From patchwork Wed Mar 8 10:01:08 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: 2809 Delivered-To: ffmpegpatchwork@gmail.com Received: by 10.103.50.79 with SMTP id y76csp953539vsy; Wed, 8 Mar 2017 02:03:54 -0800 (PST) X-Received: by 10.28.181.80 with SMTP id e77mr22135263wmf.57.1488967434225; Wed, 08 Mar 2017 02:03:54 -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 l17si3664772wra.244.2017.03.08.02.03.53; Wed, 08 Mar 2017 02:03:54 -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 B70DA6882FF; Wed, 8 Mar 2017 12:01:28 +0200 (EET) X-Original-To: ffmpeg-devel@ffmpeg.org Delivered-To: ffmpeg-devel@ffmpeg.org Received: from mail-lf0-f44.google.com (mail-lf0-f44.google.com [209.85.215.44]) by ffbox0-bg.mplayerhq.hu (Postfix) with ESMTPS id 930486882EA for ; Wed, 8 Mar 2017 12:01:26 +0200 (EET) Received: by mail-lf0-f44.google.com with SMTP id k202so12375025lfe.1 for ; Wed, 08 Mar 2017 02:01:39 -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=VYCLwg9yQnCS6u++j52A0u1TMUV52jQHjZm0M5PkbwU=; b=AMTgylw75HIFJANhZrFE9Ov04mnC2TsUja3DyMsMoG0w64koWRnb8nFVuyj8vm+ZBF pRH860dlHQGoqDQprxVPMgdopfx/i5ypBEeo5UJ7B7XgQMAvVgdinLJr+u90DBdHSpf7 x3RpYTq29g4GVeqeWAdLGqo5N874HQVROObTk8QWw6ozwCld+tb2MZxu6TqKKNp39ozl bXtfkYeNf4xYh/K1/8OpBmggpo9NLKsHr1PGtghQb0ydzP1dBZXR9dLsMSAlapbZw1eZ ZT8vYSKfgMmdr42mvUCgYANZ547pJtKRtBGTJFDlv1M6vzeZ1OT5zKWv52BAccejzhSy bsXA== 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=VYCLwg9yQnCS6u++j52A0u1TMUV52jQHjZm0M5PkbwU=; b=b+rOynW5Khoc0pZ2B95YPkRp/6rNx39BTTXqvAQKixKwPwaRVMoW/a8tv9nWVP+dMp bTTTuxbI35KDo0Nx+Fl05XT92UHCcx/Fbisix5uA8+KQN1RmzexUS7cYrnqsvAtas/7h cF/tWQ2NUAvOIRV9pB20FTvKcI8iqVuQ3jJWgnlFsN50Yl/lF3IUkBxdpM5nXMn6RfjC 3icjEXk+kFNpCob/gLpoi9AhOl8Fn+G/omm9xjhghTmN1dUND10KDFT3iGL4SSEpW2YE rkZS3Kme86EVCyA4CdT7J0r716Sd6Yv+OUA4PotgJdr7/zl0TxaRuaYBTWZk1PFwJXmt RQwQ== X-Gm-Message-State: AMke39moQXeiuPGNs9If0FwpZQzlYN32/z2Ptea72ug7SREGPEb0EggZNQCWIgvb/b/W5g== X-Received: by 10.46.87.66 with SMTP id r2mr1801001ljd.4.1488967298888; Wed, 08 Mar 2017 02:01:38 -0800 (PST) Received: from localhost.localdomain ([2001:470:28:852:7d47:68e:13e8:4933]) by smtp.gmail.com with ESMTPSA id m127sm513064lfg.58.2017.03.08.02.01.38 for (version=TLS1_2 cipher=ECDHE-RSA-AES128-SHA bits=128/128); Wed, 08 Mar 2017 02:01:38 -0800 (PST) From: =?UTF-8?q?Martin=20Storsj=C3=B6?