8a44a7a0a661384af157679980a69a4476d45ae0
3 # Copyright (C) 2014 Intel Corporation
5 # Permission is hereby granted, free of charge, to any person obtaining a
6 # copy of this software and associated documentation files (the "Software"),
7 # to deal in the Software without restriction, including without limitation
8 # the rights to use, copy, modify, merge, publish, distribute, sublicense,
9 # and/or sell copies of the Software, and to permit persons to whom the
10 # Software is furnished to do so, subject to the following conditions:
12 # The above copyright notice and this permission notice (including the next
13 # paragraph) shall be included in all copies or substantial portions of the
16 # THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 # IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 # FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
19 # THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20 # LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
21 # FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
25 # Jason Ekstrand (jason@jlekstrand.net)
29 # Convenience variables
35 # Written in the form (<search>, <replace>) where <search> is an expression
36 # and <replace> is either an expression or a value. An expression is
37 # defined as a tuple of the form (<op>, <src0>, <src1>, <src2>, <src3>)
38 # where each source is either an expression or a value. A value can be
39 # either a numeric constant or a string representing a variable name.
41 # Variable names are specified as "[#]name[@type]" where "#" inicates that
42 # the given variable will only match constants and the type indicates that
43 # the given variable will only match values from ALU instructions with the
46 # For constants, you have to be careful to make sure that it is the right
47 # type because python is unaware of the source and destination types of the
51 (('fneg', ('fneg', a
)), a
),
52 (('ineg', ('ineg', a
)), a
),
53 (('fabs', ('fabs', a
)), ('fabs', a
)),
54 (('fabs', ('fneg', a
)), ('fabs', a
)),
55 (('iabs', ('iabs', a
)), ('iabs', a
)),
56 (('iabs', ('ineg', a
)), ('iabs', a
)),
57 (('fadd', a
, 0.0), a
),
59 (('usadd_4x8', a
, 0), a
),
60 (('usadd_4x8', a
, ~
0), ~
0),
61 (('fadd', ('fmul', a
, b
), ('fmul', a
, c
)), ('fmul', a
, ('fadd', b
, c
))),
62 (('iadd', ('imul', a
, b
), ('imul', a
, c
)), ('imul', a
, ('iadd', b
, c
))),
63 (('fadd', ('fneg', a
), a
), 0.0),
64 (('iadd', ('ineg', a
), a
), 0),
65 (('iadd', ('ineg', a
), ('iadd', a
, b
)), b
),
66 (('iadd', a
, ('iadd', ('ineg', a
), b
)), b
),
67 (('fadd', ('fneg', a
), ('fadd', a
, b
)), b
),
68 (('fadd', a
, ('fadd', ('fneg', a
), b
)), b
),
69 (('fmul', a
, 0.0), 0.0),
71 (('umul_unorm_4x8', a
, 0), 0),
72 (('umul_unorm_4x8', a
, ~
0), a
),
73 (('fmul', a
, 1.0), a
),
75 (('fmul', a
, -1.0), ('fneg', a
)),
76 (('imul', a
, -1), ('ineg', a
)),
77 (('ffma', 0.0, a
, b
), b
),
78 (('ffma', a
, 0.0, b
), b
),
79 (('ffma', a
, b
, 0.0), ('fmul', a
, b
)),
80 (('ffma', a
, 1.0, b
), ('fadd', a
, b
)),
81 (('ffma', 1.0, a
, b
), ('fadd', a
, b
)),
82 (('flrp', a
, b
, 0.0), a
),
83 (('flrp', a
, b
, 1.0), b
),
84 (('flrp', a
