#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 THREADS_PER_BLOCK	32
#define WPT					8
#define RTS					4

int num_devices = 0;

__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 = THREADS_PER_BLOCK * blockIdx.x + threadIdx.x;
    int global_col = THREADS_PER_BLOCK * blockIdx.y + threadIdx.y;

    __shared__ float Asub[THREADS_PER_BLOCK][THREADS_PER_BLOCK];
    __shared__ float Bsub[THREADS_PER_BLOCK][THREADS_PER_BLOCK];

    int q = K / THREADS_PER_BLOCK;
    int r = K % THREADS_PER_BLOCK;

	if(r != 0)
		q = q + 1;

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

    for (int t = 0; t < q; t++) {
        for (int w = 0; w < WPT; w++) {
            int t_row = THREADS_PER_BLOCK * t + row;
            int t_col = THREADS_PER_BLOCK * t + col;

            if (t_col < K && (global_row + (w * RTS)) < M)
                Asub[row + w * RTS][col] = A[(global_row + (w * RTS))* K + t_col];
            else
                Asub[row + w * RTS][col] = 0;

            if ((t_row + (w * RTS)) < K && global_col < N)
                Bsub[row + w * RTS][col] = B[(t_row + (w * RTS))* N + global_col];
            else
                Bsub[row + w * RTS][col] = 0;
        }

        __syncthreads();

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

        __syncthreads();
    }


    for (int w = 0; w < WPT; w++) {
        if (global_row + (w * RTS) >= M || global_col >= N) {
            return;
        }

        C[(global_row + (w * RTS)) * N + global_col] = temp[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 tpb = ((THREADS_PER_BLOCK + WPT - 1) / WPT);
	int ptb = (((Mend[i] - Mbegin[i]) + WPT - 1) / WPT);

    dim3 blockDim(tpb, THREADS_PER_BLOCK, 1);
    dim3 gridDim(((ptb + tpb - 1) / tpb), ((N + THREADS_PER_BLOCK - 1) / THREADS_PER_BLOCK), 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( 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( 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() );
  }
}