[FFmpeg-cvslog] [ffmpeg-web] branch master updated. c1acb1b web/security: Add more missing CVE#s
The branch, master has been updated via c1acb1b9bd2551a147fd422e96ed456da810aef3 (commit) from a4b40b1f993070377e98759e6db0a4d08a9649c5 (commit) - Log - commit c1acb1b9bd2551a147fd422e96ed456da810aef3 Author: Michael Niedermayer AuthorDate: Thu Nov 24 21:38:51 2022 +0100 Commit: Michael Niedermayer CommitDate: Thu Nov 24 21:38:51 2022 +0100 web/security: Add more missing CVE#s Signed-off-by: Michael Niedermayer diff --git a/src/security b/src/security index 270c455..aae87ee 100644 --- a/src/security +++ b/src/security @@ -572,6 +572,17 @@ CVE-2020-22046, 02fd294a333baaa55501eb0a26b86c99a80e4569 / 097c917c147661f5378da CVE-2020-22048, ee981f7ceb2c20dbfc5a2f5f27b0c44032eac798 / fddef964e8aa4a2c123e470db1436a082ff6bcf3, ticket/8303 +3.4.10 + +Fixes following vulnerabilities: + + +CVE-2020-20891, b239ccff7db0d418a74adcebfb1f2304f9a2f1f0 / 64a805883d7223c868a683f0030837d859edd2ab, ticket/8282 +CVE-2020-20892, 32a384519a57ad850789636c4c686091a53ce217 / 19587c9332f5be4f6bc6d7b2b8ef3fd21dfeaa01, ticket/8265 +CVE-2020-20896, b8197738d27f21583d9f83d7fa8c978d3a47af85 / dd01947397b98e94c3f2a79d5820aaf4594f4d3b, ticket/8273 +CVE-2020-20902, 04240e1d09e67c6e92189a96aeab96ef7428d942 / 2c78a76cb0443f8a12a5eadc3b58373aa2f4ab22, ticket/8176 + + 3.4.9 Fixes following vulnerabilities: @@ -587,6 +598,7 @@ CVE-2020-35965, 00115573e3030eff57847e1045ec18f0da5adb5c / 3e5959b3457f7f1856d99 CVE-2021-38114, e61b25e2557394e640a5aae901473785a4b23db5 / 7150f9575671f898382c370acae35f9087a30ba1 CVE-2021-38171, bc9e0b6cd2839acbac8da3232d715eb66857e453 / 9ffa49496d1aae4cbbb387aac28a9e061a6ab0a6 CVE-2021-38291, a4a3fd814aac900175ec4a2811cb5bf98c1ddad3 / e01d306c647b5827102260b885faa223b646d2d1, ticket/9312 +CVE-2020-23906, d46b698478f11ab85135b3cf0a7944c4dd62e37c / ec59dc73f0cc8930bf5dae389cd76d049d537ca7, ticket/8782 3.4.8 @@ -607,6 +619,10 @@ CVE-2019-12730, 59ac4182583e4791a7f98b79099916fd96beedfd / ed188f6dcdf0935c939ed CVE-2019-13390, cfa7c079f72b65bfe038af84d95d384a609d4f0a / aef24efb0c1e65097ab77a4bf9264189bdf3ace3 CVE-2019-17542, 4aaf644892843e3c68f4761725ab9435745f015c / 02f909dc24b1f05cfbba75077c7707b905e63cd2 CVE-2019-17539, c3b7afa4e917d748f0c3f8237b04ebdd99bdcacb / 8df6884832ec413cf032dfaa45c23b1c7876670c +CVE-2020-20448, 0c9ad1c746e3a8ccb7c6f292e10c8017c0a9dc3b / 8802e329c8317ca5ceb929df48a23eb0f9e852b2, ticket/7990 +CVE-2020-20448, e0d167051e93bad55a4c009399de1545aa07eeb5 / 55279d699fa64d8eb1185d8db04ab4ed92e8dea2, ticket/7990 +CVE-2020-20902, 4b4c26ca09b525168339df8697eb7f6bfe20345f / 0c61661a2cbe1b8b284c80ada1c2fdddf4992cad, ticket/8176 +CVE-2020-20902, f628f38f6e43c140167005593b447c47fd731a44 / 5f0acc5064ed501cb40d4aaccae2b3ce5c4552fd, ticket/8176 3.4.6 @@ -805,24 +821,6 @@ CVE-2017-7866, e371f031b942d73e02c090170975561fabd5c264 FFmpeg 3.2 -3.2.18 - -Fixes following vulnerabilities: - - -CVE-2020-20451, 0c949b6ebfcee1b23a5fe33a3bc8af167956ea1e / 21265f42ecb265debe9fec1dbfd0cb7de5a8aefb, ticket/8094 -CVE-2020-21041, 3d350ec7281cd0d357231fc2c99f44ebe425d586 / 5d9f44da460f781a1604d537d0555b78e29438ba, ticket/7989 -CVE-2020-22016, 02161c6ed194ddfa00fd2af7684a8099852bc3ce / 58aa0ed8f10753ee90f4a4a1f4f3da803cf7c145, ticket/8183 -CVE-2020-22020, 93ad1e4a9f17ac5145c2957bb270a454c8a0cefd / ce5274c1385d55892a692998923802023526b765, ticket/8239 -CVE-2020-22022, ea5d154845dfc1c6e550d197d7da79aee87c9f66 / 07050d7bdc32d82e53ee5bb727f5882323d00dba, ticket/8264 -CVE-2020-22025, ff73a50456b93e8d555603c093a3ebd193d0b097 / ccf4ab8c9aca0aee66bcc2914031a9c97ac0eeb8, ticket/8260 -CVE-2020-22031, 1a4d18820d551bedcfa03b7e8ca72df87d4b5cfa / 0e68e8c93f9068596484ec8ba725586860e06fc8, ticket/8243 -CVE-2020-22032, a19796a15ee0ab82e2b70d253d328111e9f916e0 / de598f82f8c3f8000e1948548e8088148e2b1f44, ticket/8275 -CVE-2020-22041, 4f566654e744c7810f4afdd91fe00fdd1ef46646 / 3488e0977c671568731afa12b811adce9d4d807f, ticket/8296 -CVE-2020-22044, 40dfd623632ed22bf3c98465ae3e68fcb1f31200 / 1d479300cbe0522c233b7d51148aea2b29bd29ad, ticket/8295 -CVE-2020-22046, 1a541dc0c5e1279251c9ed4cd416005fcca6e748 / 097c917c147661f5378dae8fe3f7e46f43236426, ticket/8294 - - 3.2.18 Fixes following vulnerabilities: @@ -846,14 +844,29 @@ CVE-2020-22046, bbc9751da67286d27f379dbe3b52ee3b55b0503e / 097c917c147661f5378da CVE-2020-22048, 64d2e0b20066058cf1c6dc3c49adab6d18d66fcc / fddef964e8aa4a2c123e470db1436a082ff6bcf3, ticket/8303 +3.2.17 + +Fixes following vulnerabilities: + + +CVE-2020-20891, f8b4426c10aa65f4c04847a50ebfdcb8782a49b7 / 64a805883d7223c868a683f0030837d859edd2ab, ticket/8282 +CVE-2020-20892, 94e502e96b0870177e0af4c1e8718ac71475e374 / 19587c9332f5be4f6bc6d7b2b8ef3fd21dfeaa01, ticket/8265 +CVE-2020-20902, abf9627f70ed8467b1646d56205e61f965f11468 / 2c78a76cb0443f8a12a5eadc3b58373aa2f4ab22, ticket/8176 + + 3.2.16 Fixes following vulnerabilities: CVE-2019-17539,
[FFmpeg-cvslog] swscale/utils: Move functions to avoid forward declarations
ffmpeg | branch: master | Andreas Rheinhardt |
Sat Nov 19 05:58:59 2022 +0100| [ff39dcb129806477e9a05c30dfdefb96f7fb0a25] |
committer: Andreas Rheinhardt
swscale/utils: Move functions to avoid forward declarations
Signed-off-by: Andreas Rheinhardt
> http://git.videolan.org/gitweb.cgi/ffmpeg.git/?a=commit;h=ff39dcb129806477e9a05c30dfdefb96f7fb0a25
---
libswscale/utils.c | 407 ++---
1 file changed, 200 insertions(+), 207 deletions(-)
diff --git a/libswscale/utils.c b/libswscale/utils.c
index fb788fc330..cdd89e4b58 100644
--- a/libswscale/utils.c
+++ b/libswscale/utils.c
@@ -59,13 +59,6 @@
#include "swscale.h"
#include "swscale_internal.h"
-static SwsVector *sws_getIdentityVec(void);
-static void sws_addVec(SwsVector *a, SwsVector *b);
-static void sws_shiftVec(SwsVector *a, int shift);
-static void sws_printVec2(SwsVector *a, AVClass *log_ctx, int log_level);
-
-static void handle_formats(SwsContext *c);
-
typedef struct FormatEntry {
uint8_t is_supported_in :1;
uint8_t is_supported_out:1;
@@ -926,6 +919,74 @@ static void fill_xyztables(struct SwsContext *c)
}
}
+static int handle_jpeg(enum AVPixelFormat *format)
+{
+switch (*format) {
+case AV_PIX_FMT_YUVJ420P:
+*format = AV_PIX_FMT_YUV420P;
+return 1;
+case AV_PIX_FMT_YUVJ411P:
+*format = AV_PIX_FMT_YUV411P;
+return 1;
+case AV_PIX_FMT_YUVJ422P:
+*format = AV_PIX_FMT_YUV422P;
+return 1;
+case AV_PIX_FMT_YUVJ444P:
+*format = AV_PIX_FMT_YUV444P;
+return 1;
+case AV_PIX_FMT_YUVJ440P:
+*format = AV_PIX_FMT_YUV440P;
+return 1;
+case AV_PIX_FMT_GRAY8:
+case AV_PIX_FMT_YA8:
+case AV_PIX_FMT_GRAY9LE:
+case AV_PIX_FMT_GRAY9BE:
+case AV_PIX_FMT_GRAY10LE:
+case AV_PIX_FMT_GRAY10BE:
+case AV_PIX_FMT_GRAY12LE:
+case AV_PIX_FMT_GRAY12BE:
+case AV_PIX_FMT_GRAY14LE:
+case AV_PIX_FMT_GRAY14BE:
+case AV_PIX_FMT_GRAY16LE:
+case AV_PIX_FMT_GRAY16BE:
+case AV_PIX_FMT_YA16BE:
+case AV_PIX_FMT_YA16LE:
+return 1;
+default:
+return 0;
+}
+}
+
+static int handle_0alpha(enum AVPixelFormat *format)
+{
+switch (*format) {
+case AV_PIX_FMT_0BGR: *format = AV_PIX_FMT_ABGR ; return 1;
+case AV_PIX_FMT_BGR0: *format = AV_PIX_FMT_BGRA ; return 4;
+case AV_PIX_FMT_0RGB: *format = AV_PIX_FMT_ARGB ; return 1;
+case AV_PIX_FMT_RGB0: *format = AV_PIX_FMT_RGBA ; return 4;
+default: return 0;
+}
+}
+
+static int handle_xyz(enum AVPixelFormat *format)
+{
+switch (*format) {
+case AV_PIX_FMT_XYZ12BE : *format = AV_PIX_FMT_RGB48BE; return 1;
+case AV_PIX_FMT_XYZ12LE : *format = AV_PIX_FMT_RGB48LE; return 1;
+default:return 0;
+}
+}
+
+static void handle_formats(SwsContext *c)
+{
+c->src0Alpha |= handle_0alpha(>srcFormat);
+c->dst0Alpha |= handle_0alpha(>dstFormat);
+c->srcXYZ|= handle_xyz(>srcFormat);
+c->dstXYZ|= handle_xyz(>dstFormat);
+if (c->srcXYZ || c->dstXYZ)
+fill_xyztables(c);
+}
+
static int range_override_needed(enum AVPixelFormat format)
{
return !isYUV(format) && !isGray(format);
@@ -1112,74 +1173,6 @@ int sws_getColorspaceDetails(struct SwsContext *c, int
**inv_table,
return 0;
}
-static int handle_jpeg(enum AVPixelFormat *format)
-{
-switch (*format) {
-case AV_PIX_FMT_YUVJ420P:
-*format = AV_PIX_FMT_YUV420P;
-return 1;
-case AV_PIX_FMT_YUVJ411P:
-*format = AV_PIX_FMT_YUV411P;
-return 1;
-case AV_PIX_FMT_YUVJ422P:
-*format = AV_PIX_FMT_YUV422P;
-return 1;
-case AV_PIX_FMT_YUVJ444P:
-*format = AV_PIX_FMT_YUV444P;
-return 1;
-case AV_PIX_FMT_YUVJ440P:
-*format = AV_PIX_FMT_YUV440P;
-return 1;
-case AV_PIX_FMT_GRAY8:
-case AV_PIX_FMT_YA8:
-case AV_PIX_FMT_GRAY9LE:
-case AV_PIX_FMT_GRAY9BE:
-case AV_PIX_FMT_GRAY10LE:
-case AV_PIX_FMT_GRAY10BE:
-case AV_PIX_FMT_GRAY12LE:
-case AV_PIX_FMT_GRAY12BE:
-case AV_PIX_FMT_GRAY14LE:
-case AV_PIX_FMT_GRAY14BE:
-case AV_PIX_FMT_GRAY16LE:
-case AV_PIX_FMT_GRAY16BE:
-case AV_PIX_FMT_YA16BE:
-case AV_PIX_FMT_YA16LE:
-return 1;
-default:
-return 0;
-}
-}
-
-static int handle_0alpha(enum AVPixelFormat *format)
-{
-switch (*format) {
-case AV_PIX_FMT_0BGR: *format = AV_PIX_FMT_ABGR ; return 1;
-case AV_PIX_FMT_BGR0: *format = AV_PIX_FMT_BGRA ; return 4;
-case AV_PIX_FMT_0RGB: *format = AV_PIX_FMT_ARGB ; return 1;
-case AV_PIX_FMT_RGB0: *format = AV_PIX_FMT_RGBA ; return 4;
-default: return 0;
-}
-}
-
-static int handle_xyz(enum AVPixelFormat *format)
-{
-switch (*format) {
-case
[FFmpeg-cvslog] swscale/utils: Fix indentation
ffmpeg | branch: master | Andreas Rheinhardt |
Mon Nov 21 00:50:29 2022 +0100| [1ff9c07fa696443d2d243e45d5794c8b19946a1b] |
committer: Andreas Rheinhardt
swscale/utils: Fix indentation
Forgotten after c1eb3e7fecdc270e03a700d61ef941600a6af491.
Signed-off-by: Andreas Rheinhardt
> http://git.videolan.org/gitweb.cgi/ffmpeg.git/?a=commit;h=1ff9c07fa696443d2d243e45d5794c8b19946a1b
---
libswscale/utils.c | 20 ++--
1 file changed, 10 insertions(+), 10 deletions(-)
diff --git a/libswscale/utils.c b/libswscale/utils.c
index 5a728afd89..90734f66ef 100644
--- a/libswscale/utils.c
+++ b/libswscale/utils.c
@@ -1307,16 +1307,16 @@ static av_cold int sws_init_single_context(SwsContext
*c, SwsFilter *srcFilter,
if (!(unscaled && sws_isSupportedEndiannessConversion(srcFormat) &&
av_pix_fmt_swap_endianness(srcFormat) == dstFormat)) {
-if (!sws_isSupportedInput(srcFormat)) {
-av_log(c, AV_LOG_ERROR, "%s is not supported as input pixel format\n",
- av_get_pix_fmt_name(srcFormat));
-return AVERROR(EINVAL);
-}
-if (!sws_isSupportedOutput(dstFormat)) {
-av_log(c, AV_LOG_ERROR, "%s is not supported as output pixel format\n",
- av_get_pix_fmt_name(dstFormat));
-return AVERROR(EINVAL);
-}
+if (!sws_isSupportedInput(srcFormat)) {
+av_log(c, AV_LOG_ERROR, "%s is not supported as input pixel
format\n",
+ av_get_pix_fmt_name(srcFormat));
+return AVERROR(EINVAL);
+}
+if (!sws_isSupportedOutput(dstFormat)) {
+av_log(c, AV_LOG_ERROR, "%s is not supported as output pixel
format\n",
+ av_get_pix_fmt_name(dstFormat));
+return AVERROR(EINVAL);
+}
}
av_assert2(desc_src && desc_dst);
___
ffmpeg-cvslog mailing list
ffmpeg-cvslog@ffmpeg.org
https://ffmpeg.org/mailman/listinfo/ffmpeg-cvslog
To unsubscribe, visit link above, or email
ffmpeg-cvslog-requ...@ffmpeg.org with subject "unsubscribe".
[FFmpeg-cvslog] swscale/utils: Derive range from YUVJ-pix-fmt only once
ffmpeg | branch: master | Andreas Rheinhardt |
Mon Nov 21 00:03:01 2022 +0100| [b2d1a258162a619187bbb0a72f7e8eb94f91cfa4] |
committer: Andreas Rheinhardt
swscale/utils: Derive range from YUVJ-pix-fmt only once
Currently, it is done once per slice-thread, leading to
one warning per slice-thread in case a YUVJ pixel format
has been originally used.
This also fixes the anomaly that said parameter are only
updated for the user-facing context (whose values are retrievable
via av_opt_get()) if slice-threading is not in use.
Fixes ticket #9860.
Signed-off-by: Andreas Rheinhardt
> http://git.videolan.org/gitweb.cgi/ffmpeg.git/?a=commit;h=b2d1a258162a619187bbb0a72f7e8eb94f91cfa4
---
libswscale/utils.c | 18 ++
1 file changed, 10 insertions(+), 8 deletions(-)
diff --git a/libswscale/utils.c b/libswscale/utils.c
index cdd89e4b58..5a728afd89 100644
--- a/libswscale/utils.c
+++ b/libswscale/utils.c
@@ -1277,8 +1277,7 @@ static av_cold int sws_init_single_context(SwsContext *c,
SwsFilter *srcFilter,
int dstH = c->dstH;
int dst_stride= FFALIGN(dstW * sizeof(int16_t) + 66, 16);
int flags, cpu_flags;
-enum AVPixelFormat srcFormat = c->srcFormat;
-enum AVPixelFormat dstFormat = c->dstFormat;
+enum AVPixelFormat srcFormat, dstFormat;
const AVPixFmtDescriptor *desc_src;
const AVPixFmtDescriptor *desc_dst;
int ret = 0;
@@ -1291,12 +1290,6 @@ static av_cold int sws_init_single_context(SwsContext
*c, SwsFilter *srcFilter,
unscaled = (srcW == dstW && srcH == dstH);
-c->srcRange |= handle_jpeg(>srcFormat);
-c->dstRange |= handle_jpeg(>dstFormat);
-
-if(srcFormat!=c->srcFormat || dstFormat!=c->dstFormat)
-av_log(c, AV_LOG_WARNING, "deprecated pixel format used, make sure you
did set range correctly\n");
-
if (!c->contrast && !c->saturation && !c->dstFormatBpp)
sws_setColorspaceDetails(c, ff_yuv2rgb_coeffs[SWS_CS_DEFAULT],
c->srcRange,
ff_yuv2rgb_coeffs[SWS_CS_DEFAULT],
@@ -2034,6 +2027,7 @@ av_cold int sws_init_context(SwsContext *c, SwsFilter
*srcFilter,
SwsFilter *dstFilter)
{
static AVOnce rgb2rgb_once = AV_ONCE_INIT;
+enum AVPixelFormat src_format, dst_format;
int ret;
c->frame_src = av_frame_alloc();
@@ -2044,6 +2038,14 @@ av_cold int sws_init_context(SwsContext *c, SwsFilter
*srcFilter,
if (ff_thread_once(_once, ff_sws_rgb2rgb_init) != 0)
return AVERROR_UNKNOWN;
+src_format = c->srcFormat;
+dst_format = c->dstFormat;
+c->srcRange |= handle_jpeg(>srcFormat);
+c->dstRange |= handle_jpeg(>dstFormat);
+
+if (src_format != c->srcFormat || dst_format != c->dstFormat)
+av_log(c, AV_LOG_WARNING, "deprecated pixel format used, make sure you
did set range correctly\n");
+
if (c->nb_threads != 1) {
ret = context_init_threaded(c, srcFilter, dstFilter);
if (ret < 0 || c->nb_threads > 1)
___
ffmpeg-cvslog mailing list
ffmpeg-cvslog@ffmpeg.org
https://ffmpeg.org/mailman/listinfo/ffmpeg-cvslog
To unsubscribe, visit link above, or email
ffmpeg-cvslog-requ...@ffmpeg.org with subject "unsubscribe".
