2 # Copyright (C) 2014 Connor Abbott
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:
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
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
24 # Connor Abbott (cwabbott0@gmail.com)
28 # Class that represents all the information we have about the opcode
29 # NOTE: this must be kept in sync with nir_op_info
32 """Class that represents all the information we have about the opcode
33 NOTE: this must be kept in sync with nir_op_info
35 def __init__(self
, name
, output_size
, output_type
, input_sizes
,
36 input_types
, is_conversion
, algebraic_properties
, const_expr
):
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.
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.
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.
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 or (output_size
== 8) or (output_size
== 16)
79 for size
in input_sizes
:
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
93 # helper variables for strings
111 _TYPE_SPLIT_RE
= re
.compile(r
'(?P<type>int|uint|float|bool)(?P<bits>\d+)?')
113 def type_has_size(type_
):
114 m
= _TYPE_SPLIT_RE
.match(type_
)
115 assert m
is not None, 'Invalid NIR type string: "{}"'.format(type_
)
116 return m
.group('bits') is not None
118 def type_size(type_
):
119 m
= _TYPE_SPLIT_RE
.match(type_
)
120 assert m
is not None, 'Invalid NIR type string: "{}"'.format(type_
)
121 assert m
.group('bits') is not None, \
122 'NIR type string has no bit size: "{}"'.format(type_
)
123 return int(m
.group('bits'))
125 def type_sizes(type_
):
126 if type_has_size(type_
):
127 return [type_size(type_
)]
128 elif type_
== 'bool':
129 return [1, 8, 16, 32]
130 elif type_
== 'float':
133 return [1, 8, 16, 32, 64]
135 def type_base_type(type_
):
136 m
= _TYPE_SPLIT_RE
.match(type_
)
137 assert m
is not None, 'Invalid NIR type string: "{}"'.format(type_
)
138 return m
.group('type')
140 # Operation where the first two sources are commutative.
142 # For 2-source operations, this just mathematical commutativity. Some
143 # 3-source operations, like ffma, are only commutative in the first two
145 _2src_commutative
= "2src_commutative "
146 associative
= "associative "
148 # global dictionary of opcodes
151 def opcode(name
, output_size
, output_type
, input_sizes
, input_types
,
152 is_conversion
, algebraic_properties
, const_expr
):
153 assert name
not in opcodes
154 opcodes
[name
] = Opcode(name
, output_size
, output_type
, input_sizes
,
155 input_types
, is_conversion
, algebraic_properties
,
158 def unop_convert(name
, out_type
, in_type
, const_expr
):
159 opcode(name
, 0, out_type
, [0], [in_type
], False, "", const_expr
)
161 def unop(name
, ty
, const_expr
):
162 opcode(name
, 0, ty
, [0], [ty
], False, "", const_expr
)
164 def unop_horiz(name
, output_size
, output_type
, input_size
, input_type
,
166 opcode(name
, output_size
, output_type
, [input_size
], [input_type
],
167 False, "", const_expr
)
169 def unop_reduce(name
, output_size
, output_type
, input_type
, prereduce_expr
,
170 reduce_expr
, final_expr
):
172 return "(" + prereduce_expr
.format(src
=src
) + ")"
174 return final_expr
.format(src
="(" + src
+ ")")
175 def reduce_(src0
, src1
):
176 return reduce_expr
.format(src0
=src0
, src1
=src1
)
177 src0
= prereduce("src0.x")
178 src1
= prereduce("src0.y")
179 src2
= prereduce("src0.z")
180 src3
= prereduce("src0.w")
181 unop_horiz(name
+ "2", output_size
, output_type
, 2, input_type
,
182 final(reduce_(src0
, src1
)))
183 unop_horiz(name
+ "3", output_size
, output_type
, 3, input_type
,
184 final(reduce_(reduce_(src0
, src1
), src2
)))
185 unop_horiz(name
+ "4", output_size
, output_type
, 4, input_type
,
186 final(reduce_(reduce_(src0
, src1
), reduce_(src2
, src3
))))
188 def unop_numeric_convert(name
, out_type
, in_type
, const_expr
):
189 opcode(name
, 0, out_type
, [0], [in_type
], True, "", const_expr
)
191 unop("mov", tuint
, "src0")
193 unop("ineg", tint
, "-src0")
194 unop("fneg", tfloat
, "-src0")
195 unop("inot", tint
, "~src0") # invert every bit of the integer
196 unop("fsign", tfloat
, ("bit_size == 64 ? " +
197 "((src0 == 0.0) ? 0.0 : ((src0 > 0.0) ? 1.0 : -1.0)) : " +
198 "((src0 == 0.0f) ? 0.0f : ((src0 > 0.0f) ? 1.0f : -1.0f))"))
199 unop("isign", tint
, "(src0 == 0) ? 0 : ((src0 > 0) ? 1 : -1)")
200 unop("iabs", tint
, "(src0 < 0) ? -src0 : src0")
201 unop("fabs", tfloat
, "fabs(src0)")
202 unop("fsat", tfloat
, ("fmin(fmax(src0, 0.0), 1.0)"))
203 unop("frcp", tfloat
, "bit_size == 64 ? 1.0 / src0 : 1.0f / src0")
204 unop("frsq", tfloat
, "bit_size == 64 ? 1.0 / sqrt(src0) : 1.0f / sqrtf(src0)")
205 unop("fsqrt", tfloat
, "bit_size == 64 ? sqrt(src0) : sqrtf(src0)")
206 unop("fexp2", tfloat
, "exp2f(src0)")
207 unop("flog2", tfloat
, "log2f(src0)")
209 # Generate all of the numeric conversion opcodes
210 for src_t
in [tint
, tuint
, tfloat
, tbool
]:
212 dst_types
= [tfloat
, tint
, tbool
]
214 dst_types
= [tfloat
, tint
