void simulationStep() {
try {
// copy
auto buffer = cl::Buffer(context, CL_MEM_READ_ONLY,
sizeof(unsigned char) * 4 * fieldWidth * fieldHeight,
nullptr, nullptr);
queue.enqueueWriteBuffer(buffer, CL_TRUE, 0,
sizeof(unsigned char) * 4 * fieldWidth * fieldHeight,
visualizationBufferCPU, NULL, NULL);
// enque
stepKernel.setArg(2, buffer);
cl::NDRange global((size_t) (fieldWidth * fieldHeight));
queue.enqueueNDRangeKernel(stepKernel, cl::NullRange, global, cl::NullRange);
// read back
queue.enqueueReadBuffer(visualizationBufferGPU, CL_TRUE, 0,
sizeof(unsigned char) * 4 * fieldWidth * fieldHeight,
visualizationBufferCPU, NULL, NULL);
// finish
queue.finish();
} catch (cl::Error err) {
std::cout << "Error: " << err.what() << "(" << err.err() << ")" << std::endl;
exit(3);
}
glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA, fieldWidth, fieldHeight, 0, GL_RGBA, GL_UNSIGNED_BYTE,
visualizationBufferCPU);
}
void PathOCLRenderThread::EnqueueAdvancePathsKernel(cl::CommandQueue &oclQueue) {
PathOCLRenderEngine *engine = (PathOCLRenderEngine *)renderEngine;
const u_int taskCount = engine->taskCount;
// Micro kernels version
oclQueue.enqueueNDRangeKernel(*advancePathsKernel_MK_RT_NEXT_VERTEX, cl::NullRange,
cl::NDRange(taskCount), cl::NDRange(advancePathsWorkGroupSize));
oclQueue.enqueueNDRangeKernel(*advancePathsKernel_MK_HIT_NOTHING, cl::NullRange,
cl::NDRange(taskCount), cl::NDRange(advancePathsWorkGroupSize));
oclQueue.enqueueNDRangeKernel(*advancePathsKernel_MK_HIT_OBJECT, cl::NullRange,
cl::NDRange(taskCount), cl::NDRange(advancePathsWorkGroupSize));
oclQueue.enqueueNDRangeKernel(*advancePathsKernel_MK_RT_DL, cl::NullRange,
cl::NDRange(taskCount), cl::NDRange(advancePathsWorkGroupSize));
oclQueue.enqueueNDRangeKernel(*advancePathsKernel_MK_DL_ILLUMINATE, cl::NullRange,
cl::NDRange(taskCount), cl::NDRange(advancePathsWorkGroupSize));
oclQueue.enqueueNDRangeKernel(*advancePathsKernel_MK_DL_SAMPLE_BSDF, cl::NullRange,
cl::NDRange(taskCount), cl::NDRange(advancePathsWorkGroupSize));
oclQueue.enqueueNDRangeKernel(*advancePathsKernel_MK_GENERATE_NEXT_VERTEX_RAY, cl::NullRange,
cl::NDRange(taskCount), cl::NDRange(advancePathsWorkGroupSize));
oclQueue.enqueueNDRangeKernel(*advancePathsKernel_MK_SPLAT_SAMPLE, cl::NullRange,
cl::NDRange(taskCount), cl::NDRange(advancePathsWorkGroupSize));
oclQueue.enqueueNDRangeKernel(*advancePathsKernel_MK_NEXT_SAMPLE, cl::NullRange,
cl::NDRange(taskCount), cl::NDRange(advancePathsWorkGroupSize));
oclQueue.enqueueNDRangeKernel(*advancePathsKernel_MK_GENERATE_CAMERA_RAY, cl::NullRange,
cl::NDRange(taskCount), cl::NDRange(advancePathsWorkGroupSize));
}
void updateParticles(float timeDelta)
{
try
{
vector<cl::Memory> glBuffers;
glBuffers.push_back(m_positions);
glBuffers.push_back(m_colors);
//this will update our system by calculating new velocity and updating the positions of our particles
//Make sure OpenGL is done using our VBOs
glFinish();
// map OpenGL buffer object for writing from OpenCL
// this passes in the vector of VBO buffer objects (position and color)
m_queue.enqueueAcquireGLObjects(&glBuffers);
m_particleKernel.setArg(5, timeDelta); //pass in the timestep
//execute the kernel
m_queue.enqueueNDRangeKernel(m_particleKernel, cl::NullRange, cl::NDRange(m_numParticles),
cl::NullRange);
//Release the VBOs so OpenGL can play with them
