src/compiler/nir/nir_opcodes.py

   1 #
   2 # Copyright (C) 2014 Connor Abbott
   3 #
   4 # Permission is hereby granted, free of charge, to any person obtaining a
   5 # copy of this software and associated documentation files (the "Software"),
   6 # to deal in the Software without restriction, including without limitation
   7 # the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8 # and/or sell copies of the Software, and to permit persons to whom the
   9 # Software is furnished to do so, subject to the following conditions:
  10 #
  11 # The above copyright notice and this permission notice (including the next
  12 # paragraph) shall be included in all copies or substantial portions of the
  13 # Software.
  14 #
  15 # THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16 # IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17 # FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  18 # THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  19 # LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  20 # FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  21 # IN THE SOFTWARE.
  22 #
  23 # Authors:
  24 #    Connor Abbott (cwabbott0@gmail.com)
  25
  26 import re
  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, is_conversion, 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       - is_conversion is true if this opcode represents a type conversion
  43       - algebraic_properties is a space-seperated string, where nir_op_is_ is
  44         prepended before each entry
  45       - const_expr is an expression or series of statements that computes the
  46         constant value of the opcode given the constant values of its inputs.
  47
  48       Constant expressions are formed from the variables src0, src1, ...,
  49       src(N-1), where N is the number of arguments.  The output of the
  50       expression should be stored in the dst variable.  Per-component input
  51       and output variables will be scalars and non-per-component input and
  52       output variables will be a struct with fields named x, y, z, and w
  53       all of the correct type.  Input and output variables can be assumed
  54       to already be of the correct type and need no conversion.  In
  55       particular, the conversion from the C bool type to/from  NIR_TRUE and
  56       NIR_FALSE happens automatically.
  57
  58       For per-component instructions, the entire expression will be
  59       executed once for each component.  For non-per-component
  60       instructions, the expression is expected to store the correct values
  61       in dst.x, dst.y, etc.  If "dst" does not exist anywhere in the
  62       constant expression, an assignment to dst will happen automatically
  63       and the result will be equivalent to "dst = <expression>" for
  64       per-component instructions and "dst.x = dst.y = ... = <expression>"
  65       for non-per-component instructions.
  66       """
  67       assert isinstance(name, str)
  68       assert isinstance(output_size, int)
  69       assert isinstance(output_type, str)
  70       assert isinstance(input_sizes, list)
  71       assert isinstance(input_sizes[0], int)
  72       assert isinstance(input_types, list)
  73       assert isinstance(input_types[0], str)
  74       assert isinstance(is_conversion, bool)
  75       assert isinstance(algebraic_properties, str)
  76       assert isinstance(const_expr, str)
  77       assert len(input_sizes) == len(input_types)
  78       assert 0 <= output_size <= 4
  79       for size in input_sizes:
  80          assert 0 <= size <= 4
  81          if output_size != 0:
  82             assert size != 0
  83       self.name = name
  84       self.num_inputs = len(input_sizes)
  85       self.output_size = output_size
  86       self.output_type = output_type
  87       self.input_sizes = input_sizes
  88       self.input_types = input_types
  89       self.is_conversion = is_conversion
  90       self.algebraic_properties = algebraic_properties
  91       self.const_expr = const_expr
  92
  93 # helper variables for strings
  94 tfloat = "float"
  95 tint = "int"
  96 tbool = "bool"
  97 tbool1 = "bool1"
  98 tbool32 = "bool32"
  99 tuint = "uint"
 100 tuint16 = "uint16"
 101 tfloat32 = "float32"
 102 tint32 = "int32"
 103 tuint32 = "uint32"
 104 tint64 = "int64"
 105 tuint64 = "uint64"
 106 tfloat64 = "float64"
 107
 108 _TYPE_SPLIT_RE = re.compile(r'(?P<type>int|uint|float|bool)(?P<bits>\d+)?')
 109
 110 def type_has_size(type_):
 111     m = _TYPE_SPLIT_RE.match(type_)
 112     assert m is not None, 'Invalid NIR type string: "{}"'.format(type_)
 113     return m.group('bits') is not None
 114
 115 def type_size(type_):
 116     m = _TYPE_SPLIT_RE.match(type_)
 117     assert m is not None, 'Invalid NIR type string: "{}"'.format(type_)
 118     assert m.group('bits') is not None, \
 119            'NIR type string has no bit size: "{}"'.format(type_)
 120     return int(m.group('bits'))
 121
 122 def type_sizes(type_):
 123     if type_has_size(type_):
 124         return [type_size(type_)]
 125     elif type_ == 'bool':
 126         return [1, 32]
 127     elif type_ == 'float':
 128         return [16, 32, 64]
 129     else:
 130         return [1, 8, 16, 32, 64]
 131
 132 def type_base_type(type_):
 133     m = _TYPE_SPLIT_RE.match(type_)
 134     assert m is not None, 'Invalid NIR type string: "{}"'.format(type_)
 135     return m.group('type')
 136
 137 # Operation where the first two sources are commutative.
 138 #
 139 # For 2-source operations, this just mathematical commutativity.  Some
 140 # 3-source operations, like ffma, are only commutative in the first two
 141 # sources.
