#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;

__global__ void sgemm(float *A, float *B, float *C, int M, int N, int K) {
#if 1
#define TS 32
#define WPT 16
#define RTS (TS/WPT)

  int r = threadIdx.x; //get_local_id(0);
  int c = threadIdx.y; //get_local_id(1);
  int g_r = (blockDim.x * WPT) * blockIdx.x + r; //TS * get_group_id(0) + r;
  int g_c = blockDim.y * blockIdx.y + c; //TS * get_group_id(1) + c;

  __shared__ float PA[TS][TS];
  __shared__ float PB[TS][TS];

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

  int n_tiles = (K + TS - 1) / TS;
  for (int t = 0; t < n_tiles; t++) {
    for (int w = 0; w < WPT; w++) {
      int t_r = TS * t + r;
      int t_c = TS * t + c;
      PA[r + w * RTS][c] = ((g_r + w * RTS) >= M || t_c >= K)? 0 : A[(g_r + w * RTS) * K + t_c];
      PB[r + w * RTS][c] = ((t_r + w * RTS) >= K || g_c >= N)? 0 : B[(t_r + w * RTS) * N + g_c];
    }

		__syncthreads();

    for (int k = 0; k < TS; k++) {
      for (int w = 0; w < WPT; w++) {
        tmp[w] += PA[r + w * RTS][k] * PB[k][c];
      }
    }

		__syncthreads();
  }

  for (int w = 0; w < WPT; w++) {
    if ((g_r + w * RTS) >= M || g_c >= N) continue;
    C[(g_r + w * RTS) * N + g_c] = tmp[w];
  }

#else
  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];
  }
#endif
}

// 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 TS 32
#define WPT 16

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;

  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() );
  }
}