Fix Aila & Laine layout gpu traversal.

tiny_bvh_fenster.cpp

@@ -1,7 +1,7 @@
 #include "external/fenster.h" // https://github.com/zserge/fenster
 
 // #define USE_EMBREE // enable to verify correct implementation, win64 only for now.
-// #define LOADSPONZA
+#define LOADSPONZA
 
 #define TINYBVH_IMPLEMENTATION
 #include "tiny_bvh.h"
@@ -65,7 +65,7 @@ void Init()
 	s.seekp( 0 );
 	s.read( (char*)&verts, 4 );
 	printf( "Loading triangle data (%i tris).\n", verts );
-	verts *= 3, triangles = (bvhvec4*)default_malloc( verts * 16 );
+	verts *= 3, triangles = (bvhvec4*)malloc64( verts * 16 );
 	s.read( (char*)triangles, verts * 16 );
 #else
 	// generate a sphere flake scene
@@ -95,6 +95,7 @@ void Init()
 	// build a BVH over the scene
 #if defined(BVH_USEAVX)
 	bvh.BuildAVX( triangles, verts / 3 );
+	bvh.Convert( BVH::WALD_32BYTE, BVH::AILA_LAINE );
 #elif defined(BVH_USENEON)
 	bvh.BuildNEON( triangles, verts / 3 );
 #else
@@ -152,7 +153,7 @@ void Tick( uint32_t* buf )
 		rays[i].hit.prim = rayhit.hit.primID, rays[i].hit.t = rayhit.ray.tfar;
 	}
 #else
-	for (int i = 0; i < N; i++) bvh.Intersect( rays[i] );
+	for (int i = 0; i < N; i++) bvh.Intersect( rays[i], BVH::AILA_LAINE );
 #endif
 
 	// visualize result

tiny_bvh_speedtest.cpp

@@ -23,7 +23,7 @@
 #define TRAVERSE_2WAY_MT
 #define TRAVERSE_2WAY_MT_PACKET
 #define TRAVERSE_2WAY_MT_DIVERGENT
-// #define TRAVERSE_OPTIMIZED_ST
+#define TRAVERSE_OPTIMIZED_ST
 // #define EMBREE_BUILD // win64-only for now.
 // #define EMBREE_TRAVERSE // win64-only for now.
 
@@ -90,6 +90,21 @@ struct Timer
 	std::chrono::high_resolution_clock::time_point start;
 };
 
+#if 0
+void dump_bitmap( Ray* rays )
+{
+	unsigned char pixel[SCRWIDTH * SCRHEIGHT]; float sum;
+	for (int s, x, y, i = 0, ty = 0; ty < SCRHEIGHT / 4; ty++) for (int tx = 0; tx < SCRWIDTH / 4; tx++)
+		for (y = 0; y < 4; y++) for (s = 0, x = 0; x < 4; x++)
+		{
+			for (s = 0, sum = 0; s < 16; s++, i++) sum += rays[i].hit.t == 1e30f ? 0 : rays[i].hit.t;
+			pixel[tx * 4 + x + (ty * 4 + y) * SCRWIDTH] = (unsigned char)((int)(sum * 0.01f) & 255);
+		}
+	FILE* f = fopen( "testimage.raw", "wb" );
+	fwrite( pixel, 1, SCRWIDTH * SCRHEIGHT, f ); // for debugging, forget fclose
+}
+#endif
+
 void sphere_flake( float x, float y, float z, float s, int d = 0 )
 {
 	// procedural tesselated sphere flake object
@@ -306,7 +321,7 @@ int main()
 	printf( "- CPU, coherent,   basic 2-way layout, ST: " );
 	t.reset();
 	for (int pass = 0; pass < 3; pass++)
-		for (int i = 0; i < N; i += 8 ) bvh.Intersect( rays[i] );
+		for (int i = 0; i < N; i += 8) bvh.Intersect( rays[i] );
 	float traceTimeST = t.elapsed() / 3.0f;
 	mrays = (float)(N / 8) / traceTimeST;
 	printf( "%8.1fms for %6.2fM rays => %6.2fMRay/s\n", traceTimeST * 1000, (float)(N / 8) * 1e-6f, mrays * 1e-6f );
@@ -347,7 +362,7 @@ int main()
 	tinyocl::Buffer rayData( N * sizeof( tinybvh::Ray ), rays );
 	rayData.CopyToDevice();
 	// create an event to time the OpenCL kernel
-	cl_event event; 
+	cl_event event;
 	cl_ulong startTime, endTime;
 	// start timer and start kernel on gpu
 	t.reset();
@@ -356,9 +371,9 @@ int main()
 	for (int pass = 0; pass < 8; pass++)
 	{
 		ailalaine_kernel.Run( N, 64, 0, &event ); // for now, todo.
-		clWaitForEvents(1, &event ); // OpenCL kernsl run asynchronously
-		clGetEventProfilingInfo( event, CL_PROFILING_COMMAND_START, sizeof( cl_ulong ), &startTime, 0 ); 
-		clGetEventProfilingInfo( event, CL_PROFILING_COMMAND_END, sizeof( cl_ulong ), &endTime, 0 ); 
+		clWaitForEvents( 1, &event ); // OpenCL kernsl run asynchronously
+		clGetEventProfilingInfo( event, CL_PROFILING_COMMAND_START, sizeof( cl_ulong ), &startTime, 0 );
+		clGetEventProfilingInfo( event, CL_PROFILING_COMMAND_END, sizeof( cl_ulong ), &endTime, 0 );
 		traceTimeGPU += (endTime - startTime) * 1e-9f; // event timing is in nanoseconds
 	}
 	// get results from GPU - this also syncs the queue.
@@ -367,6 +382,7 @@ int main()
 	traceTimeGPU /= 8.0f;
 	mrays = (float)N / traceTimeGPU;
 	printf( "%8.1fms for %6.2fM rays => %6.2fMRay/s\n", traceTimeGPU * 1000, (float)N * 1e-6f, mrays * 1e-6f );
+	dump_bitmap( rays );
 
