Fix segfault in qr(dA) \ b

[luraess]

Problem

On Julia 1.10+, adjoint(Q::AbstractQ) returns LinearAlgebra.AdjointQ, not Adjoint, making the six existing lmul!/rmul! overloads on Adjoint{T,<:QRPackedQ} / Transpose{T,<:QRPackedQ} dead code. Julia 1.12 then added:

lmul!(adjQ::AdjointQ{<:Any,<:QRPackedQ{T,<:StridedMatrix}}, B::StridedVecOrMat{T})

Since ROCArray <: StridedArray, this matches GPU arrays and calls CPU OpenBLAS with a GPU pointer → segfault.

qr(dA) \ b was also broken independently: stdlib's \(Factorization, b) allocates the output with zeros(T, n) — always a CPU array — then passes it through ldiv! mixing GPU factors with CPU data → segfault.

Fix

• Replace the six dead Adjoint/Transpose-based overloads with AdjointQ-based ones, widening B to StridedROCVecOrMat to also cover SubArray views produced by stdlib internals.
• Add Base.:\ overloads for QR{T,<:ROCMatrix,<:ROCVector} that call ldiv!(F, copy(b)) directly, bypassing the zeros(T, n) CPU allocation.
• Widen ormqr! / LAPACK.ormqr! to accept StridedROCVecOrMat.

Tests

• qr(dA) \ b and qr(dA) \ B (tall matrix, vector and matrix RHS)
• Explicit lmul!/rmul! for Q, Q', and transpose(Q) (real types)
• ldiv!(qr(dA), dx) and ldiv!(dy, qr(dA), dx)

[mfherbst]

Awesome, thanks @luraess !

[github-actions]

AMDGPU.jl Benchmarks

Details

Benchmark suite	Current: 876515b	Previous: 4ba6e0b	Ratio
amdgpu/synchronization/context/device	600 ns	620 ns	0.97
amdgpu/synchronization/stream/blocking	250 ns	250 ns	1
amdgpu/synchronization/stream/nonblocking	340 ns	350 ns	0.97
array/accumulate/Float32/1d	84371 ns	86201 ns	0.98
array/accumulate/Float32/dims=1	263373 ns	319185 ns	0.83
array/accumulate/Float32/dims=1L	135672 ns	134892 ns	1.01
array/accumulate/Float32/dims=2	133672 ns	129962 ns	1.03
array/accumulate/Float32/dims=2L	2806689 ns	2803440 ns	1.00
array/accumulate/Int64/1d	95751 ns	97561 ns	0.98
array/accumulate/Int64/dims=1	245044 ns	243253 ns	1.01
array/accumulate/Int64/dims=1L	167512 ns	166972 ns	1.00
array/accumulate/Int64/dims=2	126312 ns	126982 ns	0.99
array/accumulate/Int64/dims=2L	2985611 ns	2991421 ns	1.00
array/broadcast	88781 ns	91982 ns	0.97
array/construct	1740 ns	1690 ns	1.03
array/copy	39211 ns	37151 ns	1.06
array/copyto!/cpu_to_gpu	182582 ns	182603 ns	1.00
array/copyto!/gpu_to_cpu	181833 ns	182843 ns	0.99
array/copyto!/gpu_to_gpu	82551 ns	61041 ns	1.35
array/iteration/findall/bool	180702 ns	179112 ns	1.01
array/iteration/findall/int	198143 ns	187843 ns	1.05
array/iteration/findfirst/bool	117522 ns	119701 ns	0.98
array/iteration/findfirst/int	115812 ns	116352 ns	1.00
array/iteration/findmin/1d	169172 ns	169822 ns	1.00
array/iteration/findmin/2d	155302 ns	155632 ns	1.00
array/iteration/logical	349604 ns	351185 ns	1.00
array/iteration/scalar	289124 ns	292104 ns	0.99
array/permutedims/2d	73961 ns	72651 ns	1.02
array/permutedims/3d	73041 ns	73951 ns	0.99
array/permutedims/4d	75891 ns	75441 ns	1.01
array/random/rand/Float32	50491 ns	54361 ns	0.93
array/random/rand/Int64	56910 ns	55600 ns	1.02
array/random/rand!/Float32	125862 ns	83711 ns	1.50
array/random/rand!/Int64	96432 ns	146362 ns	0.66
array/random/randn/Float32	90511 ns	93562 ns	0.97
array/random/randn!/Float32	123812 ns	110852 ns	1.12
array/reductions/mapreduce/Float32/1d	129542 ns	133492 ns	0.97
array/reductions/mapreduce/Float32/dims=1	94982 ns	95501 ns	0.99
array/reductions/mapreduce/Float32/dims=1L	774391 ns	774131 ns	1.00
array/reductions/mapreduce/Float32/dims=2	96771 ns	97341 ns	0.99
array/reductions/mapreduce/Float32/dims=2L	301224 ns	297434 ns	1.01
array/reductions/mapreduce/Int64/1d	133642 ns	133462 ns	1.00
array/reductions/mapreduce/Int64/dims=1	95391 ns	94862 ns	1.01
array/reductions/mapreduce/Int64/dims=1L	783531 ns	781601 ns	1.00
array/reductions/mapreduce/Int64/dims=2	96841 ns	95151 ns	1.02
array/reductions/mapreduce/Int64/dims=2L	299504 ns	298665 ns	1.00
array/reductions/reduce/Float32/1d	133552 ns	129141 ns	1.03
array/reductions/reduce/Float32/dims=1	94731 ns	95271 ns	0.99
array/reductions/reduce/Float32/dims=1L	774031 ns	772951 ns	1.00
array/reductions/reduce/Float32/dims=2	96971 ns	97282 ns	1.00
array/reductions/reduce/Float32/dims=2L	296904 ns	306574 ns	0.97
array/reductions/reduce/Int64/1d	133322 ns	133571 ns	1.00
array/reductions/reduce/Int64/dims=1	95271 ns	95052 ns	1.00
array/reductions/reduce/Int64/dims=1L	780231 ns	781570 ns	1.00
array/reductions/reduce/Int64/dims=2	96781 ns	96001 ns	1.01
array/reductions/reduce/Int64/dims=2L	297644 ns	297294 ns	1.00
array/reverse/1d	43131 ns	43120 ns	1.00
array/reverse/1dL	75091 ns	75241 ns	1.00
array/reverse/1dL_inplace	100991 ns	165292 ns	0.61
array/reverse/1d_inplace	72581 ns	111651 ns	0.65
array/reverse/2d	50971 ns	50591 ns	1.01
array/reverse/2dL	101601 ns	101221 ns	1.00
array/reverse/2dL_inplace	104521 ns	178993 ns	0.58
array/reverse/2d_inplace	82771 ns	136752 ns	0.61
array/sorting/1d	342115 ns	342394 ns	1.00
integration/byval/reference	39010 ns	39281 ns	0.99
integration/byval/slices=1	40430 ns	40281 ns	1.00
integration/byval/slices=2	129382 ns	144452 ns	0.90
integration/byval/slices=3	246053 ns	237824 ns	1.03
integration/volumerhs	5030180 ns	5035301 ns	1.00
kernel/indexing	65921 ns	58331 ns	1.13
kernel/indexing_checked	92051 ns	46671 ns	1.97
kernel/launch	1350 ns	1300 ns	1.04
kernel/rand	195823 ns	127022 ns	1.54
latency/import	1587127731 ns	1567060274 ns	1.01
latency/precompile	36233947041 ns	36346768726 ns	1.00
latency/ttfp	2156775563 ns	2138341786 ns	1.01

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