* [FFmpeg-devel] [PATCH v2 1/3] avfilter/vf_colordetect: add new color range detection filter
2025-07-16 16:24 [FFmpeg-devel] [PATCH v2 0/3] avfilter: add vf_colordetect filter Niklas Haas
@ 2025-07-16 16:24 ` Niklas Haas
2025-07-16 16:24 ` [FFmpeg-devel] [PATCH v2 2/3] tests/checkasm: add check for vf_colordetect Niklas Haas
2025-07-16 16:24 ` [FFmpeg-devel] [PATCH v2 3/3] avfilter/vf_colordetect: add x86 SIMD implementation Niklas Haas
2 siblings, 0 replies; 11+ messages in thread
From: Niklas Haas @ 2025-07-16 16:24 UTC (permalink / raw)
To: ffmpeg-devel; +Cc: Niklas Haas
From: Niklas Haas <git@haasn.dev>
This filter can detect various properties about the image, including
whether or not there are out-of-range values, or whether the input appears
to use straight or premultiplied alpha.
Of course, these can only be heuristics, with "undetermined" as the base
case. While we can definitely prove the existence of full range or
straight alpha colors, we can never infer the opposite.
---
doc/filters.texi | 27 ++++
libavfilter/Makefile | 1 +
libavfilter/allfilters.c | 1 +
libavfilter/vf_colordetect.c | 252 +++++++++++++++++++++++++++++++++++
libavfilter/vf_colordetect.h | 149 +++++++++++++++++++++
5 files changed, 430 insertions(+)
create mode 100644 libavfilter/vf_colordetect.c
create mode 100644 libavfilter/vf_colordetect.h
diff --git a/doc/filters.texi b/doc/filters.texi
index ed2956fe75..74e9e71559 100644
--- a/doc/filters.texi
+++ b/doc/filters.texi
@@ -9753,6 +9753,33 @@ colorchannelmixer=.393:.769:.189:0:.349:.686:.168:0:.272:.534:.131
This filter supports the all above options as @ref{commands}.
+@section colordetect
+Analyze the video frames to determine the effective value range and alpha
+mode.
+
+The filter accepts the following options:
+
+@table @option
+@item mode
+Set of properties to detect. Unavailable properties, such as alpha mode for
+an input image without an alpha channel, will be ignored automatically.
+
+Accepts a combination of the following flags:
+
+@table @samp
+@item color_range
+Detect if the source countains luma pixels outside the limited (MPEG) range,
+which indicates that this is a full range YUV source.
+@item alpha_mode
+Detect if the source contains color values above the alpha channel, which
+indicates that the alpha channel is independent (straight), rather than
+premultiplied.
+@item all
+Enable detection of all of the above properties. This is the default.
+@end table
+
+@end table
+
@section colorize
Overlay a solid color on the video stream.
diff --git a/libavfilter/Makefile b/libavfilter/Makefile
index 9e9153f5b0..e19f67a3a7 100644
--- a/libavfilter/Makefile
+++ b/libavfilter/Makefile
@@ -237,6 +237,7 @@ OBJS-$(CONFIG_COLORBALANCE_FILTER) += vf_colorbalance.o
OBJS-$(CONFIG_COLORCHANNELMIXER_FILTER) += vf_colorchannelmixer.o
OBJS-$(CONFIG_COLORCONTRAST_FILTER) += vf_colorcontrast.o
OBJS-$(CONFIG_COLORCORRECT_FILTER) += vf_colorcorrect.o
+OBJS-$(CONFIG_COLORDETECT_FILTER) += vf_colordetect.o
OBJS-$(CONFIG_COLORIZE_FILTER) += vf_colorize.o
OBJS-$(CONFIG_COLORKEY_FILTER) += vf_colorkey.o
OBJS-$(CONFIG_COLORKEY_OPENCL_FILTER) += vf_colorkey_opencl.o opencl.o \
diff --git a/libavfilter/allfilters.c b/libavfilter/allfilters.c
index 409099bf1f..f3c2092b15 100644
--- a/libavfilter/allfilters.c
+++ b/libavfilter/allfilters.c
@@ -218,6 +218,7 @@ extern const FFFilter ff_vf_colorbalance;
extern const FFFilter ff_vf_colorchannelmixer;
extern const FFFilter ff_vf_colorcontrast;
extern const FFFilter ff_vf_colorcorrect;
+extern const FFFilter ff_vf_colordetect;
extern const FFFilter ff_vf_colorize;
extern const FFFilter ff_vf_colorkey;
extern const FFFilter ff_vf_colorkey_opencl;
diff --git a/libavfilter/vf_colordetect.c b/libavfilter/vf_colordetect.c
new file mode 100644
index 0000000000..0fb892634f
--- /dev/null
+++ b/libavfilter/vf_colordetect.c
@@ -0,0 +1,252 @@
+/*
+ * Copyright (c) 2025 Niklas Haas
+ *
+ * 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
+ */
+
+/**
+ * @file
+ * Video color space detector, tries to auto-detect YUV range and alpha mode.
+ */
+
+#include <stdbool.h>
+#include <stdatomic.h>
+
+#include "config.h"
+
+#include "libavutil/mem.h"
+#include "libavutil/opt.h"
+#include "libavutil/pixdesc.h"
+
+#include "avfilter.h"
+#include "filters.h"
+#include "formats.h"
+#include "video.h"
+
+#include "vf_colordetect.h"
+
+enum AlphaMode {
+ ALPHA_NONE = -1,
+ ALPHA_UNDETERMINED = 0,
+ ALPHA_STRAIGHT,
+ /* No way to positively identify premultiplied alpha */
+};
+
+enum ColorDetectMode {
+ COLOR_DETECT_COLOR_RANGE = 1 << 0,
+ COLOR_DETECT_ALPHA_MODE = 1 << 1,
+};
+
+typedef struct ColorDetectContext {
+ const AVClass *class;
+ FFColorDetectDSPContext dsp;
+ unsigned mode;
+
+ const AVPixFmtDescriptor *desc;
+ int nb_threads;
+ int depth;
+ int idx_a;
+ int mpeg_min;
+ int mpeg_max;
+
+ atomic_int detected_range; // enum AVColorRange
+ atomic_int detected_alpha; // enum AlphaMode
+} ColorDetectContext;
+
+#define OFFSET(x) offsetof(ColorDetectContext, x)
+#define FLAGS AV_OPT_FLAG_VIDEO_PARAM|AV_OPT_FLAG_FILTERING_PARAM
+
+static const AVOption colordetect_options[] = {
+ { "mode", "Image properties to detect", OFFSET(mode), AV_OPT_TYPE_FLAGS, {.i64 = -1}, 0, UINT_MAX, FLAGS, .unit = "mode" },
+ { "color_range", "Detect (YUV) color range", 0, AV_OPT_TYPE_CONST, {.i64 = COLOR_DETECT_COLOR_RANGE}, 0, 0, FLAGS, .unit = "mode" },
+ { "alpha_mode", "Detect alpha mode", 0, AV_OPT_TYPE_CONST, {.i64 = COLOR_DETECT_ALPHA_MODE }, 0, 0, FLAGS, .unit = "mode" },
+ { "all", "Detect all supported properties", 0, AV_OPT_TYPE_CONST, {.i64 = -1}, 0, 0, FLAGS, .unit = "mode" },
+ { NULL }
+};
+
+AVFILTER_DEFINE_CLASS(colordetect);
+
+static int query_format(const AVFilterContext *ctx,
+ AVFilterFormatsConfig **cfg_in,
+ AVFilterFormatsConfig **cfg_out)
+{
+ int want_flags = AV_PIX_FMT_FLAG_PLANAR;
+ int reject_flags = AV_PIX_FMT_FLAG_PAL | AV_PIX_FMT_FLAG_HWACCEL |
+ AV_PIX_FMT_FLAG_BITSTREAM | AV_PIX_FMT_FLAG_FLOAT |
+ AV_PIX_FMT_FLAG_BAYER | AV_PIX_FMT_FLAG_XYZ;
+
+ if (HAVE_BIGENDIAN) {
+ want_flags |= AV_PIX_FMT_FLAG_BE;
+ } else {
+ reject_flags |= AV_PIX_FMT_FLAG_BE;
+ }
+
+ AVFilterFormats *formats = ff_formats_pixdesc_filter(want_flags, reject_flags);
+ return ff_set_common_formats2(ctx, cfg_in, cfg_out, formats);
+}
+
+static int config_input(AVFilterLink *inlink)
+{
+ AVFilterContext *ctx = inlink->dst;
+ ColorDetectContext *s = ctx->priv;
+ const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(inlink->format);
+ const int depth = desc->comp[0].depth;
+ const int mpeg_min = 16 << (depth - 8);
+ const int mpeg_max = 235 << (depth - 8);
+ if (depth > 16) /* not currently possible; prevent future bugs */
+ return AVERROR(ENOTSUP);
+
+ s->desc = desc;
+ s->depth = depth;
+ s->mpeg_min = mpeg_min;
+ s->mpeg_max = mpeg_max;
+ s->nb_threads = ff_filter_get_nb_threads(ctx);
+
+ if (desc->flags & AV_PIX_FMT_FLAG_RGB) {
+ atomic_init(&s->detected_range, AVCOL_RANGE_JPEG);
+ } else {
+ atomic_init(&s->detected_range, AVCOL_RANGE_UNSPECIFIED);
+ }
+
+ if (desc->flags & AV_PIX_FMT_FLAG_ALPHA) {
+ s->idx_a = desc->comp[desc->nb_components - 1].plane;
+ atomic_init(&s->detected_alpha, ALPHA_UNDETERMINED);
+ } else {
+ atomic_init(&s->detected_alpha, ALPHA_NONE);
+ }
+
+ ff_color_detect_dsp_init(&s->dsp, depth, inlink->color_range);
+ return 0;
+}
+
+static int detect_range(AVFilterContext *ctx, void *arg,
+ int jobnr, int nb_jobs)
+{
+ ColorDetectContext *s = ctx->priv;