m_queue.enqueueReleaseGLObjects(&glBuffers, NULL);
m_queue.finish();
}
catch(cl::Error &error)
{
LOG_ERROR << error.what() << "(" << oclErrorString(error.err()) << ")";
}
}
void procOCL_OCV(int tex, int w, int h)
{
int64_t t = getTimeMs();
cl::ImageGL imgIn (theContext, CL_MEM_READ_ONLY, GL_TEXTURE_2D, 0, tex);
std::vector < cl::Memory > images(1, imgIn);
theQueue.enqueueAcquireGLObjects(&images);
theQueue.finish();
cv::UMat uIn, uOut, uTmp;
cv::ocl::convertFromImage(imgIn(), uIn);
LOGD("loading texture data to OpenCV UMat costs %d ms", getTimeInterval(t));
theQueue.enqueueReleaseGLObjects(&images);
t = getTimeMs();
//cv::blur(uIn, uOut, cv::Size(5, 5));
cv::Laplacian(uIn, uTmp, CV_8U);
cv:multiply(uTmp, 10, uOut);
cv::ocl::finish();
LOGD("OpenCV processing costs %d ms", getTimeInterval(t));
t = getTimeMs();
cl::ImageGL imgOut(theContext, CL_MEM_WRITE_ONLY, GL_TEXTURE_2D, 0, tex);
images.clear();
images.push_back(imgOut);
theQueue.enqueueAcquireGLObjects(&images);
cl_mem clBuffer = (cl_mem)uOut.handle(cv::ACCESS_READ);
cl_command_queue q = (cl_command_queue)cv::ocl::Queue::getDefault().ptr();
size_t offset = 0;
size_t origin[3] = { 0, 0, 0 };
size_t region[3] = { w, h, 1 };
CV_Assert(clEnqueueCopyBufferToImage (q, clBuffer, imgOut(), offset, origin, region, 0, NULL, NULL) == CL_SUCCESS);
theQueue.enqueueReleaseGLObjects(&images);
cv::ocl::finish();
LOGD("uploading results to texture costs %d ms", getTimeInterval(t));
}
void MetaBallsApp::update()
{
std::vector<cl::Memory> acquire( { mClParticleBuf, mClMarchingRenderBuffer, mClMarchingDebugBuffer } );
mClCommandQueue.enqueueAcquireGLObjects( &acquire );
updateParticles();
updateMarching();
mClCommandQueue.enqueueReleaseGLObjects( &acquire );
}
void PTWeekend::draw()
{
/*
* BEGIN - each frame part
*/
/* Enqueue kernel for execution */
glm::vec3 origin,lower_left, hor, ver;
float theta = camera.getFov() * M_PI / 180.0f;
float half_height = tan(theta / 2.0f);
float half_width = camera.getAspectRatio() * half_height;
origin = camera.getEyePoint();
glm::vec3 u, v, w;
w = -glm::normalize(camera.getViewDirection()); //odd...
u = glm::normalize(glm::cross(glm::vec3(0,1,0), w));
v = glm::cross(w, u);
lower_left = origin - half_width * u - half_height * v - w;
hor = 2.0f * half_width * u;
ver = 2.0f * half_height * v;
pt_assert(cl_set_pinhole_cam_arg(origin, lower_left, hor, ver, cam_buffer, cmd_queue), "Could not fill camera buffer");
clStatus = cmd_queue.enqueueAcquireGLObjects(&img_buffer, NULL, NULL);
pt_assert(clStatus, "Could not acquire gl objects");
cl::Event profiling_evt;
clStatus = cmd_queue.enqueueNDRangeKernel(kernel,
cl::NDRange(0,0),
cl::NDRange(img_width, img_height),
cl::NDRange(local_width,local_height),
NULL,
&profiling_evt);
profiling_evt.wait();
pt_assert(clStatus, "Could not enqueue the kernel");
clStatus = cmd_queue.enqueueReleaseGLObjects(&img_buffer, NULL, NULL);
pt_assert(clStatus, "Could not release gl objects");
cmd_queue.finish();
cl_ulong time_start = profiling_evt.getProfilingInfo<CL_PROFILING_COMMAND_START>();
cl_ulong time_end = profiling_evt.getProfilingInfo<CL_PROFILING_COMMAND_END>();
cl_ulong total_time = time_end - time_start;
std::cout << "Total time: " << total_time * 0.001 * 0.001 << " ms \n";
/*
* END - each frame part
*/
gl::draw(imgTex, Rectf(0, 0, getWindowWidth(), getWindowHeight()));
}