, a
, b
), a
),
85 (('flrp', 0.0, a
, b
), ('fmul', a
, b
)),
86 (('flrp', a
, b
, ('b2f', c
)), ('bcsel', c
, b
, a
), 'options->lower_flrp'),
87 (('flrp', a
, b
, c
), ('fadd', ('fmul', c
, ('fsub', b
, a
)), a
), 'options->lower_flrp'),
88 (('ffract', a
), ('fsub', a
, ('ffloor', a
)), 'options->lower_ffract'),
89 (('fadd', ('fmul', a
, ('fadd', 1.0, ('fneg', ('b2f', c
)))), ('fmul', b
, ('b2f', c
))), ('bcsel', c
, b
, a
), 'options->lower_flrp'),
90 (('fadd', ('fmul', a
, ('fadd', 1.0, ('fneg', c
))), ('fmul', b
, c
)), ('flrp', a
, b
, c
), '!options->lower_flrp'),
91 (('fadd', a
, ('fmul', ('b2f', c
), ('fadd', b
, ('fneg', a
)))), ('bcsel', c
, b
, a
), 'options->lower_flrp'),
92 (('fadd', a
, ('fmul', c
, ('fadd', b
, ('fneg', a
)))), ('flrp', a
, b
, c
), '!options->lower_flrp'),
93 (('ffma', a
, b
, c
), ('fadd', ('fmul', a
, b
), c
), 'options->lower_ffma'),
94 (('fadd', ('fmul', a
, b
), c
), ('ffma', a
, b
, c
), '!options->lower_ffma'),
95 # Comparison simplifications
96 (('inot', ('flt', a
, b
)), ('fge', a
, b
)),
97 (('inot', ('fge', a
, b
)), ('flt', a
, b
)),
98 (('inot', ('feq', a
, b
)), ('fne', a
, b
)),
99 (('inot', ('fne', a
, b
)), ('feq', a
, b
)),
100 (('inot', ('ilt', a
, b
)), ('ige', a
, b
)),
101 (('inot', ('ige', a
, b
)), ('ilt', a
, b
)),
102 (('inot', ('ieq', a
, b
)), ('ine', a
, b
)),
103 (('inot', ('ine', a
, b
)), ('ieq', a
, b
)),
104 (('fge', ('fneg', ('fabs', a
)), 0.0), ('feq', a
, 0.0)),
105 (('bcsel', ('flt', a
, b
), a
, b
), ('fmin', a
, b
)),
106 (('bcsel', ('flt', a
, b
), b
, a
), ('fmax', a
, b
)),
107 (('bcsel', ('inot', 'a@bool'), b
, c
), ('bcsel', a
, c
, b
)),
108 (('bcsel', a
, ('bcsel', a
, b
, c
), d
), ('bcsel', a
, b
, d
)),
115 (('fmin', ('fmax', a
, 0.0), 1.0), ('fsat', a
), '!options->lower_fsat'),
116 (('fmax', ('fmin', a
, 1.0), 0.0), ('fsat', a
), '!options->lower_fsat'),
117 (('fsat', a
), ('fmin', ('fmax', a
, 0.0), 1.0), 'options->lower_fsat'),
118 (('fsat', ('fsat', a
)), ('fsat', a
)),
119 (('fmin', ('fmax', ('fmin', ('fmax', a
, 0.0), 1.0), 0.0), 1.0), ('fmin', ('fmax', a
, 0.0), 1.0)),
120 (('ior', ('flt', a
, b
), ('flt', a
, c
)), ('flt', a
, ('fmax', b
, c
))),
121 (('ior', ('flt', a
, c
), ('flt', b
, c
)), ('flt', ('fmin', a
, b
), c
)),
122 (('ior', ('fge', a
, b
), ('fge', a
, c
)), ('fge', a
, ('fmin', b
, c
))),
123 (('ior', ('fge', a
, c
), ('fge', b
, c
)), ('fge', ('fmax', a
, b
), c
)),
124 (('slt', a
, b
), ('b2f', ('flt', a
, b
)), 'options->lower_scmp'),
125 (('sge', a
, b
), ('b2f', ('fge', a
, b
)), 'options->lower_scmp'),
126 (('seq', a
, b
), ('b2f', ('feq', a
, b
)), 'options->lower_scmp'),
127 (('sne', a
, b
), ('b2f', ('fne', a
, b
)), 'options->lower_scmp'),
128 (('fne', ('fneg', a
), a
), ('fne', a
, 0.0)),
129 (('feq', ('fneg', a
), a
), ('feq', a
, 0.0)),
131 (('imul', ('b2i', a
), ('b2i', b
)), ('b2i', ('iand', a
, b
))),
132 (('fmul', ('b2f', a
), ('b2f', b
)), ('b2f', ('iand', a
, b
))),
133 (('fsat', ('fadd', ('b2f', a
), ('b2f', b
))), ('b2f', ('ior', a
, b
))),
134 (('iand', 'a@bool', 1.0), ('b2f', a
)),
135 (('flt', ('fneg', ('b2f', a
)), 0), a
), # Generated by TGSI KILL_IF.
