cozis/Lina
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

numpy performance reached on reasonable matrix size

Browse Source
This commit is contained in:
rdgarce
2023-08-18 13:09:43 +02:00
parent 4b3127d84d
commit 44e68aa0a9
6 changed files with 360 additions and 34 deletions
Download Patch File Download Diff File Expand all files Collapse all files
benchmarks/bench_dot.c
+129 -32
View File
@@ -3,38 +3,43 @@
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <assert.h>
#include "../src/lina.h"
#define A_ROWS 1000llu
#define A_COLS 146llu
#define A_ROWS 960llu
#define A_COLS 960llu
#define B_ROWS 146llu
#define B_COLS 1024llu
#define B_ROWS 960llu
#define B_COLS 960llu
uint64_t nanos();
int saveMatrixToStream(FILE *fp, double *A, int width, int height, char **error);
static uint64_t nanos();
int main()
{
uint64_t ops = A_ROWS*B_COLS*2*A_COLS;
uint64_t start,stop,lina_dot_time, lina_dot_mod1_time, lina_dot_mod1_1_time, lina_dot_mod2_time;
double *A = (double *)malloc(sizeof(double)*A_ROWS*A_COLS);
double *B = (double *)malloc(sizeof(double)*B_ROWS*B_COLS);
uint64_t start,stop,lina_dot_time, lina_dot1_time, lina_dot2_time, lina_dot3_time, lina_dot4_time;
double *A = (double *)aligned_alloc(32,sizeof(double)*A_ROWS*A_COLS);
double *B = (double *)aligned_alloc(32,sizeof(double)*B_ROWS*B_COLS);
double *C1 = (double *)malloc(sizeof(double)*A_ROWS*B_COLS);
double *C2 = (double *)malloc(sizeof(double)*A_ROWS*B_COLS);
double *C3 = (double *)malloc(sizeof(double)*A_ROWS*B_COLS);
double *C4 = (double *)malloc(sizeof(double)*A_ROWS*B_COLS);
double *C1 = (double *)aligned_alloc(32,sizeof(double)*A_ROWS*B_COLS);
double *C2 = (double *)aligned_alloc(32,sizeof(double)*A_ROWS*B_COLS);
double *C3 = (double *)aligned_alloc(32,sizeof(double)*A_ROWS*B_COLS);
double *C4 = (double *)aligned_alloc(32,sizeof(double)*A_ROWS*B_COLS);
double *C5 = (double *)aligned_alloc(32,sizeof(double)*A_ROWS*B_COLS);
for (int i = 0; i < A_ROWS*A_COLS; i++)
A[i] = (double)(rand()%10);
A[i] = (double)(rand()%2);
for (int i = 0; i < B_ROWS*B_COLS; i++)
B[i] = (double)(rand()%10);
B[i] = (double)(rand()%2);
for (int i = 0; i < A_ROWS*B_COLS; i++)
{
C1[i] = (double)(rand()%10);
C2[i] = (double)(rand()%10);
C3[i] = (double)(rand()%10);
C4[i] = (double)(rand()%10);
C1[i] = (double)(rand()%2);
C2[i] = (double)(rand()%2);
C3[i] = (double)(rand()%2);
C4[i] = (double)(rand()%2);
C5[i] = (double)(rand()%2);
}
start = nanos();
@@ -44,36 +49,78 @@ int main()
lina_dot_time = stop-start;
start = nanos();
lina_dot_mod1(A,B,C2,A_ROWS,A_COLS,B_COLS);
lina_dot1(A,B,C2,A_ROWS,A_COLS,B_COLS);
stop = nanos();
lina_dot_mod1_time = stop-start;
lina_dot1_time = stop-start;
start = nanos();
lina_dot_mod2(A,B,C3,A_ROWS,A_COLS,B_COLS);
lina_dot2(A,B,C3,A_ROWS,A_COLS,B_COLS);
stop = nanos();
lina_dot_mod2_time = stop-start;
lina_dot2_time = stop-start;
start = nanos();
lina_dot_mod2_old(A,B,C4,A_ROWS,A_COLS,B_COLS);
lina_dot3(A,B,C4,A_ROWS,A_COLS,B_COLS);
stop = nanos();
lina_dot_mod1_1_time = stop-start;
lina_dot3_time = stop-start;
if(!memcmp(C1,C2,sizeof(double)*A_ROWS*B_COLS) && !memcmp(C2,C3,sizeof(double)*A_ROWS*B_COLS) && !memcmp(C3,C4,sizeof(double)*A_ROWS*B_COLS))
