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("flog", tfloat
, "logf(src0)") # log base e
157 unop("fexp2", tfloat
, "exp2f(src0)")
158 unop("flog2", tfloat
, "log2f(src0)")
159 unop_convert("f2i", tfloat
, tint
, "src0") # Float-to-integer conversion.
160 unop_convert("f2u", tfloat
, tunsigned
, "src0") # Float-to-unsigned conversion
161 unop_convert("i2f", tint
, tfloat
, "src0") # Integer-to-float conversion.
162 # Float-to-boolean conversion
163 unop_convert("f2b", tfloat
, tbool
, "src0 != 0.0f")
164 # Boolean-to-float conversion
165 unop_convert("b2f", tbool
, tfloat
, "src0 ? 1.0f : 0.0f")
166 # Int-to-boolean conversion
167 unop_convert("i2b", tint
, tbool
, "src0 != 0")
168 unop_convert("b2i", tbool
, tint
, "src0 ? 1 : 0") # Boolean-to-int conversion
169 unop_convert("u2f", tunsigned
, tfloat
, "src0") #Unsigned-to-float conversion.
171 unop_reduce("bany", 1, tbool
, tbool
, "{src}", "{src0} || {src1}", "{src}")
172 unop_reduce("ball", 1, tbool
, tbool
, "{src}", "{src0} && {src1}", "{src}")
173 unop_reduce("fany", 1, tfloat
, tfloat
, "{src} != 0.0f", "{src0} || {src1}",
174 "{src} ? 1.0f : 0.0f")
175 unop_reduce("fall", 1, tfloat
, tfloat
, "{src} != 0.0f", "{src0} && {src1}",
176 "{src} ? 1.0f : 0.0f")
178 # Unary floating-point rounding operations.
181 unop("ftrunc", tfloat
, "truncf(src0)")
182 unop("fceil", tfloat
, "ceilf(src0)")
183 unop("ffloor", tfloat
, "floorf(src0)")
184 unop("ffract", tfloat
, "src0 - floorf(src0)")
185 unop("fround_even", tfloat
, "_mesa_roundevenf(src0)")
188 # Trigonometric operations.
191 unop("fsin", tfloat
, "sinf(src0)")
192 unop("fcos", tfloat
, "cosf(src0)")
195 # Partial derivatives.
198 unop("fddx", tfloat
, "0.0f") # the derivative of a constant is 0.
199 unop("fddy", tfloat
, "0.0f")
200 unop("fddx_fine", tfloat
, "0.0f")
201 unop("fddy_fine", tfloat
, "0.0f")
202 unop("fddx_coarse", tfloat
, "0.0f")
203 unop("fddy_coarse", tfloat
, "0.0f")
206 # Floating point pack and unpack operations.
209 unop_horiz("pack_" + fmt
+ "_2x16", 1, tunsigned
, 2, tfloat
, """
210 dst.x = (uint32_t) pack_fmt_1x16(src0.x);
211 dst.x |= ((uint32_t) pack_fmt_1x16(src0.y)) << 16;
212 """.replace("fmt", fmt
))
215 unop_horiz("pack_" + fmt
+ "_4x8", 1, tunsigned
, 4, tfloat
, """
216 dst.x = (uint32_t) pack_fmt_1x8(src0.x);
217 dst.x |= ((uint32_t) pack_fmt_1x8(src0.y)) << 8;
218 dst.x |= ((uint32_t) pack_fmt_1x8(src0.z)) << 16;
219 dst.x |= ((uint32_t) pack_fmt_1x8(src0.w)) << 24;
220 """.replace("fmt", fmt
))
222 def unpack_2x16(fmt
):
223 unop_horiz("unpack_" + fmt
+ "_2x16", 2, tfloat
, 1, tunsigned
, """
224 dst.x = unpack_fmt_1x16((uint16_t)(src0.x & 0xffff));
225 dst.y = unpack_fmt_1x16((uint16_t)(src0.x << 16));
226 """.replace("fmt", fmt
))
229 unop_horiz("unpack_" + fmt
+ "_4x8", 4, tfloat
, 1, tunsigned
, """
230 dst.x = unpack_fmt_1x8((uint8_t)(src0.x & 0xff));
231 dst.y = unpack_fmt_1x8((uint8_t)((src0.x >> 8) & 0xff));
232 dst.z = unpack_fmt_1x8((uint8_t)((src0.x >> 16) & 0xff));
233 dst.w = unpack_fmt_1x8((uint8_t)(src0.x >> 24));
234 """.replace("fmt", fmt
))
249 # Lowered floating point unpacking operations.
