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

optimized saxpy and daxpy for sandybridge

This commit is contained in:
Werner Saar 2015-04-28 10:18:32 +02:00
parent e42652f772
commit e216f686cb
4 changed files with 107 additions and 77 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
kernel/x86_64/daxpy.c
View File

@ -79,11 +79,7 @@ int CNAME(BLASLONG n, BLASLONG dummy0, BLASLONG dummy1, FLOAT da, FLOAT *x, BLAS
if ( (inc_x == 1) && (inc_y == 1) ) if ( (inc_x == 1) && (inc_y == 1) )
{ {
#if defined(SANDYBRIDGE)
BLASLONG n1 = n & -32;
#else
BLASLONG n1 = n & -16; BLASLONG n1 = n & -16;
#endif
if ( n1 ) if ( n1 )
daxpy_kernel_8(n1, x, y , &da ); daxpy_kernel_8(n1, x, y , &da );

87
kernel/x86_64/daxpy_microk_sandy-2.c
View File

@ -37,48 +37,67 @@ static void daxpy_kernel_8( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha)
__asm__ __volatile__ __asm__ __volatile__
( (
"vbroadcastsd (%4), %%ymm0 \n\t" // alpha "vbroadcastsd (%4), %%ymm0 \n\t" // alpha
".align 16 \n\t"
"1: \n\t"
"vmovups (%3,%0,8), %%ymm8 \n\t" "vmovups (%3,%0,8), %%ymm8 \n\t"
"vmovups 32(%3,%0,8), %%ymm9 \n\t" "vmovups 32(%3,%0,8), %%ymm9 \n\t"
"vmovups 64(%3,%0,8), %%ymm10 \n\t" "vmovups 64(%3,%0,8), %%ymm10 \n\t"
"vmovups 96(%3,%0,8), %%ymm11 \n\t" "vmovups 96(%3,%0,8), %%ymm11 \n\t"
"vmovups 128(%3,%0,8), %%ymm12 \n\t" "vmovups (%2,%0,8), %%ymm4 \n\t"
"vmovups 160(%3,%0,8), %%ymm13 \n\t" "vmovups 32(%2,%0,8), %%ymm5 \n\t"
"vmovups 192(%3,%0,8), %%ymm14 \n\t" "vmovups 64(%2,%0,8), %%ymm6 \n\t"
"vmovups 224(%3,%0,8), %%ymm15 \n\t" "vmovups 96(%2,%0,8), %%ymm7 \n\t"
"vmulpd (%2,%0,8), %%ymm0, %%ymm1 \n\t" "addq $16, %0 \n\t"
"vmulpd 32(%2,%0,8), %%ymm0, %%ymm2 \n\t" "subq $16, %1 \n\t"
"vaddpd %%ymm8 , %%ymm1, %%ymm8 \n\t" "jz 2f \n\t"
"vmulpd 64(%2,%0,8), %%ymm0, %%ymm3 \n\t"
"vaddpd %%ymm9 , %%ymm2, %%ymm9 \n\t"
"vmulpd 96(%2,%0,8), %%ymm0, %%ymm4 \n\t"
"vaddpd %%ymm10, %%ymm3, %%ymm10 \n\t"
"vmulpd 128(%2,%0,8), %%ymm0, %%ymm5 \n\t"
"vaddpd %%ymm11, %%ymm4, %%ymm11 \n\t"
"vmulpd 160(%2,%0,8), %%ymm0, %%ymm6 \n\t"
"vaddpd %%ymm12, %%ymm5, %%ymm12 \n\t"
"vmulpd 192(%2,%0,8), %%ymm0, %%ymm7 \n\t"
"vmulpd 224(%2,%0,8), %%ymm0, %%ymm1 \n\t"
"vaddpd %%ymm13, %%ymm6, %%ymm13 \n\t" ".align 16 \n\t"
"vmovups %%ymm8 , (%3,%0,8) \n\t" "1: \n\t"
"vaddpd %%ymm14, %%ymm7, %%ymm14 \n\t"
"vmovups %%ymm9 , 32(%3,%0,8) \n\t"
"vaddpd %%ymm15, %%ymm1, %%ymm15 \n\t"
"vmovups %%ymm10, 64(%3,%0,8) \n\t"
"vmovups %%ymm11, 96(%3,%0,8) \n\t"
"vmovups %%ymm12,128(%3,%0,8) \n\t"
"vmovups %%ymm13,160(%3,%0,8) \n\t"
"vmovups %%ymm14,192(%3,%0,8) \n\t"
"vmovups %%ymm15,224(%3,%0,8) \n\t"
"addq $32, %0 \n\t" "vmulpd %%ymm4, %%ymm0, %%ymm4 \n\t"
"subq $32, %1 \n\t" "vaddpd %%ymm8 , %%ymm4, %%ymm12 \n\t"
"vmulpd %%ymm5, %%ymm0, %%ymm5 \n\t"
"vaddpd %%ymm9 , %%ymm5, %%ymm13 \n\t"
"vmulpd %%ymm6, %%ymm0, %%ymm6 \n\t"
"vaddpd %%ymm10, %%ymm6, %%ymm14 \n\t"
"vmulpd %%ymm7, %%ymm0, %%ymm7 \n\t"
"vaddpd %%ymm11, %%ymm7, %%ymm15 \n\t"
"vmovups (%3,%0,8), %%ymm8 \n\t"
"vmovups 32(%3,%0,8), %%ymm9 \n\t"
"vmovups 64(%3,%0,8), %%ymm10 \n\t"
"vmovups 96(%3,%0,8), %%ymm11 \n\t"
"vmovups (%2,%0,8), %%ymm4 \n\t"
"vmovups 32(%2,%0,8), %%ymm5 \n\t"
"vmovups 64(%2,%0,8), %%ymm6 \n\t"
"vmovups 96(%2,%0,8), %%ymm7 \n\t"
"vmovups %%ymm12, -128(%3,%0,8) \n\t"
"vmovups %%ymm13, -96(%3,%0,8) \n\t"
"vmovups %%ymm14, -64(%3,%0,8) \n\t"
"vmovups %%ymm15, -32(%3,%0,8) \n\t"
"addq $16, %0 \n\t"
"subq $16, %1 \n\t"
"jnz 1b \n\t" "jnz 1b \n\t"
"vzeroupper \n\t"
"2: \n\t"
"vmulpd %%ymm4, %%ymm0, %%ymm4 \n\t"
"vmulpd %%ymm5, %%ymm0, %%ymm5 \n\t"
"vmulpd %%ymm6, %%ymm0, %%ymm6 \n\t"
"vmulpd %%ymm7, %%ymm0, %%ymm7 \n\t"
"vaddpd %%ymm8 , %%ymm4, %%ymm12 \n\t"
"vaddpd %%ymm9 , %%ymm5, %%ymm13 \n\t"
"vaddpd %%ymm10, %%ymm6, %%ymm14 \n\t"
"vaddpd %%ymm11, %%ymm7, %%ymm15 \n\t"
"vmovups %%ymm12, -128(%3,%0,8) \n\t"
"vmovups %%ymm13, -96(%3,%0,8) \n\t"
"vmovups %%ymm14, -64(%3,%0,8) \n\t"
"vmovups %%ymm15, -32(%3,%0,8) \n\t"
"vzeroupper \n\t"
: :
: :

