build: add amd hip sdk support for Windows #21595

Steven Xiao · 2026-01-26T18:51:19Z

Steven Xiao commented

2026-01-26 18:51:19 +00:00

Add support for AMD's HIP (Heterogeneous-compute Interface for Portability)
SDK on Windows, enabling GPU-accelerated video processing on AMD GPUs.

This commit introduces the build infrastructure for HIP, similar to the
existing CUDA support. HIP allows writing portable GPU code that can run
on AMD GPUs with RDNA2 architecture or newer.

Build Instructions (MSYS2 MinGW):

Install MSYS2 from https://www.msys2.org/
Install AMD HIP SDK from:
https://www.amd.com/en/developer/resources/rocm-hub/hip-sdk.html

Default installation path: C:\Program Files\AMD\ROCm<version>
Ensure hipcc is available: C:\Program Files\AMD\ROCm<version>\bin\hipcc.bat
Add HIP SDK to PATH (in MSYS2 terminal):
export PATH="/c/Program Files/AMD/ROCm/6.4/bin:$PATH"

(Adjust version number as needed)
Configure and Build:

./configure --enable-hip --enable-nonfree
make -j

Note: The default target architecture is gfx900. To target a specific
GPU architecture, use --hipccflags="--offload-arch=" where
is your GPU's architecture code (e.g., gfx1030 for RDNA2, gfx1100 for RDNA3).

External Dependencies:
Header files:
- hip/hip_runtime.h (from HIP SDK, or compat/hip/hip_runtime.h)
Tools:
- hipcc (HIP compiler, from AMD HIP SDK)
Runtime:
- AMD GPU driver with ROCm/HIP support
- amdhip64.dll (HIP runtime library)

Notes:

HIP enables GPU-accelerated code that can run on AMD GPUs.
The --enable-hip option enables the hipcc compiler for building
GPU kernels (.hip files) into GPU code objects (.hsaco files).
HIP support provides infrastructure for GPU-accelerated filters.
Actual GPU filters must implement HIP kernels similar to existing
CUDA filters (scale_cuda, yadif_cuda, etc.).

Add support for AMD's HIP (Heterogeneous-compute Interface for Portability) SDK on Windows, enabling GPU-accelerated video processing on AMD GPUs. This commit introduces the build infrastructure for HIP, similar to the existing CUDA support. HIP allows writing portable GPU code that can run on AMD GPUs with RDNA2 architecture or newer. Build Instructions (MSYS2 MinGW): 1. Install MSYS2 from https://www.msys2.org/ 2. Install AMD HIP SDK from: https://www.amd.com/en/developer/resources/rocm-hub/hip-sdk.html Default installation path: C:\Program Files\AMD\ROCm\<version>\ Ensure hipcc is available: C:\Program Files\AMD\ROCm\<version>\bin\hipcc.bat 3. Add HIP SDK to PATH (in MSYS2 terminal): export PATH="/c/Program Files/AMD/ROCm/6.4/bin:$PATH" (Adjust version number as needed) 4. Configure and Build: ./configure --enable-hip --enable-nonfree make -j Note: The default target architecture is gfx900. To target a specific GPU architecture, use --hipccflags="--offload-arch=<arch>" where <arch> is your GPU's architecture code (e.g., gfx1030 for RDNA2, gfx1100 for RDNA3). External Dependencies: Header files: - hip/hip_runtime.h (from HIP SDK, or compat/hip/hip_runtime.h) Tools: - hipcc (HIP compiler, from AMD HIP SDK) Runtime: - AMD GPU driver with ROCm/HIP support - amdhip64.dll (HIP runtime library) Notes: - HIP enables GPU-accelerated code that can run on AMD GPUs. - The --enable-hip option enables the hipcc compiler for building GPU kernels (.hip files) into GPU code objects (.hsaco files). - HIP support provides infrastructure for GPU-accelerated filters. Actual GPU filters must implement HIP kernels similar to existing CUDA filters (scale_cuda, yadif_cuda, etc.).

