Skip to content

Commit c3cd6e7

Browse files
xianyixianyi
committed
Merge pull request #434 from wernsaar/develop
A lot of performance enhancements
2 parents 5c0d0ec + 11eab4c commit c3cd6e7

31 files changed

+3318
-45
lines changed

‎kernel/x86_64/KERNEL‎

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -377,7 +377,7 @@ SGEMVNKERNEL = sgemv_n.c
377377
endif
378378

379379
ifndef SGEMVTKERNEL
380-
SGEMVTKERNEL = ../arm/gemv_t.c
380+
SGEMVTKERNEL = sgemv_t.c
381381
endif
382382

383383
ifndef DGEMVNKERNEL

‎kernel/x86_64/KERNEL.BULLDOZER‎

Lines changed: 2 additions & 8 deletions
Original file line numberDiff line numberDiff line change
@@ -1,14 +1,8 @@
1-
ifdef OS_WINDOWS
21
SGEMVNKERNEL = sgemv_n.c
3-
SGEMVTKERNEL = ../arm/gemv_t.c
4-
else
5-
SGEMVNKERNEL = sgemv_n.c
6-
SGEMVTKERNEL = sgemv_t_avx.c
7-
endif
8-
2+
SGEMVTKERNEL = sgemv_t.c
93

104
ZGEMVNKERNEL = zgemv_n_dup.S
11-
ZGEMVTKERNEL = zgemv_t.S
5+
ZGEMVTKERNEL = zgemv_t.c
126

137
DGEMVNKERNEL = dgemv_n_bulldozer.S
148
DGEMVTKERNEL = dgemv_t_bulldozer.S

‎kernel/x86_64/KERNEL.HASWELL‎

Lines changed: 9 additions & 6 deletions
Original file line numberDiff line numberDiff line change
@@ -1,11 +1,14 @@
1-
ifdef OS_WINDOWS
21
SGEMVNKERNEL = sgemv_n.c
3-
SGEMVTKERNEL = ../arm/gemv_t.c
4-
else
5-
SGEMVNKERNEL = sgemv_n.c
6-
SGEMVTKERNEL = sgemv_t_avx.c
7-
endif
2+
SGEMVTKERNEL = sgemv_t.c
3+
4+
DGEMVNKERNEL = dgemv_n.c
5+
DGEMVTKERNEL = dgemv_t.c
6+
7+
ZGEMVNKERNEL = zgemv_n.c
8+
ZGEMVTKERNEL = zgemv_t.c
89

10+
CGEMVNKERNEL = cgemv_n.c
11+
CGEMVTKERNEL = cgemv_t.c
912

1013
SGEMMKERNEL = sgemm_kernel_16x4_haswell.S
1114
SGEMMINCOPY = ../generic/gemm_ncopy_16.c

‎kernel/x86_64/KERNEL.NEHALEM‎

Lines changed: 1 addition & 7 deletions
Original file line numberDiff line numberDiff line change
@@ -1,11 +1,5 @@
1-
ifdef OS_WINDOWS
21
SGEMVNKERNEL = sgemv_n.c
3-
SGEMVTKERNEL = ../arm/gemv_t.c
4-
else
5-
SGEMVNKERNEL = sgemv_n.c
6-
SGEMVTKERNEL = ../arm/gemv_t.c
7-
endif
8-
2+
SGEMVTKERNEL = sgemv_t.c
93

104
SGEMMKERNEL = gemm_kernel_4x8_nehalem.S
115
SGEMMINCOPY = gemm_ncopy_4.S

‎kernel/x86_64/KERNEL.PILEDRIVER‎

Lines changed: 1 addition & 7 deletions
Original file line numberDiff line numberDiff line change
@@ -1,11 +1,5 @@
1-
ifdef OS_WINDOWS
21
SGEMVNKERNEL = sgemv_n.c
3-
SGEMVTKERNEL = ../arm/gemv_t.c
4-
else
5-
SGEMVNKERNEL = sgemv_n.c
6-
SGEMVTKERNEL = sgemv_t_avx.c
7-
endif
8-
2+
SGEMVTKERNEL = sgemv_t.c
93

104
ZGEMVNKERNEL = zgemv_n_dup.S
115
ZGEMVTKERNEL = zgemv_t.S

‎kernel/x86_64/KERNEL.SANDYBRIDGE‎

Lines changed: 3 additions & 6 deletions
Original file line numberDiff line numberDiff line change
@@ -1,10 +1,7 @@
1-
ifdef OS_WINDOWS
21
SGEMVNKERNEL = sgemv_n.c
3-
SGEMVTKERNEL = ../arm/gemv_t.c
4-
else
5-
SGEMVNKERNEL = sgemv_n.c
6-
SGEMVTKERNEL = sgemv_t_avx.c
7-
endif
2+
SGEMVTKERNEL = sgemv_t.c
3+
4+
ZGEMVNKERNEL = zgemv_n.c
85

