SWDEV-470698 - fix formatting, add format check workflow (#657)
Cette révision appartient à :
Danylo Lytovchenko
révisé par
GitHub
GitHub
Parent
5840940caa
révision
f7338717ae
@@ -35,8 +35,7 @@ std::condition_variable cv;
|
||||
* unsigned int flags);`
|
||||
* - Create an instance of userObject to manage lifetime of a resource.
|
||||
*/
|
||||
template<typename T>
|
||||
__global__ void KernelFn(T *Ad, int clockrate, int WaitSecs) {
|
||||
template <typename T> __global__ void KernelFn(T* Ad, int clockrate, int WaitSecs) {
|
||||
uint64_t num_cycles = (uint64_t)clockrate;
|
||||
num_cycles = num_cycles * 1000 * WaitSecs;
|
||||
uint64_t start = clock64(), cycles = 0;
|
||||
@@ -53,7 +52,7 @@ __global__ void KernelFn(T *Ad, int clockrate, int WaitSecs) {
|
||||
}
|
||||
void threadFunc_dltMemory() {
|
||||
std::unique_lock lk(m);
|
||||
cv.wait(lk, []{ return setVar; });
|
||||
cv.wait(lk, [] { return setVar; });
|
||||
REQUIRE(globalPtr != nullptr);
|
||||
HIP_CHECK(hipHostFree(globalPtr));
|
||||
setVar = false;
|
||||
@@ -64,31 +63,25 @@ void destroyPinnedObj(void* ptr) {
|
||||
cv.notify_one();
|
||||
}
|
||||
template <typename T>
|
||||
void hipUserObjectCreate_int_float_Objects(T* hostArr,
|
||||
T* devArr, void destroyObj(void*)) {
|
||||
void hipUserObjectCreate_int_float_Objects(T* hostArr, T* devArr, void destroyObj(void*)) {
|
||||
int clockrate = 0;
|
||||
HIP_CHECK(hipDeviceGetAttribute(&clockrate,
|
||||
hipDeviceAttributeMemoryClockRate, 0));
|
||||
HIP_CHECK(hipDeviceGetAttribute(&clockrate, hipDeviceAttributeMemoryClockRate, 0));
|
||||
hipGraph_t graph = nullptr;
|
||||
hipStream_t stream;
|
||||
HIP_CHECK(hipStreamCreate(&stream));
|
||||
HIP_CHECK(hipStreamBeginCapture(stream,
|
||||
hipStreamCaptureModeGlobal));
|
||||
HIP_CHECK(hipMemcpyAsync(devArr, hostArr, sizeof(int),
|
||||
hipMemcpyHostToDevice, stream));
|
||||
KernelFn<<< 1, 1, 0, stream>>>(devArr, clockrate, 5);
|
||||
HIP_CHECK(hipMemcpyAsync(hostArr, devArr, sizeof(int),
|
||||
hipMemcpyDeviceToHost, stream));
|
||||
HIP_CHECK(hipStreamBeginCapture(stream, hipStreamCaptureModeGlobal));
|
||||
HIP_CHECK(hipMemcpyAsync(devArr, hostArr, sizeof(int), hipMemcpyHostToDevice, stream));
|
||||
KernelFn<<<1, 1, 0, stream>>>(devArr, clockrate, 5);
|
||||
HIP_CHECK(hipMemcpyAsync(hostArr, devArr, sizeof(int), hipMemcpyDeviceToHost, stream));
|
||||
HIP_CHECK(hipStreamEndCapture(stream, &graph));
|
||||
REQUIRE(graph != nullptr);
|
||||
hipUserObject_t Uobj;
|
||||
int refCount = 1;
|
||||
HIP_CHECK(hipUserObjectCreate(&Uobj, hostArr, destroyObj, refCount,
|
||||
hipUserObjectNoDestructorSync));
|
||||
HIP_CHECK(
|
||||
hipUserObjectCreate(&Uobj, hostArr, destroyObj, refCount, hipUserObjectNoDestructorSync));
|
||||
