src/glsl/nir/nir_opcodes.py

   1 #! /usr/bin/env python
   2 #
   3 # Copyright (C) 2014 Connor Abbott
   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 #    Connor Abbott (cwabbott0@gmail.com)
  26
  27
  28 # Class that represents all the information we have about the opcode
  29 # NOTE: this must be kept in sync with nir_op_info
  30
  31 class Opcode(object):
  32    """Class that represents all the information we have about the opcode
  33    NOTE: this must be kept in sync with nir_op_info
  34    """
  35    def __init__(self, name, output_size, output_type, input_sizes,
  36                 input_types, algebraic_properties, const_expr):
  37       """Parameters:
  38
  39       - name is the name of the opcode (prepend nir_op_ for the enum name)
  40       - all types are strings that get nir_type_ prepended to them
  41       - input_types is a list of types
  42       - algebraic_properties is a space-seperated string, where nir_op_is_ is
  43         prepended before each entry
  44       - const_expr is an expression or series of statements that computes the
  45         constant value of the opcode given the constant values of its inputs.
  46
  47       Constant expressions are formed from the variables src0, src1, ...,
  48       src(N-1), where N is the number of arguments.  The output of the
  49       expression should be stored in the dst variable.  Per-component input
  50       and output variables will be scalars and non-per-component input and
  51       output variables will be a struct with fields named x, y, z, and w
  52       all of the correct type.  Input and output variables can be assumed
  53       to already be of the correct type and need no conversion.  In
  54       particular, the conversion from the C bool type to/from  NIR_TRUE and
  55       NIR_FALSE happens automatically.
  56
  57       For per-component instructions, the entire expression will be
  58       executed once for each component.  For non-per-component
  59       instructions, the expression is expected to store the correct values
  60       in dst.x, dst.y, etc.  If "dst" does not exist anywhere in the
  61       constant expression, an assignment to dst will happen automatically
  62       and the result will be equivalent to "dst = <expression>" for
  63       per-component instructions and "dst.x = dst.y = ... = <expression>"
  64       for non-per-component instructions.
  65       """
  66       assert isinstance(name, str)
  67       assert isinstance(output_size, int)
  68       assert isinstance(output_type, str)
  69       assert isinstance(input_sizes, list)
  70       assert isinstance(input_sizes[0], int)
  71       assert isinstance(input_types, list)
  72       assert isinstance(input_types[0], str)
  73       assert isinstance(algebraic_properties, str)
  74       assert isinstance(const_expr, str)
  75       assert len(input_sizes) == len(input_types)
  76       assert 0 <= output_size <= 4
  77       for size in input_sizes:
  78          assert 0 <= size <= 4
  79          if output_size != 0:
  80             assert size != 0
  81       self.name = name
  82       self.num_inputs = len(input_sizes)
  83       self.output_size = output_size
  84       self.output_type = output_type
  85       self.input_sizes = input_sizes
  86       self.input_types = input_types
  87       self.algebraic_properties = algebraic_properties
  88       self.const_expr = const_expr
  89
  90 # helper variables for strings
  91 tfloat = "float"
  92 tint = "int"
  93 tbool = "bool"
  94 tunsigned = "unsigned"
  95
  96 commutative = "commutative "
  97 associative = "associative "
  98
  99 # global dictionary of opcodes
 100 opcodes = {}
 101
 102 def opcode(name, output_size, output_type, input_sizes, input_types,
 103            algebraic_properties, const_expr):
 104    assert name not in opcodes
 105    opcodes[name] = Opcode(name, output_size, output_type, input_sizes,
 106                           input_types, algebraic_properties, const_expr)
 107
 108 def unop_convert(name, in_type, out_type, const_expr):
 109    opcode(name, 0, out_type, [0], [in_type], "", const_expr)
 110
 111 def unop(name, ty, const_expr):
 112    opcode(name, 0, ty, [0], [ty], "", const_expr)
 113
 114 def unop_horiz(name, output_size, output_type, input_size, input_type,
 115                const_expr):
 116    opcode(name, output_size, output_type, [input_size], [input_type], "",
 117           const_expr)
 118
 119 def unop_reduce(name, output_size, output_type, input_type, prereduce_expr,
 120                 reduce_expr, final_expr):
 121    def prereduce(src):
 122       return "(" + prereduce_expr.format(src=src) + ")"
 123    def final(src):
 124       return final_expr.format(src="(" + src + ")")
 125    def reduce_(src0, src1):
 126       return reduce_expr.format(src0=src0, src1=src1)
 127    src0 = prereduce("src0.x")
 128    src1 = prereduce("src0.y")
 129    src2 = prereduce("src0.z")
 130    src3 = prereduce("src0.w")
 131    unop_horiz(name + "2", output_size, output_type, 2, input_type,
 132               final(reduce_(src0, src1)))
 133    unop_horiz(name + "3", output_size, output_type, 3, input_type,
 134               final(reduce_(reduce_(src0, src1), src2)))
 135    unop_horiz(name + "4", output_size, output_type, 4, input_type,
 136               final(reduce_(reduce_(src0, src1), reduce_(src2, src3))))
 137
 138
 139 # These two move instructions differ in what modifiers they support and what
 140 # the negate modifier means. Otherwise, they are identical.
