Skip to content

Reduce the diff with the pytrch:main branch #3

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
gouzil wants to merge 3 commits into
base: paddle
Choose a base branch
from
+18 −23
Open

Reduce the diff with the pytrch:main branch #3

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
3a6ca0c
reduce narrow, see: paddlepaddle/paddle#77544
gouzil Mar 29, 2026
22b2c3e
feat: support BackendSelect
gouzil Apr 4, 2026
c348bb8
fix: enable audio frame decoding check in getFramesPlayedInRangeAudio
gouzil Apr 5, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
6 changes: 2 additions & 4 deletions src/torchcodec/_core/AVIOTensorContext.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -123,10 +123,8 @@ AVIOToTensorContext::AVIOToTensorContext()
}

torch::Tensor AVIOToTensorContext::getOutputTensor() {
throw std::runtime_error(
"AVIOToTensorContext::getOutputTensor is not implemented yet.");
// return tensorContext_.data.narrow(
// /*dim=*/0, /*start=*/0, /*length=*/tensorContext_.max_pos);
return tensorContext_.data.narrow(
/*dim=*/0, /*start=*/0, /*length=*/tensorContext_.max_pos);
}

} // namespace facebook::torchcodec
25 changes: 11 additions & 14 deletions src/torchcodec/_core/SingleStreamDecoder.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1030,15 +1030,15 @@ AudioFramesOutput SingleStreamDecoder::getFramesPlayedInRangeAudio(
frames.push_back(*lastSamples);
}

// TORCH_CHECK(
// frames.size() > 0 && firstFramePtsSeconds.has_value(),
// "No audio frames were decoded. ",
// "This is probably because start_seconds is too high(",
// startSeconds,
// "),",
// "or because stop_seconds(",
// stopSecondsOptional,
// ") is too low.");
TORCH_CHECK(
frames.size() > 0 && firstFramePtsSeconds.has_value(),
"No audio frames were decoded. ",
"This is probably because start_seconds is too high(",
startSeconds,
"),",
"or because stop_seconds(",
stopSecondsOptional,
") is too low.");

return AudioFramesOutput{torch::cat(frames, 1), *firstFramePtsSeconds};
}
Expand Down Expand Up @@ -1419,11 +1419,8 @@ std::optional<torch::Tensor> SingleStreamDecoder::maybeFlushSwrBuffers() {
auto actualNumRemainingSamples = swr_convert(
swrContext_.get(), outputBuffers.data(), numRemainingSamples, nullptr, 0);
Copy link

Copilot AI Apr 4, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

swr_convert() can return a negative value on error. Passing that directly as the length to lastSamples.narrow(...) will throw an unrelated/cryptic error (or potentially create an invalid slice). Add an explicit check for actualNumRemainingSamples < 0 and surface a clear failure (e.g., TORCH_CHECK(actualNumRemainingSamples >= 0, ...)) before narrowing; optionally also handle the == 0 case explicitly.

Suggested change
swrContext_.get(), outputBuffers.data(), numRemainingSamples, nullptr, 0);
swrContext_.get(), outputBuffers.data(), numRemainingSamples, nullptr, 0);
TORCH_CHECK(
actualNumRemainingSamples >= 0,
"swr_convert() failed while flushing buffered audio samples: ",
actualNumRemainingSamples);
if (actualNumRemainingSamples == 0) {
return std::nullopt;
}

Copilot uses AI. Check for mistakes.

throw std::runtime_error(
"SingleStreamDecoder::maybeFlushSwrBuffers is not implemented yet.");

// return lastSamples.narrow(
// /*dim=*/1, /*start=*/0, /*length=*/actualNumRemainingSamples);
return lastSamples.narrow(
/*dim=*/1, /*start=*/0, /*length=*/actualNumRemainingSamples);
}

// --------------------------------------------------------------------------
Expand Down
10 changes: 5 additions & 5 deletions src/torchcodec/_core/custom_ops.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -10,12 +10,12 @@
#include <string>
#include "c10/core/SymIntArrayRef.h"
#include "c10/util/Exception.h"
#include "torch/library.h"
#include "src/torchcodec/_core/AVIOFileLikeContext.h"
#include "src/torchcodec/_core/AVIOTensorContext.h"
#include "src/torchcodec/_core/Encoder.h"
#include "src/torchcodec/_core/SingleStreamDecoder.h"
#include "src/torchcodec/_core/ValidationUtils.h"
#include "torch/library.h"

namespace facebook::torchcodec {

Expand Down Expand Up @@ -118,7 +118,7 @@ OpsFrameOutput makeOpsFrameOutput(FrameOutput& frame) {
// frame.data,
// torch::tensor(frame.ptsSeconds, torch::dtype(torch::kFloat64)),
// torch::tensor(frame.durationSeconds, torch::dtype(torch::kFloat64)));
return std::make_tuple(
return std::make_tuple(
frame.data,
torch::full({}, frame.ptsSeconds, torch::kFloat64),
torch::full({}, frame.durationSeconds, torch::kFloat64));
Expand Down Expand Up @@ -920,15 +920,15 @@ void scan_all_streams_to_update_metadata(at::Tensor& decoder) {
videoDecoder->scanFileAndUpdateMetadataAndIndex();
}

TORCH_LIBRARY_IMPL(torchcodec_ns, CPU, m) {
TORCH_LIBRARY_IMPL(torchcodec_ns, BackendSelect, m) {
m.impl("create_from_file", &create_from_file);
m.impl("create_from_tensor", &create_from_tensor);
m.impl("_create_from_file_like", &_create_from_file_like);
Comment on lines +923 to 926
Copy link

Copilot AI Apr 4, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Registering create_from_tensor under BackendSelect is risky because it has a Tensor argument and the implementation assumes host-accessible memory (no video_tensor.is_cpu() check; it is passed into AVIOFromTensorContext which reads via data_ptr()/memcpy). With this dispatch key change, CUDA tensors may no longer fail dispatch and could reach this CPU-only code path. Consider registering create_from_tensor under CPU (and only keeping no-Tensor factory ops like create_from_file / _get_json_ffmpeg_library_versions under BackendSelect), or add a strict CPU device check for video_tensor if BackendSelect is intentional.

Copilot uses AI. Check for mistakes.
m.impl(
"_get_json_ffmpeg_library_versions", &_get_json_ffmpeg_library_versions);
// }
}

// TORCH_LIBRARY_IMPL(torchcodec_ns, CPU, m) {
TORCH_LIBRARY_IMPL(torchcodec_ns, CPU, m) {
m.impl("encode_audio_to_file", &encode_audio_to_file);
m.impl("encode_audio_to_tensor", &encode_audio_to_tensor);
m.impl("_encode_audio_to_file_like", &_encode_audio_to_file_like);
Expand Down
Loading