b = 'b'
c = 'c'
d = 'd'
+e = 'e'
# Written in the form (<search>, <replace>) where <search> is an expression
# and <replace> is either an expression or a value. An expression is
(('~flrp', a, b, 1.0), b),
(('~flrp', a, a, b), a),
(('~flrp', 0.0, a, b), ('fmul', a, b)),
+
+ # flrp(a, a + b, c) => a + flrp(0, b, c) => a + (b * c)
+ (('~flrp', a, ('fadd(is_used_once)', a, b), c), ('fadd', ('fmul', b, c), a)),
+ (('~flrp@32', a, ('fadd', a, b), c), ('fadd', ('fmul', b, c), a), 'options->lower_flrp32'),
+ (('~flrp@64', a, ('fadd', a, b), c), ('fadd', ('fmul', b, c), a), 'options->lower_flrp64'),
+
+ (('~flrp@32', ('fadd', a, b), ('fadd', a, c), d), ('fadd', ('flrp', b, c, d), a), 'options->lower_flrp32'),
+ (('~flrp@64', ('fadd', a, b), ('fadd', a, c), d), ('fadd', ('flrp', b, c, d), a), 'options->lower_flrp64'),
+
+ (('~flrp@32', a, ('fmul(is_used_once)', a, b), c), ('fmul', ('flrp', 1.0, b, c), a), 'options->lower_flrp32'),
+ (('~flrp@64', a, ('fmul(is_used_once)', a, b), c), ('fmul', ('flrp', 1.0, b, c), a), 'options->lower_flrp64'),
+
+ (('~flrp', ('fmul(is_used_once)', a, b), ('fmul(is_used_once)', a, c), d), ('fmul', ('flrp', b, c, d), a)),
+
(('~flrp', a, b, ('b2f', 'c@1')), ('bcsel', c, b, a), 'options->lower_flrp32'),
(('~flrp', a, 0.0, c), ('fadd', ('fmul', ('fneg', a), c), a)),
- (('flrp@32', a, b, c), ('fadd', ('fmul', c, ('fsub', b, a)), a), 'options->lower_flrp32'),
- (('flrp@64', a, b, c), ('fadd', ('fmul', c, ('fsub', b, a)), a), 'options->lower_flrp64'),
+ (('ftrunc', a), ('bcsel', ('flt', a, 0.0), ('fneg', ('ffloor', ('fabs', a))), ('ffloor', ('fabs', a))), 'options->lower_ftrunc'),
(('ffloor', a), ('fsub', a, ('ffract', a)), 'options->lower_ffloor'),
(('fadd', a, ('fneg', ('ffract', a))), ('ffloor', a), '!options->lower_ffloor'),
(('ffract', a), ('fsub', a, ('ffloor', a)), 'options->lower_ffract'),
(('ffma', a, b, c), ('fadd', ('fmul', a, b), c), 'options->lower_ffma'),
(('~fadd', ('fmul', a, b), c), ('ffma', a, b, c), 'options->fuse_ffma'),
+ (('~fmul', ('fadd', ('iand', ('ineg', ('b2i32', 'a@bool')), ('fmul', b, c)), '#d'), '#e'),
+ ('bcsel', a, ('fmul', ('fadd', ('fmul', b, c), d), e), ('fmul', d, e))),
+
(('fdot4', ('vec4', a, b, c, 1.0), d), ('fdph', ('vec3', a, b, c), d)),
(('fdot4', ('vec4', a, 0.0, 0.0, 0.0), b), ('fmul', a, b)),
(('fdot4', ('vec4', a, b, 0.0, 0.0), c), ('fdot2', ('vec2', a, b), c)),
(('fge', ('fneg', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), 0.0), ('inot', ('ior', a, b))),
(('fge', 0.0, ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), ('inot', ('ior', a, b))),
+ (('flt', a, ('fneg', a)), ('flt', a, 0.0)),
+ (('fge', a, ('fneg', a)), ('fge', a, 0.0)),
+
# Some optimizations (below) convert things like (a < b || c < b) into
# (min(a, c) < b). However, this interfers with the previous optimizations
# that try to remove comparisons with negated sums of b2f. This just
(('bcsel', a, ('bcsel(is_used_once)', b, c, d), ('bcsel', b, c, 'e')), ('bcsel', b, c, ('bcsel', a, d, 'e'))),
(('bcsel', a, ('bcsel', b, c, d), ('bcsel(is_used_once)', b, 'e', d)), ('bcsel', b, ('bcsel', a, c, 'e'), d)),
(('bcsel', a, ('bcsel(is_used_once)', b, c, d), ('bcsel', b, 'e', d)), ('bcsel', b, ('bcsel', a, c, 'e'), d)),
- (('bcsel', a, True, 'b@bool'), ('ior', a, b)),
+ (('bcsel', a, True, b), ('ior', a, b)),
+ (('bcsel', a, a, b), ('ior', a, b)),
+ (('bcsel', a, b, False), ('iand', a, b)),
+ (('bcsel', a, b, a), ('iand', a, b)),
(('fmin', a, a), a),
(('fmax', a, a), a),
(('imin', a, a), a),
(('ior', ('uge', 1, a), ('ieq', a, 2)), ('uge', 2, a)),
(('ior', ('uge', 2, a), ('ieq', a, 3)), ('uge', 3, a)),
- (('ior', 'a@bool', ('ieq', a, False)), True),
+ # The (i2f32, ...) part is an open-coded fsign. When that is combined with
+ # the bcsel, it's basically copysign(1.0, a). There is no copysign in NIR,
+ # so emit an open-coded version of that.
