Repository navigation
Build test values from bits, not convert, and run the cuTile tests against cuTile main - #20
Merged
Merged
Conversation
The tests defined `convert` between UInt8 and their stand-in float types as a bit cast, so UInt8 literals could stand for bit patterns. `convert` converts values, and in 0.3 bits go through `raw_bits` and `from_raw_bits`, so the tests now build values with `float4`, `native4` or `from_raw_bits`, and read bits back with `raw_bits`. The cross-type constructor and broadcast tests convert real values, from integers to Bool and from Bool to UInt8. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01GqhLuuGKG92oXMxGJxVHDf
Microfloats 0.2.9 supports BitPacking 0.3. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01GqhLuuGKG92oXMxGJxVHDf
Codecov Report✅ All modified and coverable lines are covered by tests. Additional details and impacted files@@ Coverage Diff @@
## main #20 +/- ##
==========================================
+ Coverage 87.64% 94.41% +6.76%
==========================================
Files 6 6
Lines 340 340
==========================================
+ Hits 298 321 +23
+ Misses 42 19 -23 ☔ View full report in Codecov by Harness. 🚀 New features to boost your workflow:
|
AntonOresten
force-pushed
the
bitcast-converts
branch
from
November 7, 2026 12:28
22a6f7c to
b36b5dd
Compare
…ease Julia cuTile 1.0.0 requires CUDA_Compiler_jll 0.5, which holds CUDACore at 6.3 and LLVM.jl at 9, and LLVM.jl 9 has no libLLVMExtra for Julia 1.14. cuTile main resolves CUDACore 6.5 and LLVM.jl 10, and the extension tests pass against it unchanged. On 1.14-alpha cuTile main loads but fails in its own precompile workload, so the tests skip prerelease Julia, as they skip 1.10. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01GqhLuuGKG92oXMxGJxVHDf
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Removes the bit-cast
convertmethods from the tests, and runs the cuTile extension tests in CI again, against cuTilemain.Tests without bit-cast
convertThe tests defined
convertbetweenUInt8and their stand-in float types as a bit cast, so aUInt8literal could stand for a bit pattern. That passedraw_bitsandfrom_raw_bitsoff as the constructors' value conversion, which is what 0.3 removed from the API.convertconverts values: the realFloat4_E2M1FNin Microfloats turns0x03into 3.0, not into the bit pattern0x3, which is 1.5.test/runtests.jl, the stand-ins' bit-cast constructors, and the two methods intest/unused_bits.jl.float4,native4orfrom_raw_bits, and read bits withraw_bits.Booland fromBooltoUInt8. This covers the constructor,Narrow(T).(x),dest .= xandT.(narr).NVector{Float4_E2M1FN,2}(0x0a, 0x03)is now aMethodError, as it should be for an element type that can't be converted from integers.cuTile tests in CI
They're on again, now that Microfloats 0.2.9 supports BitPacking 0.3.
main. cuTile 1.0.0 requiresCUDA_Compiler_jll0.5, which holds CUDACore at 6.3 and LLVM.jl at 9. cuTilemainresolves CUDACore 6.5.1 and LLVM.jl 10.4.3, and the extension tests pass against it unchanged. Pin the next cuTile release exactly once there is one, since the codegen tests match emitted Tile IR.libLLVMExtrafor 1.14, which is what failed theprejob. LLVM.jl 10 loads on 1.14.0-alpha2, but cuTilemainthen fails in its own precompile workload (axis order must be a tuple, got Core.BindingPartition). The tests skip prerelease Julia, as they skip 1.10, with an@info.Tests
mainand Microfloats 0.2.9 (4813).🤖 Generated with Claude Code
https://claude.ai/code/session_01GqhLuuGKG92oXMxGJxVHDf