This issue was brought to my attention in private communication with regard to OpenCL’s rotate() function, but some experiments show it affects cuda::std::rotl() and cuda::std::rotr() as well. Best I can tell, in all GPU architectures currently supported by CUDA, the GPU hardware funnel-shift instr…

yes, it makes a lot of sense. The implementation was too conservative. I already opened a PR for this issue Optimize `cuda::std::rotl/rotr` by fbusato · Pull Request #9352 · NVIDIA/cccl · GitHub . Should be integrated in CTK 13.5

Thanks for the report Norbert, I am following up internally

thanks for reporting the issue. There is one subtle difference between the proposed solution and cuda::std functions. cuda::std::rotl/rotr needs to satisfy the requirements of the respective C++ functions std::rotl - cppreference.com . In particular, the shift amount s could be negative and outside …

@fbusato I am well aware of the requirements of the C++20 standard for these rotate functions. These requirements are naturally met by the SHF hardware instructions in wrapping mode. Negative and “out-of-bounds” shift amounts are part of the simple test scaffold I included in my original post, with …

Thanks, looking forward to it. I am not sure how actively NVIDIA maintains OpenCL these days, but as I noted in the opening sentence, the issue was first noticed with regard to OpenCL’s rotate() functions, which should be optimized as well.

Thanks for looking into this guys. Here was my original post: [image] OpenCL variable-amount rotate(x, y) extra mask instruction in PTX GPU-Accelerated Libraries NVidia Developers, I have an OpenCL macro: ' define ROTL(x, s) rotate((uint)(x), (uint)(s)) It …

Suboptimal `cuda::std::rotl()` and `cuda::std::rotr()` in CUDA 13.3 and earlier versions

Accelerated Computing CUDA CUDA Programming and Performance

mix June 13, 2026, 1:47pm 7

Thanks for looking into this guys. Here was my original post:

Topic		Replies	Views
Bitwise rotation in CUDA CUDA Programming and Performance	1	9237	December 9, 2009
Builtin rotate() of 64-bit integer broken with NVIDIA CUDA 7.0 driver CUDA Programming and Performance	6	2579	June 22, 2015
ROL intrinsic function CUDA Programming and Performance	11	16596	September 29, 2007
Rotate (barrel shift) in hardware? CUDA Programming and Performance	3	6684	April 7, 2007
Problem with left shift CUDA Programming and Performance	4	3140	January 24, 2011
OpenCL variable-amount rotate(x, y) extra mask instruction in PTX GPU-Accelerated Libraries opencl	3	86	June 5, 2026
Functional Difference Emu vs. G80 (shift macro) CUDA Programming and Performance	6	6232	April 26, 2007
Problem using LFSR random number generator in CUDA FORTRAN Legacy PGI Compilers (archived)	18	11447	March 21, 2011
>=32bit shift on 32bit integer CUDA Programming and Performance cuda	4	1401	September 13, 2021
Minor Compiler Optimization Bit rotation CUDA Programming and Performance	6	2322	March 18, 2009

Suboptimal `cuda::std::rotl()` and `cuda::std::rotr()` in CUDA 13.3 and earlier versions

Related topics