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
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
108 _TYPE_SPLIT_RE
= re
.compile(r
'(?P<type>int|uint|float|bool)(?P<bits>\d+)?')
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
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'))
122 def type_sizes(type_
):
123 if type_has_size(type_
):
124 return [type_size(type_
)]
125 elif type_
== 'bool':
127 elif type_
== 'float':
130 return [1, 8, 16, 32, 64]
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')
137 # Operation where the first two sources are commutative.
139 # For 2-source operations, this just mathematical commutativity. Some
140 # 3-source operations, like ffma, are only commutative in the first two
142 _2src_commutative
= "2src_commutative "
143 associative
= "associative "
145 # global dictionary of opcodes
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
,
155 def unop_convert(name
, out_type
, in_type
, const_expr
):
156 opcode(name
, 0, out_type
, [0], [in_type
], False, "", const_expr
)
158 def unop(name
, ty
, const_expr
):
159 opcode(name
, 0, ty
, [0], [ty
], False, "", const_expr
)
161 def unop_horiz(name
, output_size
, output_type
, input_size
, input_type
,
163 opcode(name
, output_size
, output_type
, [input_size
], [input_type
],
164 False, "", const_expr
)
166 def unop_reduce(name
, output_size
, output_type
, input_type
, prereduce_expr
,
167 reduce_expr
, final_expr
):
169 return "(" + prereduce_expr
.format(src
=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
))))
185 def unop_numeric_convert(name
, out_type
, in_type
, const_expr
):
186 opcode(name
, 0, out_type
, [0], [in_type
], True, "", const_expr
)
188 unop("mov", tuint
, "src0")
190 unop("ineg", tint
, "-src0")
191 unop("fneg", tfloat
, "-src0")
192 unop("inot", tint
, "~src0") # invert every bit of the integer
193 unop("fsign", tfloat
, ("bit_size == 64 ? " +
194 "((src0 == 0.0) ? 0.0 : ((src0 > 0.0) ? 1.0 : -1.0)) : " +
195 "((src0 == 0.0f) ? 0.0f : ((src0 > 0.0f) ? 1.0f : -1.0f))"))
196 unop("isign", tint
, "(src0 == 0) ? 0 : ((src0 > 0) ? 1 : -1)")
197 unop("iabs", tint
, "(src0 < 0) ? -src0 : src0")
198 unop("fabs", tfloat
, "fabs(src0)")
199 unop("fsat", tfloat
, ("bit_size == 64 ? " +
200 "((src0 > 1.0) ? 1.0 : ((src0 <= 0.0) ? 0.0 : src0)) : " +
201 "((src0 > 1.0f) ? 1.0f : ((src0 <= 0.0f) ? 0.0f : src0))"))
202 unop("frcp", tfloat
, "bit_size == 64 ? 1.0 / src0 : 1.0f / src0")
203 unop("frsq", tfloat
, "bit_size == 64 ? 1.0 / sqrt(src0) : 1.0f / sqrtf(src0)")
204 unop("fsqrt", tfloat
, "bit_size == 64 ? sqrt(src0) : sqrtf(src0)")
205 unop("fexp2", tfloat
, "exp2f(src0)")
206 unop("flog2", tfloat
, "log2f(src0)")
208 # Generate all of the numeric conversion opcodes
209 for src_t
in [tint
, tuint
, tfloat
, tbool
]:
211 dst_types
= [tfloat
, tint
]
213 dst_types
= [tfloat
, tint
, tbool
]
215 dst_types
= [tfloat
, tuint
]
216 elif src_t
== tfloat
:
217 dst_types
= [tint
, tuint
, tfloat
, tbool
]
219 for dst_t
in dst_types
:
220 for dst_bit_size
in type_sizes(dst_t
):
221 if dst_bit_size
== 16 and dst_t
== tfloat
and src_t
== tfloat
:
222 rnd_modes
= ['_rtne', '_rtz', '']
223 for rnd_mode
in rnd_modes
:
224 if rnd_mode
== '_rtne':
227 dst = _mesa_half_to_float(_mesa_float_to_float16_rtne(src0));
232 elif rnd_mode
== '_rtz':
235 dst = _mesa_half_to_float(_mesa_float_to_float16_rtz(src0));
243 unop_numeric_convert("{0}2{1}{2}{3}".format(src_t
[0],
247 dst_t
+ str(dst_bit_size
),
249 elif dst_bit_size
== 32 and dst_t
== tfloat
and src_t
== tfloat
:
251 if (bit_size > 32 && nir_is_rounding_mode_rtz(execution_mode, 32)) {
252 dst = _mesa_double_to_float_rtz(src0);
257 unop_numeric_convert("{0}2{1}{2}".format(src_t
[0], dst_t
[0],
259 dst_t
+ str(dst_bit_size
), src_t
, conv_expr
)
261 conv_expr
= "src0 != 0" if dst_t
== tbool
else "src0"
262 unop_numeric_convert("{0}2{1}{2}".format(src_t
[0], dst_t
[0],
264 dst_t
+ str(dst_bit_size
), src_t
, conv_expr
)
267 # Unary floating-point rounding operations.
