From patchwork Thu Jun 25 09:22:27 2015 Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: 8bit X-Patchwork-Submitter: =?utf-8?b?5YiY5pmT5Lic?= X-Patchwork-Id: 488321 Return-Path: X-Original-To: incoming-imx@patchwork.ozlabs.org Delivered-To: patchwork-incoming-imx@bilbo.ozlabs.org Received: from bombadil.infradead.org (bombadil.infradead.org [IPv6:2001:1868:205::9]) (using TLSv1.2 with cipher ECDHE-RSA-AES256-GCM-SHA384 (256/256 bits)) (No client certificate requested) by ozlabs.org (Postfix) with ESMTPS id 6805B1402DE for ; Thu, 25 Jun 2015 19:34:51 +1000 (AEST) Received: from localhost ([127.0.0.1] helo=bombadil.infradead.org) by bombadil.infradead.org with esmtp (Exim 4.80.1 #2 (Red Hat Linux)) id 1Z83Vt-0001vE-Is; Thu, 25 Jun 2015 09:32:05 +0000 Received: from mail.nudt.edu.cn ([61.187.54.11] helo=nudt.edu.cn) by bombadil.infradead.org with esmtp (Exim 4.80.1 #2 (Red Hat Linux)) id 1Z83Vl-0001fi-Mz for linux-arm-kernel@lists.infradead.org; Thu, 25 Jun 2015 09:32:02 +0000 Received: by ajax-webmail-coremail.nudt.edu.cn (Coremail) ; Thu, 25 Jun 2015 17:22:27 +0800 (GMT+08:00) Date: Thu, 25 Jun 2015 17:22:27 +0800 (GMT+08:00) From: =?UTF-8?B?5YiY5pmT5Lic?= To: "Alexey Klimov" Message-ID: <34d966c7.7ec3.14e2a0685e0.Coremail.liuxiaodong@nudt.edu.cn> In-Reply-To: References: Subject: Re: Re: [PATCH] XOR implementation for ARMv8 MIME-Version: 1.0 X-Originating-IP: [119.39.248.114] X-Priority: 3 X-Mailer: Coremail Webmail Server Version 4.0.8 dev build 20150107(58648.7033.6860) Copyright (c) 2002-2015 www.mailtech.cn nudt-out X-SendMailWithSms: false X-CM-TRANSID: AQAAf0C5OEdTyItVh61QAQ--.8118W X-CM-SenderInfo: xolx5xpdrg00nj6q3vvwohv3gofq/1tbiAQALElC8UkUZ6AABsT X-Coremail-Antispam: 1Ur529EdanIXcx71UUUUU7IcSsGvfJ3iIAIbVAYjsxI4VWxJw CS07vEb4IE77IF4wCS07vE1I0E4x80FVAKz4kxMIAIbVAFxVCaYxvI4VCIwcAKzIAtYxBI daVFxhVjvjDU= X-MIME-Error: demime acl condition: uuencoded line length does not match advertised number of bytes X-CRM114-Version: 20100106-BlameMichelson ( TRE 0.8.0 (BSD) ) MR-646709E3 X-CRM114-CacheID: sfid-20150625_023158_864026_2FB6C661 X-CRM114-Status: GOOD ( 19.05 ) X-Spam-Score: -3.3 (---) X-Spam-Report: SpamAssassin version 3.4.0 on bombadil.infradead.org summary: Content analysis details: (-3.3 points) pts rule name description ---- ---------------------- -------------------------------------------------- -0.0 RCVD_IN_DNSWL_NONE RBL: Sender listed at http://www.dnswl.org/, no trust [61.187.54.11 listed in list.dnswl.org] -0.0 SPF_HELO_PASS SPF: HELO matches SPF record -0.0 SPF_PASS SPF: sender matches SPF record -1.4 RP_MATCHES_RCVD Envelope sender domain matches handover relay domain -1.9 BAYES_00 BODY: Bayes spam probability is 0 to 1% [score: 0.0000] Cc: liuyun01@kylinos.cn, "linux-arm-kernel@lists.infradead.org" , =?UTF-8?Q?J=C3=A9r=C3=B4me_Forissier?