136 (('flt', ('fsub', 0.0, ('b2f', a
)), 0), a
), # Generated by TGSI KILL_IF.
137 # Comparison with the same args. Note that these are not done for
138 # the float versions because NaN always returns false on float
140 (('ilt', a
, a
), False),
141 (('ige', a
, a
), True),
142 (('ieq', a
, a
), True),
143 (('ine', a
, a
), False),
144 (('ult', a
, a
), False),
145 (('uge', a
, a
), True),
146 # Logical and bit operations
147 (('fand', a
, 0.0), 0.0),
149 (('iand', a
, ~
0), a
),
153 (('fxor', a
, a
), 0.0),
155 (('fxor', a
, 0.0), a
),
157 (('inot', ('inot', a
)), a
),
159 (('iand', ('inot', a
), ('inot', b
)), ('inot', ('ior', a
, b
))),
160 (('ior', ('inot', a
), ('inot', b
)), ('inot', ('iand', a
, b
))),
161 # Shift optimizations
168 (('iand', 0xff, ('ushr', a
, 24)), ('ushr', a
, 24)),
169 (('iand', 0xffff, ('ushr', a
, 16)), ('ushr', a
, 16)),
170 # Exponential/logarithmic identities
171 (('fexp2', ('flog2', a
)), a
), # 2^lg2(a) = a
172 (('flog2', ('fexp2', a
)), a
), # lg2(2^a) = a
173 (('fpow', a
, b
), ('fexp2', ('fmul', ('flog2', a
), b
)), 'options->lower_fpow'), # a^b = 2^(lg2(a)*b)
174 (('fexp2', ('fmul', ('flog2', a
), b
)), ('fpow', a
, b
), '!options->lower_fpow'), # 2^(lg2(a)*b) = a^b
175 (('fexp2', ('fadd', ('fmul', ('flog2', a
), b
), ('fmul', ('flog2', c
), d
))),
176 ('fmul', ('fpow', a
, b
), ('fpow', c
, d
)), '!options->lower_fpow'), # 2^(lg2(a) * b + lg2(c) + d) = a^b * c^d
177 (('fpow', a
, 1.0), a
),
178 (('fpow', a
, 2.0), ('fmul', a
, a
)),
179 (('fpow', a
, 4.0), ('fmul', ('fmul', a
, a
), ('fmul', a
, a
))),
180 (('fpow', 2.0, a
), ('fexp2', a
)),
181 (('fpow', ('fpow', a
, 2.2), 0.454545), a
),
182 (('fpow', ('fabs', ('fpow', a
, 2.2)), 0.454545), ('fabs', a
)),
183 (('fsqrt', ('fexp2', a
)), ('fexp2', ('fmul', 0.5, a
))),
184 (('frcp', ('fexp2', a
)), ('fexp2', ('fneg', a
))),
185 (('frsq', ('fexp2', a
)), ('fexp2', ('fmul', -0.5, a
))),
186 (('flog2', ('fsqrt', a
)), ('fmul', 0.5, ('flog2', a
))),
187 (('flog2', ('frcp', a
)), ('fneg', ('flog2', a
))),
188 (('flog2', ('frsq', a
)), ('fmul', -0.5, ('flog2', a
))),
189 (('flog2', ('fpow', a
, b
)), ('fmul', b
, ('flog2', a
))),
190 (('fadd', ('flog2', a
), ('flog2', b
)), ('flog2', ('fmul', a
, b
))),
191 (('fadd', ('flog2', a
), ('fneg', ('flog2', b
))), ('flog2', ('fdiv', a
, b
))),
192 (('fmul', ('fexp2', a
), ('fexp2', b
)), ('fexp2', ('fadd', a
, b
))),
193 # Division and reciprocal
194 (('fdiv', 1.0, a
), ('frcp', a
)),
195 (('fdiv', a
, b
