Fix erf and a few other improvements

[christiangnrd]

Tried to turn of KA SpecialFunctions tests without realizing we don't have gamma implemented, but that uncovered a bug with our openlibm port of erf.

Best looked at each commit individually. I slightly refactored some tests to make it easier to mark broken, and changed the other tests around for consistency. It now also tests Float16 and Float32 instead of testing Float32 twice.

[github-actions]

Your PR requires formatting changes to meet the project's style guidelines.
Please consider running Runic (git runic main) to apply these changes.

Click here to view the suggested changes.

diff --git a/ext/SpecialFunctionsExt.jl b/ext/SpecialFunctionsExt.jl
index df00168d..8f3dc880 100644
--- a/ext/SpecialFunctionsExt.jl
+++ b/ext/SpecialFunctionsExt.jl
@@ -71,14 +71,14 @@ Metal.@device_override function SpecialFunctions.erf(x::Float32)
                 if ix < 0x04000000 	# |x|<0x1p-119
                     return (8 * x + efx8 * x) / 8	# avoid spurious underflow
                 end
-                return x + efx*x
+                return x + efx * x
             end
         end
         z = x * x
         r = pp0 + z * (pp1 + z * pp2)
         s = 1.0f0 + z * (qq1 + z * (qq2 + z * qq3))
-	    y = r / s
-	    return x + x*y
+        y = r / s
+        return x + x * y
     end
 
     if ix < 0x3fa00000 	# 0.84375 <= |x| < 1.25
diff --git a/test/device/intrinsics/math.jl b/test/device/intrinsics/math.jl
index 765d2ea3..1ea0771d 100644
--- a/test/device/intrinsics/math.jl
+++ b/test/device/intrinsics/math.jl
@@ -237,38 +237,38 @@ end
     end
 
     let # log1p
-        arr = T.(collect(LinRange(nextfloat(-1.0f0), 10.0f0, 20)))
+            arr = T.(collect(LinRange(nextfloat(-1.0f0), 10.0f0, 20)))
         buffer = MtlArray(arr)
-        cpures = log1p.(arr)
-        @test Array(log1p.(buffer)) ≈ log1p.(arr)
+            cpures = log1p.(arr)
+            @test Array(log1p.(buffer)) ≈ log1p.(arr)
     end
 
     let # erf
-        arr = T[-1.0, -0.5, 0.0, 1.0e-3, 1.0, 2.0, 5.5]
+            arr = T[-1.0, -0.5, 0.0, 1.0e-3, 1.0, 2.0, 5.5]
         buffer = MtlArray(arr)
-        cpures = SpecialFunctions.erf.(arr)
-        @test Array(SpecialFunctions.erf.(buffer)) ≈ cpures broken = (T == Float16)
+            cpures = SpecialFunctions.erf.(arr)
+            @test Array(SpecialFunctions.erf.(buffer)) ≈ cpures broken = (T == Float16)
     end
 
     let # erfc
-        arr = T.(collect(LinRange(nextfloat(-3.0f0), 3.0f0, 20)))
+            arr = T.(collect(LinRange(nextfloat(-3.0f0), 3.0f0, 20)))
         buffer = MtlArray(arr)
-        cpures = SpecialFunctions.erfc.(arr)
-        @test Array(SpecialFunctions.erfc.(buffer)) ≈ cpures broken = (T == Float16)
+            cpures = SpecialFunctions.erfc.(arr)
+            @test Array(SpecialFunctions.erfc.(buffer)) ≈ cpures broken = (T == Float16)
     end
 
     let # erfinv
-        arr = T.(collect(LinRange(-1.0f0, 1.0f0, 20)))
+            arr = T.(collect(LinRange(-1.0f0, 1.0f0, 20)))
         buffer = MtlArray(arr)
-        cpures = SpecialFunctions.erfinv.(arr)
-        @test Array(SpecialFunctions.erfinv.(buffer)) ≈ cpures
+            cpures = SpecialFunctions.erfinv.(arr)
+            @test Array(SpecialFunctions.erfinv.(buffer)) ≈ cpures
     end
 
     let # expm1
-        arr = T.(collect(LinRange(nextfloat(-88.0f0), 88.0f0, 100)))
+            arr = T.(collect(LinRange(nextfloat(-88.0f0), 88.0f0, 100)))
         buffer = MtlArray(arr)
-        cpures = expm1.(arr)
-        @test Array(expm1.(buffer)) ≈ cpures
+            cpures = expm1.(arr)
+            @test Array(expm1.(buffer)) ≈ cpures
     end
 
 

[github-actions]

