| 1 |
// Copyright (c) 2018 Google LLC |
| 2 |
// |
| 3 |
// Licensed under the Apache License, Version 2.0 (the "License"); |
| 4 |
// you may not use this file except in compliance with the License. |
| 5 |
// You may obtain a copy of the License at |
| 6 |
// |
| 7 |
// http://www.apache.org/licenses/LICENSE-2.0 |
| 8 |
// |
| 9 |
// Unless required by applicable law or agreed to in writing, software |
| 10 |
// distributed under the License is distributed on an "AS IS" BASIS, |
| 11 |
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 12 |
// See the License for the specific language governing permissions and |
| 13 |
// limitations under the License. |
| 14 |
|
| 15 |
#include "source/opt/const_folding_rules.h" |
| 16 |
|
| 17 |
#include "source/opt/ir_context.h" |
| 18 |
|
| 19 |
namespace spvtools { |
| 20 |
namespace opt { |
| 21 |
namespace { |
| 22 |
|
| 23 |
const uint32_t kExtractCompositeIdInIdx = 0; |
| 24 |
|
| 25 |
// Returns true if |type| is Float or a vector of Float. |
| 26 |
bool HasFloatingPoint(const analysis::Type* type) { |
| 27 |
if (type->AsFloat()) { |
| 28 |
return true; |
| 29 |
} else if (const analysis::Vector* vec_type = type->AsVector()) { |
| 30 |
return vec_type->element_type()->AsFloat() != nullptr; |
| 31 |
} |
| 32 |
|
| 33 |
return false; |
| 34 |
} |
| 35 |
|
| 36 |
// Folds an OpcompositeExtract where input is a composite constant. |
| 37 |
ConstantFoldingRule FoldExtractWithConstants() { |
| 38 |
return [](IRContext* context, Instruction* inst, |
| 39 |
const std::vector<const analysis::Constant*>& constants) |
| 40 |
-> const analysis::Constant* { |
| 41 |
const analysis::Constant* c = constants[kExtractCompositeIdInIdx]; |
| 42 |
if (c == nullptr) { |
| 43 |
return nullptr; |
| 44 |
} |
| 45 |
|
| 46 |
for (uint32_t i = 1; i < inst->NumInOperands(); ++i) { |
| 47 |
uint32_t element_index = inst->GetSingleWordInOperand(i); |
| 48 |
if (c->AsNullConstant()) { |
| 49 |
// Return Null for the return type. |
| 50 |
analysis::ConstantManager* const_mgr = context->get_constant_mgr(); |
| 51 |
analysis::TypeManager* type_mgr = context->get_type_mgr(); |
| 52 |
return const_mgr->GetConstant(type_mgr->GetType(inst->type_id()), {}); |
| 53 |
} |
| 54 |
|
| 55 |
auto cc = c->AsCompositeConstant(); |
| 56 |
assert(cc != nullptr); |
| 57 |
auto components = cc->GetComponents(); |
| 58 |
// Protect against invalid IR. Refuse to fold if the index is out |
| 59 |
// of bounds. |
| 60 |
if (element_index >= components.size()) return nullptr; |
| 61 |
c = components[element_index]; |
| 62 |
} |
| 63 |
return c; |
| 64 |
}; |
| 65 |
} |
| 66 |
|
| 67 |
ConstantFoldingRule FoldVectorShuffleWithConstants() { |
| 68 |
return [](IRContext* context, Instruction* inst, |
| 69 |
const std::vector<const analysis::Constant*>& constants) |
| 70 |
-> const analysis::Constant* { |
| 71 |
assert(inst->opcode() == SpvOpVectorShuffle); |
| 72 |
const analysis::Constant* c1 = constants[0]; |
| 73 |
const analysis::Constant* c2 = constants[1]; |
| 74 |
if (c1 == nullptr || c2 == nullptr) { |
| 75 |
return nullptr; |
| 76 |
} |
| 77 |
|
| 78 |
analysis::ConstantManager* const_mgr = context->get_constant_mgr(); |
| 79 |
const analysis::Type* element_type = c1->type()->AsVector()->element_type(); |
| 80 |
|
| 81 |
std::vector<const analysis::Constant*> c1_components; |
| 82 |
if (const analysis::VectorConstant* vec_const = c1->AsVectorConstant()) { |
| 83 |
c1_components = vec_const->GetComponents(); |
| 84 |
} else { |
| 85 |
assert(c1->AsNullConstant()); |
| 86 |
const analysis::Constant* element = |
| 87 |
const_mgr->GetConstant(element_type, {}); |
| 88 |
c1_components.resize(c1->type()->AsVector()->element_count(), element); |
| 89 |
} |
| 90 |
std::vector<const analysis::Constant*> c2_components; |
| 91 |
if (const analysis::VectorConstant* vec_const = c2->AsVectorConstant()) { |
| 92 |
c2_components = vec_const->GetComponents(); |
| 93 |
} else { |
| 94 |
assert(c2->AsNullConstant()); |
| 95 |
const analysis::Constant* element = |
| 96 |
const_mgr->GetConstant(element_type, {}); |
| 97 |
c2_components.resize(c2->type()->AsVector()->element_count(), element); |
| 98 |
} |
| 99 |
|
| 100 |
std::vector<uint32_t> ids; |
| 101 |
const uint32_t undef_literal_value = 0xffffffff; |
| 102 |
for (uint32_t i = 2; i < inst->NumInOperands(); ++i) { |
| 103 |
uint32_t index = inst->GetSingleWordInOperand(i); |
| 104 |
if (index == undef_literal_value) { |
| 105 |
// Don't fold shuffle with undef literal value. |
| 106 |
return nullptr; |
| 107 |
} else if (index < c1_components.size()) { |
| 108 |
Instruction* member_inst = |
| 109 |
const_mgr->GetDefiningInstruction(c1_components[index]); |
| 110 |
ids.push_back(member_inst->result_id()); |
| 111 |
} else { |
| 112 |
Instruction* member_inst = const_mgr->GetDefiningInstruction( |
| 113 |
c2_components[index - c1_components.size()]); |
| 114 |
ids.push_back(member_inst->result_id()); |
| 115 |
} |
| 116 |
} |
| 117 |
|
| 118 |
analysis::TypeManager* type_mgr = context->get_type_mgr(); |
| 119 |
return const_mgr->GetConstant(type_mgr->GetType(inst->type_id()), ids); |
| 120 |
}; |
| 121 |
} |
| 122 |
|
| 123 |
ConstantFoldingRule FoldVectorTimesScalar() { |
| 124 |
return [](IRContext* context, Instruction* inst, |
