2
from nir_opcodes import opcodes
3
from nir_opcodes import type_has_size, type_size, type_sizes, type_base_type
5
def type_add_size(type_, size):
6
if type_has_size(type_):
8
return type_ + str(size)
12
if not type_has_size(op.output_type):
13
sizes = set(type_sizes(op.output_type))
15
for input_type in op.input_types:
16
if not type_has_size(input_type):
18
sizes = set(type_sizes(input_type))
20
sizes = sizes.intersection(set(type_sizes(input_type)))
22
return sorted(list(sizes)) if sizes is not None else None
24
def get_const_field(type_):
25
if type_size(type_) == 1:
27
elif type_base_type(type_) == 'bool':
28
return 'i' + str(type_size(type_))
29
elif type_ == "float16":
32
return type_base_type(type_)[0] + str(type_size(type_))
36
* Copyright (C) 2014 Intel Corporation
38
* Permission is hereby granted, free of charge, to any person obtaining a
39
* copy of this software and associated documentation files (the "Software"),
40
* to deal in the Software without restriction, including without limitation
41
* the rights to use, copy, modify, merge, publish, distribute, sublicense,
42
* and/or sell copies of the Software, and to permit persons to whom the
43
* Software is furnished to do so, subject to the following conditions:
45
* The above copyright notice and this permission notice (including the next
46
* paragraph) shall be included in all copies or substantial portions of the
49
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
50
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
51
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
52
* THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
53
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
54
* FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
58
* Jason Ekstrand (jason@jlekstrand.net)
62
#include "util/rounding.h" /* for _mesa_roundeven */
63
#include "util/half_float.h"
64
#include "util/double.h"
65
#include "util/softfloat.h"
66
#include "util/bigmath.h"
67
#include "util/format/format_utils.h"
68
#include "nir_constant_expressions.h"
71
* \brief Checks if the provided value is a denorm and flushes it to zero.
74
constant_denorm_flush_to_zero(nir_const_value *value, unsigned bit_size)
78
if (0 == (value->u64 & 0x7ff0000000000000))
79
value->u64 &= 0x8000000000000000;
82
if (0 == (value->u32 & 0x7f800000))
83
value->u32 &= 0x80000000;
86
if (0 == (value->u16 & 0x7c00))
92
* Evaluate one component of packSnorm4x8.
95
pack_snorm_1x8(float x)
97
/* From section 8.4 of the GLSL 4.30 spec:
101
* The conversion for component c of v to fixed point is done as
104
* packSnorm4x8: round(clamp(c, -1, +1) * 127.0)
106
* We must first cast the float to an int, because casting a negative
107
* float to a uint is undefined.
109
return (uint8_t) (int)
110
_mesa_roundevenf(CLAMP(x, -1.0f, +1.0f) * 127.0f);
114
* Evaluate one component of packSnorm2x16.
117
pack_snorm_1x16(float x)
119
/* From section 8.4 of the GLSL ES 3.00 spec:
123
* The conversion for component c of v to fixed point is done as
126
* packSnorm2x16: round(clamp(c, -1, +1) * 32767.0)
128
* We must first cast the float to an int, because casting a negative
129
* float to a uint is undefined.
131
return (uint16_t) (int)
132
_mesa_roundevenf(CLAMP(x, -1.0f, +1.0f) * 32767.0f);
136
* Evaluate one component of unpackSnorm4x8.
139
unpack_snorm_1x8(uint8_t u)
141
/* From section 8.4 of the GLSL 4.30 spec:
145
* The conversion for unpacked fixed-point value f to floating point is
148
* unpackSnorm4x8: clamp(f / 127.0, -1, +1)
150
return CLAMP((int8_t) u / 127.0f, -1.0f, +1.0f);
154
* Evaluate one component of unpackSnorm2x16.
157
unpack_snorm_1x16(uint16_t u)
159
/* From section 8.4 of the GLSL ES 3.00 spec:
163
* The conversion for unpacked fixed-point value f to floating point is
166
* unpackSnorm2x16: clamp(f / 32767.0, -1, +1)
168
return CLAMP((int16_t) u / 32767.0f, -1.0f, +1.0f);
172
* Evaluate one component packUnorm4x8.
175
pack_unorm_1x8(float x)
177
/* From section 8.4 of the GLSL 4.30 spec:
181
* The conversion for component c of v to fixed point is done as
184
* packUnorm4x8: round(clamp(c, 0, +1) * 255.0)
186
return (uint8_t) (int)
187
_mesa_roundevenf(CLAMP(x, 0.0f, 1.0f) * 255.0f);
191
* Evaluate one component packUnorm2x16.