[FFmpeg-cvslog] swscale/utils: Avoid calling ff_thread_once() unnecessarily
ffmpeg | branch: master | Andreas Rheinhardt |
Sat Nov 19 05:44:44 2022 +0100| [baccc1c5417f990ebfc1b6780e2dab255a72ee3c] |
committer: Andreas Rheinhardt
swscale/utils: Avoid calling ff_thread_once() unnecessarily
Signed-off-by: Andreas Rheinhardt
> http://git.videolan.org/gitweb.cgi/ffmpeg.git/?a=commit;h=baccc1c5417f990ebfc1b6780e2dab255a72ee3c
---
libswscale/utils.c | 7 ---
1 file changed, 4 insertions(+), 3 deletions(-)
diff --git a/libswscale/utils.c b/libswscale/utils.c
index 053c6bb76b..fb788fc330 100644
--- a/libswscale/utils.c
+++ b/libswscale/utils.c
@@ -1340,13 +1340,10 @@ static av_cold int sws_init_single_context(SwsContext
*c, SwsFilter *srcFilter,
int ret = 0;
enum AVPixelFormat tmpFmt;
static const float float_mult = 1.0f / 255.0f;
-static AVOnce rgb2rgb_once = AV_ONCE_INIT;
cpu_flags = av_get_cpu_flags();
flags = c->flags;
emms_c();
-if (ff_thread_once(_once, ff_sws_rgb2rgb_init) != 0)
-return AVERROR_UNKNOWN;
unscaled = (srcW == dstW && srcH == dstH);
@@ -2043,6 +2040,7 @@ fail: // FIXME replace things by appropriate error codes
av_cold int sws_init_context(SwsContext *c, SwsFilter *srcFilter,
SwsFilter *dstFilter)
{
+static AVOnce rgb2rgb_once = AV_ONCE_INIT;
int ret;
c->frame_src = av_frame_alloc();
@@ -2050,6 +2048,9 @@ av_cold int sws_init_context(SwsContext *c, SwsFilter
*srcFilter,
if (!c->frame_src || !c->frame_dst)
return AVERROR(ENOMEM);
+if (ff_thread_once(_once, ff_sws_rgb2rgb_init) != 0)
+return AVERROR_UNKNOWN;
+
if (c->nb_threads != 1) {
ret = context_init_threaded(c, srcFilter, dstFilter);
if (ret < 0 || c->nb_threads > 1)
___
ffmpeg-cvslog mailing list
ffmpeg-cvslog@ffmpeg.org
https://ffmpeg.org/mailman/listinfo/ffmpeg-cvslog
To unsubscribe, visit link above, or email
ffmpeg-cvslog-requ...@ffmpeg.org with subject "unsubscribe".
[FFmpeg-cvslog] swscale/utils: Don't allocate AVFrames for slice contexts
ffmpeg | branch: master | Andreas Rheinhardt |
Sat Nov 19 05:42:29 2022 +0100| [8ee071122806724a00eecb6b1eff639890c4be48] |
committer: Andreas Rheinhardt
swscale/utils: Don't allocate AVFrames for slice contexts
Only the parent context's AVFrames are ever used.
Signed-off-by: Andreas Rheinhardt
> http://git.videolan.org/gitweb.cgi/ffmpeg.git/?a=commit;h=8ee071122806724a00eecb6b1eff639890c4be48
---
libswscale/utils.c | 15 +--
1 file changed, 5 insertions(+), 10 deletions(-)
diff --git a/libswscale/utils.c b/libswscale/utils.c
index c6fa07f752..053c6bb76b 100644
--- a/libswscale/utils.c
+++ b/libswscale/utils.c
@@ -1317,11 +1317,6 @@ static int context_init_threaded(SwsContext *c,
}
}
-c->frame_src = av_frame_alloc();
-c->frame_dst = av_frame_alloc();
-if (!c->frame_src || !c->frame_dst)
-return AVERROR(ENOMEM);
-
return 0;
}
@@ -1581,11 +1576,6 @@ static av_cold int sws_init_single_context(SwsContext
*c, SwsFilter *srcFilter,
if (!FF_ALLOCZ_TYPED_ARRAY(c->formatConvBuffer, FFALIGN(srcW * 2 + 78, 16)
* 2))
goto nomem;
-c->frame_src = av_frame_alloc();
-c->frame_dst = av_frame_alloc();
-if (!c->frame_src || !c->frame_dst)
-goto nomem;
-
c->srcBpc = desc_src->comp[0].depth;
if (c->srcBpc < 8)
c->srcBpc = 8;
@@ -2055,6 +2045,11 @@ av_cold int sws_init_context(SwsContext *c, SwsFilter
*srcFilter,
{
int ret;
+c->frame_src = av_frame_alloc();
+c->frame_dst = av_frame_alloc();
+if (!c->frame_src || !c->frame_dst)
+return AVERROR(ENOMEM);
+
if (c->nb_threads != 1) {
ret = context_init_threaded(c, srcFilter, dstFilter);
if (ret < 0 || c->nb_threads > 1)
___
ffmpeg-cvslog mailing list
ffmpeg-cvslog@ffmpeg.org
https://ffmpeg.org/mailman/listinfo/ffmpeg-cvslog
To unsubscribe, visit link above, or email
ffmpeg-cvslog-requ...@ffmpeg.org with subject "unsubscribe".
[FFmpeg-cvslog] swscale/utils: Factor initializing single slice context out
ffmpeg | branch: master | Andreas Rheinhardt |
Sat Nov 19 04:58:45 2022 +0100| [64ed1d40df82949114ca5c4cbf33858ae94cc7f9] |
committer: Andreas Rheinhardt
swscale/utils: Factor initializing single slice context out
Initializing slice threads currently uses the function
(sws_init_context()) that is also used for initializing
user-facing contexts with the only difference being that
nb_threads is set to one before initializing the slice contexts.
Yet sws_init_context() also initializes lots of stuff
that is not slice-dependent, i.e. (src|dst)Range. This
currently only works because the code sets these fields
to the same values for all slice contexts. This is not
nice; even worse, it entails that log messages are printed
once per slice context (and therefore fill the screen).
This commit lays the groundwork to fix this.
Signed-off-by: Andreas Rheinhardt
> http://git.videolan.org/gitweb.cgi/ffmpeg.git/?a=commit;h=64ed1d40df82949114ca5c4cbf33858ae94cc7f9
---
libswscale/utils.c | 31 +--
1 file changed, 21 insertions(+), 10 deletions(-)
diff --git a/libswscale/utils.c b/libswscale/utils.c
index 85640a143f..c6fa07f752 100644
--- a/libswscale/utils.c
+++ b/libswscale/utils.c
@@ -1268,6 +1268,9 @@ static enum AVPixelFormat alphaless_fmt(enum
AVPixelFormat fmt)
}
}
+static int sws_init_single_context(SwsContext *c, SwsFilter *srcFilter,
+ SwsFilter *dstFilter);
+
static int context_init_threaded(SwsContext *c,
SwsFilter *src_filter, SwsFilter *dst_filter)
{
@@ -1301,7 +1304,7 @@ static int context_init_threaded(SwsContext *c,
c->slice_ctx[i]->nb_threads = 1;
-ret = sws_init_context(c->slice_ctx[i], src_filter, dst_filter);
+ret = sws_init_single_context(c->slice_ctx[i], src_filter, dst_filter);
if (ret < 0)
return ret;
@@ -1322,8 +1325,8 @@ static int context_init_threaded(SwsContext *c,
return 0;
}
-av_cold int sws_init_context(SwsContext *c, SwsFilter *srcFilter,
- SwsFilter *dstFilter)
+static av_cold int sws_init_single_context(SwsContext *c, SwsFilter *srcFilter,
+ SwsFilter *dstFilter)
{
int i;
int usesVFilter, usesHFilter;
@@ -1344,13 +1347,6 @@ av_cold int sws_init_context(SwsContext *c, SwsFilter
*srcFilter,
static const float float_mult = 1.0f / 255.0f;
static AVOnce rgb2rgb_once = AV_ONCE_INIT;
-if (c->nb_threads != 1) {
-ret = context_init_threaded(c, srcFilter, dstFilter);
-if (ret < 0 || c->nb_threads > 1)
-return ret;
-// threading disabled in this build, init as single-threaded
-}
-
cpu_flags = av_get_cpu_flags();
flags = c->flags;
emms_c();
@@ -2054,6 +2050,21 @@ fail: // FIXME replace things by appropriate error codes
return ret;
}
+av_cold int sws_init_context(SwsContext *c, SwsFilter *srcFilter,
+ SwsFilter *dstFilter)
+{
+int ret;
+
+if (c->nb_threads != 1) {
+ret = context_init_threaded(c, srcFilter, dstFilter);
+if (ret < 0 || c->nb_threads > 1)
+return ret;
+// threading disabled in this build, init as single-threaded
+}
+
+return sws_init_single_context(c, srcFilter, dstFilter);
+}
+
SwsContext *sws_alloc_set_opts(int srcW, int srcH, enum AVPixelFormat
srcFormat,
int dstW, int dstH, enum AVPixelFormat
dstFormat,
int flags, const double *param)
___
ffmpeg-cvslog mailing list
ffmpeg-cvslog@ffmpeg.org
https://ffmpeg.org/mailman/listinfo/ffmpeg-cvslog
To unsubscribe, visit link above, or email
ffmpeg-cvslog-requ...@ffmpeg.org with subject "unsubscribe".
[FFmpeg-cvslog] [ffmpeg-web] branch master updated. a4b40b1 web/security: add missing CVE#s
The branch, master has been updated via a4b40b1f993070377e98759e6db0a4d08a9649c5 (commit) from 5c52853ae8867e2aad1a9f8256bfc0e00302e363 (commit) - Log - commit a4b40b1f993070377e98759e6db0a4d08a9649c5 Author: Michael Niedermayer AuthorDate: Thu Nov 24 17:59:01 2022 +0100 Commit: Michael Niedermayer CommitDate: Thu Nov 24 18:01:18 2022 +0100 web/security: add missing CVE#s diff --git a/src/security b/src/security index 44fa15e..270c455 100644 --- a/src/security +++ b/src/security @@ -11,8 +11,25 @@ Fixes following vulnerabilities: CVE-2022-2566, 6f53f0d09ea4c9c7f7354f018a87ef840315207d / c953baa084607dd1d84c3bfcce3cf6a87c3e6e05 +5.1 + +Fixes following vulnerabilities: + + +CVE-2022-1475, 757da974b21833529cc41bdcc9684c29660cdfa8, ticket/9651 + + + FFmpeg 5.0 +5.0.1 + +Fixes following vulnerabilities: + + +CVE-2022-1475, 95322e07673885ebcbb8fd54f30a9b8f17d5be6a / 757da974b21833529cc41bdcc9684c29660cdfa8, ticket/9651 + + 5.0 Fixes following vulnerabilities: @@ -32,6 +49,14 @@ CVE-2021-38291, e01d306c647b5827102260b885faa223b646d2d1 ticket/9312, FFmpeg 4.4 +4.4.2 + +Fixes following vulnerabilities: + + +CVE-2022-1475, e9e2ddbc6c78cc18b76093617f82c920e58a8d1f / 757da974b21833529cc41bdcc9684c29660cdfa8, ticket/9651 + + 4.4.1 Fixes following vulnerabilities: @@ -61,6 +86,7 @@ CVE-2020-20450, 5400e4a50c61e53e1bc50b3e77201649bbe9c510, ticket/7993 CVE-2020-21041, 5d9f44da460f781a1604d537d0555b78e29438ba, ticket/7989 CVE-2020-22038, 7c32e9cf93b712f8463573a59ed4e98fd10fa013, ticket/8285 CVE-2020-22042, 426c16d61a9b5056a157a1a2a057a4e4d13eef84, ticket/8267 +CVE-2020-23906, ec59dc73f0cc8930bf5dae389cd76d049d537ca7, ticket/8782 CVE-2020-24020, 584f396132aa19d21bb1e38ad9a5d428869290cb, ticket/8718 CVE-2021-30123, d6f293353c94c7ce200f6e0975ae3de49787f91f, ticket/8845, never affected a release CVE-2020-35964, 27a99e2c7d450fef15594671eef4465c8a166bd7 @@ -78,6 +104,13 @@ Fixes following vulnerabilities: CVE-2020-21041, 50cadf8dc52e94372a181dd60a527c55d1d155f5 / 5d9f44da460f781a1604d537d0555b78e29438ba, ticket/7989 +4.3.4 + +Fixes following vulnerabilities: + + +CVE-2022-1475, fa2e4afe8d0a23fac37392ef6506cfc9841f8d3d / 757da974b21833529cc41bdcc9684c29660cdfa8, ticket/9651 + 4.3.3 @@ -113,6 +146,7 @@ Fixes following vulnerabilities: CVE-2020-13904, a3fdeb0c3a4ecabab2c2351b86fc92004526e9cc / b5e39880fb7269b1b3577cee288e06aa3dc1dfa2 CVE-2020-14212, dd273d359e45ab69398ac0dc41206d5f1a9371bf / 0b3bd001ac1745d9d008a2d195817df57d7d1d14 +CVE-2020-23906, be84216c53a4ed81573c82320e9c4a20e9b349d9 / ec59dc73f0cc8930bf5dae389cd76d049d537ca7, ticket/8782 4.3 @@ -128,6 +162,18 @@ CVE-2020-12284, 1812352d767ccf5431aa440123e2e260a4db2726 CVE-2020-20448, 55279d699fa64d8eb1185d8db04ab4ed92e8dea2 CVE-2020-20448, 8802e329c8317ca5ceb929df48a23eb0f9e852b2, ticket/7990 CVE-2020-20451, 21265f42ecb265debe9fec1dbfd0cb7de5a8aefb, ticket/8094 +CVE-2020-20891, 64a805883d7223c868a683f0030837d859edd2ab, ticket/8282 +CVE-2020-20892, 19587c9332f5be4f6bc6d7b2b8ef3fd21dfeaa01, ticket/8265 +CVE-2020-20896, dd01947397b98e94c3f2a79d5820aaf4594f4d3b, ticket/8273 +CVE-2020-20898, 99f8d32129dd233d4eb2efa44678a0bc44869f23, ticket/8263 +CVE-2021-38090, 99f8d32129dd233d4eb2efa44678a0bc44869f23, ticket/8263, duplicate CVE# +CVE-2021-38091, 99f8d32129dd233d4eb2efa44678a0bc44869f23, ticket/8263, duplicate CVE# +CVE-2021-38092, 99f8d32129dd233d4eb2efa44678a0bc44869f23, ticket/8263, duplicate CVE# +CVE-2021-38093, 99f8d32129dd233d4eb2efa44678a0bc44869f23, ticket/8263, duplicate CVE# +CVE-2021-38094, 99f8d32129dd233d4eb2efa44678a0bc44869f23, ticket/8263, duplicate CVE# +CVE-2020-20902, 0c61661a2cbe1b8b284c80ada1c2fdddf4992cad, ticket/8176 +CVE-2020-20902, 2c78a76cb0443f8a12a5eadc3b58373aa2f4ab22, ticket/8176 +CVE-2020-20902, 5f0acc5064ed501cb40d4aaccae2b3ce5c4552fd, ticket/8176 CVE-2020-22016, 58aa0ed8f10753ee90f4a4a1f4f3da803cf7c145, ticket/8183 CVE-2020-22017, d4d6b7b0355f3597cad3b8d12911790c73b5f96d, ticket/8309 CVE-2020-22020, ce5274c1385d55892a692998923802023526b765, ticket/8239 @@ -149,6 +195,7 @@ CVE-2020-22040, 1a0c584abc9709b1d11dbafef05d22e0937d7d19, ticket/8283 CVE-2020-22041, 3488e0977c671568731afa12b811adce9d4d807f, ticket/8296 CVE-2020-22043, b288a7eb3d963a175e177b6219c8271076ee8590, ticket/8284 CVE-2020-22044, 1d479300cbe0522c233b7d51148aea2b29bd29ad, ticket/8295 +CVE-2020-22046, 097c917c147661f5378dae8fe3f7e46f43236426, ticket/8294 FFmpeg 4.2 @@ -180,6 +227,9 @@ CVE-2020-22048, 7d4c2d90b3997542a2dece32a1234f3bc3629610 / fddef964e8aa4a2c123e4 Fixes following vulnerabilities: +CVE-2020-20891, 84fdfdf8595150c04b86febd1ef2eae3878c84b8 / 64a805883d7223c868a683f0030837d859edd2ab, ticket/8282 +CVE-2020-20892, 15900ff8e68f38404bd6d392d474d99f65cdbbf9 / 19587c9332f5be4f6bc6d7b2b8ef3fd21dfeaa01, ticket/8265 +CVE-2020-20896, c4629d8abe270ec5e5d79f7d18cd0b12cd5fd797 /
[FFmpeg-cvslog] lavu: bump minor and add APIchanges entry for lavu/tx DCT
ffmpeg | branch: master | Lynne | Sun Nov 20 21:17:30 2022 +0100| [e97368eba5b48a958d3b398780e56b12db92d1a1] | committer: Lynne lavu: bump minor and add APIchanges entry for lavu/tx DCT > http://git.videolan.org/gitweb.cgi/ffmpeg.git/?a=commit;h=e97368eba5b48a958d3b398780e56b12db92d1a1 --- doc/APIchanges | 3 +++ 1 file changed, 3 insertions(+) diff --git a/doc/APIchanges b/doc/APIchanges index 038ca865ec..ab7ce15fae 100644 --- a/doc/APIchanges +++ b/doc/APIchanges @@ -14,6 +14,9 @@ libavutil: 2021-04-27 API changes, most recent first: +2022-11-xx - xx - lavu 57.43.100 - tx.h + Add AV_TX_FLOAT_DCT, AV_TX_DOUBLE_DCT and AV_TX_INT32_DCT. + 2022-xx-xx - xx - lavu 57.42.100 - dict.h Add av_dict_iterate(). ___ ffmpeg-cvslog mailing list ffmpeg-cvslog@ffmpeg.org https://ffmpeg.org/mailman/listinfo/ffmpeg-cvslog To unsubscribe, visit link above, or email ffmpeg-cvslog-requ...@ffmpeg.org with subject "unsubscribe".