| 125 |
const std::vector<const analysis::Constant*>& constants) |
| 126 |
-> const analysis::Constant* { |
| 127 |
assert(inst->opcode() == SpvOpVectorTimesScalar); |
| 128 |
analysis::ConstantManager* const_mgr = context->get_constant_mgr(); |
| 129 |
analysis::TypeManager* type_mgr = context->get_type_mgr(); |
| 130 |
|
| 131 |
if (!inst->IsFloatingPointFoldingAllowed()) { |
| 132 |
if (HasFloatingPoint(type_mgr->GetType(inst->type_id()))) { |
| 133 |
return nullptr; |
| 134 |
} |
| 135 |
} |
| 136 |
|
| 137 |
const analysis::Constant* c1 = constants[0]; |
| 138 |
const analysis::Constant* c2 = constants[1]; |
| 139 |
|
| 140 |
if (c1 && c1->IsZero()) { |
| 141 |
return c1; |
| 142 |
} |
| 143 |
|
| 144 |
if (c2 && c2->IsZero()) { |
| 145 |
// Get or create the NullConstant for this type. |
| 146 |
std::vector<uint32_t> ids; |
| 147 |
return const_mgr->GetConstant(type_mgr->GetType(inst->type_id()), ids); |
| 148 |
} |
| 149 |
|
| 150 |
if (c1 == nullptr || c2 == nullptr) { |
| 151 |
return nullptr; |
| 152 |
} |
| 153 |
|
| 154 |
// Check result type. |
| 155 |
const analysis::Type* result_type = type_mgr->GetType(inst->type_id()); |
| 156 |
const analysis::Vector* vector_type = result_type->AsVector(); |
| 157 |
assert(vector_type != nullptr); |
| 158 |
const analysis::Type* element_type = vector_type->element_type(); |
| 159 |
assert(element_type != nullptr); |
| 160 |
const analysis::Float* float_type = element_type->AsFloat(); |
| 161 |
assert(float_type != nullptr); |
| 162 |
|
| 163 |
// Check types of c1 and c2. |
| 164 |
assert(c1->type()->AsVector() == vector_type); |
| 165 |
assert(c1->type()->AsVector()->element_type() == element_type && |
| 166 |
c2->type() == element_type); |
| 167 |
|
| 168 |
// Get a float vector that is the result of vector-times-scalar. |
| 169 |
std::vector<const analysis::Constant*> c1_components = |
| 170 |
c1->GetVectorComponents(const_mgr); |
| 171 |
std::vector<uint32_t> ids; |
| 172 |
if (float_type->width() == 32) { |
| 173 |
float scalar = c2->GetFloat(); |
| 174 |
for (uint32_t i = 0; i < c1_components.size(); ++i) { |
| 175 |
utils::FloatProxy<float> result(c1_components[i]->GetFloat() * scalar); |
| 176 |
std::vector<uint32_t> words = result.GetWords(); |
| 177 |
const analysis::Constant* new_elem = |
| 178 |
const_mgr->GetConstant(float_type, words); |
| 179 |
ids.push_back(const_mgr->GetDefiningInstruction(new_elem)->result_id()); |
| 180 |
} |
| 181 |
return const_mgr->GetConstant(vector_type, ids); |
| 182 |
} else if (float_type->width() == 64) { |
| 183 |
double scalar = c2->GetDouble(); |
| 184 |
for (uint32_t i = 0; i < c1_components.size(); ++i) { |
| 185 |
utils::FloatProxy<double> result(c1_components[i]->GetDouble() * |
| 186 |
scalar); |
| 187 |
std::vector<uint32_t> words = result.GetWords(); |
| 188 |
const analysis::Constant* new_elem = |
| 189 |
const_mgr->GetConstant(float_type, words); |
| 190 |
ids.push_back(const_mgr->GetDefiningInstruction(new_elem)->result_id()); |
| 191 |
} |
| 192 |
return const_mgr->GetConstant(vector_type, ids); |
| 193 |
} |
| 194 |
return nullptr; |
| 195 |
}; |
| 196 |
} |
| 197 |
|
| 198 |
ConstantFoldingRule FoldCompositeWithConstants() { |
| 199 |
// Folds an OpCompositeConstruct where all of the inputs are constants to a |
| 200 |
// constant. A new constant is created if necessary. |
| 201 |
return [](IRContext* context, Instruction* inst, |
| 202 |
const std::vector<const analysis::Constant*>& constants) |
| 203 |
-> const analysis::Constant* { |
| 204 |
analysis::ConstantManager* const_mgr = context->get_constant_mgr(); |
| 205 |
analysis::TypeManager* type_mgr = context->get_type_mgr(); |
| 206 |
const analysis::Type* new_type = type_mgr->GetType(inst->type_id()); |
| 207 |
Instruction* type_inst = |
| 208 |
context->get_def_use_mgr()->GetDef(inst->type_id()); |
| 209 |
|
| 210 |
std::vector<uint32_t> ids; |
| 211 |
for (uint32_t i = 0; i < constants.size(); ++i) { |
| 212 |
const analysis::Constant* element_const = constants[i]; |
| 213 |
if (element_const == nullptr) { |
| 214 |
return nullptr; |
| 215 |
} |
| 216 |
|
| 217 |
uint32_t component_type_id = 0; |
| 218 |
if (type_inst->opcode() == SpvOpTypeStruct) { |
| 219 |
component_type_id = type_inst->GetSingleWordInOperand(i); |
| 220 |
} else if (type_inst->opcode() == SpvOpTypeArray) { |
| 221 |
component_type_id = type_inst->GetSingleWordInOperand(0); |
| 222 |
} |
| 223 |
|
| 224 |
uint32_t element_id = |
| 225 |
const_mgr->FindDeclaredConstant(element_const, component_type_id); |
| 226 |
if (element_id == 0) { |
| 227 |
return nullptr; |
| 228 |
} |
| 229 |
ids.push_back(element_id); |
| 230 |
} |
| 231 |
return const_mgr->GetConstant(new_type, ids); |
| 232 |
}; |
| 233 |
} |
| 234 |
|
| 235 |
// The interface for a function that returns the result of applying a scalar |
| 236 |
// floating-point binary operation on |a| and |b|. The type of the return value |
| 237 |
// will be |type|. The input constants must also be of type |type|. |
| 238 |
using UnaryScalarFoldingRule = std::function<const analysis::Constant*( |
| 239 |
const analysis::Type* result_type, const analysis::Constant* a, |
| 240 |
analysis::ConstantManager*)>; |
| 241 |
|
| 242 |
// The interface for a function that returns the result of applying a scalar |
