intel/compute-runtime
1
0
Fork 0
mirror of https://github.com/intel/compute-runtime.git synced 2025-12-27 07:44:16 +08:00
Code Issues Packages Projects Releases Wiki Activity

Update disabling caching for a resource

Browse Source 
Change-Id: I00eac0add01f75a1b82d04cf42652c15b776a457
Signed-off-by: Filip Hazubski <filip.hazubski@intel.com>
This commit is contained in:
Filip Hazubski
2019-01-30 10:57:42 +01:00
committed by sys_ocldev
parent 32ecd91401
commit d30cc221df
10 changed files with 182 additions and 48 deletions
Download Patch File Download Diff File Expand all files Collapse all files

183
unit_tests/api/cl_mem_locally_uncached_resource_tests.cpp
View File

@@ -17,71 +17,172 @@
#include "test.h"
#include "unit_tests/fixtures/hello_world_fixture.h"
#include "unit_tests/helpers/hw_parse.h"
#include "unit_tests/utilities/base_object_utils.h"
using namespace OCLRT;
namespace clMemLocallyUncachedResourceTests {
struct clMemLocallyUncachedResourceFixture : Test<HelloWorldFixture<HelloWorldFixtureFactory>>,
::testing::WithParamInterface<bool> {};
HWTEST_P(clMemLocallyUncachedResourceFixture, GivenLocallyCachedOrUncachedBufferWhenItIsSetAndQueuedThenItIsCorrectlyCached) {
template <typename FamilyType>
uint32_t argMocs(Kernel &kernel, size_t argIndex) {
using RENDER_SURFACE_STATE = typename FamilyType::RENDER_SURFACE_STATE;
auto surfaceStateHeapAddress = kernel.getSurfaceStateHeap();
auto surfaceStateHeapAddressOffset = kernel.getKernelInfo().kernelArgInfo[argIndex].offsetHeap;
auto surfaceState = reinterpret_cast<RENDER_SURFACE_STATE *>(ptrOffset(surfaceStateHeapAddress, surfaceStateHeapAddressOffset));
return surfaceState->getMemoryObjectControlState();
}
template <typename FamilyType>
uint32_t cmdQueueMocs(CommandQueue *pCmdQ) {
using STATE_BASE_ADDRESS = typename FamilyType::STATE_BASE_ADDRESS;
auto pCmdQHw = reinterpret_cast<CommandQueueHw<FamilyType> *>(pCmdQ);
auto &csr = pCmdQHw->getCommandStreamReceiver();
HardwareParse hwParse;
hwParse.parseCommands<FamilyType>(csr.getCS(0), 0);
auto itorCmd = reverse_find<STATE_BASE_ADDRESS *>(hwParse.cmdList.rbegin(), hwParse.cmdList.rend());
EXPECT_NE(hwParse.cmdList.rend(), itorCmd);
auto sba = genCmdCast<STATE_BASE_ADDRESS *>(*itorCmd);
EXPECT_NE(nullptr, sba);
const size_t n = 512;
size_t globalWorkSize[3] = {n, 1, 1};
size_t localWorkSize[3] = {256, 1, 1};
bool useUncachedFlag = GetParam();
return sba->getStatelessDataPortAccessMemoryObjectControlState();
}
const size_t n = 512;
const size_t globalWorkSize[3] = {n, 1, 1};
const size_t localWorkSize[3] = {256, 1, 1};
const cl_mem_properties_intel *propertiesCacheable = nullptr;
const cl_mem_properties_intel propertiesUncacheable[] = {CL_MEM_FLAGS_INTEL, CL_MEM_LOCALLY_UNCACHED_RESOURCE, 0};
using clMemLocallyUncachedResourceFixture = Test<HelloWorldFixture<HelloWorldFixtureFactory>>;