4
kernel/x86_64/saxpy.c
View File

@ -75,11 +75,7 @@ int CNAME(BLASLONG n, BLASLONG dummy0, BLASLONG dummy1, FLOAT da, FLOAT *x, BLAS
if ( (inc_x == 1) && (inc_y == 1) ) if ( (inc_x == 1) && (inc_y == 1) )
{ {
#if defined(SANDYBRIDGE)
BLASLONG n1 = n & -64;
#else
BLASLONG n1 = n & -32; BLASLONG n1 = n & -32;
#endif
if ( n1 ) if ( n1 )
saxpy_kernel_16(n1, x, y , &da ); saxpy_kernel_16(n1, x, y , &da );

89
kernel/x86_64/saxpy_microk_sandy-2.c
View File

@ -37,48 +37,67 @@ static void saxpy_kernel_16( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha)
__asm__ __volatile__ __asm__ __volatile__
( (
"vbroadcastss (%4), %%ymm0 \n\t" // alpha "vbroadcastss (%4), %%ymm0 \n\t" // alpha
".align 16 \n\t"
"1: \n\t"
"vmovups (%3,%0,4), %%ymm8 \n\t" "vmovups (%3,%0,4), %%ymm8 \n\t"
"vmovups 32(%3,%0,4), %%ymm9 \n\t" "vmovups 32(%3,%0,4), %%ymm9 \n\t"
"vmovups 64(%3,%0,4), %%ymm10 \n\t" "vmovups 64(%3,%0,4), %%ymm10 \n\t"
"vmovups 96(%3,%0,4), %%ymm11 \n\t" "vmovups 96(%3,%0,4), %%ymm11 \n\t"
"vmovups 128(%3,%0,4), %%ymm12 \n\t" "vmovups (%2,%0,4), %%ymm4 \n\t"
"vmovups 160(%3,%0,4), %%ymm13 \n\t" "vmovups 32(%2,%0,4), %%ymm5 \n\t"
"vmovups 192(%3,%0,4), %%ymm14 \n\t" "vmovups 64(%2,%0,4), %%ymm6 \n\t"
"vmovups 224(%3,%0,4), %%ymm15 \n\t" "vmovups 96(%2,%0,4), %%ymm7 \n\t"
"vmulps (%2,%0,4), %%ymm0, %%ymm1 \n\t" "addq $32, %0 \n\t"
"vmulps 32(%2,%0,4), %%ymm0, %%ymm2 \n\t" "subq $32, %1 \n\t"
"vaddps %%ymm8 , %%ymm1, %%ymm8 \n\t" "jz 2f \n\t"
"vmulps 64(%2,%0,4), %%ymm0, %%ymm3 \n\t"
"vaddps %%ymm9 , %%ymm2, %%ymm9 \n\t"
"vmulps 96(%2,%0,4), %%ymm0, %%ymm4 \n\t"
"vaddps %%ymm10, %%ymm3, %%ymm10 \n\t"
"vmulps 128(%2,%0,4), %%ymm0, %%ymm5 \n\t"
"vaddps %%ymm11, %%ymm4, %%ymm11 \n\t"
"vmulps 160(%2,%0,4), %%ymm0, %%ymm6 \n\t"
"vaddps %%ymm12, %%ymm5, %%ymm12 \n\t"
"vmulps 192(%2,%0,4), %%ymm0, %%ymm7 \n\t"
"vmulps 224(%2,%0,4), %%ymm0, %%ymm1 \n\t"
"vaddps %%ymm13, %%ymm6, %%ymm13 \n\t" ".align 16 \n\t"
"vmovups %%ymm8 , (%3,%0,4) \n\t" "1: \n\t"
"vaddps %%ymm14, %%ymm7, %%ymm14 \n\t"
"vmovups %%ymm9 , 32(%3,%0,4) \n\t"