Steven Xiao requested review from Gyan Doshi

2026-01-26 18:51:20 +00:00

ffmpeg-devel added the

new

label

2026-01-26 18:51:30 +00:00

Steven Xiao force-pushed build/add-amd-hip-sdk-support-for-windows from 35ffe9f2a5

Autolabel / Labeler (pull_request_target) Has been skipped

Details

Lint / Pre-Commit (pull_request) Failing after 29s

Details

Test / Fate (linux-aarch64, static, 64 bit) (pull_request) Has been cancelled

Details

Test / Fate (linux-amd64, static, 32 bit) (pull_request) Has been cancelled

Details

Test / Fate (Full, wine) (pull_request) Has been cancelled

Details

Test / Fate (linux-amd64, shared, 64 bit) (pull_request) Has been cancelled

Details

to 8cc0583eb9

Lint / Pre-Commit (pull_request) Failing after 27s

Details

Autolabel / Labeler (pull_request_target) Has been skipped

Details

Test / Fate (Full, wine) (pull_request) Has been cancelled

Details

Test / Fate (linux-amd64, static, 32 bit) (pull_request) Has been cancelled

Details

Test / Fate (linux-amd64, shared, 64 bit) (pull_request) Has been cancelled

Details

Test / Fate (linux-aarch64, static, 64 bit) (pull_request) Has been cancelled

Details

2026-01-26 18:56:47 +00:00

Compare

Timo Rothenpieler commented

2026-01-26 19:04:17 +00:00

Shader Compression unified both cuda and vulkan ones already.
So you don't need to re-add yet another one for this.

Also, this is at least missing the vpath handling in the makefiles, so it most likely breaks out-of-tree builds.

Shader Compression unified both cuda and vulkan ones already. So you don't need to re-add yet another one for this. Also, this is at least missing the vpath handling in the makefiles, so it most likely breaks out-of-tree builds.

ffmpeg-devel removed the

new

label

2026-01-26 19:04:28 +00:00

Steven Xiao force-pushed build/add-amd-hip-sdk-support-for-windows from 8cc0583eb9

Lint / Pre-Commit (pull_request) Failing after 27s

Details

Autolabel / Labeler (pull_request_target) Has been skipped

Details

Test / Fate (Full, wine) (pull_request) Has been cancelled

Details

Test / Fate (linux-amd64, static, 32 bit) (pull_request) Has been cancelled

Details

Test / Fate (linux-amd64, shared, 64 bit) (pull_request) Has been cancelled

Details

Test / Fate (linux-aarch64, static, 64 bit) (pull_request) Has been cancelled

Details

to 295b6c2cc9

Autolabel / Labeler (pull_request_target) Successful in 10s

Details

Lint / Pre-Commit (pull_request) Successful in 24s

Details

Test / Fate (linux-amd64, static, 32 bit) (pull_request) Successful in 12m9s

Details

Test / Fate (linux-aarch64, static, 64 bit) (pull_request) Successful in 11m49s

Details

Test / Fate (linux-amd64, shared, 64 bit) (pull_request) Successful in 14m20s

Details

Test / Fate (Full, wine) (pull_request) Successful in 41m7s

Details

2026-01-26 19:08:32 +00:00

Compare

Steven Xiao force-pushed build/add-amd-hip-sdk-support-for-windows from a706bb5312

Autolabel / Labeler (pull_request_target) Successful in 10s

Details

Lint / Pre-Commit (pull_request) Successful in 27s

Details

Test / Fate (linux-aarch64, static, 64 bit) (pull_request) Has been cancelled

Details

Test / Fate (linux-amd64, shared, 64 bit) (pull_request) Has been cancelled

Details

Test / Fate (linux-amd64, static, 32 bit) (pull_request) Has been cancelled

Details

Test / Fate (Full, wine) (pull_request) Has been cancelled

Details

to 253079cf93

Autolabel / Labeler (pull_request_target) Successful in 8s

Details

Lint / Pre-Commit (pull_request) Successful in 24s

Details

Test / Fate (linux-aarch64, static, 64 bit) (pull_request) Successful in 11m47s

Details

Test / Fate (linux-amd64, static, 32 bit) (pull_request) Has been cancelled

Details

Test / Fate (linux-amd64, shared, 64 bit) (pull_request) Has been cancelled

Details

Test / Fate (Full, wine) (pull_request) Has been cancelled

Details

2026-01-26 20:04:22 +00:00

Compare

Steven Xiao force-pushed build/add-amd-hip-sdk-support-for-windows from 253079cf93

