From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: X-Spam-Checker-Version: SpamAssassin 3.4.0 (2014-02-07) on aws-us-west-2-korg-lkml-1.web.codeaurora.org Received: from bombadil.infradead.org (bombadil.infradead.org [198.137.202.133]) (using TLSv1.2 with cipher ECDHE-RSA-AES256-GCM-SHA384 (256/256 bits)) (No client certificate requested) by smtp.lore.kernel.org (Postfix) with ESMTPS id AF12AC5B572 for ; Fri, 14 Aug 2026 11:08:46 +0000 (UTC) DKIM-Signature: v=1; a=rsa-sha256; q=dns/txt; c=relaxed/relaxed; d=lists.infradead.org; s=bombadil.20210309; h=Sender: Content-Transfer-Encoding:Content-Type:List-Subscribe:List-Help:List-Post: List-Archive:List-Unsubscribe:List-Id:MIME-Version:References:In-Reply-To: Message-ID:Date:Subject:Cc:To:From:Reply-To:Content-ID:Content-Description: Resent-Date:Resent-From:Resent-Sender:Resent-To:Resent-Cc:Resent-Message-ID: List-Owner; bh=CRylBI5S5jQkYRfKtWWnHQefFatYues++fLF4mD4mtU=; b=tW7aevZ3JbG61F aevmAq1/vYt1CSl+sFAt+L1+lSNd9phvaAmWOfGlC0ssoQMly17NhaPYsSJdBuMRL9G7ry/EgiREn tdDGZ/nAv919Ov8PL6UUfzcgH7s8mEiyI+0XAz4jmfww2onYK5t08jpKGiMhiuZljWHHvRyr4chtR /GxUUDZiLSWQ9wAGUrtrFRQM69tDUV+pQGLW+RlePVCDDZQdf02UAZHYjnIDxgHf3kGFZh0cdOb/X SgsHndLS5uGtO9Xpr38gEPzTFDR/QzgXrLaE0zqn7BuJA/I0Yr2KW2wrfVLjlKbm8+6e1oCNS1zNT 54YJ21wCm29X71f2QCAQ==; Received: from localhost ([::1] helo=bombadil.infradead.org) by bombadil.infradead.org with esmtp (Exim 4.99.1 #2 (Red Hat Linux)) id 1wupmE-00000002ZIT-2ddq; Fri, 14 Aug 2026 11:08:42 +0000 Received: from mail-wm1-x335.google.com ([2a00:1450:4864:20::335]) by bombadil.infradead.org with esmtps (Exim 4.99.1 #2 (Red Hat Linux)) id 1wupmB-00000002ZHQ-3GJA for linux-rockchip@lists.infradead.org; Fri, 14 Aug 2026 11:08:41 +0000 Received: by mail-wm1-x335.google.com with SMTP id 5b1f17b1804b1-4954b3c5cbeso346355e9.1 for ; Fri, 14 Aug 2026 04:08:38 -0700 (PDT) DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=gmail.com; s=20251104; t=1786705717; x=1787310517; darn=lists.infradead.org; h=content-transfer-encoding:mime-version:references:in-reply-to :message-id:date:subject:cc:to:from:from:to:cc:subject:date :message-id:reply-to:content-type; bh=7S7sFuRCjatzs09JKRbH4HGwrINbYlULF/Unkzu0baA=; b=RDqq35jUsTAvM16xD3zfMwdy50gvnedHsFD7CS3h8F5zr5z/uVSGZDljtJEMIkxn0T OjaULCgwZqIOWJCww2Ce5riZpX+9EZ5lAUJaP/U4lXltxj0Ma9OYV9ZUD9O5Pyp6vjXA yex9pPoBQaeNyooGSdERn7+V7WJ5RIXXg9FhAsM+Vk09F/XAJiI9G9UoZcCwX3xSnLNI yYTiRcGZThvkQfBe2Jdao0ycXKb6UnY+v4xitxFFu5Kl8tCSb80FqBgXI1Bre+NiOVkE pKMnG9QAE0Kqri+et91L8tsgkCf0j/9lLMABXKXSZdp0PjQmO+RsWTI53jCYcRP8Cz92 luYQ== X-Google-DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=1e100.net; s=20251104; t=1786705717; x=1787310517; h=content-transfer-encoding:mime-version:references:in-reply-to :message-id:date:subject:cc:to:from:x-gm-gg:x-gm-message-state:from :to:cc:subject:date:message-id:reply-to:content-type; bh=7S7sFuRCjatzs09JKRbH4HGwrINbYlULF/Unkzu0baA=; b=LQ0Jwc/pn0fwEpVApTBkSDDTqdp3OWlkgzmdpfz4eh4Cvv++F3vhbZ7WDF9vZA6aN2 