int clPeak::runKernelLatency(cl::CommandQueue &queue, cl::Program &prog, device_info_t &devInfo)
{
if(!isKernelLatency)
return 0;
cl::Context ctx = queue.getInfo<CL_QUEUE_CONTEXT>();
cl_uint numItems = (devInfo.maxWGSize) * (devInfo.numCUs) * FETCH_PER_WI;
cl::NDRange globalSize = (numItems / FETCH_PER_WI);
cl::NDRange localSize = devInfo.maxWGSize;
int iters = devInfo.kernelLatencyIters;
float latency;
try
{
log->print(NEWLINE TAB TAB "Kernel launch latency : ");
log->xmlOpenTag("kernel_launch_latency");
log->xmlAppendAttribs("unit", "us");
cl::Buffer inputBuf = cl::Buffer(ctx, CL_MEM_READ_ONLY, (numItems * sizeof(float)));
cl::Buffer outputBuf = cl::Buffer(ctx, CL_MEM_WRITE_ONLY, (numItems * sizeof(float)));
cl::Kernel kernel_v1(prog, "global_bandwidth_v1_local_offset");
kernel_v1.setArg(0, inputBuf), kernel_v1.setArg(1, outputBuf);
// Dummy calls
queue.enqueueNDRangeKernel(kernel_v1, cl::NullRange, globalSize, localSize);
queue.enqueueNDRangeKernel(kernel_v1, cl::NullRange, globalSize, localSize);
queue.finish();
latency = 0;
for(int i=0; i<iters; i++)
{
cl::Event timeEvent;
queue.enqueueNDRangeKernel(kernel_v1, cl::NullRange, globalSize, localSize, NULL, &timeEvent);
queue.finish();
cl_ulong start = timeEvent.getProfilingInfo<CL_PROFILING_COMMAND_QUEUED>() / 1000;
cl_ulong end = timeEvent.getProfilingInfo<CL_PROFILING_COMMAND_START>() / 1000;
latency += (float)((int)end - (int)start);
}
latency /= iters;
log->print(latency); log->print(" us" NEWLINE);
log->xmlSetContent(latency);
log->xmlCloseTag();
}
catch(cl::Error error)
{
log->print(error.err() + NEWLINE);
log->print(TAB TAB "Tests skipped" NEWLINE);
return -1;
}
return 0;
}
cl::Event RuntimeMeasurementsManager::enqueueNewMarker(cl::CommandQueue queue) {
cl::Event event;
#if !defined(CL_VERSION_1_2) || defined(CL_USE_DEPRECATED_OPENCL_1_1_APIS)
// Use deprecated API
queue.enqueueMarker(&event);
#else
queue.enqueueMarkerWithWaitList(NULL, &event)
#endif
queue.finish();
return event;
}
void initParticles(uint32_t num_particles)
{
m_geom = gl::Geometry::create();
m_geom->setPrimitiveType(GL_POINTS);
m_mesh = gl::Mesh::create(m_geom, m_pointMaterial);
m_numParticles = num_particles;
GLsizei numBytes = m_numParticles * sizeof(vec4);
m_geom->vertices().resize(m_numParticles, vec3(0));
m_geom->colors().resize(m_numParticles, vec4(1));
m_geom->point_sizes().resize(m_numParticles, 9.f);
m_geom->createGLBuffers();
m_mesh->material()->setPointSize(2.f);
scene().addObject(m_mesh);
try
{
// shared position buffer for OpenGL / OpenCL
m_positions = cl::BufferGL(m_context, CL_MEM_READ_WRITE, m_geom->vertexBuffer().id());
m_colors = cl::BufferGL(m_context, CL_MEM_READ_WRITE, m_geom->colorBuffer().id());
//create the OpenCL only arrays
m_velocities = cl::Buffer( m_context, CL_MEM_WRITE_ONLY, numBytes );
m_positionGen = cl::Buffer( m_context, CL_MEM_WRITE_ONLY, numBytes );
m_velocityGen = cl::Buffer( m_context, CL_MEM_WRITE_ONLY, numBytes );
vector<vec4> posGen, velGen;
for (int i = 0; i < m_numParticles; i++)
{
posGen.push_back( vec4(glm::ballRand(20.0f), 1.f) );
vec2 tmp = glm::linearRand(vec2(-100), vec2(100));
float life = kinski::random(2.f, 5.f);
float yVel = kinski::random<float>(5, 15);
velGen.push_back(vec4(tmp.x, yVel, tmp.y, life));
m_geom->point_sizes()[i] = kinski::random(5.f, 15.f);
}
m_geom->createGLBuffers();