), ('fmul', a
, ('frcp', b
)), 'options->lower_fdiv'),
196 (('frcp', ('frcp', a
)), a
),
197 (('frcp', ('fsqrt', a
)), ('frsq', a
)),
198 (('fsqrt', a
), ('frcp', ('frsq', a
)), 'options->lower_fsqrt'),
199 (('frcp', ('frsq', a
)), ('fsqrt', a
), '!options->lower_fsqrt'),
200 # Boolean simplifications
201 (('ieq', 'a@bool', True), a
),
202 (('ine', 'a@bool', True), ('inot', a
)),
203 (('ine', 'a@bool', False), a
),
204 (('ieq', 'a@bool', False), ('inot', 'a')),
205 (('bcsel', a
, True, False), ('ine', a
, 0)),
206 (('bcsel', a
, False, True), ('ieq', a
, 0)),
207 (('bcsel', True, b
, c
), b
),
208 (('bcsel', False, b
, c
), c
),
209 # The result of this should be hit by constant propagation and, in the
210 # next round of opt_algebraic, get picked up by one of the above two.
211 (('bcsel', '#a', b
, c
), ('bcsel', ('ine', 'a', 0), b
, c
)),
213 (('bcsel', a
, b
, b
), b
),
214 (('fcsel', a
, b
, b
), b
),
217 (('i2b', ('b2i', a
)), a
),
218 (('f2i', ('ftrunc', a
)), ('f2i', a
)),
219 (('f2u', ('ftrunc', a
)), ('f2u', a
)),
222 (('ushr', a
, 24), ('extract_u8', a
, 3), '!options->lower_extract_byte'),
223 (('iand', 0xff, ('ushr', a
, 16)), ('extract_u8', a
, 2), '!options->lower_extract_byte'),
224 (('iand', 0xff, ('ushr', a
, 8)), ('extract_u8', a
, 1), '!options->lower_extract_byte'),
225 (('iand', 0xff, a
), ('extract_u8', a
, 0), '!options->lower_extract_byte'),
228 (('ushr', a
, 16), ('extract_u16', a
, 1), '!options->lower_extract_word'),
229 (('iand', 0xffff, a
), ('extract_u16', a
, 0), '!options->lower_extract_word'),
232 (('fsub', a
, ('fsub', 0.0, b
)), ('fadd', a
, b
)),
233 (('isub', a
, ('isub', 0, b
)), ('iadd', a
, b
)),
234 (('ussub_4x8', a
, 0), a
),
235 (('ussub_4x8', a
, ~
0), 0),
236 (('fsub', a
, b
), ('fadd', a
, ('fneg', b
)), 'options->lower_sub'),
237 (('isub', a
, b
), ('iadd', a
, ('ineg', b
)), 'options->lower_sub'),
238 (('fneg', a
), ('fsub', 0.0, a
), 'options->lower_negate'),
239 (('ineg', a
), ('isub', 0, a
), 'options->lower_negate'),
240 (('fadd', a
, ('fsub', 0.0, b
)), ('fsub', a
, b
)),
241 (('iadd', a
, ('isub', 0, b
)), ('isub', a
, b
)),
242 (('fabs', ('fsub', 0.0, a
)), ('fabs', a
)),
243 (('iabs', ('isub', 0, a
)), ('iabs', a
)),
246 (('fmod', a
, b
), ('fsub', a
, ('fmul', b
, ('ffloor', ('fdiv', a
, b
)))), 'options->lower_fmod'),
247 (('uadd_carry', a
, b
), ('b2i', ('ult', ('iadd', a
, b
), a
)), 'options->lower_uadd_carry'),
248 (('usub_borrow', a
, b
), ('b2i', ('ult', a
, b
