aneshlya

LLVM intrinsics corresponding to VNNI instructions operate not on specific int8/int16 types, but on packed `int32` types. ISPC should adhere to this. It will allow the packing of types that...

After careful consideration of various design approaches of exposing VNNI functions in ISPC stdlib I'm inclined towards the following set of functions in stdlib: ``` enum ISPC_PACKED_FACTOR { ISPC_PACKED_FACTOR_BYTE, //...

`ISPC_PACKED_FACTOR_BYTE` uniform parameter would be optimized so it will not produce any additional control flow instructions. However I really like `dot2add` and `dot4add` naming which has this "factor" information compactly...

This would work! The ISPC stdlib frequently uses function overloading, and we don't typically encode type information in the naming. However, I agree that with packed types, having type information...

Could you paste CMake command of how you build ISPC? Also if you have a Dockerfile or any other script containing steps for setting of the whole build environment, please...

If we enable tests as part of the build, we should install them. The option to disable tests is `ISPCRT_BUILD_TESTS`. It can be OFF by default. But if it's ON,...

Hi @StefanBruens , We expect that lit-tests are run using the same LLVM as for ISPC build. LLVM for ISPC should be built using `alloy.py` script which will patch it...

ISPC has limits for integral types in stdlib: `INT8_MAX`/`INT8_MIN`, `INT16_MAX`/`INT16_MIN`, `INT32_MAX`/`INT32_MIN`, `INT64_MAX`/`INT64_MIN`, `UINT8_MAX`, `UINT16_MAX`, `UINT32_MAX`, `UINT64_MAX` . But doesn't have for floating point types. Plus, no mentions of it in...

I believe that the issues created by @nurmukhametov are covering all topics touched in this one so I'm closing the issue. @MkazemAkhgary, feel free to create another one if something...

Please format lit-tests with `clang-format`.