sALZrkaueCb/xfAyM/bqz447qOW4VcqIQIGHb6ab3VIopetKzHRca3cYGdX+llSNteQ/ a/bnLzkgetTV1ZK4p6l0+oifRT1ZeLobB5UcewHTxynSkAttZKRryX5fiWljfdnv3gnc L1BbeY7GnR3E3RZjAQIWrocbYcYwJJCRNR631/nipsSiJiCfoJu3U/X1OuySbt3nJcUJ LwLd3Yf1WgYwOQj1ET7ckfYN8Ew5lQLKTNmfqILRx7Rqxmw1WhbGXzyIvalj9NxRKbX2 NKXA== X-Forwarded-Encrypted: i=1; AHgh+RqKqPl/fY/nqDElVvAJD4L2WIjNC/g4s++lgnP3tgfWVufUL2BoPi4Q+lg4S9XhKlcLAmxvjJ+UmF8UTMYgzQ==@lists.infradead.org X-Gm-Message-State: AOJu0Yxj2gGCJ3mjAZC88mi1wEqzUaXKwl+JZbzpScWBiqcOOuppLfkK FbJ8VUSSCuIE9Fz9qA0m3Q6qYGQrJODdfb/VlH3mBjSp5J2ArCcZNVXc X-Gm-Gg: AR+sD13FohEILSZDw0Gc8gdi8bBH0ittdFq6JWBcanipfmaYGScDmZ89jG9vD/w8TDb IYP1d3xPPE7YPQ6G74f6KjIxno6jcMaT3hgP0CHbKOfEL8M0qupHNVAf67qnQkEjgfsEYFc19ri VGfuW9eTWIqHPEnX1G9nt1bAHCDOFU2BmxRPYNvr5s+ZRNEXbxz5RN3ccj4kdYnSOHgeUEGHZta BeqWSQXK8/YCJClE/jXW3RIpliWimjgGWnQxwCTdIz8d0XS2nrjVEKVkoLiX36KFDH2Yro7KnVu NfEz9zqN1aHi/lK6jiS+wAA1J+Wid0wTU+J8GlcGKA38Q0ihzvfvHUS7GnsVnGvARlrm//rnhjq k5Sg8oVlLZzsan6sImyJXP6n1Xm+6xJQ6rGNNCQxg0/xtMZK4vzl9YygEEBi+zUrlT6Q8YuHXBp Lj2lJlQZt1IAOPcQoEaTAOWNeTZNJ9IS+le3VwAXsjo6tidLkEv1oK71ZcyCOlF3Kcwx4a5sx0j NCvQKojSuorKpU2PWYFiBtPQDRBU59zYTZU71tJsOGDRZBTqqGUwZdhGJqy1uDn5BxfGg== X-Received: by 2002:a05:600c:1395:b0:499:77de:78ba with SMTP id 5b1f17b1804b1-499879826b2mr34045235e9.2.1786705717053; Fri, 14 Aug 2026 04:08:37 -0700 (PDT) Received: from OrangePi5-Plus.BB-HOME (20014C4E1B8D33002713106A8ABE6E8B.dsl.pool.telekom.hu. [2001:4c4e:1b8d:3300:2713:106a:8abe:6e8b]) by smtp.gmail.com with ESMTPSA id 5b1f17b1804b1-499877d7974sm37724265e9.1.2026.08.14.04.08.35 (version=TLS1_3 cipher=TLS_AES_256_GCM_SHA384 bits=256/256); Fri, 14 Aug 2026 04:08:35 -0700 (PDT) From: Igor Paunovic To: Jiaxing Hu Cc: Igor Paunovic , Tomeu Vizoso , =?UTF-8?q?Heiko=20St=C3=BCbner?= , Chaoyi Chen , alchark@flipper.net, dri-devel@lists.freedesktop.org, linux-rockchip@lists.infradead.org, linux-arm-kernel@lists.infradead.org, linux-kernel@vger.kernel.org Subject: Re: [PATCH v7 08/10] accel/rocket: add RK3576 NPU (RKNN) support Date: Fri, 14 Aug 2026 13:08:21 +0200 Message-ID: <20260814110825.14893-1-royalnet026@gmail.com> X-Mailer: git-send-email 2.53.0 In-Reply-To: <20260814082652.3852617-1-gahing@gahingwoo.com> References: <20260812094106.1391698-1-gahing@gahingwoo.com> <20260812094106.1391698-9-gahing@gahingwoo.com> <20260813095718.9747-1-royalnet026@gmail.com> <20260814082652.3852617-1-gahing@gahingwoo.com> MIME-Version: 1.0 X-CRM114-Version: 20100106-BlameMichelson ( TRE 0.9.0 (BSD) ) MR-646709E3 X-CRM114-CacheID: sfid-20260814_040839_833519_86AB602C X-CRM114-Status: UNSURE ( 9.23 ) X-CRM114-Notice: Please train this message. X-BeenThere: linux-rockchip@lists.infradead.org X-Mailman-Version: 2.1.34 Precedence: list List-Id: Upstream kernel work for Rockchip platforms List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Content-Type: text/plain; charset="us-ascii" Content-Transfer-Encoding: 7bit Sender: "Linux-rockchip" Errors-To: linux-rockchip-bounces+linux-rockchip=archiver.kernel.org@lists.infradead.org Hi Jiaxing, Here is the RK3588 column, all 27 operators, ROCKET_SEED=7, scored the way perch.py scores: a channel is good when its maxdiff (md below) against max(cpu, output zero point) is at most 1. Setup: Orange Pi 5 Plus (RK3588), kernel 7.2.0-rc6, the rocket driver from this kernel's tree rebuilt with my clocks-by-name and devfreq patches on top, Mesa at bf70ab68a21, teflon delegate, model mobilenet_v1_1_224_quant.tflite from the Mesa test suite (md5 4f348b87dca3315d2b3646cf5a3b31cf), per-operator models generated with the four byte output patch you described. The "correct hw" column is your chainmodel.py against the same model file. One difference to flag up front: against this model file chainmodel prints 36/256 for operator 8 where your table has 34/256, so our model files are not byte-identical, and the columns below should be read against each other rather than against your FINDINGS numbers. op kind correct hw RK3588 0 conv 32/32 md 1 32/32 md 1 1 depthwise 28/32 md 3 28/32 md 3 2 1x1 22/64 md 6 22/64 md 6 3 depthwise 21/64 md 13 21/64 md 13 4 1x1 18/128 md 14 8/128 md 15 5 depthwise 9/128 md 13 3/128 md 15 6 1x1 4/128 md 23 1/128 md 24 7 depthwise 7/128 md 10 3/128 md 12 8 1x1 36/256 md 7 20/256 md 17 9 depthwise 32/256 md 11 22/256 md 13 10 1x1 29/256 md 9 28/256 md 10 11 depthwise 53/256 md 8 48/256 md 11 12 1x1 166/512 md 11 121/512 md 13 13 depthwise 142/512 md 9 137/512 md 12 14 1x1 82/512 md 7 72/512 md 10 15 depthwise 154/512 md 10 115/512 md 15 16 1x1 82/512 md 7 71/512 md 10 17 depthwise 156/512 md 14 141/512 md 21 18 1x1 92/512 md 7 77/512 md 10 19 depthwise 170/512 md 7 152/512 md 12 20 1x1 102/512 md 8 84/512 md 10 21 depthwise 166/512 md 13 150/512 md 12 22 1x1 174/512 md 6 161/512 md 7 23 depthwise 293/512 md 5 261/512 md 7 24 1x1 671/1024 md 6 636/1024 md 5 25 depthwise 718/1024 md 9 684/1024 md 8 26 1x1 572/1024 md 25 574/1024 md 23 Three things stand out from here. Operators 0 through 3 score identically to your chain simulation -- same good-channel counts, same maxdiff -- including operator 3, the stride 2 depthwise with the asymmetric padding you suspect for the first RK3576 divergence. They are not byte-identical to the simulated hardware: diffing the raw tensors against the requant_hw chain shows a few hundred elements per surface already off by 1-4 at operators 0-3. That looks like the same small extra rounding difference that pushes the scores below your column from operator 4 on; through operator 3 it just stays under the maxdiff <= 1 scoring threshold. There is no md 255 anywhere. From operator 4 on, RK3588 sits somewhat below the simulation (8 vs 18 at op 4, 1 vs 4 at op 6), but the maxdiff