WASM Support

[abhi-agg]

This issue documents the possible tasks to optimize/enable the wasm port of translation engine for inference:

High Priority:

• A built-in wasm 8-bit matrix multiply primitive browsermt/bergamot-translator#205
• Compile with at least SSE4.1 for wasm (Firefox disables wasm simd for < SSE4.1)
• Enable 8-bit integer computations in Attention layer of Marian framework browsermt/marian-dev#50
• Profile and Optimize the JS bindings of the engine for performance
• Use WASI instead of Emscripten toolchain for compilation to wasm

Low priority/Untriaged:

• Use multiple threads for inference
• Enable SharedArrayBuffer on Firefox Android (required for multi-threaded wasm)
• Non-simd variant of bergamot?
• Check if all intended cmake options are being used
• Check if all intended compilation flags are being used (dependent on the toolchain)
• Enable compiling with -munimplemented-simd128 flag for wasm (is it still valid?)
• Guard the remaining try/catch/throw statements (lots of 3rd party dependencies, probably a tedious task)?
• A true exception-free (-fno-exception) compilation (lots of 3rd party dependencies, probably a tedious task)?