| 243 |
// floating-point binary operation on |a| and |b|. The type of the return value |
| 244 |
// will be |type|. The input constants must also be of type |type|. |
| 245 |
using BinaryScalarFoldingRule = std::function<const analysis::Constant*( |
| 246 |
const analysis::Type* result_type, const analysis::Constant* a, |
| 247 |
const analysis::Constant* b, analysis::ConstantManager*)>; |
| 248 |
|
| 249 |
// Returns a |ConstantFoldingRule| that folds unary floating point scalar ops |
| 250 |
// using |scalar_rule| and unary float point vectors ops by applying |
| 251 |
// |scalar_rule| to the elements of the vector. The |ConstantFoldingRule| |
| 252 |
// that is returned assumes that |constants| contains 1 entry. If they are |
| 253 |
// not |nullptr|, then their type is either |Float| or |Integer| or a |Vector| |
| 254 |
// whose element type is |Float| or |Integer|. |
| 255 |
ConstantFoldingRule FoldFPUnaryOp(UnaryScalarFoldingRule scalar_rule) { |
| 256 |
return [scalar_rule](IRContext* context, Instruction* inst, |
| 257 |
const std::vector<const analysis::Constant*>& constants) |
| 258 |
-> const analysis::Constant* { |
| 259 |
analysis::ConstantManager* const_mgr = context->get_constant_mgr(); |
| 260 |
analysis::TypeManager* type_mgr = context->get_type_mgr(); |
| 261 |
const analysis::Type* result_type = type_mgr->GetType(inst->type_id()); |
| 262 |
const analysis::Vector* vector_type = result_type->AsVector(); |
| 263 |
|
| 264 |
if (!inst->IsFloatingPointFoldingAllowed()) { |
| 265 |
return nullptr; |
| 266 |
} |
| 267 |
|
| 268 |
const analysis::Constant* arg = |
| 269 |
(inst->opcode() == SpvOpExtInst) ? constants[1] : constants[0]; |
| 270 |
|
| 271 |
if (arg == nullptr) { |
| 272 |
return nullptr; |
| 273 |
} |
| 274 |
|
| 275 |
if (vector_type != nullptr) { |
| 276 |
std::vector<const analysis::Constant*> a_components; |
| 277 |
std::vector<const analysis::Constant*> results_components; |
| 278 |
|
| 279 |
a_components = arg->GetVectorComponents(const_mgr); |
| 280 |
|
| 281 |
// Fold each component of the vector. |
| 282 |
for (uint32_t i = 0; i < a_components.size(); ++i) { |
| 283 |
results_components.push_back(scalar_rule(vector_type->element_type(), |
| 284 |
a_components[i], const_mgr)); |
| 285 |
if (results_components[i] == nullptr) { |
| 286 |
return nullptr; |
| 287 |
} |
| 288 |
} |
| 289 |
|
| 290 |
// Build the constant object and return it. |
| 291 |
std::vector<uint32_t> ids; |
| 292 |
for (const analysis::Constant* member : results_components) { |
| 293 |
ids.push_back(const_mgr->GetDefiningInstruction(member)->result_id()); |
| 294 |
} |
| 295 |
return const_mgr->GetConstant(vector_type, ids); |
| 296 |
} else { |
| 297 |
return scalar_rule(result_type, arg, const_mgr); |
| 298 |
} |
| 299 |
}; |
| 300 |
} |
| 301 |
|
| 302 |
// Returns the result of folding the constants in |constants| according the |
| 303 |
// |scalar_rule|. If |result_type| is a vector, then |scalar_rule| is applied |
| 304 |
// per component. |
| 305 |
const analysis::Constant* FoldFPBinaryOp( |
| 306 |
BinaryScalarFoldingRule scalar_rule, uint32_t result_type_id, |
| 307 |
const std::vector<const analysis::Constant*>& constants, |
| 308 |
IRContext* context) { |
| 309 |
analysis::ConstantManager* const_mgr = context->get_constant_mgr(); |
| 310 |
analysis::TypeManager* type_mgr = context->get_type_mgr(); |
| 311 |
const analysis::Type* result_type = type_mgr->GetType(result_type_id); |
| 312 |
const analysis::Vector* vector_type = result_type->AsVector(); |
| 313 |
|
| 314 |
if (constants[0] == nullptr || constants[1] == nullptr) { |
| 315 |
return nullptr; |
| 316 |
} |
| 317 |
|
| 318 |
if (vector_type != nullptr) { |
| 319 |
std::vector<const analysis::Constant*> a_components; |
| 320 |
std::vector<const analysis::Constant*> b_components; |
| 321 |
std::vector<const analysis::Constant*> results_components; |
| 322 |
|
| 323 |
a_components = constants[0]->GetVectorComponents(const_mgr); |
| 324 |
b_components = constants[1]->GetVectorComponents(const_mgr); |
| 325 |
|
| 326 |
// Fold each component of the vector. |
| 327 |
for (uint32_t i = 0; i < a_components.size(); ++i) { |
| 328 |
results_components.push_back(scalar_rule(vector_type->element_type(), |
| 329 |
a_components[i], b_components[i], |
| 330 |
const_mgr)); |
| 331 |
if (results_components[i] == nullptr) { |
| 332 |
return nullptr; |
| 333 |
} |
| 334 |
} |
| 335 |
|
| 336 |
// Build the constant object and return it. |
| 337 |
std::vector<uint32_t> ids; |
| 338 |
for (const analysis::Constant* member : results_components) { |
| 339 |
ids.push_back(const_mgr->GetDefiningInstruction(member)->result_id()); |
| 340 |
} |
| 341 |
return const_mgr->GetConstant(vector_type, ids); |
| 342 |
} else { |
| 343 |
return scalar_rule(result_type, constants[0], constants[1], const_mgr); |
| 344 |
} |
| 345 |
} |
| 346 |
|
| 347 |
// Returns a |ConstantFoldingRule| that folds floating point scalars using |
| 348 |
// |scalar_rule| and vectors of floating point by applying |scalar_rule| to the |
| 349 |
// elements of the vector. The |ConstantFoldingRule| that is returned assumes |
| 350 |
// that |constants| contains 2 entries. If they are not |nullptr|, then their |