)), 'options->lower_usub_borrow'),
250 (('bitfield_insert', 'base', 'insert', 'offset', 'bits'),
251 ('bcsel', ('ilt', 31, 'bits'), 'insert',
252 ('bfi', ('bfm', 'bits', 'offset'), 'insert', 'base')),
253 'options->lower_bitfield_insert'),
255 (('ibitfield_extract', 'value', 'offset', 'bits'),
256 ('bcsel', ('ilt', 31, 'bits'), 'value',
257 ('ibfe', 'value', 'offset', 'bits')),
258 'options->lower_bitfield_extract'),
260 (('ubitfield_extract', 'value', 'offset', 'bits'),
261 ('bcsel', ('ult', 31, 'bits'), 'value',
262 ('ubfe', 'value', 'offset', 'bits')),
263 'options->lower_bitfield_extract'),
265 (('extract_i8', a
, b
),
266 ('ishr', ('ishl', a
, ('imul', ('isub', 3, b
), 8)), 24),
267 'options->lower_extract_byte'),
269 (('extract_u8', a
, b
),
270 ('iand', ('ushr', a
, ('imul', b
, 8)), 0xff),
271 'options->lower_extract_byte'),
273 (('extract_i16', a
, b
),
274 ('ishr', ('ishl', a
, ('imul', ('isub', 1, b
), 16)), 16),
275 'options->lower_extract_word'),
277 (('extract_u16', a
, b
),
278 ('iand', ('ushr', a
, ('imul', b
, 16)), 0xffff),
279 'options->lower_extract_word'),
281 (('pack_unorm_2x16', 'v'),
282 ('pack_uvec2_to_uint',
283 ('f2u', ('fround_even', ('fmul', ('fsat', 'v'), 65535.0)))),
284 'options->lower_pack_unorm_2x16'),
286 (('pack_unorm_4x8', 'v'),
287 ('pack_uvec4_to_uint',
288 ('f2u', ('fround_even', ('fmul', ('fsat', 'v'), 255.0)))),
289 'options->lower_pack_unorm_4x8'),
291 (('pack_snorm_2x16', 'v'),
292 ('pack_uvec2_to_uint',
293 ('f2i', ('fround_even', ('fmul', ('fmin', 1.0, ('fmax', -1.0, 'v')), 32767.0)))),
294 'options->lower_pack_snorm_2x16'),
296 (('pack_snorm_4x8', 'v'),
297 ('pack_uvec4_to_uint',
298 ('f2i', ('fround_even', ('fmul', ('fmin', 1.0, ('fmax', -1.0, 'v')), 127.0)))),
299 'options->lower_pack_snorm_4x8'),
301 (('unpack_unorm_2x16', 'v'),
302 ('fdiv', ('u2f', ('vec2', ('extract_u16', 'v', 0),
303 ('extract_u16', 'v', 1))),
305 'options->lower_unpack_unorm_2x16'),
307 (('unpack_unorm_4x8', 'v'),
308 ('fdiv', ('u2f', ('vec4', ('extract_u8', 'v', 0),
309 ('extract_u8', 'v', 1),
310 ('extract_u8', 'v', 2),
311 ('extract_u8', 'v', 3))),
313 'options->lower_unpack_unorm_4x8'),
315 (('unpack_snorm_2x16', 'v'),
316 ('fmin', 1.0, ('fmax', -1.0, ('fdiv', ('i2f', ('vec2', ('extract_i16', 'v', 0),
317 ('extract_i16', 'v', 1))),
319 'options->lower_unpack_snorm_2x16'),
321 (('unpack_snorm_4x8', 'v'),
322 ('fmin', 1.0, ('fmax', -1.0, ('fdiv', ('i2f', ('vec4', ('extract_i8', 'v', 0),
323 ('extract_i8', 'v', 1),