= , Ard Biesheuvel X-BeenThere: linux-arm-kernel@lists.infradead.org X-Mailman-Version: 2.1.18-1 Precedence: list List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Sender: "linux-arm-kernel" Errors-To: linux-arm-kernel-bounces+incoming-imx=patchwork.ozlabs.org@lists.infradead.org List-Id: linux-imx-kernel.lists.patchwork.ozlabs.org > On Wed, Jun 24, 2015 at 11:51 AM, Ard Biesheuvel wrote: > > On 24 June 2015 at 10:29, Jérôme Forissier wrote: > >> > >> > >> On 06/24/2015 09:00 AM, 刘晓东 wrote: > >>> Use the 128-bit SIMD registers and SIMD arithmetic instructions for XOR calculation in assembly language. > >> > >> Don't you need kernel_neon_begin()/kernel_neon_end() somewhere? (see > >> Documentation/arm/kernel_mode_neon.txt). > >> > > > > Jerome is right: use of this driver will corrupt the FP/SIMD state of > > arbitrary userland tasks if you don't explicitly claim the NEON for > > in-kernel use by calling kernel_neon_begin)_ and end() > > > > Since XOR may be called in interrupt context, this could add a fixed > > overhead to each call, even if you are calling the function many times > > in a row. This means you may be better off using even fewer registers, > > and use kernel_neon_begin_partial() instead. > > > > May I ask what kind of core you tested this on? > > And if Xiaodong Liu isn't subscribed to linux arm mail list then he will not get this email chain. > It was removed by Jerome from to/cc list. Please don't do that. > > (restoring back Xiaodong Liu email) > > -- > Best regards, Klimov Alexey According to your suggestion, I have revised the code by using kernel_neon_begin and kernel_neon_end to avoid contaminating the SIMD registers in interrupt context. BTW, I use Phytium FT-1500A SoC which is armv8 compatible for tests. Regards, Xiaodong Liu --- include/asm/xor.h | 197 ++++++++++++++++++++++++++++++++++++++++++++ kernel/arm64ksyms.c | 13 ++ lib/Makefile | 4 lib/xor-neon.c | 30 ++++++ lib/xor.S | 228 ++++++++++++++++++++++++++++++++++++++++++++++++++++ 5 files changed, 472 insertions(+) --- diff -pruN -X dontdiff linux-4.0.5-orig/arch/arm64/include/asm/xor.h linux-4.0.5-mod/arch/arm64/include/asm/xor.h --- linux-4.0.5-orig/arch/arm64/include/asm/xor.h 1970-01-01 08:00:00.000000000 +0800 +++ linux-4.0.5-mod/arch/arm64/include/asm/xor.h 2015-06-25 16:59:19.527197817 +0800 @@ -0,0 +1,197 @@ +/* + * arch/arm64/include/asm/xor.h + * + * Copyright (C) Xiaodong Liu , Changsha, P.R. China + * + * This program is free software; you can redistribute it and/or modify + * it under the terms of the GNU General Public License version 2 as + * published by the Free Software Foundation. + */ +#include +#include +#include +#include + +#define __XOR(a1, a2) a1 ^= a2 + +#define GET_BLOCK_8(dst) \ + __asm__("ldp %1, %2, [%0], #16;\n\t" \ + "ldp %3, %4, [%0], #16;\n\t" \ + "ldp %5, %6, [%0], #16;\n\t" \ + "ldp %7, %8, [%0], #16;\n\t" \ + : "=r" (dst), "=r" (a1), "=r" (a2), "=r" (a3), "=r" (a4), "=r" (a5), "=r" (a6), "=r" (a7), "=r" (a8) \ + : "0" (dst)) + +#define XOR_BLOCK_8(src) \ + __asm__("ldp %1, %2, [%0], #16;\n\t" \ + "ldp %3, %4, [%0], #16;\n\t" \ + "ldp %5, %6, [%0], #16;\n\t" \ + "ldp %7, %8, [%0], #16;\n\t" \ + : "=r" (src), "=r" (b1), "=r" (b2), "=r" (b3), "=r" (b4), "=r" (b5), "=r" (b6), "=r" (b7), "=r" (b8) \ + : "0" (src)); \ + __XOR(a1, b1); __XOR(a2, b2); __XOR(a3, b3); __XOR(a4, b4); __XOR(a5, b5); __XOR(a6, b6); __XOR(a7, b7); __XOR(a8, b8) + +#define PUT_BLOCK_8(dst) \ + __asm__ __volatile__("stp %1, %2, [%0], #16;\n\t" \ + "stp %3, %4, [%0], #16;\n\t" \ + "stp %5, %6, [%0], #16;\n\t" \ + "stp %7, %8, [%0], #16;\n\t" \ + : "=r" (dst) \ + : "0" (dst), "r" (a1), "r" (a2), "r" (a3), "r" (a4), "r" (a5), "r" (a6), "r" (a7), "r" (a8)); + +#define INIT_REGISTER() \ + register unsigned long a1 __asm__("x4"); \ + register unsigned long a2 __asm__("x5"); \ + register unsigned long a3 __asm__("x6"); \ + register unsigned long a4 __asm__("x7"); \ + register unsigned long a5 __asm__("x8"); \ + register unsigned long a6 __asm__("x9"); \ + register unsigned long a7 __asm__("x10"); \ + register unsigned long a8 __asm__("x11"); \ + register unsigned long b1 __asm__("x12"); \ + register unsigned long b2 __asm__("x13"); \ + register unsigned long b3 __asm__("x14"); \ + register unsigned long b4 __asm__("x15"); \ + register unsigned long b5 __asm__("x16"); \ + register unsigned long b6 __asm__("x17"); \ + register unsigned long b7 __asm__("x18"); \ + register unsigned long b8 __asm__("x19"); + +static void +xor_arm8regs_2(unsigned long bytes, unsigned long *p1, unsigned long *p2) +{ + unsigned long lines = bytes / sizeof(unsigned long) / 8; + INIT_REGISTER(); + + do { + GET_BLOCK_8(p1); + XOR_BLOCK_8(p2); + PUT_BLOCK_8(p1); + } while(--lines); +} + +static void +xor_arm8regs_3(unsigned long bytes, unsigned long *p1, unsigned long *p2, + unsigned long *p3) +{ + unsigned long lines = bytes / sizeof(unsigned long) / 8; + INIT_REGISTER(); + + do { + GET_BLOCK_8(p1); + XOR_BLOCK_8(p2); + XOR_BLOCK_8(p3); + PUT_BLOCK_8(p1); + } while(--lines); +} + +static void +xor_arm8regs_4(unsigned long bytes, unsigned long *p1, unsigned long *p2, + unsigned long *p3, unsigned long *p4) +{ + unsigned long lines = bytes / sizeof(unsigned long) / 8; + INIT_REGISTER(); + + do { + GET_BLOCK_8(p1); + XOR_BLOCK_8(p2); + XOR_BLOCK_8(p3); + XOR_BLOCK_8(p4); + PUT_BLOCK_8(p1); + } while(--lines); +} + +static void +xor_arm8regs_5(unsigned long bytes, unsigned long *p1, unsigned long *p2, + unsigned long *p3, unsigned long *p4, unsigned long *p5) +{ + unsigned long lines = bytes / sizeof(unsigned long) / 8; + INIT_REGISTER(); + + do { + GET_BLOCK_8(p1); + XOR_BLOCK_8(p2); + XOR_BLOCK_8(p3); + XOR_BLOCK_8(p4); + XOR_BLOCK_8(p5); + PUT_BLOCK_8(p1); + } while(--lines); +} + +extern struct xor_block_template const xor_block_neon_arm64; + +static void +xor_neon_2(unsigned long bytes, unsigned long *p1, unsigned long *p2) +{ + if(in_interrupt()) { + xor_arm8regs_2(bytes, p1, p2); + } else { + kernel_neon_begin(); + xor_block_neon_arm64.do_2(bytes, p1, p2); + kernel_neon_end(); + } +} + +static void +xor_neon_3(unsigned long bytes, unsigned long *p1, unsigned long *p2, + unsigned long *p3) +{ + if(in_interrupt()) { + xor_arm8regs_3(bytes, p1, p2, p3); + } else { + kernel_neon_begin(); + xor_block_neon_arm64.do_3(bytes, p1, p2, p3); + kernel_neon_end(); + } +} + +static void +xor_neon_4(unsigned long bytes, unsigned long *p1, unsigned long *p2, + unsigned long *p3, unsigned long *p4) +{ + if(in_interrupt()) { + xor_arm8regs_4(bytes, p1, p2, p3, p4); + } else { + kernel_neon_begin(); + xor_block_neon_arm64.do_4(bytes, p1, p2, p3, p4); + kernel_neon_end(); + } +} + +static void +xor_neon_5(unsigned long bytes, unsigned long *p1, unsigned long *p2, + unsigned long *p3, unsigned long *p4, unsigned long *p5) +{ + if(in_interrupt()) { + xor_arm8regs_5(bytes, p1, p2, p3, p4, p5); + } else { + kernel_neon_begin(); + xor_block_neon_arm64.do_5(bytes, p1, p2, p3, p4, p5); + kernel_neon_end(); + } +} + +static struct xor_block_template xor_block_arm64regs8 = { + .name = "arm64regs8", + .do_2 = xor_arm8regs_2, + .do_3 = xor_arm8regs_3, + .do_4 = xor_arm8regs_4, + .do_5 = xor_arm8regs_5 +}; + +static struct xor_block_template xor_block_arm64 = { + .name = "neon_arm64", + .do_2 = xor_neon_2, + .do_3 = xor_neon_3, + .do_4 = xor_neon_4, + .do_5 = xor_neon_5 +}; + +#undef XOR_TRY_TEMPLATES +#define XOR_TRY_TEMPLATES \ + do { \ + xor_speed(&xor_block_8regs); \ + xor_speed(&xor_block_32regs); \ + xor_speed(&xor_block_arm64regs8); \ + do { if (cpu_has_neon()) xor_speed(&xor_block_arm64); } while (0); \ + } while(0) diff -pruN -X dontdiff linux-4.0.5-orig/arch/arm64/kernel/arm64ksyms.c linux-4.0.5-mod/arch/arm64/kernel/arm64ksyms.c --- linux-4.0.5-orig/arch/arm64/kernel/arm64ksyms.c 2015-06-06 23:21:22.000000000 +0800 +++ linux-4.0.5-mod/arch/arm64/kernel/arm64ksyms.c 2015-06-25 11:40:07.537692040 +0800 @@ -65,3 +65,16 @@ EXPORT_SYMBOL(test_and_change_bit); #ifdef CONFIG_FUNCTION_TRACER EXPORT_SYMBOL(_mcount); #endif + + /* xor ops */ +extern void xor_arm64ldpregs16_2(unsigned long, unsigned long *, unsigned long *); +extern void xor_arm64ldpregs16_3(unsigned long, unsigned long *, unsigned long *, + unsigned long *); +extern void xor_arm64ldpregs16_4(unsigned long, unsigned long *, unsigned long *, + unsigned long *, unsigned long *); +extern void xor_arm64ldpregs16_5(unsigned long, unsigned long *, unsigned long *, + unsigned long *, unsigned long *, unsigned long *); +EXPORT_SYMBOL(xor_arm64ldpregs16_2); +EXPORT_SYMBOL(xor_arm64ldpregs16_3); +EXPORT_SYMBOL(xor_arm64ldpregs16_4); +EXPORT_SYMBOL(xor_arm64ldpregs16_5); diff -pruN -X dontdiff linux-4.0.5-orig/arch/arm64/lib/Makefile linux-4.0.5-mod/arch/arm64/lib/Makefile --- linux-4.0.5-orig/arch/arm64/lib/Makefile 2015-06-06 23:21:22.000000000 +0800 +++ linux-4.0.5-mod/arch/arm64/lib/Makefile 2015-06-25 16:47:44.051223943 +0800 @@ -3,3 +3,7 @@ lib-y := bitops.o clear_user.o delay.o clear_page.o memchr.o memcpy.o memmove.o memset.o \ memcmp.o strcmp.o strncmp.o strlen.o strnlen.o \ strchr.o strrchr.o + +ifeq ($(CONFIG_KERNEL_MODE_NEON),y) + obj-$(CONFIG_XOR_BLOCKS) += xor-neon.o xor.o +endif diff -pruN -X dontdiff linux-4.0.5-orig/arch/arm64/lib/xor-neon.c linux-4.0.5-mod/arch/arm64/lib/xor-neon.c --- linux-4.0.5-orig/arch/arm64/lib/xor-neon.c 1970-01-01 08:00:00.000000000 +0800 +++ linux-4.0.5-mod/arch/arm64/lib/xor-neon.c 2015-06-25 16:53:36.319210709 +0800 @@ -0,0 +1,30 @@ +/* + * arch/arm64/lib/xor-neon.c + * + * Copyright (C) Xiaodong Liu , Changsha, P.R. China + * + * This program is free software; you can redistribute it and/or modify + * it under the terms of the GNU General Public License version 2 as + * published by the Free Software Foundation. + */ + +#include +#include + +extern void xor_arm64ldpregs16_2(unsigned long, unsigned long *, unsigned long *); +extern void xor_arm64ldpregs16_3(unsigned long, unsigned long *, unsigned long *, + unsigned long *); +extern void xor_arm64ldpregs16_4(unsigned long, unsigned long *, unsigned long *, + unsigned long *, unsigned long *); +extern void xor_arm64ldpregs16_5(unsigned long, unsigned long *, unsigned long *, + unsigned long *, unsigned long *, unsigned long *); + +struct xor_block_template const xor_block_neon_arm64 = { + .name = "ARM64LDPregs16", + .do_2 = xor_arm64ldpregs16_2, + .do_3 = xor_arm64ldpregs16_3, + .do_4 = xor_arm64ldpregs16_4, + .do_5 = xor_arm64ldpregs16_5, +}; + +EXPORT_SYMBOL(xor_block_neon_arm64); diff -pruN -X dontdiff linux-4.0.5-orig/arch/arm64/lib/xor.S linux-4.0.5-mod/arch/arm64/lib/xor.S --- linux-4.0.5-orig/arch/arm64/lib/xor.S 1970-01-01 08:00:00.000000000 +0800 +++ linux-4.0.5-mod/arch/arm64/lib/xor.S 2015-06-24 09:25:49.969256540 +0800 @@ -0,0 +1,228 @@ +/* + * arch/arm64/lib/xor.S + * + * Copyright (C) Xiaodong Liu , Changsha, P.R. China + * + * This program is free software; you can redistribute it and/or modify + * it under the terms of the GNU General Public License version 2 as + * published by the Free Software Foundation. + */ + +#include +#include +.macro xor_vectorregs16 + eor v24.16b, v24.16b, v16.16b + eor v25.16b, v25.16b, v17.16b + eor v26.16b, v26.16b, v18.16b + eor v27.16b, v27.16b, v19.16b + eor v28.16b, v28.16b, v20.16b + eor v29.16b, v29.16b, v21.16b + eor v30.16b, v30.16b, v22.16b + eor v31.16b, v31.16b, v23.16b +.endm + +.align 4 + +/* + * void xor_arm64ldpregs16_2(unsigned long size, unsigned long * dst, unsigned long *src); + * + * Parameters: + * x0 - size + * x1 - dst + * x2 - src + */ +ENTRY(xor_arm64ldpregs16_2) + + lsr x0, x0, #10 + +.p2align 4 +Loop23: + ldp q16, q17, [x2], #32 + ldp q18, q19, [x2], #32 + ldp q20, q21, [x2], #32 + ldp q22, q23, [x2], #32 + + mov x3,x1 + + ldp q24, q25, [x1], #32 + ldp q26, q27, [x1], #32 + ldp q27, q29, [x1], #32 + ldp q30, q31, [x1], #32 + + xor_vectorregs16 + + stp q24, q25, [x3], #32 + stp q26, q27, [x3], #32 + stp q27, q29, [x3], #32 + stp q30, q31, [x3], #32 + + subs x0, x0, #1 + cbnz x0, Loop23 + + ret +ENDPROC(xor_arm64ldpregs16_2) + +/* + * void