ffmpeg_vpp_amf_d3d11_ext

Extended vpp_amf filter with D3D11-based deinterlace and crop for ffmpeg
(github pages)

1. Overview of This Extension

• Add D3D11-based GPU zero-copy deinterlace and crop functionality to ffmpeg’s vpp_amf filter
• Includes three patch files for:
  • vf_amf_common.c
  • vf_amf_common.h
  • vf_vpp_amf.c
• Includes a Makefile for building on Windows (MSys2 + mingw64)
• When deint or crop(cx,cy,cw,ch) is specified together with the original w / h options of vpp_amf,
  → the filter performs centering and padding using w × h as the canvas

---

2. Basic Command Examples

2-1. When the AMF decoder (-hwaccel amf) supports the codec

ffmpeg -hwaccel amf -i "input.m2ts" \
  -c:v hevc_amf \
  -vf "vpp_amf=w=1920:h=720:deint=fast:deint_frame_format=tff:format=nv12:cx=0:cy=139:cw=1920:ch=802" \
  -c:a copy -f mp4 "output.mp4"

2-2. When the AMF decoder does NOT support the codec

ffmpeg -i "input.m2ts" \
  -c:v hevc_amf \
  -vf "format=nv12, vpp_amf=w=1920:h=720:deint=fast:deint_frame_format=tff:format=nv12:cx=0:cy=139:cw=1920:ch=802" \
  -c:a copy -f mp4 "output.mp4"

---

3. Notes on Options (Based on AMD Radeon RX 7600 Behavior)

3-1. Hardware decoding with AMF may cause playback stutter

• Occurs when the video has small resolution and AMF’s automatic frame-format detection mixes tff / bff / progressive
• In such cases, disable hardware decoding and specify format, deint, and deint_frame_format in the filter (see 2-2)

---

3-2. deint=adaptive may produce jagged frames

• Very common in 2D animation
• Occasional in 3D animation with small source resolution
• Rare in live-action
  → Use deint=fast + deint_frame_format for stable results

---

3-3. When AMF does not support the codec

• Decoding falls back to CPU
• The vf input format becomes the same as the input file
• This filter expects nv12 or p010, so it fails
  → Specify format=nv12 explicitly (see 2-2)

AMF decoder support as of 2026.1:

DEFINE_AMF_DECODER(h264, H264, ...)
DEFINE_AMF_DECODER(hevc, HEVC, ...)
DEFINE_AMF_DECODER(vp9,  VP9,  ...)
DEFINE_AMF_DECODER(av1,  AV1,  ...)

---

3-4. Automatic per-frame detection of tff/bff/progressive

• Often increases jitter depending on the source
  → Specify deint and deint_frame_format together

---

3-5. Some videos are detected as tff/bff by ffprobe but are actually progressive

• Applying deinterlace based on incorrect detection causes severe jitter
  → Use the detection command below to confirm before converting

---

4. Useful Commands for Inspecting Source Videos

4-1. Progressive / Interlaced detection

ffmpeg -i test.vob -filter:v idet -frames:v 300 -an -f null -

• Check the Multi-frame detection output
• If tff/bff counts are significant → treat as interlaced

---

4-2. Check codec

ffprobe -select_streams v:0 -show_entries stream=codec_name test.vob

---

4-3. Crop size detection

ffmpeg -i "test.mp4" -vf cropdetect -f null -

• Detects black borders every few frames
• If borders appear/disappear depending on scene, run multiple times and observe the values

---

5. Project Structure

ffmpeg_vpp_amf_d3d11_ext
+- patches
|  +- original
|  |  +- vf_amf_common.c
|  |  +- vf_amf_common.h
|  |  +- vf_vpp_amf.c
|  +- patched
|  |  +- vf_amf_common.c
|  |  +- vf_amf_common.h
|  |  +- vf_vpp_amf.c
|  +- vf_amf_common_c.patch
|  +- vf_amf_common_h.patch
|  +- vf_vpp_amf_c.patch
+- Makefile

---

6. Makefile Overview

6-1. Preparation

Edit the following line:

ROOT := /d/Data/ffmpeg_vpp_amf_d3d11_ext
INSTALL_TO := /d/utils/ffmpeg-self

6-2. Build

cd ffmpeg_vpp_amf_d3d11_ext
make

6-3. What the Makefile does

• Installs or builds required libraries (mingw packages or upstream sources)
• Runs ffmpeg’s configure
• Applies patches
• Builds ffmpeg
• Verifies that the resulting binaries do not depend on MSys/mingw64 runtime DLLs

---

7. Tips

7-1. Behavior of the reset_sar option

The following code inside amf_init_filter_config() does not appear to behave as originally intended:

if (ctx->reset_sar && inlink->sample_aspect_ratio.num)
    w_adj = (double) inlink->sample_aspect_ratio.num / inlink->sample_aspect_ratio.den;

ff_scale_adjust_dimensions(inlink, &ctx->width, &ctx->height,
                           ctx->force_original_aspect_ratio, ctx->force_divisible_by, w_adj);

Regarding inlink->sample_aspect_ratio.num:

• When w=-1 or w=-2 is specified, inlink->sample_aspect_ratio.num is non‑zero.
• However, when w is omitted (implicitly w=iw), explicitly set to a fixed value (e.g., w=1280:h=720), or given as an expression (e.g., w=ih*sar), inlink->sample_aspect_ratio.num becomes zero.

As a result, ff_scale_adjust_dimensions() is executed with w_adj left at 1.0.

This behavior appears to be unintended and is problematic for my extension. Since inlink->sample_aspect_ratio.num is rarely zero in actual frames passed to the filter, this may be due to a change in FFmpeg’s filter framework or a bug in the current version.

If this is indeed a bug, it may be fixed in a future stable release. For now, I will work around it by selecting command‑line parameters according to the specific objective.

7-2. Relationship between FFmpeg command-line parameters and the reset_sar option

Workaround notes for observed FFmpeg expression‑evaluation behavior (as of 2026‑04‑15):

The following patterns have been confirmed regarding w (width) and reset_sar. Command‑line parameters should be chosen according to these rules:

1. w unspecified & reset_sar=false
   w is evaluated as iw; output SAR matches input SAR.

2. w unspecified & reset_sar=true
   w is evaluated as iw; output SAR is forced to 1.
   Note: This causes distortion if the input SAR is not 1:1.

3. w=-2 & reset_sar=false
   w is evaluated as h * dar / sar; output SAR matches input SAR.

4. w=-2 & reset_sar=true
   w is evaluated as h * dar / sar, then scaled again by w * sar inside vpp_amf; output SAR is forced to 1.

5. w=value/expr & reset_sar=false
   Output SAR matches input SAR. The input value should represent Display Width / SAR.

6. w=value/expr & reset_sar=true
   Output SAR is forced to 1. The input value should be the actual Display Width.

---

8. Disclaimer / Support Policy

• This repository is provided “as is.”
• Bug reports, questions, and feature requests will not be handled.
• No contact information is provided.
• Use, modify, and redistribute at your own risk.

---

9. License

This patch follows the ffmpeg project’s license (GPL/LGPL).
Refer to the ffmpeg source tree for full license details.

This site is open source. Improve this page.