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: f41f373	Previous: bc81d40	Ratio
array/accumulate/Float32/1d	98915 ns	98667 ns	1.00
array/accumulate/Float32/dims=1	74908 ns	74079 ns	1.01
array/accumulate/Float32/dims=1L	1599681 ns	1599572 ns	1.00
array/accumulate/Float32/dims=2	140661 ns	139645 ns	1.01
array/accumulate/Float32/dims=2L	660907 ns	660578 ns	1.00
array/accumulate/Int64/1d	118396 ns	118201 ns	1.00
array/accumulate/Int64/dims=1	79173 ns	79258 ns	1.00
array/accumulate/Int64/dims=1L	1716336 ns	1715309 ns	1.00
array/accumulate/Int64/dims=2	153488 ns	153091 ns	1.00
array/accumulate/Int64/dims=2L	987716 ns	987286 ns	1.00
array/broadcast	18273 ns	18223 ns	1.00
array/construct	1078.1 ns	1098.7 ns	0.98
array/copy	16657 ns	16605 ns	1.00
array/copyto!/cpu_to_gpu	208688 ns	206529 ns	1.01
array/copyto!/gpu_to_cpu	240551 ns	239820 ns	1.00
array/copyto!/gpu_to_gpu	8901 ns	8819 ns	1.01
array/iteration/findall/bool	132729 ns	133125 ns	1.00
array/iteration/findall/int	146300 ns	146351 ns	1.00
array/iteration/findfirst/bool	68470 ns	69131 ns	0.99
array/iteration/findfirst/int	69955 ns	70202 ns	1.00
array/iteration/findmin/1d	65436 ns	62891 ns	1.04
array/iteration/findmin/2d	99866 ns	99773 ns	1.00
array/iteration/logical	188593 ns	188295 ns	1.00
array/iteration/scalar	63264 ns	63035 ns	1.00
array/permutedims/2d	49404 ns	48726 ns	1.01
array/permutedims/3d	51027 ns	50552 ns	1.01
array/permutedims/4d	50367 ns	50019 ns	1.01
array/random/rand/Float32	11843 ns	11718 ns	1.01
array/random/rand/Int64	23589 ns	21972 ns	1.07
array/random/rand!/Float32	7898.5 ns	7684.75 ns	1.03
array/random/rand!/Int64	19613 ns	17879 ns	1.10
array/random/randn/Float32	33933 ns	33563 ns	1.01
array/random/randn!/Float32	23887 ns	24207 ns	0.99
array/reductions/mapreduce/Float32/1d	31916 ns	31970 ns	1.00
array/reductions/mapreduce/Float32/dims=1	37777 ns	37560 ns	1.01
array/reductions/mapreduce/Float32/dims=1L	50858 ns	50748 ns	1.00
array/reductions/mapreduce/Float32/dims=2	54915 ns	54859 ns	1.00
array/reductions/mapreduce/Float32/dims=2L	66733 ns	66258 ns	1.01
array/reductions/mapreduce/Int64/1d	38705 ns	38667 ns	1.00
array/reductions/mapreduce/Int64/dims=1	40340 ns	40569 ns	0.99
array/reductions/mapreduce/Int64/dims=1L	88805 ns	88767 ns	1.00
array/reductions/mapreduce/Int64/dims=2	57540 ns	57046 ns	1.01
array/reductions/mapreduce/Int64/dims=2L	83176 ns	83249 ns	1.00
array/reductions/reduce/Float32/1d	31908 ns	31908 ns	1
array/reductions/reduce/Float32/dims=1	37850 ns	37834 ns	1.00
array/reductions/reduce/Float32/dims=1L	50960 ns	50899 ns	1.00
array/reductions/reduce/Float32/dims=2	55366 ns	54907 ns	1.01
array/reductions/reduce/Float32/dims=2L	68197 ns	68257 ns	1.00
array/reductions/reduce/Int64/1d	39430 ns	38884 ns	1.01
array/reductions/reduce/Int64/dims=1	40342 ns	40427 ns	1.00
array/reductions/reduce/Int64/dims=1L	88629 ns	88694 ns	1.00
array/reductions/reduce/Int64/dims=2	57334 ns	57232 ns	1.00
array/reductions/reduce/Int64/dims=2L	82733 ns	82964 ns	1.00
array/reverse/1d	16912 ns	17058 ns	0.99
array/reverse/1dL	69640 ns	69721 ns	1.00
array/reverse/1dL_inplace	67005 ns	66952 ns	1.00
array/reverse/1d_inplace	9601.666666666666 ns	9893.666666666666 ns	0.97
array/reverse/2d	19775 ns	19541 ns	1.01
array/reverse/2dL	73114 ns	72830 ns	1.00
array/reverse/2dL_inplace	66691 ns	66851 ns	1.00
array/reverse/2d_inplace	9972 ns	10080 ns	0.99
array/sorting/1d	2655411 ns	2667782 ns	1.00
array/sorting/2d	1039053 ns	1038338 ns	1.00
array/sorting/by	3193248 ns	3193204 ns	1.00
cuda/synchronization/context/auto	1042.9 ns	1031.5 ns	1.01
cuda/synchronization/context/blocking	813.9222222222222 ns	791.2282608695652 ns	1.03
cuda/synchronization/context/nonblocking	5797.833333333333 ns	5839.5 ns	0.99
cuda/synchronization/stream/auto	886 ns	880.5576923076923 ns	1.01
cuda/synchronization/stream/blocking	685.4183006535948 ns	677.7908496732026 ns	1.01
cuda/synchronization/stream/nonblocking	5644.166666666667 ns	5561.333333333333 ns	1.01
integration/byval/reference	147575 ns	147498 ns	1.00
integration/byval/slices=1	149611 ns	149705 ns	1.00
integration/byval/slices=2	292499 ns	292393 ns	1.00
integration/byval/slices=3	435006 ns	435188 ns	1.00
integration/cudadevrt	104601 ns	104517 ns	1.00
integration/volumerhs	9305642 ns	9306993 ns	1.00
kernel/indexing	12363 ns	12494 ns	0.99
kernel/indexing_checked	13190 ns	13141 ns	1.00
kernel/launch	2010 ns	2103.777777777778 ns	0.96
kernel/occupancy	642.6686746987951 ns	683.375 ns	0.94
kernel/rand	14890 ns	13705 ns	1.09
latency/import	3889951490 ns	3893015807 ns	1.00
latency/precompile	4658193847 ns	4656329210 ns	1.00
latency/ttfp	5403708128 ns	5439496872 ns	0.99

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