PR #24380 opened by mskpluk
URL: https://code.ffmpeg.org/FFmpeg/FFmpeg/pulls/24380
Patch URL: https://code.ffmpeg.org/FFmpeg/FFmpeg/pulls/24380.patch
Commit 87b7e578ec ("avcodec/amfenc: add encoder average QP stats") enables
`AMF_VIDEO_ENCODER_*_STATISTICS_FEEDBACK` on every submitted surface. The AMF
property defaults to false and asking the encoder to collect per-frame
statistics is not free: the hardware has to gather the average QP over all CTBs
of every picture.
On a Radeon RX 9070 XT (Navi48, VCN5) driven through the open source Mesa/RADV
Vulkan backend the cost is roughly 48% additional wall time per frame, for a bit
identical bitstream. Measured with a minimal build (`--enable-amf`, AMF headers
v1.5.2), 1080p23.976, 4353 frames, `hevc_amf`, `-rc cqp -qp_i 24 -qp_p 24`,
software decode. Median of repeated runs on an otherwise idle machine:
| build | preset | wall | fps |
|---|---|---|---|
| n8.1.2 | `-quality speed` | 9.27 s | 469.6 |
| n9.0.1 | `-quality speed` | 13.78 s | 315.9 |
| n9.0.1 + this patch | `-quality speed` | **9.27 s** | **469.6** |
| n9.0.1 + this patch, `-qp_stats 1` | `-quality speed` | 13.78 s | 315.9 |
| n8.1.2 | `-quality quality` | 16.29 s | 267.2 |
| n9.0.1 | `-quality quality` | 24.05 s | 181.0 |
| n9.0.1 + this patch | `-quality quality` | **16.29 s** | **267.2** |
| n9.0.1 + this patch, `-qp_stats 1` | `-quality quality` | 23.80 s | 182.9 |
The SHA256 of the encoded video stream is identical across all runs, with and
without the feedback enabled. Per-process DRM fdinfo shows the VCN encode engine
going from 69% to 92% busy at the quality preset once the feedback is disabled —
the engine was being starved, not doing more work.
Other encoders that call `ff_encode_add_stats_side_data()` (libx264, libx265,
libsvtav1, ...) get the quantiser for free from data they already have, so
reporting it unconditionally costs them nothing. That is not the case here,
which is why this is made opt-in rather than always on.
Add a `qp_stats` AVOption, defaulting to false, restoring pre-9.0 throughput by
default. Users who want the average QP reported in the ffmpeg CLI progress line,
or as `AV_PKT_DATA_QUALITY_STATS` side data, can enable it with `-qp_stats 1`.
From 29806a702e24d0cfa7343811617c956cab519e6a Mon Sep 17 00:00:00 2001
From: Marcin Krycki <[email protected]>
Date: Sun, 6 Sep 2026 14:36:27 +0100
Subject: [PATCH] avcodec/amfenc: make encoder statistics feedback opt-in
Commit 87b7e578ec ("avcodec/amfenc: add encoder average QP stats") enables
AMF_VIDEO_ENCODER_*_STATISTICS_FEEDBACK on every submitted surface. The AMF
property defaults to false and asking the encoder to collect per-frame
statistics is not free: the hardware has to gather the average QP over all
CTBs of every picture.
On a Radeon RX 9070 XT (Navi48, VCN5) driven through the open source
Mesa/RADV Vulkan backend the cost is roughly 48% additional wall time per
frame, for a bit identical bitstream. Measured with a minimal build
(--enable-amf, AMF headers v1.5.2), 1080p23.976, 4353 frames, hevc_amf,
-rc cqp -qp_i 24 -qp_p 24, software decode. Median of repeated runs on an
otherwise idle machine:
n8.1.2 -quality speed 9.27 s 469.6 fps
n9.0.1 -quality speed 13.78 s 315.9 fps
n9.0.1 + this patch -quality speed 9.27 s 469.6 fps
n9.0.1 + this patch, -qp_stats 1 13.78 s 315.9 fps
n8.1.2 -quality quality 16.29 s 267.2 fps
n9.0.1 -quality quality 24.05 s 181.0 fps
n9.0.1 + this patch -quality quality 16.29 s 267.2 fps
n9.0.1 + this patch, -qp_stats 1 23.80 s 182.9 fps
The SHA256 of the encoded video stream is identical across all runs, with
and without the feedback enabled. Per-process DRM fdinfo shows the VCN
encode engine going from 69% to 92% busy at the quality preset once the
feedback is disabled - the engine was being starved, not doing more work.