96

107
SGEMMKERNEL = sgemm_kernel_16x4_sandy.S

‎kernel/x86_64/cgemv_n.c‎

Lines changed: 255 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,255 @@
1+
/***************************************************************************
2+
Copyright (c) 2014, The OpenBLAS Project
3+
All rights reserved.
4+
Redistribution and use in source and binary forms, with or without
5+
modification, are permitted provided that the following conditions are
6+
met:
7+
1. Redistributions of source code must retain the above copyright
8+
notice, this list of conditions and the following disclaimer.
9+
2. Redistributions in binary form must reproduce the above copyright
10+
notice, this list of conditions and the following disclaimer in
11+
the documentation and/or other materials provided with the
12+
distribution.
13+
3. Neither the name of the OpenBLAS project nor the names of
14+
its contributors may be used to endorse or promote products
15+
derived from this software without specific prior written permission.
16+
THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
17+
AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
18+
IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
19+
ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
20+
LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
21+
DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
22+
SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
23+
CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
24+
OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
25+
USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
26+
*****************************************************************************/
27+
28+
#include <stdlib.h>
29+
#include <stdio.h>
30+
#include "common.h"
31+
32+
#if defined(HASWELL)
33+
#include "cgemv_n_microk_haswell-2.c"
34+
#endif
35+
36+
37+
#define NBMAX 2048
38+
39+
#ifndef HAVE_KERNEL_16x4
40+
41+
static void cgemv_kernel_16x4(BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
42+
{
43+
BLASLONG i;
44+
FLOAT *a0,*a1,*a2,*a3;
45+
a0 = ap[0];
46+
a1 = ap[1];
47+
a2 = ap[2];
48+
a3 = ap[3];
49+
50+
for ( i=0; i< 2*n; i+=2 )
51+
{
52+
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
53+
y[i] += a0[i]*x[0] - a0[i+1] * x[1];
54+
y[i+1] += a0[i]*x[1] + a0[i+1] * x[0];
55+
y[i] += a1[i]*x[2] - a1[i+1] * x[3];
56+
y[i+1] += a1[i]*x[3] + a1[i+1] * x[2];
57+
y[i] += a2[i]*x[4] - a2[i+1] * x[5];
58+
y[i+1] += a2[i]*x[5] + a2[i+1] * x[4];
59+
y[i] += a3[i]*x[6] - a3[i+1] * x[7];
60+
y[i+1] += a3[i]*x[7] + a3[i+1] * x[6];
61+
#else
62+
y[i] += a0[i]*x[0] + a0[i+1] * x[1];
63+
y[i+1] += a0[i]*x[1] - a0[i+1] * x[0];
64+
y[i] += a1[i]*x[2] + a1[i+1] * x[3];
65+
y[i+1] += a1[i]*x[3] - a1[i+1] * x[2];
66+
y[i] += a2[i]*x[4] + a2[i+1] * x[5];
67+
y[i+1] += a2[i]*x[5] - a2[i+1] * x[4];
68+
y[i] += a3[i]*x[6] + a3[i+1] * x[7];
69+
y[i+1] += a3[i]*x[7] - a3[i+1] * x[6];
70+
#endif
71+
}
72+
}
73+
74+
#endif
75+
76+
static void cgemv_kernel_16x1(BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y)
77+
{
78+
BLASLONG i;
79+
FLOAT *a0;
80+
a0 = ap;
81+
82+
for ( i=0; i< 2*n; i+=2 )
83+
{
84+
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
85+
y[i] += a0[i]*x[0] - a0[i+1] * x[1];
86+
y[i+1] += a0[i]*x[1] + a0[i+1] * x[0];
87+
#else
88+
y[i] += a0[i]*x[0] + a0[i+1] * x[1];
89+
y[i+1] += a0[i]*x[1] - a0[i+1] * x[0];
90+
#endif
91+
92+
}
93+
}
94+
95+
96+
static void zero_y(BLASLONG n, FLOAT *dest)
97+
{
98+