HIP_CHECK(hipGraphRetainUserObject(graph, Uobj, refCount, 0));
|
||||
hipGraphExec_t graph_instance;
|
||||
HIP_CHECK(hipGraphInstantiate(&graph_instance, graph,
|
||||
nullptr, nullptr, 0));
|
||||
HIP_CHECK(hipGraphInstantiate(&graph_instance, graph, nullptr, nullptr, 0));
|
||||
HIP_CHECK(hipGraphDestroy(graph));
|
||||
SECTION("graph_instance is destroyed before async launch completes") {
|
||||
HIP_CHECK(hipGraphLaunch(graph_instance, stream));
|
||||
@@ -124,11 +117,11 @@ void hipUserObjectCreate_int_float_Objects(T* hostArr,
|
||||
TEST_CASE("Unit_hipGraphUserObj_Int_float_Objects") {
|
||||
SECTION("Called with Int Obj") {
|
||||
std::thread t1(threadFunc_dltMemory);
|
||||
int *hostArr = nullptr;
|
||||
int* hostArr = nullptr;
|
||||
HIP_CHECK(hipHostMalloc(&hostArr, sizeof(int)));
|
||||
REQUIRE(hostArr != nullptr);
|
||||
*hostArr = 1111;
|
||||
int *devArr = nullptr;
|
||||
int* devArr = nullptr;
|
||||
HIP_CHECK(hipMalloc(&devArr, sizeof(int)));
|
||||
REQUIRE(devArr != nullptr);
|
||||
hipUserObjectCreate_int_float_Objects(hostArr, devArr, destroyPinnedObj);
|
||||
@@ -137,11 +130,11 @@ TEST_CASE("Unit_hipGraphUserObj_Int_float_Objects") {
|
||||
}
|
||||
SECTION("Called with float Obj") {
|
||||
std::thread t1(threadFunc_dltMemory);
|
||||
float *hostArr = nullptr;
|
||||
float* hostArr = nullptr;
|
||||
HIP_CHECK(hipHostMalloc(&hostArr, sizeof(float)));
|
||||
REQUIRE(hostArr != nullptr);
|
||||
*hostArr = 1111.0f;
|
||||
float *devArr = nullptr;
|
||||
float* devArr = nullptr;
|
||||
HIP_CHECK(hipMalloc(&devArr, sizeof(float)));
|
||||
REQUIRE(devArr != nullptr);
|
||||
hipUserObjectCreate_int_float_Objects(hostArr, devArr, destroyPinnedObj);
|
||||
@@ -171,10 +164,9 @@ void destroyHostRegObj(void* ptr) {
|
||||
*/
|
||||
TEST_CASE("Unit_hipGraphUserObj_HostRegister") {
|
||||
int clockrate = 0;
|
||||
HIP_CHECK(hipDeviceGetAttribute(&clockrate,
|
||||
hipDeviceAttributeMemoryClockRate, 0));
|
||||
int *A_h = new int();
|
||||
int *A_d = nullptr;
|
||||
HIP_CHECK(hipDeviceGetAttribute(&clockrate, hipDeviceAttributeMemoryClockRate, 0));
|
||||
int* A_h = new int();
|
||||
int* A_d = nullptr;
|
||||
HIP_CHECK(hipHostRegister(A_h, sizeof(int), 0));
|
||||
REQUIRE(A_h != nullptr);
|
||||
*A_h = 1;
|
||||
@@ -183,19 +175,17 @@ TEST_CASE("Unit_hipGraphUserObj_HostRegister") {
|
||||
hipGraph_t graph = nullptr;
|
||||
hipStream_t stream;
|
||||
HIP_CHECK(hipStreamCreate(&stream));
|
||||
HIP_CHECK(hipStreamBeginCapture(stream,
|
||||
hipStreamCaptureModeGlobal));
|
||||
KernelFn<<< 1, 1, 0, stream>>>(A_d, clockrate, 5);
|
||||
HIP_CHECK(hipStreamBeginCapture(stream, hipStreamCaptureModeGlobal));
|
||||
KernelFn<<<1, 1, 0, stream>>>(A_d, clockrate, 5);
|
||||
HIP_CHECK(hipStreamEndCapture(stream, &graph));
