chundoong-lab-ta/SamsungDS22/submissions/final/junjip.lee/B_bak/cudabak/convolution.cu

295 lines
8.3 KiB
Plaintext

#include "convolution.h"
#include <mpi.h>
#include <stdio.h>
#include "util.h"
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 num_threads = 40;
// 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];
#if 1
__global__ void cuda_conv(
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){
const int globalRow = blockDim.x * blockIdx.x + threadIdx.x;
const int globalCol = blockDim.y * blockIdx.y + threadIdx.y;
int OH, OW;
OH = (_H + 2 *_pad - _dilation * (_R - 1) -1) / _stride + 1;
OW = (_W + 2 *_pad - _dilation * (_S - 1) -1) / _stride + 1;
int n, k , w;
w= globalCol;
n = w/ (_K * OW);
w= w- n *(_K * OW);
k= w/ OW;
w= w-k*OW;
int col = w;
int row = globalRow;
if(globalRow>=OH || globalCol >= _N*_K*OW) return;
int start_row = row * _stride - _pad;
int start_col = col * _stride - _pad;
float o =0.0f;
for (int c = 0; c<_C;c++){
for(int i=0; i<_R;i++){
for(int j=0;j<_S;j++){
int h = start_row + i *_dilation;
int w = start_col + j *_dilation;
if(h<0|| w<0 || h>=_H || w>=_W) continue;
float in = _input[n*_C*_W*_H + c*_W*_H + h*_W+w];
float fil = _filter[k*_C*_R*_S + c*_R*_S + i*_S+j];
o += in * fil;
}
}
}
_output[n*_K*OH*OW + k*OH*OW + row*OW+col]=o;;
}
#if 0
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) {
int size[2];
MPI_Request request;
MPI_Status status;
input = _input;
output = _output;
filter = _filter;
if (mpi_world_size == 2) size[1] = _N / 2;
else size[1] = 0;
size[0] = N - size[1];
OH = (H + 2 * pad - dilation * (R - 1) - 1) / stride + 1;
OW = (W + 2 * pad - dilation * (S - 1) - 1) / stride + 1;
if (mpi_rank == 0 && mpi_world_size == 2) {
MPI_Isend(&input[size[0]*C*H*W], size[1]*C*H*W, MPI_FLOAT, 1, 0, MPI_COMM_WORLD, &request);
MPI_Isend(filter, K*C*R*S, MPI_FLOAT, 1, 0, MPI_COMM_WORLD, &request);
} else if (mpi_world_size == 2) {
alloc_tensor(&input, size[1], C, H, W);
alloc_tensor(&output, size[1], K, OH, OW);
alloc_tensor(&filter, K, C, R, S);
MPI_Recv(input, size[1]*C*H*W, MPI_FLOAT, 0, 0, MPI_COMM_WORLD, &status);
MPI_Recv(filter, K*C*R*S, MPI_FLOAT, 0, 0, MPI_COMM_WORLD, &status);
}
/*
#pragma omp parallel for num_threads(num_threads) schedule(dynamic)
for (int n = 0; n < size[mpi_rank]; ++n) {
#pragma omp parallel for num_threads(num_threads) schedule(dynamic)
for (int k = 0; k < K; ++k) {
#pragma omp parallel for num_threads(num_threads) schedule(dynamic)
for (int oh = 0; oh < OH; ++oh) {
*/
#pragma omp parallel for num_threads(num_threads) collapse(3) schedule(dynamic)
for (int n = 0; n < size[mpi_rank]; ++n) {
for (int k = 0; k < K; ++k) {
for (int oh = 0; oh < OH; ++oh) {
#pragma omp parallel for num_threads(num_threads) schedule(dynamic)
for (int ow = 0; ow < OW; ++ow) {
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;
} // s
} // r
} // c
output[n * K * OH * OW + k * OH * OW + oh * OW + ow] = o;
} //ow
} // oh
} // k
} // n
if (mpi_rank == 0 && mpi_world_size == 2) {
MPI_Recv(&output[size[0]*K*OH*OW], size[1]*K*OH*OW, MPI_FLOAT, 1, 0, MPI_COMM_WORLD, &status);
} else if(mpi_world_size == 2){
MPI_Isend(output, size[1]*K*OH*OW, MPI_FLOAT, 0, 0, MPI_COMM_WORLD, &request);
}
}
#endif
#if 0
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;
if (mpi_rank == 0) {
for (int n = 0; n < N; ++n) {
for (int k = 0; k < K; ++k) {
for (int oh = 0; oh < OH; ++oh) {
for (int ow = 0; ow < OW; ++ow) {
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;
}
}
}
}
}
}
#endif
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;
CUDA_CALL( cudaGetDeviceCount(&num_devices) );
// num_devices=2;
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++) {
NN[i] = (N / num_devices) * i;
NN[i] = (N / num_devices) * (i + 1);
}
NN[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]) * 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)) );
}
*/
for (int i = 0 ; i < num_devices ; i++) {
CUDA_CALL( cudaSetDevice(i) );
CUDA_CALL( cudaMalloc(&in_d[i], NN[i]*C*H*W*sizeof(float)) );
CUDA_CALL( cudaMalloc(&out_d[i], NN[i]*K*OH*OW*sizeof(float)) );
CUDA_CALL( cudaMalloc(&fil_d[i], K*C*R*S*sizeof(float)) );
}
for (int i = 0; i < num_devices; i++) {
CUDA_CALL( cudaMemcpy(in_d[i], input + NN[i]*C*H*W*,
NN[i]*C*H*W* sizeof(float),
cudaMemcpyHostToDevice) );
CUDA_CALL( cudaMemcpy(fil_d[i], filter, K * CRS * sizeof(float), cudaMemcpyHostToDevice) );
}
// Upload A and B matrix to every GPU
for (int i = 0; i < num_devices; i++) {
CUDA_CALL( cudaSetDevice(i) );//jjlee
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 convolution_final(
int _N, int _C, int _H, int _W,
int _K, int _R, int _S,
int _pad, int _dilation, int _stride) {
// Download C matrix from GPUs
for (int i = 0; i < num_devices; i++) {
CUDA_CALL( cudaSetDevice(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() );
}
}