amilanov-amd
345 строки
14 KiB
C++
345 строки
14 KiB
C++
/*
|
|
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.
|
|
*/
|
|
|
|
/**
|
|
Testcase Scenarios of hipGraphAddMemcpyNodeToSymbol API:
|
|
|
|
Functional :
|
|
|
|
1. Allocate global symbol memory, add the MemcpyNodeToSymbol
|
|
node to the graph and verify for different memory kinds
|
|
2. Allocate const memory add the MemcpyNodeToSymbol node to
|
|
the graph and verify for different memory kinds
|
|
3. Allocate global symbol memory and device memory in GPU-0
|
|
and perform MemcpyToSymbol from peer GPU by adding it to the graph node.
|
|
4. Allocate const symbol memory and device memory in GPU-0
|
|
and perform MemcpyToSymbol from peer GPU by adding it to the graph node.
|
|
5. Allocate global memory, Add MemcpyToSymbolNode,KernelNode and memcpynode and validating
|
|
the behaviour
|
|
|
|
Negative :
|
|
|
|
1) Pass nullptr to graph node
|
|
2) Pass nullptr to graph
|
|
3) Pass nullptr to dependencies
|
|
4) Pass invalid numDependencies
|
|
5) Pass nullptr to dst
|
|
6) Pass nullptr to symbol
|
|
7) Pass invalid count
|
|
8) Pass offset+count greater than allocated size
|
|
9) Pass unintialized graph
|
|
*/
|
|
|
|
#include <hip_test_common.hh>
|
|
#include <hip_test_checkers.hh>
|
|
#include <limits>
|
|
#define SIZE 256
|
|
|
|
__device__ int globalIn[SIZE];
|
|
__device__ __constant__ int globalConst[SIZE];
|
|
|
|
__global__ void MemcpyToSymbolKernel(int* B_d) {
|
|
for (int i = 0; i < SIZE; i++) {
|
|
B_d[i] = globalIn[i];
|
|
}
|
|
}
|
|
|
|
/* This testcase verifies negative scenarios of
|
|
hipGraphAddMemcpyNodeToSymbol API */
|
|
TEST_CASE("Unit_hipGraphAddMemcpyNodeToSymbol_Negative") {
|
|
constexpr size_t Nbytes = SIZE * sizeof(int);
|
|
int* A_d{nullptr};
|
|
int *A_h{nullptr}, *B_h{nullptr};
|
|
HipTest::initArrays<int>(&A_d, nullptr, nullptr, &A_h, &B_h, nullptr, SIZE, false);
|
|
|
|
hipGraph_t graph;
|
|
hipGraphNode_t 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 HT_NVIDIA
|
|
hipGraphNode_t memcpyToSymbolNode;
|
|
SECTION("Passing nullptr to graph") {
|
|
REQUIRE(hipGraphAddMemcpyNodeToSymbol(&memcpyToSymbolNode, nullptr, dependencies.data(),
|
|
dependencies.size(), HIP_SYMBOL(globalIn), A_h, Nbytes, 0,
|
|
hipMemcpyDeviceToDevice) == hipErrorInvalidValue);
|
|
}
|
|
|
|
SECTION("Passing nullptr to graph node") {
|
|
REQUIRE(hipGraphAddMemcpyNodeToSymbol(nullptr, graph, dependencies.data(), dependencies.size(),
|
|
HIP_SYMBOL(globalIn), A_d, Nbytes, 0,
|
|
hipMemcpyDeviceToDevice) == hipErrorInvalidValue);
|
|
}
|
|
|
|
SECTION("Passing size > 1 and dependencies as nullptr") {
|
|
REQUIRE(hipGraphAddMemcpyNodeToSymbol(&memcpyToSymbolNode, graph, nullptr, 1,