 141 unop("fmov", tfloat, "src0")
 142 unop("imov", tint, "src0")
 143
 144 unop("ineg", tint, "-src0")
 145 unop("fneg", tfloat, "-src0")
 146 unop("inot", tint, "~src0") # invert every bit of the integer
 147 unop("fnot", tfloat, "(src0 == 0.0f) ? 1.0f : 0.0f")
 148 unop("fsign", tfloat, "(src0 == 0.0f) ? 0.0f : ((src0 > 0.0f) ? 1.0f : -1.0f)")
 149 unop("isign", tint, "(src0 == 0) ? 0 : ((src0 > 0) ? 1 : -1)")
 150 unop("iabs", tint, "(src0 < 0) ? -src0 : src0")
 151 unop("fabs", tfloat, "fabsf(src0)")
 152 unop("fsat", tfloat, "(src0 > 1.0f) ? 1.0f : ((src0 <= 0.0f) ? 0.0f : src0)")
 153 unop("frcp", tfloat, "1.0f / src0")
 154 unop("frsq", tfloat, "1.0f / sqrtf(src0)")
 155 unop("fsqrt", tfloat, "sqrtf(src0)")
 156 unop("fexp2", tfloat, "exp2f(src0)")
 157 unop("flog2", tfloat, "log2f(src0)")
 158 unop_convert("f2i", tfloat, tint, "src0") # Float-to-integer conversion.
 159 unop_convert("f2u", tfloat, tunsigned, "src0") # Float-to-unsigned conversion
 160 unop_convert("i2f", tint, tfloat, "src0") # Integer-to-float conversion.
 161 # Float-to-boolean conversion
 162 unop_convert("f2b", tfloat, tbool, "src0 != 0.0f")
 163 # Boolean-to-float conversion
 164 unop_convert("b2f", tbool, tfloat, "src0 ? 1.0f : 0.0f")
 165 # Int-to-boolean conversion
 166 unop_convert("i2b", tint, tbool, "src0 != 0")
 167 unop_convert("b2i", tbool, tint, "src0 ? 1 : 0") # Boolean-to-int conversion
 168 unop_convert("u2f", tunsigned, tfloat, "src0") #Unsigned-to-float conversion.
 169
 170 unop_reduce("bany", 1, tbool, tbool, "{src}", "{src0} || {src1}", "{src}")
 171 unop_reduce("ball", 1, tbool, tbool, "{src}", "{src0} && {src1}", "{src}")
 172 unop_reduce("fany", 1, tfloat, tfloat, "{src} != 0.0f", "{src0} || {src1}",
 173             "{src} ? 1.0f : 0.0f")
 174 unop_reduce("fall", 1, tfloat, tfloat, "{src} != 0.0f", "{src0} && {src1}",
 175             "{src} ? 1.0f : 0.0f")
 176
 177 # Unary floating-point rounding operations.
 178
 179
 180 unop("ftrunc", tfloat, "truncf(src0)")
 181 unop("fceil", tfloat, "ceilf(src0)")
 182 unop("ffloor", tfloat, "floorf(src0)")
 183 unop("ffract", tfloat, "src0 - floorf(src0)")
 184 unop("fround_even", tfloat, "_mesa_roundevenf(src0)")
 185
 186
 187 # Trigonometric operations.
