rekkonnect / garyon Goto Github PK

Currently, due to unknown reasons, running on an AMD Ryzen 7 1800X CPU, the benchmarks have concluded the following results:

These results indicate that the functions that use Vector128 and downcast the values into smaller sized values (such as Int64 to Int32, Int16 to Byte, etc.) are less performant than their non-generic variants.

Furthermore, any Vector128 functions related to converting from/to Double are at risk of being even less optimal than the manual unoptimized conversion. The same applies to Vector256 functions that convert into Int64, however are quite more performant than the ones converting from/to Double.

There might need be inspection of this, as it could be processor-specific. More benchmark runs would need to be performed for validation of this hypothesis.

SIMD operations have been under the umbrella of an incomplete API, and they should eventually be at least polished for the rest of the package to be cleanly usable. Pointer-related APIs should be kept in favor of older target frameworks (see #11), while Spans will be recommended for newer ones.

Older frameworks suffer from the lack of certain APIs, so there must be carefully defined conditions for building

• .NET Framework
  • 4.5
  • 4.6.1
  • 4.7.2
  • 4.8
• .NET Standard
  • 2.0
  • 2.1
• .NET Core
  • 3.1

Most functions named as "Store" perform a hidden conversion which is not indicated in the function names, in contrast to the intrinsics' function names which are trying to be as elaborate as possible. They should be renamed to "ConvertStore" to indicate that functionality.

The cause seems to be the mask that performs the operation being larger than the operation itself for custom-sized structs of such sizes, since the mask is considered a ulong*, exceeding the range of the structs themselves.

This is why the 8 ignored tests are breaking, after inspection of the breaking error while debugging the tests.