summaryrefslogtreecommitdiff
diff options
context:
space:
mode:
authorJason Ekstrand <jason.ekstrand@intel.com>2016-04-27 11:18:04 -0700
committerJason Ekstrand <jason.ekstrand@intel.com>2016-04-28 21:36:52 -0700
commit6d4a42674534ad2de254a330835344a5845428ef (patch)
tree8f9f25f7de8570f0d3bef8020f21fcc4358d2b79
parentf0af5b87ec140875dbc99d579b9b17ab935e27be (diff)
nir/algebraic: Support lowering for both 64 and 32-bit ldexp
Reviewed-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
-rw-r--r--src/compiler/nir/nir_opt_algebraic.py31
1 files changed, 22 insertions, 9 deletions
diff --git a/src/compiler/nir/nir_opt_algebraic.py b/src/compiler/nir/nir_opt_algebraic.py
index b5d90b22606..f181044faf2 100644
--- a/src/compiler/nir/nir_opt_algebraic.py
+++ b/src/compiler/nir/nir_opt_algebraic.py
@@ -381,11 +381,16 @@ optimizations = [
'options->lower_unpack_snorm_4x8'),
]
-def fexp2i(exp):
- # We assume that exp is already in the range [-126, 127].
- return ('ishl', ('iadd', exp, 127), 23)
-
-def ldexp32(f, exp):
+def fexp2i(exp, bits):
+ # We assume that exp is already in the right range.
+ if bits == 32:
+ return ('ishl', ('iadd', exp, 127), 23)
+ elif bits == 64:
+ return ('pack_double_2x32_split', 0, ('ishl', ('iadd', exp, 1023), 20))
+ else:
+ assert False
+
+def ldexp(f, exp, bits):
# First, we clamp exp to a reasonable range. The maximum possible range
# for a normal exponent is [-126, 127] and, throwing in denormals, you get
# a maximum range of [-149, 127]. This means that we can potentially have
@@ -396,7 +401,12 @@ def ldexp32(f, exp):
# handles a range on exp of [-252, 254] which allows you to create any
# value (including denorms if the hardware supports it) and to adjust the
# exponent of any normal value to anything you want.
- exp = ('imin', ('imax', exp, -252), 254)
+ if bits == 32:
+ exp = ('imin', ('imax', exp, -252), 254)
+ elif bits == 64:
+ exp = ('imin', ('imax', exp, -2044), 2046)
+ else:
+ assert False
# Now we compute two powers of 2, one for exp/2 and one for exp-exp/2.
# (We use ishr which isn't the same for -1, but the -1 case still works
@@ -406,11 +416,14 @@ def ldexp32(f, exp):
# that you can get with normalized values. Instead, we create two powers
# of two and multiply by them each in turn. That way the effective range
# of our exponent is doubled.
- pow2_1 = fexp2i(('ishr', exp, 1))
- pow2_2 = fexp2i(('isub', exp, ('ishr', exp, 1)))
+ pow2_1 = fexp2i(('ishr', exp, 1), bits)
+ pow2_2 = fexp2i(('isub', exp, ('ishr', exp, 1)), bits)
return ('fmul', ('fmul', f, pow2_1), pow2_2)
-optimizations += [(('ldexp@32', 'x', 'exp'), ldexp32('x', 'exp'))]
+optimizations += [
+ (('ldexp@32', 'x', 'exp'), ldexp('x', 'exp', 32)),
+ (('ldexp@64', 'x', 'exp'), ldexp('x', 'exp', 64)),
+]
# Unreal Engine 4 demo applications open-codes bitfieldReverse()
def bitfield_reverse(u):