src/compiler/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 = "bool32"
  94 tuint = "uint"
  95 tfloat32 = "float32"
  96 tint32 = "int32"
  97 tuint32 = "uint32"
  98 tuint64 = "uint64"
  99 tfloat64 = "float64"
 100
 101 commutative = "commutative "
 102 associative = "associative "
 103
 104 # global dictionary of opcodes
 105 opcodes = {}
 106
 107 def opcode(name, output_size, output_type, input_sizes, input_types,
 108            algebraic_properties, const_expr):
 109    assert name not in opcodes
 110    opcodes[name] = Opcode(name, output_size, output_type, input_sizes,
 111                           input_types, algebraic_properties, const_expr)
 112
 113 def unop_convert(name, out_type, in_type, const_expr):
 114    opcode(name, 0, out_type, [0], [in_type], "", const_expr)
 115
 116 def unop(name, ty, const_expr):
 117    opcode(name, 0, ty, [0], [ty], "", const_expr)
 118
 119 def unop_horiz(name, output_size, output_type, input_size, input_type,
 120                const_expr):
 121    opcode(name, output_size, output_type, [input_size], [input_type], "",
 122           const_expr)
 123
 124 def unop_reduce(name, output_size, output_type, input_type, prereduce_expr,
 125                 reduce_expr, final_expr):
 126    def prereduce(src):
 127       return "(" + prereduce_expr.format(src=src) + ")"
 128    def final(src):
 129       return final_expr.format(src="(" + src + ")")
 130    def reduce_(src0, src1):
 131       return reduce_expr.format(src0=src0, src1=src1)
 132    src0 = prereduce("src0.x")
 133    src1 = prereduce("src0.y")
 134    src2 = prereduce("src0.z")
 135    src3 = prereduce("src0.w")
 136    unop_horiz(name + "2", output_size, output_type, 2, input_type,
 137               final(reduce_(src0, src1)))
 138    unop_horiz(name + "3", output_size, output_type, 3, input_type,
 139               final(reduce_(reduce_(src0, src1), src2)))
 140    unop_horiz(name + "4", output_size, output_type, 4, input_type,
 141               final(reduce_(reduce_(src0, src1), reduce_(src2, src3))))
 142
 143
 144 # These two move instructions differ in what modifiers they support and what
 145 # the negate modifier means. Otherwise, they are identical.
 146 unop("fmov", tfloat, "src0")
 147 unop("imov", tint, "src0")
 148
 149 unop("ineg", tint, "-src0")
 150 unop("fneg", tfloat, "-src0")
 151 unop("inot", tint, "~src0") # invert every bit of the integer
 152 unop("fnot", tfloat, "(src0 == 0.0f) ? 1.0f : 0.0f")
 153 unop("fsign", tfloat, "(src0 == 0.0f) ? 0.0f : ((src0 > 0.0f) ? 1.0f : -1.0f)")
 154 unop("isign", tint, "(src0 == 0) ? 0 : ((src0 > 0) ? 1 : -1)")
 155 unop("iabs", tint, "(src0 < 0) ? -src0 : src0")
 156 unop("fabs", tfloat, "fabsf(src0)")
 157 unop("fsat", tfloat, "(src0 > 1.0f) ? 1.0f : ((src0 <= 0.0f) ? 0.0f : src0)")
 158 unop("frcp", tfloat, "1.0f / src0")
 159 unop("frsq", tfloat, "1.0f / sqrtf(src0)")
 160 unop("fsqrt", tfloat, "sqrtf(src0)")
 161 unop("fexp2", tfloat, "exp2f(src0)")
 162 unop("flog2", tfloat, "log2f(src0)")
 163 unop_convert("f2i", tint32, tfloat32, "src0") # Float-to-integer conversion.
 164 unop_convert("f2u", tuint32, tfloat32, "src0") # Float-to-unsigned conversion
 165 unop_convert("d2i", tint32, tfloat64, "src0") # Double-to-integer conversion.
 166 unop_convert("d2u", tuint32, tfloat64, "src0") # Double-to-unsigned conversion.
