AI/rocm-systems
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Merge 'develop' into 'amd-staging'

Browse Source 
Change-Id: I97006cfdc6f8fc70f0eabc388117920eb12028d6
This commit is contained in:
Jenkins
2023-05-03 11:10:58 +00:00
parent 3c958a5022 b5d2ea28e1
commit fad814bed8
9 changed files with 283 additions and 290 deletions
Download Patch File Download Diff File Expand all files Collapse all files
.jenkins/jenkinsfile
+23 -33
View File
@@ -2,53 +2,41 @@ def hipBuildTest(String backendLabel) {
node(backendLabel) {
stage("SYNC - ${backendLabel}") {
// Checkout hip repository with the PR patch
// Checkout hip-tests repository with the PR patch
dir("${WORKSPACE}/hip-tests") {
checkout scm
env.HIP_TESTS_DIR = "${WORKSPACE}" + "/hip-tests"
}
// Clone HIP repository
// Clone hip repository
dir("${WORKSPACE}/hip") {
git branch: 'develop',
url: 'https://github.com/ROCm-Developer-Tools/HIP'
url: 'https://github.com/ROCm-Developer-Tools/hip'
env.HIP_DIR = "${WORKSPACE}" + "/hip"
}
// Clone hipamd repository
dir("${WORKSPACE}/hipamd") {
// Clone clr repository
dir("${WORKSPACE}/clr") {
git branch: 'develop',
url: 'https://github.com/ROCm-Developer-Tools/hipamd'
env.HIPAMD_DIR = "${WORKSPACE}" + "/hipamd"
}
// Clone vdi and opencl for only amd backend server
if (backendLabel =~ /.*amd.*/) {
dir("${WORKSPACE}/ROCm-OpenCL-Runtime") {
git branch:'develop',
url: 'https://github.com/RadeonOpenCompute/ROCm-OpenCL-Runtime'
env.OPENCL_DIR = "${WORKSPACE}" + "/ROCm-OpenCL-Runtime"
}
dir("${WORKSPACE}/ROCclr") {
git branch:'develop',
url: 'https://github.com/ROCm-Developer-Tools/ROCclr'
env.ROCclr_DIR = "${WORKSPACE}" + "/ROCclr"
}
credentialsId: 'branch-credentials',
url: 'https://github.com/ROCm-Developer-Tools/clr'
env.CLR_DIR = "${WORKSPACE}" + "/clr"
}
}
stage("BUILD HIP - ${backendLabel}") {
// Running the build on hipamd workspace
dir("${WORKSPACE}/hipamd") {
// Running the build on clr workspace
dir("${WORKSPACE}/clr") {
sh """#!/usr/bin/env bash
set -x
rm -rf build
mkdir -p build
cd build
# Check if backend label contains string "amd" or backend host is a server with amd gpu
if [[ $backendLabel =~ amd ]]; then
cmake -DHIP_PATH=\$PWD/install -DHIP_COMMON_DIR=\$HIP_DIR -DAMD_OPENCL_PATH=\$OPENCL_DIR -DROCCLR_PATH=\$ROCclr_DIR -DCMAKE_PREFIX_PATH="/opt/rocm/" -DCMAKE_INSTALL_PREFIX=\$PWD/install ..
cmake -DCLR_BUILD_HIP=ON -DHIP_PATH=\$PWD/install -DHIP_COMMON_DIR=\$HIP_DIR -DCMAKE_PREFIX_PATH="/opt/rocm/" -DCMAKE_INSTALL_PREFIX=\$PWD/install ..
else
cmake -DHIP_PLATFORM=nvidia -DHIP_COMMON_DIR=\$HIP_DIR -DCMAKE_INSTALL_PREFIX=\$PWD/install ..
cmake -DCLR_BUILD_HIP=ON -DHIP_PLATFORM=nvidia -DHIP_COMMON_DIR=\$HIP_DIR -DCMAKE_INSTALL_PREFIX=\$PWD/install ..
