fdo-mirrors/mesa
1
0
Fork 0
mirror of https://gitlab.freedesktop.org/mesa/mesa.git synced 2026-05-07 09:18:04 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions 4

pan/compiler: Clamp fp16 ldexp exponent range
Some checks are pending
macOS-CI / macOS-CI (dri) (push) Waiting to run
Details
macOS-CI / macOS-CI (xlib) (push) Waiting to run
Details

Browse source 
Fix OpenCL-CTS error in `math_brute_force/test_bruteforce -w ldexp`

Valhall LDEXP.v2f16 takes a 16-bit exponent, while NIR ldexp uses a
32-bit exponent. Truncating large exponents can flip overflow into
underflow or leave huge 16-bit exponents to hardware behavior that does
not match OpenCL's expected signed infinity/zero results.

Clamp the exponent to a range sufficient to overflow or underflow all
fp16 values before lowering to ldexp16_pan.

Signed-off-by: Eric Guo <eric.guo@nxp.com>
Reviewed-by: Lorenzo Rossi <lorenzo.rossi@collabora.com>
Reviewed-by: Christoph Pillmayer <christoph.pillmayer@arm.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/41234>
This commit is contained in:
Eric Guo 2026-04-28 17:02:22 +08:00 • committed by Marge Bot
parent 081feabf9c
commit 352a8d6beb
1 changed files with 7 additions and 11 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

18
src/panfrost/compiler/bifrost/bifrost_nir_algebraic.py
View file

@ -131,18 +131,14 @@ for bsz in [8, 16, 32]:
# than -126 (single- precision) or -1022 (double-precision), the value
# returned may be flushed to zero."
#
# So we can't just truncate the exponent. Overflow is undefined behavior, but
# we need to return signed zero on underflow. If exp32 < INT16_MIN, we can use
# any 16-bit exponent that's sufficiently small to send all f16 values to zero.
#
# If we test exp32 < INT16_MIN directly, the comparison could not be
# vectorized, so instead we test the upper half.
# TODO: some possible values for -127 can be encoded as small
# immediates on valhall. None of the usable immediates have replicated
# i16 lanes, but for example 0xFAFCFDFE would be {-1284,-514}, both of
# which are small enough.
# So we can't just truncate the exponent. Overflow is undefined behavior for
# GLSL, but OpenCL expects us to return signed infinity, and we need to return
# signed zero on underflow. Clamp to a range that's sufficient to overflow or
# underflow all f16 values, avoiding implementation-defined behaviour for huge
# exponents in LDEXP.v2f16.
algebraic_late += [
(('ldexp', 'a@16', b), ('ldexp16_pan', a, ('b16csel', ('ilt16', ('unpack_32_2x16_split_y', b), -1), -127, ('i2i16', b))))
(('ldexp', 'a@16', b),
('ldexp16_pan', a, ('i2i16', ('imin', ('imax', b, -127), 127))))
]