270 unop("ftrunc", tfloat
, "bit_size == 64 ? trunc(src0) : truncf(src0)")
271 unop("fceil", tfloat
, "bit_size == 64 ? ceil(src0) : ceilf(src0)")
272 unop("ffloor", tfloat
, "bit_size == 64 ? floor(src0) : floorf(src0)")
273 unop("ffract", tfloat
, "src0 - (bit_size == 64 ? floor(src0) : floorf(src0))")
274 unop("fround_even", tfloat
, "bit_size == 64 ? _mesa_roundeven(src0) : _mesa_roundevenf(src0)")
276 unop("fquantize2f16", tfloat
, "(fabs(src0) < ldexpf(1.0, -14)) ? copysignf(0.0f, src0) : _mesa_half_to_float(_mesa_float_to_half(src0))")
278 # Trigonometric operations.
281 unop("fsin", tfloat
, "bit_size == 64 ? sin(src0) : sinf(src0)")
282 unop("fcos", tfloat
, "bit_size == 64 ? cos(src0) : cosf(src0)")
285 unop_convert("frexp_exp", tint32
, tfloat
, "frexp(src0, &dst);")
286 unop_convert("frexp_sig", tfloat
, tfloat
, "int n; dst = frexp(src0, &n);")
288 # Partial derivatives.
291 unop("fddx", tfloat
, "0.0") # the derivative of a constant is 0.
292 unop("fddy", tfloat
, "0.0")
293 unop("fddx_fine", tfloat
, "0.0")
294 unop("fddy_fine", tfloat
, "0.0")
295 unop("fddx_coarse", tfloat
, "0.0")
296 unop("fddy_coarse", tfloat
, "0.0")
299 # Floating point pack and unpack operations.
302 unop_horiz("pack_" + fmt
+ "_2x16", 1, tuint32
, 2, tfloat32
, """
303 dst.x = (uint32_t) pack_fmt_1x16(src0.x);
304 dst.x |= ((uint32_t) pack_fmt_1x16(src0.y)) << 16;
305 """.replace("fmt", fmt
))
308 unop_horiz("pack_" + fmt
+ "_4x8", 1, tuint32
, 4, tfloat32
, """
309 dst.x = (uint32_t) pack_fmt_1x8(src0.x);
310 dst.x |= ((uint32_t) pack_fmt_1x8(src0.y)) << 8;
311 dst.x |= ((uint32_t) pack_fmt_1x8(src0.z)) << 16;
312 dst.x |= ((uint32_t) pack_fmt_1x8(src0.w)) << 24;
313 """.replace("fmt", fmt
))
315 def unpack_2x16(fmt
):
316 unop_horiz("unpack_" + fmt
+ "_2x16", 2, tfloat32
, 1, tuint32
, """
317 dst.x = unpack_fmt_1x16((uint16_t)(src0.x & 0xffff));
318 dst.y = unpack_fmt_1x16((uint16_t)(src0.x << 16));
319 """.replace("fmt", fmt
))
322 unop_horiz("unpack_" + fmt
+ "_4x8", 4, tfloat32
, 1, tuint32
, """
323 dst.x = unpack_fmt_1x8((uint8_t)(src0.x & 0xff));
324 dst.y = unpack_fmt_1x8((uint8_t)((src0.x >> 8) & 0xff));
325 dst.z = unpack_fmt_1x8((uint8_t)((src0.x >> 16) & 0xff));
326 dst.w = unpack_fmt_1x8((uint8_t)(src0.x >> 24));
327 """.replace("fmt", fmt
))
341 unop_horiz("pack_uvec2_to_uint", 1, tuint32
, 2, tuint32
, """
342 dst.x = (src0.x & 0xffff) | (src0.y << 16);
345 unop_horiz("pack_uvec4_to_uint", 1, tuint32
, 4, tuint32
, """
346 dst.x = (src0.x << 0) |
352 unop_horiz("pack_32_2x16", 1, tuint32
, 2, tuint16
,
353 "dst.x = src0.x | ((uint32_t)src0.y << 16);")
355 unop_horiz("pack_64_2x32", 1, tuint64
, 2, tuint32
,
356 "dst.x = src0.x | ((uint64_t)src0.y << 32);")
358 unop_horiz("pack_64_4x16", 1, tuint64
, 4, tuint16
,
359 "dst.x = src0.x | ((uint64_t)src0.y << 16) | ((uint64_t)src0.z << 32) | ((uint64_t)src0.w << 48);")
361 unop_horiz("unpack_64_2x32", 2, tuint32
, 1, tuint64
,
362 "dst.x = src0.x; dst.y = src0.x >> 32;")
364 unop_horiz("unpack_64_4x16", 4, tuint16
, 1, tuint64
,
365 "dst.x = src0.x; dst.y = src0.x >> 16; dst.z = src0.x >> 32; dst.w = src0.w >> 48;")
367 unop_horiz("unpack_32_2x16", 2, tuint16
, 1, tuint32
,
368 "dst.x = src0.x; dst.y = src0.x >> 16;")
370 unop_horiz("unpack_half_2x16_flush_to_zero", 2, tfloat32
, 1, tuint32
, """
371 dst.x = unpack_half_1x16_flush_to_zero((uint16_t)(src0.x & 0xffff));
372 dst.y = unpack_half_1x16_flush_to_zero((uint16_t)(src0.x << 16));
375 # Lowered floating point unpacking operations.
