Add narrow-precision floats to cudaDataType

[AntonOresten]

Adds narrow-precision floats to the cudaDataType enum based on library_types.h in CUDA 13:

    CUDA_R_8F_E4M3 = 28, /* real as a nv_fp8_e4m3 */
    CUDA_R_8F_UE4M3 = CUDA_R_8F_E4M3, /* real as an unsigned nv_fp8_e4m3 */
    CUDA_R_8F_E5M2 = 29, /* real as a nv_fp8_e5m2 */
    CUDA_R_8F_UE8M0 = 30,  /* real as an exponent-only unsigned nv_fp8_e8m0 */
    CUDA_R_6F_E2M3  = 31,  /* real as a nv_fp6_e2m3 */
    CUDA_R_6F_E3M2  = 32,  /* real as a nv_fp6_e3m2 */
    CUDA_R_4F_E2M1  = 33,  /* real as a nv_fp4_e2m1 */

R_8F_UE4M3 is an alias of R_8F_E4M3, differing semantically in that the sign is meaningless in the NVFP4 block-scaling format where every element has a sign anyway.

Implementations like DLFP8Types.jl and Microfloats.jl could define e.g. Base.convert(::Type{CUDACore.cudaDataType}, ::Type{Float8_E5M2}) = CUDACore.R_8F_E5M2 in an extension, or there could be some jltype_to_cudaDataType function.

It doesn't work the other way though, since each cudaDataType can only map to one julia type, but not sure if this is a problem.

[github-actions]

