More LoopVectorization tests & checks #57

mcabbott · 2020-12-12T15:57:18Z

This is part II of #53.

Closes #77, closes #75, closes #72.

mcabbott · 2020-12-12T19:16:15Z

CI on Julia 1.4 picks CUDA v0.1.0 which fails:

WARNING: using CUDA.CUDA in module Main conflicts with an existing identifier.
ERROR: LoadError: importing CUDA into Main conflicts with an existing identifier

Perhaps because it has SpecialFunctions v1.1.0, and KernelAbstractions v0.2.4?

Testing locally on 1.4, the resolver picks CUDA v1.3.3 & it passes. It also has KernelAbstractions v0.4.5, and SpecialFunctions v0.10.3.

Maybe adding [compat] CUDA = "1, 2" as in #55 will solve this.

DilumAluthge · 2020-12-13T00:05:50Z

I would say: merge #55 first, and then rebase this PR on master.

mcabbott · 2020-12-13T10:18:01Z

I was going to suggest the reverse... but either will work ultimately.

I am keen to keep tests on 1.4, I dropped 1.3 when I got tired of fighting the resolver. But 1.4 forces LoopVectorization 0.8, which it seems a bit soon to drop completely, given that this package can't bound the version actually used outside of tests.

There are, however, still some test bugs to track down. The recent 1.4 pass here still has some tests disabled.

There is also a surprising slowdown, somehow LV takes 2410.5 seconds on Julia 1.4 (LV 0.8), vs 6149.6 seconds on 1.5 (0.9), there aren't that many tests disabled. From JuliaSIMD/LoopVectorization.jl#171 I learn that this may be caused by coverage, I wonder if that can be run more selectively?

chriselrod · 2020-12-13T10:49:23Z

LoopVectorization's tests are 99.9%+ compilation.

That may be a Julia 1.4 vs 1.5 difference. I recall 1.4 being faster than 1.5 on the same LoopVectorization version.

Running tests for both LoopVectorization 0.8.26 and 0.9.6 on Julia 1.5, I get (clipping off the gemm tests):
0.8.26:

