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("fexp2", tfloat
, "exp2f(src0)")
157 unop("flog2", tfloat
, "log2f(src0)")
158 unop_convert("f2i", tfloat
, tint
, "src0") # Float-to-integer conversion.
159 unop_convert("f2u", tfloat
, tunsigned
, "src0") # Float-to-unsigned conversion
160 unop_convert("i2f", tint
, tfloat
, "src0") # Integer-to-float conversion.
161 # Float-to-boolean conversion
162 unop_convert("f2b", tfloat
, tbool
, "src0 != 0.0f")
163 # Boolean-to-float conversion
164 unop_convert("b2f", tbool
, tfloat
, "src0 ? 1.0f : 0.0f")
165 # Int-to-boolean conversion
166 unop_convert("i2b", tint
, tbool
, "src0 != 0")
167 unop_convert("b2i", tbool
, tint
, "src0 ? 1 : 0") # Boolean-to-int conversion
168 unop_convert("u2f", tunsigned
, tfloat
, "src0") #Unsigned-to-float conversion.
170 unop_reduce("bany", 1, tbool
, tbool
, "{src}", "{src0} || {src1}", "{src}")
171 unop_reduce("ball", 1, tbool
, tbool
, "{src}", "{src0} && {src1}", "{src}")
172 unop_reduce("fany", 1, tfloat
, tfloat
, "{src} != 0.0f", "{src0} || {src1}",
173 "{src} ? 1.0f : 0.0f")
174 unop_reduce("fall", 1, tfloat
, tfloat
, "{src} != 0.0f", "{src0} && {src1}",
175 "{src} ? 1.0f : 0.0f")
177 # Unary floating-point rounding operations.
180 unop("ftrunc", tfloat
, "truncf(src0)")
181 unop("fceil", tfloat
, "ceilf(src0)")
182 unop("ffloor", tfloat
, "floorf(src0)")
183 unop("ffract", tfloat
, "src0 - floorf(src0)")
184 unop("fround_even", tfloat
, "_mesa_roundevenf(src0)")
187 # Trigonometric operations.
190 unop("fsin", tfloat
, "sinf(src0)")
191 unop("fcos", tfloat
, "cosf(src0)")
194 # Partial derivatives.
197 unop("fddx", tfloat
, "0.0f") # the derivative of a constant is 0.
198 unop("fddy", tfloat
, "0.0f")
199 unop("fddx_fine", tfloat
, "0.0f")
200 unop("fddy_fine", tfloat
, "0.0f")
201 unop("fddx_coarse", tfloat
, "0.0f")
202 unop("fddy_coarse", tfloat
, "0.0f")
205 # Floating point pack and unpack operations.
208 unop_horiz("pack_" + fmt
+ "_2x16", 1, tunsigned
, 2, tfloat
, """
209 dst.x = (uint32_t) pack_fmt_1x16(src0.x);
210 dst.x |= ((uint32_t) pack_fmt_1x16(src0.y)) << 16;
211 """.replace("fmt", fmt
))
214 unop_horiz("pack_" + fmt
+ "_4x8", 1, tunsigned
, 4, tfloat
, """
215 dst.x = (uint32_t) pack_fmt_1x8(src0.x);
216 dst.x |= ((uint32_t) pack_fmt_1x8(src0.y)) << 8;
217 dst.x |= ((uint32_t) pack_fmt_1x8(src0.z)) << 16;
218 dst.x |= ((uint32_t) pack_fmt_1x8(src0.w)) << 24;
219 """.replace("fmt", fmt
))
221 def unpack_2x16(fmt
):
222 unop_horiz("unpack_" + fmt
+ "_2x16", 2, tfloat
, 1, tunsigned
, """
223 dst.x = unpack_fmt_1x16((uint16_t)(src0.x & 0xffff));
224 dst.y = unpack_fmt_1x16((uint16_t)(src0.x << 16));
225 """.replace("fmt", fmt
))
228 unop_horiz("unpack_" + fmt
+ "_4x8", 4, tfloat
, 1, tunsigned
, """
229 dst.x = unpack_fmt_1x8((uint8_t)(src0.x & 0xff));
230 dst.y = unpack_fmt_1x8((uint8_t)((src0.x >> 8) & 0xff));
231 dst.z = unpack_fmt_1x8((uint8_t)((src0.x >> 16) & 0xff));
232 dst.w = unpack_fmt_1x8((uint8_t)(src0.x >> 24));
233 """.replace("fmt", fmt
))
248 # Lowered floating point unpacking operations.