 142 _2src_commutative = "2src_commutative "
 143 associative = "associative "
 144
 145 # global dictionary of opcodes
 146 opcodes = {}
 147
 148 def opcode(name, output_size, output_type, input_sizes, input_types,
 149            is_conversion, algebraic_properties, const_expr):
 150    assert name not in opcodes
 151    opcodes[name] = Opcode(name, output_size, output_type, input_sizes,
 152                           input_types, is_conversion, algebraic_properties,
 153                           const_expr)
 154
 155 def unop_convert(name, out_type, in_type, const_expr):
 156    opcode(name, 0, out_type, [0], [in_type], False, "", const_expr)
 157
 158 def unop(name, ty, const_expr):
 159    opcode(name, 0, ty, [0], [ty], False, "", const_expr)
 160
 161 def unop_horiz(name, output_size, output_type, input_size, input_type,
 162                const_expr):
 163    opcode(name, output_size, output_type, [input_size], [input_type],
 164           False, "", const_expr)
 165
 166 def unop_reduce(name, output_size, output_type, input_type, prereduce_expr,
 167                 reduce_expr, final_expr):
 168    def prereduce(src):
 169       return "(" + prereduce_expr.format(src=src) + ")"
 170    def final(src):
 171       return final_expr.format(src="(" + src + ")")
 172    def reduce_(src0, src1):
 173       return reduce_expr.format(src0=src0, src1=src1)
 174    src0 = prereduce("src0.x")
 175    src1 = prereduce("src0.y")
 176    src2 = prereduce("src0.z")
 177    src3 = prereduce("src0.w")
 178    unop_horiz(name + "2", output_size, output_type, 2, input_type,
 179               final(reduce_(src0, src1)))
 180    unop_horiz(name + "3", output_size, output_type, 3, input_type,
 181               final(reduce_(reduce_(src0, src1), src2)))
 182    unop_horiz(name + "4", output_size, output_type, 4, input_type,
 183               final(reduce_(reduce_(src0, src1), reduce_(src2, src3))))
 184
 185 def unop_numeric_convert(name, out_type, in_type, const_expr):
 186    opcode(name, 0, out_type, [0], [in_type], True, "", const_expr)
 187
 188 unop("mov", tuint, "src0")
 189
 190 unop("ineg", tint, "-src0")
 191 unop("fneg", tfloat, "-src0")
 192 unop("inot", tint, "~src0") # invert every bit of the integer
 193 unop("fnot", tfloat, ("bit_size == 64 ? ((src0 == 0.0) ? 1.0 : 0.0f) : " +
 194                       "((src0 == 0.0f) ? 1.0f : 0.0f)"))
 195 unop("fsign", tfloat, ("bit_size == 64 ? " +
 196                        "((src0 == 0.0) ? 0.0 : ((src0 > 0.0) ? 1.0 : -1.0)) : " +
 197                        "((src0 == 0.0f) ? 0.0f : ((src0 > 0.0f) ? 1.0f : -1.0f))"))
 198 unop("isign", tint, "(src0 == 0) ? 0 : ((src0 > 0) ? 1 : -1)")
 199 unop("iabs", tint, "(src0 < 0) ? -src0 : src0")
 200 unop("fabs", tfloat, "fabs(src0)")
 201 unop("fsat", tfloat, ("bit_size == 64 ? " +
 202                       "((src0 > 1.0) ? 1.0 : ((src0 <= 0.0) ? 0.0 : src0)) : " +
 203                       "((src0 > 1.0f) ? 1.0f : ((src0 <= 0.0f) ? 0.0f : src0))"))
 204 unop("frcp", tfloat, "bit_size == 64 ? 1.0 / src0 : 1.0f / src0")
 205 unop("frsq", tfloat, "bit_size == 64 ? 1.0 / sqrt(src0) : 1.0f / sqrtf(src0)")
 206 unop("fsqrt", tfloat, "bit_size == 64 ? sqrt(src0) : sqrtf(src0)")
 207 unop("fexp2", tfloat, "exp2f(src0)")
 208 unop("flog2", tfloat, "log2f(src0)")
 209
 210 # Generate all of the numeric conversion opcodes
 211 for src_t in [tint, tuint, tfloat, tbool]:
 212    if src_t == tbool:
 213       dst_types = [tfloat, tint]
 214    elif src_t == tint:
 215       dst_types = [tfloat, tint, tbool]
 216    elif src_t == tuint:
 217       dst_types = [tfloat, tuint]
 218    elif src_t == tfloat:
 219       dst_types = [tint, tuint, tfloat, tbool]
 220
 221    for dst_t in dst_types:
 222       for bit_size in type_sizes(dst_t):
 223           if bit_size == 16 and dst_t == tfloat and src_t == tfloat:
 224               rnd_modes = ['_rtne', '_rtz', '']
 225               for rnd_mode in rnd_modes:
 226                   unop_numeric_convert("{0}2{1}{2}{3}".format(src_t[0], dst_t[0],
 227                                                               bit_size, rnd_mode),
 228                                        dst_t + str(bit_size), src_t, "src0")
 229           else:
 230               conv_expr = "src0 != 0" if dst_t == tbool else "src0"
 231               unop_numeric_convert("{0}2{1}{2}".format(src_t[0], dst_t[0], bit_size),
 232                                    dst_t + str(bit_size), src_t, conv_expr)
 233
 234
 235 # Unary floating-point rounding operations.
 236
 237
 238 unop("ftrunc", tfloat, "bit_size == 64 ? trunc(src0) : truncf(src0)")
 239 unop("fceil", tfloat, "bit_size == 64 ? ceil(src0) : ceilf(src0)")
 240 unop("ffloor", tfloat, "bit_size == 64 ? floor(src0) : floorf(src0)")
 241 unop("ffract", tfloat, "src0 - (bit_size == 64 ? floor(src0) : floorf(src0))")
 242 unop("fround_even", tfloat, "bit_size == 64 ? _mesa_roundeven(src0) : _mesa_roundevenf(src0)")
 243
 244 unop("fquantize2f16", tfloat, "(fabs(src0) < ldexpf(1.0, -14)) ? copysignf(0.0f, src0) : _mesa_half_to_float(_mesa_float_to_half(src0))")
 245
 246 # Trigonometric operations.
 247
 248
 249 unop("fsin", tfloat, "bit_size == 64 ? sin(src0) : sinf(src0)")
 250 unop("fcos", tfloat, "bit_size == 64 ? cos(src0) : cosf(src0)")
 251
 252 # dfrexp
 253 unop_convert("frexp_exp", tint32, tfloat, "frexp(src0, &dst);")
 254 unop_convert("frexp_sig", tfloat, tfloat, "int n; dst = frexp(src0, &n);")
 255
 256 # Partial derivatives.
 257
 258
 259 unop("fddx", tfloat, "0.0") # the derivative of a constant is 0.
 260 unop("fddy", tfloat, "0.0")
 261 unop("fddx_fine", tfloat, "0.0")
 262 unop("fddy_fine", tfloat, "0.0")
 263 unop("fddx_coarse", tfloat, "0.0")
 264 unop("fddy_coarse", tfloat, "0.0")
 265
 266
 267 # Floating point pack and unpack operations.