[FFmpeg-cvslog] binkaudio: convert to lavu/tx
ffmpeg | branch: master | Lynne | Sun Nov 20 04:57:00 2022
+0100| [ca8aaf24dfd28ceb4709fc518b3c95b7fce07dcc] | committer: Lynne
binkaudio: convert to lavu/tx
> http://git.videolan.org/gitweb.cgi/ffmpeg.git/?a=commit;h=ca8aaf24dfd28ceb4709fc518b3c95b7fce07dcc
---
configure | 4 ++--
libavcodec/binkaudio.c | 42 +-
2 files changed, 23 insertions(+), 23 deletions(-)
diff --git a/configure b/configure
index 25cd712b16..5cb80bf1ea 100755
--- a/configure
+++ b/configure
@@ -2772,8 +2772,8 @@ asv2_encoder_select="aandcttables bswapdsp fdctdsp
pixblockdsp"
atrac1_decoder_select="sinewin"
av1_decoder_select="av1_frame_split_bsf cbs_av1"
bink_decoder_select="blockdsp hpeldsp"
-binkaudio_dct_decoder_select="dct wma_freqs"
-binkaudio_rdft_decoder_select="rdft wma_freqs"
+binkaudio_dct_decoder_select="wma_freqs"
+binkaudio_rdft_decoder_select="wma_freqs"
cavs_decoder_select="blockdsp golomb h264chroma idctdsp qpeldsp videodsp"
clearvideo_decoder_select="idctdsp"
cllc_decoder_select="bswapdsp"
diff --git a/libavcodec/binkaudio.c b/libavcodec/binkaudio.c
index 43dca1f565..046bf93207 100644
--- a/libavcodec/binkaudio.c
+++ b/libavcodec/binkaudio.c
@@ -33,15 +33,14 @@
#include "libavutil/channel_layout.h"
#include "libavutil/intfloat.h"
#include "libavutil/mem_internal.h"
+#include "libavutil/tx.h"
#define BITSTREAM_READER_LE
#include "avcodec.h"
-#include "dct.h"
#include "decode.h"
#include "get_bits.h"
#include "codec_internal.h"
#include "internal.h"
-#include "rdft.h"
#include "wma_freqs.h"
#define MAX_DCT_CHANNELS 6
@@ -63,10 +62,8 @@ typedef struct BinkAudioContext {
float previous[MAX_DCT_CHANNELS][BINK_BLOCK_MAX_SIZE / 16]; ///< coeffs
from previous audio block
float quant_table[96];
AVPacket *pkt;
-union {
-RDFTContext rdft;
-DCTContext dct;
-} trans;
+AVTXContext *tx;
+av_tx_fn tx_fn;
} BinkAudioContext;
@@ -138,12 +135,15 @@ static av_cold int decode_init(AVCodecContext *avctx)
s->first = 1;
-if (CONFIG_BINKAUDIO_RDFT_DECODER && avctx->codec->id ==
AV_CODEC_ID_BINKAUDIO_RDFT)
-ret = ff_rdft_init(>trans.rdft, frame_len_bits, DFT_C2R);
-else if (CONFIG_BINKAUDIO_DCT_DECODER)
-ret = ff_dct_init(>trans.dct, frame_len_bits, DCT_III);
-else
+if (CONFIG_BINKAUDIO_RDFT_DECODER && avctx->codec->id ==
AV_CODEC_ID_BINKAUDIO_RDFT) {
+float scale = 0.5;
+ret = av_tx_init(>tx, >tx_fn, AV_TX_FLOAT_RDFT, 1, 1 <<
frame_len_bits, , 0);
+} else if (CONFIG_BINKAUDIO_DCT_DECODER) {
+float scale = 1.0 / (1 << frame_len_bits);
+ret = av_tx_init(>tx, >tx_fn, AV_TX_FLOAT_DCT, 1, 1 <<
(frame_len_bits - 1), , 0);
+} else {
av_assert0(0);
+}
if (ret < 0)
return ret;
@@ -177,13 +177,12 @@ static int decode_block(BinkAudioContext *s, float **out,
int use_dct,
float q, quant[25];
int width, coeff;
GetBitContext *gb = >gb;
+LOCAL_ALIGNED_32(float, coeffs, [4098]);
if (use_dct)
skip_bits(gb, 2);
for (ch = 0; ch < channels; ch++) {
-FFTSample *coeffs = out[ch + ch_offset];
-
if (s->version_b) {
if (get_bits_left(gb) < 64)
return AVERROR_INVALIDDATA;
@@ -251,10 +250,15 @@ static int decode_block(BinkAudioContext *s, float **out,
int use_dct,
if (CONFIG_BINKAUDIO_DCT_DECODER && use_dct) {
coeffs[0] /= 0.5;
-s->trans.dct.dct_calc(>trans.dct, coeffs);
+s->tx_fn(s->tx, out[ch + ch_offset], coeffs, sizeof(float));
+} else if (CONFIG_BINKAUDIO_RDFT_DECODER) {
+for (int i = 2; i < s->frame_len; i += 2)
+coeffs[i + 1] *= -1;
+
+coeffs[s->frame_len + 0] = coeffs[1];
+coeffs[s->frame_len + 1] = coeffs[1] = 0;
+s->tx_fn(s->tx, out[ch + ch_offset], coeffs,
sizeof(AVComplexFloat));
}
-else if (CONFIG_BINKAUDIO_RDFT_DECODER)
-s->trans.rdft.rdft_calc(>trans.rdft, coeffs);
}
for (ch = 0; ch < channels; ch++) {
@@ -278,11 +282,7 @@ static int decode_block(BinkAudioContext *s, float **out,
int use_dct,
static av_cold int decode_end(AVCodecContext *avctx)
{
BinkAudioContext * s = avctx->priv_data;
-if (CONFIG_BINKAUDIO_RDFT_DECODER && avctx->codec->id ==
AV_CODEC_ID_BINKAUDIO_RDFT)
-ff_rdft_end(>trans.rdft);
-else if (CONFIG_BINKAUDIO_DCT_DECODER)
-ff_dct_end(>trans.dct);
-
+av_tx_uninit(>tx);
return 0;
}
___
ffmpeg-cvslog mailing list
ffmpeg-cvslog@ffmpeg.org
https://ffmpeg.org/mailman/listinfo/ffmpeg-cvslog
To unsubscribe, visit link above, or email
ffmpeg-cvslog-requ...@ffmpeg.org with subject "unsubscribe".
[FFmpeg-cvslog] lavu/tx: clarify stride for RDFT transforms
ffmpeg | branch: master | Lynne | Thu Nov 24 15:56:01 2022
+0100| [93c30bd6f0846898bb3e7172bb5de65f2d0f33ce] | committer: Lynne
lavu/tx: clarify stride for RDFT transforms
> http://git.videolan.org/gitweb.cgi/ffmpeg.git/?a=commit;h=93c30bd6f0846898bb3e7172bb5de65f2d0f33ce
---
libavutil/tx.h | 4 +++-
1 file changed, 3 insertions(+), 1 deletion(-)
diff --git a/libavutil/tx.h b/libavutil/tx.h
index cd772ad903..758f634b73 100644
--- a/libavutil/tx.h
+++ b/libavutil/tx.h
@@ -75,7 +75,9 @@ enum AVTXType {
* the double variant, it's a 'double'. If scale is NULL, 1.0 will be used
* as a default.
*
- * The stride parameter must be set to the size of a single sample in
bytes.
+ * For forward transforms (R2C), stride must be the spacing between two
+ * samples in bytes. For inverse transforms, the stride must be set
+ * to the spacing between two complex values in bytes.
*
* The forward transform performs a real-to-complex DFT of N samples to
* N/2+1 complex values.
___
ffmpeg-cvslog mailing list
ffmpeg-cvslog@ffmpeg.org
https://ffmpeg.org/mailman/listinfo/ffmpeg-cvslog
To unsubscribe, visit link above, or email
ffmpeg-cvslog-requ...@ffmpeg.org with subject "unsubscribe".
[FFmpeg-cvslog] lavu/tx: add DCT-III implementation
ffmpeg | branch: master | Lynne | Sun Nov 20 03:44:29 2022
+0100| [a56d7e0ca3be82cb5155ab0cf8206fc8b8d6861d] | committer: Lynne
lavu/tx: add DCT-III implementation
> http://git.videolan.org/gitweb.cgi/ffmpeg.git/?a=commit;h=a56d7e0ca3be82cb5155ab0cf8206fc8b8d6861d
---
libavutil/tx_template.c | 72 +++--
1 file changed, 70 insertions(+), 2 deletions(-)
diff --git a/libavutil/tx_template.c b/libavutil/tx_template.c
index 5d73809b58..1de92b9786 100644
--- a/libavutil/tx_template.c
+++ b/libavutil/tx_template.c
@@ -1737,6 +1737,11 @@ static av_cold int TX_NAME(ff_tx_dct_init)(AVTXContext
*s,
TXSample *tab;
SCALE_TYPE rsc = *((SCALE_TYPE *)scale);
+if (inv) {
+len *= 2;
+s->len *= 2;
+rsc *= 0.5;
+}
if ((ret = ff_tx_init_subtx(s, TX_TYPE(RDFT), flags, NULL, len, inv,
)))
return ret;
@@ -1752,8 +1757,13 @@ static av_cold int TX_NAME(ff_tx_dct_init)(AVTXContext
*s,
for (int i = 0; i < len; i++)
tab[i] = RESCALE(cos(i*freq)*(!inv + 1));
-for (int i = 0; i < len/2; i++)
-tab[len + i] = RESCALE(cos((len - 2*i - 1)*freq));
+if (inv) {
+for (int i = 0; i < len/2; i++)
+tab[len + i] = RESCALE(0.5 / sin((2*i + 1)*freq));
+} else {
+for (int i = 0; i < len/2; i++)
+tab[len + i] = RESCALE(cos((len - 2*i - 1)*freq));
+}
return 0;
}
@@ -1818,6 +1828,49 @@ static void TX_NAME(ff_tx_dctII)(AVTXContext *s, void
*_dst,
dst[1] = next;
}
+static void TX_NAME(ff_tx_dctIII)(AVTXContext *s, void *_dst,
+ void *_src, ptrdiff_t stride)
+{
+TXSample *dst = _dst;
+TXSample *src = _src;
+const int len = s->len;
+const int len2 = len >> 1;
+const TXSample *exp = (void *)s->exp;
+#ifdef TX_INT32
+int64_t tmp1, tmp2 = src[len - 1];
+tmp2 = (2*tmp2 + 0x4000) >> 31;
+#else
+TXSample tmp1, tmp2 = 2*src[len - 1];
+#endif
+
+src[len] = tmp2;
+
+for (int i = len - 2; i >= 2; i -= 2) {
+TXSample val1 = src[i - 0];
+TXSample val2 = src[i - 1] - src[i + 1];
+
+CMUL(src[i + 1], src[i], exp[len - i], exp[i], val1, val2);
+}
+
+s->fn[0](>sub[0], dst, src, sizeof(float));
+
+for (int i = 0; i < len2; i++) {
+TXSample in1 = dst[i];
+TXSample in2 = dst[len - i - 1];
+TXSample c = exp[len + i];
+
+tmp1 = in1 + in2;
+tmp2 = in1 - in2;
+tmp2 *= c;
+#ifdef TX_INT32
+tmp2 = (tmp2 + 0x4000) >> 31;
+#endif
+
+dst[i]= tmp1 + tmp2;
+dst[len - i - 1] = tmp1 - tmp2;
+}
+}
+
static const FFTXCodelet TX_NAME(ff_tx_dctII_def) = {
.name = TX_NAME_STR("dctII"),
.function = TX_NAME(ff_tx_dctII),
@@ -1832,6 +1885,20 @@ static const FFTXCodelet TX_NAME(ff_tx_dctII_def) = {
.prio = FF_TX_PRIO_BASE,
};
+static const FFTXCodelet TX_NAME(ff_tx_dctIII_def) = {
+.name = TX_NAME_STR("dctIII"),
+.function = TX_NAME(ff_tx_dctIII),
+.type = TX_TYPE(DCT),
+.flags = AV_TX_UNALIGNED | AV_TX_INPLACE |
+ FF_TX_OUT_OF_PLACE | FF_TX_INVERSE_ONLY,
+.factors= { 2, TX_FACTOR_ANY },
+.min_len= 2,
+.max_len= TX_LEN_UNLIMITED,
+.init = TX_NAME(ff_tx_dct_init),
+.cpu_flags = FF_TX_CPU_FLAGS_ALL,
+.prio = FF_TX_PRIO_BASE,
+};
+
int TX_TAB(ff_tx_mdct_gen_exp)(AVTXContext *s, int *pre_tab)
{
int off = 0;
@@ -1920,6 +1987,7 @@ const FFTXCodelet * const TX_NAME(ff_tx_codelet_list)[] =
{
_NAME(ff_tx_rdft_r2c_def),
_NAME(ff_tx_rdft_c2r_def),
_NAME(ff_tx_dctII_def),
+_NAME(ff_tx_dctIII_def),
NULL,
};
___
ffmpeg-cvslog mailing list
ffmpeg-cvslog@ffmpeg.org
https://ffmpeg.org/mailman/listinfo/ffmpeg-cvslog
To unsubscribe, visit link above, or email
ffmpeg-cvslog-requ...@ffmpeg.org with subject "unsubscribe".
[FFmpeg-cvslog] lavu/tx: fix last coefficient scaling for R2C transforms
ffmpeg | branch: master | Lynne | Sat Nov 19 14:16:30 2022 +0100| [43d285a40f11e15839b784c85bbbcc7fafd135b5] | committer: Lynne lavu/tx: fix last coefficient scaling for R2C transforms This was a typo. > http://git.videolan.org/gitweb.cgi/ffmpeg.git/?a=commit;h=43d285a40f11e15839b784c85bbbcc7fafd135b5 --- libavutil/tx_template.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/libavutil/tx_template.c b/libavutil/tx_template.c index 59295e06ca..f33bcf85de 100644 --- a/libavutil/tx_template.c +++ b/libavutil/tx_template.c @@ -1622,7 +1622,7 @@ static av_cold int TX_NAME(ff_tx_rdft_init)(AVTXContext *s, m = (inv ? 2*s->scale_d : s->scale_d); *tab++ = RESCALE((inv ? 0.5 : 1.0) * m); -*tab++ = RESCALE(inv ? 0.5*m : 1.0); +*tab++ = RESCALE(inv ? 0.5*m : 1.0*m); *tab++ = RESCALE( m); *tab++ = RESCALE(-m); ___ ffmpeg-cvslog mailing list ffmpeg-cvslog@ffmpeg.org https://ffmpeg.org/mailman/listinfo/ffmpeg-cvslog To unsubscribe, visit link above, or email ffmpeg-cvslog-requ...@ffmpeg.org with subject "unsubscribe".
[FFmpeg-cvslog] lavu/tx: generalize PFA FFTs
ffmpeg | branch: master | Lynne | Sat Nov 19 01:00:36 2022
+0100| [8547123f3b13da5a135994f4253686e67953be55] | committer: Lynne
lavu/tx: generalize PFA FFTs
This commit permits any stacking of FFTs of any size.
> http://git.videolan.org/gitweb.cgi/ffmpeg.git/?a=commit;h=8547123f3b13da5a135994f4253686e67953be55
---
libavutil/tx_priv.h | 3 +-
libavutil/tx_template.c | 217 +++-
2 files changed, 163 insertions(+), 57 deletions(-)
diff --git a/libavutil/tx_priv.h b/libavutil/tx_priv.h
index f11061d051..8a79cf0dd3 100644
--- a/libavutil/tx_priv.h
+++ b/libavutil/tx_priv.h
@@ -232,7 +232,8 @@ struct AVTXContext {
intlen; /* Length of the transform */
intinv; /* If transform is inverse */
int *map; /* Lookup table(s) */
-TXComplex *exp; /* Any non-pre-baked multiplication
factors needed */
+TXComplex *exp; /* Any non-pre-baked multiplication
factors,
+ * or extra temporary buffer */
TXComplex *tmp; /* Temporary buffer, if needed */
AVTXContext *sub; /* Subtransform context(s), if needed
*/
diff --git a/libavutil/tx_template.c b/libavutil/tx_template.c
index 38ab517f66..59295e06ca 100644
--- a/libavutil/tx_template.c
+++ b/libavutil/tx_template.c
@@ -949,74 +949,182 @@ static av_cold int
TX_NAME(ff_tx_fft_pfa_init)(AVTXContext *s,
int len, int inv,
const void *scale)
{
-int ret;
-int sub_len = len / cd->factors[0];
-FFTXCodeletOptions sub_opts = { .invert_lookup = 0 };
+int ret, *tmp, ps = flags & FF_TX_PRESHUFFLE;
+FFTXCodeletOptions sub_opts = { .map_dir = FF_TX_MAP_GATHER };
+size_t extra_tmp_len = 0;
+int len_list[TX_MAX_DECOMPOSITIONS];
-flags &= ~FF_TX_OUT_OF_PLACE; /* We want the subtransform to be */
-flags |= AV_TX_INPLACE; /* in-place */
-flags |= FF_TX_PRESHUFFLE; /* This function handles the permute step */
+if ((ret = ff_tx_decompose_length(len_list, TX_TYPE(FFT), len, inv)) < 0)
+return ret;
-if ((ret = ff_tx_init_subtx(s, TX_TYPE(FFT), flags, _opts,
-sub_len, inv, scale)))
+/* Two iterations to test both orderings. */
+for (int i = 0; i < ret; i++) {
+int len1 = len_list[i];
+int len2 = len / len1;
+
+/* Our ptwo transforms don't support striding the output. */
+if (len2 & (len2 - 1))
+FFSWAP(int, len1, len2);
+
+ff_tx_clear_ctx(s);
+
+/* First transform */
+sub_opts.map_dir = FF_TX_MAP_GATHER;
+flags &= ~AV_TX_INPLACE;
+flags |= FF_TX_OUT_OF_PLACE;
+flags |= FF_TX_PRESHUFFLE; /* This function handles the permute step
*/
+ret = ff_tx_init_subtx(s, TX_TYPE(FFT), flags, _opts,
+ len1, inv, scale);
+
+if (ret == AVERROR(ENOMEM)) {
+return ret;
+} else if (ret < 0) { /* Try again without a preshuffle flag */
+flags &= ~FF_TX_PRESHUFFLE;
+ret = ff_tx_init_subtx(s, TX_TYPE(FFT), flags, _opts,
+ len1, inv, scale);
+if (ret == AVERROR(ENOMEM))
+return ret;
+else if (ret < 0)
+continue;
+}
+
+/* Second transform. */
+sub_opts.map_dir = FF_TX_MAP_SCATTER;
+flags |= FF_TX_PRESHUFFLE;
+retry:
+flags &= ~FF_TX_OUT_OF_PLACE;
+flags |= AV_TX_INPLACE;
+ret = ff_tx_init_subtx(s, TX_TYPE(FFT), flags, _opts,
+ len2, inv, scale);
+
+if (ret == AVERROR(ENOMEM)) {
+return ret;
+} else if (ret < 0) { /* Try again with an out-of-place transform */
+flags |= FF_TX_OUT_OF_PLACE;
+flags &= ~AV_TX_INPLACE;
+ret = ff_tx_init_subtx(s, TX_TYPE(FFT), flags, _opts,
+ len2, inv, scale);
+if (ret == AVERROR(ENOMEM)) {
+return ret;
+} else if (ret < 0) {
+if (flags & FF_TX_PRESHUFFLE) { /* Retry again without a
preshuf flag */
+flags &= ~FF_TX_PRESHUFFLE;
+goto retry;
+} else {
+continue;
+}
+}
+}
+
+/* Success */
+break;
+}
+
+/* If nothing was sucessful, error out */
+if (ret < 0)
return ret;
/* Generate PFA map */
if ((ret = ff_tx_gen_compound_mapping(s, opts, 0,
- cd->factors[0], sub_len)))
+ s->sub[0].len, s->sub[1].len)))
return ret;
if (!(s->tmp =
[FFmpeg-cvslog] lavu/tx: add DCT-II implementation
ffmpeg | branch: master | Lynne | Sat Nov 19 14:20:23 2022
+0100| [504b7bec1a7a46ffbfd0c605fdd984df36dc9871] | committer: Lynne
lavu/tx: add DCT-II implementation
> http://git.videolan.org/gitweb.cgi/ffmpeg.git/?a=commit;h=504b7bec1a7a46ffbfd0c605fdd984df36dc9871
---
libavutil/tx.h | 14 +++
libavutil/tx_template.c | 108
2 files changed, 122 insertions(+)
diff --git a/libavutil/tx.h b/libavutil/tx.h
index 758f634b73..064edbc097 100644
--- a/libavutil/tx.h
+++ b/libavutil/tx.h
@@ -91,6 +91,20 @@ enum AVTXType {
AV_TX_DOUBLE_RDFT = 7,
AV_TX_INT32_RDFT = 8,
+/**
+ * Real to real (DCT) transforms.
