AI/rocm-systems
2
0
Форкнуть 0
Код Задачи Запросы на слияние Действия Пакеты Проекты Релизы Вики Активность

SWDEV-299127 - Merge 'develop' into 'amd-staging'

Просмотр исходного кода 
Change-Id: I126fc5009ed46697b2af46b4726902895251b0b4
Этот коммит содержится в:
Jenkins
2022-07-26 19:10:35 -04:00
родитель 3aa38b08b4 1c4822f24a
Коммит 2d920e7053
13 изменённых файлов: 1115 добавлений и 32 удалений
Скачать Patch файл Скачать Diff файл Показать все файлы Свернуть все файлы
.jenkins/Jenkinsfile
поставляемый
+4 -4
Просмотреть файл
@@ -28,7 +28,7 @@ def hipBuildTest(String backendLabel) {
}
}
}
stage("Build - HIT framework") {
stage("Build - HIT framework ${backendLabel}") {
// Running the build on hipamd workspace
dir("${WORKSPACE}/hipamd") {
sh """#!/usr/bin/env bash
@@ -52,7 +52,7 @@ def hipBuildTest(String backendLabel) {
}
}
timeout(time: 1, unit: 'HOURS') {
stage('HIP Unit Tests - HIT framework') {
stage('HIP Unit Tests - HIT framework ${backendLabel}') {
dir("${WORKSPACE}/hipamd/build") {
sh """#!/usr/bin/env bash
set -x
@@ -67,7 +67,7 @@ def hipBuildTest(String backendLabel) {
}
}
}
stage("Build - Catch2 framework") {
stage("Build - Catch2 framework ${backendLabel}") {
// Running the build on hipamd workspace
dir("${WORKSPACE}/hipamd") {
sh """#!/usr/bin/env bash
@@ -94,7 +94,7 @@ def hipBuildTest(String backendLabel) {
}
}
timeout(time: 1, unit: 'HOURS') {
stage('HIP Unit Tests - Catch2 framework') {
stage('HIP Unit Tests - Catch2 framework ${backendLabel}') {
dir("${WORKSPACE}/hipamd/build") {
sh """#!/usr/bin/env bash
set -x
bin/hipcc.pl
-5
Просмотреть файл
@@ -187,11 +187,6 @@ if ($HIP_PLATFORM eq "amd") {
print ("HIP_CLANG_TARGET=$HIP_CLANG_TARGET\n");
}
if ($isWindows) {
$HIPCXXFLAGS .= " -std=c++14 -fms-extensions -fms-compatibility";
} else {
$HIPCXXFLAGS .= " -std=c++11";
}
$HIPCXXFLAGS .= " -isystem \"$HIP_CLANG_INCLUDE_PATH/..\"";
$HIPCFLAGS .= " -isystem \"$HIP_CLANG_INCLUDE_PATH/..\"";
$HIPLDFLAGS .= " -L\"$HIP_LIB_PATH\"";
include/hip/driver_types.h
+1 -1
Просмотреть файл
@@ -142,7 +142,7 @@ typedef enum HIPresourcetype_enum {
HIP_RESOURCE_TYPE_MIPMAPPED_ARRAY = 0x01, /**< Mipmapped array resource */
HIP_RESOURCE_TYPE_LINEAR = 0x02, /**< Linear resource */
HIP_RESOURCE_TYPE_PITCH2D = 0x03 /**< Pitch 2D resource */
} HIPresourcetype;
} HIPresourcetype, hipResourcetype;
/**
* hip address modes
*/
include/hip/hip_runtime_api.h
+21 -3
Просмотреть файл
@@ -1842,7 +1842,7 @@ hipError_t hipGetLastError(void);
*/
hipError_t hipPeekAtLastError(void);
/**
* @brief Return name of the specified error code in text form.
* @brief Return hip error as text string form.
*
* @param hip_error Error code to convert to name.
* @return const char pointer to the NULL-terminated error name
@@ -1856,11 +1856,29 @@ const char* hipGetErrorName(hipError_t hip_error);
* @param hipError Error code to convert to string.
* @return const char pointer to the NULL-terminated error string
*
* @warning : on HCC, this function returns the name of the error (same as hipGetErrorName)
*
* @see hipGetErrorName, hipGetLastError, hipPeakAtLastError, hipError_t
*/
const char* hipGetErrorString(hipError_t hipError);
/**
* @brief Return hip error as text string form.
*
* @param [in] hipError Error code to convert to string.
* @param [out] const char pointer to the NULL-terminated error string
* @return #hipSuccess, #hipErrorInvalidValue
*
* @see hipGetErrorName, hipGetLastError, hipPeakAtLastError, hipError_t
*/
hipError_t hipDrvGetErrorName(hipError_t hipError, const char** errorString);
/**
* @brief Return handy text string message to explain the error which occurred
*
* @param [in] hipError Error code to convert to string.
* @param [out] const char pointer to the NULL-terminated error string
* @return #hipSuccess, #hipErrorInvalidValue
*
* @see hipGetErrorName, hipGetLastError, hipPeakAtLastError, hipError_t
*/
hipError_t hipDrvGetErrorString(hipError_t hipError, const char** errorString);
// end doxygen Error
/**
* @}
tests/catch/hipTestMain/config/config_amd_windows.json
+13 -12
Просмотреть файл
@@ -5,19 +5,11 @@
"Unit_printf_flags",
"Unit_printf_specifier",
"Unit_hipTextureMipmapObj2D_Check",
"Unit_hipManagedKeyword_MultiGpu",
"Unit_hipGraphAddHostNode_ClonedGraphwithHostNode",
"Unit_hipEventIpc",
"Unit_hipMalloc3D_Negative",
"Unit_hipPointerGetAttribute_MappedMem",
"Unit_hipStreamBeginCapture_BasicFunctional",
"Unit_hipStreamBeginCapture_hipStreamPerThread",
"Unit_hiprtc_functional",
"Unit_hipStreamValue_Write",
"Unit_hipStreamGetCaptureInfo_BasicFunctional",
"Unit_hipStreamGetCaptureInfo_hipStreamPerThread",
"Unit_hipStreamGetCaptureInfo_UniqueID",
"Unit_hipStreamGetCaptureInfo_ArgValidation",
"Unit_hipMemPoolApi_Basic",
"Unit_hipMemPoolApi_BasicAlloc",
"Unit_hipMemPoolApi_BasicTrim",
@@ -41,9 +33,18 @@
"Unit_hipArrayCreate_happy - float2",
"Unit_hipArrayCreate_happy - float4",
"Unit_hipMemVmm_Basic",
"Unit_hipGraphKernelNodeSetParams_Functional",
"Unit_hipGraphExecKernelNodeSetParams_Negative",
"Unit_hipGraphExecKernelNodeSetParams_Functional",
"Unit_hipGraphMemcpyNodeSetParams1D_Functional"
"Unit_hipGraphExecMemcpyNodeSetParamsToSymbol_Functional",
"Unit_hipMallocManaged_MultiChunkMultiDevice",
"Unit_hipMallocManaged_TwoPointers - int",