Other encoders that call ff_encode_add_stats_side_data() (libx264, libx265,
libsvtav1, ...) get the quantiser for free from data they already have, so
reporting it unconditionally costs them nothing. That is not the case here,
which is why this is made opt-in rather than always on.
Add a qp_stats AVOption, defaulting to false, restoring pre-9.0 throughput
by default. Users who want the average QP reported in the ffmpeg CLI
progress line, or as AV_PKT_DATA_QUALITY_STATS side data, can enable it
with -qp_stats 1.
---
libavcodec/amfenc.c | 14 ++++++++++----
libavcodec/amfenc.h | 1 +
libavcodec/amfenc_av1.c | 1 +
libavcodec/amfenc_h264.c | 1 +
libavcodec/amfenc_hevc.c | 1 +
5 files changed, 14 insertions(+), 4 deletions(-)
diff --git a/libavcodec/amfenc.c b/libavcodec/amfenc.c
index 3ffca750d8..20bc22d98c 100644
--- a/libavcodec/amfenc.c
+++ b/libavcodec/amfenc.c
@@ -250,7 +250,7 @@ static int amf_copy_buffer(AVCodecContext *avctx, AVPacket
*pkt, AMFBuffer *buff
break;
}
- if (average_qp >= 0) {
+ if (ctx->qp_stats && average_qp >= 0) {
ff_encode_add_stats_side_data(pkt, average_qp * FF_QP2LAMBDA, NULL, 0,
pict_type);
}
@@ -475,7 +475,9 @@ static int amf_submit_frame(AVCodecContext *avctx, AVFrame
*frame, AMFSurface
switch (avctx->codec->id) {
case AV_CODEC_ID_H264:
- AMF_ASSIGN_PROPERTY_BOOL(res, surface,
AMF_VIDEO_ENCODER_STATISTICS_FEEDBACK, 1);
+ if (ctx->qp_stats) {
+ AMF_ASSIGN_PROPERTY_BOOL(res, surface,
AMF_VIDEO_ENCODER_STATISTICS_FEEDBACK, 1);
+ }
AMF_ASSIGN_PROPERTY_INT64(res, surface, AMF_VIDEO_ENCODER_INSERT_AUD,
!!ctx->aud);
switch (frame->pict_type) {
case AV_PICTURE_TYPE_I:
@@ -496,7 +498,9 @@ static int amf_submit_frame(AVCodecContext *avctx, AVFrame
*frame, AMFSurface
}
break;
case AV_CODEC_ID_HEVC:
- AMF_ASSIGN_PROPERTY_BOOL(res, surface,
AMF_VIDEO_ENCODER_HEVC_STATISTICS_FEEDBACK, 1);
+ if (ctx->qp_stats) {
+ AMF_ASSIGN_PROPERTY_BOOL(res, surface,
AMF_VIDEO_ENCODER_HEVC_STATISTICS_FEEDBACK, 1);
+ }
AMF_ASSIGN_PROPERTY_INT64(res, surface,
AMF_VIDEO_ENCODER_HEVC_INSERT_AUD, !!ctx->aud);
switch (frame->pict_type) {
case AV_PICTURE_TYPE_I:
@@ -513,7 +517,9 @@ static int amf_submit_frame(AVCodecContext *avctx, AVFrame
*frame, AMFSurface
}
break;
case AV_CODEC_ID_AV1:
- AMF_ASSIGN_PROPERTY_BOOL(res, surface,
AMF_VIDEO_ENCODER_AV1_STATISTICS_FEEDBACK, 1);
+ if (ctx->qp_stats) {
+ AMF_ASSIGN_PROPERTY_BOOL(res, surface,
AMF_VIDEO_ENCODER_AV1_STATISTICS_FEEDBACK, 1);
+ }
if (frame->pict_type == AV_PICTURE_TYPE_I) {
if (ctx->forced_idr) {
AMF_ASSIGN_PROPERTY_INT64(res, surface,
AMF_VIDEO_ENCODER_AV1_FORCE_INSERT_SEQUENCE_HEADER, 1);
diff --git a/libavcodec/amfenc.h b/libavcodec/amfenc.h
index 1571541b9b..6a380208d1 100644
--- a/libavcodec/amfenc.h
+++ b/libavcodec/amfenc.h
@@ -100,6 +100,7 @@ typedef struct AMFEncoderContext {
int qvbr_quality_level;
int hw_high_motion_quality_boost;
int forced_idr;
+ int qp_stats;
// HEVC - specific options