m_queue.enqueueWriteBuffer(m_velocities, CL_TRUE, 0, numBytes, &velGen[0]);
m_queue.enqueueWriteBuffer(m_positionGen, CL_TRUE, 0, numBytes, &posGen[0]);
m_queue.enqueueWriteBuffer(m_velocityGen, CL_TRUE, 0, numBytes, &velGen[0]);
m_particleKernel.setArg(0, m_positions);
m_particleKernel.setArg(1, m_colors);
m_particleKernel.setArg(2, m_velocities);
m_particleKernel.setArg(3, m_positionGen);
m_particleKernel.setArg(4, m_velocityGen);
}
catch(cl::Error &error)
{
LOG_ERROR << error.what() << "(" << oclErrorString(error.err()) << ")";
}
}
void helper(uint32_t* out, int osize, uint8_t* in, int isize, int w, int h, int choice)
{
int set_size=8;
try {
cl::Buffer bufferIn = cl::Buffer(gContext, CL_MEM_READ_WRITE | CL_MEM_COPY_HOST_PTR,
isize*sizeof(cl_uchar), in, NULL);
cl::Buffer bufferOut = cl::Buffer(gContext, CL_MEM_READ_WRITE, osize*sizeof(cl_uchar4));
cl::Buffer bufferOut2= cl::Buffer(gContext, CL_MEM_READ_WRITE, osize*sizeof(cl_uchar4));
gNV21Kernel.setArg(2,w);
gNV21Kernel.setArg(3,h);
gNV21Kernel.setArg(1,bufferIn);
gNV21Kernel.setArg(0,bufferOut);
gQueue.enqueueNDRangeKernel(gNV21Kernel,
cl::NullRange,
cl::NDRange( (int)ceil((float)w/16.0f)*16,(int)ceil((float)h/16.0f)*16),
cl::NDRange(set_size,set_size),
NULL,
NULL);
if (choice==1) {
gLaplacianK.setArg(2,w);
gLaplacianK.setArg(3,h);
gLaplacianK.setArg(1,bufferOut);
gLaplacianK.setArg(0,bufferOut2);
gQueue.enqueueNDRangeKernel(gLaplacianK,
cl::NullRange,
cl::NDRange( (int)ceil((float)w/16.0f)*16,(int)ceil((float)h/16.0f)*16),
cl::NDRange(set_size,set_size),
NULL,
NULL);
}
else if (choice>1) {
gNegative.setArg(2,w);
gNegative.setArg(3,h);
gNegative.setArg(1,bufferOut);
gNegative.setArg(0,bufferOut2);
gQueue.enqueueNDRangeKernel(gNegative,
cl::NullRange,
cl::NDRange( (int)ceil((float)w/16.0f)*16,(int)ceil((float)h/16.0f)*16),
cl::NDRange(set_size,set_size),
NULL,
NULL);
}
gQueue.enqueueReadBuffer(bufferOut2, CL_TRUE, 0, osize*sizeof(cl_uchar4), out);
}
catch (cl::Error e) {
LOGI("@oclDecoder: %s %d \n",e.what(),e.err());
}
}
/**
* generate 64 bit unsigned random numbers in device global memory
*@param tinymt_status device global memories
*@param total_num total number of work items
*@param local_num number of local work items
*@param data_size number of data to generate
*/
static void generate_uint64(Buffer& tinymt_status,
int total_num,
int local_num,
int data_size)
{
#if defined(DEBUG)
cout << "generate_uint64 start" << endl;
#endif
int min_size = total_num;
if (data_size % min_size != 0) {
data_size = (data_size / min_size + 1) * min_size;
}
Kernel uint_kernel(program, "tinymt_uint64_kernel");
Buffer output_buffer(context,
CL_MEM_READ_WRITE,
data_size * sizeof(uint64_t));
uint_kernel.setArg(0, tinymt_status);
uint_kernel.setArg(1, output_buffer);
uint_kernel.setArg(2, data_size / total_num);
NDRange global(total_num);
NDRange local(local_num);
Event generate_event;
#if defined(DEBUG)
cout << "generate_uint64 enque kernel start" << endl;
#endif
queue.enqueueNDRangeKernel(uint_kernel,
NullRange,
global,
local,
NULL,
&generate_event);
uint64_t * output = new uint64_t[data_size];
generate_event.wait();
queue.enqueueReadBuffer(output_buffer,
CL_TRUE,
0,
data_size * sizeof(uint64_t),
output);
check_data(output, data_size, total_num);
#if defined(DEBUG)
print_uint64(output, data_size, total_num);
#endif
double time = get_time(generate_event);