"Unit_hipMallocManaged_TwoPointers - float",
"Unit_hipMallocManaged_TwoPointers - double",
"Unit_hipMallocManaged_DeviceContextChange - unsigned char",
"Unit_hipMallocManaged_DeviceContextChange - int",
"Unit_hipMallocManaged_DeviceContextChange - float",
"Unit_hipMallocManaged_DeviceContextChange - double",
"Unit_hipGraphNodeGetDependentNodes_Functional",
"Unit_hipGraphNodeGetDependentNodes_ParamValidation",
"Unit_hipGraphNodeGetDependencies_Functional",
"Unit_hipGraphNodeGetDependencies_ParamValidation"
]
}
tests/catch/multiproc/hipHostMallocTestsMproc.cc
+6 -2
Просмотреть файл
@@ -41,7 +41,7 @@ Testcase Scenarios :
* Defines
*/
#define LIB_ROCMSMI "librocm_smi64.so"
#define ALLOC_SIZE (1*1024*1024)
#define ALLOC_SIZE (30*1024*1024)
/**
* Global variables
@@ -221,7 +221,11 @@ static bool validatePageTableAllocations(const char *devList, int visDevCnt) {
}
for (indx = 0; indx < numdev; indx++) {
if (current[indx] - prev[indx])
// For visible hip devices, there should be increase in VRAM usage
// due to page table allocations
// For NON visible hip devices, there can be reduction in VRAM usage
// due to removal of page tables from them
if (current[indx] > prev[indx])
changeCnt++;
}
tests/catch/unit/graph/CMakeLists.txt
+3
Просмотреть файл
@@ -73,6 +73,9 @@ set(TEST_SRC
hipGraphExecKernelNodeSetParams.cc
hipGraphLaunch.cc
hipGraphMemcpyNodeSetParams1D.cc
hipGraphExecMemcpyNodeSetParamsToSymbol.cc
hipGraphNodeGetDependentNodes.cc
hipGraphNodeGetDependencies.cc
)
hip_add_exe_to_target(NAME GraphsTest
tests/catch/unit/graph/hipGraphExecMemcpyNodeSetParamsToSymbol.cc
+316
Просмотреть файл
@@ -0,0 +1,316 @@
/*
Copyright (c) 2022 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.
*/
/**
Testcase Scenarios of hipGraphExecMemcpyNodeSetParamsToSymbol API:
Functional :
1) Allocate global symbol memory, Instantiate a graph with memcpy node,
obtain executable graph and update the node params with set exec api call.
Make sure they are taking effect.
2) Allocate const symbol memory, Instantiate a graph with memcpy node,
obtain executable graph and update the node params with set exec api call.
Make sure they are taking effect.
Negative :
1) Pass hGraphExec as nullptr and check if api returns error.
2) Pass GraphNode as nullptr and check if api returns error.
3) Pass symbol ptr as nullptr, api expected to return error code.
4) Pass source ptr as nullptr, api expected to return error code.
5) Pass count as zero, api expected to return error code.
6) Pass offset+count greater than allocated size, api expected to return error code.
7) Pass same symbol pointer as source ptr and destination ptr, api expected to return error code.
8) Pass Pass both dstn ptr and source ptr as 2 different symbol ptr, api expected to return error code.
9) Copy from device ptr to host ptr but pass kind as different, api expected to return error code.
10) Check with other graph node but pass same graphExec, api expected to return error code.
*/
#include <hip_test_common.hh>
#include <hip_test_checkers.hh>
#include <limits>
#define SIZE 256
__device__ int globalIn[SIZE], globalOut[SIZE];
__device__ __constant__ int globalConst[SIZE];
__global__ void MemcpyToSymbolExecKernel(int* B_d) {
for (int i = 0 ; i < SIZE; i++) {
B_d[i] = globalIn[i];
}
}
__global__ void MemcpyToConstSymbolExecKernel(int* B_d) {
for (int i = 0 ; i < SIZE; i++) {
B_d[i] = globalConst[i];
}
}
/* This testcase verifies negative scenarios of
hipGraphExecMemcpyNodeSetParamsToSymbol API */
TEST_CASE("Unit_hipGraphExecMemcpyNodeSetParamsToSymbol_Negative") {
constexpr size_t Nbytes = SIZE * sizeof(int);
int *A_d{nullptr}, *B_d{nullptr}, *C_d{nullptr};
int *A_h{nullptr}, *B_h{nullptr};
HipTest::initArrays<int>(&A_d, &B_d, &C_d,
&A_h, &B_h, nullptr, SIZE, false);
hipGraph_t graph;
hipError_t ret;
hipGraphExec_t graphExec;
hipGraphNode_t memcpyToSymbolNode, memcpyH2D;
std::vector<hipGraphNode_t> dependencies;
HIP_CHECK(hipGraphCreate(&graph, 0));
// Adding MemcpyNode
HIP_CHECK(hipGraphAddMemcpyNode1D(&memcpyH2D, graph, nullptr, 0, A_d, A_h,
Nbytes, hipMemcpyHostToDevice));
dependencies.push_back(memcpyH2D);
HIP_CHECK(hipGraphAddMemcpyNodeToSymbol(&memcpyToSymbolNode, graph,
dependencies.data(),
dependencies.size(),
HIP_SYMBOL(globalIn),
C_d, Nbytes, 0,
hipMemcpyDeviceToDevice));
dependencies.clear();
dependencies.push_back(memcpyToSymbolNode);
HIP_CHECK(hipGraphMemcpyNodeSetParamsToSymbol(memcpyToSymbolNode,
HIP_SYMBOL(globalIn), A_d,
Nbytes, 0,
hipMemcpyDeviceToDevice));
// Instantiate the graph
HIP_CHECK(hipGraphInstantiate(&graphExec, graph, nullptr, nullptr, 0));
SECTION("Pass hGraphExec as nullptr") {
ret = hipGraphExecMemcpyNodeSetParamsToSymbol(nullptr,
memcpyToSymbolNode,
HIP_SYMBOL(globalIn),
B_d, Nbytes, 0,
hipMemcpyDeviceToDevice);
REQUIRE(hipErrorInvalidValue == ret);
}
SECTION("Pass GraphNode as nullptr") {
ret = hipGraphExecMemcpyNodeSetParamsToSymbol(graphExec,
nullptr,
HIP_SYMBOL(globalIn),
B_d, Nbytes, 0,
hipMemcpyDeviceToDevice);
REQUIRE(hipErrorInvalidValue == ret);
}
SECTION("Pass symbol ptr as nullptr") {
ret = hipGraphExecMemcpyNodeSetParamsToSymbol(graphExec,
memcpyToSymbolNode,
nullptr,
B_d, Nbytes, 0,
hipMemcpyDeviceToDevice);
REQUIRE(hipErrorInvalidSymbol == ret);