252 unop_horiz("unpack_half_2x16_split_x", 1, tfloat
, 1, tunsigned
,
253 "unpack_half_1x16((uint16_t)(src0.x & 0xffff))")
254 unop_horiz("unpack_half_2x16_split_y", 1, tfloat
, 1, tunsigned
,
255 "unpack_half_1x16((uint16_t)(src0.x >> 16))")
258 # Bit operations, part of ARB_gpu_shader5.
261 unop("bitfield_reverse", tunsigned
, """
262 /* we're not winning any awards for speed here, but that's ok */
264 for (unsigned bit = 0; bit < 32; bit++)
265 dst |= ((src0 >> bit) & 1) << (31 - bit);
267 unop("bit_count", tunsigned
, """
269 for (unsigned bit = 0; bit < 32; bit++) {
270 if ((src0 >> bit) & 1)
275 unop_convert("ufind_msb", tunsigned
, tint
, """
277 for (int bit = 31; bit > 0; bit--) {
278 if ((src0 >> bit) & 1) {
285 unop("ifind_msb", tint
, """
287 for (int bit = 31; bit >= 0; bit--) {
288 /* If src0 < 0, we're looking for the first 0 bit.
289 * if src0 >= 0, we're looking for the first 1 bit.
291 if ((((src0 >> bit) & 1) && (src0 >= 0)) ||
292 (!((src0 >> bit) & 1) && (src0 < 0))) {
299 unop("find_lsb", tint
, """
301 for (unsigned bit = 0; bit < 32; bit++) {
302 if ((src0 >> bit) & 1) {
310 for i
in xrange(1, 5):
311 for j
in xrange(1, 5):
312 unop_horiz("fnoise{0}_{1}".format(i
, j
), i
, tfloat
, j
, tfloat
, "0.0f")
314 def binop_convert(name
, out_type
, in_type
, alg_props
, const_expr
):
315 opcode(name
, 0, out_type
, [0, 0], [in_type
, in_type
], alg_props
, const_expr
)
317 def binop(name
, ty
, alg_props
, const_expr
):
318 binop_convert(name
, ty
, ty
, alg_props
, const_expr
)
320 def binop_compare(name
, ty
, alg_props
, const_expr
):
321 binop_convert(name
, tbool
, ty
, alg_props
, const_expr
)
323 def binop_horiz(name
, out_size
, out_type
, src1_size
, src1_type
, src2_size
,
324 src2_type
, const_expr
):
325 opcode(name
, out_size
, out_type
, [src1_size
, src2_size
], [src1_type
, src2_type
],
328 def binop_reduce(name
, output_size
, output_type
, src_type
, prereduce_expr
,
329 reduce_expr
, final_expr
):
331 return final_expr
.format(src
= "(" + src
+ ")")
332 def reduce_(src0
, src1
):
333 return reduce_expr
.format(src0
=src0
, src1
=src1
)
334 def prereduce(src0
, src1
):
335 return "(" + prereduce_expr
.format(src0
=src0
, src1
=src1
) + ")"
336 src0
= prereduce("src0.x", "src1.x")
337 src1
= prereduce("src0.y", "src1.y")
338 src2
= prereduce("src0.z", "src1.z")
339 src3
= prereduce("src0.w", "src1.w")
340 opcode(name
+ "2", output_size
, output_type
,
341 [2, 2], [src_type
, src_type
], commutative
,
342 final(reduce_(src0
, src1
)))
343 opcode(name
+ "3", output_size
, output_type
,
344 [3, 3], [src_type
, src_type
], commutative
,
345 final(reduce_(reduce_(src0
, src1
), src2
)))
346 opcode(name
+ "4", output_size
, output_type
,
347 [4, 4], [src_type
, src_type
], commutative
,
348 final(reduce_(reduce_(src0
, src1
), reduce_(src2
, src3
))))
350 binop("fadd", tfloat
, commutative
+ associative
, "src0 + src1")
351 binop("iadd", tint
, commutative
+ associative
, "src0 + src1")
352 binop("fsub", tfloat
, "", "src0 - src1")
353 binop("isub", tint
, "", "src0 - src1")
355 binop("fmul", tfloat
, commutative
+ associative
, "src0 * src1")
356 # low 32-bits of signed/unsigned integer multiply
357 binop("imul", tint
, commutative
+ associative
, "src0 * src1")
358 # high 32-bits of signed integer multiply
359 binop("imul_high", tint
, commutative
,
360 "(int32_t)(((int64_t) src0 * (int64_t) src1) >> 32)")
361 # high 32-bits of unsigned integer multiply
362 binop("umul_high", tunsigned
, commutative
,
363 "(uint32_t)(((uint64_t) src0 * (uint64_t) src1) >> 32)")
365 binop("fdiv", tfloat
, "", "src0 / src1")
366 binop("idiv", tint
, "", "src0 / src1")
367 binop("udiv", tunsigned
, "", "src0 / src1")
369 # returns a boolean representing the carry resulting from the addition of
370 # the two unsigned arguments.
