intel/compute-runtime
1
0
Fork 0
mirror of https://github.com/intel/compute-runtime.git synced 2026-01-11 16:45:25 +08:00
Code Issues Packages Projects Releases Wiki Activity

Copy host ptr on cpu if possible in clCreateBuffer

Browse Source 
use cpu copy with locked pointer if possible
because this is faster than copy on gpu
limit to buffers of size at most 64kb

Related-To: NEO-7332

Signed-off-by: Dominik Dabek <dominik.dabek@intel.com>
This commit is contained in:
Dominik Dabek
2022-09-21 12:39:39 +00:00
committed by Compute-Runtime-Automation
parent 7ded401615
commit d8b7d56160
7 changed files with 166 additions and 14 deletions
Download Patch File Download Diff File Expand all files Collapse all files

33
opencl/source/mem_obj/buffer.cpp
View File

@@ -7,6 +7,7 @@
#include "opencl/source/mem_obj/buffer.h"
#include "shared/source/command_container/implicit_scaling.h"
#include "shared/source/command_stream/command_stream_receiver.h"
#include "shared/source/debug_settings/debug_settings_manager.h"
#include "shared/source/execution_environment/root_device_environment.h"
@@ -410,19 +411,33 @@ Buffer *Buffer::create(Context *context,
auto &device = pBuffer->getContext()->getDevice(0u)->getDevice();
auto &hwInfo = device.getHardwareInfo();
auto hwInfoConfig = HwInfoConfig::get(hwInfo.platform.eProductFamily);
auto blitMemoryToAllocationResult = BlitOperationResult::Unsupported;
if (hwInfoConfig->isBlitterFullySupported(hwInfo) && isLocalMemory) {
blitMemoryToAllocationResult = BlitHelperFunctions::blitMemoryToAllocation(device, allocationInfo.memory, pBuffer->getOffset(), hostPtr, {size, 1, 1});
bool copyOnCpuAllowed = false == ImplicitScalingHelper::isImplicitScalingEnabled(device.getDeviceBitfield(), true) &&
size <= Buffer::maxBufferSizeForCopyOnCpu &&
!isCompressionEnabled &&
hwInfoConfig->getLocalMemoryAccessMode(hwInfo) != LocalMemoryAccessMode::CpuAccessDisallowed;
if (DebugManager.flags.CopyHostPtrOnCpu.get() != -1) {
copyOnCpuAllowed = DebugManager.flags.CopyHostPtrOnCpu.get() == 1;
}
if (auto lockedPointer = copyOnCpuAllowed ? device.getMemoryManager()->lockResource(allocationInfo.memory) : nullptr) {
memcpy_s(ptrOffset(lockedPointer, pBuffer->getOffset()), size, hostPtr, size);
allocationInfo.memory->setAubWritable(true, GraphicsAllocation::defaultBank);
allocationInfo.memory->setTbxWritable(true, GraphicsAllocation::defaultBank);
copyExecuted = true;
} else {
auto blitMemoryToAllocationResult = BlitOperationResult::Unsupported;
if (blitMemoryToAllocationResult != BlitOperationResult::Success) {
auto cmdQ = context->getSpecialQueue(rootDeviceIndex);
if (CL_SUCCESS != cmdQ->enqueueWriteBuffer(pBuffer, CL_TRUE, 0, size, hostPtr, allocationInfo.mapAllocation, 0, nullptr, nullptr)) {
errcodeRet = CL_OUT_OF_RESOURCES;
if (hwInfoConfig->isBlitterFullySupported(hwInfo) && isLocalMemory) {
blitMemoryToAllocationResult = BlitHelperFunctions::blitMemoryToAllocation(device, allocationInfo.memory, pBuffer->getOffset(), hostPtr, {size, 1, 1});
}
if (blitMemoryToAllocationResult != BlitOperationResult::Success) {
auto cmdQ = context->getSpecialQueue(rootDeviceIndex);
if (CL_SUCCESS != cmdQ->enqueueWriteBuffer(pBuffer, CL_TRUE, 0, size, hostPtr, allocationInfo.mapAllocation, 0, nullptr, nullptr)) {
errcodeRet = CL_OUT_OF_RESOURCES;
}
}
copyExecuted = true;
}
copyExecuted = true;
} else {
memcpy_s(allocationInfo.memory->getUnderlyingBuffer(), size, hostPtr, size);
copyExecuted = true;

1
opencl/source/mem_obj/buffer.h
View File

@@ -58,6 +58,7 @@ extern ValidateInputAndCreateBufferFunc validateInputAndCreateBuffer;
class Buffer : public MemObj {
public:
constexpr static size_t maxBufferSizeForReadWriteOnCpu = 10 * MB;
constexpr static size_t maxBufferSizeForCopyOnCpu = 64 * KB;
constexpr static cl_ulong maskMagic = 0xFFFFFFFFFFFFFFFFLL;
constexpr static cl_ulong objectMagic = MemObj::objectMagic | 0x02;
bool forceDisallowCPUCopy = false;