}
SECTION("Pass source ptr as nullptr") {
ret = hipGraphExecMemcpyNodeSetParamsToSymbol(graphExec,
memcpyToSymbolNode,
HIP_SYMBOL(globalIn),
nullptr, Nbytes, 0,
hipMemcpyDeviceToDevice);
REQUIRE(hipErrorInvalidValue == ret);
}
SECTION("Pass count as zero") {
ret = hipGraphExecMemcpyNodeSetParamsToSymbol(graphExec,
memcpyToSymbolNode,
HIP_SYMBOL(globalIn),
B_d, 0, 0,
hipMemcpyDeviceToDevice);
REQUIRE(hipErrorInvalidValue == ret);
}
SECTION("Pass offset+count greater than allocated size") {
ret = hipGraphExecMemcpyNodeSetParamsToSymbol(graphExec,
memcpyToSymbolNode,
HIP_SYMBOL(globalIn),
B_d, Nbytes, 10,
hipMemcpyDeviceToDevice);
REQUIRE(hipErrorInvalidValue == ret);
}
SECTION("Pass same symbol pointer as source ptr and destination ptr") {
ret = hipGraphExecMemcpyNodeSetParamsToSymbol(graphExec,
memcpyToSymbolNode,
HIP_SYMBOL(globalIn),
HIP_SYMBOL(globalIn),
Nbytes, 0,
hipMemcpyDeviceToDevice);
REQUIRE(hipErrorInvalidValue == ret);
}
SECTION("Pass both dstn ptr and source ptr as 2 different symbol ptr") {
ret = hipGraphExecMemcpyNodeSetParamsToSymbol(graphExec,
memcpyToSymbolNode,
HIP_SYMBOL(globalIn),
HIP_SYMBOL(globalOut),
Nbytes, 0,
hipMemcpyDeviceToDevice);
REQUIRE(hipErrorInvalidValue == ret);
}
SECTION("Copy from device ptr to host ptr but pass kind as different") {
ret = hipGraphExecMemcpyNodeSetParamsToSymbol(graphExec,
memcpyToSymbolNode,
HIP_SYMBOL(globalIn),
B_h,
Nbytes, 0,
hipMemcpyDeviceToHost);
REQUIRE(hipSuccess != ret);
}
SECTION("Check with other graph node but pass same graphExec") {
hipGraph_t graph1;
hipGraphNode_t memcpyToSymbolNode1{};
HIP_CHECK(hipGraphCreate(&graph1, 0));
HIP_CHECK(hipGraphAddMemcpyNodeToSymbol(&memcpyToSymbolNode1, graph1,
nullptr,
0,
HIP_SYMBOL(globalOut),
A_d, Nbytes, 0,
hipMemcpyDeviceToDevice));
ret = hipGraphExecMemcpyNodeSetParamsToSymbol(graphExec,
memcpyToSymbolNode1,
HIP_SYMBOL(globalOut),
HIP_SYMBOL(globalIn),
Nbytes, 0,
hipMemcpyDeviceToDevice);
REQUIRE(hipErrorInvalidValue == ret);
}
HipTest::freeArrays<int>(A_d, B_d, C_d,
A_h, B_h, nullptr, false);
HIP_CHECK(hipGraphExecDestroy(graphExec));
HIP_CHECK(hipGraphDestroy(graph));
}
static
void hipGraphExecMemcpyNodeSetParamsToSymbol_GlobalMem(bool useConstVar) {
constexpr size_t Nbytes = SIZE * sizeof(int);
constexpr auto blocksPerCU = 6; // to hide latency
constexpr auto threadsPerBlock = 256;
unsigned blocks = HipTest::setNumBlocks(blocksPerCU, threadsPerBlock, SIZE);
hipGraphNode_t memcpytosymbolkernel, memcpyD2H_B;
hipKernelNodeParams kernelNodeParams{};
int *A_d{nullptr}, *B_d{nullptr}, *C_d{nullptr};
int *A_h{nullptr}, *B_h{nullptr};
HipTest::initArrays<int>(&A_d, &B_d, &C_d,
&A_h, &B_h, nullptr, SIZE, false);
hipGraph_t graph;
hipGraphExec_t graphExec;
hipGraphNode_t memcpyToSymbolNode, memcpyH2D_A;
std::vector<hipGraphNode_t> dependencies;
HIP_CHECK(hipGraphCreate(&graph, 0));
// Adding MemcpyNode
HIP_CHECK(hipGraphAddMemcpyNode1D(&memcpyH2D_A, graph, nullptr, 0, A_d, A_h,
Nbytes, hipMemcpyHostToDevice));
dependencies.push_back(memcpyH2D_A);
if (useConstVar) {
HIP_CHECK(hipGraphAddMemcpyNodeToSymbol(&memcpyToSymbolNode, graph,
dependencies.data(),
dependencies.size(),
HIP_SYMBOL(globalConst),
C_d, Nbytes, 0,
hipMemcpyDeviceToDevice));
} else {
HIP_CHECK(hipGraphAddMemcpyNodeToSymbol(&memcpyToSymbolNode, graph,
dependencies.data(),
dependencies.size(),
HIP_SYMBOL(globalIn),
C_d, Nbytes, 0,
hipMemcpyDeviceToDevice));
}
dependencies.clear();
dependencies.push_back(memcpyToSymbolNode);
// Adding Kernel node
void* kernelArgs1[] = {&B_d};
if (useConstVar) {
kernelNodeParams.func =
reinterpret_cast<void *>(MemcpyToConstSymbolExecKernel);
} else {
kernelNodeParams.func = reinterpret_cast<void *>(MemcpyToSymbolExecKernel);
}
kernelNodeParams.gridDim = dim3(blocks);
kernelNodeParams.blockDim = dim3(threadsPerBlock);
kernelNodeParams.sharedMemBytes = 0;
kernelNodeParams.kernelParams = reinterpret_cast<void**>(kernelArgs1);
kernelNodeParams.extra = nullptr;
HIP_CHECK(hipGraphAddKernelNode(&memcpytosymbolkernel, graph,
dependencies.data(), dependencies.size(),
&kernelNodeParams));
dependencies.clear();
dependencies.push_back(memcpytosymbolkernel);
// Adding MemcpyNode
HIP_CHECK(hipGraphAddMemcpyNode1D(&memcpyD2H_B, graph, dependencies.data(),
dependencies.size(), B_h, B_d,
Nbytes, hipMemcpyDeviceToHost));
// Instantiate the graph
HIP_CHECK(hipGraphInstantiate(&graphExec, graph, nullptr, nullptr, 0));
// Update the node with source pointer from C_d to A_d
if (useConstVar) {
HIP_CHECK(hipGraphExecMemcpyNodeSetParamsToSymbol(graphExec,
memcpyToSymbolNode,
HIP_SYMBOL(globalConst), A_d,
Nbytes, 0,
hipMemcpyDeviceToDevice));
} else {
HIP_CHECK(hipGraphExecMemcpyNodeSetParamsToSymbol(graphExec,
memcpyToSymbolNode,
HIP_SYMBOL(globalIn), A_d,
Nbytes, 0,
hipMemcpyDeviceToDevice));
}
HIP_CHECK(hipGraphLaunch(graphExec, 0));
// Validating the result
for (int i = 0; i < SIZE; i++) {
if (B_h[i] != A_h[i]) {
WARN("Validation failed B_h[i] " << B_h[i] << "A_h[i] " << A_h[i]);
REQUIRE(false);
}
}
HipTest::freeArrays<int>(A_d, B_d, C_d,
A_h, B_h, nullptr, false);
HIP_CHECK(hipGraphExecDestroy(graphExec));
HIP_CHECK(hipGraphDestroy(graph));
}
/* Test verifies hipGraphExecMemcpyNodeSetParamsToSymbol Functional scenario.