|
||||
REQUIRE(graph != nullptr);
|
||||
hipUserObject_t Uobj;
|
||||
int refCount = 1;
|
||||
HIP_CHECK(hipUserObjectCreate(&Uobj, A_h, destroyHostRegObj, refCount,
|
||||
hipUserObjectNoDestructorSync));
|
||||
HIP_CHECK(
|
||||
hipUserObjectCreate(&Uobj, A_h, destroyHostRegObj, refCount, hipUserObjectNoDestructorSync));
|
||||
HIP_CHECK(hipGraphRetainUserObject(graph, Uobj, refCount, 0));
|
||||
hipGraphExec_t graph_instance;
|
||||
HIP_CHECK(hipGraphInstantiate(&graph_instance, graph,
|
||||
nullptr, nullptr, 0));
|
||||
HIP_CHECK(hipGraphInstantiate(&graph_instance, graph, nullptr, nullptr, 0));
|
||||
HIP_CHECK(hipGraphDestroy(graph));
|
||||
HIP_CHECK(hipGraphLaunch(graph_instance, stream));
|
||||
HIP_CHECK(hipGraphExecDestroy(graph_instance));
|
||||
@@ -205,9 +195,7 @@ TEST_CASE("Unit_hipGraphUserObj_HostRegister") {
|
||||
HIP_CHECK(hipStreamDestroy(stream));
|
||||
HIP_CHECK(hipHostUnregister(A_h));
|
||||
}
|
||||
template <typename T>
|
||||
__global__ void StructClassKernelFn(T *Obj,
|
||||
int clockrate, int WaitSecs) {
|
||||
template <typename T> __global__ void StructClassKernelFn(T* Obj, int clockrate, int WaitSecs) {
|
||||
uint64_t num_cycles = (uint64_t)clockrate;
|
||||
num_cycles = num_cycles * 1000 * WaitSecs;
|
||||
uint64_t start = clock64(), cycles = 0;
|
||||
@@ -216,31 +204,25 @@ __global__ void StructClassKernelFn(T *Obj,
|
||||
}
|
||||
Obj->count = 9999;
|
||||
}
|
||||
template <typename T>
|
||||
void hipUserObjectCreate_Struct_Class_Objects(T* Obj_h, T* Obj_d) {
|
||||
template <typename T> void hipUserObjectCreate_Struct_Class_Objects(T* Obj_h, T* Obj_d) {
|
||||
int clockrate = 0;
|
||||
HIP_CHECK(hipDeviceGetAttribute(&clockrate,
|
||||
hipDeviceAttributeMemoryClockRate, 0));
|
||||
HIP_CHECK(hipDeviceGetAttribute(&clockrate, hipDeviceAttributeMemoryClockRate, 0));
|
||||
hipGraph_t graph = nullptr;
|
||||
hipStream_t stream;
|
||||
HIP_CHECK(hipStreamCreate(&stream));
|
||||
HIP_CHECK(hipStreamBeginCapture(stream,
|
||||
hipStreamCaptureModeGlobal));
|
||||
HIP_CHECK(hipMemcpyAsync(Obj_d, Obj_h, sizeof(BoxStruct),
|
||||
hipMemcpyHostToDevice, stream));
|
||||
StructClassKernelFn<<< 1, 1, 0, stream>>>(Obj_d, clockrate, 5);
|
||||
HIP_CHECK(hipMemcpyAsync(Obj_h, Obj_d, sizeof(BoxStruct),
|
||||
hipMemcpyDeviceToHost, stream));
|
||||
HIP_CHECK(hipStreamBeginCapture(stream, hipStreamCaptureModeGlobal));
|
||||
HIP_CHECK(hipMemcpyAsync(Obj_d, Obj_h, sizeof(BoxStruct), hipMemcpyHostToDevice, stream));
|
||||
StructClassKernelFn<<<1, 1, 0, stream>>>(Obj_d, clockrate, 5);
|
||||
HIP_CHECK(hipMemcpyAsync(Obj_h, Obj_d, sizeof(BoxStruct), hipMemcpyDeviceToHost, stream));
|
||||
HIP_CHECK(hipStreamEndCapture(stream, &graph));
|
||||
REQUIRE(graph != nullptr);