HWTEST_F(clMemLocallyUncachedResourceFixture, GivenAtLeastOneLocallyUncacheableResourceWhenSettingKernelArgumentsThenKernelIsUncacheable) {
cl_int retVal = CL_SUCCESS;
std::unique_ptr<Kernel> kernel(Kernel::create(pProgram, *pProgram->getKernelInfo("CopyBuffer"), &retVal));
EXPECT_EQ(CL_SUCCESS, retVal);
cl_mem_properties_intel propertiesUncached[] = {CL_MEM_FLAGS_INTEL, CL_MEM_LOCALLY_UNCACHED_RESOURCE, 0};
cl_mem_properties_intel *properties = (useUncachedFlag ? propertiesUncached : nullptr);
auto buffer1 = clCreateBufferWithPropertiesINTEL(context, properties, n * sizeof(float), nullptr, nullptr);
auto buffer2 = clCreateBufferWithPropertiesINTEL(context, properties, n * sizeof(float), nullptr, nullptr);
auto bufferCacheable1 = clCreateBufferWithPropertiesINTEL(context, propertiesCacheable, n * sizeof(float), nullptr, nullptr);
auto pBufferCacheable1 = clUniquePtr(castToObject<Buffer>(bufferCacheable1));
auto bufferCacheable2 = clCreateBufferWithPropertiesINTEL(context, propertiesCacheable, n * sizeof(float), nullptr, nullptr);
auto pBufferCacheable2 = clUniquePtr(castToObject<Buffer>(bufferCacheable2));
retVal = clSetKernelArg(kernel.get(), 0, sizeof(cl_mem), &buffer1);
EXPECT_EQ(CL_SUCCESS, retVal);
auto surfaceStateHeapAddress = kernel.get()->getSurfaceStateHeap();
auto surfaceStateHeapAddressOffset = kernel.get()->getKernelInfo().kernelArgInfo[0].offsetHeap;
auto surfaceState = reinterpret_cast<RENDER_SURFACE_STATE *>(ptrOffset(surfaceStateHeapAddress, surfaceStateHeapAddressOffset));
auto expectedMocs = pDevice->getGmmHelper()->getMOCS(useUncachedFlag ? GMM_RESOURCE_USAGE_OCL_BUFFER_CACHELINE_MISALIGNED
: GMM_RESOURCE_USAGE_OCL_BUFFER);
EXPECT_EQ(expectedMocs, surfaceState->getMemoryObjectControlState());
auto bufferUncacheable1 = clCreateBufferWithPropertiesINTEL(context, propertiesUncacheable, n * sizeof(float), nullptr, nullptr);
auto pBufferUncacheable1 = clUniquePtr(castToObject<Buffer>(bufferUncacheable1));
auto bufferUncacheable2 = clCreateBufferWithPropertiesINTEL(context, propertiesUncacheable, n * sizeof(float), nullptr, nullptr);
auto pBufferUncacheable2 = clUniquePtr(castToObject<Buffer>(bufferUncacheable2));
retVal = clSetKernelArg(kernel.get(), 1, sizeof(cl_mem), &buffer2);
auto mocsCacheable = kernel->getDevice().getGmmHelper()->getMOCS(GMM_RESOURCE_USAGE_OCL_BUFFER);
auto mocsUncacheable = kernel->getDevice().getGmmHelper()->getMOCS(GMM_RESOURCE_USAGE_OCL_BUFFER_CACHELINE_MISALIGNED);
retVal = clSetKernelArg(kernel.get(), 0, sizeof(cl_mem), &bufferCacheable1);
EXPECT_EQ(CL_SUCCESS, retVal);
surfaceStateHeapAddressOffset = kernel.get()->getKernelInfo().kernelArgInfo[1].offsetHeap;
surfaceState = reinterpret_cast<RENDER_SURFACE_STATE *>(ptrOffset(surfaceStateHeapAddress, surfaceStateHeapAddressOffset));
EXPECT_EQ(expectedMocs, surfaceState->getMemoryObjectControlState());
EXPECT_EQ(mocsCacheable, argMocs<FamilyType>(*kernel, 0));