start = nanos();
lina_dot4(A,B,C5,A_ROWS,A_COLS,B_COLS);
stop = nanos();
lina_dot4_time = stop-start;
if(!memcmp(C1,C2,sizeof(double)*A_ROWS*B_COLS)
&& !memcmp(C2,C3,sizeof(double)*A_ROWS*B_COLS)
&& !memcmp(C3,C4,sizeof(double)*A_ROWS*B_COLS)
&& !memcmp(C4,C5,sizeof(double)*A_ROWS*B_COLS))
{
printf( "lina_dot : %f GFLOPS\n"
"lina_dot_mod1: %f GFLOPS\n"
"lina_dot_mod2: %f GFLOPS\n"
"lina_dot_mod2_old: %f GFLOPS\n", (double)ops/lina_dot_time,
(double)ops/lina_dot_mod1_time,
(double)ops/lina_dot_mod2_time,
(double)ops/lina_dot_mod1_1_time);
printf( "lina_dot : %f GFLOPS\n"
"lina_dot1: %f GFLOPS\n"
"lina_dot2: %f GFLOPS\n"
"lina_dot3: %f GFLOPS\n"
"lina_dot4: %f GFLOPS\n", (double)ops/lina_dot_time,
(double)ops/lina_dot1_time,
(double)ops/lina_dot2_time,
(double)ops/lina_dot3_time,
(double)ops/lina_dot4_time);
FILE *fp = fopen("lina_dots_success.txt", "w");
if (!fp)
return -1;
saveMatrixToStream(fp,C1,A_ROWS,A_COLS,NULL);
fprintf(fp,"\nFINE\n");
saveMatrixToStream(fp,C2,A_ROWS,A_COLS,NULL);
fprintf(fp,"\nFINE\n");
saveMatrixToStream(fp,C3,A_ROWS,A_COLS,NULL);
fprintf(fp,"\nFINE\n");
saveMatrixToStream(fp,C4,A_ROWS,A_COLS,NULL);
fprintf(fp,"\nFINE\n");
saveMatrixToStream(fp,C5,A_ROWS,A_COLS,NULL);
fclose(fp);
}
else
{
printf("ERRORE: i prodotti matriciali sono diversi!\n");
FILE *fp = fopen("lina_dots_error.txt", "w");
if (!fp)
return -1;
saveMatrixToStream(fp,C1,A_ROWS,A_COLS,NULL);
fprintf(fp,"\nFINE\n");
saveMatrixToStream(fp,C2,A_ROWS,A_COLS,NULL);
fprintf(fp,"\nFINE\n");
saveMatrixToStream(fp,C3,A_ROWS,A_COLS,NULL);
fprintf(fp,"\nFINE\n");
saveMatrixToStream(fp,C4,A_ROWS,A_COLS,NULL);
fprintf(fp,"\nFINE\n");
saveMatrixToStream(fp,C5,A_ROWS,A_COLS,NULL);
fclose(fp);
}
free(A);
free(B);
@@ -81,4 +128,54 @@ int main()
free(C2);
free(C3);
free(C4);
free(C5);
}
static uint64_t nanos()
{
struct timespec time;
clock_gettime(CLOCK_MONOTONIC, &time);
return (uint64_t)time.tv_sec*1000000000 + (uint64_t)time.tv_nsec;
}
int saveMatrixToStream(FILE *fp, double *A, int width, int height, char **error)
{
assert(A != NULL);
char *dummy;
if (error == NULL)
error = &dummy;
else
*error = NULL;
if (width < 1) {
*error = "The provided width is less than one";
return -1;
}
if (height < 1) {
*error = "The provided height is less than one";
return -1;
}
if (fp == NULL)
fp = stdout;
putc('[',fp);
for (int i = 0; i < height-1; i++) {
for (int j = 0; j < width-1; j++)
fprintf(fp, "%.1f ", A[i*width + j]);
fprintf(fp, "%.1f,\n", A[i*width + width-1]);
}
for (int j = 0; j < width-1; j++)
fprintf(fp, "%.1f ", A[(height-1)*width + j]);
fprintf(fp, "%.1f", A[(height-1)*width + width-1]);
putc(']',fp);
return 0;
}
benchmarks/benchmarks.sh
+3
View File
@@ -0,0 +1,3 @@
gcc bench_dot.c ../src/lina.c -O3 -march=native -ffast-math -funroll-loops -o bench_dot
./bench_dot
python3 py_dot.py
benchmarks/py_dot.py
+22
View File
@@ -0,0 +1,22 @@
import os
os.environ['OMP_NUM_THREADS'] = '1'
import numpy as np
import time
N = 1024
if __name__ == "__main__":
A = np.random.randn(N,N).astype(np.float64)
B = np.random.randn(N,N).astype(np.float64)
start = time.monotonic()
C = A @ B
stop = time.monotonic()
s = stop-start
ops = 2*N*N*N
print(f"NUMPY: {ops/s * 1e-9} GFLOPS\n")