From 6d0db0151f0ee4c2ac9561d4b1ece7db1d96fdef Mon Sep 17 00:00:00 2001 From: Werner Saar Date: Thu, 16 Apr 2015 11:19:37 +0200 Subject: [PATCH] added optimized zaxpy-kernels --- kernel/x86_64/KERNEL.HASWELL | 1 + kernel/x86_64/KERNEL.SANDYBRIDGE | 1 + kernel/x86_64/zaxpy.c | 12 +- kernel/x86_64/zaxpy_microk_bulldozer-2.c | 173 +++++++++++++++++-------- kernel/x86_64/zaxpy_microk_haswell-2.c | 132 +++++++++++++++++++ kernel/x86_64/zaxpy_microk_sandy-2.c | 198 ++++++++++++++++++++++++++++ kernel/x86_64/zaxpy_microk_steamroller-2.c | 200 +++++++++++++++++++++++++++++ lapack-netlib/TESTING/sep.in | 4 +- lapack-netlib/TESTING/zctest.in | 2 +- 9 files changed, 663 insertions(+), 60 deletions(-) create mode 100644 kernel/x86_64/zaxpy_microk_haswell-2.c create mode 100644 kernel/x86_64/zaxpy_microk_sandy-2.c create mode 100644 kernel/x86_64/zaxpy_microk_steamroller-2.c diff --git a/kernel/x86_64/KERNEL.HASWELL b/kernel/x86_64/KERNEL.HASWELL index 9cce777..36f0762 100644 --- a/kernel/x86_64/KERNEL.HASWELL +++ b/kernel/x86_64/KERNEL.HASWELL @@ -18,6 +18,7 @@ ZDOTKERNEL = zdot.c SAXPYKERNEL = saxpy.c DAXPYKERNEL = daxpy.c CAXPYKERNEL = caxpy.c +ZAXPYKERNEL = zaxpy.c SGEMMKERNEL = sgemm_kernel_16x4_haswell.S SGEMMINCOPY = ../generic/gemm_ncopy_16.c diff --git a/kernel/x86_64/KERNEL.SANDYBRIDGE b/kernel/x86_64/KERNEL.SANDYBRIDGE index b783c9a..9ae5681 100644 --- a/kernel/x86_64/KERNEL.SANDYBRIDGE +++ b/kernel/x86_64/KERNEL.SANDYBRIDGE @@ -12,6 +12,7 @@ ZDOTKERNEL = zdot.c SAXPYKERNEL = saxpy.c DAXPYKERNEL = daxpy.c CAXPYKERNEL = caxpy.c +ZAXPYKERNEL = zaxpy.c SGEMMKERNEL = sgemm_kernel_16x4_sandy.S SGEMMINCOPY = ../generic/gemm_ncopy_16.c diff --git a/kernel/x86_64/zaxpy.c b/kernel/x86_64/zaxpy.c index 52a25c7..1aa95d2 100644 --- a/kernel/x86_64/zaxpy.c +++ b/kernel/x86_64/zaxpy.c @@ -29,8 +29,14 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. #include "common.h" -#if defined(BULLDOZER) || defined(PILEDRIVER) || defined(STEAMROLLER) +#if defined(BULLDOZER) #include "zaxpy_microk_bulldozer-2.c" +#elif defined(PILEDRIVER) || defined(STEAMROLLER) +#include "zaxpy_microk_steamroller-2.c" +#elif defined(HASWELL) +#include "zaxpy_microk_haswell-2.c" +#elif defined(SANDYBRIDGE) +#include "zaxpy_microk_sandy-2.c" #endif @@ -78,13 +84,13 @@ int CNAME(BLASLONG n, BLASLONG dummy0, BLASLONG dummy1, FLOAT da_r, FLOAT da_i, if ( (inc_x == 1) && (inc_y == 1) ) { - int n1 = n & -4; + int n1 = n & -16; if ( n1 ) { da[0] = da_r; da[1] = da_i; - zaxpy_kernel_4(n1, x, y , &da ); + zaxpy_kernel_4(n1, x, y , da ); ix = 2 * n1; } i = n1; diff --git a/kernel/x86_64/zaxpy_microk_bulldozer-2.c b/kernel/x86_64/zaxpy_microk_bulldozer-2.c index f9732cd..0e15761 100644 --- a/kernel/x86_64/zaxpy_microk_bulldozer-2.c +++ b/kernel/x86_64/zaxpy_microk_bulldozer-2.c @@ -31,89 +31,154 @@ static void zaxpy_kernel_4( BLASLONG n, FLOAT *x, FLOAT *y , FLOAT *alpha) __att static void zaxpy_kernel_4( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha) { +#if !defined(CONJ) + FLOAT mvec[2] = { -1.0, 1.0 }; +#else + FLOAT mvec[2] = { 1.0, -1.0 }; +#endif BLASLONG register i = 0; + if ( n < 384 ) + { + __asm__ __volatile__ ( + "vzeroupper \n\t" "vmovddup (%4), %%xmm0 \n\t" // real