nir: Add optimization for doing removing f16/f32 conversions

This eliminates conversions between f16 and f32 where possible. We can
always remove an upcast followed by a down cast, that is:

  f2f16 ( f2f32 (a) )  ->  a
  f2fmp ( f2f32 (a) )  ->  a

In the other direction, f2f16 loses precision and can't be undone by a
f2f32.  However, by definition it's always safe to elminate f2fmp:

  f2f32 ( f2fmp (a) )  ->  a

v2. [Neil Roberts (nroberts@igalia.com)]

Reviewed-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/merge_requests/3822>

diff --git a/src/compiler/nir/nir_opt_algebraic.py b/src/compiler/nir/nir_opt_algebraic.py
index 281b3a61e787682ac97bed069ba2d88057e1d213..5e586a303deb866fb2992a84f812098d38707029 100644 (file)
--- a/src/compiler/nir/nir_opt_algebraic.py
+++ b/src/compiler/nir/nir_opt_algebraic.py
@@ -831,6 +831,13 @@ optimizations.extend([
    (('~f2u32', ('i2f', 'a@32')), a),
    (('~f2u32', ('u2f', 'a@32')), a),
 
+   # Conversions from float16 to float32 and back can always be removed
+   (('f2f16', ('f2f32', 'a@16')), a),
+   (('f2fmp', ('f2f32', 'a@16')), a),
+   # Conversions to float16 would be lossy so they should only be removed if
+   # the instruction was generated by the precision lowering pass.
+   (('f2f32', ('f2fmp', 'a@32')), a),
+
    (('ffloor', 'a(is_integral)'), a),
    (('fceil', 'a(is_integral)'), a),
    (('ftrunc', 'a(is_integral)'), a),