#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 i = blockDim.x * blockIdx.x + threadIdx.x;
  //int j = blockDim.y * blockIdx.y + threadIdx.y;
  //if (i >= M || j >= N)
  //  return;

  //C[i * N + j] = 0;
  //for (int k = 0; k < K; ++k) {
  //  C[i * N + j] += A[i * K + k] * B[k * N + j];
  //}
  const int row = threadIdx.x;
  const int col = threadIdx.y;
  const int i = (blockDim.x*WPT)*blockIdx.x+threadIdx.x;
  const int j = blockDim.y*blockIdx.y+threadIdx.y;
 
  __shared__ float Asub[TS][TS];
  __shared__ float Bsub[TS][TS];

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

  //const int numTiles = (K%TS==0)? K/TS: K/TS+1;
  const int numTiles = (K+TS-1)/TS;
  for(int t=0; t<numTiles; t++)
  {
	const int tileRow = TS*t + row;
	const int tileCol = TS*t + col;
 	for(int w=0; w<WPT; w++)
  	{
		if(tileCol >=K || (i+w*RTS) >= M)
		       Asub[row+w*RTS][col] = 0.0f;
		else	
		       Asub[row+w*RTS][col] = A[(i+w*RTS)*K+tileCol];

		if(tileRow+w*RTS >=K || j >= N)
		       Bsub[row+w*RTS][col] = 0.0f;
		else
		       Bsub[row+w*RTS][col] = B[(tileRow+w*RTS)*N+j];

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

	for(int w=0; w<WPT;w++)
	{
		if((i+w*RTS) >= M || j >= N) continue;
		C[(i+w*RTS)*N+j] = cc[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];
#define THREADS_PER_BLOCK 512

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 size = (i==num_devices-1)? ((M/num_devices)+(M%num_devices)):M/num_devices; 
    //dim3 blockDim(1, 1, 1);   //the number of threads in x,y,z per block?
    dim3 blockDim(TS/WPT, TS, 1);   //the number of threads in x,y,z per block?
    //dim3 gridDim(Mend[i] - Mbegin[i], N, 1);  //block?
    //dim3 gridDim((((Mend[i] - Mbegin[i]+TS-1)/TS)+WPT-1/WPT), (N+TS-1)/TS, 1);  //block?
    dim3 gridDim((((Mend[i] - Mbegin[i]+TS-1)/TS)), (N+TS-1)/TS, 1);  //block?

    CUDA_CALL( cudaSetDevice(i) );
    sgemm<<<gridDim, blockDim>>>(a_d[i], b_d[i], c_d[i], M, N, K);
    //sgemm<<<gridDim, blockDim>>>(a_d[i], b_d[i], c_d[i], size, 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( cudaDeviceSynchronize() );
  }
}