[torchcodec] Add support for Nvidia GPU Decoding #58
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
facebook-github-bot
added
the
CLA Signed
|
This pull request was exported from Phabricator. Differential Revision: D59121006 |
|
This pull request was exported from Phabricator. Differential Revision: D59121006 |
5 similar comments
|
This pull request was exported from Phabricator. Differential Revision: D59121006 |
|
This pull request was exported from Phabricator. Differential Revision: D59121006 |
|
This pull request was exported from Phabricator. Differential Revision: D59121006 |
|
This pull request was exported from Phabricator. Differential Revision: D59121006 |
|
This pull request was exported from Phabricator. Differential Revision: D59121006 |
|
This pull request was exported from Phabricator. Differential Revision: D59121006 |
3 similar comments
|
This pull request was exported from Phabricator. Differential Revision: D59121006 |
|
This pull request was exported from Phabricator. Differential Revision: D59121006 |
|
This pull request was exported from Phabricator. Differential Revision: D59121006 |
ahmadsharif1
added a commit
to ahmadsharif1/torchcodec
that referenced
this pull request
Jul 31, 2024
Summary: X-link: pytorch#58 1. Add CUDA support to VideoDecoder.cpp. This is done by checking what device is passed into the options and using CUDA if the device type is cuda. 2. Add -DENABLE_CUDA flag in cmake. 3. Check ENABLE_CUDA environment variable in setup.py and pass it down to cmake if it is present. 4. Add a unit test to demonstrate that CUDA decoding does work. This uses a different tensor than the one from CPU decoding because hardware decoding is intrinsically a bit inaccurate. I generated the reference tensor by dumping the tensor from the GPU on my devVM. It is possible different Nvidia hardware show different outputs. How to test this in a more robust way is TBD. 5. Added a new parameter for cuda device index for `add_video_stream`. If this is present, we will use it to do hardware decoding on a CUDA device. There is a whole bunch of TODOs: 1. Currently GPU utilization is only 7-8% when decoding the video. We need to get this higher. 2. Speed it up compared to CPU implementation. Currently this is slower than CPU decoding even for HD videos (probably because we can't hide the CPU to GPU memcpy). However, decode+resize is faster as the benchmark says. Differential Revision: D59121006
|
This pull request was exported from Phabricator. Differential Revision: D59121006 |
4 similar comments
|
This pull request was exported from Phabricator. Differential Revision: D59121006 |
|
This pull request was exported from Phabricator. Differential Revision: D59121006 |
|
This pull request was exported from Phabricator. Differential Revision: D59121006 |
|
This pull request was exported from Phabricator. Differential Revision: D59121006 |
ahmadsharif1
added a commit
to ahmadsharif1/torchcodec
that referenced
this pull request
Jul 31, 2024
Summary: Pull Request resolved: pytorch#58 X-link: pytorch#58 1. Add CUDA support to VideoDecoder.cpp. This is done by checking what device is passed into the options and using CUDA if the device type is cuda. 2. Add -DENABLE_CUDA flag in cmake. 3. Check ENABLE_CUDA environment variable in setup.py and pass it down to cmake if it is present. 4. Add a unit test to demonstrate that CUDA decoding does work. This uses a different tensor than the one from CPU decoding because hardware decoding is intrinsically a bit inaccurate. I generated the reference tensor by dumping the tensor from the GPU on my devVM. It is possible different Nvidia hardware show different outputs. How to test this in a more robust way is TBD. 5. Added a new parameter for cuda device index for `add_video_stream`. If this is present, we will use it to do hardware decoding on a CUDA device. There is a whole bunch of TODOs: 1. Currently GPU utilization is only 7-8% when decoding the video. We need to get this higher. 2. Speed it up compared to CPU implementation. Currently this is slower than CPU decoding even for HD videos (probably because we can't hide the CPU to GPU memcpy). However, decode+resize is faster as the benchmark says. Reviewed By: scotts Differential Revision: D59121006
ahmadsharif1
added a commit
to ahmadsharif1/torchcodec
that referenced
this pull request
Jul 31, 2024
