add support for half-precision gemm

[bjarthur]

is Float16 the right type to use in the ccall?

also, what do you want to do about tests? cublas only supports half precision for gemm, not any of the other functions (e.g. symm, trsm, hemm, etc.). i'd have to split that big for-loop apart and duplicate the array construction.

for size(A)==(64,1,1000), size(B)==(1,64,1000), size(C)==(64,64,1000) the execution time of CUBLAS.gemm_strided_batched! is 3/4 as long compared to float32 on an RTX8000, and several fold slower on a 1080Ti.

fixes #1076

[maleadt]

This function is deprecated; see deprecated.jl.

[maleadt]

Float16 gemm already works, I assume you intend to add support for the batched APIs?

also, what do you want to do about tests? cublas only supports half precision for gemm, not any of the other functions (e.g. symm, trsm, hemm, etc.). i'd have to split that big for-loop apart and duplicate the array construction.

Moving those test out to a self-contained testset that covers more types seems like the easiest solution. And the CUBLAS tests desperately need to be refactored like that anyway.

[bjarthur]

i've removed Hgemm_v2 and added tests.

[maleadt]

Thanks, looks good!
The Windows CI failure seems related though, Float16 tests will need to be gated by the device capability.

[bjarthur]

the failing windows build lists these:

  │ │ CUDA toolkit 11.4.0, artifact installation
  | │ CUDA driver 11.3.0
  | │ NVIDIA driver 466.27.0
  | │
  | │ Libraries:
  │ │ - CUBLAS: 11.5.2
  | │ - CURAND: 10.2.5
  | │ - CUFFT: 10.5.0
  | │ - CUSOLVER: 11.2.0
  | │ - CUSPARSE: 11.6.0
  | │ - CUPTI: 14.0.0
  | │ - NVML: 11.0.0+466.27
  | │ - CUDNN: 8.20.0 (for CUDA 11.3.0)
  | │ - CUTENSOR: 1.3.0 (for CUDA 11.2.0)
  | │
  | │ Toolchain:
  | │ - Julia: 1.6.2
  | │ - LLVM: 11.0.1
  | │ - PTX ISA support: 3.2, 4.0, 4.1, 4.2, 4.3, 5.0, 6.0, 6.1, 6.3, 6.4, 6.5, 7.0
  | │ - Device capability support: sm_35, sm_37, sm_50, sm_52, sm_53, sm_60, sm_61, sm_62, sm_70, sm_72, sm_75, sm_80

the passing 11.4 build is identical except the NVIDIA driver is 460.73, as is the NVML +. on what should i gate?

[maleadt]

You left out the important part, listing the actual device and its capability :-) Float16 generally requires capability 5.3 or higher, while ths GTX 970 is sm_52.

[codecov]

Codecov Report

Merging #1080 (d807b5b) into master (5710e25) will not change coverage.
The diff coverage is n/a.

❗ Current head d807b5b differs from pull request most recent head 381ac17. Consider uploading reports for the commit 381ac17 to get more accurate results

@@           Coverage Diff           @@
##           master    #1080   +/-   ##
=======================================
  Coverage   80.00%   80.00%           
=======================================
  Files         118      118           
  Lines        7656     7656           
=======================================
  Hits         6125     6125           
  Misses       1531     1531           

Impacted Files	Coverage Δ
lib/cublas/wrappers.jl	92.37% <ø> (ø)

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 5710e25...381ac17. Read the comment docs.

[bjarthur]

all tests pass now

[maleadt]

Great, thanks!