#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 TS (32)
#define WPT (16)
#define RTS (2)

#define MAX_NUM_GPU 4
int num_devices = 0;

__global__ void sgemm(float *A, float *B, float *C, int M, int N, int K, int numTiles) {
  const int row = threadIdx.x;
  const int col = threadIdx.y;
  const int globalRow = TS * blockIdx.x + threadIdx.x;
  const int globalCol = TS * blockIdx.y + threadIdx.y;

  __shared__ float Asub[TS][TS];
  __shared__ float Bsub[TS][TS];

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

  for(int t=0; t<numTiles; t++) {
    for(int w=0; w<WPT; w++) {
      const int tiledRow = TS*t + row;
      const int tiledCol = TS*t + col;
      if((globalRow + w*RTS) < M && tiledCol < K)
        Asub[w*RTS + row][col] = A[(globalRow + w*RTS)*K + tiledCol];
      else
        Asub[w*RTS + row][col] = 0.0f;
      if((tiledRow + w*RTS) < K && globalCol < N)
	Bsub[row + w*RTS][col] = B[(tiledRow + w*RTS)*N + globalCol];
      else
	Bsub[row + w*RTS][col] = 0.0f;
    }

    __syncthreads();

    for(int k=0; k<TS; k++) {
      for(int w=0; w<WPT; w++) {
	acc[w] += Asub[row + w*RTS][k] * Bsub[k][col];
      }
    }

    __syncthreads();
  }

  for(int w=0; w<WPT; w++) {
    if((globalRow + w*RTS) < M && globalCol < N)
      C[(globalRow + w*RTS)*N + globalCol] = acc[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 rows[MAX_NUM_GPU]; 

void mat_mul(float *_A, float *_B, float *_C, int _M, int _N, int _K) {
  int M_XL = (M+TS-1)/TS*TS;
  int N_XL = (N+TS-1)/TS*TS;
  int K_XL = (K+TS-1)/TS*TS;
  int numTiles = K_XL/TS;

  // Launch kernel on every GPU
  for (int i = 0; i < num_devices; i++) {
    int rows_XL = (rows[i]+TS-1)/TS*TS;
    dim3 blockDim(RTS, TS, 1);
    dim3 gridDim(rows_XL/TS, N_XL/TS, 1);

    CUDA_CALL( cudaSetDevice(i) );
    sgemm<<<gridDim, blockDim>>>(a_d[i], b_d[i], c_d[i], rows[i], N, K, numTiles);
  }

  // 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++) {
    rows[i] = i==0? M/num_devices + M%num_devices : M/num_devices;
  }

  // Allocate device memory for each GPU
  for (int i = 0; i < num_devices; i++) {
    CUDA_CALL( cudaSetDevice(i) );
    CUDA_CALL( cudaMalloc(&a_d[i], (rows[i]) * K * sizeof(float)) );
    CUDA_CALL( cudaMalloc(&b_d[i], K * N * sizeof(float)) );
    CUDA_CALL( cudaMalloc(&c_d[i], (rows[i]) * N * sizeof(float)) );
  }

  // Upload A and B matrix to every GPU
  int offset = 0;
  for (int i = 0; i < num_devices; i++) {
    CUDA_CALL( cudaMemcpy(a_d[i], A + offset * K,
               (rows[i]) * K * sizeof(float),
               cudaMemcpyHostToDevice) );
    CUDA_CALL( cudaMemcpy(b_d[i], B, K * N * sizeof(float), cudaMemcpyHostToDevice) );
    offset += rows[i];
  }

  // 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
  int offset = 0;
  for (int i = 0; i < num_devices; i++) {
    CUDA_CALL( cudaMemcpy(C + offset * N, c_d[i],
               (rows[i]) * N * sizeof(float),
               cudaMemcpyDeviceToHost) );
    offset += rows[i];
  }

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