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

added optimized sgemv_t kernel for haswell

tags/v0.2.12^2
wernsaar 11 years ago
parent
2be5c7a640
commit
cbbc80aad3
2 changed files with 212 additions and 39 deletions
Split View
Diff Options
Show Stats Download Patch File Download Diff File
  1. +64
    -39
      kernel/x86_64/sgemv_n_microk_haswell-4.c
  2. +148
    -0
      kernel/x86_64/sgemv_t_microk_haswell-4.c

+ 64
- 39
kernel/x86_64/sgemv_n_microk_haswell-4.c View File

@@ -28,9 +28,9 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.


#define HAVE_KERNEL_4x8 1
static void sgemv_kernel_4x8( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, BLASLONG lda4) __attribute__ ((noinline));
static void sgemv_kernel_4x8( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, BLASLONG lda4, FLOAT *alpha) __attribute__ ((noinline));

static void sgemv_kernel_4x8( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, BLASLONG lda4)
static void sgemv_kernel_4x8( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, BLASLONG lda4, FLOAT *alpha)
{

BLASLONG register i = 0;
@@ -47,10 +47,13 @@ static void sgemv_kernel_4x8( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, BLASLO
"vbroadcastss 24(%2), %%ymm2 \n\t" // x6
"vbroadcastss 28(%2), %%ymm3 \n\t" // x7

"vbroadcastss (%9), %%ymm6 \n\t" // alpha

"testq $0x04, %1 \n\t"
"jz .L08LABEL%= \n\t"

"vmovups (%3,%0,4), %%xmm4 \n\t" // 4 * y
"vmovups (%3,%0,4), %%xmm7 \n\t" // 4 * y
"vxorps %%xmm4 , %%xmm4, %%xmm4 \n\t"
"vxorps %%xmm5 , %%xmm5, %%xmm5 \n\t"

"vfmadd231ps (%4,%0,4), %%xmm12, %%xmm4 \n\t"
@@ -64,6 +67,8 @@ static void sgemv_kernel_4x8( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, BLASLO
"vfmadd231ps (%7,%8,4), %%xmm3 , %%xmm5 \n\t"

"vaddps %%xmm4 , %%xmm5 , %%xmm5 \n\t"
"vmulps %%xmm6 , %%xmm5 , %%xmm5 \n\t"
"vaddps %%xmm7 , %%xmm5 , %%xmm5 \n\t"

"vmovups %%xmm5, (%3,%0,4) \n\t" // 4 * y

@@ -76,7 +81,8 @@ static void sgemv_kernel_4x8( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, BLASLO
"testq $0x08, %1 \n\t"
"jz .L16LABEL%= \n\t"

"vmovups (%3,%0,4), %%ymm4 \n\t" // 8 * y
"vmovups (%3,%0,4), %%ymm7 \n\t" // 8 * y
"vxorps %%ymm4 , %%ymm4, %%ymm4 \n\t"
"vxorps %%ymm5 , %%ymm5, %%ymm5 \n\t"

"vfmadd231ps (%4,%0,4), %%ymm12, %%ymm4 \n\t"
@@ -90,6 +96,9 @@ static void sgemv_kernel_4x8( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, BLASLO
"vfmadd231ps (%7,%8,4), %%ymm3 , %%ymm5 \n\t"

"vaddps %%ymm4 , %%ymm5 , %%ymm5 \n\t"
"vmulps %%ymm6 , %%ymm5 , %%ymm5 \n\t"
"vaddps %%ymm7 , %%ymm5 , %%ymm5 \n\t"


"vmovups %%ymm5, (%3,%0,4) \n\t" // 8 * y

@@ -105,46 +114,43 @@ static void sgemv_kernel_4x8( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, BLASLO

".align 16 \n\t"
".L01LOOP%=: \n\t"
// "prefetcht0 192(%3,%0,4) \n\t"
"vmovups (%3,%0,4), %%ymm4 \n\t" // 8 * y
"vmovups 32(%3,%0,4), %%ymm5 \n\t" // 8 * y