, tbool
]
216 dst_types
= [tfloat
, tuint
]
217 elif src_t
== tfloat
:
218 dst_types
= [tint
, tuint
, tfloat
, tbool
]
220 for dst_t
in dst_types
:
221 for dst_bit_size
in type_sizes(dst_t
):
222 if dst_bit_size
== 16 and dst_t
== tfloat
and src_t
== tfloat
:
223 rnd_modes
= ['_rtne', '_rtz', '']
224 for rnd_mode
in rnd_modes
:
225 if rnd_mode
== '_rtne':
228 dst = _mesa_half_to_float(_mesa_float_to_float16_rtne(src0));
233 elif rnd_mode
== '_rtz':
236 dst = _mesa_half_to_float(_mesa_float_to_float16_rtz(src0));
244 unop_numeric_convert("{0}2{1}{2}{3}".format(src_t
[0],
248 dst_t
+ str(dst_bit_size
),
250 elif dst_bit_size
== 32 and dst_t
== tfloat
and src_t
== tfloat
:
252 if (bit_size > 32 && nir_is_rounding_mode_rtz(execution_mode, 32)) {
253 dst = _mesa_double_to_float_rtz(src0);
258 unop_numeric_convert("{0}2{1}{2}".format(src_t
[0], dst_t
[0],
260 dst_t
+ str(dst_bit_size
), src_t
, conv_expr
)
262 conv_expr
= "src0 != 0" if dst_t
== tbool
else "src0"
263 unop_numeric_convert("{0}2{1}{2}".format(src_t
[0], dst_t
[0],
265 dst_t
+ str(dst_bit_size
), src_t
, conv_expr
)
267 # Special opcode that is the same as f2f16 except that it is safe to remove it
268 # if the result is immediately converted back to float32 again. This is
269 # generated as part of the precision lowering pass. mp stands for medium
271 unop_numeric_convert("f2fmp", tfloat16
, tfloat
, opcodes
["f2f16"].const_expr
)
273 # Unary floating-point rounding operations.
276 unop("ftrunc", tfloat
, "bit_size == 64 ? trunc(src0) : truncf(src0)")
277 unop("fceil", tfloat
, "bit_size == 64 ? ceil(src0) : ceilf(src0)")
278 unop("ffloor", tfloat
, "bit_size == 64 ? floor(src0) : floorf(src0)")
279 unop("ffract", tfloat
, "src0 - (bit_size == 64 ? floor(src0) : floorf(src0))")
280 unop("fround_even", tfloat
, "bit_size == 64 ? _mesa_roundeven(src0) : _mesa_roundevenf(src0)")
282 unop("fquantize2f16", tfloat
, "(fabs(src0) < ldexpf(1.0, -14)) ? copysignf(0.0f, src0) : _mesa_half_to_float(_mesa_float_to_half(src0))")
284 # Trigonometric operations.
287 unop("fsin", tfloat
, "bit_size == 64 ? sin(src0) : sinf(src0)")
288 unop("fcos", tfloat
, "bit_size == 64 ? cos(src0) : cosf(src0)")
291 unop_convert("frexp_exp", tint32
, tfloat
, "frexp(src0, &dst);")
292 unop_convert("frexp_sig", tfloat
, tfloat
, "int n; dst = frexp(src0, &n);")
294 # Partial derivatives.
297 unop("fddx", tfloat
, "0.0") # the derivative of a constant is 0.
298 unop("fddy", tfloat
, "0.0")
299 unop("fddx_fine", tfloat
, "0.0")
300 unop("fddy_fine", tfloat
, "0.0")
301 unop("fddx_coarse", tfloat
, "0.0")
302 unop("fddy_coarse", tfloat
, "0.0")
305 # Floating point pack and unpack operations.
308 unop_horiz("pack_" + fmt
+ "_2x16", 1, tuint32
, 2, tfloat32
, """
309 dst.x = (uint32_t) pack_fmt_1x16(src0.x);
310 dst.x |= ((uint32_t) pack_fmt_1x16(src0.y)) << 16;
311 """.replace("fmt", fmt
))
314 unop_horiz("pack_" + fmt
+ "_4x8", 1, tuint32
, 4, tfloat32
, """
315 dst.x = (uint32_t) pack_fmt_1x8(src0.x);
316 dst.x |= ((uint32_t) pack_fmt_1x8(src0.y)) << 8;
317 dst.x |= ((uint32_t) pack_fmt_1x8(src0.z)) << 16;
318 dst.x |= ((uint32_t) pack_fmt_1x8(src0.w)) << 24;
319 """.replace("fmt", fmt
))
321 def unpack_2x16(fmt
):
322 unop_horiz("unpack_" + fmt
+ "_2x16", 2, tfloat32
, 1, tuint32
, """
323 dst.x = unpack_fmt_1x16((uint16_t)(src0.x & 0xffff));
324 dst.y = unpack_fmt_1x16((uint16_t)(src0.x << 16));
325 """.replace("fmt", fmt
))
328 unop_horiz("unpack_" + fmt
+ "_4x8", 4, tfloat32
, 1, tuint32
, """
329 dst.x = unpack_fmt_1x8((uint8_t)(src0.x & 0xff));
330 dst.y = unpack_fmt_1x8((uint8_t)((src0.x >> 8) & 0xff));
331 dst.z = unpack_fmt_1x8((uint8_t)((src0.x >> 16) & 0xff));
332 dst.w = unpack_fmt_1x8((uint8_t)(src0.x >> 24));
333 """.replace("fmt", fmt
))
347 unop_horiz("pack_uvec2_to_uint", 1, tuint32
, 2, tuint32
, """
348 dst.x = (src0.x & 0xffff) | (src0.y << 16);
351 unop_horiz("pack_uvec4_to_uint", 1, tuint32
, 4, tuint32
, """
352 dst.x = (src0.x << 0) |
358 unop_horiz("pack_32_2x16", 1, tuint32
, 2, tuint16
,
359 "dst.x = src0.x | ((uint32_t)src0.y << 16);")
361 unop_horiz("pack_64_2x32", 1, tuint64
, 2, tuint32
,
362 "dst.x = src0.x | ((uint64_t)src0.y << 32);")
364 unop_horiz("pack_64_4x16", 1, tuint64
, 4, tuint16
,
365 "dst.x = src0.x | ((uint64_t)src0.y << 16) | ((uint64_t)src0.z << 32) | ((uint64_t)src0.w << 48);")
367 unop_horiz("unpack_64_2x32", 2, tuint32
, 1, tuint64
,
368 "dst.x = src0.x; dst.y = src0.x >> 32;")
370 unop_horiz("unpack_64_4x16", 4, tuint16
, 1, tuint64
,
371 "dst.x = src0.x; dst.y = src0.x >> 16; dst.z = src0.x >> 32; dst.w = src0.w >> 48;")
373 unop_horiz("unpack_32_2x16", 2, tuint16
, 1, tuint32
,
374 "dst.x = src0.x; dst.y = src0.x >> 16;")
376 unop_horiz("unpack_half_2x16_flush_to_zero", 2, tfloat32
, 1, tuint32
, """
377 dst.x = unpack_half_1x16_flush_to_zero((uint16_t)(src0.x & 0xffff));
378 dst.y = unpack_half_1x16_flush_to_zero((uint16_t)(src0.x << 16));
381 # Lowered floating point unpacking operations.