Autolabel / Labeler (pull_request_target) Successful in 8s

Details

Lint / Pre-Commit (pull_request) Successful in 24s

Details

Test / Fate (linux-aarch64, static, 64 bit) (pull_request) Successful in 11m47s

Details

Test / Fate (linux-amd64, static, 32 bit) (pull_request) Has been cancelled

Details

Test / Fate (linux-amd64, shared, 64 bit) (pull_request) Has been cancelled

Details

Test / Fate (Full, wine) (pull_request) Has been cancelled

Details

to 816ff702fa

Lint / Pre-Commit (pull_request) Successful in 24s

Details

Test / Fate (linux-aarch64, static, 64 bit) (pull_request) Successful in 11m55s

Details

Test / Fate (linux-amd64, static, 32 bit) (pull_request) Successful in 12m4s

Details

Test / Fate (linux-amd64, shared, 64 bit) (pull_request) Successful in 14m53s

Details

Test / Fate (Full, wine) (pull_request) Successful in 41m25s

Details

Autolabel / Labeler (pull_request_target) Has been skipped

Details

2026-01-26 20:17:19 +00:00

Compare

Steven Xiao commented

2026-01-26 21:05:49 +00:00

@BtbN wrote in #21595 (comment):

Shader Compression unified both cuda and vulkan ones already. So you don't need to re-add yet another one for this.

Also, this is at least missing the vpath handling in the makefiles, so it most likely breaks out-of-tree builds.

@BtbN Changed from CONFIG_HSACO_COMPRESSION to CONFIG_SHADER_COMPRESSION in ffbuild/common.mak to align with cuda and vulkan.

Also, add vpath handling for out-of-tree builds in Makefile.

@BtbN wrote in https://code.ffmpeg.org/FFmpeg/FFmpeg/pulls/21595#issuecomment-23724: > Shader Compression unified both cuda and vulkan ones already. So you don't need to re-add yet another one for this. > > Also, this is at least missing the vpath handling in the makefiles, so it most likely breaks out-of-tree builds. @BtbN Changed from CONFIG_HSACO_COMPRESSION to CONFIG_SHADER_COMPRESSION in ffbuild/common.mak to align with cuda and vulkan. Also, add vpath handling for out-of-tree builds in Makefile.

Gyan Doshi reviewed

2026-01-27 04:49:31 +00:00

doc/build_with_hip_sdk.txt Outdated

					
				@ -0,0 +1,91 @@

				Building FFmpeg with AMD HIP SDK Support (Windows)

This entire file is redundant with the addition in general_contents.texi. Should be removed.

The file "general_contents.texi" was removed.

For posterity's sake, actually the file doc/build_with_hip_sdk.txt was removed, not doc/general_contents.texi

For posterity's sake, actually the file `doc/build_with_hip_sdk.txt` was removed, not `doc/general_contents.texi`

Yes, the file "doc/build_with_hip_sdk.txt" was removed, which is the actual we decided to remove. :)

GyanD marked this conversation as resolved

Zhao Zhili commented

2026-01-27 06:31:48 +00:00

Please ensure a thorough manual review of all AI-generated code before submission. Document on how to use pacman to install gcc and make in commit message is far from necessary, and add an option const named 8 with value 8 is unnecessary.

AI-generated code tends to be verbose; it requires the developer's own judgment and cleanup.

Please ensure a thorough manual review of all AI-generated code before submission. Document on how to use pacman to install gcc and make in commit message is far from necessary, and add an option const named 8 with value 8 is unnecessary. AI-generated code tends to be verbose; it requires the developer's own judgment and cleanup.

Jun Zhao requested changes

2026-01-27 08:59:38 +00:00

configure Outdated

					
				@ -7069,0 +7124,4 @@

				hipccflags="$hipccflags -std=c++17"

				if enabled x86_64 || enabled ppc64 || enabled aarch64; then

Has it actually been verified on a PPC64 machine?