| 351 |
// type is either |Float| or a |Vector| whose element type is |Float|. |
| 352 |
ConstantFoldingRule FoldFPBinaryOp(BinaryScalarFoldingRule scalar_rule) { |
| 353 |
return [scalar_rule](IRContext* context, Instruction* inst, |
| 354 |
const std::vector<const analysis::Constant*>& constants) |
| 355 |
-> const analysis::Constant* { |
| 356 |
if (!inst->IsFloatingPointFoldingAllowed()) { |
| 357 |
return nullptr; |
| 358 |
} |
| 359 |
if (inst->opcode() == SpvOpExtInst) { |
| 360 |
return FoldFPBinaryOp(scalar_rule, inst->type_id(), |
| 361 |
{constants[1], constants[2]}, context); |
| 362 |
} |
| 363 |
return FoldFPBinaryOp(scalar_rule, inst->type_id(), constants, context); |
| 364 |
}; |
| 365 |
} |
| 366 |
|
| 367 |
// This macro defines a |UnaryScalarFoldingRule| that performs float to |
| 368 |
// integer conversion. |
| 369 |
// TODO(greg-lunarg): Support for 64-bit integer types. |
| 370 |
UnaryScalarFoldingRule FoldFToIOp() { |
| 371 |
return [](const analysis::Type* result_type, const analysis::Constant* a, |
| 372 |
analysis::ConstantManager* const_mgr) -> const analysis::Constant* { |
| 373 |
assert(result_type != nullptr && a != nullptr); |
| 374 |
const analysis::Integer* integer_type = result_type->AsInteger(); |
| 375 |
const analysis::Float* float_type = a->type()->AsFloat(); |
| 376 |
assert(float_type != nullptr); |
| 377 |
assert(integer_type != nullptr); |
| 378 |
if (integer_type->width() != 32) return nullptr; |
| 379 |
if (float_type->width() == 32) { |
| 380 |
float fa = a->GetFloat(); |
| 381 |
uint32_t result = integer_type->IsSigned() |
| 382 |
? static_cast<uint32_t>(static_cast<int32_t>(fa)) |
| 383 |
: static_cast<uint32_t>(fa); |
| 384 |
std::vector<uint32_t> words = {result}; |
| 385 |
return const_mgr->GetConstant(result_type, words); |
| 386 |
} else if (float_type->width() == 64) { |
| 387 |
double fa = a->GetDouble(); |
| 388 |
uint32_t result = integer_type->IsSigned() |
| 389 |
? static_cast<uint32_t>(static_cast<int32_t>(fa)) |
| 390 |
: static_cast<uint32_t>(fa); |
| 391 |
std::vector<uint32_t> words = {result}; |
| 392 |
return const_mgr->GetConstant(result_type, words); |
| 393 |
} |
| 394 |
return nullptr; |
| 395 |
}; |
| 396 |
} |
| 397 |
|
| 398 |
// This function defines a |UnaryScalarFoldingRule| that performs integer to |
| 399 |
// float conversion. |
| 400 |
// TODO(greg-lunarg): Support for 64-bit integer types. |
| 401 |
UnaryScalarFoldingRule FoldIToFOp() { |
| 402 |
return [](const analysis::Type* result_type, const analysis::Constant* a, |
| 403 |
analysis::ConstantManager* const_mgr) -> const analysis::Constant* { |
| 404 |
assert(result_type != nullptr && a != nullptr); |
| 405 |
const analysis::Integer* integer_type = a->type()->AsInteger(); |
| 406 |
const analysis::Float* float_type = result_type->AsFloat(); |
| 407 |
assert(float_type != nullptr); |
| 408 |
assert(integer_type != nullptr); |
| 409 |
if (integer_type->width() != 32) return nullptr; |
| 410 |
uint32_t ua = a->GetU32(); |
| 411 |
if (float_type->width() == 32) { |
| 412 |
float result_val = integer_type->IsSigned() |
| 413 |
? static_cast<float>(static_cast<int32_t>(ua)) |
| 414 |
: static_cast<float>(ua); |
| 415 |
utils::FloatProxy<float> result(result_val); |
| 416 |
std::vector<uint32_t> words = {result.data()}; |
| 417 |
return const_mgr->GetConstant(result_type, words); |
| 418 |
} else if (float_type->width() == 64) { |
| 419 |
double result_val = integer_type->IsSigned() |
| 420 |
? static_cast<double>(static_cast<int32_t>(ua)) |
| 421 |
: static_cast<double>(ua); |
| 422 |
utils::FloatProxy<double> result(result_val); |
| 423 |
std::vector<uint32_t> words = result.GetWords(); |
| 424 |
return const_mgr->GetConstant(result_type, words); |
| 425 |
} |
| 426 |
return nullptr; |
| 427 |
}; |
| 428 |
} |
| 429 |
|
| 430 |
// This defines a |UnaryScalarFoldingRule| that performs |OpQuantizeToF16|. |
| 431 |
UnaryScalarFoldingRule FoldQuantizeToF16Scalar() { |
| 432 |
return [](const analysis::Type* result_type, const analysis::Constant* a, |
| 433 |
analysis::ConstantManager* const_mgr) -> const analysis::Constant* { |
| 434 |
assert(result_type != nullptr && a != nullptr); |
| 435 |
const analysis::Float* float_type = a->type()->AsFloat(); |
| 436 |
assert(float_type != nullptr); |
| 437 |
if (float_type->width() != 32) { |
| 438 |
return nullptr; |
| 439 |
} |
| 440 |
|
| 441 |
float fa = a->GetFloat(); |
| 442 |
utils::HexFloat<utils::FloatProxy<float>> orignal(fa); |
| 443 |
utils::HexFloat<utils::FloatProxy<utils::Float16>> quantized(0); |
| 444 |
utils::HexFloat<utils::FloatProxy<float>> result(0.0f); |
| 445 |
orignal.castTo(quantized, utils::round_direction::kToZero); |
| 446 |
quantized.castTo(result, utils::round_direction::kToZero); |
| 447 |
std::vector<uint32_t> words = {result.getBits()}; |
| 448 |
return const_mgr->GetConstant(result_type, words); |
| 449 |
}; |
| 450 |
} |
| 451 |
|
| 452 |
// This macro defines a |BinaryScalarFoldingRule| that applies |op|. The |
| 453 |
// operator |op| must work for both float and double, and use syntax "f1 op f2". |
| 454 |
#define FOLD_FPARITH_OP(op) \ |
| 455 |
[](const analysis::Type* result_type_in_macro, const analysis::Constant* a, \ |
| 456 |
const analysis::Constant* b, \ |