383 unop_convert("unpack_half_2x16_split_x", tfloat32
, tuint32
,
384 "unpack_half_1x16((uint16_t)(src0 & 0xffff))")
385 unop_convert("unpack_half_2x16_split_y", tfloat32
, tuint32
,
386 "unpack_half_1x16((uint16_t)(src0 >> 16))")
388 unop_convert("unpack_half_2x16_split_x_flush_to_zero", tfloat32
, tuint32
,
389 "unpack_half_1x16_flush_to_zero((uint16_t)(src0 & 0xffff))")
390 unop_convert("unpack_half_2x16_split_y_flush_to_zero", tfloat32
, tuint32
,
391 "unpack_half_1x16_flush_to_zero((uint16_t)(src0 >> 16))")
393 unop_convert("unpack_32_2x16_split_x", tuint16
, tuint32
, "src0")
394 unop_convert("unpack_32_2x16_split_y", tuint16
, tuint32
, "src0 >> 16")
396 unop_convert("unpack_64_2x32_split_x", tuint32
, tuint64
, "src0")
397 unop_convert("unpack_64_2x32_split_y", tuint32
, tuint64
, "src0 >> 32")
399 # Bit operations, part of ARB_gpu_shader5.
402 unop("bitfield_reverse", tuint32
, """
403 /* we're not winning any awards for speed here, but that's ok */
405 for (unsigned bit = 0; bit < 32; bit++)
406 dst |= ((src0 >> bit) & 1) << (31 - bit);
408 unop_convert("bit_count", tuint32
, tuint
, """
410 for (unsigned bit = 0; bit < bit_size; bit++) {
411 if ((src0 >> bit) & 1)
416 unop_convert("ufind_msb", tint32
, tuint
, """
418 for (int bit = bit_size - 1; bit >= 0; bit--) {
419 if ((src0 >> bit) & 1) {
426 unop("uclz", tuint32
, """
428 for (bit = bit_size - 1; bit >= 0; bit--) {
429 if ((src0 & (1u << bit)) != 0)
432 dst = (unsigned)(31 - bit);
435 unop("ifind_msb", tint32
, """
437 for (int bit = 31; bit >= 0; bit--) {
438 /* If src0 < 0, we're looking for the first 0 bit.
439 * if src0 >= 0, we're looking for the first 1 bit.
441 if ((((src0 >> bit) & 1) && (src0 >= 0)) ||
442 (!((src0 >> bit) & 1) && (src0 < 0))) {
449 unop_convert("find_lsb", tint32
, tint
, """
451 for (unsigned bit = 0; bit < bit_size; bit++) {
452 if ((src0 >> bit) & 1) {
459 # AMD_gcn_shader extended instructions
460 unop_horiz("cube_face_coord", 2, tfloat32
, 3, tfloat32
, """
462 float absX = fabsf(src0.x);
463 float absY = fabsf(src0.y);
464 float absZ = fabsf(src0.z);
467 if (absX >= absY && absX >= absZ) { ma = 2 * src0.x; }
468 if (absY >= absX && absY >= absZ) { ma = 2 * src0.y; }
469 if (absZ >= absX && absZ >= absY) { ma = 2 * src0.z; }
471 if (src0.x >= 0 && absX >= absY && absX >= absZ) { dst.x = -src0.z; dst.y = -src0.y; }
472 if (src0.x < 0 && absX >= absY && absX >= absZ) { dst.x = src0.z; dst.y = -src0.y; }
473 if (src0.y >= 0 && absY >= absX && absY >= absZ) { dst.x = src0.x; dst.y = src0.z; }
474 if (src0.y < 0 && absY >= absX && absY >= absZ) { dst.x = src0.x; dst.y = -src0.z; }
475 if (src0.z >= 0 && absZ >= absX && absZ >= absY) { dst.x = src0.x; dst.y = -src0.y; }
476 if (src0.z < 0 && absZ >= absX && absZ >= absY) { dst.x = -src0.x; dst.y = -src0.y; }
478 dst.x = dst.x / ma + 0.5;
479 dst.y = dst.y / ma + 0.5;
482 unop_horiz("cube_face_index", 1, tfloat32
, 3, tfloat32
, """
483 float absX = fabsf(src0.x);
484 float absY = fabsf(src0.y);
485 float absZ = fabsf(src0.z);
486 if (src0.x >= 0 && absX >= absY && absX >= absZ) dst.x = 0;
487 if (src0.x < 0 && absX >= absY && absX >= absZ) dst.x = 1;
488 if (src0.y >= 0 && absY >= absX && absY >= absZ) dst.x = 2;
489 if (src0.y < 0 && absY >= absX && absY >= absZ) dst.x = 3;
490 if (src0.z >= 0 && absZ >= absX && absZ >= absY) dst.x = 4;
491 if (src0.z < 0 && absZ >= absX && absZ >= absY) dst.x = 5;
494 # Sum of vector components
495 unop_reduce("fsum", 1, tfloat
, tfloat
, "{src}", "{src0} + {src1}", "{src}")
497 def binop_convert(name
, out_type
, in_type
, alg_props
, const_expr
):
498 opcode(name
, 0, out_type
, [0, 0], [in_type
, in_type
],
499 False, alg_props
, const_expr
)
501 def binop(name
, ty
, alg_props
, const_expr
):
502 binop_convert(name
, ty
, ty
, alg_props
, const_expr
)
504 def binop_compare(name
, ty
, alg_props
, const_expr
):
505 binop_convert(name
, tbool1
, ty
, alg_props
, const_expr
)
507 def binop_compare8(name
, ty
, alg_props
, const_expr
):
508 binop_convert(name
, tbool8
, ty
, alg_props
, const_expr
)
510 def binop_compare16(name
, ty
, alg_props
, const_expr
):
511 binop_convert(name
, tbool16
, ty
, alg_props
, const_expr
)
513 def binop_compare32(name
, ty
, alg_props
, const_expr
):
514 binop_convert(name
, tbool32
, ty
, alg_props
, const_expr
)
516 def binop_compare_all_sizes(name
, ty
, alg_props
, const_expr
):
