Go version
go1.27
Output of go env in your module/workspace:
GOARCH='amd64'
GOEXPERIMENT='simd'
What did you do?
I added the missing 8-bit interleave operations to simd/archsimd on amd64:
InterleaveLo and InterleaveHi for Int8x16 and Uint8x16.InterleaveLoGrouped and InterleaveHiGrouped for Int8x32, Int8x64, Uint8x32, and Uint8x64.
Implementation: https://go-review.googlesource.com/c/go/+/796340
What did you see happen?
The 8-bit vector types do not provide the corresponding interleave operations on amd64, although these operations exist for wider integer element types.
In particular, the following methods are missing:
InterleaveLo and InterleaveHi for Int8x16 and Uint8x16.InterleaveLoGrouped and InterleaveHiGrouped for Int8x32, Int8x64, Uint8x32, and Uint8x64.
What did you expect to see?
I expected the 8-bit vector types to provide the same interleave operations as the wider integer vector types.
On amd64, these operations map directly to VPUNPCKLBW and VPUNPCKHBW, so the current omission appears unintended. Adding them makes the API consistent across signed and unsigned integer vector types and enables byte-oriented SIMD algorithms without manual emulation.
Go version
go1.27
Output of
go envin your module/workspace:What did you do?
I added the missing 8-bit interleave operations to
simd/archsimdon amd64:InterleaveLoandInterleaveHiforInt8x16andUint8x16.InterleaveLoGroupedandInterleaveHiGroupedforInt8x32,Int8x64,Uint8x32, andUint8x64.Implementation: https://go-review.googlesource.com/c/go/+/796340
What did you see happen?
The 8-bit vector types do not provide the corresponding interleave operations on amd64, although these operations exist for wider integer element types.
In particular, the following methods are missing:
InterleaveLoandInterleaveHiforInt8x16andUint8x16.InterleaveLoGroupedandInterleaveHiGroupedforInt8x32,Int8x64,Uint8x32, andUint8x64.What did you expect to see?
I expected the 8-bit vector types to provide the same interleave operations as the wider integer vector types.
On amd64, these operations map directly to
VPUNPCKLBWandVPUNPCKHBW, so the current omission appears unintended. Adding them makes the API consistent across signed and unsigned integer vector types and enables byte-oriented SIMD algorithms without manual emulation.