fi
make -j\$(nproc)
make install -j\$(nproc)
@@ -59,7 +47,7 @@ def hipBuildTest(String backendLabel) {
stage("BUILD HIP TESTS - ${backendLabel}") {
// Running the build on HIP TESTS workspace
dir("${WORKSPACE}/hip-tests") {
env.HIP_PATH = "${HIPAMD_DIR}" + "/build/install"
env.HIP_PATH = "${CLR_DIR}" + "/build/install"
sh """#!/usr/bin/env bash
set -x
rm -rf build
@@ -78,17 +66,19 @@ def hipBuildTest(String backendLabel) {
}
}
stage("TEST - ${backendLabel}") {
dir("${WORKSPACE}/hip-tests") {
sh """#!/usr/bin/env bash
set -x
cd build
timeout(time: 1, unit: 'HOURS') {
stage("TEST - ${backendLabel}") {
dir("${WORKSPACE}/hip-tests") {
sh """#!/usr/bin/env bash
set -x
cd build
if [[ $backendLabel =~ amd ]]; then
ctest
ctest --overwrite BuildDirectory=. --output-junit hiptest_output_catch_amd.xml
else
ctest -E 'Unit_hipMemcpyHtoD_Positive_Synchronization_Behavior|Unit_hipMemcpy_Positive_Synchronization_Behavior|Unit_hipFreeNegativeHost'
ctest --overwrite BuildDirectory=. --output-junit hiptest_output_catch_nvidia.xml -E 'Unit_hipMemcpyHtoD_Positive_Synchronization_Behavior|Unit_hipMemcpy_Positive_Synchronization_Behavior|Unit_hipFreeNegativeHost'
fi
"""
}
}
}
}
catch/hipTestMain/config/config_amd_linux_common.json
+14 -2
View File
@@ -55,9 +55,21 @@
"Unit_hipDeviceSetSharedMemConfig_Negative_Parameters",
"Disabling tests tracked with SWDEV-394083",
"Unit_hipDeviceSynchronize_Positive_Nullstream",
"Disabling tests tracked with SWDEV-393637",
"Disabling tests tracked with SWDEV-393637",
"Unit_hipDeviceSynchronize_Functional",
"Unit_hipMemset3DSync",
"Unit_hipStreamAddCallback_StrmSyncTiming"
"Unit_hipStreamAddCallback_StrmSyncTiming",
"Disabling test tracked SWDEV-394199",
"Unit_hipStreamCreateWithPriority_MulthreadNonblockingflag",
"Disabling test tracked SWDEV-395683",
"Unit_hipStreamPerThread_MultiThread",
"Disabling tests tracked with SWDEV-389647..",
"Unit_hipMemcpy2DToArrayAsync_Positive_Synchronization_Behavior",
"Disabling test tracked SWDEV-391555",
"Unit_hipMemcpyPeer_Positive_ZeroSize",
"Unit_hipMemcpyPeerAsync_Positive_ZeroSize",
"Disabling test tracked SWDEV-391718",
"Unit_hipMemRangeGetAttribute_TstCountParam"
]
}
catch/hipTestMain/config/config_amd_windows_common.json
+16 -1
View File
@@ -147,6 +147,21 @@
"Unit_hipGraphMemsetNodeSetParams_Positive_Basic - uint8_t",
"Unit_hipStreamWaitEvent_UninitializedStream_Negative",
"Unit_hipDeviceSetSharedMemConfig_Negative_Parameters",
"Unit_hipDeviceGetUuid_Positive"
"Unit_hipDeviceGetUuid_Positive",
"Disabling test tracked SWDEV-394199",
"Unit_hipStreamCreateWithPriority_MulthreadNonblockingflag",
"SWDEV-396618 hipEventElapsedTime returns sucess",
"Unit_hipEventElapsedTime_NotReady_Negative",
"SWDEV-396617 ExecMemcpyNodeSetParamsFromSymbol fails in direction",
"Unit_hipGraphExecMemcpyNodeSetParamsFromSymbol_Negative_Parameters",
"SWDEV-396616 hipMemMap returns invalid error",
"Unit_hipMemVmm_Basic",
"SWDEV-396615 mGPUs not considered correctly",
"Unit_hipManagedKeyword_MultiGpu",
"Disabling test tracked SWDEV-391555",
"Unit_hipMemcpyPeer_Positive_ZeroSize",
"Unit_hipMemcpyPeerAsync_Positive_ZeroSize",
"Disabling test tracked SWDEV-391718",
"Unit_hipMemRangeGetAttribute_TstCountParam"
]
}
catch/unit/device/hipDeviceSetGetSharedMemConfig.cc
-5
View File
@@ -43,13 +43,8 @@ TEST_CASE("Unit_hipDeviceSetSharedMemConfig_Positive_Basic") {
}
TEST_CASE("Unit_hipDeviceSetSharedMemConfig_Negative_Parameters") {
#if HT_AMD
HIP_CHECK_ERROR(hipDeviceSetSharedMemConfig(static_cast<hipSharedMemConfig>(-1)),
hipErrorNotSupported);
#elif HT_NVIDIA
HIP_CHECK_ERROR(hipDeviceSetSharedMemConfig(static_cast<hipSharedMemConfig>(-1)),
hipErrorInvalidValue);
#endif
}
TEST_CASE("Unit_hipDeviceGetSharedMemConfig_Positive_Default") {
catch/unit/graph/hipGraphInstantiateWithFlags.cc
+47 -5
View File
@@ -1,5 +1,5 @@
/*
Copyright (c) 2022 Advanced Micro Devices, Inc. All rights reserved.