#= /home/chriselrod/.julia/packages/LoopVectorization/pHMnJ/test/printmethods.jl:2 =# @__LINE__() = 2
  2.300284 seconds (5.02 M allocations: 256.232 MiB, 2.01% gc time)
(Float64, #= /home/chriselrod/.julia/packages/LoopVectorization/pHMnJ/test/fallback.jl:4 =# @__LINE__()) = (Float64, 4)
  6.730764 seconds (12.41 M allocations: 626.128 MiB, 3.69% gc time)
  0.035563 seconds (69.39 k allocations: 3.810 MiB)
  2.574874 seconds (8.74 M allocations: 446.962 MiB, 3.17% gc time)
  0.517646 seconds (2.19 M allocations: 110.348 MiB, 3.07% gc time)
(T, #= /home/chriselrod/.julia/packages/LoopVectorization/pHMnJ/test/offsetarrays.jl:204 =# @__LINE__()) = (Float32, 204)
r = -1:1
r = -2:2
(T, #= /home/chriselrod/.julia/packages/LoopVectorization/pHMnJ/test/offsetarrays.jl:204 =# @__LINE__()) = (Float64, 204)
r = -1:1
r = -2:2
186.278399 seconds (323.83 M allocations: 24.462 GiB, 5.23% gc time)
(Float64, #= /home/chriselrod/.julia/packages/LoopVectorization/pHMnJ/test/tensors.jl:51 =# @__LINE__()) = (Float64, 51)
  5.293765 seconds (13.67 M allocations: 746.867 MiB, 5.85% gc time)
(T, #= /home/chriselrod/.julia/packages/LoopVectorization/pHMnJ/test/map.jl:4 =# @__LINE__()) = (Float32, 4)
(T, #= /home/chriselrod/.julia/packages/LoopVectorization/pHMnJ/test/map.jl:4 =# @__LINE__()) = (Float64, 4)
  1.959536 seconds (7.39 M allocations: 376.356 MiB, 4.18% gc time)
(T, #= /home/chriselrod/.julia/packages/LoopVectorization/pHMnJ/test/filter.jl:4 =# @__LINE__()) = (Float32, 4)
(T, #= /home/chriselrod/.julia/packages/LoopVectorization/pHMnJ/test/filter.jl:4 =# @__LINE__()) = (Float64, 4)
(T, #= /home/chriselrod/.julia/packages/LoopVectorization/pHMnJ/test/filter.jl:4 =# @__LINE__()) = (Int32, 4)
(T, #= /home/chriselrod/.julia/packages/LoopVectorization/pHMnJ/test/filter.jl:4 =# @__LINE__()) = (Int64, 4)
  0.306723 seconds (724.39 k allocations: 37.850 MiB, 2.75% gc time)
(T, #= /home/chriselrod/.julia/packages/LoopVectorization/pHMnJ/test/mapreduce.jl:19 =# @__LINE__()) = (Int32, 19)
(T, #= /home/chriselrod/.julia/packages/LoopVectorization/pHMnJ/test/mapreduce.jl:19 =# @__LINE__()) = (Int64, 19)
(T, #= /home/chriselrod/.julia/packages/LoopVectorization/pHMnJ/test/mapreduce.jl:19 =# @__LINE__()) = (Float32, 19)
(T, #= /home/chriselrod/.julia/packages/LoopVectorization/pHMnJ/test/mapreduce.jl:19 =# @__LINE__()) = (Float64, 19)
 49.398566 seconds (448.45 M allocations: 29.106 GiB, 10.44% gc time)
(T, #= /home/chriselrod/.julia/packages/LoopVectorization/pHMnJ/test/ifelsemasks.jl:366 =# @__LINE__()) = (Float32, 366)
(T, #= /home/chriselrod/.julia/packages/LoopVectorization/pHMnJ/test/ifelsemasks.jl:366 =# @__LINE__()) = (Float64, 366)
(T, #= /home/chriselrod/.julia/packages/LoopVectorization/pHMnJ/test/ifelsemasks.jl:366 =# @__LINE__()) = (Int32, 366)
(T, #= /home/chriselrod/.julia/packages/LoopVectorization/pHMnJ/test/ifelsemasks.jl:366 =# @__LINE__()) = (Int64, 366)
 20.996625 seconds (56.05 M allocations: 2.809 GiB, 8.57% gc time)
