Try fast linear indexes for KA

[vchuravy]

I am unsure why we couldn't do that in the beginning.

[maleadt]

Will have to wait for #2593 to get merged.

[github-actions]

CUDA.jl Benchmarks

Benchmark suite	Current: c23ab7f	Previous: 14ae82d	Ratio
latency/precompile	45424277593.5 ns	45345622059 ns	1.00
latency/ttfp	6384836106.5 ns	6434638936 ns	0.99
latency/import	3034614553 ns	3051828695.5 ns	0.99
integration/volumerhs	9567744 ns	9568259 ns	1.00
integration/byval/slices=1	146777 ns	146590 ns	1.00
integration/byval/slices=3	425370 ns	425823 ns	1.00
integration/byval/reference	144803 ns	144766 ns	1.00
integration/byval/slices=2	286177 ns	286423 ns	1.00
integration/cudadevrt	103434 ns	103488.5 ns	1.00
kernel/indexing	13938 ns	14282.5 ns	0.98
kernel/indexing_checked	15060 ns	15333 ns	0.98
kernel/occupancy	701.1063829787234 ns	720.4492753623189 ns	0.97
kernel/launch	2124.1111111111113 ns	2130.5 ns	1.00
kernel/rand	16334 ns	17397 ns	0.94
array/reverse/1d	19520 ns	19471 ns	1.00
array/reverse/2d	24603 ns	24536 ns	1.00
array/reverse/1d_inplace	10031.666666666666 ns	10836.333333333334 ns	0.93
array/reverse/2d_inplace	11528 ns	11284 ns	1.02
array/copy	20270 ns	20310 ns	1.00
array/iteration/findall/int	159097 ns	158042 ns	1.01
array/iteration/findall/bool	139369 ns	138224 ns	1.01
array/iteration/findfirst/int	153853 ns	154038.5 ns	1.00
array/iteration/findfirst/bool	154627.5 ns	155126 ns	1.00
array/iteration/scalar	75657 ns	76714 ns	0.99
array/iteration/logical	207799 ns	214056.5 ns	0.97
array/iteration/findmin/1d	41128 ns	41628 ns	0.99
array/iteration/findmin/2d	94766 ns	94463 ns	1.00
array/reductions/reduce/1d	38659 ns	51305 ns	0.75
array/reductions/reduce/2d	44155.5 ns	42302 ns	1.04
array/reductions/mapreduce/1d	37246.5 ns	44898.5 ns	0.83
array/reductions/mapreduce/2d	51913.5 ns	52966.5 ns	0.98
array/broadcast	21698 ns	21607 ns	1.00
array/copyto!/gpu_to_gpu	11663 ns	13399 ns	0.87
array/copyto!/cpu_to_gpu	213248 ns	213579.5 ns	1.00
array/copyto!/gpu_to_cpu	246883 ns	245985.5 ns	1.00
array/accumulate/1d	108538 ns	109003 ns	1.00
array/accumulate/2d	79961 ns	79807 ns	1.00
array/construct	1197.35 ns	1147.9 ns	1.04
array/random/randn/Float32	43009 ns	43138 ns	1.00
array/random/randn!/Float32	26240 ns	26215 ns	1.00
array/random/rand!/Int64	27084 ns	27096 ns	1.00
array/random/rand!/Float32	8824.666666666666 ns	8869.333333333334 ns	0.99
array/random/rand/Int64	29684 ns	29884 ns	0.99
array/random/rand/Float32	12772 ns	12925 ns	0.99
array/permutedims/4d	65015 ns	67255 ns	0.97
array/permutedims/2d	56278 ns	56783 ns	0.99
array/permutedims/3d	60503.5 ns	58969.5 ns	1.03
array/sorting/1d	2920400.5 ns	2933376.5 ns	1.00
array/sorting/by	3499981 ns	3499572.5 ns	1.00
array/sorting/2d	1084450 ns	1084491.5 ns	1.00
cuda/synchronization/stream/auto	1027.8 ns	1039.3 ns	0.99
cuda/synchronization/stream/nonblocking	6532.2 ns	6569.6 ns	0.99
cuda/synchronization/stream/blocking	804.8350515463917 ns	796.7647058823529 ns	1.01
cuda/synchronization/context/auto	1166.8 ns	1224.5 ns	0.95
cuda/synchronization/context/nonblocking	6741.4 ns	6745.4 ns	1.00
cuda/synchronization/context/blocking	891.9583333333334 ns	915.2391304347826 ns	0.97

