From 5a71bce3713ce0b074b1ad33f2c5e9c6bcddde2c Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Cl=C3=A9ment=20B=C5=93sch?= Date: Sun, 6 May 2018 10:54:49 +0200 Subject: [PATCH] lavfi/nlmeans: add AArch64 SIMD for compute_safe_ssd_integral_image ssd_integral_image_c: 49204.6 ssd_integral_image_neon: 28346.8 --- libavfilter/aarch64/Makefile | 3 + libavfilter/aarch64/vf_nlmeans_init.c | 33 +++++++++++ libavfilter/aarch64/vf_nlmeans_neon.S | 80 +++++++++++++++++++++++++++ libavfilter/tests/integral.c | 6 +- libavfilter/vf_nlmeans.c | 26 ++++++--- libavfilter/vf_nlmeans.h | 35 ++++++++++++ 6 files changed, 175 insertions(+), 8 deletions(-) create mode 100644 libavfilter/aarch64/Makefile create mode 100644 libavfilter/aarch64/vf_nlmeans_init.c create mode 100644 libavfilter/aarch64/vf_nlmeans_neon.S create mode 100644 libavfilter/vf_nlmeans.h diff --git a/libavfilter/aarch64/Makefile b/libavfilter/aarch64/Makefile new file mode 100644 index 0000000000..b58daa3a3f --- /dev/null +++ b/libavfilter/aarch64/Makefile @@ -0,0 +1,3 @@ +OBJS-$(CONFIG_NLMEANS_FILTER) += aarch64/vf_nlmeans_init.o + +NEON-OBJS-$(CONFIG_NLMEANS_FILTER) += aarch64/vf_nlmeans_neon.o diff --git a/libavfilter/aarch64/vf_nlmeans_init.c b/libavfilter/aarch64/vf_nlmeans_init.c new file mode 100644 index 0000000000..a1edefb144 --- /dev/null +++ b/libavfilter/aarch64/vf_nlmeans_init.c @@ -0,0 +1,33 @@ +/* + * This file is part of FFmpeg. + * + * FFmpeg 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. + * + * FFmpeg 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 FFmpeg; if not, write to the Free Software + * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA + */ + +#include "libavutil/aarch64/cpu.h" +#include "libavfilter/vf_nlmeans.h" + +void ff_compute_safe_ssd_integral_image_neon(uint32_t *dst, ptrdiff_t dst_linesize_32, + const uint8_t *s1, ptrdiff_t linesize1, + const uint8_t *s2, ptrdiff_t linesize2, + int w, int h); + +av_cold void ff_nlmeans_init_aarch64(NLMeansDSPContext *dsp) +{ + int cpu_flags = av_get_cpu_flags(); + + if (have_neon(cpu_flags)) + dsp->compute_safe_ssd_integral_image = ff_compute_safe_ssd_integral_image_neon; +} diff --git a/libavfilter/aarch64/vf_nlmeans_neon.S b/libavfilter/aarch64/vf_nlmeans_neon.S new file mode 100644 index 0000000000..6308a428db --- /dev/null +++ b/libavfilter/aarch64/vf_nlmeans_neon.S @@ -0,0 +1,80 @@ +/* + * Copyright (c) 2018 Clément Bœsch + * + * This file is part of FFmpeg. + * + * FFmpeg 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. + * + * FFmpeg 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 FFmpeg; if not, write to the Free Software + * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA + */ + +#include "libavutil/aarch64/asm.S" + +// acc_sum_store(ABCD) = {X+A, X+A+B, X+A+B+C, X+A+B+C+D} +.macro acc_sum_store x, xb + dup v24.4S, v24.4S[3] // ...X -> XXXX + ext v25.16B, v26.16B, \xb, #12 // ext(0000,ABCD,12)=0ABC + add v24.4S, v24.4S, \x // XXXX+ABCD={X+A,X+B,X+C,X+D} + add v24.4S, v24.4S, v25.4S // {X+A,X+B+A,X+C+B,X+D+C} (+0ABC) + ext v25.16B, v26.16B, v25.16B, #12 // ext(0000,0ABC,12)=00AB + add v24.4S, v24.4S, v25.4S // {X+A,X+B+A,X+C+B+A,X+D+C+B} (+00AB) + ext v25.16B, v26.16B, v25.16B, #12 // ext(0000,00AB,12)=000A + add v24.4S, v24.4S, v25.4S // {X+A,X+B+A,X+C+B+A,X+D+C+B+A} (+000A) + st1 {v24.4S}, [x0], #16 // write 4x32-bit final values +.endm + +function ff_compute_safe_ssd_integral_image_neon, export=1 + movi v26.4S, #0 // used as zero for the "rotations" in acc_sum_store + sub x3, x3, w6, UXTW // s1 