 167 unop_convert("i2f", tfloat32, tint32, "src0") # Integer-to-float conversion.
 168 unop_convert("i2d", tfloat64, tint32, "src0") # Integer-to-double conversion.
 169 # Float-to-boolean conversion
 170 unop_convert("f2b", tbool, tfloat32, "src0 != 0.0f")
 171 unop_convert("d2b", tbool, tfloat64, "src0 != 0.0")
 172 # Boolean-to-float conversion
 173 unop_convert("b2f", tfloat32, tbool, "src0 ? 1.0f : 0.0f")
 174 # Int-to-boolean conversion
 175 unop_convert("i2b", tbool, tint32, "src0 != 0")
 176 unop_convert("b2i", tint32, tbool, "src0 ? 1 : 0") # Boolean-to-int conversion
 177 unop_convert("u2f", tfloat32, tuint32, "src0") # Unsigned-to-float conversion.
 178 unop_convert("u2d", tfloat64, tuint32, "src0") # Unsigned-to-double conversion.
 179 # double-to-float conversion
 180 unop_convert("d2f", tfloat32, tfloat64, "src0") # Single to double precision
 181 unop_convert("f2d", tfloat64, tfloat32, "src0") # Double to single precision
 182
 183 # Unary floating-point rounding operations.
 184
 185
 186 unop("ftrunc", tfloat, "bit_size == 64 ? trunc(src0) : truncf(src0)")
 187 unop("fceil", tfloat, "bit_size == 64 ? ceil(src0) : ceilf(src0)")
 188 unop("ffloor", tfloat, "bit_size == 64 ? floor(src0) : floorf(src0)")
 189 unop("ffract", tfloat, "src0 - (bit_size == 64 ? floor(src0) : floorf(src0))")
 190 unop("fround_even", tfloat, "bit_size == 64 ? _mesa_roundeven(src0) : _mesa_roundevenf(src0)")
 191
 192 unop("fquantize2f16", tfloat, "(fabs(src0) < ldexpf(1.0, -14)) ? copysignf(0.0f, src0) : _mesa_half_to_float(_mesa_float_to_half(src0))")
 193
 194 # Trigonometric operations.
 195
 196
 197 unop("fsin", tfloat, "bit_size == 64 ? sin(src0) : sinf(src0)")
 198 unop("fcos", tfloat, "bit_size == 64 ? cos(src0) : cosf(src0)")
 199
 200
 201 # Partial derivatives.
 202
 203
 204 unop("fddx", tfloat, "0.0") # the derivative of a constant is 0.
 205 unop("fddy", tfloat, "0.0")
 206 unop("fddx_fine", tfloat, "0.0")
 207 unop("fddy_fine", tfloat, "0.0")
 208 unop("fddx_coarse", tfloat, "0.0")
 209 unop("fddy_coarse", tfloat, "0.0")
 210
 211
 212 # Floating point pack and unpack operations.