// "prefetcht0 192(%4,%0,4) \n\t"
"vxorps %%ymm4 , %%ymm4, %%ymm4 \n\t"
"vxorps %%ymm5 , %%ymm5, %%ymm5 \n\t"
"vmovups (%3,%0,4), %%ymm8 \n\t" // 8 * y
"vmovups 32(%3,%0,4), %%ymm9 \n\t" // 8 * y

"vfmadd231ps (%4,%0,4), %%ymm12, %%ymm4 \n\t"
"vfmadd231ps 32(%4,%0,4), %%ymm12, %%ymm5 \n\t"
// "prefetcht0 192(%5,%0,4) \n\t"
"vfmadd231ps (%5,%0,4), %%ymm13, %%ymm4 \n\t"
"vfmadd231ps 32(%5,%0,4), %%ymm13, %%ymm5 \n\t"
// "prefetcht0 192(%6,%0,4) \n\t"
"vfmadd231ps (%6,%0,4), %%ymm14, %%ymm4 \n\t"
"vfmadd231ps 32(%6,%0,4), %%ymm14, %%ymm5 \n\t"
// "prefetcht0 192(%7,%0,4) \n\t"
"vfmadd231ps (%7,%0,4), %%ymm15, %%ymm4 \n\t"
"vfmadd231ps 32(%7,%0,4), %%ymm15, %%ymm5 \n\t"

// "prefetcht0 192(%4,%8,4) \n\t"
"vfmadd231ps (%4,%8,4), %%ymm0 , %%ymm4 \n\t"
"addq $16, %0 \n\t"
"addq $16, %0 \n\t"
"vfmadd231ps 32(%4,%8,4), %%ymm0 , %%ymm5 \n\t"
// "prefetcht0 192(%5,%8,4) \n\t"
"vfmadd231ps (%5,%8,4), %%ymm1 , %%ymm4 \n\t"
"vfmadd231ps 32(%5,%8,4), %%ymm1 , %%ymm5 \n\t"
// "prefetcht0 192(%6,%8,4) \n\t"
"vfmadd231ps (%6,%8,4), %%ymm2 , %%ymm4 \n\t"
"vfmadd231ps 32(%6,%8,4), %%ymm2 , %%ymm5 \n\t"
// "prefetcht0 192(%7,%8,4) \n\t"
"vfmadd231ps (%7,%8,4), %%ymm3 , %%ymm4 \n\t"
"vfmadd231ps 32(%7,%8,4), %%ymm3 , %%ymm5 \n\t"

"vfmadd231ps %%ymm6 , %%ymm4 , %%ymm8 \n\t"
"vfmadd231ps %%ymm6 , %%ymm5 , %%ymm9 \n\t"

"addq $16, %8 \n\t"
"vmovups %%ymm4,-64(%3,%0,4) \n\t" // 8 * y
"vmovups %%ymm8,-64(%3,%0,4) \n\t" // 8 * y
"subq $16, %1 \n\t"
"vmovups %%ymm5,-32(%3,%0,4) \n\t" // 8 * y
"vmovups %%ymm9,-32(%3,%0,4) \n\t" // 8 * y

"jnz .L01LOOP%= \n\t"

".L16END%=: \n\t"
"vzeroupper \n\t"
"vzeroupper \n\t"

:
:
@@ -156,11 +162,14 @@ static void sgemv_kernel_4x8( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, BLASLO
"r" (ap[1]), // 5
"r" (ap[2]), // 6
"r" (ap[3]), // 7
"r" (lda4) // 8
"r" (lda4), // 8
"r" (alpha) // 9
: "cc",
"%xmm0", "%xmm1",
"%xmm2", "%xmm3",
"%xmm4", "%xmm5",
"%xmm6", "%xmm7",
"%xmm8", "%xmm9",
"%xmm12", "%xmm13", "%xmm14", "%xmm15",
"memory"
);
@@ -170,9 +179,9 @@ static void sgemv_kernel_4x8( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, BLASLO


#define HAVE_KERNEL_4x4 1
static void sgemv_kernel_4x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y) __attribute__ ((noinline));
static void sgemv_kernel_4x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, FLOAT *alpha) __attribute__ ((noinline));

static void sgemv_kernel_4x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
static void sgemv_kernel_4x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, FLOAT *alpha)
{