Verified on "x86_64" architecture, but not PPC64. Removed the other options.

Steven Xiao commented

2026-01-27 15:59:39 +00:00

@quink wrote in #21595 (comment):

Please ensure a thorough manual review of all AI-generated code before submission. Document on how to use pacman to install gcc and make in commit message is far from necessary, and add an option const named 8 with value 8 is unnecessary.

AI-generated code tends to be verbose; it requires the developer's own judgment and cleanup.

@quink Please take a reference of this document https://github.com/GPUOpen-LibrariesAndSDKs/AMF/wiki/Build%20FFmpeg%20with%20AMF%20Support#3-build-ffmpeg-with-amf-for-windows. This document was created by me 4 years ago. I did the work step by step, then summarize this document. From my practice, those steps such as using pacman to install dependent packages are necessary. This is not related AI or human. This is really my experience and my practices.

Maybe for some people, this is not necessary, and take it as a common knowledge.

If reviewer think "using pacman to install dependent packages" is not necessary, of course I'd like to delete this part.

@quink wrote in https://code.ffmpeg.org/FFmpeg/FFmpeg/pulls/21595#issuecomment-23753: > Please ensure a thorough manual review of all AI-generated code before submission. Document on how to use pacman to install gcc and make in commit message is far from necessary, and add an option const named 8 with value 8 is unnecessary. > > AI-generated code tends to be verbose; it requires the developer's own judgment and cleanup. @quink Please take a reference of this document https://github.com/GPUOpen-LibrariesAndSDKs/AMF/wiki/Build%20FFmpeg%20with%20AMF%20Support#3-build-ffmpeg-with-amf-for-windows. This document was created by me 4 years ago. I did the work step by step, then summarize this document. From my practice, those steps such as using pacman to install dependent packages are necessary. This is not related AI or human. This is really my experience and my practices. Maybe for some people, this is not necessary, and take it as a common knowledge. If reviewer think "using pacman to install dependent packages" is not necessary, of course I'd like to delete this part.

Steven Xiao requested review from Zhao Zhili

2026-01-27 16:07:36 +00:00

Zhao Zhili commented

2026-01-27 16:31:03 +00:00

@younengxiao wrote in #21595 (comment):

@quink wrote in #21595 (comment):

Please ensure a thorough manual review of all AI-generated code before submission. Document on how to use pacman to install gcc and make in commit message is far from necessary, and add an option const named 8 with value 8 is unnecessary.
AI-generated code tends to be verbose; it requires the developer's own judgment and cleanup.

@quink Please take a reference of this document https://github.com/GPUOpen-LibrariesAndSDKs/AMF/wiki/Build%20FFmpeg%20with%20AMF%20Support#3-build-ffmpeg-with-amf-for-windows. This document was created by me 4 years ago. I did the work step by step, then summarize this document. From my practice, those steps such as using pacman to install dependent packages are necessary. This is not related AI or human. This is really my experience and my practices.

Maybe for some people, this is not necessary, and take it as a common knowledge.

If reviewer think "using pacman to install dependent packages" is not necessary, of course I'd like to delete this part.

I am sorry for the misunderstanding. I never realized there were developers who couldn't distinguish between a wiki page and a git commit message.

@younengxiao wrote in https://code.ffmpeg.org/FFmpeg/FFmpeg/pulls/21595#issuecomment-23785: > @quink wrote in #21595 (comment): > > > Please ensure a thorough manual review of all AI-generated code before submission. Document on how to use pacman to install gcc and make in commit message is far from necessary, and add an option const named 8 with value 8 is unnecessary. > > AI-generated code tends to be verbose; it requires the developer's own judgment and cleanup. > > @quink Please take a reference of this document https://github.com/GPUOpen-LibrariesAndSDKs/AMF/wiki/Build%20FFmpeg%20with%20AMF%20Support#3-build-ffmpeg-with-amf-for-windows. This document was created by me 4 years ago. I did the work step by step, then summarize this document. From my practice, those steps such as using pacman to install dependent packages are necessary. This is not related AI or human. This is really my experience and my practices. > > Maybe for some people, this is not necessary, and take it as a common knowledge. > > If reviewer think "using pacman to install dependent packages" is not necessary, of course I'd like to delete this part. I am sorry for the misunderstanding. I never realized there were developers who couldn't distinguish between a wiki page and a git commit message.

