From: wernsaar Date: Fri, 22 Nov 2013 16:21:10 +0000 (+0100) Subject: added nrm2 kernel for ARMV6 X-Git-Tag: v0.2.9.rc1~7^2~6^2~32 X-Git-Url: http://review.tizen.org/git/?a=commitdiff_plain;h=9a0f9789296e764c5fe92c614c7cc9d655061bb3;p=platform%2Fupstream%2Fopenblas.git added nrm2 kernel for ARMV6 --- diff --git a/kernel/arm/KERNEL.ARMV6 b/kernel/arm/KERNEL.ARMV6 index 17a4f92..03eff6e 100644 --- a/kernel/arm/KERNEL.ARMV6 +++ b/kernel/arm/KERNEL.ARMV6 @@ -50,11 +50,10 @@ DDOTKERNEL = ddot_vfp.S CDOTKERNEL = cdot_vfp.S ZDOTKERNEL = zdot_vfp.S - -SNRM2KERNEL = nrm2.c -DNRM2KERNEL = nrm2.c -CNRM2KERNEL = znrm2.c -ZNRM2KERNEL = znrm2.c +SNRM2KERNEL = nrm2_vfp.S +DNRM2KERNEL = nrm2_vfp.S +CNRM2KERNEL = nrm2_vfp.S +ZNRM2KERNEL = nrm2_vfp.S SROTKERNEL = rot_vfp.S DROTKERNEL = rot_vfp.S diff --git a/kernel/arm/nrm2_vfp.S b/kernel/arm/nrm2_vfp.S new file mode 100644 index 0000000..4c62917 --- /dev/null +++ b/kernel/arm/nrm2_vfp.S @@ -0,0 +1,565 @@ +/*************************************************************************** +Copyright (c) 2013, 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. +*****************************************************************************/ + +/************************************************************************************** +* 2013/11/22 Saar +* BLASTEST : OK +* CTEST : OK +* TEST : OK +* +**************************************************************************************/ + +#define ASSEMBLER +#include "common.h" + +#define STACKSIZE 256 + +#define N r0 +#define X r1 +#define INC_X r2 + +#define I r12 + +#define X_PRE 512 + +/************************************************************************************** +* Macro definitions +**************************************************************************************/ + + +#if !defined(COMPLEX) + +#if defined(DOUBLE) + + +.macro KERNEL_F1 + + fldmiad X!, { d4 } + vcmpe.f64 d4, d6 // compare with 0.0 + vmrs APSR_nzcv, fpscr + beq KERNEL_F1_NEXT_\@ + vabs.f64 d4, d4 + vcmpe.f64 d0, d4 // compare with scale + vmrs APSR_nzcv, fpscr + vdivge.f64 d2 , d4, d0 // scale >= x ? x / scale + vmlage.f64 d1 , d2 , d2 // ssq += ( x/scale ) * ( x/scale ) + bge KERNEL_F1_NEXT_\@ + vdiv.f64 d2 , d0, d4 // scale / x + vmul.f64 d2 , d2, d2 // ( scale / x ) * ( scale / x ) + vmul.f64 d3 , d1, d2 // ssq * ( scale / x ) * ( scale / x ) + vadd.f64 d1 , d3, d7 // ssq = 1 + ssq * ( scale / x ) * ( scale / x ) + vmov.f64 d0 , d4 // scale = x + +KERNEL_F1_NEXT_\@: + +.endm + +.macro