1) Allocate global symbol memory, Instantiate a graph with memcpy node,
obtain executable graph and update the node params with set exec api call.
Make sure they are taking effect.
2) Allocate const symbol memory, Instantiate a graph with memcpy node,
obtain executable graph and update the node params with set exec api call.
Make sure they are taking effect.
*/
TEST_CASE("Unit_hipGraphExecMemcpyNodeSetParamsToSymbol_Functional") {
SECTION("Check and update with Global Device Symbol Memory") {
hipGraphExecMemcpyNodeSetParamsToSymbol_GlobalMem(false);
}
SECTION("Check and update with Constant Global Device Symbol Memory") {
hipGraphExecMemcpyNodeSetParamsToSymbol_GlobalMem(true);
}
}
tests/catch/unit/graph/hipGraphNodeGetDependencies.cc
+368
Просмотреть файл
@@ -0,0 +1,368 @@
/*
Copyright (c) 2022 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.
*/
/**
Testcase Scenarios
------------------
Functional:
1) Create a graph and add nodes with dependencies. Try to fetch dependencies of a node added and verify the api returns the node's
dependencies that were defined.
2) When pDependencies is passed as nullptr, verify pNumDependencies returns actual number of dependencies of node.
3) When pNumDependencies is higher than the actual number of dependencies, the remaining entries in pDependencies will be set to NULL,
and the number of nodes actually obtained will be returned in pNumDependencies.
4) When pNumDependencies is lesser than the actual number of dependencies, api should return the requested number of dependencies.
Argument Validation:
1) Verify the api returns pNumDependencies(0) when node passed is a root node.
2) Pass node as nullptr and verify api doesn’t crash, returns error code.
3) Pass pNumDependencies as nullptr and verify api doesn’t crash, returns error code.
4) Pass node as un-initialized/invalid parameter and verify api returns error code.
*/
#include <hip_test_common.hh>
#include <hip_test_checkers.hh>
#include <hip_test_kernels.hh>
static __global__ void updateResult(int* C_d, int* Res_d, int val,
int64_t NELEM) {
size_t offset = (blockIdx.x * blockDim.x + threadIdx.x);
size_t stride = blockDim.x * gridDim.x;
for (int64_t i = NELEM - stride + offset; i >= 0; i -= stride) {
Res_d[i] = C_d[i] + val;
}
}
static __global__ void vectorSum(const int* A_d, const int* B_d,
const int* C_d, int* Res_d, size_t NELEM) {
size_t offset = (blockIdx.x * blockDim.x + threadIdx.x);
size_t stride = blockDim.x * gridDim.x;
for (size_t i = offset; i < NELEM; i += stride) {
Res_d[i] = A_d[i] + B_d[i] + C_d[i];
}
}
/**
* Verify api when GetDependencies is requested
* for actual number of nodes.
*/
static void queryActualNumOfDeps(const std::vector<hipGraphNode_t> &Nlist,
hipGraphNode_t kernel_vecAdd, size_t numDeps) {
hipGraphNode_t* depnodes;
int numBytes = sizeof(hipGraphNode_t) * numDeps;
depnodes = reinterpret_cast<hipGraphNode_t *>(malloc(numBytes));
REQUIRE(depnodes != nullptr);
HIP_CHECK(hipGraphNodeGetDependencies(kernel_vecAdd, depnodes, &numDeps));
REQUIRE(numDeps == Nlist.size());
// Verify all dependencies are present in the node entries returned
for (auto Node : Nlist) {
bool found = false;
for (size_t i = 0; i < numDeps; i++) {
if (Node == depnodes[i]) {
found = true;
break;
}
}
if (!found) {
INFO("Dependency node " << Node << " not present in returned list");
REQUIRE(false);
}
}
free(depnodes);
}
/**
* Verify api when GetDependencies queried
* for greater number than actual number of nodes.
*/
static void queryGreaterNumOfDeps(const std::vector<hipGraphNode_t> &Nlist,
hipGraphNode_t kernel_vecAdd, size_t numDeps) {
constexpr auto addlEntries = 4;
hipGraphNode_t* depnodes;
size_t totDeps = numDeps + addlEntries;
int numBytes = sizeof(hipGraphNode_t) * totDeps;
depnodes = reinterpret_cast<hipGraphNode_t *>(malloc(numBytes));
REQUIRE(depnodes != nullptr);
HIP_CHECK(hipGraphNodeGetDependencies(kernel_vecAdd, depnodes, &totDeps));
REQUIRE(totDeps == Nlist.size());
for (auto i = numDeps; i < numDeps + addlEntries; i++) {
REQUIRE(depnodes[i] == nullptr);
}
// Verify all dependencies are present in the node entries returned
for (auto Node : Nlist) {
bool found = false;
for (size_t i = 0; i < numDeps; i++) {
if (Node == depnodes[i]) {
found = true;
break;
}
}
if (!found) {
INFO("Dependency node " << Node << " not present in returned list");
REQUIRE(false);
}
}
free(depnodes);
}
/**
* Verify api when GetDependencies queried
* for lesser number than actual number of nodes.