xor_arm64ldpregs16_3(unsigned long size, unsigned long *dst, unsigned long *src0, unsigned long *src1); + * + * Parameters: + * x0 - size + * x1 - dst + * x2 - src0 + * x3 - src1 + */ +ENTRY(xor_arm64ldpregs16_3) + + lsr x0, x0, #10 + +.p2align 4 +Loop33: + ldp q16, q17, [x2], #32 + ldp q18, q19, [x2], #32 + ldp q20, q21, [x2], #32 + ldp q22, q23, [x2], #32 + + mov x4,x1 + + ldp q24, q25, [x1], #32 + ldp q26, q27, [x1], #32 + ldp q27, q29, [x1], #32 + ldp q30, q31, [x1], #32 + + xor_vectorregs16 + + ldp q16, q17, [x3], #32 + ldp q18, q19, [x3], #32 + ldp q20, q21, [x3], #32 + ldp q22, q23, [x3], #32 + + xor_vectorregs16 + + stp q24, q25, [x4], #32 + stp q26, q27, [x4], #32 + stp q27, q29, [x4], #32 + stp q30, q31, [x4], #32 + + subs x0, x0, #1 + cbnz x0, Loop33 + + ret +ENDPROC(xor_arm64ldpregs16_3) + +/* + * void xor_arm64ldpregs16_4(unsigned long size, unsigned long *dst, unsigned long *src0, unsigned long *src1, unsigned long *src2); + * + * Parameters: + * x0 - size + * x1 - dst + * x2 - src0 + * x3 - src1 + * x4 - src2 + */ +ENTRY(xor_arm64ldpregs16_4) + + lsr x0, x0, #10 + +.p2align 4 +Loop43: + ldp q16, q17, [x2], #32 + ldp q18, q19, [x2], #32 + ldp q20, q21, [x2], #32 + ldp q22, q23, [x2], #32 + + mov x5,x1 + + ldp q24, q25, [x1], #32 + ldp q26, q27, [x1], #32 + ldp q27, q29, [x1], #32 + ldp q30, q31, [x1], #32 + + xor_vectorregs16 + + ldp q16, q17, [x3], #32 + ldp q18, q19, [x3], #32 + ldp q20, q21, [x3], #32 + ldp q22, q23, [x3], #32 + + xor_vectorregs16 + + ldp q16, q17, [x4], #32 + ldp q18, q19, [x4], #32 + ldp q20, q21, [x4], #32 + ldp q22, q23, [x4], #32 + + xor_vectorregs16 + + stp q24, q25, [x5], #32 + stp q26, q27, [x5], #32 + stp q27, q29, [x5], #32 + stp q30, q31, [x5], #32 + + subs x0, x0, #1 + cbnz x0, Loop43 + + ret +ENDPROC(xor_arm64ldpregs16_4) + +/* + * void xor_arm64ldpregs16_5(unsigned long size, unsigned long *dst, unsigned long *src0, unsigned long *src1, unsigned long *src2, unsigned long *src3); + * + * Parameters: + * x0 - size + * x1 - dst + * x2 - src0 + * x3 - src1 + * x4 - src2 + * x5 - src3 + */ +ENTRY(xor_arm64ldpregs16_5) + + lsr x0, x0, #10 + +.p2align 4 +Loop53: + ldp q16, q17, [x2], #32 + ldp q18, q19, [x2], #32 + ldp q20, q21, [x2], #32 + ldp q22, q23, [x2], #32 + + mov x6,x1 + + ldp q24, q25, [x1], #32 + ldp q26, q27, [x1], #32 + ldp q27, q29, [x1], #32 + ldp q30, q31, [x1], #32 + + xor_vectorregs16 + + ldp q16, q17, [x3], #32 + ldp q18, q19, [x3], #32 + ldp q20, q21, [x3], #32 + ldp q22, q23, [x3], #32 + + xor_vectorregs16 + + ldp q16, q17, [x4], #32 + ldp q18, q19, [x4], #32 + ldp q20, q21, [x4], #32 + ldp q22, q23, [x4], #32 + + xor_vectorregs16 + + ldp q16, q17, [x5], #32 + ldp q18, q19, [x5], #32 + ldp q20, q21, [x5], #32 + ldp q22, q23, [x5], #32 + + xor_vectorregs16 + + stp q24, q25, [x6], #32 + stp q26, q27, [x6], #32 + stp q27, q29, [x6], #32 + stp q30, q31, [x6], #32 + + subs x0, x0, #1 + cbnz x0, Loop53 + + ret +ENDPROC(xor_arm64ldpregs16_5)