Add scatter for CUDA support #296
Conversation
|
I faced some issue while running into these error: julia> using NNlibCUDA
[ Info: Precompiling NNlibCUDA [a00861dc-f156-4864-bf3c-e6376f28a68d]
ERROR: LoadError: ArgumentError: Package CompilerSupportLibraries_jll does not have JLLWrappers in its dependencies:
- If you have CompilerSupportLibraries_jll checked out for development and have
added JLLWrappers as a dependency but haven't updated your primary
environment's manifest file, try `Pkg.resolve()`.
- Otherwise you may need to report an issue with CompilerSupportLibraries_jllIt doesn't resolve when I execute Any idea? c.c.: @CarloLucibello |
| 1 => cu([3 3 4 4 5; | ||
| 5 5 6 6 7]), | ||
| ) | ||
| srcs = Dict( |
There was a problem hiding this comment.
Could we represent them like for loops with multiple iterators?
There was a problem hiding this comment.
I put key for the testing arrays because it indicates which test case is being tested. It can surely be simplified into for loops with arrays. But, do developers need to know the test case and their corresponding values?
|
Try with ] add CUDA#master |
|
@CarloLucibello I tried |
|
Also need Julia 1.6 |
|
Sorry, using v1.6 and CUDA#master still cannot resolve the error. |
|
@yuehhua I can't reproduce your problem. For me moving to the lib/NNlibCUDA directory, opening julia 1.6, |
|
It seems that there are some pre-compiling errors in my side. [yuehhua@yuehhua-GL65 NNlib.jl]$ cd lib/NNlibCUDA/
[yuehhua@yuehhua-GL65 NNlibCUDA]$ julia-1.6.0 --proj
_
_ _ _(_)_ | Documentation: https://docs.julialang.org
(_) | (_) (_) |
_ _ _| |_ __ _ | Type "?" for help, "]?" for Pkg help.
| | | | | | |/ _` | |
| | |_| | | | (_| | | Version 1.6.0-rc3 (2021-03-16)
_/ |\__'_|_|_|\__'_| | Official https://julialang.org/ release
|__/ |
(NNlibCUDA) pkg> up
Updating registry at `~/.julia/registries/General`
Updating git-repo `https://github.com/JuliaGPU/CUDA.jl.git`
Updating `/media/yuehhua/Workbench/workspace/NNlib.jl/lib/NNlibCUDA/Project.toml`
[052768ef] ~ CUDA v2.6.0 `https://github.com/JuliaGPU/CUDA.jl.git#master` ⇒ v2.6.0 `https://github.com/JuliaGPU/CUDA.jl.git#master`
Updating `/media/yuehhua/Workbench/workspace/NNlib.jl/lib/NNlibCUDA/Manifest.toml`
[052768ef] ~ CUDA v2.6.0 `https://github.com/JuliaGPU/CUDA.jl.git#master` ⇒ v2.6.0 `https://github.com/JuliaGPU/CUDA.jl.git#master`
Precompiling project...
Progress [========================================>] 18/18
✗ MbedTLS_jll
✗ Zlib_jll
✗ CompilerSupportLibraries_jll
✗ CUDA
✗ NNlibCUDA
13 dependencies successfully precompiled in 18 seconds (10 already precompiled)
5 dependencies errored |
|
I can reproduce now. Seems to be due to some problem in CUDA.jl, I filed an issue JuliaGPU/CUDA.jl#787 |
|
So not a bug, you just need |
|
I merged #306, so things should be fine if you rebase on master and use NNlibCUDA's manifest. |
|
based on the buildkite run, it seems that some of the gpu tests fail because they fallback to NNlib's scatter implementation |
yuehhua
force-pushed
the
scatter-cuda
branch
2 times, most recently
from
787fe7b to
294ee92
Compare
try to unify the scatter API add const MAX_THREADS and correct dynamic call atomic functions
|
Current PR has been moved to FluxML/NNlibCUDA.jl#1, so close this one. |
No description provided.