(T, #= /home/chriselrod/.julia/packages/LoopVectorization/pHMnJ/test/dot.jl:234 =# @__LINE__()) = (Float32, 234)
(T, #= /home/chriselrod/.julia/packages/LoopVectorization/pHMnJ/test/dot.jl:234 =# @__LINE__()) = (Float64, 234)
(T, #= /home/chriselrod/.julia/packages/LoopVectorization/pHMnJ/test/dot.jl:234 =# @__LINE__()) = (Int32, 234)
(T, #= /home/chriselrod/.julia/packages/LoopVectorization/pHMnJ/test/dot.jl:234 =# @__LINE__()) = (Int64, 234)
 12.680781 seconds (44.60 M allocations: 2.271 GiB, 4.08% gc time)
(T, #= /home/chriselrod/.julia/packages/LoopVectorization/pHMnJ/test/special.jl:339 =# @__LINE__()) = (Float32, 339)
(T, #= /home/chriselrod/.julia/packages/LoopVectorization/pHMnJ/test/special.jl:339 =# @__LINE__()) = (Float64, 339)
  4.091667 seconds (13.12 M allocations: 633.796 MiB, 2.57% gc time)
(T, #= /home/chriselrod/.julia/packages/LoopVectorization/pHMnJ/test/gemv.jl:211 =# @__LINE__()) = (Float32, 211)
(T, #= /home/chriselrod/.julia/packages/LoopVectorization/pHMnJ/test/gemv.jl:211 =# @__LINE__()) = (Float64, 211)
(T, #= /home/chriselrod/.julia/packages/LoopVectorization/pHMnJ/test/gemv.jl:211 =# @__LINE__()) = (Int32, 211)
(T, #= /home/chriselrod/.julia/packages/LoopVectorization/pHMnJ/test/gemv.jl:211 =# @__LINE__()) = (Int64, 211)
 15.978444 seconds (52.62 M allocations: 2.395 GiB, 2.73% gc time)
(T, #= /home/chriselrod/.julia/packages/LoopVectorization/pHMnJ/test/miscellaneous.jl:789 =# @__LINE__()) = (Float32, 789)
(T, #= /home/chriselrod/.julia/packages/LoopVectorization/pHMnJ/test/miscellaneous.jl:789 =# @__LINE__()) = (Float64, 789)
(T, #= /home/chriselrod/.julia/packages/LoopVectorization/pHMnJ/test/miscellaneous.jl:1070 =# @__LINE__()) = (Float32, 1070)
(T, #= /home/chriselrod/.julia/packages/LoopVectorization/pHMnJ/test/miscellaneous.jl:1070 =# @__LINE__()) = (Float64, 1070)
 30.148872 seconds (127.87 M allocations: 6.813 GiB, 8.33% gc time)
(T, #= /home/chriselrod/.julia/packages/LoopVectorization/pHMnJ/test/copy.jl:129 =# @__LINE__()) = (Float32, 129)
(T, #= /home/chriselrod/.julia/packages/LoopVectorization/pHMnJ/test/copy.jl:129 =# @__LINE__()) = (Float64, 129)
(T, #= /home/chriselrod/.julia/packages/LoopVectorization/pHMnJ/test/copy.jl:129 =# @__LINE__()) = (Int32, 129)
(T, #= /home/chriselrod/.julia/packages/LoopVectorization/pHMnJ/test/copy.jl:129 =# @__LINE__()) = (Int64, 129)
  3.025664 seconds (9.11 M allocations: 447.853 MiB, 4.00% gc time)
(T, #= /home/chriselrod/.julia/packages/LoopVectorization/pHMnJ/test/broadcast.jl:5 =# @__LINE__()) = (Float32, 5)
(T, #= /home/chriselrod/.julia/packages/LoopVectorization/pHMnJ/test/broadcast.jl:5 =# @__LINE__()) = (Float64, 5)
(T, #= /home/chriselrod/.julia/packages/LoopVectorization/pHMnJ/test/broadcast.jl:5 =# @__LINE__()) = (Int32, 5)
(T, #= /home/chriselrod/.julia/packages/LoopVectorization/pHMnJ/test/broadcast.jl:5 =# @__LINE__()) = (Int64, 5)
 98.177554 seconds (140.41 M allocations: 7.892 GiB, 4.08% gc time)

