TheRock/rocm-systems/projects/hip-tests/catch/unit/module/hipGetFuncBySymbol_exe.cc

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

#include <hip/hip_runtime.h>

#define ARR_SIZE (32 * 32)
#define SIZE (ARR_SIZE * sizeof(int))

#define HIP_CHECK(error)                                                                           \
  {                                                                                                \
    hipError_t localError = error;                                                                 \
    if ((localError != hipSuccess) && (localError != hipErrorPeerAccessAlreadyEnabled)) {          \
      printf("error: '%s'(%d) from %s at %s:%d\n", hipGetErrorString(localError), localError,      \
             #error, __FUNCTION__, __LINE__);                                                      \
      exit(0);                                                                                     \
    }                                                                                              \
  }

/**
 * Sample Kernel to be used for functional test cases
 */
__global__ void hipKernel(int* a) {
  int offset = blockDim.x * blockIdx.x + threadIdx.x;
  int stride = blockDim.x * gridDim.x;

  for (int i = offset; i < ARR_SIZE; i += stride) {
    a[i] += a[i];
  }
}

/**
 * This function gets the function pointer and launches the kernel and
 * verifies the reult
 */
int main() {
  hipFunction_t funcPointer;

  if (hipGetFuncBySymbol(&funcPointer, reinterpret_cast<const void*>(hipKernel)) != hipSuccess) {
    return -1;
  }

  int* h_a = reinterpret_cast<int*>(malloc(SIZE));
  if (h_a == nullptr) {
    return -1;
  }

  int* output_ref = reinterpret_cast<int*>(malloc(SIZE));
  if (output_ref == nullptr) {
    return -1;
  }

  for (int i = 0; i < ARR_SIZE; i++) {
    h_a[i] = 2;
    output_ref[i] = 4;
  }

  int* d_a = nullptr;
  HIP_CHECK(hipMalloc(&d_a, SIZE));
  if (d_a == nullptr) {
    return -1;
  }
  HIP_CHECK(hipMemcpy(d_a, h_a, SIZE, hipMemcpyHostToDevice));

  dim3 blocksPerGrid(1, 1, 1);
  dim3 threadsPerBlock(1, 1, 64);

  void* kernelParam[] = {d_a};
  auto size = sizeof(kernelParam);
  void* kernel_parameter[] = {HIP_LAUNCH_PARAM_BUFFER_POINTER, &kernelParam,
                              HIP_LAUNCH_PARAM_BUFFER_SIZE, &size, HIP_LAUNCH_PARAM_END};

  if (hipModuleLaunchKernel(funcPointer, blocksPerGrid.x, blocksPerGrid.y, blocksPerGrid.z,
                            threadsPerBlock.x, threadsPerBlock.y, threadsPerBlock.z, 0, 0, nullptr,
                            kernel_parameter) != hipSuccess) {
    return -1;
  }

  HIP_CHECK(hipMemcpy(h_a, d_a, SIZE, hipMemcpyDeviceToHost));

  for (int i = 0; i < ARR_SIZE; i++) {
    if (h_a[i] != output_ref[i]) {
      return -1;
    }
  }

  free(h_a);
  free(output_ref);
  HIP_CHECK(hipFree(d_a));

  return 0;
}