 268
 269 def pack_2x16(fmt):
 270    unop_horiz("pack_" + fmt + "_2x16", 1, tuint32, 2, tfloat32, """
 271 dst.x = (uint32_t) pack_fmt_1x16(src0.x);
 272 dst.x |= ((uint32_t) pack_fmt_1x16(src0.y)) << 16;
 273 """.replace("fmt", fmt))
 274
 275 def pack_4x8(fmt):
 276    unop_horiz("pack_" + fmt + "_4x8", 1, tuint32, 4, tfloat32, """
 277 dst.x = (uint32_t) pack_fmt_1x8(src0.x);
 278 dst.x |= ((uint32_t) pack_fmt_1x8(src0.y)) << 8;
 279 dst.x |= ((uint32_t) pack_fmt_1x8(src0.z)) << 16;
 280 dst.x |= ((uint32_t) pack_fmt_1x8(src0.w)) << 24;
 281 """.replace("fmt", fmt))
 282
 283 def unpack_2x16(fmt):
 284    unop_horiz("unpack_" + fmt + "_2x16", 2, tfloat32, 1, tuint32, """
 285 dst.x = unpack_fmt_1x16((uint16_t)(src0.x & 0xffff));
 286 dst.y = unpack_fmt_1x16((uint16_t)(src0.x << 16));
 287 """.replace("fmt", fmt))
 288
 289 def unpack_4x8(fmt):
 290    unop_horiz("unpack_" + fmt + "_4x8", 4, tfloat32, 1, tuint32, """
 291 dst.x = unpack_fmt_1x8((uint8_t)(src0.x & 0xff));
 292 dst.y = unpack_fmt_1x8((uint8_t)((src0.x >> 8) & 0xff));
 293 dst.z = unpack_fmt_1x8((uint8_t)((src0.x >> 16) & 0xff));
 294 dst.w = unpack_fmt_1x8((uint8_t)(src0.x >> 24));
 295 """.replace("fmt", fmt))
 296
 297
 298 pack_2x16("snorm")
 299 pack_4x8("snorm")
 300 pack_2x16("unorm")
 301 pack_4x8("unorm")
 302 pack_2x16("half")
 303 unpack_2x16("snorm")
 304 unpack_4x8("snorm")
 305 unpack_2x16("unorm")
 306 unpack_4x8("unorm")
 307 unpack_2x16("half")
 308
 309 unop_horiz("pack_uvec2_to_uint", 1, tuint32, 2, tuint32, """
 310 dst.x = (src0.x & 0xffff) | (src0.y << 16);
 311 """)
 312
 313 unop_horiz("pack_uvec4_to_uint", 1, tuint32, 4, tuint32, """
 314 dst.x = (src0.x <<  0) |
 315         (src0.y <<  8) |
 316         (src0.z << 16) |
 317         (src0.w << 24);
 318 """)
 319
 320 unop_horiz("pack_32_2x16", 1, tuint32, 2, tuint16,
 321            "dst.x = src0.x | ((uint32_t)src0.y << 16);")
 322
 323 unop_horiz("pack_64_2x32", 1, tuint64, 2, tuint32,
 324            "dst.x = src0.x | ((uint64_t)src0.y << 32);")
 325
 326 unop_horiz("pack_64_4x16", 1, tuint64, 4, tuint16,
 327            "dst.x = src0.x | ((uint64_t)src0.y << 16) | ((uint64_t)src0.z << 32) | ((uint64_t)src0.w << 48);")
 328
 329 unop_horiz("unpack_64_2x32", 2, tuint32, 1, tuint64,
 330            "dst.x = src0.x; dst.y = src0.x >> 32;")
 331
 332 unop_horiz("unpack_64_4x16", 4, tuint16, 1, tuint64,
 333            "dst.x = src0.x; dst.y = src0.x >> 16; dst.z = src0.x >> 32; dst.w = src0.w >> 48;")
 334
 335 unop_horiz("unpack_32_2x16", 2, tuint16, 1, tuint32,
 336            "dst.x = src0.x; dst.y = src0.x >> 16;")
 337
 338 # Lowered floating point unpacking operations.
 339
 340
 341 unop_convert("unpack_half_2x16_split_x", tfloat32, tuint32,
 342              "unpack_half_1x16((uint16_t)(src0 & 0xffff))")
 343 unop_convert("unpack_half_2x16_split_y", tfloat32, tuint32,
 344              "unpack_half_1x16((uint16_t)(src0 >> 16))")
 345
 346 unop_convert("unpack_32_2x16_split_x", tuint16, tuint32, "src0")
 347 unop_convert("unpack_32_2x16_split_y", tuint16, tuint32, "src0 >> 16")
 348
 349 unop_convert("unpack_64_2x32_split_x", tuint32, tuint64, "src0")
 350 unop_convert("unpack_64_2x32_split_y", tuint32, tuint64, "src0 >> 32")
 351
 352 # Bit operations, part of ARB_gpu_shader5.
 353
 354
 355 unop("bitfield_reverse", tuint32, """
 356 /* we're not winning any awards for speed here, but that's ok */
 357 dst = 0;
 358 for (unsigned bit = 0; bit < 32; bit++)
 359    dst |= ((src0 >> bit) & 1) << (31 - bit);
 360 """)
 361 unop_convert("bit_count", tuint32, tuint, """
 362 dst = 0;
 363 for (unsigned bit = 0; bit < bit_size; bit++) {
 364    if ((src0 >> bit) & 1)
 365       dst++;
 366 }
 367 """)
 368
 369 unop_convert("ufind_msb", tint32, tuint, """
 370 dst = -1;
 371 for (int bit = bit_size - 1; bit >= 0; bit--) {
 372    if ((src0 >> bit) & 1) {
 373       dst = bit;
 374       break;
 375    }
 376 }
 377 """)
 378
 379 unop("ifind_msb", tint32, """
 380 dst = -1;
 381 for (int bit = 31; bit >= 0; bit--) {
 382    /* If src0 < 0, we're looking for the first 0 bit.
 383     * if src0 >= 0, we're looking for the first 1 bit.