*/
static void queryLesserNumOfDeps(const std::vector<hipGraphNode_t> &Nlist,
hipGraphNode_t kernel_vecAdd, size_t numDeps) {
size_t totDeps = numDeps - 1;
hipGraphNode_t* depnodes;
int numBytes = sizeof(hipGraphNode_t) * totDeps;
size_t count{};
depnodes = reinterpret_cast<hipGraphNode_t *>(malloc(numBytes));
REQUIRE(depnodes != nullptr);
HIP_CHECK(hipGraphNodeGetDependencies(kernel_vecAdd, depnodes, &totDeps));
REQUIRE(totDeps == Nlist.size() - 1);
// Verify all dependencies are present in the node entries returned
for (auto Node : Nlist) {
for (size_t i = 0; i < totDeps; i++) {
if (Node == depnodes[i]) {
count++;
break;
}
}
}
REQUIRE(count == totDeps);
free(depnodes);
}
/**
* Functional Test for getting dependencies of node in graph and verifying execution
*/
TEST_CASE("Unit_hipGraphNodeGetDependencies_Functional") {
constexpr size_t N = 1024;
constexpr size_t Nbytes = N * sizeof(int);
constexpr auto blocksPerCU = 6; // to hide latency
constexpr auto threadsPerBlock = 256;
hipGraphNode_t kernel_vecSqr{}, kernel_vecAdd{};
hipGraphNode_t kernelmod1{}, kernelmod2{}, kernelmod3{};
hipGraphNode_t memcpyD2H{}, memcpyH2D_A{};
hipKernelNodeParams kernelNodeParams{};
hipGraph_t graph{};
size_t numDeps{};
hipStream_t streamForGraph;
int *A_d, *C_d;
int *A_h, *C_h;
int *Res1_d, *Res2_d, *Res3_d;
int *Sum_d, *Sum_h;
hipGraphExec_t graphExec;
size_t NElem{N};
HIP_CHECK(hipStreamCreate(&streamForGraph));
HIP_CHECK(hipGraphCreate(&graph, 0));
HipTest::initArrays<int>(&A_d, &C_d, &Sum_d, &A_h, &C_h, &Sum_h, N);
HipTest::initArrays<int>(&Res1_d, &Res2_d, &Res3_d,
nullptr, nullptr, nullptr, N);
unsigned blocks = HipTest::setNumBlocks(blocksPerCU, threadsPerBlock, N);
// Initialize input buffer and vecsqr result
for (size_t i = 0; i < N; ++i) {
A_h[i] = i + 1;
C_h[i] = A_h[i] * A_h[i];
}
HIP_CHECK(hipGraphAddMemcpyNode1D(&memcpyH2D_A, graph, nullptr, 0, A_d, A_h,
Nbytes, hipMemcpyHostToDevice));
void* kernelArgsVS[] = {&A_d, &C_d, reinterpret_cast<void *>(&NElem)};
memset(&kernelNodeParams, 0, sizeof(kernelNodeParams));
kernelNodeParams.func =
reinterpret_cast<void *>(HipTest::vector_square<int>);
kernelNodeParams.gridDim = dim3(blocks);
kernelNodeParams.blockDim = dim3(threadsPerBlock);
kernelNodeParams.sharedMemBytes = 0;
kernelNodeParams.kernelParams = reinterpret_cast<void**>(kernelArgsVS);
kernelNodeParams.extra = nullptr;
HIP_CHECK(hipGraphAddKernelNode(&kernel_vecSqr, graph, &memcpyH2D_A, 1,
&kernelNodeParams));
// Create multiple nodes dependent on vecSqr node.
// Dependent nodes takes vecSqr input and computes output independently.
std::vector<hipGraphNode_t> nodelist;
int incValue1{1};
void* kernelArgs1[] = {&C_d, &Res1_d, &incValue1,
reinterpret_cast<void *>(&NElem)};
memset(&kernelNodeParams, 0, sizeof(kernelNodeParams));
kernelNodeParams.func =
reinterpret_cast<void *>(updateResult);
kernelNodeParams.gridDim = dim3(blocks);
kernelNodeParams.blockDim = dim3(threadsPerBlock);
kernelNodeParams.sharedMemBytes = 0;
kernelNodeParams.kernelParams = reinterpret_cast<void**>(kernelArgs1);
kernelNodeParams.extra = nullptr;
HIP_CHECK(hipGraphAddKernelNode(&kernelmod1, graph, &kernel_vecSqr, 1,
&kernelNodeParams));
nodelist.push_back(kernelmod1);
int incValue2{2};
void* kernelArgs2[] = {&C_d, &Res2_d, &incValue2,
reinterpret_cast<void *>(&NElem)};
memset(&kernelNodeParams, 0, sizeof(kernelNodeParams));
kernelNodeParams.func =
reinterpret_cast<void *>(updateResult);
kernelNodeParams.gridDim = dim3(blocks);
kernelNodeParams.blockDim = dim3(threadsPerBlock);
kernelNodeParams.sharedMemBytes = 0;
kernelNodeParams.kernelParams = reinterpret_cast<void**>(kernelArgs2);
kernelNodeParams.extra = nullptr;
HIP_CHECK(hipGraphAddKernelNode(&kernelmod2, graph, &kernel_vecSqr, 1,
&kernelNodeParams));
nodelist.push_back(kernelmod2);
int incValue3{3};
void* kernelArgs3[] = {&C_d, &Res3_d, &incValue3,
reinterpret_cast<void *>(&NElem)};
memset(&kernelNodeParams, 0, sizeof(kernelNodeParams));
kernelNodeParams.func =
reinterpret_cast<void *>(updateResult);
kernelNodeParams.gridDim = dim3(blocks);
kernelNodeParams.blockDim = dim3(threadsPerBlock);
kernelNodeParams.sharedMemBytes = 0;
kernelNodeParams.kernelParams = reinterpret_cast<void**>(kernelArgs3);
kernelNodeParams.extra = nullptr;
HIP_CHECK(hipGraphAddKernelNode(&kernelmod3, graph, &kernel_vecSqr, 1,
&kernelNodeParams));
nodelist.push_back(kernelmod3);
// Compute sum from all dependent nodes
void* kernelArgsAdd[] = {&Res1_d, &Res2_d, &Res3_d, &Sum_d,
reinterpret_cast<void *>(&NElem)};
memset(&kernelNodeParams, 0, sizeof(kernelNodeParams));
kernelNodeParams.func =
reinterpret_cast<void *>(vectorSum);
kernelNodeParams.gridDim = dim3(blocks);
kernelNodeParams.blockDim = dim3(threadsPerBlock);
kernelNodeParams.sharedMemBytes = 0;
kernelNodeParams.kernelParams = reinterpret_cast<void**>(kernelArgsAdd);
kernelNodeParams.extra = nullptr;
HIP_CHECK(hipGraphAddKernelNode(&kernel_vecAdd, graph,
nodelist.data(), nodelist.size(),
&kernelNodeParams));
HIP_CHECK(hipGraphNodeGetDependencies(kernel_vecAdd, nullptr, &numDeps));
REQUIRE(numDeps == nodelist.size());
// Verify api When Dependencies are requested for actual number of nodes.
queryActualNumOfDeps(nodelist, kernel_vecAdd, numDeps);
// Verify api When Dependencies are requested for more than
// actual number of nodes.
queryGreaterNumOfDeps(nodelist, kernel_vecAdd, numDeps);
// Verify api When Dependencies are requested for less than
// actual number of nodes.
queryLesserNumOfDeps(nodelist, kernel_vecAdd, numDeps);
HIP_CHECK(hipGraphAddMemcpyNode1D(&memcpyD2H, graph, &kernel_vecAdd, 1,
Sum_h, Sum_d,
Nbytes, hipMemcpyDeviceToHost));
// Instantiate and launch the graph
HIP_CHECK(hipGraphInstantiate(&graphExec, graph, nullptr, nullptr, 0));
HIP_CHECK(hipGraphLaunch(graphExec, streamForGraph));
HIP_CHECK(hipStreamSynchronize(streamForGraph));
// Validate the computation
for (size_t i = 0; i < N; i++) {
if ( Sum_h[i] != ( (C_h[i] + incValue1)
+ (C_h[i] + incValue2)
+ (C_h[i] + incValue3) ) ) {
INFO("Sum not matching at " << i << " Sum_h[i] " << Sum_h[i]
<< " C_h[i] " << C_h[i]);
REQUIRE(false);
}
}
HipTest::freeArrays<int>(A_d, C_d, Sum_d, A_h, C_h, Sum_h, false);
HipTest::freeArrays<int>(Res1_d, Res2_d, Res3_d,
nullptr, nullptr, nullptr, false);
HIP_CHECK(hipGraphExecDestroy(graphExec));
HIP_CHECK(hipGraphDestroy(graph));
HIP_CHECK(hipStreamDestroy(streamForGraph));
}
/**
* Test performs api parameter validation by passing various values
* as input and output parameters and validates the behavior.