+ *
+ * The forward transform is a DCT-II.
+ * The inverse transform is a DCT-III.
+ *
+ * The input array is always overwritten. DCT-III requires that the
+ * input be padded with 2 extra samples. Stride must be set to the
+ * spacing between two samples in bytes.
+ */
+AV_TX_FLOAT_DCT = 9,
+AV_TX_DOUBLE_DCT = 10,
+AV_TX_INT32_DCT = 11,
+
/* Not part of the API, do not use */
AV_TX_NB,
};
diff --git a/libavutil/tx_template.c b/libavutil/tx_template.c
index f33bcf85de..5d73809b58 100644
--- a/libavutil/tx_template.c
+++ b/libavutil/tx_template.c
@@ -1725,6 +1725,113 @@ static const FFTXCodelet TX_NAME(ff_tx_rdft_c2r_def) = {
.prio = FF_TX_PRIO_BASE,
};
+static av_cold int TX_NAME(ff_tx_dct_init)(AVTXContext *s,
+ const FFTXCodelet *cd,
+ uint64_t flags,
+ FFTXCodeletOptions *opts,
+ int len, int inv,
+ const void *scale)
+{
+int ret;
+double freq;
+TXSample *tab;
+SCALE_TYPE rsc = *((SCALE_TYPE *)scale);
+
+
+if ((ret = ff_tx_init_subtx(s, TX_TYPE(RDFT), flags, NULL, len, inv,
)))
+return ret;
+
+s->exp = av_malloc((len/2)*3*sizeof(TXSample));
+if (!s->exp)
+return AVERROR(ENOMEM);
+
+tab = (TXSample *)s->exp;
+
+freq = M_PI/(len*2);
+
+for (int i = 0; i < len; i++)
+tab[i] = RESCALE(cos(i*freq)*(!inv + 1));
+
+for (int i = 0; i < len/2; i++)
+tab[len + i] = RESCALE(cos((len - 2*i - 1)*freq));
+
+return 0;
+}
+
+static void TX_NAME(ff_tx_dctII)(AVTXContext *s, void *_dst,
+ void *_src, ptrdiff_t stride)
+{
+TXSample *dst = _dst;
+TXSample *src = _src;
+const int len = s->len;
+const int len2 = len >> 1;
+const TXSample *exp = (void *)s->exp;
+TXSample next;
+#ifdef TX_INT32
+int64_t tmp1, tmp2;
+#else
+TXSample tmp1, tmp2;
+#endif
+
+for (int i = 0; i < len2; i++) {
+TXSample in1 = src[i];
+TXSample in2 = src[len - i - 1];
+TXSample s= exp[len + i];
+
+#ifdef TX_INT32
+tmp1 = in1 + in2;
+tmp2 = in1 - in2;
+
+tmp1 >>= 1;
+tmp2 *= s;
+
+tmp2 = (tmp2 + 0x4000) >> 31;
+#else
+tmp1 = (in1 + in2)*0.5;
+tmp2 = (in1 - in2)*s;
+#endif
+
+src[i] = tmp1 + tmp2;
+src[len - i - 1] = tmp1 - tmp2;
+}
+
+s->fn[0](>sub[0], dst, src, sizeof(TXComplex));
+
+next = dst[len];
+
+for (int i = len - 2; i > 0; i -= 2) {
+TXSample tmp;
+
+CMUL(tmp, dst[i], exp[len - i], exp[i], dst[i + 0], dst[i + 1]);
+
+dst[i + 1] = next;
+
+next += tmp;
+}
+
+#ifdef TX_INT32
+tmp1 = ((int64_t)exp[0]) * ((int64_t)dst[0]);
+dst[0] = (tmp1 + 0x4000) >> 31;
+#else
+dst[0] = exp[0] * dst[0];
+#endif
+dst[1] = next;
+}
+
+static const FFTXCodelet TX_NAME(ff_tx_dctII_def) = {
+.name = TX_NAME_STR("dctII"),
+.function = TX_NAME(ff_tx_dctII),
+.type = TX_TYPE(DCT),
+.flags = AV_TX_UNALIGNED | AV_TX_INPLACE |
+ FF_TX_OUT_OF_PLACE | FF_TX_FORWARD_ONLY,
+.factors= { 2, TX_FACTOR_ANY },
+.min_len= 2,
+.max_len= TX_LEN_UNLIMITED,
+.init = TX_NAME(ff_tx_dct_init),
+.cpu_flags = FF_TX_CPU_FLAGS_ALL,
+.prio = FF_TX_PRIO_BASE,
+};
+
int TX_TAB(ff_tx_mdct_gen_exp)(AVTXContext *s, int *pre_tab)
{
int off = 0;
@@ -1812,6 +1919,7 @@ const FFTXCodelet * const TX_NAME(ff_tx_codelet_list)[] =
{
_NAME(ff_tx_mdct_inv_full_def),
_NAME(ff_tx_rdft_r2c_def),
_NAME(ff_tx_rdft_c2r_def),
+_NAME(ff_tx_dctII_def),
NULL,
};
___
ffmpeg-cvslog mailing list
ffmpeg-cvslog@ffmpeg.org
https://ffmpeg.org/mailman/listinfo/ffmpeg-cvslog
To unsubscribe, visit link above, or email
ffmpeg-cvslog-requ...@ffmpeg.org with subject "unsubscribe".
[FFmpeg-cvslog] lavu/tx: refactor and separate codelet list and prio code
ffmpeg | branch: master | Lynne | Thu Nov 17 22:14:53 2022
+0100| [1c8d77a2bfa239621b63c4553c6221560b1ee298] | committer: Lynne
lavu/tx: refactor and separate codelet list and prio code
> http://git.videolan.org/gitweb.cgi/ffmpeg.git/?a=commit;h=1c8d77a2bfa239621b63c4553c6221560b1ee298
---
libavutil/tx.c | 125 ++---
1 file changed, 66 insertions(+), 59 deletions(-)
diff --git a/libavutil/tx.c b/libavutil/tx.c
index 319392788f..ff81d235ba 100644
--- a/libavutil/tx.c
+++ b/libavutil/tx.c
@@ -300,6 +300,67 @@ static const FFTXCodelet * const ff_tx_null_list[] = {
NULL,
};
+/* Array of all compiled codelet lists. Order is irrelevant. */
+static const FFTXCodelet * const * const codelet_list[] = {
+ff_tx_codelet_list_float_c,
+ff_tx_codelet_list_double_c,
+ff_tx_codelet_list_int32_c,
+ff_tx_null_list,
+#if HAVE_X86ASM
+ff_tx_codelet_list_float_x86,
+#endif
+#if ARCH_AARCH64
+ff_tx_codelet_list_float_aarch64,
+#endif
+};
+static const int codelet_list_num = FF_ARRAY_ELEMS(codelet_list);
+
+static const int cpu_slow_mask = AV_CPU_FLAG_SSE2SLOW | AV_CPU_FLAG_SSE3SLOW |
+ AV_CPU_FLAG_ATOM | AV_CPU_FLAG_SSSE3SLOW |
+ AV_CPU_FLAG_AVXSLOW |
AV_CPU_FLAG_SLOW_GATHER;
+
+static const int cpu_slow_penalties[][2] = {
+{ AV_CPU_FLAG_SSE2SLOW,1 + 64 },
+{ AV_CPU_FLAG_SSE3SLOW,1 + 64 },
+{ AV_CPU_FLAG_SSSE3SLOW, 1 + 64 },
+{ AV_CPU_FLAG_ATOM,1 + 128 },
+{ AV_CPU_FLAG_AVXSLOW, 1 + 128 },
+{ AV_CPU_FLAG_SLOW_GATHER, 1 + 32 },
+};
+
+static int get_codelet_prio(const FFTXCodelet *cd, int cpu_flags, int len)
+{
+int prio = cd->prio;
+int max_factor = 0;
+
+/* If the CPU has a SLOW flag, and the instruction is also flagged
+ * as being slow for such, reduce its priority */
+for (int i = 0; i < FF_ARRAY_ELEMS(cpu_slow_penalties); i++) {
+if ((cpu_flags & cd->cpu_flags) & cpu_slow_penalties[i][0])
+prio -= cpu_slow_penalties[i][1];
+}
+
+/* Prioritize aligned-only codelets */
+if ((cd->flags & FF_TX_ALIGNED) && !(cd->flags & AV_TX_UNALIGNED))
+prio += 64;
+
+/* Codelets for specific lengths are generally faster */
+if ((len == cd->min_len) && (len == cd->max_len))
+prio += 64;
+
+/* Forward-only or inverse-only transforms are generally better */
+if ((cd->flags & (FF_TX_FORWARD_ONLY | FF_TX_INVERSE_ONLY)))
+prio += 64;
+
+/* Larger factors are generally better */
+for (int i = 0; i < TX_MAX_SUB; i++)
+max_factor = FFMAX(cd->factors[i], max_factor);
+if (max_factor)
+prio += 16*max_factor;
+
+return prio;
+}
+
#if !CONFIG_SMALL
static void print_flags(AVBPrint *bp, uint64_t f)
{
@@ -465,41 +526,15 @@ av_cold int ff_tx_init_subtx(AVTXContext *s, enum
AVTXType type,
AVTXContext *sub = NULL;
TXCodeletMatch *cd_tmp, *cd_matches = NULL;
unsigned int cd_matches_size = 0;
+int codelet_list_idx = codelet_list_num;
int nb_cd_matches = 0;
#if !CONFIG_SMALL
AVBPrint bp = { 0 };
#endif
-/* Array of all compiled codelet lists. Order is irrelevant. */
-const FFTXCodelet * const * const codelet_list[] = {
-ff_tx_codelet_list_float_c,
-ff_tx_codelet_list_double_c,
-ff_tx_codelet_list_int32_c,
-ff_tx_null_list,
-#if HAVE_X86ASM
-ff_tx_codelet_list_float_x86,
-#endif
-#if ARCH_AARCH64
-ff_tx_codelet_list_float_aarch64,
-#endif
-};
-int codelet_list_num = FF_ARRAY_ELEMS(codelet_list);
-
/* We still accept functions marked with SLOW, even if the CPU is
* marked with the same flag, but we give them lower priority. */
const int cpu_flags = av_get_cpu_flags();
-const int slow_mask = AV_CPU_FLAG_SSE2SLOW | AV_CPU_FLAG_SSE3SLOW |
- AV_CPU_FLAG_ATOM | AV_CPU_FLAG_SSSE3SLOW |
- AV_CPU_FLAG_AVXSLOW | AV_CPU_FLAG_SLOW_GATHER;
-
-static const int slow_penalties[][2] = {
-{ AV_CPU_FLAG_SSE2SLOW,1 + 64 },
-{ AV_CPU_FLAG_SSE3SLOW,1 + 64 },
-{ AV_CPU_FLAG_SSSE3SLOW, 1 + 64 },
-{ AV_CPU_FLAG_ATOM,1 + 128 },
-{ AV_CPU_FLAG_AVXSLOW, 1 + 128 },
-{ AV_CPU_FLAG_SLOW_GATHER, 1 + 32 },
-};
/* Flags the transform wants */
uint64_t req_flags = flags;
@@ -519,13 +554,11 @@ av_cold int ff_tx_init_subtx(AVTXContext *s, enum
AVTXType type,
/* Loop through all codelets in all codelet lists to find matches
* to the requirements */
-while (codelet_list_num--) {
-const FFTXCodelet * const * list = codelet_list[codelet_list_num];
+while (codelet_list_idx--) {
+const FFTXCodelet * const * list = codelet_list[codelet_list_idx];
const FFTXCodelet *cd = NULL;
while ((cd = *list++)) {
-int max_factor = 0;
-
/*
[FFmpeg-cvslog] lavu/tx: add length decomposition function
ffmpeg | branch: master | Lynne | Sat Nov 19 00:58:13 2022
+0100| [7f019e77586ab62d37856d09071fcdeef880bcd9] | committer: Lynne
lavu/tx: add length decomposition function
Rather than using a list of lengths supported, this goes a step beyond
and uses all registered codelets to come up with a good decomposition.
> http://git.videolan.org/gitweb.cgi/ffmpeg.git/?a=commit;h=7f019e77586ab62d37856d09071fcdeef880bcd9
---
libavutil/tx.c | 143
libavutil/tx_priv.h | 8 +++
2 files changed, 151 insertions(+)
diff --git a/libavutil/tx.c b/libavutil/tx.c
index 8027e983ba..c692239656 100644
--- a/libavutil/tx.c
+++ b/libavutil/tx.c
@@ -17,6 +17,7 @@
*/
#include "avassert.h"
+#include "intmath.h"
#include "cpu.h"
#include "qsort.h"
#include "bprint.h"
@@ -395,6 +396,148 @@ static int get_codelet_prio(const FFTXCodelet *cd, int
cpu_flags, int len)
return prio;
}
+typedef struct FFTXLenDecomp {
+int len;
+int len2;
+int prio;
+const FFTXCodelet *cd;
+} FFTXLenDecomp;
+
+static int cmp_decomp(FFTXLenDecomp *a, FFTXLenDecomp *b)
+{
+return FFDIFFSIGN(b->prio, a->prio);
+}
+
+int ff_tx_decompose_length(int dst[TX_MAX_DECOMPOSITIONS], enum AVTXType type,
+ int len, int inv)
+{
+int nb_decomp = 0;
+FFTXLenDecomp ld[TX_MAX_DECOMPOSITIONS];
+int codelet_list_idx = codelet_list_num;
+
+const int cpu_flags = av_get_cpu_flags();
+
+/* Loop through all codelets in all codelet lists to find matches
+ * to the requirements */
+while (codelet_list_idx--) {
+const FFTXCodelet * const * list = codelet_list[codelet_list_idx];
+const FFTXCodelet *cd = NULL;
+
+while ((cd = *list++)) {
+int fl = len;
+int skip = 0, prio;
+int factors_product = 1, factors_mod = 0;
+
+if (nb_decomp >= TX_MAX_DECOMPOSITIONS)
+goto sort;
+
+/* Check if the type matches */
+if (cd->type != TX_TYPE_ANY && type != cd->type)
+continue;
+
+/* Check direction for non-orthogonal codelets */
+if (((cd->flags & FF_TX_FORWARD_ONLY) && inv) ||
+((cd->flags & (FF_TX_INVERSE_ONLY | AV_TX_FULL_IMDCT)) &&
!inv))
+continue;
+
+/* Check if the CPU supports the required ISA */
+if (cd->cpu_flags != FF_TX_CPU_FLAGS_ALL &&
+!(cpu_flags & (cd->cpu_flags & ~cpu_slow_mask)))
+continue;
+
+for (int i = 0; i < TX_MAX_FACTORS; i++) {
+if (!cd->factors[i] || (fl == 1))
+break;
+
+if (cd->factors[i] == TX_FACTOR_ANY) {
+factors_mod++;
+factors_product *= fl;
+} else if (!(fl % cd->factors[i])) {
+factors_mod++;
+if (cd->factors[i] == 2) {
+int b = ff_ctz(fl);
+fl >>= b;
+factors_product <<= b;
+} else {
+do {
+fl /= cd->factors[i];
+factors_product *= cd->factors[i];
+} while (!(fl % cd->factors[i]));
+}
+}
+}
+
+/* Disqualify if factor requirements are not satisfied or if
trivial */
+if ((factors_mod < cd->nb_factors) || (len == factors_product))
+continue;
+
+if (av_gcd(factors_product, fl) != 1)
+continue;
+
+/* Check if length is supported and factorization was successful */
+if ((factors_product < cd->min_len) ||
+(cd->max_len != TX_LEN_UNLIMITED && (factors_product >
cd->max_len)))
+continue;
+
+prio = get_codelet_prio(cd, cpu_flags, factors_product) *
factors_product;
+
+/* Check for duplicates */
+for (int i = 0; i < nb_decomp; i++) {
+if (factors_product == ld[i].len) {
+/* Update priority if new one is higher */
+if (prio > ld[i].prio)
+ld[i].prio = prio;
+skip = 1;
+break;
+}
+}
+
+/* Add decomposition if unique */
+if (!skip) {
+ld[nb_decomp].cd = cd;
+ld[nb_decomp].len = factors_product;
+ld[nb_decomp].len2 = fl;
+ld[nb_decomp].prio = prio;
+nb_decomp++;
+}
+}
+}
+
+if (!nb_decomp)
+return AVERROR(EINVAL);
+
+sort:
+AV_QSORT(ld, nb_decomp, FFTXLenDecomp, cmp_decomp);
+
+for (int i = 0; i < nb_decomp; i++) {
+if (ld[i].cd->nb_factors > 1)
+dst[i] = ld[i].len2;
+else
+dst[i] = ld[i].len;
+}
+
+
[FFmpeg-cvslog] lavu/tx: refactor to explicitly track and convert lookup table order
ffmpeg | branch: master | Lynne | Sat Nov 19 00:47:45 2022
+0100| [87bae6b0189d5cb71b836890078f96a4d1abd277] | committer: Lynne
lavu/tx: refactor to explicitly track and convert lookup table order
Necessary for generalizing PFAs.
