#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 8
#define RTS (TS/WPT)

#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 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];
//  }  
  // Thread identifiers
  const int row = threadIdx.x; 
  const int col = threadIdx.y; 
  const int global_row = (blockDim.x * WPT) * blockIdx.x + threadIdx.x; 
  const int global_col = blockDim.y * blockIdx.y + threadIdx.y; 

  // Local memory to fit a tile of TS * TS elements of A and B
  __shared__ float Asub[TS][TS];
  __shared__ float Bsub[TS][TS];

  // initialize the accumulation registers
  float accumulation_val[WPT];
  for(int w = 0; w < WPT; w++) {
	  accumulation_val[w] = 0.0f; 
  }

  // const int num_tiles = (K % TS) > 0 ? K/TS + 1 : K / TS;
  // Loop over all tiles
  const int num_tiles = (K + TS - 1) / TS;
  for(int t = 0; t < num_tiles; t++){
  // Load one tile of A and B into local memory
	  for(int w = 0; w < WPT; w++){
		  const int t_row = TS * t + row;
		  const int t_col = TS * t + col;

		  if(global_row + w * RTS >= M || t_col >= K) {Asub[row + w * RTS][col]=0.0f;}
		  else 
		  Asub[row + w * RTS][col] = A[(global_row + w * RTS) * K + t_col];
		  
		  if(t_row + w * RTS >= K || global_col >= N) {Bsub[row + w * RTS][col]=0.0f;}
		  else 
		  Bsub[row + w * RTS][col] = B[(t_row + w * RTS) * N + global_col];
	 }
	  __syncthreads();

  // Perform the computation for a single tile
  	for (int k = 0; k < TS; k++) {  
	   for(int w = 0; w < WPT; w++) { 
    		accumulation_val[w] += Asub[row + w * RTS][k] * Bsub[k][col];
	   }
	}
	  __syncthreads();
  }
  // Store the final results
	for(int w = 0; w < WPT; w++) {
		if(global_row + w * RTS >= M || global_col >= N) continue;
		else C[(global_row + w * RTS) * N + global_col] = accumulation_val[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++) {
  // dim3 blockDim(1, 1, 1);
	  dim3 blockDim(TS / WPT, TS, 1);
	  dim3 gridDim( ((Mend[i] - Mbegin[i] + TS - 1) / TS), (N + TS - 1) / TS, 1);
  //dim3 gridDim(Mend[i] - Mbegin[i], N, 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) );

  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() );
  }
}