Steven Xiao commented

2026-01-27 16:42:43 +00:00

@quink wrote in #21595 (comment):

@younengxiao wrote in #21595 (comment):

@quink wrote in #21595 (comment):

Please ensure a thorough manual review of all AI-generated code before submission. Document on how to use pacman to install gcc and make in commit message is far from necessary, and add an option const named 8 with value 8 is unnecessary.
AI-generated code tends to be verbose; it requires the developer's own judgment and cleanup.

@quink Please take a reference of this document https://github.com/GPUOpen-LibrariesAndSDKs/AMF/wiki/Build%20FFmpeg%20with%20AMF%20Support#3-build-ffmpeg-with-amf-for-windows. This document was created by me 4 years ago. I did the work step by step, then summarize this document. From my practice, those steps such as using pacman to install dependent packages are necessary. This is not related AI or human. This is really my experience and my practices.
Maybe for some people, this is not necessary, and take it as a common knowledge.
If reviewer think "using pacman to install dependent packages" is not necessary, of course I'd like to delete this part.

I am sorry for the misunderstanding. I never realized there were developers who couldn't distinguish between a wiki page and a git commit message.

For some modules, if the build steps are complicated, it is a good idea to put the build steps into the commit message. This will help people who want to build ffmpeg on their own. Please take reference this pull request f4e0664fd1. I have suffered to take long time to find way to build some specific ffmpeg modules. This is the reason I also the reason I put the build steps in the commit message, and I also post page https://github.com/GPUOpen-LibrariesAndSDKs/AMF/wiki/Build%20FFmpeg%20with%20AMF%20Support.

@quink wrote in https://code.ffmpeg.org/FFmpeg/FFmpeg/pulls/21595#issuecomment-23787: > @younengxiao wrote in #21595 (comment): > > > @quink wrote in #21595 (comment): > > > Please ensure a thorough manual review of all AI-generated code before submission. Document on how to use pacman to install gcc and make in commit message is far from necessary, and add an option const named 8 with value 8 is unnecessary. > > > AI-generated code tends to be verbose; it requires the developer's own judgment and cleanup. > > > > > > @quink Please take a reference of this document https://github.com/GPUOpen-LibrariesAndSDKs/AMF/wiki/Build%20FFmpeg%20with%20AMF%20Support#3-build-ffmpeg-with-amf-for-windows. This document was created by me 4 years ago. I did the work step by step, then summarize this document. From my practice, those steps such as using pacman to install dependent packages are necessary. This is not related AI or human. This is really my experience and my practices. > > Maybe for some people, this is not necessary, and take it as a common knowledge. > > If reviewer think "using pacman to install dependent packages" is not necessary, of course I'd like to delete this part. > > I am sorry for the misunderstanding. I never realized there were developers who couldn't distinguish between a wiki page and a git commit message. For some modules, if the build steps are complicated, it is a good idea to put the build steps into the commit message. This will help people who want to build ffmpeg on their own. Please take reference this pull request https://code.ffmpeg.org/FFmpeg/FFmpeg/commit/f4e0664fd1bc94ee1a1afbd20c5a32ac33d93a2d. I have suffered to take long time to find way to build some specific ffmpeg modules. This is the reason I also the reason I put the build steps in the commit message, and I also post page https://github.com/GPUOpen-LibrariesAndSDKs/AMF/wiki/Build%20FFmpeg%20with%20AMF%20Support.