Copyright (c) 2023 Advanced Micro Devices, Inc. All rights reserved.
Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
@@ -20,20 +20,16 @@ THE SOFTWARE.
/*
hipGraphInstantiateWithFlags(hipGraphExec_t* pGraphExec, hipGraph_t graph, unsigned long long flags);
Testcase Scenarios of hipGraphInstantiateWithFlags API:
Negative:
1) Pass nullptr to pGraphExec
2) Pass nullptr to graph
4) Pass invalid flag
Functional:
1) Create dependencies graph and instantiate the graph
2) Create graph in one GPU device and instantiate, launch in peer GPU device
3) Create stream capture graph and instantite the graph
4) Create stream capture graph in one GPU device and instantite the graph launch
in peer GPU device
Mapping is missing for NVIDIA platform hence skipping the testcases
*/
@@ -311,3 +307,49 @@ TEST_CASE("Unit_hipGraphInstantiateWithFlags_StreamCaptureDeviceContextChg") {
SUCCEED("skipped the testcase as no of devices is less than 2");
}
}
/* Create graph and add memAlloc node, but no corresponding memFree node to it.
Instantiate graph with flag - hipGraphInstantiateFlagAutoFreeOnLaunch
Launch and check graph execution should work properly and
free memory allocated by memAlloc call manually using hipFree api.
Note - This test case is just to check if hipGraphInstantiateFlagAutoFreeOnLaunch
is not resulting in compilation error or api failure. Real functional test
will be added once the feature is fully implemented.
*/
TEST_CASE("Unit_hipGraphInstantiateWithFlags_FlagAutoFreeOnLaunch_check") {
constexpr size_t size = 512 * 1024 * 1024;
constexpr size_t Nbytes = size * sizeof(int);
hipGraph_t graph;
hipGraphExec_t graphExec;
hipStream_t stream;
hipGraphNode_t allocNodeA;
hipMemAllocNodeParams allocParam;
HIP_CHECK(hipGraphCreate(&graph, 0));
HIP_CHECK(hipStreamCreate(&stream));
memset(&allocParam, 0, sizeof(allocParam));
allocParam.bytesize = Nbytes;
allocParam.poolProps.allocType = hipMemAllocationTypePinned;
allocParam.poolProps.location.id = 0;
allocParam.poolProps.location.type = hipMemLocationTypeDevice;
HIP_CHECK(hipGraphAddMemAllocNode(&allocNodeA, graph, nullptr,
0, &allocParam));
REQUIRE(allocParam.dptr != nullptr);
int *A_d = reinterpret_cast<int *>(allocParam.dptr);
// Instantiate with Flag and launch the graph
HIP_CHECK(hipGraphInstantiateWithFlags(&graphExec, graph,
hipGraphInstantiateFlagAutoFreeOnLaunch));
HIP_CHECK(hipGraphLaunch(graphExec, stream));
HIP_CHECK(hipStreamSynchronize(stream));
HIP_CHECK(hipFree(A_d)); // free allocMemory manually
HIP_CHECK(hipGraphDestroy(graph));
HIP_CHECK(hipGraphExecDestroy(graphExec));
HIP_CHECK(hipStreamDestroy(stream));
}
catch/unit/memory/CMakeLists.txt
-1
View File
@@ -51,7 +51,6 @@ set(TEST_SRC
hipMallocManaged_MultiScenario.cc
hipMemsetNegative.cc
hipMemset.cc
hipMemsetAsyncMultiThread.cc
hipMemset3D.cc
hipMemset2D.cc
hipHostMallocTests.cc
catch/unit/memory/hipMemsetAsyncMultiThread.cc
-243
View File
@@ -1,243 +0,0 @@
/*
* Copyright (c) 2021 Advanced Micro Devices, Inc. All rights reserved.
