From patchwork Wed Mar 8 10:00:48 2017 Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: 7bit X-Patchwork-Submitter: =?utf-8?q?Martin_Storsj=C3=B6?= X-Patchwork-Id: 2802 Delivered-To: ffmpegpatchwork@gmail.com Received: by 10.103.50.79 with SMTP id y76csp952957vsy; Wed, 8 Mar 2017 02:02:22 -0800 (PST) X-Received: by 10.223.175.238 with SMTP id y46mr4223161wrd.63.1488967342607; Wed, 08 Mar 2017 02:02:22 -0800 (PST) Return-Path: Received: from ffbox0-bg.mplayerhq.hu (ffbox0-bg.ffmpeg.org. [79.124.17.100]) by mx.google.com with ESMTP id s10si3693914wrb.43.2017.03.08.02.02.22; Wed, 08 Mar 2017 02:02:22 -0800 (PST) Received-SPF: pass (google.com: domain of ffmpeg-devel-bounces@ffmpeg.org designates 79.124.17.100 as permitted sender) client-ip=79.124.17.100; Authentication-Results: mx.google.com; dkim=neutral (body hash did not verify) header.i=@martin-st.20150623.gappssmtp.com; spf=pass (google.com: domain of ffmpeg-devel-bounces@ffmpeg.org designates 79.124.17.100 as permitted sender) smtp.mailfrom=ffmpeg-devel-bounces@ffmpeg.org Received: from [127.0.1.1] (localhost [127.0.0.1]) by ffbox0-bg.mplayerhq.hu (Postfix) with ESMTP id E226C6882B5; Wed, 8 Mar 2017 12:01:18 +0200 (EET) X-Original-To: ffmpeg-devel@ffmpeg.org Delivered-To: ffmpeg-devel@ffmpeg.org Received: from mail-lf0-f53.google.com (mail-lf0-f53.google.com [209.85.215.53]) by ffbox0-bg.mplayerhq.hu (Postfix) with ESMTPS id 95D496805A7 for ; Wed, 8 Mar 2017 12:01:10 +0200 (EET) Received: by mail-lf0-f53.google.com with SMTP id a6so12435225lfa.0 for ; Wed, 08 Mar 2017 02:01:24 -0800 (PST) DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=martin-st.20150623.gappssmtp.com; s=20150623; h=from:to:subject:date:message-id:in-reply-to:references; bh=+XnCGd3J+C0m9uGPzLypSQf5aY5Gwe6xFzlL7p2WRNw=; b=oInPWEYa8e7ZsQP5YWgkIN0F9A1WHo6Zphwevd0vAAYKCnL5R7IF1cgVVszHljjUVO B4wcPN9X3+Xp7kr6ZEHaL4fDoHjs1kneOyNpCE5l/gCz53FnKjvt9cXSopF+4gZXJRby 0eYigXS9npGJnPuMPIsBKTyqK6AgmYUp+UBamGH7cwyYZwyhRpd+LmwEFYi6L1s67D4U fGSpECnc25AwbFT3GbcWZVyZIEKv5fhrxZGXQelI9hpK61ggl2N9f0h5pMRcbz4+qf5K A+9ofoeMUc30j9qJHHZqRZlUjTOeDPSkg2g3r3OIOunOINN98fiyuuCgtBkRzEiUGxPS G0IQ== X-Google-DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=1e100.net; s=20161025; h=x-gm-message-state:from:to:subject:date:message-id:in-reply-to :references; bh=+XnCGd3J+C0m9uGPzLypSQf5aY5Gwe6xFzlL7p2WRNw=; b=XdssXbu1QgXt8oZhWRtpoy5yJ9MoXSwwsHfn0RRk46FK953QhTb+LF78MsLiVP1L1Y nwpWEs6XCKm7giz4qWzdqeWeepVf6e87ks9octCn2dz/aH1jegjACWOMYUO1VWG/CJR7 1QUQaNQksoOIoB3MX6W3ekXiV56JIoRMUVlHw7hi49Jv4/I6D4hiKFMQ27CtX3YM1bCN +YySmr8VYsICNfGDlc2dxlot5pAcrXbk1MUM6f9Fciur+v+vO+X2bt5GDIBByJXS5qGT wM2O+BDhno1EWrksIrIJOhecXsYaN0LIvKPtn6CeE8w9yqKF+vPHO3KXRD0+SdKYZiR9 qaBA== X-Gm-Message-State: AMke39kkXbqSOu8sid6z8/9XWIOifcYzwcwYfAyylWVwXVV5rfFRwZttWnN7UEeFzow6tg== X-Received: by 10.25.235.77 with SMTP id j74mr1427969lfh.60.1488967282928; Wed, 08 Mar 2017 02:01:22 -0800 (PST) Received: from localhost.localdomain ([2001:470:28:852:7d47:68e:13e8:4933]) by smtp.gmail.com with ESMTPSA id m127sm513064lfg.58.2017.03.08.02.01.22 for (version=TLS1_2 cipher=ECDHE-RSA-AES128-SHA bits=128/128); Wed, 08 Mar 2017 02:01:22 -0800 (PST) From: =?UTF-8?q?Martin=20Storsj=C3=B6?