| 457 |
analysis::ConstantManager* const_mgr_in_macro) \ |
| 458 |
-> const analysis::Constant* { \ |
| 459 |
assert(result_type_in_macro != nullptr && a != nullptr && b != nullptr); \ |
| 460 |
assert(result_type_in_macro == a->type() && \ |
| 461 |
result_type_in_macro == b->type()); \ |
| 462 |
const analysis::Float* float_type_in_macro = \ |
| 463 |
result_type_in_macro->AsFloat(); \ |
| 464 |
assert(float_type_in_macro != nullptr); \ |
| 465 |
if (float_type_in_macro->width() == 32) { \ |
| 466 |
float fa = a->GetFloat(); \ |
| 467 |
float fb = b->GetFloat(); \ |
| 468 |
utils::FloatProxy<float> result_in_macro(fa op fb); \ |
| 469 |
std::vector<uint32_t> words_in_macro = result_in_macro.GetWords(); \ |
| 470 |
return const_mgr_in_macro->GetConstant(result_type_in_macro, \ |
| 471 |
words_in_macro); \ |
| 472 |
} else if (float_type_in_macro->width() == 64) { \ |
| 473 |
double fa = a->GetDouble(); \ |
| 474 |
double fb = b->GetDouble(); \ |
| 475 |
utils::FloatProxy<double> result_in_macro(fa op fb); \ |
| 476 |
std::vector<uint32_t> words_in_macro = result_in_macro.GetWords(); \ |
| 477 |
return const_mgr_in_macro->GetConstant(result_type_in_macro, \ |
| 478 |
words_in_macro); \ |
| 479 |
} \ |
| 480 |
return nullptr; \ |
| 481 |
} |
| 482 |
|
| 483 |
// Define the folding rule for conversion between floating point and integer |
| 484 |
ConstantFoldingRule FoldFToI() { return FoldFPUnaryOp(FoldFToIOp()); } |
| 485 |
ConstantFoldingRule FoldIToF() { return FoldFPUnaryOp(FoldIToFOp()); } |
| 486 |
ConstantFoldingRule FoldQuantizeToF16() { |
| 487 |
return FoldFPUnaryOp(FoldQuantizeToF16Scalar()); |
| 488 |
} |
| 489 |
|
| 490 |
// Define the folding rules for subtraction, addition, multiplication, and |
| 491 |
// division for floating point values. |
| 492 |
ConstantFoldingRule FoldFSub() { return FoldFPBinaryOp(FOLD_FPARITH_OP(-)); } |
| 493 |
ConstantFoldingRule FoldFAdd() { return FoldFPBinaryOp(FOLD_FPARITH_OP(+)); } |
| 494 |
ConstantFoldingRule FoldFMul() { return FoldFPBinaryOp(FOLD_FPARITH_OP(*)); } |
| 495 |
ConstantFoldingRule FoldFDiv() { return FoldFPBinaryOp(FOLD_FPARITH_OP(/)); } |
| 496 |
|
| 497 |
bool CompareFloatingPoint(bool op_result, bool op_unordered, |
| 498 |
bool need_ordered) { |
| 499 |
if (need_ordered) { |
| 500 |
// operands are ordered and Operand 1 is |op| Operand 2 |
| 501 |
return !op_unordered && op_result; |
| 502 |
} else { |
| 503 |
// operands are unordered or Operand 1 is |op| Operand 2 |
| 504 |
return op_unordered || op_result; |
| 505 |
} |
| 506 |
} |
| 507 |
|
| 508 |
// This macro defines a |BinaryScalarFoldingRule| that applies |op|. The |
| 509 |
// operator |op| must work for both float and double, and use syntax "f1 op f2". |
| 510 |
#define FOLD_FPCMP_OP(op, ord) \ |
| 511 |
[](const analysis::Type* result_type, const analysis::Constant* a, \ |
| 512 |
const analysis::Constant* b, \ |
| 513 |
analysis::ConstantManager* const_mgr) -> const analysis::Constant* { \ |
| 514 |
assert(result_type != nullptr && a != nullptr && b != nullptr); \ |
| 515 |
assert(result_type->AsBool()); \ |
| 516 |
assert(a->type() == b->type()); \ |
| 517 |
const analysis::Float* float_type = a->type()->AsFloat(); \ |
| 518 |
assert(float_type != nullptr); \ |
| 519 |
if (float_type->width() == 32) { \ |
| 520 |
float fa = a->GetFloat(); \ |
| 521 |
float fb = b->GetFloat(); \ |
| 522 |
bool result = CompareFloatingPoint( \ |
| 523 |
fa op fb, std::isnan(fa) || std::isnan(fb), ord); \ |
| 524 |
std::vector<uint32_t> words = {uint32_t(result)}; \ |
| 525 |
return const_mgr->GetConstant(result_type, words); \ |
| 526 |
} else if (float_type->width() == 64) { \ |
| 527 |
double fa = a->GetDouble(); \ |
| 528 |
double fb = b->GetDouble(); \ |
| 529 |
bool result = CompareFloatingPoint( \ |
| 530 |
fa op fb, std::isnan(fa) || std::isnan(fb), ord); \ |
| 531 |
std::vector<uint32_t> words = {uint32_t(result)}; \ |
| 532 |
return const_mgr->GetConstant(result_type, words); \ |
| 533 |
} \ |
| 534 |
return nullptr; \ |
| 535 |
} |
| 536 |
|
| 537 |
// Define the folding rules for ordered and unordered comparison for floating |
| 538 |
// point values. |
| 539 |
ConstantFoldingRule FoldFOrdEqual() { |
| 540 |
return FoldFPBinaryOp(FOLD_FPCMP_OP(==, true)); |
| 541 |
} |
| 542 |
ConstantFoldingRule FoldFUnordEqual() { |
| 543 |
return FoldFPBinaryOp(FOLD_FPCMP_OP(==, false)); |
| 544 |
} |
| 545 |
ConstantFoldingRule FoldFOrdNotEqual() { |
| 546 |
return FoldFPBinaryOp(FOLD_FPCMP_OP(!=, true)); |
| 547 |
} |
| 548 |
ConstantFoldingRule FoldFUnordNotEqual() { |
| 549 |
return FoldFPBinaryOp(FOLD_FPCMP_OP(!=, false)); |
| 550 |
} |
| 551 |
ConstantFoldingRule FoldFOrdLessThan() { |
| 552 |
return FoldFPBinaryOp(FOLD_FPCMP_OP(<, true)); |
| 553 |
} |
| 554 |
ConstantFoldingRule FoldFUnordLessThan() { |
| 555 |
return FoldFPBinaryOp(FOLD_FPCMP_OP(<, false)); |
| 556 |
} |
| 557 |
ConstantFoldingRule FoldFOrdGreaterThan() { |
| 558 |
return FoldFPBinaryOp(FOLD_FPCMP_OP(>, true)); |
| 559 |
} |
| 560 |
ConstantFoldingRule FoldFUnordGreaterThan() { |
| 561 |
return FoldFPBinaryOp(FOLD_FPCMP_OP(>, false)); |
| 562 |
} |
| 563 |
ConstantFoldingRule FoldFOrdLessThanEqual() { |
| 564 |
return FoldFPBinaryOp(FOLD_FPCMP_OP(<=, true)); |
