b = 'b'
c = 'c'
+algebraic = [
+ (('pack_unorm_4x8', a), ('pack_32_4x8', ('f2u8', ('fround_even', ('fmul', ('fsat', a), 255.0))))),
+
+ # Allows us to schedule as a multiply by 2
+ (('~fadd', ('fadd', a, b), a), ('fadd', ('fadd', a, a), b)),
+]
+
algebraic_late = [
# ineg must be lowered late, but only for integers; floats will try to
# have modifiers attached... hence why this has to be here rather than
# Fuse clamp_positive. This should probably be shared with Utgard/bifrost
(('fmax', a, 0.0), ('fclamp_pos', a)),
+
+ (('ishl', 'a@16', b), ('u2u16', ('ishl', ('u2u32', a), b))),
+ (('ishr', 'a@16', b), ('i2i16', ('ishr', ('i2i32', a), b))),
+ (('ushr', 'a@16', b), ('u2u16', ('ushr', ('u2u32', a), b))),
+
+ (('ishl', 'a@8', b), ('u2u8', ('u2u16', ('ishl', ('u2u32', ('u2u16', a)), b)))),
+ (('ishr', 'a@8', b), ('i2i8', ('i2i16', ('ishr', ('i2i32', ('i2i16', a)), b)))),
+ (('ushr', 'a@8', b), ('u2u8', ('u2u16', ('ushr', ('u2u32', ('u2u16', a)), b)))),
+
+ # Canonical form. The scheduler will convert back if it makes sense.
+ (('fmul', a, 2.0), ('fadd', a, a))
]
# Midgard is able to type convert down by only one "step" per instruction; if
-# NIR wants more than one step, we need to break up into multiple instructions
+# NIR wants more than one step, we need to break up into multiple instructions.
+# Nevertheless, we can do both a size step and a floating/int step at once.
converts = []
while srcsz <= srcsz_max:
# Size converter lowering is only needed if src and dst sizes are
# spaced by a factor > 2.
- # Type converter lowering is needed as soon as src_size != dst_size
- if srcsz != dstsz and ((srcsz * 2 != dstsz and srcsz != dstsz * 2) or op[0] != op[2]):
+ if srcsz != dstsz and (srcsz * 2 != dstsz and srcsz != dstsz * 2):
cursz = srcsz
rule = a
# When converting down we first do the type conversion followed
print('#include "midgard_nir.h"')
+ print(nir_algebraic.AlgebraicPass("midgard_nir_lower_algebraic_early",
+ algebraic).render())
+
print(nir_algebraic.AlgebraicPass("midgard_nir_lower_algebraic_late",
algebraic_late + converts + constant_switch).render())