+ const AVFrame *in = arg;
+ const ptrdiff_t stride = in->linesize[0];
+ const int y_start = (in->height * jobnr) / nb_jobs;
+ const int y_end = (in->height * (jobnr + 1)) / nb_jobs;
+ const int h_slice = y_end - y_start;
+
+ if (s->dsp.detect_range(in->data[0] + y_start * stride, stride,
+ in->width, h_slice, s->mpeg_min, s->mpeg_max))
+ atomic_store(&s->detected_range, AVCOL_RANGE_JPEG);
+
+ return 0;
+}
+
+static int detect_alpha(AVFilterContext *ctx, void *arg,
+ int jobnr, int nb_jobs)
+{
+ ColorDetectContext *s = ctx->priv;
+ const AVFrame *in = arg;
+ const int w = in->width;
+ const int h = in->height;
+ const int y_start = (h * jobnr) / nb_jobs;
+ const int y_end = (h * (jobnr + 1)) / nb_jobs;
+ const int h_slice = y_end - y_start;
+
+ const int nb_planes = (s->desc->flags & AV_PIX_FMT_FLAG_RGB) ? 3 : 1;
+ const ptrdiff_t alpha_stride = in->linesize[s->idx_a];
+ const uint8_t *alpha = in->data[s->idx_a] + y_start * alpha_stride;
+
+ const int p = (1 << s->depth) - 1;
+ const int q = s->mpeg_max - s->mpeg_min;
+ const int k = s->mpeg_min * p + 128;
+
+ for (int i = 0; i < nb_planes; i++) {
+ const ptrdiff_t stride = in->linesize[i];
+ if (s->dsp.detect_alpha(in->data[i] + y_start * stride, stride,
+ alpha, alpha_stride, w, h_slice, p, q, k)) {
+ atomic_store(&s->detected_alpha, ALPHA_STRAIGHT);
+ return 0;
+ }
+ }
+
+ return 0;
+}
+
+static int filter_frame(AVFilterLink *inlink, AVFrame *in)
+{
+ AVFilterContext *ctx = inlink->dst;
+ ColorDetectContext *s = ctx->priv;
+ const int nb_threads = FFMIN(inlink->h, s->nb_threads);
+
+ if (s->mode & COLOR_DETECT_COLOR_RANGE && s->detected_range == AVCOL_RANGE_UNSPECIFIED)
+ ff_filter_execute(ctx, detect_range, in, NULL, nb_threads);
+ if (s->mode & COLOR_DETECT_ALPHA_MODE && s->detected_alpha == ALPHA_UNDETERMINED)
+ ff_filter_execute(ctx, detect_alpha, in, NULL, nb_threads);
+
+ return ff_filter_frame(inlink->dst->outputs[0], in);
+}
+
+static av_cold void uninit(AVFilterContext *ctx)
+{
+ ColorDetectContext *s = ctx->priv;
+ if (!s->mode)
+ return;
+
+ av_log(ctx, AV_LOG_INFO, "Detected color properties:\n");
+ if (s->mode & COLOR_DETECT_COLOR_RANGE) {
+ av_log(ctx, AV_LOG_INFO, " Color range: %s\n",
+ s->detected_range == AVCOL_RANGE_JPEG ? "JPEG / full range"
+ : "undetermined");
+ }
+
+ if (s->mode & COLOR_DETECT_ALPHA_MODE) {
+ av_log(ctx, AV_LOG_INFO, " Alpha mode: %s\n",
+ s->detected_alpha == ALPHA_NONE ? "none" :
+ s->detected_alpha == ALPHA_STRAIGHT ? "straight / independent"
+ : "undetermined");
+ }
+}
+
+av_cold void ff_color_detect_dsp_init(FFColorDetectDSPContext *dsp, int depth,
+ enum AVColorRange color_range)
+{
+ if (!dsp->detect_range)
+ dsp->detect_range = depth > 8 ? ff_detect_range16_c : ff_detect_range_c;
+ if (!dsp->detect_alpha) {
+ if (color_range == AVCOL_RANGE_JPEG) {
+ dsp->detect_alpha = depth > 8 ? ff_detect_alpha16_full_c : ff_detect_alpha_full_c;
+ } else {
+ dsp->detect_alpha = depth > 8 ? ff_detect_alpha16_limited_c : ff_detect_alpha_limited_c;
+ }
+ }
+}
+
+static const AVFilterPad colordetect_inputs[] = {
+ {
+ .name = "default",
+ .type = AVMEDIA_TYPE_VIDEO,
+ .config_props = config_input,
+ .filter_frame = filter_frame,
+ },
+};
+
+const FFFilter ff_vf_colordetect = {
+ .p.name = "colordetect",
+ .p.description = NULL_IF_CONFIG_SMALL("Detect video color properties."),
+ .p.priv_class = &colordetect_class,
+ .p.flags = AVFILTER_FLAG_SLICE_THREADS | AVFILTER_FLAG_METADATA_ONLY,
+ .priv_size = sizeof(ColorDetectContext),
+ FILTER_INPUTS(colordetect_inputs),
+ FILTER_OUTPUTS(ff_video_default_filterpad),
+ FILTER_QUERY_FUNC2(query_format),
+ .uninit = uninit,
+};
diff --git a/libavfilter/vf_colordetect.h b/libavfilter/vf_colordetect.h
new file mode 100644
index 0000000000..8998ed83d4
--- /dev/null
+++ b/libavfilter/vf_colordetect.h
@@ -0,0 +1,149 @@
+/*
+ * 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_VF_COLORDETECT_H
+#define AVFILTER_VF_COLORDETECT_H
+
+#include <stddef.h>
+#include <stdint.h>
+
+#include <libavutil/macros.h>
+#include <libavutil/pixfmt.h>
+
+typedef struct FFColorDetectDSPContext {
+ /* Returns 1 if an out-of-range value was detected, 0 otherwise */
+ int (*detect_range)(const uint8_t *data, ptrdiff_t stride,
+ ptrdiff_t width, ptrdiff_t height,
+ int mpeg_min, int mpeg_max);
+
+ /* Returns 1 if the color value exceeds the alpha value, 0 otherwise */
+ int (*detect_alpha)(const uint8_t *color, ptrdiff_t color_stride,
+ const uint8_t *alpha, ptrdiff_t alpha_stride,
+ ptrdiff_t width, ptrdiff_t height,
+ int p, int q, int k);
+} FFColorDetectDSPContext;
+
+void ff_color_detect_dsp_init(FFColorDetectDSPContext *dsp, int depth,
+ enum AVColorRange color_range);
+
+static inline int ff_detect_range_c(const uint8_t *data, ptrdiff_t stride,
+ ptrdiff_t width, ptrdiff_t height,
+ int mpeg_min, int mpeg_max)
+{
+ while (height--) {
+ for (int x = 0; x < width; x++) {
+ const uint8_t val = data[x];
+ if (val < mpeg_min || val > mpeg_max)
+ return 1;
+ }
+ data += stride;
+ }
+
+ return 0;
+}
+
+static inline int ff_detect_range16_c(const uint8_t *data, ptrdiff_t stride,
+ ptrdiff_t width, ptrdiff_t height,
+ int mpeg_min, int mpeg_max)
+{
+ while (height--) {
+ const uint16_t *data16 = (const uint16_t *) data;
+ for (int x = 0; x < width; x++) {
+ const uint16_t val = data16[x];
+ if (val < mpeg_min || val > mpeg_max)
+ return 1;
+ }
+ data += stride;
+ }
+
+ return 0;
+}
+
+static inline int
+ff_detect_alpha_full_c(const uint8_t *color, ptrdiff_t color_stride,
+ const uint8_t *alpha, ptrdiff_t alpha_stride,
+ ptrdiff_t width, ptrdiff_t height,
+ int p, int q, int k)
+{
+ while (height--) {
+ for (int x = 0; x < width; x++) {
+ if (color[x] > alpha[x])
+ return 1;
+ }
+ color += color_stride;
+ alpha += alpha_stride;
+ }
+ return 0;
+}
+
+static inline int
+ff_detect_alpha_limited_c(const uint8_t *color, ptrdiff_t color_stride,
+ const uint8_t *alpha, ptrdiff_t alpha_stride,
+ ptrdiff_t width, ptrdiff_t height,
+ int p, int q, int k)
+{
+ while (height--) {
+ for (int x = 0; x < width; x++) {
+ if (p * color[x] - k > q * alpha[x])
+ return 1;
+ }
+ color += color_stride;
+ alpha += alpha_stride;
+ }
+ return 0;
+}
+
+static inline int
+ff_detect_alpha16_full_c(const uint8_t *color, ptrdiff_t color_stride,
+ const uint8_t *alpha, ptrdiff_t alpha_stride,
+ ptrdiff_t width, ptrdiff_t height,
+ int p, int q, int k)
+{
+ while (height--) {
+ const uint16_t *color16 = (const uint16_t *) color;
+ const uint16_t *alpha16 = (const uint16_t *) alpha;
+ for (int x = 0; x < width; x++) {
+ if (color16[x] > alpha16[x])
+ return 1;
+ }
+ color += color_stride;
+ alpha += alpha_stride;
+ }
+ return 0;
+}
+
+static inline int
+ff_detect_alpha16_limited_c(const uint8_t *color, ptrdiff_t color_stride,
+ const uint8_t *alpha, ptrdiff_t alpha_stride,
+ ptrdiff_t width, ptrdiff_t height,
+ int p, int q, int k)
+{
+ while (height--) {
+ const uint16_t *color16 = (const uint16_t *) color;
+ const uint16_t *alpha16 = (const uint16_t *) alpha;
+ for (int x = 0; x < width; x++) {
+ if ((int64_t) p * color16[x] - k > (int64_t) q * alpha16[x])
+ return 1;
+ }
+ color += color_stride;
+ alpha += alpha_stride;
+ }
+ return 0;
+}
+
+#endif /* AVFILTER_VF_COLORDETECT_H */
--
2.50.1
_______________________________________________
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
https://ffmpeg.org/mailman/listinfo/ffmpeg-devel
To unsubscribe, visit link above, or email
ffmpeg-devel-request@ffmpeg.org with subject "unsubscribe".