|
|
HIP_SYMBOL(globalIn), A_d, Nbytes, 0,
|
|
hipMemcpyDeviceToDevice) == hipErrorInvalidValue);
|
|
}
|
|
|
|
SECTION("Passing invalid dependencies size") {
|
|
REQUIRE(hipGraphAddMemcpyNodeToSymbol(&memcpyToSymbolNode, graph, dependencies.data(), 10,
|
|
HIP_SYMBOL(globalIn), A_d, Nbytes, 0,
|
|
hipMemcpyDeviceToDevice) == hipErrorInvalidValue);
|
|
}
|
|
|
|
SECTION("Passing nullptr to dst") {
|
|
REQUIRE(hipGraphAddMemcpyNodeToSymbol(&memcpyToSymbolNode, graph, dependencies.data(),
|
|
dependencies.size(), nullptr, A_d, Nbytes, 0,
|
|
hipMemcpyDeviceToDevice) == hipErrorInvalidSymbol);
|
|
}
|
|
|
|
SECTION("Passing nullptr to source") {
|
|
REQUIRE(hipGraphAddMemcpyNodeToSymbol(&memcpyToSymbolNode, graph, dependencies.data(),
|
|
dependencies.size(), HIP_SYMBOL(globalIn), nullptr,
|
|
Nbytes, 0,
|
|
hipMemcpyDeviceToDevice) == hipErrorInvalidValue);
|
|
}
|
|
|
|
SECTION("Passing offset+size > max size") {
|
|
REQUIRE(hipGraphAddMemcpyNodeToSymbol(&memcpyToSymbolNode, graph, dependencies.data(),
|
|
dependencies.size(), HIP_SYMBOL(globalIn), A_d, Nbytes,
|
|
10, hipMemcpyDeviceToDevice) == hipErrorInvalidValue);
|
|
}
|
|
|
|
SECTION("Passing Max count") {
|
|
REQUIRE(hipGraphAddMemcpyNodeToSymbol(&memcpyToSymbolNode, graph, dependencies.data(),
|
|
dependencies.size(), HIP_SYMBOL(globalIn), A_d,
|
|
std::numeric_limits<int>::max(), 0,
|
|
hipMemcpyDeviceToDevice) == hipErrorInvalidValue);
|
|
}
|
|
|
|
SECTION("Pass Unintialized graph") {
|
|
hipGraph_t unint_graph;
|
|
REQUIRE(hipGraphAddMemcpyNodeToSymbol(&memcpyToSymbolNode, unint_graph, dependencies.data(),
|
|
dependencies.size(), HIP_SYMBOL(globalIn), A_d, Nbytes, 0,
|
|
hipMemcpyDeviceToDevice) == hipErrorInvalidValue);
|
|
}
|
|
#endif
|
|
|
|
HipTest::freeArrays<int>(A_d, nullptr, nullptr, A_h, B_h, nullptr, false);
|
|
HIP_CHECK(hipGraphDestroy(graph));
|
|
}
|
|
/*
|
|
This function is used to verify the following scenarios
|
|
1. Create global variable, allocate Memory in GPU-0 and create dependency graph of
|
|
hipGraphAddMemcpyNodeToSymbol API in GPU-1 and validate the result
|
|
2. Allocate global memory, Create dependency graph and validate the result on GPU-0
|
|
3. Allocate global const memory, Create dependency graph and validate the result on GPU-0
|
|
4. Create global const variable, allocate Memory in GPU-0 and create dependency graph of
|
|
hipGraphAddMemcpyNodeToSymbol API in GPU-1 and validate the result
|
|
*/
|
|
void hipGraphAddMemcpyNodeToSymbol_GlobalMemory(bool device_ctxchg = false,
|
|
bool const_device_var = false) {
|
|
constexpr size_t Nbytes = SIZE * sizeof(int);
|
|
int* A_d{nullptr};
|
|
