chundoong-lab-ta/SamsungDS22/submissions/HW6/jh8846.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 4
#define K_MAX 8192
int num_devices = 0;

#define TILE_WIDTH 16

__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;
       	int tx = threadIdx.x, ty = threadIdx.y;
       	int Row = by * TILE_WIDTH + ty;
       	int Col = bx * TILE_WIDTH + tx;

	if(K >= K_MAX)
	{
		 if (Row > M || Col > N) return;
	}

	float Pvalue = 0;
	int end_m = (K-1)/TILE_WIDTH+1;

	for (int m = 0; m < end_m; ++m) 
	{
       		if (Row < M && m*TILE_WIDTH+tx < K)
        		 	ds_M[ty][tx] = A[Row*K + m*TILE_WIDTH+tx];
       		else
          			ds_M[ty][tx] = 0;
       		if (Col < N && m*TILE_WIDTH+ty < K)
          			ds_N[ty][tx] = B[(m*TILE_WIDTH+ty)*N+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();
    	}

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

/*
int ROW = blockDim.x * blockIdx.x + threadIdx.x;
  int COL = blockDim.y * blockIdx.y + threadIdx.y;
  if (ROW >= M || COL >= N)
    return;

  if(ROW<M && COL<N)
  {
	float value = 0.0f;
	for(int i=0; i<N; i++)
	{
		value += A[ROW*K + i] * B[i*N + COL];
	}
	C[ROW*N + COL] = value;
  }

*/

/*
  float sum = 0.0f;
  for(int k=0; k<M; ++k)
  {
	float a = A[i*K + k];
	float b = B[k*N + j];
	sum += a*b;
  }

  C[i*N + j] = sum;  
*/

/*
  C[i * N + j] = 0;
  for (int k = 0; k < K; ++k) {
    C[i * N + j] += A[i * K + k] * B[k * N + j];
  }
*/


// 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 dimGrid((N-1)/TILE_WIDTH+1, (M-1)/TILE_WIDTH+1, 1);
    	dim3 dimBlock(TILE_WIDTH, TILE_WIDTH, 1);
	//dim3 blockDim(16, 16, 1);
    	//dim3 gridDim(ceil(K/16.0), ceil(M/16.0), 1);

    CUDA_CALL( cudaSetDevice(i) );
    sgemm<<<dimGrid, dimBlock>>>(a_d[i], b_d[i], c_d[i], (Mend[i] - Mbegin[i]), N, K);
  }

  // DO NOT REMOVE; NEEDED FOR TIME MEASURE
  for (int i = 0; i < num_devices; 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) );

  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( 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) );

	CUDA_CALL(cudaFree(a_d[i]));
	CUDA_CALL(cudaFree(b_d[i]));
	CUDA_CALL(cudaFree(c_d[i]));
  }

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