TheRock/rocm-systems/projects/hip-tests/samples/2_Cookbook/8_peer2peer/peer2peer.cpp

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

#include <iostream>
#include <hip/hip_runtime.h>
#include <assert.h>
#define WIDTH 32

#define NUM (WIDTH * WIDTH)

#define THREADS_PER_BLOCK_X 4
#define THREADS_PER_BLOCK_Y 4
#define THREADS_PER_BLOCK_Z 1

using namespace std;

#define KNRM "\x1B[0m"
#define KRED "\x1B[31m"

#define failed(...)                                                                                \
  printf("%serror: ", KRED);                                                                       \
  printf(__VA_ARGS__);                                                                             \
  printf("\n");                                                                                    \
  printf("error: TEST FAILED\n%s", KNRM);                                                          \
  abort();

#define HIPCHECK(error)                                                                            \
  {                                                                                                \
    hipError_t localError = error;                                                                 \
    if ((localError != hipSuccess) && (localError != hipErrorPeerAccessAlreadyEnabled) &&          \
        (localError != hipErrorPeerAccessNotEnabled)) {                                            \
      printf("%serror: '%s'(%d) from %s at %s:%d%s\n", KRED, hipGetErrorString(localError),        \
             localError, #error, __FILE__, __LINE__, KNRM);                                        \
      failed("API returned error code.");                                                          \
    }                                                                                              \
  }

void checkPeer2PeerSupport() {
  int gpuCount;
  int canAccessPeer;

  HIPCHECK(hipGetDeviceCount(&gpuCount));

  for (int currentGpu = 0; currentGpu < gpuCount; currentGpu++) {
    HIPCHECK(hipSetDevice(currentGpu));

    for (int peerGpu = 0; peerGpu < currentGpu; peerGpu++) {
      if (currentGpu != peerGpu) {
        HIPCHECK(hipDeviceCanAccessPeer(&canAccessPeer, currentGpu, peerGpu));
        printf("currentGpu#%d canAccessPeer: peerGpu#%d=%d\n", currentGpu, peerGpu, canAccessPeer);
      }

      HIPCHECK(hipSetDevice(peerGpu));
      HIPCHECK(hipDeviceReset());
    }
    HIPCHECK(hipSetDevice(currentGpu));
    HIPCHECK(hipDeviceReset());
  }
}

void enablePeer2Peer(int currentGpu, int peerGpu) {
  int canAccessPeer;

  // Must be on a multi-gpu system:
  assert(currentGpu != peerGpu);

  HIPCHECK(hipSetDevice(currentGpu));
  hipDeviceCanAccessPeer(&canAccessPeer, currentGpu, peerGpu);

  if (canAccessPeer == 1) {
    HIPCHECK(hipDeviceEnablePeerAccess(peerGpu, 0));
  } else
    printf("peer2peer transfer not possible between the selected gpu devices");
}

void disablePeer2Peer(int currentGpu, int peerGpu) {
  int canAccessPeer;

  // Must be on a multi-gpu system:
  assert(currentGpu != peerGpu);

  HIPCHECK(hipSetDevice(currentGpu));
  hipDeviceCanAccessPeer(&canAccessPeer, currentGpu, peerGpu);

  if (canAccessPeer == 1) {
    HIPCHECK(hipDeviceDisablePeerAccess(peerGpu));
  } else
    printf("peer2peer disable not required");
}


__global__ void matrixTranspose_static_shared(float* out, float* in, const int width) {
  __shared__ float sharedMem[WIDTH * WIDTH];

  int x = blockDim.x * blockIdx.x + threadIdx.x;
  int y = blockDim.y * blockIdx.y + threadIdx.y;

  sharedMem[y * width + x] = in[x * width + y];

  __syncthreads();

  out[y * width + x] = sharedMem[y * width + x];
}

__global__ void matrixTranspose_dynamic_shared(float* out, float* in, const int width) {
  extern __shared__ float sharedMem[];

  int x = blockDim.x * blockIdx.x + threadIdx.x;
  int y = blockDim.y * blockIdx.y + threadIdx.y;

  sharedMem[y * width + x] = in[x * width + y];