= To: ffmpeg-devel@ffmpeg.org Date: Wed, 8 Mar 2017 12:00:48 +0200 Message-Id: <1488967274-8143-8-git-send-email-martin@martin.st> X-Mailer: git-send-email 2.7.4 In-Reply-To: <1488967274-8143-1-git-send-email-martin@martin.st> References: <1488967274-8143-1-git-send-email-martin@martin.st> Subject: [FFmpeg-devel] [PATCH 08/34] aarch64: vp9itxfm: Do separate functions for half/quarter idct16 and idct32 X-BeenThere: ffmpeg-devel@ffmpeg.org X-Mailman-Version: 2.1.20 Precedence: list List-Id: FFmpeg development discussions and patches List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Reply-To: FFmpeg development discussions and patches MIME-Version: 1.0 Errors-To: ffmpeg-devel-bounces@ffmpeg.org Sender: "ffmpeg-devel" This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 14740 bytes to 24292 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: vp9_inv_dct_dct_16x16_sub1_add_neon: 236.7 vp9_inv_dct_dct_16x16_sub2_add_neon: 1051.0 vp9_inv_dct_dct_16x16_sub4_add_neon: 1051.0 vp9_inv_dct_dct_16x16_sub8_add_neon: 1051.0 vp9_inv_dct_dct_16x16_sub12_add_neon: 1387.4 vp9_inv_dct_dct_16x16_sub16_add_neon: 1387.6 vp9_inv_dct_dct_32x32_sub1_add_neon: 554.1 vp9_inv_dct_dct_32x32_sub2_add_neon: 5198.5 vp9_inv_dct_dct_32x32_sub4_add_neon: 5198.6 vp9_inv_dct_dct_32x32_sub8_add_neon: 5196.3 vp9_inv_dct_dct_32x32_sub12_add_neon: 6183.4 vp9_inv_dct_dct_32x32_sub16_add_neon: 6174.3 vp9_inv_dct_dct_32x32_sub20_add_neon: 7151.4 vp9_inv_dct_dct_32x32_sub24_add_neon: 7145.3 vp9_inv_dct_dct_32x32_sub28_add_neon: 8119.3 vp9_inv_dct_dct_32x32_sub32_add_neon: 8118.7 After: vp9_inv_dct_dct_16x16_sub1_add_neon: 236.7 vp9_inv_dct_dct_16x16_sub2_add_neon: 640.8 vp9_inv_dct_dct_16x16_sub4_add_neon: 639.0 vp9_inv_dct_dct_16x16_sub8_add_neon: 842.0 vp9_inv_dct_dct_16x16_sub12_add_neon: 1388.3 vp9_inv_dct_dct_16x16_sub16_add_neon: 1389.3 vp9_inv_dct_dct_32x32_sub1_add_neon: 554.1 vp9_inv_dct_dct_32x32_sub2_add_neon: 3685.5 vp9_inv_dct_dct_32x32_sub4_add_neon: 3685.1 vp9_inv_dct_dct_32x32_sub8_add_neon: 3684.4 vp9_inv_dct_dct_32x32_sub12_add_neon: 5312.2 vp9_inv_dct_dct_32x32_sub16_add_neon: 5315.4 vp9_inv_dct_dct_32x32_sub20_add_neon: 7154.9 vp9_inv_dct_dct_32x32_sub24_add_neon: 7154.5 vp9_inv_dct_dct_32x32_sub28_add_neon: 8126.6 vp9_inv_dct_dct_32x32_sub32_add_neon: 8127.2 This is cherrypicked from libav commit a63da4511d0fee66695ff4afd264ba1dbf1e812d. --- libavcodec/aarch64/vp9itxfm_neon.S | 525 ++++++++++++++++++++++++++++++++----- 1 file changed, 466 insertions(+), 59 deletions(-) diff --git a/libavcodec/aarch64/vp9itxfm_neon.S b/libavcodec/aarch64/vp9itxfm_neon.S index e45d385..3eb999a 100644 --- a/libavcodec/aarch64/vp9itxfm_neon.S +++ b/libavcodec/aarch64/vp9itxfm_neon.S @@ -75,6 +75,17 @@ endconst .endif .endm +// Same as dmbutterfly0 above, but treating the input in in2 as zero, +// writing the same output into both out1 and out2. +.macro dmbutterfly0_h out1, out2, in1, in2, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6 + smull \tmp1\().4s, \in1\().4h, v0.h[0] + smull2 \tmp2\().4s, \in1\().8h, v0.h[0] + rshrn \out1\().4h, \tmp1\().4s, #14 + rshrn2 \out1\().8h, \tmp2\().4s, #14 + rshrn \out2\().4h, \tmp1\().4s, #14 + rshrn2 \out2\().8h, \tmp2\().4s, #14 +.endm + // out1,out2 = in1 * coef1 - in2 * coef2 // out3,out4 = in1 * coef2 + in2 * coef1 // out are 4 x .4s registers, in are 2 x .8h registers @@ -104,6 +115,43 @@ endconst rshrn2 \inout2\().8h, \tmp4\().4s, #14 .endm +// Same as dmbutterfly above, but treating the input in inout2 as zero +.macro dmbutterfly_h1 inout1, inout2, coef1, coef2, tmp1, tmp2, tmp3, tmp4 + smull \tmp1\().4s, \inout1\().4h, \coef1 + smull2 \tmp2\().4s, \inout1\().8h, \coef1 + smull \tmp3\().4s, \inout1\().4h, \coef2 + smull2 \tmp4\().4s, \inout1\().8h, \coef2 + rshrn \inout1\().4h, \tmp1\().4s, #14 + rshrn2 \inout1\().8h, \tmp2\().4s, #14 + rshrn \inout2\().4h, \tmp3\().4s, #14 + rshrn2 \inout2\().8h, \tmp4\().4s, #14 +.endm + +// Same as dmbutterfly above, but treating the input in inout1 as zero +.macro dmbutterfly_h2 inout1, inout2, coef1, coef2, tmp1, tmp2, tmp3, tmp4 + smull \tmp1\().4s, \inout2\().4h, \coef2 + smull2 \tmp2\().4s, \inout2\().8h, \coef2 + smull \tmp3\().4s, \inout2\().4h, \coef1 + smull2 \tmp4\().4s, \inout2\().8h, \coef1 + neg \tmp1\().4s, \tmp1\().4s + neg \tmp2\().4s, \tmp2\().4s + rshrn \inout2\().4h, \tmp3\().4s, #14 + rshrn2 \inout2\().8h, \tmp4\().4s, #14 + rshrn \inout1\().4h, \tmp1\().4s, #14 + rshrn2 \inout1\().8h, \tmp2\().4s, #14 +.endm + +.macro dsmull_h out1, out2, in, coef + smull \out1\().4s, \in\().4h, \coef + smull2 \out2\().4s, \in\().8h, \coef +.endm + +.macro drshrn_h out, in1, in2, shift + rshrn \out\().4h, \in1\().4s, \shift + rshrn2 \out\().8h, \in2\().4s, \shift +.endm + + // out1 = in1 + in2 // out2 = in1 - in2 .macro butterfly_8h out1, out2, in1, in2 @@ -463,6 +511,30 @@ function idct16x16_dc_add_neon ret endfunc +.macro idct16_end + butterfly_8h v18, v7, v4, v7 // v18 = t0a, v7 = t7a + butterfly_8h v19, v22, v5, v22 // v19 = t1a, v22 = t6 + butterfly_8h v4, v26, v20, v26 // v4 = t2a, v26 = t5 + butterfly_8h v5, v6, v28, v6 // v5 = t3a, v6 = t4 + butterfly_8h v20, v28, v16, v24 // v20 = t8a, v28 = t11a + butterfly_8h v24, v21, v23, v21 // v24 = t9, v21 = t10 + butterfly_8h v23, v27, v25, v27 // v23 = t14, v27 = t13 + butterfly_8h v25, v29, v29, v17 // v25 = t15a, v29 = t12a + + dmbutterfly0 v2, v3, v27, v21, v2, v3, v16, v17, v30, v31 // v2 = t13a, v3 = t10a + dmbutterfly0 v28, v27, v29, v28, v21, v29, v16, v17, v30, v31 // v28 = t12, v27 = t11 + + butterfly_8h v16, v31, v18, v25 // v16 = out[0], v31 = out[15] + butterfly_8h v17, v30, v19, v23 // v17 = out[1], v30 = out[14] + butterfly_8h_r v25, v22, v22, v24 // v25 = out[9], v22 = out[6] + butterfly_8h v23, v24, v7, v20 // v23 = out[7], v24 = out[8] + butterfly_8h v18, v29, v4, v2 // v18 = out[2], v29 = out[13] + butterfly_8h v19, v28, v5, v28 // v19 = out[3], v28 = out[12] + butterfly_8h v20, v27, v6, v27 // v20 = out[4], v27 = out[11] + butterfly_8h v21, v26, v26, v3 // v21 = out[5], v26 = out[10] + ret +.endm + function idct16 dmbutterfly0 v16, v24, v16, v24, v2, v3, v4, v5, v6, v7 // v16 = t0a, v24 = t1a dmbutterfly v20, v28, v0.h[1], v0.h[2], v2, v3, v4, v5 // v20 = t2a, v28 = t3a @@ -485,28 +557,65 @@ function idct16 dmbutterfly0 v22, v26, v22, v26, v2, v3, v18, v19, v30, v31 // v22 = t6a, v26 = t5a dmbutterfly v23, v25, v0.h[1], v0.h[2], v18, v19, v30, v31 // v23 = t9a, v25 = t14a dmbutterfly v27, v21, v0.h[1], v0.h[2], v18, v19, v30, v31, neg=1 // v27 = t13a, v21 = t10a + idct16_end +endfunc - butterfly_8h v18, v7, v4, v7 // v18 = t0a, v7 = t7a - butterfly_8h v19, v22, v5, v22 // v19 = t1a, v22 = t6 - butterfly_8h v4, v26, v20, v26 // v4 = t2a, v26 = t5 - butterfly_8h v5, v6, v28, v6 // v5 = t3a, v6 = t4 - butterfly_8h v20, v28, v16, v24 // v20 = t8a, v28 = t11a - butterfly_8h v24, v21, v23, v21 // v24 = t9, v21 = t10 - butterfly_8h v23, v27, v25, v27 // v23 = t14, v27 = t13 - butterfly_8h v25, v29, v29, v17 // v25 = t15a, v29 = t12a +function idct16_half + dmbutterfly0_h v16, v24, v16, v24, v2, v3, v4, v5, v6, v7 // v16 = t0a, v24 = t1a + dmbutterfly_h1 v20, v28, v0.h[1], v0.h[2], v2, v3, v4, v5 // v20 = t2a, v28 = t3a + dmbutterfly_h1 v18, v30, v0.h[3], v0.h[4], v2, v3, v4, v5 // v18 = t4a, v30 = t7a + dmbutterfly_h2 v26, v22, v0.h[5], v0.h[6], v2, v3, v4, v5 // v26 = t5a, v22 = t6a + dmbutterfly_h1 v17, v31, v0.h[7], v1.h[0], v2, v3, v4, v5 // v17 = t8a, v31 = t15a + dmbutterfly_h2 v25, v23, v1.h[1], v1.h[2], v2, v3, v4, v5 // v25 = t9a, v23 = t14a + dmbutterfly_h1 v21, v27, v1.h[3], v1.h[4], v2, v3, v4, v5 // v21 = t10a, v27 = t13a + dmbutterfly_h2 v29, v19, v1.h[5], v1.h[6], v2, v3, v4, v5 // v29 = t11a, v19 = t12a - dmbutterfly0 v2, v3, v27, v21, v2, v3, v16, v17, v30, v31 // v2 = t13a, v3 = t10a - dmbutterfly0 v28, v27, v29, v28, v21, v29, v16, v17, v30, v31 // v28 = t12, v27 = t11 + butterfly_8h v4, v28, v16, v28 // v4 = t0, v28 = t3 + butterfly_8h v5, v20, v24, v20 // v5 = t1, v20 = t2 + butterfly_8h v6, v26, v18, v26 // v6 = t4, v26 = t5 + butterfly_8h v7, v22, v30, v22 // v7 = t7, v22 = t6 + butterfly_8h v16, v25, v17, v25 // v16 = t8, v25 = t9 + butterfly_8h v24, v21, v29, v21 // v24 = t11, v21 = t10 + butterfly_8h v17, v27, v19, v27 // v17 = t12, v27 = t13 + butterfly_8h v29, v23, v31, v23 // v29 = t15, v23 = t14 - butterfly_8h v16, v31, v18, v25 // v16 = out[0], v31 = out[15] - butterfly_8h v17, v30, v19, v23 // v17 = out[1], v30 = out[14] - butterfly_8h_r v25, v22, v22, v24 // v25 = out[9], v22 = out[6] - butterfly_8h v23, v24, v7, v20 // v23 = out[7], v24 = out[8] - butterfly_8h v18, v29, v4, v2 // v18 = out[2], v29 = out[13] - butterfly_8h v19, v28, v5, v28 // v19 = out[3], v28 = out[12] - butterfly_8h v20, v27, v6, v27 // v20 = out[4], v27 = out[11] - butterfly_8h v21, v26, v26, v3 // v21 = out[5], v26 = out[10] - ret + dmbutterfly0 v22, v26, v22, v26, v2, v3, v18, v19, v30, v31 // v22 = t6a, v26 = t5a + dmbutterfly v23, v25, v0.h[1], v0.h[2], v18, v19, v30, v31 // v23 = t9a, v25 = t14a + dmbutterfly v27, v21, v0.h[1], v0.h[2], v18, v19, v30, v31, neg=1 // v27 = t13a, v21 = t10a + idct16_end +endfunc + +function idct16_quarter + dsmull_h v24, v25, v19, v1.h[6] + dsmull_h v4, v5, v17, v0.h[7] + dsmull_h v7, v6, v18, v0.h[4] + dsmull_h v30, v31, v18, v0.h[3] + neg v24.4s, v24.4s + neg v25.4s, v25.4s + dsmull_h v29, v28, v17, v1.h[0] + dsmull_h v26, v27, v19, v1.h[5] + dsmull_h v22, v23, v16, v0.h[0] + drshrn_h v24, v24, v25, #14 + drshrn_h v16, v4, v5, #14 + drshrn_h v7, v7, v6, #14 + drshrn_h v6, v30, v31, #14 + drshrn_h v29, v29, v28, #14 + drshrn_h v17, v26, v27, #14 + drshrn_h v28, v22, v23, #14 + + dmbutterfly_l v20, v21, v22, v23, v17, v24, v0.h[1], v0.h[2] + dmbutterfly_l v18, v19, v30, v31, v29, v16, v0.h[1], v0.h[2] + neg v22.4s, v22.4s + neg v23.4s, v23.4s + drshrn_h v27, v20, v21, #14 + drshrn_h v21, v22, v23, #14 + drshrn_h v23, v18, v19, #14 + drshrn_h v25, v30, v31, #14 + mov v4.16b, v28.16b + mov v5.16b, v28.16b + dmbutterfly0 v22, v26, v7, v6, v18, v19, v30, v31 + mov v20.16b, v28.16b + idct16_end endfunc function iadst16 @@ -756,6 +865,13 @@ function ff_vp9_\txfm1\()_\txfm2\()_16x16_add_neon, export=1 .endif mov x9, #32 +.ifc \txfm1\()_\txfm2,idct_idct + cmp w3, #10 + b.le idct16x16_quarter_add_neon + cmp w3, #38 + b.le idct16x16_half_add_neon +.endif + .irp i, 0, 8 add x0, sp, #(\i*32) .ifc \txfm1\()_\txfm2,idct_idct @@ -812,6 +928,116 @@ itxfm_func16x16 iadst, idct itxfm_func16x16 idct, iadst itxfm_func16x16 iadst, iadst +function idct16_1d_8x16_pass1_quarter_neon + mov x14, x30 + movi v2.8h, #0 +.irp i, 16, 17, 18, 