377 unop_convert("unpack_half_2x16_split_x", tfloat32
, tuint32
,
378 "unpack_half_1x16((uint16_t)(src0 & 0xffff))")
379 unop_convert("unpack_half_2x16_split_y", tfloat32
, tuint32
,
380 "unpack_half_1x16((uint16_t)(src0 >> 16))")
382 unop_convert("unpack_half_2x16_split_x_flush_to_zero", tfloat32
, tuint32
,
383 "unpack_half_1x16_flush_to_zero((uint16_t)(src0 & 0xffff))")
384 unop_convert("unpack_half_2x16_split_y_flush_to_zero", tfloat32
, tuint32
,
385 "unpack_half_1x16_flush_to_zero((uint16_t)(src0 >> 16))")
387 unop_convert("unpack_32_2x16_split_x", tuint16
, tuint32
, "src0")
388 unop_convert("unpack_32_2x16_split_y", tuint16
, tuint32
, "src0 >> 16")
390 unop_convert("unpack_64_2x32_split_x", tuint32
, tuint64
, "src0")
391 unop_convert("unpack_64_2x32_split_y", tuint32
, tuint64
, "src0 >> 32")
393 # Bit operations, part of ARB_gpu_shader5.
396 unop("bitfield_reverse", tuint32
, """
397 /* we're not winning any awards for speed here, but that's ok */
399 for (unsigned bit = 0; bit < 32; bit++)
400 dst |= ((src0 >> bit) & 1) << (31 - bit);
402 unop_convert("bit_count", tuint32
, tuint
, """
404 for (unsigned bit = 0; bit < bit_size; bit++) {
405 if ((src0 >> bit) & 1)
410 unop_convert("ufind_msb", tint32
, tuint
, """
412 for (int bit = bit_size - 1; bit >= 0; bit--) {
413 if ((src0 >> bit) & 1) {
420 unop("ifind_msb", tint32
, """
422 for (int bit = 31; bit >= 0; bit--) {
423 /* If src0 < 0, we're looking for the first 0 bit.
424 * if src0 >= 0, we're looking for the first 1 bit.
426 if ((((src0 >> bit) & 1) && (src0 >= 0)) ||
427 (!((src0 >> bit) & 1) && (src0 < 0))) {
434 unop_convert("find_lsb", tint32
, tint
, """
436 for (unsigned bit = 0; bit < bit_size; bit++) {
437 if ((src0 >> bit) & 1) {
445 for i
in range(1, 5):
446 for j
in range(1, 5):
447 unop_horiz("fnoise{0}_{1}".format(i
, j
), i
, tfloat
, j
, tfloat
, "0.0f")
450 # AMD_gcn_shader extended instructions
451 unop_horiz("cube_face_coord", 2, tfloat32
, 3, tfloat32
, """
453 float absX = fabs(src0.x);
454 float absY = fabs(src0.y);
455 float absZ = fabs(src0.z);
458 if (absX >= absY && absX >= absZ) { ma = 2 * src0.x; }
459 if (absY >= absX && absY >= absZ) { ma = 2 * src0.y; }
460 if (absZ >= absX && absZ >= absY) { ma = 2 * src0.z; }
462 if (src0.x >= 0 && absX >= absY && absX >= absZ) { dst.x = -src0.z; dst.y = -src0.y; }
463 if (src0.x < 0 && absX >= absY && absX >= absZ) { dst.x = src0.z; dst.y = -src0.y; }
464 if (src0.y >= 0 && absY >= absX && absY >= absZ) { dst.x = src0.x; dst.y = src0.z; }
465 if (src0.y < 0 && absY >= absX && absY >= absZ) { dst.x = src0.x; dst.y = -src0.z; }
466 if (src0.z >= 0 && absZ >= absX && absZ >= absY) { dst.x = src0.x; dst.y = -src0.y; }
467 if (src0.z < 0 && absZ >= absX && absZ >= absY) { dst.x = -src0.x; dst.y = -src0.y; }
469 dst.x = dst.x / ma + 0.5;
470 dst.y = dst.y / ma + 0.5;
473 unop_horiz("cube_face_index", 1, tfloat32
, 3, tfloat32
, """
474 float absX = fabs(src0.x);
475 float absY = fabs(src0.y);
476 float absZ = fabs(src0.z);
477 if (src0.x >= 0 && absX >= absY && absX >= absZ) dst.x = 0;
478 if (src0.x < 0 && absX >= absY && absX >= absZ) dst.x = 1;
479 if (src0.y >= 0 && absY >= absX && absY >= absZ) dst.x = 2;
480 if (src0.y < 0 && absY >= absX && absY >= absZ) dst.x = 3;
481 if (src0.z >= 0 && absZ >= absX && absZ >= absY) dst.x = 4;
482 if (src0.z < 0 && absZ >= absX && absZ >= absY) dst.x = 5;
485 # Sum of vector components
486 unop_reduce("fsum", 1, tfloat
, tfloat
, "{src}", "{src0} + {src1}", "{src}")
488 def binop_convert(name
, out_type
, in_type
, alg_props
, const_expr
):
489 opcode(name
, 0, out_type
, [0, 0], [in_type
, in_type
],
490 False, alg_props
, const_expr
)
492 def binop(name