  __syncthreads();

  out[y * width + x] = sharedMem[y * width + x];
}

int main() {
  checkPeer2PeerSupport();

  int gpuCount;
  int currentGpu, peerGpu;

  HIPCHECK(hipGetDeviceCount(&gpuCount));

  if (gpuCount < 2) {
    printf("Peer2Peer application requires atleast 2 gpu devices");
    return 0;
  }

  currentGpu = 0;
  peerGpu = (currentGpu + 1);

  printf("currentGpu=%d peerGpu=%d (Total no. of gpu = %d)\n", currentGpu, peerGpu, gpuCount);

  float *data[2], *TransposeMatrix[2], *gpuTransposeMatrix[2], *randArray;

  int width = WIDTH;

  randArray = (float*)malloc(NUM * sizeof(float));

  for (int i = 0; i < NUM; i++) {
    randArray[i] = (float)i * 1.0f;
  }

  enablePeer2Peer(currentGpu, peerGpu);

  HIPCHECK(hipSetDevice(currentGpu));
  TransposeMatrix[0] = (float*)malloc(NUM * sizeof(float));
  hipMalloc((void**)&gpuTransposeMatrix[0], NUM * sizeof(float));
  hipMalloc((void**)&data[0], NUM * sizeof(float));
  hipMemcpy(data[0], randArray, NUM * sizeof(float), hipMemcpyHostToDevice);

  hipLaunchKernelGGL(
      matrixTranspose_static_shared, dim3(WIDTH / THREADS_PER_BLOCK_X, WIDTH / THREADS_PER_BLOCK_Y),
      dim3(THREADS_PER_BLOCK_X, THREADS_PER_BLOCK_Y), 0, 0, gpuTransposeMatrix[0], data[0], width);
  HIPCHECK(hipDeviceSynchronize());

  HIPCHECK(hipSetDevice(peerGpu));
  TransposeMatrix[1] = (float*)malloc(NUM * sizeof(float));
  hipMalloc((void**)&gpuTransposeMatrix[1], NUM * sizeof(float));
  hipMalloc((void**)&data[1], NUM * sizeof(float));
  hipMemcpy(data[1], gpuTransposeMatrix[0], NUM * sizeof(float), hipMemcpyDeviceToDevice);

  hipLaunchKernelGGL(matrixTranspose_dynamic_shared,
                     dim3(WIDTH / THREADS_PER_BLOCK_X, WIDTH / THREADS_PER_BLOCK_Y),
                     dim3(THREADS_PER_BLOCK_X, THREADS_PER_BLOCK_Y), sizeof(float) * WIDTH * WIDTH,
                     0, gpuTransposeMatrix[1], data[1], width);

  hipMemcpy(TransposeMatrix[1], gpuTransposeMatrix[1], NUM * sizeof(float), hipMemcpyDeviceToHost);

  hipDeviceSynchronize();

  disablePeer2Peer(currentGpu, peerGpu);

  // verify the results
  int errors = 0;
  double eps = 1.0E-6;
  for (int i = 0; i < NUM; i++) {
    if (std::abs(randArray[i] - TransposeMatrix[1][i]) > eps) {
      printf("%d cpu: %f gpu peered data  %f\n", i, randArray[i], TransposeMatrix[1][i]);
      errors++;
    }
  }
  if (errors != 0) {
    printf("FAILED: %d errors\n", errors);
  } else {
    printf("Peer2Peer PASSED!\n");
  }

  free(randArray);
  for (int i = 0; i < 2; i++) {
    hipFree(data[i]);
    hipFree(gpuTransposeMatrix[i]);
    free(TransposeMatrix[i]);
  }

  HIPCHECK(hipSetDevice(peerGpu));
  HIPCHECK(hipDeviceReset());

  HIPCHECK(hipSetDevice(currentGpu));
  HIPCHECK(hipDeviceReset());

  return 0;
}