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

fix: Use getGpuAddressToPatch() of pattern allocation

Browse Source 
When using FillBuffer builtin, use GPU address to patch (i.e. gpuAddress
plus allocationOffset minus gpuBaseAddress) instead of gpuAddress.
This change fixes a bug - possible abort due to huge difference between
gpuAddress() and gpuAddressToPatch() values.

Related-To: NEO-7973
Signed-off-by: Kacper Nowak <kacper.nowak@intel.com>
This commit is contained in:
Kacper Nowak
2023-10-17 02:40:07 +00:00
committed by Compute-Runtime-Automation
parent 750b5ba89a
commit 26666d8c6f
3 changed files with 39 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files

35
opencl/test/unit_test/command_queue/enqueue_fill_buffer_tests.cpp
View File

@@ -1,5 +1,5 @@
/*
* Copyright (C) 2018-2022 Intel Corporation
* Copyright (C) 2018-2023 Intel Corporation
*
* SPDX-License-Identifier: MIT
*
@@ -523,6 +523,39 @@ HWTEST_F(EnqueueFillBufferCmdTests, givenEnqueueFillBufferWhenPatternAllocationI
EXPECT_EQ(AllocationType::FILL_PATTERN, patternAllocation->getAllocationType());
}
HWTEST_F(EnqueueFillBufferCmdTests, whenFillingBufferThenUseGpuAddressForPatchingOfPatternAllocation) {
auto patternAllocation = context.getMemoryManager()->allocateGraphicsMemoryWithProperties(MockAllocationProperties{context.getDevice(0)->getRootDeviceIndex(), EnqueueFillBufferTraits::patternSize});
// Set gpuBaseAddress and offset so gpuAddress != gpuAddressToPatch
if (0u == patternAllocation->getGpuBaseAddress()) {
patternAllocation->setGpuBaseAddress(4096u);
}
patternAllocation->setAllocationOffset(10u);
EnqueueFillBufferHelper<>::enqueueFillBuffer(pCmdQ, buffer);
auto &builder = BuiltInDispatchBuilderOp::getBuiltinDispatchInfoBuilder(EBuiltInOps::FillBuffer, pCmdQ->getClDevice());
ASSERT_NE(nullptr, &builder);
BuiltinOpParams dc;
MemObj patternMemObj(&this->context, 0, {}, 0, 0, alignUp(EnqueueFillBufferTraits::patternSize, 4), patternAllocation->getUnderlyingBuffer(),
patternAllocation->getUnderlyingBuffer(), GraphicsAllocationHelper::toMultiGraphicsAllocation(patternAllocation), false, false, true);
dc.srcMemObj = &patternMemObj;
dc.dstMemObj = buffer;
dc.dstOffset = {EnqueueFillBufferTraits::offset, 0, 0};
dc.size = {EnqueueFillBufferTraits::size, 0, 0};
MultiDispatchInfo multiDispatchInfo(dc);
builder.buildDispatchInfos(multiDispatchInfo);
EXPECT_NE(0u, multiDispatchInfo.size());
auto kernel = multiDispatchInfo.begin()->getKernel();
auto patternArgIndex = 2;
const auto &patternArg = kernel->getKernelArguments().at(patternArgIndex);
EXPECT_EQ(patternAllocation->getGpuAddressToPatch(), reinterpret_cast<uint64_t>(patternArg.value));
context.getMemoryManager()->freeGraphicsMemory(patternAllocation);
}
struct EnqueueFillBufferHw : public ::testing::Test {
void SetUp() override {

8
opencl/test/unit_test/command_queue/enqueue_svm_mem_fill_tests.cpp
View File

@@ -55,7 +55,7 @@ struct BaseEnqueueSvmMemFillFixture : public ClDeviceFixture,
using BaseEnqueueSvmMemFillTest = Test<BaseEnqueueSvmMemFillFixture>;
HWTEST_F(BaseEnqueueSvmMemFillTest, givenEnqueueSVMMemFillWhenUsingFillBufferBuilderThenUseGpuAddressOfPatternSVMAllocation) {
HWTEST_F(BaseEnqueueSvmMemFillTest, givenEnqueueSVMMemFillWhenUsingFillBufferBuilderThenUseGpuAddressForPatchingOfPatternSVMAllocation) {
struct MockFillBufferBuilder : MockBuiltinDispatchInfoBuilder {
MockFillBufferBuilder(BuiltIns &kernelLib, ClDevice &clDevice, BuiltinDispatchInfoBuilder *origBuilder, const void *pattern, size_t patternSize)
: MockBuiltinDispatchInfoBuilder(kernelLib, clDevice, origBuilder),
@@ -82,10 +82,10 @@ HWTEST_F(BaseEnqueueSvmMemFillTest, givenEnqueueSVMMemFillWhenUsingFillBufferBui
size_t patternSize = 0x10u;
auto patternAllocation = static_cast<MockGraphicsAllocation *>(context->getMemoryManager()->allocateGraphicsMemoryWithProperties({pCmdQ->getDevice().getRootDeviceIndex(), 2 * patternSize, AllocationType::FILL_PATTERN, pCmdQ->getDevice().getDeviceBitfield()}));
// offset cpuPtr so cpuPtr != gpuAddress in order to ensure that setArgSVM will be called using gpu address of the pattern allocation
// offset cpuPtr so cpuPtr != gpuAddress (for patching) in order to ensure that setArgSVM will be called using gpu address of the pattern allocation
auto origCpuPtr = patternAllocation->cpuPtr;
patternAllocation->cpuPtr = ptrOffset(patternAllocation->cpuPtr, patternSize);
ASSERT_NE((uint64_t)patternAllocation->cpuPtr, patternAllocation->getGpuAddress());
ASSERT_NE((uint64_t)patternAllocation->cpuPtr, patternAllocation->getGpuAddressToPatch());
auto internalAllocStorage = pCmdQ->getGpgpuCommandStreamReceiver().getInternalAllocationStorage();
internalAllocStorage->storeAllocation(std::unique_ptr<GraphicsAllocation>(patternAllocation), REUSABLE_ALLOCATION);
@@ -111,7 +111,7 @@ HWTEST_F(BaseEnqueueSvmMemFillTest, givenEnqueueSVMMemFillWhenUsingFillBufferBui
auto patternArgIndex = 2;
const auto &patternArg = kernel->getKernelArguments().at(patternArgIndex);
EXPECT_EQ(patternAllocation->getGpuAddress(), reinterpret_cast<uint64_t>(patternArg.value));
EXPECT_EQ(patternAllocation->getGpuAddressToPatch(), reinterpret_cast<uint64_t>(patternArg.value));
patternAllocation->cpuPtr = origCpuPtr;
}