#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 TILE_WIDTH 16
#define TS 32                        // The square-root of the 2D tile-size (== work-group dims)
#define WPT 8                        // The amount of work-per-thread, i.e. the thread-coarsening factor
#define RTS (TS/WPT)                 // The reduced tile-size in one dimension

#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 sm_M[TILE_WIDTH][TILE_WIDTH];
    __shared__ float sm_N[TILE_WIDTH][TILE_WIDTH];
    int blockx = blockIdx.x, blocky = blockIdx.y,
       threadx = threadIdx.x, thready = threadIdx.y,
       Row = blocky * TILE_WIDTH + thready,
       Col = blockx * TILE_WIDTH + threadx;
    float Pval = 0;

    for (int m = 0; m < (K-1)/TILE_WIDTH+1; ++m) {
       if (Row < M && m*TILE_WIDTH+threadx < K)
          sm_M[thready][threadx] = A[Row*K + m*TILE_WIDTH+threadx];
       else
          sm_M[thready][threadx] = 0;
       if (Col < N && m*TILE_WIDTH+thready < K)
          sm_N[thready][threadx] = B[(m*TILE_WIDTH+thready)*N+Col];
       else
          sm_N[thready][threadx] = 0;

       __syncthreads();
       for (int k = 0; k < TILE_WIDTH; ++k)
          Pval += sm_M[thready][k] * sm_N[k][threadx];
       __syncthreads();
    }
    if (Row < M && Col < N)
       C[Row*N+Col] = Pval;
}

// 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++) {

    int MSize = Mend[i] - Mbegin[i];
    dim3 gridDim((N-1)/TILE_WIDTH+1, (MSize-1)/TILE_WIDTH+1, 1);
    dim3 blockDim(TILE_WIDTH, TILE_WIDTH, 1);    
    
    CUDA_CALL( cudaSetDevice(i) );
    sgemm<<<gridDim, blockDim>>>(a_d[i], b_d[i], c_d[i], MSize , 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) );
  //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( 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() );
  }
}