never exceeds 24 across all 27 operators and the deep layers track the simulation closely (574 vs 572 at op 26). A control run with ROCKET_SEED=11 keeps the same character: operator 0 still 32/32, no saturated maxdiff anywhere, worst case md 35 at operator 6. So from the RK3588 side your read looks right: the deep-layer compounding is the reference artifact, and the RK3576 collapse from operator 4 with maxdiff 255 has no counterpart here. Igor _______________________________________________ Linux-rockchip mailing list Linux-rockchip@lists.infradead.org http://lists.infradead.org/mailman/listinfo/linux-rockchip From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: X-Spam-Checker-Version: SpamAssassin 3.4.0 (2014-02-07) on aws-us-west-2-korg-lkml-1.web.codeaurora.org Received: from gabe.freedesktop.org (gabe.freedesktop.org [131.252.210.177]) (using TLSv1.2 with cipher ECDHE-RSA-AES256-GCM-SHA384 (256/256 bits)) (No client certificate requested) by smtp.lore.kernel.org (Postfix) with ESMTPS id EEEF7C5B572 for ; Fri, 14 Aug 2026 11:08:40 +0000 (UTC) Received: from gabe.freedesktop.org (localhost [127.0.0.1]) by gabe.freedesktop.org (Postfix) with ESMTP id 10D4410E0F4; Fri, 14 Aug 2026 11:08:40 +0000 (UTC) Authentication-Results: gabe.freedesktop.org; dkim=pass (2048-bit key; unprotected) header.d=gmail.com header.i=@gmail.com header.b="d1rLy+rW"; dkim-atps=neutral Received: from mail-wm1-f41.google.com (mail-wm1-f41.google.com [209.85.128.41]) by gabe.freedesktop.org (Postfix) with ESMTPS id 05F1410E0F4 for ; Fri, 14 Aug 2026 11:08:38 +0000 (UTC) Received: by mail-wm1-f41.google.com with SMTP id 5b1f17b1804b1-49987367394so241945e9.0 for ; Fri, 14 Aug 2026 04:08:38 -0700 (PDT) DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=gmail.com; s=20251104; t=1786705717; x=1787310517; darn=lists.freedesktop.org; h=content-transfer-encoding:mime-version:references:in-reply-to :message-id:date:subject:cc:to:from:from:to:cc:subject:date :message-id:reply-to:content-type; bh=7S7sFuRCjatzs09JKRbH4HGwrINbYlULF/Unkzu0baA=; b=d1rLy+rW7FPupv0ISEYu/TyL/NCMmOEWM+7hGM43wrWuXui4ACNKf6XSoiwMzL0XQp x3/Os1cv+CTH1hBM/ZmQBstd/GXLq4UaT0br8sWP0etaCXcZ+ZN6eDwOrYdF8cLWVDtw aR/gfawlasmCKNTNtnrGCg9BQzEfNpa664/iopq4fc5rcNWqczJpeU9wb0gDI0ogBCl8 rekLLsq9fmn6OrmcCvuTYKyoWoL8A2DToILL0mYPSPeOpBNohZgjn+Ye3AZ/6KW5+jNL nlh67FHzkXRoGYzSPJWH3MoXFBDaQDT9n4nxFTNUKIpKcCJHs6EnpA+f3NdnfpxmTMcl kqLA== X-Google-DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=1e100.net; s=20251104; t=1786705717; x=1787310517; h=content-transfer-encoding:mime-version:references:in-reply-to :message-id:date:subject:cc:to:from:x-gm-gg:x-gm-message-state:from :to:cc:subject:date:message-id:reply-to:content-type; bh=7S7sFuRCjatzs09JKRbH4HGwrINbYlULF/Unkzu0baA=; b=DV35Ay0577UbWiuj352BeYxRm+U36vWPAFcFM/R2nUqChTuoMJLbYc5XL6fV5d1lkp KO84RPv1rBruaoZIgtzPMZOQARWNjMyxTqbp3qDE9qWZSGZyt3TrIwB9/XwXQWKmx7F6 