 188
 189
 190 unop("fsin", tfloat, "sinf(src0)")
 191 unop("fcos", tfloat, "cosf(src0)")
 192
 193
 194 # Partial derivatives.
 195
 196
 197 unop("fddx", tfloat, "0.0f") # the derivative of a constant is 0.
 198 unop("fddy", tfloat, "0.0f")
 199 unop("fddx_fine", tfloat, "0.0f")
 200 unop("fddy_fine", tfloat, "0.0f")
 201 unop("fddx_coarse", tfloat, "0.0f")
 202 unop("fddy_coarse", tfloat, "0.0f")
 203
 204
 205 # Floating point pack and unpack operations.
 206
 207 def pack_2x16(fmt):
 208    unop_horiz("pack_" + fmt + "_2x16", 1, tunsigned, 2, tfloat, """
 209 dst.x = (uint32_t) pack_fmt_1x16(src0.x);
 210 dst.x |= ((uint32_t) pack_fmt_1x16(src0.y)) << 16;
 211 """.replace("fmt", fmt))
 212
 213 def pack_4x8(fmt):
 214    unop_horiz("pack_" + fmt + "_4x8", 1, tunsigned, 4, tfloat, """
 215 dst.x = (uint32_t) pack_fmt_1x8(src0.x);
 216 dst.x |= ((uint32_t) pack_fmt_1x8(src0.y)) << 8;
 217 dst.x |= ((uint32_t) pack_fmt_1x8(src0.z)) << 16;
 218 dst.x |= ((uint32_t) pack_fmt_1x8(src0.w)) << 24;
 219 """.replace("fmt", fmt))
 220
 221 def unpack_2x16(fmt):
 222    unop_horiz("unpack_" + fmt + "_2x16", 2, tfloat, 1, tunsigned, """
 223 dst.x = unpack_fmt_1x16((uint16_t)(src0.x & 0xffff));
 224 dst.y = unpack_fmt_1x16((uint16_t)(src0.x << 16));
 225 """.replace("fmt", fmt))
 226
 227 def unpack_4x8(fmt):
 228    unop_horiz("unpack_" + fmt + "_4x8", 4, tfloat, 1, tunsigned, """
 229 dst.x = unpack_fmt_1x8((uint8_t)(src0.x & 0xff));
 230 dst.y = unpack_fmt_1x8((uint8_t)((src0.x >> 8) & 0xff));
 231 dst.z = unpack_fmt_1x8((uint8_t)((src0.x >> 16) & 0xff));
 232 dst.w = unpack_fmt_1x8((uint8_t)(src0.x >> 24));
 233 """.replace("fmt", fmt))
 234
 235
 236 pack_2x16("snorm")
 237 pack_4x8("snorm")
 238 pack_2x16("unorm")
 239 pack_4x8("unorm")
 240 pack_2x16("half")
 241 unpack_2x16("snorm")
 242 unpack_4x8("snorm")
 243 unpack_2x16("unorm")
 244 unpack_4x8("unorm")
 245 unpack_2x16("half")
 246
 247
 248 # Lowered floating point unpacking operations.
 249
 250
 251 unop_horiz("unpack_half_2x16_split_x", 1, tfloat, 1, tunsigned,
 252            "unpack_half_1x16((uint16_t)(src0.x & 0xffff))")
 253 unop_horiz("unpack_half_2x16_split_y", 1, tfloat, 1, tunsigned,
 254            "unpack_half_1x16((uint16_t)(src0.x >> 16))")
 255
 256
 257 # Bit operations, part of ARB_gpu_shader5.
 258
 259
 260 unop("bitfield_reverse", tunsigned, """
 261 /* we're not winning any awards for speed here, but that's ok */
 262 dst = 0;
 263 for (unsigned bit = 0; bit < 32; bit++)
 264    dst |= ((src0 >> bit) & 1) << (31 - bit);
 265 """)
 266 unop("bit_count", tunsigned, """
 267 dst = 0;
 268 for (unsigned bit = 0; bit < 32; bit++) {
 269    if ((src0 >> bit) & 1)
 270       dst++;
 271 }
 272 """)
 273
 274 unop_convert("ufind_msb", tunsigned, tint, """
 275 dst = -1;
 276 for (int bit = 31; bit > 0; bit--) {
 277    if ((src0 >> bit) & 1) {
 278       dst = bit;
 279       break;
 280    }
 281 }
 282 """)
 283
 284 unop("ifind_msb", tint, """
 285 dst = -1;
 286 for (int bit = 31; bit >= 0; bit--) {
 287    /* If src0 < 0, we're looking for the first 0 bit.