retVal = clSetKernelArg(kernel.get(), 1, sizeof(cl_mem), &bufferCacheable2);
EXPECT_EQ(CL_SUCCESS, retVal);
EXPECT_EQ(mocsCacheable, argMocs<FamilyType>(*kernel, 1));
EXPECT_TRUE(kernel->isPatched());
retVal = clEnqueueNDRangeKernel(pCmdQ, kernel.get(), 1, nullptr, globalWorkSize, localWorkSize, 0, nullptr, nullptr);
EXPECT_EQ(CL_SUCCESS, retVal);
EXPECT_EQ(mocsCacheable, cmdQueueMocs<FamilyType>(pCmdQ));
auto pCmdQHw = reinterpret_cast<CommandQueueHw<FamilyType> *>(pCmdQ);
ASSERT_NE(nullptr, pCmdQHw);
auto &csr = pCmdQHw->getCommandStreamReceiver();
HardwareParse hwParse;
hwParse.parseCommands<FamilyType>(csr.getCS(0), 0);
auto itorCmd = find<STATE_BASE_ADDRESS *>(hwParse.cmdList.begin(), hwParse.cmdList.end());
EXPECT_NE(hwParse.cmdList.end(), itorCmd);
auto sba = genCmdCast<STATE_BASE_ADDRESS *>(*itorCmd);
ASSERT_NE(nullptr, sba);
EXPECT_EQ(expectedMocs, sba->getStatelessDataPortAccessMemoryObjectControlState());
retVal = clReleaseMemObject(buffer1);
retVal = clSetKernelArg(kernel.get(), 0, sizeof(cl_mem), &bufferUncacheable1);
EXPECT_EQ(CL_SUCCESS, retVal);
retVal = clReleaseMemObject(buffer2);
EXPECT_EQ(mocsUncacheable, argMocs<FamilyType>(*kernel, 0));
EXPECT_TRUE(kernel->isPatched());
retVal = clEnqueueNDRangeKernel(pCmdQ, kernel.get(), 1, nullptr, globalWorkSize, localWorkSize, 0, nullptr, nullptr);
EXPECT_EQ(CL_SUCCESS, retVal);
EXPECT_EQ(mocsUncacheable, cmdQueueMocs<FamilyType>(pCmdQ));
retVal = clSetKernelArg(kernel.get(), 1, sizeof(cl_mem), &bufferUncacheable2);
EXPECT_EQ(CL_SUCCESS, retVal);
EXPECT_EQ(mocsUncacheable, argMocs<FamilyType>(*kernel, 0));
EXPECT_TRUE(kernel->isPatched());
retVal = clEnqueueNDRangeKernel(pCmdQ, kernel.get(), 1, nullptr, globalWorkSize, localWorkSize, 0, nullptr, nullptr);
EXPECT_EQ(CL_SUCCESS, retVal);
EXPECT_EQ(mocsUncacheable, cmdQueueMocs<FamilyType>(pCmdQ));
retVal = clSetKernelArg(kernel.get(), 0, sizeof(cl_mem), &bufferCacheable1);
EXPECT_EQ(CL_SUCCESS, retVal);
EXPECT_EQ(mocsCacheable, argMocs<FamilyType>(*kernel, 0));
EXPECT_TRUE(kernel->isPatched());
retVal = clEnqueueNDRangeKernel(pCmdQ, kernel.get(), 1, nullptr, globalWorkSize, localWorkSize, 0, nullptr, nullptr);
EXPECT_EQ(CL_SUCCESS, retVal);
EXPECT_EQ(mocsUncacheable, cmdQueueMocs<FamilyType>(pCmdQ));
retVal = clSetKernelArg(kernel.get(), 1, sizeof(cl_mem), &bufferCacheable2);
EXPECT_EQ(CL_SUCCESS, retVal);
EXPECT_EQ(mocsCacheable, argMocs<FamilyType>(*kernel, 1));
EXPECT_TRUE(kernel->isPatched());
retVal = clEnqueueNDRangeKernel(pCmdQ, kernel.get(), 1, nullptr, globalWorkSize, localWorkSize, 0, nullptr, nullptr);
EXPECT_EQ(CL_SUCCESS, retVal);
EXPECT_EQ(mocsCacheable, cmdQueueMocs<FamilyType>(pCmdQ));
}
INSTANTIATE_TEST_CASE_P(clMemLocallyUncachedResourceTest,
clMemLocallyUncachedResourceFixture,
::testing::Bool());
HWTEST_F(clMemLocallyUncachedResourceFixture, WhenUnsettingUncacheableResourceFromKernelThanKernelContinuesToCorrectlySetMocs) {
cl_int retVal = CL_SUCCESS;
