#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 TS 32
#define WPT 16
#define RTS (TS / WPT)
__global__ void sgemm(float *A, float *B, float *C, int M, int N, int K) {
  int row = threadIdx.x;
  int col = threadIdx.y;
  int global_row = WPT * blockDim.x * blockIdx.x + row; // row index of C
  int global_col = blockDim.y * blockIdx.y + col; // column index of C

  __shared__ float A_SUB[TS][TS];
  __shared__ float B_SUB[TS][TS];

	float iv[WPT];

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

  int _K = (K + TS - 1) / TS * TS;

  for (int t = 0; t < _K / TS; t++) {
    for (int w = 0; w < WPT; w++) {
      const int t_row = TS * t + row;
      const int t_col = TS * t + col;
      const int indexA = (global_row + w*RTS) * K + t_col;
      const int indexB = (t_row + w*RTS) * N + global_col;

      A_SUB[row + w*RTS][col] = (global_row + w*RTS < M && t_col < K) ? A[indexA] : 0.0f;
      B_SUB[row + w*RTS][col] = (t_row + w*RTS < K && global_col < N) ? B[indexB] : 0.0f;
    }

    __syncthreads();
    for (int k = 0; k < TS; k++) {
      for (int w = 0; w < WPT; w++) {
        iv[w] += A_SUB[row + w * RTS][k] * B_SUB[k][col];
      }
		}
    __syncthreads();
	}

  if (global_row >= M || global_col >= N) return; // boundary check

  for (int w = 0; w < WPT; w++) {
    C[(global_row + w * RTS)* N + global_col] = iv[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++) {
    int MTS = ((TS + WPT - 1) / WPT * WPT) / WPT; // modified TS;
    int MGWS = ((Mend[i] - Mbegin[i] + WPT - 1) / WPT * WPT) / WPT;
    size_t gws[2] = {(size_t)MGWS, (size_t)N}, lws[2] = {(size_t)MTS, (size_t)TS};
    for (int j = 0; j < 2; ++j) {
      gws[j] = (gws[j] + lws[j] - 1) / lws[j] * lws[j];
    }
    dim3 blockDim(lws[0], lws[1], 1);
    dim3 gridDim(gws[0]/lws[0], gws[1]/lws[1], 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( 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);
  }

  // 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( cudaSetDevice(i) );
    CUDA_CALL( cudaDeviceSynchronize() );
  }
}