372 binop_convert("uadd_carry", tbool
, tunsigned
, commutative
, "src0 + src1 < src0")
374 # returns a boolean representing the borrow resulting from the subtraction
375 # of the two unsigned arguments.
377 binop_convert("usub_borrow", tbool
, tunsigned
, "", "src1 < src0")
379 binop("fmod", tfloat
, "", "src0 - src1 * floorf(src0 / src1)")
380 binop("umod", tunsigned
, "", "src1 == 0 ? 0 : src0 % src1")
387 # these integer-aware comparisons return a boolean (0 or ~0)
389 binop_compare("flt", tfloat
, "", "src0 < src1")
390 binop_compare("fge", tfloat
, "", "src0 >= src1")
391 binop_compare("feq", tfloat
, commutative
, "src0 == src1")
392 binop_compare("fne", tfloat
, commutative
, "src0 != src1")
393 binop_compare("ilt", tint
, "", "src0 < src1")
394 binop_compare("ige", tint
, "", "src0 >= src1")
395 binop_compare("ieq", tint
, commutative
, "src0 == src1")
396 binop_compare("ine", tint
, commutative
, "src0 != src1")
397 binop_compare("ult", tunsigned
, "", "src0 < src1")
398 binop_compare("uge", tunsigned
, "", "src0 >= src1")
400 # integer-aware GLSL-style comparisons that compare floats and ints
402 binop_reduce("ball_fequal", 1, tbool
, tfloat
, "{src0} == {src1}",
403 "{src0} && {src1}", "{src}")
404 binop_reduce("bany_fnequal", 1, tbool
, tfloat
, "{src0} != {src1}",
405 "{src0} || {src1}", "{src}")
406 binop_reduce("ball_iequal", 1, tbool
, tint
, "{src0} == {src1}",
407 "{src0} && {src1}", "{src}")
408 binop_reduce("bany_inequal", 1, tbool
, tint
, "{src0} != {src1}",
409 "{src0} || {src1}", "{src}")
411 # non-integer-aware GLSL-style comparisons that return 0.0 or 1.0
413 binop_reduce("fall_equal", 1, tfloat
, tfloat
, "{src0} == {src1}",
414 "{src0} && {src1}", "{src} ? 1.0f : 0.0f")
415 binop_reduce("fany_nequal", 1, tfloat
, tfloat
, "{src0} != {src1}",
416 "{src0} || {src1}", "{src} ? 1.0f : 0.0f")
418 # These comparisons for integer-less hardware return 1.0 and 0.0 for true
419 # and false respectively
421 binop("slt", tfloat
, "", "(src0 < src1) ? 1.0f : 0.0f") # Set on Less Than
422 binop("sge", tfloat
, "", "(src0 >= src1) ? 1.0f : 0.0f") # Set on Greater or Equal
423 binop("seq", tfloat
, commutative
, "(src0 == src1) ? 1.0f : 0.0f") # Set on Equal
424 binop("sne", tfloat
, commutative
, "(src0 != src1) ? 1.0f : 0.0f") # Set on Not Equal
427 binop("ishl", tint
, "", "src0 << src1")
428 binop("ishr", tint
, "", "src0 >> src1")
429 binop("ushr", tunsigned
, "", "src0 >> src1")
431 # bitwise logic operators
433 # These are also used as boolean and, or, xor for hardware supporting
437 binop("iand", tunsigned
, commutative
+ associative
, "src0 & src1")
438 binop("ior", tunsigned
, commutative
+ associative
, "src0 | src1")
439 binop("ixor", tunsigned
, commutative
+ associative
, "src0 ^ src1")
442 # floating point logic operators
444 # These use (src != 0.0) for testing the truth of the input, and output 1.0
445 # for true and 0.0 for false
447 binop("fand", tfloat
, commutative
,
448 "((src0 != 0.0f) && (src1 != 0.0f)) ? 1.0f : 0.0f")
449 binop("for", tfloat
, commutative
,
450 "((src0 != 0.0f) || (src1 != 0.0f)) ? 1.0f : 0.0f")
451 binop("fxor", tfloat
, commutative
,
452 "(src0 != 0.0f && src1 == 0.0f) || (src0 == 0.0f && src1 != 0.0f) ? 1.0f : 0.0f")
454 binop_reduce("fdot", 1, tfloat
, tfloat
, "{src0} * {src1}", "{src0} + {src1}",
457 binop("fmin", tfloat
, "", "fminf(src0, src1)")
458 binop("imin", tint
, commutative
+ associative