19 + load_clear \i, x2, x9 +.endr + + bl idct16_quarter + + // Do two 8x8 transposes. Originally, v16-v31 contain the + // 16 rows. Afterwards, v16-v23 and v24-v31 contain the two + // transposed 8x8 blocks. + transpose_8x8H v16, v17, v18, v19, v20, v21, v22, v23, v2, v3 + transpose_8x8H v24, v25, v26, v27, v28, v29, v30, v31, v2, v3 + + // Store the transposed 8x8 blocks horizontally. + // The first 8x8 block is kept in registers for the second pass, + // store the rest in the temp buffer. + // Since only a 4x4 part of the input was nonzero, this means that + // only 4 rows are nonzero after transposing, and the second pass + // only reads the topmost 4 rows. Therefore only store the topmost + // 4 rows. + add x0, x0, #16 +.irp i, 24, 25, 26, 27 + store \i, x0, x9 +.endr + br x14 +endfunc + +function idct16_1d_8x16_pass2_quarter_neon + mov x14, x30 + cbz x3, 1f +.irp i, 16, 17, 18, 19 + load \i, x2, x9 +.endr +1: + + add x3, x0, x1 + lsl x1, x1, #1 + bl idct16_quarter + + load_add_store v16.8h, v17.8h, v18.8h, v19.8h, v20.8h, v21.8h, v22.8h, v23.8h, v16.8b, v17.8b + load_add_store v24.8h, v25.8h, v26.8h, v27.8h, v28.8h, v29.8h, v30.8h, v31.8h, v16.8b, v17.8b + + br x14 +endfunc + +function idct16_1d_8x16_pass1_half_neon + mov x14, x30 + movi v2.8h, #0 +.irp i, 16, 17, 18, 19, 20, 21, 22, 23 + load_clear \i, x2, x9 +.endr + + bl idct16_half + + // Do two 8x8 transposes. Originally, v16-v31 contain the + // 16 rows. Afterwards, v16-v23 and v24-v31 contain the two + // transposed 8x8 blocks. + transpose_8x8H v16, v17, v18, v19, v20, v21, v22, v23, v2, v3 + transpose_8x8H v24, v25, v26, v27, v28, v29, v30, v31, v2, v3 + + // Store the transposed 8x8 blocks horizontally. + // The first 8x8 block is kept in registers for the second pass, + // store the rest in the temp buffer. + add x0, x0, #16 +.irp i, 24, 25, 26, 27, 28, 29, 30, 31 + store \i, x0, x9 +.endr + br x14 +endfunc + +function idct16_1d_8x16_pass2_half_neon + mov x14, x30 + cbz x3, 1f +.irp i, 16, 17, 18, 19, 20, 21, 22, 23 + load \i, x2, x9 +.endr +1: + + add x3, x0, x1 + lsl x1, x1, #1 + bl idct16_half + + load_add_store v16.8h, v17.8h, v18.8h, v19.8h, v20.8h, v21.8h, v22.8h, v23.8h, v16.8b, v17.8b + load_add_store v24.8h, v25.8h, v26.8h, v27.8h, v28.8h, v29.8h, v30.8h, v31.8h, v16.8b, v17.8b + + br x14 +endfunc + +.macro idct16_partial size +function idct16x16_\size\()_add_neon + add x0, sp, #(0*32) + add x2, x6, #(0*2) + bl idct16_1d_8x16_pass1_\size\()_neon +.irp i, 0, 8 + add x0, x4, #(\i) + mov x1, x5 + add x2, sp, #(\i*2) + mov x3, #\i + bl idct16_1d_8x16_pass2_\size\()_neon +.endr + + add sp, sp, #512 + br x15 +endfunc +.endm + +idct16_partial quarter +idct16_partial half function idct32x32_dc_add_neon movrel x4, idct_coeffs @@ -848,6 +1074,37 @@ function idct32x32_dc_add_neon ret endfunc +.macro idct32_end + butterfly_8h v16, v5, v4, v5 // v16 = t16a, v5 = t19a + butterfly_8h v17, v20, v23, v20 // v17 = t17, v20 = t18 + butterfly_8h v18, v6, v7, v6 // v18 = t23a, v6 = t20a + butterfly_8h v19, v21, v22, v21 // v19 = t22, v21 = t21 + butterfly_8h v4, v28, v28, v30 // v4 = t24a, v28 = t27a + butterfly_8h v23, v26, v25, v26 // v23 = t25, v26 = t26 + butterfly_8h v7, v3, v29, v31 // v7 = t31a, v3 = t28a + butterfly_8h v22, v27, v24, v27 // v22 = t30, v27 = t29 + + dmbutterfly v27, v20, v0.h[1], v0.h[2], v24, v25, v30, v31 // v27 = t18a, v20 = t29a + dmbutterfly v3, v5, v0.h[1], v0.h[2], v24, v25, v30, v31 // v3 = t19, v5 = t28 + dmbutterfly v28, v6, v0.h[1], v0.h[2], v24, v25, v30, v31, neg=1 // v28 = t27, v6 = t20 + dmbutterfly v26, v21, v0.h[1], v0.h[2], v24, v25, v30, v31, neg=1 // v26 = t26a, v21 = t21a + + butterfly_8h v31, v24, v7, v4 // v31 = t31, v24 = t24 + butterfly_8h v30, v25, v22, v23 // v30 = t30a, v25 = t25a + butterfly_8h_r v23, v16, v16, v18 // v23 = t23, v16 = t16 + butterfly_8h_r v22, v17, v17, v19 // v22 = t22a, v17 = t17a + butterfly_8h v18, v21, v27, v21 // v18 = t18, v21 = t21 + butterfly_8h_r v27, v28, v5, v28 // v27 = t27a, v28 = t28a + butterfly_8h v29, v26, v20, v26 // v29 = t29, v26 = t26 + butterfly_8h v19, v20, v3, v6 // v19 = t19a, v20 = t20 + + dmbutterfly0 v27, v20, v27, v20, v2, v3, v4, v5, v6, v7 // v27 = t27, v20 = t20 + dmbutterfly0 v26, v21, v26, v21, v2, v3, v4, v5, v6, v7 // v26 = t26a, v21 = t21a + dmbutterfly0 v25, v22, v25, v22, v2, v3, v4, v5, v6, v7 // v25 = t25, v22 = t22 + dmbutterfly0 v24, v23, v24, v23, v2, v3, v4, v5, v6, v7 // v24 = t24a, v23 = t23a + ret +.endm + function idct32_odd ld1 {v0.8h,v1.8h}, [x11] @@ -875,37 +1132,88 @@ function idct32_odd dmbutterfly v27, v20, v0.h[3], v0.h[4], v16, v17, v18, v19, neg=1 // v27 = t29a, v20 = t18a dmbutterfly v21, v26, v0.h[5], v0.h[6], v16, v17, v18, v19 // v21 = t21a, v26 = t26a dmbutterfly v25, v22, v0.h[5], v0.h[6], v16, v17, v18, v19, neg=1 // v25 = t25a, v22 = t22a + idct32_end +endfunc - butterfly_8h v16, v5, v4, v5 // v16 = t16a, v5 = t19a - butterfly_8h v17, v20, v23, v20 // v17 = t17, v20 = t18 - butterfly_8h v18, v6, v7, v6 // v18 = t23a, v6 = t20a - butterfly_8h v19, v21, v22, v21 // v19 = t22, v21 = t21 - butterfly_8h v4, v28, v28, v30 // v4 = t24a, v28 = t27a - butterfly_8h v23, v26, v25, v26 // v23 = t25, v26 = t26 - butterfly_8h v7, v3, v29, v31 // v7 = t31a, v3 = t28a - butterfly_8h v22, v27, v24, v27 // v22 = t30, v27 = t29 +function idct32_odd_half + ld1 {v0.8h,v1.8h}, [x11] - dmbutterfly v27, v20, v0.h[1], v0.h[2], v24, v25, v30, v31 // v27 = t18a, v20 = t29a - dmbutterfly v3, v5, v0.h[1], v0.h[2], v24, v25, v30, v31 // v3 = t19, v5 = t28 - dmbutterfly v28, v6, v0.h[1], v0.h[2], v24, v25, v30, v31, neg=1 // v28 = t27, v6 = t20 - dmbutterfly v26, v21, v0.h[1], v0.h[2], v24, v25, v30, v31, neg=1 // v26 = t26a, v21 = t21a + dmbutterfly_h1 v16, v31, v0.h[0], v0.h[1], v4, v5, v6, v7 // v16 = t16a, v31 = t31a + dmbutterfly_h2 v24, v23, v0.h[2], v0.h[3], v4, v5, v6, v7 // v24 = t17a, v23 = t30a + dmbutterfly_h1 v20, v27, v0.h[4], v0.h[5], v4, v5, v6, v7 // v20 = t18a, v27 = t29a + dmbutterfly_h2 v28, v19, v0.h[6], v0.h[7], v4, v5, v6, v7 // v28 = t19a, v19 = t28a + dmbutterfly_h1 v18, v29, v1.h[0], v1.h[1], v4, v5, v6, v7 // v18 = t20a, v29 = t27a + dmbutterfly_h2 v26, v21, v1.h[2], v1.h[3], v4, v5, v6, v7 // v26 = t21a, v21 = t26a + dmbutterfly_h1 v22, v25, v1.h[4], v1.h[5], v4, v5, v6, v7 // v22 = t22a, v25 = t25a + dmbutterfly_h2 v30, v17, v1.h[6], v1.h[7], v4, v5, v6, v7 // v30 = t23a, v17 = t24a - butterfly_8h v31, v24, v7, v4 // v31 = t31, v24 = t24 - butterfly_8h v30, v25, v22, v23 // v30 = t30a, v25 = t25a - butterfly_8h_r v23, v16, v16, v18 // v23 = t23, v16 = t16 - butterfly_8h_r v22, v17, v17, v19 // v22 = t22a, v17 = t17a - butterfly_8h v18, v21, v27, v21 // v18 = t18, v21 = t21 - butterfly_8h_r v27, v28, v5, v28 // v27 = t27a, v28 = t28a - butterfly_8h v29, v26, v20, v26 // v29 = t29, v26 = t26 - butterfly_8h v19, v20, v3, v6 // v19 = t19a, v20 = t20 + ld1 {v0.8h}, [x10] - dmbutterfly0 