Re:Optimizing compilers reload vector constants needlessly
by janwas on Thursday December 08, 2022 @01:53AM
(#33905024)
> Also if I try to isolate the SIMD bits in C++ I’ll lose the opportunity to have them be inlined which will defeat the optimization purpose.
Agreed. Usually the interface would be something like RunEntireAlgorithm(), not DotProduct().
> For those that are new to this, can you give an example of a kind of computation or algorithm which is well-served by your project but not possible with vector extensions
Sure. Vector extensions are OKish for simple math but JPEG XL includes nontrivial cross-lane operations such as
transpose and boundary handling for convolution.
__builtin_shufflevector requires a known vector length, and can be pessimized (fusing two into one general all-to-all permute which is more expensive than two simple shuffles).
Also, vqsort (https://github.com/google/highway/tree/master/hwy/contrib/so...) almost entirely consists of
operations not supported by the extensions, and actually works out of the box on variable-length RISC-V and SVE, which compiler extensions cannot.
Agreed. Usually the interface would be something like RunEntireAlgorithm(), not DotProduct().
> For those that are new to this, can you give an example of a kind of computation or algorithm which is well-served by your project but not possible with vector extensions
Sure. Vector extensions are OKish for simple math but JPEG XL includes nontrivial cross-lane operations such as transpose and boundary handling for convolution. __builtin_shufflevector requires a known vector length, and can be pessimized (fusing two into one general all-to-all permute which is more expensive than two simple shuffles).
Also, vqsort (https://github.com/google/highway/tree/master/hwy/contrib/so...) almost entirely consists of operations not supported by the extensions, and actually works out of the box on variable-length RISC-V and SVE, which compiler extensions cannot.