* Permission is hereby granted, free of charge, to any person obtaining a copy
* of this software and associated documentation files (the "Software"), to deal
* 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
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNNESS 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
* THE SOFTWARE.
*/
/*
* Test that validates functionality of hipmemsetAsync apis over multi threads
*/
#include <hip_test_common.hh>
#include <hip_test_checkers.hh>
#define NUM_THREADS 20
#define ITER 10
#define N (4*1024*1024)
template <typename T>
class MemSetAsyncMthreadTest {
public:
T *A_h, *A_d, *B_h;
T memSetVal;
size_t Nbytes;
bool testResult = true;
int validateCount = 0;
hipStream_t stream;
void memAllocate(T memSetValue) {
memSetVal = memSetValue;
Nbytes = N * sizeof(T);
A_h = reinterpret_cast<T*>(malloc(Nbytes));
HIP_ASSERT(A_h != nullptr);
HIP_CHECK(hipMalloc(&A_d, Nbytes));
B_h = reinterpret_cast<T*>(malloc(Nbytes));
HIP_ASSERT(B_h != nullptr);
HIP_CHECK(hipStreamCreate(&stream));
}
void threadCompleteStatus() {
for (int k = 0 ; k < N ; k++) {
if ((A_h[k] == memSetVal) && (B_h[k] == memSetVal)) {
validateCount+= 1;
}
}
}
bool resultAfterAllIterations() {
memDeallocate();
testResult = (validateCount == (ITER * N)) ? true: false;
return testResult;
}
void memDeallocate() {
HIP_CHECK(hipFree(A_d));
free(A_h);
free(B_h);
HIP_CHECK(hipStreamDestroy(stream));
}
};
template <typename T>
void queueJobsForhipMemsetAsync(T* A_d, T* A_h, T memSetVal, size_t Nbytes,
hipStream_t stream) {
HIPCHECK(hipMemsetAsync(A_d, memSetVal, N, stream));
HIPCHECK(hipMemcpyAsync(A_h, A_d, Nbytes, hipMemcpyDeviceToHost, stream));
}
template <typename T>
void queueJobsForhipMemsetD32Async(T* A_d, T* A_h, T memSetVal, size_t Nbytes,
hipStream_t stream) {
HIPCHECK(hipMemsetD32Async((hipDeviceptr_t)A_d, memSetVal, N, stream));
HIPCHECK(hipMemcpyAsync(A_h, A_d, Nbytes, hipMemcpyDeviceToHost, stream));
}
template <typename T>
void queueJobsForhipMemsetD16Async(T* A_d, T* A_h, T memSetVal, size_t Nbytes,
hipStream_t stream) {
HIPCHECK(hipMemsetD16Async((hipDeviceptr_t)A_d, memSetVal, N, stream));
HIPCHECK(hipMemcpyAsync(A_h, A_d, Nbytes, hipMemcpyDeviceToHost, stream));
}
template <typename T>
void queueJobsForhipMemsetD8Async(T* A_d, T* A_h, T memSetVal, size_t Nbytes,
hipStream_t stream) {
HIPCHECK(hipMemsetD8Async((hipDeviceptr_t)A_d, memSetVal, N, stream));
HIPCHECK(hipMemcpyAsync(A_h, A_d, Nbytes, hipMemcpyDeviceToHost, stream));
}
/* Queue hipMemsetAsync jobs on multiple threads and verify they all
* finished on all threads successfully
*/
bool testhipMemsetAsyncWithMultiThread() {
MemSetAsyncMthreadTest <char> obj;
constexpr char memsetval = 0x42;
obj.memAllocate(memsetval);
std::thread t[NUM_THREADS];
for (int i = 0 ; i < ITER ; i++) {
for (int k = 0 ; k < NUM_THREADS ; k++) {
if (k%2) {
t[k] = std::thread(queueJobsForhipMemsetAsync<char>, obj.A_d, obj.A_h,
obj.memSetVal, obj.Nbytes, obj.stream);
} else {
t[k] = std::thread(queueJobsForhipMemsetAsync<char>, obj.A_d, obj.B_h,
obj.memSetVal, obj.Nbytes, obj.stream);
}
}
for (int j = 0 ; j < NUM_THREADS ; j++) {
t[j].join();
}
HIP_CHECK(hipStreamSynchronize(obj.stream));
