1 #! /usr/bin/env python 2 # 3 # Copyright (C) 2014 Intel Corporation 4 # 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: 11 # 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 14 # Software. 15 # 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 22 # IN THE SOFTWARE. 23 # 24 # Authors: 25 # Jason Ekstrand (jason (at] jlekstrand.net) 26 27 import nir_algebraic 28 29 # Convenience variables 30 a = 'a' 31 b = 'b' 32 c = 'c' 33 d = 'd' 34 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. 40 # 41 # If the opcode in a search expression is prefixed by a '~' character, this 42 # indicates that the operation is inexact. Such operations will only get 43 # applied to SSA values that do not have the exact bit set. This should be 44 # used by by any optimizations that are not bit-for-bit exact. It should not, 45 # however, be used for backend-requested lowering operations as those need to 46 # happen regardless of precision. 47 # 48 # Variable names are specified as "[#]name[@type][(cond)]" where "#" inicates 49 # that the given variable will only match constants and the type indicates that 50 # the given variable will only match values from ALU instructions with the 51 # given output type, and (cond) specifies an additional condition function 52 # (see nir_search_helpers.h). 53 # 54 # For constants, you have to be careful to make sure that it is the right 55 # type because python is unaware of the source and destination types of the 56 # opcodes. 57 # 58 # All expression types can have a bit-size specified. For opcodes, this 59 # looks like "op@32", for variables it is "a@32" or "a@uint32" to specify a 60 # type and size, and for literals, you can write "2.0@32". In the search half 61 # of the expression this indicates that it should only match that particular 62 # bit-size. In the replace half of the expression this indicates that the 63 # constructed value should have that bit-size. 64 65 optimizations = [ 66 67 (('imul', a, '#b@32(is_pos_power_of_two)'), ('ishl', a, ('find_lsb', b))), 68 (('imul', a, '#b@32(is_neg_power_of_two)'), ('ineg', ('ishl', a, ('find_lsb', ('iabs', b))))), 69 (('udiv', a, 1), a), 70 (('idiv', a, 1), a), 71 (('umod', a, 1), 0), 72 (('imod', a, 1), 0), 73 (('udiv', a, '#b@32(is_pos_power_of_two)'), ('ushr', a, ('find_lsb', b))), 74 (('idiv', a, '#b@32(is_pos_power_of_two)'), ('imul', ('isign', a), ('ushr', ('iabs', a), ('find_lsb', b))), 'options->lower_idiv'), 75 (('idiv', a, '#b@32(is_neg_power_of_two)'), ('ineg', ('imul', ('isign', a), ('ushr', ('iabs', a), ('find_lsb', ('iabs', b))))), 'options->lower_idiv'), 76 (('umod', a, '#b(is_pos_power_of_two)'), ('iand', a, ('isub', b, 1))), 77 78 (('fneg', ('fneg', a)), a), 79 (('ineg', ('ineg', a)), a), 80 (('fabs', ('fabs', a)), ('fabs', a)), 81 (('fabs', ('fneg', a)), ('fabs', a)), 82 (('fabs', ('u2f', a)), ('u2f', a)), 83 (('iabs', ('iabs', a)), ('iabs', a)), 84 (('iabs', ('ineg', a)), ('iabs', a)), 85 (('~fadd', a, 