3 # Copyright (C) 2014 Connor Abbott
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:
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
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
25 # 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
, 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 - 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.
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.
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.
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
:
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
90 # helper variables for strings
94 tunsigned
= "unsigned"
96 commutative
= "commutative "
97 associative
= "associative "
99 # global dictionary of opcodes
102 def opcode(name
, output_size
, output_type
, input_sizes
, input_types
,
103 algebraic_properties
, const_expr
):
104 assert name
not in opcodes
105 opcodes
[name
] = Opcode(name
, output_size
, output_type
, input_sizes
,
106 input_types
, algebraic_properties
, const_expr
)
108 def unop_convert(name
, in_type
, out_type
, const_expr
):
109 opcode(name
, 0, out_type
, [0], [in_type
], "", const_expr
)
111 def unop(name
, ty
, const_expr
):
112 opcode(name
, 0, ty
, [0], [ty
], "", const_expr
)
114 def unop_horiz(name
, output_size
, output_type
, input_size
, input_type
,
116 opcode(name
, output_size
, output_type
, [input_size
], [input_type
], "",
119 def unop_reduce(name
, output_size
, output_type
, input_type
, prereduce_expr
,
120 reduce_expr
, final_expr
):
122 return "(" + prereduce_expr
.format(src
=src
) + ")"
124 return final_expr
.format(src
="(" + src
+ ")")
125 def reduce_(src0
, src1
):
126 return reduce_expr
.format(src0
=src0
, src1
=src1
)
127 src0
= prereduce("src0.x")
128 src1
= prereduce("src0.y")
129 src2
= prereduce("src0.z")
130 src3
= prereduce("src0.w")
131 unop_horiz(name
+ "2", output_size
, output_type
, 2, input_type
,
132 final(reduce_(src0
, src1
)))
133 unop_horiz(name
+ "3", output_size
, output_type
, 3, input_type
,
134 final(reduce_(reduce_(src0
, src1
), src2
)))
135 unop_horiz(name
+ "4", output_size
, output_type
, 4, input_type
,
136 final(reduce_(reduce_(src0
, src1
), reduce_(src2
, src3
))))
139 # These two move instructions differ in what modifiers they support and what
140 # the negate modifier means. Otherwise, they are identical.
141 unop("fmov", tfloat
, "src0")
142 unop("imov", tint
, "src0")
144 unop("ineg", tint
, "-src0")
145 unop("fneg", tfloat
, "-src0")
146 unop("inot", tint
, "~src0") # invert every bit of the integer
147 unop("fnot", tfloat
, "(src0 == 0.0f) ? 1.0f : 0.0f")
148 unop("fsign", tfloat
, "(src0 == 0.0f) ? 0.0f : ((src0 > 0.0f) ? 1.0f : -1.0f)")
149 unop("isign", tint
, "(src0 == 0) ? 0 : ((src0 > 0) ? 1 : -1)")
150 unop("iabs", tint
, "(src0 < 0) ? -src0 : src0")
151 unop("fabs", tfloat
, "fabsf(src0)")
152 unop("fsat", tfloat
, "(src0 > 1.0f) ? 1.0f : ((src0 <= 0.0f) ? 0.0f : src0)")
153 unop("frcp", tfloat
, "1.0f / src0")
154 unop("frsq", tfloat
, "1.0f / sqrtf(src0)")
155 unop("fsqrt", tfloat
, "sqrtf(src0)")
156 unop("fexp", tfloat
, "expf(src0)") # < e^x
157 unop("flog", tfloat
, "logf(src0)") # log base e
158 unop("fexp2", tfloat
, "exp2f(src0)")
159 unop("flog2", tfloat
, "log2f(src0)")
160 unop_convert("f2i", tfloat
, tint
, "src0") # Float-to-integer conversion.
161 unop_convert("f2u", tfloat
, tunsigned
, "src0") # Float-to-unsigned conversion
162 unop_convert("i2f", tint
, tfloat
, "src0") # Integer-to-float conversion.
163 # Float-to-boolean conversion
164 unop_convert("f2b", tfloat
, tbool
, "src0 == 0.0f")
165 # Boolean-to-float conversion
166 unop_convert("b2f", tbool
, tfloat
, "src0 ? 1.0f : 0.0f")
167 # Int-to-boolean conversion
168 unop_convert("i2b", tint
, tbool
, "src0 == 0")
169 unop_convert("b2i", tbool
, tint
, "src0 ? 0 : -1") # Boolean-to-int conversion
170 unop_convert("u2f", tunsigned
, tfloat
, "src0") #Unsigned-to-float conversion.