^ permalink raw reply [flat|nested] 11+ messages in thread* [FFmpeg-devel] [PATCH v2 2/3] tests/checkasm: add check for vf_colordetect
2025-07-16 16:24 [FFmpeg-devel] [PATCH v2 0/3] avfilter: add vf_colordetect filter Niklas Haas
2025-07-16 16:24 ` [FFmpeg-devel] [PATCH v2 1/3] avfilter/vf_colordetect: add new color range detection filter Niklas Haas
@ 2025-07-16 16:24 ` Niklas Haas
2025-07-16 22:48 ` James Almer
2025-07-16 16:24 ` [FFmpeg-devel] [PATCH v2 3/3] avfilter/vf_colordetect: add x86 SIMD implementation Niklas Haas
2 siblings, 1 reply; 11+ messages in thread
From: Niklas Haas @ 2025-07-16 16:24 UTC (permalink / raw)
To: ffmpeg-devel; +Cc: Niklas Haas
From: Niklas Haas <git@haasn.dev>
---
tests/checkasm/Makefile | 1 +
tests/checkasm/checkasm.c | 3 +
tests/checkasm/checkasm.h | 1 +
tests/checkasm/vf_colordetect.c | 137 ++++++++++++++++++++++++++++++++
4 files changed, 142 insertions(+)
create mode 100644 tests/checkasm/vf_colordetect.c
diff --git a/tests/checkasm/Makefile b/tests/checkasm/Makefile
index afd62d95ba..d8328ecc68 100644
--- a/tests/checkasm/Makefile
+++ b/tests/checkasm/Makefile
@@ -55,6 +55,7 @@ CHECKASMOBJS-$(CONFIG_AVCODEC) += $(AVCODECOBJS-yes)
AVFILTEROBJS-$(CONFIG_AFIR_FILTER) += af_afir.o
AVFILTEROBJS-$(CONFIG_BLEND_FILTER) += vf_blend.o
AVFILTEROBJS-$(CONFIG_BWDIF_FILTER) += vf_bwdif.o
+AVFILTEROBJS-$(CONFIG_COLORDETECT_FILTER)+= vf_colordetect.o
AVFILTEROBJS-$(CONFIG_COLORSPACE_FILTER) += vf_colorspace.o
AVFILTEROBJS-$(CONFIG_EQ_FILTER) += vf_eq.o
AVFILTEROBJS-$(CONFIG_GBLUR_FILTER) += vf_gblur.o
diff --git a/tests/checkasm/checkasm.c b/tests/checkasm/checkasm.c
index f4e3d4f433..e01994768e 100644
--- a/tests/checkasm/checkasm.c
+++ b/tests/checkasm/checkasm.c
@@ -269,6 +269,9 @@ static const struct {
#if CONFIG_BWDIF_FILTER
{ "vf_bwdif", checkasm_check_vf_bwdif },
#endif
+ #if CONFIG_COLORDETECT_FILTER
+ { "vf_colordetect", checkasm_check_colordetect },
+ #endif
#if CONFIG_COLORSPACE_FILTER
{ "vf_colorspace", checkasm_check_colorspace },
#endif
diff --git a/tests/checkasm/checkasm.h b/tests/checkasm/checkasm.h
index e829942d58..0b3e238318 100644
--- a/tests/checkasm/checkasm.h
+++ b/tests/checkasm/checkasm.h
@@ -89,6 +89,7 @@ void checkasm_check_av_tx(void);
void checkasm_check_blend(void);
void checkasm_check_blockdsp(void);
void checkasm_check_bswapdsp(void);
+void checkasm_check_colordetect(void);
void checkasm_check_colorspace(void);
void checkasm_check_diracdsp(void);
void checkasm_check_exrdsp(void);
diff --git a/tests/checkasm/vf_colordetect.c b/tests/checkasm/vf_colordetect.c
new file mode 100644
index 0000000000..af02529376
--- /dev/null
+++ b/tests/checkasm/vf_colordetect.c
@@ -0,0 +1,137 @@
+/*
+ * This file is part of FFmpeg.
+ *
+ * FFmpeg is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation; either version 2 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 General Public License for more details.
+ *
+ * You should have received a copy of the GNU 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 <string.h>
+#include "checkasm.h"
+
+#include "libavfilter/vf_colordetect.h"
+#include "libavutil/mem_internal.h"
+
+#define WIDTH 256
+#define HEIGHT 16
+#define STRIDE (WIDTH + 32)
+
+static void check_range_detect(int depth)
+{
+ const int mpeg_min = 16 << (depth - 8);
+ const int mpeg_max = 235 << (depth - 8);
+
+ FFColorDetectDSPContext dsp = {0};
+ ff_color_detect_dsp_init(&dsp, depth, AVCOL_RANGE_UNSPECIFIED);
+
+ declare_func(int, const uint8_t *, ptrdiff_t, ptrdiff_t, ptrdiff_t, int, int);
+
+ /* Initialize to 128, which should always return 0 */
+ LOCAL_ALIGNED_32(uint8_t, in, [HEIGHT * STRIDE]);
+ memset(in, 0x80, HEIGHT * STRIDE);
+
+ /* Place an out-of-range value in a random position near the center */
+ const int h2 = HEIGHT >> 1;
+ int idx0 = ((rnd() % h2) + h2) * STRIDE + (rnd() % WIDTH);
+ if (depth > 8) {
+ idx0 &= ~1;
+ in[idx0] = in[idx0 + 1] = 0;
+ } else {
+ in[idx0] = 0;
+ }
+
+ int w = WIDTH;
+ if (depth > 8)
+ w /= 2;
+
+ if (check_func(dsp.detect_range, "rangedetect%d", depth)) {
+ /* Test increasing height, to ensure we hit the placed 0 eventually */
+ for (int h = 1; h <= HEIGHT; h++) {
+ int res_ref = call_ref(in, STRIDE, w, h, mpeg_min, mpeg_max);
+ int res_new = call_new(in, STRIDE, w, h, mpeg_min, mpeg_max);
+ if (res_ref != res_new)
+ fail();
+ }
+
+ /* Test performance of base case without any out-of-range values */
+ memset(in, 0x80, HEIGHT * STRIDE);
+ bench_new(in, STRIDE, w, HEIGHT, mpeg_min, mpeg_max);
+ }
+}
+
+static void check_alpha_detect(int depth, enum AVColorRange range)
+{
+ const int mpeg_min = 16 << (depth - 8);
+ const int mpeg_max = 235 << (depth - 8);
+ const int p = (1 << depth) - 1;
+ const int q = mpeg_max - mpeg_min;
+ const int k = mpeg_min * p + 128;
+
+ FFColorDetectDSPContext dsp = {0};
+ ff_color_detect_dsp_init(&dsp, depth, range);
+
+ declare_func(int, const uint8_t *, ptrdiff_t, const uint8_t *, ptrdiff_t,
+ ptrdiff_t, ptrdiff_t, int p, int q, int k);
+
+ LOCAL_ALIGNED_32(uint8_t, luma, [HEIGHT * STRIDE]);
+ LOCAL_ALIGNED_32(uint8_t, alpha, [HEIGHT * STRIDE]);
+ memset(luma, 0x80, HEIGHT * STRIDE);
+ memset(alpha, 0xFF, HEIGHT * STRIDE);
+
+ /* Try and force overflow */
+ if (depth > 8 && range == AVCOL_RANGE_MPEG) {
+ ((uint16_t *) luma)[0] = 235 << (depth - 8);
+ ((uint16_t *) luma)[1] = 16 << (depth - 8);
+ } else {
+ luma[0] = 235;
+ luma[1] = 16;
+ }
+
+ /* Place an out-of-range value in a random position near the center */
+ const int h2 = HEIGHT >> 1;
+ int idx0 = ((rnd() % h2) + h2) * STRIDE + (rnd() % WIDTH);
+ if (depth > 8) {
+ idx0 &= ~1;
+ alpha[idx0] = alpha[idx0 + 1] = 0;
+ } else {
+ alpha[idx0] = 0;
+ }
+
+ int w = WIDTH;
+ if (depth > 8)
+ w /= 2;
+
+ if (check_func(dsp.detect_alpha, "alphadetect%d_%s", depth, range == AVCOL_RANGE_JPEG ? "full" : "limited")) {
+ /* Test increasing height, to ensure we hit the placed 0 eventually */
+ for (int h = 1; h <= HEIGHT; h++) {
+ int res_ref = call_ref(luma, STRIDE, alpha, STRIDE, w, h, p, q, k);
+ int res_new = call_new(luma, STRIDE, alpha, STRIDE, w, h, p, q, k);
+ if (res_ref != res_new)
+ fail();
+ }
+
+ /* Test performance of base case without any out-of-range values */
+ memset(alpha, 0xFF, HEIGHT * STRIDE);
+ bench_new(luma, STRIDE, alpha, STRIDE, w, HEIGHT, p, q, k);
+ }
+}
+
+void checkasm_check_colordetect(void)
+{
+ for (int depth = 8; depth <= 16; depth += 8) {
+ check_range_detect(depth);
+ check_alpha_detect(depth, AVCOL_RANGE_JPEG);
+ check_alpha_detect(depth, AVCOL_RANGE_MPEG);
+ report("colordetect%d", depth);
+ }
+}
--
2.50.1
_______________________________________________
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
https://ffmpeg.org/mailman/listinfo/ffmpeg-devel
To unsubscribe, visit link above, or email
ffmpeg-devel-request@ffmpeg.org with subject "unsubscribe".