251 unop_horiz("unpack_half_2x16_split_x", 1, tfloat
, 1, tunsigned
,
252 "unpack_half_1x16((uint16_t)(src0.x & 0xffff))")
253 unop_horiz("unpack_half_2x16_split_y", 1, tfloat
, 1, tunsigned
,
254 "unpack_half_1x16((uint16_t)(src0.x >> 16))")
257 # Bit operations, part of ARB_gpu_shader5.
260 unop("bitfield_reverse", tunsigned
, """
261 /* we're not winning any awards for speed here, but that's ok */
263 for (unsigned bit = 0; bit < 32; bit++)
264 dst |= ((src0 >> bit) & 1) << (31 - bit);
266 unop("bit_count", tunsigned
, """
268 for (unsigned bit = 0; bit < 32; bit++) {
269 if ((src0 >> bit) & 1)
274 unop_convert("ufind_msb", tunsigned
, tint
, """
276 for (int bit = 31; bit > 0; bit--) {
277 if ((src0 >> bit) & 1) {
284 unop("ifind_msb", tint
, """
286 for (int bit = 31; bit >= 0; bit--) {
287 /* If src0 < 0, we're looking for the first 0 bit.
288 * if src0 >= 0, we're looking for the first 1 bit.
290 if ((((src0 >> bit) & 1) && (src0 >= 0)) ||
291 (!((src0 >> bit) & 1) && (src0 < 0))) {
298 unop("find_lsb", tint
, """
300 for (unsigned bit = 0; bit < 32; bit++) {
301 if ((src0 >> bit) & 1) {
309 for i
in xrange(1, 5):
310 for j
in xrange(1, 5):
311 unop_horiz("fnoise{0}_{1}".format(i
, j
), i
, tfloat
, j
, tfloat
, "0.0f")
313 def binop_convert(name
, out_type
, in_type
, alg_props
, const_expr
):
314 opcode(name
, 0, out_type
, [0, 0], [in_type
, in_type
], alg_props
, const_expr
)
316 def binop(name
, ty
, alg_props
, const_expr
):
317 binop_convert(name
, ty
, ty
, alg_props
, const_expr
)
319 def binop_compare(name
, ty
, alg_props
, const_expr
):
320 binop_convert(name
, tbool
, ty
, alg_props
, const_expr
)
322 def binop_horiz(name
, out_size
, out_type
, src1_size
, src1_type
, src2_size
,
323 src2_type
, const_expr
):
324 opcode(name
, out_size
, out_type
, [src1_size
, src2_size
], [src1_type
, src2_type
],
327 def binop_reduce(name
, output_size
, output_type
, src_type
, prereduce_expr
,
328 reduce_expr
, final_expr
):
330 return final_expr
.format(src
= "(" + src
+ ")")
331 def reduce_(src0
, src1
):
332 return reduce_expr
.format(src0
=src0
, src1
=src1
)
333 def prereduce(src0
, src1
):
334 return "(" + prereduce_expr
.format(src0
=src0
, src1
=src1
) + ")"
335 src0
= prereduce("src0.x", "src1.x")
336 src1
= prereduce("src0.y", "src1.y")
337 src2
= prereduce("src0.z", "src1.z")
338 src3
= prereduce("src0.w", "src1.w")
339 opcode(name
+ "2", output_size
, output_type
,
340 [2, 2], [src_type
, src_type
], commutative
,
341 final(reduce_(src0
, src1
)))
342 opcode(name
+ "3", output_size
, output_type
,
343 [3, 3], [src_type
, src_type
], commutative
,
344 final(reduce_(reduce_(src0
, src1
), src2
)))
345 opcode(name
+ "4", output_size
, output_type
,
346 [4, 4], [src_type
, src_type
], commutative
,
347 final(reduce_(reduce_(src0
, src1
), reduce_(src2
, src3
))))
349 binop("fadd", tfloat
, commutative
+ associative
, "src0 + src1")
350 binop("iadd", tint
, commutative
+ associative
, "src0 + src1")
351 binop("fsub", tfloat
, "", "src0 - src1")
352 binop("isub", tint
, "", "src0 - src1")
354 binop("fmul", tfloat
, commutative
+ associative
, "src0 * src1")
355 # low 32-bits of signed/unsigned integer multiply
356 binop("imul", tint
, commutative
+ associative
, "src0 * src1")
357 # high 32-bits of signed integer multiply
358 binop("imul_high", tint
, commutative
,
359 "(int32_t)(((int64_t) src0 * (int64_t) src1) >> 32)")
360 # high 32-bits of unsigned integer multiply
361 binop("umul_high", tunsigned
, commutative
,
362 "(uint32_t)(((uint64_t) src0 * (uint64_t) src1) >> 32)")
364 binop("fdiv", tfloat
, "", "src0 / src1")
365 binop("idiv", tint
, "", "src0 / src1")
366 binop("udiv", tunsigned
, "", "src0 / src1")
368 # returns a boolean representing the carry resulting from the addition of
369 # the two unsigned arguments.
