Add PTXFDivFastPass to lower fdiv fast to NVPTX approximate division#800
Merged
Conversation
The LLVM NVPTX backend handles fdiv fast for Float32 (→ div.approx.ftz.f32) but has no fast path for Float64. This IR-level pass covers both: - Float32: replaces fdiv with __nv_fast_fdividef (libdevice) - Float64: replaces fdiv with rcp.approx.ftz.d + Newton refinement, matching CUDA.jl's inv_fast(::Float64) algorithm The pass fires when the instruction carries the afn fast-math flag (set by @fastmath) or when target.fastmath=true. It follows the NVVMReflectPass pattern already in ptx.jl. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Codecov Report❌ Patch coverage is
Additional details and impacted files@@ Coverage Diff @@
## main #800 +/- ##
==========================================
+ Coverage 75.41% 75.69% +0.27%
==========================================
Files 25 25
Lines 3930 3983 +53
==========================================
+ Hits 2964 3015 +51
- Misses 966 968 +2 ☔ View full report in Codecov by Sentry. 🚀 New features to boost your workflow:
|
Member
|
Rebased and added a Float32 path that plays nicely with #804, so that we can avoid these overrides in CUDA.jl. |
Drop the `target.fastmath` check (`apply_fastmath!` stamps `afn` already), and emit NVPTX intrinsics directly so the f32 rewrite doesn't depend on libdevice being linked. f32 picks the FTZ variant from the function's `denormal-fp-math-f32` attribute. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This was referenced May 20, 2026
maleadt
added a commit
to JuliaGPU/CUDA.jl
that referenced
this pull request
May 21, 2026
Building on JuliaGPU/GPUCompiler.jl#805, JuliaGPU/GPUCompiler.jl#804, JuliaGPU/GPUCompiler.jl#800, avoid some of the uses of `libdevice`'s intrinsics, instead emitting vanilla LLVM IR and having GPUCompiler.jl post-process it into what we need in PTX. This has many advantages, including (potentially) better optimization, compatibility with LLVM tools like Enzyme, etc. Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
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
2 participants
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.
Overarching goal is to move the fast math handling from CUDA.jl to the GPUCompiler backend.
The LLVM NVPTX backend handles fdiv fast for Float32 (→ div.approx.ftz.f32)
but has no fast path for Float64. This IR-level pass covers both:
matching CUDA.jl's inv_fast(::Float64) algorithm
The pass fires when the instruction carries the afn fast-math flag (set by
@fastmath) or when target.fastmath=true. It follows the NVVMReflectPass
pattern already in ptx.jl.
Co-Authored-By: Claude Sonnet 4.6 noreply@anthropic.com