= To: ffmpeg-devel@ffmpeg.org Date: Wed, 8 Mar 2017 12:01:08 +0200 Message-Id: <1488967274-8143-28-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 28/34] arm: vp9lpf: Implement the mix2_44 function with one single filter pass 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" For this case, with 8 inputs but only changing 4 of them, we can fit all 16 input pixels into a q register, and still have enough temporary registers for doing the loop filter. The wd=8 filters would require too many temporary registers for processing all 16 pixels at once though. Before: Cortex A7 A8 A9 A53 vp9_loop_filter_mix2_v_44_16_neon: 289.7 256.2 237.5 181.2 After: vp9_loop_filter_mix2_v_44_16_neon: 221.2 150.5 177.7 138.0 This is cherrypicked from libav commit 575e31e931e4178e9f1e24407503c9b4ec0ef9ba. --- libavcodec/arm/vp9dsp_init_arm.c | 7 +- libavcodec/arm/vp9lpf_neon.S | 191 +++++++++++++++++++++++++++++++++++++++ 2 files changed, 195 insertions(+), 3 deletions(-) diff --git a/libavcodec/arm/vp9dsp_init_arm.c b/libavcodec/arm/vp9dsp_init_arm.c index f7b539e..4c57fd6 100644 --- a/libavcodec/arm/vp9dsp_init_arm.c +++ b/libavcodec/arm/vp9dsp_init_arm.c @@ -195,6 +195,8 @@ define_loop_filters(8, 8); define_loop_filters(16, 8); define_loop_filters(16, 16); +define_loop_filters(44, 16); + #define lf_mix_fn(dir, wd1, wd2, stridea) \ static void loop_filter_##dir##_##wd1##wd2##_16_neon(uint8_t *dst, \ ptrdiff_t stride, \ @@ -208,7 +210,6 @@ static void loop_filter_##dir##_##wd1##wd2##_16_neon(uint8_t *dst, lf_mix_fn(h, wd1, wd2, stride) \ lf_mix_fn(v, wd1, wd2, sizeof(uint8_t)) -lf_mix_fns(4, 4) lf_mix_fns(4, 8) lf_mix_fns(8, 4) lf_mix_fns(8, 8) @@ -228,8 +229,8 @@ static av_cold void vp9dsp_loopfilter_init_arm(VP9DSPContext *dsp) dsp->loop_filter_16[0] = ff_vp9_loop_filter_h_16_16_neon; dsp->loop_filter_16[1] = ff_vp9_loop_filter_v_16_16_neon; - dsp->loop_filter_mix2[0][0][0] = loop_filter_h_44_16_neon; - dsp->loop_filter_mix2[0][0][1] = loop_filter_v_44_16_neon; + dsp->loop_filter_mix2[0][0][0] = ff_vp9_loop_filter_h_44_16_neon; + dsp->loop_filter_mix2[0][0][1] = ff_vp9_loop_filter_v_44_16_neon; dsp->loop_filter_mix2[0][1][0] = loop_filter_h_48_16_neon; dsp->loop_filter_mix2[0][1][1] = loop_filter_v_48_16_neon; dsp->loop_filter_mix2[1][0][0] = loop_filter_h_84_16_neon; diff --git a/libavcodec/arm/vp9lpf_neon.S b/libavcodec/arm/vp9lpf_neon.S index 2d91092..8d44d58 100644 --- a/libavcodec/arm/vp9lpf_neon.S +++ b/libavcodec/arm/vp9lpf_neon.S @@ -44,6 +44,109 @@ vtrn.8 \r2, \r3 .endm +@ The input to and output from this macro is in the registers q8-q15, +@ and q0-q7 are used as scratch registers. +@ p3 = q8, p0 = q11, q0 = q12, q3 = q15 +.macro loop_filter_q + vdup.u8 d0, r2 @ E + lsr r2, r2, #8 + vdup.u8 d2, r3 @ I + lsr r3, r3, #8 + vdup.u8 d1, r2 @ E + vdup.u8 d3, r3 @ I + + vabd.u8 q2, q8, q9 @ abs(p3 - p2) + vabd.u8 q3, q9, q10 @ abs(p2 - p1) + vabd.u8 q4, q10, q11 @ abs(p1 - p0) + vabd.u8 q5, q12, q13 @ abs(q0 - q1) + vabd.u8 q6, q13, q14 @ abs(q1 - q2) + vabd.u8 