obj.threadCompleteStatus();
}
return obj.resultAfterAllIterations();
}
bool testhipMemsetD32AsyncWithMultiThread() {
MemSetAsyncMthreadTest <int32_t> obj;
constexpr int memsetD32val = 0xDEADBEEF;
obj.memAllocate(memsetD32val);
std::thread t[NUM_THREADS];
for (int i = 0 ; i < ITER ; i++) {
for (int k = 0 ; k < NUM_THREADS ; k++) {
if (k%2) {
t[k] = std::thread(queueJobsForhipMemsetD32Async<int32_t>, obj.A_d,
obj.A_h, obj.memSetVal, obj.Nbytes, obj.stream);
} else {
t[k] = std::thread(queueJobsForhipMemsetD32Async<int32_t>, obj.A_d,
obj.B_h, obj.memSetVal, obj.Nbytes, obj.stream);
}
}
for (int j = 0 ; j < NUM_THREADS ; j++) {
t[j].join();
}
HIP_CHECK(hipStreamSynchronize(obj.stream));
obj.threadCompleteStatus();
}
return obj.resultAfterAllIterations();
}
bool testhipMemsetD16AsyncWithMultiThread() {
MemSetAsyncMthreadTest <int16_t> obj;
constexpr int16_t memsetD16val = 0xDEAD;
obj.memAllocate(memsetD16val);
std::thread t[NUM_THREADS];
for (int i = 0 ; i < ITER ; i++) {
for (int k = 0 ; k < NUM_THREADS ; k++) {
if (k%2) {
t[k] = std::thread(queueJobsForhipMemsetD16Async<int16_t>, obj.A_d,
obj.A_h, obj.memSetVal, obj.Nbytes, obj.stream);
} else {
t[k] = std::thread(queueJobsForhipMemsetD16Async<int16_t>, obj.A_d,
obj.B_h, obj.memSetVal, obj.Nbytes, obj.stream);
}
}
for (int j = 0 ; j < NUM_THREADS ; j++) {
t[j].join();
}
HIP_CHECK(hipStreamSynchronize(obj.stream));
obj.threadCompleteStatus();
}
return obj.resultAfterAllIterations();
}
bool testhipMemsetD8AsyncWithMultiThread() {
MemSetAsyncMthreadTest <char> obj;
constexpr char memsetD8val = 0xDE;
obj.memAllocate(memsetD8val);
std::thread t[NUM_THREADS];
for (int i = 0 ; i < ITER ; i++) {
for (int k = 0 ; k < NUM_THREADS ; k++) {
if (k%2) {
t[k] = std::thread(queueJobsForhipMemsetD8Async<char>, obj.A_d,
obj.A_h, obj.memSetVal, obj.Nbytes, obj.stream);
} else {
t[k] = std::thread(queueJobsForhipMemsetD8Async<char>, obj.A_d,
obj.B_h, obj.memSetVal, obj.Nbytes, obj.stream);
}
}
for (int j = 0 ; j < NUM_THREADS ; j++) {
t[j].join();
}
HIP_CHECK(hipStreamSynchronize(obj.stream));
obj.threadCompleteStatus();
}
return obj.resultAfterAllIterations();
}
/*
* Test that validates functionality of hipmemsetAsync apis over multi threads
*/
TEST_CASE("Unit_hipMemsetAsync_QueueJobsMultithreaded") {
bool ret;
SECTION("hipMemsetAsync With MultiThread") {
ret = testhipMemsetAsyncWithMultiThread();
REQUIRE(ret == true);
}
SECTION("hipMemsetD32Async With MultiThread") {
ret = testhipMemsetD32AsyncWithMultiThread();
REQUIRE(ret == true);
}
SECTION("hipMemsetD16Async With MultiThread") {
ret = testhipMemsetD16AsyncWithMultiThread();
REQUIRE(ret == true);
}
SECTION("hipMemsetD8Async With MultiThread") {
ret = testhipMemsetD8AsyncWithMultiThread();
REQUIRE(ret == true);
}
}
catch/unit/multiThread/CMakeLists.txt
+1
View File
@@ -1,5 +1,6 @@
# Common Tests - Test independent of all platforms
set(TEST_SRC
hipMemsetAsyncMultiThread.cc
hipMultiThreadDevice.cc
hipMultiThreadStreams1.cc
hipMultiThreadStreams2.cc
catch/unit/multiThread/hipMemsetAsyncMultiThread.cc
+182
View File
@@ -0,0 +1,182 @@
/*
* Copyright (c) 2023 Advanced Micro Devices, Inc. All rights reserved.