| 565 |
} |
| 566 |
ConstantFoldingRule FoldFUnordLessThanEqual() { |
| 567 |
return FoldFPBinaryOp(FOLD_FPCMP_OP(<=, false)); |
| 568 |
} |
| 569 |
ConstantFoldingRule FoldFOrdGreaterThanEqual() { |
| 570 |
return FoldFPBinaryOp(FOLD_FPCMP_OP(>=, true)); |
| 571 |
} |
| 572 |
ConstantFoldingRule FoldFUnordGreaterThanEqual() { |
| 573 |
return FoldFPBinaryOp(FOLD_FPCMP_OP(>=, false)); |
| 574 |
} |
| 575 |
|
| 576 |
// Folds an OpDot where all of the inputs are constants to a |
| 577 |
// constant. A new constant is created if necessary. |
| 578 |
ConstantFoldingRule FoldOpDotWithConstants() { |
| 579 |
return [](IRContext* context, Instruction* inst, |
| 580 |
const std::vector<const analysis::Constant*>& constants) |
| 581 |
-> const analysis::Constant* { |
| 582 |
analysis::ConstantManager* const_mgr = context->get_constant_mgr(); |
| 583 |
analysis::TypeManager* type_mgr = context->get_type_mgr(); |
| 584 |
const analysis::Type* new_type = type_mgr->GetType(inst->type_id()); |
| 585 |
assert(new_type->AsFloat() && "OpDot should have a float return type."); |
| 586 |
const analysis::Float* float_type = new_type->AsFloat(); |
| 587 |
|
| 588 |
if (!inst->IsFloatingPointFoldingAllowed()) { |
| 589 |
return nullptr; |
| 590 |
} |
| 591 |
|
| 592 |
// If one of the operands is 0, then the result is 0. |
| 593 |
bool has_zero_operand = false; |
| 594 |
|
| 595 |
for (int i = 0; i < 2; ++i) { |
| 596 |
if (constants[i]) { |
| 597 |
if (constants[i]->AsNullConstant() || |
| 598 |
constants[i]->AsVectorConstant()->IsZero()) { |
| 599 |
has_zero_operand = true; |
| 600 |
break; |
| 601 |
} |
| 602 |
} |
| 603 |
} |
| 604 |
|
| 605 |
if (has_zero_operand) { |
| 606 |
if (float_type->width() == 32) { |
| 607 |
utils::FloatProxy<float> result(0.0f); |
| 608 |
std::vector<uint32_t> words = result.GetWords(); |
| 609 |
return const_mgr->GetConstant(float_type, words); |
| 610 |
} |
| 611 |
if (float_type->width() == 64) { |
| 612 |
utils::FloatProxy<double> result(0.0); |
| 613 |
std::vector<uint32_t> words = result.GetWords(); |
| 614 |
return const_mgr->GetConstant(float_type, words); |
| 615 |
} |
| 616 |
return nullptr; |
| 617 |
} |
| 618 |
|
| 619 |
if (constants[0] == nullptr || constants[1] == nullptr) { |
| 620 |
return nullptr; |
| 621 |
} |
| 622 |
|
| 623 |
std::vector<const analysis::Constant*> a_components; |
| 624 |
std::vector<const analysis::Constant*> b_components; |
| 625 |
|
| 626 |
a_components = constants[0]->GetVectorComponents(const_mgr); |
| 627 |
b_components = constants[1]->GetVectorComponents(const_mgr); |
| 628 |
|
| 629 |
utils::FloatProxy<double> result(0.0); |
| 630 |
std::vector<uint32_t> words = result.GetWords(); |
| 631 |
const analysis::Constant* result_const = |
| 632 |
const_mgr->GetConstant(float_type, words); |
| 633 |
for (uint32_t i = 0; i < a_components.size() && result_const != nullptr; |
| 634 |
++i) { |
| 635 |
if (a_components[i] == nullptr || b_components[i] == nullptr) { |
| 636 |
return nullptr; |
| 637 |
} |
| 638 |
|
| 639 |
const analysis::Constant* component = FOLD_FPARITH_OP(*)( |
| 640 |
new_type, a_components[i], b_components[i], const_mgr); |
| 641 |
if (component == nullptr) { |
| 642 |
return nullptr; |
| 643 |
} |
| 644 |
result_const = |
| 645 |
FOLD_FPARITH_OP(+)(new_type, result_const, component, const_mgr); |
| 646 |
} |
| 647 |
return result_const; |
| 648 |
}; |
| 649 |
} |
| 650 |
|
| 651 |
// This function defines a |UnaryScalarFoldingRule| that subtracts the constant |
| 652 |
// from zero. |
| 653 |
UnaryScalarFoldingRule FoldFNegateOp() { |
| 654 |
return [](const analysis::Type* result_type, const analysis::Constant* a, |
| 655 |
analysis::ConstantManager* const_mgr) -> const analysis::Constant* { |
| 656 |
assert(result_type != nullptr && a != nullptr); |
| 657 |
assert(result_type == a->type()); |
| 658 |
const analysis::Float* float_type = result_type->AsFloat(); |
| 659 |
assert(float_type != nullptr); |
| 660 |
if (float_type->width() == 32) { |
| 661 |
float fa = a->GetFloat(); |
| 662 |
utils::FloatProxy<float> result(-fa); |
| 663 |
std::vector<uint32_t> words = result.GetWords(); |
| 664 |
return const_mgr->GetConstant(result_type, words); |
| 665 |
} else if (float_type->width() == 64) { |
| 666 |
double da = a->GetDouble(); |
| 667 |
utils::FloatProxy<double> result(-da); |
| 668 |
std::vector<uint32_t> words = result.GetWords(); |
| 669 |
return const_mgr->GetConstant(result_type, words); |
| 670 |
} |
| 671 |
return nullptr; |
| 672 |
}; |
| 673 |
} |
| 674 |
|
| 675 |
ConstantFoldingRule FoldFNegate() { return FoldFPUnaryOp(FoldFNegateOp()); } |
| 676 |
|
| 677 |
ConstantFoldingRule FoldFClampFeedingCompare(uint32_t cmp_opcode) { |
| 678 |
return [cmp_opcode](IRContext* context, Instruction* inst, |
| 679 |
const std::vector<const analysis::Constant*>& constants) |
| 680 |
-> const analysis::Constant* { |
| 681 |
analysis::ConstantManager* const_mgr = context->get_constant_mgr(); |
| 682 |
analysis::DefUseManager* def_use_mgr = context->get_def_use_mgr(); |
| 683 |
|
| 684 |
if (!inst->IsFloatingPointFoldingAllowed()) { |
| 685 |
return nullptr; |
| 686 |
} |
| 687 |
|
| 688 |
uint32_t non_const_idx = (constants[0] ? 1 : 0); |
| 689 |
uint32_t operand_id = inst->GetSingleWordInOperand(non_const_idx); |
