#include "convolution.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 conv_kernel(float *input, float *filter, float *output,
    int N, int C, int H, int W,
    int K, int R, int S,
    int pad, int dilation, int stride) {
  
  int n = blockIdx.x;
  int k = blockIdx.y;;
  int oh = blockIdx.z;
  int ow = threadIdx.x;
  int OH = (H + 2 * pad - dilation * (R - 1) - 1) / stride + 1;
  int OW = (W + 2 * pad - dilation * (S - 1) - 1) / stride + 1;

  float o = 0.f;
  for (int c = 0; c < C; ++c) {
    for (int r = 0; r < R; ++r) {
      for (int s = 0; s < S; ++s) {
        int h = oh * stride - pad + r * dilation;
        int w = ow * stride - pad + s * dilation;
        if (h < 0 || h >= H || w < 0 || w >= W) continue;
        float i = input[n * C * H * W + c * H * W + h * W + w];
        float f = filter[k * C * R * S + c * R * S + r * S + s];
        o += i * f;
      }
    }
  }
  output[n * K * OH * OW + k * OH * OW + oh * OW + ow] = o;

}

// 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 N, C, H, W;
static int K, R, S;
static int OH, OW;
static int pad;
static int dilation;
static int stride;
static int Mbegin[MAX_NUM_GPU], Mend[MAX_NUM_GPU];


void cuda_conv() {

  // Launch kernel on every GPU
  for (int i = 0; i < num_devices; i++) {
    dim3 blockDim(OW, 1, 1);
    dim3 gridDim(Mend[i] - Mbegin[i], K, OH);

    CUDA_CALL( cudaSetDevice(i) );
    conv_kernel<<<gridDim, blockDim>>>(a_d[i], b_d[i], c_d[i], Mend[i]-Mbegin[i], C, H, W, K, R, S, pad, dilation, stride);
  }

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

void cuda_mem_init(int _N, int _C, int _H, int _W, int _K, int _R, int _S, int _pad, int _dilation, int _stride) {
  N = _N; C = _C; H = _H; W = _W;
  K = _K; R = _R; S = _S;
  pad = _pad;
  dilation = _dilation;
  stride = _stride;
  OH = (H + 2 * pad - dilation * (R - 1) - 1) / stride + 1;
  OW = (W + 2 * pad - dilation * (S - 1) - 1) / stride + 1;

  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] = (N / num_devices) * i;
    Mend[i] = (N / num_devices) * (i + 1);
  }
  Mend[num_devices - 1] = N;

  // 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]) * C * H * W * sizeof(float)) );
    CUDA_CALL( cudaMalloc(&b_d[i], K * C * R * S * sizeof(float)) );
    CUDA_CALL( cudaMalloc(&c_d[i], (Mend[i] - Mbegin[i]) * K * OH * OW * sizeof(float)) );
  }
}

void cuda_conv_init(float *input, float *filter, float *output) {
  // Upload A and B matrix to every GPU
  for (int i = 0; i < num_devices; i++) {
    CUDA_CALL( cudaMemcpy(a_d[i], input + Mbegin[i] * C * H * W,
               (Mend[i] - Mbegin[i]) * C * H * W * sizeof(float),
               cudaMemcpyHostToDevice) );
    CUDA_CALL( cudaMemcpy(b_d[i], filter, K * C * R * S * sizeof(float), cudaMemcpyHostToDevice) );
  }

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

void cuda_conv_final(float *output) {
  // Do any post-matmul cleanup work here.

  // Download C matrix from GPUs
  for (int i = 0; i < num_devices; i++) {
    CUDA_CALL( cudaMemcpy(output + Mbegin[i] * K * OH * OW, c_d[i],
               (Mend[i] - Mbegin[i]) * K * OH * OW * sizeof(float),
               cudaMemcpyDeviceToHost) );
  }

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