On 30/09/2026 10:35, Richard Biener wrote:
On Tue, Sep 29, 2026 at 4:44 PM Andrew Stubbs <[email protected]> wrote:
The "simt" mode (no "-directed") is planned to integrate SIMT
parallelization with OpenMP threads so that the "omp simd" directive is
not needed and SIMT is invisible to the end user.
Is it planned to have SIMT vs. non-SIMT switchable per loop (nest) or per
function (given OMP will outline the nest anyway)? If you mark the --param
as Optimization it would be in the per-function optimization node ...
Sadly SIMT requires a whole different multilib, with a different ABI, so
SIMT/SIMD is all or nothing.
In theory, SIMD code could call the SIMT library, if we adjusted the ABI
accordingly, but that library can have no SIMD inside it, so library
calls would be slower. And, we'd need to do something very ugly to
enable the SIMD "clones" that SIMD code-gen expects to find in the
libraries.
Andrew