q7, q14, q15 @ abs(q2 - q3) + vmax.u8 q2, q2, q3 + vmax.u8 q3, q4, q5 + vmax.u8 q4, q6, q7 + vabd.u8 q5, q11, q12 @ abs(p0 - q0) + vmax.u8 q2, q2, q3 + vqadd.u8 q5, q5, q5 @ abs(p0 - q0) * 2 + vabd.u8 q7, q10, q13 @ abs(p1 - q1) + vmax.u8 q2, q2, q4 @ max(abs(p3 - p2), ..., abs(q2 - q3)) + vshr.u8 q7, q7, #1 + vcle.u8 q2, q2, q1 @ max(abs()) <= I + vqadd.u8 q5, q5, q7 @ abs(p0 - q0) * 2 + abs(p1 - q1) >> 1 + vcle.u8 q5, q5, q0 + vand q2, q2, q5 @ fm + + vshrn.u16 d10, q2, #4 + vmov r2, r3, d10 + orrs r2, r2, r3 + @ If no pixels need filtering, just exit as soon as possible + beq 9f + + @ Calculate the normal inner loop filter for 2 or 4 pixels + ldr r3, [sp, #64] + vabd.u8 q3, q10, q11 @ abs(p1 - p0) + vabd.u8 q4, q13, q12 @ abs(q1 - q0) + + vsubl.u8 q5, d20, d26 @ p1 - q1 + vsubl.u8 q6, d21, d27 @ p1 - q1 + vmax.u8 q3, q3, q4 @ max(abs(p1 - p0), abs(q1 - q0)) + vqmovn.s16 d10, q5 @ av_clip_int8p(p1 - q1) + vqmovn.s16 d11, q6 @ av_clip_int8p(p1 - q1) + vdup.u8 d8, r3 @ H + lsr r3, r3, #8 + vdup.u8 d9, r3 @ H + vsubl.u8 q6, d24, d22 @ q0 - p0 + vsubl.u8 q7, d25, d23 @ q0 - p0 + vcle.u8 q3, q3, q4 @ hev + vmov.s16 q0, #3 + vand q3, q3, q2 @ !hev && fm && !flat8in + + vmul.s16 q6, q6, q0 @ 3 * (q0 - p0) + vmul.s16 q7, q7, q0 @ 3 * (q0 - p0) + vbic q5, q5, q3 @ if (!hev) av_clip_int8 = 0 + vaddw.s8 q6, q6, d10 @ 3 * (q0 - p0) [+ av_clip_int8(p1 - q1)] + vaddw.s8 q7, q7, d11 @ 3 * (q0 - p0) [+ av_clip_int8(p1 - q1)] + vmov.s8 q5, #4 + vqmovn.s16 d12, q6 + vqmovn.s16 d13, q7 @ av_clip_int8(3 * (q0 - p0) [+ av_clip_int8(p1 - q1)], BIT_DEPTH - 1) = f + vmov.s8 q0, #3 + + vqadd.s8 q5, q6, q5 @ FFMIN(f + 4, 127) + vqadd.s8 q0, q6, q0 @ FFMIN(f + 3, 127) + vmovl.u8 q6, d22 @ p0 + vmovl.u8 q7, d23 @ p0 + vshr.s8 q5, q5, #3 @ f1 + vshr.s8 q0, q0, #3 @ f2 + + vaddw.s8 q6, q6, d0 @ p0 + f2 + vaddw.s8 q7, q7, d1 @ p0 + f2 + vqmovun.s16 d0, q6 @ out p0 + vmovl.u8 q6, d24 @ q0 + vqmovun.s16 d1, q7 @ out p0 + vmovl.u8 q7, d25 @ q0 + vsubw.s8 q6, q6, d10 @ q0 - f1 + vsubw.s8 q7, q7, d11 @ q0 - f1 + vqmovun.s16 d12, q6 @ out q0 + vqmovun.s16 d13, q7 @ out q0 + vrshr.s8 q5, q5, #1 @ f = (f1 + 1) >> 1 + vbit q11, q0, q2 @ if (fm && !flat8in) + vbit q12, q6, q2 + + vmovl.u8 q0, d20 @ p1 + vmovl.u8 q2, d21 @ p1 + vmovl.u8 q6, d26 @ q1 + vmovl.u8 q7, d27 @ q1 + vaddw.s8 q0, q0, d10 @ p1 + f + vaddw.s8 q2, q2, d11 @ p1 + f + vsubw.s8 q6, q6, d10 @ q1 - f + vsubw.s8 q7, q7, d11 @ q1 - f + vqmovun.s16 d0, q0 @ out p1 + vqmovun.s16 d1, q2 @ out p1 + vqmovun.s16 d12, q6 @ out q1 + vqmovun.s16 d13, q7 @ out q1 + vbit q10, q0, q3 @ if (!hev && fm && !flat8in) + vbit q13, q6, q3 +.endm + @ The input to and output from this macro is in the registers