Encode into file-like

Author: NicolasHug

src/torchcodec/_core/AVIOContextHolder.cpp

@@ -23,10 +23,10 @@ void AVIOContextHolder::createAVIOContext(
       buffer != nullptr,
       "Failed to allocate buffer of size " + std::to_string(bufferSize));
 
-  TORCH_CHECK(
-      (seek != nullptr) && ((write != nullptr) ^ (read != nullptr)),
-      "seek method must be defined, and either write or read must be defined. "
-      "But not both!")
+  //   TORCH_CHECK(
+  //       (seek != nullptr) && ((write != nullptr) ^ (read != nullptr)),
+  //       "seek method must be defined, and either write or read must be
+  //       defined. " "But not both!")
   avioContext_.reset(avioAllocContext(
       buffer,
       bufferSize,

src/torchcodec/_core/AVIOFileLikeContext.cpp

@@ -23,7 +23,7 @@ AVIOFileLikeContext::AVIOFileLikeContext(py::object fileLike)
         py::hasattr(fileLike, "seek"),
         "File like object must implement a seek method.");
   }
-  createAVIOContext(&read, nullptr, &seek, &fileLike_);
+  createAVIOContext(&read, &write, &seek, &fileLike_);
 }
 
 int AVIOFileLikeContext::read(void* opaque, uint8_t* buf, int buf_size) {
@@ -77,4 +77,12 @@ int64_t AVIOFileLikeContext::seek(void* opaque, int64_t offset, int whence) {
   return py::cast<int64_t>((*fileLike)->attr("seek")(offset, whence));
 }
 
+int AVIOFileLikeContext::write(void* opaque, const uint8_t* buf, int buf_size) {
+  auto fileLike = static_cast<UniquePyObject*>(opaque);
+  py::gil_scoped_acquire gil;
+  py::bytes bytes_obj(reinterpret_cast<const char*>(buf), buf_size);
+
+  return py::cast<int64_t>((*fileLike)->attr("write")(bytes_obj));
+}
+
 } // namespace facebook::torchcodec

src/torchcodec/_core/AVIOFileLikeContext.h

@@ -24,6 +24,7 @@ class AVIOFileLikeContext : public AVIOContextHolder {
  private:
   static int read(void* opaque, uint8_t* buf, int buf_size);
   static int64_t seek(void* opaque, int64_t offset, int whence);
+  static int write(void* opaque, const uint8_t* buf, int buf_size);
 
   // Note that we dynamically allocate the Python object because we need to
   // strictly control when its destructor is called. We must hold the GIL

src/torchcodec/_core/Encoder.cpp

@@ -1,6 +1,7 @@
 #include <sstream>
 
 #include "src/torchcodec/_core/AVIOBytesContext.h"
+#include "src/torchcodec/_core/AVIOContextHolder.h"
 #include "src/torchcodec/_core/Encoder.h"
 #include "torch/types.h"
 
@@ -148,6 +149,31 @@ AudioEncoder::AudioEncoder(
   initializeEncoder(sampleRate, bitRate);
 }
 
+// TODO this sucks, shouldn't need 2 separate constructors for AVIOContextHolder
+AudioEncoder::AudioEncoder(
+    const torch::Tensor wf,
+    int sampleRate,
+    std::string_view formatName,
+    std::unique_ptr<AVIOContextHolder> avioContextHolder,
+    std::optional<int64_t> bitRate)
+    : wf_(validateWf(wf)), avioContextHolderrrr_(std::move(avioContextHolder)) {
+  setFFmpegLogLevel();
+  AVFormatContext* avFormatContext = nullptr;
+  int status = avformat_alloc_output_context2(
+      &avFormatContext, nullptr, formatName.data(), nullptr);
+
+  TORCH_CHECK(
+      avFormatContext != nullptr,
+      "Couldn't allocate AVFormatContext. ",
+      "Check the desired extension? ",
+      getFFMPEGErrorStringFromErrorCode(status));
+  avFormatContext_.reset(avFormatContext);
+
+  avFormatContext_->pb = avioContextHolderrrr_->getAVIOContext();
+
+  initializeEncoder(sampleRate, bitRate);
+}
+
 void AudioEncoder::initializeEncoder(
     int sampleRate,
     std::optional<int64_t> bitRate) {

