void OpenCLBVHKernel::FreeBuffers()
{
delete kernel;
kernel = NULL;
OpenCLDeviceDescription *deviceDesc = device->GetDeviceDesc();
deviceDesc->FreeMemory(vertsBuff->getInfo<CL_MEM_SIZE>());
delete vertsBuff;
vertsBuff = NULL;
deviceDesc->FreeMemory(trisBuff->getInfo<CL_MEM_SIZE>());
delete trisBuff;
trisBuff = NULL;
deviceDesc->FreeMemory(bvhBuff->getInfo<CL_MEM_SIZE>());
delete bvhBuff;
bvhBuff = NULL;
}
OpenCLKernel *BVHAccel::NewOpenCLKernel(OpenCLIntersectionDevice *dev,
unsigned int stackSize, bool disableImageStorage) const
{
OpenCLBVHKernel *kernel = new OpenCLBVHKernel(dev);
const Context *deviceContext = dev->GetContext();
cl::Context &oclContext = dev->GetOpenCLContext();
const std::string &deviceName(dev->GetName());
OpenCLDeviceDescription *deviceDesc = dev->GetDeviceDesc();
LR_LOG(deviceContext, "[OpenCL device::" << deviceName <<
"] Vertices buffer size: " <<
(sizeof(Point) * preprocessedMesh->GetTotalVertexCount() / 1024) <<
"Kbytes");
cl::Buffer *vertsBuff = new cl::Buffer(oclContext,
CL_MEM_READ_ONLY | CL_MEM_COPY_HOST_PTR,
sizeof(Point) * preprocessedMesh->GetTotalVertexCount(),
preprocessedMesh->GetVertices());
deviceDesc->AllocMemory(vertsBuff->getInfo<CL_MEM_SIZE>());
LR_LOG(deviceContext, "[OpenCL device::" << deviceName <<
"] Triangle indices buffer size: " <<
(sizeof(Triangle) * preprocessedMesh->GetTotalTriangleCount() / 1024) <<
"Kbytes");
cl::Buffer *trisBuff = new cl::Buffer(oclContext,
CL_MEM_READ_ONLY | CL_MEM_COPY_HOST_PTR,
sizeof(Triangle) * preprocessedMesh->GetTotalTriangleCount(),
preprocessedMesh->GetTriangles());
deviceDesc->AllocMemory(trisBuff->getInfo<CL_MEM_SIZE>());
LR_LOG(deviceContext, "[OpenCL device::" << deviceName <<
"] BVH buffer size: " <<
(sizeof(BVHAccelArrayNode) * nNodes / 1024) <<
"Kbytes");
cl::Buffer *bvhBuff = new cl::Buffer(oclContext,
CL_MEM_READ_ONLY | CL_MEM_COPY_HOST_PTR,
sizeof(BVHAccelArrayNode) * nNodes,
(void*)bvhTree);
deviceDesc->AllocMemory(bvhBuff->getInfo<CL_MEM_SIZE>());
kernel->SetBuffers(vertsBuff, preprocessedMesh->GetTotalTriangleCount(),
trisBuff, nNodes, bvhBuff);
return kernel;
}
OCLRenderEngine::OCLRenderEngine(RenderConfig *rcfg, Film *flm,
boost::mutex *flmMutex, bool fatal) : RenderEngine(rcfg, flm, flmMutex) {
#if !defined(LUXRAYS_DISABLE_OPENCL)
const Properties &cfg = renderConfig->cfg;
const bool useCPUs = (cfg.GetInt("opencl.cpu.use", 1) != 0);
const bool useGPUs = (cfg.GetInt("opencl.gpu.use", 1) != 0);
const u_int forceGPUWorkSize = cfg.GetInt("opencl.gpu.workgroup.size", 64);
const u_int forceCPUWorkSize = cfg.GetInt("opencl.cpu.workgroup.size", 1);
const string oclDeviceConfig = cfg.GetString("opencl.devices.select", "");
// Start OpenCL devices
std::vector<DeviceDescription *> descs = ctx->GetAvailableDeviceDescriptions();
DeviceDescription::Filter(DEVICE_TYPE_OPENCL_ALL, descs);
// Device info
bool haveSelectionString = (oclDeviceConfig.length() > 0);
if (haveSelectionString && (oclDeviceConfig.length() != descs.size())) {
stringstream ss;
ss << "OpenCL device selection string has the wrong length, must be " <<
descs.size() << " instead of " << oclDeviceConfig.length();
throw runtime_error(ss.str().c_str());
}
for (size_t i = 0; i < descs.size(); ++i) {
OpenCLDeviceDescription *desc = static_cast<OpenCLDeviceDescription *>(descs[i]);
if (haveSelectionString) {
if (oclDeviceConfig.at(i) == '1') {