* Permission is hereby granted, free of charge, to any person obtaining a copy
* of this software and associated documentation files (the "Software"), to deal
* 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
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNNESS 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 THE SOFTWARE.
*/
#include <hip_test_checkers.hh>
#include <hip_test_common.hh>
#include <algorithm>
#include <limits>
#include <memory>
#include <random>
#include <type_traits>
#include <utility>
#include <vector>
enum TestType { SameStream = 0, DifferentStreams };
// Func -> Memset/MemsetD[8/16/32], T - type of data to be worked on
// basically each thread memsets value present in input one by one on its
// allocated buffer
template <typename Func, typename T>
void threadCall(Func f, hipStream_t stream) {
// Should match hipMemsetAsync or hipMemsetD*Async arguments
static_assert(
(std::is_same<Func, hipError_t (*)(void *, int, size_t, hipStream_t)>::
value || // hipMemsetAsync
std::is_same<Func,
hipError_t (*)(hipDeviceptr_t, int, size_t,
hipStream_t)>::value || // hipMemsetD32Async
std::is_same<Func,
hipError_t (*)(hipDeviceptr_t, unsigned short, size_t,
hipStream_t)>::value || // hipMemsetD16Async
std::is_same<Func,
hipError_t (*)(hipDeviceptr_t, unsigned char, size_t,
hipStream_t)>::value) && // hipMemsetD8Async
"Func f should be hipMemsetAsync or hipMemsetD*Async");
constexpr bool cast_2_void =
std::is_same<Func,
hipError_t (*)(void *, int, size_t, hipStream_t)>::value;
// Use the unsiged type, since memset concerns with set bit values over a mem
// address
typedef typename std::make_unsigned<T>::type unsigned_t;
unsigned_t min = 0;
unsigned_t max = std::numeric_limits<unsigned_t>::max();
std::mt19937_64 engine(std::random_device{}());
auto distribution = std::uniform_int_distribution<unsigned int>(
min, max); // this needs to be unsigned because windows does not treats
// char as numeric types
T *ptr{nullptr};
constexpr size_t size = 1024;
constexpr size_t iter = 1024;
HIP_CHECK_THREAD(hipMalloc(&ptr, sizeof(T) * size));
hipEvent_t event{};
HIP_CHECK_THREAD(hipEventCreate(&event));
union overlay_val_t {
T t_val;
unsigned_t u_val;
} overlay_val;
std::vector<T> dst(size, 0);
for (size_t i = 0; i < iter; i++) {
overlay_val.u_val = static_cast<unsigned_t>(
distribution(engine)); // generate an unsigned int number
if constexpr (cast_2_void) {
HIP_CHECK_THREAD(f((void *)ptr, overlay_val.t_val, size, stream));
} else {
HIP_CHECK_THREAD(
f(*(hipDeviceptr_t *)&ptr, overlay_val.t_val, size, stream));
}
HIP_CHECK_THREAD(hipMemcpyAsync(dst.data(), ptr, size * sizeof(T),
hipMemcpyDeviceToHost, stream));
HIP_CHECK_THREAD(hipEventRecord(event, stream));
HIP_CHECK_THREAD(
hipStreamWaitEvent(stream, event, 0)); // wait till memcpy is done
REQUIRE_THREAD(std::all_of(dst.begin(), dst.end(), [&](T v) {
// If this test ever fails, add prints here on mismatch