 384     */
 385    if ((((src0 >> bit) & 1) && (src0 >= 0)) ||
 386       (!((src0 >> bit) & 1) && (src0 < 0))) {
 387       dst = bit;
 388       break;
 389    }
 390 }
 391 """)
 392
 393 unop_convert("find_lsb", tint32, tint, """
 394 dst = -1;
 395 for (unsigned bit = 0; bit < bit_size; bit++) {
 396    if ((src0 >> bit) & 1) {
 397       dst = bit;
 398       break;
 399    }
 400 }
 401 """)
 402
 403
 404 for i in range(1, 5):
 405    for j in range(1, 5):
 406       unop_horiz("fnoise{0}_{1}".format(i, j), i, tfloat, j, tfloat, "0.0f")
 407
 408
 409 # AMD_gcn_shader extended instructions
 410 unop_horiz("cube_face_coord", 2, tfloat32, 3, tfloat32, """
 411 dst.x = dst.y = 0.0;
 412 float absX = fabs(src0.x);
 413 float absY = fabs(src0.y);
 414 float absZ = fabs(src0.z);
 415
 416 float ma = 0.0;
 417 if (absX >= absY && absX >= absZ) { ma = 2 * src0.x; }
 418 if (absY >= absX && absY >= absZ) { ma = 2 * src0.y; }
 419 if (absZ >= absX && absZ >= absY) { ma = 2 * src0.z; }
 420
 421 if (src0.x >= 0 && absX >= absY && absX >= absZ) { dst.x = -src0.z; dst.y = -src0.y; }
 422 if (src0.x < 0 && absX >= absY && absX >= absZ) { dst.x = src0.z; dst.y = -src0.y; }
 423 if (src0.y >= 0 && absY >= absX && absY >= absZ) { dst.x = src0.x; dst.y = src0.z; }
 424 if (src0.y < 0 && absY >= absX && absY >= absZ) { dst.x = src0.x; dst.y = -src0.z; }
 425 if (src0.z >= 0 && absZ >= absX && absZ >= absY) { dst.x = src0.x; dst.y = -src0.y; }
 426 if (src0.z < 0 && absZ >= absX && absZ >= absY) { dst.x = -src0.x; dst.y = -src0.y; }
 427
 428 dst.x = dst.x / ma + 0.5;
 429 dst.y = dst.y / ma + 0.5;
 430 """)
 431
 432 unop_horiz("cube_face_index", 1, tfloat32, 3, tfloat32, """
 433 float absX = fabs(src0.x);
 434 float absY = fabs(src0.y);
 435 float absZ = fabs(src0.z);
 436 if (src0.x >= 0 && absX >= absY && absX >= absZ) dst.x = 0;
 437 if (src0.x < 0 && absX >= absY && absX >= absZ) dst.x = 1;
 438 if (src0.y >= 0 && absY >= absX && absY >= absZ) dst.x = 2;
 439 if (src0.y < 0 && absY >= absX && absY >= absZ) dst.x = 3;
 440 if (src0.z >= 0 && absZ >= absX && absZ >= absY) dst.x = 4;
 441 if (src0.z < 0 && absZ >= absX && absZ >= absY) dst.x = 5;
 442 """)
 443
 444
 445 def binop_convert(name, out_type, in_type, alg_props, const_expr):
 446    opcode(name, 0, out_type, [0, 0], [in_type, in_type],
 447           False, alg_props, const_expr)
 448
 449 def binop(name, ty, alg_props, const_expr):
 450    binop_convert(name, ty, ty, alg_props, const_expr)
 451
 452 def binop_compare(name, ty, alg_props, const_expr):
 453    binop_convert(name, tbool1, ty, alg_props, const_expr)
 454
 455 def binop_compare32(name, ty, alg_props, const_expr):
 456    binop_convert(name, tbool32, ty, alg_props, const_expr)
 457
 458 def binop_horiz(name, out_size, out_type, src1_size, src1_type, src2_size,
 459                 src2_type, const_expr):
 460    opcode(name, out_size, out_type, [src1_size, src2_size], [src1_type, src2_type],
 461           False, "", const_expr)
 462
 463 def binop_reduce(name, output_size, output_type, src_type, prereduce_expr,
 464                  reduce_expr, final_expr):
 465    def final(src):
 466       return final_expr.format(src= "(" + src + ")")
 467    def reduce_(src0, src1):
 468       return reduce_expr.format(src0=src0, src1=src1)
 469    def prereduce(src0, src1):
 470       return "(" + prereduce_expr.format(src0=src0, src1=src1) + ")"
 471    src0 = prereduce("src0.x", "src1.x")
 472    src1 = prereduce("src0.y", "src1.y")
 473    src2 = prereduce("src0.z", "src1.z")
 474    src3 = prereduce("src0.w", "src1.w")
 475    opcode(name + "2", output_size, output_type,
 476           [2, 2], [src_type, src_type], False, _2src_commutative,
 477           final(reduce_(src0, src1)))
 478    opcode(name + "3", output_size, output_type,
 479           [3, 3], [src_type, src_type], False, _2src_commutative,
 480           final(reduce_(reduce_(src0, src1), src2)))
 481    opcode(name + "4", output_size, output_type,
 482           [4, 4], [src_type, src_type], False, _2src_commutative,
 483           final(reduce_(reduce_(src0, src1), reduce_(src2, src3))))
 484
 485 binop("fadd", tfloat, _2src_commutative + associative, "src0 + src1")
 486 binop("iadd", tint, _2src_commutative + associative, "src0 + src1")
 487 binop("iadd_sat", tint, _2src_commutative, """
 488       src1 > 0 ?
 489          (src0 + src1 < src0 ? (1ull << (bit_size - 1)) - 1 : src0 + src1) :
 490          (src0 < src0 + src1 ? (1ull << (bit_size - 1))     : src0 + src1)
 491 """)
 492 binop("uadd_sat", tuint, _2src_commutative,
 493       "(src0 + src1) < src0 ? MAX_UINT_FOR_SIZE(sizeof(src0) * 8) : (src0 + src1)")
 494 binop("isub_sat", tint, "", """
 495       src1 < 0 ?
