Replacing throw with TORCH_CHECK

[Krishn1412]

Replaced all the occurrences of throw in the cpp code with TORCH_CHECK.

TORCH_CHECK throws a runtime error, so the cpp test suite had to be updated as well.

[scotts]

I find this boolean expression much harder to reason about than the previous one. It's wordier, but let's instead do:

if (seekMode_ == SeekMode::approximate) {
  TORCH_CHECK(
    streamMetadata.averageFpsFromHeader.has_value(),
    "Same message as before");
}

I prefer that, as it reads closer to what we would say out loud: if the seek mode is approximate, assert that it has a value for fps.

[scotts]

We have to revert this check back to what it was previously. We turn the EndOfFileException into a Python IndexError in the Python-to-C++ layer:

torchcodec/src/torchcodec/_core/custom_ops.cpp

Lines 303 to 307 in f4a351c

	try {
	result = videoDecoder->getNextFrame();
	} catch (const SingleStreamDecoder::EndOfFileException& e) {
	C10_THROW_ERROR(IndexError, e.what());
	}

[Krishn1412]

Should I change the custom_ops.cpp file instead to handle it using TORCH_CHECK runtime errors?

[scotts]

No, we should keep what's going on in custom_ops.cpp as-is. It has the behavior we want, and we're already using the most relevant PyTorch APIs to achieve it.

[Krishn1412]

got it!

[scotts]

@Krishn1412, thanks for working on this! You're getting test failures because you'll need to pull changes from main - we changed a bunch of variable names in the past few days. There's also likely going to be some test failures related to not throwing the EndOfFileException that will need to be fixed.

[scotts]

@Krishn1412, it looks like we're still getting some errors related to the exceptions we're now throwing versus the errors we expect to see. If you follow the contributing guide, you can actually run all of these tests locally, too.