intel/compute-runtime
1
0
Fork 0
mirror of https://github.com/intel/compute-runtime.git synced 2026-01-03 06:49:52 +08:00
Code Issues Packages Projects Releases Wiki Activity

Enable sharedMemory path for write buffer

Browse Source 
Signed-off-by: Andrzej Koska <andrzej.koska@intel.com>
Related-To: NEO-5360
This commit is contained in:
Andrzej Koska
2021-01-25 09:48:31 +00:00
committed by Compute-Runtime-Automation
parent 4417537eb5
commit 17bbe80271
3 changed files with 56 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
opencl/source/command_queue/enqueue_read_buffer.h
View File

@@ -60,7 +60,6 @@ cl_int CommandQueueHw<GfxFamily>::enqueueReadBuffer(
if ((svmEntry->gpuAllocations.getGraphicsAllocation(rootDeviceIndex)->getGpuAddress() + svmEntry->size) < (castToUint64(ptr) + size)) {
return CL_INVALID_OPERATION;
}
mapAllocation = svmEntry->cpuAllocation ? svmEntry->cpuAllocation : svmEntry->gpuAllocations.getGraphicsAllocation(rootDeviceIndex);
if (isCpuCopyAllowed) {
if (svmEntry->memoryType == DEVICE_UNIFIED_MEMORY) {
@@ -99,7 +98,7 @@ cl_int CommandQueueHw<GfxFamily>::enqueueReadBuffer(
surfaces[1] = &mapSurface;
mapSurface.setGraphicsAllocation(mapAllocation);
//get offset between base cpu ptr of map allocation and dst ptr
if (memoryType != DEVICE_UNIFIED_MEMORY) {
if ((memoryType != DEVICE_UNIFIED_MEMORY) && (memoryType != SHARED_UNIFIED_MEMORY)) {
size_t dstOffset = ptrDiff(dstPtr, mapAllocation->getUnderlyingBuffer());
dstPtr = reinterpret_cast<void *>(mapAllocation->getGpuAddress() + dstOffset);
}

3
opencl/source/command_queue/enqueue_write_buffer.h
View File

@@ -56,7 +56,6 @@ cl_int CommandQueueHw<GfxFamily>::enqueueWriteBuffer(
isCpuCopyAllowed = false;
}
}
mapAllocation = svmEntry->cpuAllocation ? svmEntry->cpuAllocation : svmEntry->gpuAllocations.getGraphicsAllocation(rootDeviceIndex);
}
}
@@ -91,7 +90,7 @@ cl_int CommandQueueHw<GfxFamily>::enqueueWriteBuffer(
surfaces[1] = &mapSurface;
mapSurface.setGraphicsAllocation(mapAllocation);
//get offset between base cpu ptr of map allocation and dst ptr
if (memoryType != DEVICE_UNIFIED_MEMORY) {
if ((memoryType != DEVICE_UNIFIED_MEMORY) && (memoryType != SHARED_UNIFIED_MEMORY)) {
size_t srcOffset = ptrDiff(srcPtr, mapAllocation->getUnderlyingBuffer());
srcPtr = reinterpret_cast<void *>(mapAllocation->getGpuAddress() + srcOffset);
}

54
opencl/test/unit_test/memory_manager/unified_memory_manager_tests.cpp
View File

@@ -1003,4 +1003,58 @@ HWTEST_F(UnfiedSharedMemoryHWTest, givenDeviceUsmAllocationWhenReadBufferThenCpu
gpuAllocation->setCpuPtrAndGpuAddress(cpuPtr, gpuAddress);
delete buffer;
clMemFreeINTEL(&mockContext, deviceMemory);
}
HWTEST_F(UnfiedSharedMemoryHWTest, givenSharedUsmAllocationWhenWriteBufferThenCpuPtrIsNotUsed) {
SVMAllocsManager::UnifiedMemoryProperties unifiedMemoryProperties(InternalMemoryType::SHARED_UNIFIED_MEMORY,
mockContext.getRootDeviceIndices(), mockContext.getDeviceBitfields());
auto sharedMemory = mockContext.getSVMAllocsManager()->createUnifiedMemoryAllocation(4096u, unifiedMemoryProperties);
auto svmAllocation = mockContext.getSVMAllocsManager()->getSVMAlloc(sharedMemory);
GraphicsAllocation *gpuAllocation = svmAllocation->gpuAllocations.getGraphicsAllocation(mockContext.getDevice(0)->getRootDeviceIndex());
char *cpuPtr = static_cast<char *>(gpuAllocation->getUnderlyingBuffer());
auto gpuAddress = gpuAllocation->getGpuAddress();
void *gpuPtr = reinterpret_cast<void *>(gpuAddress);
char *shiftedPtr = cpuPtr + 0x10;
gpuAllocation->setCpuPtrAndGpuAddress(shiftedPtr, reinterpret_cast<uint64_t>(shiftedPtr));
cl_mem_flags flags = 0;
auto status = CL_INVALID_PLATFORM;
auto buffer = Buffer::create(&mockContext, flags, 4096u, nullptr, status);
ASSERT_EQ(CL_SUCCESS, status);
TestCommandQueueHw<FamilyType> myCmdQ(&mockContext, mockContext.getDevice(0u), 0);
myCmdQ.enqueueWriteBuffer(buffer, false, 0u, 4096u, sharedMemory, nullptr, 0u, nullptr, nullptr);
EXPECT_EQ(gpuPtr, myCmdQ.srcPtr);
gpuAllocation->setCpuPtrAndGpuAddress(cpuPtr, gpuAddress);
delete buffer;
clMemFreeINTEL(&mockContext, sharedMemory);
}
HWTEST_F(UnfiedSharedMemoryHWTest, givenSharedUsmAllocationWhenReadBufferThenCpuPtrIsNotUsed) {
SVMAllocsManager::UnifiedMemoryProperties unifiedMemoryProperties(InternalMemoryType::SHARED_UNIFIED_MEMORY,
mockContext.getRootDeviceIndices(), mockContext.getDeviceBitfields());
auto sharedMemory = mockContext.getSVMAllocsManager()->createUnifiedMemoryAllocation(4096u, unifiedMemoryProperties);
auto svmAllocation = mockContext.getSVMAllocsManager()->getSVMAlloc(sharedMemory);
GraphicsAllocation *gpuAllocation = svmAllocation->gpuAllocations.getGraphicsAllocation(mockContext.getDevice(0)->getRootDeviceIndex());
char *cpuPtr = static_cast<char *>(gpuAllocation->getUnderlyingBuffer());
auto gpuAddress = gpuAllocation->getGpuAddress();
void *gpuPtr = reinterpret_cast<void *>(gpuAddress);
char *shiftedPtr = cpuPtr + 0x10;
gpuAllocation->setCpuPtrAndGpuAddress(shiftedPtr, reinterpret_cast<uint64_t>(shiftedPtr));
cl_mem_flags flags = 0;
auto status = CL_INVALID_PLATFORM;
auto buffer = Buffer::create(&mockContext, flags, 4096u, nullptr, status);
ASSERT_EQ(CL_SUCCESS, status);
TestCommandQueueHw<FamilyType> myCmdQ(&mockContext, mockContext.getDevice(0u), 0);
myCmdQ.enqueueReadBuffer(buffer, false, 0u, 4096u, sharedMemory, nullptr, 0u, nullptr, nullptr);
EXPECT_EQ(gpuPtr, myCmdQ.dstPtr);
gpuAllocation->setCpuPtrAndGpuAddress(cpuPtr, gpuAddress);
delete buffer;
clMemFreeINTEL(&mockContext, sharedMemory);
}