0.0), a), 86 (('iadd', a, 0), a), 87 (('usadd_4x8', a, 0), a), 88 (('usadd_4x8', a, ~0), ~0), 89 (('~fadd', ('fmul', a, b), ('fmul', a, c)), ('fmul', a, ('fadd', b, c))), 90 (('iadd', ('imul', a, b), ('imul', a, c)), ('imul', a, ('iadd', b, c))), 91 (('~fadd', ('fneg', a), a), 0.0), 92 (('iadd', ('ineg', a), a), 0), 93 (('iadd', ('ineg', a), ('iadd', a, b)), b), 94 (('iadd', a, ('iadd', ('ineg', a), b)), b), 95 (('~fadd', ('fneg', a), ('fadd', a, b)), b), 96 (('~fadd', a, ('fadd', ('fneg', a), b)), b), 97 (('~fmul', a, 0.0), 0.0), 98 (('imul', a, 0), 0), 99 (('umul_unorm_4x8', a, 0), 0), 100 (('umul_unorm_4x8', a, ~0), a), 101 (('fmul', a, 1.0), a), 102 (('imul', a, 1), a), 103 (('fmul', a, -1.0), ('fneg', a)), 104 (('imul', a, -1), ('ineg', a)), 105 (('~ffma', 0.0, a, b), b), 106 (('~ffma', a, 0.0, b), b), 107 (('~ffma', a, b, 0.0), ('fmul', a, b)), 108 (('ffma', a, 1.0, b), ('fadd', a, b)), 109 (('ffma', 1.0, a, b), ('fadd', a, b)), 110 (('~flrp', a, b, 0.0), a), 111 (('~flrp', a, b, 1.0), b), 112 (('~flrp', a, a, b), a), 113 (('~flrp', 0.0, a, b), ('fmul', a, b)), 114 (('~flrp', a, b, ('b2f', c)), ('bcsel', c, b, a), 'options->lower_flrp32'), 115 (('~flrp', a, 0.0, c), ('fadd', ('fmul', ('fneg', a), c), a)), 116 (('flrp@32', a, b, c), ('fadd', ('fmul', c, ('fsub', b, a)), a), 'options->lower_flrp32'), 117 (('flrp@64', a, b, c), ('fadd', ('fmul', c, ('fsub', b, a)), a), 'options->lower_flrp64'), 118 (('ffract', a), ('fsub', a, ('ffloor', a)), 'options->lower_ffract'), 119 (('~fadd', ('fmul', a, ('fadd', 1.0, ('fneg', ('b2f', c)))), ('fmul', b, ('b2f', c))), ('bcsel', c, b, a), 'options->lower_flrp32'), 120 (('~fadd@32', ('fmul', a, ('fadd', 1.0, ('fneg', c ))), ('fmul', b, c )), ('flrp', a, b, c), '!options->lower_flrp32'), 121 (('~fadd@64', ('fmul', a, ('fadd', 1.0, ('fneg', c ))), ('fmul', b, c )), ('flrp', a, b, c), '!options->lower_flrp64'), 122 (('~fadd', a, ('fmul', ('b2f', c), ('fadd', b, ('fneg', a)))), ('bcsel', c, b, a), 'options->lower_flrp32'), 123 (('~fadd@32', a, ('fmul', c , ('fadd', b, ('fneg', a)))), ('flrp', a, b, c), '!options->lower_flrp32'), 124 (('~fadd@64', a, ('fmul', c , ('fadd', b, ('fneg', a)))), ('flrp', a, b, c), '!options->lower_flrp64'), 125 (('ffma', a, b, c), ('fadd', ('fmul', a, b), c), 'options->lower_ffma'), 126 (('~fadd', ('fmul', a, b), c), ('ffma', a, b, c), 'options->fuse_ffma'), 127 128 # (a * #b + #c) << #d 129 # ((a * #b) << #d) + (#c << #d) 130 # (a * (#b << #d)) + (#c << #d) 131 (('ishl', ('iadd', ('imul', a, '#b'), '#c'), '#d'), 132 ('iadd', ('imul', a, ('ishl', b, d)), ('ishl', c, d))), 133 134 # (a * #b) << #c 135 # a * (#b << #c) 136 (('ishl', ('imul', a, '#b'), '#c'), ('imul', a, ('ishl', b, c))), 137 138 # Comparison simplifications 139 (('~inot', ('flt', a, b)), ('fge', a, b)), 140 (('~inot', ('fge', a, b)), ('flt', a, b)), 141 (('~inot', ('feq', a, b)), ('fne', a, b)), 142 (('~inot', ('fne', a, b)), ('feq', a, b)), 143 (('inot', ('ilt', a, b)), ('ige', a, b)), 