CUDA.jl Benchmarks

Details

Benchmark suite	Current: 603e01b	Previous: fbb9098	Ratio
array/accumulate/Float32/1d	98650 ns	98020 ns	1.01
array/accumulate/Float32/dims=1	74613 ns	73968 ns	1.01
array/accumulate/Float32/dims=1L	1599478 ns	1598577 ns	1.00
array/accumulate/Float32/dims=2	140349 ns	139829 ns	1.00
array/accumulate/Float32/dims=2L	660575 ns	660479 ns	1.00
array/accumulate/Int64/1d	118541 ns	118236 ns	1.00
array/accumulate/Int64/dims=1	78538 ns	78818 ns	1.00
array/accumulate/Int64/dims=1L	1716721 ns	1717142 ns	1.00
array/accumulate/Int64/dims=2	153234 ns	151913 ns	1.01
array/accumulate/Int64/dims=2L	986806 ns	987639 ns	1.00
array/broadcast	18155 ns	18185 ns	1.00
array/construct	915.21875 ns	900.1904761904761 ns	1.02
array/copy	16092 ns	16081 ns	1.00
array/copyto!/cpu_to_gpu	209944 ns	210623 ns	1.00
array/copyto!/gpu_to_cpu	241959 ns	242406 ns	1.00
array/copyto!/gpu_to_gpu	10027 ns	8883.666666666666 ns	1.13
array/iteration/findall/bool	132638 ns	132352 ns	1.00
array/iteration/findall/int	147033 ns	146447 ns	1.00
array/iteration/findfirst/bool	68498 ns	67784 ns	1.01
array/iteration/findfirst/int	69743 ns	69220 ns	1.01
array/iteration/findmin/1d	62791 ns	62612 ns	1.00
array/iteration/findmin/2d	100452 ns	99990 ns	1.00
array/iteration/logical	187356 ns	186900 ns	1.00
array/iteration/scalar	61927 ns	62246 ns	0.99
array/permutedims/2d	49356 ns	49143 ns	1.00
array/permutedims/3d	50426 ns	49912 ns	1.01
array/permutedims/4d	49671 ns	50252 ns	0.99
array/random/rand/Float32	10979 ns	10722 ns	1.02
array/random/rand/Int64	21140 ns	20792 ns	1.02
array/random/rand!/Float32	7629 ns	7728 ns	0.99
array/random/rand!/Int64	20269 ns	20553 ns	0.99
array/random/randn/Float32	33042 ns	33301 ns	0.99
array/random/randn!/Float32	23523 ns	24939 ns	0.94
array/reductions/mapreduce/Float32/1d	31277 ns	31290 ns	1.00
array/reductions/mapreduce/Float32/dims=1	37569 ns	37312 ns	1.01
array/reductions/mapreduce/Float32/dims=1L	50264 ns	50036 ns	1.00
array/reductions/mapreduce/Float32/dims=2	54868 ns	54292 ns	1.01
array/reductions/mapreduce/Float32/dims=2L	66489 ns	65880 ns	1.01
array/reductions/mapreduce/Int64/1d	38437 ns	39323 ns	0.98
array/reductions/mapreduce/Int64/dims=1	40493 ns	40145 ns	1.01
array/reductions/mapreduce/Int64/dims=1L	88023 ns	87779 ns	1.00
array/reductions/mapreduce/Int64/dims=2	57305 ns	57093 ns	1.00
array/reductions/mapreduce/Int64/dims=2L	82670 ns	82500 ns	1.00
array/reductions/reduce/Float32/1d	31665 ns	31207 ns	1.01
array/reductions/reduce/Float32/dims=1	37442 ns	37284 ns	1.00
array/reductions/reduce/Float32/dims=1L	50207 ns	50034 ns	1.00
array/reductions/reduce/Float32/dims=2	54870 ns	54677 ns	1.00
array/reductions/reduce/Float32/dims=2L	67734 ns	67287 ns	1.01
array/reductions/reduce/Int64/1d	38022 ns	38695 ns	0.98
array/reductions/reduce/Int64/dims=1	39940 ns	39938 ns	1.00
array/reductions/reduce/Int64/dims=1L	88256 ns	87933 ns	1.00
array/reductions/reduce/Int64/dims=2	57034 ns	56801 ns	1.00
array/reductions/reduce/Int64/dims=2L	82019 ns	82417 ns	1.00
array/reverse/1d	16257 ns	16212 ns	1.00
array/reverse/1dL	68999 ns	69012 ns	1.00
array/reverse/1dL_inplace	67048 ns	66891 ns	1.00
array/reverse/1d_inplace	9634.666666666666 ns	9703.666666666666 ns	0.99
array/reverse/2d	19422 ns	19393 ns	1.00
array/reverse/2dL	72757 ns	72842 ns	1.00
array/reverse/2dL_inplace	66722 ns	66766 ns	1.00
array/reverse/2d_inplace	10315 ns	10134 ns	1.02
array/sorting/1d	2654767 ns	2658542 ns	1.00
array/sorting/2d	1037891 ns	1039211 ns	1.00
array/sorting/by	3192053 ns	3193897 ns	1.00
cuda/synchronization/context/auto	1028.9 ns	1040 ns	0.99
cuda/synchronization/context/blocking	783.5858585858585 ns	795.5444444444445 ns	0.98
cuda/synchronization/context/nonblocking	5870 ns	5738.666666666667 ns	1.02
cuda/synchronization/stream/auto	879.8301886792453 ns	889.2083333333334 ns	0.99
cuda/synchronization/stream/blocking	675.8141025641025 ns	676.8387096774194 ns	1.00
cuda/synchronization/stream/nonblocking	5741.714285714285 ns	5562.285714285715 ns	1.03
integration/byval/reference	147462 ns	147385 ns	1.00
integration/byval/slices=1	149274 ns	149634 ns	1.00
integration/byval/slices=2	292306 ns	292176 ns	1.00
integration/byval/slices=3	434580 ns	435096 ns	1.00
integration/cudadevrt	104483 ns	104391 ns	1.00
integration/volumerhs	9307671 ns	9304645 ns	1.00
kernel/indexing	12574 ns	12267 ns	1.03
kernel/indexing_checked	13247 ns	13063 ns	1.01
kernel/launch	2014.111111111111 ns	1980.4 ns	1.02
kernel/occupancy	722.2867647058823 ns	638.797619047619 ns	1.13
kernel/rand	13564 ns	14927 ns	0.91
latency/import	3912069945 ns	3890391684 ns	1.01
latency/precompile	4687020413 ns	4668709973 ns	1.00
latency/ttfp	4916672311 ns	4883527794 ns	1.01

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

[AntonOresten]

@maleadt How, and where, could packages define DataType to cudaDataType mappings? Is this considered public API? If not, would it need to be in extensions of CUDACore?

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 17.42%. Comparing base (fbb9098) to head (603e01b).

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #3180      +/-   ##
==========================================
- Coverage   17.44%   17.42%   -0.03%     
==========================================
  Files         124      124              
  Lines        9883     9883              
==========================================
- Hits         1724     1722       -2     
- Misses       8159     8161       +2     

☔ 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.