#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;	// ORG
int num_devices = 4;		// DEBUG

/*
// ORIGINAL VERSION
__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];
  }
}
*/


// 41490 GFLOPS: TS(64), WPT(8)
#define TS 64
#define WPT 8

// 35000 GFLOPS: TS(72), WPT(8)
//#define TS 72
//#define WPT 8

// 30490 GFLOPS: TS(64), WPT(16)
//#define TS 64
//#define WPT 16

// 30000 GFLOPS: TS(64), WPT(4)
//#define TS 64
//#define WPT 4

// 28000 GFLOPS: TS(72), WPT(16)
//#define TS 72
//#define WPT 16

// 17000 GFLOPS: TS(16), WPT(8)
//#define TS 16
//#define WPT 8


#define RTS TS/WPT


// VERSION: PARTITIONED FOR M 
__global__ void sgemm(float *A, float *B, float *C, int M, int N, int K) {
	int bx = blockIdx.x;
	int by = blockIdx.y;
	int tx = threadIdx.x;
	int ty = threadIdx.y;
	int row = threadIdx.y;
	int col = threadIdx.x;
	//int global_row = (blockDim.y*WPT)*blockIdx.y + threadIdx.y;
	//int global_col = (blockDim.x)*blockIdx.x + threadIdx.x;
	int global_row = (blockDim.y*WPT)*by + ty;
	int global_col = (blockDim.x)*bx + tx;

	__shared__ float Asub[TS][TS];
	__shared__ float Bsub[TS][TS];

	float intermediate_val[WPT];

//	printf("DEBUG: (bx,by) = (%d, %d), (tx,ty)=(%d, %d), (bsx,bsy)=(%d,%d), (global_row, global_col)= (%d,%d) \n",bx, by, tx,ty, blockDim.x, blockDim.y, global_row, global_col);

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

	const int num_tiles = (K % TS) ? K/TS +1 : K/TS;      // aligned for TS(tile size)
	for(int t=0; t < num_tiles; t++)  {
		for(int w=0; w < WPT; w++) {
			const int t_row = TS * t + row;
			const int t_col = TS * t + col;

			// for zero padding for aligned area
			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];

			// for zero padding for aligned area
			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();

		for(int k=0; k<TS; k++) {
				for(int w=0; w < WPT; w++) {
					intermediate_val[w] += Asub[row+w*RTS][k] * Bsub[k][col];
				}
		}

		__syncthreads();

	}

	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] = intermediate_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];
//static int Mi[MAX_NUM_GPU];

static int Msz;

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++) {
		int bsx, bsy;
		bsx = TS;
		bsy = TS/WPT;
    dim3 blockDim(bsx, bsy, 1);
		
		Msz = (Mend[i] - Mbegin[i]);

		int gx, gy;
		gx = N;
		gy = (Msz+WPT-1)/WPT;

		// aligned for block
		gx = (gx + bsx -1)/bsx;
		gy = (gy + bsy -1)/bsy;

    dim3 gridDim(gx, gy, 1);


    CUDA_CALL( cudaSetDevice(i) );
    sgemm<<<gridDim, blockDim>>>(a_d[i], b_d[i], c_d[i], Msz, 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() );
  }
}