Handle hipTENSOR 2.4: fix the version decode, reject the layouts it miscomputes

[luraess]

hipTENSOR 2.4 (the version in the ROCm 10.0 bundles) no longer uses the mode labels to line a contraction's operands up. It reads every descriptor positionally, and computes the right result only when the operands are packed column-major and indexed as

• A = [M…, K…] — A's free modes, then the contracted ones
• B = [N…, K…]
• C = D = [M…, N…] — A's free modes, then B's

Every other layout is silently miscomputed: hiptensorContract returns HIPTENSOR_STATUS_SUCCESS and writes a plausible but wrong result. hipTENSOR 2.2 handled arbitrary mode orders, which is what the wrapper was written against.

Of the 120 shapes test/hiptensor/contractions.jl generates, 115 came out wrong on an MI300A, with relative errors of 0.3–1.2. Reordering the descriptors alone fixes only 80 of them; all 120 are correct once every operand is packed and canonical.

Two related changes in 2.4:

• HIPTENSOR_OP_CONJ on a contraction's input is now rejected — hiptensorCreatePlan fails with HIPTENSOR_STATUS_EXECUTION_FAILED. On 2.2 it was accepted and silently ignored.
• hiptensorGetVersion() returns 2004000, i.e. major * 1_000_000 + minor * 1_000 + patch. version() decoded it cuTENSOR-style and reported the library as v"200.40.0".

Changes

• fix the version() decode, so hipTENSOR 2.4.0 reports v"2.4.0"
• add check_contraction_layout, called from plan_contraction: on hipTENSOR ≥ 2.4, reject any layout the library would miscompute, naming the index order each operand needs. Below 2.4 it returns immediately, so 2.2 is untouched.
• document both restrictions in docs/src/libraries/tensor.md
• tests use random mode orders below 2.4 and the canonical layout at or above it, expect hipTENSORError from OP_CONJ on 2.4, and gain a test asserting the guard fires

Making arbitrary layouts work on 2.4 would mean materialising permuted copies of every operand — a copy of A, B and C on most calls — so that is left until upstream says whether the layout requirement is intended.

Testing

Verified on an MI300A (gfx942) against hipTENSOR 2.4.0 from the ROCm 10.0 gfx94X-dcgpu bundle: test/hiptensor/contractions.jl passes 433/433, no failures or broken tests.

Not tested against hipTENSOR 2.2. That path is unchanged by construction.

[github-actions]