d16-d31, @ and d0-d7 are used as scratch registers. @ p7 = d16 .. p3 = d20, p0 = d23, q0 = d24, q3 = d27, q7 = d31 @@ -455,6 +558,94 @@ function ff_vp9_loop_filter_h_4_8_neon, export=1 bx lr endfunc +function ff_vp9_loop_filter_v_44_16_neon, export=1 + vpush {q4-q7} + sub r12, r0, r1, lsl #2 + vld1.8 {q8}, [r12,:128], r1 @ p3 + vld1.8 {q12}, [r0, :128], r1 @ q0 + vld1.8 {q9}, [r12,:128], r1 @ p2 + vld1.8 {q13}, [r0, :128], r1 @ q1 + vld1.8 {q10}, [r12,:128], r1 @ p1 + vld1.8 {q14}, [r0, :128], r1 @ q2 + vld1.8 {q11}, [r12,:128], r1 @ p0 + vld1.8 {q15}, [r0, :128], r1 @ q3 + sub r0, r0, r1, lsl #2 + sub r12, r12, r1, lsl #1 + + loop_filter_q + + vst1.8 {q10}, [r12,:128], r1 + vst1.8 {q12}, [r0, :128], r1 + vst1.8 {q11}, [r12,:128], r1 + vst1.8 {q13}, [r0, :128], r1 +9: + vpop {q4-q7} + bx lr +endfunc + +function ff_vp9_loop_filter_h_44_16_neon, export=1 + vpush {q4-q7} + sub r12, r0, #4 + add r0, r12, r1, lsl #2 + vld1.8 {d16}, [r12], r1 + vld1.8 {d24}, [r0], r1 + vld1.8 {d18}, [r12], r1 + vld1.8 {d26}, [r0], r1 + vld1.8 {d20}, [r12], r1 + vld1.8 {d28}, [r0], r1 + vld1.8 {d22}, [r12], r1 + vld1.8 {d30}, [r0], r1 + mov r12, r0 + add r0, r0, r1, lsl #2 + vld1.8 {d17}, [r12], r1 + vld1.8 {d25}, [r0], r1 + vld1.8 {d19}, [r12], r1 + vld1.8 {d27}, [r0], r1 + vld1.8 {d21}, [r12], r1 + vld1.8 {d29}, [r0], r1 + vld1.8 {d23}, [r12], r1 + vld1.8 {d31}, [r0], r1 + + @ Transpose the 16x8 pixels, as two 8x8 parts + transpose_8x8 q8, q9, q10, q11, q12, q13, q14, q15 + + loop_filter_q + + sub r12, r0, r1, lsl #4 + add r0, r12, r1, lsl #3 + @ Move r0/r12 forward by 2 pixels; we don't need to rewrite the + @ outermost 2 pixels since they aren't changed. + add r12, r12, #2 + add r0, r0, #2 + + @ We only will write the mid 4 pixels back; after the loop filter, + @ these are in q10, q11, q12, q13, ordered as rows (16x4 pixels). + @ We need to transpose them to columns, done with a 4x4 transpose + @ (which in practice is four 4x4 transposes of the 4x4 blocks of + @ the 16x4 pixels; into 4x16 pixels). + transpose_4x4 q10, q11, q12, q13 + + vst1.32 {d20[0]}, [r12], r1 + vst1.32 {d21[0]}, [r0], r1 + vst1.32 {d22[0]}, [r12], r1 + vst1.32 {d23[0]}, [r0], r1 + vst1.32 {d24[0]}, [r12], r1 + vst1.32 {d25[0]}, [r0], r1 + vst1.32 {d26[0]}, [r12], r1 + vst1.32 {d27[0]}, [r0], r1 + vst1.32 {d20[1]}, [r12], r1 + vst1.32 {d21[1]}, [r0], r1 + vst1.32 {d22[1]}, [r12], r1 + vst1.32 {d23[1]}, [r0], r1 + vst1.32 {d24[1]}, [r12], r1 + vst1.32 {d25[1]}, [r0], r1 + vst1.32 {d26[1]}, [r12], r1 + vst1.32 {d27[1]}, [r0], r1 +9: + vpop {q4-q7} + bx lr +endfunc + function ff_vp9_loop_filter_v_8_8_neon, export=1 sub r12, r0, r1, lsl #2 vld1.8 {d20}, [r12,:64], r1 @ p3