Changed TensorSpan and ReadOnlyTensorSpan layout for better performance. #103244
Conversation
michaelgsharp
requested review from
stephentoub,
eiriktsarpalis,
ericstj and
tannergooding
|
Note regarding the |
1 similar comment
|
Note regarding the |
|
Tagging subscribers to this area: @dotnet/area-system-numerics-tensors |
src/libraries/System.Numerics.Tensors/src/System/Numerics/Tensors/netcore/TensorShape.cs
Outdated
Show resolved
Hide resolved
src/libraries/System.Numerics.Tensors/src/System/Numerics/Tensors/netcore/TensorShape.cs
Outdated
Show resolved
Hide resolved
src/libraries/System.Numerics.Tensors/src/System/Numerics/Tensors/netcore/ReadOnlyTensorSpan.cs
Outdated
Show resolved
Hide resolved
|
@tannergooding did you have any more feedback on this one? |
|
|
||
| namespace System.Numerics.Tensors | ||
| { | ||
| internal readonly struct TensorShape |
There was a problem hiding this comment.
Could this also be a ref struct?
There was a problem hiding this comment.
We want to be able to use the same struct as part of Tensor<T> so we can avoid additional allocations for common cases there as well.
| { | ||
| internal readonly nint[]? _metadata; // 8 bytes | ||
|
|
||
| internal readonly nint _memoryLength; // 8 bytes |
There was a problem hiding this comment.
Given the layout and size concerns of this change, does the new type require an explicit StructLayoutAttribute annotation? Are we ok with using the default layout? Is there an expectation that the fields will be consumed by unmanaged code.
There was a problem hiding this comment.
This struct contains a managed field (nint[]? _metadata) and as such is not blittable and will never have StructLayout respected, the runtime will treat it as Auto regardless (you can technically mark it as LayoutKind.Explicit and that would be respected, but that's a de-optimization and would cause many other issues).
There was a problem hiding this comment.
This struct contains a managed field (
nint[]? _metadata) and as such is not blittable and will never haveStructLayoutrespected, the runtime will treat it asAutoregardless (you can technically mark it asLayoutKind.Explicitand that would be respected, but that's a de-optimization and would cause many other issues).
Does Mono behave the same? I think it might respect Sequential for managed types.
There was a problem hiding this comment.
Does Mono behave the same? I think it might respect Sequential for managed types.
That doesn't really matter. RyuJIT doesn't and so code cannot depend on the managed side being sequential, that is part of why we must use InlineArray (rather than declaring n-sequential fields).
This is likewise an internal struct, so it can't be used by developers in interop scenarios. Due to it being a struct containing a managed field, it likewise couldn't be used in interop without marshalling and so there is zero benefit to changing the layout from its default.
There was a problem hiding this comment.
Does Mono behave the same? I think it might respect Sequential for managed types.
That doesn't really matter. RyuJIT doesn't and so code cannot depend on the managed side being sequential, that is part of why we must use
InlineArray(rather than declaring n-sequential fields).This is likewise an internal struct, so it can't be used by developers in interop scenarios. Due to it being a struct containing a managed field, it likewise couldn't be used in interop without marshalling and so there is zero benefit to changing the layout from its default.
Yeah but I've meant that Auto could maybe still be beneficial for Mono.
There was a problem hiding this comment.
Mono should likely consider adjusting their layout algorithm to match the RyuJIT algorithm in that scenario.
There are thousands of structs across the BCL and we are not going to go and annotate every single one of them to be explicitly Auto because they contain some managed field.
| @@ -109,12 +103,8 @@ public ReadOnlyTensorSpan(T[]? array, int start, scoped ReadOnlySpan<nint> lengt | |||
| ThrowHelper.ThrowArgument_InvalidStridesAndLengths(); | |||
| } | |||
|
|
|||
| _flattenedLength = linearLength; | |||
| _memoryLength = array.Length; | |||
| _shape = new TensorShape(array.Length - start, lengths, strides); | |||
There was a problem hiding this comment.
nit: I wonder if some of the validation logic could be in a helper on TensorShape
-- Not needed to fix in this PR, just a possible simplification opportunity
src/libraries/System.Numerics.Tensors/src/System/Numerics/Tensors/netcore/ReadOnlyTensorSpan.cs
Outdated
Show resolved
Hide resolved
src/libraries/System.Numerics.Tensors/src/System/Numerics/Tensors/netcore/ReadOnlyTensorSpan.cs
Outdated
Show resolved
Hide resolved
src/libraries/System.Numerics.Tensors/src/System/Numerics/Tensors/netcore/ReadOnlyTensorSpan.cs
Show resolved
Hide resolved
src/libraries/System.Numerics.Tensors/src/System/Numerics/Tensors/netcore/ReadOnlyTensorSpan.cs
Outdated
Show resolved
Hide resolved
src/libraries/System.Numerics.Tensors/src/System/Numerics/Tensors/netcore/ReadOnlyTensorSpan.cs
Outdated
Show resolved
Hide resolved
src/libraries/System.Numerics.Tensors/src/System/Numerics/Tensors/netcore/TensorShape.cs
Outdated
Show resolved
Hide resolved
michaelgsharp
force-pushed
the
tensor-struct
branch
from
d275019
to
63bc01d
Compare
|
/ba-g suppressing unrelated test failure to merge ahead of snap. |
…ce. (dotnet#103244) * TensorShape * fixed TensorShape issue * removed extra inline buffer type. Fixed tests. Added large dimension testing * removed typo * adding unsaved files * changes to use const for stack alloc comparison * changes from pr comments
Changed TensorSpan and ReadOnlyTensorSpan layout for better performance. This included making a new span to hold all the shared data between them and allowing for allocation free behavior for dimensions less than 6.