233 lines
7.2 KiB
Plaintext
233 lines
7.2 KiB
Plaintext
#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 TS 32
|
|
#define WPT 32
|
|
#define RTS 1
|
|
#define TSM TS
|
|
#define TSN TS
|
|
#define TSK TS
|
|
#define CEIL_DIV(x,y) (((x) + (y) - 1) / (y))
|
|
#define PADDINGX TS
|
|
#define PADDINGY TS
|
|
|
|
int num_devices = 0;
|
|
|
|
__global__ void sgemm(float *A, float *B, float *C, int M, int N, int K) {
|
|
const int row = threadIdx.x;
|
|
const int col = threadIdx.y;
|
|
const int global_row = ( blockDim.x * WPT ) * blockIdx.x + threadIdx.x;
|
|
const int global_col = blockDim.y * blockIdx.y + threadIdx.y;
|
|
|
|
__shared__ float Asub[TS][TS];
|
|
__shared__ float Bsub[TS][TS];
|
|
|
|
float Csub[WPT];
|
|
for (int w = 0; w < WPT; w++) {
|
|
Csub[w] = 0.0f;
|
|
}
|
|
|
|
const int num_tiles = (K + TS - 1) / TS;
|
|
|
|
for (int t = 0; t < num_tiles; t++) {
|
|
#pragma unroll
|
|
for (int w = 0; w < WPT; w++) {
|
|
const int t_row = TS * t + row;
|
|
const int t_col = TS * t + col;
|
|
Asub[row + w*RTS][col] = A[(global_row + w*RTS) * K + t_col];
|
|
Bsub[row + w*RTS][col] = B[(t_row + w*RTS) * N + global_col];
|
|
}
|
|
|
|
__syncthreads();
|
|
|
|
for (int k = 0; k < TS; k++) {
|
|
for (int w = 0; w < WPT; w++) {
|
|
Csub[w] += Asub[row + w*RTS][k] * Bsub[k][col];
|
|
}
|
|
}
|
|
|
|
__syncthreads();
|
|
}
|
|
|
|
#pragma unroll
|
|
for (int w = 0; w < WPT; w++) {
|
|
C[(global_row + w*RTS) * N + global_col] = Csub[w];
|
|
}
|
|
}
|
|
|
|
// Pad the P * Q matrix with zeroes to form a P_XL * Q_XL matrix
|
|
__global__ void paddingAddZeroes(const int P, const int Q,
|
|
const float* input,
|
|
const int P_XL, const int Q_XL,
|
|
float* output) {
|
|
|
|
const int tx = blockIdx.x * blockDim.x + threadIdx.x;
|
|
const int ty = blockIdx.y * blockDim.y + threadIdx.y;
|
|
|
|
// Check whether we are within bounds of the XL matrix
|
|
if (tx < P_XL && ty < Q_XL) {
|
|
// Copy the input or pad a zero
|
|
float value;
|
|
if (tx < P && ty < Q) {
|
|
value = input[tx*Q + ty];
|
|
}
|
|
else {
|
|
value = 0.0f;
|
|
}
|
|
|
|
output[tx*Q_XL + ty] = value;
|
|
}
|
|
}
|
|
|
|
// Remove padded values from a P_XL * Q_XL matrix to form a P * Q matrix
|
|
__global__ void paddingRemoveZeroes(const int P_XL, const int Q_XL,
|
|
const float* input,
|
|
const int P, const int Q,
|
|
float* output) {
|
|
|
|
const int tx = blockIdx.x * blockDim.x + threadIdx.x;
|
|
const int ty = blockIdx.y * blockDim.y + threadIdx.y;
|
|
|
|
// Only store the result if within P * Q bounds
|
|
if (tx < P && ty < Q) {
|
|
output[tx*Q + ty] = input[tx*Q_XL + ty];
|
|
}
|
|
}
|
|
|
|
// 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 float *a_ad[MAX_NUM_GPU];
