#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 NUM_WORK_ITEM (32)
#define VECTOR_WIDTH (16)
#define RTS (NUM_WORK_ITEM / VECTOR_WIDTH)
#define ALIGN_UP(_X, _Y) (((_X) + (_Y) - 1) & ~((_Y) - 1))
int num_devices = 0;

__global__ void sgemm(float *__restrict A, float *__restrict B, float *__restrict C, int M, int N, int K, int NON_OPTIMAL) 
{
// blockDim.x -> NUM_WORK_ITEM
  const int i = threadIdx.x; // row index of C
  const int j = threadIdx.y; // column index of C
  const int global_row = NUM_WORK_ITEM * blockIdx.x + i;
  const int global_col = NUM_WORK_ITEM * blockIdx.y + j;
  float intermediate_val[16] = {0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f};
    
  __shared__ float tileA[NUM_WORK_ITEM][NUM_WORK_ITEM];
  __shared__ float tileB[NUM_WORK_ITEM][NUM_WORK_ITEM];

  if (NON_OPTIMAL == 0)
  {
    const int num_tiles = K / NUM_WORK_ITEM;
    
    // printf("i : %d, j : %d, global_row : %d, global_col : %d\n", i, j, global_row, global_col);
    for (int t = 0; t < num_tiles; t++)
    {
      for (int w = 0; w < VECTOR_WIDTH; w++)
      {
        const int t_row = NUM_WORK_ITEM * t + i;
        const int t_col = NUM_WORK_ITEM * t + j;

        tileA[i + w * RTS][j] = A[((global_row + w * RTS)) * K + t_col];
        tileB[i + w * RTS][j] = B[((t_row + w * RTS)) * N + global_col];
      }
      
      __syncthreads();

      for (int k = 0; k < NUM_WORK_ITEM; k++)
      {
        for (int w = 0; w < VECTOR_WIDTH; w++)
        {
          intermediate_val[w] += tileA[i + w * RTS][k] * tileB[k][j];
        }
      }

      __syncthreads();
    }

    for (int w = 0; w < VECTOR_WIDTH; w++)
    {
      C[(global_row + w * RTS) * N + global_col] = intermediate_val[w];
    }
  }
  else
  {   
    const int num_tiles = (K + NUM_WORK_ITEM - 1)  / NUM_WORK_ITEM;
    
    // printf("i : %d, j : %d, global_row : %d, global_col : %d\n", i, j, global_row, global_col);
    for (int t = 0; t < num_tiles; t++)
    {
      for (int w = 0; w < VECTOR_WIDTH; w++)
      {
        const int t_row = NUM_WORK_ITEM * t + i;
        const int t_col = NUM_WORK_ITEM * t + j;
        if (global_row + w * RTS >= M || t_col >= K) 
        {
          tileA[i + w * RTS][j] = 0.0f;
        }
		    else 
        {
          tileA[i + w * RTS][j] = A[((global_row + w * RTS)) * K + t_col];
        }

        if (t_row + w * RTS >= K || global_col >= N) 
        {
          tileB[i + w * RTS][j] = 0.0f;
        }
        else 
        {
          tileB[i + w * RTS][j] = B[((t_row + w * RTS)) * N + global_col];
        }
      }
      
      __syncthreads();

      for (int k = 0; k < NUM_WORK_ITEM; k++)
      {
        for (int w = 0; w < VECTOR_WIDTH; w++)
        {
          intermediate_val[w] += tileA[i + w * RTS][k] * tileB[k][j];
        }
      }

      __syncthreads();
    }

    for (int w = 0; w < VECTOR_WIDTH; w++)
    {
      if(global_row + w * RTS >= M || global_col >= N)
      {
        break;
      } 
		  else 
      {
        C[(global_row + w * RTS) * N + global_col] = intermediate_val[w];
      }
    }
    // printf("C[%d] = , %+.3f, %+.3f, %+.3f, %+.3f,...\n", (global_row) * N + global_col, intermediate_val[0],intermediate_val[1],intermediate_val[2],intermediate_val[3]);

  }
}

// 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], Msize[MAX_NUM_GPU];
static int NON_OPTIMAL;

void mat_mul(float *_A, float *_B, float *_C, int _M, int _N, int _K) {
  
  if (_M % (NUM_WORK_ITEM * num_devices) != 0
    || _N % NUM_WORK_ITEM != 0
    || _K % NUM_WORK_ITEM != 0)
    {
      NON_OPTIMAL = 1; 
    }
    else
    {
      NON_OPTIMAL = 0;
    }

  // Launch kernel on every GPU
  for (int i = 0; i < num_devices; i++) {
    dim3 blockDim(RTS, NUM_WORK_ITEM, 1);
    dim3 gridDim(ALIGN_UP(Msize[i], NUM_WORK_ITEM) / NUM_WORK_ITEM , \
                  ALIGN_UP(N, NUM_WORK_ITEM) / NUM_WORK_ITEM, 1);

    CUDA_CALL( cudaSetDevice(i) );
    sgemm<<<gridDim, blockDim>>>(a_d[i], b_d[i], c_d[i], Msize[i], N, K, NON_OPTIMAL);
  }

  // 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);
    Msize[i] = Mend[i] - Mbegin[i];
  }
  Mend[num_devices - 1] = M;
  Msize[num_devices - 1] = Mend[num_devices - 1] - Mbegin[num_devices - 1];

  // Allocate device memory for each GPU
  for (int i = 0; i < num_devices; i++) {
    CUDA_CALL( cudaSetDevice(i) );
    CUDA_CALL( cudaMalloc(&a_d[i], (Msize[i]) * K * sizeof(float)) );
    CUDA_CALL( cudaMalloc(&b_d[i], K * N * sizeof(float)) );
    CUDA_CALL( cudaMalloc(&c_d[i], (Msize[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,
               (Msize[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],
               (Msize[i]) * N * sizeof(float),
               cudaMemcpyDeviceToHost) );
  }

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