144 (('inot', ('ige', a, b)), ('ilt', a, b)), 145 (('inot', ('ieq', a, b)), ('ine', a, b)), 146 (('inot', ('ine', a, b)), ('ieq', a, b)), 147 148 # 0.0 >= b2f(a) 149 # b2f(a) <= 0.0 150 # b2f(a) == 0.0 because b2f(a) can only be 0 or 1 151 # inot(a) 152 (('fge', 0.0, ('b2f', a)), ('inot', a)), 153 154 (('fge', ('fneg', ('b2f', a)), 0.0), ('inot', a)), 155 156 # 0.0 < fabs(a) 157 # fabs(a) > 0.0 158 # fabs(a) != 0.0 because fabs(a) must be >= 0 159 # a != 0.0 160 (('flt', 0.0, ('fabs', a)), ('fne', a, 0.0)), 161 162 (('fge', ('fneg', ('fabs', a)), 0.0), ('feq', a, 0.0)), 163 (('bcsel', ('flt', b, a), b, a), ('fmin', a, b)), 164 (('bcsel', ('flt', a, b), b, a), ('fmax', a, b)), 165 (('bcsel', ('inot', a), b, c), ('bcsel', a, c, b)), 166 (('bcsel', a, ('bcsel', a, b, c), d), ('bcsel', a, b, d)), 167 (('bcsel', a, True, 'b@bool'), ('ior', a, b)), 168 (('fmin', a, a), a), 169 (('fmax', a, a), a), 170 (('imin', a, a), a), 171 (('imax', a, a), a), 172 (('umin', a, a), a), 173 (('umax', a, a), a), 174 (('~fmin', ('fmax', a, 0.0), 1.0), ('fsat', a), '!options->lower_fsat'), 175 (('~fmax', ('fmin', a, 1.0), 0.0), ('fsat', a), '!options->lower_fsat'), 176 (('fsat', a), ('fmin', ('fmax', a, 0.0), 1.0), 'options->lower_fsat'), 177 (('fsat', ('fsat', a)), ('fsat', a)), 178 (('fmin', ('fmax', ('fmin', ('fmax', a, b), c), b), c), ('fmin', ('fmax', a, b), c)), 179 (('imin', ('imax', ('imin', ('imax', a, b), c), b), c), ('imin', ('imax', a, b), c)), 180 (('umin', ('umax', ('umin', ('umax', a, b), c), b), c), ('umin', ('umax', a, b), c)), 181 (('fmax', ('fsat', a), '#b@32(is_zero_to_one)'), ('fsat', ('fmax', a, b))), 182 (('fmin', ('fsat', a), '#b@32(is_zero_to_one)'), ('fsat', ('fmin', a, b))), 183 (('extract_u8', ('imin', ('imax', a, 0), 0xff), 0), ('imin', ('imax', a, 0), 0xff)), 184 (('~ior', ('flt', a, b), ('flt', a, c)), ('flt', a, ('fmax', b, c))), 185 (('~ior', ('flt', a, c), ('flt', b, c)), ('flt', ('fmin', a, b), c)), 186 (('~ior', ('fge', a, b), ('fge', a, c)), ('fge', a, ('fmin', b, c))), 187 (('~ior', ('fge', a, c), ('fge', b, c)), ('fge', ('fmax', a, b), c)), 188 (('fabs', ('slt', a, b)), ('slt', a, b)), 189 (('fabs', ('sge', a, b)), ('sge', a, b)), 190 (('fabs', ('seq', a, b)), ('seq', a, b)), 191 (('fabs', ('sne', a, b)), ('sne', a, b)), 192 (('slt', a, b), ('b2f', ('flt', a, b)), 'options->lower_scmp'), 193 (('sge', a, b), ('b2f', ('fge', a, b)), 'options->lower_scmp'), 194 (('seq', a, b), ('b2f', ('feq', a, b)), 'options->lower_scmp'), 195 (('sne', a, b), ('b2f', ('fne', a, b)), 'options->lower_scmp'), 196 (('fne', ('fneg', a), a), ('fne', a, 0.0)), 197 (('feq', ('fneg', a), a), ('feq', a, 0.0)), 198 # Emulating booleans 199 (('imul', ('b2i', a), ('b2i', b)), ('b2i', ('iand', a, b))), 200 (('fmul', ('b2f', a), ('b2f', b)), ('b2f', ('iand', a, b))), 201 (('fsat', ('fadd', ('b2f', a), ('b2f', b))), ('b2f', ('ior', a, b))), 202 (('iand', 'a@bool', 1.0), ('b2f', a)), 203 # True/False are ~0 and 0 in NIR. b2i of True is 1, and -1 is ~0 (True). 