#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 TILE_SIZE 16

__global__ void sgemm(float *A, float *B, float *C, int M, int N, int K) {
  int bx = blockIdx.x, by = blockIdx.y;
  int tx = threadIdx.x, ty = threadIdx.y;

  __shared__ float As[TILE_SIZE][TILE_SIZE];
  __shared__ float Bs[TILE_SIZE][TILE_SIZE];

  int aBegin = K * TILE_SIZE * by;
  int aEnd = aBegin + K - 1;
  int aStep = TILE_SIZE;

  int bBegin = TILE_SIZE * bx;
  int bStep = TILE_SIZE * N;

  float Csub = 0;

  for (int i = aBegin, j = bBegin; i <= aEnd; i += aStep, j += bStep) {
    As[ty][tx] = A[i + K * ty + tx];
    Bs[tx][ty] = B[j + N * tx + ty];

    __syncthreads();

    for (int k = 0; k < TILE_SIZE; ++k) {
      Csub += As[ty][k]*Bs[k][tx];
    }
    
    __syncthreads();
  }
  int cIdx = N * TILE_SIZE * by + TILE_SIZE * bx;
  C[cIdx + N * ty + tx] = Csub;
}

// 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
  // FIXME: Every GPU redundantly calculate all elements of C
  for (int i = 0; i < num_devices; i++) {
    dim3 blockDim(TILE_SIZE, TILE_SIZE, 1);
    dim3 gridDim(N / TILE_SIZE, (Mend[i] - Mbegin[i]) / TILE_SIZE, 1);

    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;

  cudaGetDeviceCount(&num_devices);

  printf("Using %d devices\n", num_devices);
  for (int i = 0; i < num_devices; i++) {
    cudaDeviceProp prop;
    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++) {
    cudaSetDevice(i);
    cudaMalloc(&a_d[i], (Mend[i] - Mbegin[i]) * K * sizeof(float));
    cudaMalloc(&b_d[i], K * N * sizeof(float));
    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++) {
    cudaMemcpy(a_d[i], A + Mbegin[i] * K,
               (Mend[i] - Mbegin[i]) * K * sizeof(float),
               cudaMemcpyHostToDevice);
    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++) {
    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 GPU 0
  // FIXME: Result from the other GPUs are ignored
  for (int i = 0; i < num_devices; i++) {
    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++) {
    cudaDeviceSynchronize();
  }
}