|
||||
hipUserObject_t Uobj;
|
||||
int refCount = 1;
|
||||
HIP_CHECK(hipUserObjectCreate(&Uobj, Obj_h, destroyPinnedObj, refCount,
|
||||
hipUserObjectNoDestructorSync));
|
||||
HIP_CHECK(
|
||||
hipUserObjectCreate(&Uobj, Obj_h, destroyPinnedObj, refCount, hipUserObjectNoDestructorSync));
|
||||
HIP_CHECK(hipGraphRetainUserObject(graph, Uobj, refCount, 0));
|
||||
hipGraphExec_t graph_instance;
|
||||
HIP_CHECK(hipGraphInstantiate(&graph_instance, graph,
|
||||
nullptr, nullptr, 0));
|
||||
HIP_CHECK(hipGraphInstantiate(&graph_instance, graph, nullptr, nullptr, 0));
|
||||
HIP_CHECK(hipGraphDestroy(graph));
|
||||
HIP_CHECK(hipGraphLaunch(graph_instance, stream));
|
||||
HIP_CHECK(hipGraphExecDestroy(graph_instance));
|
||||
@@ -277,8 +259,7 @@ TEST_CASE("Unit_hipGraphUserObj_Struct_Class_Ojects") {
|
||||
BoxStruct* structObj_d;
|
||||
HIP_CHECK(hipMalloc(&structObj_d, sizeof(BoxStruct)));
|
||||
REQUIRE(structObj_d != nullptr);
|
||||
hipUserObjectCreate_Struct_Class_Objects<BoxStruct>(structObj_h,
|
||||
structObj_d);
|
||||
hipUserObjectCreate_Struct_Class_Objects<BoxStruct>(structObj_h, structObj_d);
|
||||
t1.join();
|
||||
}
|
||||
SECTION("Called with Class Object") {
|
||||
@@ -290,8 +271,7 @@ TEST_CASE("Unit_hipGraphUserObj_Struct_Class_Ojects") {
|
||||
BoxClass* classObj_d;
|
||||
HIP_CHECK(hipMalloc(&classObj_d, sizeof(BoxClass)));
|
||||
REQUIRE(classObj_d != nullptr);
|
||||
hipUserObjectCreate_Struct_Class_Objects<BoxClass>(classObj_h,
|
||||
classObj_d);
|
||||
hipUserObjectCreate_Struct_Class_Objects<BoxClass>(classObj_h, classObj_d);
|
||||
t1.join();
|
||||
}
|
||||
}
|
||||
@@ -313,45 +293,39 @@ TEST_CASE("Unit_hipGraphUserObj_Struct_Class_Ojects") {
|
||||
*/
|
||||
TEST_CASE("Unit_hipGraphUserObj_ClonedGraph") {
|
||||
int clockrate = 0;
|
||||
HIP_CHECK(hipDeviceGetAttribute(&clockrate,
|
||||
hipDeviceAttributeMemoryClockRate, 0));
|
||||
HIP_CHECK(hipDeviceGetAttribute(&clockrate, hipDeviceAttributeMemoryClockRate, 0));
|
||||
std::thread t1(threadFunc_dltMemory);
|
||||
int *hostArr = nullptr;
|
||||
int* hostArr = nullptr;
|
||||
HIP_CHECK(hipHostMalloc(&hostArr, sizeof(int)));
|
||||
REQUIRE(hostArr != nullptr);
|
||||
*hostArr = 1111;
|
||||
int *devArr = nullptr;
|
||||
int* devArr = nullptr;
|
||||
HIP_CHECK(hipMalloc(&devArr, sizeof(int)));
|
||||
REQUIRE(devArr != nullptr);
|
||||
hipGraph_t graph = nullptr, clonedgraph = nullptr;
|
||||
hipStream_t stream;
|
||||
HIP_CHECK(hipStreamCreate(&stream));
|
||||
HIP_CHECK(hipStreamBeginCapture(stream,
|
||||
hipStreamCaptureModeGlobal));
|
||||
HIP_CHECK(hipMemcpyAsync(devArr, hostArr, sizeof(int),
|
||||
hipMemcpyHostToDevice, stream));
|
||||
KernelFn<<< 1, 1, 0, stream>>>(devArr, clockrate, 5);
|
||||