371 binop_convert("uadd_carry", tbool
, tunsigned
, commutative
, "src0 + src1 < src0")
373 # returns a boolean representing the borrow resulting from the subtraction
374 # of the two unsigned arguments.
376 binop_convert("usub_borrow", tbool
, tunsigned
, "", "src1 < src0")
378 binop("fmod", tfloat
, "", "src0 - src1 * floorf(src0 / src1)")
379 binop("umod", tunsigned
, "", "src1 == 0 ? 0 : src0 % src1")
386 # these integer-aware comparisons return a boolean (0 or ~0)
388 binop_compare("flt", tfloat
, "", "src0 < src1")
389 binop_compare("fge", tfloat
, "", "src0 >= src1")
390 binop_compare("feq", tfloat
, commutative
, "src0 == src1")
391 binop_compare("fne", tfloat
, commutative
, "src0 != src1")
392 binop_compare("ilt", tint
, "", "src0 < src1")
393 binop_compare("ige", tint
, "", "src0 >= src1")
394 binop_compare("ieq", tint
, commutative
, "src0 == src1")
395 binop_compare("ine", tint
, commutative
, "src0 != src1")
396 binop_compare("ult", tunsigned
, "", "src0 < src1")
397 binop_compare("uge", tunsigned
, "", "src0 >= src1")
399 # integer-aware GLSL-style comparisons that compare floats and ints
401 binop_reduce("ball_fequal", 1, tbool
, tfloat
, "{src0} == {src1}",
402 "{src0} && {src1}", "{src}")
403 binop_reduce("bany_fnequal", 1, tbool
, tfloat
, "{src0} != {src1}",
404 "{src0} || {src1}", "{src}")
405 binop_reduce("ball_iequal", 1, tbool
, tint
, "{src0} == {src1}",
406 "{src0} && {src1}", "{src}")
407 binop_reduce("bany_inequal", 1, tbool
, tint
, "{src0} != {src1}",
408 "{src0} || {src1}", "{src}")
410 # non-integer-aware GLSL-style comparisons that return 0.0 or 1.0
412 binop_reduce("fall_equal", 1, tfloat
, tfloat
, "{src0} == {src1}",
413 "{src0} && {src1}", "{src} ? 1.0f : 0.0f")
414 binop_reduce("fany_nequal", 1, tfloat
, tfloat
, "{src0} != {src1}",
415 "{src0} || {src1}", "{src} ? 1.0f : 0.0f")
417 # These comparisons for integer-less hardware return 1.0 and 0.0 for true
418 # and false respectively
420 binop("slt", tfloat
, "", "(src0 < src1) ? 1.0f : 0.0f") # Set on Less Than
421 binop("sge", tfloat
, "", "(src0 >= src1) ? 1.0f : 0.0f") # Set on Greater or Equal
422 binop("seq", tfloat
, commutative
, "(src0 == src1) ? 1.0f : 0.0f") # Set on Equal
423 binop("sne", tfloat
, commutative
, "(src0 != src1) ? 1.0f : 0.0f") # Set on Not Equal
426 binop("ishl", tint
, "", "src0 << src1")
427 binop("ishr", tint
, "", "src0 >> src1")
428 binop("ushr", tunsigned
, "", "src0 >> src1")
430 # bitwise logic operators
432 # These are also used as boolean and, or, xor for hardware supporting
436 binop("iand", tunsigned
, commutative
+ associative
, "src0 & src1")
437 binop("ior", tunsigned
, commutative
+ associative
, "src0 | src1")
438 binop("ixor", tunsigned
, commutative
+ associative
, "src0 ^ src1")
441 # floating point logic operators
443 # These use (src != 0.0) for testing the truth of the input, and output 1.0
444 # for true and 0.0 for false
446 binop("fand", tfloat
, commutative
,
447 "((src0 != 0.0f) && (src1 != 0.0f)) ? 1.0f : 0.0f")
448 binop("for", tfloat
, commutative
,
449 "((src0 != 0.0f) || (src1 != 0.0f)) ? 1.0f : 0.0f")
450 binop("fxor", tfloat
, commutative
,
451 "(src0 != 0.0f && src1 == 0.0f) || (src0 == 0.0f && src1 != 0.0f) ? 1.0f : 0.0f")
453 binop_reduce("fdot", 1, tfloat
, tfloat
, "{src0} * {src1}", "{src0} + {src1}",
456 binop("fmin", tfloat
, "", "fminf(src0, src1)")
457 binop("imin", tint
, commutative
+ associative
