From 14303717dcddae43cdc55beb0685dae86f566fd8 Mon Sep 17 00:00:00 2001
From: Joseph Redmon <pjreddie@gmail.com>
Date: Sat, 25 Oct 2014 18:57:26 +0000
Subject: [PATCH] Fast, needs to be faster
---
src/gemm.c | 146 ++++++++++++++++++++++++++++++++++++++++++++++--
1 files changed, 138 insertions(+), 8 deletions(-)
diff --git a/src/gemm.c b/src/gemm.c
index fa78daf..2e53b31 100644
--- a/src/gemm.c
+++ b/src/gemm.c
@@ -1,5 +1,5 @@
#include "mini_blas.h"
-#include <clBLAS.h>
+#include "utils.h"
void gemm(int TA, int TB, int M, int N, int K, float ALPHA,
float *A, int lda,
@@ -104,6 +104,7 @@
#include "opencl.h"
#include <math.h>
+#include <clBLAS.h>
#define STR_HELPER(x) #x
#define STR(x) STR_HELPER(x)
@@ -111,7 +112,7 @@
#ifdef __APPLE__
#define BLOCK 1
#else
-#define BLOCK 8
+#define BLOCK 16
#endif
cl_kernel get_gemm_kernel()
@@ -125,6 +126,44 @@
return gemm_kernel;
}
+cl_kernel get_gemm_nt_kernel()
+{
+ static int init = 0;
+ static cl_kernel gemm_kernel;
+ if(!init){
+ gemm_kernel = get_kernel("src/gemm_new.cl", "gemm_nt", "-D BLOCK=" STR(BLOCK) );
+ init = 1;
+ }
+ return gemm_kernel;
+}
+
+cl_kernel get_gemm_tn_kernel()
+{
+ static int init = 0;
+ static cl_kernel gemm_kernel;
+ if(!init){
+ gemm_kernel = get_kernel("src/gemm_new.cl", "gemm_tn", "-D BLOCK=" STR(BLOCK) );
+ init = 1;
+ }
+ return gemm_kernel;
+}
+
+cl_kernel get_gemm_nn_kernel()
+{
+ static int init = 0;
+ static cl_kernel gemm_kernel;
+ if(!init){
+ gemm_kernel = get_kernel("src/gemm_new.cl", "gemm_nn", "-D BLOCK=" STR(BLOCK) );
+ init = 1;
+ }
+ return gemm_kernel;
+}
+
+void gemm_ongpu_new(int TA, int TB, int M, int N, int K, float ALPHA,
+ cl_mem A_gpu, int lda,
+ cl_mem B_gpu, int ldb,
+ float BETA,
+ cl_mem C_gpu, int ldc);
void gemm_ongpu_old(int TA, int TB, int M, int N, int K, float ALPHA,
cl_mem A_gpu, int lda,
cl_mem B_gpu, int ldb,
@@ -137,10 +176,51 @@
float BETA,
cl_mem C_gpu, int ldc)
{
+ /*
cl_setup();
- //cl.error = clblasSgemm(clblasRowMajor, TA?clblasTrans:clblasNoTrans, TB?clblasTrans:clblasNoTrans,M, N, K,ALPHA, A_gpu, 0, lda,B_gpu, 0, ldb,BETA, C_gpu, 0, ldc,1, &queue, 0, NULL, &event);
- //check_error(cl);
- gemm_ongpu_old(TA, TB, M, N, K, ALPHA, A_gpu, lda, B_gpu, ldb, BETA, C_gpu, ldc);
+ cl_command_queue queue = cl.queue;
+ cl_event event;
+ cl.error = clblasSgemm(clblasRowMajor, TA?clblasTrans:clblasNoTrans, TB?clblasTrans:clblasNoTrans,M, N, K,ALPHA, A_gpu, 0, lda,B_gpu, 0, ldb,BETA, C_gpu, 0, ldc,1, &queue, 0, NULL, &event);
+ */
+
+ gemm_ongpu_new(TA, TB, M, N, K, ALPHA, A_gpu, lda, B_gpu, ldb, BETA, C_gpu, ldc);
+}
+
+void gemm_ongpu_new(int TA, int TB, int M, int N, int K, float ALPHA,
+ cl_mem A_gpu, int lda,
+ cl_mem B_gpu, int ldb,
+ float BETA,
+ cl_mem C_gpu, int ldc)
+{
+ //printf("gpu: %d %d %d %d %d\n",TA, TB, M, N, K);
+ cl_setup();
+ cl_kernel gemm_kernel = get_gemm_kernel();
+ if(!TA && !TB) gemm_kernel = get_gemm_nn_kernel();
+ if(!TA && TB) gemm_kernel = get_gemm_nt_kernel();
+ if(TA && !TB) gemm_kernel = get_gemm_tn_kernel();
+ cl_command_queue queue = cl.queue;
+
+ cl_uint i = 0;
+ cl.error = clSetKernelArg(gemm_kernel, i++, sizeof(TA), (void*) &TA);
+ cl.error = clSetKernelArg(gemm_kernel, i++, sizeof(TB), (void*) &TB);
+ cl.error = clSetKernelArg(gemm_kernel, i++, sizeof(M), (void*) &M);
+ cl.error = clSetKernelArg(gemm_kernel, i++, sizeof(N), (void*) &N);
+ cl.error = clSetKernelArg(gemm_kernel, i++, sizeof(K), (void*) &K);
+ cl.error = clSetKernelArg(gemm_kernel, i++, sizeof(ALPHA), (void*) &ALPHA);
