OSchip
/
OpenBLAS
Not watched
1
0
Fork 0
Code
Releases 66 Wiki evaluate Activity
Issues 0 Pull Requests 0 Datasets Model Cloudbrain HPC
Browse Source

optimized zgemv_t for bulldozer

tags/v0.2.12^2
wernsaar 11 years ago
parent
a0fb68ab42
commit
a8b0812feb
1 changed files with 236 additions and 2 deletions
Unified View
Diff Options
Show Stats Download Patch File Download Diff File
  1. +236
    -2
      kernel/x86_64/zgemv_t_microk_bulldozer-4.c

+ 236
- 2
kernel/x86_64/zgemv_t_microk_bulldozer-4.c View File

@@ -31,8 +31,6 @@ static void zgemv_kernel_4x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, FLOAT
static void zgemv_kernel_4x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, FLOAT *alpha) static void zgemv_kernel_4x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, FLOAT *alpha)
{ {


if ( n < 4 ) return;

BLASLONG register i = 0; BLASLONG register i = 0;


__asm__ __volatile__ __asm__ __volatile__
@@ -220,4 +218,240 @@ static void zgemv_kernel_4x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, FLOAT


} }


#define HAVE_KERNEL_4x2 1
static void zgemv_kernel_4x2( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, FLOAT *alpha) __attribute__ ((noinline));

static void zgemv_kernel_4x2( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, FLOAT *alpha)
{

BLASLONG register i = 0;

__asm__ __volatile__
(
"vzeroupper \n\t"

"vxorpd %%xmm8 , %%xmm8 , %%xmm8 \n\t" // temp
"vxorpd %%xmm9 , %%xmm9 , %%xmm9 \n\t" // temp
"vxorpd %%xmm10, %%xmm10, %%xmm10 \n\t" // temp
"vxorpd %%xmm11, %%xmm11, %%xmm11 \n\t" // temp

".align 16 \n\t"
".L01LOOP%=: \n\t"

"vmovddup (%2,%0,8), %%xmm0 \n\t" // real value from x0
"vmovddup 8(%2,%0,8), %%xmm1 \n\t" // imag value from x0

"prefetcht0 192(%4,%0,8) \n\t"
"vmovups (%4,%0,8), %%xmm4 \n\t" // 1 complex values from a0
"prefetcht0 192(%5,%0,8) \n\t"
"vmovups (%5,%0,8), %%xmm5 \n\t" // 1 complex values from a1

"vfmaddpd %%xmm8 , %%xmm4 , %%xmm0, %%xmm8 \n\t" // ar0*xr0,al0*xr0
"vfmaddpd %%xmm9 , %%xmm4 , %%xmm1, %%xmm9 \n\t" // ar0*xl0,al0*xl0
"vfmaddpd %%xmm10, %%xmm5 , %%xmm0, %%xmm10 \n\t" // ar0*xr0,al0*xr0
"vfmaddpd %%xmm11, %%xmm5 , %%xmm1, %%xmm11 \n\t" // ar0*xl0,al0*xl0

"vmovddup 16(%2,%0,8), %%xmm0 \n\t" // real value from x0
"vmovddup 24(%2,%0,8), %%xmm1 \n\t" // imag value from x0

"vmovups 16(%4,%0,8), %%xmm4 \n\t" // 1 complex values from a0
"vmovups 16(%5,%0,8), %%xmm5 \n\t" // 1 complex values from a1

"vfmaddpd %%xmm8 , %%xmm4 , %%xmm0, %%xmm8 \n\t" // ar0*xr0,al0*xr0
"vfmaddpd %%xmm9 , %%xmm4 , %%xmm1, %%xmm9 \n\t" // ar0*xl0,al0*xl0
"vfmaddpd %%xmm10, %%xmm5 , %%xmm0, %%xmm10 \n\t" // ar0*xr0,al0*xr0
"vfmaddpd %%xmm11, %%xmm5 , %%xmm1, %%xmm11 \n\t" // ar0*xl0,al0*xl0

"vmovddup 32(%2,%0,8), %%xmm0 \n\t" // real value from x0
"vmovddup 40(%2,%0,8), %%xmm1 \n\t" // imag value from x0

"vmovups 32(%4,%0,8), %%xmm4 \n\t" // 1 complex values from a0
"vmovups 32(%5,%0,8), %%xmm5 \n\t" // 1 complex values from a1

"vfmaddpd %%xmm8 , %%xmm4 , %%xmm0, %%xmm8 \n\t" // ar0*xr0,al0*xr0
"vfmaddpd %%xmm9 , %%xmm4 , %%xmm1, %%xmm9 \n\t" // ar0*xl0,al0*xl0
"vfmaddpd %%xmm10, %%xmm5 , %%xmm0, %%xmm10 \n\t" // ar0*xr0,al0*xr0
"vfmaddpd %%xmm11, %%xmm5 , %%xmm1, %%xmm11 \n\t" // ar0*xl0,al0*xl0