Steven Xiao referenced this pull request

2026-01-27 17:03:25 +00:00

avfilter/vf_mestimate_d3d12: remove redundant mb_size constants #21593

Steven Xiao force-pushed build/add-amd-hip-sdk-support-for-windows from 816ff702fa

Lint / Pre-Commit (pull_request) Successful in 24s

Details

Test / Fate (linux-aarch64, static, 64 bit) (pull_request) Successful in 11m55s

Details

Test / Fate (linux-amd64, static, 32 bit) (pull_request) Successful in 12m4s

Details

Test / Fate (linux-amd64, shared, 64 bit) (pull_request) Successful in 14m53s

Details

Test / Fate (Full, wine) (pull_request) Successful in 41m25s

Details

Autolabel / Labeler (pull_request_target) Has been skipped

Details

to 7d950339ea

Lint / Pre-Commit (pull_request) Successful in 23s

Details

Test / Fate (linux-aarch64, static, 64 bit) (pull_request) Successful in 11m45s

Details

Test / Fate (linux-amd64, shared, 64 bit) (pull_request) Successful in 15m15s

Details

Test / Fate (linux-amd64, static, 32 bit) (pull_request) Successful in 15m35s

Details

Test / Fate (Full, wine) (pull_request) Successful in 24m56s

Details

Autolabel / Labeler (pull_request_target) Has been skipped

Details

2026-01-27 18:36:55 +00:00

Compare

Steven Xiao commented

2026-01-27 19:15:57 +00:00

@quink wrote in #21595 (comment):

Please ensure a thorough manual review of all AI-generated code before submission. Document on how to use pacman to install gcc and make in commit message is far from necessary, and add an option const named 8 with value 8 is unnecessary.

AI-generated code tends to be verbose; it requires the developer's own judgment and cleanup.

As suggested, removed the commit message part which uses pacman to install dependent packages.

@quink wrote in https://code.ffmpeg.org/FFmpeg/FFmpeg/pulls/21595#issuecomment-23753: > Please ensure a thorough manual review of all AI-generated code before submission. Document on how to use pacman to install gcc and make in commit message is far from necessary, and add an option const named 8 with value 8 is unnecessary. > > AI-generated code tends to be verbose; it requires the developer's own judgment and cleanup. As suggested, removed the commit message part which uses pacman to install dependent packages.

Steven Xiao requested review from Jun Zhao

2026-01-27 19:16:19 +00:00

Jun Zhao commented

2026-01-28 10:47:14 +00:00

@younengxiao wrote in #21595 (comment):

@quink wrote in #21595 (comment):

@younengxiao wrote in #21595 (comment):

@quink wrote in #21595 (comment):

Please ensure a thorough manual review of all AI-generated code before submission. Document on how to use pacman to install gcc and make in commit message is far from necessary, and add an option const named 8 with value 8 is unnecessary.
AI-generated code tends to be verbose; it requires the developer's own judgment and cleanup.

@quink Please take a reference of this document https://github.com/GPUOpen-LibrariesAndSDKs/AMF/wiki/Build%20FFmpeg%20with%20AMF%20Support#3-build-ffmpeg-with-amf-for-windows. This document was created by me 4 years ago. I did the work step by step, then summarize this document. From my practice, those steps such as using pacman to install dependent packages are necessary. This is not related AI or human. This is really my experience and my practices.
Maybe for some people, this is not necessary, and take it as a common knowledge.
If reviewer think "using pacman to install dependent packages" is not necessary, of course I'd like to delete this part.

I am sorry for the misunderstanding. I never realized there were developers who couldn't distinguish between a wiki page and a git commit message.

For some modules, if the build steps are complicated, it is a good idea to put the build steps into the commit message. This will help people who want to build ffmpeg on their own. Please take reference this pull request f4e0664fd1. I have suffered to take long time to find way to build some specific ffmpeg modules. This is the reason I also the reason I put the build steps in the commit message, and I also post page https://github.com/GPUOpen-LibrariesAndSDKs/AMF/wiki/Build%20FFmpeg%20with%20AMF%20Support.

In fact, I believe you've just provided a bad case. The commit message should be clear and concise, explaining the purpose of this commit and what specific actions were taken.

No one would look up how to compile and set up the environment via Git commit messages.