 #endif
 
@@ -382,7 +398,7 @@ int main()
 	gpu4Nodes.CopyToDevice();
 #ifndef GPU_2WAY // otherwise these already exist.
 	// create an event to time the OpenCL kernel
-	cl_event event; 
+	cl_event event;
 	cl_ulong startTime, endTime;
 	// create rays and send them to the gpu side
 	tinyocl::Buffer rayData( N * sizeof( tinybvh::Ray ), rays );
@@ -395,9 +411,9 @@ int main()
 	for (int pass = 0; pass < 8; pass++)
 	{
 		gpu4way_kernel.Run( N, 64, 0, &event ); // for now, todo.
-		clWaitForEvents(1, &event ); // OpenCL kernsl run asynchronously
-		clGetEventProfilingInfo( event, CL_PROFILING_COMMAND_START, sizeof( cl_ulong ), &startTime, 0 ); 
-		clGetEventProfilingInfo( event, CL_PROFILING_COMMAND_END, sizeof( cl_ulong ), &endTime, 0 ); 
+		clWaitForEvents( 1, &event ); // OpenCL kernsl run asynchronously
+		clGetEventProfilingInfo( event, CL_PROFILING_COMMAND_START, sizeof( cl_ulong ), &startTime, 0 );
+		clGetEventProfilingInfo( event, CL_PROFILING_COMMAND_END, sizeof( cl_ulong ), &endTime, 0 );
 		traceTimeGPU4 += (endTime - startTime) * 1e-9f; // event timing is in nanoseconds
 	}
 	// get results from GPU - this also syncs the queue.

traverse.cl

@@ -23,8 +23,8 @@ void kernel traverse_ailalaine( global struct BVHNodeAlt* altNode, global unsign
 	const float3 O = rayData[threadId].O.xyz;
 	const float3 D = rayData[threadId].D.xyz;
 	const float3 rD = rayData[threadId].rD.xyz;
-	float t = 1e30f; // ignoring value set in ray to spare one memory transaction.
 	float4 hit;
+	hit.x = 1e30f; // ignoring value set in ray to spare one memory transaction.
 	// traverse BVH
 	unsigned node = 0, stack[64], stackPtr = 0;
 	while (1)
@@ -49,12 +49,12 @@ void kernel traverse_ailalaine( global struct BVHNodeAlt* altNode, global unsign
 				const float f = 1 / a;
 				const float3 s = O - tri[0].xyz;
 				const float u = f * dot( s, h );
-				if (u < 0 && u > 1) continue;
+				if (u < 0 || u > 1) continue;
 				const float3 q = cross( s, edge1.xyz );
 				const float v = f * dot( D, q );
-				if (v < 0 && u + v > 1) continue;
+				if (v < 0 || u + v > 1) continue;
 				const float d = f * dot( edge2.xyz, q );
-				if (d > 0.0f && d < t) hit = (float4)(t = d, u, v, as_float( triIdx ));
+				if (d > 0.0f && d < hit.x) hit = (float4)(d, u, v, as_float( triIdx ));
 			}
 			if (stackPtr == 0) break;
 			node = stack[--stackPtr];
@@ -68,8 +68,8 @@ void kernel traverse_ailalaine( global struct BVHNodeAlt* altNode, global unsign
 		const float3 mintb = fmin( t1b, t2b ), maxtb = fmax( t1b, t2b );
 		const float tmina = fmax( fmax( fmax( minta.x, minta.y ), minta.z ), 0 );
 		const float tminb = fmax( fmax( fmax( mintb.x, mintb.y ), mintb.z ), 0 );
-		const float tmaxa = fmin( fmin( fmin( maxta.x, maxta.y ), maxta.z ), t );
-		const float tmaxb = fmin( fmin( fmin( maxtb.x, maxtb.y ), maxtb.z ), t );
+		const float tmaxa = fmin( fmin( fmin( maxta.x, maxta.y ), maxta.z ), hit.x );
+		const float tmaxb = fmin( fmin( fmin( maxtb.x, maxtb.y ), maxtb.z ), hit.x );
 		float dist1 = tmina > tmaxa ? 1e30f : tmina;
 		float dist2 = tminb > tmaxb ? 1e30f : tminb;
 		// traverse nearest child first