cout << "generate time:" << time * 1000 << "ms" << endl;
delete[] output;
#if defined(DEBUG)
cout << "generate_uint64 end" << endl;
#endif
}
void copyFromDevice(cl::CommandQueue &queue)
{
if(m_pElts==NULL)
throw cl::Error(CL_INVALID_MEM_OBJECT, "copyFromDevice - Buffer is not initialised.");
queue.enqueueReadBuffer(m_buffer, CL_TRUE, 0, m_cb, m_pElts);
}
void kernel(cl::Buffer& devOut, cl::CommandQueue& queue)
{
static std::once_flag compileFlag;
static cl::Program prog;
static cl::Kernel kern;
std::call_once(compileFlag,
[queue]() {
prog = cl::Program(queue.getInfo<CL_QUEUE_CONTEXT>(), fractal_ocl_kernel, true);
kern = cl::Kernel(prog, "julia");
});
//auto juliaOp = cl::make_kernel<Buffer, unsigned, unsigned>(kern);
static const NDRange local(8, 8);
NDRange global(local[0] * divup(DIMX, local[0]),
local[1] * divup(DIMY, local[1]));
kern.setArg(0, devOut);
kern.setArg(1, DIMX);
kern.setArg(2, DIMY);
queue.enqueueNDRangeKernel(kern, cl::NullRange, global, local);
//juliaOp(EnqueueArgs(queue, global, local), devOut, DIMX, DIMY);
}
bool runTestType(cl::Context context, cl::CommandQueue queue)
{
cl_uint size = 1024 * 2 + 15;
std::vector<T> input(size);
std::cout << "##Testing scan for " << input.size() << " elements and type "
<< magnet::CL::detail::traits<T>::kernel_type();
for(size_t i = 0; i < input.size(); ++i)
input[i] = i+1;
// create input buffer using pinned memory
cl::Buffer bufferIn(context, CL_MEM_ALLOC_HOST_PTR |
CL_MEM_COPY_HOST_PTR | CL_MEM_READ_WRITE,
sizeof(T) * input.size(), &input[0])
;
magnet::CL::scan<T> scanFunctor;
scanFunctor.build(queue, context);
scanFunctor(bufferIn, bufferIn);
std::vector<T> output(size);
queue.enqueueReadBuffer(bufferIn, CL_TRUE, 0, input.size() *
sizeof(T), &output[0]);
bool failed = !testOutput(input, output);
std::cout << (failed ? " FAILED" : " PASSED") << std::endl;
return failed;
}
cl::Event runKernel(const cl::CommandQueue& queue, const cl::Kernel& kernel, const cl::NDRange& globalSize, const cl::NDRange& groupSize, std::vector<cl::Event>& events)
{
cl::Event event;
queue.enqueueNDRangeKernel(kernel, cl::NullRange, globalSize, groupSize, &events, &event);
events.push_back(event);
return event;
}
void runTestType(cl::Context context, cl::CommandQueue queue)
{
cl_uint size = 2 << 10;
std::vector<T> input(size);
std::cout << "##Testing bitonic sort for " << input.size() << " elements and type "
<< magnet::CL::detail::traits<T>::kernel_type()
<< std::endl;
for(size_t i = 0; i < input.size(); ++i)
input[i] = input.size() - i - 1;
// create input buffer using pinned memory
cl::Buffer bufferIn(context, CL_MEM_ALLOC_HOST_PTR |
CL_MEM_COPY_HOST_PTR | CL_MEM_READ_WRITE,
sizeof(T) * input.size(), &input[0])
;
magnet::CL::bitonicSort<T> bitonicSortFunctor;
bitonicSortFunctor.build(queue, context);
bitonicSortFunctor(bufferIn);
std::vector<T> output(size);
queue.enqueueReadBuffer(bufferIn, CL_TRUE, 0, input.size() *
sizeof(T), &output[0]);
if (!testOutput(input, output))
M_throw() << "Incorrect output for size "
<< input.size()
<< " and type "
<< magnet::CL::detail::traits<T>::kernel_type();
}
inline void OpenCL::addkernelarg(std::size_t i, std::vector<T> const & arg, cl::Kernel & kernel,cl::CommandQueue &quene) const
{
cl::Buffer buffer(this->context,CL_MEM_READ_WRITE,arg.size()*sizeof(T));
// std::cout << "enqeue\n";