| 690 |
Instruction* operand_inst = def_use_mgr->GetDef(operand_id); |
| 691 |
|
| 692 |
analysis::TypeManager* type_mgr = context->get_type_mgr(); |
| 693 |
const analysis::Type* operand_type = |
| 694 |
type_mgr->GetType(operand_inst->type_id()); |
| 695 |
|
| 696 |
if (!operand_type->AsFloat()) { |
| 697 |
return nullptr; |
| 698 |
} |
| 699 |
|
| 700 |
if (operand_type->AsFloat()->width() != 32 && |
| 701 |
operand_type->AsFloat()->width() != 64) { |
| 702 |
return nullptr; |
| 703 |
} |
| 704 |
|
| 705 |
if (operand_inst->opcode() != SpvOpExtInst) { |
| 706 |
return nullptr; |
| 707 |
} |
| 708 |
|
| 709 |
if (operand_inst->GetSingleWordInOperand(1) != GLSLstd450FClamp) { |
| 710 |
return nullptr; |
| 711 |
} |
| 712 |
|
| 713 |
if (constants[1] == nullptr && constants[0] == nullptr) { |
| 714 |
return nullptr; |
| 715 |
} |
| 716 |
|
| 717 |
uint32_t max_id = operand_inst->GetSingleWordInOperand(4); |
| 718 |
const analysis::Constant* max_const = |
| 719 |
const_mgr->FindDeclaredConstant(max_id); |
| 720 |
|
| 721 |
uint32_t min_id = operand_inst->GetSingleWordInOperand(3); |
| 722 |
const analysis::Constant* min_const = |
| 723 |
const_mgr->FindDeclaredConstant(min_id); |
| 724 |
|
| 725 |
bool found_result = false; |
| 726 |
bool result = false; |
| 727 |
|
| 728 |
switch (cmp_opcode) { |
| 729 |
case SpvOpFOrdLessThan: |
| 730 |
case SpvOpFUnordLessThan: |
| 731 |
case SpvOpFOrdGreaterThanEqual: |
| 732 |
case SpvOpFUnordGreaterThanEqual: |
| 733 |
if (constants[0]) { |
| 734 |
if (min_const) { |
| 735 |
if (constants[0]->GetValueAsDouble() < |
| 736 |
min_const->GetValueAsDouble()) { |
| 737 |
found_result = true; |
| 738 |
result = (cmp_opcode == SpvOpFOrdLessThan || |
| 739 |
cmp_opcode == SpvOpFUnordLessThan); |
| 740 |
} |
| 741 |
} |
| 742 |
if (max_const) { |
| 743 |
if (constants[0]->GetValueAsDouble() >= |
| 744 |
max_const->GetValueAsDouble()) { |
| 745 |
found_result = true; |
| 746 |
result = !(cmp_opcode == SpvOpFOrdLessThan || |
| 747 |
cmp_opcode == SpvOpFUnordLessThan); |
| 748 |
} |
| 749 |
} |
| 750 |
} |
| 751 |
|
| 752 |
if (constants[1]) { |
| 753 |
if (max_const) { |
| 754 |
if (max_const->GetValueAsDouble() < |
| 755 |
constants[1]->GetValueAsDouble()) { |
| 756 |
found_result = true; |
| 757 |
result = (cmp_opcode == SpvOpFOrdLessThan || |
| 758 |
cmp_opcode == SpvOpFUnordLessThan); |
| 759 |
} |
| 760 |
} |
| 761 |
|
| 762 |
if (min_const) { |
| 763 |
if (min_const->GetValueAsDouble() >= |
| 764 |
constants[1]->GetValueAsDouble()) { |
| 765 |
found_result = true; |
| 766 |
result = !(cmp_opcode == SpvOpFOrdLessThan || |
| 767 |
cmp_opcode == SpvOpFUnordLessThan); |
| 768 |
} |
| 769 |
} |
| 770 |
} |
| 771 |
break; |
| 772 |
case SpvOpFOrdGreaterThan: |
| 773 |
case SpvOpFUnordGreaterThan: |
| 774 |
case SpvOpFOrdLessThanEqual: |
| 775 |
case SpvOpFUnordLessThanEqual: |
| 776 |
if (constants[0]) { |
| 777 |
if (min_const) { |
| 778 |
if (constants[0]->GetValueAsDouble() <= |
| 779 |
min_const->GetValueAsDouble()) { |
| 780 |
found_result = true; |
| 781 |
result = (cmp_opcode == SpvOpFOrdLessThanEqual || |
| 782 |
cmp_opcode == SpvOpFUnordLessThanEqual); |
| 783 |
} |
| 784 |
} |
| 785 |
if (max_const) { |
| 786 |
if (constants[0]->GetValueAsDouble() > |
| 787 |
max_const->GetValueAsDouble()) { |
| 788 |
found_result = true; |
| 789 |
result = !(cmp_opcode == SpvOpFOrdLessThanEqual || |
| 790 |
cmp_opcode == SpvOpFUnordLessThanEqual); |
| 791 |
} |
| 792 |
} |
| 793 |
} |
| 794 |
|
| 795 |
if (constants[1]) { |
| 796 |
if (max_const) { |
| 797 |
if (max_const->GetValueAsDouble() <= |
| 798 |
constants[1]->GetValueAsDouble()) { |
| 799 |
found_result = true; |
| 800 |
result = (cmp_opcode == SpvOpFOrdLessThanEqual || |
| 801 |
cmp_opcode == SpvOpFUnordLessThanEqual); |
| 802 |
} |
| 803 |
} |
| 804 |
|
| 805 |
if (min_const) { |
| 806 |
if (min_const->GetValueAsDouble() > |
| 807 |
constants[1]->GetValueAsDouble()) { |
| 808 |
found_result = true; |
| 809 |
result = !(cmp_opcode == SpvOpFOrdLessThanEqual || |
| 810 |
cmp_opcode == SpvOpFUnordLessThanEqual); |
| 811 |
} |
| 812 |
} |
| 813 |
} |
| 814 |
break; |
| 815 |
default: |
| 816 |
return nullptr; |
| 817 |
} |
| 818 |
|
| 819 |
if (!found_result) { |
| 820 |
return nullptr; |
| 821 |
} |
| 822 |
|
| 823 |
const analysis::Type* bool_type = |
| 824 |
context->get_type_mgr()->GetType(inst->type_id()); |
| 825 |
const analysis::Constant* result_const = |
| 826 |
const_mgr->GetConstant(bool_type, {static_cast<uint32_t>(result)}); |
| 827 |
assert(result_const); |
| 828 |
return result_const; |
| 829 |
}; |
| 830 |
} |
| 831 |
|
| 832 |
ConstantFoldingRule FoldFMix() { |
| 833 |
return [](IRContext* context, Instruction* inst, |
| 834 |
const std::vector<const analysis::Constant*>& constants) |
| 835 |
-> const analysis::Constant* { |
| 836 |
analysis::ConstantManager* const_mgr = context->get_constant_mgr(); |
| 837 |
assert(inst->opcode() == SpvOpExtInst && |
| 838 |
"Expecting an extended instruction."); |
| 839 |
assert(inst->GetSingleWordInOperand(0) == |
| 840 |
context->get_feature_mgr()->GetExtInstImportId_GLSLstd450() && |
| 841 |
"Expecting a GLSLstd450 extended instruction."); |
| 842 |
assert(inst->GetSingleWordInOperand(1) == GLSLstd450FMix && |