, ty
, alg_props
, const_expr
):
493 binop_convert(name
, ty
, ty
, alg_props
, const_expr
)
495 def binop_compare(name
, ty
, alg_props
, const_expr
):
496 binop_convert(name
, tbool1
, ty
, alg_props
, const_expr
)
498 def binop_compare32(name
, ty
, alg_props
, const_expr
):
499 binop_convert(name
, tbool32
, ty
, alg_props
, const_expr
)
501 def binop_horiz(name
, out_size
, out_type
, src1_size
, src1_type
, src2_size
,
502 src2_type
, const_expr
):
503 opcode(name
, out_size
, out_type
, [src1_size
, src2_size
], [src1_type
, src2_type
],
504 False, "", const_expr
)
506 def binop_reduce(name
, output_size
, output_type
, src_type
, prereduce_expr
,
507 reduce_expr
, final_expr
):
509 return final_expr
.format(src
= "(" + src
+ ")")
510 def reduce_(src0
, src1
):
511 return reduce_expr
.format(src0
=src0
, src1
=src1
)
512 def prereduce(src0
, src1
):
513 return "(" + prereduce_expr
.format(src0
=src0
, src1
=src1
) + ")"
514 src0
= prereduce("src0.x", "src1.x")
515 src1
= prereduce("src0.y", "src1.y")
516 src2
= prereduce("src0.z", "src1.z")
517 src3
= prereduce("src0.w", "src1.w")
518 opcode(name
+ "2", output_size
, output_type
,
519 [2, 2], [src_type
, src_type
], False, _2src_commutative
,
520 final(reduce_(src0
, src1
)))
521 opcode(name
+ "3", output_size
, output_type
,
522 [3, 3], [src_type
, src_type
], False, _2src_commutative
,
523 final(reduce_(reduce_(src0
, src1
), src2
)))
524 opcode(name
+ "4", output_size
, output_type
,
525 [4, 4], [src_type
, src_type
], False, _2src_commutative
,
526 final(reduce_(reduce_(src0
, src1
), reduce_(src2
, src3
))))
528 binop("fadd", tfloat
, _2src_commutative
+ associative
,"""
529 if (nir_is_rounding_mode_rtz(execution_mode, bit_size)) {
531 dst = _mesa_double_add_rtz(src0, src1);
533 dst = _mesa_double_to_float_rtz((double)src0 + (double)src1);
538 binop("iadd", tint
, _2src_commutative
+ associative
, "src0 + src1")
539 binop("iadd_sat", tint
, _2src_commutative
, """
541 (src0 + src1 < src0 ? (1ull << (bit_size - 1)) - 1 : src0 + src1) :
542 (src0 < src0 + src1 ? (1ull << (bit_size - 1)) : src0 + src1)
544 binop("uadd_sat", tuint
, _2src_commutative
,
545 "(src0 + src1) < src0 ? MAX_UINT_FOR_SIZE(sizeof(src0) * 8) : (src0 + src1)")
546 binop("isub_sat", tint
, "", """
548 (src0 - src1 < src0 ? (1ull << (bit_size - 1)) - 1 : src0 - src1) :
549 (src0 < src0 - src1 ? (1ull << (bit_size - 1)) : src0 - src1)
551 binop("usub_sat", tuint
, "", "src0 < src1 ? 0 : src0 - src1")
553 binop("fsub", tfloat
, "", """
554 if (nir_is_rounding_mode_rtz(execution_mode, bit_size)) {
556 dst = _mesa_double_sub_rtz(src0, src1);
558 dst = _mesa_double_to_float_rtz((double)src0 - (double)src1);
563 binop("isub", tint
, "", "src0 - src1")
565 binop("fmul", tfloat
, _2src_commutative
+ associative
, """
566 if (nir_is_rounding_mode_rtz(execution_mode, bit_size)) {
568 dst = _mesa_double_mul_rtz(src0, src1);
570 dst = _mesa_double_to_float_rtz((double)src0 * (double)src1);
575 # low 32-bits of signed/unsigned integer multiply
576 binop("imul", tint
, _2src_commutative
+ associative
, "src0 * src1")
578 # Generate 64 bit result from 2 32 bits quantity
579 binop_convert("imul_2x32_64", tint64
, tint32
, _2src_commutative
,
580 "(int64_t)src0 * (int64_t)src1")
581 binop_convert("umul_2x32_64", tuint64
, tuint32
, _2src_commutative
,
582 "(uint64_t)src0 * (uint64_t)src1")
584 # high 32-bits of signed integer multiply
585 binop("imul_high", tint
, _2src_commutative
, """
586 if (bit_size == 64) {
587 /* We need to do a full 128-bit x 128-bit multiply in order for the sign
588 * extension to work properly. The casts are kind-of annoying but needed
589 * to prevent compiler warnings.
