floraachy
/
OpenBLAS
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

optimized sgemv_t_4.c for small size

This commit is contained in:
wernsaar 2014-08-31 13:23:44 +02:00
parent e2fc8c8c2c
commit 848c0f16f7
1 changed files with 108 additions and 42 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

150
kernel/x86_64/sgemv_t_4.c
View File

@ -64,6 +64,8 @@ static void sgemv_kernel_4x4(BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
#endif
static void sgemv_kernel_4x1(BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y) __attribute__ ((noinline));
static void sgemv_kernel_4x1(BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y)
{
BLASLONG i;
@ -71,11 +73,51 @@ static void sgemv_kernel_4x1(BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y)
a0 = ap;
FLOAT temp = 0.0;
if (n <=0 ) return;
/*
for ( i=0; i< n; i+=4 )
{
temp += a0[i]*x[i] + a0[i+1]*x[i+1] + a0[i+2]*x[i+2] + a0[i+3]*x[i+3];
}
*y = temp;
*/
i=0;
__asm__ __volatile__
(
"xorps %%xmm10 , %%xmm10 \n\t"
".align 16 \n\t"
".L01LOOP%=: \n\t"
"movups (%3,%0,4) , %%xmm12 \n\t"
"movups (%4,%0,4) , %%xmm11 \n\t"
"mulps %%xmm11 , %%xmm12 \n\t"
"addq $4 , %0 \n\t"
"addps %%xmm12 , %%xmm10 \n\t"
"subq $4 , %1 \n\t"
"jnz .L01LOOP%= \n\t"
"haddps %%xmm10, %%xmm10 \n\t"
"haddps %%xmm10, %%xmm10 \n\t"
"movss %%xmm10, (%2) \n\t"
:
:
"r" (i), // 0
"r" (n), // 1
"r" (y), // 2
"r" (ap), // 3
"r" (x) // 4
: "cc",
"%xmm10", "%xmm11", "%xmm12",
"memory"
);
}
static void copy_x(BLASLONG n, FLOAT *src, FLOAT *dest, BLASLONG inc_src)
@ -89,6 +131,57 @@ static void copy_x(BLASLONG n, FLOAT *src, FLOAT *dest, BLASLONG inc_src)
}
}
static void add_y(BLASLONG n, FLOAT da , FLOAT *src, FLOAT *dest, BLASLONG inc_dest) __attribute__ ((noinline));
static void add_y(BLASLONG n, FLOAT da , FLOAT *src, FLOAT *dest, BLASLONG inc_dest)
{
BLASLONG i;
if ( inc_dest != 1 )
{
for ( i=0; i<n; i++ )
{
*dest += src[i] * da;
dest += inc_dest;
}
return;
}
i=0;
__asm__ __volatile__
(
"movss (%2) , %%xmm10 \n\t"
"shufps $0 , %%xmm10 , %%xmm10 \n\t"
".align 16 \n\t"
".L01LOOP%=: \n\t"
"movups (%3,%0,4) , %%xmm12 \n\t"
"movups (%4,%0,4) , %%xmm11 \n\t"
"mulps %%xmm10 , %%xmm12 \n\t"
"addq $4 , %0 \n\t"
"addps %%xmm11 , %%xmm12 \n\t"
"subq $4 , %1 \n\t"
"movups %%xmm12, -16(%4,%0,4) \n\t"
"jnz .L01LOOP%= \n\t"
:
:
"r" (i), // 0
"r" (n), // 1
"r" (&da), // 2
"r" (src), // 3
"r" (dest) // 4
: "cc",
"%xmm10", "%xmm11", "%xmm12",
"memory"
);
}
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)
{
@ -103,11 +196,13 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha, FLOAT *a, BLASLO
BLASLONG m3;
BLASLONG n2;
FLOAT ybuffer[4],*xbuffer;
FLOAT *ytemp;
if ( m < 1 ) return(0);
if ( n < 1 ) return(0);
xbuffer = buffer;
ytemp = buffer + NBMAX;
n1 = n >> 2 ;
n2 = n & 3 ;
@ -140,65 +235,36 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha, FLOAT *a, BLASLO
FLOAT *ap[4];
FLOAT *yp;
BLASLONG register lda4 = 4 * lda;
ap[0] = a_ptr;
ap[1] = a_ptr + lda;
ap[2] = ap[1] + lda;
ap[3] = ap[2] + lda;
yp = ytemp;
for( i = 0; i < n1 ; i++)
{
sgemv_kernel_4x4(NB,ap,xbuffer,ybuffer);
sgemv_kernel_4x4(NB,ap,xbuffer,yp);
ap[0] += lda4 ;
ap[1] += lda4 ;
ap[2] += lda4 ;
ap[3] += lda4 ;
a_ptr += lda4 ;
if ( inc_y == 1 )
{
__asm__ __volatile__
(
"movss (%0) , %%xmm10 \n\t"
"shufps $0 , %%xmm10 , %%xmm10 \n\t"
"movups (%1) , %%xmm12 \n\t"
"movups (%2) , %%xmm11 \n\t"
"mulps %%xmm10 , %%xmm12 \n\t"
"addps %%xmm11 , %%xmm12 \n\t"
"movups %%xmm12, (%2) \n\t"
:
:
"r" (&alpha), // 0
"r" (ybuffer), // 1
"r" (y_ptr) // 2
:
"%xmm10", "%xmm11", "%xmm12",
"memory"
);
y_ptr += 4;
}
else
{
*y_ptr += ybuffer[0]*alpha;
y_ptr += inc_y;
*y_ptr += ybuffer[1]*alpha;
y_ptr += inc_y;
*y_ptr += ybuffer[2]*alpha;
y_ptr += inc_y;
*y_ptr += ybuffer[3]*alpha;
y_ptr += inc_y;
}
yp += 4;
}
if ( n1 > 0 )
{
add_y(n1*4, alpha, ytemp, y_ptr, inc_y );
y_ptr += n1 * inc_y * 4;
a_ptr += n1 * lda4 ;
}
for( i = 0; i < n2 ; i++)
{
sgemv_kernel_4x1(NB,a_ptr,xbuffer,ybuffer);
a_ptr += 1 * lda;
*y_ptr += ybuffer[0]*alpha;
a_ptr += lda;
*y_ptr += ybuffer[0] * alpha;
y_ptr += inc_y;
}