> http://git.videolan.org/gitweb.cgi/ffmpeg.git/?a=commit;h=87bae6b0189d5cb71b836890078f96a4d1abd277
---
libavutil/aarch64/tx_float_init.c | 5 +-
libavutil/tx.c| 109 +-
libavutil/tx_priv.h | 52 --
libavutil/tx_template.c | 49 -
libavutil/x86/tx_float_init.c | 46
5 files changed, 181 insertions(+), 80 deletions(-)
diff --git a/libavutil/aarch64/tx_float_init.c
b/libavutil/aarch64/tx_float_init.c
index e7b73b4bf9..8300472c4c 100644
--- a/libavutil/aarch64/tx_float_init.c
+++ b/libavutil/aarch64/tx_float_init.c
@@ -37,12 +37,11 @@ static av_cold int neon_init(AVTXContext *s, const
FFTXCodelet *cd,
uint64_t flags, FFTXCodeletOptions *opts,
int len, int inv, const void *scale)
{
-const int inv_lookup = opts ? opts->invert_lookup : 1;
ff_tx_init_tabs_float(len);
if (cd->max_len == 2)
-return ff_tx_gen_ptwo_revtab(s, inv_lookup);
+return ff_tx_gen_ptwo_revtab(s, opts);
else
-return ff_tx_gen_split_radix_parity_revtab(s, len, inv, inv_lookup, 8,
0);
+return ff_tx_gen_split_radix_parity_revtab(s, len, inv, opts, 8, 0);
}
const FFTXCodelet * const ff_tx_codelet_list_float_aarch64[] = {
diff --git a/libavutil/tx.c b/libavutil/tx.c
index ff81d235ba..8027e983ba 100644
--- a/libavutil/tx.c
+++ b/libavutil/tx.c
@@ -39,11 +39,41 @@ static av_always_inline int mulinv(int n, int m)
return 0;
}
+int ff_tx_gen_pfa_input_map(AVTXContext *s, FFTXCodeletOptions *opts,
+int d1, int d2)
+{
+const int sl = d1*d2;
+
+s->map = av_malloc(s->len*sizeof(*s->map));
+if (!s->map)
+return AVERROR(ENOMEM);
+
+for (int k = 0; k < s->len; k += sl) {
+if (s->inv || (opts && opts->map_dir == FF_TX_MAP_SCATTER)) {
+for (int m = 0; m < d2; m++)
+for (int n = 0; n < d1; n++)
+s->map[k + ((m*d1 + n*d2) % (sl))] = m*d1 + n;
+} else {
+for (int m = 0; m < d2; m++)
+for (int n = 0; n < d1; n++)
+s->map[k + m*d1 + n] = (m*d1 + n*d2) % (sl);
+}
+
+if (s->inv)
+for (int w = 1; w <= ((sl) >> 1); w++)
+FFSWAP(int, s->map[k + w], s->map[k + sl - w]);
+}
+
+s->map_dir = opts ? opts->map_dir : FF_TX_MAP_GATHER;
+
+return 0;
+}
+
/* Guaranteed to work for any n, m where gcd(n, m) == 1 */
-int ff_tx_gen_compound_mapping(AVTXContext *s, int n, int m)
+int ff_tx_gen_compound_mapping(AVTXContext *s, FFTXCodeletOptions *opts,
+ int inv, int n, int m)
{
int *in_map, *out_map;
-const int inv = s->inv;
const int len = n*m;/* Will not be equal to s->len for MDCTs */
int m_inv, n_inv;
@@ -61,14 +91,22 @@ int ff_tx_gen_compound_mapping(AVTXContext *s, int n, int m)
out_map = s->map + len;
/* Ruritanian map for input, CRT map for output, can be swapped */
-for (int j = 0; j < m; j++) {
-for (int i = 0; i < n; i++) {
-in_map[j*n + i] = (i*m + j*n) % len;
-out_map[(i*m*m_inv + j*n*n_inv) % len] = i*m + j;
+if (opts && opts->map_dir == FF_TX_MAP_SCATTER) {
+for (int j = 0; j < m; j++) {
+for (int i = 0; i < n; i++) {
+in_map[(i*m + j*n) % len] = j*n + i;
+out_map[(i*m*m_inv + j*n*n_inv) % len] = i*m + j;
+}
+}
+} else {
+for (int j = 0; j < m; j++) {
+for (int i = 0; i < n; i++) {
+in_map[j*n + i] = (i*m + j*n) % len;
+out_map[(i*m*m_inv + j*n*n_inv) % len] = i*m + j;
+}
}
}
-/* Change transform direction by reversing all ACs */
if (inv) {
for (int i = 0; i < m; i++) {
int *in = _map[i*n + 1]; /* Skip the DC */
@@ -77,17 +115,7 @@ int ff_tx_gen_compound_mapping(AVTXContext *s, int n, int m)
}
}
-/* Our 15-point transform is also a compound one, so embed its input map */
-if (n == 15) {
-for (int k = 0; k < m; k++) {
-int tmp[15];
-memcpy(tmp, _map[k*15], 15*sizeof(*tmp));
-for (int i = 0; i < 5; i++) {
-for (int j = 0; j < 3; j++)
-in_map[k*15 + i*3 + j] = tmp[(i*3 + j*5) % 15];
-}
-}
-}
+s->map_dir = opts ? opts->map_dir : FF_TX_MAP_GATHER;
return 0;
}
@@ -103,21 +131,23 @@ static inline int split_radix_permutation(int i, int len,
int inv)
return split_radix_permutation(i, len, inv) * 4 + 1 - 2*(!(i & len) ^ inv);
}
-int
[FFmpeg-cvslog] lavu/tx: add ff_tx_clear_ctx()
ffmpeg | branch: master | Lynne | Sat Oct 1 12:20:10 2022
+0200| [dd77e61182865e396195a19b1e6ec697717cef56] | committer: Lynne
lavu/tx: add ff_tx_clear_ctx()
This function allows implementations to clean up a context after
successfully initializing subcontexts.
> http://git.videolan.org/gitweb.cgi/ffmpeg.git/?a=commit;h=dd77e61182865e396195a19b1e6ec697717cef56
---
libavutil/tx.c | 27 +++
libavutil/tx_priv.h | 3 +++
2 files changed, 22 insertions(+), 8 deletions(-)
diff --git a/libavutil/tx.c b/libavutil/tx.c
index 246a7aa980..13fb54f916 100644
--- a/libavutil/tx.c
+++ b/libavutil/tx.c
@@ -225,24 +225,35 @@ int ff_tx_gen_split_radix_parity_revtab(AVTXContext *s,
int len, int inv,
return 0;
}
-static void reset_ctx(AVTXContext *s)
+static void reset_ctx(AVTXContext *s, int free_sub)
{
if (!s)
return;
if (s->sub)
-for (int i = 0; i < s->nb_sub; i++)
-reset_ctx(>sub[i]);
+for (int i = 0; i < TX_MAX_SUB; i++)
+reset_ctx(>sub[i], free_sub + 1);
-if (s->cd_self->uninit)
+if (s->cd_self && s->cd_self->uninit)
s->cd_self->uninit(s);
-av_freep(>sub);
+if (free_sub)
+av_freep(>sub);
+
av_freep(>map);
av_freep(>exp);
av_freep(>tmp);
-memset(s, 0, sizeof(*s));
+/* Nothing else needs to be reset, it gets overwritten if another
+ * ff_tx_init_subtx() call is made. */
+s->nb_sub = 0;
+s->opaque = NULL;
+memset(s->fn, 0, sizeof(*s->fn));
+}
+
+void ff_tx_clear_ctx(AVTXContext *s)
+{
+reset_ctx(s, 0);
}
av_cold void av_tx_uninit(AVTXContext **ctx)
@@ -250,7 +261,7 @@ av_cold void av_tx_uninit(AVTXContext **ctx)
if (!(*ctx))
return;
-reset_ctx(*ctx);
+reset_ctx(*ctx, 1);
av_freep(ctx);
}
@@ -635,7 +646,7 @@ av_cold int ff_tx_init_subtx(AVTXContext *s, enum AVTXType
type,
s->fn[s->nb_sub] = NULL;
s->cd[s->nb_sub] = NULL;
-reset_ctx(sctx);
+reset_ctx(sctx, 0);
if (ret == AVERROR(ENOMEM))
break;
}
diff --git a/libavutil/tx_priv.h b/libavutil/tx_priv.h
index 56e78631ba..d9e38ba19b 100644
--- a/libavutil/tx_priv.h
+++ b/libavutil/tx_priv.h
@@ -240,6 +240,9 @@ int ff_tx_init_subtx(AVTXContext *s, enum AVTXType type,
uint64_t flags, FFTXCodeletOptions *opts,
int len, int inv, const void *scale);
+/* Clear the context by freeing all tables, maps and subtransforms. */
+void ff_tx_clear_ctx(AVTXContext *s);
+
/*
* Generates the PFA permutation table into AVTXContext->pfatab. The end table
* is appended to the start table.
___
ffmpeg-cvslog mailing list
ffmpeg-cvslog@ffmpeg.org
https://ffmpeg.org/mailman/listinfo/ffmpeg-cvslog
To unsubscribe, visit link above, or email
ffmpeg-cvslog-requ...@ffmpeg.org with subject "unsubscribe".
[FFmpeg-cvslog] lavu/tx: improve transform tree logging
ffmpeg | branch: master | Lynne | Sat Oct 1 12:35:14 2022
+0200| [958b3760b54245b934054f8aa72a608bdb2a48b8] | committer: Lynne
lavu/tx: improve transform tree logging
Now prints the actual codelet size used, as well as the number of
allowed factors.
> http://git.videolan.org/gitweb.cgi/ffmpeg.git/?a=commit;h=958b3760b54245b934054f8aa72a608bdb2a48b8
---
libavutil/tx.c | 56 +++-
1 file changed, 35 insertions(+), 21 deletions(-)
diff --git a/libavutil/tx.c b/libavutil/tx.c
index a1173f6137..319392788f 100644
--- a/libavutil/tx.c
+++ b/libavutil/tx.c
@@ -343,7 +343,7 @@ static void print_type(AVBPrint *bp, enum AVTXType type)
"unknown");
}
-static void print_cd_info(const FFTXCodelet *cd, int prio, int print_prio)
+static void print_cd_info(const FFTXCodelet *cd, int prio, int len, int
print_prio)
{
AVBPrint bp = { 0 };
av_bprint_init(, 0, AV_BPRINT_SIZE_AUTOMATIC);
@@ -353,27 +353,41 @@ static void print_cd_info(const FFTXCodelet *cd, int
prio, int print_prio)
print_type(, cd->type);
av_bprintf(, ", len: ");
-if (cd->min_len != cd->max_len)
-av_bprintf(, "[%i, ", cd->min_len);
-
-if (cd->max_len == TX_LEN_UNLIMITED)
-av_bprintf(, "∞");
-else
-av_bprintf(, "%i", cd->max_len);
-
-av_bprintf(, "%s, factors: [", cd->min_len != cd->max_len ? "]" : "");
-for (int i = 0; i < TX_MAX_SUB; i++) {
-if (i && cd->factors[i])
-av_bprintf(, ", ");
-if (cd->factors[i] == TX_FACTOR_ANY)
-av_bprintf(, "any");
-else if (cd->factors[i])
-av_bprintf(, "%i", cd->factors[i]);
+if (!len) {
+if (cd->min_len != cd->max_len)
+av_bprintf(, "[%i, ", cd->min_len);
+
+if (cd->max_len == TX_LEN_UNLIMITED)
+av_bprintf(, "∞");
else
-break;
+av_bprintf(, "%i", cd->max_len);
+} else {
+av_bprintf(, "%i", len);
}
-av_bprintf(, "], ");
+if (cd->factors[1]) {
+av_bprintf(, "%s, factors", !len && cd->min_len != cd->max_len ?
"]" : "");
+if (!cd->nb_factors)
+av_bprintf(, ": [");
+else
+av_bprintf(, "[%i]: [", cd->nb_factors);
+
+for (int i = 0; i < TX_MAX_FACTORS; i++) {
+if (i && cd->factors[i])
+av_bprintf(, ", ");
+if (cd->factors[i] == TX_FACTOR_ANY)
+av_bprintf(, "any");
+else if (cd->factors[i])
+av_bprintf(, "%i", cd->factors[i]);
+else
+break;
+}
+
+av_bprintf(, "], ");
+} else {
+av_bprintf(, "%s, factor: %i, ",
+ !len && cd->min_len != cd->max_len ? "]" : "",
cd->factors[0]);
+}
print_flags(, cd->flags);
if (print_prio)
@@ -389,7 +403,7 @@ static void print_tx_structure(AVTXContext *s, int depth)
for (int i = 0; i <= depth; i++)
av_log(NULL, AV_LOG_DEBUG, "");
-print_cd_info(cd, cd->prio, 0);
+print_cd_info(cd, cd->prio, s->len, 0);
for (int i = 0; i < s->nb_sub; i++)
print_tx_structure(>sub[i], depth + 1);
@@ -604,7 +618,7 @@ av_cold int ff_tx_init_subtx(AVTXContext *s, enum AVTXType
type,
for (int i = 0; i < nb_cd_matches; i++) {
av_log(NULL, AV_LOG_DEBUG, "%i: ", i + 1);
-print_cd_info(cd_matches[i].cd, cd_matches[i].prio, 1);
+print_cd_info(cd_matches[i].cd, cd_matches[i].prio, 0, 1);
}
#endif
___
ffmpeg-cvslog mailing list
ffmpeg-cvslog@ffmpeg.org
https://ffmpeg.org/mailman/listinfo/ffmpeg-cvslog
To unsubscribe, visit link above, or email
ffmpeg-cvslog-requ...@ffmpeg.org with subject "unsubscribe".
[FFmpeg-cvslog] lavu/tx: allow codelets to specify a minimum number of matching factors
ffmpeg | branch: master | Lynne | Sat Oct 1 12:21:28 2022
+0200| [6ddd10c3e2d63d1ad1ea1034b0e3862107a27063] | committer: Lynne
lavu/tx: allow codelets to specify a minimum number of matching factors
> http://git.videolan.org/gitweb.cgi/ffmpeg.git/?a=commit;h=6ddd10c3e2d63d1ad1ea1034b0e3862107a27063
---
libavutil/tx.c | 30 +-
libavutil/tx_priv.h | 11 +--
libavutil/tx_template.c | 18 ++
3 files changed, 40 insertions(+), 19 deletions(-)
diff --git a/libavutil/tx.c b/libavutil/tx.c
index 13fb54f916..a1173f6137 100644
--- a/libavutil/tx.c
+++ b/libavutil/tx.c
@@ -409,42 +409,38 @@ static int cmp_matches(TXCodeletMatch *a, TXCodeletMatch
*b)
/* We want all factors to completely cover the length */
static inline int check_cd_factors(const FFTXCodelet *cd, int len)
{
-int all_flag = 0;
+int matches = 0, any_flag = 0;
-for (int i = 0; i < TX_MAX_SUB; i++) {
+for (int i = 0; i < TX_MAX_FACTORS; i++) {
int factor = cd->factors[i];
-/* Conditions satisfied */
-if (len == 1)
-return 1;
-
-/* No more factors */
-if (!factor) {
-break;
-} else if (factor == TX_FACTOR_ANY) {
-all_flag = 1;
+if (factor == TX_FACTOR_ANY) {
+any_flag = 1;
+matches++;
continue;
-}
-
-if (factor == 2) { /* Fast path */
+} else if (len <= 1 || !factor) {
+break;
+} else if (factor == 2) { /* Fast path */
int bits_2 = ff_ctz(len);
if (!bits_2)
-return 0; /* Factor not supported */
+continue; /* Factor not supported */
len >>= bits_2;
+matches++;
} else {
int res = len % factor;
if (res)
-return 0; /* Factor not supported */
+continue; /* Factor not supported */
while (!res) {
len /= factor;
res = len % factor;
}
+matches++;
}
}
-return all_flag || (len == 1);
+return (cd->nb_factors <= matches) && (any_flag || len == 1);
}
av_cold int ff_tx_init_subtx(AVTXContext *s, enum AVTXType type,
diff --git a/libavutil/tx_priv.h b/libavutil/tx_priv.h
index d9e38ba19b..80d045f6af 100644
--- a/libavutil/tx_priv.h
+++ b/libavutil/tx_priv.h
@@ -71,7 +71,8 @@ typedef void TXComplex;
.function = TX_FN_NAME(fn, suffix),
\
.type = TX_TYPE(tx_type),
\
.flags = FF_TX_ALIGNED | FF_TX_OUT_OF_PLACE | cd_flags,
\
-.factors= { f1, f2 },
\
+.factors= { (f1), (f2) },
\
+.nb_factors = !!(f1) + !!(f2),
\
.min_len= len_min,
\
.max_len= len_max,
\
.init = init_fn,
\
@@ -163,6 +164,9 @@ typedef struct FFTXCodeletOptions {
invert the lookup direction for the map
generated */
} FFTXCodeletOptions;
+/* Maximum number of factors a codelet may have. Arbitrary. */
+#define TX_MAX_FACTORS 16
+
/* Maximum amount of subtransform functions, subtransforms and factors.
Arbitrary. */
#define TX_MAX_SUB 4
@@ -175,13 +179,16 @@ typedef struct FFTXCodelet {
uint64_t flags; /* A combination of AVTXFlags and codelet
* flags that describe its properties. */
-int factors[TX_MAX_SUB]; /* Length factors */
+int factors[TX_MAX_FACTORS]; /* Length factors. MUST be coprime. */
#define TX_FACTOR_ANY -1 /* When used alone, signals that the codelet
* supports all factors. Otherwise, if other
* factors are present, it signals that
whatever
* remains will be supported, as long as the
* other factors are a component of the
length */
+int nb_factors; /* Minimum number of factors that have to
+ * be a modulo of the length. Must not be 0.
*/
+
int min_len; /* Minimum length of transform, must be >= 1
*/
int max_len; /* Maximum length of transform */
#define TX_LEN_UNLIMITED -1 /* Special length value to permit all
lengths */
diff --git a/libavutil/tx_template.c b/libavutil/tx_template.c
index 228209521b..c157719d73 100644
--- a/libavutil/tx_template.c
+++ b/libavutil/tx_template.c
@@ -518,6 +518,7 @@ static const FFTXCodelet TX_NAME(ff_tx_fft##n##_ns_def) =
[FFmpeg-cvslog] x86/tx_float_init: properly specify the supported factors of 15xM FFTs
ffmpeg | branch: master | Lynne | Sat Oct 1 12:44:06 2022
+0200| [92100eee5b588e60b8fe3e14d35766e8bab2] | committer: Lynne
x86/tx_float_init: properly specify the supported factors of 15xM FFTs
Only powers of two are currently supported.
> http://git.videolan.org/gitweb.cgi/ffmpeg.git/?a=commit;h=92100eee5b588e60b8fe3e14d35766e8bab2
---
libavutil/x86/tx_float_init.c | 6 +++---
1 file changed, 3 insertions(+), 3 deletions(-)
diff --git a/libavutil/x86/tx_float_init.c b/libavutil/x86/tx_float_init.c
index 523b43a689..97ee44defa 100644
--- a/libavutil/x86/tx_float_init.c
+++ b/libavutil/x86/tx_float_init.c
@@ -290,11 +290,11 @@ const FFTXCodelet * const ff_tx_codelet_list_float_x86[]
= {
TX_DEF(fft_sr_ns, FFT, 64, 131072, 2, 0, 384, b8_i2, avx2, AVX2,
AV_TX_INPLACE | FF_TX_PRESHUFFLE,
AV_CPU_FLAG_AVXSLOW | AV_CPU_FLAG_SLOW_GATHER),
-TX_DEF(fft_pfa_15xM, FFT, 60, TX_LEN_UNLIMITED, 15, TX_FACTOR_ANY, 320,
fft_pfa_init, avx2, AVX2,
+TX_DEF(fft_pfa_15xM, FFT, 60, TX_LEN_UNLIMITED, 15, 2, 320, fft_pfa_init,
avx2, AVX2,
AV_TX_INPLACE, AV_CPU_FLAG_AVXSLOW | AV_CPU_FLAG_SLOW_GATHER),
-TX_DEF(fft_pfa_15xM_asm, FFT, 60, TX_LEN_UNLIMITED, 15, TX_FACTOR_ANY,
384, fft_pfa_init, avx2, AVX2,
+TX_DEF(fft_pfa_15xM_asm, FFT, 60, TX_LEN_UNLIMITED, 15, 2, 384,
fft_pfa_init, avx2, AVX2,
AV_TX_INPLACE | FF_TX_PRESHUFFLE | FF_TX_ASM_CALL,
AV_CPU_FLAG_AVXSLOW | AV_CPU_FLAG_SLOW_GATHER),
-TX_DEF(fft_pfa_15xM_ns, FFT, 60, TX_LEN_UNLIMITED, 15, TX_FACTOR_ANY, 384,
fft_pfa_init, avx2, AVX2,
+TX_DEF(fft_pfa_15xM_ns, FFT, 60, TX_LEN_UNLIMITED, 15, 2, 384,
fft_pfa_init, avx2, AVX2,
AV_TX_INPLACE | FF_TX_PRESHUFFLE, AV_CPU_FLAG_AVXSLOW |
AV_CPU_FLAG_SLOW_GATHER),
TX_DEF(mdct_inv, MDCT, 16, TX_LEN_UNLIMITED, 2, TX_FACTOR_ANY, 384,
m_inv_init, avx2, AVX2,
___
ffmpeg-cvslog mailing list
ffmpeg-cvslog@ffmpeg.org
https://ffmpeg.org/mailman/listinfo/ffmpeg-cvslog
To unsubscribe, visit link above, or email
ffmpeg-cvslog-requ...@ffmpeg.org with subject "unsubscribe".