return v == overlay_val.t_val;
}));
}
HIP_CHECK_THREAD(hipEventDestroy(event));
HIP_CHECK_THREAD(hipFree(ptr));
}
// Func -> Memset/MemsetD[8/16/32], T - type of data to be worked on
template <typename Func, typename T> void launchThreads(Func f, TestType type) {
static_assert(!std::is_pointer<T>::value && "Argument cant be a pointer");
// Should match hipMemsetAsync or hipMemsetD*Async arguments
static_assert(
(std::is_same<Func, hipError_t (*)(void *, int, size_t, hipStream_t)>::
value || // hipMemsetAsync
std::is_same<Func,
hipError_t (*)(hipDeviceptr_t, int, size_t,
hipStream_t)>::value || // hipMemsetD32Async
std::is_same<Func,
hipError_t (*)(hipDeviceptr_t, unsigned short, size_t,
hipStream_t)>::value || // hipMemsetD16Async
std::is_same<Func,
hipError_t (*)(hipDeviceptr_t, unsigned char, size_t,
hipStream_t)>::value) && // hipMemsetD8Async
"Func f should be hipMemsetAsync or hipMemsetD*Async");
const size_t num_threads =
(std::thread::hardware_concurrency() > 8)
? (((std::thread::hardware_concurrency() / 4) >= 127)
? 127
: (std::thread::hardware_concurrency() / 4))
: 2; // thread count between 2 - 127
const size_t num_streams = (type == SameStream) ? 1 : num_threads;
std::vector<hipStream_t> streams(num_streams, nullptr);
for (size_t i = 0; i < num_streams; i++) {
HIP_CHECK(hipStreamCreate(&streams[i]));
REQUIRE(streams[i] != nullptr);
}
std::vector<std::thread> thread_pool;
thread_pool.reserve(num_threads);
auto thread_func = threadCall<Func, T>;
for (size_t i = 0; i < num_threads; i++) {
auto stream = (type == SameStream) ? streams[0] : streams[i];
thread_pool.emplace_back(std::thread(thread_func, f, stream));
}
for (size_t i = 0; i < num_threads; i++) {
thread_pool[i].join();
}
HIP_CHECK_THREAD_FINALIZE(); // Make sure all thread have exited properly
for (size_t i = 0; i < num_streams; i++) {
HIP_CHECK(hipStreamDestroy(streams[i]));
}
}
TEST_CASE("Unit_hipMemsetAsync_QueueJobsMultithreaded") {
using hipMemsetAsync_t =
hipError_t (*)(void *, int, const size_t, hipStream_t);
using hipMemsetAsyncD8_t =
hipError_t (*)(hipDeviceptr_t, unsigned char, const size_t, hipStream_t);
using hipMemsetAsyncD16_t =
hipError_t (*)(hipDeviceptr_t, unsigned short, const size_t, hipStream_t);
using hipMemsetAsyncD32_t =
hipError_t (*)(hipDeviceptr_t, int, const size_t, hipStream_t);
launchThreads<hipMemsetAsync_t, char>(hipMemsetAsync, SameStream);
launchThreads<hipMemsetAsync_t, char>(hipMemsetAsync, DifferentStreams);
launchThreads<hipMemsetAsyncD8_t, unsigned char>(hipMemsetD8Async,
SameStream);
launchThreads<hipMemsetAsyncD8_t, unsigned char>(hipMemsetD8Async,
DifferentStreams);
launchThreads<hipMemsetAsyncD16_t, unsigned short>(hipMemsetD16Async,
SameStream);
launchThreads<hipMemsetAsyncD16_t, unsigned short>(hipMemsetD16Async,
DifferentStreams);
launchThreads<hipMemsetAsyncD32_t, int>(hipMemsetD32Async, SameStream);
launchThreads<hipMemsetAsyncD32_t, int>(hipMemsetD32Async, DifferentStreams);
}