* Test will include both negative and positive scenarios.
*/
TEST_CASE("Unit_hipGraphNodeGetDependencies_ParamValidation") {
hipGraph_t graph{};
const int numBytes = 100;
size_t numDeps{1};
hipGraphNode_t memsetNode{}, depnodes{};
hipError_t ret{};
char *A_d;
HIP_CHECK(hipGraphCreate(&graph, 0));
HIP_CHECK(hipMalloc(&A_d, numBytes));
hipMemsetParams memsetParams{};
memsetParams.dst = reinterpret_cast<void*>(A_d);
memsetParams.value = 1;
memsetParams.pitch = 0;
memsetParams.elementSize = sizeof(char);
memsetParams.width = numBytes * sizeof(char);
memsetParams.height = 1;
HIP_CHECK(hipGraphAddMemsetNode(&memsetNode, graph, nullptr,
0, &memsetParams));
SECTION("node passed is a root node") {
ret = hipGraphNodeGetDependencies(memsetNode, &depnodes, &numDeps);
// Api expected to return success and no dependencies.
REQUIRE(ret == hipSuccess);
REQUIRE(numDeps == 0);
}
SECTION("node as nullptr") {
ret = hipGraphNodeGetDependencies(nullptr, &depnodes, &numDeps);
REQUIRE(ret == hipErrorInvalidValue);
}
SECTION("NumDependencies as nullptr") {
ret = hipGraphNodeGetDependencies(memsetNode, &depnodes, nullptr);
REQUIRE(ret == hipErrorInvalidValue);
}
SECTION("node as un-initialized/invalid parameter") {
hipGraphNode_t uninit_node{};
ret = hipGraphNodeGetDependencies(uninit_node, &depnodes, &numDeps);
REQUIRE(ret == hipErrorInvalidValue);
}
HIP_CHECK(hipGraphDestroy(graph));
HIP_CHECK(hipFree(A_d));
}
tests/catch/unit/graph/hipGraphNodeGetDependentNodes.cc
+366
Просмотреть файл
@@ -0,0 +1,366 @@
/*
Copyright (c) 2022 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.
*/
/**
Testcase Scenarios
------------------
Functional:
1) Create a graph and add nodes with dependencies. Query for dependent nodes of the node passed and verify the result with dependencies defined.
2) When pDependentNodes is passed as nullptr, verify pNumDependentNodes returns the number of dependent nodes.
3) When pNumDependentNodes is higher than the actual number of dependent nodes, the remaining entries in pDependentNodes will be set to NULL,
and the number of nodes actually obtained will be returned in pNumDependentNodes.
4) When pNumDependentNodes is lesser than the actual number of dependent nodes, api should return the requested number of nodes in pDependentNodes.
Argument Validation:
1) Add a single node in graph and pass the node to api. Verify the api returns dependent nodes as 0.
2) Pass node as nullptr and verify api doesn’t crash, returns error code.
3) Pass pNumDependentNodes as nullptr and verify api doesn’t crash, returns error code.
4) Pass node as un-initialized/invalid parameter and verify api returns error code.
*/
#include <hip_test_common.hh>
#include <hip_test_checkers.hh>
#include <hip_test_kernels.hh>
static __global__ void updateResult(int* C_d, int* Res_d, int val,
int64_t NELEM) {
size_t offset = (blockIdx.x * blockDim.x + threadIdx.x);
size_t stride = blockDim.x * gridDim.x;
for (int64_t i = NELEM - stride + offset; i >= 0; i -= stride) {
Res_d[i] = C_d[i] + val;
}
}
static __global__ void vectorSum(const int* A_d, const int* B_d,
const int* C_d, int* Res_d, size_t NELEM) {
size_t offset = (blockIdx.x * blockDim.x + threadIdx.x);
size_t stride = blockDim.x * gridDim.x;
for (size_t i = offset; i < NELEM; i += stride) {
Res_d[i] = A_d[i] + B_d[i] + C_d[i];
}
}
/**
* Verify api when GetDependent nodes is requested
* for actual number of nodes.
*/
static void queryActualNumOfDepNodes(const std::vector<hipGraphNode_t> &Nlist,
hipGraphNode_t kernel_vecSqr, size_t numDeps) {
hipGraphNode_t* depnodes;
int numBytes = sizeof(hipGraphNode_t) * numDeps;
depnodes = reinterpret_cast<hipGraphNode_t *>(malloc(numBytes));
REQUIRE(depnodes != nullptr);
HIP_CHECK(hipGraphNodeGetDependentNodes(kernel_vecSqr, depnodes, &numDeps));
REQUIRE(numDeps == Nlist.size());
// Verify all dependent nodes are present in the node entries returned
for (auto Node : Nlist) {
bool found = false;
for (size_t i = 0; i < numDeps; i++) {
if (Node == depnodes[i]) {
found = true;
break;
}
}
if (!found) {
INFO("Dependent node " << Node << " not present in returned list");
REQUIRE(false);
}
}
free(depnodes);
}
/**
* Verify api when GetDependent nodes queried
* for greater number than actual number of nodes.
*/
static void queryGreaterNumOfDepNodes(const std::vector<hipGraphNode_t> &Nlist,
hipGraphNode_t kernel_vecSqr, size_t numDeps) {
constexpr auto addlEntries = 4;
hipGraphNode_t* depnodes;
size_t totDeps = numDeps + addlEntries;
int numBytes = sizeof(hipGraphNode_t) * totDeps;
depnodes = reinterpret_cast<hipGraphNode_t *>(malloc(numBytes));
REQUIRE(depnodes != nullptr);
HIP_CHECK(hipGraphNodeGetDependentNodes(kernel_vecSqr, depnodes, &totDeps));
REQUIRE(totDeps == Nlist.size());
for (auto i = numDeps; i < numDeps + addlEntries; i++) {
REQUIRE(depnodes[i] == nullptr);
}
// Verify all dependent nodes are present in the node entries returned
for (auto Node : Nlist) {
bool found = false;
for (size_t i = 0; i < numDeps; i++) {
if (Node == depnodes[i]) {
found = true;
break;
}
}
if (!found) {
INFO("Dependent node " << Node << " not present in returned list");
REQUIRE(false);
}
}
free(depnodes);
}
/**
* Verify api when GetDependent nodes queried
* for lesser number than actual number of nodes.