 496          (src0 - src1 < src0 ? (1ull << (bit_size - 1)) - 1 : src0 - src1) :
 497          (src0 < src0 - src1 ? (1ull << (bit_size - 1))     : src0 - src1)
 498 """)
 499 binop("usub_sat", tuint, "", "src0 < src1 ? 0 : src0 - src1")
 500
 501 binop("fsub", tfloat, "", "src0 - src1")
 502 binop("isub", tint, "", "src0 - src1")
 503
 504 binop("fmul", tfloat, _2src_commutative + associative, "src0 * src1")
 505 # low 32-bits of signed/unsigned integer multiply
 506 binop("imul", tint, _2src_commutative + associative, "src0 * src1")
 507
 508 # Generate 64 bit result from 2 32 bits quantity
 509 binop_convert("imul_2x32_64", tint64, tint32, _2src_commutative,
 510               "(int64_t)src0 * (int64_t)src1")
 511 binop_convert("umul_2x32_64", tuint64, tuint32, _2src_commutative,
 512               "(uint64_t)src0 * (uint64_t)src1")
 513
 514 # high 32-bits of signed integer multiply
 515 binop("imul_high", tint, _2src_commutative, """
 516 if (bit_size == 64) {
 517    /* We need to do a full 128-bit x 128-bit multiply in order for the sign
 518     * extension to work properly.  The casts are kind-of annoying but needed
 519     * to prevent compiler warnings.
 520     */
 521    uint32_t src0_u32[4] = {
 522       src0,
 523       (int64_t)src0 >> 32,
 524       (int64_t)src0 >> 63,
 525       (int64_t)src0 >> 63,
 526    };
 527    uint32_t src1_u32[4] = {
 528       src1,
 529       (int64_t)src1 >> 32,
 530       (int64_t)src1 >> 63,
 531       (int64_t)src1 >> 63,
 532    };
 533    uint32_t prod_u32[4];
 534    ubm_mul_u32arr(prod_u32, src0_u32, src1_u32);
 535    dst = (uint64_t)prod_u32[2] | ((uint64_t)prod_u32[3] << 32);
 536 } else {
 537    dst = ((int64_t)src0 * (int64_t)src1) >> bit_size;
 538 }
 539 """)
 540
 541 # high 32-bits of unsigned integer multiply
 542 binop("umul_high", tuint, _2src_commutative, """
 543 if (bit_size == 64) {
 544    /* The casts are kind-of annoying but needed to prevent compiler warnings. */
 545    uint32_t src0_u32[2] = { src0, (uint64_t)src0 >> 32 };
 546    uint32_t src1_u32[2] = { src1, (uint64_t)src1 >> 32 };
 547    uint32_t prod_u32[4];
 548    ubm_mul_u32arr(prod_u32, src0_u32, src1_u32);
 549    dst = (uint64_t)prod_u32[2] | ((uint64_t)prod_u32[3] << 32);
 550 } else {
 551    dst = ((uint64_t)src0 * (uint64_t)src1) >> bit_size;
 552 }
 553 """)
 554
 555 # low 32-bits of unsigned integer multiply
 556 binop("umul_low", tuint32, _2src_commutative, """
 557 uint64_t mask = (1 << (bit_size / 2)) - 1;
 558 dst = ((uint64_t)src0 & mask) * ((uint64_t)src1 & mask);
 559 """)
 560
 561
 562 binop("fdiv", tfloat, "", "src0 / src1")
 563 binop("idiv", tint, "", "src1 == 0 ? 0 : (src0 / src1)")
 564 binop("udiv", tuint, "", "src1 == 0 ? 0 : (src0 / src1)")
 565
 566 # returns a boolean representing the carry resulting from the addition of
 567 # the two unsigned arguments.
 568
 569 binop_convert("uadd_carry", tuint, tuint, _2src_commutative, "src0 + src1 < src0")
 570
 571 # returns a boolean representing the borrow resulting from the subtraction
 572 # of the two unsigned arguments.
 573
 574 binop_convert("usub_borrow", tuint, tuint, "", "src0 < src1")
 575
 576 # hadd: (a + b) >> 1 (without overflow)
 577 # x + y = x - (x & ~y) + (x & ~y) + y - (~x & y) + (~x & y)
 578 #       =      (x & y) + (x & ~y) +      (x & y) + (~x & y)
 579 #       = 2 *  (x & y) + (x & ~y) +                (~x & y)
 580 #       =     ((x & y) << 1) + (x ^ y)
 581 #
 582 # Since we know that the bottom bit of (x & y) << 1 is zero,
 583 #
 584 # (x + y) >> 1 = (((x & y) << 1) + (x ^ y)) >> 1
 585 #              =   (x & y) +      ((x ^ y)  >> 1)
 586 binop("ihadd", tint, _2src_commutative, "(src0 & src1) + ((src0 ^ src1) >> 1)")
 587 binop("uhadd", tuint, _2src_commutative, "(src0 & src1) + ((src0 ^ src1) >> 1)")
 588
 589 # rhadd: (a + b + 1) >> 1 (without overflow)
 590 # x + y + 1 = x + (~x & y) - (~x & y) + y + (x & ~y) - (x & ~y) + 1
 591 #           =      (x | y) - (~x & y) +      (x | y) - (x & ~y) + 1
 592 #           = 2 *  (x | y) - ((~x & y) +               (x & ~y)) + 1
 593 #           =     ((x | y) << 1) - (x ^ y) + 1
 594 #
 595 # Since we know that the bottom bit of (x & y) << 1 is zero,
 596 #
 597 # (x + y + 1) >> 1 = (x | y) + (-(x ^ y) + 1) >> 1)
 598 #                  = (x | y) -  ((x ^ y)      >> 1)
 599 binop("irhadd", tint, _2src_commutative, "(src0 | src1) + ((src0 ^ src1) >> 1)")
 600 binop("urhadd", tuint, _2src_commutative, "(src0 | src1) + ((src0 ^ src1) >> 1)")
 601
 602 binop("umod", tuint, "", "src1 == 0 ? 0 : src0 % src1")
 603
 604 # For signed integers, there are several different possible definitions of
 605 # "modulus" or "remainder".  We follow the conventions used by LLVM and
 606 # SPIR-V.  The irem opcode implements the standard C/C++ signed "%"
 607 # operation while the imod opcode implements the more mathematical
 608 # "modulus" operation.  For details on the difference, see
 609 #
 610 # http://mathforum.org/library/drmath/view/52343.html
 611
 612 binop("irem", tint, "", "src1 == 0 ? 0 : src0 % src1")
 613 binop("imod", tint, "",
 614       "src1 == 0 ? 0 : ((src0 % src1 == 0 || (src0 >= 0) == (src1 >= 0)) ?"
