chundoong-lab-ta/SamsungDS22/submissions/HW6/hongpooh.kim/mat_mul.cu

#include "mat_mul.h"

#include <cstdio>
#include <cuda_runtime.h>

#define CUDA_CALL(f)                                                           \
  {                                                                            \
    cudaError_t err = (f);                                                     \
    if (err != cudaSuccess) {                                                  \
      fprintf(stderr, "CUDA error at [%s:%d] %d %s\n", __FILE__, __LINE__,     \
              err, cudaGetErrorString(err));                                   \
      exit(1);                                                                 \
    }                                                                          \
  }

//#define MAX_NUM_GPU 1
#define MAX_NUM_GPU 4
#define TS  16  // Tile Width 16

// Array of device (GPU) pointers
static float *a_d[MAX_NUM_GPU];
static float *b_d[MAX_NUM_GPU];
static float *c_d[MAX_NUM_GPU];
static int M, N, K;
static int Mbegin[MAX_NUM_GPU], Mend[MAX_NUM_GPU];

int num_devices = 0;
int size;


__global__ void sgemm(float *A, float *B, float *C, int M, int N, int K) {
/* 1. original code - about 40 GFLOPS 
  int i = blockDim.x * blockIdx.x + threadIdx.x;
  int j = blockDim.y * blockIdx.y + threadIdx.y;
  if (i >= M || j >= N)
    return;

  C[i * N + j] = 0;
  for (int k = 0; k < K; ++k) {
    C[i * N + j] += A[i * K + k] * B[k * N + j];
  }
}
*/
    __shared__ float Asub[TS][TS];
    __shared__ float Bsub[TS][TS];
    int bx = blockIdx.x, by = blockIdx.y;
    int tx = threadIdx.x, ty = threadIdx.y;
   //test_only: int by = blockIdx.x, bx = blockIdx.y;
   //test_only: int ty = threadIdx.x, tx = threadIdx.y;
    int Row = by * TS + ty;
    int Col = bx * TS + tx;
    float Pvalue = 0;

    //const int num_tiles = M / TS;
    //for (int m = 0; m < (M-1)/TS + 1; m++) {
    for (int m = 0; m < (K - 1)/TS + 1; m++) {
        if (Row < M && m*TS + tx < K) {
            Asub[ty][tx] = A[Row*K + m*TS + tx];
        } else {
            Asub[ty][tx] = 0.0f;
        }

        if (Col < N && m*TS + ty < K) {
            Bsub[ty][tx] = B[(m*TS + ty)*N + Col];
        } else {
            Bsub[ty][tx] = 0x0f;
        }

        __syncthreads();
        for (int k = 0; k < TS; k++) {
            Pvalue += Asub[ty][k] * Bsub[k][tx];
        }
        __syncthreads();
    }
    if (Row < M && Col < N) {
        C[Row*N + Col] = Pvalue;
    }
}


void mat_mul(float *_A, float *_B, float *_C, int _M, int _N, int _K) {

  // Launch kernel on every GPU
  for (int i = 0; i < num_devices; i++) {
    size = (i == num_devices - 1) ? (M-(M/num_devices*(num_devices-1))) : M/num_devices;
    //dim3 blockDim(1, 1, 1);
    dim3 blockDim(TS, TS, 1);
    //dim3 gridDim(Mend[i] - Mbegin[i], N, 1);
    //dim3 gridDim((N-1)/TS + 1, (M-1)/TS + 1, 1);
    dim3 gridDim((N-1)/TS + 1, (size-1)/TS + 1, 1);

    CUDA_CALL( cudaSetDevice(i) );
    sgemm<<<gridDim, blockDim>>>(a_d[i], b_d[i], c_d[i], M, N, K);
  }

  // DO NOT REMOVE; NEEDED FOR TIME MEASURE
  for (int i = 0; i < num_devices; i++) {
    //CUDA_CALL( cudaSetDevice(i) );
    CUDA_CALL( cudaDeviceSynchronize() );
  }
}

void mat_mul_init(float *A, float *B, float *C, int _M, int _N, int _K) {
  M = _M, N = _N, K = _K;

  CUDA_CALL( cudaGetDeviceCount(&num_devices) );

  // temp: for Single GPU test
  //num_devices = 1;

  printf("Using %d devices\n", num_devices);
  for (int i = 0; i < num_devices; i++) {
    cudaDeviceProp prop;
    CUDA_CALL( cudaGetDeviceProperties(&prop, i) );

    // Try printing more detailed information here
    printf("[GPU %d] %s\n", i, prop.name);
  }

  if (num_devices <= 0) {
    printf("No CUDA device found. Aborting\n");
    exit(1);
  }

  // Setup problem size for each GPU
  for (int i = 0; i < num_devices; i++) {
    Mbegin[i] = (M / num_devices) * i;
    Mend[i] = (M / num_devices) * (i + 1);
  }
  Mend[num_devices - 1] = M;

  // Allocate device memory for each GPU
  for (int i = 0; i < num_devices; i++) {
    CUDA_CALL( cudaSetDevice(i) );
    CUDA_CALL( cudaMalloc(&a_d[i], (Mend[i] - Mbegin[i])*K*sizeof(float)) );
    CUDA_CALL( cudaMalloc(&b_d[i], K*N*sizeof(float)) );
    CUDA_CALL( cudaMalloc(&c_d[i], (Mend[i] - Mbegin[i])*N*sizeof(float)) );
  }

  // Upload A and B matrix to every GPU
  for (int i = 0; i < num_devices; i++) {
    CUDA_CALL( cudaMemcpy(a_d[i], A + Mbegin[i]*K,
               (Mend[i] - Mbegin[i])*K*sizeof(float), cudaMemcpyHostToDevice) );
    CUDA_CALL( cudaMemcpy(b_d[i], B, K*N*sizeof(float), cudaMemcpyHostToDevice) );
  }

  // DO NOT REMOVE; NEEDED FOR TIME MEASURE
  for (int i = 0; i < num_devices; i++) {
    CUDA_CALL( cudaSetDevice(i) );
    CUDA_CALL( cudaDeviceSynchronize() );
  }
}

void mat_mul_final(float *A, float *B, float *C, int M, int N, int K) {
  // Do any post-matmul cleanup work here.

  // Download C matrix from GPUs
  for (int i = 0; i < num_devices; i++) {
    CUDA_CALL( cudaMemcpy(C + Mbegin[i]*N, c_d[i],
               (Mend[i] - Mbegin[i])*N*sizeof(float), cudaMemcpyDeviceToHost) );
  }

  // DO NOT REMOVE; NEEDED FOR TIME MEASURE
  for (int i = 0; i < num_devices; i++) {
    CUDA_CALL( cudaDeviceSynchronize() );
  }
}