0.9.6:

#= /home/chriselrod/.julia/packages/LoopVectorization/5Eosy/test/printmethods.jl:2 =# @__LINE__() = 2
  2.346917 seconds (5.03 M allocations: 256.636 MiB, 2.05% gc time)
  0.010064 seconds (9.74 k allocations: 628.211 KiB)
(Float64, #= /home/chriselrod/.julia/packages/LoopVectorization/5Eosy/test/fallback.jl:4 =# @__LINE__()) = (Float64, 4)
  9.985701 seconds (19.52 M allocations: 987.222 MiB, 4.43% gc time)
  0.032851 seconds (67.26 k allocations: 3.726 MiB)
  2.540316 seconds (9.26 M allocations: 477.631 MiB, 3.24% gc time)
  0.834916 seconds (3.25 M allocations: 166.502 MiB, 2.88% gc time)
(T, #= /home/chriselrod/.julia/packages/LoopVectorization/5Eosy/test/offsetarrays.jl:211 =# @__LINE__()) = (Float32, 211)
r = -1:1
r = -2:2
(T, #= /home/chriselrod/.julia/packages/LoopVectorization/5Eosy/test/offsetarrays.jl:211 =# @__LINE__()) = (Float64, 211)
r = -1:1
r = -2:2
  5.230486 seconds (16.37 M allocations: 794.435 MiB, 2.26% gc time)
(Float64, #= /home/chriselrod/.julia/packages/LoopVectorization/5Eosy/test/tensors.jl:51 =# @__LINE__()) = (Float64, 51)
  5.337574 seconds (16.69 M allocations: 905.243 MiB, 9.21% gc time)
(T, #= /home/chriselrod/.julia/packages/LoopVectorization/5Eosy/test/map.jl:4 =# @__LINE__()) = (Float32, 4)
(T, #= /home/chriselrod/.julia/packages/LoopVectorization/5Eosy/test/map.jl:4 =# @__LINE__()) = (Float64, 4)
  2.939322 seconds (10.46 M allocations: 519.786 MiB, 4.44% gc time)
(T, #= /home/chriselrod/.julia/packages/LoopVectorization/5Eosy/test/filter.jl:4 =# @__LINE__()) = (Float32, 4)
(T, #= /home/chriselrod/.julia/packages/LoopVectorization/5Eosy/test/filter.jl:4 =# @__LINE__()) = (Float64, 4)
(T, #= /home/chriselrod/.julia/packages/LoopVectorization/5Eosy/test/filter.jl:4 =# @__LINE__()) = (Int32, 4)
(T, #= /home/chriselrod/.julia/packages/LoopVectorization/5Eosy/test/filter.jl:4 =# @__LINE__()) = (Int64, 4)
  0.476314 seconds (1.19 M allocations: 62.302 MiB, 1.93% gc time)
(T, #= /home/chriselrod/.julia/packages/LoopVectorization/5Eosy/test/mapreduce.jl:19 =# @__LINE__()) = (Int32, 19)
(T, #= /home/chriselrod/.julia/packages/LoopVectorization/5Eosy/test/mapreduce.jl:19 =# @__LINE__()) = (Int64, 19)
(T, #= /home/chriselrod/.julia/packages/LoopVectorization/5Eosy/test/mapreduce.jl:19 =# @__LINE__()) = (Float32, 19)
(T, #= /home/chriselrod/.julia/packages/LoopVectorization/5Eosy/test/mapreduce.jl:19 =# @__LINE__()) = (Float64, 19)
 49.636117 seconds (485.14 M allocations: 30.130 GiB, 10.02% gc time)
(T, #= /home/chriselrod/.julia/packages/LoopVectorization/5Eosy/test/ifelsemasks.jl:366 =# @__LINE__()) = (Float32, 366)
(T, #= /home/chriselrod/.julia/packages/LoopVectorization/5Eosy/test/ifelsemasks.jl:366 =# @__LINE__()) = (Float64, 366)
(T, #= /home/chriselrod/.julia/packages/LoopVectorization/5Eosy/test/ifelsemasks.jl:366 =# @__LINE__()) = (Int32, 366)
(T, #= /home/chriselrod/.julia/packages/LoopVectorization/5Eosy/test/ifelsemasks.jl:366 =# @__LINE__()) = (Int64, 366)
 20.505279 seconds (60.46 M allocations: 3.053 GiB, 9.45% gc time)
(T, #= /home/chriselrod/.julia/packages/LoopVectorization/5Eosy/test/dot.jl:234 =# @__LINE__()) = (Float32, 234)
(T, #= /home/chriselrod/.julia/packages/LoopVectorization/5Eosy/test/dot.jl:234 =# @__LINE__()) = (Float64, 234)
(T, #= /home/chriselrod/.julia/packages/LoopVectorization/5Eosy/test/dot.jl:234 =# @__LINE__()) = (Int32, 234)
(T, #= /home/chriselrod/.julia/packages/LoopVectorization/5Eosy/test/dot.jl:234 =# @__LINE__()) = (Int64, 234)
 13.784316 seconds (50.99 M allocations: 2.600 GiB, 5.45% gc time)
(T, #= /home/chriselrod/.julia/packages/LoopVectorization/5Eosy/test/special.jl:339 =# @__LINE__()) = (Float32, 339)
(T, #= /home/chriselrod/.julia/packages/LoopVectorization/5Eosy/test/special.jl:339 =# @__LINE__()) = (Float64, 339)
  4.632350 seconds (15.96 M allocations: 765.493 MiB, 3.58% gc time)
(T, #= /home/chriselrod/.julia/packages/LoopVectorization/5Eosy/test/gemv.jl:211 =# @__LINE__()) = (Float32, 211)
(T, #= /home/chriselrod/.julia/packages/LoopVectorization/5Eosy/test/gemv.jl:211 =# @__LINE__()) = (Float64, 211)
(T, #= /home/chriselrod/.julia/packages/LoopVectorization/5Eosy/test/gemv.jl:211 =# @__LINE__()) = (Int32, 211)
(T, #= /home/chriselrod/.julia/packages/LoopVectorization/5Eosy/test/gemv.jl:211 =# @__LINE__()) = (Int64, 211)
 16.232091 seconds (53.92 M allocations: 2.471 GiB, 3.38% gc time)
(T, #= /home/chriselrod/.julia/packages/LoopVectorization/5Eosy/test/miscellaneous.jl:792 =# @__LINE__()) = (Float32, 792)
(T, #= /home/chriselrod/.julia/packages/LoopVectorization/5Eosy/test/miscellaneous.jl:792 =# @__LINE__()) = (Float64, 792)
(T, #= /home/chriselrod/.julia/packages/LoopVectorization/5Eosy/test/miscellaneous.jl:1075 =# @__LINE__()) = (Float32, 1075)
(T, #= /home/chriselrod/.julia/packages/LoopVectorization/5Eosy/test/miscellaneous.jl:1075 =# @__LINE__()) = (Float64, 1075)
 33.904497 seconds (138.12 M allocations: 7.276 GiB, 8.18% gc time)
(T, #= /home/chriselrod/.julia/packages/LoopVectorization/5Eosy/test/copy.jl:129 =# @__LINE__()) = (Float32, 129)
(T, #= /home/chriselrod/.julia/packages/LoopVectorization/5Eosy/test/copy.jl:129 =# @__LINE__()) = (Float64, 129)
(T, #= /home/chriselrod/.julia/packages/LoopVectorization/5Eosy/test/copy.jl:129 =# @__LINE__()) = (Int32, 129)
(T, #= /home/chriselrod/.julia/packages/LoopVectorization/5Eosy/test/copy.jl:129 =# @__LINE__()) = (Int64, 129)
  2.892922 seconds (9.38 M allocations: 471.123 MiB, 4.23% gc time)
(T, #= /home/chriselrod/.julia/packages/LoopVectorization/5Eosy/test/broadcast.jl:8 =# @__LINE__()) = (Float32, 8)
(T, #= /home/chriselrod/.julia/packages/LoopVectorization/5Eosy/test/broadcast.jl:8 =# @__LINE__()) = (Float64, 8)
(T, #= /home/chriselrod/.julia/packages/LoopVectorization/5Eosy/test/broadcast.jl:8 =# @__LINE__()) = (Int32, 8)
(T, #= /home/chriselrod/.julia/packages/LoopVectorization/5Eosy/test/broadcast.jl:8 =# @__LINE__()) = (Int64, 8)
101.429469 seconds (151.64 M allocations: 8.432 GiB, 3.62% gc time)

