bool Program::build(const Epic::Core::Array<Device>& devices,
const Epic::Core::ASCIIString& options) const
{
Array<cl_device_id> deviceArray;
size_t numDevices = 0;
if(devices.isEmpty()) {
numDevices = devices.count();
deviceArray = deviceIDHelper(devices);
}
cl_int err = clBuildProgram(programHandle(), numDevices, deviceArray.data(), options.data(), nullptr, nullptr);
/* Skip CL_BUILD_PROGRAM_FAILURE error, will be handled later in this code. */
if((err != CL_BUILD_PROGRAM_FAILURE) && (err != CL_SUCCESS)) {
throw OpenCLException(err);
}
Epic::Core::Array<Device> devs;
bool ret = true;
if(devices.isEmpty()) {
devs = this->devices();
} else {
devs = devices;
}
for(size_t i = 0; i < devs.count(); i++) {
cl_build_status status = buildStatus(devs[i]);
cout << "Build log for device " << devs[i].deviceID() << endl;
cout << buildLog(devs[i]) << endl;
while(status == CL_BUILD_IN_PROGRESS) {
status = buildStatus(devs[i]);
}
if(status == CL_BUILD_SUCCESS) {
cout << "Program compilation successful" << endl;
} else {
cout << "Program compilation failed" << endl;
}
ret &= (status == CL_BUILD_SUCCESS);
}
return ret;
}
/*! \brief Checks whether the build process was successfull or not.*/
void ocl::Program::checkBuild(cl_int buildErr) const
{
// Exiting the program is not acceptable.
if ( buildErr == CL_SUCCESS )
return;
std::ostringstream oss;
bool headerSet = false;
for ( auto const& device : _context->devices() )
{
cl_build_status buildStatus = CL_SUCCESS;
clGetProgramBuildInfo( _id, device.id(), CL_PROGRAM_BUILD_STATUS, sizeof buildStatus, &buildStatus, nullptr );
if ( buildStatus != CL_SUCCESS )
{
if ( !headerSet )
{
oss << "Program failed to build.\n";
headerSet = true;
}
size_t size = 0u;
clGetProgramBuildInfo( _id, device.id(), CL_PROGRAM_BUILD_LOG, 0u, nullptr, &size );
std::unique_ptr< cl_char[] > buildLog( new cl_char[size] );
clGetProgramBuildInfo( _id, device.id(), CL_PROGRAM_BUILD_LOG, size, buildLog.get(), nullptr );
oss << "Device " << device.name() << " Build Log:\n" << buildLog.get() << '\n';
}
}
if ( !oss.str().empty() )
throw std::runtime_error( oss.str() );
}
// Compile the kernels that this plan uses, and store into the plan
clfftStatus FFTAction::compileKernels( const cl_command_queue commQueueFFT, const clfftPlanHandle plHandle, FFTPlan* fftPlan )
{
cl_int status = 0;
size_t deviceListSize = 0;
FFTRepo& fftRepo = FFTRepo::getInstance( );
// create a cl program executable for the device associated with command queue
// Get the device
cl_device_id &q_device = fftPlan->bakeDevice;
cl_program program;
if( fftRepo.getclProgram( this->getGenerator(), this->getSignatureData(), program, q_device, fftPlan->context ) == CLFFT_INVALID_PROGRAM )
{
FFTBinaryLookup lookup (this->getGenerator(), plHandle, fftPlan->context, q_device);
lookup.variantRaw(this->getSignatureData(), this->getSignatureData()->datasize);
if (lookup.found())
{
#if FFT_CACHE_DEBUG
// debug message in debug mode to ensure that the cache is used
fprintf(stderr, "Kernel loaded from cache\n");
#endif
program = lookup.getProgram();
}
else
{
#if FFT_CACHE_DEBUG
fprintf(stderr, "Kernel built from source\n");
#endif
// If the user wishes us to write the kernels out to disk, we do so
if( fftRepo.setupData.debugFlags & CLFFT_DUMP_PROGRAMS )
{
OPENCL_V( writeKernel( plHandle, this->getGenerator(), this->getSignatureData(), fftPlan->context, fftPlan->bakeDevice ), _T( "writeKernel failed." ) );
}
std::string programCode;
OPENCL_V( fftRepo.getProgramCode( this->getGenerator(), this->getSignatureData(), programCode, q_device, fftPlan->context ), _T( "fftRepo.getProgramCode failed." ) );
const char* source = programCode.c_str();
program = clCreateProgramWithSource( fftPlan->context, 1, &source, NULL, &status );
OPENCL_V( status, _T( "clCreateProgramWithSource failed." ) );
// create a cl program executable for the device associated with command queue
#if defined(DEBUGGING)
status = clBuildProgram( program, 1, &q_device, "-g -cl-opt-disable", NULL, NULL); // good for debugging kernels
// if you have trouble creating smbols that GDB can pick up to set a breakpoint after kernels are loaded into memory
// this can be used to stop execution to allow you to set a breakpoint in a kernel after kernel symbols are in memory.