^ permalink raw reply [flat|nested] 11+ messages in thread* Re: [FFmpeg-devel] [PATCH v2 2/3] tests/checkasm: add check for vf_colordetect
2025-07-16 16:24 ` [FFmpeg-devel] [PATCH v2 2/3] tests/checkasm: add check for vf_colordetect Niklas Haas
@ 2025-07-16 22:48 ` James Almer
0 siblings, 0 replies; 11+ messages in thread
From: James Almer @ 2025-07-16 22:48 UTC (permalink / raw)
To: ffmpeg-devel
[-- Attachment #1.1.1: Type: text/plain, Size: 7885 bytes --]
On 7/16/2025 1:24 PM, Niklas Haas wrote:
> From: Niklas Haas <git@haasn.dev>
>
> ---
> tests/checkasm/Makefile | 1 +
> tests/checkasm/checkasm.c | 3 +
> tests/checkasm/checkasm.h | 1 +
> tests/checkasm/vf_colordetect.c | 137 ++++++++++++++++++++++++++++++++
> 4 files changed, 142 insertions(+)
> create mode 100644 tests/checkasm/vf_colordetect.c
Missing a fate-checkasm-vf_colordetect entry in tests/fate/checkasm.mak
>
> diff --git a/tests/checkasm/Makefile b/tests/checkasm/Makefile
> index afd62d95ba..d8328ecc68 100644
> --- a/tests/checkasm/Makefile
> +++ b/tests/checkasm/Makefile
> @@ -55,6 +55,7 @@ CHECKASMOBJS-$(CONFIG_AVCODEC) += $(AVCODECOBJS-yes)
> AVFILTEROBJS-$(CONFIG_AFIR_FILTER) += af_afir.o
> AVFILTEROBJS-$(CONFIG_BLEND_FILTER) += vf_blend.o
> AVFILTEROBJS-$(CONFIG_BWDIF_FILTER) += vf_bwdif.o
> +AVFILTEROBJS-$(CONFIG_COLORDETECT_FILTER)+= vf_colordetect.o
> AVFILTEROBJS-$(CONFIG_COLORSPACE_FILTER) += vf_colorspace.o
> AVFILTEROBJS-$(CONFIG_EQ_FILTER) += vf_eq.o
> AVFILTEROBJS-$(CONFIG_GBLUR_FILTER) += vf_gblur.o
> diff --git a/tests/checkasm/checkasm.c b/tests/checkasm/checkasm.c
> index f4e3d4f433..e01994768e 100644
> --- a/tests/checkasm/checkasm.c
> +++ b/tests/checkasm/checkasm.c
> @@ -269,6 +269,9 @@ static const struct {
> #if CONFIG_BWDIF_FILTER
> { "vf_bwdif", checkasm_check_vf_bwdif },
> #endif
> + #if CONFIG_COLORDETECT_FILTER
> + { "vf_colordetect", checkasm_check_colordetect },
> + #endif
> #if CONFIG_COLORSPACE_FILTER
> { "vf_colorspace", checkasm_check_colorspace },
> #endif
> diff --git a/tests/checkasm/checkasm.h b/tests/checkasm/checkasm.h
> index e829942d58..0b3e238318 100644
> --- a/tests/checkasm/checkasm.h
> +++ b/tests/checkasm/checkasm.h
> @@ -89,6 +89,7 @@ void checkasm_check_av_tx(void);
> void checkasm_check_blend(void);
> void checkasm_check_blockdsp(void);
> void checkasm_check_bswapdsp(void);
> +void checkasm_check_colordetect(void);
> void checkasm_check_colorspace(void);
> void checkasm_check_diracdsp(void);
> void checkasm_check_exrdsp(void);
> diff --git a/tests/checkasm/vf_colordetect.c b/tests/checkasm/vf_colordetect.c
> new file mode 100644
> index 0000000000..af02529376
> --- /dev/null
> +++ b/tests/checkasm/vf_colordetect.c
> @@ -0,0 +1,137 @@
> +/*
> + * This file is part of FFmpeg.
> + *
> + * FFmpeg is free software; you can redistribute it and/or modify
> + * it under the terms of the GNU General Public License as published by
> + * the Free Software Foundation; either version 2 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 General Public License for more details.
> + *
> + * You should have received a copy of the GNU 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 <string.h>
> +#include "checkasm.h"
> +
> +#include "libavfilter/vf_colordetect.h"
> +#include "libavutil/mem_internal.h"
> +
> +#define WIDTH 256
> +#define HEIGHT 16
> +#define STRIDE (WIDTH + 32)
> +
> +static void check_range_detect(int depth)
> +{
> + const int mpeg_min = 16 << (depth - 8);
> + const int mpeg_max = 235 << (depth - 8);
> +
> + FFColorDetectDSPContext dsp = {0};
> + ff_color_detect_dsp_init(&dsp, depth, AVCOL_RANGE_UNSPECIFIED);
> +
> + declare_func(int, const uint8_t *, ptrdiff_t, ptrdiff_t, ptrdiff_t, int, int);
> +
> + /* Initialize to 128, which should always return 0 */
> + LOCAL_ALIGNED_32(uint8_t, in, [HEIGHT * STRIDE]);
> + memset(in, 0x80, HEIGHT * STRIDE);
> +
> + /* Place an out-of-range value in a random position near the center */
> + const int h2 = HEIGHT >> 1;
> + int idx0 = ((rnd() % h2) + h2) * STRIDE + (rnd() % WIDTH);
> + if (depth > 8) {
> + idx0 &= ~1;
> + in[idx0] = in[idx0 + 1] = 0;
> + } else {
> + in[idx0] = 0;
> + }
> +
> + int w = WIDTH;
> + if (depth > 8)
> + w /= 2;
> +
> + if (check_func(dsp.detect_range, "rangedetect%d", depth)) {
detect_range_%d
> + /* Test increasing height, to ensure we hit the placed 0 eventually */
> + for (int h = 1; h <= HEIGHT; h++) {
> + int res_ref = call_ref(in, STRIDE, w, h, mpeg_min, mpeg_max);
> + int res_new = call_new(in, STRIDE, w, h, mpeg_min, mpeg_max);
> + if (res_ref != res_new)
> + fail();
> + }
> +
> + /* Test performance of base case without any out-of-range values */
> + memset(in, 0x80, HEIGHT * STRIDE);
> + bench_new(in, STRIDE, w, HEIGHT, mpeg_min, mpeg_max);
> + }
> +}
> +
> +static void check_alpha_detect(int depth, enum AVColorRange range)
> +{
> + const int mpeg_min = 16 << (depth - 8);
> + const int mpeg_max = 235 << (depth - 8);
> + const int p = (1 << depth) - 1;
> + const int q = mpeg_max - mpeg_min;
> + const int k = mpeg_min * p + 128;
> +
> + FFColorDetectDSPContext dsp = {0};
> + ff_color_detect_dsp_init(&dsp, depth, range);
> +
> + declare_func(int, const uint8_t *, ptrdiff_t, const uint8_t *, ptrdiff_t,
> + ptrdiff_t, ptrdiff_t, int p, int q, int k);
> +
> + LOCAL_ALIGNED_32(uint8_t, luma, [HEIGHT * STRIDE]);
> + LOCAL_ALIGNED_32(uint8_t, alpha, [HEIGHT * STRIDE]);
> + memset(luma, 0x80, HEIGHT * STRIDE);
> + memset(alpha, 0xFF, HEIGHT * STRIDE);
> +
> + /* Try and force overflow */
> + if (depth > 8 && range == AVCOL_RANGE_MPEG) {
> + ((uint16_t *) luma)[0] = 235 << (depth - 8);
> + ((uint16_t *) luma)[1] = 16 << (depth - 8);
> + } else {
> + luma[0] = 235;
> + luma[1] = 16;
> + }
> +
> + /* Place an out-of-range value in a random position near the center */
> + const int h2 = HEIGHT >> 1;
> + int idx0 = ((rnd() % h2) + h2) * STRIDE + (rnd() % WIDTH);
> + if (depth > 8) {
> + idx0 &= ~1;
> + alpha[idx0] = alpha[idx0 + 1] = 0;
> + } else {
> + alpha[idx0] = 0;
> + }
> +
> + int w = WIDTH;
> + if (depth > 8)
> + w /= 2;
> +
> + if (check_func(dsp.detect_alpha, "alphadetect%d_%s", depth, range == AVCOL_RANGE_JPEG ? "full" : "limited")) {
detect_alpha_%d_%s
> + /* Test increasing height, to ensure we hit the placed 0 eventually */
> + for (int h = 1; h <= HEIGHT; h++) {
