From patchwork Tue May 31 21:35:02 2022 Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: 7bit X-Patchwork-Submitter: Adhemerval Zanella Netto X-Patchwork-Id: 54600 Return-Path: X-Original-To: patchwork@sourceware.org Delivered-To: patchwork@sourceware.org Received: from server2.sourceware.org (localhost [IPv6:::1]) by sourceware.org (Postfix) with ESMTP id 34C513834F04 for ; Tue, 31 May 2022 21:39:03 +0000 (GMT) DKIM-Filter: OpenDKIM Filter v2.11.0 sourceware.org 34C513834F04 DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=sourceware.org; s=default; t=1654033143; bh=5R09Tm75NwPSi/5MvsMuVQEuUMZUXwmCRnaGL5LBX18=; h=To:Subject:Date:In-Reply-To:References:List-Id:List-Unsubscribe: List-Archive:List-Post:List-Help:List-Subscribe:From:Reply-To: From; b=I1r8LePbC0k+exYVvKRzuNJqE9aMQYUHxTFDEeOTS3TDSHYGP0UsTC/++Fs2ji3Qo jHRgLquRMjmm904yYgAliaaFZsqwoxi2Ga/R6ffOsaOjpr9K+0qRiItNb9OKQJqqeM r7IgO4N6+ONVhy6JGdRyzGxKlFijtHzTii9eEtz8= X-Original-To: libc-alpha@sourceware.org Delivered-To: libc-alpha@sourceware.org Received: from mail-oa1-x34.google.com (mail-oa1-x34.google.com [IPv6:2001:4860:4864:20::34]) by sourceware.org (Postfix) with ESMTPS id 7CB2A3836E78 for ; Tue, 31 May 2022 21:35:15 +0000 (GMT) DMARC-Filter: OpenDMARC Filter v1.4.1 sourceware.org 7CB2A3836E78 Received: by mail-oa1-x34.google.com with SMTP id 586e51a60fabf-e5e433d66dso161543fac.5 for ; Tue, 31 May 2022 14:35:15 -0700 (PDT) X-Google-DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=1e100.net; s=20210112; h=x-gm-message-state:from:to:subject:date:message-id:in-reply-to :references:mime-version:content-transfer-encoding; bh=5R09Tm75NwPSi/5MvsMuVQEuUMZUXwmCRnaGL5LBX18=; b=B199bbt/8+zWyKGs+mBvQ9t1RyWllQ26yMlwRJWd9RpXbkG4M0RecBeGVznWR1Gtz2 hQrWmgEl+VydGqFhJQSeZmisPsTiEPp9oi9pReH3qA5fURWH5FslFTt8yWuL3On+hzsM 9fmwCXPGTFwsknfFGIxikaVD8Q/RQMOE/YzUMpAKv5DjqmXjHdZAgeHdPkMaFlLd3s7w gSIejDsD/qdZ+GVB7k02d6G4rNpvrsTSHuNmGuSJaqNvwAluULpw7GVBD/aMQegLwKj8 L2A1MJE1NbJdWDkNtElSwUEIAue7NQfuN7CQVBY7N05N9e3CgYjT/xjDqlahUwq8kdSo nyLA== X-Gm-Message-State: AOAM5336ZmBqyvmZqACbIfsGmb6WnBiSsRhiFZ5x7m3IcQnXhMvhpont 69enjOflG1aS7JiqMPOkE8vLS548Tvwq9Q== X-Google-Smtp-Source: ABdhPJzpj9BsCFWGK78D07g06ZsiMrBUVlYyPPKNC3sYfm1qNZbj1SJHyejr1MJYhHxtbxei+ujd/g== X-Received: by 2002:a05:6870:51cc:b0:e1:e6ee:448f with SMTP id b12-20020a05687051cc00b000e1e6ee448fmr14472977oaj.136.1654032914086; Tue, 31 May 2022 14:35:14 -0700 (PDT) Received: from birita.. ([2804:431:c7ca:e39c:980f:eaf1:7a30:10ef]) by smtp.gmail.com with ESMTPSA id