, "src1 > src0 ? src0 : src1")
458 binop("umin", tunsigned
, commutative
+ associative
, "src1 > src0 ? src0 : src1")
459 binop("fmax", tfloat
, "", "fmaxf(src0, src1)")
460 binop("imax", tint
, commutative
+ associative
, "src1 > src0 ? src1 : src0")
461 binop("umax", tunsigned
, commutative
+ associative
, "src1 > src0 ? src1 : src0")
463 binop("fpow", tfloat
, "", "powf(src0, src1)")
465 binop_horiz("pack_half_2x16_split", 1, tunsigned
, 1, tfloat
, 1, tfloat
,
466 "pack_half_1x16(src0.x) | (pack_half_1x16(src1.x) << 16)")
468 binop_convert("bfm", tunsigned
, tint
, "", """
469 int offset = src0, bits = src1;
470 if (offset < 0 || bits < 0 || offset + bits > 32)
471 dst = 0; /* undefined per the spec */
473 dst = ((1 << bits)- 1) << offset;
476 opcode("ldexp", 0, tfloat
, [0, 0], [tfloat
, tint
], "", """
477 dst = ldexp(src0, src1);
478 /* flush denormals to zero. */
480 dst = copysign(0.0f, src0);
483 # Combines the first component of each input to make a 2-component vector.
485 binop_horiz("vec2", 2, tunsigned
, 1, tunsigned
, 1, tunsigned
, """
490 def triop(name
, ty
, const_expr
):
491 opcode(name
, 0, ty
, [0, 0, 0], [ty
, ty
, ty
], "", const_expr
)
492 def triop_horiz(name
, output_size
, src1_size
, src2_size
, src3_size
, const_expr
):
493 opcode(name
, output_size
, tunsigned
,
494 [src1_size
, src2_size
, src3_size
],
495 [tunsigned
, tunsigned
, tunsigned
], "", const_expr
)
497 triop("ffma", tfloat
, "src0 * src1 + src2")
499 triop("flrp", tfloat
, "src0 * (1 - src2) + src1 * src2")
503 # A vector conditional select instruction (like ?:, but operating per-
504 # component on vectors). There are two versions, one for floating point
505 # bools (0.0 vs 1.0) and one for integer bools (0 vs ~0).
508 triop("fcsel", tfloat
, "(src0 != 0.0f) ? src1 : src2")
509 opcode("bcsel", 0, tunsigned
, [0, 0, 0],
510 [tbool
, tunsigned
, tunsigned
], "", "src0 ? src1 : src2")
512 triop("bfi", tunsigned
, """
513 unsigned mask = src0, insert = src1 & mask, base = src2;
522 dst = (base & ~mask) | insert;
526 opcode("ubitfield_extract", 0, tunsigned
,
527 [0, 1, 1], [tunsigned
, tint
, tint
], "", """
528 unsigned base = src0;
529 int offset = src1.x, bits = src2.x;
532 } else if (bits < 0 || offset < 0 || offset + bits > 32) {
533 dst = 0; /* undefined per the spec */
535 dst = (base >> offset) & ((1 << bits) - 1);
538 opcode("ibitfield_extract", 0, tint
,
539 [0, 1, 1], [tint
, tint
, tint
], "", """
541 int offset = src1.x, bits = src2.x;
544 } else if (offset < 0 || bits < 0 || offset + bits > 32) {
547 dst = (base << (32 - offset - bits)) >> offset; /* use sign-extending shift */
551 # Combines the first component of each input to make a 3-component vector.
553 triop_horiz("vec3", 3, 1, 1, 1, """
559 def quadop_horiz(name
, output_size
, src1_size
, src2_size
, src3_size
,
560 src4_size
, const_expr
):
561 opcode(name
, output_size
, tunsigned
,
562 [src1_size
, src2_size
, src3_size
, src4_size
],
563 [tunsigned
, tunsigned
, tunsigned
, tunsigned
],
566 opcode("bitfield_insert", 0, tunsigned
, [0, 0, 1, 1],
567 [tunsigned
, tunsigned
, tint
, tint
], "", """
568 unsigned base = src0, insert = src1;
569 int offset = src2.x, bits = src3.x;
572 } else if (offset < 0 || bits < 0 || bits + offset > 32) {
575 unsigned mask = ((1 << bits) - 1) << offset;
576 dst = (base & ~mask) | ((insert << bits) & mask);
580 quadop_horiz("vec4", 4, 1, 1, 1, 1, """