, "src1 > src0 ? src0 : src1")
459 binop("umin", tunsigned
, commutative
+ associative
, "src1 > src0 ? src0 : src1")
460 binop("fmax", tfloat
, "", "fmaxf(src0, src1)")
461 binop("imax", tint
, commutative
+ associative
, "src1 > src0 ? src1 : src0")
462 binop("umax", tunsigned
, commutative
+ associative
, "src1 > src0 ? src1 : src0")
464 binop("fpow", tfloat
, "", "powf(src0, src1)")
466 binop_horiz("pack_half_2x16_split", 1, tunsigned
, 1, tfloat
, 1, tfloat
,
467 "pack_half_1x16(src0.x) | (pack_half_1x16(src1.x) << 16)")
469 binop_convert("bfm", tunsigned
, tint
, "", """
470 int offset = src0, bits = src1;
471 if (offset < 0 || bits < 0 || offset + bits > 32)
472 dst = 0; /* undefined per the spec */
474 dst = ((1 << bits)- 1) << offset;
477 opcode("ldexp", 0, tfloat
, [0, 0], [tfloat
, tint
], "", """
478 dst = ldexp(src0, src1);
479 /* flush denormals to zero. */
481 dst = copysign(0.0f, src0);
484 # Combines the first component of each input to make a 2-component vector.
486 binop_horiz("vec2", 2, tunsigned
, 1, tunsigned
, 1, tunsigned
, """
491 def triop(name
, ty
, const_expr
):
492 opcode(name
, 0, ty
, [0, 0, 0], [ty
, ty
, ty
], "", const_expr
)
493 def triop_horiz(name
, output_size
, src1_size
, src2_size
, src3_size
, const_expr
):
494 opcode(name
, output_size
, tunsigned
,
495 [src1_size
, src2_size
, src3_size
],
496 [tunsigned
, tunsigned
, tunsigned
], "", const_expr
)
498 triop("ffma", tfloat
, "src0 * src1 + src2")
500 triop("flrp", tfloat
, "src0 * (1 - src2) + src1 * src2")
504 # A vector conditional select instruction (like ?:, but operating per-
505 # component on vectors). There are two versions, one for floating point
506 # bools (0.0 vs 1.0) and one for integer bools (0 vs ~0).
509 triop("fcsel", tfloat
, "(src0 != 0.0f) ? src1 : src2")
510 opcode("bcsel", 0, tunsigned
, [0, 0, 0],
511 [tbool
, tunsigned
, tunsigned
], "", "src0 ? src1 : src2")
513 triop("bfi", tunsigned
, """
514 unsigned mask = src0, insert = src1 & mask, base = src2;
523 dst = (base & ~mask) | insert;
527 opcode("ubitfield_extract", 0, tunsigned
,
528 [0, 1, 1], [tunsigned
, tint
, tint
], "", """
529 unsigned base = src0;
530 int offset = src1.x, bits = src2.x;
533 } else if (bits < 0 || offset < 0 || offset + bits > 32) {
534 dst = 0; /* undefined per the spec */
536 dst = (base >> offset) & ((1 << bits) - 1);
539 opcode("ibitfield_extract", 0, tint
,
540 [0, 1, 1], [tint
, tint
, tint
], "", """
542 int offset = src1.x, bits = src2.x;
545 } else if (offset < 0 || bits < 0 || offset + bits > 32) {
548 dst = (base << (32 - offset - bits)) >> offset; /* use sign-extending shift */
552 # Combines the first component of each input to make a 3-component vector.
554 triop_horiz("vec3", 3, 1, 1, 1, """
560 def quadop_horiz(name
, output_size
, src1_size
, src2_size
, src3_size
,
561 src4_size
, const_expr
):
562 opcode(name
, output_size
, tunsigned
,
563 [src1_size
, src2_size
, src3_size
, src4_size
],
564 [tunsigned
, tunsigned
, tunsigned
, tunsigned
],
567 opcode("bitfield_insert", 0, tunsigned
, [0, 0, 1, 1],
568 [tunsigned
, tunsigned
, tint
, tint
], "", """
569 unsigned base = src0, insert = src1;
570 int offset = src2.x, bits = src3.x;
573 } else if (offset < 0 || bits < 0 || bits + offset > 32) {
576 unsigned mask = ((1 << bits) - 1) << offset;
577 dst = (base & ~mask) | ((insert << bits) & mask);
581 quadop_horiz("vec4", 4, 1, 1, 1, 1, """