padding (s1_linesize - w) + sub x5, x5, w6, UXTW // s2 padding (s2_linesize - w) + sub x9, x0, x1, UXTW #2 // dst_top + sub x1, x1, w6, UXTW // dst padding (dst_linesize_32 - w) + lsl x1, x1, #2 // dst padding expressed in bytes +1: mov w10, w6 // width copy for each line + sub x0, x0, #16 // beginning of the dst line minus 4 sums + sub x8, x9, #4 // dst_top-1 + ld1 {v24.4S}, [x0], #16 // load ...X (contextual last sums) +2: ld1 {v0.16B}, [x2], #16 // s1[x + 0..15] + ld1 {v1.16B}, [x4], #16 // s2[x + 0..15] + ld1 {v16.4S,v17.4S}, [x8], #32 // dst_top[x + 0..7 - 1] + usubl v2.8H, v0.8B, v1.8B // d[x + 0..7] = s1[x + 0..7] - s2[x + 0..7] + usubl2 v3.8H, v0.16B, v1.16B // d[x + 8..15] = s1[x + 8..15] - s2[x + 8..15] + ld1 {v18.4S,v19.4S}, [x8], #32 // dst_top[x + 8..15 - 1] + smull v4.4S, v2.4H, v2.4H // d[x + 0..3]^2 + smull2 v5.4S, v2.8H, v2.8H // d[x + 4..7]^2 + ld1 {v20.4S,v21.4S}, [x9], #32 // dst_top[x + 0..7] + smull v6.4S, v3.4H, v3.4H // d[x + 8..11]^2 + smull2 v7.4S, v3.8H, v3.8H // d[x + 12..15]^2 + ld1 {v22.4S,v23.4S}, [x9], #32 // dst_top[x + 8..15] + sub v0.4S, v20.4S, v16.4S // dst_top[x + 0..3] - dst_top[x + 0..3 - 1] + sub v1.4S, v21.4S, v17.4S // dst_top[x + 4..7] - dst_top[x + 4..7 - 1] + add v0.4S, v0.4S, v4.4S // + d[x + 0..3]^2 + add v1.4S, v1.4S, v5.4S // + d[x + 4..7]^2 + sub v2.4S, v22.4S, v18.4S // dst_top[x + 8..11] - dst_top[x + 8..11 - 1] + sub v3.4S, v23.4S, v19.4S // dst_top[x + 12..15] - dst_top[x + 12..15 - 1] + add v2.4S, v2.4S, v6.4S // + d[x + 8..11]^2 + add v3.4S, v3.4S, v7.4S // + d[x + 12..15]^2 + acc_sum_store v0.4S, v0.16B // accumulate and store dst[ 0..3] + acc_sum_store v1.4S, v1.16B // accumulate and store dst[ 4..7] + acc_sum_store v2.4S, v2.16B // accumulate and store dst[ 8..11] + acc_sum_store v3.4S, v3.16B // accumulate and store dst[12..15] + subs w10, w10, #16 // width dec + b.ne 2b // loop til next line + add x2, x2, x3 // skip to next line (s1) + add x4, x4, x5 // skip to next line (s2) + add x0, x0, x1 // skip to next line (dst) + add x9, x9, x1 // skip to next line (dst_top) + subs w7, w7, #1 // height dec + b.ne 1b + ret +endfunc diff --git a/libavfilter/tests/integral.c b/libavfilter/tests/integral.c index 049fefae83..2a8e8ff55f 100644 --- a/libavfilter/tests/integral.c +++ b/libavfilter/tests/integral.c @@ -57,6 +57,10 @@ int main(void) uint32_t *ii_start = ii + ii_lz_32 + 1; // skip top 0-line and left 0-column uint32_t *ii_start2 = ii2 + ii_lz_32 + 1; // skip top 0-line and left 0-column + NLMeansDSPContext dsp = {0}; + + ff_nlmeans_init(&dsp); + if (!ii || !ii2) return -1; @@ -64,7 +68,7 @@ int main(void) for (xoff = -e; xoff <= e; xoff++) { printf("xoff=%d yoff=%d\n", xoff, yoff); - compute_ssd_integral_image(ii_start, ii_lz_32, + compute_ssd_integral_image(&dsp, ii_start, ii_lz_32, src, lz, xoff, yoff, e, w, h); display_integral(ii_start, ii_w, ii_h, ii_lz_32); diff --git a/libavfilter/vf_nlmeans.c b/libavfilter/vf_nlmeans.c index b081a4e5af..c30e44498f 100644 --- a/libavfilter/vf_nlmeans.c +++ b/libavfilter/vf_nlmeans.c @@ -20,7 +20,6 @@ /** * @todo - * - SIMD for compute_safe_ssd_integral_image * - SIMD for final weighted averaging * - better automatic defaults? see "Parameters" @ http://www.ipol.im/pub/art/2011/bcm_nlm/ * - temporal support (probably doesn't need any displacement according to @@ -37,6 +36,7 @@ #include "avfilter.h" #include "formats.h" #include "internal.h" +#include "vf_nlmeans.h" #include "video.h" struct weighted_avg { @@ -66,6 +66,7 @@ typedef struct NLMeansContext { double