1 /***************************************************************************
2 Copyright (c) 2017, The OpenBLAS Project
4 Redistribution and use in source and binary forms, with or without
5 modification, are permitted provided that the following conditions are
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
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 *****************************************************************************/
33 #define HAVE_KERNEL_4x4_VEC 1
34 #define HAVE_KERNEL_4x2_VEC 1
35 #define HAVE_KERNEL_4x1_VEC 1
37 #if defined(HAVE_KERNEL_4x4_VEC) || defined(HAVE_KERNEL_4x2_VEC) || defined(HAVE_KERNEL_4x1_VEC)
38 #include <vecintrin.h>
41 #ifdef HAVE_KERNEL_4x4
43 #elif HAVE_KERNEL_4x4_VEC
45 static void dgemv_kernel_4x4(BLASLONG n, FLOAT **ap, FLOAT *xo, FLOAT *y, FLOAT *alpha)
53 __vector double v_x0 = {x0,x0};
54 __vector double v_x1 = {x1,x1};
55 __vector double v_x2 = {x2,x2};
56 __vector double v_x3 = {x3,x3};
57 __vector double* v_y =(__vector double*)y;
58 __vector double* va0 = (__vector double*)ap[0];
59 __vector double* va1 = (__vector double*)ap[1];
60 __vector double* va2 = (__vector double*)ap[2];
61 __vector double* va3 = (__vector double*)ap[3];
63 for ( i=0; i< n/2; i+=2 )
65 v_y[i] += v_x0 * va0[i] + v_x1 * va1[i] + v_x2 * va2[i] + v_x3 * va3[i] ;
66 v_y[i+1] += v_x0 * va0[i+1] + v_x1 * va1[i+1] + v_x2 * va2[i+1] + v_x3 * va3[i+1] ;
72 static void dgemv_kernel_4x4(BLASLONG n, FLOAT **ap, FLOAT *xo, FLOAT *y, FLOAT *alpha)
75 FLOAT *a0,*a1,*a2,*a3;
76 FLOAT x[4] __attribute__ ((aligned (16)));
83 x[i] = xo[i] * *alpha;
85 for ( i=0; i< n; i+=4 )
87 y[i] += a0[i]*x[0] + a1[i]*x[1] + a2[i]*x[2] + a3[i]*x[3];
88 y[i+1] += a0[i+1]*x[0] + a1[i+1]*x[1] + a2[i+1]*x[2] + a3[i+1]*x[3];
89 y[i+2] += a0[i+2]*x[0] + a1[i+2]*x[1] + a2[i+2]*x[2] + a3[i+2]*x[3];
90 y[i+3] += a0[i+3]*x[0] + a1[i+3]*x[1] + a2[i+3]*x[2] + a3[i+3]*x[3];
97 #ifdef HAVE_KERNEL_4x2
99 #elif HAVE_KERNEL_4x2_VEC
101 static void dgemv_kernel_4x2(BLASLONG n, FLOAT **ap, FLOAT *xo, FLOAT *y, FLOAT *alpha)
107 __vector double v_x0 = {x0,x0};
108 __vector double v_x1 = {x1,x1};
109 __vector double* v_y =(__vector double*)y;
110 __vector double* va0 = (__vector double*)ap[0];
111 __vector double* va1 = (__vector double*)ap[1];
113 for ( i=0; i< n/2; i+=2 )
115 v_y[i] += v_x0 * va0[i] + v_x1 * va1[i] ;
116 v_y[i+1] += v_x0 * va0[i+1] + v_x1 * va1[i+1] ;
121 static void dgemv_kernel_4x2(BLASLONG n, FLOAT **ap, FLOAT *xo, FLOAT *y, FLOAT *alpha)
125 FLOAT x[4] __attribute__ ((aligned (16)));
130 x[i] = xo[i] * *alpha;
132 for ( i=0; i< n; i+=4 )
134 y[i] += a0[i]*x[0] + a1[i]*x[1];
135 y[i+1] += a0[i+1]*x[0] + a1[i+1]*x[1];
136 y[i+2] += a0[i+2]*x[0] + a1[i+2]*x[1];
137 y[i+3] += a0[i+3]*x[0] + a1[i+3]*x[1];
144 #ifdef HAVE_KERNEL_4x1
146 #elif HAVE_KERNEL_4x1_VEC
147 static void dgemv_kernel_4x1(BLASLONG n, FLOAT *ap, FLOAT *xo, FLOAT *y, FLOAT *alpha)
153 __vector double v_x0 = {x0,x0};
154 __vector double* v_y =(__vector double*)y;
155 __vector double* va0 = (__vector double*)ap;
157 for ( i=0; i< n/2; i+=2 )
