Implemented MPSGraphs.graph_conv!

[lolzdev]

This should allow closing #210

[github-actions]

Metal Benchmarks

Details

Benchmark suite	Current: 79afbf7	Previous: 1d2f000	Ratio
latency/precompile	26423261083 ns	25549419083 ns	1.03
latency/ttfp	2446354000 ns	2346831687.5 ns	1.04
latency/import	1499449000 ns	1427666042 ns	1.05
integration/metaldevrt	520209 ns	877750 ns	0.59
integration/byval/slices=1	1193708 ns	1568625 ns	0.76
integration/byval/slices=3	7839459 ns	8402792 ns	0.93
integration/byval/reference	1191042 ns	1559958 ns	0.76
integration/byval/slices=2	2116542 ns	2629875 ns	0.80
kernel/indexing	261958 ns	627417 ns	0.42
kernel/indexing_checked	274416 ns	608750 ns	0.45
kernel/launch	12834 ns	12667 ns	1.01
kernel/rand	294459 ns	576167 ns	0.51
array/construct	6833 ns	6500 ns	1.05
array/broadcast	286500 ns	606708 ns	0.47
array/random/randn/Float32	494125 ns	1011104 ns	0.49
array/random/randn!/Float32	420292 ns	753875 ns	0.56
array/random/rand!/Int64	317875 ns	548708 ns	0.58
array/random/rand!/Float32	278416 ns	586208.5 ns	0.47
array/random/rand/Int64	480541 ns	789709 ns	0.61
array/random/rand/Float32	353833.5 ns	645000 ns	0.55
array/accumulate/Int64/1d	964083 ns	1260667 ns	0.76
array/accumulate/Int64/dims=1	1044271 ns	1859104.5 ns	0.56
array/accumulate/Int64/dims=2	1375875 ns	2179083 ns	0.63
array/accumulate/Int64/dims=1L	9593146 ns	11673271 ns	0.82
array/accumulate/Int64/dims=2L	7797584 ns	9628146 ns	0.81
array/accumulate/Float32/1d	801083 ns	1121395.5 ns	0.71
array/accumulate/Float32/dims=1	905541.5 ns	1571667 ns	0.58
array/accumulate/Float32/dims=2	1185459 ns	1889459 ns	0.63
array/accumulate/Float32/dims=1L	8556479 ns	9834209 ns	0.87
array/accumulate/Float32/dims=2L	4416624.5 ns	7249666.5 ns	0.61
array/reductions/reduce/Int64/1d	751583 ns	1386875 ns	0.54
array/reductions/reduce/Int64/dims=1	690667 ns	1117250 ns	0.62
array/reductions/reduce/Int64/dims=2	726250 ns	1152958 ns	0.63
array/reductions/reduce/Int64/dims=1L	1162854 ns	2013209 ns	0.58
array/reductions/reduce/Int64/dims=2L	2227750 ns	4244083 ns	0.52
array/reductions/reduce/Float32/1d	564250 ns	988750 ns	0.57
array/reductions/reduce/Float32/dims=1	403750 ns	843520.5 ns	0.48
array/reductions/reduce/Float32/dims=2	434041 ns	857917 ns	0.51
array/reductions/reduce/Float32/dims=1L	682417 ns	1326625 ns	0.51
array/reductions/reduce/Float32/dims=2L	1245770.5 ns	1810667 ns	0.69
array/reductions/mapreduce/Int64/1d	754167 ns	1356437.5 ns	0.56
array/reductions/mapreduce/Int64/dims=1	690459 ns	1102166.5 ns	0.63
array/reductions/mapreduce/Int64/dims=2	726583 ns	1149750 ns	0.63
array/reductions/mapreduce/Int64/dims=1L	1145334 ns	1988375 ns	0.58
array/reductions/mapreduce/Int64/dims=2L	1856375 ns	3626916 ns	0.51
array/reductions/mapreduce/Float32/1d	580583 ns	1055917 ns	0.55
array/reductions/mapreduce/Float32/dims=1	404292 ns	847396 ns	0.48
array/reductions/mapreduce/Float32/dims=2	430667 ns	860979.5 ns	0.50
array/reductions/mapreduce/Float32/dims=1L	681208 ns	1333042 ns	0.51
array/reductions/mapreduce/Float32/dims=2L	1250500 ns	1898125 ns	0.66