 213
 214 def pack_2x16(fmt):
 215    unop_horiz("pack_" + fmt + "_2x16", 1, tuint32, 2, tfloat32, """
 216 dst.x = (uint32_t) pack_fmt_1x16(src0.x);
 217 dst.x |= ((uint32_t) pack_fmt_1x16(src0.y)) << 16;
 218 """.replace("fmt", fmt))
 219
 220 def pack_4x8(fmt):
 221    unop_horiz("pack_" + fmt + "_4x8", 1, tuint32, 4, tfloat32, """
 222 dst.x = (uint32_t) pack_fmt_1x8(src0.x);
 223 dst.x |= ((uint32_t) pack_fmt_1x8(src0.y)) << 8;
 224 dst.x |= ((uint32_t) pack_fmt_1x8(src0.z)) << 16;
 225 dst.x |= ((uint32_t) pack_fmt_1x8(src0.w)) << 24;
 226 """.replace("fmt", fmt))
 227
 228 def unpack_2x16(fmt):
 229    unop_horiz("unpack_" + fmt + "_2x16", 2, tfloat32, 1, tuint32, """
 230 dst.x = unpack_fmt_1x16((uint16_t)(src0.x & 0xffff));
 231 dst.y = unpack_fmt_1x16((uint16_t)(src0.x << 16));
 232 """.replace("fmt", fmt))
 233
 234 def unpack_4x8(fmt):
 235    unop_horiz("unpack_" + fmt + "_4x8", 4, tfloat32, 1, tuint32, """
 236 dst.x = unpack_fmt_1x8((uint8_t)(src0.x & 0xff));
 237 dst.y = unpack_fmt_1x8((uint8_t)((src0.x >> 8) & 0xff));
 238 dst.z = unpack_fmt_1x8((uint8_t)((src0.x >> 16) & 0xff));
 239 dst.w = unpack_fmt_1x8((uint8_t)(src0.x >> 24));
 240 """.replace("fmt", fmt))
 241
 242
 243 pack_2x16("snorm")
 244 pack_4x8("snorm")
 245 pack_2x16("unorm")
 246 pack_4x8("unorm")
 247 pack_2x16("half")
 248 unpack_2x16("snorm")
 249 unpack_4x8("snorm")
 250 unpack_2x16("unorm")
 251 unpack_4x8("unorm")
 252 unpack_2x16("half")
 253
 254 unop_horiz("pack_uvec2_to_uint", 1, tuint32, 2, tuint32, """
 255 dst.x = (src0.x & 0xffff) | (src0.y >> 16);
 256 """)
 257
 258 unop_horiz("pack_uvec4_to_uint", 1, tuint32, 4, tuint32, """
 259 dst.x = (src0.x <<  0) |
 260         (src0.y <<  8) |
 261         (src0.z << 16) |
 262         (src0.w << 24);
 263 """)
 264
 265 unop_horiz("pack_double_2x32", 1, tuint64, 2, tuint32, """
 266 union {
 267     uint64_t u64;
 268     struct {
 269         uint32_t i1;
 270         uint32_t i2;
 271     };
 272 } di;
 273
 274 di.i1 = src0.x;
 275 di.i2 = src0.y;
 276 dst.x = di.u64;
 277 """)
 278
 279 unop_horiz("unpack_double_2x32", 2, tuint32, 1, tuint64, """
 280 union {
 281     uint64_t u64;
 282     struct {
 283         uint32_t i1;
 284         uint32_t i2;
 285     };
 286 } di;
 287
 288 di.u64 = src0.x;
 289 dst.x = di.i1;
 290 dst.y = di.i2;
 291 """)
 292
 293 # Lowered floating point unpacking operations.
 294
 295
 296 unop_horiz("unpack_half_2x16_split_x", 1, tfloat32, 1, tuint32,
 297            "unpack_half_1x16((uint16_t)(src0.x & 0xffff))")
 298 unop_horiz("unpack_half_2x16_split_y", 1, tfloat32, 1, tuint32,
 299            "unpack_half_1x16((uint16_t)(src0.x >> 16))")
 300
 301 unop_convert("unpack_double_2x32_split_x", tuint32, tuint64, """
 302 union {
 303     uint64_t u64;
 304     struct {
 305         uint32_t x;
 306         uint32_t y;
 307     };
 308 } di;
 309 di.u64 = src0;
 310 dst = di.x;
 311 """)
 312
 313 unop_convert("unpack_double_2x32_split_y", tuint32, tuint64, """
 314 union {
 315     uint64_t u64;
 316     struct {
 317         uint32_t x;
 318         uint32_t y;
 319     };
 320 } di;
 321 di.u64 = src0;
 322 dst = di.y;
 323 """)
 324
 325 # Bit operations, part of ARB_gpu_shader5.