> + int res_ref = call_ref(luma, STRIDE, alpha, STRIDE, w, h, p, q, k);
> + int res_new = call_new(luma, STRIDE, alpha, STRIDE, w, h, p, q, k);
> + if (res_ref != res_new)
> + fail();
> + }
> +
> + /* Test performance of base case without any out-of-range values */
> + memset(alpha, 0xFF, HEIGHT * STRIDE);
> + bench_new(luma, STRIDE, alpha, STRIDE, w, HEIGHT, p, q, k);
> + }
> +}
> +
> +void checkasm_check_colordetect(void)
> +{
> + for (int depth = 8; depth <= 16; depth += 8) {
> + check_range_detect(depth);
> + check_alpha_detect(depth, AVCOL_RANGE_JPEG);
> + check_alpha_detect(depth, AVCOL_RANGE_MPEG);
> + report("colordetect%d", depth);
check_range_detect(depth);
report("range_detect_%d", depth);
check_alpha_detect(depth, AVCOL_RANGE_JPEG);
check_alpha_detect(depth, AVCOL_RANGE_MPEG);
report("alpha_detect_%d", depth);
> + }
> +}
[-- Attachment #1.2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 495 bytes --]
[-- Attachment #2: Type: text/plain, Size: 251 bytes --]
_______________________________________________
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
https://ffmpeg.org/mailman/listinfo/ffmpeg-devel
To unsubscribe, visit link above, or email
ffmpeg-devel-request@ffmpeg.org with subject "unsubscribe".
^ permalink raw reply [flat|nested] 11+ messages in thread
* [FFmpeg-devel] [PATCH v2 3/3] avfilter/vf_colordetect: add x86 SIMD implementation
2025-07-16 16:24 [FFmpeg-devel] [PATCH v2 0/3] avfilter: add vf_colordetect filter Niklas Haas
2025-07-16 16:24 ` [FFmpeg-devel] [PATCH v2 1/3] avfilter/vf_colordetect: add new color range detection filter Niklas Haas
2025-07-16 16:24 ` [FFmpeg-devel] [PATCH v2 2/3] tests/checkasm: add check for vf_colordetect Niklas Haas
@ 2025-07-16 16:24 ` Niklas Haas
2025-07-16 20:06 ` Henrik Gramner via ffmpeg-devel
2025-07-16 20:25 ` James Almer
2 siblings, 2 replies; 11+ messages in thread
From: Niklas Haas @ 2025-07-16 16:24 UTC (permalink / raw)
To: ffmpeg-devel; +Cc: Niklas Haas
From: Niklas Haas <git@haasn.dev>
alphadetect8_full_c: 5658.2 ( 1.00x)
alphadetect8_full_avx2: 215.1 (26.31x)
alphadetect8_full_avx512: 133.5 (42.40x)
alphadetect8_limited_c: 7391.5 ( 1.00x)
alphadetect8_limited_avx2: 649.3 (11.38x)
alphadetect8_limited_avx512: 330.5 (22.36x)
alphadetect16_full_c: 3027.4 ( 1.00x)
alphadetect16_full_avx2: 209.4 (14.46x)
alphadetect16_full_avx512: 141.4 (21.41x)
alphadetect16_limited_c: 3880.9 ( 1.00x)
alphadetect16_limited_avx2: 734.9 ( 5.28x)
alphadetect16_limited_avx512: 349.2 (11.11x)
rangedetect8_c: 5854.2 ( 1.00x)
rangedetect8_avx2: 138.9 (42.15x)
rangedetect8_avx512: 106.2 (55.12x)
rangedetect16_c: 4122.0 ( 1.00x)
rangedetect16_avx2: 138.6 (29.74x)
rangedetect16_avx512: 104.1 (39.60x)
---
libavfilter/vf_colordetect.c | 4 +
libavfilter/vf_colordetect.h | 3 +
libavfilter/x86/Makefile | 2 +
libavfilter/x86/vf_colordetect.asm | 154 ++++++++++++++++++++++++++
libavfilter/x86/vf_colordetect_init.c | 105 ++++++++++++++++++
5 files changed, 268 insertions(+)
create mode 100644 libavfilter/x86/vf_colordetect.asm
create mode 100644 libavfilter/x86/vf_colordetect_init.c
diff --git a/libavfilter/vf_colordetect.c b/libavfilter/vf_colordetect.c
index 0fb892634f..642bec62c6 100644
--- a/libavfilter/vf_colordetect.c
+++ b/libavfilter/vf_colordetect.c
@@ -219,6 +219,10 @@ static av_cold void uninit(AVFilterContext *ctx)
av_cold void ff_color_detect_dsp_init(FFColorDetectDSPContext *dsp, int depth,
enum AVColorRange color_range)
{
+#if ARCH_X86
+ ff_color_detect_dsp_init_x86(dsp, depth, color_range);
+#endif
+
if (!dsp->detect_range)
dsp->detect_range = depth > 8 ? ff_detect_range16_c : ff_detect_range_c;
if (!dsp->detect_alpha) {
diff --git a/libavfilter/vf_colordetect.h b/libavfilter/vf_colordetect.h
index 8998ed83d4..78d296433d 100644
--- a/libavfilter/vf_colordetect.h
+++ b/libavfilter/vf_colordetect.h
@@ -41,6 +41,9 @@ typedef struct FFColorDetectDSPContext {
void ff_color_detect_dsp_init(FFColorDetectDSPContext *dsp, int depth,
enum AVColorRange color_range);
+void ff_color_detect_dsp_init_x86(FFColorDetectDSPContext *dsp, int depth,
+ enum AVColorRange color_range);
+
static inline int ff_detect_range_c(const uint8_t *data, ptrdiff_t stride,
ptrdiff_t width, ptrdiff_t height,
int mpeg_min, int mpeg_max)
diff --git a/libavfilter/x86/Makefile b/libavfilter/x86/Makefile
index 0efe3f8d2c..3a7f761ad4 100644
--- a/libavfilter/x86/Makefile
+++ b/libavfilter/x86/Makefile
@@ -5,6 +5,7 @@ OBJS-$(CONFIG_ANLMDN_FILTER) += x86/af_anlmdn_init.o
OBJS-$(CONFIG_ATADENOISE_FILTER) += x86/vf_atadenoise_init.o
OBJS-$(CONFIG_BLEND_FILTER) += x86/vf_blend_init.o
OBJS-$(CONFIG_BWDIF_FILTER) += x86/vf_bwdif_init.o
+OBJS-$(CONFIG_COLORDETECT_FILTER) += x86/vf_colordetect_init.o
OBJS-$(CONFIG_COLORSPACE_FILTER) += x86/colorspacedsp_init.o
OBJS-$(CONFIG_CONVOLUTION_FILTER) += x86/vf_convolution_init.o
OBJS-$(CONFIG_EBUR128_FILTER) += x86/f_ebur128_init.o
@@ -51,6 +52,7 @@ X86ASM-OBJS-$(CONFIG_ANLMDN_FILTER) += x86/af_anlmdn.o
X86ASM-OBJS-$(CONFIG_ATADENOISE_FILTER) += x86/vf_atadenoise.o
X86ASM-OBJS-$(CONFIG_BLEND_FILTER) += x86/vf_blend.o
X86ASM-OBJS-$(CONFIG_BWDIF_FILTER) += x86/vf_bwdif.o
+X86ASM-OBJS-$(CONFIG_COLORDETECT_FILTER) += x86/vf_colordetect.o
X86ASM-OBJS-$(CONFIG_COLORSPACE_FILTER) += x86/colorspacedsp.o
X86ASM-OBJS-$(CONFIG_CONVOLUTION_FILTER) += x86/vf_convolution.o
X86ASM-OBJS-$(CONFIG_EBUR128_FILTER) += x86/f_ebur128.o
diff --git a/libavfilter/x86/vf_colordetect.asm b/libavfilter/x86/vf_colordetect.asm
new file mode 100644
index 0000000000..38ffbff757
--- /dev/null
+++ b/libavfilter/x86/vf_colordetect.asm
@@ -0,0 +1,154 @@
+;*****************************************************************************
+;* x86-optimized functions for blackdetect filter
+;*
+;* Copyright (C) 2025 Niklas Haas
+;*
+;* 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/x86/x86util.asm"
+
+SECTION .text
+
+%macro detect_range_fn 1 ; suffix
+cglobal detect_range%1, 6, 7, 5, data, stride, width, height, mpeg_min, mpeg_max, x
+ movd xm0, mpeg_mind
+ movd xm1, mpeg_maxd
+ vpbroadcast%1 m0, xm0
+ vpbroadcast%1 m1, xm1
+ add dataq, widthq
+ neg widthq
+.lineloop:
+ mova m2, m0
+ mova m3, m1
+ mov xq, widthq
+ .loop:
+ movu m4, [dataq + xq]
+ pminu%1 m2, m4
+ pmaxu%1 m3, m4
+ add xq, mmsize
+ jl .loop
+
