SWDEV-240806 - [hip-graph] hip_memory and hip_kernel separate command creation from enqueue
hipMemcpy1D all variants, hipMemcpy3D, hipMemset, hipMemset3D and hipLaunchKernel Change-Id: Ieeb767cf87aa0546bc3b3d5d2e01ab618e0d285c
This commit is contained in:
Anusha GodavarthySurya
کامیت شده توسط
Anusha Godavarthy Surya
Anusha Godavarthy Surya
والد
60e2b1299c
کامیت
d86b4696ca
+449
-411
تفاوت فایلی نمایش داده نمی شود زیرا این فایل بسیار بزرگ است
Diff را بارگزاری کن
+107
-63
@@ -209,134 +209,178 @@ hipError_t hipFuncSetSharedMemConfig ( const void* func, hipSharedMemConfig conf
|
||||
HIP_RETURN(hipSuccess);
|
||||
}
|
||||
|
||||
hipError_t ihipModuleLaunchKernel(hipFunction_t f, uint32_t globalWorkSizeX,
|
||||
uint32_t globalWorkSizeY, uint32_t globalWorkSizeZ,
|
||||
uint32_t blockDimX, uint32_t blockDimY, uint32_t blockDimZ,
|
||||
uint32_t sharedMemBytes, hipStream_t hStream,
|
||||
void **kernelParams, void **extra,
|
||||
hipEvent_t startEvent, hipEvent_t stopEvent, uint32_t flags = 0,
|
||||
uint32_t params = 0, uint32_t gridId = 0, uint32_t numGrids = 0,
|
||||
uint64_t prevGridSum = 0, uint64_t allGridSum = 0, uint32_t firstDevice = 0) {
|
||||
HIP_INIT_API(ihipModuleLaunchKernel, f, globalWorkSizeX, globalWorkSizeY, globalWorkSizeZ,
|
||||
blockDimX, blockDimY, blockDimZ, sharedMemBytes, hStream, kernelParams, extra, startEvent,
|
||||
stopEvent, flags, params);
|
||||
|
||||
HIP_RETURN_ONFAIL(PlatformState::instance().initStatManagedVarDevicePtr(ihipGetDevice()));
|
||||
|
||||
inline hipError_t ihipLaunchKernel_validate(hipFunction_t f, uint32_t globalWorkSizeX,
|
||||
uint32_t globalWorkSizeY, uint32_t globalWorkSizeZ,
|
||||
uint32_t blockDimX, uint32_t blockDimY,
|
||||
uint32_t blockDimZ, uint32_t sharedMemBytes,
|
||||
void** kernelParams, void** extra,
|
||||
uint32_t params = 0) {
|
||||
if (f == nullptr) {
|
||||
LogPrintfError("%s", "Function passed is null");
|
||||
return hipErrorInvalidImage;
|
||||
}
|
||||
if ((kernelParams != nullptr) && (extra != nullptr)) {
|
||||
LogPrintfError(
|
||||
"%s", "Both, kernelParams and extra Params are provided, only one should be provided");
|
||||
LogPrintfError("%s",
|
||||
"Both, kernelParams and extra Params are provided, only one should be provided");
|
||||
return hipErrorInvalidValue;
|
||||
}
|
||||
if (globalWorkSizeX == 0 || globalWorkSizeY == 0 || globalWorkSizeZ == 0 ||
|
||||
blockDimX == 0 || blockDimY == 0 || blockDimZ == 0) {
|
||||
if (globalWorkSizeX == 0 || globalWorkSizeY == 0 || globalWorkSizeZ == 0 || blockDimX == 0 ||
|
||||
blockDimY == 0 || blockDimZ == 0) {
|
||||
return hipErrorInvalidValue;
|
||||
}
|
||||
|
||||
hip::DeviceFunc* function = hip::DeviceFunc::asFunction(f);
|
||||
amd::Kernel* kernel = function->kernel();
|
||||
|
||||
amd::ScopedLock lock(function->dflock_);
|
||||
|
||||
hip::Event* eStart = reinterpret_cast<hip::Event*>(startEvent);
|
||||
hip::Event* eStop = reinterpret_cast<hip::Event*>(stopEvent);
|
||||
amd::HostQueue* queue = hip::getQueue(hStream);
|