159 v_y[i] += v_x0 * va0[i] ;
160 v_y[i+1] += v_x0 * va0[i+1] ;
167 static void dgemv_kernel_4x1(BLASLONG n, FLOAT *ap, FLOAT *xo, FLOAT *y, FLOAT *alpha)
171 FLOAT x[4] __attribute__ ((aligned (16)));
175 x[i] = xo[i] * *alpha;
177 for ( i=0; i< n; i+=4 )
180 y[i+1] += a0[i+1]*x[0];
181 y[i+2] += a0[i+2]*x[0];
182 y[i+3] += a0[i+3]*x[0];
191 static void add_y(BLASLONG n, FLOAT *src, FLOAT *dest, BLASLONG inc_dest) __attribute__ ((noinline));
193 static void add_y(BLASLONG n, FLOAT *src, FLOAT *dest, BLASLONG inc_dest)
197 for ( i=0; i<n; i++ ){
206 int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha, FLOAT *a, BLASLONG lda, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT *buffer)
219 BLASLONG lda4 = lda << 2;
220 FLOAT xbuffer[8],*ybuffer;
222 if ( m < 1 ) return(0);
223 if ( n < 1 ) return(0);
232 m2 = (m & (NBMAX-1)) - m3 ;
238 while ( NB == NBMAX )
244 if ( m2 == 0 ) break;
257 memset(ybuffer,0,NB*8);
265 for( i = 0; i < n1 ; i++)
267 dgemv_kernel_4x4(NB,ap,x_ptr,ybuffer,&alpha);
278 dgemv_kernel_4x2(NB,ap,x_ptr,ybuffer,&alpha);
286 dgemv_kernel_4x1(NB,a_ptr,x_ptr,ybuffer,&alpha);
297 for( i = 0; i < n1 ; i++)
299 xbuffer[0] = x_ptr[0];
301 xbuffer[1] = x_ptr[0];
303 xbuffer[2] = x_ptr[0];
305 xbuffer[3] = x_ptr[0];
307 dgemv_kernel_4x4(NB,ap,xbuffer,ybuffer,&alpha);
315 for( i = 0; i < n2 ; i++)
317 xbuffer[0] = x_ptr[0];
319 dgemv_kernel_4x1(NB,a_ptr,xbuffer,ybuffer,&alpha);
329 add_y(NB,ybuffer,y_ptr,inc_y);
337 if ( m3 == 0 ) return(0);
346 if ( lda == 3 && inc_x ==1 )
349 for( i = 0; i < ( n & -4 ); i+=4 )
352 temp0 += a_ptr[0] * x_ptr[0] + a_ptr[3] * x_ptr[1];
353 temp1 += a_ptr[1] * x_ptr[0] + a_ptr[4] * x_ptr[1];
354 temp2 += a_ptr[2] * x_ptr[0] + a_ptr[5] * x_ptr[1];
356 temp0 += a_ptr[6] * x_ptr[2] + a_ptr[9] * x_ptr[3];
357 temp1 += a_ptr[7] * x_ptr[2] + a_ptr[10] * x_ptr[3];
358 temp2 += a_ptr[8] * x_ptr[2] + a_ptr[11] * x_ptr[3];
366 temp0 += a_ptr[0] * x_ptr[0];
367 temp1 += a_ptr[1] * x_ptr[0];
368 temp2 += a_ptr[2] * x_ptr[0];
377 for( i = 0; i < n; i++ )
379 temp0 += a_ptr[0] * x_ptr[0];
380 temp1 += a_ptr[1] * x_ptr[0];
381 temp2 += a_ptr[2] * x_ptr[0];
389 y_ptr[0] += alpha * temp0;
391 y_ptr[0] += alpha * temp1;
393 y_ptr[0] += alpha * temp2;
404 if ( lda == 2 && inc_x ==1 )
407 for( i = 0; i < (n & -4) ; i+=4 )
409 temp0 += a_ptr[0] * x_ptr[0] + a_ptr[2] * x_ptr[1];
410 temp1 += a_ptr[1] * x_ptr[0] + a_ptr[3] * x_ptr[1];
411 temp0 += a_ptr[4] * x_ptr[2] + a_ptr[6] * x_ptr[3];
412 temp1 += a_ptr[5] * x_ptr[2] + a_ptr[7] * x_ptr[3];
421 temp0 += a_ptr[0] * x_ptr[0];
422 temp1 += a_ptr[1] * x_ptr[0];
431 for( i = 0; i < n; i++ )
433 temp0 += a_ptr[0] * x_ptr[0];
434 temp1 += a_ptr[1] * x_ptr[0];
442 y_ptr[0] += alpha * temp0;
444 y_ptr[0] += alpha * temp1;
453 if ( lda == 1 && inc_x ==1 )
456 for( i = 0; i < (n & -4); i+=4 )
458 temp += a_ptr[i] * x_ptr[i] + a_ptr[i+1] * x_ptr[i+1] + a_ptr[i+2] * x_ptr[i+2] + a_ptr[i+3] * x_ptr[i+3];
464 temp += a_ptr[i] * x_ptr[i];
471 for( i = 0; i < n; i++ )
473 temp += a_ptr[0] * x_ptr[0];
479 y_ptr[0] += alpha * temp;