+ ; test if the data is out of range
+ pxor m2, m0
+%if cpuflag(avx512)
+ vpternlogq m2, m3, m1, 0xF6 ; m2 |= m3 ^ m1
+ vptestmq k1, m2, m2
+ kortestb k1, k1
+%else
+ pxor m3, m1
+ por m2, m3
+ ptest m2, m2
+%endif
+ jnz .end
+ add dataq, strideq
+ dec heightq
+ jg .lineloop
+.end:
+ setnz al
+ movzx rax, al
+ RET
+%endmacro
+
+%macro detect_alpha_fn 3 ; suffix, hsuffix, range
+cglobal detect_alpha%1_%3, 6, 7, 6, color, color_stride, alpha, alpha_stride, width, height, x
+ pxor m0, m0
+ add colorq, widthq
+ add alphaq, widthq
+ neg widthq
+%ifidn %3, limited
+%if ARCH_X86_64
+ movq xm3, r6mp ; p
+ movq xm4, r7mp ; q
+ movq xm5, r8mp ; k
+%else
+ movd xm3, r6mp ; p
+ movd xm4, r7mp ; q
+ movd xm5, r8mp ; k
+%endif
+ vpbroadcast%2 m3, xm3
+ vpbroadcast%2 m4, xm4
+ vpbroadcast%2 m5, xm5
+%endif
+.lineloop:
+ mov xq, widthq
+ .loop:
+ %ifidn %3, full
+ movu m1, [colorq + xq]
+ movu m2, [alphaq + xq]
+ pmaxu%1 m1, m2
+ %else
+ pmovzx%1%2 m1, [colorq + xq]
+ pmovzx%1%2 m2, [alphaq + xq]
+ pmull%2 m1, m3
+ pmull%2 m2, m4
+ %ifidn %1, b
+ psubusw m1, m5
+ %else
+ pmaxu%2 m1, m5
+ psub%2 m1, m5
+ %endif
+ pmaxu%2 m1, m2
+ %endif
+ %if cpuflag(avx512)
+ vpternlogq m0, m1, m2, 0xF6 ; m0 |= m1 ^ m2
+ %else
+ pxor m1, m2
+ por m0, m1
+ %endif
+ %ifidn %3, full
+ add xq, mmsize
+ %else
+ add xq, mmsize >> 1
+ %endif
+ jl .loop
+
+%if cpuflag(avx512)
+ vptestmq k1, m0, m0
+ kortestb k1, k1
+%else
+ ptest m0, m0
+%endif
+ jnz .found
+
+ add colorq, color_strideq
+ add alphaq, alpha_strideq
+ dec heightq
+ jg .lineloop
+ xor rax, rax
+ RET
+
+.found:
+ mov rax, 1
+ RET
+%endmacro
+
+INIT_YMM avx2
+detect_range_fn b
+detect_range_fn w
+detect_alpha_fn b, w, full
+detect_alpha_fn w, d, full
+detect_alpha_fn b, w, limited
+detect_alpha_fn w, d, limited
+
+INIT_ZMM avx512
+detect_range_fn b
+detect_range_fn w
+detect_alpha_fn b, w, full
+detect_alpha_fn w, d, full
+detect_alpha_fn b, w, limited
+detect_alpha_fn w, d, limited
diff --git a/libavfilter/x86/vf_colordetect_init.c b/libavfilter/x86/vf_colordetect_init.c
new file mode 100644
index 0000000000..62a7e87388
--- /dev/null
+++ b/libavfilter/x86/vf_colordetect_init.c
@@ -0,0 +1,105 @@
+/*
+ * Copyright (c) 2025 Niklas Haas
+ *
+ * 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/attributes.h"
+#include "libavutil/x86/cpu.h"
+#include "libavfilter/vf_colordetect.h"
+
+#define DETECT_RANGE_FUNC(FUNC_NAME, ASM_FUNC_NAME, C_FUNC_NAME, SHIFT, MMSIZE) \
+int ASM_FUNC_NAME(const uint8_t *src, ptrdiff_t stride, \
+ ptrdiff_t width, ptrdiff_t height, int min, int max); \
+ \
+static int FUNC_NAME(const uint8_t *src, ptrdiff_t stride, \
+ ptrdiff_t width, ptrdiff_t height, int min, int max) \
+{ \
+ ptrdiff_t bytes = (width << SHIFT) & ~(MMSIZE - 1); \
+ int ret = ASM_FUNC_NAME(src, stride, bytes, height, min, max); \
+ if (ret) \
+ return ret; \
+ \
+ return C_FUNC_NAME(src + bytes, stride, width - (bytes >> SHIFT), \
+ height, min, max); \
+}
+
+#define DETECT_ALPHA_FUNC(FUNC_NAME, ASM_FUNC_NAME, C_FUNC_NAME, SHIFT, MMSIZE) \
+int ASM_FUNC_NAME(const uint8_t *color, ptrdiff_t color_stride, \
+ const uint8_t *alpha, ptrdiff_t alpha_stride, \
+ ptrdiff_t width, ptrdiff_t height, int p, int q, int k); \
+ \
+static int FUNC_NAME(const uint8_t *color, ptrdiff_t color_stride, \
+ const uint8_t *alpha, ptrdiff_t alpha_stride, \
+ ptrdiff_t width, ptrdiff_t height, int p, int q, int k) \
+{ \
+ ptrdiff_t bytes = (width << SHIFT) & ~(MMSIZE - 1); \
+ int ret = ASM_FUNC_NAME(color, color_stride, alpha, alpha_stride, \
+ bytes, height, p, q, k); \
+ if (ret) \
+ return ret; \
+ \
+ return C_FUNC_NAME(color + bytes, color_stride, alpha + bytes, alpha_stride,\
+ width - (bytes >> SHIFT), height, p, q, k); \
+}
+
+#if HAVE_X86ASM
+#if HAVE_AVX512_EXTERNAL
+DETECT_RANGE_FUNC(detect_range_avx512, ff_detect_rangeb_avx512, ff_detect_range_c, 0, 64)
+DETECT_RANGE_FUNC(detect_range16_avx512, ff_detect_rangew_avx512, ff_detect_range16_c, 1, 64)
+DETECT_ALPHA_FUNC(detect_alpha_full_avx512, ff_detect_alphab_full_avx512, ff_detect_alpha_full_c, 0, 64)
+DETECT_ALPHA_FUNC(detect_alpha16_full_avx512, ff_detect_alphaw_full_avx512, ff_detect_alpha16_full_c, 1, 64)
+DETECT_ALPHA_FUNC(detect_alpha_limited_avx512, ff_detect_alphab_limited_avx512, ff_detect_alpha_limited_c, 0, 64)
+DETECT_ALPHA_FUNC(detect_alpha16_limited_avx512, ff_detect_alphaw_limited_avx512, ff_detect_alpha16_limited_c, 1, 64)
+#endif
+#if HAVE_AVX2_EXTERNAL
+DETECT_RANGE_FUNC(detect_range_avx2, ff_detect_rangeb_avx2, ff_detect_range_c, 0, 32)
+DETECT_RANGE_FUNC(detect_range16_avx2, ff_detect_rangew_avx2, ff_detect_range16_c, 1, 32)
+DETECT_ALPHA_FUNC(detect_alpha_full_avx2, ff_detect_alphab_full_avx2, ff_detect_alpha_full_c, 0, 32)
+DETECT_ALPHA_FUNC(detect_alpha16_full_avx2, ff_detect_alphaw_full_avx2, ff_detect_alpha16_full_c, 1, 32)
+DETECT_ALPHA_FUNC(detect_alpha_limited_avx2, ff_detect_alphab_limited_avx2, ff_detect_alpha_limited_c, 0, 32)
+DETECT_ALPHA_FUNC(detect_alpha16_limited_avx2, ff_detect_alphaw_limited_avx2, ff_detect_alpha16_limited_c, 1, 32)
+#endif
+#endif
+
+av_cold void ff_color_detect_dsp_init_x86(FFColorDetectDSPContext *dsp, int depth,
+ enum AVColorRange color_range)
+{
+#if HAVE_X86ASM
+ int cpu_flags = av_get_cpu_flags();
+#if HAVE_AVX2_EXTERNAL
+ if (EXTERNAL_AVX2(cpu_flags)) {
+ dsp->detect_range = depth > 8 ? detect_range16_avx2 : detect_range_avx2;
+ if (color_range == AVCOL_RANGE_JPEG) {
+ dsp->detect_alpha = depth > 8 ? detect_alpha16_full_avx2 : detect_alpha_full_avx2;
+ } else {
+ dsp->detect_alpha = depth > 8 ? detect_alpha16_limited_avx2 : detect_alpha_limited_avx2;
+ }
+ }
+#endif
+#if HAVE_AVX512_EXTERNAL
+ if (EXTERNAL_AVX512(cpu_flags)) {
+ dsp->detect_range = depth > 8 ? detect_range16_avx512 : detect_range_avx512;
+ if (color_range == AVCOL_RANGE_JPEG) {
+ dsp->detect_alpha = depth > 8 ? detect_alpha16_full_avx512 : detect_alpha_full_avx512;
+ } else {
+ dsp->detect_alpha = depth > 8 ? detect_alpha16_limited_avx512 : detect_alpha_limited_avx512;
+ }
+ }
+#endif
+#endif
+}
--
2.50.1
_______________________________________________
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
https://ffmpeg.org/mailman/listinfo/ffmpeg-devel
To unsubscribe, visit link above, or email
ffmpeg-devel-request@ffmpeg.org with subject "unsubscribe".