*/
static void queryLesserNumOfDepNodes(const std::vector<hipGraphNode_t> &Nlist,
hipGraphNode_t kernel_vecSqr, size_t numDeps) {
size_t totDeps = numDeps - 1;
hipGraphNode_t* depnodes;
int numBytes = sizeof(hipGraphNode_t) * totDeps;
size_t count{};
depnodes = reinterpret_cast<hipGraphNode_t *>(malloc(numBytes));
REQUIRE(depnodes != nullptr);
HIP_CHECK(hipGraphNodeGetDependentNodes(kernel_vecSqr, depnodes, &totDeps));
REQUIRE(totDeps == Nlist.size() - 1);
// Verify all dependent nodes are present in the node entries returned
for (auto Node : Nlist) {
for (size_t i = 0; i < totDeps; i++) {
if (Node == depnodes[i]) {
count++;
break;
}
}
}
REQUIRE(count == totDeps);
free(depnodes);
}
/**
* Functional Test for getting dependent nodes in graph and verifying execution
*/
TEST_CASE("Unit_hipGraphNodeGetDependentNodes_Functional") {
constexpr size_t N = 1024;
constexpr size_t Nbytes = N * sizeof(int);
constexpr auto blocksPerCU = 6; // to hide latency
constexpr auto threadsPerBlock = 256;
hipGraphNode_t kernel_vecSqr{}, kernel_vecAdd{};
hipGraphNode_t kernelmod1{}, kernelmod2{}, kernelmod3{};
hipGraphNode_t memcpyD2H{}, memcpyH2D_A{};
hipKernelNodeParams kernelNodeParams{};
hipGraph_t graph{};
size_t numDeps{};
hipStream_t streamForGraph;
int *A_d, *C_d;
int *A_h, *C_h;
int *Res1_d, *Res2_d, *Res3_d;
int *Sum_d, *Sum_h;
hipGraphExec_t graphExec;
size_t NElem{N};
HIP_CHECK(hipStreamCreate(&streamForGraph));
HIP_CHECK(hipGraphCreate(&graph, 0));
HipTest::initArrays<int>(&A_d, &C_d, &Sum_d, &A_h, &C_h, &Sum_h, N);
HipTest::initArrays<int>(&Res1_d, &Res2_d, &Res3_d,
nullptr, nullptr, nullptr, N);
unsigned blocks = HipTest::setNumBlocks(blocksPerCU, threadsPerBlock, N);
// Initialize input buffer and vecsqr result
for (size_t i = 0; i < N; ++i) {
A_h[i] = i + 1;
C_h[i] = A_h[i] * A_h[i];
}
HIP_CHECK(hipGraphAddMemcpyNode1D(&memcpyH2D_A, graph, nullptr, 0, A_d, A_h,
Nbytes, hipMemcpyHostToDevice));
void* kernelArgsVS[] = {&A_d, &C_d, reinterpret_cast<void *>(&NElem)};
memset(&kernelNodeParams, 0, sizeof(kernelNodeParams));
kernelNodeParams.func =
reinterpret_cast<void *>(HipTest::vector_square<int>);
kernelNodeParams.gridDim = dim3(blocks);
kernelNodeParams.blockDim = dim3(threadsPerBlock);
kernelNodeParams.sharedMemBytes = 0;
kernelNodeParams.kernelParams = reinterpret_cast<void**>(kernelArgsVS);
kernelNodeParams.extra = nullptr;
HIP_CHECK(hipGraphAddKernelNode(&kernel_vecSqr, graph, &memcpyH2D_A, 1,
&kernelNodeParams));
// Create multiple nodes dependent on vecSqr node.
// Dependent nodes takes vecSqr input and computes output independently.
std::vector<hipGraphNode_t> nodelist;
int incValue1{1};
void* kernelArgs1[] = {&C_d, &Res1_d, &incValue1,
reinterpret_cast<void *>(&NElem)};
memset(&kernelNodeParams, 0, sizeof(kernelNodeParams));
kernelNodeParams.func =
reinterpret_cast<void *>(updateResult);
kernelNodeParams.gridDim = dim3(blocks);
kernelNodeParams.blockDim = dim3(threadsPerBlock);
kernelNodeParams.sharedMemBytes = 0;
kernelNodeParams.kernelParams = reinterpret_cast<void**>(kernelArgs1);
kernelNodeParams.extra = nullptr;
HIP_CHECK(hipGraphAddKernelNode(&kernelmod1, graph, &kernel_vecSqr, 1,
&kernelNodeParams));
nodelist.push_back(kernelmod1);
int incValue2{2};
void* kernelArgs2[] = {&C_d, &Res2_d, &incValue2,
reinterpret_cast<void *>(&NElem)};
memset(&kernelNodeParams, 0, sizeof(kernelNodeParams));
kernelNodeParams.func =
reinterpret_cast<void *>(updateResult);
kernelNodeParams.gridDim = dim3(blocks);
kernelNodeParams.blockDim = dim3(threadsPerBlock);
kernelNodeParams.sharedMemBytes = 0;
kernelNodeParams.kernelParams = reinterpret_cast<void**>(kernelArgs2);
kernelNodeParams.extra = nullptr;
HIP_CHECK(hipGraphAddKernelNode(&kernelmod2, graph, &kernel_vecSqr, 1,
&kernelNodeParams));
nodelist.push_back(kernelmod2);
int incValue3{3};
void* kernelArgs3[] = {&C_d, &Res3_d, &incValue3,
reinterpret_cast<void *>(&NElem)};
memset(&kernelNodeParams, 0, sizeof(kernelNodeParams));
kernelNodeParams.func =
reinterpret_cast<void *>(updateResult);
kernelNodeParams.gridDim = dim3(blocks);
kernelNodeParams.blockDim = dim3(threadsPerBlock);
kernelNodeParams.sharedMemBytes = 0;
kernelNodeParams.kernelParams = reinterpret_cast<void**>(kernelArgs3);
kernelNodeParams.extra = nullptr;
HIP_CHECK(hipGraphAddKernelNode(&kernelmod3, graph, &kernel_vecSqr, 1,
&kernelNodeParams));
nodelist.push_back(kernelmod3);
HIP_CHECK(hipGraphNodeGetDependentNodes(kernel_vecSqr, nullptr, &numDeps));
REQUIRE(numDeps == nodelist.size());
// Verify api When Dependent nodes are requested for actual number of nodes.
queryActualNumOfDepNodes(nodelist, kernel_vecSqr, numDeps);
// Verify api When Dependent nodes are requested for more than
// actual number of nodes.
