From nobody Sun Nov 24 07:51:35 2024 Delivered-To: importer@patchew.org Authentication-Results: mx.zohomail.com; dkim=pass; spf=pass (zohomail.com: domain of gnu.org designates 209.51.188.17 as permitted sender) smtp.mailfrom=qemu-devel-bounces+importer=patchew.org@nongnu.org; dmarc=pass(p=none dis=none) header.from=linaro.org ARC-Seal: i=1; a=rsa-sha256; t=1725541568; cv=none; d=zohomail.com; s=zohoarc; b=WLSjiZAHB1pqDRrgTJHMHiT674bB/cHsNWobUS/bOPoba9wXH+KqlLmolOFzZy5Xod/I4imFUY/q6jRRolVNZ7yI7JKNy8BsXlgFvxV77SNRmU8/Bdnc9NTJCSmxzEVqsXVHqC9FKhWLzzCvQeuzyyC9eQCrGR4GsEJpXGPg35Q= ARC-Message-Signature: i=1; a=rsa-sha256; c=relaxed/relaxed; d=zohomail.com; s=zohoarc; t=1725541568; h=Content-Transfer-Encoding:Date:Date:From:From:In-Reply-To:List-Subscribe:List-Post:List-Id:List-Archive:List-Help:List-Unsubscribe:MIME-Version:Message-ID:References:Sender:Subject:Subject:To:To:Message-Id:Reply-To:Cc; bh=g5k9uorZoB64FeJd18+6sOiVbENoo0GfinS46arCmIY=; b=O5G+fL5eBFiJhzRLRBcI/y9inX+3Oz2FiUlld3R6bL8BqImjCE8BUuf+x4HdBTIW4nOGTUuU9bmT3bEHHvhOT2rF2WsbsZU3xrKifJdaVItjJSspcCcl8wxOjnnrvm8aV9OLTJyr0XIT+VU8W0EPSZ1FF7qBxwBl2QuTJL1xACw= ARC-Authentication-Results: i=1; mx.zohomail.com; dkim=pass; spf=pass (zohomail.com: domain of gnu.org designates 209.51.188.17 as permitted sender) smtp.mailfrom=qemu-devel-bounces+importer=patchew.org@nongnu.org; dmarc=pass header.from= (p=none dis=none) Return-Path: Received: from lists.gnu.org (lists.gnu.org [209.51.188.17]) by mx.zohomail.com with SMTPS id 1725541568326360.15042177815974; Thu, 5 Sep 2024 06:06:08 -0700 (PDT) Received: from localhost ([::1] helo=lists1p.gnu.org) by lists.gnu.org with esmtp (Exim 4.90_1) (envelope-from ) id 1smC72-0001Lw-5Z; Thu, 05 Sep 2024 09:01:24 -0400 Received: from eggs.gnu.org ([2001:470:142:3::10]) by lists.gnu.org with esmtps (TLS1.2:ECDHE_RSA_AES_256_GCM_SHA384:256) (Exim 4.90_1) (envelope-from ) id 1smC6w-000137-DQ for qemu-devel@nongnu.org; Thu, 05 Sep 2024 09:01:19 -0400 Received: from mail-wm1-x32c.google.com ([2a00:1450:4864:20::32c]) by eggs.gnu.org with esmtps (TLS1.2:ECDHE_RSA_AES_128_GCM_SHA256:128) (Exim 4.90_1) (envelope-from ) id 1smC6s-0001jE-6T for qemu-devel@nongnu.org; Thu, 05 Sep 2024 09:01:18 -0400 Received: by mail-wm1-x32c.google.com with SMTP id 5b1f17b1804b1-42bac9469e8so5914745e9.3 for ; Thu, 05 Sep 2024 06:01:08 -0700 (PDT) Received: from orth.archaic.org.uk (orth.archaic.org.uk. [2001:8b0:1d0::2]) by smtp.gmail.com with ESMTPSA id 5b1f17b1804b1-42bb6e355dfsm232251365e9.46.2024.09.05.06.01.06 for (version=TLS1_3 cipher=TLS_AES_256_GCM_SHA384 bits=256/256); Thu, 05 Sep 2024 06:01:07 -0700 (PDT) DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=linaro.org; s=google; t=1725541268; x=1726146068; darn=nongnu.org; h=content-transfer-encoding:mime-version:references:in-reply-to :message-id:date:subject:to:from:from:to:cc:subject:date:message-id :reply-to; bh=g5k9uorZoB64FeJd18+6sOiVbENoo0GfinS46arCmIY=; b=IUUjiZPcTzX4UE8kCZ5cyEjymQ60F+TrIA0fxt9tXDNgYHhwzW9+lx2qgMF9zOgN0t 3fcQMLs7+pF6fbNOtbg3on91B5Y5y9TkGO0LVWBxOUpND4QvosHF1mEU3SDerhqA1z0W n8FE4fztwPwqnZqnrzzmMkaOwegEsvEaf9lQM4o79bC8cWJ/egw4OfjXiUyHRwmWZ1ps 9pNRzzAxN3YgxWouxeDpEHuUqU+wu0Yf7AG96JJchLqU32AoJX8AWjrO4KgJfos9rx93 yneB/J7EReiQg/+idOW7mbFdsMD8Gd68u9Y1KnAwy+zs6Zm10rhm1a13/tAjT0GZyNg5 U7JA== X-Google-DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=1e100.net; s=20230601; t=1725541268; x=1726146068; h=content-transfer-encoding:mime-version:references:in-reply-to :message-id:date:subject:to:from:x-gm-message-state:from:to:cc :subject:date:message-id:reply-to; bh=g5k9uorZoB64FeJd18+6sOiVbENoo0GfinS46arCmIY=; b=hu2Vmet7JQqLFDjyGJ6NYOOkWRbB1MpbD81E3GkWtikS8zU9tle/keYnilpiehrH5Z DAyzsG0NIY7vT7/IcwjYW8+4oGL5t9o0My37ismblWNzSYY7PIW4eOdIIqPRnqw/xZKX VfjAdfP6LJjKOypTpYs4WeRok6UUur9AH+kkcbC1PFCJlTv0r1nPagOxfiyldJxuu6tN npnjPVSiJ34iPG+OT/upDq7GHPnQYG8haqFlBQmwruwPVB5tl1d2erejLQqRFiGaVV5R YMTDOe8fODYwB/RLLTQiPTzOHmVFZc9zRcYe6re0e6cs/8fNfkxv2XitLP7wrhogQre9 ag7A== X-Gm-Message-State: AOJu0YynIYIyeCGuaZCUBe/Skl8UZm0pHR7HwVpYCoA6JxfLRuM6DKaG 6YTrWQ6OeOmW2TGNwjd5ZxHNil6CTLyCS17SSIzFzPThhhu++YzJ9qsrQ0IbsmKViWWOswBFsGe T X-Google-Smtp-Source: AGHT+IE5bzwX33KMXrSjNSD9GRFmUc3PpxdnL8ox2bVMonFFXz03qx7TbQlv3A7FAMJFBe1B66NXWw== X-Received: by 2002:a05:600c:3b08:b0:426:5f8f:51a4 with SMTP id 5b1f17b1804b1-42c954503femr43438655e9.12.1725541267345; Thu, 05 Sep 2024 06:01:07 -0700 (PDT) From: Peter Maydell To: qemu-devel@nongnu.org Subject: [PULL 06/25] target/arm: Prepare bfdotadd() callers for FEAT_EBF support Date: Thu, 5 Sep 2024 14:00:41 +0100 Message-Id: <20240905130100.298768-7-peter.maydell@linaro.org> X-Mailer: git-send-email 2.34.1 In-Reply-To: <20240905130100.298768-1-peter.maydell@linaro.org> References: <20240905130100.298768-1-peter.maydell@linaro.org> MIME-Version: 1.0 Content-Transfer-Encoding: quoted-printable Received-SPF: pass (zohomail.com: domain of gnu.org designates 209.51.188.17 as permitted sender) client-ip=209.51.188.17; envelope-from=qemu-devel-bounces+importer=patchew.org@nongnu.org; helo=lists.gnu.org; Received-SPF: pass client-ip=2a00:1450:4864:20::32c; envelope-from=peter.maydell@linaro.org; helo=mail-wm1-x32c.google.com X-Spam_score_int: -20 X-Spam_score: -2.1 X-Spam_bar: -- X-Spam_report: (-2.1 / 5.0 