^ permalink raw reply [flat|nested] 11+ messages in thread* Re: [FFmpeg-devel] [PATCH v2 3/3] avfilter/vf_colordetect: add x86 SIMD implementation
2025-07-16 16:24 ` [FFmpeg-devel] [PATCH v2 3/3] avfilter/vf_colordetect: add x86 SIMD implementation Niklas Haas
@ 2025-07-16 20:06 ` Henrik Gramner via ffmpeg-devel
2025-07-17 9:37 ` Niklas Haas
2025-07-16 20:25 ` James Almer
1 sibling, 1 reply; 11+ messages in thread
From: Henrik Gramner via ffmpeg-devel @ 2025-07-16 20:06 UTC (permalink / raw)
To: FFmpeg development discussions and patches; +Cc: Henrik Gramner
On Wed, Jul 16, 2025 at 6:26 PM Niklas Haas <ffmpeg@haasn.xyz> wrote:
> +cglobal detect_range%1, 6, 7, 5, data, stride, width, height, mpeg_min, mpeg_max, x
> + movd xm0, mpeg_mind
> + movd xm1, mpeg_maxd
> + vpbroadcast%1 m0, xm0
> + vpbroadcast%1 m1, xm1
You could perhaps also do something like the following to shave off a
few instructions:
cglobal detect_range%1, 4, 7, 5, data, stride, width, height,
mpeg_min, mpeg_max, x
%if UNIX64 && notcpuflag(avx512)
movd xm0, mpeg_mind
movd xm1, mpeg_maxd
vpbroadcast%1 m0, xm0
vpbroadcast%1 m1, xm1
%else
vpbroadcast%1 m0, mpeg_minm
vpbroadcast%1 m0, mpeg_maxm
%endif
_______________________________________________
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
https://ffmpeg.org/mailman/listinfo/ffmpeg-devel
To unsubscribe, visit link above, or email
ffmpeg-devel-request@ffmpeg.org with subject "unsubscribe".
^ permalink raw reply [flat|nested] 11+ messages in thread* Re: [FFmpeg-devel] [PATCH v2 3/3] avfilter/vf_colordetect: add x86 SIMD implementation
2025-07-16 20:06 ` Henrik Gramner via ffmpeg-devel
@ 2025-07-17 9:37 ` Niklas Haas
0 siblings, 0 replies; 11+ messages in thread
From: Niklas Haas @ 2025-07-17 9:37 UTC (permalink / raw)
To: Henrik Gramner via ffmpeg-devel,
FFmpeg development discussions and patches
Cc: Henrik Gramner
On Wed, 16 Jul 2025 22:06:28 +0200 Henrik Gramner via ffmpeg-devel <ffmpeg-devel@ffmpeg.org> wrote:
> On Wed, Jul 16, 2025 at 6:26 PM Niklas Haas <ffmpeg@haasn.xyz> wrote:
> > +cglobal detect_range%1, 6, 7, 5, data, stride, width, height, mpeg_min, mpeg_max, x
> > + movd xm0, mpeg_mind
> > + movd xm1, mpeg_maxd
> > + vpbroadcast%1 m0, xm0
> > + vpbroadcast%1 m1, xm1
>
> You could perhaps also do something like the following to shave off a
> few instructions:
>
> cglobal detect_range%1, 4, 7, 5, data, stride, width, height,
> mpeg_min, mpeg_max, x
> %if UNIX64 && notcpuflag(avx512)
> movd xm0, mpeg_mind
> movd xm1, mpeg_maxd
> vpbroadcast%1 m0, xm0
> vpbroadcast%1 m1, xm1
> %else
> vpbroadcast%1 m0, mpeg_minm
> vpbroadcast%1 m0, mpeg_maxm
> %endif
Broadcasting directly from an address works on AVX2 as well, no? Why the
conditional here?
> _______________________________________________
> ffmpeg-devel mailing list
> ffmpeg-devel@ffmpeg.org
> https://ffmpeg.org/mailman/listinfo/ffmpeg-devel
>
> To unsubscribe, visit link above, or email
> ffmpeg-devel-request@ffmpeg.org with subject "unsubscribe".
_______________________________________________
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
https://ffmpeg.org/mailman/listinfo/ffmpeg-devel
To unsubscribe, visit link above, or email
ffmpeg-devel-request@ffmpeg.org with subject "unsubscribe".
^ permalink raw reply [flat|nested] 11+ messages in thread
* Re: [FFmpeg-devel] [PATCH v2 3/3] avfilter/vf_colordetect: add x86 SIMD implementation
2025-07-16 16:24 ` [FFmpeg-devel] [PATCH v2 3/3] avfilter/vf_colordetect: add x86 SIMD implementation Niklas Haas
2025-07-16 20:06 ` Henrik Gramner via ffmpeg-devel
@ 2025-07-16 20:25 ` James Almer
2025-07-16 21:49 ` Niklas Haas
2025-07-17 9:41 ` Niklas Haas
1 sibling, 2 replies; 11+ messages in thread
From: James Almer @ 2025-07-16 20:25 UTC (permalink / raw)
To: ffmpeg-devel
[-- Attachment #1.1.1: Type: text/plain, Size: 1051 bytes --]
On 7/16/2025 1:24 PM, Niklas Haas wrote:
> +cglobal detect_alpha%1_%3, 6, 7, 6, color, color_stride, alpha, alpha_stride, width, height, x
> + pxor m0, m0
> + add colorq, widthq
> + add alphaq, widthq
> + neg widthq
> +%ifidn %3, limited
> +%if ARCH_X86_64
> + movq xm3, r6mp ; p
> + movq xm4, r7mp ; q
> + movq xm5, r8mp ; k
They are int arguments, so no point using movq.
> +%else
> + movd xm3, r6mp ; p
> + movd xm4, r7mp ; q
> + movd xm5, r8mp ; k
> +%endif
> + vpbroadcast%2 m3, xm3
> + vpbroadcast%2 m4, xm4
> + vpbroadcast%2 m5, xm5
> +%endif
> +.lineloop:
> + mov xq, widthq
> + .loop:
> + %ifidn %3, full
> + movu m1, [colorq + xq]
> + movu m2, [alphaq + xq]
> + pmaxu%1 m1, m2
> + %else
> + pmovzx%1%2 m1, [colorq + xq]
> + pmovzx%1%2 m2, [alphaq + xq]
> + pmull%2 m1, m3
> + pmull%2 m2, m4
pmulld is painfully slow, so try to instead use pmaddwd. It should be
doable because the input are words.
[-- Attachment #1.2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 495 bytes --]
[-- Attachment #2: Type: text/plain, Size: 251 bytes --]
_______________________________________________
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
https://ffmpeg.org/mailman/listinfo/ffmpeg-devel
To unsubscribe, visit link above, or email
ffmpeg-devel-request@ffmpeg.org with subject "unsubscribe".
^ permalink raw reply [flat|nested] 11+ messages in thread
* Re: [FFmpeg-devel] [PATCH v2 3/3] avfilter/vf_colordetect: add x86 SIMD implementation
2025-07-16 20:25 ` James Almer
@ 2025-07-16 21:49 ` Niklas Haas
2025-07-16 22:21 ` James Almer
2025-07-17 9:41 ` Niklas Haas
1 sibling, 1 reply; 11+ messages in thread
From: Niklas Haas @ 2025-07-16 21:49 UTC (permalink / raw)
To: ffmpeg-devel
On Wed, 16 Jul 2025 17:25:12 -0300 James Almer <jamrial@gmail.com> wrote:
> On 7/16/2025 1:24 PM, Niklas Haas wrote:
> > +cglobal detect_alpha%1_%3, 6, 7, 6, color, color_stride, alpha, alpha_stride, width, height, x
> > + pxor m0, m0
> > + add colorq, widthq
> > + add alphaq, widthq
> > + neg widthq
> > +%ifidn %3, limited
> > +%if ARCH_X86_64
> > + movq xm3, r6mp ; p
> > + movq xm4, r7mp ; q
> > + movq xm5, r8mp ; k
>
> They are int arguments, so no point using movq.