"vmovddup 48(%2,%0,8), %%xmm0 \n\t" // real value from x0
"vmovddup 56(%2,%0,8), %%xmm1 \n\t" // imag value from x0

"vmovups 48(%4,%0,8), %%xmm4 \n\t" // 1 complex values from a0
"vmovups 48(%5,%0,8), %%xmm5 \n\t" // 1 complex values from a1

"vfmaddpd %%xmm8 , %%xmm4 , %%xmm0, %%xmm8 \n\t" // ar0*xr0,al0*xr0
"vfmaddpd %%xmm9 , %%xmm4 , %%xmm1, %%xmm9 \n\t" // ar0*xl0,al0*xl0
"vfmaddpd %%xmm10, %%xmm5 , %%xmm0, %%xmm10 \n\t" // ar0*xr0,al0*xr0
"vfmaddpd %%xmm11, %%xmm5 , %%xmm1, %%xmm11 \n\t" // ar0*xl0,al0*xl0

"addq $8 , %0 \n\t"
"subq $4 , %1 \n\t"
"jnz .L01LOOP%= \n\t"

"vmovddup (%6) , %%xmm0 \n\t" // value from alpha
"vmovddup 8(%6) , %%xmm1 \n\t" // value from alpha

#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
"vpermilpd $0x1 , %%xmm9 , %%xmm9 \n\t"
"vpermilpd $0x1 , %%xmm11, %%xmm11 \n\t"
"vaddsubpd %%xmm9 , %%xmm8, %%xmm8 \n\t"
"vaddsubpd %%xmm11, %%xmm10, %%xmm10 \n\t"
#else
"vpermilpd $0x1 , %%xmm8 , %%xmm8 \n\t"
"vpermilpd $0x1 , %%xmm10, %%xmm10 \n\t"
"vaddsubpd %%xmm8 , %%xmm9 , %%xmm8 \n\t"
"vaddsubpd %%xmm10, %%xmm11, %%xmm10 \n\t"
"vpermilpd $0x1 , %%xmm8 , %%xmm8 \n\t"
"vpermilpd $0x1 , %%xmm10, %%xmm10 \n\t"
#endif

"vmulpd %%xmm8 , %%xmm1 , %%xmm9 \n\t" // t_r * alpha_i , t_i * alpha_i
"vmulpd %%xmm8 , %%xmm0 , %%xmm8 \n\t" // t_r * alpha_r , t_i * alpha_r
"vmulpd %%xmm10, %%xmm1 , %%xmm11 \n\t" // t_r * alpha_i , t_i * alpha_i
"vmulpd %%xmm10, %%xmm0 , %%xmm10 \n\t" // t_r * alpha_r , t_i * alpha_r

#if !defined(XCONJ)
"vpermilpd $0x1 , %%xmm9 , %%xmm9 \n\t"
"vpermilpd $0x1 , %%xmm11, %%xmm11 \n\t"
"vaddsubpd %%xmm9 , %%xmm8, %%xmm8 \n\t"
"vaddsubpd %%xmm11, %%xmm10, %%xmm10 \n\t"
#else
"vpermilpd $0x1 , %%xmm8 , %%xmm8 \n\t"
"vpermilpd $0x1 , %%xmm10, %%xmm10 \n\t"
"vaddsubpd %%xmm8 , %%xmm9 , %%xmm8 \n\t"
"vaddsubpd %%xmm10, %%xmm11, %%xmm10 \n\t"
"vpermilpd $0x1 , %%xmm8 , %%xmm8 \n\t"
"vpermilpd $0x1 , %%xmm10, %%xmm10 \n\t"
#endif

"vaddpd (%3) , %%xmm8 , %%xmm8 \n\t"
"vaddpd 16(%3) , %%xmm10, %%xmm10 \n\t"

"vmovups %%xmm8 , (%3) \n\t"
"vmovups %%xmm10, 16(%3) \n\t"

"vzeroupper \n\t"

:
:
"r" (i), // 0
"r" (n), // 1
"r" (x), // 2
"r" (y), // 3
"r" (ap[0]), // 4
"r" (ap[1]), // 5
"r" (alpha) // 6
: "cc",
"%xmm0", "%xmm1", "%xmm2", "%xmm3",
"%xmm4", "%xmm5", "%xmm6", "%xmm7",
"%xmm8", "%xmm9", "%xmm10", "%xmm11",
"%xmm12", "%xmm13", "%xmm14", "%xmm15",
"memory"
);

}



#define HAVE_KERNEL_4x1 1
static void zgemv_kernel_4x1( BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y, FLOAT *alpha) __attribute__ ((noinline));

static void zgemv_kernel_4x1( BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y, FLOAT *alpha)
{

BLASLONG register i = 0;