quene.enqueueWriteBuffer(buffer,CL_FALSE,0,sizeof(T)*arg.size(),&(arg[0]));
kernel.setArg(i,buffer);
}
inline void OpenCL::addkernelarg(std::size_t i, T const (& arg)[N], cl::Kernel & kernel,cl::CommandQueue &quene) const
{
cl::Buffer buffer(this->context,CL_MEM_READ_WRITE,N*sizeof(T));
// std::cout << "enqeue\n";
quene.enqueueWriteBuffer(buffer,CL_FALSE,0,sizeof(T)*N,&arg);
kernel.setArg(i,buffer);
}
CL::Event OGLSharedFramebuffer::release(CL::CommandQueue& queue, const CL::Event& evt)
{
if (_shared) {
CL::Event e = queue.enq_GL_release(_cl_buffer->get(),
"release framebuffer", evt);
return e;
} else {
assert(_local);
CL::Event e = queue.enq_read_buffer(*_cl_buffer, _local, _tex_buffer.get_size(),
"read framebuffer", evt);
queue.wait_for_events(e);
_tex_buffer.load(_local);
return CL::Event();
}
}
void procOCL_I2I(int texIn, int texOut, int w, int h)
{
if(!haveOpenCL) return;
LOGD("procOCL_I2I(%d, %d, %d, %d)", texIn, texOut, w, h);
cl::ImageGL imgIn (theContext, CL_MEM_READ_ONLY, GL_TEXTURE_2D, 0, texIn);
cl::ImageGL imgOut(theContext, CL_MEM_WRITE_ONLY, GL_TEXTURE_2D, 0, texOut);
std::vector < cl::Memory > images;
images.push_back(imgIn);
images.push_back(imgOut);
int64_t t = getTimeMs();
theQueue.enqueueAcquireGLObjects(&images);
theQueue.finish();
LOGD("enqueueAcquireGLObjects() costs %d ms", getTimeInterval(t));
t = getTimeMs();
cl::Kernel Laplacian(theProgI2I, "Laplacian"); //TODO: may be done once
Laplacian.setArg(0, imgIn);
Laplacian.setArg(1, imgOut);
theQueue.finish();
LOGD("Kernel() costs %d ms", getTimeInterval(t));
t = getTimeMs();
theQueue.enqueueNDRangeKernel(Laplacian, cl::NullRange, cl::NDRange(w, h), cl::NullRange);
theQueue.finish();
LOGD("enqueueNDRangeKernel() costs %d ms", getTimeInterval(t));
t = getTimeMs();
theQueue.enqueueReleaseGLObjects(&images);
theQueue.finish();
LOGD("enqueueReleaseGLObjects() costs %d ms", getTimeInterval(t));
}
void read(const cl::CommandQueue &q, size_t offset, size_t size, T *host,
bool blocking = false) const
{
if (size)
q.enqueueReadBuffer(
buffer, blocking ? CL_TRUE : CL_FALSE,
sizeof(T) * offset, sizeof(T) * size, host
);
}
real L2Norm(const Buffer3D & in,cl::CommandQueue & q)
{
cl::Buffer ans (CLContextLoader::getContext(),CL_MEM_READ_WRITE,sizeof(real)*in.width()*in.height()*in.depth());
CLContextLoader::getRedL2NormKer().setArg(0,in());
CLContextLoader::getRedL2NormKer().setArg(1,ans());
q.enqueueNDRangeKernel(CLContextLoader::getRedL2NormKer(),
cl::NDRange(0),
cl::NDRange(in.width()*in.height()*in.depth()),
getBestWorkspaceDim(cl::NDRange(in.width()*in.height()*in.depth())));
ans = performReduction(ans,CLContextLoader::getRedSumAllKer(),q,in.width()*in.height()*in.depth());
real res;
q.enqueueReadBuffer(ans,true,0,sizeof(real),&res);
return sqrt(res);
}
cl::Event copyToDeviceAsync(cl::CommandQueue &queue)
{
if(m_pElts==NULL)
throw cl::Error(CL_INVALID_MEM_OBJECT, "copyToDevice - Buffer is not initialised.");
cl::Event complete;
queue.enqueueWriteBuffer(m_buffer, CL_FALSE, 0, m_cb, m_pElts, NULL, &complete);
return complete;
}
CL::Event OGLSharedFramebuffer::acquire(CL::CommandQueue& queue, const CL::Event& e)
{
if (_shared) {
return queue.enq_GL_acquire(_cl_buffer->get(),
"acquire framebuffer", e);
} else {
return e;
}
}
/**
* initialize tinymt status in device global memory
* using 1 parameter for 1 generator.