517 binop_compare(name
, ty
, alg_props
, const_expr
)
518 binop_compare8(name
+ "8", ty
, alg_props
, const_expr
)
519 binop_compare16(name
+ "16", ty
, alg_props
, const_expr
)
520 binop_compare32(name
+ "32", ty
, alg_props
, const_expr
)
522 def binop_horiz(name
, out_size
, out_type
, src1_size
, src1_type
, src2_size
,
523 src2_type
, const_expr
):
524 opcode(name
, out_size
, out_type
, [src1_size
, src2_size
], [src1_type
, src2_type
],
525 False, "", const_expr
)
527 def binop_reduce(name
, output_size
, output_type
, src_type
, prereduce_expr
,
528 reduce_expr
, final_expr
):
530 return final_expr
.format(src
= "(" + src
+ ")")
531 def reduce_(src0
, src1
):
532 return reduce_expr
.format(src0
=src0
, src1
=src1
)
533 def prereduce(src0
, src1
):
534 return "(" + prereduce_expr
.format(src0
=src0
, src1
=src1
) + ")"
535 src0
= prereduce("src0.x", "src1.x")
536 src1
= prereduce("src0.y", "src1.y")
537 src2
= prereduce("src0.z", "src1.z")
538 src3
= prereduce("src0.w", "src1.w")
539 opcode(name
+ "2", output_size
, output_type
,
540 [2, 2], [src_type
, src_type
], False, _2src_commutative
,
541 final(reduce_(src0
, src1
)))
542 opcode(name
+ "3", output_size
, output_type
,
543 [3, 3], [src_type
, src_type
], False, _2src_commutative
,
544 final(reduce_(reduce_(src0
, src1
), src2
)))
545 opcode(name
+ "4", output_size
, output_type
,
546 [4, 4], [src_type
, src_type
], False, _2src_commutative
,
547 final(reduce_(reduce_(src0
, src1
), reduce_(src2
, src3
))))
549 def binop_reduce_all_sizes(name
, output_size
, src_type
, prereduce_expr
,
550 reduce_expr
, final_expr
):
551 binop_reduce(name
, output_size
, tbool1
, src_type
,
552 prereduce_expr
, reduce_expr
, final_expr
)
553 binop_reduce("b8" + name
[1:], output_size
, tbool8
, src_type
,
554 prereduce_expr
, reduce_expr
, final_expr
)
555 binop_reduce("b16" + name
[1:], output_size
, tbool16
, src_type
,
556 prereduce_expr
, reduce_expr
, final_expr
)
557 binop_reduce("b32" + name
[1:], output_size
, tbool32
, src_type
,
558 prereduce_expr
, reduce_expr
, final_expr
)
560 binop("fadd", tfloat
, _2src_commutative
+ associative
,"""
561 if (nir_is_rounding_mode_rtz(execution_mode, bit_size)) {
563 dst = _mesa_double_add_rtz(src0, src1);
565 dst = _mesa_double_to_float_rtz((double)src0 + (double)src1);
570 binop("iadd", tint
, _2src_commutative
+ associative
, "src0 + src1")
571 binop("iadd_sat", tint
, _2src_commutative
, """
573 (src0 + src1 < src0 ? (1ull << (bit_size - 1)) - 1 : src0 + src1) :
574 (src0 < src0 + src1 ? (1ull << (bit_size - 1)) : src0 + src1)
576 binop("uadd_sat", tuint
, _2src_commutative
,
577 "(src0 + src1) < src0 ? MAX_UINT_FOR_SIZE(sizeof(src0) * 8) : (src0 + src1)")
578 binop("isub_sat", tint
, "", """
580 (src0 - src1 < src0 ? (1ull << (bit_size - 1)) - 1 : src0 - src1) :
581 (src0 < src0 - src1 ? (1ull << (bit_size - 1)) : src0 - src1)
583 binop("usub_sat", tuint
, "", "src0 < src1 ? 0 : src0 - src1")
585 binop("fsub", tfloat
, "", """
586 if (nir_is_rounding_mode_rtz(execution_mode, bit_size)) {
588 dst = _mesa_double_sub_rtz(src0, src1);
590 dst = _mesa_double_to_float_rtz((double)src0 - (double)src1);
595 binop("isub", tint
, "", "src0 - src1")
596 binop_convert("uabs_isub", tuint
, tint
, "", """
597 src1 > src0 ? (uint64_t) src1 - (uint64_t) src0
598 : (uint64_t) src0 - (uint64_t) src1
600 binop("uabs_usub", tuint
, "", "(src1 > src0) ? (src1 - src0) : (src0 - src1)")
602 binop("fmul", tfloat
, _2src_commutative
+ associative
, """
603 if (nir_is_rounding_mode_rtz(execution_mode, bit_size)) {
605 dst = _mesa_double_mul_rtz(src0, src1);
607 dst = _mesa_double_to_float_rtz((double)src0 * (double)src1);
612 # low 32-bits of signed/unsigned integer multiply
613 binop("imul", tint
, _2src_commutative
+ associative
, "src0 * src1")
615 # Generate 64 bit result from 2 32 bits quantity
616 binop_convert("imul_2x32_64", tint64
, tint32
, _2src_commutative
,
617 "(int64_t)src0 * (int64_t)src1")
618 binop_convert("umul_2x32_64", tuint64
, tuint32
, _2src_commutative
,
619 "(uint64_t)src0 * (uint64_t)src1")
621 # high 32-bits of signed integer multiply
622 binop("imul_high", tint
, _2src_commutative
, """
623 if (bit_size == 64) {
624 /* We need to do a full 128-bit x 128-bit multiply in order for the sign
625 * extension to work properly. The casts are kind-of annoying but needed
626 * to prevent compiler warnings.