Metal Benchmarks

Details

Benchmark suite	Current: 37c7915	Previous: a5b56dc	Ratio
private array/construct	26257 ns	24288.25 ns	1.08
private array/broadcast	462750 ns	462666 ns	1.00
private array/random/randn/Float32	923792 ns	809041.5 ns	1.14
private array/random/randn!/Float32	594541 ns	634417 ns	0.94
private array/random/rand!/Int64	555396 ns	566041 ns	0.98
private array/random/rand!/Float32	562291 ns	605750 ns	0.93
private array/random/rand/Int64	875000.5 ns	779834 ns	1.12
private array/random/rand/Float32	832437.5 ns	655667 ns	1.27
private array/copyto!/gpu_to_gpu	547187.5 ns	657625 ns	0.83
private array/copyto!/cpu_to_gpu	630896 ns	637291.5 ns	0.99
private array/copyto!/gpu_to_cpu	632625 ns	834166.5 ns	0.76
private array/accumulate/1d	1426917 ns	1334500 ns	1.07
private array/accumulate/2d	1495562.5 ns	1396958 ns	1.07
private array/iteration/findall/int	2275166.5 ns	2077292 ns	1.10
private array/iteration/findall/bool	2030584 ns	1843270.5 ns	1.10
private array/iteration/findfirst/int	1804812.5 ns	1714708 ns	1.05
private array/iteration/findfirst/bool	1715500 ns	1664854 ns	1.03
private array/iteration/scalar	2566708.5 ns	3557500 ns	0.72
private array/iteration/logical	3554000 ns	3207375 ns	1.11
private array/iteration/findmin/1d	1867792 ns	1757000 ns	1.06
private array/iteration/findmin/2d	1428666.5 ns	1357875 ns	1.05
private array/reductions/reduce/1d	910500 ns	1041270.5 ns	0.87
private array/reductions/reduce/2d	686125 ns	655187.5 ns	1.05
private array/reductions/mapreduce/1d	992478.5 ns	1044250 ns	0.95
private array/reductions/mapreduce/2d	682375 ns	665167 ns	1.03
private array/permutedims/4d	2636208 ns	2525375 ns	1.04
private array/permutedims/2d	1079917 ns	1028666 ns	1.05
private array/permutedims/3d	1798583 ns	1591000 ns	1.13
private array/copy	847083 ns	581875 ns	1.46
latency/precompile	9897694833 ns	9742903667 ns	1.02
latency/ttfp	3855217666.5 ns	3743066375 ns	1.03
latency/import	1284400395.5 ns	1260251146 ns	1.02
integration/metaldevrt	752958 ns	725167 ns	1.04
integration/byval/slices=1	1644562.5 ns	1637687.5 ns	1.00
integration/byval/slices=3	19545937.5 ns	9931917 ns	1.97
integration/byval/reference	1653354 ns	1568625 ns	1.05
integration/byval/slices=2	2801166 ns	2585875 ns	1.08
kernel/indexing	455875 ns	451375 ns	1.01
kernel/indexing_checked	463000 ns	455625 ns	1.02
kernel/launch	9076.5 ns	40828.125 ns	0.22
metal/synchronization/stream	15041 ns	14500 ns	1.04
metal/synchronization/context	15458 ns	14875 ns	1.04
shared array/construct	24275 ns	24486.083333333336 ns	0.99
shared array/broadcast	455333 ns	460791 ns	0.99
shared array/random/randn/Float32	917000 ns	804834 ns	1.14
shared array/random/randn!/Float32	608208 ns	637417 ns	0.95
shared array/random/rand!/Int64	553458 ns	564000 ns	0.98
shared array/random/rand!/Float32	555271 ns	606125 ns	0.92
shared array/random/rand/Int64	873833 ns	750521 ns	1.16
shared array/random/rand/Float32	800813 ns	625979.5 ns	1.28
shared array/copyto!/gpu_to_gpu	80000 ns	78667 ns	1.02
shared array/copyto!/cpu_to_gpu	79959 ns	81458.5 ns	0.98
shared array/copyto!/gpu_to_cpu	79542 ns	83625 ns	0.95
shared array/accumulate/1d	1431417 ns	1356083.5 ns	1.06
shared array/accumulate/2d	1486875 ns	1392416 ns	1.07
shared array/iteration/findall/int	2013083 ns	1859000 ns	1.08
shared array/iteration/findall/bool	1722542 ns	1594250 ns	1.08
shared array/iteration/findfirst/int	1506521 ns	1405917 ns	1.07
shared array/iteration/findfirst/bool	1434583 ns	1375333 ns	1.04
shared array/iteration/scalar	161000 ns	157041 ns	1.03
shared array/iteration/logical	3249333 ns	2995999.5 ns	1.08
shared array/iteration/findmin/1d	1580083 ns	1473166.5 ns	1.07
shared array/iteration/findmin/2d	1437875 ns	1367666 ns	1.05
shared array/reductions/reduce/1d	749708.5 ns	733833 ns	1.02
shared array/reductions/reduce/2d	708333 ns	672875 ns	1.05
shared array/reductions/mapreduce/1d	732333 ns	742208 ns	0.99
shared array/reductions/mapreduce/2d	705000 ns	666458.5 ns	1.06
shared array/permutedims/4d	2662188 ns	2549000 ns	1.04
shared array/permutedims/2d	1094438 ns	1027917 ns	1.06
shared array/permutedims/3d	1785187.5 ns	1583875 ns	1.13
shared array/copy	213292 ns	245500 ns	0.87

This comment was automatically generated by workflow using github-action-benchmark.

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 80.47%. Comparing base (a5b56dc) to head (37c7915).
Report is 1 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #582      +/-   ##
==========================================
- Coverage   80.73%   80.47%   -0.26%     
==========================================
  Files          61       61              
  Lines        2657     2658       +1     
==========================================
- Hits         2145     2139       -6     
- Misses        512      519       +7     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.