KERNEL_F8 + + pld [ X, #X_PRE ] + KERNEL_F1 + KERNEL_F1 + KERNEL_F1 + KERNEL_F1 + pld [ X, #X_PRE ] + KERNEL_F1 + KERNEL_F1 + KERNEL_F1 + KERNEL_F1 + +.endm + +.macro KERNEL_S1 + + fldmiad X, { d4 } + vcmpe.f64 d4, d6 // compare with 0.0 + vmrs APSR_nzcv, fpscr + beq KERNEL_S1_NEXT + vabs.f64 d4, d4 + vcmpe.f64 d0, d4 // compare with scale + vmrs APSR_nzcv, fpscr + vdivge.f64 d2 , d4, d0 // scale >= x ? x / scale + vmlage.f64 d1 , d2 , d2 // ssq += ( x/scale ) * ( x/scale ) + bge KERNEL_S1_NEXT + vdiv.f64 d2 , d0, d4 // scale / x + vmul.f64 d2 , d2, d2 // ( scale / x ) * ( scale / x ) + vmul.f64 d3 , d1, d2 // ssq * ( scale / x ) * ( scale / x ) + vadd.f64 d1 , d3, d7 // ssq = 1 + ssq * ( scale / x ) * ( scale / x ) + vmov.f64 d0 , d4 // scale = x + +KERNEL_S1_NEXT: + + add X, X, INC_X + +.endm + +#else + +.macro KERNEL_F1 + + fldmias X!, { s4 } + vcmpe.f32 s4, s6 // compare with 0.0 + vmrs APSR_nzcv, fpscr + beq KERNEL_F1_NEXT_\@ + vabs.f32 s4, s4 + vcmpe.f32 s0, s4 // compare with scale + vmrs APSR_nzcv, fpscr + vdivge.f32 s2 , s4, s0 // scale >= x ? x / scale + vmlage.f32 s1 , s2 , s2 // ssq += ( x/scale ) * ( x/scale ) + bge KERNEL_F1_NEXT_\@ + vdiv.f32 s2 , s0, s4 // scale / x + vmul.f32 s2 , s2, s2 // ( scale / x ) * ( scale / x ) + vmul.f32 s3 , s1, s2 // ssq * ( scale / x ) * ( scale / x ) + vadd.f32 s1 , s3, s7 // ssq = 1 + ssq * ( scale / x ) * ( scale / x ) + vmov.f32 s0 , s4 // scale = x + +KERNEL_F1_NEXT_\@: + +.endm + +.macro KERNEL_F8 + + pld [ X, #X_PRE ] + KERNEL_F1 + KERNEL_F1 + KERNEL_F1 + KERNEL_F1 + + KERNEL_F1 + KERNEL_F1 + KERNEL_F1 + KERNEL_F1 + +.endm + +.macro KERNEL_S1 + + fldmias X, { s4 } + vcmpe.f32 s4, s6 // compare with 0.0 + vmrs APSR_nzcv, fpscr + beq KERNEL_S1_NEXT + vabs.f32 s4, s4 + vcmpe.f32 s0, s4 // compare with scale + vmrs APSR_nzcv, fpscr + vdivge.f32 s2 , s4, s0 // scale >= x ? x / scale + vmlage.f32 s1 , s2 , s2 // ssq += ( x/scale ) * ( x/scale ) + bge KERNEL_S1_NEXT + vdiv.f32 s2 , s0, s4 // scale / x + vmul.f32 s2 , s2, s2 // ( scale / x ) * ( scale / x ) + vmul.f32 s3 , s1, s2 // ssq * ( scale / x ) * ( scale / x ) + vadd.f32 s1 , s3, s7 // ssq = 1 + ssq * ( scale / x ) * ( scale / x ) + vmov.f32 s0 , s4 // scale = x + +KERNEL_S1_NEXT: + + add X, X, INC_X + +.endm + + + + +#endif + +#else + +#if defined(DOUBLE) + +.macro KERNEL_F1 + + fldmiad X!, { d4 - d5 } + + vcmpe.f64 d4, d6 // compare with 0.0 + vmrs APSR_nzcv, fpscr + beq KERNEL_F1_NEXT_\@ + vabs.f64 d4, d4 + vcmpe.f64 d0, d4 // compare