__asm__ __volatile__
(
"vzeroupper \n\t"

"vxorpd %%xmm8 , %%xmm8 , %%xmm8 \n\t" // temp
"vxorpd %%xmm9 , %%xmm9 , %%xmm9 \n\t" // temp

".align 16 \n\t"
".L01LOOP%=: \n\t"

"vmovddup (%2,%0,8), %%xmm0 \n\t" // real value from x0
"vmovddup 8(%2,%0,8), %%xmm1 \n\t" // imag value from x0

"prefetcht0 192(%4,%0,8) \n\t"
"vmovups (%4,%0,8), %%xmm4 \n\t" // 1 complex values from a0
"vmovups 16(%4,%0,8), %%xmm5 \n\t" // 1 complex values from a0

"vmovddup 16(%2,%0,8), %%xmm2 \n\t" // real value from x0
"vmovddup 24(%2,%0,8), %%xmm3 \n\t" // imag value from x0

"vfmaddpd %%xmm8 , %%xmm4 , %%xmm0, %%xmm8 \n\t" // ar0*xr0,al0*xr0
"vfmaddpd %%xmm9 , %%xmm4 , %%xmm1, %%xmm9 \n\t" // ar0*xl0,al0*xl0

"vmovddup 32(%2,%0,8), %%xmm0 \n\t" // real value from x0
"vmovddup 40(%2,%0,8), %%xmm1 \n\t" // imag value from x0

"vfmaddpd %%xmm8 , %%xmm5 , %%xmm2, %%xmm8 \n\t" // ar0*xr0,al0*xr0
"vfmaddpd %%xmm9 , %%xmm5 , %%xmm3, %%xmm9 \n\t" // ar0*xl0,al0*xl0

"vmovups 32(%4,%0,8), %%xmm4 \n\t" // 1 complex values from a0
"vmovups 48(%4,%0,8), %%xmm5 \n\t" // 1 complex values from a0

"vmovddup 48(%2,%0,8), %%xmm2 \n\t" // real value from x0
"vmovddup 56(%2,%0,8), %%xmm3 \n\t" // imag value from x0

"addq $8 , %0 \n\t"
"vfmaddpd %%xmm8 , %%xmm4 , %%xmm0, %%xmm8 \n\t" // ar0*xr0,al0*xr0
"vfmaddpd %%xmm9 , %%xmm4 , %%xmm1, %%xmm9 \n\t" // ar0*xl0,al0*xl0

"subq $4 , %1 \n\t"
"vfmaddpd %%xmm8 , %%xmm5 , %%xmm2, %%xmm8 \n\t" // ar0*xr0,al0*xr0
"vfmaddpd %%xmm9 , %%xmm5 , %%xmm3, %%xmm9 \n\t" // ar0*xl0,al0*xl0

"jnz .L01LOOP%= \n\t"

"vmovddup (%5) , %%xmm0 \n\t" // value from alpha
"vmovddup 8(%5) , %%xmm1 \n\t" // value from alpha

#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
"vpermilpd $0x1 , %%xmm9 , %%xmm9 \n\t"
"vaddsubpd %%xmm9 , %%xmm8, %%xmm8 \n\t"
#else
"vpermilpd $0x1 , %%xmm8 , %%xmm8 \n\t"
"vaddsubpd %%xmm8 , %%xmm9 , %%xmm8 \n\t"
"vpermilpd $0x1 , %%xmm8 , %%xmm8 \n\t"
#endif

"vmulpd %%xmm8 , %%xmm1 , %%xmm9 \n\t" // t_r * alpha_i , t_i * alpha_i
"vmulpd %%xmm8 , %%xmm0 , %%xmm8 \n\t" // t_r * alpha_r , t_i * alpha_r

#if !defined(XCONJ)
"vpermilpd $0x1 , %%xmm9 , %%xmm9 \n\t"
"vaddsubpd %%xmm9 , %%xmm8, %%xmm8 \n\t"
#else
"vpermilpd $0x1 , %%xmm8 , %%xmm8 \n\t"
"vaddsubpd %%xmm8 , %%xmm9 , %%xmm8 \n\t"
"vpermilpd $0x1 , %%xmm8 , %%xmm8 \n\t"
#endif

"vaddpd (%3) , %%xmm8 , %%xmm8 \n\t"

"vmovups %%xmm8 , (%3) \n\t"

"vzeroupper \n\t"

:
:
"r" (i), // 0
"r" (n), // 1
"r" (x), // 2
"r" (y), // 3
"r" (ap), // 4
"r" (alpha) // 5
: "cc",
"%xmm0", "%xmm1", "%xmm2", "%xmm3",
"%xmm4", "%xmm5", "%xmm6", "%xmm7",
"%xmm8", "%xmm9", "%xmm10", "%xmm11",
"%xmm12", "%xmm13", "%xmm14", "%xmm15",
"memory"
);

}




Write Preview
Loading…
Cancel
Save