EXSWHTEC-169 - Implement additional tests for Kernel Graph Node APIs (#7)
- Tidy up hipGraphAddKernelNode tests - Tidy up hipGraphKernelNodeGetParams tests - Tidy up hipGraphKernelNodeSetParams tests - Tidy up hipGraphExecKernelNodeSetParams tests. - Disable failing test sections on AMD.
이 커밋은 다음에 포함됨:
Mirza Halilčević
GitHub
@@ -6,25 +6,27 @@ in the Software without restriction, including without limitation the rights
|
||||
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
|
||||
copies of the Software, and to permit persons to whom the Software is
|
||||
furnished to do so, subject to the following conditions:
|
||||
|
||||
The above copyright notice and this permission notice shall be included in
|
||||
all copies or substantial portions of the Software.
|
||||
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANNTY OF ANY KIND, EXPRESS OR
|
||||
|
||||
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
|
||||
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
|
||||
FITNNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
|
||||
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
|
||||
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
|
||||
LIABILITY, WHETHER INN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
|
||||
OUT OF OR INN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
|
||||
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
|
||||
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
|
||||
THE SOFTWARE.
|
||||
*/
|
||||
|
||||
/**
|
||||
Testcase Scenarios :
|
||||
Test Case Scenarios :
|
||||
Negative -
|
||||
1) Pass hGraphExec as nullptr and verify api returns error code.
|
||||
2) Pass node as nullptr and verify api returns error code.
|
||||
3) Pass NodeParams as un-initialized structure object and verify api returns error code.
|
||||
4) Pass pNodeParams as nullptr and verify api returns error code.
|
||||
5) Pass NodeParams:func datamember as nullptr and verify api returns error code.
|
||||
5) Pass NodeParams:func data member as nullptr and verify api returns error code.
|
||||
Functional -
|
||||
1) Instantiate a graph with kernel node, obtain executable graph and update
|
||||
the kernel node params with set and check it is taking effect.
|
||||
@@ -39,12 +41,11 @@ Functional -
|
||||
*/
|
||||
TEST_CASE("Unit_hipGraphExecKernelNodeSetParams_Negative") {
|
||||
constexpr size_t N = 1024;
|
||||
constexpr size_t Nbytes = N * sizeof(int);
|
||||
constexpr auto blocksPerCU = 6; // to hide latency
|
||||
constexpr auto threadsPerBlock = 256;
|
||||
unsigned blocks = HipTest::setNumBlocks(blocksPerCU, threadsPerBlock, N);
|
||||
hipGraph_t graph;
|
||||
hipError_t ret;
|
||||
hipGraphNode_t memcpyNode, kNode{};
|
||||
hipGraphNode_t kNode{};
|
||||
hipKernelNodeParams kNodeParams{};
|
||||
hipStream_t streamForGraph;
|
||||
int *A_d, *B_d, *C_d;
|
||||
@@ -55,57 +56,67 @@ TEST_CASE("Unit_hipGraphExecKernelNodeSetParams_Negative") {
|
||||
|
||||
HIP_CHECK(hipStreamCreate(&streamForGraph));
|
||||
HipTest::initArrays(&A_d, &B_d, &C_d, &A_h, &B_h, &C_h, N, false);
|
||||
unsigned blocks = HipTest::setNumBlocks(blocksPerCU, threadsPerBlock, N);
|
||||
|
||||
HIP_CHECK(hipGraphCreate(&graph, 0));
|
||||
HIP_CHECK(hipGraphAddMemcpyNode1D(&memcpyNode, graph, nullptr, 0, A_d, A_h,
|
||||
Nbytes, hipMemcpyHostToDevice));
|
||||
dependencies.push_back(memcpyNode);
|
||||
HIP_CHECK(hipGraphAddMemcpyNode1D(&memcpyNode, graph, nullptr, 0, B_d, B_h,
|
||||
Nbytes, hipMemcpyHostToDevice));
|
||||
dependencies.push_back(memcpyNode);
|
||||
|
||||
void* kernelArgs[] = {&A_d, &B_d, &C_d, reinterpret_cast<void *>(&NElem)};
|
||||
kNodeParams.func = reinterpret_cast<void *>(HipTest::vectorADD<int>);
|
||||