172 unop_reduce("bany", 1, tbool
, tbool
, "{src}", "{src0} || {src1}", "{src}")
173 unop_reduce("ball", 1, tbool
, tbool
, "{src}", "{src0} && {src1}", "{src}")
174 unop_reduce("fany", 1, tfloat
, tfloat
, "{src} != 0.0f", "{src0} || {src1}",
175 "{src} ? 1.0f : 0.0f")
176 unop_reduce("fall", 1, tfloat
, tfloat
, "{src} != 0.0f", "{src0} && {src1}",
177 "{src} ? 1.0f : 0.0f")
179 # Unary floating-point rounding operations.
182 unop("ftrunc", tfloat
, "truncf(src0)")
183 unop("fceil", tfloat
, "ceilf(src0)")
184 unop("ffloor", tfloat
, "floorf(src0)")
185 unop("ffract", tfloat
, "src0 - floorf(src0)")
186 unop("fround_even", tfloat
, "_mesa_round_to_even(src0)")
189 # Trigonometric operations.
192 unop("fsin", tfloat
, "sinf(src0)")
193 unop("fcos", tfloat
, "cosf(src0)")
194 unop("fsin_reduced", tfloat
, "sinf(src0)")
195 unop("fcos_reduced", tfloat
, "cosf(src0)")
198 # Partial derivatives.
201 unop("fddx", tfloat
, "0.0f") # the derivative of a constant is 0.
202 unop("fddy", tfloat
, "0.0f")
203 unop("fddx_fine", tfloat
, "0.0f")
204 unop("fddy_fine", tfloat
, "0.0f")
205 unop("fddx_coarse", tfloat
, "0.0f")
206 unop("fddy_coarse", tfloat
, "0.0f")
209 # Floating point pack and unpack operations.
212 unop_horiz("pack_" + fmt
+ "_2x16", 1, tunsigned
, 2, tfloat
, """
213 dst.x = (uint32_t) pack_fmt_1x16(src0.x);
214 dst.x |= ((uint32_t) pack_fmt_1x16(src0.y)) << 16;
215 """.replace("fmt", fmt
))
218 unop_horiz("pack_" + fmt
+ "_4x8", 1, tunsigned
, 4, tfloat
, """
219 dst.x = (uint32_t) pack_fmt_1x8(src0.x);
220 dst.x |= ((uint32_t) pack_fmt_1x8(src0.y)) << 8;
221 dst.x |= ((uint32_t) pack_fmt_1x8(src0.z)) << 16;
222 dst.x |= ((uint32_t) pack_fmt_1x8(src0.w)) << 24;
223 """.replace("fmt", fmt
))
225 def unpack_2x16(fmt
):
226 unop_horiz("unpack_" + fmt
+ "_2x16", 2, tfloat
, 1, tunsigned
, """
227 dst.x = unpack_fmt_1x16((uint16_t)(src0.x & 0xffff));
228 dst.y = unpack_fmt_1x16((uint16_t)(src0.x << 16));
229 """.replace("fmt", fmt
))
232 unop_horiz("unpack_" + fmt
+ "_4x8", 4, tfloat
, 1, tunsigned
, """
233 dst.x = unpack_fmt_1x8((uint8_t)(src0.x & 0xff));
234 dst.y = unpack_fmt_1x8((uint8_t)((src0.x >> 8) & 0xff));
235 dst.z = unpack_fmt_1x8((uint8_t)((src0.x >> 16) & 0xff));
236 dst.w = unpack_fmt_1x8((uint8_t)(src0.x >> 24));
237 """.replace("fmt", fmt
))
252 # Lowered floating point unpacking operations.
255 unop_horiz("unpack_half_2x16_split_x", 1, tfloat
, 1, tunsigned
,
256 "unpack_half_1x16((uint16_t)(src0.x & 0xffff))")
257 unop_horiz("unpack_half_2x16_split_y", 1, tfloat
, 1, tunsigned
,
258 "unpack_half_1x16((uint16_t)(src0.x >> 16))")
261 # Bit operations, part of ARB_gpu_shader5.
