TheRock/rocm-systems/projects/hip-tests/catch/unit/graph/graph_dependency_common.hh

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

#pragma once

#include <hip/hip_runtime_api.h>
#include <hip_test_common.hh>
#include <hip_test_kernels.hh>
#include <resource_guards.hh>

template <typename T> __global__ void updateResult(T* C_d, T* Res_d, T val,
                                                  int NELEM) {
  size_t offset = (blockIdx.x * blockDim.x + threadIdx.x);
  size_t stride = blockDim.x * gridDim.x;

  for (int i = NELEM - stride + offset; i >= 0; i -= stride) {
    Res_d[i] = C_d[i] + val;
  }
}

template <typename T> __global__ void vectorSum(const T* A_d, const T* B_d,
                                 const T* C_d, T* 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];
  }
}

template <typename T>
void graphNodesCommon(hipGraph_t& graph, T* hostMem1, T* devMem1, T* hostMem2, T* devMem2,
                        T* hostMem3, T* devMem3, size_t N, std::vector<hipGraphNode_t>& from,
                        std::vector<hipGraphNode_t>& to, std::vector<hipGraphNode_t>& nodelist) {
  size_t Nbytes = N * sizeof(T);
  constexpr auto blocksPerCU = 6;  // to hide latency
  constexpr auto threadsPerBlock = 256;
  hipGraphNode_t memset_A, memset_B, memsetKer_C;
  hipGraphNode_t memcpyH2D_A, memcpyH2D_B, memcpyD2H_C;
  hipGraphNode_t kernel_vecAdd;
  hipKernelNodeParams kernelNodeParams{};
  hipMemsetParams memsetParams{};
  int memsetVal{};
  size_t NElem{N};

  unsigned blocks = HipTest::setNumBlocks(blocksPerCU, threadsPerBlock, N);

  memset(&memsetParams, 0, sizeof(memsetParams));
  memsetParams.dst = reinterpret_cast<void*>(devMem1);
  memsetParams.value = 0;
  memsetParams.pitch = 0;
  memsetParams.elementSize = sizeof(T);
  memsetParams.width = N;
  memsetParams.height = 1;
  HIP_CHECK(hipGraphAddMemsetNode(&memset_A, graph, nullptr, 0, &memsetParams));

  from.push_back(memset_A);

  memset(&memsetParams, 0, sizeof(memsetParams));
  memsetParams.dst = reinterpret_cast<void*>(devMem2);
  memsetParams.value = 0;
  memsetParams.pitch = 0;
  memsetParams.elementSize = sizeof(T);
  memsetParams.width = N;
  memsetParams.height = 1;
  HIP_CHECK(hipGraphAddMemsetNode(&memset_B, graph, nullptr, 0, &memsetParams));

  from.push_back(memset_B);

  void* kernelArgs1[] = {&devMem3, &memsetVal, reinterpret_cast<void*>(&NElem)};
  kernelNodeParams.func = reinterpret_cast<void*>(HipTest::memsetReverse<uint32_t>);
  kernelNodeParams.gridDim = dim3(blocks);
  kernelNodeParams.blockDim = dim3(threadsPerBlock);
  kernelNodeParams.sharedMemBytes = 0;
  kernelNodeParams.kernelParams = reinterpret_cast<void**>(kernelArgs1);
  kernelNodeParams.extra = nullptr;
  HIP_CHECK(hipGraphAddKernelNode(&memsetKer_C, graph, nullptr, 0, &kernelNodeParams));

  HIP_CHECK(hipGraphAddMemcpyNode1D(&memcpyH2D_A, graph, nullptr, 0, devMem1, hostMem1, Nbytes,
                                    hipMemcpyHostToDevice));

  from.push_back(memcpyH2D_A);
  to.push_back(memcpyH2D_A);

  HIP_CHECK(hipGraphAddMemcpyNode1D(&memcpyH2D_B, graph, nullptr, 0, devMem2, hostMem2, Nbytes,
                                    hipMemcpyHostToDevice));

  from.push_back(memcpyH2D_B);
  to.push_back(memcpyH2D_B);
  from.push_back(memsetKer_C);