part of alpha "vmovddup 8(%4), %%xmm1 \n\t" // imag part of alpha +#if !defined(CONJ) + "vmulpd (%5), %%xmm1 , %%xmm1 \n\t" +#else + "vmulpd (%5), %%xmm0 , %%xmm0 \n\t" +#endif ".align 16 \n\t" "1: \n\t" - "prefetcht0 768(%2,%0,8) \n\t" "vmovups (%2,%0,8), %%xmm5 \n\t" // 1 complex values from x + ".align 2 \n\t" "vmovups 16(%2,%0,8), %%xmm7 \n\t" // 1 complex values from x "vmovups 32(%2,%0,8), %%xmm9 \n\t" // 1 complex values from x "vmovups 48(%2,%0,8), %%xmm11 \n\t" // 1 complex values from x - "prefetcht0 768(%3,%0,8) \n\t" -#if !defined(CONJ) - "vfmaddpd (%3,%0,8), %%xmm0 , %%xmm5, %%xmm12 \n\t" - "vpermilpd $0x1 , %%xmm5 , %%xmm4 \n\t" // exchange real and imag part - "vmulpd %%xmm1, %%xmm4 , %%xmm4 \n\t" + "vmovups 64(%2,%0,8), %%xmm12 \n\t" // 1 complex values from x + "vmovups 80(%2,%0,8), %%xmm13 \n\t" // 1 complex values from x + "vmovups 96(%2,%0,8), %%xmm14 \n\t" // 1 complex values from x + "vmovups 112(%2,%0,8), %%xmm15 \n\t" // 1 complex values from x - "vfmaddpd 16(%3,%0,8), %%xmm0 , %%xmm7, %%xmm13 \n\t" + "vpermilpd $0x1 , %%xmm5 , %%xmm4 \n\t" // exchange real and imag part "vpermilpd $0x1 , %%xmm7 , %%xmm6 \n\t" // exchange real and imag part - "vmulpd %%xmm1, %%xmm6 , %%xmm6 \n\t" - - "vfmaddpd 32(%3,%0,8), %%xmm0 , %%xmm9, %%xmm14 \n\t" "vpermilpd $0x1 , %%xmm9 , %%xmm8 \n\t" // exchange real and imag part - "vmulpd %%xmm1, %%xmm8 , %%xmm8 \n\t" - - "vfmaddpd 48(%3,%0,8), %%xmm0 , %%xmm11,%%xmm15 \n\t" "vpermilpd $0x1 , %%xmm11, %%xmm10 \n\t" // exchange real and imag part - "vmulpd %%xmm1, %%xmm10, %%xmm10 \n\t" - - "vaddsubpd %%xmm4, %%xmm12, %%xmm12 \n\t" - "vaddsubpd %%xmm6, %%xmm13, %%xmm13 \n\t" - "vaddsubpd %%xmm8, %%xmm14, %%xmm14 \n\t" - "vaddsubpd %%xmm10,%%xmm15, %%xmm15 \n\t" -#else - - "vmulpd %%xmm0, %%xmm5, %%xmm4 \n\t" // a_r*x_r, a_r*x_i - "vmulpd %%xmm1, %%xmm5, %%xmm5 \n\t" // a_i*x_r, a_i*x_i - "vmulpd %%xmm0, %%xmm7, %%xmm6 \n\t" // a_r*x_r, a_r*x_i - "vmulpd %%xmm1, %%xmm7, %%xmm7 \n\t" // a_i*x_r, a_i*x_i - "vmulpd %%xmm0, %%xmm9, %%xmm8 \n\t" // a_r*x_r, a_r*x_i - "vmulpd %%xmm1, %%xmm9, %%xmm9 \n\t" // a_i*x_r, a_i*x_i - "vmulpd %%xmm0, %%xmm11, %%xmm10 \n\t" // a_r*x_r, a_r*x_i - "vmulpd %%xmm1, %%xmm11, %%xmm11 \n\t" // a_i*x_r, a_i*x_i + "vfmaddpd (%3,%0,8), %%xmm0 , %%xmm5, %%xmm5 \n\t" + ".align 2 \n\t" + "vfmaddpd 16(%3,%0,8), %%xmm0 , %%xmm7, %%xmm7 \n\t" + "vfmaddpd 32(%3,%0,8), %%xmm0 , %%xmm9, %%xmm9 \n\t" + "vfmaddpd 48(%3,%0,8), %%xmm0 , %%xmm11,%%xmm11 \n\t" + + "vfmaddpd %%xmm5 , %%xmm1 , %%xmm4 , %%xmm5 \n\t" + "vfmaddpd %%xmm7 , %%xmm1 , %%xmm6 , %%xmm7 \n\t" + "vfmaddpd %%xmm9 , %%xmm1 , %%xmm8 , %%xmm9 \n\t" + "vfmaddpd %%xmm11, %%xmm1 , %%xmm10, %%xmm11 \n\t" + + "vpermilpd $0x1 , %%xmm12, %%xmm4 \n\t" // exchange real and imag part + "vpermilpd $0x1 , %%xmm13, %%xmm6 \n\t" // exchange real and imag part + "vpermilpd $0x1 , %%xmm14, %%xmm8 \n\t" // exchange real and imag part + "vpermilpd $0x1 , %%xmm15, %%xmm10 \n\t" // exchange real and imag part + + "vfmaddpd 64(%3,%0,8), %%xmm0 , %%xmm12, %%xmm12 \n\t" + "vfmaddpd 80(%3,%0,8), %%xmm0 , %%xmm13, %%xmm13 \n\t" + "vfmaddpd 96(%3,%0,8), %%xmm0 , %%xmm14, %%xmm14 \n\t" + "vfmaddpd 112(%3,%0,8), %%xmm0 , %%xmm15, %%xmm15 \n\t" + + "vfmaddpd %%xmm12, %%xmm1 , %%xmm4 , %%xmm12 \n\t" + "vfmaddpd %%xmm13, %%xmm1 , %%xmm6 , %%xmm13 \n\t" + "vfmaddpd %%xmm14, %%xmm1 , %%xmm8 , %%xmm14 \n\t" + "vfmaddpd %%xmm15, %%xmm1 , %%xmm10, %%xmm15 \n\t" + + "vmovups %%xmm5 , (%3,%0,8) \n\t" + ".align 2 \n\t" + "vmovups %%xmm7 , 16(%3,%0,8) \n\t" + "vmovups %%xmm9 , 32(%3,%0,8) \n\t" + "vmovups %%xmm11, 48(%3,%0,8) \n\t" + "vmovups %%xmm12, 64(%3,%0,8) \n\t" + "vmovups %%xmm13, 80(%3,%0,8) \n\t" + "vmovups %%xmm14, 96(%3,%0,8) \n\t" + "vmovups %%xmm15,112(%3,%0,8) \n\t" + + "addq $16, %0 \n\t" + "subq $8 , %1 \n\t" + "jnz 1b \n\t" + "vzeroupper \n\t" - "vpermilpd $0x1 , %%xmm4 , %%xmm4 \n\t" // exchange real and imag part - "vaddsubpd %%xmm4 ,%%xmm5 , %%xmm4 \n\t" - "vpermilpd $0x1 , %%xmm4 , %%xmm4 \n\t" // exchange real and imag part + : + : + "r" (i), // 0 + "r" (n), // 1 + "r" (x), // 2 + "r" (y), // 3 + "r" (alpha), // 4 + "r" (mvec) // 5 + : "cc", + "%xmm0", "%xmm1", + "%xmm4", "%xmm5", "%xmm6", "%xmm7", + "%xmm8", "%xmm9", "%xmm10", "%xmm11", + "%xmm12", "%xmm13", "%xmm14", "%xmm15", + "memory" + ); + return; + } - "vpermilpd $0x1 , %%xmm6 , %%xmm6 \n\t" // exchange real and imag part - "vaddsubpd %%xmm6 ,%%xmm7 , %%xmm6 \n\t" - "vpermilpd $0x1 , %%xmm6 , %%xmm6 \n\t" // exchange real and imag part + __asm__ __volatile__ + ( + "vzeroupper \n\t" + "vmovddup (%4), %%xmm0 \n\t" // real part of alpha + "vmovddup 8(%4), %%xmm1 \n\t" // imag part of alpha +#if !defined(CONJ) + "vmulpd (%5), %%xmm1 , %%xmm1 \n\t" +#else + "vmulpd (%5), %%xmm0 , %%xmm0 \n\t" +#endif - "vpermilpd $0x1 , %%xmm8 , %%xmm8 \n\t" // exchange real and imag part - "vaddsubpd %%xmm8 ,%%xmm9 , %%xmm8 \n\t" - "vpermilpd $0x1 , %%xmm8 , %%xmm8 \n\t" // exchange real and imag part + ".align 16 \n\t" + "1: \n\t" - "vpermilpd $0x1 , %%xmm10, %%xmm10 \n\t" // exchange real and imag part - "vaddsubpd %%xmm10,%%xmm11, %%xmm10 \n\t" - "vpermilpd $0x1 , %%xmm10, %%xmm10 \n\t" // exchange real and imag part + "prefetcht0 512(%2,%0,8) \n\t" + "vmovups (%2,%0,8), %%xmm5 \n\t" // 1 complex values from x + ".align 2 \n\t" + "vmovups 16(%2,%0,8), %%xmm7 \n\t" // 1 complex values from x + "vmovups 32(%2,%0,8), %%xmm9 \n\t" // 1 complex values from x + "vmovups 48(%2,%0,8), %%xmm11 \n\t" // 1 complex values from x - "vaddpd (%3,%0,8) ,%%xmm4 , %%xmm12 \n\t" - "vaddpd 16(%3,%0,8) ,%%xmm6 , %%xmm13 \n\t" - "vaddpd 32(%3,%0,8) ,%%xmm8 , %%xmm14 \n\t" - "vaddpd 48(%3,%0,8) ,%%xmm10, %%xmm15 \n\t" + "vpermilpd $0x1 , %%xmm5 , %%xmm4 \n\t" // exchange real and imag part + "vpermilpd $0x1 , %%xmm7 , %%xmm6 \n\t" // exchange real and imag part + "vpermilpd $0x1 , %%xmm9 , %%xmm8 \n\t" // exchange real and imag part + "vpermilpd $0x1 , %%xmm11, %%xmm10 \n\t" // exchange real and imag part + "prefetcht0 512(%3,%0,8) \n\t" + "vfmaddpd (%3,%0,8), %%xmm0 , %%xmm5, %%xmm5 \n\t" + ".align 2 \n\t" + "vfmaddpd 16(%3,%0,8), %%xmm0 , %%xmm7, %%xmm7 \n\t" + "vfmaddpd 32(%3,%0,8), %%xmm0 , %%xmm9, %%xmm9 \n\t" + "vfmaddpd 48(%3,%0,8), %%xmm0 , %%xmm11,%%xmm11 \n\t" -#endif + "vfmaddpd %%xmm5 , %%xmm1 , %%xmm4 , %%xmm5 \n\t" + "vfmaddpd %%xmm7 , %%xmm1 , %%xmm6 , %%xmm7 \n\t" + "vfmaddpd %%xmm9 , %%xmm1 , %%xmm8 , %%xmm9 \n\t" + "vfmaddpd %%xmm11, %%xmm1 , %%xmm10, %%xmm11 \n\t" - "vmovups %%xmm12, (%3,%0,8) \n\t" - "vmovups %%xmm13, 16(%3,%0,8) \n\t" - "vmovups %%xmm14, 32(%3,%0,8) \n\t" - "vmovups %%xmm15, 48(%3,%0,8) \n\t" + "vmovups %%xmm5 , (%3,%0,8) \n\t" + ".align 2 \n\t" + "vmovups %%xmm7 , 16(%3,%0,8) \n\t" + "vmovups %%xmm9 , 32(%3,%0,8) \n\t" + "vmovups %%xmm11, 48(%3,%0,8) \n\t" "addq $8 , %0 \n\t" - "subq $4 , %1 \n\t" + "subq $4, %1 \n\t" "jnz 1b \n\t" + "vzeroupper \n\t" : : @@ -121,15 +186,15 @@ static void zaxpy_kernel_4( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha) "r" (n), // 1 "r" (x), // 2 "r" (y), // 3 - "r" (alpha) // 4 + "r" (alpha), // 4 + "r" (mvec) // 5 : "cc", "%xmm0", "%xmm1", "%xmm4", "%xmm5", "%xmm6", "%xmm7", "%xmm8", "%xmm9", "%xmm10", "%xmm11", - "%xmm12", "%xmm13", "%xmm14", "%xmm15", "memory" ); -} +} diff --git a/kernel/x86_64/zaxpy_microk_haswell-2.c b/kernel/x86_64/zaxpy_microk_haswell-2.c new file mode 100644 index 0000000..e7e5595 --- /dev/null +++ b/kernel/x86_64/zaxpy_microk_haswell-2.c @@ -0,0 +1,132 @@ +/*************************************************************************** +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_4 1 +static void zaxpy_kernel_4( BLASLONG n, FLOAT *x, FLOAT *y , FLOAT *alpha) __attribute__ ((noinline)); + +static void zaxpy_kernel_4( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha) +{ + +#if !defined(CONJ) + FLOAT mvec[4] = { -1.0, 1.0, -1.0, 1.0 }; +#else + FLOAT mvec[4] = { 1.0, -1.0, 1.0, -1.0 }; +#endif + + BLASLONG register i = 0; + + __asm__ __volatile__ + ( + "vzeroupper \n\t" + "vbroadcastsd (%4), %%ymm0 \n\t" // real part of alpha + "vbroadcastsd 8(%4), %%ymm1 \n\t" // imag part of alpha +#if !defined(CONJ) + "vmulpd (%5), %%ymm1 , %%ymm1 \n\t" +#else + "vmulpd (%5), %%ymm0 , %%ymm0 \n\t" +#endif + + ".align 16 \n\t" + "1: \n\t" + + "vmovups (%2,%0,8), %%ymm5 \n\t" // 2 complex values from x + ".align 2 \n\t" + "vmovups 32(%2,%0,8), %%ymm7 \n\t" // 2 complex values from x + "vmovups 64(%2,%0,8), %%ymm9 \n\t" // 2 complex values from x + "vmovups 96(%2,%0,8), %%ymm11 \n\t" // 2 complex values from x + + "vmovups 128(%2,%0,8), %%ymm12 \n\t" // 2 complex values from x + "vmovups 160(%2,%0,8), %%ymm13 \n\t" // 2 complex values from x + "vmovups 192(%2,%0,8), %%ymm14 \n\t" // 2 complex values from x + "vmovups 224(%2,%0,8), %%ymm15 \n\t" // 2 complex values from x + + "vpermilpd $0x5 , %%ymm5 , %%ymm4 \n\t" // exchange real and imag part + "vpermilpd $0x5 , %%ymm7 , %%ymm6 \n\t" // exchange real and imag part + "vpermilpd $0x5 , %%ymm9 , %%ymm8 \n\t" // exchange real and imag part + "vpermilpd $0x5 , %%ymm11, %%ymm10 \n\t" // exchange real and imag part + + "vfmadd213pd (%3,%0,8), %%ymm0 , %%ymm5 \n\t" + ".align 2 \n\t" + "vfmadd213pd 32(%3,%0,8), %%ymm0 , %%ymm7 \n\t" + "vfmadd213pd 64(%3,%0,8), %%ymm0 , %%ymm9 \n\t" + "vfmadd213pd 96(%3,%0,8), %%ymm0 , %%ymm11 \n\t" + + "vfmadd231pd %%ymm1 , %%ymm4 , %%ymm5 \n\t" + "vfmadd231pd %%ymm1 , %%ymm6 , %%ymm7 \n\t" + "vfmadd231pd %%ymm1 , %%ymm8 , %%ymm9 \n\t" + "vfmadd231pd %%ymm1 , %%ymm10, %%ymm11 \n\t" + + "vpermilpd $0x5 , %%ymm12, %%ymm4 \n\t" // exchange real and imag part + "vpermilpd $0x5 , %%ymm13, %%ymm6 \n\t" // exchange real and imag part + "vpermilpd $0x5 , %%ymm14, %%ymm8 \n\t" // exchange real and imag part + "vpermilpd $0x5 , %%ymm15, %%ymm10 \n\t" // exchange real and imag part + + "vfmadd213pd 128(%3,%0,8), %%ymm0 , %%ymm12 \n\t" + "vfmadd213pd 160(%3,%0,8), %%ymm0 , %%ymm13 \n\t" + "vfmadd213pd 192(%3,%0,8), %%ymm0 , %%ymm14 \n\t" + "vfmadd213pd 224(%3,%0,8), %%ymm0 , %%ymm15 \n\t" + + "vfmadd231pd %%ymm1 , %%ymm4 , %%ymm12 \n\t" + "vfmadd231pd %%ymm1 , %%ymm6 , %%ymm13 \n\t" + "vfmadd231pd %%ymm1 , %%ymm8 , %%ymm14 \n\t" + "vfmadd231pd %%ymm1 , %%ymm10, %%ymm15 \n\t" + + "vmovups %%ymm5 , (%3,%0,8) \n\t" + ".align 2 \n\t" + "vmovups %%ymm7 , 32(%3,%0,8) \n\t" + "vmovups %%ymm9 , 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" + "subq $16, %1 \n\t" + "jnz 1b \n\t" + "vzeroupper \n\t" + + : + : + "r" (i), // 0 + "r" (n), // 1 + "r" (x), // 2 + "r" (y), // 3 + "r" (alpha), // 4 + "r" (mvec) // 5 + : "cc", + "%xmm0", "%xmm1", + "%xmm4", "%xmm5", "%xmm6", "%xmm7", + "%xmm8", "%xmm9", "%xmm10", "%xmm11", + "%xmm12", "%xmm13", "%xmm14", "%xmm15", + "memory" + ); + + +} + diff --git a/kernel/x86_64/zaxpy_microk_sandy-2.c b/kernel/x86_64/zaxpy_microk_sandy-2.c new file mode 100644 index 0000000..8b0a7ed --- /dev/null +++ b/kernel/x86_64/zaxpy_microk_sandy-2.c @@ -0,0 +1,198 @@ +/*************************************************************************** +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_4 1 +static void zaxpy_kernel_4( BLASLONG n, FLOAT *x, FLOAT *y , FLOAT *alpha) __attribute__ ((noinline)); + +static void zaxpy_kernel_4( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha) +{ + +#if !defined(CONJ) + FLOAT mvec[4] = { -1.0, 1.0, -1.0, 1.0 }; +#else + FLOAT mvec[4] = { 1.0, -1.0, 1.0, -1.0 }; +#endif + + BLASLONG register i = 0; + + if ( n < 1280 ) + { + + + __asm__ __volatile__ + ( + "vzeroupper \n\t" + "vbroadcastsd (%4), %%ymm0 \n\t" // real part of alpha + "vbroadcastsd 8(%4), %%ymm1 \n\t" // imag part of alpha +#if !defined(CONJ) + "vmulpd (%5), %%ymm1 , %%ymm1 \n\t" +#else + "vmulpd (%5), %%ymm0 , %%ymm0 \n\t" +#endif + + ".align 16 \n\t" + "1: \n\t" + + "vmovups (%2,%0,8), %%ymm5 \n\t" // 4 complex values from x + ".align 2 \n\t" + "vmovups 32(%2,%0,8), %%ymm7 \n\t" // 4 complex values from x + "vmovups 64(%2,%0,8), %%ymm9 \n\t" // 4 complex values from x + "vmovups 96(%2,%0,8), %%ymm11 \n\t" // 4 complex values from x + + "vpermilpd $0x5 , %%ymm5 , %%ymm4 \n\t" // exchange real and imag part + "vpermilpd $0x5 , %%ymm7 , %%ymm6 \n\t" // exchange real and imag part + "vpermilpd $0x5 , %%ymm9 , %%ymm8 \n\t" // exchange real and imag part + "vpermilpd $0x5 , %%ymm11, %%ymm10 \n\t" // exchange real and imag part + + "vmulpd %%ymm5 , %%ymm0 , %%ymm5 \n\t" + "vmulpd %%ymm7 , %%ymm0 , %%ymm7 \n\t" + "vmulpd %%ymm9 , %%ymm0 , %%ymm9 \n\t" + "vmulpd %%ymm11, %%ymm0 , %%ymm11 \n\t" + + "vaddpd (%3,%0,8), %%ymm5 , %%ymm5 \n\t" + "vaddpd 32(%3,%0,8), %%ymm7 , %%ymm7 \n\t" + "vaddpd 64(%3,%0,8), %%ymm9 , %%ymm9 \n\t" + "vaddpd 96(%3,%0,8), %%ymm11, %%ymm11 \n\t" + + "vmulpd %%ymm4 , %%ymm1 , %%ymm4 \n\t" + "vmulpd %%ymm6 , %%ymm1 , %%ymm6 \n\t" + "vmulpd %%ymm8 , %%ymm1 , %%ymm8 \n\t" + "vmulpd %%ymm10, %%ymm1 , %%ymm10 \n\t" + + "vaddpd %%ymm4 , %%ymm5 , %%ymm5 \n\t" + "vaddpd %%ymm6 , %%ymm7 , %%ymm7 \n\t" + "vaddpd %%ymm8 , %%ymm9 , %%ymm9 \n\t" + "vaddpd %%ymm10, %%ymm11, %%ymm11 \n\t" + + "vmovups %%ymm5 , (%3,%0,8) \n\t" + ".align 2 \n\t" + "vmovups %%ymm7 , 32(%3,%0,8) \n\t" + "vmovups %%ymm9 , 64(%3,%0,8) \n\t" + "vmovups %%ymm11, 96(%3,%0,8) \n\t" + + "addq $16, %0 \n\t" + "subq $8 , %1 \n\t" + "jnz 1b \n\t" + "vzeroupper \n\t" + + : + : + "r" (i), // 0 + "r" (n), // 1 + "r" (x), // 2 + "r" (y), // 3 + "r" (alpha), // 4 + "r" (mvec) // 5 + : "cc", + "%xmm0", "%xmm1", + "%xmm4", "%xmm5", "%xmm6", "%xmm7", + "%xmm8", "%xmm9", "%xmm10", "%xmm11", + "memory" + ); + return; + } + + __asm__ __volatile__ + ( + "vzeroupper \n\t" + "vbroadcastsd (%4), %%ymm0 \n\t" // real part of alpha + "vbroadcastsd 8(%4), %%ymm1 \n\t" // imag part of alpha +#if !defined(CONJ) + "vmulpd (%5), %%ymm1 , %%ymm1 \n\t" +#else + "vmulpd (%5), %%ymm0 , %%ymm0 \n\t" +#endif + + ".align 16 \n\t" + "1: \n\t" + + "prefetcht0 512(%2,%0,8) \n\t" + "prefetcht0 576(%2,%0,8) \n\t" + "vmovups (%2,%0,8), %%ymm5 \n\t" // 4 complex values from x + ".align 2 \n\t" + "vmovups 32(%2,%0,8), %%ymm7 \n\t" // 4 complex values from x + "vmovups 64(%2,%0,8), %%ymm9 \n\t" // 4 complex values from x + "vmovups 96(%2,%0,8), %%ymm11 \n\t" // 4 complex values from x + + "vpermilpd $0x5 , %%ymm5 , %%ymm4 \n\t" // exchange real and imag part + "vpermilpd $0x5 , %%ymm7 , %%ymm6 \n\t" // exchange real and imag part + "vpermilpd $0x5 , %%ymm9 , %%ymm8 \n\t" // exchange real and imag part + "vpermilpd $0x5 , %%ymm11, %%ymm10 \n\t" // exchange real and imag part + + "vmulpd %%ymm5 , %%ymm0 , %%ymm5 \n\t" + "vmulpd %%ymm7 , %%ymm0 , %%ymm7 \n\t" + "vmulpd %%ymm9 , %%ymm0 , %%ymm9 \n\t" + "vmulpd %%ymm11, %%ymm0 , %%ymm11 \n\t" + + "prefetcht0 512(%3,%0,8) \n\t" + "prefetcht0 576(%3,%0,8) \n\t" + "vaddpd (%3,%0,8), %%ymm5 , %%ymm5 \n\t" + "vaddpd 32(%3,%0,8), %%ymm7 , %%ymm7 \n\t" + "vaddpd 64(%3,%0,8), %%ymm9 , %%ymm9 \n\t" + "vaddpd 96(%3,%0,8), %%ymm11, %%ymm11 \n\t" + + "vmulpd %%ymm4 , %%ymm1 , %%ymm4 \n\t" + "vmulpd %%ymm6 , %%ymm1 , %%ymm6 \n\t" + "vmulpd %%ymm8 , %%ymm1 , %%ymm8 \n\t" + "vmulpd %%ymm10, %%ymm1 , %%ymm10 \n\t" + + "vaddpd %%ymm4 , %%ymm5 , %%ymm5 \n\t" + "vaddpd %%ymm6 , %%ymm7 , %%ymm7 \n\t" + "vaddpd %%ymm8 , %%ymm9 , %%ymm9 \n\t" + "vaddpd %%ymm10, %%ymm11, %%ymm11 \n\t" + + "vmovups %%ymm5 , (%3,%0,8) \n\t" + ".align 2 \n\t" + "vmovups %%ymm7 , 32(%3,%0,8) \n\t" + "vmovups %%ymm9 , 64(%3,%0,8) \n\t" + "vmovups %%ymm11, 96(%3,%0,8) \n\t" + + "addq $16, %0 \n\t" + "subq $8 , %1 \n\t" + "jnz 1b \n\t" + "vzeroupper \n\t" + + : + : + "r" (i), // 0 + "r" (n), // 1 + "r" (x), // 2 + "r" (y), // 3 + "r" (alpha), // 4 + "r" (mvec) // 5 + : "cc", + "%xmm0", "%xmm1", + "%xmm4", "%xmm5", "%xmm6", "%xmm7", + "%xmm8", "%xmm9", "%xmm10", "%xmm11", + "memory" + ); + + + + +} + diff --git a/kernel/x86_64/zaxpy_microk_steamroller-2.c b/kernel/x86_64/zaxpy_microk_steamroller-2.c new file mode 100644 index 0000000..728d092 --- /dev/null +++ b/kernel/x86_64/zaxpy_microk_steamroller-2.c @@ -0,0 +1,200 @@ +/*************************************************************************** +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_4 1 +static void zaxpy_kernel_4( BLASLONG n, FLOAT *x, FLOAT *y , FLOAT *alpha) __attribute__ ((noinline)); + +static void zaxpy_kernel_4( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha) +{ + +#if !defined(CONJ) + FLOAT mvec[2] = { -1.0, 1.0 }; +#else + FLOAT mvec[2] = { 1.0, -1.0 }; +#endif + + BLASLONG register i = 0; + + if ( n < 640 ) + { + + __asm__ __volatile__ + ( + "vzeroupper \n\t" + "vmovddup (%4), %%xmm0 \n\t" // real part of alpha + "vmovddup 8(%4), %%xmm1 \n\t" // imag part of alpha +#if !defined(CONJ) + "vmulpd (%5), %%xmm1 , %%xmm1 \n\t" +#else + "vmulpd (%5), %%xmm0 , %%xmm0 \n\t" +#endif + + ".align 16 \n\t" + "1: \n\t" + + "vmovups (%2,%0,8), %%xmm5 \n\t" // 2 complex values from x + ".align 2 \n\t" + "vmovups 16(%2,%0,8), %%xmm7 \n\t" // 2 complex values from x + "vmovups 32(%2,%0,8), %%xmm9 \n\t" // 2 complex values from x + "vmovups 48(%2,%0,8), %%xmm11 \n\t" // 2 complex values from x + + "vmovups 64(%2,%0,8), %%xmm12 \n\t" // 2 complex values from x + "vmovups 80(%2,%0,8), %%xmm13 \n\t" // 2 complex values from x + "vmovups 96(%2,%0,8), %%xmm14 \n\t" // 2 complex values from x + "vmovups 112(%2,%0,8), %%xmm15 \n\t" // 2 complex values from x + + "vpermilpd $0x1 , %%xmm5 , %%xmm4 \n\t" // exchange real and imag part + "vpermilpd $0x1 , %%xmm7 , %%xmm6 \n\t" // exchange real and imag part + "vpermilpd $0x1 , %%xmm9 , %%xmm8 \n\t" // exchange real and imag part + "vpermilpd $0x1 , %%xmm11, %%xmm10 \n\t" // exchange real and imag part + + "vfmadd213pd (%3,%0,8), %%xmm0 , %%xmm5 \n\t" + ".align 2 \n\t" + "vfmadd213pd 16(%3,%0,8), %%xmm0 , %%xmm7 \n\t" + "vfmadd213pd 32(%3,%0,8), %%xmm0 , %%xmm9 \n\t" + "vfmadd213pd 48(%3,%0,8), %%xmm0 , %%xmm11 \n\t" + + "vfmadd231pd %%xmm1 , %%xmm4 , %%xmm5 \n\t" + "vfmadd231pd %%xmm1 , %%xmm6 , %%xmm7 \n\t" + "vfmadd231pd %%xmm1 , %%xmm8 , %%xmm9 \n\t" + "vfmadd231pd %%xmm1 , %%xmm10, %%xmm11 \n\t" + + "vpermilpd $0x1 , %%xmm12, %%xmm4 \n\t" // exchange real and imag part + "vpermilpd $0x1 , %%xmm13, %%xmm6 \n\t" // exchange real and imag part + "vpermilpd $0x1 , %%xmm14, %%xmm8 \n\t" // exchange real and imag part + "vpermilpd $0x1 , %%xmm15, %%xmm10 \n\t" // exchange real and imag part + + "vfmadd213pd 64(%3,%0,8), %%xmm0 , %%xmm12 \n\t" + "vfmadd213pd 80(%3,%0,8), %%xmm0 , %%xmm13 \n\t" + "vfmadd213pd 96(%3,%0,8), %%xmm0 , %%xmm14 \n\t" + "vfmadd213pd 112(%3,%0,8), %%xmm0 , %%xmm15 \n\t" + + "vfmadd231pd %%xmm1 , %%xmm4 , %%xmm12 \n\t" + "vfmadd231pd %%xmm1 , %%xmm6 , %%xmm13 \n\t" + "vfmadd231pd %%xmm1 , %%xmm8 , %%xmm14 \n\t" + "vfmadd231pd %%xmm1 , %%xmm10, %%xmm15 \n\t" + + "vmovups %%xmm5 , (%3,%0,8) \n\t" + ".align 2 \n\t" + "vmovups %%xmm7 , 16(%3,%0,8) \n\t" + "vmovups %%xmm9 , 32(%3,%0,8) \n\t" + "vmovups %%xmm11, 48(%3,%0,8) \n\t" + "vmovups %%xmm12, 64(%3,%0,8) \n\t" + "vmovups %%xmm13, 80(%3,%0,8) \n\t" + "vmovups %%xmm14, 96(%3,%0,8) \n\t" + "vmovups %%xmm15,112(%3,%0,8) \n\t" + + "addq $16, %0 \n\t" + "subq $8 , %1 \n\t" + "jnz 1b \n\t" + "vzeroupper \n\t" + + : + : + "r" (i), // 0 + "r" (n), // 1 + "r" (x), // 2 + "r" (y), // 3 + "r" (alpha), // 4 + "r" (mvec) // 5 + : "cc", + "%xmm0", "%xmm1", + "%xmm4", "%xmm5", "%xmm6", "%xmm7", + "%xmm8", "%xmm9", "%xmm10", "%xmm11", + "%xmm12", "%xmm13", "%xmm14", "%xmm15", + "memory" + ); + return; + } + + __asm__ __volatile__ + ( + "vzeroupper \n\t" + "vmovddup (%4), %%xmm0 \n\t" // real part of alpha + "vmovddup 8(%4), %%xmm1 \n\t" // imag part of alpha +#if !defined(CONJ) + "vmulpd (%5), %%xmm1 , %%xmm1 \n\t" +#else + "vmulpd (%5), %%xmm0 , %%xmm0 \n\t" +#endif + + ".align 16 \n\t" + "1: \n\t" + + "prefetcht0 512(%2,%0,8) \n\t" + "vmovups (%2,%0,8), %%xmm5 \n\t" // 2 complex values from x + ".align 2 \n\t" + "vmovups 16(%2,%0,8), %%xmm7 \n\t" // 2 complex values from x + "vmovups 32(%2,%0,8), %%xmm9 \n\t" // 2 complex values from x + "vmovups 48(%2,%0,8), %%xmm11 \n\t" // 2 complex values from x + + "vpermilpd $0x1 , %%xmm5 , %%xmm4 \n\t" // exchange real and imag part + "vpermilpd $0x1 , %%xmm7 , %%xmm6 \n\t" // exchange real and imag part + "vpermilpd $0x1 , %%xmm9 , %%xmm8 \n\t" // exchange real and imag part + "vpermilpd $0x1 , %%xmm11, %%xmm10 \n\t" // exchange real and imag part + + "prefetcht0 512(%3,%0,8) \n\t" + "vfmadd213pd (%3,%0,8), %%xmm0 , %%xmm5 \n\t" + ".align 2 \n\t" + "vfmadd213pd 16(%3,%0,8), %%xmm0 , %%xmm7 \n\t" + "vfmadd213pd 32(%3,%0,8), %%xmm0 , %%xmm9 \n\t" + "vfmadd213pd 48(%3,%0,8), %%xmm0 , %%xmm11 \n\t" + + "vfmadd231pd %%xmm1 , %%xmm4 , %%xmm5 \n\t" + "vfmadd231pd %%xmm1 , %%xmm6 , %%xmm7 \n\t" + "vfmadd231pd %%xmm1 , %%xmm8 , %%xmm9 \n\t" + "vfmadd231pd %%xmm1 , %%xmm10, %%xmm11 \n\t" + + "vmovups %%xmm5 , (%3,%0,8) \n\t" + ".align 2 \n\t" + "vmovups %%xmm7 , 16(%3,%0,8) \n\t" + "vmovups %%xmm9 , 32(%3,%0,8) \n\t" + "vmovups %%xmm11, 48(%3,%0,8) \n\t" + + "addq $8 , %0 \n\t" + "subq $4 , %1 \n\t" + "jnz 1b \n\t" + "vzeroupper \n\t" + + : + : + "r" (i), // 0 + "r" (n), // 1 + "r" (x), // 2 + "r" (y), // 3 + "r" (alpha), // 4 + "r" (mvec) // 5 + : "cc", + "%xmm0", "%xmm1", + "%xmm4", "%xmm5", "%xmm6", "%xmm7", + "%xmm8", "%xmm9", "%xmm10", "%xmm11", + "memory" + ); + + +} + diff --git a/lapack-netlib/TESTING/sep.in b/lapack-netlib/TESTING/sep.in index c71a754..6ef4329 100644 --- a/lapack-netlib/TESTING/sep.in +++ b/lapack-netlib/TESTING/sep.in @@ -1,11 +1,11 @@ SEP: Data file for testing Symmetric Eigenvalue Problem routines 8 Number of values of N -0 1 2 3 5 19 20 21 Values of N (dimension) +0 1 2 3 5 18 19 21 Values of N (dimension) 5 Number of values of NB 1 3 3 3 10 Values of NB (blocksize) 2 2 2 2 2 Values of NBMIN (minimum blocksize) 1 0 5 9 1 Values of NX (crossover point) -170.0 Threshold value +300.0 Threshold value T Put T to test the LAPACK routines T Put T to test the driver routines T Put T to test the error exits diff --git a/lapack-netlib/TESTING/zctest.in b/lapack-netlib/TESTING/zctest.in index ef88cc0..48e88ec 100644 --- a/lapack-netlib/TESTING/zctest.in +++ b/lapack-netlib/TESTING/zctest.in @@ -1,6 +1,6 @@ Data file for testing ZCGESV/ZCPOSV LAPACK routines 11 Number of values of M -0 1 2 13 17 45 78 91 101 120 132 Values of M (row dimension) +0 1 2 13 17 45 78 91 101 121 132 Values of M (row dimension) 4 Number of values of NRHS 1 2 15 16 Values of NRHS (number of right hand sides) 30.0 Threshold value of test ratio -- 2.7.4