HIP_CHECK(hipMemcpyAsync(hostArr, devArr, sizeof(int),
|
||||
hipMemcpyDeviceToHost, stream));
|
||||
HIP_CHECK(hipStreamBeginCapture(stream, hipStreamCaptureModeGlobal));
|
||||
HIP_CHECK(hipMemcpyAsync(devArr, hostArr, sizeof(int), hipMemcpyHostToDevice, stream));
|
||||
KernelFn<<<1, 1, 0, stream>>>(devArr, clockrate, 5);
|
||||
HIP_CHECK(hipMemcpyAsync(hostArr, devArr, sizeof(int), hipMemcpyDeviceToHost, stream));
|
||||
HIP_CHECK(hipStreamEndCapture(stream, &graph));
|
||||
REQUIRE(graph != nullptr);
|
||||
hipUserObject_t Uobj;
|
||||
int refCount = 1;
|
||||
HIP_CHECK(hipUserObjectCreate(&Uobj, hostArr, destroyPinnedObj, refCount,
|
||||
hipUserObjectNoDestructorSync));
|
||||
hipUserObjectNoDestructorSync));
|
||||
HIP_CHECK(hipGraphRetainUserObject(graph, Uobj, refCount, 0));
|
||||
hipGraphExec_t originalGraphInstance, clonedGraphInstance;
|
||||
// Instantiate and launch the original graph
|
||||
HIP_CHECK(hipGraphInstantiate(&originalGraphInstance, graph,
|
||||
nullptr, nullptr, 0));
|
||||
HIP_CHECK(hipGraphInstantiate(&originalGraphInstance, graph, nullptr, nullptr, 0));
|
||||
HIP_CHECK(hipGraphLaunch(originalGraphInstance, stream));
|
||||
HIP_CHECK(hipGraphExecDestroy(originalGraphInstance));
|
||||
REQUIRE(*hostArr == 1111);
|
||||
HIP_CHECK(hipGraphClone(&clonedgraph, graph));
|
||||
REQUIRE(clonedgraph != nullptr);
|
||||
// Instantiate and launch the cloned graph
|
||||
HIP_CHECK(hipGraphInstantiate(&clonedGraphInstance, clonedgraph,
|
||||
nullptr, nullptr, 0));
|
||||
HIP_CHECK(hipGraphInstantiate(&clonedGraphInstance, clonedgraph, nullptr, nullptr, 0));
|
||||
HIP_CHECK(hipGraphDestroy(graph));
|
||||
HIP_CHECK(hipGraphDestroy(clonedgraph));
|
||||
HIP_CHECK(hipGraphLaunch(clonedGraphInstance, stream));
|
||||
@@ -363,7 +337,7 @@ TEST_CASE("Unit_hipGraphUserObj_ClonedGraph") {
|
||||
HIP_CHECK(hipStreamDestroy(stream));
|
||||
HIP_CHECK(hipFree(devArr));
|
||||
}
|
||||
__global__ void ManualGraphKernelFn(int *Ad, int clockrate, int WaitSecs) {
|
||||
__global__ void ManualGraphKernelFn(int* Ad, int clockrate, int WaitSecs) {
|
||||
uint64_t num_cycles = (uint64_t)clockrate;
|
||||
num_cycles = num_cycles * 1000 * WaitSecs;
|
||||
uint64_t start = clock64(), cycles = 0;
|
||||
@@ -390,25 +364,23 @@ __global__ void ManualGraphKernelFn(int *Ad, int clockrate, int WaitSecs) {
|
||||
*/
|
||||
TEST_CASE("Unit_hipGraphUserObj_ManualGraph") {
|
||||
int clockrate = 0;
|
||||
HIP_CHECK(hipDeviceGetAttribute(&clockrate,
|
||||
hipDeviceAttributeMemoryClockRate, 0));
|
||||
HIP_CHECK(hipDeviceGetAttribute(&clockrate, hipDeviceAttributeMemoryClockRate, 0));
|
||||
std::thread t1(threadFunc_dltMemory);
|
||||
hipGraph_t graph;
|
||||
hipGraphNode_t memcpyNode, kNode;
|
||||