[FFmpeg-cvslog] x86/tx_float: implement striding in fft_15xM
ffmpeg | branch: master | Lynne | Fri Sep 30 11:00:44 2022 +0200| [fab97faf02118240c28695c1a6401e7bcc4b21a8] | committer: Lynne x86/tx_float: implement striding in fft_15xM > http://git.videolan.org/gitweb.cgi/ffmpeg.git/?a=commit;h=fab97faf02118240c28695c1a6401e7bcc4b21a8 --- libavutil/x86/tx_float.asm | 45 + 1 file changed, 29 insertions(+), 16 deletions(-) diff --git a/libavutil/x86/tx_float.asm b/libavutil/x86/tx_float.asm index 6f83555ce5..2ad84c2885 100644 --- a/libavutil/x86/tx_float.asm +++ b/libavutil/x86/tx_float.asm @@ -24,7 +24,7 @@ ; Intra-asm call convention: ; 320 bytes of stack available ; 14 GPRs available (last 4 must not be clobbered) -; Additionally, don't clobber ctx, in, out, len, lut +; Additionally, don't clobber ctx, in, out, stride, len, lut ; All vector regs available ; TODO: @@ -863,7 +863,7 @@ FFT4_FN inv, 1, 1 %macro FFT8_SSE_FN 1 INIT_XMM sse3 %if %1 -cglobal fft8_asm_float, 0, 0, 0, ctx, out, in, tmp +cglobal fft8_asm_float, 0, 0, 0, ctx, out, in, stride, tmp movaps m0, [inq + 0*mmsize] movaps m1, [inq + 1*mmsize] movaps m2, [inq + 2*mmsize] @@ -896,7 +896,7 @@ cglobal fft8_float, 4, 4, 6, ctx, out, in, tmp %endif %if %1 -cglobal fft8_ns_float, 4, 4, 6, ctx, out, in, tmp +cglobal fft8_ns_float, 4, 5, 6, ctx, out, in, stride, tmp call mangle(ff_tx_fft8_asm_float_sse3) RET %endif @@ -908,7 +908,7 @@ FFT8_SSE_FN 1 %macro FFT8_AVX_FN 1 INIT_YMM avx %if %1 -cglobal fft8_asm_float, 0, 0, 0, ctx, out, in, tmp +cglobal fft8_asm_float, 0, 0, 0, ctx, out, in, stride, tmp movaps m0, [inq + 0*mmsize] movaps m1, [inq + 1*mmsize] %else @@ -936,7 +936,7 @@ cglobal fft8_float, 4, 4, 4, ctx, out, in, tmp %endif %if %1 -cglobal fft8_ns_float, 4, 4, 4, ctx, out, in, tmp +cglobal fft8_ns_float, 4, 5, 4, ctx, out, in, stride, tmp call mangle(ff_tx_fft8_asm_float_avx) RET %endif @@ -948,7 +948,7 @@ FFT8_AVX_FN 1 %macro FFT16_FN 2 INIT_YMM %1 %if %2 -cglobal fft16_asm_float, 0, 0, 0, ctx, out, in, tmp +cglobal fft16_asm_float, 0, 0, 0, ctx, out, in, stride, tmp movaps m0, [inq + 0*mmsize] movaps m1, [inq + 1*mmsize] movaps m2, [inq + 2*mmsize] @@ -985,7 +985,7 @@ cglobal fft16_float, 4, 4, 8, ctx, out, in, tmp %endif %if %2 -cglobal fft16_ns_float, 4, 4, 8, ctx, out, in, tmp +cglobal fft16_ns_float, 4, 5, 8, ctx, out, in, stride, tmp call mangle(ff_tx_fft16_asm_float_ %+ %1) RET %endif @@ -999,7 +999,7 @@ FFT16_FN fma3, 1 %macro FFT32_FN 2 INIT_YMM %1 %if %2 -cglobal fft32_asm_float, 0, 0, 0, ctx, out, in, tmp +cglobal fft32_asm_float, 0, 0, 0, ctx, out, in, stride, tmp movaps m4, [inq + 4*mmsize] movaps m5, [inq + 5*mmsize] movaps m6, [inq + 6*mmsize] @@ -1069,7 +1069,7 @@ cglobal fft32_float, 4, 4, 16, ctx, out, in, tmp %endif %if %2 -cglobal fft32_ns_float, 4, 4, 16, ctx, out, in, tmp +cglobal fft32_ns_float, 4, 5, 16, ctx, out, in, stride, tmp call mangle(ff_tx_fft32_asm_float_ %+ %1) RET %endif @@ -1123,9 +1123,9 @@ ALIGN 16 %macro FFT_SPLIT_RADIX_FN 2 INIT_YMM %1 %if %2 -cglobal fft_sr_asm_float, 0, 0, 0, ctx, out, in, tmp, len, lut, itab, rtab, tgt, off +cglobal fft_sr_asm_float, 0, 0, 0, ctx, out, in, stride, len, lut, itab, rtab, tgt, tmp %else -cglobal fft_sr_float, 4, 10, 16, 272, ctx, out, in, tmp, len, lut, itab, rtab, tgt, off +cglobal fft_sr_float, 4, 10, 16, 272, ctx, out, in, stride, len, lut, itab, rtab, tgt, tmp movsxd lenq, dword [ctxq + AVTXContext.len] mov lutq, [ctxq + AVTXContext.map] %endif @@ -1391,12 +1391,15 @@ FFT_SPLIT_RADIX_DEF 131072 ; Final synthesis + deinterleaving code ;=== .deinterleave: +%if %2 +PUSH strideq +%endif mov tgtq, lenq imul tmpq, lenq, 2 -lea offq, [4*lenq + tmpq] +lea strideq, [4*lenq + tmpq] .synth_deinterleave: -SPLIT_RADIX_COMBINE_DEINTERLEAVE_FULL tmpq, offq +SPLIT_RADIX_COMBINE_DEINTERLEAVE_FULL tmpq, strideq add outq, 8*mmsize add rtabq, 4*mmsize sub itabq, 4*mmsize @@ -1404,6 +1407,7 @@ FFT_SPLIT_RADIX_DEF 131072 jg .synth_deinterleave %if %2 +POP strideq sub outq, tmpq neg tmpq lea inq, [inq + tmpq*4] @@ -1706,6 +1710,7 @@ cglobal mdct_inv_float, 4, 14, 16, 320, ctx, out, in, stride, len, lut, exp, t1, jge .stride4_pre .transform: +mov strideq, 2*4 mov t4q, ctxq ; backup original context mov t5q, [ctxq + AVTXContext.fn] ; subtransform's jump point mov ctxq, [ctxq + AVTXContext.sub] @@ -1767,7 +1772,7 @@ IMDCT_FN avx2 %macro PFA_15_FN 2 INIT_YMM %1 %if %2 -cglobal fft_pfa_15xM_asm_float, 0, 0, 0, ctx, out, in, stride, len, lut, buf, map, tgt, tmp, \ +cglobal fft_pfa_15xM_asm_float, 0, 8, 0, ctx, out, in, stride, len, lut, buf, map, tgt, tmp, \ tgt5, stride3,
[FFmpeg-cvslog] x86/tx_float: optimize and macro out FFT15
ffmpeg | branch: master | Lynne | Tue Sep 27 04:47:46 2022 +0200| [877e575b5d44adc252d4434d2ec53232b2000956] | committer: Lynne x86/tx_float: optimize and macro out FFT15 > http://git.videolan.org/gitweb.cgi/ffmpeg.git/?a=commit;h=877e575b5d44adc252d4434d2ec53232b2000956 --- libavutil/x86/tx_float.asm | 277 +++-- 1 file changed, 143 insertions(+), 134 deletions(-) diff --git a/libavutil/x86/tx_float.asm b/libavutil/x86/tx_float.asm index 5ed0007530..0061829581 100644 --- a/libavutil/x86/tx_float.asm +++ b/libavutil/x86/tx_float.asm @@ -91,7 +91,7 @@ s16_perm: dd 0, 1, 2, 3, 1, 0, 3, 2 s15_perm: dd 0, 6, 5, 3, 2, 4, 7, 1 -mask_mmpp: dd NEG, NEG, NEG, NEG, NEG, NEG, POS, POS +mask_mmpp: dd NEG, NEG, POS, POS, NEG, NEG, NEG, NEG mask_pppm: dd NEG, NEG, NEG, NEG, POS, POS, POS, NEG mask_ppmpmmpm: dd POS, POS, NEG, POS, NEG, NEG, POS, NEG mask_mppmmpmp: dd NEG, POS, POS, NEG, NEG, POS, NEG, POS @@ -307,6 +307,132 @@ SECTION .text %undef perm %endmacro +; Single 15-point complex FFT +; Input: +; xm0 must contain in[0,1].reim +; m2 - in[3-6].reim +; m3 - in[7-11].reim +; m4 - in[12-15].reim +; xm5 must contain in[2].reimreim +; +; Output: +; m0, m1, m2 - ACs +; xm14 - out[0] +; xm15 - out[10, 5] +%macro FFT15 0 +shufps xm1, xm0, xm0, q3223 ; in[1].imrereim +shufps xm0, xm0, xm0, q1001 ; in[0].imrereim + +xorps xm1, xm11 +addps xm1, xm0 ; pc[0,1].imre + +shufps xm0, xm1, xm1, q3232 ; pc[1].reimreim +addps xm0, xm5 ; dc[0].reimreim + +mulps xm1, xm9 ; tab[0123]*pc[01] + +shufpd xm6, xm1, xm1, 01b; pc[1,0].reim +xorps xm1, xm11 +addps xm1, xm1, xm6 +addsubps xm1, xm5, xm1 ; dc[1,2].reim + +subps m7, m2, m3 ; q[0-3].imre +addps m6, m2, m3 ; q[4-7] +shufps m7, m7, m7, q2301 ; q[0-3].reim + +addps m5, m4, m6 ; y[0-3] + +vperm2f128 m14, m9, m9, 0x11 ; tab[23232323] +vbroadcastsd m15, xm9; tab[01010101] + +mulps m6, m14 +mulps m7, m15 + +subps m2, m6, m7 ; k[0-3] +addps m3, m6, m7 ; k[4-7] + +shufps m12, m11, m11, q3232 ; + +addsubps m6, m4, m2 ; k[0-3] +addsubps m7, m4, m3 ; k[4-7] + +; 15pt from here on +vpermpd m2, m5, q0123; y[3-0] +vpermpd m3, m6, q0123; k[3-0] +vpermpd m4, m7, q0123; k[7-4] + +xorps m5, m12 +xorps m6, m12 +xorps m7, m12 + +addps m2, m5 ; t[0-3] +addps m3, m6 ; t[4-7] +addps m4, m7 ; t[8-11] + +movlhps xm14, xm2; out[0] +unpcklpd xm15, xm3, xm4 ; out[10,5] +unpckhpd xm5, xm3, xm4 ; out[10,5] + +addps xm14, xm2 ; out[0] +addps xm15, xm5 ; out[10,5] +addps xm14, xm0 ; out[0] +addps xm15, xm1 ; out[10,5] + +shufps m12, m10, m10, q3232 ; tab5 4 5 4 5 8 9 8 9 +shufps m13, m10, m10, q1010 ; tab5 6 7 6 7 10 11 10 11 + +mulps m5, m2, m12; t[0-3] +mulps m6, m3, m12; t[4-7] +mulps m7, m4, m12; t[8-11] + +mulps m2, m13; r[0-3] +mulps m3, m13; r[4-7] +mulps m4, m13; r[8-11] + +shufps m5, m5, m5, q1032 ; t[1,0,3,2].reim +shufps m6, m6, m6, q1032 ; t[5,4,7,6].reim +shufps m7, m7, m7, q1032 ; t[9,8,11,10].reim + +vperm2f128 m13, m11, m11, 0x01 ; mmpp +shufps m12, m11, m11, q3232 ; + +xorps m5, m13 +xorps m6, m13 +xorps m7, m13 + +addps m2, m5 ; r[0,1,2,3] +addps m3, m6 ; r[4,5,6,7] +addps m4, m7 ; r[8,9,10,11] + +shufps m5, m2, m2, q2301 +shufps m6, m3, m3, q2301 +shufps m7, m4, m4, q2301 + +xorps m2, m12 +xorps m3, m12 +xorps m4, m12 + +vpermpd m5, m5, q0123 +vpermpd m6, m6, q0123 +vpermpd m7, m7, q0123 + +addps m5, m2 +addps m6, m3 +addps m7, m4 + +vpermps m5, m8, m5 +vpermps m6, m8, m6 +vpermps m7, m8, m7 + +vbroadcastsd m0, xm0 ; dc[0] +vpermpd m2, m1, q; dc[2] +vbroadcastsd m1, xm1 ; dc[1] + +addps m0, m5 +addps m1, m6 +addps m2, m7 +%endmacro + ; Cobmines m0...m8 (tx1[even, even, odd, odd], tx2,3[even], tx2,3[odd]) coeffs ; Uses all 16 of registers. ; Output is slightly permuted such that tx2,3's coefficients are interleaved @@ -1610,11 +1736,10 @@ cglobal fft_pfa_15xM_float, 4, 14, 16, 320, ctx, out, in, stride, len, lut, buf, imul stride3q, strideq, 3 imul stride5q, strideq, 5 -movaps m13, [mask_mmpp]
[FFmpeg-cvslog] lavu/tx: drop requirement of input == output for in-place transforms
ffmpeg | branch: master | Lynne | Thu Nov 17 20:08:50 2022
+0100| [d4e39cae2e250a6fb9ed3a3a5a93694f4d445165] | committer: Lynne
lavu/tx: drop requirement of input == output for in-place transforms
No longer necessary.
> http://git.videolan.org/gitweb.cgi/ffmpeg.git/?a=commit;h=d4e39cae2e250a6fb9ed3a3a5a93694f4d445165
---
libavutil/tx.h | 5 ++---
1 file changed, 2 insertions(+), 3 deletions(-)
diff --git a/libavutil/tx.h b/libavutil/tx.h
index 3de2f7231b..cd772ad903 100644
--- a/libavutil/tx.h
+++ b/libavutil/tx.h
@@ -115,9 +115,8 @@ typedef void (*av_tx_fn)(AVTXContext *s, void *out, void
*in, ptrdiff_t stride);
*/
enum AVTXFlags {
/**
- * Performs an in-place transformation on the input. The output argument
- * of av_tn_fn() MUST match the input. May be unsupported or slower for
some
- * transform types.
+ * Allows for in-place transformations, where input == output.
+ * May be unsupported or slower for some transform types.
*/
AV_TX_INPLACE = 1ULL << 0,
___
ffmpeg-cvslog mailing list
ffmpeg-cvslog@ffmpeg.org
https://ffmpeg.org/mailman/listinfo/ffmpeg-cvslog
To unsubscribe, visit link above, or email
ffmpeg-cvslog-requ...@ffmpeg.org with subject "unsubscribe".
[FFmpeg-cvslog] lavu/tx: support out-of-place transforms in fft_inplace
ffmpeg | branch: master | Lynne | Thu Nov 17 20:06:43 2022
+0100| [fff3e1d8489ee83949f67faba8908755846a6f4f] | committer: Lynne
lavu/tx: support out-of-place transforms in fft_inplace
This makes testing easier, as a unified path can be used for in/out of
place transforms.
> http://git.videolan.org/gitweb.cgi/ffmpeg.git/?a=commit;h=fff3e1d8489ee83949f67faba8908755846a6f4f
---
libavutil/tx_template.c | 13 +++--
1 file changed, 7 insertions(+), 6 deletions(-)
diff --git a/libavutil/tx_template.c b/libavutil/tx_template.c
index 2a8afcb02a..5274133ec4 100644
--- a/libavutil/tx_template.c
+++ b/libavutil/tx_template.c
@@ -773,6 +773,7 @@ static void TX_NAME(ff_tx_fft)(AVTXContext *s, void *_dst,
static void TX_NAME(ff_tx_fft_inplace)(AVTXContext *s, void *_dst,
void *_src, ptrdiff_t stride)
{
+TXComplex *src = _src;
TXComplex *dst = _dst;
TXComplex tmp;
const int *map = s->sub->map;
@@ -781,16 +782,16 @@ static void TX_NAME(ff_tx_fft_inplace)(AVTXContext *s,
void *_dst,
src_idx = *inplace_idx++;
do {
-tmp = dst[src_idx];
+tmp = src[src_idx];
dst_idx = map[src_idx];
do {
-FFSWAP(TXComplex, tmp, dst[dst_idx]);
+FFSWAP(TXComplex, tmp, src[dst_idx]);
dst_idx = map[dst_idx];
} while (dst_idx != src_idx); /* Can be > as well, but was less
predictable */
-dst[dst_idx] = tmp;
+src[dst_idx] = tmp;
} while ((src_idx = *inplace_idx++));
-s->fn[0](>sub[0], dst, dst, stride);
+s->fn[0](>sub[0], dst, src, stride);
}
static const FFTXCodelet TX_NAME(ff_tx_fft_def) = {
@@ -810,13 +811,13 @@ static const FFTXCodelet TX_NAME(ff_tx_fft_inplace_def) =
{
.name = TX_NAME_STR("fft_inplace"),
.function = TX_NAME(ff_tx_fft_inplace),
.type = TX_TYPE(FFT),
-.flags = AV_TX_UNALIGNED | AV_TX_INPLACE,
+.flags = AV_TX_UNALIGNED | FF_TX_OUT_OF_PLACE | AV_TX_INPLACE,
.factors[0] = TX_FACTOR_ANY,
.min_len= 2,
.max_len= TX_LEN_UNLIMITED,
.init = TX_NAME(ff_tx_fft_init),
.cpu_flags = FF_TX_CPU_FLAGS_ALL,
-.prio = FF_TX_PRIO_BASE,
+.prio = FF_TX_PRIO_BASE - 512,
};
static av_cold int TX_NAME(ff_tx_fft_init_naive_small)(AVTXContext *s,
___
ffmpeg-cvslog mailing list
ffmpeg-cvslog@ffmpeg.org
https://ffmpeg.org/mailman/listinfo/ffmpeg-cvslog
To unsubscribe, visit link above, or email
ffmpeg-cvslog-requ...@ffmpeg.org with subject "unsubscribe".
[FFmpeg-cvslog] x86/tx_float: add a standalone 15-point AVX2 transform
ffmpeg | branch: master | Lynne | Wed Sep 28 06:46:57 2022
+0200| [cc1df4045eba7273b573ecb40380f000144d] | committer: Lynne
x86/tx_float: add a standalone 15-point AVX2 transform
Enables its use everywhere else in the framework.