QfEMyc8U1TNwg64dyChMKfOz9ty74R70RGz9jLZ2SobnbgsUSGDD+yoq96AUfIaEBpRj hJ5oAOLfYHZlx1l9PG7TFPiFVslh9qPfBbWYVb4X7i2opz7xRAdijbM8J+yaF5qr+nfz cyqTxu4yKDyQk9aYYb66e26DTVlu2ibNfLAt6XrpfRZrei66wKgpr9wiUufi4wDJPuVc gSPQ== X-Forwarded-Encrypted: i=1; AHgh+RosMk1qnwPYhxcv+mK/g+WMHpyUmWOvdXIprMNhBwKIPpwOO2LKu6qc2SSJRKAStwY9OQ2lvxBJgW0=@lists.freedesktop.org X-Gm-Message-State: AOJu0YxMjndpUPYigmUNUmmjRf1V/TYx/Fm+C9iNLkCSxYmo/L+jX7DR iX2JTX18hA0RL4rGF3F0SLFrxSF8Q/Z6L7hcE71sEFq9GXS4nfp32WED X-Gm-Gg: AR+sD13w3ZCajCRIAif4bZqSuzV158FB8rB8rWJoNP+0jTlwdcyTw4SbuavawPqfn/y 1L3E7++BpFvLpqeCRXwd5mn9mbfOJOj7TEHTCSvZK0jvhwGxouWBpEFDb07BovWt9p/FD8Sx9eQ c/eVTNolDZoG34oyc2C+iGOhn4GhbYZhn0weIxXdeSotbhzAmNqm4GtnDPwbvcB5CgRpL0eJkj3 SBzcNfLTFq+1RTTUUttiP3Dzqpi3TMW3Dla+m4WYbL79NL/XJqlavu4Ou+vZsDXIxCa+lfQLn5G yOZR1nbvIbzlgfP2C7hV9uG3rjHt+Xj0DB380N5utzlvlFll4GBqw2fnXYVcL/S6E3zJ3WbThTX aZUBlEm4lzNfpm8rOEjWGjx5OOY/sz/Pi4bN//2EsJ6gwXg3eJMBN4snbdQWI83lCz4/oCVpK7b jO38tCWGZm+uEQBiYN8dUDniN6BAzEF6zc0DlPyM9QBe9XZEGulcnwJ1saI9/SdLiFLAKaJBkCD CMvzYPsv00aG/RCJHScL9wYKUOhqzYjKPd6dOpdYB3pkL1BfdfZl3skUgAkQRwM/C7/9Q== X-Received: by 2002:a05:600c:1395:b0:499:77de:78ba with SMTP id 5b1f17b1804b1-499879826b2mr34045235e9.2.1786705717053; Fri, 14 Aug 2026 04:08:37 -0700 (PDT) Received: from OrangePi5-Plus.BB-HOME (20014C4E1B8D33002713106A8ABE6E8B.dsl.pool.telekom.hu. [2001:4c4e:1b8d:3300:2713:106a:8abe:6e8b]) by smtp.gmail.com with ESMTPSA id 5b1f17b1804b1-499877d7974sm37724265e9.1.2026.08.14.04.08.35 (version=TLS1_3 cipher=TLS_AES_256_GCM_SHA384 bits=256/256); Fri, 14 Aug 2026 04:08:35 -0700 (PDT) From: Igor Paunovic To: Jiaxing Hu Cc: Igor Paunovic , Tomeu Vizoso , =?UTF-8?q?Heiko=20St=C3=BCbner?= , Chaoyi Chen , alchark@flipper.net, dri-devel@lists.freedesktop.org, linux-rockchip@lists.infradead.org, linux-arm-kernel@lists.infradead.org, linux-kernel@vger.kernel.org Subject: Re: [PATCH v7 08/10] accel/rocket: add RK3576 NPU (RKNN) support Date: Fri, 14 Aug 2026 13:08:21 +0200 Message-ID: <20260814110825.14893-1-royalnet026@gmail.com> X-Mailer: git-send-email 2.53.0 In-Reply-To: <20260814082652.3852617-1-gahing@gahingwoo.com> References: <20260812094106.1391698-1-gahing@gahingwoo.com> <20260812094106.1391698-9-gahing@gahingwoo.com> <20260813095718.9747-1-royalnet026@gmail.com> <20260814082652.3852617-1-gahing@gahingwoo.com> MIME-Version: 1.0 Content-Transfer-Encoding: 8bit X-BeenThere: dri-devel@lists.freedesktop.org X-Mailman-Version: 2.1.29 Precedence: list List-Id: Direct Rendering Infrastructure - Development List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Errors-To: dri-devel-bounces@lists.freedesktop.org Sender: "dri-devel" Hi Jiaxing, Here is the RK3588 column, all 27 operators, ROCKET_SEED=7, scored the way perch.py scores: a channel is good when its maxdiff (md below) against max(cpu, output zero point) is at most 1. Setup: Orange Pi 5 Plus (RK3588), kernel 