264 unop("bitfield_reverse", tunsigned
, """
265 /* we're not winning any awards for speed here, but that's ok */
267 for (unsigned bit = 0; bit < 32; bit++)
268 dst |= ((src0 >> bit) & 1) << (31 - bit);
270 unop("bit_count", tunsigned
, """
272 for (unsigned bit = 0; bit < 32; bit++) {
273 if ((src0 >> bit) & 1)
278 unop_convert("ufind_msb", tunsigned
, tint
, """
280 for (int bit = 31; bit > 0; bit--) {
281 if ((src0 >> bit) & 1) {
288 unop("ifind_msb", tint
, """
290 for (int bit = 31; bit >= 0; bit--) {
291 /* If src0 < 0, we're looking for the first 0 bit.
292 * if src0 >= 0, we're looking for the first 1 bit.
294 if ((((src0 >> bit) & 1) && (src0 >= 0)) ||
295 (!((src0 >> bit) & 1) && (src0 < 0))) {
302 unop("find_lsb", tint
, """
304 for (unsigned bit = 0; bit < 32; bit++) {
305 if ((src0 >> bit) & 1) {
313 for i
in xrange(1, 5):
314 for j
in xrange(1, 5):
315 unop_horiz("fnoise{0}_{1}".format(i
, j
), i
, tfloat
, j
, tfloat
, "0.0f")
317 def binop_convert(name
, out_type
, in_type
, alg_props
, const_expr
):
318 opcode(name
, 0, out_type
, [0, 0], [in_type
, in_type
], alg_props
, const_expr
)
320 def binop(name
, ty
, alg_props
, const_expr
):
321 binop_convert(name
, ty
, ty
, alg_props
, const_expr
)
323 def binop_compare(name
, ty
, alg_props
, const_expr
):
324 binop_convert(name
, tbool
, ty
, alg_props
, const_expr
)
326 def binop_horiz(name
, out_size
, out_type
, src1_size
, src1_type
, src2_size
,
327 src2_type
, const_expr
):
328 opcode(name
, out_size
, out_type
, [src1_size
, src2_size
], [src1_type
, src2_type
],
331 def binop_reduce(name
, output_size
, output_type
, src_type
, prereduce_expr
,
332 reduce_expr
, final_expr
):
334 return final_expr
.format(src
= "(" + src
+ ")")
335 def reduce_(src0
, src1
):
336 return reduce_expr
.format(src0
=src0
, src1
=src1
)
337 def prereduce(src0
, src1
):
338 return "(" + prereduce_expr
.format(src0
=src0
, src1
=src1
) + ")"
339 src0
= prereduce("src0.x", "src1.x")
340 src1
= prereduce("src0.y", "src1.y")
341 src2
= prereduce("src0.z", "src1.z")
342 src3
= prereduce("src0.w", "src1.w")
343 opcode(name
+ "2", output_size
, output_type
,
344 [2, 2], [src_type
, src_type
], commutative
,
345 final(reduce_(src0
, src1
)))
346 opcode(name
+ "3", output_size
, output_type
,
347 [3, 3], [src_type
, src_type
], commutative
,
348 final(reduce_(reduce_(src0
, src1
), src2
)))
349 opcode(name
+ "4", output_size
, output_type
,
350 [4, 4], [src_type
, src_type
], commutative
,
351 final(reduce_(reduce_(src0
, src1
), reduce_(src2
, src3
))))
353 binop("fadd", tfloat
, commutative
+ associative
, "src0 + src1")
354 binop("iadd", tint
, commutative
+ associative
, "src0 + src1")
355 binop("fsub", tfloat
, "", "src0 - src1")
356 binop("isub", tint
, "", "src0 - src1")
358 binop("fmul", tfloat
, commutative
+ associative
, "src0 * src1")
359 # low 32-bits of signed/unsigned integer multiply
360 binop("imul", tint
, commutative
+ associative
, "src0 * src1")
361 # high 32-bits of signed integer multiply
362 binop("imul_high", tint
, commutative
,
363 "(int32_t)(((int64_t) src0 * (int64_t) src1) >> 32)")
364 # high 32-bits of unsigned integer multiply
365 binop("umul_high", tunsigned
, commutative
,
366 "(uint32_t)(((uint64_t) src0 * (uint64_t) src1) >> 32)")
368 binop("fdiv", tfloat
, "", "src0 / src1")
369 binop("idiv", tint
, "", "src0 / src1")
370 binop("udiv", tunsigned
, "", "src0 / src1")
372 # returns a boolean representing the carry resulting from the addition of
373 # the two unsigned arguments.