queryGreaterNumOfDepNodes(nodelist, kernel_vecSqr, numDeps);
// Verify api When Dependent nodes are requested for less than
// actual number of nodes.
queryLesserNumOfDepNodes(nodelist, kernel_vecSqr, numDeps);
// Compute sum from all dependent nodes
void* kernelArgsAdd[] = {&Res1_d, &Res2_d, &Res3_d, &Sum_d,
reinterpret_cast<void *>(&NElem)};
memset(&kernelNodeParams, 0, sizeof(kernelNodeParams));
kernelNodeParams.func =
reinterpret_cast<void *>(vectorSum);
kernelNodeParams.gridDim = dim3(blocks);
kernelNodeParams.blockDim = dim3(threadsPerBlock);
kernelNodeParams.sharedMemBytes = 0;
kernelNodeParams.kernelParams = reinterpret_cast<void**>(kernelArgsAdd);
kernelNodeParams.extra = nullptr;
HIP_CHECK(hipGraphAddKernelNode(&kernel_vecAdd, graph,
nodelist.data(), nodelist.size(),
&kernelNodeParams));
HIP_CHECK(hipGraphAddMemcpyNode1D(&memcpyD2H, graph, &kernel_vecAdd, 1,
Sum_h, Sum_d,
Nbytes, hipMemcpyDeviceToHost));
// Instantiate and launch the graph
HIP_CHECK(hipGraphInstantiate(&graphExec, graph, nullptr, nullptr, 0));
HIP_CHECK(hipGraphLaunch(graphExec, streamForGraph));
HIP_CHECK(hipStreamSynchronize(streamForGraph));
// Validate the computation
for (size_t i = 0; i < N; i++) {
if ( Sum_h[i] != ( (C_h[i] + incValue1)
+ (C_h[i] + incValue2)
+ (C_h[i] + incValue3) ) ) {
INFO("Sum not matching at " << i << " Sum_h[i] " << Sum_h[i]
<< " C_h[i] " << C_h[i]);
REQUIRE(false);
}
}
HipTest::freeArrays<int>(A_d, C_d, Sum_d, A_h, C_h, Sum_h, false);
HipTest::freeArrays<int>(Res1_d, Res2_d, Res3_d,
nullptr, nullptr, nullptr, false);
HIP_CHECK(hipGraphExecDestroy(graphExec));
HIP_CHECK(hipGraphDestroy(graph));
HIP_CHECK(hipStreamDestroy(streamForGraph));
}
/**
* Test performs api parameter validation by passing various values
* as input and output parameters and validates the behavior.
* Test will include both negative and positive scenarios.
*/
TEST_CASE("Unit_hipGraphNodeGetDependentNodes_ParamValidation") {
hipGraph_t graph{};
const int numBytes = 100;
size_t numDeps{1};
hipGraphNode_t memsetNode{}, depnodes{};
hipError_t ret{};
char *A_d;
HIP_CHECK(hipGraphCreate(&graph, 0));
HIP_CHECK(hipMalloc(&A_d, numBytes));
hipMemsetParams memsetParams{};
memsetParams.dst = reinterpret_cast<void*>(A_d);
memsetParams.value = 1;
memsetParams.pitch = 0;
memsetParams.elementSize = sizeof(char);
memsetParams.width = numBytes * sizeof(char);
memsetParams.height = 1;
HIP_CHECK(hipGraphAddMemsetNode(&memsetNode, graph, nullptr,
0, &memsetParams));
SECTION("single node in graph") {
ret = hipGraphNodeGetDependentNodes(memsetNode, &depnodes, &numDeps);
// Api expected to return success and no dependent nodes.
REQUIRE(ret == hipSuccess);
REQUIRE(numDeps == 0);
}
SECTION("node as nullptr") {
ret = hipGraphNodeGetDependentNodes(nullptr, &depnodes, &numDeps);
REQUIRE(ret == hipErrorInvalidValue);
}
SECTION("NumDependentNodes as nullptr") {
ret = hipGraphNodeGetDependentNodes(memsetNode, &depnodes, nullptr);
REQUIRE(ret == hipErrorInvalidValue);
}
SECTION("node as un-initialized/invalid parameter") {
hipGraphNode_t uninit_node{};
ret = hipGraphNodeGetDependentNodes(uninit_node, &depnodes, &numDeps);
REQUIRE(ret == hipErrorInvalidValue);
}
HIP_CHECK(hipGraphDestroy(graph));
HIP_CHECK(hipFree(A_d));
}
tests/catch/unit/memory/hipMemcpy3D.cc
+6 -2
Просмотреть файл
@@ -552,9 +552,13 @@ void Memcpy3D<T>::simple_Memcpy3D() {
*/
TEMPLATE_TEST_CASE("Unit_hipMemcpy3D_Basic", "[hipMemcpy3D]",
int, unsigned int, float) {
int numDevices = 0;
auto i = GENERATE(10, 100, 1024, 10*1024);
int device = -1;
HIP_CHECK(hipGetDevice(&device));
hipDeviceProp_t prop;
HIP_CHECK(hipGetDeviceProperties(&prop,device));
auto i = GENERATE_COPY(10, 100, 1024, prop.maxTexture3D[0]);
auto j = GENERATE(10, 100);
int numDevices = 0;
HIP_CHECK(hipGetDeviceCount(&numDevices));
if (numDevices > 1) {
if (std::is_same<TestType, float>::value) {
tests/catch/unit/memory/hipMemcpy3DAsync.cc
+5 -1
Просмотреть файл
@@ -662,7 +662,11 @@ TEMPLATE_TEST_CASE("Unit_hipMemcpy3DAsync_Basic",
int, unsigned int, float) {
int numDevices = 0;
HIP_CHECK(hipGetDeviceCount(&numDevices));
auto i = GENERATE(10, 100, 1024, 10*1024);
int device = -1;
HIP_CHECK(hipGetDevice(&device));
hipDeviceProp_t prop;
HIP_CHECK(hipGetDeviceProperties(&prop,device));
auto i = GENERATE_COPY(10, 100, 1024, prop.maxTexture3D[0]);
auto j = GENERATE(10, 100);
if (numDevices > 1) {
if (std::is_same<TestType, int>::value) {
tests/src/runtimeApi/memory/hipHostMallocTests.cpp
+6 -2
Просмотреть файл
@@ -61,7 +61,7 @@ Testcase Scenarios :
*/
#define LIB_ROCMSMI "librocm_smi64.so"
#define NUM_BYTES 1000
#define ALLOC_SIZE (1*1024*1024)
#define ALLOC_SIZE (30*1024*1024)
/**
@@ -246,7 +246,11 @@ bool validatePageTableAllocations(const char *devList, int devCnt) {
}
for (indx = 0; indx < numdev; indx++) {
if (current[indx] - prev[indx])
// For visible hip devices, there should be increase in VRAM usage
// due to page table allocations
// For NON visible hip devices, there can be reduction in VRAM usage
// due to removal of page tables from them
if (current[indx] > prev[indx])
changeCnt++;
}