 288     * if src0 >= 0, we're looking for the first 1 bit.
 289     */
 290    if ((((src0 >> bit) & 1) && (src0 >= 0)) ||
 291       (!((src0 >> bit) & 1) && (src0 < 0))) {
 292       dst = bit;
 293       break;
 294    }
 295 }
 296 """)
 297
 298 unop("find_lsb", tint, """
 299 dst = -1;
 300 for (unsigned bit = 0; bit < 32; bit++) {
 301    if ((src0 >> bit) & 1) {
 302       dst = bit;
 303       break;
 304    }
 305 }
 306 """)
 307
 308
 309 for i in xrange(1, 5):
 310    for j in xrange(1, 5):
 311       unop_horiz("fnoise{0}_{1}".format(i, j), i, tfloat, j, tfloat, "0.0f")
 312
 313 def binop_convert(name, out_type, in_type, alg_props, const_expr):
 314    opcode(name, 0, out_type, [0, 0], [in_type, in_type], alg_props, const_expr)
 315
 316 def binop(name, ty, alg_props, const_expr):
 317    binop_convert(name, ty, ty, alg_props, const_expr)
 318
 319 def binop_compare(name, ty, alg_props, const_expr):
 320    binop_convert(name, tbool, ty, alg_props, const_expr)
 321
 322 def binop_horiz(name, out_size, out_type, src1_size, src1_type, src2_size,
 323                 src2_type, const_expr):
 324    opcode(name, out_size, out_type, [src1_size, src2_size], [src1_type, src2_type],
 325           "", const_expr)
 326
 327 def binop_reduce(name, output_size, output_type, src_type, prereduce_expr,
 328                  reduce_expr, final_expr):
 329    def final(src):
 330       return final_expr.format(src= "(" + src + ")")
 331    def reduce_(src0, src1):
 332       return reduce_expr.format(src0=src0, src1=src1)
 333    def prereduce(src0, src1):
 334       return "(" + prereduce_expr.format(src0=src0, src1=src1) + ")"
 335    src0 = prereduce("src0.x", "src1.x")
 336    src1 = prereduce("src0.y", "src1.y")
 337    src2 = prereduce("src0.z", "src1.z")
 338    src3 = prereduce("src0.w", "src1.w")
 339    opcode(name + "2", output_size, output_type,
 340           [2, 2], [src_type, src_type], commutative,
 341           final(reduce_(src0, src1)))
 342    opcode(name + "3", output_size, output_type,
 343           [3, 3], [src_type, src_type], commutative,
 344           final(reduce_(reduce_(src0, src1), src2)))
 345    opcode(name + "4", output_size, output_type,
 346           [4, 4], [src_type, src_type], commutative,
 347           final(reduce_(reduce_(src0, src1), reduce_(src2, src3))))
 348
 349 binop("fadd", tfloat, commutative + associative, "src0 + src1")
 350 binop("iadd", tint, commutative + associative, "src0 + src1")
 351 binop("fsub", tfloat, "", "src0 - src1")
 352 binop("isub", tint, "", "src0 - src1")
 353
 354 binop("fmul", tfloat, commutative + associative, "src0 * src1")
 355 # low 32-bits of signed/unsigned integer multiply
 356 binop("imul", tint, commutative + associative, "src0 * src1")
 357 # high 32-bits of signed integer multiply
 358 binop("imul_high", tint, commutative,
 359       "(int32_t)(((int64_t) src0 * (int64_t) src1) >> 32)")
 360 # high 32-bits of unsigned integer multiply
 361 binop("umul_high", tunsigned, commutative,
 362       "(uint32_t)(((uint64_t) src0 * (uint64_t) src1) >> 32)")
 363
 364 binop("fdiv", tfloat, "", "src0 / src1")
 365 binop("idiv", tint, "", "src0 / src1")
 366 binop("udiv", tunsigned, "", "src0 / src1")
 367
 368 # returns a boolean representing the carry resulting from the addition of
 369 # the two unsigned arguments.
