Skip to content

CSA: request ex2.approx.ftz through fastmath= so the compressor builds at the cutlass-dsl floor #463

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
Anerudhan merged 1 commit into from
Jul 31, 2026
Merged

CSA: request ex2.approx.ftz through fastmath= so the compressor builds at the cutlass-dsl floor #463

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
11 changes: 10 additions & 1 deletion python/cudnn/csa/compressor/compressor_sm100_r128.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -125,9 +125,18 @@
_L2E_LO = 1.925963033500011e-08 # fp32(log2(e) - _L2E_HI); residual ~4e-16


# ``cute.math.exp2``'s ``approx``/``ftz`` keywords are newer than the
# ``nvidia-cutlass-dsl>=4.5.0`` floor this project's dependency spec resolves to, so the
# same instruction is requested through ``fastmath=True``, which both versions accept.
# Both spellings lower to ``ex2.approx.ftz.f32`` -- the form the tolerance contract is
# calibrated on: on 4.6.1 the 16-kernel ``reg_probe_csa_compressor_r128.py`` PTX is
# byte-identical either way (1104 ``ex2.approx.ftz.f32``, same registers, no spills), and
# on 4.5.0 the same 1104 instructions are emitted.


def _exp_fast(x):
y = _ffma_rn(x, cutlass.Float32(_L2E_LO), _fmul_rn(x, cutlass.Float32(_L2E_HI)))
return cute_math.exp2(y, approx=True, ftz=True)
return cute_math.exp2(y, fastmath=True)


@cute.kernel
Expand Down