375 binop_convert("uadd_carry", tbool
, tunsigned
, commutative
, "src0 + src1 < src0")
377 # returns a boolean representing the borrow resulting from the subtraction
378 # of the two unsigned arguments.
380 binop_convert("usub_borrow", tbool
, tunsigned
, "", "src1 < src0")
382 binop("fmod", tfloat
, "", "src0 - src1 * floorf(src0 / src1)")
383 binop("umod", tunsigned
, "", "src1 == 0 ? 0 : src0 % src1")
390 # these integer-aware comparisons return a boolean (0 or ~0)
392 binop_compare("flt", tfloat
, "", "src0 < src1")
393 binop_compare("fge", tfloat
, "", "src0 >= src1")
394 binop_compare("feq", tfloat
, commutative
, "src0 == src1")
395 binop_compare("fne", tfloat
, commutative
, "src0 != src1")
396 binop_compare("ilt", tint
, "", "src0 < src1")
397 binop_compare("ige", tint
, "", "src0 >= src1")
398 binop_compare("ieq", tint
, commutative
, "src0 == src1")
399 binop_compare("ine", tint
, commutative
, "src0 != src1")
400 binop_compare("ult", tunsigned
, "", "src0 < src1")
401 binop_compare("uge", tunsigned
, "", "src0 >= src1")
403 # integer-aware GLSL-style comparisons that compare floats and ints
405 binop_reduce("ball_fequal", 1, tbool
, tfloat
, "{src0} == {src1}",
406 "{src0} && {src1}", "{src}")
407 binop_reduce("bany_fnequal", 1, tbool
, tfloat
, "{src0} != {src1}",
408 "{src0} || {src1}", "{src}")
409 binop_reduce("ball_iequal", 1, tbool
, tint
, "{src0} == {src1}",
410 "{src0} && {src1}", "{src}")
411 binop_reduce("bany_inequal", 1, tbool
, tint
, "{src0} != {src1}",
412 "{src0} || {src1}", "{src}")
414 # non-integer-aware GLSL-style comparisons that return 0.0 or 1.0
416 binop_reduce("fall_equal", 1, tfloat
, tfloat
, "{src0} == {src1}",
417 "{src0} && {src1}", "{src} ? 1.0f : 0.0f")
418 binop_reduce("fany_nequal", 1, tfloat
, tfloat
, "{src0} != {src1}",
419 "{src0} || {src1}", "{src} ? 1.0f : 0.0f")
421 # These comparisons for integer-less hardware return 1.0 and 0.0 for true
422 # and false respectively
424 binop("slt", tfloat
, "", "(src0 < src1) ? 1.0f : 0.0f") # Set on Less Than
425 binop("sge", tfloat
, "", "(src0 >= src1) ? 1.0f : 0.0f") # Set on Greater or Equal
426 binop("seq", tfloat
, commutative
, "(src0 == src1) ? 1.0f : 0.0f") # Set on Equal
427 binop("sne", tfloat
, commutative
, "(src0 != src1) ? 1.0f : 0.0f") # Set on Not Equal
430 binop("ishl", tint
, "", "src0 << src1")
431 binop("ishr", tint
, "", "src0 >> src1")
432 binop("ushr", tunsigned
, "", "src0 >> src1")
434 # bitwise logic operators
436 # These are also used as boolean and, or, xor for hardware supporting
440 binop("iand", tunsigned
, commutative
+ associative
, "src0 & src1")
441 binop("ior", tunsigned
, commutative
+ associative
, "src0 | src1")
442 binop("ixor", tunsigned
, commutative
+ associative
, "src0 ^ src1")
445 # floating point logic operators
447 # These use (src != 0.0) for testing the truth of the input, and output 1.0
448 # for true and 0.0 for false
450 binop("fand", tfloat
, commutative
,
451 "((src0 != 0.0f) && (src1 != 0.0f)) ? 1.0f : 0.0f")
452 binop("for", tfloat
, commutative
,
453 "((src0 != 0.0f) || (src1 != 0.0f)) ? 1.0f : 0.0f")
454 binop("fxor", tfloat
, commutative
,
455 "(src0 != 0.0f && src1 == 0.0f) || (src0 == 0.0f && src1 != 0.0f) ? 1.0f : 0.0f")
457 binop_reduce("fdot", 1, tfloat
, tfloat
, "{src0} * {src1}", "{src0} + {src1}",
460 binop("fmin", tfloat
, "", "fminf(src0, src1)")
461 binop("imin", tint
, commutative
+ associative
, "src1 > src0 ? src0 : src1")
462 binop("umin", tunsigned
, commutative
+ associative
, "src1 > src0 ? src0 : src1")
463 binop("fmax", tfloat
, "", "fmaxf(src0, src1)")
464 binop("imax", tint
, commutative
+ associative
, "src1 > src0 ? src1 : src0")
465 binop("umax", tunsigned
, commutative
+ associative
, "src1 > src0 ? src1 : src0")
467 binop("fpow", tfloat
, "", "powf(src0, src1)")
469 binop_horiz("pack_half_2x16_split", 1, tunsigned
, 1, tfloat
, 1, tfloat
,
470 "pack_half_1x16(src0.x) | (pack_half_1x16(src1.x) << 16)")
472 binop_convert("bfm", tunsigned
, tint
, "", """
473 int offset = src0, bits = src1;
474 if (offset < 0 || bits < 0 || offset + bits > 32)
475 dst = 0; /* undefined per the spec */
477 dst = ((1 << bits)- 1) << offset;
480 opcode("ldexp", 0, tunsigned
, [0, 0], [tfloat
, tint
], "", """
481 dst = ldexp(src0, src1);
482 /* flush denormals to zero. */
484 dst = copysign(0.0f, src0);
487 # Combines the first component of each input to make a 2-component vector.