 326
 327
 328 unop("bitfield_reverse", tuint32, """
 329 /* we're not winning any awards for speed here, but that's ok */
 330 dst = 0;
 331 for (unsigned bit = 0; bit < 32; bit++)
 332    dst |= ((src0 >> bit) & 1) << (31 - bit);
 333 """)
 334 unop("bit_count", tuint32, """
 335 dst = 0;
 336 for (unsigned bit = 0; bit < 32; bit++) {
 337    if ((src0 >> bit) & 1)
 338       dst++;
 339 }
 340 """)
 341
 342 unop_convert("ufind_msb", tint32, tuint32, """
 343 dst = -1;
 344 for (int bit = 31; bit > 0; bit--) {
 345    if ((src0 >> bit) & 1) {
 346       dst = bit;
 347       break;
 348    }
 349 }
 350 """)
 351
 352 unop("ifind_msb", tint32, """
 353 dst = -1;
 354 for (int bit = 31; bit >= 0; bit--) {
 355    /* If src0 < 0, we're looking for the first 0 bit.
 356     * if src0 >= 0, we're looking for the first 1 bit.
 357     */
 358    if ((((src0 >> bit) & 1) && (src0 >= 0)) ||
 359       (!((src0 >> bit) & 1) && (src0 < 0))) {
 360       dst = bit;
 361       break;
 362    }
 363 }
 364 """)
 365
 366 unop("find_lsb", tint32, """
 367 dst = -1;
 368 for (unsigned bit = 0; bit < 32; bit++) {
 369    if ((src0 >> bit) & 1) {
 370       dst = bit;
 371       break;
 372    }
 373 }
 374 """)
 375
 376
 377 for i in xrange(1, 5):
 378    for j in xrange(1, 5):
 379       unop_horiz("fnoise{0}_{1}".format(i, j), i, tfloat, j, tfloat, "0.0f")
 380
 381 def binop_convert(name, out_type, in_type, alg_props, const_expr):
 382    opcode(name, 0, out_type, [0, 0], [in_type, in_type], alg_props, const_expr)
 383
 384 def binop(name, ty, alg_props, const_expr):
 385    binop_convert(name, ty, ty, alg_props, const_expr)
 386
 387 def binop_compare(name, ty, alg_props, const_expr):
 388    binop_convert(name, tbool, ty, alg_props, const_expr)
 389
 390 def binop_horiz(name, out_size, out_type, src1_size, src1_type, src2_size,
 391                 src2_type, const_expr):
 392    opcode(name, out_size, out_type, [src1_size, src2_size], [src1_type, src2_type],
 393           "", const_expr)
 394
 395 def binop_reduce(name, output_size, output_type, src_type, prereduce_expr,
 396                  reduce_expr, final_expr):
 397    def final(src):
 398       return final_expr.format(src= "(" + src + ")")
 399    def reduce_(src0, src1):
 400       return reduce_expr.format(src0=src0, src1=src1)
 401    def prereduce(src0, src1):
 402       return "(" + prereduce_expr.format(src0=src0, src1=src1) + ")"
 403    src0 = prereduce("src0.x", "src1.x")
 404    src1 = prereduce("src0.y", "src1.y")
 405    src2 = prereduce("src0.z", "src1.z")
 406    src3 = prereduce("src0.w", "src1.w")
 407    opcode(name + "2", output_size, output_type,
 408           [2, 2], [src_type, src_type], commutative,
 409           final(reduce_(src0, src1)))
 410    opcode(name + "3", output_size, output_type,
 411           [3, 3], [src_type, src_type], commutative,
 412           final(reduce_(reduce_(src0, src1), src2)))
 413    opcode(name + "4", output_size, output_type,
 414           [4, 4], [src_type, src_type], commutative,
 415           final(reduce_(reduce_(src0, src1), reduce_(src2, src3))))
 416
 417 binop("fadd", tfloat, commutative + associative, "src0 + src1")
 418 binop("iadd", tint, commutative + associative, "src0 + src1")
 419 binop("fsub", tfloat, "", "src0 - src1")
 420 binop("isub", tint, "", "src0 - src1")
 421
 422 binop("fmul", tfloat, commutative + associative, "src0 * src1")
 423 # low 32-bits of signed/unsigned integer multiply
 424 binop("imul", tint, commutative + associative, "src0 * src1")
 425 # high 32-bits of signed integer multiply
 426 binop("imul_high", tint32, commutative,
 427       "(int32_t)(((int64_t) src0 * (int64_t) src1) >> 32)")
 428 # high 32-bits of unsigned integer multiply
 429 binop("umul_high", tuint32, commutative,
 430       "(uint32_t)(((uint64_t) src0 * (uint64_t) src1) >> 32)")
 431
 432 binop("fdiv", tfloat, "", "src0 / src1")
 433 binop("idiv", tint, "", "src0 / src1")
 434 binop("udiv", tuint, "", "src0 / src1")
 435
 436 # returns a boolean representing the carry resulting from the addition of
 437 # the two unsigned arguments.
