chundoong-lab-ta/SamsungDS22/submissions/HW6/junjip.lee/mat_mul.cu

#include "mat_mul.h"

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

#define TS 4
#define WPT 8

#define TILE_WIDTH 16

#define RTS (TS/WPT)
#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 4
int num_devices = 0;

__global__ void sgemm(float *A, float *B, float *C, int M, int N, int K) {

 __shared__ float ds_M[TILE_WIDTH][TILE_WIDTH];
    __shared__ float ds_N[TILE_WIDTH][TILE_WIDTH];
    int bx = blockIdx.x, by = blockIdx.y,
       tx = threadIdx.x, ty = threadIdx.y,
       Row = by * TILE_WIDTH + ty,
       Col = bx * TILE_WIDTH + tx;
    float Pvalue = 0;
 //   float Pvalue[WPT];

  //  for (int w=0; w<WPT; w++) {

  //      acc[w] = 0.0f;

  //  }
  
   int numARows=M;
   int numAColumns=K;
   int numBRows=K;
   int numBColumns=N;
   int numCRows=M;
   int numCColumns=N;

  const int numTiles = (K+TILE_WIDTH-1)/TILE_WIDTH;


 // printf("Row %d Col %d numCColumns %d M %d N %d K %d \n",Row,Col,numCColumns,M,N,K);
  //  for (int t=0; t<numTiles; t++) {

    for (int m = 0; m < (numAColumns-1)/TILE_WIDTH+1; ++m) {
       if (Row < numARows && m*TILE_WIDTH+tx < numAColumns)
          ds_M[ty][tx] = A[Row*numAColumns + m*TILE_WIDTH+tx];
       else
          ds_M[ty][tx] = 0;
       if (Col < numBColumns && m*TILE_WIDTH+ty < numBRows)
          ds_N[ty][tx] = B[(m*TILE_WIDTH+ty)*numBColumns+Col];
       else
          ds_N[ty][tx] = 0;

       __syncthreads();
       for (int k = 0; k < TILE_WIDTH; ++k)
          Pvalue += ds_M[ty][k] * ds_N[k][tx];
       __syncthreads();

    //  }//  for (int t=0; t<numTiles; t++) 

    }
    if (Row < numCRows && Col < numCColumns)
       C[Row*numCColumns+Col] = Pvalue;


}

// 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];

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++) {
    //dim3 blockDim(1, 1, 1);
    //dim3 gridDim(Mend[i] - Mbegin[i], N, 1);
   // dim3 gridDim((_N-1)/TILE_WIDTH+1, (M-1)/TILE_WIDTH+1, 1);

dim3 gridDim((_N-1)/TILE_WIDTH+1, ( Mend[i] - Mbegin[i]-1)/TILE_WIDTH+1, 1);

 //     dim3 gridDim((Mend[i] - Mbegin[i])/TILE_WIDTH+1, (Mend[i] - Mbegin[i])/TILE_WIDTH+1, 1);
  
    dim3 blockDim(TILE_WIDTH, TILE_WIDTH, 1);

//jung dim3 blockDim(TS/WPT, TS, 1);
//jung dim3 gridDim(((Mend[i] - Mbegin[i]+TS-1)/TS), (N+TS-1)/TS, 1);

//  dim3 blockDim(TS/WPT, TS, 1);
//  dim3 gridDim(((Mend[i] - Mbegin[i]+TS-1)/TS), (N+TS-1)/TS, 1);
    CUDA_CALL( cudaSetDevice(i) );
    sgemm<<<gridDim, blockDim>>>(a_d[i], b_d[i], c_d[i], Mend[i] - Mbegin[i], N, K);


//sgemm<<<gridDim, blockDim>>>(a_d[i], b_d[i], c_d[i], M_siz[i], N, K);

  }

  // DO NOT REMOVE; NEEDED FOR TIME MEASURE
  for (int i = 0; i < num_devices; i++) {
      CUDA_CALL( cudaSetDevice(i) );//jjlee
    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) );
 // num_devices=2;

  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);
 //   printf(" [%d] Mbegin %d Mend %d \n",i,Mbegin[i],Mend[i]);
  }
  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( cudaSetDevice(i) );//jjlee
    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( 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( cudaSetDevice(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() );
  }
}