TheRock/rocm-systems/projects/hip-tests/catch/unit/graph/hipGraphKernelNodeSetParams.cc

/*
 * Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
 *
 * SPDX-License-Identifier: MIT
 */

/**
Test Case Scenarios :
Negative -
1) Pass node as nullptr and verify api returns error code.
2) Pass pNodeParams as nullptr and verify api returns error code.
Functional -
1) Add kernel node to graph with certain kernel params, now update the kernel
   node params with set and check taking effect after launching graph.
2) Add kernel node to graph with certain kernel params, now get kernel node parameters
   with hipGraphKernelNodeGetParams, then update the kernel node params with
   hipGraphKernelNodeSetParams, finally check taking effect after launching graph.
*/

#include <hip_test_checkers.hh>
#include <hip_test_common.hh>
#include <hip_test_kernels.hh>

/* Test verifies hipGraphKernelNodeSetParams API Negative scenarios.
 */
HIP_TEST_CASE(Unit_hipGraphKernelNodeSetParams_Negative) {
  constexpr int N = 1024;
  size_t NElem{N};
  constexpr auto blocksPerCU = 6;  // to hide latency
  constexpr auto threadsPerBlock = 256;
  unsigned blocks = HipTest::setNumBlocks(blocksPerCU, threadsPerBlock, N);
  int *A_d, *B_d, *C_d;
  hipGraph_t graph;
  hipGraphNode_t kNode;
  hipKernelNodeParams kNodeParams{};

  HIP_CHECK(hipMalloc(&A_d, sizeof(int) * N));
  HIP_CHECK(hipMalloc(&B_d, sizeof(int) * N));
  HIP_CHECK(hipMalloc(&C_d, sizeof(int) * N));
  HIP_CHECK(hipGraphCreate(&graph, 0));

  void* kernelArgs[] = {&A_d, &B_d, &C_d, reinterpret_cast<void*>(&NElem)};
  kNodeParams.func = reinterpret_cast<void*>(HipTest::vectorADD<int>);
  kNodeParams.gridDim = dim3(blocks);
  kNodeParams.blockDim = dim3(threadsPerBlock);
  kNodeParams.kernelParams = reinterpret_cast<void**>(kernelArgs);
  HIP_CHECK(hipGraphAddKernelNode(&kNode, graph, nullptr, 0, &kNodeParams));

  SECTION("Pass node as nullptr") {
    HIP_CHECK_ERROR(hipGraphKernelNodeSetParams(nullptr, &kNodeParams), hipErrorInvalidValue);
  }

  SECTION("Pass kNodeParams as nullptr") {
    HIP_CHECK_ERROR(hipGraphKernelNodeSetParams(kNode, nullptr), hipErrorInvalidValue);
  }

#if HT_NVIDIA  // on AMD this returns hipErrorInvalidValue
  SECTION("Pass NodeParams func data member as nullptr") {
    kNodeParams.func = nullptr;
    HIP_CHECK_ERROR(hipGraphKernelNodeSetParams(kNode, &kNodeParams),
                    hipErrorInvalidDeviceFunction);
  }
#endif

#if HT_NVIDIA  // segfaults on AMD
  SECTION("Pass kernelParams data member as nullptr") {
    kNodeParams.kernelParams = nullptr;
    HIP_CHECK_ERROR(hipGraphKernelNodeSetParams(kNode, &kNodeParams), hipErrorInvalidValue);
  }
#endif

#if HT_NVIDIA  // segfaults on AMD
  SECTION("node is not a kernel node") {
    hipGraphNode_t empty_node;
    HIP_CHECK(hipGraphAddEmptyNode(&empty_node, graph, nullptr, 0));
    HIP_CHECK_ERROR(hipGraphKernelNodeSetParams(empty_node, &kNodeParams), hipErrorInvalidValue);
  }
#endif

  HIP_CHECK(hipFree(A_d));
  HIP_CHECK(hipFree(B_d));
  HIP_CHECK(hipFree(C_d));
  HIP_CHECK(hipGraphDestroy(graph));
}

/**
 * Functional Test for API Set Kernel Params
 */
HIP_TEST_CASE(Unit_hipGraphKernelNodeSetParams_Functional) {
  constexpr size_t N = 1024;
  constexpr size_t Nbytes = N * sizeof(int);
  constexpr auto blocksPerCU = 6;  // to hide latency
  constexpr auto threadsPerBlock = 256;
  unsigned blocks = HipTest::setNumBlocks(blocksPerCU, threadsPerBlock, N);
  hipGraph_t graph;
  hipGraphNode_t memcpyNode, kNode;
  hipKernelNodeParams kNodeParams{}, kNodeParams1{};
  hipStream_t streamForGraph;
  int *A_d, *B_d, *C_d;
  int *A_h, *B_h, *C_h;
  std::vector<hipGraphNode_t> dependencies;
  hipGraphExec_t graphExec;
  size_t NElem{N};

