fdo-mirrors/mesa
1
0
Fork 0
mirror of https://gitlab.freedesktop.org/mesa/mesa.git synced 2026-06-09 23:08:18 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions 2

nir/opt_algebraic: Optimize mediump fadd/fmul done in highp
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 
When we do f2fmp(fadd(f2f32(a), f2f32(b))) we can always optimize it to
fadd(a, b) and obtain the same result minus an intermediate rounding
step, same for fmul.

I verified this on CPU using a custom script with Berkley SoftFloat
implementation, the results there are bit-for-bit identical except for
NaN representations.

Signed-off-by: Lorenzo Rossi <lorenzo.rossi@collabora.com>
Reviewed-by: Mel Henning <mhenning@darkrefraction.com>
Reviewed-by: Ashley Smith <ashley.smith@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/41419>
This commit is contained in:
Lorenzo Rossi 2026-05-06 18:18:51 +02:00 • committed by Marge Bot
parent e07bc4dbbf
commit 830cd2dc47
1 changed files with 8 additions and 5 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

13
src/compiler/nir/nir_opt_algebraic.py
View file

@ -2000,11 +2000,6 @@ optimizations.extend([
(('f2u32', ('f2fmp', 'a@32')), ('f2u32', a), 'true', TestStatus.UNSUPPORTED),
(('i2f32', ('i2imp', 'a@32')), ('i2f32', a), 'true', TestStatus.UNSUPPORTED),
# f16 -> f2f32 -> fmul32 -> f2fmp when the second operand is a constant
# The optimization only works when the constant can be safely represented with 16 bits
(('f2fmp', ('fmul(is_used_once,contract)', ('f2f32', 'a@16'), '#b(is_representable_as_f16)')), ('fmul', a, ('f2fmp', b)), 'true', TestStatus.UNSUPPORTED),
(('f2fmp', ('fadd(is_used_once,contract)', ('f2f32', 'a@16'), '#b(is_representable_as_f16)')), ('fadd', a, ('f2fmp', b)), 'true', TestStatus.UNSUPPORTED),
(('ffloor', 'a(is_integral)'), a),
(('fceil', 'a(is_integral)'), a),
(('ftrunc', 'a(is_integral)'), a),
@ -2053,6 +2048,14 @@ optimizations.extend([
(('ult', a, 0), False),
])
# Doing a 32-bit fadd/fmul on 16-bit operands is equivalent to a 16-bit
# fadd/fmul, contract is required for the double rounding.
for op in ['fadd', 'fmul']:
optimizations += [
(('f2fmp', (f'{op}(is_used_once,contract)', ('f2f32', 'a@16'), '#b(is_representable_as_f16)')), (op, a, ('f2fmp', b)), 'true', TestStatus.UNSUPPORTED),
(('f2fmp', (f'{op}(is_used_once,contract)', ('f2f32', 'a@16'), ('f2f32', 'b@16'))), (op, a, b), 'true', TestStatus.UNSUPPORTED),
]
for bits in [16, 32, 64]:
cond = '!(options->lower_doubles_options & nir_lower_fp64_full_software)' if bits == 64 else 'true'
bcsel = 'bcsel@{}'.format(bits)