*@param tinymt_status internal state of kernel side tinymt
*@param total total number of work items
*@param local_item number of local work items
*@param seed seed for initialization
*/
static void initialize_by_seed(Buffer& tinymt_status,
int total,
int local_item,
uint32_t seed)
{
#if defined(DEBUG)
cout << "initialize_by_seed start" << endl;
#endif
Kernel init_kernel(program, "tinymt_init_seed_kernel");
init_kernel.setArg(0, tinymt_status);
init_kernel.setArg(1, seed);
NDRange global(total);
NDRange local(local_item);
Event event;
#if defined(DEBUG)
cout << "global:" << dec << total << endl;
cout << "group:" << dec << (total / local_item) << endl;
cout << "local:" << dec << local_item << endl;
#endif
queue.enqueueNDRangeKernel(init_kernel,
NullRange,
global,
local,
NULL,
&event);
double time = get_time(event);
tinymt32j_t status[total];
queue.enqueueReadBuffer(tinymt_status,
CL_TRUE,
0,
sizeof(tinymt32j_t) * total,
status);
cout << "initializing time = " << time * 1000 << "ms" << endl;
#if defined(DEBUG)
cout << "status[0].s0:" << hex << status[0].s0 << endl;
cout << "status[0].s1:" << hex << status[0].s1 << endl;
cout << "status[0].s2:" << hex << status[0].s2 << endl;
cout << "status[0].s3:" << hex << status[0].s3 << endl;
#endif
check_status(status, total);
#if defined(DEBUG)
cout << "initialize_by_seed end" << endl;
#endif
}
void sumTest(cl::Buffer queue_data, cl::Buffer queue_metadata,
cl::Buffer& device_result, int iterations,
ProgramCache& cache,
cl::CommandQueue& queue)
{
cl::Context context = queue.getInfo<CL_QUEUE_CONTEXT>();
std::vector<std::string> sources;
sources.push_back("ParallelQueue");
sources.push_back("ParallelQueueTests");
cl::Program& program = cache.getProgram(sources);
cl::Kernel sum_test_kernel(program, "sum_test");
cl::Device device = queue.getInfo<CL_QUEUE_DEVICE>();
int warp_size = sum_test_kernel
.getWorkGroupInfo<CL_KERNEL_PREFERRED_WORK_GROUP_SIZE_MULTIPLE>(device);
std::cout << "warp size: " << warp_size << std::endl;
int max_group_size = device.getInfo<CL_DEVICE_MAX_WORK_ITEM_SIZES>()[0];
int queue_num_threads = 512;
if(queue_num_threads > max_group_size)
queue_num_threads = max_group_size;
cl::LocalSpaceArg local_queue
= cl::__local(sizeof(int) * queue_num_threads * 2);
cl::LocalSpaceArg reduction_buffer
= cl::__local(sizeof(int) * queue_num_threads);
cl::LocalSpaceArg got_work
= cl::__local(sizeof(int));
cl::LocalSpaceArg prefix_sum_input
= cl::__local(sizeof(int) * queue_num_threads);
cl::LocalSpaceArg prefix_sum_output
= cl::__local(sizeof(int) * queue_num_threads);
sum_test_kernel.setArg(0, queue_data);
sum_test_kernel.setArg(1, queue_metadata);
sum_test_kernel.setArg(2, device_result);
sum_test_kernel.setArg(3, iterations);
sum_test_kernel.setArg(4, local_queue);
sum_test_kernel.setArg(5, reduction_buffer);
sum_test_kernel.setArg(6, got_work);
sum_test_kernel.setArg(7, prefix_sum_input);
sum_test_kernel.setArg(8, prefix_sum_output);
cl::NDRange nullRange;
cl::NDRange global(queue_num_threads, 1);
cl::NDRange local(queue_num_threads, 1);
cl_int status = queue.enqueueNDRangeKernel(sum_test_kernel,
nullRange, global, local);
}
CL::Event Framebuffer::clear(CL::CommandQueue& queue, const CL::Event& e)
{
_clear_kernel.set_arg(0, _cl_buffer->get());
vec4 color = config.clear_color();
_clear_kernel.set_arg(1, vec4(powf(color.x, 2.2),
powf(color.y, 2.2),
powf(color.z, 2.2), 1000));
return queue.enq_kernel(_clear_kernel, _size.x * _size.y, 256,
"clear framebuffer", e);
}
/**
* initialize tinymt status in device global memory
* using 1 parameter for all generators.