> http://git.videolan.org/gitweb.cgi/ffmpeg.git/?a=commit;h=cc1df4045eba7273b573ecb40380f000144d
---
libavutil/x86/tx_float.asm| 63 +++
libavutil/x86/tx_float_init.c | 54 +
2 files changed, 117 insertions(+)
diff --git a/libavutil/x86/tx_float.asm b/libavutil/x86/tx_float.asm
index 0061829581..6f83555ce5 100644
--- a/libavutil/x86/tx_float.asm
+++ b/libavutil/x86/tx_float.asm
@@ -1515,6 +1515,69 @@ FFT_SPLIT_RADIX_FN avx2, 1
%endif
%endif
+%macro FFT15_FN 2
+INIT_YMM avx2
+cglobal fft15_ %+ %2, 4, 10, 16, ctx, out, in, stride, len, lut, tmp, tgt5,
stride3, stride5
+mov lutq, [ctxq + AVTXContext.map]
+
+imul stride3q, strideq, 3
+imul stride5q, strideq, 5
+
+movaps m11, [mask_mmpp] ; mmpp
+movaps m10, [tab_53_float] ; tab5
+movaps xm9, [tab_53_float + 32] ; tab3
+vpermpd m9, m9, q1110; tab[23232323]
+movaps m8, [s15_perm]
+
+%if %1
+movups xm0, [inq]
+movddup xm5, [inq + 16]
+movups m2, [inq + mmsize*0 + 24]
+movups m3, [inq + mmsize*1 + 24]
+movups m4, [inq + mmsize*2 + 24]
+%else
+LOAD64_LUT xm0, inq, lutq, 0, tmpq, m14, xm15
+LOAD64_LUT m2, inq, lutq, (mmsize/2)*0 + 12, tmpq, m6, m7
+LOAD64_LUT m3, inq, lutq, (mmsize/2)*1 + 12, tmpq, m14, m15
+LOAD64_LUT m4, inq, lutq, (mmsize/2)*2 + 12, tmpq, m6, m7
+mov tmpd, [lutq + 8]
+movddup xm5, [inq + tmpq*8]
+%endif
+
+FFT15
+
+lea tgt5q, [outq + stride5q]
+lea tmpq, [outq + stride5q*2]
+
+movhps [outq], xm14 ; out[0]
+movhps [outq + stride5q*1], xm15 ; out[5]
+movlps [outq + stride5q*2], xm15 ; out[10]
+
+vextractf128 xm3, m0, 1
+vextractf128 xm4, m1, 1
+vextractf128 xm5, m2, 1
+
+movlps [outq + strideq*1], xm1
+movhps [outq + strideq*2], xm2
+movlps [outq + stride3q*1], xm3
+movhps [outq + strideq*4], xm4
+movlps [outq + stride3q*2], xm0
+movlps [outq + strideq*8], xm5
+movhps [outq + stride3q*4], xm0
+movhps [tgt5q + strideq*2], xm1
+movhps [tgt5q + strideq*4], xm3
+movlps [tmpq + strideq*1], xm2
+movlps [tmpq + stride3q*1], xm4
+movhps [tmpq + strideq*4], xm5
+
+RET
+%endmacro
+
+%if ARCH_X86_64 && HAVE_AVX2_EXTERNAL
+FFT15_FN 0, float
+FFT15_FN 1, ns_float
+%endif
+
%macro IMDCT_FN 1
INIT_YMM %1
cglobal mdct_inv_float, 4, 14, 16, 320, ctx, out, in, stride, len, lut, exp,
t1, t2, t3, \
diff --git a/libavutil/x86/tx_float_init.c b/libavutil/x86/tx_float_init.c
index 8e2babb539..523b43a689 100644
--- a/libavutil/x86/tx_float_init.c
+++ b/libavutil/x86/tx_float_init.c
@@ -30,6 +30,8 @@ TX_DECL_FN(fft8, sse3)
TX_DECL_FN(fft8_ns, sse3)
TX_DECL_FN(fft8, avx)
TX_DECL_FN(fft8_ns, avx)
+TX_DECL_FN(fft15, avx2)
+TX_DECL_FN(fft15_ns, avx2)
TX_DECL_FN(fft16, avx)
TX_DECL_FN(fft16_ns, avx)
TX_DECL_FN(fft16, fma3)
@@ -85,6 +87,53 @@ static av_cold int b ##basis## _i ##interleave(AVTXContext
*s, \
DECL_INIT_FN(8, 0)
DECL_INIT_FN(8, 2)
+static av_cold int factor_init(AVTXContext *s, const FFTXCodelet *cd,
+ uint64_t flags, FFTXCodeletOptions *opts,
+ int len, int inv, const void *scale)
+{
+TX_TAB(ff_tx_init_tabs)(len);
+
+s->map = av_malloc(len*sizeof(s->map));
+s->map[0] = 0; /* DC is always at the start */
+if (inv) /* Reversing the ACs flips the transform direction */
+for (int i = 1; i < len; i++)
+s->map[i] = len - i;
+else
+for (int i = 1; i < len; i++)
+s->map[i] = i;
+
+if (len == 15) {
+int cnt = 0, tmp[15];
+
+/* Our 15-point transform is actually a 5x3 PFA, so embed its input
map. */
+memcpy(tmp, s->map, 15*sizeof(*tmp));
+for (int i = 0; i < 5; i++)
+for (int j = 0; j < 3; j++)
+s->map[i*3 + j] = tmp[(i*3 + j*5) % 15];
+
+/* Special 15-point assembly permutation */
+memcpy(tmp, s->map, 15*sizeof(*tmp));
+for (int i = 1; i < 15; i += 3) {
+s->map[cnt] = tmp[i];
+cnt++;
+}
+for (int i = 2; i < 15; i += 3) {
+s->map[cnt] = tmp[i];
+cnt++;
+}
+for (int i = 0; i < 15; i += 3) {
+s->map[cnt] = tmp[i];
+cnt++;
+}
+memmove(>map[7], >map[6], 4*sizeof(int));
+memmove(>map[3], >map[1], 4*sizeof(int));
+s->map[1] = tmp[2];
+s->map[2] = tmp[0];
+}
+
+return 0;
+}
+
static av_cold int m_inv_init(AVTXContext *s, const FFTXCodelet *cd,
uint64_t flags, FFTXCodeletOptions *opts,
[FFmpeg-cvslog] lavu/tx: add fft_inplace_small transforms
ffmpeg | branch: master | Lynne | Thu Nov 17 20:10:45 2022
+0100| [68cabf875015610decda7e564dc5697f6c21f707] | committer: Lynne
lavu/tx: add fft_inplace_small transforms
This is much faster than the loop.
> http://git.videolan.org/gitweb.cgi/ffmpeg.git/?a=commit;h=68cabf875015610decda7e564dc5697f6c21f707
---
libavutil/tx_template.c | 34 +++---
1 file changed, 31 insertions(+), 3 deletions(-)
diff --git a/libavutil/tx_template.c b/libavutil/tx_template.c
index 5274133ec4..747731a06d 100644
--- a/libavutil/tx_template.c
+++ b/libavutil/tx_template.c
@@ -754,20 +754,34 @@ static av_cold int TX_NAME(ff_tx_fft_init)(AVTXContext *s,
return 0;
}
+static av_cold int TX_NAME(ff_tx_fft_inplace_small_init)(AVTXContext *s,
+ const FFTXCodelet *cd,
+ uint64_t flags,
+ FFTXCodeletOptions
*opts,
+ int len, int inv,
+ const void *scale)
+{
+if (!(s->tmp = av_malloc(len*sizeof(*s->tmp
+return AVERROR(ENOMEM);
+flags &= ~AV_TX_INPLACE;
+return TX_NAME(ff_tx_fft_init)(s, cd, flags, opts, len, inv, scale);
+}
+
static void TX_NAME(ff_tx_fft)(AVTXContext *s, void *_dst,
void *_src, ptrdiff_t stride)
{
TXComplex *src = _src;
-TXComplex *dst = _dst;
+TXComplex *dst1 = s->flags & AV_TX_INPLACE ? s->tmp : _dst;
+TXComplex *dst2 = _dst;
int *map = s->sub[0].map;
int len = s->len;
/* Compilers can't vectorize this anyway without assuming AVX2, which they
* generally don't, at least without -march=native -mtune=native */
for (int i = 0; i < len; i++)
-dst[i] = src[map[i]];
+dst1[i] = src[map[i]];
-s->fn[0](>sub[0], dst, dst, stride);
+s->fn[0](>sub[0], dst2, dst1, stride);
}
static void TX_NAME(ff_tx_fft_inplace)(AVTXContext *s, void *_dst,
@@ -807,6 +821,19 @@ static const FFTXCodelet TX_NAME(ff_tx_fft_def) = {
.prio = FF_TX_PRIO_BASE,
};
+static const FFTXCodelet TX_NAME(ff_tx_fft_inplace_small_def) = {
+.name = TX_NAME_STR("fft_inplace_small"),
+.function = TX_NAME(ff_tx_fft),
+.type = TX_TYPE(FFT),
+.flags = AV_TX_UNALIGNED | FF_TX_OUT_OF_PLACE | AV_TX_INPLACE,
+.factors[0] = TX_FACTOR_ANY,
+.min_len= 2,
+.max_len= 65536,
+.init = TX_NAME(ff_tx_fft_inplace_small_init),
+.cpu_flags = FF_TX_CPU_FLAGS_ALL,
+.prio = FF_TX_PRIO_BASE - 256,
+};
+
static const FFTXCodelet TX_NAME(ff_tx_fft_inplace_def) = {
.name = TX_NAME_STR("fft_inplace"),
.function = TX_NAME(ff_tx_fft_inplace),
@@ -1638,6 +1665,7 @@ const FFTXCodelet * const TX_NAME(ff_tx_codelet_list)[] =
{
/* Standalone transforms */
_NAME(ff_tx_fft_def),
_NAME(ff_tx_fft_inplace_def),
+_NAME(ff_tx_fft_inplace_small_def),
_NAME(ff_tx_fft_pfa_3xM_def),
_NAME(ff_tx_fft_pfa_5xM_def),
_NAME(ff_tx_fft_pfa_7xM_def),
___
ffmpeg-cvslog mailing list
ffmpeg-cvslog@ffmpeg.org
https://ffmpeg.org/mailman/listinfo/ffmpeg-cvslog
To unsubscribe, visit link above, or email
ffmpeg-cvslog-requ...@ffmpeg.org with subject "unsubscribe".
[FFmpeg-cvslog] lavu/tx: make C ptwo transforms in+out of place
ffmpeg | branch: master | Lynne | Thu Nov 17 20:03:09 2022
+0100| [d260796f119682274c83e2f1465f56f3e314c4a4] | committer: Lynne
lavu/tx: make C ptwo transforms in+out of place
We assume that _all_ in-place transforms can operate out of place,
which isn't true, because the C ptwo transforms were always in-place (dst).
> http://git.videolan.org/gitweb.cgi/ffmpeg.git/?a=commit;h=d260796f119682274c83e2f1465f56f3e314c4a4
---
libavutil/tx_template.c | 117 +---
1 file changed, 61 insertions(+), 56 deletions(-)
diff --git a/libavutil/tx_template.c b/libavutil/tx_template.c
index f53a241248..2a8afcb02a 100644
--- a/libavutil/tx_template.c
+++ b/libavutil/tx_template.c
@@ -611,8 +611,8 @@ static const FFTXCodelet TX_NAME(ff_tx_fft##n##_ns_def) = {
\
.name = TX_NAME_STR("fft" #n "_ns"), \
.function = TX_NAME(ff_tx_fft##n##_ns), \
.type = TX_TYPE(FFT), \
-.flags = AV_TX_INPLACE | AV_TX_UNALIGNED | \
- FF_TX_PRESHUFFLE, \
+.flags = FF_TX_OUT_OF_PLACE | AV_TX_INPLACE | \
+ AV_TX_UNALIGNED | FF_TX_PRESHUFFLE, \
.factors[0] = 2,\
.min_len= n,\
.max_len= n,\
@@ -621,70 +621,75 @@ static const FFTXCodelet TX_NAME(ff_tx_fft##n##_ns_def) =
{ \
.prio = FF_TX_PRIO_BASE, \
};
-#define DECL_SR_CODELET(n, n2, n4) \
-static void TX_NAME(ff_tx_fft##n##_ns)(AVTXContext *s, void *dst,\
-void *src, ptrdiff_t stride) \
-{\
-TXComplex *z = dst; \
-const TXSample *cos = TX_TAB(ff_tx_tab_##n); \
- \
-TX_NAME(ff_tx_fft##n2##_ns)(s, z,z,stride); \
-TX_NAME(ff_tx_fft##n4##_ns)(s, z + n4*2, z + n4*2, stride); \
-TX_NAME(ff_tx_fft##n4##_ns)(s, z + n4*3, z + n4*3, stride); \
-TX_NAME(ff_tx_fft_sr_combine)(z, cos, n4 >> 1); \
-}\
- \
+#define DECL_SR_CODELET(n, n2, n4)\
+static void TX_NAME(ff_tx_fft##n##_ns)(AVTXContext *s, void *_dst,\
+void *_src, ptrdiff_t stride) \
+{ \
+TXComplex *src = _src;\
+TXComplex *dst = _dst;\
+const TXSample *cos = TX_TAB(ff_tx_tab_##n); \
+ \
+TX_NAME(ff_tx_fft##n2##_ns)(s, dst,src,stride); \
+TX_NAME(ff_tx_fft##n4##_ns)(s, dst + n4*2, src + n4*2, stride); \
+TX_NAME(ff_tx_fft##n4##_ns)(s, dst + n4*3, src + n4*3, stride); \
+TX_NAME(ff_tx_fft_sr_combine)(dst, cos, n4 >> 1); \
+} \
+ \
DECL_SR_CODELET_DEF(n)
-static void TX_NAME(ff_tx_fft2_ns)(AVTXContext *s, void *dst,
- void *src, ptrdiff_t stride)
+static void TX_NAME(ff_tx_fft2_ns)(AVTXContext *s, void *_dst,
+ void *_src, ptrdiff_t stride)
{
-TXComplex *z = dst;
+TXComplex *src = _src;
+TXComplex *dst = _dst;
TXComplex tmp;
-BF(tmp.re, z[0].re, z[0].re, z[1].re);
-BF(tmp.im, z[0].im, z[0].im, z[1].im);
-z[1] = tmp;
+BF(tmp.re, dst[0].re, src[0].re, src[1].re);
+BF(tmp.im, dst[0].im, src[0].im, src[1].im);
+dst[1] = tmp;
}
-static void TX_NAME(ff_tx_fft4_ns)(AVTXContext *s, void *dst,
- void *src, ptrdiff_t stride)
+static void TX_NAME(ff_tx_fft4_ns)(AVTXContext *s, void *_dst,
+ void *_src, ptrdiff_t stride)
{
-TXComplex *z = dst;
+TXComplex *src = _src;
+TXComplex *dst = _dst;
TXSample t1, t2, t3, t4, t5, t6, t7, t8;
-BF(t3, t1, z[0].re, z[1].re);
-BF(t8, t6, z[3].re, z[2].re);
-BF(z[2].re, z[0].re, t1, t6);
-BF(t4, t2, z[0].im, z[1].im);
-BF(t7, t5, z[2].im, z[3].im);
-BF(z[3].im, z[1].im, t4, t8);
-BF(z[3].re, z[1].re, t3, t7);
-BF(z[2].im, z[0].im, t2, t5);
+BF(t3, t1, src[0].re, src[1].re);
+BF(t8, t6, src[3].re, src[2].re);
+BF(dst[2].re, dst[0].re, t1, t6);
+BF(t4, t2, src[0].im, src[1].im);
+
[FFmpeg-cvslog] lavu/tx: support output stride in naive transforms
ffmpeg | branch: master | Lynne | Wed Sep 28 12:59:08 2022
+0200| [fbe4fd992f4327fcf17b2a76a823c38945b0ea13] | committer: Lynne
lavu/tx: support output stride in naive transforms
Allows them to be used in general PFAs.
> http://git.videolan.org/gitweb.cgi/ffmpeg.git/?a=commit;h=fbe4fd992f4327fcf17b2a76a823c38945b0ea13
---
libavutil/tx_template.c | 8 ++--
1 file changed, 6 insertions(+), 2 deletions(-)
diff --git a/libavutil/tx_template.c b/libavutil/tx_template.c
index 747731a06d..228209521b 100644
--- a/libavutil/tx_template.c
+++ b/libavutil/tx_template.c
@@ -880,6 +880,8 @@ static void TX_NAME(ff_tx_fft_naive)(AVTXContext *s, void
*_dst, void *_src,
const int n = s->len;
double phase = s->inv ? 2.0*M_PI/n : -2.0*M_PI/n;
+stride /= sizeof(*dst);
+
for (int i = 0; i < n; i++) {
TXComplex tmp = { 0 };
for (int j = 0; j < n; j++) {
@@ -893,7 +895,7 @@ static void TX_NAME(ff_tx_fft_naive)(AVTXContext *s, void
*_dst, void *_src,
tmp.re += res.re;
tmp.im += res.im;
}
-dst[i] = tmp;
+dst[i*stride] = tmp;
}
}
@@ -904,6 +906,8 @@ static void TX_NAME(ff_tx_fft_naive_small)(AVTXContext *s,
void *_dst, void *_sr
TXComplex *dst = _dst;
const int n = s->len;
+stride /= sizeof(*dst);
+
for (int i = 0; i < n; i++) {
TXComplex tmp = { 0 };
for (int j = 0; j < n; j++) {
@@ -913,7 +917,7 @@ static void TX_NAME(ff_tx_fft_naive_small)(AVTXContext *s,
void *_dst, void *_sr
tmp.re += res.re;
tmp.im += res.im;
}
-dst[i] = tmp;
+dst[i*stride] = tmp;
}
}
___
ffmpeg-cvslog mailing list
ffmpeg-cvslog@ffmpeg.org
https://ffmpeg.org/mailman/listinfo/ffmpeg-cvslog
To unsubscribe, visit link above, or email
ffmpeg-cvslog-requ...@ffmpeg.org with subject "unsubscribe".
[FFmpeg-cvslog] lavu/tx: add naive_small FFT
ffmpeg | branch: master | Lynne | Sun Sep 25 08:19:17 2022
+0200| [37008dc4026c6a460c454a95f3f2766afbc702e3] | committer: Lynne
lavu/tx: add naive_small FFT
The same as naive but with precomputed tables. Makes it more useful
for odd-factors we don't support yet.
> http://git.videolan.org/gitweb.cgi/ffmpeg.git/?a=commit;h=37008dc4026c6a460c454a95f3f2766afbc702e3
---
libavutil/tx_template.c | 63 +++--
1 file changed, 61 insertions(+), 2 deletions(-)
diff --git a/libavutil/tx_template.c b/libavutil/tx_template.c
index d72281f09c..f53a241248 100644
--- a/libavutil/tx_template.c
+++ b/libavutil/tx_template.c
@@ -814,6 +814,31 @@ static const FFTXCodelet TX_NAME(ff_tx_fft_inplace_def) = {
.prio = FF_TX_PRIO_BASE,
};
+static av_cold int TX_NAME(ff_tx_fft_init_naive_small)(AVTXContext *s,
+ const FFTXCodelet *cd,
+ uint64_t flags,
+ FFTXCodeletOptions
*opts,
+ int len, int inv,
+ const void *scale)
+{
+const double phase = s->inv ? 2.0*M_PI/len : -2.0*M_PI/len;
+
+if (!(s->exp = av_malloc(len*len*sizeof(*s->exp
+return AVERROR(ENOMEM);
+
+for (int i = 0; i < len; i++) {
+for (int j = 0; j < len; j++) {
+const double factor = phase*i*j;
+s->exp[i*j] = (TXComplex){
+RESCALE(cos(factor)),
+RESCALE(sin(factor)),
+};
+}
+}
+
+return 0;
+}
+
static void TX_NAME(ff_tx_fft_naive)(AVTXContext *s, void *_dst, void *_src,
ptrdiff_t stride)
{
@@ -822,9 +847,9 @@ static void TX_NAME(ff_tx_fft_naive)(AVTXContext *s, void
*_dst, void *_src,
const int n = s->len;
double phase = s->inv ? 2.0*M_PI/n : -2.0*M_PI/n;
-for(int i = 0; i < n; i++) {
+for (int i = 0; i < n; i++) {
TXComplex tmp = { 0 };
-for(int j = 0; j < n; j++) {
+for (int j = 0; j < n; j++) {
const double factor = phase*i*j;
const TXComplex mult = {
RESCALE(cos(factor)),
@@ -839,6 +864,39 @@ static void TX_NAME(ff_tx_fft_naive)(AVTXContext *s, void
*_dst, void *_src,
}
}
+static void TX_NAME(ff_tx_fft_naive_small)(AVTXContext *s, void *_dst, void
*_src,
+ ptrdiff_t stride)
+{
+TXComplex *src = _src;
+TXComplex *dst = _dst;
+const int n = s->len;
+
+for (int i = 0; i < n; i++) {
+TXComplex tmp = { 0 };
+for (int j = 0; j < n; j++) {
+TXComplex res;
+const TXComplex mult = s->exp[i*j];
+CMUL3(res, src[j], mult);
+tmp.re += res.re;
+tmp.im += res.im;
+}
+dst[i] = tmp;
+}
+}
+
+static const FFTXCodelet TX_NAME(ff_tx_fft_naive_small_def) = {
+.name = TX_NAME_STR("fft_naive_small"),
+.function = TX_NAME(ff_tx_fft_naive_small),
+.type = TX_TYPE(FFT),
+.flags = AV_TX_UNALIGNED | FF_TX_OUT_OF_PLACE,
+.factors[0] = TX_FACTOR_ANY,
+.min_len= 2,
+.max_len= 1024,
+.init = TX_NAME(ff_tx_fft_init_naive_small),
+.cpu_flags = FF_TX_CPU_FLAGS_ALL,
+.prio = FF_TX_PRIO_MIN/2,
+};
+
static const FFTXCodelet TX_NAME(ff_tx_fft_naive_def) = {
.name = TX_NAME_STR("fft_naive"),
.function = TX_NAME(ff_tx_fft_naive),
@@ -1580,6 +1638,7 @@ const FFTXCodelet * const TX_NAME(ff_tx_codelet_list)[] =
{
_NAME(ff_tx_fft_pfa_9xM_def),
_NAME(ff_tx_fft_pfa_15xM_def),
_NAME(ff_tx_fft_naive_def),
+_NAME(ff_tx_fft_naive_small_def),
_NAME(ff_tx_mdct_fwd_def),
_NAME(ff_tx_mdct_inv_def),
_NAME(ff_tx_mdct_pfa_3xM_fwd_def),
___
ffmpeg-cvslog mailing list
ffmpeg-cvslog@ffmpeg.org
https://ffmpeg.org/mailman/listinfo/ffmpeg-cvslog
To unsubscribe, visit link above, or email
ffmpeg-cvslog-requ...@ffmpeg.org with subject "unsubscribe".