||||
const amd::Device& device = queue->vdev()->device();
|
||||
|
||||
if (extra != nullptr) {
|
||||
if (extra[0] != HIP_LAUNCH_PARAM_BUFFER_POINTER || extra[2] != HIP_LAUNCH_PARAM_BUFFER_SIZE ||
|
||||
extra[4] != HIP_LAUNCH_PARAM_END) {
|
||||
return hipErrorNotInitialized;
|
||||
}
|
||||
}
|
||||
const amd::Device* device = hip::getCurrentDevice()->devices()[0];
|
||||
// Make sure dispatch doesn't exceed max workgroup size limit
|
||||
if (blockDimX * blockDimY * blockDimZ > device.info().maxWorkGroupSize_) {
|
||||
if (blockDimX * blockDimY * blockDimZ > device->info().maxWorkGroupSize_) {
|
||||
return hipErrorInvalidConfiguration;
|
||||
}
|
||||
|
||||
hip::DeviceFunc* function = hip::DeviceFunc::asFunction(f);
|
||||
amd::Kernel* kernel = function->kernel();
|
||||
// Make sure the launch params are not larger than if specified launch_bounds
|
||||
// If it exceeds, then print a warning and continue for now
|
||||
if (blockDimX * blockDimY * blockDimZ > kernel->getDeviceKernel(device)->workGroupInfo()->size_) {
|
||||
// If it exceeds, then print a warning and continue for now
|
||||
if (blockDimX * blockDimY * blockDimZ >
|
||||
kernel->getDeviceKernel(*device)->workGroupInfo()->size_) {
|
||||
LogPrintfWarning("%s", "Launch params are larger than launch bounds");
|
||||
}
|
||||
|
||||
|
||||
if (params & amd::NDRangeKernelCommand::CooperativeGroups) {
|
||||
if (!device.info().cooperativeGroups_) {
|
||||
if (!device->info().cooperativeGroups_) {
|
||||
return hipErrorLaunchFailure;
|
||||
}
|
||||
int num_blocks = 0;
|
||||
int max_blocks_per_grid = 0;
|
||||
int best_block_size = 0;
|
||||
int block_size = blockDimX * blockDimY * blockDimZ;
|
||||
hip_impl::ihipOccupancyMaxActiveBlocksPerMultiprocessor(
|
||||
&num_blocks, &max_blocks_per_grid, &best_block_size, device, f, block_size, sharedMemBytes, true);
|
||||
hip_impl::ihipOccupancyMaxActiveBlocksPerMultiprocessor(&num_blocks, &max_blocks_per_grid,
|
||||
&best_block_size, *device, f,
|
||||
block_size, sharedMemBytes, true);
|
||||
if (((globalWorkSizeX * globalWorkSizeY * globalWorkSizeZ) / block_size) >
|
||||
unsigned(max_blocks_per_grid)) {
|
||||
unsigned(max_blocks_per_grid)) {
|
||||
return hipErrorCooperativeLaunchTooLarge;
|
||||
}
|
||||
}
|
||||
if (params & amd::NDRangeKernelCommand::CooperativeMultiDeviceGroups) {
|
||||
if (!device.info().cooperativeMultiDeviceGroups_) {
|
||||
if (!device->info().cooperativeMultiDeviceGroups_) {
|
||||
return hipErrorLaunchFailure;
|
||||
}
|
||||
}
|
||||
if (!queue) {
|
||||
return hipErrorOutOfMemory;
|
||||
}
|
||||
|
||||
size_t globalWorkOffset[3] = {0};
|
||||
size_t globalWorkSize[3] = { globalWorkSizeX, globalWorkSizeY, globalWorkSizeZ };
|
||||
size_t localWorkSize[3] = { blockDimX, blockDimY, blockDimZ };
|
||||
amd::NDRangeContainer ndrange(3, globalWorkOffset, globalWorkSize, localWorkSize);
|
||||
amd::Command::EventWaitList waitList;
|
||||
bool profileNDRange = false;
|
||||
address kernargs = nullptr;
|
||||
|
||||