194
pack_unorm_1x16(float x)
196
/* From section 8.4 of the GLSL ES 3.00 spec:
200
* The conversion for component c of v to fixed point is done as
203
* packUnorm2x16: round(clamp(c, 0, +1) * 65535.0)
205
return (uint16_t) (int)
206
_mesa_roundevenf(CLAMP(x, 0.0f, 1.0f) * 65535.0f);
210
* Evaluate one component of unpackUnorm4x8.
213
unpack_unorm_1x8(uint8_t u)
215
/* From section 8.4 of the GLSL 4.30 spec:
219
* The conversion for unpacked fixed-point value f to floating point is
222
* unpackUnorm4x8: f / 255.0
224
return (float) u / 255.0f;
228
* Evaluate one component of unpackUnorm2x16.
231
unpack_unorm_1x16(uint16_t u)
233
/* From section 8.4 of the GLSL ES 3.00 spec:
237
* The conversion for unpacked fixed-point value f to floating point is
240
* unpackUnorm2x16: f / 65535.0
242
return (float) u / 65535.0f;
246
* Evaluate one component of packHalf2x16.
249
pack_half_1x16(float x)
251
return _mesa_float_to_half(x);
255
* Evaluate one component of unpackHalf2x16.
258
unpack_half_1x16_flush_to_zero(uint16_t u)
260
if (0 == (u & 0x7c00))
262
return _mesa_half_to_float(u);
266
* Evaluate one component of unpackHalf2x16.
269
unpack_half_1x16(uint16_t u)
271
return _mesa_half_to_float(u);
274
/* Some typed vector structures to make things like src0.y work */
275
typedef int8_t int1_t;
276
typedef uint8_t uint1_t;
277
typedef float float16_t;
278
typedef float float32_t;
279
typedef double float64_t;
280
typedef bool bool1_t;
281
typedef bool bool8_t;
282
typedef bool bool16_t;
283
typedef bool bool32_t;
284
typedef bool bool64_t;
285
% for type in ["float", "int", "uint", "bool"]:
286
% for width in type_sizes(type):
287
struct ${type}${width}_vec {
308
<%def name="evaluate_op(op, bit_size, execution_mode)">
310
output_type = type_add_size(op.output_type, bit_size)
311
input_types = [type_add_size(type_, bit_size) for type_ in op.input_types]
314
## For each non-per-component input, create a variable srcN that
315
## contains x, y, z, and w elements which are filled in with the
316
## appropriately-typed values.
317
% for j in range(op.num_inputs):
318
% if op.input_sizes[j] == 0:
320
% elif "src" + str(j) not in op.const_expr:
321
## Avoid unused variable warnings
325
const struct ${input_types[j]}_vec src${j} = {
326
% for k in range(op.input_sizes[j]):
327
% if input_types[j] == "int1":
328
/* 1-bit integers use a 0/-1 convention */
329
-(int1_t)_src[${j}][${k}].b,
330
% elif input_types[j] == "float16":
331
_mesa_half_to_float(_src[${j}][${k}].u16),
333
_src[${j}][${k}].${get_const_field(input_types[j])},
336
% for k in range(op.input_sizes[j], 16):
342
% if op.output_size == 0:
343
## For per-component instructions, we need to iterate over the
344
## components and apply the constant expression one component
346
for (unsigned _i = 0; _i < num_components; _i++) {
347
## For each per-component input, create a variable srcN that
348
## contains the value of the current (_i'th) component.
349
% for j in range(op.num_inputs):
350
% if op.input_sizes[j] != 0:
352
% elif "src" + str(j) not in op.const_expr:
353
## Avoid unused variable warnings
355
% elif input_types[j] == "int1":
356
/* 1-bit integers use a 0/-1 convention */
357
const int1_t src${j} = -(int1_t)_src[${j}][_i].b;
358
% elif input_types[j] == "float16":
359
const float src${j} =
360
_mesa_half_to_float(_src[${j}][_i].u16);
362
const ${input_types[j]}_t src${j} =
363
_src[${j}][_i].${get_const_field(input_types[j])};