  HIP_CHECK(hipStreamCreate(&streamForGraph));
  HipTest::initArrays(&A_d, &B_d, &C_d, &A_h, &B_h, &C_h, N, false);
  HIP_CHECK(hipGraphCreate(&graph, 0));

  HIP_CHECK(hipGraphAddMemcpyNode1D(&memcpyNode, graph, nullptr, 0, A_d, A_h, Nbytes,
                                    hipMemcpyHostToDevice));
  dependencies.push_back(memcpyNode);
  HIP_CHECK(hipGraphAddMemcpyNode1D(&memcpyNode, graph, nullptr, 0, B_d, B_h, Nbytes,
                                    hipMemcpyHostToDevice));
  dependencies.push_back(memcpyNode);

  void* kernelArgs[] = {&A_d, &B_d, &C_d, reinterpret_cast<void*>(&NElem)};
  kNodeParams.func = reinterpret_cast<void*>(HipTest::vectorADD<int>);
  kNodeParams.gridDim = dim3(blocks);
  kNodeParams.blockDim = dim3(threadsPerBlock);
  kNodeParams.kernelParams = reinterpret_cast<void**>(kernelArgs);
  HIP_CHECK(
      hipGraphAddKernelNode(&kNode, graph, dependencies.data(), dependencies.size(), &kNodeParams));

  kNodeParams1.func = reinterpret_cast<void*>(HipTest::vectorSUB<int>);
  kNodeParams1.gridDim = dim3(blocks);
  kNodeParams1.blockDim = dim3(threadsPerBlock);
  kNodeParams1.kernelParams = reinterpret_cast<void**>(kernelArgs);
  HIP_CHECK(hipGraphKernelNodeSetParams(kNode, &kNodeParams1));

  dependencies.clear();
  dependencies.push_back(kNode);
  HIP_CHECK(hipGraphAddMemcpyNode1D(&memcpyNode, graph, dependencies.data(), dependencies.size(),
                                    C_h, C_d, Nbytes, hipMemcpyDeviceToHost));

  // Instantiate and launch the graph
  HIP_CHECK(hipGraphInstantiate(&graphExec, graph, NULL, NULL, 0));
  HIP_CHECK(hipGraphLaunch(graphExec, streamForGraph));
  HIP_CHECK(hipStreamSynchronize(streamForGraph));

  // Verify graph execution result
  HipTest::checkVectorSUB<int>(A_h, B_h, C_h, N);

  HipTest::freeArrays(A_d, B_d, C_d, A_h, B_h, C_h, false);
  HIP_CHECK(hipGraphExecDestroy(graphExec));
  HIP_CHECK(hipGraphDestroy(graph));
  HIP_CHECK(hipStreamDestroy(streamForGraph));
}

static __global__ void ker_vec_add(int* A, int* B) {
  int i = threadIdx.x + blockDim.x * blockIdx.x;
  A[i] = A[i] + B[i];
}

static __global__ void ker_vec_sub(int* A, int* B) {
  int i = threadIdx.x + blockDim.x * blockIdx.x;
  A[i] = A[i] - B[i];
}

// Internal class for creating nested graphs.
class GraphKernelNodeGetSetParam {
  const int N = 1024;
  size_t Nbytes;
  const int threadsPerBlock = 256;
  const int blocks = (N / threadsPerBlock);
  hipGraphNode_t memcpyH2D_A1, memcpyH2D_A2, memcpyD2H_A3, vec_maths;
  hipGraph_t graph;
  hipKernelNodeParams kerNodeParams{};
  int *A1_d, *A2_d, *A1_h, *A2_h, *A3_h;