628 uint32_t src0_u32[4] = {
634 uint32_t src1_u32[4] = {
640 uint32_t prod_u32[4];
641 ubm_mul_u32arr(prod_u32, src0_u32, src1_u32);
642 dst = (uint64_t)prod_u32[2] | ((uint64_t)prod_u32[3] << 32);
644 dst = ((int64_t)src0 * (int64_t)src1) >> bit_size;
648 # high 32-bits of unsigned integer multiply
649 binop("umul_high", tuint
, _2src_commutative
, """
650 if (bit_size == 64) {
651 /* The casts are kind-of annoying but needed to prevent compiler warnings. */
652 uint32_t src0_u32[2] = { src0, (uint64_t)src0 >> 32 };
653 uint32_t src1_u32[2] = { src1, (uint64_t)src1 >> 32 };
654 uint32_t prod_u32[4];
655 ubm_mul_u32arr(prod_u32, src0_u32, src1_u32);
656 dst = (uint64_t)prod_u32[2] | ((uint64_t)prod_u32[3] << 32);
658 dst = ((uint64_t)src0 * (uint64_t)src1) >> bit_size;
662 # low 32-bits of unsigned integer multiply
663 binop("umul_low", tuint32
, _2src_commutative
, """
664 uint64_t mask = (1 << (bit_size / 2)) - 1;
665 dst = ((uint64_t)src0 & mask) * ((uint64_t)src1 & mask);
668 # Multiply 32-bits with low 16-bits.
669 binop("imul_32x16", tint32
, "", "src0 * (int16_t) src1")
670 binop("umul_32x16", tuint32
, "", "src0 * (uint16_t) src1")
672 binop("fdiv", tfloat
, "", "src0 / src1")
673 binop("idiv", tint
, "", "src1 == 0 ? 0 : (src0 / src1)")
674 binop("udiv", tuint
, "", "src1 == 0 ? 0 : (src0 / src1)")
676 # returns a boolean representing the carry resulting from the addition of
677 # the two unsigned arguments.
679 binop_convert("uadd_carry", tuint
, tuint
, _2src_commutative
, "src0 + src1 < src0")
681 # returns a boolean representing the borrow resulting from the subtraction
682 # of the two unsigned arguments.
684 binop_convert("usub_borrow", tuint
, tuint
, "", "src0 < src1")
686 # hadd: (a + b) >> 1 (without overflow)
687 # x + y = x - (x & ~y) + (x & ~y) + y - (~x & y) + (~x & y)
688 # = (x & y) + (x & ~y) + (x & y) + (~x & y)
689 # = 2 * (x & y) + (x & ~y) + (~x & y)
690 # = ((x & y) << 1) + (x ^ y)
692 # Since we know that the bottom bit of (x & y) << 1 is zero,
694 # (x + y) >> 1 = (((x & y) << 1) + (x ^ y)) >> 1
695 # = (x & y) + ((x ^ y) >> 1)
696 binop("ihadd", tint
, _2src_commutative
, "(src0 & src1) + ((src0 ^ src1) >> 1)")
697 binop("uhadd", tuint
, _2src_commutative
, "(src0 & src1) + ((src0 ^ src1) >> 1)")
699 # rhadd: (a + b + 1) >> 1 (without overflow)
700 # x + y + 1 = x + (~x & y) - (~x & y) + y + (x & ~y) - (x & ~y) + 1
701 # = (x | y) - (~x & y) + (x | y) - (x & ~y) + 1
702 # = 2 * (x | y) - ((~x & y) + (x & ~y)) + 1
703 # = ((x | y) << 1) - (x ^ y) + 1
705 # Since we know that the bottom bit of (x & y) << 1 is zero,
707 # (x + y + 1) >> 1 = (x | y) + (-(x ^ y) + 1) >> 1)
708 # = (x | y) - ((x ^ y) >> 1)
709 binop("irhadd", tint
, _2src_commutative
, "(src0 | src1) + ((src0 ^ src1) >> 1)")
710 binop("urhadd", tuint
, _2src_commutative
, "(src0 | src1) + ((src0 ^ src1) >> 1)")
712 binop("umod", tuint
, "", "src1 == 0 ? 0 : src0 % src1")
714 # For signed integers, there are several different possible definitions of
715 # "modulus" or "remainder". We follow the conventions used by LLVM and
716 # SPIR-V. The irem opcode implements the standard C/C++ signed "%"
717 # operation while the imod opcode implements the more mathematical
718 # "modulus" operation. For details on the difference, see
720 # http://mathforum.org/library/drmath/view/52343.html
722 binop("irem", tint
, "", "src1 == 0 ? 0 : src0 % src1")
723 binop("imod", tint
, "",
724 "src1 == 0 ? 0 : ((src0 % src1 == 0 || (src0 >= 0) == (src1 >= 0)) ?"
