Repository metrics
- Stars
- (1,408 stars)
- PR merge metrics
- (Avg merge 5d 5h) (16 merged PRs in 30d)
Description
Is your feature request related to a problem? Please describe.
CUDA.jl offers __byte_perm, which is very similar to the prmt PTX intrinsic, but lacks one important feature: One cannot use it for sign extension. From the prmt documentation:
"The msb defines if the byte value should be copied, or if the sign (msb of the byte) should be replicated over all 8 bits of the target position (sign extend of the byte value); msb=0 means copy the literal value; msb=1 means replicate the sign."
This feature is missing from the __byte_perm intrinsic.
Describe the solution you'd like
I would like CUDA.jl to offer access to the prmt PTX intrinsic, presumably under a different name (maybe prmt).
Describe alternatives you've considered
I am currently manually defining this function:
function cuda_prmt(x::UInt32, y::UInt32, op::UInt32)
LLVM.Interop.@asmcall("prmt.b32 \$0, \$1, \$2, \$3;", "=r,r,r,i", UInt32, Tuple{UInt32,UInt32,UInt32}, x, y, op)
end
Additional context