Summary: Pull Request resolved: pytorch#137 Pull Request resolved: pytorch#58 X-link: pytorch#58 1. Add CUDA support to VideoDecoder.cpp. This is done by checking what device is passed into the options and using CUDA if the device type is cuda. 2. Add -DENABLE_CUDA flag in cmake. 3. Check ENABLE_CUDA environment variable in setup.py and pass it down to cmake if it is present. 4. Add a unit test to demonstrate that CUDA decoding does work. This uses a different tensor than the one from CPU decoding because hardware decoding is intrinsically a bit inaccurate. I generated the reference tensor by dumping the tensor from the GPU on my devVM. It is possible different Nvidia hardware show different outputs. How to test this in a more robust way is TBD. 5. Added a new parameter for cuda device index for `add_video_stream`. If this is present, we will use it to do hardware decoding on a CUDA device. There is a whole bunch of TODOs: 1. Currently GPU utilization is only 7-8% when decoding the video. We need to get this higher. 2. Speed it up compared to CPU implementation. Currently this is slower than CPU decoding even for HD videos (probably because we can't hide the CPU to GPU memcpy). However, decode+resize is faster as the benchmark says. Reviewed By: scotts Differential Revision: D59121006
ahmadsharif1
added a commit
to ahmadsharif1/torchcodec
that referenced
this pull request
Aug 1, 2024
Summary: Pull Request resolved: pytorch#137 Pull Request resolved: pytorch#58 X-link: pytorch#58 1. Add CUDA support to VideoDecoder.cpp. This is done by checking what device is passed into the options and using CUDA if the device type is cuda. 2. Add -DENABLE_CUDA flag in cmake. 3. Check ENABLE_CUDA environment variable in setup.py and pass it down to cmake if it is present. 4. Add a unit test to demonstrate that CUDA decoding does work. This uses a different tensor than the one from CPU decoding because hardware decoding is intrinsically a bit inaccurate. I generated the reference tensor by dumping the tensor from the GPU on my devVM. It is possible different Nvidia hardware show different outputs. How to test this in a more robust way is TBD. 5. Added a new parameter for cuda device index for `add_video_stream`. If this is present, we will use it to do hardware decoding on a CUDA device. There is a whole bunch of TODOs: 1. Currently GPU utilization is only 7-8% when decoding the video. We need to get this higher. 2. Speed it up compared to CPU implementation. Currently this is slower than CPU decoding even for HD videos (probably because we can't hide the CPU to GPU memcpy). However, decode+resize is faster as the benchmark says. Reviewed By: scotts Differential Revision: D59121006
ahmadsharif1
added a commit
to ahmadsharif1/torchcodec
that referenced
this pull request
Aug 1, 2024
Summary: Pull Request resolved: pytorch#137 Pull Request resolved: pytorch#58 X-link: pytorch#58 1. Add CUDA support to VideoDecoder.cpp. This is done by checking what device is passed into the options and using CUDA if the device type is cuda. 2. Add -DENABLE_CUDA flag in cmake. 3. Check ENABLE_CUDA environment variable in setup.py and pass it down to cmake if it is present. 4. Add a unit test to demonstrate that CUDA decoding does work. This uses a different tensor than the one from CPU decoding because hardware decoding is intrinsically a bit inaccurate. I generated the reference tensor by dumping the tensor from the GPU on my devVM. It is possible different Nvidia hardware show different outputs. How to test this in a more robust way is TBD. 5. Added a new parameter for cuda device index for `add_video_stream`. If this is present, we will use it to do hardware decoding on a CUDA device. There is a whole bunch of TODOs: 1. Currently GPU utilization is only 7-8% when decoding the video. We need to get this higher. 2. Speed it up compared to CPU implementation. Currently this is slower than CPU decoding even for HD videos (probably because we can't hide the CPU to GPU memcpy). However, decode+resize is faster as the benchmark says. Reviewed By: scotts Differential Revision: D59121006
ahmadsharif1
added a commit
to ahmadsharif1/torchcodec
that referenced
this pull request
Aug 1, 2024
