Open avik-pal opened 3 weeks ago
Attention: Patch coverage is 8.33333%
with 11 lines
in your changes missing coverage. Please review.
Project coverage is 68.91%. Comparing base (
c185f04
) to head (f056a2d
).
Files | Patch % | Lines |
---|---|---|
src/impl/batchnorm.jl | 0.00% | 6 Missing :warning: |
src/impl/groupnorm.jl | 0.00% | 4 Missing :warning: |
src/impl/normalization.jl | 50.00% | 1 Missing :warning: |
:exclamation: There is a different number of reports uploaded between BASE (c185f04) and HEAD (f056a2d). Click for more details.
HEAD has 19 uploads less than BASE
| Flag | BASE (c185f04) | HEAD (f056a2d) | |------|------|------| ||37|18|
:umbrella: View full report in Codecov by Sentry.
:loudspeaker: Have feedback on the report? Share it here.
Currently, this is a performance disaster. Locally, I see like slowdowns of atleast 5-10x. Let's see the numbers on the dedicated benchmarks.
The main pro of this approach is that the maintenance burden significantly goes down. Now how can we solve this? (Probably this is better off as a KA Issue)
Finer control of
CPU
backend from KA:@simd
and@simd ivdep
loop info. Either by default or by supplying to the backend object -- https://github.com/JuliaGPU/KernelAbstractions.jl/pull/436