int *A_h{nullptr}, *B_h{nullptr};
|
|
HipTest::initArrays<int>(&A_d, nullptr, nullptr, &A_h, &B_h, nullptr, SIZE, false);
|
|
|
|
hipGraph_t graph;
|
|
hipGraphExec_t graphExec;
|
|
hipGraphNode_t memcpyToSymbolNode, memcpyFromSymbolNode, memcpyH2D_A;
|
|
std::vector<hipGraphNode_t> dependencies;
|
|
HIP_CHECK(hipGraphCreate(&graph, 0));
|
|
|
|
if (device_ctxchg) {
|
|
HIP_CHECK(hipSetDevice(1));
|
|
HIP_CHECK(hipDeviceEnablePeerAccess(0, 0));
|
|
}
|
|
// Adding MemcpyNode
|
|
HIP_CHECK(hipGraphAddMemcpyNode1D(&memcpyH2D_A, graph, nullptr, 0, A_d, A_h, Nbytes,
|
|
hipMemcpyHostToDevice));
|
|
dependencies.push_back(memcpyH2D_A);
|
|
|
|
// Adding MemcpyNodeToSymbol
|
|
|
|
if (const_device_var) {
|
|
HIP_CHECK(hipGraphAddMemcpyNodeToSymbol(&memcpyToSymbolNode, graph, dependencies.data(),
|
|
dependencies.size(), HIP_SYMBOL(globalConst), A_d,
|
|
Nbytes, 0, hipMemcpyDeviceToDevice));
|
|
} else {
|
|
HIP_CHECK(hipGraphAddMemcpyNodeToSymbol(&memcpyToSymbolNode, graph, dependencies.data(),
|
|
dependencies.size(), HIP_SYMBOL(globalIn), A_d, Nbytes,
|
|
0, hipMemcpyDeviceToDevice));
|
|
}
|
|
dependencies.clear();
|
|
dependencies.push_back(memcpyToSymbolNode);
|
|
|
|
// Adding MemcpyNodeFromSymbol
|
|
if (const_device_var) {
|
|
HIP_CHECK(hipGraphAddMemcpyNodeFromSymbol(&memcpyFromSymbolNode, graph, dependencies.data(),
|
|
dependencies.size(), B_h, HIP_SYMBOL(globalConst),
|
|
Nbytes, 0, hipMemcpyDeviceToHost));
|
|
} else {
|
|
HIP_CHECK(hipGraphAddMemcpyNodeFromSymbol(&memcpyFromSymbolNode, graph, dependencies.data(),
|
|
dependencies.size(), B_h, HIP_SYMBOL(globalIn),
|
|
Nbytes, 0, hipMemcpyDeviceToHost));
|
|
}
|
|
|
|
// Instantiate and launch the graph
|
|
HIP_CHECK(hipGraphInstantiate(&graphExec, graph, nullptr, nullptr, 0));
|
|
HIP_CHECK(hipGraphLaunch(graphExec, 0));
|
|
HIP_CHECK(hipStreamSynchronize(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, nullptr, nullptr, A_h, B_h, nullptr, false);
|
|
HIP_CHECK(hipGraphExecDestroy(graphExec));
|
|
HIP_CHECK(hipGraphDestroy(graph));
|
|
}
|
|
/*
|
|
This testcase verifies allocating global symbol memory,
|
|
add the MemcpyNodeToSymbol node to the graph and
|
|
erifying the result
|
|
*/
|
|
TEST_CASE("Unit_hipGraphAddMemcpyNodeToSymbol_GlobalMemory") {
|
|
hipGraphAddMemcpyNodeToSymbol_GlobalMemory(false, false);
|
|
}
|
|
|
|
/*
|
|
This testcase verifies allocating global const symbol memory,
|
|
add the MemcpyNodeToSymbol node to the graph and
|
|
verifying the result
|
|
*/
|
|
TEST_CASE("Unit_hipGraphAddMemcpyNodeToSymbol_GlobalConstMemory") {
|
|
hipGraphAddMemcpyNodeToSymbol_GlobalMemory(false, true);
|
|
}
|
|
|
|
#if HT_NVIDIA
|
|
/*
|
|