hipKernelNodeParams kNodeParams{};
|
||||
hipStream_t stream;
|
||||
int *hostArr = nullptr;
|
||||
int* hostArr = nullptr;
|
||||
HIP_CHECK(hipHostMalloc(&hostArr, sizeof(int)));
|
||||
REQUIRE(hostArr != nullptr);
|
||||
*hostArr = 1111;
|
||||
int *devArr = nullptr;
|
||||
int* devArr = nullptr;
|
||||
HIP_CHECK(hipMalloc(&devArr, sizeof(int)));
|
||||
REQUIRE(devArr != nullptr);
|
||||
std::vector<hipGraphNode_t> dependencies;
|
||||
HIP_CHECK(hipStreamCreate(&stream));
|
||||
HIP_CHECK(hipGraphCreate(&graph, 0));
|
||||
HIP_CHECK(hipGraphAddMemcpyNode1D(&memcpyNode, graph, nullptr, 0, devArr,
|
||||
hostArr, sizeof(int),
|
||||
HIP_CHECK(hipGraphAddMemcpyNode1D(&memcpyNode, graph, nullptr, 0, devArr, hostArr, sizeof(int),
|
||||
hipMemcpyHostToDevice));
|
||||
dependencies.push_back(memcpyNode);
|
||||
kNodeParams.func = reinterpret_cast<void*>(ManualGraphKernelFn);
|
||||
@@ -417,23 +389,20 @@ TEST_CASE("Unit_hipGraphUserObj_ManualGraph") {
|
||||
kNodeParams.gridDim = dim3(1);
|
||||
kNodeParams.blockDim = dim3(1);
|
||||
kNodeParams.kernelParams = kernelArgs;
|
||||
HIP_CHECK(hipGraphAddKernelNode(&kNode, graph, dependencies.data(),
|
||||
dependencies.size(), &kNodeParams));
|
||||
HIP_CHECK(
|
||||
hipGraphAddKernelNode(&kNode, graph, dependencies.data(), dependencies.size(), &kNodeParams));
|
||||
dependencies.clear();
|
||||
dependencies.push_back(kNode);
|
||||
HIP_CHECK(hipGraphAddMemcpyNode1D(&memcpyNode, graph, dependencies.data(),
|
||||
dependencies.size(),
|
||||
hostArr, devArr, sizeof(int),
|
||||
hipMemcpyDeviceToHost));
|
||||
HIP_CHECK(hipGraphAddMemcpyNode1D(&memcpyNode, graph, dependencies.data(), dependencies.size(),
|
||||
hostArr, devArr, sizeof(int), hipMemcpyDeviceToHost));
|
||||
REQUIRE(graph != nullptr);
|
||||
hipUserObject_t Uobj;
|
||||
int refCount = 1;
|
||||
HIP_CHECK(hipUserObjectCreate(&Uobj, hostArr, destroyPinnedObj, refCount,
|
||||
hipUserObjectNoDestructorSync));
|
||||
hipUserObjectNoDestructorSync));
|
||||
HIP_CHECK(hipGraphRetainUserObject(graph, Uobj, refCount, 0));
|
||||
hipGraphExec_t graph_instance;
|
||||
HIP_CHECK(hipGraphInstantiate(&graph_instance, graph,
|
||||
nullptr, nullptr, 0));
|
||||
HIP_CHECK(hipGraphInstantiate(&graph_instance, graph, nullptr, nullptr, 0));
|
||||
HIP_CHECK(hipGraphDestroy(graph));
|
||||
HIP_CHECK(hipGraphLaunch(graph_instance, stream));
|
||||
HIP_CHECK(hipGraphExecDestroy(graph_instance));
|
||||
@@ -445,7 +414,6 @@ TEST_CASE("Unit_hipGraphUserObj_ManualGraph") {
|
||||
HIP_CHECK(hipFree(devArr));
|
||||
}
|
||||
/**
|
||||
* End doxygen group GraphTest.
|
||||
* @}
|
||||
*/
|
||||
|
||||
* End doxygen group GraphTest.
|
||||
* @}
|
||||
*/
|
||||
|
||||
Référencer dans un nouveau ticket
Bloquer un utilisateur