From d2b20c5c51b1f174d40dd8879b8a7387ce5a5416 Mon Sep 17 00:00:00 2001 From: wernsaar Date: Mon, 25 Nov 2013 12:25:58 +0100 Subject: [PATCH] add optimized axpy kernel --- kernel/arm/KERNEL.ARMV6 | 8 +- kernel/arm/KERNEL.ARMV7 | 8 +- kernel/arm/axpy_vfp.S | 503 ++++++++++++++++++++++++++++++++++++++++++++++++ 3 files changed, 511 insertions(+), 8 deletions(-) create mode 100644 kernel/arm/axpy_vfp.S diff --git a/kernel/arm/KERNEL.ARMV6 b/kernel/arm/KERNEL.ARMV6 index 9dad061..79fb5f9 100644 --- a/kernel/arm/KERNEL.ARMV6 +++ b/kernel/arm/KERNEL.ARMV6 @@ -35,10 +35,10 @@ DASUMKERNEL = asum_vfp.S CASUMKERNEL = asum_vfp.S ZASUMKERNEL = asum_vfp.S -SAXPYKERNEL = axpy.c -DAXPYKERNEL = axpy.c -CAXPYKERNEL = zaxpy.c -ZAXPYKERNEL = zaxpy.c +SAXPYKERNEL = axpy_vfp.S +DAXPYKERNEL = axpy_vfp.S +CAXPYKERNEL = axpy_vfp.S +ZAXPYKERNEL = axpy_vfp.S SCOPYKERNEL = scopy_vfp.S DCOPYKERNEL = dcopy_vfp.S diff --git a/kernel/arm/KERNEL.ARMV7 b/kernel/arm/KERNEL.ARMV7 index 07d38d1..1433270 100644 --- a/kernel/arm/KERNEL.ARMV7 +++ b/kernel/arm/KERNEL.ARMV7 @@ -40,10 +40,10 @@ DASUMKERNEL = asum_vfp.S CASUMKERNEL = asum_vfp.S ZASUMKERNEL = asum_vfp.S -SAXPYKERNEL = ../arm/axpy.c -DAXPYKERNEL = ../arm/axpy.c -CAXPYKERNEL = ../arm/zaxpy.c -ZAXPYKERNEL = ../arm/zaxpy.c +SAXPYKERNEL = axpy_vfp.S +DAXPYKERNEL = axpy_vfp.S +CAXPYKERNEL = axpy_vfp.S +ZAXPYKERNEL = axpy_vfp.S SCOPYKERNEL = scopy_vfp.S DCOPYKERNEL = dcopy_vfp.S diff --git a/kernel/arm/axpy_vfp.S b/kernel/arm/axpy_vfp.S new file mode 100644 index 0000000..acc5757 --- /dev/null +++ b/kernel/arm/axpy_vfp.S @@ -0,0 +1,503 @@ +/*************************************************************************** +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. +*****************************************************************************/ + +/************************************************************************************** +* 2013/11/14 Saar +* BLASTEST : xOK +* CTEST : xOK +* TEST : xOK +* +**************************************************************************************/ + +#define ASSEMBLER +#include "common.h" + +#define STACKSIZE 256 + +#define OLD_INC_X [fp, #0 ] +#define OLD_Y [fp, #4 ] +#define OLD_INC_Y [fp, #8 ] + + +#define N r0 +#define Y r1 +#define INC_X r2 +#define X r3 +#define INC_Y r4 + +#define I r12 + +#define X_PRE 512 + +/************************************************************************************** +* Macro definitions +**************************************************************************************/ + +/*****************************************************************************************/ + +#if !defined(CONJ) + +#if defined(DOUBLE) + +#define FMAC_R1 fmacd +#define FMAC_R2 fnmacd +#define FMAC_I1 fmacd +#define FMAC_I2 fmacd + +#else + +#define FMAC_R1 fmacs +#define FMAC_R2 fnmacs +#define FMAC_I1 fmacs +#define FMAC_I2 fmacs + +#endif + +#else // CONJ + +#if defined(DOUBLE) + +#define FMAC_R1 fmacd +#define FMAC_R2 fmacd +#define FMAC_I1 fnmacd +#define FMAC_I2 fmacd + +#else + +#define FMAC_R1 fmacs +#define FMAC_R2 fmacs +#define FMAC_I1 fnmacs +#define FMAC_I2 fmacs + +#endif + +#endif + + +#if !defined(COMPLEX) + +#if defined(DOUBLE) + +.macro KERNEL_F4 + + pld [ X, #X_PRE ] + fldmiad X!, { d4 - d7 } + pld [ Y, #X_PRE ] + fldmiad Y , { d8 - d11 } + fmacd d8 , d0, d4 + fstmiad Y!, { d8 } + fmacd d9 , d0, d5 + fstmiad Y!, { d9 } + fmacd d10, d0, d6 + fstmiad Y!, { d10 } + fmacd d11, d0, d7 + fstmiad Y!, { d11 } + + +.endm + + +.macro KERNEL_F1 + + fldmiad X!, { d4 } + fldmiad Y , { d8 } + fmacd d8 , d0, d4 + fstmiad Y!, { d8 } + +.endm + +.macro KERNEL_S1 + + fldmiad X , { d4 } + fldmiad Y , { d8 } + fmacd d8 , d0, d4 + fstmiad Y , { d8 } + add X, X, INC_X + add Y, Y, INC_Y + +.endm + +#else + +.macro KERNEL_F4 + + fldmias X!, { s4 - s7 } + fldmias Y , { s8 - s11 } + fmacs s8 , s0, s4 + fstmias Y!, { s8 } + fmacs s9 , s0, s5 + fstmias Y!, { s9 } + fmacs s10, s0, s6 + fstmias Y!, { s10 } + fmacs s11, s0, s7 + fstmias Y!, { s11 } + + +.endm + + +.macro KERNEL_F1 + + fldmias X!, { s4 } + fldmias Y , { s8 } + fmacs s8 , s0, s4 + fstmias Y!, { s8 } + +.endm + +.macro KERNEL_S1 + + fldmias X , { s4 } + fldmias Y , { s8 } + fmacs s8 , s0, s4 + fstmias Y , { s8 } + add X, X, INC_X + add Y, Y, INC_Y + +.endm + + +#endif + +#else + +#if defined(DOUBLE) + +.macro KERNEL_F4 + + pld [ X, #X_PRE ] + fldmiad X!, { d4 - d7 } + pld [ Y, #X_PRE ] + fldmiad Y , { d8 - d11 } + + FMAC_R1 d8 , d0, d4 + FMAC_R2 d8 , d1, d5 + FMAC_I1 d9 , d0, d5 + FMAC_I2 d9 , d1, d4 + fstmiad Y!, { d8 } + fstmiad Y!, { d9 } + + FMAC_R1 d10, d0, d6 + FMAC_R2 d10, d1, d7 + FMAC_I1 d11, d0, d7 + FMAC_I2 d11, d1, d6 + fstmiad Y!, { d10 } + fstmiad Y!, { d11 } + + pld [ X, #X_PRE ] + fldmiad X!, { d4 - d7 } + pld [ Y, #X_PRE ] + fldmiad Y , { d8 - d11 } + + FMAC_R1 d8 , d0, d4 + FMAC_R2 d8 , d1, d5 + FMAC_I1 d9 , d0, d5 + FMAC_I2 d9 , d1, d4 + fstmiad Y!, { d8 } + fstmiad Y!, { d9 } + + FMAC_R1 d10, d0, d6 + FMAC_R2 d10, d1, d7 + FMAC_I1 d11, d0, d7 + FMAC_I2 d11, d1, d6 + fstmiad Y!, { d10 } + fstmiad Y!, { d11 } + + + + + +.endm + + +.macro KERNEL_F1 + + fldmiad X!, { d4 - d5 } + fldmiad Y , { d8 - d9 } + + FMAC_R1 d8 , d0, d4 + FMAC_R2 d8 , d1, d5 + FMAC_I1 d9 , d0, d5 + FMAC_I2 d9 , d1, d4 + fstmiad Y!, { d8 } + fstmiad Y!, { d9 } + + + +.endm + +.macro KERNEL_S1 + + fldmiad X , { d4 - d5 } + fldmiad Y , { d8 - d9 } + + FMAC_R1 d8 , d0, d4 + FMAC_R2 d8 , d1, d5 + FMAC_I1 d9 , d0, d5 + FMAC_I2 d9 , d1, d4 + fstmiad Y , { d8 - d9 } + + add X, X, INC_X + add Y, Y, INC_Y + +.endm + + + +#else + +.macro KERNEL_F4 + + pld [ X, #X_PRE ] + fldmias X!, { s4 - s7 } + pld [ Y, #X_PRE ] + fldmias Y , { s8 - s11 } + + FMAC_R1 s8 , s0, s4 + FMAC_R2 s8 , s1, s5 + FMAC_I1 s9 , s0, s5 + FMAC_I2 s9 , s1, s4 + fstmias Y!, { s8 } + fstmias Y!, { s9 } + + FMAC_R1 s10, s0, s6 + FMAC_R2 s10, s1, s7 + FMAC_I1 s11, s0, s7 + FMAC_I2 s11, s1, s6 + fstmias Y!, { s10 } + fstmias Y!, { s11 } + + fldmias X!, { s4 - s7 } + fldmias Y , { s8 - s11 } + + FMAC_R1 s8 , s0, s4 + FMAC_R2 s8 , s1, s5 + FMAC_I1 