JuliaSIMD / JuliaSIMD/LoopVectorization.jl
Handle moving operations in reductions
- Dominant language
- Julia
- Stars
- 789
- Forks
- 73
- PR merge metrics
- No merged PRs in 30d
Description
This either `UndefVarError`s or (with a reordering in lowering) produces the wrong answer:
```julia
function doubleadd_loop_avxtest(x,a)
s = zero(eltype(x))
@avx for i in 1:length(x)
s += 2*x[i] + a
end
return s
end
```
because it transforms this into something like
```julia
stemp = s + a
for i in ...
sreduct = 2*x[i] + stemp
end
```
Where
1. `sreduct` isn't being accumulated across the loop's iterations.
2. `a` (after being hoisted out) is missing a constant multiple of `length(x) / W` where `W` is the SIMD vector width.
I should address both of these issues.
Contributor guide
No contributing guide indexed for this repository
Research direction
Start with the @avx reduction and lowering paths, using the doubleadd_loop_avxtest example in the issue to reproduce the failure. Trace how the reduction and hoisted a are transformed; done means the reduction accumulates across iterations and the hoisted value includes the required multiple of length(x) / W.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- julia
- Domain
- compilers, performance
- Issue type
- Bug
- Difficulty
- 4/5
- Estimated time
- 3-5 days
- Activity status
- Stale
- Clarity
- Mostly clear
- Newbie friendliness
- 42/100