[FFmpeg-cvslog] lavu/tx: generalize single-factor transforms
ffmpeg | branch: master | Lynne | Sat Sep 24 06:49:16 2022
+0200| [45bd4bf79f9b69ac4cec1bd00c433407b3aa7ae4] | committer: Lynne
lavu/tx: generalize single-factor transforms
Not that useful, but it gives us fast small odd-length transforms.
> http://git.videolan.org/gitweb.cgi/ffmpeg.git/?a=commit;h=45bd4bf79f9b69ac4cec1bd00c433407b3aa7ae4
---
libavutil/tx.c | 4 ++--
libavutil/tx_priv.h | 2 +-
libavutil/tx_template.c | 46 +++---
3 files changed, 26 insertions(+), 26 deletions(-)
diff --git a/libavutil/tx.c b/libavutil/tx.c
index 556fcbb94b..246a7aa980 100644
--- a/libavutil/tx.c
+++ b/libavutil/tx.c
@@ -121,9 +121,9 @@ int ff_tx_gen_ptwo_revtab(AVTXContext *s, int invert_lookup)
return 0;
}
-int ff_tx_gen_ptwo_inplace_revtab_idx(AVTXContext *s)
+int ff_tx_gen_inplace_map(AVTXContext *s, int len)
{
-int *src_map, out_map_idx = 0, len = s->len;
+int *src_map, out_map_idx = 0;
if (!s->sub || !s->sub->map)
return AVERROR(EINVAL);
diff --git a/libavutil/tx_priv.h b/libavutil/tx_priv.h
index fb61119009..56e78631ba 100644
--- a/libavutil/tx_priv.h
+++ b/libavutil/tx_priv.h
@@ -259,7 +259,7 @@ int ff_tx_gen_ptwo_revtab(AVTXContext *s, int
invert_lookup);
* specific order, allows the revtab to be done in-place. The sub-transform
* and its map should already be initialized.
*/
-int ff_tx_gen_ptwo_inplace_revtab_idx(AVTXContext *s);
+int ff_tx_gen_inplace_map(AVTXContext *s, int len);
/*
* This generates a parity-based revtab of length len and direction inv.
diff --git a/libavutil/tx_template.c b/libavutil/tx_template.c
index 6c8d0a1ebc..b547800447 100644
--- a/libavutil/tx_template.c
+++ b/libavutil/tx_template.c
@@ -650,12 +650,12 @@ DECL_SR_CODELET(32768,16384,8192)
DECL_SR_CODELET(65536,32768,16384)
DECL_SR_CODELET(131072,65536,32768)
-static av_cold int TX_NAME(ff_tx_fft_sr_init)(AVTXContext *s,
- const FFTXCodelet *cd,
- uint64_t flags,
- FFTXCodeletOptions *opts,
- int len, int inv,
- const void *scale)
+static av_cold int TX_NAME(ff_tx_fft_init)(AVTXContext *s,
+ const FFTXCodelet *cd,
+ uint64_t flags,
+ FFTXCodeletOptions *opts,
+ int len, int inv,
+ const void *scale)
{
int ret;
int is_inplace = !!(flags & AV_TX_INPLACE);
@@ -668,14 +668,14 @@ static av_cold int TX_NAME(ff_tx_fft_sr_init)(AVTXContext
*s,
if ((ret = ff_tx_init_subtx(s, TX_TYPE(FFT), flags, _opts, len, inv,
scale)))
return ret;
-if (is_inplace && (ret = ff_tx_gen_ptwo_inplace_revtab_idx(s)))
+if (is_inplace && (ret = ff_tx_gen_inplace_map(s, len)))
return ret;
return 0;
}
-static void TX_NAME(ff_tx_fft_sr)(AVTXContext *s, void *_dst,
- void *_src, ptrdiff_t stride)
+static void TX_NAME(ff_tx_fft)(AVTXContext *s, void *_dst,
+ void *_src, ptrdiff_t stride)
{
TXComplex *src = _src;
TXComplex *dst = _dst;
@@ -690,8 +690,8 @@ static void TX_NAME(ff_tx_fft_sr)(AVTXContext *s, void
*_dst,
s->fn[0](>sub[0], dst, dst, stride);
}
-static void TX_NAME(ff_tx_fft_sr_inplace)(AVTXContext *s, void *_dst,
- void *_src, ptrdiff_t stride)
+static void TX_NAME(ff_tx_fft_inplace)(AVTXContext *s, void *_dst,
+ void *_src, ptrdiff_t stride)
{
TXComplex *dst = _dst;
TXComplex tmp;
@@ -713,28 +713,28 @@ static void TX_NAME(ff_tx_fft_sr_inplace)(AVTXContext *s,
void *_dst,
s->fn[0](>sub[0], dst, dst, stride);
}
-static const FFTXCodelet TX_NAME(ff_tx_fft_sr_def) = {
-.name = TX_NAME_STR("fft_sr"),
-.function = TX_NAME(ff_tx_fft_sr),
+static const FFTXCodelet TX_NAME(ff_tx_fft_def) = {
+.name = TX_NAME_STR("fft"),
+.function = TX_NAME(ff_tx_fft),
.type = TX_TYPE(FFT),
.flags = AV_TX_UNALIGNED | FF_TX_OUT_OF_PLACE,
-.factors[0] = 2,
+.factors[0] = TX_FACTOR_ANY,
.min_len= 2,
.max_len= TX_LEN_UNLIMITED,
-.init = TX_NAME(ff_tx_fft_sr_init),
+.init = TX_NAME(ff_tx_fft_init),
.cpu_flags = FF_TX_CPU_FLAGS_ALL,
.prio = FF_TX_PRIO_BASE,
};
-static const FFTXCodelet TX_NAME(ff_tx_fft_sr_inplace_def) = {
-.name = TX_NAME_STR("fft_sr_inplace"),
-.function = TX_NAME(ff_tx_fft_sr_inplace),
+static const FFTXCodelet TX_NAME(ff_tx_fft_inplace_def) = {
+.name = TX_NAME_STR("fft_inplace"),
+.function = TX_NAME(ff_tx_fft_inplace),
.type = TX_TYPE(FFT),
[FFmpeg-cvslog] lavu/tx: make prime factor transforms truly in-place
ffmpeg | branch: master | Lynne | Sat Sep 24 06:47:21 2022
+0200| [79f11e24098d6392015656897bc7842c9d2aea43] | committer: Lynne
lavu/tx: make prime factor transforms truly in-place
They all overwrote in[0] and then used it as a DC.
> http://git.videolan.org/gitweb.cgi/ffmpeg.git/?a=commit;h=79f11e24098d6392015656897bc7842c9d2aea43
---
libavutil/tx_template.c | 108 +---
1 file changed, 56 insertions(+), 52 deletions(-)
diff --git a/libavutil/tx_template.c b/libavutil/tx_template.c
index 666af5e496..6c8d0a1ebc 100644
--- a/libavutil/tx_template.c
+++ b/libavutil/tx_template.c
@@ -171,36 +171,37 @@ av_cold void TX_TAB(ff_tx_init_tabs)(int len)
static av_always_inline void fft3(TXComplex *out, TXComplex *in,
ptrdiff_t stride)
{
-TXComplex tmp[2];
+TXComplex tmp[3];
const TXSample *tab = TX_TAB(ff_tx_tab_53);
#ifdef TX_INT32
int64_t mtmp[4];
#endif
-BF(tmp[0].re, tmp[1].im, in[1].im, in[2].im);
-BF(tmp[0].im, tmp[1].re, in[1].re, in[2].re);
+tmp[0] = in[0];
+BF(tmp[1].re, tmp[2].im, in[1].im, in[2].im);
+BF(tmp[1].im, tmp[2].re, in[1].re, in[2].re);
-out[0*stride].re = in[0].re + tmp[1].re;
-out[0*stride].im = in[0].im + tmp[1].im;
+out[0*stride].re = tmp[0].re + tmp[2].re;
+out[0*stride].im = tmp[0].im + tmp[2].im;
#ifdef TX_INT32
-mtmp[0] = (int64_t)tab[ 8] * tmp[0].re;
-mtmp[1] = (int64_t)tab[ 9] * tmp[0].im;
-mtmp[2] = (int64_t)tab[10] * tmp[1].re;
-mtmp[3] = (int64_t)tab[10] * tmp[1].im;
-out[1*stride].re = in[0].re - (mtmp[2] + mtmp[0] + 0x4000 >> 31);
-out[1*stride].im = in[0].im - (mtmp[3] - mtmp[1] + 0x4000 >> 31);
-out[2*stride].re = in[0].re - (mtmp[2] - mtmp[0] + 0x4000 >> 31);
-out[2*stride].im = in[0].im - (mtmp[3] + mtmp[1] + 0x4000 >> 31);
+mtmp[0] = (int64_t)tab[ 8] * tmp[1].re;
+mtmp[1] = (int64_t)tab[ 9] * tmp[1].im;
+mtmp[2] = (int64_t)tab[10] * tmp[2].re;
+mtmp[3] = (int64_t)tab[10] * tmp[2].im;
+out[1*stride].re = tmp[0].re - (mtmp[2] + mtmp[0] + 0x4000 >> 31);
+out[1*stride].im = tmp[0].im - (mtmp[3] - mtmp[1] + 0x4000 >> 31);
+out[2*stride].re = tmp[0].re - (mtmp[2] - mtmp[0] + 0x4000 >> 31);
+out[2*stride].im = tmp[0].im - (mtmp[3] + mtmp[1] + 0x4000 >> 31);
#else
-tmp[0].re = tab[ 8] * tmp[0].re;
-tmp[0].im = tab[ 9] * tmp[0].im;
-tmp[1].re = tab[10] * tmp[1].re;
-tmp[1].im = tab[10] * tmp[1].im;
-out[1*stride].re = in[0].re - tmp[1].re + tmp[0].re;
-out[1*stride].im = in[0].im - tmp[1].im - tmp[0].im;
-out[2*stride].re = in[0].re - tmp[1].re - tmp[0].re;
-out[2*stride].im = in[0].im - tmp[1].im + tmp[0].im;
+tmp[1].re = tab[ 8] * tmp[1].re;
+tmp[1].im = tab[ 9] * tmp[1].im;
+tmp[2].re = tab[10] * tmp[2].re;
+tmp[2].im = tab[10] * tmp[2].im;
+out[1*stride].re = tmp[0].re - tmp[2].re + tmp[1].re;
+out[1*stride].im = tmp[0].im - tmp[2].im - tmp[1].im;
+out[2*stride].re = tmp[0].re - tmp[2].re - tmp[1].re;
+out[2*stride].im = tmp[0].im - tmp[2].im + tmp[1].im;
#endif
}
@@ -208,16 +209,17 @@ static av_always_inline void fft3(TXComplex *out,
TXComplex *in,
static av_always_inline void NAME(TXComplex *out, TXComplex *in,\
ptrdiff_t stride) \
{ \
-TXComplex z0[4], t[6]; \
+TXComplex dc, z0[4], t[6]; \
const TXSample *tab = TX_TAB(ff_tx_tab_53); \
\
+dc = in[0]; \
BF(t[1].im, t[0].re, in[1].re, in[4].re); \
BF(t[1].re, t[0].im, in[1].im, in[4].im); \
BF(t[3].im, t[2].re, in[2].re, in[3].re); \
BF(t[3].re, t[2].im, in[2].im, in[3].im); \
\
-out[D0*stride].re = in[0].re + t[0].re + t[2].re; \
-out[D0*stride].im = in[0].im + t[0].im + t[2].im; \
+out[D0*stride].re = dc.re + t[0].re + t[2].re; \
+out[D0*stride].im = dc.im + t[0].im + t[2].im; \
\
SMUL(t[4].re, t[0].re, tab[0], tab[2], t[2].re, t[0].re); \
SMUL(t[4].im, t[0].im, tab[0], tab[2], t[2].im, t[0].im); \
@@ -229,14 +231,14 @@ static av_always_inline void NAME(TXComplex *out,
TXComplex *in,\
BF(z0[2].re, z0[1].re, t[4].re, t[5].re); \
BF(z0[2].im, z0[1].im, t[4].im, t[5].im); \
\
-out[D1*stride].re =
[FFmpeg-cvslog] lavu/tx: list all odd-length FFT factors as regular codelets
ffmpeg | branch: master | Lynne | Sat Sep 24 06:50:17 2022
+0200| [e8a9b7b29877db9e3887562007df7a53325b67d1] | committer: Lynne
lavu/tx: list all odd-length FFT factors as regular codelets
Allows them to be picked just like any other transform.
> http://git.videolan.org/gitweb.cgi/ffmpeg.git/?a=commit;h=e8a9b7b29877db9e3887562007df7a53325b67d1
---
libavutil/tx_template.c | 88 +
1 file changed, 88 insertions(+)
diff --git a/libavutil/tx_template.c b/libavutil/tx_template.c
index b547800447..d72281f09c 100644
--- a/libavutil/tx_template.c
+++ b/libavutil/tx_template.c
@@ -472,6 +472,81 @@ static av_always_inline void fft15(TXComplex *out,
TXComplex *in,
fft5_m3(out, tmp + 10, stride);
}
+static av_cold int TX_NAME(ff_tx_fft_factor_init)(AVTXContext *s,
+ const FFTXCodelet *cd,
+ uint64_t flags,
+ FFTXCodeletOptions *opts,
+ int len, int inv,
+ const void *scale)
+{
+TX_TAB(ff_tx_init_tabs)(len);
+
+if (flags & FF_TX_PRESHUFFLE) {
+s->map = av_malloc(len*sizeof(s->map));
+s->map[0] = 0; /* DC is always at the start */
+if (inv) /* Reversing the ACs flips the transform direction */
+for (int i = 1; i < len; i++)
+s->map[i] = len - i;
+else
+for (int i = 1; i < len; i++)
+s->map[i] = i;
+}
+
+/* Our 15-point transform is actually a 5x3 PFA, so embed its input map. */
+if (len == 15) {
+int tmp[15];
+memcpy(tmp, s->map, 15*sizeof(*tmp));
+for (int i = 0; i < 5; i++) {
+for (int j = 0; j < 3; j++)
+s->map[i*3 + j] = tmp[(i*3 + j*5) % 15];
+}
+}
+
+return 0;
+}
+
+#define DECL_FACTOR_S(n)
\
+static void TX_NAME(ff_tx_fft##n)(AVTXContext *s, void *dst,
\
+ void *src, ptrdiff_t stride)
\
+{
\
+fft##n((TXComplex *)dst, (TXComplex *)src, stride / sizeof(TXComplex));
\
+}
\
+static const FFTXCodelet TX_NAME(ff_tx_fft##n##_ns_def) = {
\
+.name = TX_NAME_STR("fft" #n "_ns"),
\
+.function = TX_NAME(ff_tx_fft##n),
\
+.type = TX_TYPE(FFT),
\
+.flags = AV_TX_INPLACE | FF_TX_OUT_OF_PLACE |
\
+ AV_TX_UNALIGNED | FF_TX_PRESHUFFLE,
\
+.factors[0] = n,
\
+.min_len= n,
\
+.max_len= n,
\
+.init = TX_NAME(ff_tx_fft_factor_init),
\
+.cpu_flags = FF_TX_CPU_FLAGS_ALL,
\
+.prio = FF_TX_PRIO_BASE,
\
+};
+
+#define DECL_FACTOR_F(n)
\
+DECL_FACTOR_S(n)
\
+static const FFTXCodelet TX_NAME(ff_tx_fft##n##_fwd_def) = {
\
+.name = TX_NAME_STR("fft" #n "_fwd"),
\
+.function = TX_NAME(ff_tx_fft##n),
\
+.type = TX_TYPE(FFT),
\
+.flags = AV_TX_INPLACE | FF_TX_OUT_OF_PLACE |
\
+ AV_TX_UNALIGNED | FF_TX_FORWARD_ONLY,
\
+.factors[0] = n,
\
+.min_len= n,
\
+.max_len= n,
\
+.init = TX_NAME(ff_tx_fft_factor_init),
\
+.cpu_flags = FF_TX_CPU_FLAGS_ALL,
\
+.prio = FF_TX_PRIO_BASE,
\
+};
+
+DECL_FACTOR_F(3)
+DECL_FACTOR_F(5)
+DECL_FACTOR_F(7)
+DECL_FACTOR_F(9)
+DECL_FACTOR_S(15)
+
#define BUTTERFLIES(a0, a1, a2, a3)\
do { \
r0=a0.re; \
@@ -1483,6 +1558,19 @@ const FFTXCodelet * const TX_NAME(ff_tx_codelet_list)[]
= {
_NAME(ff_tx_fft65536_ns_def),
[FFmpeg-cvslog] fate/aacenc: increase tolerance for ln-128k test
ffmpeg | branch: master | Lynne | Thu Nov 17 12:39:47 2022 +0100| [d556f6fa9bda60af411415a6aaa9fa26fe2ebf45] | committer: Lynne fate/aacenc: increase tolerance for ln-128k test The encoder is sensitive to changes in precision, and its test target was a compromise. It was already close to failing on x87 FPUs. ff_mdct_init used double precision entirely from the scale to computing the MDCT exp tables. av_tx_init uses single-precision for the scale, with a small input change which was enough to tip the test into failing on x87 FPUs. Increase the fuzz factor in line with other AAC encoder tests to fix. > http://git.videolan.org/gitweb.cgi/ffmpeg.git/?a=commit;h=d556f6fa9bda60af411415a6aaa9fa26fe2ebf45 --- tests/fate/aac.mak | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/tests/fate/aac.mak b/tests/fate/aac.mak index 1743428f54..4f8d1cdcea 100644 --- a/tests/fate/aac.mak +++ b/tests/fate/aac.mak @@ -174,7 +174,7 @@ fate-aac-ln-encode-128k: REF = $(SAMPLES)/audio-reference/luckynight_2ch_44kHz_s fate-aac-ln-encode-128k: CMP_SHIFT = -4096 fate-aac-ln-encode-128k: CMP_TARGET = 622 fate-aac-ln-encode-128k: SIZE_TOLERANCE = 3560 -fate-aac-ln-encode-128k: FUZZ = 5 +fate-aac-ln-encode-128k: FUZZ = 10 FATE_AAC_ENCODE += fate-aac-pns-encode fate-aac-pns-encode: CMD = enc_dec_pcm adts wav s16le $(TARGET_SAMPLES)/audio-reference/luckynight_2ch_44kHz_s16.wav -c:a aac -aac_coder fast -aac_pns 1 -aac_is 0 -aac_ms 0 -aac_tns 0 -b:a 128k -cutoff 22050 -fflags +bitexact -flags +bitexact ___ ffmpeg-cvslog mailing list ffmpeg-cvslog@ffmpeg.org https://ffmpeg.org/mailman/listinfo/ffmpeg-cvslog To unsubscribe, visit link above, or email ffmpeg-cvslog-requ...@ffmpeg.org with subject "unsubscribe".
