TheRock/rocm-systems/projects/hip-tests/catch/unit/dynamicLoading/liblazyLoad.cc

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

#include <hip/hip_runtime.h>
#include <hip/hip_runtime_api.h>

#include <algorithm>  // std::min

#define CHECK_RET_VAL(cmd)                                                                         \
  {                                                                                                \
    hipError_t error = cmd;                                                                        \
    if (error != hipSuccess) {                                                                     \
      fprintf(stderr, "error: '%s'(%d) at %s:%d\n", hipGetErrorString(error), error, __FILE__,     \
              __LINE__);                                                                           \
      exit(EXIT_FAILURE);                                                                          \
    }                                                                                              \
  }

__global__ static void addition(float* C, float* A, float* B, size_t N) {
  size_t offset = blockIdx.x * blockDim.x + threadIdx.x;
  size_t stride = blockDim.x * gridDim.x;

  for (size_t i = offset; i < N; i += stride) {
    C[i] = A[i] + B[i];
  }
}

static int vectorAddKernelTest() {
  int testResult = 1;
  float *A_d, *B_d, *C_d;
  float *A_h, *B_h, *C_h;
  size_t N = 1000000;
  size_t Nbytes = N * sizeof(float);
  static int device = 0;

  CHECK_RET_VAL(hipSetDevice(device));
  hipDeviceProp_t props;
  CHECK_RET_VAL(hipGetDeviceProperties(&props, device));
  A_h = reinterpret_cast<float*>(malloc(Nbytes));
  CHECK_RET_VAL(A_h == nullptr ? hipErrorOutOfMemory : hipSuccess);
  B_h = reinterpret_cast<float*>(malloc(Nbytes));
  CHECK_RET_VAL(B_h == nullptr ? hipErrorOutOfMemory : hipSuccess);
  C_h = reinterpret_cast<float*>(malloc(Nbytes));
  CHECK_RET_VAL(C_h == nullptr ? hipErrorOutOfMemory : hipSuccess);

  // Fill with Phi + i
  for (size_t i = 0; i < N; i++) {
    A_h[i] = 1.618f + i;
    B_h[i] = 1.618f + i;
  }

  CHECK_RET_VAL(hipMalloc(&A_d, Nbytes));
  CHECK_RET_VAL(hipMalloc(&B_d, Nbytes));
  CHECK_RET_VAL(hipMalloc(&C_d, Nbytes));
  CHECK_RET_VAL(hipMemcpy(A_d, A_h, Nbytes, hipMemcpyHostToDevice));
  CHECK_RET_VAL(hipMemcpy(B_d, B_h, Nbytes, hipMemcpyHostToDevice));

  const unsigned blocks = 512;
  const unsigned threadsPerBlock = 256;
  hipLaunchKernelGGL(addition, dim3(blocks), dim3(threadsPerBlock), 0, 0, C_d, A_d, B_d, N);
  CHECK_RET_VAL(hipMemcpy(C_h, C_d, Nbytes, hipMemcpyDeviceToHost));

  for (size_t i = 0; i < N; i++) {
    if (C_h[i] != (A_h[i] + B_h[i])) {
      testResult = 0;
      break;
    }
  }

  CHECK_RET_VAL(hipFree(A_d));
  CHECK_RET_VAL(hipFree(B_d));
  CHECK_RET_VAL(hipFree(C_d));

  free(A_h);
  free(B_h);
  free(C_h);
  return testResult;
}

#include "hip/hip_cooperative_groups.h"

namespace cg = cooperative_groups;

static const uint BufferSizeInDwords = 448 * 1024 * 1024;