 438
 439 binop_convert("uadd_carry", tuint, tuint, commutative, "src0 + src1 < src0")
 440
 441 # returns a boolean representing the borrow resulting from the subtraction
 442 # of the two unsigned arguments.
 443
 444 binop_convert("usub_borrow", tuint, tuint, "", "src0 < src1")
 445
 446 binop("fmod", tfloat, "", "src0 - src1 * floorf(src0 / src1)")
 447 binop("umod", tuint, "", "src1 == 0 ? 0 : src0 % src1")
 448
 449 #
 450 # Comparisons
 451 #
 452
 453
 454 # these integer-aware comparisons return a boolean (0 or ~0)
 455
 456 binop_compare("flt", tfloat, "", "src0 < src1")
 457 binop_compare("fge", tfloat, "", "src0 >= src1")
 458 binop_compare("feq", tfloat, commutative, "src0 == src1")
 459 binop_compare("fne", tfloat, commutative, "src0 != src1")
 460 binop_compare("ilt", tint, "", "src0 < src1")
 461 binop_compare("ige", tint, "", "src0 >= src1")
 462 binop_compare("ieq", tint, commutative, "src0 == src1")
 463 binop_compare("ine", tint, commutative, "src0 != src1")
 464 binop_compare("ult", tuint, "", "src0 < src1")
 465 binop_compare("uge", tuint, "", "src0 >= src1")
 466
 467 # integer-aware GLSL-style comparisons that compare floats and ints
 468
 469 binop_reduce("ball_fequal",  1, tbool, tfloat, "{src0} == {src1}",
 470              "{src0} && {src1}", "{src}")
 471 binop_reduce("bany_fnequal", 1, tbool, tfloat, "{src0} != {src1}",
 472              "{src0} || {src1}", "{src}")
 473 binop_reduce("ball_iequal",  1, tbool, tint, "{src0} == {src1}",
 474              "{src0} && {src1}", "{src}")
 475 binop_reduce("bany_inequal", 1, tbool, tint, "{src0} != {src1}",
 476              "{src0} || {src1}", "{src}")
 477
 478 # non-integer-aware GLSL-style comparisons that return 0.0 or 1.0
 479
 480 binop_reduce("fall_equal",  1, tfloat32, tfloat32, "{src0} == {src1}",
 481              "{src0} && {src1}", "{src} ? 1.0f : 0.0f")
 482 binop_reduce("fany_nequal", 1, tfloat32, tfloat32, "{src0} != {src1}",
 483              "{src0} || {src1}", "{src} ? 1.0f : 0.0f")
 484
 485 # These comparisons for integer-less hardware return 1.0 and 0.0 for true
 486 # and false respectively
 487
 488 binop("slt", tfloat32, "", "(src0 < src1) ? 1.0f : 0.0f") # Set on Less Than
 489 binop("sge", tfloat32, "", "(src0 >= src1) ? 1.0f : 0.0f") # Set on Greater or Equal
 490 binop("seq", tfloat32, commutative, "(src0 == src1) ? 1.0f : 0.0f") # Set on Equal
 491 binop("sne", tfloat32, commutative, "(src0 != src1) ? 1.0f : 0.0f") # Set on Not Equal
 492
 493
 494 binop("ishl", tint, "", "src0 << src1")
 495 binop("ishr", tint, "", "src0 >> src1")
 496 binop("ushr", tuint, "", "src0 >> src1")
 497
 498 # bitwise logic operators
 499 #
 500 # These are also used as boolean and, or, xor for hardware supporting
 501 # integers.