367
## Create an appropriately-typed variable dst and assign the
368
## result of the const_expr to it. If const_expr already contains
369
## writes to dst, just include const_expr directly.
370
% if "dst" in op.const_expr:
371
${output_type}_t dst;
375
${output_type}_t dst = ${op.const_expr};
378
## Store the current component of the actual destination to the
380
% if output_type == "int1" or output_type == "uint1":
381
/* 1-bit integers get truncated */
382
_dst_val[_i].b = dst & 1;
383
% elif output_type.startswith("bool"):
384
## Sanitize the C value to a proper NIR 0/-1 bool
385
_dst_val[_i].${get_const_field(output_type)} = -(int)dst;
386
% elif output_type == "float16":
387
if (nir_is_rounding_mode_rtz(execution_mode, 16)) {
388
_dst_val[_i].u16 = _mesa_float_to_float16_rtz(dst);
390
_dst_val[_i].u16 = _mesa_float_to_float16_rtne(dst);
393
_dst_val[_i].${get_const_field(output_type)} = dst;
396
% if op.name != "fquantize2f16" and type_base_type(output_type) == "float":
397
% if type_has_size(output_type):
398
if (nir_is_denorm_flush_to_zero(execution_mode, ${type_size(output_type)})) {
399
constant_denorm_flush_to_zero(&_dst_val[_i], ${type_size(output_type)});
402
if (nir_is_denorm_flush_to_zero(execution_mode, ${bit_size})) {
403
constant_denorm_flush_to_zero(&_dst_val[i], bit_size);
409
## In the non-per-component case, create a struct dst with
410
## appropriately-typed elements x, y, z, and w and assign the result
411
## of the const_expr to all components of dst, or include the
412
## const_expr directly if it writes to dst already.
413
struct ${output_type}_vec dst;
415
% if "dst" in op.const_expr:
418
## Splat the value to all components. This way expressions which
419
## write the same value to all components don't need to explicitly
421
dst.x = dst.y = dst.z = dst.w = ${op.const_expr};
424
## For each component in the destination, copy the value of dst to
425
## the actual destination.
426
% for k in range(op.output_size):
427
% if output_type == "int1" or output_type == "uint1":
428
/* 1-bit integers get truncated */
429
_dst_val[${k}].b = dst.${"xyzwefghijklmnop"[k]} & 1;
430
% elif output_type.startswith("bool"):
431
## Sanitize the C value to a proper NIR 0/-1 bool
432
_dst_val[${k}].${get_const_field(output_type)} = -(int)dst.${"xyzwefghijklmnop"[k]};
433
% elif output_type == "float16":
434
if (nir_is_rounding_mode_rtz(execution_mode, 16)) {
435
_dst_val[${k}].u16 = _mesa_float_to_float16_rtz(dst.${"xyzwefghijklmnop"[k]});
437
_dst_val[${k}].u16 = _mesa_float_to_float16_rtne(dst.${"xyzwefghijklmnop"[k]});
440
_dst_val[${k}].${get_const_field(output_type)} = dst.${"xyzwefghijklmnop"[k]};
443
% if op.name != "fquantize2f16" and type_base_type(output_type) == "float":
444
% if type_has_size(output_type):
445
if (nir_is_denorm_flush_to_zero(execution_mode, ${type_size(output_type)})) {
446
constant_denorm_flush_to_zero(&_dst_val[${k}], ${type_size(output_type)});
449
if (nir_is_denorm_flush_to_zero(execution_mode, ${bit_size})) {
450
constant_denorm_flush_to_zero(&_dst_val[${k}], bit_size);
458
% for name, op in sorted(opcodes.items()):
459
% if op.name == "fsat":
460
#if defined(_MSC_VER) && (defined(_M_ARM64) || defined(_M_ARM64EC))
461
#pragma optimize("", off) /* Temporary work-around for MSVC compiler bug, present in VS2019 16.9.2 */
465
evaluate_${name}(nir_const_value *_dst_val,
466
UNUSED unsigned num_components,
467
${"UNUSED" if op_bit_sizes(op) is None else ""} unsigned bit_size,
468
UNUSED nir_const_value **_src,
469
UNUSED unsigned execution_mode)
471
% if op_bit_sizes(op) is not None:
473
% for bit_size in op_bit_sizes(op):
475
${evaluate_op(op, bit_size, execution_mode)}
481
unreachable("unknown bit width");
484
${evaluate_op(op, 0, execution_mode)}
487
% if op.name == "fsat":
488
#if defined(_MSC_VER) && (defined(_M_ARM64) || defined(_M_ARM64EC))
489
#pragma optimize("", on) /* Temporary work-around for MSVC compiler bug, present in VS2019 16.9.2 */
495
nir_eval_const_opcode(nir_op op, nir_const_value *dest,
496
unsigned num_components, unsigned bit_width,
497
nir_const_value **src,
498
unsigned float_controls_execution_mode)
501
% for name in sorted(opcodes.keys()):
503
evaluate_${name}(dest, num_components, bit_width, src, float_controls_execution_mode);
507
unreachable("shouldn't get here");
511
from mako.template import Template
513
print(Template(template).render(opcodes=opcodes, type_sizes=type_sizes,
514
type_base_type=type_base_type,
516
type_has_size=type_has_size,
517
type_add_size=type_add_size,
518
op_bit_sizes=op_bit_sizes,
519
get_const_field=get_const_field))