v27, v20, v27, v20, v2, v3, v4, v5, v6, v7 // v27 = t27, v20 = t20 - dmbutterfly0 v26, v21, v26, v21, v2, v3, v4, v5, v6, v7 // v26 = t26a, v21 = t21a - dmbutterfly0 v25, v22, v25, v22, v2, v3, v4, v5, v6, v7 // v25 = t25, v22 = t22 - dmbutterfly0 v24, v23, v24, v23, v2, v3, v4, v5, v6, v7 // v24 = t24a, v23 = t23a - ret + butterfly_8h v4, v24, v16, v24 // v4 = t16, v24 = t17 + butterfly_8h v5, v20, v28, v20 // v5 = t19, v20 = t18 + butterfly_8h v6, v26, v18, v26 // v6 = t20, v26 = t21 + butterfly_8h v7, v22, v30, v22 // v7 = t23, v22 = t22 + butterfly_8h v28, v25, v17, v25 // v28 = t24, v25 = t25 + butterfly_8h v30, v21, v29, v21 // v30 = t27, v21 = t26 + butterfly_8h v29, v23, v31, v23 // v29 = t31, v23 = t30 + butterfly_8h v31, v27, v19, v27 // v31 = t28, v27 = t29 + + dmbutterfly v23, v24, v0.h[3], v0.h[4], v16, v17, v18, v19 // v23 = t17a, v24 = t30a + dmbutterfly v27, v20, v0.h[3], v0.h[4], v16, v17, v18, v19, neg=1 // v27 = t29a, v20 = t18a + dmbutterfly v21, v26, v0.h[5], v0.h[6], v16, v17, v18, v19 // v21 = t21a, v26 = t26a + dmbutterfly v25, v22, v0.h[5], v0.h[6], v16, v17, v18, v19, neg=1 // v25 = t25a, v22 = t22a + idct32_end endfunc +function idct32_odd_quarter + ld1 {v0.8h,v1.8h}, [x11] + + dsmull_h v4, v5, v16, v0.h[0] + dsmull_h v28, v29, v19, v0.h[7] + dsmull_h v30, v31, v16, v0.h[1] + dsmull_h v22, v23, v17, v1.h[6] + dsmull_h v7, v6, v17, v1.h[7] + dsmull_h v26, v27, v19, v0.h[6] + dsmull_h v20, v21, v18, v1.h[0] + dsmull_h v24, v25, v18, v1.h[1] + + ld1 {v0.8h}, [x10] + + neg v28.4s, v28.4s + neg v29.4s, v29.4s + neg v7.4s, v7.4s + neg v6.4s, v6.4s + + drshrn_h v4, v4, v5, #14 + drshrn_h v5, v28, v29, #14 + drshrn_h v29, v30, v31, #14 + drshrn_h v28, v22, v23, #14 + drshrn_h v7, v7, v6, #14 + drshrn_h v31, v26, v27, #14 + drshrn_h v6, v20, v21, #14 + drshrn_h v30, v24, v25, #14 + + dmbutterfly_l v16, v17, v18, v19, v29, v4, v0.h[3], v0.h[4] + dmbutterfly_l v27, v26, v20, v21, v31, v5, v0.h[3], v0.h[4] + drshrn_h v23, v16, v17, #14 + drshrn_h v24, v18, v19, #14 + neg v20.4s, v20.4s + neg v21.4s, v21.4s + drshrn_h v27, v27, v26, #14 + drshrn_h v20, v20, v21, #14 + dmbutterfly_l v16, v17, v18, v19, v30, v6, v0.h[5], v0.h[6] + drshrn_h v21, v16, v17, #14 + drshrn_h v26, v18, v19, #14 + dmbutterfly_l v16, v17, v18, v19, v28, v7, v0.h[5], v0.h[6] + drshrn_h v25, v16, v17, #14 + neg v18.4s, v18.4s + neg v19.4s, v19.4s + drshrn_h v22, v18, v19, #14 + + idct32_end +endfunc + +.macro idct32_funcs suffix // Do an 32-point IDCT of a 8x32 slice out of a 32x32 matrix. // The 32-point IDCT can be decomposed into two 16-point IDCTs; // a normal IDCT16 with every other input component (the even ones, with @@ -917,19 +1225,30 @@ endfunc // x9 = double input stride // x10 = idct_coeffs // x11 = idct_coeffs + 32 -function idct32_1d_8x32_pass1_neon +function idct32_1d_8x32_pass1\suffix\()_neon mov x14, x30 ld1 {v0.8h,v1.8h}, [x10] - movi v4.8h, #0 + movi v2.8h, #0 // v16 = IN(0), v17 = IN(2) ... v31 = IN(30) +.ifb \suffix .irp i, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31 - ld1 {v\i\().8h}, [x2] - st1 {v4.8h}, [x2], x9 + load_clear \i, x2, x9 .endr +.endif +.ifc \suffix,_quarter +.irp i, 16, 17, 18, 19 + load_clear \i, x2, x9 +.endr +.endif +.ifc \suffix,_half +.irp i, 16, 17, 18, 19, 20, 21, 22, 23 + load_clear \i, x2, x9 +.endr +.endif - bl idct16 + bl idct16\suffix // Do two 8x8 transposes. Originally, v16-v31 contain the // 16 rows. Afterwards, v16-v23 and