@younengxiao wrote in https://code.ffmpeg.org/FFmpeg/FFmpeg/pulls/21595#issuecomment-23794: > @quink wrote in #21595 (comment): > > > @younengxiao wrote in #21595 (comment): > > > @quink wrote in #21595 (comment): > > > > Please ensure a thorough manual review of all AI-generated code before submission. Document on how to use pacman to install gcc and make in commit message is far from necessary, and add an option const named 8 with value 8 is unnecessary. > > > > AI-generated code tends to be verbose; it requires the developer's own judgment and cleanup. > > > > > > > > > @quink Please take a reference of this document https://github.com/GPUOpen-LibrariesAndSDKs/AMF/wiki/Build%20FFmpeg%20with%20AMF%20Support#3-build-ffmpeg-with-amf-for-windows. This document was created by me 4 years ago. I did the work step by step, then summarize this document. From my practice, those steps such as using pacman to install dependent packages are necessary. This is not related AI or human. This is really my experience and my practices. > > > Maybe for some people, this is not necessary, and take it as a common knowledge. > > > If reviewer think "using pacman to install dependent packages" is not necessary, of course I'd like to delete this part. > > > > > > I am sorry for the misunderstanding. I never realized there were developers who couldn't distinguish between a wiki page and a git commit message. > > For some modules, if the build steps are complicated, it is a good idea to put the build steps into the commit message. This will help people who want to build ffmpeg on their own. Please take reference this pull request [`f4e0664fd1`](https://code.ffmpeg.org/FFmpeg/FFmpeg/commit/f4e0664fd1bc94ee1a1afbd20c5a32ac33d93a2d). I have suffered to take long time to find way to build some specific ffmpeg modules. This is the reason I also the reason I put the build steps in the commit message, and I also post page https://github.com/GPUOpen-LibrariesAndSDKs/AMF/wiki/Build%20FFmpeg%20with%20AMF%20Support. In fact, I believe you've just provided a bad case. The commit message should be clear and concise, explaining the purpose of this commit and what specific actions were taken. No one would look up how to compile and set up the environment via Git commit messages.

Gyan Doshi commented

2026-01-28 10:53:50 +00:00

The commit message is not a place for end-user guidance. It should just describe the changes made and why.

Lynne reviewed

2026-01-28 13:15:50 +00:00

Lynne left a comment

I don't agree we should add support for this.

Unlike CUDA, which has plenty of users, as well as a whole ecosystem with decoders, encoders and filters, HIP would only be usable for filters. Our OpenCL code is already bitrotting after no one's touched it for years.

Furthermore, Vulkan covers all the hardware that HIP supports, is more open, and more available.

I don't agree we should add support for this. Unlike CUDA, which has plenty of users, as well as a whole ecosystem with decoders, encoders and filters, HIP would only be usable for filters. Our OpenCL code is already bitrotting after no one's touched it for years. Furthermore, Vulkan covers all the hardware that HIP supports, is more open, and more available.

Lynne closed this pull request

2026-01-28 13:15:54 +00:00

Steven Xiao commented

2026-01-28 14:35:27 +00:00

@GyanD @quink @mypopydev Thanks for the discussion about commit message. and we reach agreement what commit message should be. That makes sense 👍.

Steven Xiao commented

2026-01-28 15:14:21 +00:00

@Lynne wrote in #21595 (comment):

I don't agree we should add support for this.

Unlike CUDA, which has plenty of users, as well as a whole ecosystem with decoders, encoders and filters, HIP would only be usable for filters. Our OpenCL code is already bitrotting after no one's touched it for years.

Furthermore, Vulkan covers all the hardware that HIP supports, is more open, and more available.

I understand your concern about the scope and ecosystem of HIP at current time, and I respect your final decision.

@Lynne wrote in https://code.ffmpeg.org/FFmpeg/FFmpeg/pulls/21595#issuecomment-23849: > I don't agree we should add support for this. > > Unlike CUDA, which has plenty of users, as well as a whole ecosystem with decoders, encoders and filters, HIP would only be usable for filters. Our OpenCL code is already bitrotting after no one's touched it for years. > > Furthermore, Vulkan covers all the hardware that HIP supports, is more open, and more available. I understand your concern about the scope and ecosystem of HIP at current time, and I respect your final decision.