Added ability to accumulate in FP16. Convert BF16 to FP32. For FP16 and BF16 GEMM in RISC-V (BF16 now works for pre-RVA23)

[ChipKerchner]

Added ability to accumulate in FP16 for GEMM. Widens once at the end of loops.

Testing LLVM FP16 LMUL1 VLEN256 GEMM 1 0 0  512  512  512   1  2.0  1.0  1

Total time =         24948910

Testing LLVM FP16_N LMUL1 VLEN256 GEMM 1 0 0  512  512  512   1  2.0  1.0  1

Total time =         18968190

Accumulation differences are about 4 epsilons - compared to the widening (previous) version. But the performance it up to 2.7X faster - Note: BananaPi shows only 1.85X faster.

BF16 is 1.5X faster and now works for RVA22 systems like BananaPi

[ChipKerchner]

Unfortunately BF16 only has widening MADD instructions. So the same changes can not be made for BF16.

[ChipKerchner]

These are for VLEN = 256 only currently

[ChipKerchner]

It now works for VLEN = 128.

[ChipKerchner]

Main loop now uses LMUL = 2

[ChipKerchner]

Even faster!!!

Testing LLVM FP16_N LMUL1 VLEN256 GEMM 1 0 0  512  512  512   1  2.0  1.0  1

Total time =         13400067

[ChipKerchner]

Convert inputs from BF16 to FP32 and use FP32 vector madds. 18% faster.

[ChipKerchner]

Convert BF16 values once (and vectorized) - 3-4% faster.

[ChipKerchner]

Latest BF16 version is 48% faster than the current version. And is 3% faster than the FP32 .

It uses 1.5X more memory unfortunately.

It would be possible to make this version to work on the BananaPi - which doesn't support BF16 vector instructions. The conversions could be done with uint32 conversion and shift. And fixing the remaining BF16 vectors to not uses vector BF16 MADDs.

[ChipKerchner]

BF16 GEMM now works for pre-RVA23 systems like BananaPi

[ChipKerchner]

This is ready to go unless someone has comments.