BLASLONG register i = 0;
@@ -185,17 +194,25 @@ static void sgemv_kernel_4x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
"vbroadcastss 8(%2), %%ymm14 \n\t" // x2
"vbroadcastss 12(%2), %%ymm15 \n\t" // x3

"vbroadcastss (%8), %%ymm6 \n\t" // alpha

"testq $0x04, %1 \n\t"
"jz .L08LABEL%= \n\t"

"vmovups (%3,%0,4), %%xmm4 \n\t" // 4 * y
"vxorps %%ymm4 , %%ymm4, %%ymm4 \n\t"
"vxorps %%ymm5 , %%ymm5, %%ymm5 \n\t"
"vmovups (%3,%0,4), %%xmm7 \n\t" // 4 * y

"vfmadd231ps (%4,%0,4), %%xmm12, %%xmm4 \n\t"
"vfmadd231ps (%5,%0,4), %%xmm13, %%xmm4 \n\t"
"vfmadd231ps (%5,%0,4), %%xmm13, %%xmm5 \n\t"
"vfmadd231ps (%6,%0,4), %%xmm14, %%xmm4 \n\t"
"vfmadd231ps (%7,%0,4), %%xmm15, %%xmm4 \n\t"
"vfmadd231ps (%7,%0,4), %%xmm15, %%xmm5 \n\t"

"vaddps %%xmm4 , %%xmm5 , %%xmm5 \n\t"
"vmulps %%xmm6 , %%xmm5 , %%xmm5 \n\t"
"vaddps %%xmm7 , %%xmm5 , %%xmm5 \n\t"

"vmovups %%xmm4, (%3,%0,4) \n\t" // 4 * y
"vmovups %%xmm5, (%3,%0,4) \n\t" // 4 * y

"addq $4 , %0 \n\t"
"subq $4 , %1 \n\t"
@@ -205,14 +222,20 @@ static void sgemv_kernel_4x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
"testq $0x08, %1 \n\t"
"jz .L16LABEL%= \n\t"

"vmovups (%3,%0,4), %%ymm4 \n\t" // 8 * y
"vxorps %%ymm4 , %%ymm4, %%ymm4 \n\t"
"vxorps %%ymm5 , %%ymm5, %%ymm5 \n\t"
"vmovups (%3,%0,4), %%ymm7 \n\t" // 8 * y

"vfmadd231ps (%4,%0,4), %%ymm12, %%ymm4 \n\t"
"vfmadd231ps (%5,%0,4), %%ymm13, %%ymm4 \n\t"
"vfmadd231ps (%5,%0,4), %%ymm13, %%ymm5 \n\t"
"vfmadd231ps (%6,%0,4), %%ymm14, %%ymm4 \n\t"
"vfmadd231ps (%7,%0,4), %%ymm15, %%ymm4 \n\t"
"vfmadd231ps (%7,%0,4), %%ymm15, %%ymm5 \n\t"

"vaddps %%ymm4 , %%ymm5 , %%ymm5 \n\t"
"vmulps %%ymm6 , %%ymm5 , %%ymm5 \n\t"
"vaddps %%ymm7 , %%ymm5 , %%ymm5 \n\t"

"vmovups %%ymm4, (%3,%0,4) \n\t" // 8 * y
"vmovups %%ymm5, (%3,%0,4) \n\t" // 8 * y

"addq $8 , %0 \n\t"
"subq $8 , %1 \n\t"
@@ -225,24 +248,23 @@ static void sgemv_kernel_4x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)

".align 16 \n\t"
".L01LOOP%=: \n\t"
"vmovups (%3,%0,4), %%ymm4 \n\t" // 8 * y
"vmovups 32(%3,%0,4), %%ymm5 \n\t" // 8 * y
"vmovups (%3,%0,4), %%ymm8 \n\t" // 8 * y
"vmovups 32(%3,%0,4), %%ymm9 \n\t" // 8 * y

