Add nested conversion of Float64 to Float32

[papo1011]

Fixes inconsistent behavior (#650):
-mtl(rand(SVector{2,Float64}, n)) raised an error
-mtl(rand(Float64, n)) silently converted to Float32

Now no error raised, instead convert Float64 -> Float32:

mtl( rand( SVector{2, Float64}, 1000 ) )
# MtlVector{SVector{2, Float32}, Metal.PrivateStorage}

[github-actions]

Note: Please ignore this comment unless asked otherwise by a reviewer.

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/src/array.jl b/src/array.jl
index 0f8c9cc7..ea4ec7d7 100644
--- a/src/array.jl
+++ b/src/array.jl
@@ -485,26 +485,26 @@ Adapt.adapt_storage(::MtlArrayAdaptor, x::Float64) = Float32(x)
 Adapt.adapt_storage(::MtlArrayAdaptor, x::Complex{Float64}) = ComplexF32(x)
 
 # AbstractFloat → Float32
-Adapt.adapt_storage(::MtlArrayAdaptor{S}, xs::AbstractArray{T,N}) where {T<:AbstractFloat,N,S} =
+Adapt.adapt_storage(::MtlArrayAdaptor{S}, xs::AbstractArray{T, N}) where {T <: AbstractFloat, N, S} =
     isbits(xs) ? xs : MtlArray{Float32,N,S}(xs)
 
 # Float16 — preserve (more specific)
-Adapt.adapt_storage(::MtlArrayAdaptor{S}, xs::AbstractArray{T,N}) where {T<:Float16,N,S} =
-    isbits(xs) ? xs : MtlArray{T,N,S}(xs)
+Adapt.adapt_storage(::MtlArrayAdaptor{S}, xs::AbstractArray{T, N}) where {T <: Float16, N, S} =
+    isbits(xs) ? xs : MtlArray{T, N, S}(xs)
 
 # Complex{AbstractFloat} → ComplexF32
-Adapt.adapt_storage(::MtlArrayAdaptor{S}, xs::AbstractArray{T,N}) where {T<:Complex{<:AbstractFloat},N,S} =
+Adapt.adapt_storage(::MtlArrayAdaptor{S}, xs::AbstractArray{T, N}) where {T <: Complex{<:AbstractFloat}, N, S} =
     isbits(xs) ? xs : MtlArray{ComplexF32,N,S}(xs)
 
 # Complex{Float16} — preserve
-Adapt.adapt_storage(::MtlArrayAdaptor{S}, xs::AbstractArray{T,N}) where {T<:Complex{Float16},N,S} =
-    isbits(xs) ? xs : MtlArray{T,N,S}(xs)
+Adapt.adapt_storage(::MtlArrayAdaptor{S}, xs::AbstractArray{T, N}) where {T <: Complex{Float16}, N, S} =
+    isbits(xs) ? xs : MtlArray{T, N, S}(xs)
 
 # Generic — descend via adapt to handle composite types
-function Adapt.adapt_storage(to::MtlArrayAdaptor{S}, xs::AbstractArray{T,N}) where {T,N,S}
+function Adapt.adapt_storage(to::MtlArrayAdaptor{S}, xs::AbstractArray{T, N}) where {T, N, S}
     isbits(xs) && return xs
     adapted = map(x -> adapt(to, x), xs)
-    MtlArray{eltype(adapted),N,S}(adapted)
+    return MtlArray{eltype(adapted), N, S}(adapted)
 end
 
 """

[codecov]

Codecov Report

❌ Patch coverage is 90.00000% with 1 line in your changes missing coverage. Please review.
✅ Project coverage is 82.26%. Comparing base (1d2f000) to head (f83fed5).
⚠️ Report is 3 commits behind head on main.

Files with missing lines	Patch %	Lines
src/array.jl	90.00%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #748      +/-   ##
==========================================
+ Coverage   82.01%   82.26%   +0.25%     
==========================================
  Files          62       62              
  Lines        2874     2881       +7     
==========================================
+ Hits         2357     2370      +13     
+ Misses        517      511       -6     

☔ 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.

[github-actions]