 502
 503
 504 binop("iand", tuint, commutative + associative, "src0 & src1")
 505 binop("ior", tuint, commutative + associative, "src0 | src1")
 506 binop("ixor", tuint, commutative + associative, "src0 ^ src1")
 507
 508
 509 # floating point logic operators
 510 #
 511 # These use (src != 0.0) for testing the truth of the input, and output 1.0
 512 # for true and 0.0 for false
 513
 514 binop("fand", tfloat32, commutative,
 515       "((src0 != 0.0f) && (src1 != 0.0f)) ? 1.0f : 0.0f")
 516 binop("for", tfloat32, commutative,
 517       "((src0 != 0.0f) || (src1 != 0.0f)) ? 1.0f : 0.0f")
 518 binop("fxor", tfloat32, commutative,
 519       "(src0 != 0.0f && src1 == 0.0f) || (src0 == 0.0f && src1 != 0.0f) ? 1.0f : 0.0f")
 520
 521 binop_reduce("fdot", 1, tfloat, tfloat, "{src0} * {src1}", "{src0} + {src1}",
 522              "{src}")
 523
 524 binop_reduce("fdot_replicated", 4, tfloat, tfloat,
 525              "{src0} * {src1}", "{src0} + {src1}", "{src}")
 526
 527 opcode("fdph", 1, tfloat, [3, 4], [tfloat, tfloat], "",
 528        "src0.x * src1.x + src0.y * src1.y + src0.z * src1.z + src1.w")
 529 opcode("fdph_replicated", 4, tfloat, [3, 4], [tfloat, tfloat], "",
 530        "src0.x * src1.x + src0.y * src1.y + src0.z * src1.z + src1.w")
 531
 532 binop("fmin", tfloat, "", "fminf(src0, src1)")
 533 binop("imin", tint, commutative + associative, "src1 > src0 ? src0 : src1")
 534 binop("umin", tuint, commutative + associative, "src1 > src0 ? src0 : src1")
 535 binop("fmax", tfloat, "", "fmaxf(src0, src1)")
 536 binop("imax", tint, commutative + associative, "src1 > src0 ? src1 : src0")
 537 binop("umax", tuint, commutative + associative, "src1 > src0 ? src1 : src0")
 538
 539 # Saturated vector add for 4 8bit ints.
 540 binop("usadd_4x8", tint32, commutative + associative, """
 541 dst = 0;
 542 for (int i = 0; i < 32; i += 8) {
 543    dst |= MIN2(((src0 >> i) & 0xff) + ((src1 >> i) & 0xff), 0xff) << i;
 544 }
 545 """)
 546
 547 # Saturated vector subtract for 4 8bit ints.
 548 binop("ussub_4x8", tint32, "", """
 549 dst = 0;
 550 for (int i = 0; i < 32; i += 8) {
 551    int src0_chan = (src0 >> i) & 0xff;
 552    int src1_chan = (src1 >> i) & 0xff;
 553    if (src0_chan > src1_chan)
 554       dst |= (src0_chan - src1_chan) << i;
 555 }
 556 """)
 557
 558 # vector min for 4 8bit ints.
 559 binop("umin_4x8", tint32, commutative + associative, """
 560 dst = 0;
 561 for (int i = 0; i < 32; i += 8) {
 562    dst |= MIN2((src0 >> i) & 0xff, (src1 >> i) & 0xff) << i;
 563 }
 564 """)
 565
 566 # vector max for 4 8bit ints.
 567 binop("umax_4x8", tint32, commutative + associative, """
 568 dst = 0;
 569 for (int i = 0; i < 32; i += 8) {
 570    dst |= MAX2((src0 >> i) & 0xff, (src1 >> i) & 0xff) << i;
 571 }
 572 """)
 573
 574 # unorm multiply: (a * b) / 255.