This testcase verifies allocating global symbol memory and device variables
|
|
in GPU-0 and add the MemcpyNodeToSymbol node to the graph and
|
|
verifying the result in GPU-1
|
|
*/
|
|
TEST_CASE("Unit_hipGraphAddMemcpyNodeToSymbol_GlobalMemoryPeerDevice") {
|
|
int numDevices = 0;
|
|
int canAccessPeer = 0;
|
|
HIP_CHECK(hipGetDeviceCount(&numDevices));
|
|
if (numDevices > 1) {
|
|
hipDeviceCanAccessPeer(&canAccessPeer, 0, 1);
|
|
if (canAccessPeer) {
|
|
hipGraphAddMemcpyNodeToSymbol_GlobalMemory(true, false);
|
|
} else {
|
|
SUCCEED("Machine does not seem to have P2P");
|
|
}
|
|
} else {
|
|
SUCCEED("skipped the testcase as no of devices is less than 2");
|
|
}
|
|
}
|
|
/*
|
|
This testcase verifies allocating global const symbol memory and device variables
|
|
in GPU-0 and add the MemcpyNodeToSymbol node to the graph and
|
|
verifying the result in GPU-1
|
|
*/
|
|
TEST_CASE("Unit_hipGraphAddMemcpyNodeToSymbol_GlobalConstMemoryPeerDevice",
|
|
"[multigpu]") {
|
|
int numDevices = 0;
|
|
int canAccessPeer = 0;
|
|
HIP_CHECK(hipGetDeviceCount(&numDevices));
|
|
if (numDevices > 1) {
|
|
hipDeviceCanAccessPeer(&canAccessPeer, 0, 1);
|
|
if (canAccessPeer) {
|
|
hipGraphAddMemcpyNodeToSymbol_GlobalMemory(true, true);
|
|
} else {
|
|
SUCCEED("Machine does not seem to have P2P");
|
|
}
|
|
} else {
|
|
SUCCEED("skipped the testcase as no of devices is less than 2");
|
|
}
|
|
}
|
|
#endif
|
|
/*
|
|
This testcaser verifies allocating global memory,
|
|
Add MemcpyToSymbolNode,KernelNode and memcpynode and validating
|
|
the behaviour
|
|
*/
|
|
TEST_CASE("Unit_hipGraphAddMemcpyNodeToSymbol_MemcpyToSymbolNodeWithKernel",
|
|
"[multigpu]") {
|
|
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};
|
|
int *A_h{nullptr}, *B_h{nullptr};
|
|
HipTest::initArrays<int>(&A_d, &B_d, nullptr, &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);
|
|
|
|
HIP_CHECK(hipGraphAddMemcpyNodeToSymbol(&memcpyToSymbolNode, graph, dependencies.data(),
|
|
dependencies.size(), HIP_SYMBOL(globalIn), A_d, Nbytes, 0,
|
|
hipMemcpyDeviceToDevice));
|
|
dependencies.clear();
|
|
dependencies.push_back(memcpyToSymbolNode);
|
|
|
|
// Adding Kernel node
|
|
void* kernelArgs1[] = {&B_d};
|
|
kernelNodeParams.func = reinterpret_cast<void*>(MemcpyToSymbolKernel);
|
|
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 and launch the graph
|
|
HIP_CHECK(hipGraphInstantiate(&graphExec, graph, nullptr, nullptr, 0));
|
|
HIP_CHECK(hipGraphLaunch(graphExec, 0));
|
|
HIP_CHECK(hipStreamSynchronize(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, nullptr, A_h, B_h, nullptr, false);
|
|
HIP_CHECK(hipGraphExecDestroy(graphExec));
|
|
HIP_CHECK(hipGraphDestroy(graph));
|
|
}
|