324 ('extract_i8', 'v', 2),
325 ('extract_i8', 'v', 3))),
327 'options->lower_unpack_snorm_4x8'),
330 # Unreal Engine 4 demo applications open-codes bitfieldReverse()
331 def bitfield_reverse(u
):
332 step1
= ('ior', ('ishl', u
, 16), ('ushr', u
, 16))
333 step2
= ('ior', ('ishl', ('iand', step1
, 0x00ff00ff), 8), ('ushr', ('iand', step1
, 0xff00ff00), 8))
334 step3
= ('ior', ('ishl', ('iand', step2
, 0x0f0f0f0f), 4), ('ushr', ('iand', step2
, 0xf0f0f0f0), 4))
335 step4
= ('ior', ('ishl', ('iand', step3
, 0x33333333), 2), ('ushr', ('iand', step3
, 0xcccccccc), 2))
336 step5
= ('ior', ('ishl', ('iand', step4
, 0x55555555), 1), ('ushr', ('iand', step4
, 0xaaaaaaaa), 1))
340 optimizations
+= [(bitfield_reverse('x'), ('bitfield_reverse', 'x'))]
343 # Add optimizations to handle the case where the result of a ternary is
344 # compared to a constant. This way we can take things like
350 # a ? (0 > 0) : (1 > 0)
352 # which constant folding will eat for lunch. The resulting ternary will
353 # further get cleaned up by the boolean reductions above and we will be
354 # left with just the original variable "a".
355 for op
in ['flt', 'fge', 'feq', 'fne',
356 'ilt', 'ige', 'ieq', 'ine', 'ult', 'uge']:
358 ((op
, ('bcsel', 'a', '#b', '#c'), '#d'),
359 ('bcsel', 'a', (op
, 'b', 'd'), (op
, 'c', 'd'))),
360 ((op
, '#d', ('bcsel', a
, '#b', '#c')),
361 ('bcsel', 'a', (op
, 'd', 'b'), (op
, 'd', 'c'))),
364 # This section contains "late" optimizations that should be run after the
365 # regular optimizations have finished. Optimizations should go here if
366 # they help code generation but do not necessarily produce code that is
367 # more easily optimizable.
368 late_optimizations
= [
369 (('flt', ('fadd', a
, b
), 0.0), ('flt', a
, ('fneg', b
))),
370 (('fge', ('fadd', a
, b
), 0.0), ('fge', a
, ('fneg', b
))),
371 (('feq', ('fadd', a
, b
), 0.0), ('feq', a
, ('fneg', b
))),
372 (('fne', ('fadd', a
, b
), 0.0), ('fne', a
, ('fneg', b
))),
373 (('fdot2', a
, b
), ('fdot_replicated2', a
, b
), 'options->fdot_replicates'),
374 (('fdot3', a
, b
), ('fdot_replicated3', a
, b
), 'options->fdot_replicates'),
375 (('fdot4', a
, b
), ('fdot_replicated4', a
, b
), 'options->fdot_replicates'),
376 (('fdph', a
, b
), ('fdph_replicated', a
, b
), 'options->fdot_replicates'),
379 print nir_algebraic
.AlgebraicPass("nir_opt_algebraic", optimizations
).render()
380 print nir_algebraic
.AlgebraicPass("nir_opt_algebraic_late",
381 late_optimizations
).render()