"prefetcht0 192(%4,%0,4) \n\t"
"vfmadd231ps (%4,%0,4), %%ymm12, %%ymm4 \n\t"
"vfmadd231ps 32(%4,%0,4), %%ymm12, %%ymm5 \n\t"
"prefetcht0 192(%5,%0,4) \n\t"
"vfmadd231ps (%5,%0,4), %%ymm13, %%ymm4 \n\t"
"vfmadd231ps 32(%5,%0,4), %%ymm13, %%ymm5 \n\t"
"prefetcht0 192(%6,%0,4) \n\t"
"vfmadd231ps (%6,%0,4), %%ymm14, %%ymm4 \n\t"
"vfmadd231ps 32(%6,%0,4), %%ymm14, %%ymm5 \n\t"
"prefetcht0 192(%7,%0,4) \n\t"
"vfmadd231ps (%7,%0,4), %%ymm15, %%ymm4 \n\t"
"vfmadd231ps 32(%7,%0,4), %%ymm15, %%ymm5 \n\t"

"vmovups %%ymm4, (%3,%0,4) \n\t" // 8 * y
"vmovups %%ymm5, 32(%3,%0,4) \n\t" // 8 * y
"vfmadd231ps %%ymm6 , %%ymm4 , %%ymm8 \n\t"
"vfmadd231ps %%ymm6 , %%ymm5 , %%ymm9 \n\t"

"vmovups %%ymm8, (%3,%0,4) \n\t" // 8 * y
"vmovups %%ymm9, 32(%3,%0,4) \n\t" // 8 * y

"addq $16, %0 \n\t"
"subq $16, %1 \n\t"
@@ -260,9 +282,12 @@ static void sgemv_kernel_4x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
"r" (ap[0]), // 4
"r" (ap[1]), // 5
"r" (ap[2]), // 6
"r" (ap[3]) // 7
"r" (ap[3]), // 7
"r" (alpha) // 8
: "cc",
"%xmm4", "%xmm5",
"%xmm6", "%xmm7",
"%xmm8", "%xmm9",
"%xmm12", "%xmm13", "%xmm14", "%xmm15",
"memory"
);


+ 148
- 0
kernel/x86_64/sgemv_t_microk_haswell-4.c View File

@@ -0,0 +1,148 @@
/***************************************************************************
Copyright (c) 2014, The OpenBLAS Project
All rights reserved.
Redistribution and use in source and binary forms, with or without
modification, are permitted provided that the following conditions are
met:
1. Redistributions of source code must retain the above copyright
notice, this list of conditions and the following disclaimer.
2. Redistributions in binary form must reproduce the above copyright
notice, this list of conditions and the following disclaimer in
the documentation and/or other materials provided with the
distribution.
3. Neither the name of the OpenBLAS project nor the names of
its contributors may be used to endorse or promote products
derived from this software without specific prior written permission.
THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*****************************************************************************/

#define HAVE_KERNEL_4x4 1
static void sgemv_kernel_4x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y) __attribute__ ((noinline));

static void sgemv_kernel_4x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
{

BLASLONG register i = 0;

__asm__ __volatile__
(
"vzeroupper \n\t"
"vxorps %%ymm4 , %%ymm4, %%ymm4 \n\t"
"vxorps %%ymm5 , %%ymm5, %%ymm5 \n\t"
"vxorps %%ymm6 , %%ymm6, %%ymm6 \n\t"
"vxorps %%ymm7 , %%ymm7, %%ymm7 \n\t"

"testq $0x04, %1 \n\t"
"jz .L08LABEL%= \n\t"

"vmovups (%2,%0,4), %%xmm12 \n\t" // 4 * x

"vfmadd231ps (%4,%0,4), %%xmm12, %%xmm4 \n\t"
"vfmadd231ps (%5,%0,4), %%xmm12, %%xmm5 \n\t"
"vfmadd231ps (%6,%0,4), %%xmm12, %%xmm6 \n\t"
"vfmadd231ps (%7,%0,4), %%xmm12, %%xmm7 \n\t"

"addq $4 , %0 \n\t"
"subq $4 , %1 \n\t"