+ cl.error = clSetKernelArg(gemm_kernel, i++, sizeof(A_gpu), (void*) &A_gpu);
+ cl.error = clSetKernelArg(gemm_kernel, i++, sizeof(lda), (void*) &lda);
+ cl.error = clSetKernelArg(gemm_kernel, i++, sizeof(B_gpu), (void*) &B_gpu);
+ cl.error = clSetKernelArg(gemm_kernel, i++, sizeof(ldb), (void*) &ldb);
+ cl.error = clSetKernelArg(gemm_kernel, i++, sizeof(BETA), (void*) &BETA);
+ cl.error = clSetKernelArg(gemm_kernel, i++, sizeof(C_gpu), (void*) &C_gpu);
+ cl.error = clSetKernelArg(gemm_kernel, i++, sizeof(ldc), (void*) &ldc);
+ check_error(cl);
+
+ const size_t global_size[] = {ceil((float)N/BLOCK)*BLOCK, ceil((float)M/BLOCK)*BLOCK};
+ const size_t local_size[] = {BLOCK, BLOCK};
+
+ clEnqueueNDRangeKernel(queue, gemm_kernel, 2, 0, global_size, local_size, 0, 0, 0);
+ check_error(cl);
}
void gemm_ongpu_old(int TA, int TB, int M, int N, int K, float ALPHA,
@@ -170,7 +250,7 @@
cl.error = clSetKernelArg(gemm_kernel, i++, sizeof(ldc), (void*) &ldc);
check_error(cl);
- const size_t global_size[] = {ceil((float)M/BLOCK)*BLOCK, ceil((float)N/BLOCK)*BLOCK};
+ const size_t global_size[] = {ceil((float)N/BLOCK)*BLOCK, ceil((float)M/BLOCK)*BLOCK};
const size_t local_size[] = {BLOCK, BLOCK};
clEnqueueNDRangeKernel(queue, gemm_kernel, 2, 0, global_size, local_size, 0, 0, 0);
@@ -235,7 +315,7 @@
float *c = random_matrix(m,n);
int i;
clock_t start = clock(), end;
- for(i = 0; i<10; ++i){
+ for(i = 0; i<32; ++i){
gemm_gpu(TA,TB,m,n,k,1,a,lda,b,ldb,1,c,n);
}
end = clock();
@@ -245,6 +325,39 @@
free(c);
}
+void time_ongpu(int TA, int TB, int m, int k, int n)
+{
+ int iter = 100;
+ float *a = random_matrix(m,k);
+ float *b = random_matrix(k,n);
+
+ int lda = (!TA)?k:m;
+ int ldb = (!TB)?n:k;
+
+ float *c = random_matrix(m,n);
+
+ cl_mem a_cl = cl_make_array(a, m*k);
+ cl_mem b_cl = cl_make_array(b, k*n);
+ cl_mem c_cl = cl_make_array(c, m*n);
+
+ int i;
+ clock_t start = clock(), end;
+ for(i = 0; i<iter; ++i){
+ gemm_ongpu(TA,TB,m,n,k,1,a_cl,lda,b_cl,ldb,1,c_cl,n);
+ }
+ int flop = m*n*(2*k+3)*iter;
+ float gflop = flop/pow(10., 9);
+ end = clock();
+ float seconds = sec(end-start);
+ printf("Matrix Multiplication %dx%d * %dx%d, TA=%d, TB=%d: %lf s, %lf GFLOPS\n",m,k,k,n, TA, TB, seconds, gflop/seconds);
+ clReleaseMemObject(a_cl);
+ clReleaseMemObject(b_cl);
+ clReleaseMemObject(c_cl);
+ free(a);
+ free(b);
+ free(c);
+}
+
void test_gpu_accuracy(int TA, int TB, int m, int k, int n)
{
srand(0);
@@ -272,14 +385,16 @@
//printf("%f %f\n", c[i], c_gpu[i]);
sse += pow(c[i]-c_gpu[i], 2);
}
- printf("Matrix Multiplication %dx%d * %dx%d, TA=%d, TB=%d: %g MSE\n",m,k,k,n, TA, TB, sse/(m*n));
+ printf("Matrix Multiplication %dx%d * %dx%d, TA=%d, TB=%d: %g SSE\n",m,k,k,n, TA, TB, sse/(m*n));
free(a);
free(b);
free(c);
+ free(c_gpu);
}
void test_gpu_blas()
{
+ /*
test_gpu_accuracy(0,0,10,576,75);
test_gpu_accuracy(0,0,17,10,10);
@@ -291,6 +406,21 @@
test_gpu_accuracy(1,0,1000,10,100);
test_gpu_accuracy(0,1,1000,10,100);
test_gpu_accuracy(1,1,1000,10,100);
+ */
+ test_gpu_accuracy(0,0,131,4093,1199);
+ test_gpu_accuracy(0,1,131,4093,1199);
+ test_gpu_accuracy(1,0,131,4093,1199);
+ test_gpu_accuracy(1,1,131,4093,1199);
+
+ time_ongpu(0,0,1024,1024,1024);
+ time_ongpu(0,1,1024,1024,1024);
+ time_ongpu(1,0,1024,1024,1024);
+ time_ongpu(1,1,1024,1024,1024);
+
+ time_ongpu(0,0,128,4096,1200);
+ time_ongpu(0,1,128,4096,1200);
+ time_ongpu(1,0,128,4096,1200);
+ time_ongpu(1,1,128,4096,1200);
/*
time_gpu_random_matrix(0,0,1000,1000,100);
--
Gitblit v1.10.0