252 lines
8.3 KiB
Plaintext
252 lines
8.3 KiB
Plaintext
|
#include "convolution.h"
|
||
|
|
#include <mpi.h>
|
||
|
|
#include "util.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_NODE 2
|
||
|
|
#define MAX_NUM_GPU 4
|
||
|
|
int num_devices = 0;
|
||
|
|
|
||
|
|
#define TS 32
|
||
|
|
__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 OH, int OW,
|
||
|
|
int pad, int dilation, int stride) {
|
||
|
|
int x = threadIdx.x;
|
||
|
|
int y = threadIdx.y;
|
||
|
|
int global_x = blockDim.x * blockIdx.x + x;
|
||
|
|
int global_y = blockDim.y * blockIdx.y + y;
|
||
|
|
|
||
|
|
if (global_x >= N * OW || global_y >= K * OH) return; // boundary check
|
||
|
|
|
||
|
|
int n = global_x / OW;
|
||
|
|
int ow = global_x % OW;
|
||
|
|
int k = global_y / OH;
|
||
|
|
int oh = global_y % OH;
|
||
|
|
|
||
|
|
float o = 0.f;
|
||
|
|
for (int c = 0; c < C; ++c) {
|
||
|
|
for (int r = 0; r < R; ++r) {
|
||
|
|
int h = oh * stride - pad + r * dilation;
|
||
|
|
if (h < 0 || h >= H) continue;
|
||
|
|
for (int s = 0; s < S; ++s) {
|
||
|
|
int w = ow * stride - pad + s * dilation;
|
||
|
|
if (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;
|
||
|
|
}
|
||
|
|
|
||
|
|
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 get_size_per_rank(int rank) {
|
||
|
|
const int NN = N / mpi_world_size;
|
||
|
|
|
||
|
|
if (rank == -1) return 0;
|
||
|
|
else if (rank != 0) return NN;
|
||
|
|
else return (N - (mpi_world_size - 1) * NN);
|
||
|
|
}
|
||
|
|
|
||
|
|
int get_begin_index(int rank) {
|
||
|
|
if (rank == 0) return 0;
|
||
|
|
else if (rank == 1) return get_size_per_rank(0);
|
||
|
|
else return get_begin_index(rank - 1) + get_size_per_rank(rank - 1);
|
||
|
|
}
|
||
|
|
|
||
|
|
// Array of device (GPU) pointers
|
||
|
|
static float *a_d[MAX_NUM_NODE][MAX_NUM_GPU];
|
||
|
|
static float *b_d[MAX_NUM_NODE][MAX_NUM_GPU];
|
||
|
|
static float *c_d[MAX_NUM_NODE][MAX_NUM_GPU];
|
||
|
|
static int Nbegin[MAX_NUM_NODE][MAX_NUM_GPU], Nend[MAX_NUM_NODE][MAX_NUM_GPU];
|
||
|
|
static int Nsize[MAX_NUM_NODE][MAX_NUM_GPU];
|
||
|
|
|
||
|
|
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;
|
||
|
|
|
||
|
|
if (mpi_rank != 0) {
|
||
|
|
alloc_tensor(&input, N, C, H, W);
|
||
|
|
alloc_tensor(&output, N, K, OH, OW);
|
||
|
|
alloc_tensor(&filter, K, C, R, S);
|
||
|
|
}
|
||
|
|
MPI_Barrier(MPI_COMM_WORLD);
|
||
|
|
|
||
|
|
MPI_Bcast(filter, K * C * R * S, MPI_FLOAT, 0, MPI_COMM_WORLD);
|
||
|
|
if (mpi_rank == 0) {
|
||
|
|
for (int i = 1; i < mpi_world_size; ++i) {
|
||
|
|
MPI_Request request;
|
||
|
|
