#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
#define TS 16

int num_devices = 0;

__global__ void sgemm(float *A, float *B, float *C, int M, int N, int K) {

  __shared__ float Asub[TS][TS];
  __shared__ float Bsub[TS][TS];
				 
  int bx = blockIdx.x, by = blockIdx.y,											
      tx = threadIdx.x, ty = threadIdx.y,
      Row = by * TS + ty,
      Col = bx * TS + tx;
  float Pvalue = 0;

  int numARows=M;
  int numAColumns=K;
  int numBRows=K;
  int numBColumns=N;
  int numCRows=M;
  int numCColumns=N;

  const int numTiles = (K+TS-1)/TS;
  for (int m = 0; m < (numAColumns-1)/TS+1; ++m) {
    if (Row < numARows && m*TS+tx < numAColumns)
      Asub[ty][tx] = A[Row*numAColumns + m*TS+tx];
    else
      Asub[ty][tx] = 0;
  
    if (Col < numBColumns && m*TS+ty < numBRows)
      Bsub[ty][tx] = B[(m*TS+ty)*numBColumns+Col];
    else
      Bsub[ty][tx] = 0;

    __syncthreads();
    for (int k = 0; k < TS; ++k)
      Pvalue += Asub[ty][k] * Bsub[k][tx];
    __syncthreads();
	}

  if (Row < numCRows && Col < numCColumns)
    C[Row*numCColumns+Col] = Pvalue;
}

// 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 Mstart[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(TS, TS, 1);
    dim3 gridDim((_N-1)/TS+1, ( Mend[i] - Mstart[i]-1)/TS+1, 1);

    CUDA_CALL( cudaSetDevice(i) );
    sgemm<<<gridDim, blockDim>>>(a_d[i], b_d[i], c_d[i], Mend[i] - Mstart[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++) {
    Mstart[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] - Mstart[i]) * K * sizeof(float)) );
    CUDA_CALL( cudaMalloc(&b_d[i], K * N * sizeof(float)) );
    CUDA_CALL( cudaMalloc(&c_d[i], (Mend[i] - Mstart[i]) * N * sizeof(float)) );
    //printf("numD=%d, start=%d, end=%d\n", num_devices, Mstart[i], Mend[i]);
  }

  // Upload A and B matrix to every GPU
  for (int i = 0; i < num_devices; i++) {							  
    CUDA_CALL( cudaMemcpy(a_d[i], A + Mstart[i] * K,
               (Mend[i] - Mstart[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 + Mstart[i] * N, c_d[i],
               (Mend[i] - Mstart[i]) * N * sizeof(float),
               cudaMemcpyDeviceToHost) );
  }

  // DO NOT REMOVE; NEEDED FOR TIME MEASURE
  for (int i = 0; i < num_devices; i++) {
    CUDA_CALL( cudaDeviceSynchronize() );
  }
}