// 'extra' is a struct that contains the following info: {
|
||||
// HIP_LAUNCH_PARAM_BUFFER_POINTER, kernargs,
|
||||
// HIP_LAUNCH_PARAM_BUFFER_SIZE, &kernargs_size,
|
||||
// HIP_LAUNCH_PARAM_END }
|
||||
if (extra != nullptr) {
|
||||
if (extra[0] != HIP_LAUNCH_PARAM_BUFFER_POINTER ||
|
||||
extra[2] != HIP_LAUNCH_PARAM_BUFFER_SIZE || extra[4] != HIP_LAUNCH_PARAM_END) {
|
||||
if (extra[0] != HIP_LAUNCH_PARAM_BUFFER_POINTER || extra[2] != HIP_LAUNCH_PARAM_BUFFER_SIZE ||
|
||||
extra[4] != HIP_LAUNCH_PARAM_END) {
|
||||
return hipErrorNotInitialized;
|
||||
}
|
||||
kernargs = reinterpret_cast<address>(extra[1]);
|
||||
}
|
||||
|
||||
const amd::KernelSignature& signature = kernel->signature();
|
||||
for (size_t i = 0; i < signature.numParameters(); ++i) {
|
||||
const amd::KernelParameterDescriptor& desc = signature.at(i);
|
||||
const amd::KernelSignature& signature = kernel->signature();
|
||||
for (size_t i = 0; i < signature.numParameters(); ++i) {
|
||||
const amd::KernelParameterDescriptor& desc = signature.at(i);
|
||||
if (kernelParams == nullptr) {
|
||||
assert(kernargs != nullptr);
|
||||
kernel->parameters().set(i, desc.size_, kernargs + desc.offset_,
|
||||
desc.type_ == T_POINTER/*svmBound*/);
|
||||
desc.type_ == T_POINTER /*svmBound*/);
|
||||
} else {
|
||||
assert(extra == nullptr);
|
||||
kernel->parameters().set(i, desc.size_, kernelParams[i],
|
||||
desc.type_ == T_POINTER/*svmBound*/);
|
||||
desc.type_ == T_POINTER /*svmBound*/);
|
||||
}
|
||||
}
|
||||
return hipSuccess;
|
||||
}
|
||||
|
||||
hipError_t ihipLaunchKernelCommand(amd::Command*& command, hipFunction_t f,
|
||||
uint32_t globalWorkSizeX, uint32_t globalWorkSizeY,
|
||||
uint32_t globalWorkSizeZ, uint32_t blockDimX, uint32_t blockDimY,
|
||||
uint32_t blockDimZ, uint32_t sharedMemBytes,
|
||||
amd::HostQueue* queue, void** kernelParams, void** extra,
|
||||
hipEvent_t startEvent = nullptr, hipEvent_t stopEvent = nullptr,
|
||||
uint32_t flags = 0, uint32_t params = 0, uint32_t gridId = 0,
|
||||
uint32_t numGrids = 0, uint64_t prevGridSum = 0,
|
||||
uint64_t allGridSum = 0, uint32_t firstDevice = 0) {
|
||||
hip::DeviceFunc* function = hip::DeviceFunc::asFunction(f);
|
||||
amd::Kernel* kernel = function->kernel();
|
||||
|
||||
size_t globalWorkOffset[3] = {0};
|
||||
size_t globalWorkSize[3] = {globalWorkSizeX, globalWorkSizeY, globalWorkSizeZ};
|
||||
size_t localWorkSize[3] = {blockDimX, blockDimY, blockDimZ};
|
||||
amd::NDRangeContainer ndrange(3, globalWorkOffset, globalWorkSize, localWorkSize);
|
||||
amd::Command::EventWaitList waitList;
|
||||
bool profileNDRange = false;
|
||||
address kernargs = nullptr;
|
||||
|
||||
profileNDRange = (startEvent != nullptr || stopEvent != nullptr);
|
||||
|
||||
// Flag set to 1 signifies that kernel can be launched in anyorder
|
||||
if (flags & hipExtAnyOrderLaunch) {
|
||||
params |= amd::NDRangeKernelCommand::AnyOrderLaunch;
|
||||
params |= amd::NDRangeKernelCommand::AnyOrderLaunch;
|
||||
}
|
||||
|
||||