Gemm total, and test total:
0.8.26:

179.480013 seconds (388.27 M allocations: 23.585 GiB, 5.36% gc time)
Test Summary:        | Pass  Total
LoopVectorization.jl | 1724   1724
620.346936 seconds (1.66 G allocations: 103.007 GiB, 5.61% gc time)
    Testing LoopVectorization tests passed

0.9.6:

174.757699 seconds (417.31 M allocations: 25.917 GiB, 5.65% gc time)
Test Summary:        |  Pass  Total
LoopVectorization.jl | 20429  20429
447.869647 seconds (1.47 G allocations: 85.253 GiB, 5.85% gc time)
    Testing LoopVectorization tests passed

Overall, locally, the tests are almost 3 minutes (25%) faster.
This is on a computer with AVX512; compile times are much better without it.
Additionally, all tests were enabled. When on GitHub CI, several of the tests are disabled starting somewhere in the 0.9 series.

So, I think regressions are Julia-related.

mcabbott · 2020-12-13T10:58:43Z

Thanks for digging, as always.

Am I right to think that these local tests don't include coverage?

And, is there or should there be a Julia issue about such regressions? They could be inevitable consequence of progress elsewhere, but could also be bugs.

chriselrod · 2020-12-13T11:17:57Z

Yes, it was without coverage.

There's another package making a lot of use of DifferentialEquations.jl and ForwardDiff.jl that hit a pretty severe regression in compile times in some benchmarks on Julia 1.6, that some Julia core folks are looking into.