725 " src0 % src1 : src0 % src1 + src1)")
726 binop("fmod", tfloat
, "", "src0 - src1 * floorf(src0 / src1)")
727 binop("frem", tfloat
, "", "src0 - src1 * truncf(src0 / src1)")
734 # these integer-aware comparisons return a boolean (0 or ~0)
736 binop_compare_all_sizes("flt", tfloat
, "", "src0 < src1")
737 binop_compare_all_sizes("fge", tfloat
, "", "src0 >= src1")
738 binop_compare_all_sizes("feq", tfloat
, _2src_commutative
, "src0 == src1")
739 binop_compare_all_sizes("fne", tfloat
, _2src_commutative
, "src0 != src1")
740 binop_compare_all_sizes("ilt", tint
, "", "src0 < src1")
741 binop_compare_all_sizes("ige", tint
, "", "src0 >= src1")
742 binop_compare_all_sizes("ieq", tint
, _2src_commutative
, "src0 == src1")
743 binop_compare_all_sizes("ine", tint
, _2src_commutative
, "src0 != src1")
744 binop_compare_all_sizes("ult", tuint
, "", "src0 < src1")
745 binop_compare_all_sizes("uge", tuint
, "", "src0 >= src1")
747 # integer-aware GLSL-style comparisons that compare floats and ints
749 binop_reduce_all_sizes("ball_fequal", 1, tfloat
, "{src0} == {src1}",
750 "{src0} && {src1}", "{src}")
751 binop_reduce_all_sizes("bany_fnequal", 1, tfloat
, "{src0} != {src1}",
752 "{src0} || {src1}", "{src}")
753 binop_reduce_all_sizes("ball_iequal", 1, tint
, "{src0} == {src1}",
754 "{src0} && {src1}", "{src}")
755 binop_reduce_all_sizes("bany_inequal", 1, tint
, "{src0} != {src1}",
756 "{src0} || {src1}", "{src}")
758 # non-integer-aware GLSL-style comparisons that return 0.0 or 1.0
760 binop_reduce("fall_equal", 1, tfloat32
, tfloat32
, "{src0} == {src1}",
761 "{src0} && {src1}", "{src} ? 1.0f : 0.0f")
762 binop_reduce("fany_nequal", 1, tfloat32
, tfloat32
, "{src0} != {src1}",
763 "{src0} || {src1}", "{src} ? 1.0f : 0.0f")
765 # These comparisons for integer-less hardware return 1.0 and 0.0 for true
766 # and false respectively
768 binop("slt", tfloat32
, "", "(src0 < src1) ? 1.0f : 0.0f") # Set on Less Than
769 binop("sge", tfloat
, "", "(src0 >= src1) ? 1.0f : 0.0f") # Set on Greater or Equal
770 binop("seq", tfloat32
, _2src_commutative
, "(src0 == src1) ? 1.0f : 0.0f") # Set on Equal
771 binop("sne", tfloat32
, _2src_commutative
, "(src0 != src1) ? 1.0f : 0.0f") # Set on Not Equal
773 # SPIRV shifts are undefined for shift-operands >= bitsize,
774 # but SM5 shifts are defined to use the least significant bits, only
775 # The NIR definition is according to the SM5 specification.
776 opcode("ishl", 0, tint
, [0, 0], [tint
, tuint32
], False, "",
777 "src0 << (src1 & (sizeof(src0) * 8 - 1))")
778 opcode("ishr", 0, tint
, [0, 0], [tint
, tuint32
], False, "",
779 "src0 >> (src1 & (sizeof(src0) * 8 - 1))")
780 opcode("ushr", 0, tuint
, [0, 0], [tuint
, tuint32
], False, "",
781 "src0 >> (src1 & (sizeof(src0) * 8 - 1))")
783 opcode("urol", 0, tuint
, [0, 0], [tuint
, tuint32
], False, "", """
784 uint32_t rotate_mask = sizeof(src0) * 8 - 1;
785 dst = (src0 << (src1 & rotate_mask)) |
786 (src0 >> (-src1 & rotate_mask));
788 opcode("uror", 0, tuint
, [0, 0], [tuint
, tuint32
], False, "", """
789 uint32_t rotate_mask = sizeof(src0) * 8 - 1;
790 dst = (src0 >> (src1 & rotate_mask)) |
791 (src0 << (-src1 & rotate_mask));
794 # bitwise logic operators
796 # These are also used as boolean and, or, xor for hardware supporting
800 binop("iand", tuint
, _2src_commutative
+ associative
, "src0 & src1")
801 binop("ior", tuint
, _2src_commutative
+ associative
, "src0 | src1")
802 binop("ixor", tuint
, _2src_commutative
+ associative
, "src0 ^ src1")
805 binop_reduce("fdot", 1, tfloat
, tfloat
, "{src0} * {src1}", "{src0} + {src1}",
808 binop_reduce("fdot_replicated", 4, tfloat
, tfloat
,
809 "{src0} * {src1}", "{src0} + {src1}", "{src}")
811 opcode("fdph", 1, tfloat
, [3, 4], [tfloat
, tfloat
], False, "",
812 "src0.x * src1.x + src0.y * src1.y + src0.z * src1.z + src1.w")
813 opcode("fdph_replicated", 4, tfloat
, [3, 4], [tfloat
, tfloat
], False, "",
814 "src0.x * src1.x + src0.y * src1.y + src0.z * src1.z + src1.w")
816 binop("fmin", tfloat
, _2src_commutative
+ associative
, "fmin(src0, src1)")
817 binop("imin", tint
, _2src_commutative
+ associative
, "src1 > src0 ? src0 : src1")
818 binop("umin", tuint
, _2src_commutative
+ associative
, "src1 > src0 ? src0 : src1")
819 binop("fmax", tfloat
, _2src_commutative
+ associative
, "fmax(src0, src1)")
820 binop("imax", tint
, _2src_commutative
+ associative
, "src1 > src0 ? src1 : src0")
821 binop("umax", tuint
, _2src_commutative
+ associative
, "src1 > src0 ? src1 : src0")
823 # Saturated vector add for 4 8bit ints.