src/torchcodec/_core/Encoder.h

@@ -1,6 +1,7 @@
 #pragma once
 #include <torch/types.h>
 #include "src/torchcodec/_core/AVIOBytesContext.h"
+#include "src/torchcodec/_core/AVIOContextHolder.h"
 #include "src/torchcodec/_core/FFMPEGCommon.h"
 
 namespace facebook::torchcodec {
@@ -28,6 +29,12 @@ class AudioEncoder {
       std::string_view formatName,
       std::unique_ptr<AVIOToTensorContext> avioContextHolder,
       std::optional<int64_t> bitRate = std::nullopt);
+  AudioEncoder(
+      const torch::Tensor wf,
+      int sampleRate,
+      std::string_view formatName,
+      std::unique_ptr<AVIOContextHolder> avioContextHolder,
+      std::optional<int64_t> bitRate = std::nullopt);
   void encode();
   torch::Tensor encodeToTensor();
 
@@ -49,6 +56,7 @@ class AudioEncoder {
 
   // Stores the AVIOContext for the output tensor buffer.
   std::unique_ptr<AVIOToTensorContext> avioContextHolder_;
+  std::unique_ptr<AVIOContextHolder> avioContextHolderrrr_; // EWWWWW
 
   bool encodeWasCalled_ = false;
 };

src/torchcodec/_core/__init__.py

@@ -23,6 +23,7 @@
     create_from_file_like,
     create_from_tensor,
     encode_audio_to_file,
+    encode_audio_to_file_like,
     encode_audio_to_tensor,
     get_ffmpeg_library_versions,
     get_frame_at_index,

src/torchcodec/_core/ops.py

@@ -153,6 +153,17 @@ def create_from_file_like(
     return _convert_to_tensor(_pybind_ops.create_from_file_like(file_like, seek_mode))
 
 
+def encode_audio_to_file_like(
+    file_like: Union[io.RawIOBase, io.BufferedReader],
+    wf: torch.Tensor,
+    sample_rate: int,
+    format: str,
+    bit_rate: Optional[int] = None,
+):
+    assert _pybind_ops is not None
+    _pybind_ops.encode_audio_to_file_like(file_like, wf, sample_rate, format, bit_rate)
+
+
 # ==============================
 # Abstract impl for the operators. Needed by torch.compile.
 # ==============================

src/torchcodec/_core/pybind_ops.cpp

@@ -10,6 +10,7 @@
 #include <string>
 
 #include "src/torchcodec/_core/AVIOFileLikeContext.h"
+#include "src/torchcodec/_core/Encoder.h"
 #include "src/torchcodec/_core/SingleStreamDecoder.h"
 
 namespace py = pybind11;
@@ -38,8 +39,26 @@ int64_t create_from_file_like(
   return reinterpret_cast<int64_t>(decoder);
 }
 
+void encode_audio_to_file_like(
+    py::object file_like,
+    // const at::Tensor wf,
+    [[maybe_unused]] int wf,
+    int64_t sample_rate,
+    std::string_view format,
+    std::optional<int64_t> bit_rate = std::nullopt) {
+  auto avioContextHolder = std::make_unique<AVIOFileLikeContext>(file_like);
+  AudioEncoder(
+      torch::empty({2, 1000}, torch::kFloat32),
+      sample_rate, // TODO need validateSampleRate
+      format,
+      std::move(avioContextHolder),
+      bit_rate)
+      .encode();
+}
+
 PYBIND11_MODULE(decoder_core_pybind_ops, m) {
   m.def("create_from_file_like", &create_from_file_like);
+  m.def("encode_audio_to_file_like", &encode_audio_to_file_like);
 }
 
 } // namespace facebook::torchcodec