Array + Diagonal failure

[vchuravy]

using LinearAlgebra
using AMDGPU

            AT = ROCArray
            n = 128
            A = AT(rand(Float32, (n,n)))
            d = AT(rand(Float32, n))
            D = Diagonal(d)
            B = A + D
            collect(B) ≈ collect(A) + collect(D)

julia> collect(B) .≈ collect(A) + collect(D)
128×128 BitMatrix:
 1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  …  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  1
 1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1     0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  1
 1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1     0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  1  0  1
 1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1     0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  1  0  1
 1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1     0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  1  0  1  0  1
 1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  …  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  1  0  1  0  1
 ⋮              ⋮              ⋮              ⋮              ⋮              ⋱        ⋮              ⋮              ⋮              ⋮              ⋮     
 1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1     1  0  1  0  1  0  1  0  1  0  1  0  1  0  1  0  1  0  1  0  1  0  1  0  1
 1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1     1  0  1  0  1  0  1  0  1  0  1  0  1  0  1  0  1  0  1  0  1  0  1  0  1
 1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  …  1  0  1  0  1  0  1  0  1  0  1  0  1  0  1  0  1  0  1  0  1  0  1  0  1
 1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1     1  0  1  0  1  0  1  0  1  0  1  0  1  0  1  0  1  0  1  0  1  0  1  0  1
 1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1     1  0  1  0  1  0  1  0  1  0  1  0  1  0  1  0  1  0  1  0  1  0  1  0  1

julia> A + D
128×128 ROCMatrix{Float32}:
 1.02106     0.0921212  0.279667    0.266254  0.985505   0.174462  0.740154   …  -1.71465f38   -1.71465f38   -1.71465f38   -1.71465f38   0.644539
 0.00559413  1.56396    0.658443    0.456375  0.0258112  0.208886  0.537066      -1.71465f38   -1.71465f38   -1.71465f38   -1.71465f38   0.163299
 0.958081    0.841537   0.994774    0.200046  0.808281   0.113412  0.748186       0.0           0.0           0.300136      0.0          0.632184
 0.672491    0.97533    0.125371    0.71167   0.586824   0.630625  0.566512       0.0           0.0           0.279454      0.0          0.729033
 0.138161    0.348633   0.959647    0.120322  0.831747   0.52287   0.677146       0.76792       0.0           0.541541      0.0          0.44386
 0.81552     0.554076   0.601219    0.583099  0.406613   0.50956   0.0149934  …   0.59062       0.0           0.211465      0.0          0.425578
 0.585995    0.618082   0.875589    0.689613  0.342014   0.951317  1.47807        0.996348      0.0           0.472449      0.0          0.810768
 0.648932    0.869043   0.00508428  0.481415  0.594282   0.954457  0.594401       0.804497      0.0           0.706352      0.0          0.227551
 0.950201    0.798405   0.751419    0.431937  0.392388   0.167172  0.996166       0.437932     -5.1515f-9     0.563016      5.89816f-34  0.889226
 0.966268    0.604634   0.464085    0.72118   0.0115883  0.173688  0.0464908      0.193675     -4.392f-19     0.599506      6.0f-45      0.3995
 ⋮                                                       ⋮                    ⋱                               ⋮                          
 0.463826    0.301494   0.426812    0.231694  0.49323    0.658562  0.627726       0.398807      9.49213f-37   0.491222    NaN            0.764548
 0.111097    0.168921   0.435184    0.803009  0.22741    0.325752  0.650371   …   0.312124     -2.26569f38    0.870685    NaN            0.277152
 0.0783372   0.802934   0.973153    0.457144  0.892521   0.739615  0.0400927      0.768561     -1.9459f38     0.712432    NaN            0.238672
 0.235187    0.72732    0.214835    0.183458  0.343981   0.182465  0.526841       0.807266     -1.9459f38     0.279783    NaN            0.584277
 0.788909    0.325859   0.229556    0.362396  0.110689   0.535112  0.878508       0.990009    NaN             0.754234    NaN            0.720826
 0.460831    0.487539   0.191096    0.74637   0.925338   0.238311  0.499134       0.610513     -2.29238f38    0.726175    NaN            0.858912
 0.972706    0.259302   0.480788    0.25185   0.677029   0.156736  0.14716    …   0.339408    NaN             1.05132     NaN            0.330685
 0.446682    0.439278   0.630035    0.893114  0.324872   0.422373  0.0394168      0.957419    NaN             0.442314    NaN            0.733302
 0.0891961   0.112863   0.858945    0.131734  0.760471   0.675587  0.446861       0.923527    NaN             0.538234    NaN            0.867137

This is on #163

[vchuravy]

Looks like grid-stride broadcast JuliaGPU/GPUArrays.jl#367

[vchuravy]

julia> b .= 0
(groupsize, gridsize) = (threads, blocks * threads) = (256, 8192)
128×128 ROCMatrix{Float32}:
 0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  …  -0.372549  -0.372549  -0.372549  -0.372549  -0.372549  -0.372549  0.0
 0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0     -0.372549  -0.372549  -0.372549  -0.372549  -0.372549  -0.372549  0.0
 0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0     -0.372549  -0.372549  -0.372549  -0.372549   0.0       -0.372549  0.0
 0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0     -0.372549  -0.372549  -0.372549  -0.372549   0.0       -0.372549  0.0
 0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0     -0.372549  -0.372549   0.0       -0.372549   0.0       -0.372549  0.0
 0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  …  -0.372549  -0.372549   0.0       -0.372549   0.0       -0.372549  0.0
 0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0      0.0       -0.372549   0.0       -0.372549   0.0       -0.372549  0.0
 0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0      0.0       -0.372549   0.0       -0.372549   0.0       -0.372549  0.0
 0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0      0.0       -0.372549   0.0       -0.372549   0.0       -0.372549  0.0
 0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0      0.0       -0.372549   0.0       -0.372549   0.0       -0.372549  0.0
 ⋮                        ⋮                        ⋮                        ⋱                                               ⋮                    
 0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0      0.0       -0.372549   0.0       -0.372549   0.0       -0.372549  0.0
 0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  …   0.0       -0.372549   0.0       -0.372549   0.0       -0.372549  0.0
 0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0      0.0       -0.372549   0.0       -0.372549   0.0       -0.372549  0.0
 0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0      0.0       -0.372549   0.0       -0.372549   0.0       -0.372549  0.0
 0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0      0.0       -0.372549   0.0       -0.372549   0.0       -0.372549  0.0
 0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0      0.0       -0.372549   0.0       -0.372549   0.0       -0.372549  0.0
 0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  …   0.0       -0.372549   0.0       -0.372549   0.0       -0.372549  0.0
 0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0      0.0       -0.372549   0.0       -0.372549   0.0       -0.372549  0.0
 0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0      0.0       -0.372549   0.0       -0.372549   0.0       -0.372549  0.0

julia> 128*128
16384

So yeah something not quite right.