591 uint32_t src0_u32[4] = {
597 uint32_t src1_u32[4] = {
603 uint32_t prod_u32[4];
604 ubm_mul_u32arr(prod_u32, src0_u32, src1_u32);
605 dst = (uint64_t)prod_u32[2] | ((uint64_t)prod_u32[3] << 32);
607 dst = ((int64_t)src0 * (int64_t)src1) >> bit_size;
611 # high 32-bits of unsigned integer multiply
612 binop("umul_high", tuint
, _2src_commutative
, """
613 if (bit_size == 64) {
614 /* The casts are kind-of annoying but needed to prevent compiler warnings. */
615 uint32_t src0_u32[2] = { src0, (uint64_t)src0 >> 32 };
616 uint32_t src1_u32[2] = { src1, (uint64_t)src1 >> 32 };
617 uint32_t prod_u32[4];
618 ubm_mul_u32arr(prod_u32, src0_u32, src1_u32);
619 dst = (uint64_t)prod_u32[2] | ((uint64_t)prod_u32[3] << 32);
621 dst = ((uint64_t)src0 * (uint64_t)src1) >> bit_size;
625 # low 32-bits of unsigned integer multiply
626 binop("umul_low", tuint32
, _2src_commutative
, """
627 uint64_t mask = (1 << (bit_size / 2)) - 1;
628 dst = ((uint64_t)src0 & mask) * ((uint64_t)src1 & mask);
632 binop("fdiv", tfloat
, "", "src0 / src1")
633 binop("idiv", tint
, "", "src1 == 0 ? 0 : (src0 / src1)")
634 binop("udiv", tuint
, "", "src1 == 0 ? 0 : (src0 / src1)")
636 # returns a boolean representing the carry resulting from the addition of
637 # the two unsigned arguments.
639 binop_convert("uadd_carry", tuint
, tuint
, _2src_commutative
, "src0 + src1 < src0")
641 # returns a boolean representing the borrow resulting from the subtraction
642 # of the two unsigned arguments.
644 binop_convert("usub_borrow", tuint
, tuint
, "", "src0 < src1")
646 # hadd: (a + b) >> 1 (without overflow)
647 # x + y = x - (x & ~y) + (x & ~y) + y - (~x & y) + (~x & y)
648 # = (x & y) + (x & ~y) + (x & y) + (~x & y)
649 # = 2 * (x & y) + (x & ~y) + (~x & y)
650 # = ((x & y) << 1) + (x ^ y)
652 # Since we know that the bottom bit of (x & y) << 1 is zero,
654 # (x + y) >> 1 = (((x & y) << 1) + (x ^ y)) >> 1
655 # = (x & y) + ((x ^ y) >> 1)
656 binop("ihadd", tint
, _2src_commutative
, "(src0 & src1) + ((src0 ^ src1) >> 1)")
657 binop("uhadd", tuint
, _2src_commutative
, "(src0 & src1) + ((src0 ^ src1) >> 1)")
659 # rhadd: (a + b + 1) >> 1 (without overflow)
660 # x + y + 1 = x + (~x & y) - (~x & y) + y + (x & ~y) - (x & ~y) + 1
661 # = (x | y) - (~x & y) + (x | y) - (x & ~y) + 1
662 # = 2 * (x | y) - ((~x & y) + (x & ~y)) + 1
663 # = ((x | y) << 1) - (x ^ y) + 1
665 # Since we know that the bottom bit of (x & y) << 1 is zero,
667 # (x + y + 1) >> 1 = (x | y) + (-(x ^ y) + 1) >> 1)
668 # = (x | y) - ((x ^ y) >> 1)
669 binop("irhadd", tint
, _2src_commutative
, "(src0 | src1) + ((src0 ^ src1) >> 1)")
670 binop("urhadd", tuint
, _2src_commutative
, "(src0 | src1) + ((src0 ^ src1) >> 1)")
672 binop("umod", tuint
, "", "src1 == 0 ? 0 : src0 % src1")
674 # For signed integers, there are several different possible definitions of
675 # "modulus" or "remainder". We follow the conventions used by LLVM and
676 # SPIR-V. The irem opcode implements the standard C/C++ signed "%"
677 # operation while the imod opcode implements the more mathematical
678 # "modulus" operation. For details on the difference, see
680 # http://mathforum.org/library/drmath/view/52343.html
682 binop("irem", tint
, "", "src1 == 0 ? 0 : src0 % src1")
683 binop("imod", tint
, "",
684 "src1 == 0 ? 0 : ((src0 % src1 == 0 || (src0 >= 0) == (src1 >= 0)) ?"
