#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 TileSize 32
#define SubWorkSize 16
#define Offset (TileSize/SubWorkSize)

int num_devices = 0;

__global__ void sgemm(float *A, float *B, float *C, int M, int N, int K) {
  //const int g_row = blockDim.x * SubWorkSize * blockIdx.x + threadIdx.x;
  const int g_row = TileSize * blockIdx.x + threadIdx.x;
  //const int g_col = blockDim.y * blockIdx.y + threadIdx.y;
  const int g_col = TileSize * blockIdx.y + threadIdx.y;

  const int l_row = threadIdx.x;
  const int l_col = threadIdx.y;

  __shared__ float Asub[TileSize][TileSize];
  __shared__ float Bsub[TileSize][TileSize];

  float acc[SubWorkSize];
  for (int w = 0; w < SubWorkSize; w++) {
    acc[w] = 0.0f;
  }

  const int num_tiles = (K + TileSize - 1) / TileSize;
  for (int t = 0; t < num_tiles; t++) {
    for (int w = 0; w < SubWorkSize; w++) {
        const int t_row = TileSize * t + l_row;
        const int t_col = TileSize * t + l_col;

        int A_row_bound = g_row + w*Offset;
        int B_row_bound = t_row + w*Offset;

        if ((A_row_bound < M) && (t_col < K))
           Asub[l_row + w*Offset][l_col] = A[(g_row + w*Offset)*K + t_col];        else Asub[l_row + w*Offset][l_col] = 0.0f;

        if ((B_row_bound < K) && (g_col < N))
           Bsub[l_row + w*Offset][l_col] = B[(t_row + w*Offset)*N + g_col];        else Bsub[l_row + w*Offset][l_col] = 0.0f;
    }
    __syncthreads();

    for (int k=0; k < TileSize; k++) {
        for (int w=0; w < SubWorkSize; w++) {
            acc[w] += Asub[l_row + w*Offset][k] * Bsub[k][l_col];
        }
    }
    __syncthreads();
    // barrier(CLK_LOCAL_MEM_FENCE);
  }

  for (int w=0; w < SubWorkSize; w++) {
        int C_row_bound = g_row + w*Offset;

        if ((C_row_bound < M) && (g_col < N))
           C[(g_row + w*Offset)*N + g_col] = acc[w];
        else return;
  }

}

// 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(TileSize/SubWorkSize, TileSize, 1);
    int nBlock_row = (Mend[i] - Mbegin[i] + TileSize - 1) / TileSize;
    int nBlock_col = (N + TileSize -1) / TileSize;
    // printf("nBlock_row[%d]: %d, Mbegin[%d]: %d, Mend[%d]: %d\n", i, nBlock_row, i, Mbegin[i], i, Mend[i]);
    // printf("BlockDim.x: %d, BlockDim.y: %d \n", blockDim.x, blockDim.y);
    //dim3 gridDim(Mend[i] - Mbegin[i], N, 1);
    dim3 gridDim(nBlock_row, nBlock_col, 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( 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) ); 	// for multiple GPUs
  // num_devices = 1; 	// for performance check in a GPU

  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() );
  }
}