intel/compute-runtime
1
0
Fork 0
mirror of https://github.com/intel/compute-runtime.git synced 2025-12-21 01:04:57 +08:00
Code Issues Packages Projects Releases Wiki Activity

Correct patching private surface in cloned kernel

Browse Source 
Related-To: NEO-5081
Signed-off-by: Mateusz Jablonski <mateusz.jablonski@intel.com>
This commit is contained in:
Mateusz Jablonski
2021-09-07 11:10:18 +00:00
committed by Compute-Runtime-Automation
parent caddc63eec
commit 7f920139b4
3 changed files with 65 additions and 24 deletions
Download Patch File Download Diff File Expand all files Collapse all files

54
opencl/source/kernel/kernel.cpp
View File

@@ -202,25 +202,9 @@ cl_int Kernel::initialize() {
localBindingTableOffset = kernelDescriptor.payloadMappings.bindingTable.tableOffset;
// patch crossthread data and ssh with inline surfaces, if necessary
auto perHwThreadPrivateMemorySize = kernelDescriptor.kernelAttributes.perHwThreadPrivateMemorySize;
if (perHwThreadPrivateMemorySize) {
privateSurfaceSize = KernelHelper::getPrivateSurfaceSize(perHwThreadPrivateMemorySize, pClDevice->getSharedDeviceInfo().computeUnitsUsedForScratch);
DEBUG_BREAK_IF(privateSurfaceSize == 0);
if (privateSurfaceSize > std::numeric_limits<uint32_t>::max()) {
return CL_OUT_OF_RESOURCES;
}
privateSurface = executionEnvironment.memoryManager->allocateGraphicsMemoryWithProperties(
{rootDeviceIndex,
static_cast<size_t>(privateSurfaceSize),
GraphicsAllocation::AllocationType::PRIVATE_SURFACE,
pClDevice->getDeviceBitfield()});
if (privateSurface == nullptr) {
return CL_OUT_OF_RESOURCES;
}
const auto &privateMemoryAddress = kernelDescriptor.payloadMappings.implicitArgs.privateMemoryAddress;
patchWithImplicitSurface(reinterpret_cast<void *>(privateSurface->getGpuAddressToPatch()), *privateSurface, privateMemoryAddress);
auto status = patchPrivateSurface();
if (CL_SUCCESS != status) {
return status;
}
if (isValidOffset(kernelDescriptor.payloadMappings.implicitArgs.globalConstantsSurfaceAddress.stateless)) {
@@ -334,12 +318,44 @@ cl_int Kernel::initialize() {
return CL_SUCCESS;
}
cl_int Kernel::patchPrivateSurface() {
auto pClDevice = &getDevice();
auto rootDeviceIndex = pClDevice->getRootDeviceIndex();
auto &kernelDescriptor = kernelInfo.kernelDescriptor;
auto perHwThreadPrivateMemorySize = kernelDescriptor.kernelAttributes.perHwThreadPrivateMemorySize;
if (perHwThreadPrivateMemorySize) {
if (!privateSurface) {
privateSurfaceSize = KernelHelper::getPrivateSurfaceSize(perHwThreadPrivateMemorySize, pClDevice->getSharedDeviceInfo().computeUnitsUsedForScratch);
DEBUG_BREAK_IF(privateSurfaceSize == 0);
if (privateSurfaceSize > std::numeric_limits<uint32_t>::max()) {
return CL_OUT_OF_RESOURCES;
}
privateSurface = executionEnvironment.memoryManager->allocateGraphicsMemoryWithProperties(
{rootDeviceIndex,
static_cast<size_t>(privateSurfaceSize),
GraphicsAllocation::AllocationType::PRIVATE_SURFACE,
pClDevice->getDeviceBitfield()});
if (privateSurface == nullptr) {
return CL_OUT_OF_RESOURCES;
}
}
const auto &privateMemoryAddress = kernelDescriptor.payloadMappings.implicitArgs.privateMemoryAddress;
patchWithImplicitSurface(reinterpret_cast<void *>(privateSurface->getGpuAddressToPatch()), *privateSurface, privateMemoryAddress);
}
return CL_SUCCESS;
}
cl_int Kernel::cloneKernel(Kernel *pSourceKernel) {
// copy cross thread data to store arguments set to source kernel with clSetKernelArg on immediate data (non-pointer types)
memcpy_s(crossThreadData, crossThreadDataSize,
pSourceKernel->crossThreadData, pSourceKernel->crossThreadDataSize);
DEBUG_BREAK_IF(pSourceKernel->crossThreadDataSize != crossThreadDataSize);
[[maybe_unused]] auto status = patchPrivateSurface();
DEBUG_BREAK_IF(status != CL_SUCCESS);
// copy arguments set to source kernel with clSetKernelArg or clSetKernelArgSVMPointer
for (uint32_t i = 0; i < pSourceKernel->kernelArguments.size(); i++) {
if (0 == pSourceKernel->getKernelArgInfo(i).size) {