 615       "                 src0 % src1 : src0 % src1 + src1)")
 616 binop("fmod", tfloat, "", "src0 - src1 * floorf(src0 / src1)")
 617 binop("frem", tfloat, "", "src0 - src1 * truncf(src0 / src1)")
 618
 619 #
 620 # Comparisons
 621 #
 622
 623
 624 # these integer-aware comparisons return a boolean (0 or ~0)
 625
 626 binop_compare("flt", tfloat, "", "src0 < src1")
 627 binop_compare("fge", tfloat, "", "src0 >= src1")
 628 binop_compare("feq", tfloat, _2src_commutative, "src0 == src1")
 629 binop_compare("fne", tfloat, _2src_commutative, "src0 != src1")
 630 binop_compare("ilt", tint, "", "src0 < src1")
 631 binop_compare("ige", tint, "", "src0 >= src1")
 632 binop_compare("ieq", tint, _2src_commutative, "src0 == src1")
 633 binop_compare("ine", tint, _2src_commutative, "src0 != src1")
 634 binop_compare("ult", tuint, "", "src0 < src1")
 635 binop_compare("uge", tuint, "", "src0 >= src1")
 636 binop_compare32("flt32", tfloat, "", "src0 < src1")
 637 binop_compare32("fge32", tfloat, "", "src0 >= src1")
 638 binop_compare32("feq32", tfloat, _2src_commutative, "src0 == src1")
 639 binop_compare32("fne32", tfloat, _2src_commutative, "src0 != src1")
 640 binop_compare32("ilt32", tint, "", "src0 < src1")
 641 binop_compare32("ige32", tint, "", "src0 >= src1")
 642 binop_compare32("ieq32", tint, _2src_commutative, "src0 == src1")
 643 binop_compare32("ine32", tint, _2src_commutative, "src0 != src1")
 644 binop_compare32("ult32", tuint, "", "src0 < src1")
 645 binop_compare32("uge32", tuint, "", "src0 >= src1")
 646
 647 # integer-aware GLSL-style comparisons that compare floats and ints
 648
 649 binop_reduce("ball_fequal",  1, tbool1, tfloat, "{src0} == {src1}",
 650              "{src0} && {src1}", "{src}")
 651 binop_reduce("bany_fnequal", 1, tbool1, tfloat, "{src0} != {src1}",
 652              "{src0} || {src1}", "{src}")
 653 binop_reduce("ball_iequal",  1, tbool1, tint, "{src0} == {src1}",
 654              "{src0} && {src1}", "{src}")
 655 binop_reduce("bany_inequal", 1, tbool1, tint, "{src0} != {src1}",
 656              "{src0} || {src1}", "{src}")
 657
 658 binop_reduce("b32all_fequal",  1, tbool32, tfloat, "{src0} == {src1}",
 659              "{src0} && {src1}", "{src}")
 660 binop_reduce("b32any_fnequal", 1, tbool32, tfloat, "{src0} != {src1}",
 661              "{src0} || {src1}", "{src}")
 662 binop_reduce("b32all_iequal",  1, tbool32, tint, "{src0} == {src1}",
 663              "{src0} && {src1}", "{src}")
 664 binop_reduce("b32any_inequal", 1, tbool32, tint, "{src0} != {src1}",
 665              "{src0} || {src1}", "{src}")
 666
 667 # non-integer-aware GLSL-style comparisons that return 0.0 or 1.0
 668
 669 binop_reduce("fall_equal",  1, tfloat32, tfloat32, "{src0} == {src1}",
 670              "{src0} && {src1}", "{src} ? 1.0f : 0.0f")
 671 binop_reduce("fany_nequal", 1, tfloat32, tfloat32, "{src0} != {src1}",
 672              "{src0} || {src1}", "{src} ? 1.0f : 0.0f")
 673
 674 # These comparisons for integer-less hardware return 1.0 and 0.0 for true
 675 # and false respectively
 676
 677 binop("slt", tfloat32, "", "(src0 < src1) ? 1.0f : 0.0f") # Set on Less Than
 678 binop("sge", tfloat, "", "(src0 >= src1) ? 1.0f : 0.0f") # Set on Greater or Equal
 679 binop("seq", tfloat32, _2src_commutative, "(src0 == src1) ? 1.0f : 0.0f") # Set on Equal
 680 binop("sne", tfloat32, _2src_commutative, "(src0 != src1) ? 1.0f : 0.0f") # Set on Not Equal
 681
 682 # SPIRV shifts are undefined for shift-operands >= bitsize,
 683 # but SM5 shifts are defined to use the least significant bits, only
 684 # The NIR definition is according to the SM5 specification.
 685 opcode("ishl", 0, tint, [0, 0], [tint, tuint32], False, "",
 686        "src0 << (src1 & (sizeof(src0) * 8 - 1))")
 687 opcode("ishr", 0, tint, [0, 0], [tint, tuint32], False, "",
 688        "src0 >> (src1 & (sizeof(src0) * 8 - 1))")
 689 opcode("ushr", 0, tuint, [0, 0], [tuint, tuint32], False, "",
 690        "src0 >> (src1 & (sizeof(src0) * 8 - 1))")
 691
 692 # bitwise logic operators
 693 #
 694 # These are also used as boolean and, or, xor for hardware supporting
 695 # integers.