amd::NDRangeKernelCommand* command = new amd::NDRangeKernelCommand(
|
||||
*queue, waitList, *kernel, ndrange, sharedMemBytes,
|
||||
params, gridId, numGrids, prevGridSum, allGridSum, firstDevice, profileNDRange);
|
||||
if (!command) {
|
||||
amd::NDRangeKernelCommand* kernelCommand = new amd::NDRangeKernelCommand(
|
||||
*queue, waitList, *kernel, ndrange, sharedMemBytes, params, gridId, numGrids, prevGridSum,
|
||||
allGridSum, firstDevice, profileNDRange);
|
||||
if (!kernelCommand) {
|
||||
return hipErrorOutOfMemory;
|
||||
}
|
||||
|
||||
// Capture the kernel arguments
|
||||
if (CL_SUCCESS != command->captureAndValidate()) {
|
||||
delete command;
|
||||
if (CL_SUCCESS != kernelCommand->captureAndValidate()) {
|
||||
delete kernelCommand;
|
||||
return hipErrorOutOfMemory;
|
||||
}
|
||||
command = kernelCommand;
|
||||
return hipSuccess;
|
||||
}
|
||||
|
||||
hipError_t ihipModuleLaunchKernel(hipFunction_t f, uint32_t globalWorkSizeX,
|
||||
uint32_t globalWorkSizeY, uint32_t globalWorkSizeZ,
|
||||
uint32_t blockDimX, uint32_t blockDimY, uint32_t blockDimZ,
|
||||
uint32_t sharedMemBytes, hipStream_t hStream, void** kernelParams,
|
||||
void** extra, hipEvent_t startEvent, hipEvent_t stopEvent,
|
||||
uint32_t flags = 0, uint32_t params = 0, uint32_t gridId = 0,
|
||||
uint32_t numGrids = 0, uint64_t prevGridSum = 0,
|
||||
uint64_t allGridSum = 0, uint32_t firstDevice = 0) {
|
||||
HIP_INIT_API(ihipModuleLaunchKernel, f, globalWorkSizeX, globalWorkSizeY, globalWorkSizeZ,
|
||||
blockDimX, blockDimY, blockDimZ, sharedMemBytes, hStream, kernelParams, extra,
|
||||
startEvent, stopEvent, flags, params);
|
||||
|
||||
HIP_RETURN_ONFAIL(PlatformState::instance().initStatManagedVarDevicePtr(ihipGetDevice()));
|
||||
if (f == nullptr) {
|
||||
LogPrintfError("%s", "Function passed is null");
|
||||
return hipErrorInvalidImage;
|
||||
}
|
||||
hip::DeviceFunc* function = hip::DeviceFunc::asFunction(f);
|
||||
amd::Kernel* kernel = function->kernel();
|
||||
amd::ScopedLock lock(function->dflock_);
|
||||
|
||||
hipError_t status =
|
||||
ihipLaunchKernel_validate(f, globalWorkSizeX, globalWorkSizeY, globalWorkSizeZ, blockDimX,
|
||||
blockDimY, blockDimZ, sharedMemBytes, kernelParams, extra, params);
|
||||
if (status != hipSuccess) {
|
||||
return status;
|
||||
}
|
||||
amd::Command* command = nullptr;
|
||||
amd::HostQueue* queue = hip::getQueue(hStream);
|
||||
status = ihipLaunchKernelCommand(command, f, globalWorkSizeX, globalWorkSizeY, globalWorkSizeZ,
|
||||
blockDimX, blockDimY, blockDimZ, sharedMemBytes, queue,
|
||||
kernelParams, extra, startEvent, stopEvent, flags, params,
|
||||
gridId, numGrids, prevGridSum, allGridSum, firstDevice);
|
||||
if (status != hipSuccess) {
|
||||
return status;
|
||||
}
|
||||
|
||||
hip::Event* eStart = reinterpret_cast<hip::Event*>(startEvent);
|
||||
hip::Event* eStop = reinterpret_cast<hip::Event*>(stopEvent);
|
||||
command->enqueue();
|
||||
|
||||
if (startEvent != nullptr) {
|
||||
|
||||
مرجع در شماره جدید
Block a user