I'm not aware of an issue, but it probably wouldn't hurt.
I suspect LLVM's to blame.
I'm building Julia with ENABLE_TIMINGS defined in src/options.h, in hopes of getting output like this.

Couldn't find any documentation about it anywhere, so I'm not sure what to do from here.

EDIT: While building Julia from source, it keeps dumping these summaries now. So, I guess it'll do that automatically.
EDIT:

640.984369 seconds (379.33 M allocations: 25.899 GiB, 1.58% gc time, 99.55% compilation time)
Test Summary:        |  Pass  Total
LoopVectorization.jl | 20429  20429
1572.714641 seconds (1.61 G allocations: 107.668 GiB, 2.26% gc time, 99.16% compilation time)
ROOT                      :  0.07 %   3378911825
GC                        :  2.26 %   106785484285
LOWERING                  :  2.82 %   133262113044
PARSING                   :  0.01 %   445595991
INFERENCE                 :  5.07 %   239589938800
CODEGEN                   : 75.02 %   3544299937574
METHOD_LOOKUP_SLOW        :  0.01 %   393193036
METHOD_LOOKUP_FAST        :  0.46 %   21748429238
LLVM_OPT                  : 10.95 %   517271539743
LLVM_MODULE_FINISH        :  0.11 %   5067346249
METHOD_MATCH              :  0.30 %   14070103600
TYPE_CACHE_LOOKUP         :  0.78 %   36937850996
TYPE_CACHE_INSERT         :  0.00 %   85812602
STAGED_FUNCTION           :  1.28 %   60412449084
MACRO_INVOCATION          :  0.00 %   167520658
AST_COMPRESS              :  0.37 %   17331975652
AST_UNCOMPRESS            :  0.30 %   14386285566
SYSIMG_LOAD               :  0.00 %   155666605
ADD_METHOD                :  0.01 %   677747888
LOAD_MODULE               :  0.00 %   149572157
INIT_MODULE               :  0.00 %   12006711
     Testing LoopVectorization tests passed

So most of the time is spent in codegen. I'm not really sure what that means.

Also, I discovered a regression where @avx wasn't actually applying to the offset arrays. That's why that test was so much faster. Fixing it with a new release of ArrayInterface.

chriselrod · 2020-12-13T18:20:45Z

src/macro.jl

@@ -494,7 +494,8 @@ padmodclamp_pair(A, inds, store) = begin
        elseif ex.args[1] == :pad && length(ex.args) >= 2
            i = ex.args[2]
            if !all(==(0), ex.args[3:end]) || length(ex.args) == 2
-                push!(nopadif, :($i ∈ $axes($A,$d)))
+                # push!(nopadif, :($i ∈ $axes($A,$d)))
+                push!(nopadif, :($i >= first(axes($A,$d))), :($i <= Base.last(axes($A,$d)))) # allows avx? Weirdly, deleting "Base." causes errors


Weirdly, deleting "Base." causes errors

That's odd, wouldn't mind an issue.

Will do, I was trying to isolate it a bit, but so far the simple ones all work just fine!

DilumAluthge · 2020-12-21T00:01:49Z

FYI, on Julia 1.5, JULIA_NUM_THREADS=6 will give you 2 threads (since GHA only has 2 CPUs per runner). But on Julia 1.6, JULIA_NUM_THREADS=6 will give you 6 threads.

test/group-3.jl

Previous CI run passed, will this one?

mcabbott · 2021-01-23T13:06:57Z

Let's call this done.

Gradient checks using LoopVectorization are disabled completely on CI. They sometimes passed, sometimes failed, with strange errors.
Gradients for max/min using LoopVectorization are also disabled, they need some work.

mcabbott mentioned this pull request Dec 12, 2020

Several tweaks to the CI and testing configurations #55

Merged

chriselrod reviewed Dec 13, 2020

View reviewed changes

Michael Abbott added 12 commits December 17, 2020 02:03

only load Dual overloads with LV 0.8

4b91f76

tweak tests of onlyone(Mask ,...)

65449e1

tweak

0714e64

gradient fiddling

877903f

turn on some more gradient tests

fa7d3a7

make testsets

0f7cb66

convolutions...