s31-20020a056830439f00b0060613c844adsm6890802otv.10.2022.05.31.14.35.12 (version=TLS1_3 cipher=TLS_AES_256_GCM_SHA384 bits=256/256); Tue, 31 May 2022 14:35:13 -0700 (PDT) To: libc-alpha@sourceware.org, "H . J . Lu" Subject: [PATCH 5/5] math: Use generic sinf implementation for SSE2 version Date: Tue, 31 May 2022 18:35:02 -0300 Message-Id: <20220531213502.227895-6-adhemerval.zanella@linaro.org> X-Mailer: git-send-email 2.34.1 In-Reply-To: <20220531213502.227895-1-adhemerval.zanella@linaro.org> References: <20220531213502.227895-1-adhemerval.zanella@linaro.org> MIME-Version: 1.0 X-Spam-Status: No, score=-12.6 required=5.0 tests=BAYES_00, DKIM_SIGNED, DKIM_VALID, DKIM_VALID_AU, DKIM_VALID_EF, GIT_PATCH_0, KAM_SHORT, RCVD_IN_DNSWL_NONE, SPF_HELO_NONE, SPF_PASS, TXREP, T_SCC_BODY_TEXT_LINE autolearn=ham autolearn_force=no version=3.4.6 X-Spam-Checker-Version: SpamAssassin 3.4.6 (2021-04-09) on server2.sourceware.org X-BeenThere: libc-alpha@sourceware.org X-Mailman-Version: 2.1.29 Precedence: list List-Id: Libc-alpha mailing list List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , X-Patchwork-Original-From: Adhemerval Zanella via Libc-alpha From: Adhemerval Zanella Netto Reply-To: Adhemerval Zanella Errors-To: libc-alpha-bounces+patchwork=sourceware.org@sourceware.org Sender: "Libc-alpha" The generic implementation shows slight better performance (gcc 11.2.1 on a Ryzen 9 5900X): * s_sincosf-sse2.S: "sincosf": { "workload-random": { "duration": 3.89961e+09, "iterations": 9.5472e+07, "reciprocal-throughput": 40.8429, "latency": 40.8483, "max-throughput": 2.4484e+07, "min-throughput": 2.44808e+07 } } * generic s_cossinf.c: "sincosf": { "workload-random": { "duration": 3.71953e+09, "iterations": 1.48512e+08, "reciprocal-throughput": 25.0515, "latency": 25.0391, "max-throughput": 3.99177e+07, "min-throughput": 3.99375e+07 } } Checked on i686-linux-gnu. --- sysdeps/i386/i686/fpu/multiarch/Makefile | 1 + .../i386/i686/fpu/multiarch/libm-test-ulps | 3 + .../i386/i686/fpu/multiarch/s_sincosf-sse2.S | 585 ------------------ .../i386/i686/fpu/multiarch/s_sincosf-sse2.c | 3 + 4 files changed, 7 insertions(+), 585 deletions(-) delete mode 100644 sysdeps/i386/i686/fpu/multiarch/s_sincosf-sse2.S create mode 100644 sysdeps/i386/i686/fpu/multiarch/s_sincosf-sse2.c diff --git a/sysdeps/i386/i686/fpu/multiarch/Makefile b/sysdeps/i386/i686/fpu/multiarch/Makefile index c2cd5cefef..c8b0a6528a 100644 --- a/sysdeps/i386/i686/fpu/multiarch/Makefile +++ b/sysdeps/i386/i686/fpu/multiarch/Makefile @@ -9,4 +9,5 @@ CFLAGS-e_logf-sse2.c = -msse2 -mfpmath=sse CFLAGS-e_powf-sse2.c = -msse2 -mfpmath=sse CFLAGS-s_cosf-sse2.c = -msse2 -mfpmath=sse CFLAGS-s_sinf-sse2.c = -msse2 -mfpmath=sse +CFLAGS-s_sincosf-sse2.c = -msse2 -mfpmath=sse endif diff --git a/sysdeps/i386/i686/fpu/multiarch/libm-test-ulps b/sysdeps/i386/i686/fpu/multiarch/libm-test-ulps index 0357d7bd14..284e3fa275 100644 --- a/sysdeps/i386/i686/fpu/multiarch/libm-test-ulps +++ b/sysdeps/i386/i686/fpu/multiarch/libm-test-ulps @@ -1492,16 +1492,19 @@ ldouble: 1 Function: "sincos_downward": double: 1 +float: 1 float128: 3 ldouble: 3 Function: "sincos_towardzero": double: 1 +float: 1 float128: 2 ldouble: 2 Function: "sincos_upward": double: 1 +float: 1 float128: 3 ldouble: 3 diff --git a/sysdeps/i386/i686/fpu/multiarch/s_sincosf-sse2.S b/sysdeps/i386/i686/fpu/multiarch/s_sincosf-sse2.S deleted file mode 100644 index 71e222b930..0000000000 --- a/sysdeps/i386/i686/fpu/multiarch/s_sincosf-sse2.S +++ /dev/null @@ -1,585 +0,0 @@ -/* Optimized with sse2 version of sincosf - Copyright (C) 2012-2022 Free Software Foundation, Inc. - This file is part of the GNU C Library. - - The GNU C Library is free software; you can redistribute it and/or - modify it under the terms of the GNU Lesser General Public - License as published by the Free Software Foundation; either - version 2.1 of the License, or (at your option) any later version. - - The GNU C Library is distributed in the hope that it will be useful, - but WITHOUT ANY WARRANTY; without even the implied warranty of - MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU - Lesser General Public License for more details. - - You should have received a copy of the GNU Lesser General Public - License along with the GNU C Library; if not, see - . */ - -#include -#include - -/* Short algorithm description: - * - * 1) if |x|==0: sin(x)=x, - * cos(x)=1. - * 2) if |x|<2^-27: sin(x)=x-x*DP_SMALL, raising underflow only when needed, - * cos(x)=1-|x|. - * 3) if |x|<2^-5 : sin(x)=x+x*x^2*DP_SIN2_0+x^5*DP_SIN2_1, - * cos(x)=1+1*x^2*DP_COS2_0+x^5*DP_COS2_1 - * 4) if |x|< Pi/4: sin(x)=x+x*x^2*(S0+x^2*(S1+x^2*(S2+x^2*(S3+x^2*S4)))), - * cos(x)=1+1*x^2*(C0+x^2*(C1+x^2*(C2+x^2*(C3+x^2*C4)))). - * 5) if |x| < 9*Pi/4: - * 5.1) Range reduction: - * k=trunc(|x|/(Pi/4)), j=(k+1)&0x0e, n=k+1, t=|x|-j*Pi/4. - * 5.2) Reconstruction: - * sign_sin = sign(x) * (-1.0)^(( n >>2)&1) - * sign_cos = (-1.0)^(((n+2)>>2)&1) - * poly_sin = ((((S4*t^2 + S3)*t^2 + S2)*t^2 + S1)*t^2 + S0)*t^2*t+t - * poly_cos = ((((C4*t^2 + C3)*t^2 + C2)*t^2 + C1)*t^2 + C0)*t^2*s+s - * if(n&2 != 0) { - * using cos(t) and sin(t) polynomials for |t|= 2^23, very large args: - * 7.1) Range reduction: - * k=trunc(|x|/(Pi/4)), j=(k+1)&0xfffffffe, n=k+1, t=|x|-j*Pi/4. - * 7.2) Reconstruction same as (5.2). - * 8) if x is Inf, return x-x, and set errno=EDOM. - * 9) if x is NaN, return x-x. - * - * Special cases: - * sin/cos(+-0) = +-0/1 not raising inexact/underflow, - * sin/cos(subnormal) raises inexact/underflow, - * sin/cos(min_normalized) raises inexact/underflow, - * sin/cos(normalized) raises inexact, - * sin/cos(Inf) = NaN, raises invalid, sets errno to EDOM, - * sin/cos(NaN) = NaN. - */ - -#ifdef PIC -# define MO1(symbol) L(symbol)##@GOTOFF(%ebx) -# define MO2(symbol,reg2,_scale) L(symbol)##@GOTOFF(%ebx,reg2,_scale) -# define CFI_PUSH(REG) cfi_adjust_cfa_offset(4); cfi_rel_offset(REG,0) -# define CFI_POP(REG) cfi_adjust_cfa_offset(-4); cfi_restore(REG) -# define PUSH(REG) pushl REG; CFI_PUSH(REG) -# define POP(REG) popl REG; CFI_POP(REG) -# define ENTRANCE PUSH(%ebx); LOAD_PIC_REG(bx) -# define RETURN POP(%ebx); ret; CFI_PUSH(%ebx) -# define ARG_X 8(%esp) -# define ARG_SIN_PTR 12(%esp) -# define ARG_COS_PTR 16(%esp) -#else -# define MO1(symbol) L(symbol) -# define MO2(symbol,reg2,_scale) L(symbol)(,reg2,_scale) -# define ENTRANCE -# define RETURN ret -# define ARG_X 4(%esp) -# define ARG_SIN_PTR 8(%esp) -# define ARG_COS_PTR 12(%esp) -#endif - - .text -ENTRY(__sincosf_sse2) - /* Input: single precision x on stack at address ARG_X */ - /* pointer to sin result on stack at address ARG_SIN_PTR */ - /* pointer to cos result on stack at address ARG_COS_PTR */ - - ENTRANCE - movl ARG_X, %eax /* Bits of x */ - cvtss2sd ARG_X, %xmm0 /* DP x */ - andl $0x7fffffff, %eax /* |x| */ - - cmpl $0x3f490fdb, %eax /* |x|=Pi/4 */ - movd %eax, %xmm3 /* SP |x| */ - andpd MO1(DP_ABS_MASK),%xmm0 /* DP |x| */ - movss MO1(SP_INVPIO4), %xmm2 /* SP 1/(Pi/4) */ - - cmpl $0x40e231d6, %eax /* |x|<9*Pi/4 ? */ - jae L(large_args) - - /* Here if Pi/4<=|x|<9*Pi/4 */ - mulss %xmm3, %xmm2 /* SP |x|/(Pi/4) */ - movl ARG_X, %ecx /* Load x */ - cvttss2si %xmm2, %eax /* k, number of Pi/4 in x */ - shrl $29, %ecx /* (sign of x) << 2 */ - addl $1, %eax /* k+1 */ - movl $0x0e, %edx - andl %eax, %edx /* j = (k+1)&0x0e */ - subsd MO2(PIO4J,%edx,8), %xmm0/* t = |x| - j * Pi/4 */ - -L(reconstruction): - /* Input: %eax=n, %xmm0=t, %ecx=sign(x) */ - - movaps %xmm0, %xmm4 /* t */ - movhpd MO1(DP_ONES), %xmm4 /* 1|t */ - mulsd %xmm0, %xmm0 /* y=t^2 */ - movl $2, %edx - unpcklpd %xmm0, %xmm0 /* y|y */ - addl %eax, %edx /* k+2 */ - movaps %xmm0, %xmm1 /* y|y */ - mulpd %xmm0, %xmm0 /* z=t^4|z=t^4 */ - - movaps MO1(DP_SC4), %xmm2 /* S4 */ - mulpd %xmm0, %xmm2 /* z*S4 */ - movaps MO1(DP_SC3), %xmm3 /* S3 */ - mulpd %xmm0, %xmm3 /* z*S3 */ - xorl %eax, %ecx /* (sign_x ^ (k>>2))<<2 */ - addpd MO1(DP_SC2), %xmm2 /* S2+z*S4 */ - mulpd %xmm0, %xmm2 /* z*(S2+z*S4) */ - shrl $2, %edx /* (k+2)>>2 */ - addpd MO1(DP_SC1), %xmm3 /* S1+z*S3 */ - mulpd %xmm0, %xmm3 /* z*(S1+z*S3) */ - shrl $2, %ecx /* sign_x ^ k>>2 */ - addpd MO1(DP_SC0), %xmm2 /* S0+z*(S2+z*S4) */ - andl $1, %edx /* sign_cos = ((k+2)>>2)&1 */ - mulpd %xmm1, %xmm2 /* y*(S0+z*(S2+z*S4)) */ - andl $1, %ecx /* sign_sin = sign_x ^ ((k>>2)&1) */ - addpd %xmm2, %xmm3 /* y*(S0+y*(S1+y*(S2+y*(S3+y*S4)))) */ - mulpd %xmm4, %xmm3 /*t*y*(S0+y*(S1+y*(S2+y*(S3+y*S4))))*/ - testl $2, %eax /* n&2 != 0 ? */ - addpd %xmm4, %xmm3 /*t+t*y*(S0+y*(S1+y*(S2+y*(S3+y*S4))*/ - jnz L(sin_result_sin_poly) - -/*L(sin_result_cos_poly):*/ - /* - * Here if - * cos(x) = poly_sin * sign_cos - * sin(x) = poly_cos * sign_sin - */ - movsd MO2(DP_ONES,%ecx,8), %xmm4/* 0|sign_sin */ - movhpd MO2(DP_ONES,%edx,8), %xmm4/* sign_cos|sign_sin */ - mulpd %xmm4, %xmm3 /* result_cos|result_sin */ - movl ARG_SIN_PTR, %eax - cvtpd2ps %xmm3, %xmm0 /* SP results */ - movl ARG_COS_PTR, %ecx - movss %xmm0, (%eax) /* store sin(x) from xmm0[0] */ - shufps $1, %xmm0, %xmm0 /* move cos(x) to xmm0[0] */ - movss %xmm0, (%ecx) /* store cos(x) */ - RETURN - - .p2align 4 -L(sin_result_sin_poly): - /* - * Here if - * sin(x) = poly_sin * sign_sin - * cos(x) = poly_cos * sign_cos - */ - movsd MO2(DP_ONES,%edx,8), %xmm4/* 0|sign_cos */ - movhpd MO2(DP_ONES,%ecx,8), %xmm4/* sign_sin|sign_cos */ - mulpd %xmm4, %xmm3 /* result_sin|result_cos */ - movl ARG_SIN_PTR, %eax - cvtpd2ps %xmm3, %xmm0 /* SP results */ - movl ARG_COS_PTR, %ecx - movss %xmm0, (%ecx) /* store cos(x) from xmm0[0] */ - shufps $1, %xmm0, %xmm0 /* move sin(x) to xmm0[0] */ - movss %xmm0, (%eax) /* store sin(x) */ - RETURN - - .p2align 4 -L(large_args): - /* Here if |x|>=9*Pi/4 */ - cmpl $0x7f800000, %eax /* x is Inf or NaN ? */ - jae L(arg_inf_or_nan) - - /* Here if finite |x|>=9*Pi/4 */ - cmpl $0x4b000000, %eax /* |x|<2^23 ? */ - jae L(very_large_args) - - /* Here if 9*Pi/4<=|x|<2^23 */ - movsd MO1(DP_INVPIO4), %xmm1 /* 1/(Pi/4) */ - mulsd %xmm0, %xmm1 /* |x|/(Pi/4) */ - cvttsd2si %xmm1, %eax /* k=trunc(|x|/(Pi/4)) */ - addl $1, %eax /* k+1 */ - movl %eax, %edx - andl $0xfffffffe, %edx /* j=(k+1)&0xfffffffe */ - cvtsi2sdl %edx, %xmm4 /* DP j */ - movl ARG_X, %ecx /* Load x */ - movsd MO1(DP_PIO4HI), %xmm2 /* -PIO4HI = high part of -Pi/4 */ - shrl $29, %ecx /* (sign of x) << 2 */ - mulsd %xmm4, %xmm2 /* -j*PIO4HI */ - movsd MO1(DP_PIO4LO), %xmm3 /* -PIO4LO = low part of -Pi/4 */ - addsd %xmm2, %xmm0 /* |x| - j*PIO4HI */ - mulsd %xmm3, %xmm4 /* j*PIO4LO */ - addsd %xmm4, %xmm0 /* t = |x| - j*PIO4HI - j*PIO4LO */ - jmp L(reconstruction) - - .p2align 4 -L(very_large_args): - /* Here if finite |x|>=2^23 */ - - /* bitpos = (ix>>23) - BIAS_32 + 59; */ - shrl $23, %eax /* eb = biased exponent of x */ - subl $68, %eax /* bitpos=eb-0x7f+59, where 0x7f */ - /*is exponent bias */ - movl $28, %ecx /* %cl=28 */ - movl %eax, %edx /* bitpos copy */ - - /* j = bitpos/28; */ - div %cl /* j in register %al=%ax/%cl */ - movapd %xmm0, %xmm3 /* |x| */ - andl $0xff, %eax /* clear unneeded remainder from %ah*/ - - imull $28, %eax, %ecx /* j*28 */ - movsd MO1(DP_HI_MASK), %xmm4 /* DP_HI_MASK */ - movapd %xmm0, %xmm5 /* |x| */ - mulsd -2*8+MO2(_FPI,%eax,8), %xmm3/* tmp3 = FPI[j-2]*|x| */ - movapd %xmm0, %xmm1 /* |x| */ - mulsd -1*8+MO2(_FPI,%eax,8), %xmm5/* tmp2 = FPI[j-1]*|x| */ - mulsd 0*8+MO2(_FPI,%eax,8), %xmm0/* tmp0 = FPI[j]*|x| */ - addl $19, %ecx /* j*28+19 */ - mulsd 1*8+MO2(_FPI,%eax,8), %xmm1/* tmp1 = FPI[j+1]*|x| */ - cmpl %ecx, %edx /* bitpos>=j*28+19 ? */ - jl L(very_large_skip1) - - /* Here if bitpos>=j*28+19 */ - andpd %xmm3, %xmm4 /* HI(tmp3) */ - subsd %xmm4, %xmm3 /* tmp3 = tmp3 - HI(tmp3) */ -L(very_large_skip1): - - movsd MO1(DP_2POW52), %xmm6 - movapd %xmm5, %xmm2 /* tmp2 copy */ - addsd %xmm3, %xmm5 /* tmp5 = tmp3 + tmp2 */ - movl $1, %edx - addsd %xmm5, %xmm6 /* tmp6 = tmp5 + 2^52 */ - movsd 8+MO1(DP_2POW52), %xmm4 - movd %xmm6, %eax /* k = I64_LO(tmp6); */ - addsd %xmm6, %xmm4 /* tmp4 = tmp6 - 2^52 */ - movl ARG_X, %ecx /* Load x */ - comisd %xmm5, %xmm4 /* tmp4 > tmp5 ? */ - jbe L(very_large_skip2) - - /* Here if tmp4 > tmp5 */ - subl $1, %eax /* k-- */ - addsd 8+MO1(DP_ONES), %xmm4 /* tmp4 -= 1.0 */ -L(very_large_skip2): - - andl %eax, %edx /* k&1 */ - subsd %xmm4, %xmm3 /* tmp3 -= tmp4 */ - addsd MO2(DP_ZERONE,%edx,8), %xmm3/* t = DP_ZERONE[k&1] + tmp3 */ - addsd %xmm2, %xmm3 /* t += tmp2 */ - shrl $29, %ecx /* (sign of x) << 2 */ - addsd %xmm3, %xmm0 /* t += tmp0 */ - addl $1, %eax /* n=k+1 */ - addsd %xmm1, %xmm0 /* t += tmp1 */ - mulsd MO1(DP_PIO4), %xmm0 /* t *= PI04 */ - - jmp L(reconstruction) /* end of very_large_args peth */ - - .p2align 4 -L(arg_less_pio4): - /* Here if |x|