+ (('bcsel@32', ('feq', a, 0.0), 1.0, ('i2f32', ('iadd', ('b2i32', ('flt', 0.0, 'a@32')), ('ineg', ('b2i32', ('flt', 'a@32', 0.0)))))),
+ ('ior', 0x3f800000, ('iand', a, 0x80000000))),
+
+ (('ior', a, ('ieq', a, False)), True),
(('ior', a, ('inot', a)), -1),
+ (('ine', ('ineg', ('b2i32', 'a@1')), ('ineg', ('b2i32', 'b@1'))), ('ine', a, b)),
+ (('b2i32', ('ine', 'a@1', 'b@1')), ('b2i32', ('ixor', a, b))),
+
(('iand', ('ieq', 'a@32', 0), ('ieq', 'b@32', 0)), ('ieq', ('ior', 'a@32', 'b@32'), 0)),
# These patterns can result when (a < b || a < c) => (a < min(b, c))
# Boolean simplifications
(('i2b32(is_used_by_if)', a), ('ine32', a, 0)),
(('i2b1(is_used_by_if)', a), ('ine', a, 0)),
- (('ieq', 'a@bool', True), a),
- (('ine(is_not_used_by_if)', 'a@bool', True), ('inot', a)),
- (('ine', 'a@bool', False), a),
- (('ieq(is_not_used_by_if)', 'a@bool', False), ('inot', 'a')),
+ (('ieq', a, True), a),
+ (('ine(is_not_used_by_if)', a, True), ('inot', a)),
+ (('ine', a, False), a),
+ (('ieq(is_not_used_by_if)', a, False), ('inot', 'a')),
(('bcsel', a, True, False), a),
(('bcsel', a, False, True), ('inot', a)),
(('bcsel@32', a, 1.0, 0.0), ('b2f', a)),
('unpack_64_2x32_split_y', a)), a),
# Byte extraction
- (('ushr', ('ishl', 'a@32', 24), 24), ('extract_u8', a, 0), '!options->lower_extract_byte'),
- (('ushr', ('ishl', 'a@32', 16), 24), ('extract_u8', a, 1), '!options->lower_extract_byte'),
- (('ushr', ('ishl', 'a@32', 8), 24), ('extract_u8', a, 2), '!options->lower_extract_byte'),
+ (('ushr', 'a@16', 8), ('extract_u8', a, 1), '!options->lower_extract_byte'),
(('ushr', 'a@32', 24), ('extract_u8', a, 3), '!options->lower_extract_byte'),
- (('ishr', ('ishl', 'a@32', 24), 24), ('extract_i8', a, 0), '!options->lower_extract_byte'),
- (('ishr', ('ishl', 'a@32', 16), 24), ('extract_i8', a, 1), '!options->lower_extract_byte'),
- (('ishr', ('ishl', 'a@32', 8), 24), ('extract_i8', a, 2), '!options->lower_extract_byte'),
+ (('ushr', 'a@64', 56), ('extract_u8', a, 7), '!options->lower_extract_byte'),
+ (('ishr', 'a@16', 8), ('extract_i8', a, 1), '!options->lower_extract_byte'),
(('ishr', 'a@32', 24), ('extract_i8', a, 3), '!options->lower_extract_byte'),
- (('iand', 0xff, ('ushr', a, 16)), ('extract_u8', a, 2), '!options->lower_extract_byte'),
- (('iand', 0xff, ('ushr', a, 8)), ('extract_u8', a, 1), '!options->lower_extract_byte'),
- (('iand', 0xff, a), ('extract_u8', a, 0), '!options->lower_extract_byte'),
+ (('ishr', 'a@64', 56), ('extract_i8', a, 7), '!options->lower_extract_byte'),
+ (('iand', 0xff, a), ('extract_u8', a, 0), '!options->lower_extract_byte')
+]
+
+# After the ('extract_u8', a, 0) pattern, above, triggers, there will be
+# patterns like those below.
