+	bvh.Build( triangles, verts / 3 );
 
 #ifdef BUILD_REFERENCE