requ) BAYES_00=-1.9, DKIM_SIGNED=0.1, DKIM_VALID=-0.1, DKIM_VALID_AU=-0.1, DKIM_VALID_EF=-0.1, RCVD_IN_DNSWL_NONE=-0.0001, SPF_HELO_NONE=0.001, SPF_PASS=-0.001, T_SCC_BODY_TEXT_LINE=-0.01 autolearn=ham autolearn_force=no X-Spam_action: no action X-BeenThere: qemu-devel@nongnu.org X-Mailman-Version: 2.1.29 Precedence: list List-Id: List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Errors-To: qemu-devel-bounces+importer=patchew.org@nongnu.org Sender: qemu-devel-bounces+importer=patchew.org@nongnu.org X-ZohoMail-DKIM: pass (identity @linaro.org) X-ZM-MESSAGEID: 1725541570715116600 Content-Type: text/plain; charset="utf-8" We use bfdotadd() in four callsites for various helper functions. Currently this all assumes that we have the FPCR.EBF=3D0 semantics. For FPCR.EBF=3D1 we will need to: * call a different routine to bfdotadd() because we need to do a fused multiply-add rather than separate multiply and add steps * use a different float_status that honours the FPCR rounding mode and denormal-flushing fields * pass in an extra float_status that has been set up to perform round-to-odd rounding To prepare for this, refactor all the callsites so that instead of for (...) { x =3D bfdotadd(...); } they are: float_status fpst, fpst_odd; if (is_ebf(env, &fpst, &fpst_odd)) { for (...) { x =3D bfdotadd_ebf(..., &fpst, &fpst_odd); } } else { for (...) { x =3D bfdotadd(..., &fpst); } } For the moment the is_ebf() function always returns false, sets up fpst for EBF=3D0 semantics and never sets up fpst_odd; bfdotadd_ebf() will assert if called. We'll fill in the handling for EBF=3D1 in the next commit. This change should be a zero-behaviour-change refactor. Signed-off-by: Peter Maydell Reviewed-by: Richard Henderson --- target/arm/tcg/vec_internal.h | 37 ++++++++- target/arm/tcg/sme_helper.c | 74 ++++++++++++------ target/arm/tcg/vec_helper.c | 139 +++++++++++++++++++++++++--------- 3 files changed, 189 insertions(+), 61 deletions(-) diff --git a/target/arm/tcg/vec_internal.h b/target/arm/tcg/vec_internal.h index 3ca1b94ccf9..094f5c169ca 100644 --- a/target/arm/tcg/vec_internal.h +++ b/target/arm/tcg/vec_internal.h @@ -223,13 +223,46 @@ int64_t do_sqrdmlah_d(int64_t, int64_t, int64_t, bool= , bool); * bfdotadd: * @sum: addend * @e1, @e2: multiplicand vectors + * @fpst: floating-point status to use * * BFloat16 2-way dot product of @e1 & @e2, accumulating with @sum. * The @e1 and @e2 operands correspond to the 32-bit source vector * slots and contain two Bfloat16 values each. * - * Corresponds to the ARM pseudocode function BFDotAdd. + * Corresponds to the ARM pseudocode function BFDotAdd, specialized + * for the FPCR.EBF =3D=3D 0 case. */ -float32 bfdotadd(float32 sum, uint32_t e1, uint32_t e2); +float32 bfdotadd(float32 sum, uint32_t e1, uint32_t e2, float_status *fpst= ); +/** + * bfdotadd_ebf: + * @sum: addend + * @e1, @e2: multiplicand vectors + * @fpst: floating-point status to use + * @fpst_odd: floating-point status to use for round-to-odd operations + * + * BFloat16 2-way dot product of @e1 & @e2, accumulating with @sum. + * The @e1 and @e2 operands correspond to the 32-bit source vector + * slots and contain two Bfloat16 values each. + * + * Corresponds to the ARM pseudocode function BFDotAdd, specialized + * for the FPCR.EBF =3D=3D 1 case. + */ +float32 bfdotadd_ebf(float32 sum, uint32_t e1, uint32_t e2, + float_status *fpst, float_status *fpst_odd); + +/** + * is_ebf: + * @env: CPU state + * @statusp: pointer to floating point status to fill in + * @oddstatusp: pointer to floating point status to fill in for round-to-o= dd + * + * Determine whether a BFDotAdd operation should use FPCR.EBF =3D 0 + * or FPCR.EBF =3D 1 semantics. On return, has initialized *statusp + * and *oddstatusp to suitable float_status arguments to use with either + * bfdotadd() or bfdotadd_ebf(). + * Returns true for EBF =3D 1, false for EBF =3D 0. (The caller should use= this + * to decide whether to call bfdotadd() or bfdotadd_ebf().) + */ +bool is_ebf(CPUARMState *env, float_status *statusp, float_status *oddstat= usp); =20 #endif /* TARGET_ARM_VEC_INTERNAL_H */ diff --git a/target/arm/tcg/sme_helper.c b/target/arm/tcg/sme_helper.c index 289ffabbfbe..8cf12654e56 100644 --- a/target/arm/tcg/sme_helper.c +++ b/target/arm/tcg/sme_helper.c @@ -1085,32 +1085,62 @@ void HELPER(sme_bfmopa)(void *vza, void *vzn, void = *vzm, intptr_t row, col, oprsz =3D simd_maxsz(desc); uint32_t neg =3D simd_data(desc) * 0x80008000u; uint16_t *pn =3D vpn, *pm =3D vpm; + float_status fpst, fpst_odd; =20 - for (row =3D 0; row < oprsz; ) { - uint16_t prow =3D pn[H2(row >> 4)]; - do { - void *vza_row =3D vza + tile_vslice_offset(row); - uint32_t n =3D *(uint32_t *)(vzn + H1_4(row)); + if (is_ebf(env, &fpst, &fpst_odd)) { + for (row =3D 0; row < oprsz; ) { + uint16_t prow =3D pn[H2(row >> 4)]; + do { + void *vza_row =3D vza + tile_vslice_offset(row); + uint32_t n =3D *(uint32_t *)(vzn + H1_4(row)); =20 - n =3D f16mop_adj_pair(n, prow, neg); + n =3D f16mop_adj_pair(n, prow, neg); =20 - for (col =3D 0; col < oprsz; ) { - uint16_t pcol =3D pm[H2(col >> 4)]; - do { - if (prow & pcol & 0b0101) { - uint32_t *a =3D vza_row + H1_4(col); - uint32_t m =3D *(uint32_t *)(vzm + H1_4(col)); + for (col =3D 0; col < oprsz; ) { + uint16_t