array/private/copyto!/gpu_to_gpu	233041 ns	633020.5 ns	0.37
array/private/copyto!/cpu_to_gpu	257458 ns	804354.5 ns	0.32
array/private/copyto!/gpu_to_cpu	257000 ns	816000 ns	0.31
array/private/iteration/findall/int	1208812.5 ns	1581312.5 ns	0.76
array/private/iteration/findall/bool	1074583 ns	1404916.5 ns	0.76
array/private/iteration/findfirst/int	1196167 ns	2075167 ns	0.58
array/private/iteration/findfirst/bool	1188417 ns	2048750 ns	0.58
array/private/iteration/scalar	1725729.5 ns	4526479 ns	0.38
array/private/iteration/logical	1650375 ns	2693625 ns	0.61
array/private/iteration/findmin/1d	1422145.5 ns	2518041 ns	0.56
array/private/iteration/findmin/2d	1225250 ns	1820229.5 ns	0.67
array/private/copy	333042 ns	568854 ns	0.59
array/shared/copyto!/gpu_to_gpu	84042 ns	84291 ns	1.00
array/shared/copyto!/cpu_to_gpu	79875 ns	82875 ns	0.96
array/shared/copyto!/gpu_to_cpu	78479.5 ns	83000 ns	0.95
array/shared/iteration/findall/int	1215500 ns	1585854.5 ns	0.77
array/shared/iteration/findall/bool	1081833 ns	1421875 ns	0.76
array/shared/iteration/findfirst/int	1006125 ns	1654709 ns	0.61
array/shared/iteration/findfirst/bool	998000 ns	1643542 ns	0.61
array/shared/iteration/scalar	192000 ns	210375 ns	0.91
array/shared/iteration/logical	1447708 ns	2297959 ns	0.63
array/shared/iteration/findmin/1d	1229333 ns	2134229 ns	0.58
array/shared/iteration/findmin/2d	1201334 ns	1806042 ns	0.67
array/shared/copy	241792 ns	241812 ns	1.00
array/permutedims/4d	1726708 ns	2395583 ns	0.72
array/permutedims/2d	561417 ns	1158833 ns	0.48
array/permutedims/3d	1141042 ns	1686541 ns	0.68
metal/synchronization/stream	17334 ns	19583 ns	0.89
metal/synchronization/context	17958 ns	20291 ns	0.89

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

[codecov]

Codecov Report

❌ Patch coverage is 85.71429% with 11 lines in your changes missing coverage. Please review.
✅ Project coverage is 82.48%. Comparing base (1d2f000) to head (79afbf7).
⚠️ Report is 9 commits behind head on main.

Files with missing lines	Patch %	Lines
lib/mpsgraphs/operations.jl	82.50%	7 Missing ⚠️
lib/mpsgraphs/conv.jl	89.18%	4 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #745      +/-   ##
==========================================
+ Coverage   82.01%   82.48%   +0.47%     
==========================================
  Files          62       63       +1     
  Lines        2874     2929      +55     
==========================================
+ Hits         2357     2416      +59     
+ Misses        517      513       -4     

☔ 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.

[scheidan]

Thanks, I'm very happy to see this being tackled properly!

Just as a curiosity:

I was missing this functionality but unfortunately this is far outside my expertise. So in my desperation I asked Claude for help and surprisingly it got it to work! I'm sure it's much less elegant than this solution, but in case someone is interested:

• Changes for Metal.jl

• Changes for NNlib.jl

I tried to do it carefully, so it should be not just pure AI slop. But given my lack of understanding, I cannot guarantee anything.