|
|
static float *b_ad[MAX_NUM_GPU];
|
|
static float *c_ad[MAX_NUM_GPU];
|
|
static int M, N, K;
|
|
static int Mbegin[MAX_NUM_GPU], Mend[MAX_NUM_GPU];
|
|
|
|
static int K_XL;
|
|
static int M_XL[MAX_NUM_GPU];
|
|
static int N_XL;
|
|
|
|
void mat_mul(float *_A, float *_B, float *_C, int _M, int _N, int _K) {
|
|
|
|
// Launch kernel on every GPU
|
|
for (int i = 0; i < num_devices; i++) {
|
|
int m_size = Mend[i] - Mbegin[i];
|
|
dim3 blockDim(TS/WPT, TS, 1);
|
|
dim3 gridDim(((m_size + TS - 1) / TS), (N+TS-1)/TS, 1);
|
|
dim3 blockDimAll(TS, TS, 1);
|
|
dim3 gridDimA((M_XL[i] + TS - 1)/TS, (K_XL + TS - 1) / TS, 1);
|
|
dim3 gridDimB((K_XL + TS - 1) / TS, (N_XL + TS - 1) / TS, 1);
|
|
dim3 gridDimC((M_XL[i] + TS - 1) / TS, (N_XL + TS - 1) / TS, 1);
|
|
|
|
CUDA_CALL( cudaSetDevice(i) );
|
|
|
|
// add zero padding
|
|
paddingAddZeroes<<<gridDimA, blockDimAll>>>(m_size, K, a_d[i], M_XL[i], K_XL, a_ad[i]);
|
|
paddingAddZeroes<<<gridDimB, blockDimAll>>>(K, N, b_d[i], K_XL, N_XL, b_ad[i]);
|
|
|
|
// calculate matrix multiplication
|
|
sgemm<<<gridDim, blockDim>>>(a_ad[i], b_ad[i], c_ad[i], M_XL[i], N_XL, K_XL);
|
|
|
|
// remove zero padding
|
|
paddingRemoveZeroes<<<gridDimC, blockDimAll>>>(M_XL[i], N_XL, c_ad[i], m_size, N, c_d[i]);
|
|
}
|
|
|
|
// 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;
|
|
int slice = M / MAX_NUM_GPU, m_size;
|
|
K_XL = CEIL_DIV(K, TSK) * TSK;
|
|
N_XL = CEIL_DIV(N, TSN) * TSN;
|
|
|
|
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);
|
|
}
|
|
|
|
if (M < 4) {
|
|
num_devices = M;
|
|
}
|
|
|
|
slice = M / num_devices;
|
|
// Setup problem size for each GPU
|
|
for (int i = 0; i < num_devices; i++) {
|
|
Mbegin[i] = slice * i;
|
|
Mend[i] = (i == num_devices - 1) ? M : slice * (i + 1);
|
|
m_size = Mend[i] - Mbegin[i];
|
|
M_XL[i] = CEIL_DIV(m_size, TSM) * TSM;
|
|
}
|
|
|
|
// Allocate device memory for each GPU
|
|
for (int i = 0; i < num_devices; i++) {
|
|
m_size = Mend[i] - Mbegin[i];
|
|
CUDA_CALL( cudaSetDevice(i) );
|
|
//
|
|
CUDA_CALL( cudaMalloc(&a_d[i], m_size * K * sizeof(float)) );
|
|
CUDA_CALL( cudaMalloc(&b_d[i], K * N * sizeof(float)) );
|
|
CUDA_CALL( cudaMalloc(&c_d[i], m_size * N * sizeof(float)) );
|
|
// zero padding
|
|
CUDA_CALL( cudaMalloc(&a_ad[i], M_XL[i] * K_XL * sizeof(float)) );
|
|
CUDA_CALL( cudaMalloc(&b_ad[i], K_XL * N_XL * sizeof(float)) );
|
|
CUDA_CALL( cudaMalloc(&c_ad[i], M_XL[i] * N_XL * 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,
|
|
(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 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],
|
|
(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() );
|
|
}
|
|
}
|