s9 , s0, s5 + FMAC_I2 s9 , s1, s4 + fstmias Y!, { s8 } + fstmias Y!, { s9 } + + FMAC_R1 s10, s0, s6 + FMAC_R2 s10, s1, s7 + FMAC_I1 s11, s0, s7 + FMAC_I2 s11, s1, s6 + fstmias Y!, { s10 } + fstmias Y!, { s11 } + + + + + +.endm + + +.macro KERNEL_F1 + + fldmias X!, { s4 - s5 } + fldmias Y , { s8 - s9 } + + FMAC_R1 s8 , s0, s4 + FMAC_R2 s8 , s1, s5 + FMAC_I1 s9 , s0, s5 + FMAC_I2 s9 , s1, s4 + fstmias Y!, { s8 } + fstmias Y!, { s9 } + + + +.endm + +.macro KERNEL_S1 + + fldmias X , { s4 - s5 } + fldmias Y , { s8 - s9 } + + FMAC_R1 s8 , s0, s4 + FMAC_R2 s8 , s1, s5 + FMAC_I1 s9 , s0, s5 + FMAC_I2 s9 , s1, s4 + fstmias Y , { s8 - s9 } + + add X, X, INC_X + add Y, Y, INC_Y + +.endm + + +#endif + +#endif + +/************************************************************************************** +* End of macro definitions +**************************************************************************************/ + + PROLOGUE + + .align 5 + push {r4 , fp} + add fp, sp, #8 + sub sp, sp, #STACKSIZE // reserve stack + + ldr INC_X , OLD_INC_X + ldr Y, OLD_Y + ldr INC_Y , OLD_INC_Y + + sub r12, fp, #128 + +#if defined(DOUBLE) + vstm r12, { d8 - d15} // store floating point registers +#else + vstm r12, { s8 - s15} // store floating point registers +#endif + + cmp N, #0 + ble axpy_kernel_L999 + + cmp INC_X, #0 + beq axpy_kernel_L999 + + cmp INC_Y, #0 + beq axpy_kernel_L999 + + cmp INC_X, #1 + bne axpy_kernel_S_BEGIN + + cmp INC_Y, #1 + bne axpy_kernel_S_BEGIN + + +axpy_kernel_F_BEGIN: + + + asrs I, N, #2 // I = N / 4 + ble axpy_kernel_F1 + + .align 5 + +axpy_kernel_F4: + +#if !defined(COMPLEX) && !defined(DOUBLE) + pld [ X, #X_PRE ] + pld [ Y, #X_PRE ] +#endif + + KERNEL_F4 + + subs I, I, #1 + ble axpy_kernel_F1 + + KERNEL_F4 + + subs I, I, #1 + bne axpy_kernel_F4 + +axpy_kernel_F1: + + ands I, N, #3 + ble axpy_kernel_L999 + +axpy_kernel_F10: + + KERNEL_F1 + + subs I, I, #1 + bne axpy_kernel_F10 + + b axpy_kernel_L999 + +axpy_kernel_S_BEGIN: + +#if defined(COMPLEX) + +#if defined(DOUBLE) + lsl INC_X, INC_X, #4 // INC_X * SIZE * 2 + lsl INC_Y, INC_Y, #4 // INC_Y * SIZE * 2 +#else + lsl INC_X, INC_X, #3 // INC_X * SIZE * 2 + lsl INC_Y, INC_Y, #3 // INC_Y * SIZE * 2 +#endif + +#else + +#if defined(DOUBLE) + lsl INC_X, INC_X, #3 // INC_X * SIZE + lsl INC_Y, INC_Y, #3 // INC_Y * SIZE +#else + lsl INC_X, INC_X, #2 // INC_X * SIZE + lsl INC_Y, INC_Y, #2 // INC_Y * SIZE +#endif + +#endif + + + asrs I, N, #2 // I = N / 4 + ble axpy_kernel_S1 + + .align 5 + +axpy_kernel_S4: + + KERNEL_S1 + KERNEL_S1 + KERNEL_S1 + KERNEL_S1 + + subs I, I, #1 + bne axpy_kernel_S4 + +axpy_kernel_S1: + + ands I, N, #3 + ble axpy_kernel_L999 + +axpy_kernel_S10: + + KERNEL_S1 + + subs I, I, #1 + bne axpy_kernel_S10 + + +axpy_kernel_L999: + + sub r3, fp, #128 + +#if defined(DOUBLE) + vldm r3, { d8 - d15 } // restore floating point registers +#else + vldm r3, { s8 - s15 } // restore floating point registers +#endif + + mov r0, #0 // set return value + + sub sp, fp, #8 + pop {r4,fp} + bx lr + + EPILOGUE + -- 2.7.4