with scale + vmrs APSR_nzcv, fpscr + vdivge.f64 d2 , d4, d0 // scale >= x ? x / scale + vmlage.f64 d1 , d2 , d2 // ssq += ( x/scale ) * ( x/scale ) + bge KERNEL_F1_NEXT_\@ + vdiv.f64 d2 , d0, d4 // scale / x + vmul.f64 d2 , d2, d2 // ( scale / x ) * ( scale / x ) + vmul.f64 d3 , d1, d2 // ssq * ( scale / x ) * ( scale / x ) + vadd.f64 d1 , d3, d7 // ssq = 1 + ssq * ( scale / x ) * ( scale / x ) + vmov.f64 d0 , d4 // scale = x + +KERNEL_F1_NEXT_\@: + + vcmpe.f64 d5, d6 // compare with 0.0 + vmrs APSR_nzcv, fpscr + beq KERNEL_F1_END_\@ + vabs.f64 d5, d5 + vcmpe.f64 d0, d5 // compare with scale + vmrs APSR_nzcv, fpscr + vdivge.f64 d2 , d5, d0 // scale >= x ? x / scale + vmlage.f64 d1 , d2 , d2 // ssq += ( x/scale ) * ( x/scale ) + bge KERNEL_F1_END_\@ + vdiv.f64 d2 , d0, d5 // scale / x + vmul.f64 d2 , d2, d2 // ( scale / x ) * ( scale / x ) + vmul.f64 d3 , d1, d2 // ssq * ( scale / x ) * ( scale / x ) + vadd.f64 d1 , d3, d7 // ssq = 1 + ssq * ( scale / x ) * ( scale / x ) + vmov.f64 d0 , d5 // scale = x + +KERNEL_F1_END_\@: + + +.endm + +.macro KERNEL_F8 + + pld [ X, #X_PRE ] + KERNEL_F1 + KERNEL_F1 + pld [ X, #X_PRE ] + KERNEL_F1 + KERNEL_F1 + pld [ X, #X_PRE ] + KERNEL_F1 + KERNEL_F1 + pld [ X, #X_PRE ] + KERNEL_F1 + KERNEL_F1 + +.endm + +.macro KERNEL_S1 + + fldmiad X, { d4 - d5 } + + vcmpe.f64 d4, d6 // compare with 0.0 + vmrs APSR_nzcv, fpscr + beq KERNEL_S1_NEXT_\@ + vabs.f64 d4, d4 + vcmpe.f64 d0, d4 // compare with scale + vmrs APSR_nzcv, fpscr + vdivge.f64 d2 , d4, d0 // scale >= x ? x / scale + vmlage.f64 d1 , d2 , d2 // ssq += ( x/scale ) * ( x/scale ) + bge KERNEL_S1_NEXT_\@ + vdiv.f64 d2 , d0, d4 // scale / x + vmul.f64 d2 , d2, d2 // ( scale / x ) * ( scale / x ) + vmul.f64 d3 , d1, d2 // ssq * ( scale / x ) * ( scale / x ) + vadd.f64 d1 , d3, d7 // ssq = 1 + ssq * ( scale / x ) * ( scale / x ) + vmov.f64 d0 , d4 // scale = x + +KERNEL_S1_NEXT_\@: + + vcmpe.f64 d5, d6 // compare with 0.0 + vmrs APSR_nzcv, fpscr + beq KERNEL_S1_END_\@ + vabs.f64 d5, d5 + vcmpe.f64 d0, d5 // compare with scale + vmrs APSR_nzcv, fpscr + vdivge.f64 d2 , d5, d0 // scale >= x ? x / scale + vmlage.f64 d1 , d2 , d2 // ssq += ( x/scale ) * ( x/scale ) + bge KERNEL_S1_END_\@ + vdiv.f64 d2 , d0, d5 // scale / x + vmul.f64 d2 , d2, d2 // ( scale / x ) * ( scale / x ) + vmul.f64 d3 , d1, d2 // ssq * ( scale / x ) * ( scale / x ) + vadd.f64 d1 , d3, d7 // ssq = 1 + ssq * ( scale / x ) * ( scale / x ) + vmov.f64 d0 , d5 // scale = x + +KERNEL_S1_END_\@: + + add