This is to avoid an error:
src/libavfilter/x86/vf_colordetect.asm:145: error: invalid combination of opcode and operands
src/libavfilter/x86/vf_colordetect.asm:75: ... from macro `detect_alpha_fn' defined here
src//libavutil/x86/x86inc.asm:1518: ... from macro `movd' defined here
src//libavutil/x86/x86inc.asm:1501: ... from macro `RUN_AVX_INSTR' defined here
>
> > +%else
> > + movd xm3, r6mp ; p
> > + movd xm4, r7mp ; q
> > + movd xm5, r8mp ; k
> > +%endif
> > + vpbroadcast%2 m3, xm3
> > + vpbroadcast%2 m4, xm4
> > + vpbroadcast%2 m5, xm5
> > +%endif
> > +.lineloop:
> > + mov xq, widthq
> > + .loop:
> > + %ifidn %3, full
> > + movu m1, [colorq + xq]
> > + movu m2, [alphaq + xq]
> > + pmaxu%1 m1, m2
> > + %else
> > + pmovzx%1%2 m1, [colorq + xq]
> > + pmovzx%1%2 m2, [alphaq + xq]
> > + pmull%2 m1, m3
> > + pmull%2 m2, m4
>
> pmulld is painfully slow, so try to instead use pmaddwd. It should be
> doable because the input are words.
Will give it a try tomorrow.
>
> _______________________________________________
> ffmpeg-devel mailing list
> ffmpeg-devel@ffmpeg.org
> https://ffmpeg.org/mailman/listinfo/ffmpeg-devel
>
> To unsubscribe, visit link above, or email
> ffmpeg-devel-request@ffmpeg.org with subject "unsubscribe".
_______________________________________________
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
https://ffmpeg.org/mailman/listinfo/ffmpeg-devel
To unsubscribe, visit link above, or email
ffmpeg-devel-request@ffmpeg.org with subject "unsubscribe".
^ permalink raw reply [flat|nested] 11+ messages in thread
* Re: [FFmpeg-devel] [PATCH v2 3/3] avfilter/vf_colordetect: add x86 SIMD implementation
2025-07-16 21:49 ` Niklas Haas
@ 2025-07-16 22:21 ` James Almer
0 siblings, 0 replies; 11+ messages in thread
From: James Almer @ 2025-07-16 22:21 UTC (permalink / raw)
To: ffmpeg-devel
[-- Attachment #1.1.1: Type: text/plain, Size: 2540 bytes --]
On 7/16/2025 6:49 PM, Niklas Haas wrote:
> On Wed, 16 Jul 2025 17:25:12 -0300 James Almer <jamrial@gmail.com> wrote:
>> On 7/16/2025 1:24 PM, Niklas Haas wrote:
>>> +cglobal detect_alpha%1_%3, 6, 7, 6, color, color_stride, alpha, alpha_stride, width, height, x
>>> + pxor m0, m0
>>> + add colorq, widthq
>>> + add alphaq, widthq
>>> + neg widthq
>>> +%ifidn %3, limited
>>> +%if ARCH_X86_64
>>> + movq xm3, r6mp ; p
>>> + movq xm4, r7mp ; q
>>> + movq xm5, r8mp ; k
>>
>> They are int arguments, so no point using movq.
>
> This is to avoid an error:
>
> src/libavfilter/x86/vf_colordetect.asm:145: error: invalid combination of opcode and operands
> src/libavfilter/x86/vf_colordetect.asm:75: ... from macro `detect_alpha_fn' defined here
> src//libavutil/x86/x86inc.asm:1518: ... from macro `movd' defined here
> src//libavutil/x86/x86inc.asm:1501: ... from macro `RUN_AVX_INSTR' defined here
That's probably because of the mp suffix. If you use m, it works (as in,
"movd xm3, r6m").
In any case, you can just combine these into a single vpbroadcast%2 that
reads from memory:
%ifidn %3, limited
vpbroadcast%2 m3, r6m
vpbroadcast%2 m4, r7m
vpbroadcast%2 m5, r8m
%endif
>
>>
>>> +%else
>>> + movd xm3, r6mp ; p
>>> + movd xm4, r7mp ; q
>>> + movd xm5, r8mp ; k
>>> +%endif
>>> + vpbroadcast%2 m3, xm3
>>> + vpbroadcast%2 m4, xm4
>>> + vpbroadcast%2 m5, xm5
>>> +%endif
>>> +.lineloop:
>>> + mov xq, widthq
>>> + .loop:
>>> + %ifidn %3, full
>>> + movu m1, [colorq + xq]
>>> + movu m2, [alphaq + xq]
>>> + pmaxu%1 m1, m2
>>> + %else
>>> + pmovzx%1%2 m1, [colorq + xq]
>>> + pmovzx%1%2 m2, [alphaq + xq]
>>> + pmull%2 m1, m3
>>> + pmull%2 m2, m4
>>
>> pmulld is painfully slow, so try to instead use pmaddwd. It should be
>> doable because the input are words.
>
> Will give it a try tomorrow.
>
>>
>> _______________________________________________
>> ffmpeg-devel mailing list
>> ffmpeg-devel@ffmpeg.org
>> https://ffmpeg.org/mailman/listinfo/ffmpeg-devel
>>
>> To unsubscribe, visit link above, or email
>> ffmpeg-devel-request@ffmpeg.org with subject "unsubscribe".
> _______________________________________________
> ffmpeg-devel mailing list
> ffmpeg-devel@ffmpeg.org
> https://ffmpeg.org/mailman/listinfo/ffmpeg-devel
>
> To unsubscribe, visit link above, or email
> ffmpeg-devel-request@ffmpeg.org with subject "unsubscribe".
[-- Attachment #1.2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 495 bytes --]
[-- Attachment #2: Type: text/plain, Size: 251 bytes --]
_______________________________________________
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
https://ffmpeg.org/mailman/listinfo/ffmpeg-devel
To unsubscribe, visit link above, or email
ffmpeg-devel-request@ffmpeg.org with subject "unsubscribe".
^ permalink raw reply [flat|nested] 11+ messages in thread
* Re: [FFmpeg-devel] [PATCH v2 3/3] avfilter/vf_colordetect: add x86 SIMD implementation
2025-07-16 20:25 ` James Almer
2025-07-16 21:49 ` Niklas Haas
@ 2025-07-17 9:41 ` Niklas Haas
1 sibling, 0 replies; 11+ messages in thread
From: Niklas Haas @ 2025-07-17 9:41 UTC (permalink / raw)
To: ffmpeg-devel
On Wed, 16 Jul 2025 17:25:12 -0300 James Almer <jamrial@gmail.com> wrote:
> On 7/16/2025 1:24 PM, Niklas Haas wrote:
> > +cglobal detect_alpha%1_%3, 6, 7, 6, color, color_stride, alpha, alpha_stride, width, height, x
> > + pxor m0, m0
> > + add colorq, widthq
> > + add alphaq, widthq
> > + neg widthq
> > +%ifidn %3, limited
> > +%if ARCH_X86_64
> > + movq xm3, r6mp ; p
> > + movq xm4, r7mp ; q
> > + movq xm5, r8mp ; k
>
> They are int arguments, so no point using movq.
>
> > +%else
> > + movd xm3, r6mp ; p
> > + movd xm4, r7mp ; q
> > + movd xm5, r8mp ; k
> > +%endif
> > + vpbroadcast%2 m3, xm3
> > + vpbroadcast%2 m4, xm4
> > + vpbroadcast%2 m5, xm5
> > +%endif
> > +.lineloop:
> > + mov xq, widthq
> > + .loop:
> > + %ifidn %3, full
> > + movu m1, [colorq + xq]
> > + movu m2, [alphaq + xq]
> > + pmaxu%1 m1, m2
> > + %else
> > + pmovzx%1%2 m1, [colorq + xq]
> > + pmovzx%1%2 m2, [alphaq + xq]
> > + pmull%2 m1, m3
> > + pmull%2 m2, m4
>
> pmulld is painfully slow, so try to instead use pmaddwd. It should be
> doable because the input are words.
Wouldn't this add together adjacent pairs of words, which is not what we want?
Or do you mean I should construct two intermediate registers like
{p, 0, p, 0, p, 0, ...} and {0, p, 0, p, 0, p, ...} and then use two separate
pmaddwd instructions to process a full register?
> _______________________________________________
> ffmpeg-devel mailing list
> ffmpeg-devel@ffmpeg.org
> https://ffmpeg.org/mailman/listinfo/ffmpeg-devel
>
> To unsubscribe, visit link above, or email
> ffmpeg-devel-request@ffmpeg.org with subject "unsubscribe".
_______________________________________________
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
https://ffmpeg.org/mailman/listinfo/ffmpeg-devel
To unsubscribe, visit link above, or email
ffmpeg-devel-request@ffmpeg.org with subject "unsubscribe".
^ permalink raw reply [flat|nested] 11+ messages in thread