AMDGPU.jl Benchmarks

Details

Benchmark suite	Current: 0febfd4	Previous: 8a165c3	Ratio
amdgpu/synchronization/context/device	550 ns	545 ns	1.01
amdgpu/synchronization/stream/blocking	232.5 ns	230 ns	1.01
amdgpu/synchronization/stream/nonblocking	310 ns	305 ns	1.02
applications/bitonic_sort	1644020.5 ns	1602009.25 ns	1.03
applications/convolution	107476.5 ns	110114 ns	0.98
applications/floyd_warshall	11769463.25 ns	11791724.5 ns	1.00
applications/histogram	779653.5 ns	772268 ns	1.01
applications/prefix_sum	253046.25 ns	252573.5 ns	1.00
array/accumulate/Float32/1d	72801 ns	77508.5 ns	0.94
array/accumulate/Float32/dims=1	270053.75 ns	276093.75 ns	0.98
array/accumulate/Float32/dims=1L	80783.75 ns	88671.25 ns	0.91
array/accumulate/Float32/dims=2	71091 ns	81863.75 ns	0.87
array/accumulate/Float32/dims=2L	2759733.75 ns	2758274.75 ns	1.00
array/accumulate/Int64/1d	77183.5 ns	80568.5 ns	0.96
array/accumulate/Int64/dims=1	242468.5 ns	243593.5 ns	1.00
array/accumulate/Int64/dims=1L	84061.25 ns	84163.5 ns	1.00
array/accumulate/Int64/dims=2	83556 ns	87041.25 ns	0.96
array/accumulate/Int64/dims=2L	3065080.75 ns	3136242.5 ns	0.98
array/broadcast	73051 ns	72426 ns	1.01
array/construct	2237.5 ns	2092.75 ns	1.07
array/copy	37935.75 ns	37288 ns	1.02
array/copyto!/cpu_to_gpu	110714 ns	111639 ns	0.99
array/copyto!/gpu_to_cpu	120401.75 ns	111944 ns	1.08
array/copyto!/gpu_to_gpu	53555.75 ns	58595.75 ns	0.91
array/iteration/findall/bool	135012 ns	136929.25 ns	0.99
array/iteration/findall/int	149902 ns	149374.5 ns	1.00
array/iteration/findfirst/bool	187187.75 ns	184972.5 ns	1.01
array/iteration/findfirst/int	164017.25 ns	144249.25 ns	1.14
array/iteration/findmin/1d	109871.75 ns	120894 ns	0.91
array/iteration/findmin/2d	96051.25 ns	111526.5 ns	0.86
array/iteration/logical	240710.75 ns	244013.25 ns	0.99
array/iteration/scalar	297699.25 ns	291824 ns	1.02
array/permutedims/2d	71858.5 ns	71388.5 ns	1.01
array/permutedims/3d	71101 ns	70801 ns	1.00
array/permutedims/4d	73563.5 ns	73643.5 ns	1.00
array/random/rand/Float32	45800.5 ns	45148 ns	1.01
array/random/rand/Int64	53845.75 ns	54070.75 ns	1.00
array/random/rand!/Float32	60326 ns	64578.25 ns	0.93
array/random/rand!/Int64	66528.5 ns	52153.25 ns	1.28
array/random/randn/Float32	77931 ns	80621.25 ns	0.97
array/random/randn!/Float32	78403.75 ns	80881.25 ns	0.97
array/reductions/mapreduce/Float32/1d	87158.75 ns	98356.25 ns	0.89
array/reductions/mapreduce/Float32/dims=1	81936.25 ns	85886.25 ns	0.95
array/reductions/mapreduce/Float32/dims=1L	832621.75 ns	832188.75 ns	1.00
array/reductions/mapreduce/Float32/dims=2	78626.25 ns	77836 ns	1.01
array/reductions/mapreduce/Float32/dims=2L	144369.5 ns	143681.75 ns	1.00
array/reductions/mapreduce/Int64/1d	86563.75 ns	99344 ns	0.87
array/reductions/mapreduce/Int64/dims=1	82146.25 ns	88886.25 ns	0.92
array/reductions/mapreduce/Int64/dims=1L	832046.75 ns	830456.25 ns	1.00
array/reductions/mapreduce/Int64/dims=2	83138.75 ns	86988.75 ns	0.96
array/reductions/mapreduce/Int64/dims=2L	144652 ns	144187 ns	1.00
array/reductions/reduce/Float32/1d	87238.75 ns	98423.75 ns	0.89
array/reductions/reduce/Float32/dims=1	81931 ns	88741.25 ns	0.92
array/reductions/reduce/Float32/dims=1L	826894 ns	836601.5 ns	0.99
array/reductions/reduce/Float32/dims=2	64348.25 ns	77858.5 ns	0.83
array/reductions/reduce/Float32/dims=2L	143877 ns	143437 ns	1.00
array/reductions/reduce/Int64/1d	86956.25 ns	97399 ns	0.89
array/reductions/reduce/Int64/dims=1	81653.75 ns	89021.25 ns	0.92
array/reductions/reduce/Int64/dims=1L	830781.75 ns	830793.75 ns	1.00
array/reductions/reduce/Int64/dims=2	82531.25 ns	83643.75 ns	0.99
array/reductions/reduce/Int64/dims=2L	144274.5 ns	143402 ns	1.01
array/reverse/1d	45750.75 ns	45063.25 ns	1.02
array/reverse/1dL	75896 ns	74268.5 ns	1.02
array/reverse/1dL_inplace	80291 ns	79986.25 ns	1.00
array/reverse/1d_inplace	39923.25 ns	39733 ns	1.00
array/reverse/2d	50470.75 ns	49965.5 ns	1.01
array/reverse/2dL	82436.25 ns	83738.5 ns	0.98
array/reverse/2dL_inplace	91991.5 ns	91603.75 ns	1.00
array/reverse/2d_inplace	63406 ns	63215.75 ns	1.00
array/sorting/1d	332429.75 ns	334934.5 ns	0.99
gemm/tiled	1913989.25 ns	1881007.75 ns	1.02
gemm/tiled_unbounded	1894066.75 ns	1903723.25 ns	0.99
integration/byval/reference	39331 ns	39131 ns	1.01
integration/byval/slices=1	41271 ns	40810 ns	1.01
integration/byval/slices=2	128412 ns	158112 ns	0.81
integration/byval/slices=3	235973 ns	236294 ns	1.00
integration/volumerhs	4922089 ns	4909107 ns	1.00
kernel/indexing	30608 ns	56803.25 ns	0.54
kernel/indexing_checked	58920.75 ns	53248.25 ns	1.11
kernel/launch	1402.5 ns	1397.5 ns	1.00
kernel/rand	98511.5 ns	100381.25 ns	0.98
latency/import	1746754092 ns	1719284175 ns	1.02
latency/precompile	39936573311 ns	39908509786 ns	1.00
latency/ttfp	2313338273 ns	2329845894 ns	0.99
stencil/diffusion3d	1625170.25 ns	1625047 ns	1.00
stencil/diffusion3d_checked	1662773.25 ns	1655270 ns	1.00

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

[luraess]

Changes were comments only. Merging into main in order to have it avail for #1028

[luraess]

cscs-ci run