+for op in ('ushr', 'ishr'):
+ optimizations.extend([(('extract_u8', (op, 'a@16', 8), 0), ('extract_u8', a, 1))])
+ optimizations.extend([(('extract_u8', (op, 'a@32', 8 * i), 0), ('extract_u8', a, i)) for i in range(1, 4)])
+ optimizations.extend([(('extract_u8', (op, 'a@64', 8 * i), 0), ('extract_u8', a, i)) for i in range(1, 8)])
+optimizations.extend([(('extract_u8', ('extract_u16', a, 1), 0), ('extract_u8', a, 2))])
+
+# After the ('extract_[iu]8', a, 3) patterns, above, trigger, there will be
+# patterns like those below.
+for op in ('extract_u8', 'extract_i8'):
+ optimizations.extend([((op, ('ishl', 'a@16', 8), 1), (op, a, 0))])
+ optimizations.extend([((op, ('ishl', 'a@32', 24 - 8 * i), 3), (op, a, i)) for i in range(2, -1, -1)])
+ optimizations.extend([((op, ('ishl', 'a@64', 56 - 8 * i), 7), (op, a, i)) for i in range(6, -1, -1)])
+
+optimizations.extend([
# Word extraction
(('ushr', ('ishl', 'a@32', 16), 16), ('extract_u16', a, 0), '!options->lower_extract_word'),
(('ushr', 'a@32', 16), ('extract_u16', a, 1), '!options->lower_extract_word'),
(('bcsel', ('ine', a, -1), ('ifind_msb', a), -1), ('ifind_msb', a)),
# Misc. lowering
+ (('fmod@16', a, b), ('fsub', a, ('fmul', b, ('ffloor', ('fdiv', a, b)))), 'options->lower_fmod16'),
(('fmod@32', a, b), ('fsub', a, ('fmul', b, ('ffloor', ('fdiv', a, b)))), 'options->lower_fmod32'),
(('fmod@64', a, b), ('fsub', a, ('fmul', b, ('ffloor', ('fdiv', a, b)))), 'options->lower_fmod64'),
(('frem', a, b), ('fsub', a, ('fmul', b, ('ftrunc', ('fdiv', a, b)))), 'options->lower_fmod32'),
'options->lower_unpack_snorm_4x8'),
(('isign', a), ('imin', ('imax', a, -1), 1), 'options->lower_isign'),
-]
+ (('fsign', a), ('fsub', ('b2f', ('flt', 0.0, a)), ('b2f', ('flt', a, 0.0))), 'options->lower_fsign'),
+])
# bit_size dependent lowerings
for bit_size in [8, 16, 32, 64]:
x2yN = '{}2{}'.format(x, y)
optimizations.append(((x2yN, (b2x, a)), (b2y, a)))
+# Optimize away x2xN(a@N)
+for t in ['int', 'uint', 'float']:
+ for N in type_sizes(t):
+ x2xN = '{0}2{0}{1}'.format(t[0], N)
+ aN = 'a@{0}'.format(N)
+ optimizations.append(((x2xN, aN), a))
+
+# Optimize x2xN(y2yM(a@P)) -> y2yN(a) for integers
+# In particular, we can optimize away everything except upcast of downcast and
+# upcasts where the type differs from the other cast
+for N, M in itertools.product(type_sizes('uint'), type_sizes('uint')):
+ if N < M:
+ # The outer cast is a down-cast. It doesn't matter what the size of the
+ # argument of the inner cast is because we'll never been in the upcast
+ # of downcast case. Regardless of types, we'll always end up with y2yN
+ # in the end.
+ for x, y in itertools.product(['i', 'u'], ['i', 'u']):
+ x2xN = '{0}2{0}{1}'.format(x, N)
+ y2yM = '{0}2{0}{1}'.format(y, M)
+ y2yN = '{0}2{0}{1}'.format(y, N)
+ optimizations.append(((x2xN, (y2yM, a)), (y2yN, a)))
+ elif N > M:
+ # If the outer cast is an up-cast, we have to be more careful about the
+ # size of the argument of the inner cast and with types. In this case,
+ # the type is always the type of type up-cast which is given by the
+ # outer cast.
+ for P in type_sizes('uint'):
+ # We can't optimize away up-cast of down-cast.