489 binop_horiz("vec2", 2, tunsigned
, 1, tunsigned
, 1, tunsigned
, """
494 def triop(name
, ty
, const_expr
):
495 opcode(name
, 0, ty
, [0, 0, 0], [ty
, ty
, ty
], "", const_expr
)
496 def triop_horiz(name
, output_size
, src1_size
, src2_size
, src3_size
, const_expr
):
497 opcode(name
, output_size
, tunsigned
,
498 [src1_size
, src2_size
, src3_size
],
499 [tunsigned
, tunsigned
, tunsigned
], "", const_expr
)
501 triop("ffma", tfloat
, "src0 * src1 + src2")
503 triop("flrp", tfloat
, "src0 * (1 - src2) + src1 * src2")
507 # A vector conditional select instruction (like ?:, but operating per-
508 # component on vectors). There are two versions, one for floating point
509 # bools (0.0 vs 1.0) and one for integer bools (0 vs ~0).
512 triop("fcsel", tfloat
, "(src0 != 0.0f) ? src1 : src2")
513 opcode("bcsel", 0, tunsigned
, [0, 0, 0],
514 [tbool
, tunsigned
, tunsigned
], "", "src0 ? src1 : src2")
516 triop("bfi", tunsigned
, """
517 unsigned mask = src0, insert = src1 & mask, base = src2;
526 dst = (base & ~mask) | insert;
530 opcode("ubitfield_extract", 0, tunsigned
,
531 [0, 1, 1], [tunsigned
, tint
, tint
], "", """
532 unsigned base = src0;
533 int offset = src1.x, bits = src2.x;
536 } else if (bits < 0 || offset < 0 || offset + bits > 32) {
537 dst = 0; /* undefined per the spec */
539 dst = (base >> offset) & ((1 << bits) - 1);
542 opcode("ibitfield_extract", 0, tint
,
543 [0, 1, 1], [tint
, tint
, tint
], "", """
545 int offset = src1.x, bits = src2.x;
548 } else if (offset < 0 || bits < 0 || offset + bits > 32) {
551 dst = (base << (32 - offset - bits)) >> offset; /* use sign-extending shift */
555 # Combines the first component of each input to make a 3-component vector.
557 triop_horiz("vec3", 3, 1, 1, 1, """
563 def quadop_horiz(name
, output_size
, src1_size
, src2_size
, src3_size
,
564 src4_size
, const_expr
):
565 opcode(name
, output_size
, tunsigned
,
566 [src1_size
, src2_size
, src3_size
, src4_size
],
567 [tunsigned
, tunsigned
, tunsigned
, tunsigned
],
570 opcode("bitfield_insert", 0, tunsigned
, [0, 0, 1, 1],
571 [tunsigned
, tunsigned
, tint
, tint
], "", """
572 unsigned base = src0, insert = src1;
573 int offset = src2.x, bits = src3.x;
576 } else if (offset < 0 || bits < 0 || bits + offset > 32) {
579 unsigned mask = ((1 << bits) - 1) << offset;
580 dst = (base & ~mask) | ((insert << bits) & mask);
584 quadop_horiz("vec4", 4, 1, 1, 1, 1, """