 575 binop("umul_unorm_4x8", tint32, commutative + associative, """
 576 dst = 0;
 577 for (int i = 0; i < 32; i += 8) {
 578    int src0_chan = (src0 >> i) & 0xff;
 579    int src1_chan = (src1 >> i) & 0xff;
 580    dst |= ((src0_chan * src1_chan) / 255) << i;
 581 }
 582 """)
 583
 584 binop("fpow", tfloat, "", "bit_size == 64 ? powf(src0, src1) : pow(src0, src1)")
 585
 586 binop_horiz("pack_half_2x16_split", 1, tuint32, 1, tfloat32, 1, tfloat32,
 587             "pack_half_1x16(src0.x) | (pack_half_1x16(src1.x) << 16)")
 588
 589 binop_convert("pack_double_2x32_split", tuint64, tuint32, "", """
 590 union {
 591     uint64_t u64;
 592     struct {
 593         uint32_t x;
 594         uint32_t y;
 595     };
 596 } di;
 597 di.x = src0;
 598 di.y = src1;
 599 dst = di.u64;
 600 """)
 601
 602 # bfm implements the behavior of the first operation of the SM5 "bfi" assembly
 603 # and that of the "bfi1" i965 instruction. That is, it has undefined behavior
 604 # if either of its arguments are 32.
 605 binop_convert("bfm", tuint32, tint32, "", """
 606 int bits = src0, offset = src1;
 607 if (offset < 0 || bits < 0 || offset > 31 || bits > 31 || offset + bits > 32)
 608    dst = 0; /* undefined */
 609 else
 610    dst = ((1u << bits) - 1) << offset;
 611 """)
 612
 613 opcode("ldexp", 0, tfloat, [0, 0], [tfloat, tint], "", """
 614 dst = (bit_size == 64) ? ldexp(src0, src1) : ldexpf(src0, src1);
 615 /* flush denormals to zero. */
 616 if (!isnormal(dst))
 617    dst = copysignf(0.0f, src0);
 618 """)
 619
 620 # Combines the first component of each input to make a 2-component vector.
 621
 622 binop_horiz("vec2", 2, tuint, 1, tuint, 1, tuint, """
 623 dst.x = src0.x;
 624 dst.y = src1.x;
 625 """)
 626
 627 # Byte extraction
 628 binop("extract_u8", tuint, "", "(uint8_t)(src0 >> (src1 * 8))")
 629 binop("extract_i8", tint, "", "(int8_t)(src0 >> (src1 * 8))")
 630
 631 # Word extraction
 632 binop("extract_u16", tuint, "", "(uint16_t)(src0 >> (src1 * 16))")
 633 binop("extract_i16", tint, "", "(int16_t)(src0 >> (src1 * 16))")
 634
 635
 636 def triop(name, ty, const_expr):
 637    opcode(name, 0, ty, [0, 0, 0], [ty, ty, ty], "", const_expr)
 638 def triop_horiz(name, output_size, src1_size, src2_size, src3_size, const_expr):
 639    opcode(name, output_size, tuint,
 640    [src1_size, src2_size, src3_size],
 641    [tuint, tuint, tuint], "", const_expr)
 642
 643 triop("ffma", tfloat, "src0 * src1 + src2")
 644
 645 triop("flrp", tfloat, "src0 * (1 - src2) + src1 * src2")
 646
 647 # Conditional Select
 648 #
 649 # A vector conditional select instruction (like ?:, but operating per-
 650 # component on vectors). There are two versions, one for floating point
 651 # bools (0.0 vs 1.0) and one for integer bools (0 vs ~0).