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

[maleadt]

A handful of tests fail on this PR:

julia> CUDA.@sync sum!(CUDA.rand(Float64, (1, 1025, 2)), CUDA.rand(Float64, (1, 1025, 2)))
ERROR: a BoundsError was thrown during kernel execution on thread (513, 1, 1) in block (3, 1, 1).
Out-of-bounds array access
Stacktrace:
 [1] throw_boundserror at /home/tim/Julia/pkg/CUDA/src/device/quirks.jl:15
 [2] multiple call sites at unknown:0

ERROR: KernelException: exception thrown during kernel execution on device NVIDIA RTX 6000 Ada Generation
Stacktrace:
 [1] check_exceptions()
   @ CUDA ~/Julia/pkg/CUDA/src/compiler/exceptions.jl:39
 [2] synchronize(stream::CuStream; blocking::Bool, spin::Bool)
   @ CUDA ~/Julia/pkg/CUDA/lib/cudadrv/synchronization.jl:207
 [3] synchronize
   @ ~/Julia/pkg/CUDA/lib/cudadrv/synchronization.jl:194 [inlined]
 [4] macro expansion
   @ ~/Julia/pkg/CUDA/src/utilities.jl:36 [inlined]
 [5] top-level scope
   @ REPL[44]:1

Not a catastrophic amount though, so probably worth looking into?

Test Summary:                                  |  Pass  Error  Broken  Total  Time
  Overall                                      | 25854     15      11  25880

[maleadt]

Not [...] catastrophic though

Well, not so sure about that

julia> CUDA.@sync CUDA.ones(Float64, (1, 1025, 2));
ERROR: a BoundsError was thrown during kernel execution on thread (353, 1, 1) in block (4, 1, 1).
Out-of-bounds array access
Stacktrace:
 [1] throw_boundserror at /home/tim/Julia/pkg/CUDA/src/device/quirks.jl:15
 [2] multiple call sites at unknown:0

[maleadt]

MWE for the bounds error:

function main()
    A = CuArray{Float64}(undef, (1, 1025, 2))

    @kernel function fill_kernel!(a)
        idx = @index(Global, Linear)
        if idx >= length(a)
            if idx == length(a)+1
                @cushow threadIdx().x blockDim().x blockIdx().x gridDim().x idx
            end
        else
            a[idx] = 0f0
        end
    end

    kernel = fill_kernel!(get_backend(A))
    CUDA.@sync kernel(A; ndrange = size(A))
end

The linear index here goes out of bounds for a lot of threads, so I limited to only printing about the first one:

(threadIdx()).x = 259
(blockDim()).x = 896
(blockIdx()).x = 3
(gridDim()).x = 4
idx = 2051

The launch configuration is strange: 4 blocks of 896 threads covers 3584 items, while 3 blocks would have been sufficient by covering 2688 out of 2050, no? In any case, it's also strange that this isn't detected by the bounds check I presume @index is doing...

[maleadt]

The launch configuration is strange: 4 blocks of 896 threads covers 3584 items, while 3 blocks would have been sufficient by covering 2688 out of 2050, no?

This is how KA's launch configuration determines that:

ndrange = (1, 1025, 2)
config.threads = 896
workgroupsize = threads_to_workgroupsize(threads, ndrange) = (1, 896, 1)
KA.blocks(iterspace) = CartesianIndices((1, 2, 2))
KA.workitems(iterspace) = CartesianIndices((1, 896, 1))
blocks = length(KA.blocks(iterspace)) = 4
threads = length(KA.workitems(iterspace)) = 896

Regardless of the (somehow) missing bounds check here, it seems very wasteful to launch (1, 2, 2) = 4 groups to cover (1, 1025, 2) items if we end up generating a ND index from a linear one anyway (i.e., without having to cover each dimension individually)...

@vchuravy I'll defer to you on this.