v24-v31 contain the @@ -964,17 +1283,36 @@ function idct32_1d_8x32_pass1_neon // Move x2 back to the start of the input, and move // to the first odd row +.ifb \suffix sub x2, x2, x9, lsl #4 +.endif +.ifc \suffix,_quarter + sub x2, x2, x9, lsl #2 +.endif +.ifc \suffix,_half + sub x2, x2, x9, lsl #3 +.endif add x2, x2, #64 - movi v4.8h, #0 + movi v2.8h, #0 // v16 = IN(1), v17 = IN(3) ... v31 = IN(31) +.ifb \suffix .irp i, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31 - ld1 {v\i\().8h}, [x2] - st1 {v4.8h}, [x2], x9 + load_clear \i, x2, x9 +.endr +.endif +.ifc \suffix,_quarter +.irp i, 16, 17, 18, 19 + load_clear \i, x2, x9 +.endr +.endif +.ifc \suffix,_half +.irp i, 16, 17, 18, 19, 20, 21, 22, 23 + load_clear \i, x2, x9 .endr +.endif - bl idct32_odd + bl idct32_odd\suffix transpose_8x8H v31, v30, v29, v28, v27, v26, v25, v24, v2, v3 transpose_8x8H v23, v22, v21, v20, v19, v18, v17, v16, v2, v3 @@ -1023,33 +1361,61 @@ endfunc // x9 = double temp buffer stride // x10 = idct_coeffs // x11 = idct_coeffs + 32 -function idct32_1d_8x32_pass2_neon +function idct32_1d_8x32_pass2\suffix\()_neon mov x14, x30 ld1 {v0.8h,v1.8h}, [x10] // v16 = IN(0), v17 = IN(2) ... v31 = IN(30) +.ifb \suffix .irp i, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31 - ld1 {v\i\().8h}, [x2], x9 + load \i, x2, x9 .endr sub x2, x2, x9, lsl #4 +.endif +.ifc \suffix,_quarter +.irp i, 16, 17, 18, 19 + load \i, x2, x9 +.endr + sub x2, x2, x9, lsl #2 +.endif +.ifc \suffix,_half +.irp i, 16, 17, 18, 19, 20, 21, 22, 23 + load \i, x2, x9 +.endr + sub x2, x2, x9, lsl #3 +.endif - bl idct16 + bl idct16\suffix .irp i, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31 - st1 {v\i\().8h}, [x2], x9 + store \i, x2, x9 .endr sub x2, x2, x9, lsl #4 add x2, x2, #64 // v16 = IN(1), v17 = IN(3) ... v31 = IN(31) +.ifb \suffix .irp i, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31 - ld1 {v\i\().8h}, [x2], x9 + load \i, x2, x9 .endr sub x2, x2, x9, lsl #4 +.endif +.ifc \suffix,_quarter +.irp i, 16, 17, 18, 19 + load \i, x2, x9 +.endr + sub x2, x2, x9, lsl #2 +.endif +.ifc \suffix,_half +.irp i, 16, 17, 18, 19, 20, 21, 22, 23 + load \i, x2, x9 +.endr + sub x2, x2, x9, lsl #3 +.endif sub x2, x2, #64 - bl idct32_odd + bl idct32_odd\suffix .macro load_acc_store a, b, c, d, neg=0 .if \neg == 0 @@ -1105,6 +1471,11 @@ function idct32_1d_8x32_pass2_neon .purgem load_acc_store br x14 endfunc +.endm + +idct32_funcs +idct32_funcs _quarter +idct32_funcs _half const min_eob_idct_idct_32, align=4 .short 0, 34, 135, 336 @@ -1135,6 +1506,11 @@ function ff_vp9_idct_idct_32x32_add_neon, export=1 mov x9, #128 neg x7, x9 + cmp w3, #34 + b.le idct32x32_quarter_add_neon + cmp w3, #135 + b.le idct32x32_half_add_neon + .irp i, 0, 8, 16, 24 add x0, sp, #(\i*64) .if \i > 0 @@ -1177,3 +1553,34 @@ function ff_vp9_idct_idct_32x32_add_neon, export=1 br x15 endfunc + +.macro idct32_partial size +function idct32x32_\size\()_add_neon + add x0, sp, #(0*64) + add x2, x6, #(0*2) + bl idct32_1d_8x32_pass1_\size\()_neon +.ifc \size,half + add x0, sp, #(8*64) + add x2, x6, #(8*2) + bl idct32_1d_8x32_pass1_\size\()_neon +.endif +.irp i, 0, 8, 16, 24 + add x0, x4, #(\i) + mov x1, x5 + add x2, sp, #(\i*2) + bl idct32_1d_8x32_pass2_\size\()_neon +.endr + + add sp, sp, #2048 + + ldp d8, d9, [sp], 0x10 + ldp d10, d11, [sp], 0x10 + ldp d12, d13, [sp], 0x10 + ldp d14, d15, [sp], 0x10 + + br x15 +endfunc +.endm + +idct32_partial quarter +idct32_partial half