void* kernelArgs[] = {&A_d, &B_d, &C_d, reinterpret_cast<void*>(&NElem)};
|
||||
kNodeParams.func = reinterpret_cast<void*>(HipTest::vectorADD<int>);
|
||||
kNodeParams.gridDim = dim3(blocks);
|
||||
kNodeParams.blockDim = dim3(threadsPerBlock);
|
||||
kNodeParams.sharedMemBytes = 0;
|
||||
kNodeParams.kernelParams = reinterpret_cast<void**>(kernelArgs);
|
||||
kNodeParams.extra = nullptr;
|
||||
HIP_CHECK(hipGraphAddKernelNode(&kNode, graph, nullptr, 0, &kNodeParams));
|
||||
|
||||
hipGraphNode_t empty_node;
|
||||
HIP_CHECK(hipGraphAddEmptyNode(&empty_node, graph, &kNode, 1));
|
||||
|
||||
// Instantiate and launch the graph
|
||||
HIP_CHECK(hipGraphInstantiate(&graphExec, graph, NULL, NULL, 0));
|
||||
HIP_CHECK(hipGraphInstantiate(&graphExec, graph, nullptr, nullptr, 0));
|
||||
|
||||
SECTION("Pass hipGraphExec as nullptr") {
|
||||
ret = hipGraphExecKernelNodeSetParams(nullptr, kNode, &kNodeParams);
|
||||
REQUIRE(hipErrorInvalidValue == ret);
|
||||
HIP_CHECK_ERROR(hipGraphExecKernelNodeSetParams(nullptr, kNode, &kNodeParams),
|
||||
hipErrorInvalidValue);
|
||||
}
|
||||
|
||||
SECTION("Pass Node as nullptr") {
|
||||
ret = hipGraphExecKernelNodeSetParams(graphExec, nullptr, &kNodeParams);
|
||||
REQUIRE(hipErrorInvalidValue == ret);
|
||||
HIP_CHECK_ERROR(hipGraphExecKernelNodeSetParams(graphExec, nullptr, &kNodeParams),
|
||||
hipErrorInvalidValue);
|
||||
}
|
||||
|
||||
#if HT_AMD
|
||||
/* NodeParams null check is disabled on Nvedia as
|
||||
/* NodeParams null check is disabled on Nvidia as
|
||||
* this call gives SIGSEGV error in CUDA setup */
|
||||
SECTION("Pass NodeParams as nullptr") {
|
||||
ret = hipGraphExecKernelNodeSetParams(graphExec, kNode, nullptr);
|
||||
REQUIRE(hipErrorInvalidValue == ret);
|
||||
HIP_CHECK_ERROR(hipGraphExecKernelNodeSetParams(graphExec, kNode, nullptr),
|
||||
hipErrorInvalidValue);
|
||||
}
|
||||
#endif
|
||||
/* For below 2 scenarios -
|
||||
In AMD setup this API return - hipErrorInvalidValue and
|
||||
In CUDA setup this API return - hipErrorInvalidDeviceFunction
|
||||
As per Cuda spec API can only return "cudaSuccess, cudaErrorInvalidValue".
|
||||
*/
|
||||
SECTION("Pass NodeParams as un-initialized structure object") {
|
||||
hipKernelNodeParams kNodeParams1{};
|
||||
ret = hipGraphExecKernelNodeSetParams(graphExec, kNode, &kNodeParams1);
|
||||
REQUIRE(hipSuccess != ret);
|
||||
}
|
||||
SECTION("Pass NodeParams func datamember as nullptr") {
|
||||
|
||||
#if HT_NVIDIA // on AMD this returns hipErrorInvalidValue
|
||||
SECTION("Pass NodeParams func data member as nullptr") {
|
||||
kNodeParams.func = nullptr;
|
||||
ret = hipGraphExecKernelNodeSetParams(graphExec, kNode, &kNodeParams);
|
||||
REQUIRE(hipSuccess != ret);
|
||||
HIP_CHECK_ERROR(hipGraphExecKernelNodeSetParams(graphExec, kNode, &kNodeParams),
|
||||
hipErrorInvalidDeviceFunction);
|
||||
}
|
||||
#endif
|
||||
|
||||
#if HT_NVIDIA // segfaults on AMD
|
||||
SECTION("Pass kernelParams data member as nullptr") {
|
||||
kNodeParams.kernelParams = nullptr;
|
||||
HIP_CHECK_ERROR(hipGraphExecKernelNodeSetParams(graphExec, kNode, &kNodeParams),
|
||||
hipErrorInvalidValue);
|
||||
}
|
||||
#endif
|
||||
|
||||
#if HT_NVIDIA // segfaults on AMD
|
||||
SECTION("node is not a kernel node") {
|
||||
HIP_CHECK_ERROR(hipGraphExecKernelNodeSetParams(graphExec, empty_node, &kNodeParams),
|
||||
hipErrorInvalidValue);
|
||||
}
|
||||
#endif
|
||||
|
||||
SECTION("node is not instantiated") {
|
||||
HIP_CHECK(hipGraphAddKernelNode(&kNode, graph, nullptr, 0, &kNodeParams));
|
||||
HIP_CHECK_ERROR(hipGraphExecKernelNodeSetParams(graphExec, kNode, &kNodeParams),
|
||||
hipErrorInvalidValue);
|
||||
}
|
||||
|
||||
HipTest::freeArrays(A_d, B_d, C_d, A_h, B_h, C_h, false);
|
||||