685 " src0 % src1 : src0 % src1 + src1)")
686 binop("fmod", tfloat
, "", "src0 - src1 * floorf(src0 / src1)")
687 binop("frem", tfloat
, "", "src0 - src1 * truncf(src0 / src1)")
694 # these integer-aware comparisons return a boolean (0 or ~0)
696 binop_compare("flt", tfloat
, "", "src0 < src1")
697 binop_compare("fge", tfloat
, "", "src0 >= src1")
698 binop_compare("feq", tfloat
, _2src_commutative
, "src0 == src1")
699 binop_compare("fne", tfloat
, _2src_commutative
, "src0 != src1")
700 binop_compare("ilt", tint
, "", "src0 < src1")
701 binop_compare("ige", tint
, "", "src0 >= src1")
702 binop_compare("ieq", tint
, _2src_commutative
, "src0 == src1")
703 binop_compare("ine", tint
, _2src_commutative
, "src0 != src1")
704 binop_compare("ult", tuint
, "", "src0 < src1")
705 binop_compare("uge", tuint
, "", "src0 >= src1")
706 binop_compare32("flt32", tfloat
, "", "src0 < src1")
707 binop_compare32("fge32", tfloat
, "", "src0 >= src1")
708 binop_compare32("feq32", tfloat
, _2src_commutative
, "src0 == src1")
709 binop_compare32("fne32", tfloat
, _2src_commutative
, "src0 != src1")
710 binop_compare32("ilt32", tint
, "", "src0 < src1")
711 binop_compare32("ige32", tint
, "", "src0 >= src1")
712 binop_compare32("ieq32", tint
, _2src_commutative
, "src0 == src1")
713 binop_compare32("ine32", tint
, _2src_commutative
, "src0 != src1")
714 binop_compare32("ult32", tuint
, "", "src0 < src1")
715 binop_compare32("uge32", tuint
, "", "src0 >= src1")
717 # integer-aware GLSL-style comparisons that compare floats and ints
719 binop_reduce("ball_fequal", 1, tbool1
, tfloat
, "{src0} == {src1}",
720 "{src0} && {src1}", "{src}")
721 binop_reduce("bany_fnequal", 1, tbool1
, tfloat
, "{src0} != {src1}",
722 "{src0} || {src1}", "{src}")
723 binop_reduce("ball_iequal", 1, tbool1
, tint
, "{src0} == {src1}",
724 "{src0} && {src1}", "{src}")
725 binop_reduce("bany_inequal", 1, tbool1
, tint
, "{src0} != {src1}",
726 "{src0} || {src1}", "{src}")
728 binop_reduce("b32all_fequal", 1, tbool32
, tfloat
, "{src0} == {src1}",
729 "{src0} && {src1}", "{src}")
730 binop_reduce("b32any_fnequal", 1, tbool32
, tfloat
, "{src0} != {src1}",
731 "{src0} || {src1}", "{src}")
732 binop_reduce("b32all_iequal", 1, tbool32
, tint
, "{src0} == {src1}",
733 "{src0} && {src1}", "{src}")
734 binop_reduce("b32any_inequal", 1, tbool32
, tint
, "{src0} != {src1}",
735 "{src0} || {src1}", "{src}")
737 # non-integer-aware GLSL-style comparisons that return 0.0 or 1.0
739 binop_reduce("fall_equal", 1, tfloat32
, tfloat32
, "{src0} == {src1}",
740 "{src0} && {src1}", "{src} ? 1.0f : 0.0f")
741 binop_reduce("fany_nequal", 1, tfloat32
, tfloat32
, "{src0} != {src1}",
742 "{src0} || {src1}", "{src} ? 1.0f : 0.0f")
744 # These comparisons for integer-less hardware return 1.0 and 0.0 for true
745 # and false respectively
747 binop("slt", tfloat32
, "", "(src0 < src1) ? 1.0f : 0.0f") # Set on Less Than
748 binop("sge", tfloat
, "", "(src0 >= src1) ? 1.0f : 0.0f") # Set on Greater or Equal
749 binop("seq", tfloat32
, _2src_commutative
, "(src0 == src1) ? 1.0f : 0.0f") # Set on Equal
750 binop("sne", tfloat32
, _2src_commutative
, "(src0 != src1) ? 1.0f : 0.0f") # Set on Not Equal
752 # SPIRV shifts are undefined for shift-operands >= bitsize,
753 # but SM5 shifts are defined to use the least significant bits, only
754 # The NIR definition is according to the SM5 specification.
755 opcode("ishl", 0, tint
, [0, 0], [tint
, tuint32
], False, "",
756 "src0 << (src1 & (sizeof(src0) * 8 - 1))")
757 opcode("ishr", 0, tint
, [0, 0], [tint
, tuint32
], False, "",
758 "src0 >> (src1 & (sizeof(src0) * 8 - 1))")
759 opcode("ushr", 0, tuint
, [0, 0], [tuint
, tuint32
], False, "",
760 "src0 >> (src1 & (sizeof(src0) * 8 - 1))")
762 opcode("urol", 0, tuint
, [0, 0], [tuint
, tuint32
], False, "", """
763 uint32_t rotate_mask = sizeof(src0) * 8 - 1;
764 dst = (src0 << (src1 & rotate_mask)) |
765 (src0 >> (-src1 & rotate_mask));
767 opcode("uror", 0, tuint
, [0, 0], [tuint
, tuint32
], False, "", """
768 uint32_t rotate_mask = sizeof(src0) * 8 - 1;
769 dst = (src0 >> (src1 & rotate_mask)) |
770 (src0 << (-src1 & rotate_mask));
773 # bitwise logic operators
775 # These are also used as boolean and, or, xor for hardware supporting
779 binop("iand", tuint
, _2src_commutative
+ associative
, "src0 & src1")
780 binop("ior", tuint
, _2src_commutative
+ associative
, "src0 | src1")
781 binop("ixor", tuint
, _2src_commutative
+ associative
, "src0 ^ src1")
784 binop_reduce("fdot", 1, tfloat
, tfloat
, "{src0} * {src1}", "{src0} + {src1}",
787 binop_reduce("fdot_replicated", 4, tfloat
, tfloat
,
788 "{src0} * {src1}", "{src0} + {src1}", "{src}")
790 opcode("fdph", 1, tfloat
, [3, 4], [tfloat
, tfloat
], False, "",
791 "src0.x * src1.x + src0.y * src1.y + src0.z * src1.z + src1.w")
792 opcode("fdph_replicated", 4, tfloat
, [3, 4], [tfloat
, tfloat
], False, "",
793 "src0.x * src1.x + src0.y * src1.y + src0.z * src1.z + src1.w")
795 binop("fmin", tfloat
, "", "fmin(src0, src1)")
796 binop("imin", tint
, _2src_commutative
+ associative
, "src1 > src0 ? src0 : src1")
797 binop("umin", tuint
, _2src_commutative
+ associative
, "src1 > src0 ? src0 : src1")
798 binop("fmax", tfloat
, "", "fmax(src0, src1)")
799 binop("imax", tint
, _2src_commutative
+ associative
, "src1 > src0 ? src1 : src0")
800 binop("umax", tuint
, _2src_commutative
+ associative
, "src1 > src0 ? src1 : src0")
802 # Saturated vector add for 4 8bit ints.