pcol =3D pm[H2(col >> 4)]; + do { + if (prow & pcol & 0b0101) { + uint32_t *a =3D vza_row + H1_4(col); + uint32_t m =3D *(uint32_t *)(vzm + H1_4(col)); =20 - m =3D f16mop_adj_pair(m, pcol, 0); - *a =3D bfdotadd(*a, n, m); - } - col +=3D 4; - pcol >>=3D 4; - } while (col & 15); - } - row +=3D 4; - prow >>=3D 4; - } while (row & 15); + m =3D f16mop_adj_pair(m, pcol, 0); + *a =3D bfdotadd_ebf(*a, n, m, &fpst, &fpst_odd= ); + } + col +=3D 4; + pcol >>=3D 4; + } while (col & 15); + } + row +=3D 4; + prow >>=3D 4; + } while (row & 15); + } + } else { + for (row =3D 0; row < oprsz; ) { + uint16_t prow =3D pn[H2(row >> 4)]; + do { + void *vza_row =3D vza + tile_vslice_offset(row); + uint32_t n =3D *(uint32_t *)(vzn + H1_4(row)); + + n =3D f16mop_adj_pair(n, prow, neg); + + for (col =3D 0; col < oprsz; ) { + uint16_t pcol =3D pm[H2(col >> 4)]; + do { + if (prow & pcol & 0b0101) { + uint32_t *a =3D vza_row + H1_4(col); + uint32_t m =3D *(uint32_t *)(vzm + H1_4(col)); + + m =3D f16mop_adj_pair(m, pcol, 0); + *a =3D bfdotadd(*a, n, m, &fpst); + } + col +=3D 4; + pcol >>=3D 4; + } while (col & 15); + } + row +=3D 4; + prow >>=3D 4; + } while (row & 15); + } } } =20 diff --git a/target/arm/tcg/vec_helper.c b/target/arm/tcg/vec_helper.c index 616ec54bb77..b0de74b55f1 100644 --- a/target/arm/tcg/vec_helper.c +++ b/target/arm/tcg/vec_helper.c @@ -2790,39 +2790,58 @@ DO_MMLA_B(gvec_usmmla_b, do_usmmla_b) * BFloat16 Dot Product */ =20 -float32 bfdotadd(float32 sum, uint32_t e1, uint32_t e2) +bool is_ebf(CPUARMState *env, float_status *statusp, float_status *oddstat= usp) { /* FPCR is ignored for BFDOT and BFMMLA. */ - float_status bf_status =3D { + *statusp =3D (float_status){ .tininess_before_rounding =3D float_tininess_before_rounding, .float_rounding_mode =3D float_round_to_odd_inf, .flush_to_zero =3D true, .flush_inputs_to_zero =3D true, .default_nan_mode =3D true, }; + + return false; +} + +float32 bfdotadd(float32 sum, uint32_t e1, uint32_t e2, float_status *fpst) +{ float32 t1, t2; =20 /* * Extract each BFloat16 from the element pair, and shift * them such that they become float32. */ - t1 =3D float32_mul(e1 << 16, e2 << 16, &bf_status); - t2 =3D float32_mul(e1 & 0xffff0000u, e2 & 0xffff0000u, &bf_status); - t1 =3D float32_add(t1, t2, &bf_status); - t1 =3D float32_add(sum, t1, &bf_status); + t1 =3D float32_mul(e1 << 16, e2 << 16, fpst); + t2 =3D float32_mul(e1 & 0xffff0000u, e2 & 0xffff0000u, fpst); + t1 =3D float32_add(t1, t2, fpst); + t1 =3D float32_add(sum, t1, fpst); =20 return t1; } =20 +float32 bfdotadd_ebf(float32 sum, uint32_t e1, uint32_t e2, + float_status *fpst, float_status *fpst_odd) +{ + g_assert_not_reached(); +} + void HELPER(gvec_bfdot)(void *vd, void *vn, void *vm, void *va, CPUARMState *env, uint32_t desc) { intptr_t i, opr_sz =3D simd_oprsz(desc); float32 *d =3D vd, *a =3D va; uint32_t *n =3D vn, *m =3D vm; + float_status fpst, fpst_odd; =20 - for (i =3D 0; i < opr_sz / 4; ++i) { - d[i] =3D bfdotadd(a[i], n[i], m[i]); + if (is_ebf(env, &fpst, &fpst_odd)) { + for (i =3D 0; i < opr_sz / 4; ++i) { + d[i] =3D bfdotadd_ebf(a[i], n[i], m[i], &fpst, &fpst_odd); + } + } else { + for (i =3D 0; i < opr_sz / 4; ++i) { + d[i] =3D bfdotadd(a[i], n[i], m[i], &fpst); + } } clear_tail(d, opr_sz, simd_maxsz(desc)); } @@ -2836,12 +2855,23 @@ void HELPER(gvec_bfdot_idx)(void *vd, void *vn, voi= d *vm, intptr_t eltspersegment =3D MIN(16 / 4, elements); float32 *d =3D vd, *a =3D va; uint32_t *n =3D vn, *m =3D vm; + float_status fpst, fpst_odd; =20 - for (i =3D 0; i < elements; i +=3D eltspersegment) { - uint32_t m_idx =3D m[i + H4(index)]; + if (is_ebf(env, &fpst, &fpst_odd)) { + for (i =3D 0; i < elements; i +=3D eltspersegment) { + uint32_t m_idx =3D m[i + H4(index)]; =20 - for (j =3D i; j < i + eltspersegment; j++) { - d[j] =3D bfdotadd(a[j], n[j], m_idx); + for (j =3D i; j < i + eltspersegment; j++) { + d[j] =3D bfdotadd_ebf(a[j], n[j], m_idx, &fpst, &fpst_odd); + } + } + } else { + for (i =3D 0; i < elements; i +=3D eltspersegment) { + uint32_t m_idx =3D m[i + H4(index)]; + + for (j =3D i; j < i + eltspersegment; j++) { + d[j] =3D bfdotadd(a[j], n[j], m_idx, &fpst); + } } } clear_tail(d, opr_sz, simd_maxsz(desc)); @@ -2853,37 +2883,72 @@ void HELPER(gvec_bfmmla)(void *vd, void *vn, void *= vm, void *va, intptr_t s, opr_sz =3D simd_oprsz(desc); float32 *d =3D vd, *a =3D va; uint32_t *n =3D vn, *m =3D vm; + float_status fpst, fpst_odd; =20 - for (s =3D 0; s < opr_sz / 4; s +=3D 4) { - float32 sum00, sum01, sum10, sum11; + if (is_ebf(env, &fpst, &fpst_odd)) { + for (s =3D 0; s < opr_sz / 4; s +=3D 4) { + float32 sum00, sum01, sum10, sum11; =20 - /* - * Process the entire segment at once, writing back the - * results only after we've consumed all of the inputs. - * - * Key to indices by column: - * i j i k j k - */ - sum00 =3D a[s + H4(0 + 0)]; - sum00 =3D bfdotadd(sum00, n[s + H4(0 + 0)], m[s + H4(0 + 0)]); - sum00 =3D bfdotadd(sum00, n[s + H4(0 + 1)], m[s + H4(0 + 1)]); + /* + * Process the entire segment at once, writing back the + * results only after we've consumed all of the inputs. + * + * Key to indices by column: + * i j i k j k + */ + sum00 =3D a[s + H4(0 + 0)]; + sum00 =3D bfdotadd_ebf(sum00, n[s + H4(0 + 0)], m[s + H4(0 + 0= )], &fpst, &fpst_odd); + sum00 =3D bfdotadd_ebf(sum00, n[s + H4(0 + 1)], m[s + H4(0 + 1= )], &fpst, &fpst_odd); =20 - sum01 =3D a[s + H4(0 + 1)]; - sum01 =3D bfdotadd(sum01, n[s + H4(0 + 0)], m[s + H4(2 + 0)]); - sum01 =3D bfdotadd(sum01, n[s + H4(0 + 1)], m[s + H4(2 + 1)]); + sum01 =3D a[s + H4(0 + 1)]; + sum01 =3D bfdotadd_ebf(sum01, n[s + H4(0 + 0)], m[s + H4(2 + 0= )], &fpst, &fpst_odd); + sum01 =3D bfdotadd_ebf(sum01, n[s + H4(0 + 1)], m[s + H4(2 + 1= )], &fpst, &fpst_odd); =20 - sum10 =3D a[s + H4(2 + 0)]; - sum10 =3D bfdotadd(sum10, n[s + H4(2 + 0)], m[s + H4(0 + 0)]); - sum10 =3D bfdotadd(sum10, n[s + H4(2 + 1)], m[s + H4(0 + 1)]); + sum10 =3D a[s + H4(2 + 0)]; + sum10 =3D bfdotadd_ebf(sum10, n[s + H4(2 + 0)], m[s + H4(0 + 0= )], &fpst, &fpst_odd); + sum10 =3D bfdotadd_ebf(sum10, n[s + H4(2 + 1)], m[s + H4(0 + 1= )], &fpst, &fpst_odd); =20 - sum11 =3D a[s + H4(2 + 1)]; - sum11 =3D bfdotadd(sum11, n[s + H4(2 + 0)], m[s + H4(2 + 0)]); - sum11 =3D bfdotadd(sum11, n[s + H4(2 + 1)], m[s + H4(2 + 1)]); + sum11 =3D a[s + H4(2 + 1)]; + sum11 =3D bfdotadd_ebf(sum11, n[s + H4(2 + 0)], m[s + H4(2 + 0= )], &fpst, &fpst_odd); + sum11 =3D bfdotadd_ebf(sum11, n[s + H4(2 + 1)], m[s + H4(2 + 1= )], &fpst, &fpst_odd); =20 - d[s + H4(0 + 0)] =3D sum00; - d[s + H4(0 + 1)] =3D sum01; - d[s + H4(2 + 0)] =3D sum10; - d[s + H4(2 + 1)] =3D sum11; + d[s + H4(0 + 0)] =3D sum00; + d[s + H4(0 + 1)] =3D sum01; + d[s + H4(2 + 0)] =3D sum10; + d[s + H4(2 + 1)] =3D sum11; + } + } else { + for (s =3D 0; s < opr_sz / 4; s +=3D 4) { + float32 sum00, sum01, sum10, sum11; + + /* + * Process the entire segment at once, writing back the + * results only after we've consumed all of the inputs. + * + * Key to indices by column: + * i j i k j k + */ + sum00 =3D a[s + H4(0 + 0)]; + sum00 =3D bfdotadd(sum00, n[s + H4(0 + 0)], m[s + H4(0 + 0)], = &fpst); + sum00 =3D bfdotadd(sum00, n[s + H4(0 + 1)], m[s + H4(0 + 1)], = &fpst); + + sum01 =3D a[s + H4(0 + 1)]; + sum01 =3D bfdotadd(sum01, n[s + H4(0 + 0)], m[s + H4(2 + 0)], = &fpst); + sum01 =3D bfdotadd(sum01, n[s + H4(0 + 1)], m[s + H4(2 + 1)], = &fpst); + + sum10 =3D a[s + H4(2 + 0)]; + sum10 =3D bfdotadd(sum10, n[s + H4(2 + 0)], m[s + H4(0 + 0)], = &fpst); + sum10 =3D bfdotadd(sum10, n[s + H4(2 + 1)], m[s + H4(0 + 1)], = &fpst); + + sum11 =3D a[s + H4(2 + 1)]; + sum11 =3D bfdotadd(sum11, n[s + H4(2 + 0)], m[s + H4(2 + 0)], = &fpst); + sum11 =3D bfdotadd(sum11, n[s + H4(2 + 1)], m[s + H4(2 + 1)], = &fpst); + + d[s + H4(0 + 0)] =3D sum00; + d[s + H4(0 + 1)] =3D sum01; + d[s + H4(2 + 0)] =3D sum10; + d[s + H4(2 + 1)] =3D sum11; + } } clear_tail(d, opr_sz, simd_maxsz(desc)); } --=20 2.34.1