 696
 697
 698 binop("iand", tuint, _2src_commutative + associative, "src0 & src1")
 699 binop("ior", tuint, _2src_commutative + associative, "src0 | src1")
 700 binop("ixor", tuint, _2src_commutative + associative, "src0 ^ src1")
 701
 702
 703 # floating point logic operators
 704 #
 705 # These use (src != 0.0) for testing the truth of the input, and output 1.0
 706 # for true and 0.0 for false
 707
 708 binop("fand", tfloat32, _2src_commutative,
 709       "((src0 != 0.0f) && (src1 != 0.0f)) ? 1.0f : 0.0f")
 710 binop("for", tfloat32, _2src_commutative,
 711       "((src0 != 0.0f) || (src1 != 0.0f)) ? 1.0f : 0.0f")
 712 binop("fxor", tfloat32, _2src_commutative,
 713       "(src0 != 0.0f && src1 == 0.0f) || (src0 == 0.0f && src1 != 0.0f) ? 1.0f : 0.0f")
 714
 715 binop_reduce("fdot", 1, tfloat, tfloat, "{src0} * {src1}", "{src0} + {src1}",
 716              "{src}")
 717
 718 binop_reduce("fdot_replicated", 4, tfloat, tfloat,
 719              "{src0} * {src1}", "{src0} + {src1}", "{src}")
 720
 721 opcode("fdph", 1, tfloat, [3, 4], [tfloat, tfloat], False, "",
 722        "src0.x * src1.x + src0.y * src1.y + src0.z * src1.z + src1.w")
 723 opcode("fdph_replicated", 4, tfloat, [3, 4], [tfloat, tfloat], False, "",
 724        "src0.x * src1.x + src0.y * src1.y + src0.z * src1.z + src1.w")
 725
 726 binop("fmin", tfloat, "", "fminf(src0, src1)")
 727 binop("imin", tint, _2src_commutative + associative, "src1 > src0 ? src0 : src1")
 728 binop("umin", tuint, _2src_commutative + associative, "src1 > src0 ? src0 : src1")
 729 binop("fmax", tfloat, "", "fmaxf(src0, src1)")
 730 binop("imax", tint, _2src_commutative + associative, "src1 > src0 ? src1 : src0")
 731 binop("umax", tuint, _2src_commutative + associative, "src1 > src0 ? src1 : src0")
 732
 733 # Saturated vector add for 4 8bit ints.
 734 binop("usadd_4x8", tint32, _2src_commutative + associative, """
 735 dst = 0;
 736 for (int i = 0; i < 32; i += 8) {
 737    dst |= MIN2(((src0 >> i) & 0xff) + ((src1 >> i) & 0xff), 0xff) << i;
 738 }
 739 """)
 740
 741 # Saturated vector subtract for 4 8bit ints.
 742 binop("ussub_4x8", tint32, "", """
 743 dst = 0;
 744 for (int i = 0; i < 32; i += 8) {
 745    int src0_chan = (src0 >> i) & 0xff;
 746    int src1_chan = (src1 >> i) & 0xff;
 747    if (src0_chan > src1_chan)
 748       dst |= (src0_chan - src1_chan) << i;
 749 }
 750 """)
 751
 752 # vector min for 4 8bit ints.
 753 binop("umin_4x8", tint32, _2src_commutative + associative, """
 754 dst = 0;
 755 for (int i = 0; i < 32; i += 8) {
 756    dst |= MIN2((src0 >> i) & 0xff, (src1 >> i) & 0xff) << i;
 757 }
 758 """)
 759
 760 # vector max for 4 8bit ints.
 761 binop("umax_4x8", tint32, _2src_commutative + associative, """
 762 dst = 0;
 763 for (int i = 0; i < 32; i += 8) {
 764    dst |= MAX2((src0 >> i) & 0xff, (src1 >> i) & 0xff) << i;
 765 }
 766 """)
 767
 768 # unorm multiply: (a * b) / 255.
 769 binop("umul_unorm_4x8", tint32, _2src_commutative + associative, """
 770 dst = 0;
 771 for (int i = 0; i < 32; i += 8) {
 772    int src0_chan = (src0 >> i) & 0xff;
 773    int src1_chan = (src1 >> i) & 0xff;
 774    dst |= ((src0_chan * src1_chan) / 255) << i;
 775 }
 776 """)
 777
 778 binop("fpow", tfloat, "", "bit_size == 64 ? powf(src0, src1) : pow(src0, src1)")
 779
 780 binop_horiz("pack_half_2x16_split", 1, tuint32, 1, tfloat32, 1, tfloat32,
 781             "pack_half_1x16(src0.x) | (pack_half_1x16(src1.x) << 16)")
 782
 783 binop_convert("pack_64_2x32_split", tuint64, tuint32, "",
 784               "src0 | ((uint64_t)src1 << 32)")
 785
 786 binop_convert("pack_32_2x16_split", tuint32, tuint16, "",
 787               "src0 | ((uint32_t)src1 << 16)")
 788
 789 # bfm implements the behavior of the first operation of the SM5 "bfi" assembly
 790 # and that of the "bfi1" i965 instruction. That is, the bits and offset values
 791 # are from the low five bits of src0 and src1, respectively.
 792 binop_convert("bfm", tuint32, tint32, "", """
 793 int bits = src0 & 0x1F;
 794 int offset = src1 & 0x1F;
 795 dst = ((1u << bits) - 1) << offset;
 796 """)
 797
 798 opcode("ldexp", 0, tfloat, [0, 0], [tfloat, tint32], False, "", """
 799 dst = (bit_size == 64) ? ldexp(src0, src1) : ldexpf(src0, src1);
 800 /* flush denormals to zero. */
 801 if (!isnormal(dst))
 802    dst = copysignf(0.0f, src0);
 803 """)
 804
 805 # Combines the first component of each input to make a 2-component vector.
 806
 807 binop_horiz("vec2", 2, tuint, 1, tuint, 1, tuint, """
 808 dst.x = src0.x;
 809 dst.y = src1.x;
 810 """)
 811
 812 # Byte extraction
 813 binop("extract_u8", tuint, "", "(uint8_t)(src0 >> (src1 * 8))")
 814 binop("extract_i8", tint, "", "(int8_t)(src0 >> (src1 * 8))")
 815
 816 # Word extraction
 817 binop("extract_u16", tuint, "", "(uint16_t)(src0 >> (src1 * 16))")
 818 binop("extract_i16", tint, "", "(int16_t)(src0 >> (src1 * 16))")
 819
 820
 821 def triop(name, ty, alg_props, const_expr):
 822    opcode(name, 0, ty, [0, 0, 0], [ty, ty, ty], False, alg_props, const_expr)
 823 def triop_horiz(name, output_size, src1_size, src2_size, src3_size, const_expr):
 824    opcode(name, output_size, tuint,
 825    [src1_size, src2_size, src3_size],
 826    [tuint, tuint, tuint], False, "", const_expr)
 827
 828 triop("ffma", tfloat, _2src_commutative, "src0 * src1 + src2")
 829
 830 triop("flrp", tfloat, "", "src0 * (1 - src2) + src1 * src2")
 831
 832 # Conditional Select
 833 #
 834 # A vector conditional select instruction (like ?:, but operating per-
 835 # component on vectors). There are two versions, one for floating point
 836 # bools (0.0 vs 1.0) and one for integer bools (0 vs ~0).