 652
 653
 654 triop("fcsel", tfloat32, "(src0 != 0.0f) ? src1 : src2")
 655 opcode("bcsel", 0, tuint, [0, 0, 0],
 656       [tbool, tuint, tuint], "", "src0 ? src1 : src2")
 657
 658 # SM5 bfi assembly
 659 triop("bfi", tuint32, """
 660 unsigned mask = src0, insert = src1, base = src2;
 661 if (mask == 0) {
 662    dst = base;
 663 } else {
 664    unsigned tmp = mask;
 665    while (!(tmp & 1)) {
 666       tmp >>= 1;
 667       insert <<= 1;
 668    }
 669    dst = (base & ~mask) | (insert & mask);
 670 }
 671 """)
 672
 673 # SM5 ubfe/ibfe assembly
 674 opcode("ubfe", 0, tuint32,
 675        [0, 0, 0], [tuint32, tint32, tint32], "", """
 676 unsigned base = src0;
 677 int offset = src1, bits = src2;
 678 if (bits == 0) {
 679    dst = 0;
 680 } else if (bits < 0 || offset < 0) {
 681    dst = 0; /* undefined */
 682 } else if (offset + bits < 32) {
 683    dst = (base << (32 - bits - offset)) >> (32 - bits);
 684 } else {
 685    dst = base >> offset;
 686 }
 687 """)
 688 opcode("ibfe", 0, tint32,
 689        [0, 0, 0], [tint32, tint32, tint32], "", """
 690 int base = src0;
 691 int offset = src1, bits = src2;
 692 if (bits == 0) {
 693    dst = 0;
 694 } else if (bits < 0 || offset < 0) {
 695    dst = 0; /* undefined */
 696 } else if (offset + bits < 32) {
 697    dst = (base << (32 - bits - offset)) >> (32 - bits);
 698 } else {
 699    dst = base >> offset;
 700 }
 701 """)
 702
 703 # GLSL bitfieldExtract()
 704 opcode("ubitfield_extract", 0, tuint32,
 705        [0, 0, 0], [tuint32, tint32, tint32], "", """
 706 unsigned base = src0;
 707 int offset = src1, bits = src2;
 708 if (bits == 0) {
 709    dst = 0;
 710 } else if (bits < 0 || offset < 0 || offset + bits > 32) {
 711    dst = 0; /* undefined per the spec */
 712 } else {
 713    dst = (base >> offset) & ((1ull << bits) - 1);
 714 }
 715 """)
 716 opcode("ibitfield_extract", 0, tint32,
 717        [0, 0, 0], [tint32, tint32, tint32], "", """
 718 int base = src0;
 719 int offset = src1, bits = src2;
 720 if (bits == 0) {
 721    dst = 0;
 722 } else if (offset < 0 || bits < 0 || offset + bits > 32) {
 723    dst = 0;
 724 } else {
 725    dst = (base << (32 - offset - bits)) >> offset; /* use sign-extending shift */
 726 }
 727 """)
 728
 729 # Combines the first component of each input to make a 3-component vector.
 730
 731 triop_horiz("vec3", 3, 1, 1, 1, """
 732 dst.x = src0.x;
 733 dst.y = src1.x;
 734 dst.z = src2.x;
 735 """)
 736
 737 def quadop_horiz(name, output_size, src1_size, src2_size, src3_size,
 738                  src4_size, const_expr):
 739    opcode(name, output_size, tuint,
 740           [src1_size, src2_size, src3_size, src4_size],
 741           [tuint, tuint, tuint, tuint],
 742           "", const_expr)
 743
 744 opcode("bitfield_insert", 0, tuint32, [0, 0, 0, 0],
 745        [tuint32, tuint32, tint32, tint32], "", """
 746 unsigned base = src0, insert = src1;
 747 int offset = src2, bits = src3;
 748 if (bits == 0) {
 749    dst = 0;
 750 } else if (offset < 0 || bits < 0 || bits + offset > 32) {
 751    dst = 0;
 752 } else {
 753    unsigned mask = ((1ull << bits) - 1) << offset;
 754    dst = (base & ~mask) | ((insert << bits) & mask);
 755 }
 756 """)
 757
 758 quadop_horiz("vec4", 4, 1, 1, 1, 1, """
 759 dst.x = src0.x;
 760 dst.y = src1.x;
 761 dst.z = src2.x;
 762 dst.w = src3.x;
 763 """)
 764
 765