weight_lut[WEIGHT_LUT_SIZE]; // lookup table mapping (scaled) patch differences to their associated weights double pdiff_lut_scale; // scale factor for patch differences before looking into the LUT int max_meaningful_diff; // maximum difference considered (if the patch difference is too high we ignore the pixel) + NLMeansDSPContext dsp; } NLMeansContext; #define OFFSET(x) offsetof(NLMeansContext, x) @@ -240,7 +241,8 @@ static inline void compute_unsafe_ssd_integral_image(uint32_t *dst, ptrdiff_t ds * @param h source height * @param e research padding edge */ -static void compute_ssd_integral_image(uint32_t *ii, ptrdiff_t ii_linesize_32, +static void compute_ssd_integral_image(const NLMeansDSPContext *dsp, + uint32_t *ii, ptrdiff_t ii_linesize_32, const uint8_t *src, ptrdiff_t linesize, int offx, int offy, int e, int w, int h) { @@ -291,10 +293,10 @@ static void compute_ssd_integral_image(uint32_t *ii, ptrdiff_t ii_linesize_32, av_assert1(startx_safe - s2x >= 0); av_assert1(startx_safe - s2x < w); av_assert1(starty_safe - s2y >= 0); av_assert1(starty_safe - s2y < h); if (safe_pw && safe_ph) - compute_safe_ssd_integral_image_c(ii + starty_safe*ii_linesize_32 + startx_safe, ii_linesize_32, - src + (starty_safe - s1y) * linesize + (startx_safe - s1x), linesize, - src + (starty_safe - s2y) * linesize + (startx_safe - s2x), linesize, - safe_pw, safe_ph); + dsp->compute_safe_ssd_integral_image(ii + starty_safe*ii_linesize_32 + startx_safe, ii_linesize_32, + src + (starty_safe - s1y) * linesize + (startx_safe - s1x), linesize, + src + (starty_safe - s2y) * linesize + (startx_safe - s2x), linesize, + safe_pw, safe_ph); // right part of the integral compute_unsafe_ssd_integral_image(ii, ii_linesize_32, @@ -431,7 +433,7 @@ static int nlmeans_plane(AVFilterContext *ctx, int w, int h, int p, int r, .p = p, }; - compute_ssd_integral_image(s->ii, s->ii_lz_32, + compute_ssd_integral_image(&s->dsp, s->ii, s->ii_lz_32, src, src_linesize, offx, offy, e, w, h); ctx->internal->execute(ctx, nlmeans_slice, &td, NULL, @@ -489,6 +491,14 @@ static int filter_frame(AVFilterLink *inlink, AVFrame *in) } \ } while (0) +void ff_nlmeans_init(NLMeansDSPContext *dsp) +{ + dsp->compute_safe_ssd_integral_image = compute_safe_ssd_integral_image_c; + + if (ARCH_AARCH64) + ff_nlmeans_init_aarch64(dsp); +} + static av_cold int init(AVFilterContext *ctx) { int i; @@ -520,6 +530,8 @@ static av_cold int init(AVFilterContext *ctx) s->research_size, s->research_size, s->research_size_uv, s->research_size_uv, s->patch_size, s->patch_size, s->patch_size_uv, s->patch_size_uv); + ff_nlmeans_init(&s->dsp); + return 0; } diff --git a/libavfilter/vf_nlmeans.h b/libavfilter/vf_nlmeans.h new file mode 100644 index 0000000000..0a9aab2928 --- /dev/null +++ b/libavfilter/vf_nlmeans.h @@ -0,0 +1,35 @@ +/* + * This file is part of FFmpeg. + * + * FFmpeg 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. + * + * FFmpeg 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 FFmpeg; if not, write to the Free Software + * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA + */ + +#ifndef AVFILTER_NLMEANS_H +#define AVFILTER_NLMEANS_H + +#include +#include + +typedef struct NLMeansDSPContext { + void (*compute_safe_ssd_integral_image)(uint32_t *dst, ptrdiff_t dst_linesize_32, + const uint8_t *s1, ptrdiff_t linesize1, + const uint8_t *s2, ptrdiff_t linesize2, + int w, int h); +} NLMeansDSPContext; + +void ff_nlmeans_init(NLMeansDSPContext *dsp); +void ff_nlmeans_init_aarch64(NLMeansDSPContext *dsp); + +#endif /* AVFILTER_NLMEANS_H */