"vaddps %%ymm15, %%ymm1, %%ymm15 \n\t"
"vmovups %%ymm10, 64(%3,%0,4) \n\t"
"vmovups %%ymm11, 96(%3,%0,4) \n\t"
"vmovups %%ymm12,128(%3,%0,4) \n\t"
"vmovups %%ymm13,160(%3,%0,4) \n\t"
"vmovups %%ymm14,192(%3,%0,4) \n\t"
"vmovups %%ymm15,224(%3,%0,4) \n\t"
"addq $64, %0 \n\t" "vmulps %%ymm4, %%ymm0, %%ymm4 \n\t"
"subq $64, %1 \n\t" "vaddps %%ymm8 , %%ymm4, %%ymm12 \n\t"
"vmulps %%ymm5, %%ymm0, %%ymm5 \n\t"
"vaddps %%ymm9 , %%ymm5, %%ymm13 \n\t"
"vmulps %%ymm6, %%ymm0, %%ymm6 \n\t"
"vaddps %%ymm10, %%ymm6, %%ymm14 \n\t"
"vmulps %%ymm7, %%ymm0, %%ymm7 \n\t"
"vaddps %%ymm11, %%ymm7, %%ymm15 \n\t"
"vmovups (%3,%0,4), %%ymm8 \n\t"
"vmovups 32(%3,%0,4), %%ymm9 \n\t"
"vmovups 64(%3,%0,4), %%ymm10 \n\t"
"vmovups 96(%3,%0,4), %%ymm11 \n\t"
"vmovups (%2,%0,4), %%ymm4 \n\t"
"vmovups 32(%2,%0,4), %%ymm5 \n\t"
"vmovups 64(%2,%0,4), %%ymm6 \n\t"
"vmovups 96(%2,%0,4), %%ymm7 \n\t"
"vmovups %%ymm12, -128(%3,%0,4) \n\t"
"vmovups %%ymm13, -96(%3,%0,4) \n\t"
"vmovups %%ymm14, -64(%3,%0,4) \n\t"
"vmovups %%ymm15, -32(%3,%0,4) \n\t"
"addq $32, %0 \n\t"
"subq $32, %1 \n\t"
"jnz 1b \n\t" "jnz 1b \n\t"
"vzeroupper \n\t"
"2: \n\t"
"vmulps %%ymm4, %%ymm0, %%ymm4 \n\t"
"vmulps %%ymm5, %%ymm0, %%ymm5 \n\t"
"vmulps %%ymm6, %%ymm0, %%ymm6 \n\t"
"vmulps %%ymm7, %%ymm0, %%ymm7 \n\t"
"vaddps %%ymm8 , %%ymm4, %%ymm12 \n\t"
"vaddps %%ymm9 , %%ymm5, %%ymm13 \n\t"
"vaddps %%ymm10, %%ymm6, %%ymm14 \n\t"
"vaddps %%ymm11, %%ymm7, %%ymm15 \n\t"
"vmovups %%ymm12, -128(%3,%0,4) \n\t"
"vmovups %%ymm13, -96(%3,%0,4) \n\t"
"vmovups %%ymm14, -64(%3,%0,4) \n\t"
"vmovups %%ymm15, -32(%3,%0,4) \n\t"
"vzeroupper \n\t"
: :
: :
@ -90,7 +109,7 @@ static void saxpy_kernel_16( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha)
: "cc", : "cc",
"%xmm0", "%xmm1", "%xmm2", "%xmm3", "%xmm0", "%xmm1", "%xmm2", "%xmm3",
"%xmm4", "%xmm5", "%xmm6", "%xmm7", "%xmm4", "%xmm5", "%xmm6", "%xmm7",
"%xmm8", "%xmm9", "%xmm10", "%xmm11", "%xmm8", "%xmm9", "%xmm10", "%xmm11",
"%xmm12", "%xmm13", "%xmm14", "%xmm15", "%xmm12", "%xmm13", "%xmm14", "%xmm15",
"memory" "memory"
); );