Fix pointer types in cublasHgemm

[lpawela]

In the current main branch this fails

N = 256
A = CUDA.rand(Float16, N, N)
B = CUDA.rand(Float16, N, N)
C = CUDA.zeros(Float16, N, N)
CUDA.CUBLAS.gemm!('N', 'N', one(Float16), A, B, zero(Float16), C)

with

ERROR: MethodError: no method matching unsafe_convert(::Type{Ptr{Float16}}, ::Float16)
The function `unsafe_convert` exists, but no method is defined for this combination of argument types.

Closest candidates are:
  unsafe_convert(::Type{Cwstring}, ::Any)
   @ Base strings/cstring.jl:101
  unsafe_convert(::Type{Ptr{T}}, ::Array{T}) where T
   @ Base pointer.jl:65
  unsafe_convert(::Type{Ptr{T}}, ::Ptr{NTuple{N, T}}) where {N, T}
   @ Base refpointer.jl:188

This PR fixes this issue.

[github-actions]

CUDA.jl Benchmarks

Details

Benchmark suite	Current: 4eecb7d	Previous: 1caa3ad	Ratio
array/accumulate/Float32/1d	99564 ns	100770 ns	0.99
array/accumulate/Float32/dims=1	74994 ns	76381 ns	0.98
array/accumulate/Float32/dims=1L	1627668 ns	1630799 ns	1.00
array/accumulate/Float32/dims=2	140212 ns	141553 ns	0.99
array/accumulate/Float32/dims=2L	652281 ns	653429 ns	1.00
array/accumulate/Int64/1d	118487 ns	119282 ns	0.99
array/accumulate/Int64/dims=1	78162 ns	80346 ns	0.97
array/accumulate/Int64/dims=1L	1722317 ns	1726167 ns	1.00
array/accumulate/Int64/dims=2	153130 ns	154865 ns	0.99
array/accumulate/Int64/dims=2L	959752 ns	960762 ns	1.00
array/broadcast	18289 ns	18636 ns	0.98
array/construct	1215.8 ns	1242.5 ns	0.98
array/copy	16369 ns	16748 ns	0.98
array/copyto!/cpu_to_gpu	212420 ns	214495 ns	0.99
array/copyto!/gpu_to_cpu	278989 ns	281315 ns	0.99
array/copyto!/gpu_to_gpu	9023 ns	10356 ns	0.87
array/iteration/findall/bool	132242 ns	134688 ns	0.98
array/iteration/findall/int	146865 ns	148663 ns	0.99
array/iteration/findfirst/bool	68968 ns	70356 ns	0.98
array/iteration/findfirst/int	70412 ns	71778 ns	0.98
array/iteration/findmin/1d	66370 ns	68513 ns	0.97
array/iteration/findmin/2d	100578 ns	101687 ns	0.99
array/iteration/logical	191377 ns	195448 ns	0.98
array/iteration/scalar	64811 ns	66072 ns	0.98
array/permutedims/2d	49758 ns	50059 ns	0.99
array/permutedims/3d	51177 ns	50890 ns	1.01
array/permutedims/4d	50724 ns	51304 ns	0.99
array/random/rand/Float32	11759 ns	11937 ns	0.99
array/random/rand/Int64	22787 ns	24454 ns	0.93
array/random/rand!/Float32	7948.75 ns	8092.666666666667 ns	0.98
array/random/rand!/Int64	20658 ns	20949 ns	0.99
array/random/randn/Float32	34093 ns	36329 ns	0.94
array/random/randn!/Float32	23709 ns	24624 ns	0.96
array/reductions/mapreduce/Float32/1d	33206 ns	33603 ns	0.99
array/reductions/mapreduce/Float32/dims=1	38492 ns	38517 ns	1.00
array/reductions/mapreduce/Float32/dims=1L	50040 ns	50114 ns	1.00
array/reductions/mapreduce/Float32/dims=2	56085 ns	56468 ns	0.99
array/reductions/mapreduce/Float32/dims=2L	66831 ns	67374 ns	0.99
array/reductions/mapreduce/Int64/1d	39421 ns	39976 ns	0.99
array/reductions/mapreduce/Int64/dims=1	41529 ns	41660 ns	1.00
array/reductions/mapreduce/Int64/dims=1L	86335 ns	86590 ns	1.00
array/reductions/mapreduce/Int64/dims=2	58367 ns	59096 ns	0.99