Metal Benchmarks

Details

Benchmark suite	Current: f83fed5	Previous: 1848fba	Ratio
latency/precompile	25782881959 ns	25661029083 ns	1.00
latency/ttfp	2389108146 ns	2347953229.5 ns	1.02
latency/import	1452582834 ns	1431131708 ns	1.01
integration/metaldevrt	854062.5 ns	877062.5 ns	0.97
integration/byval/slices=1	1598791 ns	1562250 ns	1.02
integration/byval/slices=3	20475166 ns	9561167 ns	2.14
integration/byval/reference	1599416 ns	1555625 ns	1.03
integration/byval/slices=2	2734125 ns	2601916.5 ns	1.05
kernel/indexing	508208 ns	625375 ns	0.81
kernel/indexing_checked	511750 ns	659417 ns	0.78
kernel/launch	12500 ns	11541 ns	1.08
kernel/rand	516042 ns	567291 ns	0.91
array/construct	6334 ns	6333 ns	1.00
array/broadcast	510625 ns	608958 ns	0.84
array/random/randn/Float32	1040084 ns	963500 ns	1.08
array/random/randn!/Float32	729625 ns	752729 ns	0.97
array/random/rand!/Int64	538375 ns	558083 ns	0.96
array/random/rand!/Float32	540875 ns	586958 ns	0.92
array/random/rand/Int64	960708 ns	752416 ns	1.28
array/random/rand/Float32	828666 ns	662625 ns	1.25
array/accumulate/Int64/1d	1287958 ns	1244875 ns	1.03
array/accumulate/Int64/dims=1	1892459 ns	1839791 ns	1.03
array/accumulate/Int64/dims=2	2288000 ns	2174979 ns	1.05
array/accumulate/Int64/dims=1L	12240666.5 ns	11510459 ns	1.06
array/accumulate/Int64/dims=2L	9678750 ns	9796125 ns	0.99
array/accumulate/Float32/1d	1085375 ns	1111917 ns	0.98
array/accumulate/Float32/dims=1	1626000 ns	1567354.5 ns	1.04
array/accumulate/Float32/dims=2	2038750 ns	1882084 ns	1.08
array/accumulate/Float32/dims=1L	10420479.5 ns	9815541 ns	1.06
array/accumulate/Float32/dims=2L	7111666.5 ns	7235812.5 ns	0.98
array/reductions/reduce/Int64/1d	1326271 ns	1622437.5 ns	0.82
array/reductions/reduce/Int64/dims=1	1127000 ns	1148333 ns	0.98
array/reductions/reduce/Int64/dims=2	1175437.5 ns	1185375 ns	0.99
array/reductions/reduce/Int64/dims=1L	2041917 ns	2026167 ns	1.01
array/reductions/reduce/Int64/dims=2L	3755583 ns	4244833 ns	0.88
array/reductions/reduce/Float32/1d	826645.5 ns	1009292 ns	0.82
array/reductions/reduce/Float32/dims=1	806562.5 ns	851166 ns	0.95
array/reductions/reduce/Float32/dims=2	837791 ns	852062.5 ns	0.98
array/reductions/reduce/Float32/dims=1L	1359417 ns	1328000 ns	1.02
array/reductions/reduce/Float32/dims=2L	1835917 ns	1813625 ns	1.01
array/reductions/mapreduce/Int64/1d	1322875 ns	1566000 ns	0.84
array/reductions/mapreduce/Int64/dims=1	1127708 ns	1108708 ns	1.02
array/reductions/mapreduce/Int64/dims=2	1165292 ns	1153187.5 ns	1.01
array/reductions/mapreduce/Int64/dims=1L	1981417 ns	2031583 ns	0.98
array/reductions/mapreduce/Int64/dims=2L	3656937.5 ns	3617875 ns	1.01
array/reductions/mapreduce/Float32/1d	863396 ns	1047375 ns	0.82
array/reductions/mapreduce/Float32/dims=1	805500 ns	834687.5 ns	0.97
array/reductions/mapreduce/Float32/dims=2	830583 ns	873208.5 ns	0.95
array/reductions/mapreduce/Float32/dims=1L	1362584 ns	1318500 ns	1.03
array/reductions/mapreduce/Float32/dims=2L	1826271 ns	1842479 ns	0.99
array/private/copyto!/gpu_to_gpu	530834 ns	631625 ns	0.84
array/private/copyto!/cpu_to_gpu	735958.5 ns	796333 ns	0.92
array/private/copyto!/gpu_to_cpu	689312.5 ns	374541.5 ns	1.84
array/private/iteration/findall/int	1585854 ns	1555542 ns	1.02
array/private/iteration/findall/bool	1466042 ns	1407417 ns	1.04
array/private/iteration/findfirst/int	2112625 ns	2072500 ns	1.02
array/private/iteration/findfirst/bool	2043750 ns	2051292 ns	1.00
array/private/iteration/scalar	3481333.5 ns	5517729 ns	0.63
array/private/iteration/logical	2666187.5 ns	2653750 ns	1.00
array/private/iteration/findmin/1d	2583792 ns	2503750 ns	1.03
array/private/iteration/findmin/2d	1853375 ns	1792250 ns	1.03
array/private/copy	791042 ns	569166 ns	1.39
array/shared/copyto!/gpu_to_gpu	86250 ns	84166 ns	1.02
array/shared/copyto!/cpu_to_gpu	86959 ns	81417 ns	1.07
array/shared/copyto!/gpu_to_cpu	85833 ns	81875 ns	1.05
array/shared/iteration/findall/int	1580208 ns	1574708.5 ns	1.00
array/shared/iteration/findall/bool	1483625 ns	1417688 ns	1.05
array/shared/iteration/findfirst/int	1693958.5 ns	1669333.5 ns	1.01
array/shared/iteration/findfirst/bool	1654395.5 ns	1652125 ns	1.00
array/shared/iteration/scalar	204375 ns	205583.5 ns	0.99
array/shared/iteration/logical	2358021 ns	2455354.5 ns	0.96
array/shared/iteration/findmin/1d	2211229 ns	2120042 ns	1.04
array/shared/iteration/findmin/2d	1870083 ns	1807291 ns	1.03
array/shared/copy	216209 ns	242458 ns	0.89
array/permutedims/4d	2506042 ns	2399104 ns	1.04
array/permutedims/2d	1195125 ns	1179666 ns	1.01
array/permutedims/3d	1798854.5 ns	1685395.5 ns	1.07
metal/synchronization/stream	19542 ns	19209 ns	1.02
metal/synchronization/context	20000 ns	19917 ns	1.00

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

[maleadt]

Thanks. Inspecting a datatype is almost always the wrong solution though. I've made Adapt behave recursively for StaticArrays, JuliaGPU/Adapt.jl#104, and simplified the implementation here.