93f3ce5

turn grad=Dual tests on 1.4 back on

d11c3c3

tweaks

05d4180

skip 3 tests

68f2e73

oops

0be01bd

oops

9d1d51e

mcabbott force-pushed the avxci2 branch from f011733 to 9d1d51e Compare December 17, 2020 01:06

tidy up runtests after bad rebase

ba2b151

Repository owner locked and limited conversation to collaborators Dec 17, 2020

Repository owner unlocked this conversation Dec 17, 2020

mcabbott mentioned this pull request Dec 19, 2020

Use nograd keyword for functions too #71

Open

skip some tests on nightly

61a9194

mcabbott mentioned this pull request Dec 20, 2020

Some errors with interpolated functions JuliaSIMD/LoopVectorization.jl#174

Closed

Michael Abbott added 2 commits December 21, 2020 00:56

restore LV testing on nightly

abbacee

try 6 threads

87358aa

VectorizationBase 0.14

21dda72

Repository owner deleted a comment from codecov-io Dec 21, 2020

Michael Abbott added 9 commits December 21, 2020 02:27

widen signature of allzero(seen::Vec{N,Int}))

2762a4f

Zygote 0.6

1e86d2e

disable avx for min/max grad

440b838

readme

e6d5151

disable avx in min/max grad, correctly & visibly

d3d688f

versions

90bb1ba

VectorizationBase 0.14.8

4a96e91

restore Base.last hack for old LV

cedff62

skip two

f8cc7d7

mcabbott mentioned this pull request Jan 11, 2021

Update compat entries for LoopVectorization and VectorizationBase #75

Closed

Michael Abbott added 9 commits January 21, 2021 18:33

update compat bounds

4dd82ac

re-enable zygote on 1.6

a68c11b

disable anyone / onlyone, vectorised min/max grad already disabled

5819089

add some space

8178df7

add some links

3840437

one more skip

f15efed

skip Zygote on nightly

7536064

bump

e911775

bump

fd16709

mcabbott commented Jan 22, 2021

View reviewed changes

test/group-3.jl Show resolved Hide resolved

mcabbott and others added 4 commits January 22, 2021 14:31

stop CI early on multi-threaded group-3

f6a60d7

stop CI early on multi-threaded group-3, take II

e4b8c9d

trivial commit

80f2f9f

Previous CI run passed, will this one?

disable LV grad checks on CI completely

09b60f1

mcabbott merged commit 5439bc3 into master Jan 23, 2021

mcabbott deleted the avxci2 branch January 23, 2021 13:07

mcabbott mentioned this pull request Jan 23, 2021

Debug a CI crash? #79

Merged

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

More LoopVectorization tests & checks #57

More LoopVectorization tests & checks #57

mcabbott commented Dec 12, 2020 •

edited

Loading

mcabbott commented Dec 12, 2020

DilumAluthge commented Dec 13, 2020

mcabbott commented Dec 13, 2020

chriselrod commented Dec 13, 2020

mcabbott commented Dec 13, 2020

chriselrod commented Dec 13, 2020 •

edited

Loading

chriselrod Dec 13, 2020

mcabbott Dec 13, 2020

DilumAluthge commented Dec 21, 2020

mcabbott commented Jan 23, 2021 •

edited

Loading

More LoopVectorization tests & checks #57

More LoopVectorization tests & checks #57

Conversation

mcabbott commented Dec 12, 2020 • edited Loading

mcabbott commented Dec 12, 2020

DilumAluthge commented Dec 13, 2020

mcabbott commented Dec 13, 2020

chriselrod commented Dec 13, 2020

mcabbott commented Dec 13, 2020

chriselrod commented Dec 13, 2020 • edited Loading

chriselrod Dec 13, 2020

Choose a reason for hiding this comment

mcabbott Dec 13, 2020

Choose a reason for hiding this comment

DilumAluthge commented Dec 21, 2020

mcabbott commented Jan 23, 2021 • edited Loading

mcabbott commented Dec 12, 2020 •

edited

Loading

chriselrod commented Dec 13, 2020 •

edited

Loading

mcabbott commented Jan 23, 2021 •

edited

Loading