 837
 838
 839 triop("fcsel", tfloat32, "", "(src0 != 0.0f) ? src1 : src2")
 840
 841 # 3 way min/max/med
 842 triop("fmin3", tfloat, "", "fminf(src0, fminf(src1, src2))")
 843 triop("imin3", tint, "", "MIN2(src0, MIN2(src1, src2))")
 844 triop("umin3", tuint, "", "MIN2(src0, MIN2(src1, src2))")
 845
 846 triop("fmax3", tfloat, "", "fmaxf(src0, fmaxf(src1, src2))")
 847 triop("imax3", tint, "", "MAX2(src0, MAX2(src1, src2))")
 848 triop("umax3", tuint, "", "MAX2(src0, MAX2(src1, src2))")
 849
 850 triop("fmed3", tfloat, "", "fmaxf(fminf(fmaxf(src0, src1), src2), fminf(src0, src1))")
 851 triop("imed3", tint, "", "MAX2(MIN2(MAX2(src0, src1), src2), MIN2(src0, src1))")
 852 triop("umed3", tuint, "", "MAX2(MIN2(MAX2(src0, src1), src2), MIN2(src0, src1))")
 853
 854 opcode("bcsel", 0, tuint, [0, 0, 0],
 855       [tbool1, tuint, tuint], False, "", "src0 ? src1 : src2")
 856 opcode("b32csel", 0, tuint, [0, 0, 0],
 857        [tbool32, tuint, tuint], False, "", "src0 ? src1 : src2")
 858
 859 # SM5 bfi assembly
 860 triop("bfi", tuint32, "", """
 861 unsigned mask = src0, insert = src1, base = src2;
 862 if (mask == 0) {
 863    dst = base;
 864 } else {
 865    unsigned tmp = mask;
 866    while (!(tmp & 1)) {
 867       tmp >>= 1;
 868       insert <<= 1;
 869    }
 870    dst = (base & ~mask) | (insert & mask);
 871 }
 872 """)
 873
 874 # SM5 ubfe/ibfe assembly: only the 5 least significant bits of offset and bits are used.
 875 opcode("ubfe", 0, tuint32,
 876        [0, 0, 0], [tuint32, tuint32, tuint32], False, "", """
 877 unsigned base = src0;
 878 unsigned offset = src1 & 0x1F;
 879 unsigned bits = src2 & 0x1F;
 880 if (bits == 0) {
 881    dst = 0;
 882 } else if (offset + bits < 32) {
 883    dst = (base << (32 - bits - offset)) >> (32 - bits);
 884 } else {
 885    dst = base >> offset;
 886 }
 887 """)
 888 opcode("ibfe", 0, tint32,
 889        [0, 0, 0], [tint32, tuint32, tuint32], False, "", """
 890 int base = src0;
 891 unsigned offset = src1 & 0x1F;
 892 unsigned bits = src2 & 0x1F;
 893 if (bits == 0) {
 894    dst = 0;
 895 } else if (offset + bits < 32) {
 896    dst = (base << (32 - bits - offset)) >> (32 - bits);
 897 } else {
 898    dst = base >> offset;
 899 }
 900 """)
 901
 902 # GLSL bitfieldExtract()
 903 opcode("ubitfield_extract", 0, tuint32,
 904        [0, 0, 0], [tuint32, tint32, tint32], False, "", """
 905 unsigned base = src0;
 906 int offset = src1, bits = src2;
 907 if (bits == 0) {
 908    dst = 0;
 909 } else if (bits < 0 || offset < 0 || offset + bits > 32) {
 910    dst = 0; /* undefined per the spec */
 911 } else {
 912    dst = (base >> offset) & ((1ull << bits) - 1);
 913 }
 914 """)
 915 opcode("ibitfield_extract", 0, tint32,
 916        [0, 0, 0], [tint32, tint32, tint32], False, "", """
 917 int base = src0;
 918 int offset = src1, bits = src2;
 919 if (bits == 0) {
 920    dst = 0;
 921 } else if (offset < 0 || bits < 0 || offset + bits > 32) {
 922    dst = 0;
 923 } else {
 924    dst = (base << (32 - offset - bits)) >> offset; /* use sign-extending shift */
 925 }
 926 """)
 927
 928 # Combines the first component of each input to make a 3-component vector.
 929
 930 triop_horiz("vec3", 3, 1, 1, 1, """
 931 dst.x = src0.x;
 932 dst.y = src1.x;
 933 dst.z = src2.x;
 934 """)
 935
 936 def quadop_horiz(name, output_size, src1_size, src2_size, src3_size,
 937                  src4_size, const_expr):
 938    opcode(name, output_size, tuint,
 939           [src1_size, src2_size, src3_size, src4_size],
 940           [tuint, tuint, tuint, tuint],
 941           False, "", const_expr)
 942
 943 opcode("bitfield_insert", 0, tuint32, [0, 0, 0, 0],
 944        [tuint32, tuint32, tint32, tint32], False, "", """
 945 unsigned base = src0, insert = src1;
 946 int offset = src2, bits = src3;
 947 if (bits == 0) {
 948    dst = base;
 949 } else if (offset < 0 || bits < 0 || bits + offset > 32) {
 950    dst = 0;
 951 } else {
 952    unsigned mask = ((1ull << bits) - 1) << offset;
 953    dst = (base & ~mask) | ((insert << offset) & mask);
 954 }
 955 """)
 956
 957 quadop_horiz("vec4", 4, 1, 1, 1, 1, """
 958 dst.x = src0.x;
 959 dst.y = src1.x;
 960 dst.z = src2.x;
 961 dst.w = src3.x;
 962 """)
 963
 964 # ir3-specific instruction that maps directly to mul-add shift high mix,
 965 # (IMADSH_MIX16 i.e. ah * bl << 16 + c). It is used for lowering integer
 966 # multiplication (imul) on Freedreno backend..
 967 opcode("imadsh_mix16", 1, tint32,
 968        [1, 1, 1], [tint32, tint32, tint32], False, "", """
 969 dst.x = ((((src0.x & 0xffff0000) >> 16) * (src1.x & 0x0000ffff)) << 16) + src2.x;
 970 """)