204 (('ineg', ('b2i', a)), a), 205 (('flt', ('fneg', ('b2f', a)), 0), a), # Generated by TGSI KILL_IF. 206 (('flt', ('fsub', 0.0, ('b2f', a)), 0), a), # Generated by TGSI KILL_IF. 207 # Comparison with the same args. Note that these are not done for 208 # the float versions because NaN always returns false on float 209 # inequalities. 210 (('ilt', a, a), False), 211 (('ige', a, a), True), 212 (('ieq', a, a), True), 213 (('ine', a, a), False), 214 (('ult', a, a), False), 215 (('uge', a, a), True), 216 # Logical and bit operations 217 (('fand', a, 0.0), 0.0), 218 (('iand', a, a), a), 219 (('iand', a, ~0), a), 220 (('iand', a, 0), 0), 221 (('ior', a, a), a), 222 (('ior', a, 0), a), 223 (('ior', a, True), True), 224 (('fxor', a, a), 0.0), 225 (('ixor', a, a), 0), 226 (('ixor', a, 0), a), 227 (('inot', ('inot', a)), a), 228 # DeMorgan's Laws 229 (('iand', ('inot', a), ('inot', b)), ('inot', ('ior', a, b))), 230 (('ior', ('inot', a), ('inot', b)), ('inot', ('iand', a, b))), 231 # Shift optimizations 232 (('ishl', 0, a), 0), 233 (('ishl', a, 0), a), 234 (('ishr', 0, a), 0), 235 (('ishr', a, 0), a), 236 (('ushr', 0, a), 0), 237 (('ushr', a, 0), a), 238 (('iand', 0xff, ('ushr', a, 24)), ('ushr', a, 24)), 239 (('iand', 0xffff, ('ushr', a, 16)), ('ushr', a, 16)), 240 # Exponential/logarithmic identities 241 (('~fexp2', ('flog2', a)), a), # 2^lg2(a) = a 242 (('~flog2', ('fexp2', a)), a), # lg2(2^a) = a 243 (('fpow', a, b), ('fexp2', ('fmul', ('flog2', a), b)), 'options->lower_fpow'), # a^b = 2^(lg2(a)*b) 244 (('~fexp2', ('fmul', ('flog2', a), b)), ('fpow', a, b), '!options->lower_fpow'), # 2^(lg2(a)*b) = a^b 245 (('~fexp2', ('fadd', ('fmul', ('flog2', a), b), ('fmul', ('flog2', c), d))), 246 ('~fmul', ('fpow', a, b), ('fpow', c, d)), '!options->lower_fpow'), # 2^(lg2(a) * b + lg2(c) + d) = a^b * c^d 247 (('~fpow', a, 1.0), a), 248 (('~fpow', a, 2.0), ('fmul', a, a)), 249 (('~fpow', a, 4.0), ('fmul', ('fmul', a, a), ('fmul', a, a))), 250 (('~fpow', 2.0, a), ('fexp2', a)), 251 (('~fpow', ('fpow', a, 2.2), 0.454545), a), 252 (('~fpow', ('fabs', ('fpow', a, 2.2)), 0.454545), ('fabs', a)), 253 (('~fsqrt', ('fexp2', a)), ('fexp2', ('fmul', 0.5, a))), 254 (('~frcp', ('fexp2', a)), ('fexp2', ('fneg', a))), 255 (('~frsq', ('fexp2', a)), ('fexp2', ('fmul', -0.5, a))), 256 (('~flog2', ('fsqrt', a)), ('fmul', 0.5, ('flog2', a))), 257 (('~flog2', ('frcp', a)), ('fneg', ('flog2', a))), 258 (('~flog2', ('frsq', a)), ('fmul', -0.5, ('flog2', a))), 259 (('~flog2', ('fpow', a, b)), ('fmul', b, ('flog2', a))), 260 (('~fmul', ('fexp2', a), ('fexp2', b)), ('fexp2', ('fadd', a, b))), 261 # Division and reciprocal 262 (('~fdiv', 1.0, a), ('frcp', a)), 263 (('fdiv', a, b), ('fmul', a, ('frcp', b)), 'options->lower_fdiv'), 264 (('~frcp', ('frcp', a)), a), 265 (('~frcp', ('fsqrt', a)), ('frsq', a)), 266 (('fsqrt', a), ('frcp', ('frsq', a)), 'options->lower_fsqrt'), 267 (('~frcp', ('frsq', a)), ('fsqrt', a), '!options->lower_fsqrt'), 268 # Boolean simplifications 269 (('ieq', 'a@bool', True), a), 270 (('ine(is_not_used_by_if)', 'a@bool', True), ('inot', a)), 271 (('ine', 'a@bool', False), a), 272 (('ieq(is_not_used_by_if)', 'a@bool', False), ('inot', 'a')), 273 (('bcsel', a, True, False), a), 274 (('bcsel', a, False, True), ('inot', a)), 275 (('bcsel@32', a, 1.0, 0.0), ('b2f', a)), 276 (('bcsel@32', a, 0.0, 1.0), ('b2f', ('inot', a))), 277 (('bcsel@32', a, -1.0, -0.0), ('fneg', ('b2f', a))), 278 (('bcsel@32', a, -0.0, -1.0), ('fneg', ('b2f', ('inot', a)))), 279 (('bcsel', True, b, c), b), 280 (('bcsel', False, b, c), c), 281 # The result of this should be hit by constant propagation and, in the 282 # next round of opt_algebraic, get picked up by one of the above two. 283 (('bcsel', '#a', b, c), ('bcsel', ('ine', 'a', 0), b, c)), 284 285 (('bcsel', a, b, b), b), 286 (('fcsel', a, b, b), b), 287 288 # Conversions 289 (('i2b', ('b2i', a)), a), 290 (('f2i', ('ftrunc', a)), ('f2i', a)), 291 (('f2u', ('ftrunc', a)), ('f2u', a)), 292 (('i2b', ('ineg', a)), ('i2b', a)), 293 (('i2b', ('iabs', a)), ('i2b', a)), 294 (('fabs', ('b2f', a)), ('b2f', a)), 295 (('iabs', ('b2i', a)), ('b2i', a)), 296 297 # Byte extraction 298 (('ushr', a, 24), ('extract_u8', a, 3), '!options->lower_extract_byte'), 299 (('iand', 0xff, ('ushr', a, 16)), ('extract_u8', a, 2), '!options->lower_extract_byte'), 300 (('iand', 0xff, ('ushr', a, 8)), ('extract_u8', a, 1), '!options->lower_extract_byte'), 301 (('iand', 0xff, a), ('extract_u8', a, 0), '!options->lower_extract_byte'), 302 303 # Word extraction 304 (('ushr', a, 16), ('extract_u16', a, 1), '!options->lower_extract_word'), 305 (('iand', 0xffff, a), ('extract_u16', a, 0), '!options->lower_extract_word'), 306 307 # Subtracts 308 (('~fsub', a, ('fsub', 0.0, b)), ('fadd', a, b)), 309 (('isub', a, ('isub', 0, b)), ('iadd', a, b)), 310 (('ussub_4x8', a, 0), a), 311 (('ussub_4x8', a, ~0), 0), 312 (('fsub', a, b), ('fadd', a, ('fneg', b)), 'options->lower_sub'), 313 (('isub', a, b), ('iadd', a, ('ineg', b)), 'options->lower_sub'), 314 (('fneg', a), ('fsub', 0.0, a), 'options->lower_negate'), 315 (('ineg', a), ('isub', 0, a), 'options->lower_negate'), 316 (('~fadd', a, ('fsub', 0.0, b)), ('fsub', a, b)), 317 (('iadd', a, ('isub', 0, b)), ('isub', a, b)), 318 (('fabs', ('fsub', 0.0, a)), ('fabs', a)), 319 (('iabs', ('isub', 0, a)), ('iabs', a)), 320 321 # Propagate negation up multiplication chains 322 (('fmul', ('fneg', a), b), ('fneg', ('fmul', a, b))), 323 (('imul', ('ineg', a), b), ('ineg', ('imul', a, b))), 324 325 # Reassociate constants in add/mul chains so they can be folded together. 326 # For now, we only handle cases where the constants are separated by 327 # a single non-constant. We could do better eventually. 