array/reductions/mapreduce/Int64/dims=2L	82816 ns	82999 ns	1.00
array/reductions/reduce/Float32/1d	32883 ns	33460 ns	0.98
array/reductions/reduce/Float32/dims=1	38547 ns	38736 ns	1.00
array/reductions/reduce/Float32/dims=1L	50252 ns	50547 ns	0.99
array/reductions/reduce/Float32/dims=2	55902 ns	56305 ns	0.99
array/reductions/reduce/Float32/dims=2L	66816 ns	67363 ns	0.99
array/reductions/reduce/Int64/1d	39417 ns	40096 ns	0.98
array/reductions/reduce/Int64/dims=1	41196 ns	41717 ns	0.99
array/reductions/reduce/Int64/dims=1L	86315 ns	86750 ns	0.99
array/reductions/reduce/Int64/dims=2	58055 ns	58479 ns	0.99
array/reductions/reduce/Int64/dims=2L	83335 ns	83983 ns	0.99
array/reverse/1d	16550 ns	17202 ns	0.96
array/reverse/1dL	67491 ns	68074 ns	0.99
array/reverse/1dL_inplace	65180 ns	65359 ns	1.00
array/reverse/1d_inplace	8181.333333333333 ns	8376.666666666666 ns	0.98
array/reverse/2d	20037 ns	20409 ns	0.98
array/reverse/2dL	71845 ns	72447 ns	0.99
array/reverse/2dL_inplace	64984 ns	64948 ns	1.00
array/reverse/2d_inplace	9619 ns	9770 ns	0.98
array/sorting/1d	2654972 ns	2653489 ns	1.00
array/sorting/2d	1033410 ns	1034091 ns	1.00
array/sorting/by	3178646 ns	3181225 ns	1.00
cuda/synchronization/context/auto	1106.7 ns	1135.5 ns	0.97
cuda/synchronization/context/blocking	898.1860465116279 ns	939.7407407407408 ns	0.96
cuda/synchronization/context/nonblocking	6006.6 ns	6181.8 ns	0.97
cuda/synchronization/stream/auto	988.3571428571429 ns	991.6 ns	1.00
cuda/synchronization/stream/blocking	818.2359550561798 ns	841.3783783783783 ns	0.97
cuda/synchronization/stream/nonblocking	6000.4 ns	6139.8 ns	0.98
integration/byval/reference	143159 ns	143264 ns	1.00
integration/byval/slices=1	145207 ns	145361 ns	1.00
integration/byval/slices=2	283925 ns	283971 ns	1.00
integration/byval/slices=3	422283 ns	422479 ns	1.00
integration/cudadevrt	101592 ns	101731 ns	1.00
integration/volumerhs	8988193 ns	8999650 ns	1.00
kernel/indexing	12642 ns	12703 ns	1.00
kernel/indexing_checked	13516 ns	13354 ns	1.01
kernel/launch	2083.8888888888887 ns	2086 ns	1.00
kernel/occupancy	734.7027027027027 ns	693.3698630136986 ns	1.06
kernel/rand	16065 ns	15890 ns	1.01
latency/import	3854784368 ns	3853824190 ns	1.00
latency/precompile	4635091489 ns	4635422433 ns	1.00
latency/ttfp	4535766511 ns	4550893912 ns	1.00

This comment was automatically generated by workflow using github-action-benchmark.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 16.31%. Comparing base (1caa3ad) to head (4eecb7d).

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #3157      +/-   ##
==========================================
+ Coverage   10.82%   16.31%   +5.48%     
==========================================
  Files         123      124       +1     
  Lines        9479     9875     +396     
==========================================
+ Hits         1026     1611     +585     
+ Misses       8453     8264     -189     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[maleadt]

You cannot edit libcublas.jl directly, instead (or also) edit res/wrap/libcublas_epilogue.jl so that these changes don't get lost on the next toolkit release.

[lpawela]

@maleadt updated