Use ThreadSafeModule/Context on Julia 1.9 #324
Conversation
|
Maybe add a rule here to not run the Buildkite reverse CI while the PR is in draft -- those tests are fairly expensive. CUDA.jl has such a check for some of the more expensive jobs. |
|
@pchintalapudi reminder that we need a fix for |
|
We can always temporarily use LLVM.jl to render the IR on 1.9, but we need to get this fixed in time of course. |
Does this PR depend on JuliaLang/julia#44960? |
No, not yet. I pulled out the GCN changes on this branch to just focus on updates for 1.9. This will now depend on JuliaLang/julia#45251, which is the necessary fix to |
jpsamaroo
force-pushed
the
jps/tsmod-tsctx
branch
from
c8f7657 to
5926405
Compare
|
JuliaLang/julia#45252 is working out even better, passing all tests for me locally. |
jpsamaroo
force-pushed
the
jps/tsmod-tsctx
branch
from
5926405 to
21576f7
Compare
|
CI is very unhappy? |
jpsamaroo
force-pushed
the
jps/tsmod-tsctx
branch
from
21576f7 to
08dc22a
Compare
|
Ahh, yeah, I guess I just have to always leave the struct in. Pushed. |
|
Back to draft while I figure out the GCN failures... |
|
Any update? The asserts build throws an assertion:
|
jpsamaroo
force-pushed
the
jps/tsmod-tsctx
branch
from
08dc22a to
6838cc4
Compare
Codecov Report
@@ Coverage Diff @@
## master #324 +/- ##
==========================================
+ Coverage 75.52% 76.67% +1.14%
==========================================
Files 24 23 -1
Lines 2799 2379 -420
==========================================
- Hits 2114 1824 -290
+ Misses 685 555 -130
Continue to review full report at Codecov.
|
|
Not sure what's up with the macOS failure? I don't have a system to reproduce on. |
|
cc: @pchintalapudi for the mac error? |
maleadt
force-pushed
the
jps/tsmod-tsctx
branch
3 times, most recently
from
712f617 to
d220d68
Compare
|
There's a failed assertion on 1.8 (only tried macOS) too: No useful backtrace of course, this being JITLink on 1.8/LLVM 13. I'll try with a debug build. EDIT: never mind, this one is an issue with the new Metal backend. I'll have a look at the 1.9 macOS failure tomorrow. |
|
I can't reproduce the failure, but that's using a significantly more recent version of Julia: DEV-649 instead of -439 on CI. Looks like the nightlies are very much out of date? |
|
I was fed up with nightlies being out of date / unsigned / unavailable, so we're just compiling Julia from source now. Looking good, the only thing that now still needs to be addressed is #324 (comment) |
|
CI is green except for Enzyme, is that related? @vchuravy |
|
Will have to take a closer look at that, but shouldn't hold this up. Since we are testing 1.6 there |
Co-authored-by: Valentin Churavy <[email protected]>
|
|
The error: Something looks wrong with the non-RAII |
|
Don't merge this yet; this PR breaks debug info: ; julia> ptx_code_llvm(identity, Tuple{Nothing}; dump_module=true, raw=true)
warning: ignoring debug info with an invalid version (0) in
; ModuleID = 'start'
source_filename = "start"
target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v16:16:16-v32:32:32-v64:64:64-v128:128:128-n16:32:64"
target triple = "nvptx64-nvidia-cuda"
; @ operators.jl:513 within `identity`
define void @julia_identity_1531() local_unnamed_addr #0 !dbg !3 {
top:
ret void, !dbg !5
}
attributes #0 = { "probe-stack"="inline-asm" }
!llvm.dbg.cu = !{!0}
!0 = distinct !DICompileUnit(language: DW_LANG_Julia, file: !1, producer: "julia", isOptimized: true, runtimeVersion: 0, emissionKind: NoDebug, enums: !2, nameTableKind: None)
!1 = !DIFile(filename: "operators.jl", directory: ".")
!2 = !{}
!3 = distinct !DISubprogram(name: "identity", linkageName: "julia_identity_1531", scope: null, file: !1, line: 513, type: !4, scopeLine: 513, spFlags: DISPFlagDefinition | DISPFlagOptimized, unit: !0, retainedNodes: !2)
!4 = !DISubroutineType(types: !2)
!5 = !DILocation(line: 513, scope: !3)vs 1.8: ; julia> ptx_code_llvm(identity, Tuple{Nothing}; dump_module=true, raw=true)
; ModuleID = 'text'
source_filename = "text"
target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v16:16:16-v32:32:32-v64:64:64-v128:128:128-n16:32:64"
target triple = "nvptx64-nvidia-cuda"
; @ operators.jl:526 within `identity`
define void @julia_identity_754() local_unnamed_addr #0 !dbg !5 {
top:
ret void, !dbg !7
}
attributes #0 = { "probe-stack"="inline-asm" }
!llvm.module.flags = !{!0, !1}
!llvm.dbg.cu = !{!2}
!0 = !{i32 2, !"Dwarf Version", i32 4}
!1 = !{i32 2, !"Debug Info Version", i32 3}
!2 = distinct !DICompileUnit(language: DW_LANG_Julia, file: !3, producer: "julia", isOptimized: true, runtimeVersion: 0, emissionKind: NoDebug, enums: !4, nameTableKind: None)
!3 = !DIFile(filename: "operators.jl", directory: ".")
!4 = !{}
!5 = distinct !DISubprogram(name: "identity", linkageName: "julia_identity_754", scope: null, file: !3, line: 526, type: !6, scopeLine: 526, spFlags: DISPFlagDefinition | DISPFlagOptimized, unit: !2, retainedNodes: !4)
!6 = !DISubroutineType(types: !4)
!7 = !DILocation(line: 526, scope: !5) |
Shouldn't be emitted now that 1.9 doesn't set-up the module.
From changes originally by @collinwarner