int index = get_begin_index(i) * C * H * W;
|
||
|
|
int size = get_size_per_rank(i) * C * H * W;
|
||
|
|
MPI_Isend(&input[index], size, MPI_FLOAT, i, 0, MPI_COMM_WORLD, &request);
|
||
|
|
}
|
||
|
|
} else {
|
||
|
|
MPI_Request request;
|
||
|
|
int index = get_begin_index(mpi_rank) * C * H * W;
|
||
|
|
int size = get_size_per_rank(mpi_rank) * C * H * W;
|
||
|
|
|
||
|
|
MPI_Irecv(&input[index], size, MPI_FLOAT, 0, 0, MPI_COMM_WORLD, &request);
|
||
|
|
zero_tensor(output, N, K, OH, OW);
|
||
|
|
MPI_Wait(&request, MPI_STATUS_IGNORE);
|
||
|
|
}
|
||
|
|
|
||
|
|
// Upload A and B matrix to every GPU
|
||
|
|
for (int i = 0; i < num_devices; i++) {
|
||
|
|
if (Nsize[mpi_rank][i] == 0) continue;
|
||
|
|
CUDA_CALL( cudaSetDevice(i) );
|
||
|
|
CUDA_CALL( cudaMemcpy(a_d[mpi_rank][i], &input[Nbegin[mpi_rank][i] * C * H * W],
|
||
|
|
Nsize[mpi_rank][i] * C * H * W * sizeof(float),
|
||
|
|
cudaMemcpyHostToDevice) );
|
||
|
|
CUDA_CALL( cudaMemcpy(b_d[mpi_rank][i], filter, K * C * R * S * sizeof(float), cudaMemcpyHostToDevice) );
|
||
|
|
}
|
||
|
|
|
||
|
|
// DO NOT REMOVE; NEEDED FOR TIME MEASURE
|
||
|
|
for (int i = 0; i < num_devices; i++) {
|
||
|
|
if (Nsize[mpi_rank][i] == 0) continue;
|
||
|
|
CUDA_CALL( cudaSetDevice(i) );
|
||
|
|
CUDA_CALL( cudaDeviceSynchronize() );
|
||
|
|
}
|
||
|
|
// Launch kernel on every GPU
|
||
|
|
for (int i = 0; i < num_devices; i++) {
|
||
|
|
if (Nsize[mpi_rank][i] == 0) continue;
|
||
|
|
int gws[3] = {Nsize[mpi_rank][i] * OW, K * OH, 1};
|
||
|
|
int lws[3] = {TS, TS, 1};
|
||
|
|
for (int j = 0; j < 3; ++j) {
|
||
|
|
gws[j] = (gws[j] + lws[j] - 1) / lws[j] * lws[j];
|
||
|
|
}
|
||
|
|
dim3 blockDim(lws[0], lws[1], lws[2]);
|
||
|
|
dim3 gridDim(gws[0] / lws[0], gws[1] / lws[1], gws[2] / lws[2]);
|
||
|
|
|
||
|
|
CUDA_CALL( cudaSetDevice(i) );
|
||
|
|
cuda_conv<<<gridDim, blockDim>>>(a_d[mpi_rank][i], c_d[mpi_rank][i], b_d[mpi_rank][i], Nsize[mpi_rank][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++) {
|
||
|
|
if (Nsize[mpi_rank][i] == 0) continue;
|
||
|
|
CUDA_CALL( cudaSetDevice(i) );
|
||
|
|
CUDA_CALL( cudaDeviceSynchronize() );
|
||
|
|
}
|
||
|
|
|
||
|
|
// Download C matrix from GPUs
|
||
|
|
for (int i = 0; i < num_devices; i++) {
|
||
|
|
if (Nsize[mpi_rank][i] == 0) continue;
|
||
|
|
CUDA_CALL( cudaMemcpy(output + Nbegin[mpi_rank][i] * K * OH * OW, c_d[mpi_rank][i],
|
||
|
|
Nsize[mpi_rank][i] * K * OH * OW * sizeof(float),
|
||
|
|
cudaMemcpyDeviceToHost) );
|
||
|
|
}
|
||
|
|
|
||
|
|
// DO NOT REMOVE; NEEDED FOR TIME MEASURE
|
||
|
|
for (int i = 0; i < num_devices; i++) {