".L08LABEL%=: \n\t"

"testq $0x08, %1 \n\t"
"jz .L16LABEL%= \n\t"

"vmovups (%2,%0,4), %%ymm12 \n\t" // 8 * x

"vfmadd231ps (%4,%0,4), %%ymm12, %%ymm4 \n\t"
"vfmadd231ps (%5,%0,4), %%ymm12, %%ymm5 \n\t"
"vfmadd231ps (%6,%0,4), %%ymm12, %%ymm6 \n\t"
"vfmadd231ps (%7,%0,4), %%ymm12, %%ymm7 \n\t"

"addq $8 , %0 \n\t"
"subq $8 , %1 \n\t"

".L16LABEL%=: \n\t"

"cmpq $0, %1 \n\t"
"je .L16END%= \n\t"


".align 16 \n\t"
".L01LOOP%=: \n\t"
"prefetcht0 384(%2,%0,4) \n\t"
"vmovups (%2,%0,4), %%ymm12 \n\t" // 8 * x
"vmovups 32(%2,%0,4), %%ymm13 \n\t" // 8 * x

"prefetcht0 384(%4,%0,4) \n\t"
"vfmadd231ps (%4,%0,4), %%ymm12, %%ymm4 \n\t"
"vfmadd231ps (%5,%0,4), %%ymm12, %%ymm5 \n\t"
"prefetcht0 384(%5,%0,4) \n\t"
"vfmadd231ps 32(%4,%0,4), %%ymm13, %%ymm4 \n\t"
"vfmadd231ps 32(%5,%0,4), %%ymm13, %%ymm5 \n\t"
"prefetcht0 384(%6,%0,4) \n\t"
"vfmadd231ps (%6,%0,4), %%ymm12, %%ymm6 \n\t"
"vfmadd231ps (%7,%0,4), %%ymm12, %%ymm7 \n\t"
"prefetcht0 384(%7,%0,4) \n\t"
"vfmadd231ps 32(%6,%0,4), %%ymm13, %%ymm6 \n\t"
"vfmadd231ps 32(%7,%0,4), %%ymm13, %%ymm7 \n\t"

"addq $16, %0 \n\t"
"subq $16, %1 \n\t"
"jnz .L01LOOP%= \n\t"

".L16END%=: \n\t"

"vextractf128 $1 , %%ymm4, %%xmm12 \n\t"
"vextractf128 $1 , %%ymm5, %%xmm13 \n\t"
"vextractf128 $1 , %%ymm6, %%xmm14 \n\t"
"vextractf128 $1 , %%ymm7, %%xmm15 \n\t"

"vaddps %%xmm4, %%xmm12, %%xmm4 \n\t"
"vaddps %%xmm5, %%xmm13, %%xmm5 \n\t"
"vaddps %%xmm6, %%xmm14, %%xmm6 \n\t"
"vaddps %%xmm7, %%xmm15, %%xmm7 \n\t"

"vhaddps %%xmm4, %%xmm4, %%xmm4 \n\t"
"vhaddps %%xmm5, %%xmm5, %%xmm5 \n\t"
"vhaddps %%xmm6, %%xmm6, %%xmm6 \n\t"
"vhaddps %%xmm7, %%xmm7, %%xmm7 \n\t"

"vhaddps %%xmm4, %%xmm4, %%xmm4 \n\t"
"vhaddps %%xmm5, %%xmm5, %%xmm5 \n\t"
"vhaddps %%xmm6, %%xmm6, %%xmm6 \n\t"
"vhaddps %%xmm7, %%xmm7, %%xmm7 \n\t"

"vmovss %%xmm4, (%3) \n\t"
"vmovss %%xmm5, 4(%3) \n\t"
"vmovss %%xmm6, 8(%3) \n\t"
"vmovss %%xmm7, 12(%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" (ap[2]), // 6
"r" (ap[3]) // 7
: "cc",
"%xmm4", "%xmm5", "%xmm6", "%xmm7",
"%xmm12", "%xmm13", "%xmm14", "%xmm15",
"memory"
);

}



Write Preview
Loading…
Cancel
Save