  HIP_CHECK(hipGraphAddMemcpyNode1D(&memcpyD2H_C, graph, nullptr, 0, hostMem3, devMem3, Nbytes,
                                    hipMemcpyDeviceToHost));

  void* kernelArgs2[] = {&devMem1, &devMem2, &devMem3, reinterpret_cast<void*>(&NElem)};
  kernelNodeParams.func = reinterpret_cast<void*>(HipTest::vectorADD<T>);
  kernelNodeParams.gridDim = dim3(blocks);
  kernelNodeParams.blockDim = dim3(threadsPerBlock);
  kernelNodeParams.sharedMemBytes = 0;
  kernelNodeParams.kernelParams = reinterpret_cast<void**>(kernelArgs2);
  kernelNodeParams.extra = nullptr;
  HIP_CHECK(hipGraphAddKernelNode(&kernel_vecAdd, graph, nullptr, 0, &kernelNodeParams));

  from.push_back(kernel_vecAdd);
  to.push_back(kernel_vecAdd);
  to.push_back(kernel_vecAdd);
  to.push_back(kernel_vecAdd);
  to.push_back(memcpyD2H_C);


  nodelist.push_back(memset_A);
  nodelist.push_back(memset_B);
  nodelist.push_back(memsetKer_C);
  nodelist.push_back(memcpyH2D_A);
  nodelist.push_back(memcpyH2D_B);
  nodelist.push_back(kernel_vecAdd);
  nodelist.push_back(memcpyD2H_C);
}

template <typename T>
void captureNodesCommon(hipGraph_t& graph, T* hostMem1, T* devMem1, T* hostMem2, T* devMem2,
                          T* hostMem3, T* devMem3, size_t N, std::vector<hipStream_t>& streams,
                          std::vector<hipEvent_t>& events) {
  size_t Nbytes = N * sizeof(T);
  constexpr unsigned threadsPerBlock = 256;
  constexpr auto blocksPerCU = 6;  // to hide latency
  size_t NElem{N};
  int memsetVal{0};

  unsigned blocks = HipTest::setNumBlocks(blocksPerCU, threadsPerBlock, N);

  HIP_CHECK(hipStreamBeginCapture(streams[0], hipStreamCaptureModeGlobal));
  HIP_CHECK(hipEventRecord(events[0], streams[0]));
  HIP_CHECK(hipStreamWaitEvent(streams[1], events[0], 0));
  HIP_CHECK(hipStreamWaitEvent(streams[2], events[0], 0));
  // Add operations to stream3
  hipLaunchKernelGGL(HipTest::memsetReverse<T>, dim3(blocks), dim3(threadsPerBlock), 0, streams[2],
                     devMem3, memsetVal, NElem);
  HIP_CHECK(hipEventRecord(events[1], streams[2]));
  // Add operations to stream2
  HIP_CHECK(hipMemsetAsync(devMem2, 0, Nbytes, streams[1]));
  HIP_CHECK(hipMemcpyAsync(devMem2, hostMem2, Nbytes, hipMemcpyHostToDevice, streams[1]));
  HIP_CHECK(hipEventRecord(events[2], streams[1]));
  // Add operations to stream1
  HIP_CHECK(hipMemsetAsync(devMem1, 0, Nbytes, streams[0]));
  HIP_CHECK(hipMemcpyAsync(devMem1, hostMem1, Nbytes, hipMemcpyHostToDevice, streams[0]));
  HIP_CHECK(hipStreamWaitEvent(streams[0], events[2], 0));
  HIP_CHECK(hipStreamWaitEvent(streams[0], events[1], 0));
  hipLaunchKernelGGL(HipTest::vectorADD<T>, dim3(blocks), dim3(threadsPerBlock), 0, streams[0],
                     devMem1, devMem2, devMem3, NElem);
  HIP_CHECK(hipMemcpyAsync(hostMem3, devMem3, Nbytes, hipMemcpyDeviceToHost, streams[0]));
  HIP_CHECK(hipStreamEndCapture(streams[0], &graph));
}

enum class GraphGetNodesTest { equalNumNodes, lesserNumNodes, greaterNumNodes };

template <typename F>
static void validateGraphNodesCommon(
    F f, std::vector<hipGraphNode_t>& nodelist, size_t testNumNodes, GraphGetNodesTest test_type) {
  size_t numNodes = testNumNodes;
  hipGraphNode_t* nodes = new hipGraphNode_t[numNodes]{};
  int found_count{0};
  HIP_CHECK(f(nodes, &numNodes));
  // Count how many nodes from the nodelist are present
  for (auto node : nodelist) {
    for (size_t i = 0; i < numNodes; i++) {
      if (node == nodes[i]) {
        found_count++;
        break;
      }
    }
  }

  // Verify that the found number of nodes is expected
  switch (test_type) {
    case GraphGetNodesTest::equalNumNodes:
      REQUIRE(found_count == nodelist.size());
      break;
    case GraphGetNodesTest::lesserNumNodes:
      // Verify numNodes is unchanged
      REQUIRE(numNodes == testNumNodes);
      REQUIRE(found_count == testNumNodes);
      break;
    case GraphGetNodesTest::greaterNumNodes:
      // Verify numNodes is reset to actual number of nodes
      REQUIRE(numNodes == nodelist.size());
      REQUIRE(found_count == nodelist.size());
      // Verify additional entries in nodes are set to nullptr
      for (auto i = numNodes; i < testNumNodes; i++) {
        REQUIRE(nodes[i] == nullptr);
      }
  }
  delete[] nodes;
}