328 (('~fmul', '#a', ('fmul', b, '#c')), ('fmul', ('fmul', a, c), b)), 329 (('imul', '#a', ('imul', b, '#c')), ('imul', ('imul', a, c), b)), 330 (('~fadd', '#a', ('fadd', b, '#c')), ('fadd', ('fadd', a, c), b)), 331 (('iadd', '#a', ('iadd', b, '#c')), ('iadd', ('iadd', a, c), b)), 332 333 # Misc. lowering 334 (('fmod@32', a, b), ('fsub', a, ('fmul', b, ('ffloor', ('fdiv', a, b)))), 'options->lower_fmod32'), 335 (('fmod@64', a, b), ('fsub', a, ('fmul', b, ('ffloor', ('fdiv', a, b)))), 'options->lower_fmod64'), 336 (('frem', a, b), ('fsub', a, ('fmul', b, ('ftrunc', ('fdiv', a, b)))), 'options->lower_fmod32'), 337 (('uadd_carry@32', a, b), ('b2i', ('ult', ('iadd', a, b), a)), 'options->lower_uadd_carry'), 338 (('usub_borrow@32', a, b), ('b2i', ('ult', a, b)), 'options->lower_usub_borrow'), 339 340 (('bitfield_insert', 'base', 'insert', 'offset', 'bits'), 341 ('bcsel', ('ilt', 31, 'bits'), 'insert', 342 ('bfi', ('bfm', 'bits', 'offset'), 'insert', 'base')), 343 'options->lower_bitfield_insert'), 344 345 (('ibitfield_extract', 'value', 'offset', 'bits'), 346 ('bcsel', ('ilt', 31, 'bits'), 'value', 347 ('ibfe', 'value', 'offset', 'bits')), 348 'options->lower_bitfield_extract'), 349 350 (('ubitfield_extract', 'value', 'offset', 'bits'), 351 ('bcsel', ('ult', 31, 'bits'), 'value', 352 ('ubfe', 'value', 'offset', 'bits')), 353 'options->lower_bitfield_extract'), 354 355 (('extract_i8', a, b), 356 ('ishr', ('ishl', a, ('imul', ('isub', 3, b), 8)), 24), 357 'options->lower_extract_byte'), 358 359 (('extract_u8', a, b), 360 ('iand', ('ushr', a, ('imul', b, 8)), 0xff), 361 'options->lower_extract_byte'), 362 363 (('extract_i16', a, b), 364 ('ishr', ('ishl', a, ('imul', ('isub', 1, b), 16)), 16), 365 'options->lower_extract_word'), 366 367 (('extract_u16', a, b), 368 ('iand', ('ushr', a, ('imul', b, 16)), 0xffff), 369 'options->lower_extract_word'), 370 371 (('pack_unorm_2x16', 'v'), 372 ('pack_uvec2_to_uint', 373 ('f2u', ('fround_even', ('fmul', ('fsat', 'v'), 65535.0)))), 374 'options->lower_pack_unorm_2x16'), 375 376 (('pack_unorm_4x8', 'v'), 377 ('pack_uvec4_to_uint', 378 ('f2u', ('fround_even', ('fmul', ('fsat', 'v'), 255.0)))), 379 'options->lower_pack_unorm_4x8'), 380 381 (('pack_snorm_2x16', 'v'), 382 ('pack_uvec2_to_uint', 383 ('f2i', ('fround_even', ('fmul', ('fmin', 1.0, ('fmax', -1.0, 'v')), 32767.0)))), 384 'options->lower_pack_snorm_2x16'), 385 386 (('pack_snorm_4x8', 'v'), 387 ('pack_uvec4_to_uint', 388 ('f2i', ('fround_even', ('fmul', ('fmin', 1.0, ('fmax', -1.0, 'v')), 127.0)))), 389 'options->lower_pack_snorm_4x8'), 390 391 (('unpack_unorm_2x16', 'v'), 392 ('fdiv', ('u2f', ('vec2', ('extract_u16', 'v', 0), 393 ('extract_u16', 'v', 1))), 394 65535.0), 395 'options->lower_unpack_unorm_2x16'), 396 397 (('unpack_unorm_4x8', 'v'), 398 ('fdiv', ('u2f', ('vec4', ('extract_u8', 'v', 0), 399 ('extract_u8', 'v', 1), 400 ('extract_u8', 'v', 2), 401 ('extract_u8', 'v', 3))), 402 255.0), 403 'options->lower_unpack_unorm_4x8'), 404 405 (('unpack_snorm_2x16', 'v'), 406 ('fmin', 1.0, ('fmax', -1.0, ('fdiv', ('i2f', ('vec2', ('extract_i16', 'v', 0), 407 ('extract_i16', 'v', 1))), 408 32767.0))), 409 'options->lower_unpack_snorm_2x16'), 410 411 (('unpack_snorm_4x8', 'v'), 412 ('fmin', 1.0, ('fmax', -1.0, ('fdiv', ('i2f', ('vec4', ('extract_i8', 'v', 0), 413 ('extract_i8', 'v', 1), 414 ('extract_i8', 'v', 2), 415 ('extract_i8', 'v', 3))), 416 127.0))), 417 'options->lower_unpack_snorm_4x8'), 418 ] 419 420 def fexp2i(exp, bits): 421 # We assume that exp is already in the right range. 