803 binop("usadd_4x8", tint32
, _2src_commutative
+ associative
, """
805 for (int i = 0; i < 32; i += 8) {
806 dst |= MIN2(((src0 >> i) & 0xff) + ((src1 >> i) & 0xff), 0xff) << i;
810 # Saturated vector subtract for 4 8bit ints.
811 binop("ussub_4x8", tint32
, "", """
813 for (int i = 0; i < 32; i += 8) {
814 int src0_chan = (src0 >> i) & 0xff;
815 int src1_chan = (src1 >> i) & 0xff;
816 if (src0_chan > src1_chan)
817 dst |= (src0_chan - src1_chan) << i;
821 # vector min for 4 8bit ints.
822 binop("umin_4x8", tint32
, _2src_commutative
+ associative
, """
824 for (int i = 0; i < 32; i += 8) {
825 dst |= MIN2((src0 >> i) & 0xff, (src1 >> i) & 0xff) << i;
829 # vector max for 4 8bit ints.
830 binop("umax_4x8", tint32
, _2src_commutative
+ associative
, """
832 for (int i = 0; i < 32; i += 8) {
833 dst |= MAX2((src0 >> i) & 0xff, (src1 >> i) & 0xff) << i;
837 # unorm multiply: (a * b) / 255.
838 binop("umul_unorm_4x8", tint32
, _2src_commutative
+ associative
, """
840 for (int i = 0; i < 32; i += 8) {
841 int src0_chan = (src0 >> i) & 0xff;
842 int src1_chan = (src1 >> i) & 0xff;
843 dst |= ((src0_chan * src1_chan) / 255) << i;
847 binop("fpow", tfloat
, "", "bit_size == 64 ? powf(src0, src1) : pow(src0, src1)")
849 binop_horiz("pack_half_2x16_split", 1, tuint32
, 1, tfloat32
, 1, tfloat32
,
850 "pack_half_1x16(src0.x) | (pack_half_1x16(src1.x) << 16)")
852 binop_convert("pack_64_2x32_split", tuint64
, tuint32
, "",
853 "src0 | ((uint64_t)src1 << 32)")
855 binop_convert("pack_32_2x16_split", tuint32
, tuint16
, "",
856 "src0 | ((uint32_t)src1 << 16)")
858 # bfm implements the behavior of the first operation of the SM5 "bfi" assembly
859 # and that of the "bfi1" i965 instruction. That is, the bits and offset values
860 # are from the low five bits of src0 and src1, respectively.
861 binop_convert("bfm", tuint32
, tint32
, "", """
862 int bits = src0 & 0x1F;
863 int offset = src1 & 0x1F;
864 dst = ((1u << bits) - 1) << offset;
867 opcode("ldexp", 0, tfloat
, [0, 0], [tfloat
, tint32
], False, "", """
868 dst = (bit_size == 64) ? ldexp(src0, src1) : ldexpf(src0, src1);
869 /* flush denormals to zero. */
871 dst = copysignf(0.0f, src0);
874 # Combines the first component of each input to make a 2-component vector.
876 binop_horiz("vec2", 2, tuint
, 1, tuint
, 1, tuint
, """
882 binop("extract_u8", tuint
, "", "(uint8_t)(src0 >> (src1 * 8))")
883 binop("extract_i8", tint
, "", "(int8_t)(src0 >> (src1 * 8))")
886 binop("extract_u16", tuint
, "", "(uint16_t)(src0 >> (src1 * 16))")
887 binop("extract_i16", tint
, "", "(int16_t)(src0 >> (src1 * 16))")
890 def triop(name
, ty
, alg_props
, const_expr
):
891 opcode(name
, 0, ty
, [0, 0, 0], [ty
, ty
, ty
], False, alg_props
, const_expr
)
892 def triop_horiz(name
, output_size
, src1_size
, src2_size
, src3_size
, const_expr
):
893 opcode(name
, output_size
, tuint
,
894 [src1_size
, src2_size
, src3_size
],
895 [tuint
, tuint
, tuint
], False, "", const_expr
)
897 triop("ffma", tfloat
, _2src_commutative
, """
898 if (nir_is_rounding_mode_rtz(execution_mode, bit_size)) {
900 dst = _mesa_double_fma_rtz(src0, src1, src2);
901 else if (bit_size == 32)
902 dst = _mesa_float_fma_rtz(src0, src1, src2);
904 dst = _mesa_double_to_float_rtz(_mesa_double_fma_rtz(src0, src1, src2));
907 dst = fmaf(src0, src1, src2);
909 dst = fma(src0, src1, src2);
913 triop("flrp", tfloat
, "", "src0 * (1 - src2) + src1 * src2")
917 # A vector conditional select instruction (like ?:, but operating per-
918 # component on vectors). There are two versions, one for floating point
919 # bools (0.0 vs 1.0) and one for integer bools (0 vs ~0).
