[SYCL][Clang] Add support for device image compression

[uditagarwal97]

This PR adds support for device image compression for the old offloading model. I'll make another follow-up PR to extend support for the new offload model.

Design summary:

ZSTD (compression algo)   ----> LLVMSupport (Interface)  ------> clang-offload-wrapper (For compression)
 |
 ----------------------------------------------------- --------> SYCL RT (For decompression)

This PR introduces ZSTD (https://github.com/facebook/zstd) as a 3rd party dependency of DPCPP. Similar to upstream LLVM, we expect user to have zstd-dev package installed on their machine - we won't be installing zstd from sources.

How to use
To compress device images, add --offload-compress CLI option to your clang invocation. Note that we compress device images only if the size of device images exceeds a threshold, which is 512 bytes by default. Moreover, by default, we use ZSTD level 10 for compression. ZSTD compression levels provides a tradeoff between (de)compression time and compression ratio, and the compression level can be changed using --offload-compression-level=<int> CLI option.

[uditagarwal97]

Some initial performance stats:

Dataset: https://github.com/aras-p/smol-v/tree/master/tests/spirv-dumps
Dataset size: 275 SPIR-V files

Conclusion:
Overall, for SPIR-V files < 50KB, the decompression time is below 0.1ms, compression time <0.15ms, and compression ratio is ~3 (compressed image is 1/3 the original size).
For very small images (<512 bytes), I don't see much benefit of image compression.

Note:- Most of the SPIR-V files I have in the dataset are <50KB. I'm working on extending the performance evaluation to larger workloads. Also, the (de)compression performance will vary with the format of the file being compressed, so for AOT, where device images consists of target assembly, the performance stats might differ.

[jbrodman]

What happens with the PTX and AMDGPU targets? Are they covered by the "native" binary image format? Do we need additional formats?

[jbrodman]

Also guessing this feature may not make sense when combined with the native cpu device, but need to think more about that.

[uditagarwal97]

What happens with the PTX and AMDGPU targets? Are they covered by the "native" binary image format? Do we need additional formats?

I think they are covered by the "none" binary image format. This is because clang driver (in SYCL offload mode) never specifies the image format in call to clang-offload-wrapper. So, by default, the BinaryImageFormat is "none" and it is upto the SYCL runtime to determine the format (https://github.com/intel/llvm/blob/sycl/sycl/source/detail/device_binary_image.cpp#L170).

I tested my changes with PTX, and they seem to work fine, so, we'd likely not require additional formats.

[mdtoguchi]

OK for driver

[uditagarwal97]

@bso-intel @intel/llvm-reviewers-runtime ping!

[bso-intel]

When the crash occurs when compression failed, the error message that the SYCL end-user will receive may not be useful.
For example, "Failed to create ZSTD_CCtx", ""Failed to set ZSTD_c_compressionLevel", etc.
It would be much better user experience if they get a message like "Device image compression failed." + e.what().

[bso-intel]

This kind of error message is good since the SYCL end-user can understand what went wrong.

[uditagarwal97]

@bso-intel

When the crash occurs when compression failed, the error message that the SYCL end-user will receive may not be useful.
For example, "Failed to create ZSTD_CCtx", ""Failed to set ZSTD_c_compressionLevel", etc.
It would be much better user experience if they get a message like "Device image compression failed." + e.what().

In 946a738, I've wrapped zstd::compress in try/catch to throw a more meaningful error message. Note that this will only work if DPC++ is built with LLVM_ENABLE_EH.

[uditagarwal97]

add [[maybe_unused]]