824 binop("usadd_4x8", tint32
, _2src_commutative
+ associative
, """
826 for (int i = 0; i < 32; i += 8) {
827 dst |= MIN2(((src0 >> i) & 0xff) + ((src1 >> i) & 0xff), 0xff) << i;
831 # Saturated vector subtract for 4 8bit ints.
832 binop("ussub_4x8", tint32
, "", """
834 for (int i = 0; i < 32; i += 8) {
835 int src0_chan = (src0 >> i) & 0xff;
836 int src1_chan = (src1 >> i) & 0xff;
837 if (src0_chan > src1_chan)
838 dst |= (src0_chan - src1_chan) << i;
842 # vector min for 4 8bit ints.
843 binop("umin_4x8", tint32
, _2src_commutative
+ associative
, """
845 for (int i = 0; i < 32; i += 8) {
846 dst |= MIN2((src0 >> i) & 0xff, (src1 >> i) & 0xff) << i;
850 # vector max for 4 8bit ints.
851 binop("umax_4x8", tint32
, _2src_commutative
+ associative
, """
853 for (int i = 0; i < 32; i += 8) {
854 dst |= MAX2((src0 >> i) & 0xff, (src1 >> i) & 0xff) << i;
858 # unorm multiply: (a * b) / 255.
859 binop("umul_unorm_4x8", tint32
, _2src_commutative
+ associative
, """
861 for (int i = 0; i < 32; i += 8) {
862 int src0_chan = (src0 >> i) & 0xff;
863 int src1_chan = (src1 >> i) & 0xff;
864 dst |= ((src0_chan * src1_chan) / 255) << i;
868 binop("fpow", tfloat
, "", "bit_size == 64 ? powf(src0, src1) : pow(src0, src1)")
870 binop_horiz("pack_half_2x16_split", 1, tuint32
, 1, tfloat32
, 1, tfloat32
,
871 "pack_half_1x16(src0.x) | (pack_half_1x16(src1.x) << 16)")
873 binop_convert("pack_64_2x32_split", tuint64
, tuint32
, "",
874 "src0 | ((uint64_t)src1 << 32)")
876 binop_convert("pack_32_2x16_split", tuint32
, tuint16
, "",
877 "src0 | ((uint32_t)src1 << 16)")
879 # bfm implements the behavior of the first operation of the SM5 "bfi" assembly
880 # and that of the "bfi1" i965 instruction. That is, the bits and offset values
881 # are from the low five bits of src0 and src1, respectively.
882 binop_convert("bfm", tuint32
, tint32
, "", """
883 int bits = src0 & 0x1F;
884 int offset = src1 & 0x1F;
885 dst = ((1u << bits) - 1) << offset;
888 opcode("ldexp", 0, tfloat
, [0, 0], [tfloat
, tint32
], False, "", """
889 dst = (bit_size == 64) ? ldexp(src0, src1) : ldexpf(src0, src1);
890 /* flush denormals to zero. */
892 dst = copysignf(0.0f, src0);
895 # Combines the first component of each input to make a 2-component vector.
897 binop_horiz("vec2", 2, tuint
, 1, tuint
, 1, tuint
, """
903 binop("extract_u8", tuint
, "", "(uint8_t)(src0 >> (src1 * 8))")
904 binop("extract_i8", tint
, "", "(int8_t)(src0 >> (src1 * 8))")
907 binop("extract_u16", tuint
, "", "(uint16_t)(src0 >> (src1 * 16))")
908 binop("extract_i16", tint
, "", "(int16_t)(src0 >> (src1 * 16))")
911 def triop(name
, ty
, alg_props
, const_expr
):
912 opcode(name
, 0, ty
, [0, 0, 0], [ty
, ty
, ty
], False, alg_props
, const_expr
)
913 def triop_horiz(name
, output_size
, src1_size
, src2_size
, src3_size
, const_expr
):
914 opcode(name
, output_size
, tuint
,
915 [src1_size
, src2_size
, src3_size
],
916 [tuint
, tuint
, tuint
], False, "", const_expr
)
918 triop("ffma", tfloat
, _2src_commutative
, """
919 if (nir_is_rounding_mode_rtz(execution_mode, bit_size)) {
921 dst = _mesa_double_fma_rtz(src0, src1, src2);
922 else if (bit_size == 32)
923 dst = _mesa_float_fma_rtz(src0, src1, src2);
925 dst = _mesa_double_to_float_rtz(_mesa_double_fma_rtz(src0, src1, src2));
928 dst = fmaf(src0, src1, src2);
930 dst = fma(src0, src1, src2);
934 triop("flrp", tfloat
, "", "src0 * (1 - src2) + src1 * src2")
938 # A vector conditional select instruction (like ?:, but operating per-
939 # component on vectors). There are two versions, one for floating point
940 # bools (0.0 vs 1.0) and one for integer bools (0 vs ~0).
943 triop("fcsel", tfloat32
, "", "(src0 != 0.0f) ? src1 : src2")
946 triop("fmin3", tfloat
, "", "fminf(src0, fminf(src1, src2))")
947 triop("imin3", tint
, "", "MIN2(src0, MIN2(src1, src2))")
948 triop("umin3", tuint
, "", "MIN2(src0, MIN2(src1, src2))")
950 triop("fmax3", tfloat
, "", "fmaxf(src0, fmaxf(src1, src2))")
951 triop("imax3", tint
, "", "MAX2(src0, MAX2(src1, src2))")
952 triop("umax3", tuint
, "", "MAX2(src0, MAX2(src1, src2))")
954 triop("fmed3", tfloat
, "", "fmaxf(fminf(fmaxf(src0, src1), src2), fminf(src0, src1))")
955 triop("imed3", tint
, "", "MAX2(MIN2(MAX2(src0, src1), src2), MIN2(src0, src1))")
956 triop("umed3", tuint
, "", "MAX2(MIN2(MAX2(src0, src1), src2), MIN2(src0, src1))")
958 opcode("bcsel", 0, tuint
, [0, 0, 0],
959 [tbool1
, tuint
, tuint
], False, "", "src0 ? src1 : src2")
960 opcode("b8csel", 0, tuint
, [0, 0, 0],
961 [tbool8
, tuint
, tuint
], False, "", "src0 ? src1 : src2")
962 opcode("b16csel", 0, tuint
, [0, 0, 0],
963 [tbool16
, tuint
, tuint
], False, "", "src0 ? src1 : src2")
964 opcode("b32csel", 0, tuint
, [0, 0, 0],
965 [tbool32
, tuint
, tuint
], False, "", "src0 ? src1 : src2")
968 triop("bfi", tuint32
, "", """
969 unsigned mask = src0, insert = src1, base = src2;
978 dst = (base & ~mask) | (insert & mask);
983 triop("bitfield_select", tuint
, "", "(src0 & src1) | (~src0 & src2)")
985 # SM5 ubfe/ibfe assembly: only the 5 least significant bits of offset and bits are used.