7.2.0-rc6, the rocket driver from this kernel's tree rebuilt with my clocks-by-name and devfreq patches on top, Mesa at bf70ab68a21, teflon delegate, model mobilenet_v1_1_224_quant.tflite from the Mesa test suite (md5 4f348b87dca3315d2b3646cf5a3b31cf), per-operator models generated with the four byte output patch you described. The "correct hw" column is your chainmodel.py against the same model file. One difference to flag up front: against this model file chainmodel prints 36/256 for operator 8 where your table has 34/256, so our model files are not byte-identical, and the columns below should be read against each other rather than against your FINDINGS numbers. op kind correct hw RK3588 0 conv 32/32 md 1 32/32 md 1 1 depthwise 28/32 md 3 28/32 md 3 2 1x1 22/64 md 6 22/64 md 6 3 depthwise 21/64 md 13 21/64 md 13 4 1x1 18/128 md 14 8/128 md 15 5 depthwise 9/128 md 13 3/128 md 15 6 1x1 4/128 md 23 1/128 md 24 7 depthwise 7/128 md 10 3/128 md 12 8 1x1 36/256 md 7 20/256 md 17 9 depthwise 32/256 md 11 22/256 md 13 10 1x1 29/256 md 9 28/256 md 10 11 depthwise 53/256 md 8 48/256 md 11 12 1x1 166/512 md 11 121/512 md 13 13 depthwise 142/512 md 9 137/512 md 12 14 1x1 82/512 md 7 72/512 md 10 15 depthwise 154/512 md 10 115/512 md 15 16 1x1 82/512 md 7 71/512 md 10 17 depthwise 156/512 md 14 141/512 md 21 18 1x1 92/512 md 7 77/512 md 10 19 depthwise 170/512 md 7 152/512 md 12 20 1x1 102/512 md 8 84/512 md 10 21 depthwise 166/512 md 13 150/512 md 12 22 1x1 174/512 md 6 161/512 md 7 23 depthwise 293/512 md 5 261/512 md 7 24 1x1 671/1024 md 6 636/1024 md 5 25 depthwise 718/1024 md 9 684/1024 md 8 26 1x1 572/1024 md 25 574/1024 md 23 Three things stand out from here. Operators 0 through 3 score identically to your chain simulation -- same good-channel counts, same maxdiff -- including operator 3, the stride 2 depthwise with the asymmetric padding you suspect for the first RK3576 divergence. They are not byte-identical to the simulated hardware: diffing the raw tensors against the requant_hw chain shows a few hundred elements per surface already off by 1-4 at operators 0-3. That looks like the same small extra rounding difference that pushes the scores below your column from operator 4 on; through operator 3 it just stays under the maxdiff <= 1 scoring threshold. There is no md 255 anywhere. From operator 4 on, RK3588 sits somewhat below the simulation (8 vs 18 at op 4, 1 vs 4 at op 6), but the maxdiff never exceeds 24 across all 27 operators and the deep layers track the simulation closely (574 vs 572 at op 26). A control run with ROCKET_SEED=11 keeps the same character: operator 0 still 32/32, no saturated maxdiff anywhere, worst case md 35 at operator 6. So from the RK3588 side your read looks right: the deep-layer compounding is the reference artifact, and the RK3576 collapse from operator 4 with maxdiff 255 has no counterpart here. Igor