search

JuliaGPU / oneAPI.jl

Julia support for the oneAPI programming toolkit.
https://juliagpu.org/oneapi/
Other
182 stars 22 forks source link

Bump the toolchain. #405

Closed maleadt closed 6 months ago

maleadt commented 6 months ago

@amontoison: There are SPARSE failures here:

onemkl: Test Failed at /var/lib/buildkite-agent/builds/sagittarius-maleadt-net/julialang/oneapi-dot-jl/test/onemkl.jl:1141
  Expression: alpha * A * x + beta * y ≈ collect(dy)
   Evaluated: Float32[2.6881914, 3.1605413, 1.6428919, 3.1344018, 2.3153603, 1.1546891, 2.7872825, 1.7266464, 2.5613985, 1.3846693] ≈ Float32[0.37818703, 0.46414375, 0.5544208, 0.37786195, 0.09516363, 0.28792638, 0.711177, 0.5808377, 0.3629929, 0.0723116]

onemkl: Test Failed at /var/lib/buildkite-agent/builds/sagittarius-maleadt-net/julialang/oneapi-dot-jl/test/onemkl.jl:1121
  Expression: alpha * opa(A) * opb(B) + beta * C ≈ collect(dC)
   Evaluated: ComplexF32[-0.5649609f0 + 2.3203351f0im -0.66821945f0 + 0.9258573f0im; 0.8666976f0 + 3.6641912f0im 0.54576135f0 + 2.2385116f0im; … ; 0.47543937f0 + 2.0874035f0im -0.17982683f0 + 1.2065787f0im; 0.044930138f0 + 1.962906f0im -0.07940069f0 + 2.2403898f0im] ≈ ComplexF32[-0.42116433f0 + 0.77754575f0im -0.4081933f0 + 0.30477157f0im; -0.4599722f0 + 0.7409284f0im -0.039436415f0 + 0.48696488f0im; … ; 0.15226325f0 + 0.34043857f0im -0.08104593f0 + 0.05716944f0im; -0.1153074f0 + 0.1301223f0im 0.22236079f0 + 0.71008706f0im]
amontoison commented 6 months ago

@maleadt, I've commented on the tests related to sparse_symv here.

The issue seems to be upstream, but I initially suspected that only the complex versions of the routines were broken: link to the issue.

maleadt commented 6 months ago

Isn't this sparse_gemm failure with complex inputs and sparse_symv with real ones?

amontoison commented 6 months ago

Isn't this sparse_gemm failure with complex inputs and sparse_symv with real ones?

sparse_symv is not working with real and complex inputs. I was not aware of the issue with sparse_gemm but it seems to sometimes fail on complex matrices.

The routines are never tested for complex inputs... We have this kind of problem with all vendors (NVIDIA, AMD or Intel).

codecov[bot] commented 6 months ago

Codecov Report

All modified and coverable lines are covered by tests :white_check_mark:

Project coverage is 84.57%. Comparing base (2df5d3f) to head (e812aeb).

Additional details and impacted files ```diff @@ Coverage Diff @@ ## master #405 +/- ## ======================================= Coverage 84.57% 84.57% ======================================= Files 42 42 Lines 2075 2075 ======================================= Hits 1755 1755 Misses 320 320 ```

:umbrella: View full report in Codecov by Sentry.
:loudspeaker: Have feedback on the report? Share it here.