 370
 371 binop_convert("uadd_carry", tbool, tunsigned, commutative, "src0 + src1 < src0")
 372
 373 # returns a boolean representing the borrow resulting from the subtraction
 374 # of the two unsigned arguments.
 375
 376 binop_convert("usub_borrow", tbool, tunsigned, "", "src1 < src0")
 377
 378 binop("fmod", tfloat, "", "src0 - src1 * floorf(src0 / src1)")
 379 binop("umod", tunsigned, "", "src1 == 0 ? 0 : src0 % src1")
 380
 381 #
 382 # Comparisons
 383 #
 384
 385
 386 # these integer-aware comparisons return a boolean (0 or ~0)
 387
 388 binop_compare("flt", tfloat, "", "src0 < src1")
 389 binop_compare("fge", tfloat, "", "src0 >= src1")
 390 binop_compare("feq", tfloat, commutative, "src0 == src1")
 391 binop_compare("fne", tfloat, commutative, "src0 != src1")
 392 binop_compare("ilt", tint, "", "src0 < src1")
 393 binop_compare("ige", tint, "", "src0 >= src1")
 394 binop_compare("ieq", tint, commutative, "src0 == src1")
 395 binop_compare("ine", tint, commutative, "src0 != src1")
 396 binop_compare("ult", tunsigned, "", "src0 < src1")
 397 binop_compare("uge", tunsigned, "", "src0 >= src1")
 398
 399 # integer-aware GLSL-style comparisons that compare floats and ints
 400
 401 binop_reduce("ball_fequal",  1, tbool, tfloat, "{src0} == {src1}",
 402              "{src0} && {src1}", "{src}")
 403 binop_reduce("bany_fnequal", 1, tbool, tfloat, "{src0} != {src1}",
 404              "{src0} || {src1}", "{src}")
 405 binop_reduce("ball_iequal",  1, tbool, tint, "{src0} == {src1}",
 406              "{src0} && {src1}", "{src}")
 407 binop_reduce("bany_inequal", 1, tbool, tint, "{src0} != {src1}",
 408              "{src0} || {src1}", "{src}")
 409
 410 # non-integer-aware GLSL-style comparisons that return 0.0 or 1.0
 411
 412 binop_reduce("fall_equal",  1, tfloat, tfloat, "{src0} == {src1}",
 413              "{src0} && {src1}", "{src} ? 1.0f : 0.0f")
 414 binop_reduce("fany_nequal", 1, tfloat, tfloat, "{src0} != {src1}",
 415              "{src0} || {src1}", "{src} ? 1.0f : 0.0f")
 416
 417 # These comparisons for integer-less hardware return 1.0 and 0.0 for true
 418 # and false respectively
 419
 420 binop("slt", tfloat, "", "(src0 < src1) ? 1.0f : 0.0f") # Set on Less Than
 421 binop("sge", tfloat, "", "(src0 >= src1) ? 1.0f : 0.0f") # Set on Greater or Equal
 422 binop("seq", tfloat, commutative, "(src0 == src1) ? 1.0f : 0.0f") # Set on Equal
 423 binop("sne", tfloat, commutative, "(src0 != src1) ? 1.0f : 0.0f") # Set on Not Equal
 424
 425
 426 binop("ishl", tint, "", "src0 << src1")
 427 binop("ishr", tint, "", "src0 >> src1")
 428 binop("ushr", tunsigned, "", "src0 >> src1")
 429
 430 # bitwise logic operators
 431 #
 432 # These are also used as boolean and, or, xor for hardware supporting
 433 # integers.