if (desc->GetType() == DEVICE_TYPE_OPENCL_GPU)
desc->SetForceWorkGroupSize(forceGPUWorkSize);
else if (desc->GetType() == DEVICE_TYPE_OPENCL_CPU)
desc->SetForceWorkGroupSize(forceCPUWorkSize);
selectedDeviceDescs.push_back(desc);
}
} else {
if ((useCPUs && desc->GetType() == DEVICE_TYPE_OPENCL_CPU) ||
(useGPUs && desc->GetType() == DEVICE_TYPE_OPENCL_GPU)) {
if (desc->GetType() == DEVICE_TYPE_OPENCL_GPU)
desc->SetForceWorkGroupSize(forceGPUWorkSize);
else if (desc->GetType() == DEVICE_TYPE_OPENCL_CPU)
desc->SetForceWorkGroupSize(forceCPUWorkSize);
selectedDeviceDescs.push_back(descs[i]);
}
}
}
#endif
if (fatal && selectedDeviceDescs.size() == 0)
throw runtime_error("No OpenCL device selected or available");
}
void RenderingConfig::SetUpOpenCLDevices(const bool useCPUs, const bool useGPUs,
const unsigned int forceGPUWorkSize, const unsigned int oclDeviceThreads, const string &oclDeviceConfig) {
std::vector<DeviceDescription *> descs = ctx->GetAvailableDeviceDescriptions();
DeviceDescription::Filter(DEVICE_TYPE_OPENCL, descs);
// Device info
bool haveSelectionString = (oclDeviceConfig.length() > 0);
if (haveSelectionString && (oclDeviceConfig.length() != descs.size())) {
stringstream ss;
ss << "OpenCL device selection string has the wrong length, must be " <<
descs.size() << " instead of " << oclDeviceConfig.length();
throw runtime_error(ss.str().c_str());
}
std::vector<DeviceDescription *> selectedDescs;
#if !defined(LUXRAYS_DISABLE_OPENCL)
for (size_t i = 0; i < descs.size(); ++i) {
OpenCLDeviceDescription *desc = (OpenCLDeviceDescription *)descs[i];
if (haveSelectionString) {
if (oclDeviceConfig.at(i) == '1') {
if (desc->GetOpenCLType() == OCL_DEVICE_TYPE_GPU)
desc->SetForceWorkGroupSize(forceGPUWorkSize);
selectedDescs.push_back(desc);
}
} else {
if ((useCPUs && desc->GetOpenCLType() == OCL_DEVICE_TYPE_CPU) ||
(useGPUs && desc->GetOpenCLType() == OCL_DEVICE_TYPE_GPU)) {
if (desc->GetOpenCLType() == OCL_DEVICE_TYPE_GPU)
desc->SetForceWorkGroupSize(forceGPUWorkSize);
selectedDescs.push_back(descs[i]);
}
}
}
#endif
if (selectedDescs.size() == 0)
cerr << "No OpenCL device selected" << endl;
else {
#if !defined(LUXRAYS_DISABLE_OPENCL)
if (cfg.GetInt("opencl.latency.mode", 1) && (cfg.GetInt("renderengine.type", 0) == 3) &&
(cfg.GetInt("pathgpu.openglinterop.enable", 0) != 0)) {
// Ask for OpenGL interoperability on the first device
((OpenCLDeviceDescription *)selectedDescs[0])->EnableOGLInterop();
}
#endif
// Allocate devices
const size_t gpuRenderThreadCount = (oclDeviceThreads < 1) ?
(2 * selectedDescs.size()) : oclDeviceThreads;
if ((gpuRenderThreadCount == 1) && (selectedDescs.size() == 1)) {
// Optimize the special case of one render thread and one GPU
intersectionGPUDevices = ctx->AddIntersectionDevices(selectedDescs);
} else if ((gpuRenderThreadCount > 1) && (selectedDescs.size() == 1)) {
// Optimize the special case of many render thread and one GPU
intersectionGPUDevices = ctx->AddVirtualM2OIntersectionDevices(gpuRenderThreadCount, selectedDescs);
} else {
// Create and start the virtual devices (only if there is more than one GPUs)
intersectionGPUDevices = ctx->AddVirtualM2MIntersectionDevices(gpuRenderThreadCount, selectedDescs);
}
cerr << "OpenCL Devices used: ";
for (size_t i = 0; i < intersectionGPUDevices.size(); ++i)
cerr << "[" << intersectionGPUDevices[i]->GetName() << "]";
cerr << endl;
}
}
