void BVH::trace(RayBuffer& rays, RayStats* stats) const
{
for(S32 i=0;i<rays.getSize();i++)
{
Ray ray = rays.getRayForSlot(i); // takes a local copy
RayResult& result = rays.getMutableResultForSlot(i);
result.clear();
currentTreelet = -2;
uniqueTreelets.clear();
if(stats)
{
stats->platform = m_platform;
stats->numRays++;
}
traceRecursive(m_root, ray,result,rays.getNeedClosestHit(), stats);
}
}
F32 CudaTracer::traceBatch(RayBuffer& rays)
{
// No rays => done.
int numRays = rays.getSize();
if (!numRays)
return 0.0f;
// Check BVH consistency.
if (!m_bvh)
fail("CudaTracer: No BVH!");
if (m_bvh->getLayout() != getDesiredBVHLayout())
fail("CudaTracer: Incorrect BVH layout!");
// Get BVH buffers.
CUdeviceptr nodePtr = m_bvh->getNodeBuffer().getCudaPtr();
CUdeviceptr triPtr = m_bvh->getTriWoopBuffer().getCudaPtr();
Buffer& indexBuf = m_bvh->getTriIndexBuffer();
Vec2i nodeOfsA = m_bvh->getNodeSubArray(0);
Vec2i nodeOfsB = m_bvh->getNodeSubArray(1);
Vec2i nodeOfsC = m_bvh->getNodeSubArray(2);
Vec2i nodeOfsD = m_bvh->getNodeSubArray(3);
Vec2i triOfsA = m_bvh->getTriWoopSubArray(0);
Vec2i triOfsB = m_bvh->getTriWoopSubArray(1);
Vec2i triOfsC = m_bvh->getTriWoopSubArray(2);
// Compile kernel.
CudaModule* module = compileKernel();
CudaKernel kernel = module->getKernel("trace");
// Set parameters.
kernel.setParams(
numRays, // numRays
(rays.getNeedClosestHit()) ? 0 : 1, // anyHit
rays.getRayBuffer().getCudaPtr(), // rays
rays.getResultBuffer().getMutableCudaPtr(), // results
nodePtr + nodeOfsA.x, // nodesA
nodePtr + nodeOfsB.x, // nodesB
nodePtr + nodeOfsC.x, // nodesC
nodePtr + nodeOfsD.x, // nodesD
triPtr + triOfsA.x, // trisA
triPtr + triOfsB.x, // trisB
triPtr + triOfsC.x, // trisC
indexBuf.getCudaPtr()); // triIndices
// Set texture references.
module->setTexRef("t_rays", rays.getRayBuffer(), CU_AD_FORMAT_FLOAT, 4);
module->setTexRef("t_nodesA", nodePtr + nodeOfsA.x, nodeOfsA.y, CU_AD_FORMAT_FLOAT, 4);
module->setTexRef("t_nodesB", nodePtr + nodeOfsB.x, nodeOfsB.y, CU_AD_FORMAT_FLOAT, 4);
module->setTexRef("t_nodesC", nodePtr + nodeOfsC.x, nodeOfsC.y, CU_AD_FORMAT_FLOAT, 4);
module->setTexRef("t_nodesD", nodePtr + nodeOfsD.x, nodeOfsD.y, CU_AD_FORMAT_FLOAT, 4);
module->setTexRef("t_trisA", triPtr + triOfsA.x, triOfsA.y, CU_AD_FORMAT_FLOAT, 4);
module->setTexRef("t_trisB", triPtr + triOfsB.x, triOfsB.y, CU_AD_FORMAT_FLOAT, 4);
module->setTexRef("t_trisC", triPtr + triOfsC.x, triOfsC.y, CU_AD_FORMAT_FLOAT, 4);
module->setTexRef("t_triIndices", indexBuf, CU_AD_FORMAT_SIGNED_INT32, 1);
// Determine block and grid sizes.
int desiredWarps = (numRays + 31) / 32;
if (m_kernelConfig.usePersistentThreads != 0)
{
*(S32*)module->getGlobal("g_warpCounter").getMutablePtr() = 0;
desiredWarps = 720; // Tesla: 30 SMs * 24 warps, Fermi: 15 SMs * 48 warps
}
Vec2i blockSize(m_kernelConfig.blockWidth, m_kernelConfig.blockHeight);
int blockWarps = (blockSize.x * blockSize.y + 31) / 32;
int numBlocks = (desiredWarps + blockWarps - 1) / blockWarps;
// Launch.
return kernel.launchTimed(numBlocks * blockSize.x * blockSize.y, blockSize);
}