@@ -114,16 +125,15 @@ TEST_CASE("Unit_hipGraphExecKernelNodeSetParams_Negative") {
|
||||
HIP_CHECK(hipStreamDestroy(streamForGraph));
|
||||
}
|
||||
|
||||
|
||||
/**
|
||||
* Functional Test for API Exec Kernel Params
|
||||
*/
|
||||
|
||||
TEST_CASE("Unit_hipGraphExecKernelNodeSetParams_Functional") {
|
||||
constexpr size_t N = 1024;
|
||||
constexpr size_t Nbytes = N * sizeof(int);
|
||||
constexpr auto blocksPerCU = 6; // to hide latency
|
||||
constexpr auto threadsPerBlock = 256;
|
||||
unsigned blocks = HipTest::setNumBlocks(blocksPerCU, threadsPerBlock, N);
|
||||
hipGraph_t graph;
|
||||
hipGraphNode_t memcpyNode, kNode;
|
||||
hipKernelNodeParams kNodeParams{}, kNodeParams1{};
|
||||
@@ -136,43 +146,36 @@ TEST_CASE("Unit_hipGraphExecKernelNodeSetParams_Functional") {
|
||||
|
||||
HIP_CHECK(hipStreamCreate(&streamForGraph));
|
||||
HipTest::initArrays(&A_d, &B_d, &C_d, &A_h, &B_h, &C_h, N, false);
|
||||
unsigned blocks = HipTest::setNumBlocks(blocksPerCU, threadsPerBlock, N);
|
||||
|
||||
HIP_CHECK(hipGraphCreate(&graph, 0));
|
||||
HIP_CHECK(hipGraphAddMemcpyNode1D(&memcpyNode, graph, nullptr, 0, A_d, A_h,
|
||||
Nbytes, hipMemcpyHostToDevice));
|
||||
|
||||
HIP_CHECK(hipGraphAddMemcpyNode1D(&memcpyNode, graph, nullptr, 0, A_d, A_h, Nbytes,
|
||||
hipMemcpyHostToDevice));
|
||||
dependencies.push_back(memcpyNode);
|
||||
HIP_CHECK(hipGraphAddMemcpyNode1D(&memcpyNode, graph, nullptr, 0, B_d, B_h,
|
||||
Nbytes, hipMemcpyHostToDevice));
|
||||
HIP_CHECK(hipGraphAddMemcpyNode1D(&memcpyNode, graph, nullptr, 0, B_d, B_h, Nbytes,
|
||||
hipMemcpyHostToDevice));
|
||||
dependencies.push_back(memcpyNode);
|
||||
|
||||
void* kernelArgs[] = {&A_d, &B_d, &C_d, reinterpret_cast<void *>(&NElem)};
|
||||
kNodeParams.func = reinterpret_cast<void *>(HipTest::vectorADD<int>);
|
||||
void* kernelArgs[] = {&A_d, &B_d, &C_d, reinterpret_cast<void*>(&NElem)};
|
||||
kNodeParams.func = reinterpret_cast<void*>(HipTest::vectorADD<int>);
|
||||
kNodeParams.gridDim = dim3(blocks);
|
||||
kNodeParams.blockDim = dim3(threadsPerBlock);
|
||||
kNodeParams.sharedMemBytes = 0;
|
||||
kNodeParams.kernelParams = reinterpret_cast<void**>(kernelArgs);
|
||||
kNodeParams.extra = nullptr;
|
||||
HIP_CHECK(hipGraphAddKernelNode(&kNode, graph, dependencies.data(),
|
||||
dependencies.size(), &kNodeParams));
|
||||
HIP_CHECK(
|
||||
hipGraphAddKernelNode(&kNode, graph, dependencies.data(), dependencies.size(), &kNodeParams));
|
||||
|
||||
memset(&kNodeParams1, 0, sizeof(kNodeParams1));
|
||||
kNodeParams1.func = reinterpret_cast<void *>(HipTest::vectorSUB<int>);
|
||||
kNodeParams1.func = reinterpret_cast<void*>(HipTest::vectorSUB<int>);
|
||||
kNodeParams1.gridDim = dim3(blocks);
|
||||
kNodeParams1.blockDim = dim3(threadsPerBlock);
|
||||
kNodeParams1.sharedMemBytes = 0;
|
||||
kNodeParams1.kernelParams = reinterpret_cast<void**>(kernelArgs);
|
||||
kNodeParams1.extra = nullptr;
|
||||
|
||||
dependencies.clear();
|
||||
dependencies.push_back(kNode);
|
||||
HIP_CHECK(hipGraphAddMemcpyNode1D(&memcpyNode, graph, dependencies.data(),
|
||||
dependencies.size(), C_h, C_d,
|
||||
Nbytes, hipMemcpyDeviceToHost));
|
||||
HIP_CHECK(hipGraphAddMemcpyNode1D(&memcpyNode, graph, dependencies.data(), dependencies.size(),
|
||||
C_h, C_d, Nbytes, hipMemcpyDeviceToHost));
|
||||
|
||||
// Instantiate and launch the graph
|
||||
HIP_CHECK(hipGraphInstantiate(&graphExec, graph, NULL, NULL, 0));
|
||||
REQUIRE(hipSuccess == hipGraphExecKernelNodeSetParams(graphExec, kNode,
|
||||
&kNodeParams1));
|
||||
HIP_CHECK(hipGraphExecKernelNodeSetParams(graphExec, kNode, &kNodeParams1));
|
||||
HIP_CHECK(hipGraphLaunch(graphExec, streamForGraph));
|
||||
HIP_CHECK(hipStreamSynchronize(streamForGraph));
|
||||
|
||||
|
||||
새 이슈에서 참조
사용자 차단