986 opcode("ubfe", 0, tuint32
,
987 [0, 0, 0], [tuint32
, tuint32
, tuint32
], False, "", """
988 unsigned base = src0;
989 unsigned offset = src1 & 0x1F;
990 unsigned bits = src2 & 0x1F;
993 } else if (offset + bits < 32) {
994 dst = (base << (32 - bits - offset)) >> (32 - bits);
996 dst = base >> offset;
999 opcode("ibfe", 0, tint32
,
1000 [0, 0, 0], [tint32
, tuint32
, tuint32
], False, "", """
1002 unsigned offset = src1 & 0x1F;
1003 unsigned bits = src2 & 0x1F;
1006 } else if (offset + bits < 32) {
1007 dst = (base << (32 - bits - offset)) >> (32 - bits);
1009 dst = base >> offset;
1013 # GLSL bitfieldExtract()
1014 opcode("ubitfield_extract", 0, tuint32
,
1015 [0, 0, 0], [tuint32
, tint32
, tint32
], False, "", """
1016 unsigned base = src0;
1017 int offset = src1, bits = src2;
1020 } else if (bits < 0 || offset < 0 || offset + bits > 32) {
1021 dst = 0; /* undefined per the spec */
1023 dst = (base >> offset) & ((1ull << bits) - 1);
1026 opcode("ibitfield_extract", 0, tint32
,
1027 [0, 0, 0], [tint32
, tint32
, tint32
], False, "", """
1029 int offset = src1, bits = src2;
1032 } else if (offset < 0 || bits < 0 || offset + bits > 32) {
1035 dst = (base << (32 - offset - bits)) >> offset; /* use sign-extending shift */
1039 # Combines the first component of each input to make a 3-component vector.
1041 triop_horiz("vec3", 3, 1, 1, 1, """
1047 def quadop_horiz(name
, output_size
, src1_size
, src2_size
, src3_size
,
1048 src4_size
, const_expr
):
1049 opcode(name
, output_size
, tuint
,
1050 [src1_size
, src2_size
, src3_size
, src4_size
],
1051 [tuint
, tuint
, tuint
, tuint
],
1052 False, "", const_expr
)
1054 opcode("bitfield_insert", 0, tuint32
, [0, 0, 0, 0],
1055 [tuint32
, tuint32
, tint32
, tint32
], False, "", """
1056 unsigned base = src0, insert = src1;
1057 int offset = src2, bits = src3;
1060 } else if (offset < 0 || bits < 0 || bits + offset > 32) {
1063 unsigned mask = ((1ull << bits) - 1) << offset;
1064 dst = (base & ~mask) | ((insert << offset) & mask);
1068 quadop_horiz("vec4", 4, 1, 1, 1, 1, """
1075 opcode("vec8", 8, tuint
,
1076 [1] * 8, [tuint
] * 8,
1088 opcode("vec16", 16, tuint
,
1089 [1] * 16, [tuint
] * 16,
1109 # An integer multiply instruction for address calculation. This is
1110 # similar to imul, except that the results are undefined in case of
1111 # overflow. Overflow is defined according to the size of the variable
1112 # being dereferenced.
1114 # This relaxed definition, compared to imul, allows an optimization
1115 # pass to propagate bounds (ie, from an load/store intrinsic) to the
1116 # sources, such that lower precision integer multiplies can be used.
1117 # This is useful on hw that has 24b or perhaps 16b integer multiply
1119 binop("amul", tint
, _2src_commutative
+ associative
, "src0 * src1")
1121 # ir3-specific instruction that maps directly to mul-add shift high mix,
1122 # (IMADSH_MIX16 i.e. ah * bl << 16 + c). It is used for lowering integer
1123 # multiplication (imul) on Freedreno backend..
1124 opcode("imadsh_mix16", 0, tint32
,
1125 [0, 0, 0], [tint32
, tint32
, tint32
], False, "", """
1126 dst = ((((src0 & 0xffff0000) >> 16) * (src1 & 0x0000ffff)) << 16) + src2;
1129 # ir3-specific instruction that maps directly to ir3 mad.s24.
1131 # 24b multiply into 32b result (with sign extension) plus 32b int
1132 triop("imad24_ir3", tint32
, _2src_commutative
,
1133 "(((int32_t)src0 << 8) >> 8) * (((int32_t)src1 << 8) >> 8) + src2")
1135 # 24b multiply into 32b result (with sign extension)
1136 binop("imul24", tint32
, _2src_commutative
+ associative
,
1137 "(((int32_t)src0 << 8) >> 8) * (((int32_t)src1 << 8) >> 8)")
1139 # unsigned 24b multiply into 32b result plus 32b int
1140 triop("umad24", tuint32
, _2src_commutative
,
1141 "(((uint32_t)src0 << 8) >> 8) * (((uint32_t)src1 << 8) >> 8) + src2")
1143 # unsigned 24b multiply into 32b result uint
1144 binop("umul24", tint32
, _2src_commutative
+ associative
,
1145 "(((uint32_t)src0 << 8) >> 8) * (((uint32_t)src1 << 8) >> 8)")