#include "convolution.h"
#include <mpi.h>
#include <omp.h>
#include <stdio.h>
#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 MAX_NUM_NODE        4

#define THREADS_PER_BLOCK   16
#define THREADS_DIM			4
#define WPT                 8
#define RTS                 2

int num_devices = 0;

__global__ void sgemm(float *input, float *output, float *filter, 
	int N, int C, int H, int W, 
	int K, int R, int S,
	int OH, int OW,
	int pad, int dilation, int stride) {

	int row = threadIdx.x;
	int col = threadIdx.y;
	int dim = threadIdx.z;

	int oh = blockDim.x * blockIdx.x + threadIdx.x;
	int ow = blockDim.y * blockIdx.y + threadIdx.y;
	int k = blockDim.z * blockIdx.z + threadIdx.z;

	__shared__ float fsub[THREADS_DIM][THREADS_PER_BLOCK][THREADS_PER_BLOCK];

	int r_q = R / THREADS_PER_BLOCK;
	int r_r = R % THREADS_PER_BLOCK;
	if (r_r != 0)
		r_q += 1;

	int s_q = S / THREADS_PER_BLOCK;
	int s_r = S % THREADS_PER_BLOCK;
	if (s_r != 0)
		s_q += 1;


	for (int n = 0; n < N; n++) {
		float temp = 0.0f;
		for (int c = 0; c < C; c++) {
			for (int rr = 0; rr < r_q; rr++) {
				int f_row = THREADS_PER_BLOCK * rr + row;
				for (int ss = 0; ss < s_q; ss++) {
					int f_col = THREADS_PER_BLOCK * ss + col;
					if (f_row >= R || f_col >= S || k >= K)
						fsub[dim][row][col] = 0;
					else
						fsub[dim][row][col] = filter[k * C * R * S + c * R * S + f_row * S + f_col];

					__syncthreads();

					for (int r = 0; r < THREADS_PER_BLOCK; r++) {
						int h = oh * stride - pad + (THREADS_PER_BLOCK * rr + r) * dilation;
						for (int s = 0; s < THREADS_PER_BLOCK; s++) {
							int w = ow * stride - pad + (THREADS_PER_BLOCK * ss + s) * dilation;
							if (oh >= OH || ow >= OW || h < 0 || h >= H || w < 0 || w >= W) continue;
							float i = input[n * C * H * W + c * H * W + h * W + w];
							float f = fsub[dim][r][s];
							temp += i * f;
						}
					}
					//printf("temp : %f\n", temp);
					__syncthreads();
				}
			}
		}

		if (oh < OH && ow < OW && k < K)
			output[n * K * OH * OW + k * OH * OW + oh * OW + ow] = temp;
	}

}

// 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 Mbegin[MAX_NUM_GPU], Mend[MAX_NUM_GPU];

static float *input, *output, *filter;
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 mpi_rank, mpi_world_size;

int is[MAX_NUM_NODE], ie[MAX_NUM_NODE];
int count;

extern void alloc_tensor(float **t, int D0, int D1, int D2, int D3);

void convolution(
    float *_input, float *_output, float *_filter, 
    int _N, int _C, int _H, int _W,
    int _K, int _R, int _S,
    int _pad, int _dilation, int _stride) {

	input = _input;
	output = _output;
	filter = _filter;

	OH = (H + 2 * pad - dilation * (R - 1) - 1) / stride + 1;
	OW = (W + 2 * pad - dilation * (S - 1) - 1) / stride + 1;

	MPI_Status status;

	for (int i = 0; i < mpi_world_size; i++) {
		is[i] = N / mpi_world_size * i;
		ie[i] = N / mpi_world_size * (i + 1);
	}

	ie[mpi_world_size - 1] = N;

	count = ie[mpi_rank] - is[mpi_rank];

	if (mpi_rank == 0) {
		for (int i = 1; i < mpi_world_size; i++) {
			MPI_Send(&input[is[i] * C * H * W], (ie[i] - is[i]) * C * H * W, MPI_FLOAT, i, 0, MPI_COMM_WORLD);
		}
	} else {
		alloc_tensor(&input, count, C, H, W);
		alloc_tensor(&output, count, K, OH, OW);
		alloc_tensor(&filter, K, C, R, S);
		MPI_Recv(input, count * C * H * W, MPI_FLOAT, 0, 0, MPI_COMM_WORLD, &status);
	}

	MPI_Bcast(filter, K * C * R * S, MPI_FLOAT, 0, MPI_COMM_WORLD);

	// CUDA init
	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("NODE %d [GPU %d] %s\n", mpi_rank, 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] = (count / num_devices) * i;
		Mend[i] = (count / num_devices) * (i + 1);
	}
	Mend[num_devices - 1] = count;

	// 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)) );
	}

	// 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() );
	}

	// Launch kernel on every GPU
	for (int i = 0; i < num_devices; i++) {

		dim3 blockDim(THREADS_PER_BLOCK, THREADS_PER_BLOCK, THREADS_DIM);
		dim3 gridDim(((OH + THREADS_PER_BLOCK -1) / THREADS_PER_BLOCK), ((OW + THREADS_PER_BLOCK -1) / THREADS_PER_BLOCK), (K + THREADS_DIM -1) / THREADS_DIM);

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

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

}

void convolution_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;

	MPI_Comm_rank(MPI_COMM_WORLD, &mpi_rank);
	MPI_Comm_size(MPI_COMM_WORLD, &mpi_world_size);
}

void convolution_final(
    int _N, int _C, int _H, int _W,
    int _K, int _R, int _S,
    int _pad, int _dilation, int _stride) {
	  // Do any post-matmul cleanup work here.

	MPI_Status status;

	  // 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() );
	}

	if (mpi_rank == 0) {
		for (int i = 1; i < mpi_world_size; i++) {
			MPI_Recv(&output[is[i] * K * OH * OW], (ie[i] - is[i]) * K * OH * OW, MPI_FLOAT, 1, 0, MPI_COMM_WORLD, &status);
		}
	}
	else {
		MPI_Send(output, count * K * OH * OW, MPI_FLOAT, 0, 0, MPI_COMM_WORLD);
	}
}