Summary: Pull Request resolved: pytorch#137 Pull Request resolved: pytorch#58 X-link: pytorch#58 1. Add CUDA support to VideoDecoder.cpp. This is done by checking what device is passed into the options and using CUDA if the device type is cuda. 2. Add -DENABLE_CUDA flag in cmake. 3. Check ENABLE_CUDA environment variable in setup.py and pass it down to cmake if it is present. 4. Add a unit test to demonstrate that CUDA decoding does work. This uses a different tensor than the one from CPU decoding because hardware decoding is intrinsically a bit inaccurate. I generated the reference tensor by dumping the tensor from the GPU on my devVM. It is possible different Nvidia hardware show different outputs. How to test this in a more robust way is TBD. 5. Added a new parameter for cuda device index for `add_video_stream`. If this is present, we will use it to do hardware decoding on a CUDA device. There is a whole bunch of TODOs: 1. Currently GPU utilization is only 7-8% when decoding the video. We need to get this higher. 2. Speed it up compared to CPU implementation. Currently this is slower than CPU decoding even for HD videos (probably because we can't hide the CPU to GPU memcpy). However, decode+resize is faster as the benchmark says. Differential Revision: D59121006 Reviewed By: scotts
ahmadsharif1
added a commit
to ahmadsharif1/torchcodec
that referenced
this pull request
Aug 1, 2024
Summary: Pull Request resolved: pytorch#137 Pull Request resolved: pytorch#58 X-link: pytorch#58 1. Add CUDA support to VideoDecoder.cpp. This is done by checking what device is passed into the options and using CUDA if the device type is cuda. 2. Add -DENABLE_CUDA flag in cmake. 3. Check ENABLE_CUDA environment variable in setup.py and pass it down to cmake if it is present. 4. Add a unit test to demonstrate that CUDA decoding does work. This uses a different tensor than the one from CPU decoding because hardware decoding is intrinsically a bit inaccurate. I generated the reference tensor by dumping the tensor from the GPU on my devVM. It is possible different Nvidia hardware show different outputs. How to test this in a more robust way is TBD. 5. Added a new parameter for cuda device index for `add_video_stream`. If this is present, we will use it to do hardware decoding on a CUDA device. There is a whole bunch of TODOs: 1. Currently GPU utilization is only 7-8% when decoding the video. We need to get this higher. 2. Speed it up compared to CPU implementation. Currently this is slower than CPU decoding even for HD videos (probably because we can't hide the CPU to GPU memcpy). However, decode+resize is faster as the benchmark says. Reviewed By: scotts Differential Revision: D59121006
facebook-github-bot
pushed a commit
that referenced
this pull request
Aug 1, 2024
Summary: Pull Request resolved: #137 Pull Request resolved: #58 X-link: #58 1. Add CUDA support to VideoDecoder.cpp. This is done by checking what device is passed into the options and using CUDA if the device type is cuda. 2. Add -DENABLE_CUDA flag in cmake. 3. Check ENABLE_CUDA environment variable in setup.py and pass it down to cmake if it is present. 4. Add a unit test to demonstrate that CUDA decoding does work. This uses a different tensor than the one from CPU decoding because hardware decoding is intrinsically a bit inaccurate. I generated the reference tensor by dumping the tensor from the GPU on my devVM. It is possible different Nvidia hardware show different outputs. How to test this in a more robust way is TBD. 5. Added a new parameter for cuda device index for `add_video_stream`. If this is present, we will use it to do hardware decoding on a CUDA device. There is a whole bunch of TODOs: 1. Currently GPU utilization is only 7-8% when decoding the video. We need to get this higher. 2. Speed it up compared to CPU implementation. Currently this is slower than CPU decoding even for HD videos (probably because we can't hide the CPU to GPU memcpy). However, decode+resize is faster as the benchmark says. Reviewed By: scotts Differential Revision: D59121006 fbshipit-source-id: da6faa60c8de5d8e6ad90f8897d339c9979005f1
|
When it will be available for using ? |
|
It's available if you build torchcodec from source by installing pre-reqs and running: We plan to release it to pip later this year. |
|
@ahmadsharif1 Thanks for your answer! Do you plan to support RTSP gpu decoding ? |
|
@Ilyabasharov I believe RTSP should work if FFMPEG is configured correctly in your environment. You can run If you are building ffmpeg yourself you can add this to your configure line to get that:
Also, pip binaries are available now on pytorch nightly: |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Summary:
add_video_stream. If this is present, we will use it to do hardware decoding on a CUDA device.Differential Revision: D59121006