BLASLONG i;
99+
for ( i=0; i<2*n; i++ )
100+
{
101+
*dest = 0.0;
102+
dest++;
103+
}
104+
}
105+
106+
107+
108+
static void add_y(BLASLONG n, FLOAT *src, FLOAT *dest, BLASLONG inc_dest,FLOAT alpha_r, FLOAT alpha_i)
109+
{
110+
BLASLONG i;
111+
FLOAT temp_r;
112+
FLOAT temp_i;
113+
for ( i=0; i<n; i++ )
114+
{
115+
#if !defined(XCONJ)
116+
temp_r = alpha_r * src[0] - alpha_i * src[1];
117+
temp_i = alpha_r * src[1] + alpha_i * src[0];
118+
#else
119+
temp_r = alpha_r * src[0] + alpha_i * src[1];
120+
temp_i = -alpha_r * src[1] + alpha_i * src[0];
121+
#endif
122+
123+
*dest += temp_r;
124+
*(dest+1) += temp_i;
125+
126+
src+=2;
127+
dest += inc_dest;
128+
}
129+
}
130+
131+
int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha_r,FLOAT alpha_i, FLOAT *a, BLASLONG lda, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT *buffer)
132+
{
133+
BLASLONG i;
134+
BLASLONG j;
135+
FLOAT *a_ptr;
136+
FLOAT *x_ptr;
137+
FLOAT *y_ptr;
138+
FLOAT *ap[4];
139+
BLASLONG n1;
140+
BLASLONG m1;
141+
BLASLONG m2;
142+
BLASLONG n2;
143+
FLOAT xbuffer[8],*ybuffer;
144+
145+
146+
#if 0
147+
printf("%s %d %d %.16f %.16f %d %d %d\n","zgemv_n",m,n,alpha_r,alpha_i,lda,inc_x,inc_y);
148+
#endif
149+
150+
if ( m < 1 ) return(0);
151+
if ( n < 1 ) return(0);
152+
153+
ybuffer = buffer;
154+
155+
inc_x *= 2;
156+
inc_y *= 2;
157+
lda *= 2;
158+
159+
n1 = n / 4 ;
160+
n2 = n % 4 ;
161+
162+
m1 = m - ( m % 16 );
163+
m2 = (m % NBMAX) - (m % 16) ;
164+
165+
y_ptr = y;
166+
167+
BLASLONG NB = NBMAX;
168+
169+
while ( NB == NBMAX )
170+
{
171+
172+
m1 -= NB;
173+
if ( m1 < 0)
174+
{
175+
if ( m2 == 0 ) break;
176+
NB = m2;
177+
}
178+
179+
a_ptr = a;
180+
x_ptr = x;
181+
zero_y(NB,ybuffer);
182+
for( i = 0; i < n1 ; i++)
183+
{
184+
185+
xbuffer[0] = x_ptr[0];
186+
xbuffer[1] = x_ptr[1];
187+
x_ptr += inc_x;
188+
xbuffer[2] = x_ptr[0];
189+
xbuffer[3] = x_ptr[1];
190+
x_ptr += inc_x;
191+
xbuffer[4] = x_ptr[0];
192+
xbuffer[5] = x_ptr[1];
193+
x_ptr += inc_x;
194+
xbuffer[6] = x_ptr[0];
195+
xbuffer[7] = x_ptr[1];
196+
x_ptr += inc_x;
197+
198+
ap[0] = a_ptr;
199+
ap[1] = a_ptr + lda;
200+
ap[2] = ap[1] + lda;
201+
ap[3] = ap[2] + lda;
202+
cgemv_kernel_16x4(NB,ap,xbuffer,ybuffer);
203+
a_ptr += 4 * lda;
204+
}
205+
206+
for( i = 0; i < n2 ; i++)
207+
{
208+
xbuffer[0] = x_ptr[0];
209+
xbuffer[1] = x_ptr[1];
210+
x_ptr += inc_x;
211+
cgemv_kernel_16x1(NB,a_ptr,xbuffer,ybuffer);
212+
a_ptr += 1 * lda;
213+
214+
}
215+
add_y(NB,ybuffer,y_ptr,inc_y,alpha_r,alpha_i);
216+
a += 2 * NB;
217+
y_ptr += NB * inc_y;
218+
}
219+
220+
j=0;
221+
while ( j < (m % 16))
222+
{
223+
a_ptr = a;
224+
x_ptr = x;
225+
FLOAT temp_r = 0.0;
226+
FLOAT temp_i = 0.0;
227+
for( i = 0; i < n; i++ )
228+
{
229+
#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
230+
temp_r += a_ptr[0] * x_ptr[0] - a_ptr[1] * x_ptr[1];
231+
temp_i += a_ptr[0] * x_ptr[1] + a_ptr[1] * x_ptr[0];
232+
#else
233+
temp_r += a_ptr[0] * x_ptr[0] + a_ptr[1] * x_ptr[1];
234+
temp_i += a_ptr[0] * x_ptr[1] - a_ptr[1] * x_ptr[0];
235+
#endif
236+
237+
a_ptr += lda;
238+
x_ptr += inc_x;
239+
}
240+
241+
#if !defined(XCONJ)
242+
y_ptr[0] += alpha_r * temp_r - alpha_i * temp_i;
243+
y_ptr[1] += alpha_r * temp_i + alpha_i * temp_r;
244+
#else
245+
y_ptr[0] += alpha_r * temp_r + alpha_i * temp_i;
246+
y_ptr[1] -= alpha_r * temp_i - alpha_i * temp_r;
247+
#endif
248+
y_ptr += inc_y;
249+
a+=2;
250+
j++;
251+
}
252+
return(0);
253+
}
254+
255+

0 commit comments

Comments
 (0)