#ifdef DEBUG_BREAK_GDB
__debugbreak();
#endif
#else
status = clBuildProgram( program, 1, &q_device, "", NULL, NULL);
#endif
if( status != CL_SUCCESS )
{
if( status == CL_BUILD_PROGRAM_FAILURE )
{
size_t buildLogSize = 0;
OPENCL_V( clGetProgramBuildInfo( program, q_device, CL_PROGRAM_BUILD_LOG, 0, NULL, &buildLogSize ),
_T( "clGetProgramBuildInfo failed" ) );
std::vector< char > buildLog( buildLogSize );
::memset( &buildLog[ 0 ], 0x0, buildLogSize );
OPENCL_V( clGetProgramBuildInfo( program, q_device, CL_PROGRAM_BUILD_LOG, buildLogSize, &buildLog[ 0 ], NULL ),
_T( "clGetProgramBuildInfo failed" ) );
std::cerr << "\n\t\t\tBUILD LOG\n";
std::cerr << "************************************************\n";
std::cerr << &buildLog[ 0 ] << std::endl;
std::cerr << "************************************************\n";
}
OPENCL_V( status, _T( "clBuildProgram failed" ) );
}
lookup.setProgram(program, source);
lookup.populateCache();
}
fftRepo.setclProgram( this->getGenerator(), this->getSignatureData(), program, q_device, fftPlan->context );
// For real transforms we compile either forward or backward kernel
bool buildFwdKernel = buildForwardKernel();
bool buildBwdKernel = buildBackwardKernel();
// get a kernel object handle for a kernel with the given name
cl_kernel kernel;
if( buildFwdKernel )
{
if( fftRepo.getclKernel( program, CLFFT_FORWARD, kernel ) == CLFFT_INVALID_KERNEL )
{
std::string entryPoint;
OPENCL_V( fftRepo.getProgramEntryPoint( this->getGenerator(), this->getSignatureData(), CLFFT_FORWARD, entryPoint, q_device, fftPlan->context ), _T( "fftRepo.getProgramEntryPoint failed." ) );
kernel = clCreateKernel( program, entryPoint.c_str( ), &status );
OPENCL_V( status, _T( "clCreateKernel failed" ) );
fftRepo.setclKernel( program, CLFFT_FORWARD, kernel );
}
}
if( buildBwdKernel )
{
if( fftRepo.getclKernel( program, CLFFT_BACKWARD, kernel ) == CLFFT_INVALID_KERNEL )
{
std::string entryPoint;
OPENCL_V( fftRepo.getProgramEntryPoint( this->getGenerator(), this->getSignatureData(), CLFFT_BACKWARD, entryPoint, q_device, fftPlan->context ), _T( "fftRepo.getProgramEntryPoint failed." ) );
kernel = clCreateKernel( program, entryPoint.c_str( ), &status );
OPENCL_V( status, _T( "clCreateKernel failed" ) );
fftRepo.setclKernel( program, CLFFT_BACKWARD, kernel );
}
}
}
return CLFFT_SUCCESS;
}
int Parallel::setup() {
/**
* OpenCL initialization.
*/
cl_int status = Simulator::setup();
CheckStatus(status, "Simulator::setup() failed.");
cl_uint numPlatforms;
status = clGetPlatformIDs(0, NULL, &numPlatforms);
CheckStatus(status, "clGetPlatformIDs, fetching number");
DEBUG_STDOUT("Number of platforms: " << numPlatforms);
cl_platform_id platform = NULL;
if (numPlatforms > 0) {
std::unique_ptr<cl_platform_id[]> platforms (new cl_platform_id[numPlatforms]);
status = clGetPlatformIDs(numPlatforms, platforms.get(), NULL);
CheckStatus(status, "clGetPlatformIDs, fetching platforms");
for (unsigned i = 0; i < numPlatforms; ++i) {
char pbuf[100];
status = clGetPlatformInfo(platforms[i],
CL_PLATFORM_VENDOR,
sizeof(pbuf),
pbuf,
NULL);
CheckStatus(status, "clGetPlatformInfo");
}
// Just grab the first platform.