__global__ static void test_gws(uint* buf, uint bufSize, int64_t* tmpBuf, int64_t* result) {
  extern __shared__ int64_t tmp[];
  uint offset = blockIdx.x * blockDim.x + threadIdx.x;
  uint stride = gridDim.x * blockDim.x;
  cg::grid_group gg = cg::this_grid();

  int64_t sum = 0;

  for (uint i = offset; i < bufSize; i += stride) {
    sum += buf[i];
  }

  tmp[threadIdx.x] = sum;
  __syncthreads();

  if (threadIdx.x == 0) {
    sum = 0;
    for (uint i = 0; i < blockDim.x; i++) {
      sum += tmp[i];
    }
    tmpBuf[blockIdx.x] = sum;
  }

  // TODO: remove syncthreads with grid sync when compiler issue is fixed
  // gg.sync();
  __syncthreads();

  if (offset == 0) {
    for (uint i = 1; i < gridDim.x; ++i) {
      sum += tmpBuf[i];
    }
    *result = sum;
  }
}

static int cooperativeKernelTest() {
  int testResult = 1;
  uint* dA;
  int64_t* dB;
  int64_t* dC;
  int64_t* Ah;

  hipDeviceProp_t deviceProp;
  CHECK_RET_VAL(hipGetDeviceProperties(&deviceProp, 0));

  if (!deviceProp.cooperativeLaunch) {
    return testResult;
  }

  uint32_t* init = new uint32_t[BufferSizeInDwords];

  for (uint32_t i = 0; i < BufferSizeInDwords; ++i) {
    init[i] = i;
  }
  size_t SIZE = BufferSizeInDwords * sizeof(uint);

  CHECK_RET_VAL(hipMalloc(reinterpret_cast<void**>(&dA), SIZE));
  CHECK_RET_VAL(hipMalloc(reinterpret_cast<void**>(&dC), sizeof(int64_t)));
  CHECK_RET_VAL(hipMemcpy(dA, init, SIZE, hipMemcpyHostToDevice));
  Ah = reinterpret_cast<int64_t*>(malloc(sizeof(int64_t)));

  hipStream_t stream;
  CHECK_RET_VAL(hipStreamCreate(&stream));

  dim3 dimBlock = dim3(1);
  dim3 dimGrid = dim3(1);

  int numBlocks = 0;
  uint workgroups[4] = {32, 64, 128, 256};

  for (uint i = 0; i < 4; ++i) {
    dimBlock.x = workgroups[i];
    /* Calculate the device occupancy to know how many blocks can be
       run concurrently */
    CHECK_RET_VAL(hipOccupancyMaxActiveBlocksPerMultiprocessor(
        &numBlocks, test_gws, dimBlock.x * dimBlock.y * dimBlock.z, dimBlock.x * sizeof(int64_t)));
    dimGrid.x = deviceProp.multiProcessorCount * std::min(numBlocks, 32);
    CHECK_RET_VAL(hipMalloc(reinterpret_cast<void**>(&dB), dimGrid.x * sizeof(int64_t)));

    void* params[4];
    params[0] = reinterpret_cast<void*>(&dA);
    params[1] = (void*)(&BufferSizeInDwords);  // NOLINT
    params[2] = reinterpret_cast<void*>(&dB);
    params[3] = reinterpret_cast<void*>(&dC);

    CHECK_RET_VAL(hipLaunchCooperativeKernel(reinterpret_cast<void*>(test_gws), dimGrid, dimBlock,
                                             params, dimBlock.x * sizeof(int64_t), stream));

    CHECK_RET_VAL(hipMemcpy(Ah, dC, sizeof(int64_t), hipMemcpyDeviceToHost));

    if (*Ah != (((int64_t)(BufferSizeInDwords) * (BufferSizeInDwords - 1)) / 2)) {
      CHECK_RET_VAL(hipFree(dB));
      testResult = 0;
      break;

    } else {
      CHECK_RET_VAL(hipFree(dB));
    }
  }
  CHECK_RET_VAL(hipStreamDestroy(stream));
  CHECK_RET_VAL(hipFree(dC));
  CHECK_RET_VAL(hipFree(dA));
  delete[] init;
  free(Ah);
  return testResult;
}

extern "C" int lazyLoad() { return vectorAddKernelTest() & cooperativeKernelTest(); }