X, X, INC_X + +.endm + + +#else + +.macro KERNEL_F1 + + fldmias X!, { s4 - s5 } + + vcmpe.f32 s4, s6 // compare with 0.0 + vmrs APSR_nzcv, fpscr + beq KERNEL_F1_NEXT_\@ + vabs.f32 s4, s4 + vcmpe.f32 s0, s4 // compare with scale + vmrs APSR_nzcv, fpscr + vdivge.f32 s2 , s4, s0 // scale >= x ? x / scale + vmlage.f32 s1 , s2 , s2 // ssq += ( x/scale ) * ( x/scale ) + bge KERNEL_F1_NEXT_\@ + vdiv.f32 s2 , s0, s4 // scale / x + vmul.f32 s2 , s2, s2 // ( scale / x ) * ( scale / x ) + vmul.f32 s3 , s1, s2 // ssq * ( scale / x ) * ( scale / x ) + vadd.f32 s1 , s3, s7 // ssq = 1 + ssq * ( scale / x ) * ( scale / x ) + vmov.f32 s0 , s4 // scale = x + +KERNEL_F1_NEXT_\@: + + vcmpe.f32 s5, s6 // compare with 0.0 + vmrs APSR_nzcv, fpscr + beq KERNEL_F1_END_\@ + vabs.f32 s5, s5 + vcmpe.f32 s0, s5 // compare with scale + vmrs APSR_nzcv, fpscr + vdivge.f32 s2 , s5, s0 // scale >= x ? x / scale + vmlage.f32 s1 , s2 , s2 // ssq += ( x/scale ) * ( x/scale ) + bge KERNEL_F1_END_\@ + vdiv.f32 s2 , s0, s5 // scale / x + vmul.f32 s2 , s2, s2 // ( scale / x ) * ( scale / x ) + vmul.f32 s3 , s1, s2 // ssq * ( scale / x ) * ( scale / x ) + vadd.f32 s1 , s3, s7 // ssq = 1 + ssq * ( scale / x ) * ( scale / x ) + vmov.f32 s0 , s5 // scale = x + +KERNEL_F1_END_\@: + + +.endm + +.macro KERNEL_F8 + + pld [ X, #X_PRE ] + KERNEL_F1 + KERNEL_F1 + KERNEL_F1 + KERNEL_F1 + pld [ X, #X_PRE ] + KERNEL_F1 + KERNEL_F1 + KERNEL_F1 + KERNEL_F1 + +.endm + +.macro KERNEL_S1 + + fldmias X, { s4 - s5 } + + vcmpe.f32 s4, s6 // compare with 0.0 + vmrs APSR_nzcv, fpscr + beq KERNEL_S1_NEXT_\@ + vabs.f32 s4, s4 + vcmpe.f32 s0, s4 // compare with scale + vmrs APSR_nzcv, fpscr + vdivge.f32 s2 , s4, s0 // scale >= x ? x / scale + vmlage.f32 s1 , s2 , s2 // ssq += ( x/scale ) * ( x/scale ) + bge KERNEL_S1_NEXT_\@ + vdiv.f32 s2 , s0, s4 // scale / x + vmul.f32 s2 , s2, s2 // ( scale / x ) * ( scale / x ) + vmul.f32 s3 , s1, s2 // ssq * ( scale / x ) * ( scale / x ) + vadd.f32 s1 , s3, s7 // ssq = 1 + ssq * ( scale / x ) * ( scale / x ) + vmov.f32 s0 , s4 // scale = x + +KERNEL_S1_NEXT_\@: + + vcmpe.f32 s5, s6 // compare with 0.0 + vmrs APSR_nzcv, fpscr + beq KERNEL_S1_END_\@ + vabs.f32 s5, s5 + vcmpe.f32 s0, s5 // compare with scale + vmrs APSR_nzcv, fpscr + vdivge.f32 s2 , s5, s0 // scale >= x ? x / scale + vmlage.f32 s1 , s2 , s2 // ssq += ( x/scale ) * ( x/scale ) + bge KERNEL_S1_END_\@ + vdiv.f32 s2 , s0, s5 // scale / x + vmul.f32 s2 , s2, s2 // ( scale / x ) * ( scale / x ) + vmul.f32 s3 , s1, s2 // ssq * ( scale / x ) * ( scale / x ) + vadd.f32 s1 , s3, s7 // ssq = 1 + ssq * ( scale / x ) * ( scale / x ) + vmov.f32 s0 , s5 // scale = x + +KERNEL_S1_END_\@: + + add X, X, INC_X + +.endm + + + +#endif + +#endif + +/************************************************************************************** +* End of macro definitions +**************************************************************************************/ + + PROLOGUE + + b nrm2_begin + + +#if defined(COMPLEX) + +#if defined(DOUBLE) + +znrm2_one: + .word 0x00000000 + .word 0x3ff00000 + +#else + +cnrm2_one: + .word 0x3f800000 + +#endif + +#else + +#if defined(DOUBLE) + +dnrm2_one: + .word 0x00000000 + .word 0x3ff00000 + +#else + +snrm2_one: + .word 0x3f800000 + +#endif + +#endif + + + .align 5 + + +nrm2_begin: + +#if defined(COMPLEX) + +#if defined(DOUBLE) + vsub.f64 d0 , d0 , d0 // scale=0.0 + vldr.64 d1 , znrm2_one // ssq=1.0 + vmov.f64 d7 , d1 // value 1.0 + vmov.f64 d6 , d0 // value 0.0 +#else + vsub.f32 s0 , s0 , s0 // scale=0.0 + vldr.32 s1 , cnrm2_one // ssq=1.0 + vmov.f32 s7 , s1 // value 1.0 + vmov.f32 s6 , s0 // value 0.0 +#endif + +#else + +#if defined(DOUBLE) + vsub.f64 d0 , d0 , d0 // scale=0.0 + vldr.64 d1 , dnrm2_one // ssq=1.0 + vmov.f64 d7 , d1 // value 1.0 + vmov.f64 d6 , d0 // value 0.0 +#else + vsub.f32 s0 , s0 , s0 // scale=0.0 + vldr.32 s1 , snrm2_one // ssq=1.0 + vmov.f32 s7 , s1 // value 1.0 + vmov.f32 s6 , s0 // value 0.0 +#endif + + +#endif + + + cmp N, #0 + ble nrm2_kernel_L999 + + cmp INC_X, #0 + beq nrm2_kernel_L999 + + + cmp INC_X, #1 + bne nrm2_kernel_S_BEGIN + + +nrm2_kernel_F_BEGIN: + + asrs I, N, #3 // I = N / 8 + ble nrm2_kernel_F1 + +nrm2_kernel_F8: + + KERNEL_F8 + + subs I, I, #1 + bne nrm2_kernel_F8 + +nrm2_kernel_F1: + + ands I, N, #7 + ble nrm2_kernel_L999 + + +nrm2_kernel_F10: + + KERNEL_F1 + + subs I, I, #1 + bne nrm2_kernel_F10 + + b nrm2_kernel_L999 + +nrm2_kernel_S_BEGIN: + +#if defined(COMPLEX) + +#if defined(DOUBLE) + lsl INC_X, INC_X, #4 // INC_X * SIZE * 2 +#else + lsl INC_X, INC_X, #3 // INC_X * SIZE * 2 +#endif + +#else + +#if defined(DOUBLE) + lsl INC_X, INC_X, #3 // INC_X * SIZE +#else + lsl INC_X, INC_X, #2 // INC_X * SIZE +#endif + +#endif + + + +nrm2_kernel_S1: + + mov I, N + + .align 5 + +nrm2_kernel_S10: + + KERNEL_S1 + + subs I, I, #1 + bne nrm2_kernel_S10 + + +nrm2_kernel_L999: + +#if defined(DOUBLE) + vsqrt.f64 d1, d1 + vmul.f64 d0, d0, d1 +#else + vsqrt.f32 s1, s1 + vmul.f32 s0, s0, s1 +#endif + + bx lr + + EPILOGUE +