platform = platforms[0];
}
CheckConditional(platform != NULL, "platform == NULL");
cl_uint numDevices;
status = clGetDeviceIDs(platform, CL_DEVICE_TYPE_ALL, 0, NULL, &numDevices);
CheckStatus(status, "clGetDeviceIDs: fetching number");
DEBUG_STDOUT("Number of devices: " << numDevices);
cl_device_id *devices = new cl_device_id[numDevices];
status = clGetDeviceIDs(platform, CL_DEVICE_TYPE_ALL, numDevices, devices, NULL);
CheckStatus(status, "clGetDeviceIDs: fetching devices");
int deviceIndex = 0;
for (unsigned i = 0; i < numDevices; ++i) {
char pbuf[100];
status = clGetDeviceInfo(devices[i],
CL_DEVICE_NAME,
sizeof(pbuf),
pbuf,
NULL);
if (!strncmp(pbuf, "ATI", 3)) {
deviceIndex = i;
}
}
/* Create the context. */
context = clCreateContext(0, numDevices, devices, NULL, NULL, &status);
CheckConditional(context != NULL, "clCreateContextFromType");
/* Create command queue */
cl_command_queue_properties prop = CL_QUEUE_PROFILING_ENABLE;
commandQueue = clCreateCommandQueue(context, devices[deviceIndex], prop, &status);
CheckStatus(status, "clCreateCommandQueue");
/* Create a CL program using the kernel source */
SDKFile kernelFile;
std::string kernelPath = getenv("HOME") + std::string("/md-simulator/src/TestKernel.cl");
if(!kernelFile.open(kernelPath.c_str())) {
DEBUG_STDERR("Failed to load kernel file : " << kernelPath);
return MD_FAILURE;
}
const char *source = kernelFile.source().c_str();
size_t sourceSize[] = {strlen(source)};
program = clCreateProgramWithSource(context,
1,
&source,
sourceSize,
&status);
CheckStatus(status, "clCreateProgramWithSource");
/* Create a cl program executable for all the devices specified */
status = clBuildProgram(program,
numDevices,
devices,
NULL,
NULL,
NULL);
if (status != CL_SUCCESS) {
if (status == CL_BUILD_PROGRAM_FAILURE) {
cl_int logStatus;
std::unique_ptr<char[]> buildLog (nullptr);
//char *buildLog = NULL;
size_t buildLogSize = 0;
logStatus = clGetProgramBuildInfo(program,
devices[deviceIndex],
CL_PROGRAM_BUILD_LOG,
buildLogSize,
buildLog.get(),
&buildLogSize);
CheckStatus(logStatus, "clGetProgramBuildInfo");
buildLog = std::unique_ptr<char[]>(new char[buildLogSize]);
if(!buildLog) {
return MD_FAILURE;
}
std::fill_n(buildLog.get(), buildLogSize, 0);
logStatus = clGetProgramBuildInfo(program,
devices[deviceIndex],
CL_PROGRAM_BUILD_LOG,
buildLogSize,
buildLog.get(),
NULL);
CheckStatus(logStatus, "clGetProgramBuildInfo (2)");
DEBUG_STDERR("\n\t\t\tBUILD LOG\n");
DEBUG_STDERR("************************************************\n");
DEBUG_STDERR(buildLog.get());
DEBUG_STDERR("************************************************\n");
}
}
CheckStatus(status, "clBuildProgram");
/* Get a kernel object handle for a kernel with the given name */
kernel = clCreateKernel(program, "computeAccelerations", &status);
CheckStatus(status, "clCreateKernel");
/* Check group size against group size returned by kernel */
status = clGetKernelWorkGroupInfo(kernel,
devices[deviceIndex],
CL_KERNEL_WORK_GROUP_SIZE,
sizeof(size_t),
&kernelWorkGroupSize,
0);
CheckStatus(status, "clGetKernelWorkGroupInfo");
DEBUG_STDOUT("kernelWorkGroupSize: " << kernelWorkGroupSize);
/**
* Initialize some simulator data structures.
*/
global = particleCount * particleCount;
local = particleCount;
if (global * local > kernelWorkGroupSize) {
DEBUG_STDERR("WARNING - global * local > kernelWorkGroupSize; global: " << global << ", local: " << local << ", kernelWorkGroupSize: " << kernelWorkGroupSize);
return MD_FAILURE;
}
// Data holds the molecule positions.
data = std::unique_ptr<float[]> (new float[particleCount * 3]);
// Constants holds simulator constants.
constants = std::unique_ptr<float[]> (new float[NUM_CONSTANTS]);
// Copy constants to buffer;
constants[0] = epsilon;
constants[1] = sigma;
constants[2] = negForceCutoffMinusHalf;
constants[3] = forceCutoffMinusHalf;
constants[4] = wallStiffness;
// Results holds pairwise forces.
results = std::unique_ptr<float[]> (new float[particleCount * particleCount * 3]);
return MD_SUCCESS;
}