*@param tinymt_status device global memories
*@param total total number of work items
*@param local_item number of local work items
*@param seed_array seeds for initialization
*@param seed_size size of seed_array
*/
static void initialize_by_array(Buffer& tinymt_status,
int total,
int local_item,
uint64_t seed_array[],
int seed_size)
{
#if defined(DEBUG)
cout << "initialize_by_array start" << endl;
#endif
Buffer seed_array_buffer(context,
CL_MEM_READ_WRITE,
seed_size * sizeof(uint64_t));
queue.enqueueWriteBuffer(seed_array_buffer,
CL_TRUE,
0,
seed_size * sizeof(uint64_t),
seed_array);
Kernel init_kernel(program, "tinymt_init_array_kernel");
init_kernel.setArg(0, tinymt_status);
init_kernel.setArg(1, seed_array_buffer);
init_kernel.setArg(2, seed_size);
NDRange global(total);
NDRange local(local_item);
Event event;
queue.enqueueNDRangeKernel(init_kernel,
NullRange,
global,
local,
NULL,
&event);
double time = get_time(event);
tinymt64j_t status[total];
queue.enqueueReadBuffer(tinymt_status,
CL_TRUE,
0,
sizeof(tinymt64j_t) * total,
status);
cout << "initializing time = " << time * 1000 << "ms" << endl;
check_status(status, total);
#if defined(DEBUG)
cout << "initialize_by_array end" << endl;
#endif
}
void findMinSeamVert(cl::Context &ctx,
cl::CommandQueue &cmdQueue,
cl::Event &event,
std::vector<cl::Event> &deps,
cl::Buffer &energyMatrix,
cl::Buffer &vertMinEnergy,
cl::Buffer &vertMinIdx,
int width,
int height,
int pitch,
int colsRemoved) {
cl_int errNum;
errNum = findMinSeamVertKernel.setArg(0, energyMatrix);
errNum |= findMinSeamVertKernel.setArg(1, vertMinEnergy);
errNum |= findMinSeamVertKernel.setArg(2, vertMinIdx);
errNum |= findMinSeamVertKernel.setArg(3, cl::__local(256 * sizeof(float)));
errNum |= findMinSeamVertKernel.setArg(4, cl::__local(256 * sizeof(float)));
errNum |= findMinSeamVertKernel.setArg(5, width);
errNum |= findMinSeamVertKernel.setArg(6, height);
errNum |= findMinSeamVertKernel.setArg(7, pitch);
errNum |= findMinSeamVertKernel.setArg(8, colsRemoved);
if (errNum != CL_SUCCESS) {
std::cerr << "Error setting findMinSeamVert arguments." << std::endl;
exit(-1);
}
// This kernel could be written to use more than one work group, but its probably not worth it.
cl::NDRange offset = cl::NDRange(0);
cl::NDRange localWorkSize = cl::NDRange(256);
cl::NDRange globalWorkSize = cl::NDRange(256);
errNum = cmdQueue.enqueueNDRangeKernel(findMinSeamVertKernel,
offset,
globalWorkSize,
localWorkSize,
&deps,
&event);
if (errNum != CL_SUCCESS) {
std::cerr << "Error enqueuing computeSeams kernel for execution." << std::endl;
exit(-1);
}
/** DEBUG **/
// int deviceResultIdx[1];
// float deviceResultEnergy[1];
// mem::read(ctx, cmdQueue, deviceResultIdx, vertMinIdx);
// mem::read(ctx, cmdQueue, deviceResultEnergy, vertMinEnergy);
// std::cout << "deviceResultIdx = " << deviceResultIdx[0] << std::endl;
// std::cout << "deviceResultEnergy = " << deviceResultEnergy[0] << std::endl;
}
void run(T* buf)
{
cl_int err;
cl::Buffer outbuf(
m_context,
CL_MEM_WRITE_ONLY | CL_MEM_USE_HOST_PTR,
N*N*sizeof(T),
buf,
&err);
checkErr(err, "Buffer::Buffer()");
err = m_kernel.setArg(0, outbuf);
checkErr(err, "Kernel::setArg(0)");
err = m_kernel.setArg(1, N);
checkErr(err, "Kernel::setArg(1)");
err = m_kernel.setArg(2, depth);
checkErr(err, "Kernel::setArg(2)");
err = m_kernel.setArg(3, escape2);
checkErr(err, "Kernel::setArg(3)");
cl::Event event;
err = m_cmdq.enqueueNDRangeKernel(
m_kernel,
cl::NullRange,
cl::NDRange(N*N),
cl::NDRange(N, 1),
NULL,
&event);
checkErr(err, "ComamndQueue::enqueueNDRangeKernel()");
event.wait();
err = m_cmdq.enqueueReadBuffer(
outbuf,
CL_TRUE,
0,
N*N*sizeof(T),
buf);
checkErr(err, "ComamndQueue::enqueueReadBuffer()");
}