 434
 435
 436 binop("iand", tunsigned, commutative + associative, "src0 & src1")
 437 binop("ior", tunsigned, commutative + associative, "src0 | src1")
 438 binop("ixor", tunsigned, commutative + associative, "src0 ^ src1")
 439
 440
 441 # floating point logic operators
 442 #
 443 # These use (src != 0.0) for testing the truth of the input, and output 1.0
 444 # for true and 0.0 for false
 445
 446 binop("fand", tfloat, commutative,
 447       "((src0 != 0.0f) && (src1 != 0.0f)) ? 1.0f : 0.0f")
 448 binop("for", tfloat, commutative,
 449       "((src0 != 0.0f) || (src1 != 0.0f)) ? 1.0f : 0.0f")
 450 binop("fxor", tfloat, commutative,
 451       "(src0 != 0.0f && src1 == 0.0f) || (src0 == 0.0f && src1 != 0.0f) ? 1.0f : 0.0f")
 452
 453 binop_reduce("fdot", 1, tfloat, tfloat, "{src0} * {src1}", "{src0} + {src1}",
 454              "{src}")
 455
 456 binop_reduce("fdot_replicated", 4, tfloat, tfloat,
 457              "{src0} * {src1}", "{src0} + {src1}", "{src}")
 458
 459 opcode("fdph", 1, tfloat, [3, 4], [tfloat, tfloat], "",
 460        "src0.x * src1.x + src0.y * src1.y + src0.z * src1.z + src1.w")
 461 opcode("fdph_replicated", 4, tfloat, [3, 4], [tfloat, tfloat], "",
 462        "src0.x * src1.x + src0.y * src1.y + src0.z * src1.z + src1.w")
 463
 464 binop("fmin", tfloat, "", "fminf(src0, src1)")
 465 binop("imin", tint, commutative + associative, "src1 > src0 ? src0 : src1")
 466 binop("umin", tunsigned, commutative + associative, "src1 > src0 ? src0 : src1")
 467 binop("fmax", tfloat, "", "fmaxf(src0, src1)")
 468 binop("imax", tint, commutative + associative, "src1 > src0 ? src1 : src0")
 469 binop("umax", tunsigned, commutative + associative, "src1 > src0 ? src1 : src0")
 470
 471 # Saturated vector add for 4 8bit ints.
 472 binop("usadd_4x8", tint, commutative + associative, """
 473 dst = 0;
 474 for (int i = 0; i < 32; i += 8) {
 475    dst |= MIN2(((src0 >> i) & 0xff) + ((src1 >> i) & 0xff), 0xff) << i;
 476 }
 477 """)
 478
 479 # Saturated vector subtract for 4 8bit ints.
 480 binop("ussub_4x8", tint, "", """
 481 dst = 0;
 482 for (int i = 0; i < 32; i += 8) {
 483    int src0_chan = (src0 >> i) & 0xff;
 484    int src1_chan = (src1 >> i) & 0xff;
 485    if (src0_chan > src1_chan)
 486       dst |= (src0_chan - src1_chan) << i;
 487 }
 488 """)
 489
 490 # vector min for 4 8bit ints.
 491 binop("umin_4x8", tint, commutative + associative, """
 492 dst = 0;
 493 for (int i = 0; i < 32; i += 8) {
 494    dst |= MIN2((src0 >> i) & 0xff, (src1 >> i) & 0xff) << i;
 495 }
 496 """)
 497
 498 # vector max for 4 8bit ints.
 499 binop("umax_4x8", tint, commutative + associative, """
 500 dst = 0;
 501 for (int i = 0; i < 32; i += 8) {
 502    dst |= MAX2((src0 >> i) & 0xff, (src1 >> i) & 0xff) << i;
 503 }
 504 """)
 505
 506 # unorm multiply: (a * b) / 255.
 507 binop("umul_unorm_4x8", tint, commutative + associative, """
 508 dst = 0;
 509 for (int i = 0; i < 32; i += 8) {
 510    int src0_chan = (src0 >> i) & 0xff;
 511    int src1_chan = (src1 >> i) & 0xff;
 512    dst |= ((src0_chan * src1_chan) / 255) << i;
 513 }
 514 """)
 515
 516 binop("fpow", tfloat, "", "powf(src0, src1)")
 517
 518 binop_horiz("pack_half_2x16_split", 1, tunsigned, 1, tfloat, 1, tfloat,
 519             "pack_half_1x16(src0.x) | (pack_half_1x16(src1.x) << 16)")
 520
 521 binop_convert("bfm", tunsigned, tint, "", """
 522 int offset = src0, bits = src1;
 523 if (offset < 0 || bits < 0 || offset + bits > 32)
 524    dst = 0; /* undefined per the spec */
 525 else
 526    dst = ((1 << bits)- 1) << offset;
 527 """)
 528
 529 opcode("ldexp", 0, tfloat, [0, 0], [tfloat, tint], "", """
 530 dst = ldexpf(src0, src1);
 531 /* flush denormals to zero. */
 532 if (!isnormal(dst))
 533    dst = copysignf(0.0f, src0);
 534 """)
 535
 536 # Combines the first component of each input to make a 2-component vector.