422 if bits == 32: 423 return ('ishl', ('iadd', exp, 127), 23) 424 elif bits == 64: 425 return ('pack_double_2x32_split', 0, ('ishl', ('iadd', exp, 1023), 20)) 426 else: 427 assert False 428 429 def ldexp(f, exp, bits): 430 # First, we clamp exp to a reasonable range. The maximum possible range 431 # for a normal exponent is [-126, 127] and, throwing in denormals, you get 432 # a maximum range of [-149, 127]. This means that we can potentially have 433 # a swing of +-276. If you start with FLT_MAX, you actually have to do 434 # ldexp(FLT_MAX, -278) to get it to flush all the way to zero. The GLSL 435 # spec, on the other hand, only requires that we handle an exponent value 436 # in the range [-126, 128]. This implementation is *mostly* correct; it 437 # handles a range on exp of [-252, 254] which allows you to create any 438 # value (including denorms if the hardware supports it) and to adjust the 439 # exponent of any normal value to anything you want. 440 if bits == 32: 441 exp = ('imin', ('imax', exp, -252), 254) 442 elif bits == 64: 443 exp = ('imin', ('imax', exp, -2044), 2046) 444 else: 445 assert False 446 447 # Now we compute two powers of 2, one for exp/2 and one for exp-exp/2. 448 # (We use ishr which isn't the same for -1, but the -1 case still works 449 # since we use exp-exp/2 as the second exponent.) While the spec 450 # technically defines ldexp as f * 2.0^exp, simply multiplying once doesn't 451 # work with denormals and doesn't allow for the full swing in exponents 452 # that you can get with normalized values. Instead, we create two powers 453 # of two and multiply by them each in turn. That way the effective range 454 # of our exponent is doubled. 455 pow2_1 = fexp2i(('ishr', exp, 1), bits) 456 pow2_2 = fexp2i(('isub', exp, ('ishr', exp, 1)), bits) 457 return ('fmul', ('fmul', f, pow2_1), pow2_2) 458 459 optimizations += [ 460 (('ldexp@32', 'x', 'exp'), ldexp('x', 'exp', 32)), 461 (('ldexp@64', 'x', 'exp'), ldexp('x', 'exp', 64)), 462 ] 463 464 # Unreal Engine 4 demo applications open-codes bitfieldReverse() 465 def bitfield_reverse(u): 466 step1 = ('ior', ('ishl', u, 16), ('ushr', u, 16)) 467 step2 = ('ior', ('ishl', ('iand', step1, 0x00ff00ff), 8), ('ushr', ('iand', step1, 0xff00ff00), 8)) 468 step3 = ('ior', ('ishl', ('iand', step2, 0x0f0f0f0f), 4), ('ushr', ('iand', step2, 0xf0f0f0f0), 4)) 469 step4 = ('ior', ('ishl', ('iand', step3, 0x33333333), 2), ('ushr', ('iand', step3, 0xcccccccc), 2)) 470 step5 = ('ior', ('ishl', ('iand', step4, 0x55555555), 1), ('ushr', ('iand', step4, 0xaaaaaaaa), 1)) 471 472 return step5 473 474 optimizations += [(bitfield_reverse('x@32'), ('bitfield_reverse', 'x'))] 475 476 # For any float comparison operation, "cmp", if you have "a == a && a cmp b" 477 # then the "a == a" is redundant because it's equivalent to "a is not NaN" 478 # and, if a is a NaN then the second comparison will fail anyway. 