std::unique_ptr<Kernel> kernel(Kernel::create(pProgram, *pProgram->getKernelInfo("CopyBuffer"), &retVal));
EXPECT_EQ(CL_SUCCESS, retVal);
auto bufferCacheable1 = clCreateBufferWithPropertiesINTEL(context, propertiesCacheable, n * sizeof(float), nullptr, nullptr);
auto pBufferCacheable1 = clUniquePtr(castToObject<Buffer>(bufferCacheable1));
auto bufferCacheable2 = clCreateBufferWithPropertiesINTEL(context, propertiesCacheable, n * sizeof(float), nullptr, nullptr);
auto pBufferCacheable2 = clUniquePtr(castToObject<Buffer>(bufferCacheable2));
auto bufferUncacheable = clCreateBufferWithPropertiesINTEL(context, propertiesUncacheable, n * sizeof(float), nullptr, nullptr);
auto pBufferUncacheable = clUniquePtr(castToObject<Buffer>(bufferUncacheable));
auto mocsCacheable = kernel->getDevice().getGmmHelper()->getMOCS(GMM_RESOURCE_USAGE_OCL_BUFFER);
auto mocsUncacheable = kernel->getDevice().getGmmHelper()->getMOCS(GMM_RESOURCE_USAGE_OCL_BUFFER_CACHELINE_MISALIGNED);
retVal = clSetKernelArg(kernel.get(), 0, sizeof(cl_mem), &bufferCacheable1);
EXPECT_EQ(CL_SUCCESS, retVal);
EXPECT_EQ(mocsCacheable, argMocs<FamilyType>(*kernel, 0));
retVal = clSetKernelArg(kernel.get(), 1, sizeof(cl_mem), &bufferCacheable2);
EXPECT_EQ(CL_SUCCESS, retVal);
EXPECT_EQ(mocsCacheable, argMocs<FamilyType>(*kernel, 1));
EXPECT_TRUE(kernel->isPatched());
retVal = clEnqueueNDRangeKernel(pCmdQ, kernel.get(), 1, nullptr, globalWorkSize, localWorkSize, 0, nullptr, nullptr);
EXPECT_EQ(CL_SUCCESS, retVal);
EXPECT_EQ(mocsCacheable, cmdQueueMocs<FamilyType>(pCmdQ));
retVal = clSetKernelArg(kernel.get(), 0, sizeof(cl_mem), &bufferUncacheable);
EXPECT_EQ(CL_SUCCESS, retVal);
EXPECT_EQ(mocsUncacheable, argMocs<FamilyType>(*kernel, 0));
EXPECT_TRUE(kernel->isPatched());
retVal = clEnqueueNDRangeKernel(pCmdQ, kernel.get(), 1, nullptr, globalWorkSize, localWorkSize, 0, nullptr, nullptr);
EXPECT_EQ(CL_SUCCESS, retVal);
EXPECT_EQ(mocsUncacheable, cmdQueueMocs<FamilyType>(pCmdQ));
kernel->unsetArg(0);
retVal = clSetKernelArg(kernel.get(), 0, sizeof(cl_mem), &bufferCacheable1);
EXPECT_EQ(CL_SUCCESS, retVal);
EXPECT_EQ(mocsCacheable, argMocs<FamilyType>(*kernel, 0));
EXPECT_TRUE(kernel->isPatched());
retVal = clEnqueueNDRangeKernel(pCmdQ, kernel.get(), 1, nullptr, globalWorkSize, localWorkSize, 0, nullptr, nullptr);
EXPECT_EQ(CL_SUCCESS, retVal);
EXPECT_EQ(mocsCacheable, cmdQueueMocs<FamilyType>(pCmdQ));
kernel->unsetArg(0);
retVal = clSetKernelArg(kernel.get(), 0, sizeof(cl_mem), &bufferUncacheable);
EXPECT_EQ(CL_SUCCESS, retVal);
EXPECT_EQ(mocsUncacheable, argMocs<FamilyType>(*kernel, 0));
EXPECT_TRUE(kernel->isPatched());
retVal = clEnqueueNDRangeKernel(pCmdQ, kernel.get(), 1, nullptr, globalWorkSize, localWorkSize, 0, nullptr, nullptr);
EXPECT_EQ(CL_SUCCESS, retVal);
EXPECT_EQ(mocsUncacheable, cmdQueueMocs<FamilyType>(pCmdQ));
}
} // namespace clMemLocallyUncachedResourceTests