+ if M < P:
+ continue
+
+ # Because we're doing down-cast of down-cast, the types always have
+ # to match between the two casts
+ for x in ['i', 'u']:
+ x2xN = '{0}2{0}{1}'.format(x, N)
+ x2xM = '{0}2{0}{1}'.format(x, M)
+ aP = 'a@{0}'.format(P)
+ optimizations.append(((x2xN, (x2xM, aP)), (x2xN, a)))
+ else:
+ # The N == M case is handled by other optimizations
+ pass
+
def fexp2i(exp, bits):
# We assume that exp is already in the right range.
- if bits == 32:
+ if bits == 16:
+ return ('i2i16', ('ishl', ('iadd', exp, 15), 10))
+ elif bits == 32:
return ('ishl', ('iadd', exp, 127), 23)
elif bits == 64:
return ('pack_64_2x32_split', 0, ('ishl', ('iadd', exp, 1023), 20))
# 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.
- if bits == 32:
+ if bits == 16:
+ exp = ('imin', ('imax', exp, -28), 30)
+ elif bits == 32:
exp = ('imin', ('imax', exp, -252), 254)
elif bits == 64:
exp = ('imin', ('imax', exp, -2044), 2046)
return ('fmul', ('fmul', f, pow2_1), pow2_2)
optimizations += [
+ (('ldexp@16', 'x', 'exp'), ldexp('x', 'exp', 16), 'options->lower_ldexp'),
(('ldexp@32', 'x', 'exp'), ldexp('x', 'exp', 32), 'options->lower_ldexp'),
(('ldexp@64', 'x', 'exp'), ldexp('x', 'exp', 64), 'options->lower_ldexp'),
]
((op, ('bcsel(is_used_once)', a, '#b', c), '#d'), ('bcsel', a, (op, b, d), (op, c, d)))
]
+# For derivatives in compute shaders, GLSL_NV_compute_shader_derivatives
+# states:
+#
+# If neither layout qualifier is specified, derivatives in compute shaders
+# return zero, which is consistent with the handling of built-in texture
+# functions like texture() in GLSL 4.50 compute shaders.
+for op in ['fddx', 'fddx_fine', 'fddx_coarse',
+ 'fddy', 'fddy_fine', 'fddy_coarse']:
+ optimizations += [
+ ((op, 'a'), 0.0, 'info->stage == MESA_SHADER_COMPUTE && info->cs.derivative_group == DERIVATIVE_GROUP_NONE')
+]
+
# This section contains "late" optimizations that should be run before
# creating ffmas and calling regular optimizations for the final time.
# Optimizations should go here if they help code generation and conflict
(('iadd', a, ('iadd', ('ineg', a), b)), b),
(('~fadd', ('fneg', a), ('fadd', a, b)), b),
(('~fadd', a, ('fadd', ('fneg', a), b)), b),
+
+ (('~flrp@32', ('fadd(is_used_once)', a, -1.0), ('fadd(is_used_once)', a, 1.0), d), ('fadd', ('flrp', -1.0, 1.0, d), a)),
+ (('~flrp@32', ('fadd(is_used_once)', a, 1.0), ('fadd(is_used_once)', a, -1.0), d), ('fadd', ('flrp', 1.0, -1.0, d), a)),
+ (('~flrp@32', ('fadd(is_used_once)', a, '#b'), ('fadd(is_used_once)', a, '#c'), d), ('fadd', ('fmul', d, ('fadd', c, ('fneg', b))), ('fadd', a, b))),
]
# This section contains "late" optimizations that should be run after the
(('fdot4', a, b), ('fdot_replicated4', a, b), 'options->fdot_replicates'),
(('fdph', a, b), ('fdph_replicated', a, b), 'options->fdot_replicates'),
+ (('~flrp@32', ('fadd(is_used_once)', a, b), ('fadd(is_used_once)', a, c), d), ('fadd', ('flrp', b, c, d), a)),
+ (('~flrp@64', ('fadd(is_used_once)', a, b), ('fadd(is_used_once)', a, c), d), ('fadd', ('flrp', b, c, d), a)),
+
(('b2f(is_used_more_than_once)', ('inot', 'a@1')), ('bcsel', a, 0.0, 1.0)),
(('fneg(is_used_more_than_once)', ('b2f', ('inot', 'a@1'))), ('bcsel', a, -0.0, -1.0)),
(('fmin', ('fadd(is_used_once)', '#c', a), ('fadd(is_used_once)', '#c', b)), ('fadd', c, ('fmin', a, b))),
(('fmax', ('fadd(is_used_once)', '#c', a), ('fadd(is_used_once)', '#c', b)), ('fadd', c, ('fmax', a, b))),
- (('bcsel', 'a@bool', 0, ('b2f32', ('inot', 'b@bool'))), ('b2f32', ('inot', ('ior', a, b)))),
+ (('bcsel', a, 0, ('b2f32', ('inot', 'b@bool'))), ('b2f32', ('inot', ('ior', a, b)))),
]
print(nir_algebraic.AlgebraicPass("nir_opt_algebraic", optimizations).render())