#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
int num_devices = 0;


#define TILE_SIZE    32
#define WPT           8           // work per thread
#define RTS  (TILE_SIZE / WPT)

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

  const int row = threadIdx.x; // Local row ID (max: TILE_SIZE/WPT == RTS)
  const int col = threadIdx.y; // Local col ID (max: TILE_SIZE)
  const int globalRow = TILE_SIZE * blockIdx.x + row; // row index of C (N)
  const int globalCol = TILE_SIZE * blockIdx.y + col; // column index of C (M)
  const int numTiles = (K + TILE_SIZE - 1) / TILE_SIZE;

  // local memory for tile
  __shared__ float Asub[TILE_SIZE][TILE_SIZE];
  __shared__ float Bsub[TILE_SIZE][TILE_SIZE];

  // Init result memory
  float res[WPT];
  for (int i = 0; i < WPT; i++) {
      res[i] = 0.0f;
  }

  for (int t = 0; t < numTiles; t++) {
    const int tiledRow = TILE_SIZE * t + row;
    const int tiledCol = TILE_SIZE * t + col;

    // Load A and B to local memory
    for (int w = 0; w < WPT; w++) {
      if (((w * RTS + globalRow) >= M) || (tiledCol >= K)) {
        Asub[w * RTS + row][col] = 0;
      }
      else {
        Asub[w * RTS + row][col] = A[(w * RTS + globalRow) * K + tiledCol];
      }
      
      if (((w * RTS + tiledRow) >= K) || (globalCol >= N)) {
        Bsub[w * RTS + row][col] = 0;
      } 
      else {
        Bsub[w * RTS + row][col] = B[(w * RTS + tiledRow) * N + globalCol];
      }   
    }

    __syncthreads();

    // result for tile
    for (int i = 0; i < TILE_SIZE; i++) {
      for (int j = 0; j < WPT; j++) {
        res[j] += Asub[j * RTS + row][i] * Bsub[i][col];
      }
    }

    __syncthreads();
  }

  // final results in C
  for (int w = 0; w < WPT; w++) {
    if ((w * RTS + globalRow < M) && (globalCol < N)) {
      C[(w * RTS + globalRow) * N + globalCol] = res[w];
    }
  }

}


// 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(RTS, TILE_SIZE, 1);
    dim3 gridDim(((Mend[i] - Mbegin[i]) + TILE_SIZE - 1) / TILE_SIZE, (N + TILE_SIZE - 1) / TILE_SIZE, 1);

    CUDA_CALL( cudaSetDevice(i) );
    sgemm<<<gridDim, blockDim>>>(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( 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) );

  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);
  }

  // // temporary!!!
  // num_devices = 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() );
  }
}