 537
 538 binop_horiz("vec2", 2, tunsigned, 1, tunsigned, 1, tunsigned, """
 539 dst.x = src0.x;
 540 dst.y = src1.x;
 541 """)
 542
 543 def triop(name, ty, const_expr):
 544    opcode(name, 0, ty, [0, 0, 0], [ty, ty, ty], "", const_expr)
 545 def triop_horiz(name, output_size, src1_size, src2_size, src3_size, const_expr):
 546    opcode(name, output_size, tunsigned,
 547    [src1_size, src2_size, src3_size],
 548    [tunsigned, tunsigned, tunsigned], "", const_expr)
 549
 550 triop("ffma", tfloat, "src0 * src1 + src2")
 551
 552 triop("flrp", tfloat, "src0 * (1 - src2) + src1 * src2")
 553
 554 # Conditional Select
 555 #
 556 # A vector conditional select instruction (like ?:, but operating per-
 557 # component on vectors). There are two versions, one for floating point
 558 # bools (0.0 vs 1.0) and one for integer bools (0 vs ~0).
 559
 560
 561 triop("fcsel", tfloat, "(src0 != 0.0f) ? src1 : src2")
 562 opcode("bcsel", 0, tunsigned, [0, 0, 0],
 563       [tbool, tunsigned, tunsigned], "", "src0 ? src1 : src2")
 564
 565 triop("bfi", tunsigned, """
 566 unsigned mask = src0, insert = src1, base = src2;
 567 if (mask == 0) {
 568    dst = base;
 569 } else {
 570    unsigned tmp = mask;
 571    while (!(tmp & 1)) {
 572       tmp >>= 1;
 573       insert <<= 1;
 574    }
 575    dst = (base & ~mask) | (insert & mask);
 576 }
 577 """)
 578
 579 opcode("ubitfield_extract", 0, tunsigned,
 580        [0, 1, 1], [tunsigned, tint, tint], "", """
 581 unsigned base = src0;
 582 int offset = src1.x, bits = src2.x;
 583 if (bits == 0) {
 584    dst = 0;
 585 } else if (bits < 0 || offset < 0 || offset + bits > 32) {
 586    dst = 0; /* undefined per the spec */
 587 } else {
 588    dst = (base >> offset) & ((1 << bits) - 1);
 589 }
 590 """)
 591 opcode("ibitfield_extract", 0, tint,
 592        [0, 1, 1], [tint, tint, tint], "", """
 593 int base = src0;
 594 int offset = src1.x, bits = src2.x;
 595 if (bits == 0) {
 596    dst = 0;
 597 } else if (offset < 0 || bits < 0 || offset + bits > 32) {
 598    dst = 0;
 599 } else {
 600    dst = (base << (32 - offset - bits)) >> offset; /* use sign-extending shift */
 601 }
 602 """)
 603
 604 # Combines the first component of each input to make a 3-component vector.
 605
 606 triop_horiz("vec3", 3, 1, 1, 1, """
 607 dst.x = src0.x;
 608 dst.y = src1.x;
 609 dst.z = src2.x;
 610 """)
 611
 612 def quadop_horiz(name, output_size, src1_size, src2_size, src3_size,
 613                  src4_size, const_expr):
 614    opcode(name, output_size, tunsigned,
 615           [src1_size, src2_size, src3_size, src4_size],
 616           [tunsigned, tunsigned, tunsigned, tunsigned],
 617           "", const_expr)
 618
 619 opcode("bitfield_insert", 0, tunsigned, [0, 0, 1, 1],
 620        [tunsigned, tunsigned, tint, tint], "", """
 621 unsigned base = src0, insert = src1;
 622 int offset = src2.x, bits = src3.x;
 623 if (bits == 0) {
 624    dst = 0;
 625 } else if (offset < 0 || bits < 0 || bits + offset > 32) {
 626    dst = 0;
 627 } else {
 628    unsigned mask = ((1 << bits) - 1) << offset;
 629    dst = (base & ~mask) | ((insert << bits) & mask);
 630 }
 631 """)
 632
 633 quadop_horiz("vec4", 4, 1, 1, 1, 1, """
 634 dst.x = src0.x;
 635 dst.y = src1.x;
 636 dst.z = src2.x;
 637 dst.w = src3.x;
 638 """)
 639
 640