#include "mat_mul.h"
#include "util.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 (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) {
    const int row = threadIdx.x; // Local row ID (max: TS/WIDTH)
    const int col = threadIdx.y; // Local col ID (max: TS)
    const int globalRow = blockDim.x * WPT * blockIdx.x + row;
    const int globalCol = blockDim.y * blockIdx.y + col; // 0..N

  if (globalRow  >= M || globalCol >= N)   return;
    
    // Local memory to fit a tile of TS*TS elements of A and B
  __shared__ float Asub[TS][TS];
  __shared__ float Bsub[TS][TS];
  const int numTiles = K/TS;
  float sum[WPT];
  for (int w=0; w<WPT; w++) {
    sum[w] = 0.0f;
  }

  for (int t=0; t<numTiles; t++) {
    const int tiledRow = TS*t + row;
    const int tiledCol = TS*t + col;
    for (int w=0; w<WPT; w++) {
      Asub[row + w*RTS][col] = A[(globalRow + w*RTS) * K + tiledCol];
      Bsub[row + w*RTS][col] = B[(tiledRow + w*RTS) * N + globalCol];
    }
    // Synchronise to make sure the tile is loaded
    __syncthreads();

    // Perform the computation for a single tile
    for (int k=0; k<TS; k++) {
      for (int w=0; w<WPT; w++) {
        sum[w] += Asub[row + w*RTS][k] * Bsub[k][col];
      }
    }

    // Synchronise before loading the next tile
    __syncthreads();
  }

  // Store the final results in C
  for (int w=0; w<WPT; w++) {
    C[(globalRow + w*RTS) * N + globalCol] = sum[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];
static float *A, *B, *C;
static int CopyBuf = 0;

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(TS/WPT, TS, 1);
    dim3 gridDim(((Mend[i] - Mbegin[i]+TS-1)/TS), (N+TS-1)/TS, 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;
  M = (_M & (TS-1))? _M + (TS - (_M & (TS-1))) : _M;
  N = (_N & (TS-1))? _N + (TS - (_N & (TS-1))) : _N;
  K = (_K & (TS-1))? _K + (TS - (_K & (TS-1))) : _K;
  CopyBuf = 0;
  if(M == _M && K == _K){
    A = _A;
  }
  else{
    alloc_mat(&A, M, K);

    for (int i = 0; i < M; i++) { 
      for (int j = 0; j < K; j++) {
        if(i<_M && j<_K)
          A[i * K + j] = _A[i * _K + j];
        else
          A[i * K + j] = 0;
      }
    }
    CopyBuf = 1;
  }

  if(K == _K && N == _N){
    B = _B;
  }
  else{
    alloc_mat(&B, K, N);

    for (int i = 0; i < K; i++) { 
      for (int j = 0; j < N; j++) {
        if(i<_K && j<_N)
          B[i * N + j] = _B[i * _N + j];
        else
          B[i * N + j] = 0;
      }
    }
    CopyBuf = 1;
  }
  
  if(M == _M && N == _N){
    C = _C;
  }
  else{
    alloc_mat(&C, M, N);
    zero_mat(C, M, N);
    CopyBuf = 1;
  }
  /////////////////////////
  CUDA_CALL( cudaGetDeviceCount(&num_devices) );
  
  if(num_devices > MAX_NUM_GPU) num_devices = MAX_NUM_GPU;

  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) );
  }

  if(CopyBuf == 1){
    for (int i = 0; i < _M; i++) { 
      for (int j = 0; j < _N; j++) {
          _C[i * _N + j] = C[i * N + j];
      }
    }
  }
  
  // DO NOT REMOVE; NEEDED FOR TIME MEASURE
  for (int i = 0; i < num_devices; i++) {
    CUDA_CALL( cudaDeviceSynchronize() );
  }
}