|
||
|
|
if (Nsize[mpi_rank][i] == 0) continue;
|
||
|
|
CUDA_CALL( cudaSetDevice(i) );
|
||
|
|
CUDA_CALL( cudaDeviceSynchronize() );
|
||
|
|
}
|
||
|
|
|
||
|
|
if (mpi_rank == 0) {
|
||
|
|
MPI_Request request[mpi_world_size];
|
||
|
|
for (int i = 1; i < mpi_world_size; ++i) {
|
||
|
|
int index = get_begin_index(i) * K * OH * OW;
|
||
|
|
int size = get_size_per_rank(i) * K * OH * OW;
|
||
|
|
MPI_Irecv(&output[index], size, MPI_FLOAT, i, 0, MPI_COMM_WORLD, &request[i]);
|
||
|
|
}
|
||
|
|
for (int i = 1; i < mpi_world_size; ++i) {
|
||
|
|
MPI_Wait(&request[i], MPI_STATUS_IGNORE);
|
||
|
|
}
|
||
|
|
} else {
|
||
|
|
int index = get_begin_index(mpi_rank) * K * OH * OW;
|
||
|
|
int size = get_size_per_rank(mpi_rank) * K * OH * OW;
|
||
|
|
MPI_Send(&output[index], size, MPI_FLOAT, 0, 0, MPI_COMM_WORLD);
|
||
|
|
}
|
||
|
|
}
|
||
|
|
|
||
|
|
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);
|
||
|
|
|
||
|
|
CUDA_CALL( cudaGetDeviceCount(&num_devices) );
|
||
|
|
if (num_devices > MAX_NUM_GPU)
|
||
|
|
num_devices = MAX_NUM_GPU;
|
||
|
|
|
||
|
|
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);
|
||
|
|
}
|
||
|
|
|
||
|
|
OH = (H + 2 * pad - dilation * (R - 1) - 1) / stride + 1;
|
||
|
|
OW = (W + 2 * pad - dilation * (S - 1) - 1) / stride + 1;
|
||
|
|
|
||
|
|
// Setup problem size for each GPU
|
||
|
|
int NN = get_size_per_rank(mpi_rank);
|
||
|
|
int NNbegin = get_begin_index(mpi_rank);
|
||
|
|
for (int i = 0; i < num_devices; i++) {
|
||
|
|
Nbegin[mpi_rank][i] = (NN / num_devices) * i + NNbegin;
|
||
|
|
Nend[mpi_rank][i] = (NN / num_devices) * (i + 1) + NNbegin;
|
||
|
|
Nsize[mpi_rank][i] = Nend[mpi_rank][i] - Nbegin[mpi_rank][i];
|
||
|
|
}
|
||
|
|
Nend[mpi_rank][num_devices - 1] = NN + NNbegin;
|
||
|
|
Nsize[mpi_rank][num_devices - 1] = Nend[mpi_rank][num_devices - 1] - Nbegin[mpi_rank][num_devices - 1];
|
||
|
|
|
||
|
|
// Allocate device memory for each GPU
|
||
|
|
for (int i = 0; i < num_devices; i++) {
|
||
|
|
if (Nsize[mpi_rank][i] == 0) continue;
|
||
|
|
CUDA_CALL( cudaSetDevice(i) );
|
||
|
|
CUDA_CALL( cudaMalloc(&a_d[mpi_rank][i], Nsize[mpi_rank][i] * C * H * W * sizeof(float)) );
|
||
|
|
CUDA_CALL( cudaMalloc(&b_d[mpi_rank][i], K * C * R * S * sizeof(float)) );
|
||
|
|
CUDA_CALL( cudaMalloc(&c_d[mpi_rank][i], Nsize[mpi_rank][i] * K * OH * OW * sizeof(float)) );
|
||
|
|
}
|
||
|
|
}
|
||
|
|
|
||
|
|
void convolution_final(
|
||
|
|
int _N, int _C, int _H, int _W,
|
||
|
|
int _K, int _R, int _S,
|
||
|
|
int _pad, int _dilation, int _stride) {
|
||
|
|
|
||
|
|
}
|