922 triop("fcsel", tfloat32
, "", "(src0 != 0.0f) ? src1 : src2")
925 triop("fmin3", tfloat
, "", "fminf(src0, fminf(src1, src2))")
926 triop("imin3", tint
, "", "MIN2(src0, MIN2(src1, src2))")
927 triop("umin3", tuint
, "", "MIN2(src0, MIN2(src1, src2))")
929 triop("fmax3", tfloat
, "", "fmaxf(src0, fmaxf(src1, src2))")
930 triop("imax3", tint
, "", "MAX2(src0, MAX2(src1, src2))")
931 triop("umax3", tuint
, "", "MAX2(src0, MAX2(src1, src2))")
933 triop("fmed3", tfloat
, "", "fmaxf(fminf(fmaxf(src0, src1), src2), fminf(src0, src1))")
934 triop("imed3", tint
, "", "MAX2(MIN2(MAX2(src0, src1), src2), MIN2(src0, src1))")
935 triop("umed3", tuint
, "", "MAX2(MIN2(MAX2(src0, src1), src2), MIN2(src0, src1))")
937 opcode("bcsel", 0, tuint
, [0, 0, 0],
938 [tbool1
, tuint
, tuint
], False, "", "src0 ? src1 : src2")
939 opcode("b32csel", 0, tuint
, [0, 0, 0],
940 [tbool32
, tuint
, tuint
], False, "", "src0 ? src1 : src2")
943 triop("bfi", tuint32
, "", """
944 unsigned mask = src0, insert = src1, base = src2;
953 dst = (base & ~mask) | (insert & mask);
958 triop("bitfield_select", tuint
, "", "(src0 & src1) | (~src0 & src2)")
960 # SM5 ubfe/ibfe assembly: only the 5 least significant bits of offset and bits are used.
961 opcode("ubfe", 0, tuint32
,
962 [0, 0, 0], [tuint32
, tuint32
, tuint32
], False, "", """
963 unsigned base = src0;
964 unsigned offset = src1 & 0x1F;
965 unsigned bits = src2 & 0x1F;
968 } else if (offset + bits < 32) {
969 dst = (base << (32 - bits - offset)) >> (32 - bits);
971 dst = base >> offset;
974 opcode("ibfe", 0, tint32
,
975 [0, 0, 0], [tint32
, tuint32
, tuint32
], False, "", """
977 unsigned offset = src1 & 0x1F;
978 unsigned bits = src2 & 0x1F;
981 } else if (offset + bits < 32) {
982 dst = (base << (32 - bits - offset)) >> (32 - bits);
984 dst = base >> offset;
988 # GLSL bitfieldExtract()
989 opcode("ubitfield_extract", 0, tuint32
,
990 [0, 0, 0], [tuint32
, tint32
, tint32
], False, "", """
991 unsigned base = src0;
992 int offset = src1, bits = src2;
995 } else if (bits < 0 || offset < 0 || offset + bits > 32) {
996 dst = 0; /* undefined per the spec */
998 dst = (base >> offset) & ((1ull << bits) - 1);
1001 opcode("ibitfield_extract", 0, tint32
,
1002 [0, 0, 0], [tint32
, tint32
, tint32
], False, "", """
1004 int offset = src1, bits = src2;
1007 } else if (offset < 0 || bits < 0 || offset + bits > 32) {
1010 dst = (base << (32 - offset - bits)) >> offset; /* use sign-extending shift */
1014 # Combines the first component of each input to make a 3-component vector.
1016 triop_horiz("vec3", 3, 1, 1, 1, """
1022 def quadop_horiz(name
, output_size
, src1_size
, src2_size
, src3_size
,
1023 src4_size
, const_expr
):
1024 opcode(name
, output_size
, tuint
,
1025 [src1_size
, src2_size
, src3_size
, src4_size
],
1026 [tuint
, tuint
, tuint
, tuint
],
1027 False, "", const_expr
)
1029 opcode("bitfield_insert", 0, tuint32
, [0, 0, 0, 0],
1030 [tuint32
, tuint32
, tint32
, tint32
], False, "", """
1031 unsigned base = src0, insert = src1;
1032 int offset = src2, bits = src3;
1035 } else if (offset < 0 || bits < 0 || bits + offset > 32) {
1038 unsigned mask = ((1ull << bits) - 1) << offset;
1039 dst = (base & ~mask) | ((insert << offset) & mask);
1043 quadop_horiz("vec4", 4, 1, 1, 1, 1, """
1050 # ir3-specific instruction that maps directly to mul-add shift high mix,
1051 # (IMADSH_MIX16 i.e. ah * bl << 16 + c). It is used for lowering integer
1052 # multiplication (imul) on Freedreno backend..
1053 opcode("imadsh_mix16", 1, tint32
,
1054 [1, 1, 1], [tint32
, tint32
, tint32
], False, "", """
1055 dst.x = ((((src0.x & 0xffff0000) >> 16) * (src1.x & 0x0000ffff)) << 16) + src2.x;