479 for op in ['flt', 'fge', 'feq']: 480 optimizations += [ 481 (('iand', ('feq', a, a), (op, a, b)), (op, a, b)), 482 (('iand', ('feq', a, a), (op, b, a)), (op, b, a)), 483 ] 484 485 # Add optimizations to handle the case where the result of a ternary is 486 # compared to a constant. This way we can take things like 487 # 488 # (a ? 0 : 1) > 0 489 # 490 # and turn it into 491 # 492 # a ? (0 > 0) : (1 > 0) 493 # 494 # which constant folding will eat for lunch. The resulting ternary will 495 # further get cleaned up by the boolean reductions above and we will be 496 # left with just the original variable "a". 497 for op in ['flt', 'fge', 'feq', 'fne', 498 'ilt', 'ige', 'ieq', 'ine', 'ult', 'uge']: 499 optimizations += [ 500 ((op, ('bcsel', 'a', '#b', '#c'), '#d'), 501 ('bcsel', 'a', (op, 'b', 'd'), (op, 'c', 'd'))), 502 ((op, '#d', ('bcsel', a, '#b', '#c')), 503 ('bcsel', 'a', (op, 'd', 'b'), (op, 'd', 'c'))), 504 ] 505 506 # This section contains "late" optimizations that should be run after the 507 # regular optimizations have finished. Optimizations should go here if 508 # they help code generation but do not necessarily produce code that is 509 # more easily optimizable. 510 late_optimizations = [ 511 # Most of these optimizations aren't quite safe when you get infinity or 512 # Nan involved but the first one should be fine. 513 (('flt', ('fadd', a, b), 0.0), ('flt', a, ('fneg', b))), 514 (('~fge', ('fadd', a, b), 0.0), ('fge', a, ('fneg', b))), 515 (('~feq', ('fadd', a, b), 0.0), ('feq', a, ('fneg', b))), 516 (('~fne', ('fadd', a, b), 0.0), ('fne', a, ('fneg', b))), 517 518 (('fdot2', a, b), ('fdot_replicated2', a, b), 'options->fdot_replicates'), 519 (('fdot3', a, b), ('fdot_replicated3', a, b), 'options->fdot_replicates'), 520 (('fdot4', a, b), ('fdot_replicated4', a, b), 'options->fdot_replicates'), 521 (('fdph', a, b), ('fdph_replicated', a, b), 'options->fdot_replicates'), 522 523 (('b2f(is_used_more_than_once)', ('inot', a)), ('bcsel', a, 0.0, 1.0)), 524 (('fneg(is_used_more_than_once)', ('b2f', ('inot', a))), ('bcsel', a, -0.0, -1.0)), 525 526 # we do these late so that we don't get in the way of creating ffmas 527 (('fmin', ('fadd(is_used_once)', '#c', a), ('fadd(is_used_once)', '#c', b)), ('fadd', c, ('fmin', a, b))), 528 (('fmax', ('fadd(is_used_once)', '#c', a), ('fadd(is_used_once)', '#c', b)), ('fadd', c, ('fmax', a, b))), 529 ] 530 531 print nir_algebraic.AlgebraicPass("nir_opt_algebraic", optimizations).render() 532 print nir_algebraic.AlgebraicPass("nir_opt_algebraic_late", 533 late_optimizations).render() 534