diff --git a/libavcodec/amfenc_av1.c b/libavcodec/amfenc_av1.c
index af20b5d04d..a572d439a7 100644
--- a/libavcodec/amfenc_av1.c
+++ b/libavcodec/amfenc_av1.c
@@ -130,6 +130,7 @@ static const AVOption options[] = {
{ "caq", "context adaptive quantization", 0,
AV_OPT_TYPE_CONST, {.i64 = AMF_VIDEO_ENCODER_AV1_AQ_MODE_CAQ }, 0, 0, VE, .unit
= "adaptive_quantisation_mode" },
{ "forced_idr", "Force I frames to be IDR frames",
OFFSET(forced_idr), AV_OPT_TYPE_BOOL, { .i64 = 0 }, 0, 1, VE },
+ { "qp_stats", "Report average QP per frame as packet side data
(reduces throughput)", OFFSET(qp_stats), AV_OPT_TYPE_BOOL, { .i64 = 0 }, 0, 1,
VE },
{ "align", "alignment mode",
OFFSET(align), AV_OPT_TYPE_INT, {.i64 =
AMF_VIDEO_ENCODER_AV1_ALIGNMENT_MODE_NO_RESTRICTIONS },
AMF_VIDEO_ENCODER_AV1_ALIGNMENT_MODE_64X16_ONLY,
AMF_VIDEO_ENCODER_AV1_ALIGNMENT_MODE_NO_RESTRICTIONS, VE, .unit = "align" },
{ "64x16", "", 0, AV_OPT_TYPE_CONST, {.i64 =
AMF_VIDEO_ENCODER_AV1_ALIGNMENT_MODE_64X16_ONLY }, 0, 0, VE,
.unit = "align" },
diff --git a/libavcodec/amfenc_h264.c b/libavcodec/amfenc_h264.c
index 650a9bc7e9..214fac6909 100644
--- a/libavcodec/amfenc_h264.c
+++ b/libavcodec/amfenc_h264.c
@@ -140,6 +140,7 @@ static const AVOption options[] = {
{ "me_quarter_pel", "Enable ME Quarter Pixel",
OFFSET(me_quarter_pel),AV_OPT_TYPE_BOOL, { .i64 = -1 }, -1, 1, VE },
{ "forced_idr", "Force I frames to be IDR frames",
OFFSET(forced_idr) , AV_OPT_TYPE_BOOL, { .i64 = 0 }, 0, 1, VE },
+ { "qp_stats", "Report average QP per frame as packet side data
(reduces throughput)", OFFSET(qp_stats), AV_OPT_TYPE_BOOL, { .i64 = 0 }, 0, 1,
VE },
{ "aud", "Inserts AU Delimiter NAL unit", OFFSET(aud)
, AV_OPT_TYPE_BOOL, { .i64 = -1 }, -1, 1, VE },
{ "smart_access_video", "Enable Smart Access Video to enhance
performance by utilizing both APU and dGPU memory access",
OFFSET(smart_access_video), AV_OPT_TYPE_BOOL, {.i64 = -1 }, -1, 1, VE},
diff --git a/libavcodec/amfenc_hevc.c b/libavcodec/amfenc_hevc.c
index 6fa20d172e..4e280fcf15 100644
--- a/libavcodec/amfenc_hevc.c
+++ b/libavcodec/amfenc_hevc.c
@@ -111,6 +111,7 @@ static const AVOption options[] = {
{ "me_quarter_pel", "Enable ME Quarter Pixel ",
OFFSET(me_quarter_pel),AV_OPT_TYPE_BOOL,{ .i64 = -1 }, -1, 1, VE },
{ "forced_idr", "Force I frames to be IDR frames",
OFFSET(forced_idr) ,AV_OPT_TYPE_BOOL,{ .i64 = 0 }, 0, 1, VE },
+ { "qp_stats", "Report average QP per frame as packet side data
(reduces throughput)", OFFSET(qp_stats), AV_OPT_TYPE_BOOL, { .i64 = 0 }, 0, 1,
VE },
{ "aud", "Inserts AU Delimiter NAL unit",
OFFSET(aud) ,AV_OPT_TYPE_BOOL,{ .i64 = -1 }, -1, 1, VE },
{ "smart_access_video", "Enable Smart Access Video to enhance
performance by utilizing both APU and dGPU memory access",
OFFSET(smart_access_video), AV_OPT_TYPE_BOOL, {.i64 = -1 }, -1, 1, VE},
--
2.52.0
_______________________________________________
ffmpeg-devel mailing list -- [email protected]
To unsubscribe send an email to [email protected]