 public:
  // Create a nested Graph
  GraphKernelNodeGetSetParam() {
    Nbytes = N * sizeof(int);
    // Allocate device buffers
    HIP_CHECK(hipMalloc(&A1_d, Nbytes));
    HIP_CHECK(hipMalloc(&A2_d, Nbytes));
    // Allocate host buffers
    A1_h = reinterpret_cast<int*>(malloc(Nbytes));
    REQUIRE(A1_h != nullptr);
    A2_h = reinterpret_cast<int*>(malloc(Nbytes));
    REQUIRE(A2_h != nullptr);
    A3_h = reinterpret_cast<int*>(malloc(Nbytes));
    REQUIRE(A3_h != nullptr);
    // Create all the 3 level graphs
    HIP_CHECK(hipGraphCreate(&graph, 0));
    void* kernelArgs[] = {&A1_d, &A2_d};
    kerNodeParams.func = reinterpret_cast<void*>(ker_vec_add);
    kerNodeParams.gridDim = dim3(blocks);
    kerNodeParams.blockDim = dim3(threadsPerBlock);
    kerNodeParams.kernelParams = reinterpret_cast<void**>(kernelArgs);
    HIP_CHECK(hipGraphAddKernelNode(&vec_maths, graph, nullptr, 0, &kerNodeParams));
    // Add nodes to graph
    HIP_CHECK(hipGraphAddMemcpyNode1D(&memcpyH2D_A1, graph, nullptr, 0, A1_d, A1_h, Nbytes,
                                      hipMemcpyHostToDevice));
    HIP_CHECK(hipGraphAddMemcpyNode1D(&memcpyH2D_A2, graph, nullptr, 0, A2_d, A2_h, Nbytes,
                                      hipMemcpyHostToDevice));
    HIP_CHECK(hipGraphAddMemcpyNode1D(&memcpyD2H_A3, graph, nullptr, 0, A3_h, A1_d, Nbytes,
                                      hipMemcpyDeviceToHost));
    HIP_CHECK(hipGraphAddDependencies(graph, &memcpyH2D_A1, &vec_maths, 1));
    HIP_CHECK(hipGraphAddDependencies(graph, &memcpyH2D_A2, &vec_maths, 1));
    HIP_CHECK(hipGraphAddDependencies(graph, &vec_maths, &memcpyD2H_A3, 1));
  }

  // Fill Random Input Data
  void fillRandInpData() {
    for (int i = 0; i < N; i++) {
      A1_h[i] = (rand() % 256);  // NOLINT
      A2_h[i] = (rand() % 256);  // NOLINT
    }
  }

  hipGraph_t* getRootGraph() { return &graph; }

  void updateNode() {
    size_t numNodes = 0;
    HIP_CHECK(hipGraphGetNodes(graph, nullptr, &numNodes));
    hipGraphNode_t* nodes =
        reinterpret_cast<hipGraphNode_t*>(malloc(numNodes * sizeof(hipGraphNode_t)));
    HIP_CHECK(hipGraphGetNodes(graph, nodes, &numNodes));
    // Get the Graph node from the embedded graph
    size_t nodeIdx = 0;
    for (size_t idx = 0; idx < numNodes; idx++) {
      hipGraphNodeType nodeType;
      HIP_CHECK(hipGraphNodeGetType(nodes[idx], &nodeType));
      if (nodeType == hipGraphNodeTypeKernel) {
        nodeIdx = idx;
        break;
      }
    }
    hipKernelNodeParams nodeParam;
    HIP_CHECK(hipGraphKernelNodeGetParams(nodes[nodeIdx], &nodeParam));
    nodeParam.func = reinterpret_cast<void*>(ker_vec_sub);
    HIP_CHECK(hipGraphKernelNodeSetParams(nodes[nodeIdx], &nodeParam));
    free(nodes);
  }

  // Function to validate result
  void validateOutData() { HipTest::checkVectorSUB<int>(A1_h, A2_h, A3_h, N); }

  // Destroy resources
  ~GraphKernelNodeGetSetParam() {
    // Free all allocated buffers
    HIP_CHECK(hipFree(A2_d));
    HIP_CHECK(hipFree(A1_d));
    free(A3_h);
    free(A2_h);
    free(A1_h);
    HIP_CHECK(hipGraphDestroy(graph));
  }
};

HIP_TEST_CASE(Unit_hipGraphKernelNodeGetSetParams_Functional) {
  hipGraph_t* graph;
  hipStream_t streamForGraph;
  hipGraphExec_t graphExec;
  GraphKernelNodeGetSetParam GraphKernelNodeGetSetParamObj;
  graph = GraphKernelNodeGetSetParamObj.getRootGraph();
  GraphKernelNodeGetSetParamObj.updateNode();
  HIP_CHECK(hipStreamCreate(&streamForGraph));
  // Instantiate and launch the childgraph
  HIP_CHECK(hipGraphInstantiate(&graphExec, (*graph), nullptr, nullptr, 0));
  GraphKernelNodeGetSetParamObj.fillRandInpData();
  HIP_CHECK(hipGraphLaunch(graphExec, streamForGraph));
  HIP_CHECK(hipStreamSynchronize(streamForGraph));
  GraphKernelNodeGetSetParamObj.validateOutData();
  HIP_CHECK(hipStreamDestroy(streamForGraph));
  HIP_CHECK(hipGraphExecDestroy(graphExec));
}