| 843 |
"Expecting and FMix instruction."); |
| 844 |
|
| 845 |
if (!inst->IsFloatingPointFoldingAllowed()) { |
| 846 |
return nullptr; |
| 847 |
} |
| 848 |
|
| 849 |
// Make sure all FMix operands are constants. |
| 850 |
for (uint32_t i = 1; i < 4; i++) { |
| 851 |
if (constants[i] == nullptr) { |
| 852 |
return nullptr; |
| 853 |
} |
| 854 |
} |
| 855 |
|
| 856 |
const analysis::Constant* one; |
| 857 |
bool is_vector = false; |
| 858 |
const analysis::Type* result_type = constants[1]->type(); |
| 859 |
const analysis::Type* base_type = result_type; |
| 860 |
if (base_type->AsVector()) { |
| 861 |
is_vector = true; |
| 862 |
base_type = base_type->AsVector()->element_type(); |
| 863 |
} |
| 864 |
assert(base_type->AsFloat() != nullptr && |
| 865 |
"FMix is suppose to act on floats or vectors of floats."); |
| 866 |
|
| 867 |
if (base_type->AsFloat()->width() == 32) { |
| 868 |
one = const_mgr->GetConstant(base_type, |
| 869 |
utils::FloatProxy<float>(1.0f).GetWords()); |
| 870 |
} else { |
| 871 |
one = const_mgr->GetConstant(base_type, |
| 872 |
utils::FloatProxy<double>(1.0).GetWords()); |
| 873 |
} |
| 874 |
|
| 875 |
if (is_vector) { |
| 876 |
uint32_t one_id = const_mgr->GetDefiningInstruction(one)->result_id(); |
| 877 |
one = |
| 878 |
const_mgr->GetConstant(result_type, std::vector<uint32_t>(4, one_id)); |
| 879 |
} |
| 880 |
|
| 881 |
const analysis::Constant* temp1 = FoldFPBinaryOp( |
| 882 |
FOLD_FPARITH_OP(-), inst->type_id(), {one, constants[3]}, context); |
| 883 |
if (temp1 == nullptr) { |
| 884 |
return nullptr; |
| 885 |
} |
| 886 |
|
| 887 |
const analysis::Constant* temp2 = FoldFPBinaryOp( |
| 888 |
FOLD_FPARITH_OP(*), inst->type_id(), {constants[1], temp1}, context); |
| 889 |
if (temp2 == nullptr) { |
| 890 |
return nullptr; |
| 891 |
} |
| 892 |
const analysis::Constant* temp3 = |
| 893 |
FoldFPBinaryOp(FOLD_FPARITH_OP(*), inst->type_id(), |
| 894 |
{constants[2], constants[3]}, context); |
| 895 |
if (temp3 == nullptr) { |
| 896 |
return nullptr; |
| 897 |
} |
| 898 |
return FoldFPBinaryOp(FOLD_FPARITH_OP(+), inst->type_id(), {temp2, temp3}, |
| 899 |
context); |
| 900 |
}; |
| 901 |
} |
| 902 |
|
| 903 |
template <class IntType> |
| 904 |
IntType FoldIClamp(IntType x, IntType min_val, IntType max_val) { |
| 905 |
if (x < min_val) { |
| 906 |
x = min_val; |
| 907 |
} |
| 908 |
if (x > max_val) { |
| 909 |
x = max_val; |
| 910 |
} |
| 911 |
return x; |
| 912 |
} |
| 913 |
|
| 914 |
const analysis::Constant* FoldMin(const analysis::Type* result_type, |
| 915 |
const analysis::Constant* a, |
| 916 |
const analysis::Constant* b, |
| 917 |
analysis::ConstantManager*) { |
| 918 |
if (const analysis::Integer* int_type = result_type->AsInteger()) { |
| 919 |
if (int_type->width() == 32) { |
| 920 |
if (int_type->IsSigned()) { |
| 921 |
int32_t va = a->GetS32(); |
| 922 |
int32_t vb = b->GetS32(); |
| 923 |
return (va < vb ? a : b); |
| 924 |
} else { |
| 925 |
uint32_t va = a->GetU32(); |
| 926 |
uint32_t vb = b->GetU32(); |
| 927 |
return (va < vb ? a : b); |
| 928 |
} |
| 929 |
} else if (int_type->width() == 64) { |
| 930 |
if (int_type->IsSigned()) { |
| 931 |
int64_t va = a->GetS64(); |
| 932 |
int64_t vb = b->GetS64(); |
| 933 |
return (va < vb ? a : b); |
| 934 |
} else { |
| 935 |
uint64_t va = a->GetU64(); |
| 936 |
uint64_t vb = b->GetU64(); |
| 937 |
return (va < vb ? a : b); |
| 938 |
} |
| 939 |
} |
| 940 |
} else if (const analysis::Float* float_type = result_type->AsFloat()) { |
| 941 |
if (float_type->width() == 32) { |
| 942 |
float va = a->GetFloat(); |
| 943 |
float vb = b->GetFloat(); |
| 944 |
return (va < vb ? a : b); |
| 945 |
} else if (float_type->width() == 64) { |
| 946 |
double va = a->GetDouble(); |
| 947 |
double vb = b->GetDouble(); |
| 948 |
return (va < vb ? a : b); |
| 949 |
} |
| 950 |
} |
| 951 |
return nullptr; |
| 952 |
} |
| 953 |
|
| 954 |
const analysis::Constant* FoldMax(const analysis::Type* result_type, |
| 955 |
const analysis::Constant* a, |
| 956 |
const analysis::Constant* b, |
| 957 |
analysis::ConstantManager*) { |
| 958 |
if (const analysis::Integer* int_type = result_type->AsInteger()) { |
| 959 |
if (int_type->width() == 32) { |
| 960 |
if (int_type->IsSigned()) { |
| 961 |
int32_t va = a->GetS32(); |
| 962 |
int32_t vb = b->GetS32(); |
| 963 |
return (va > vb ? a : b); |
| 964 |
} else { |
| 965 |
uint32_t va = a->GetU32(); |
| 966 |
uint32_t vb = b->GetU32(); |
| 967 |
return (va > vb ? a : b); |
| 968 |
} |
| 969 |
} else if (int_type->width() == 64) { |
| 970 |
if (int_type->IsSigned()) { |
| 971 |
int64_t va = a->GetS64(); |
| 972 |
int64_t vb = b->GetS64(); |
| 973 |
return (va > vb ? a : b); |
| 974 |
} else { |
| 975 |
uint64_t va = a->GetU64(); |
| 976 |
uint64_t vb = b->GetU64(); |
| 977 |
return (va > vb ? a : b); |
| 978 |
} |
| 979 |
} |
| 980 |
} else if (const analysis::Float* float_type = result_type->AsFloat()) { |
| 981 |
if (float_type->width() == 32) { |
| 982 |
float va = a->GetFloat(); |
| 983 |
float vb = b->GetFloat(); |
| 984 |
return (va > vb ? a : b); |
| 985 |
} else if (float_type->width() == 64) { |
| 986 |
double va = a->GetDouble(); |
| 987 |
double vb = b->GetDouble(); |
| 988 |
return (va > vb ? a : b); |
| 989 |
} |
| 990 |
} |
| 991 |
return nullptr; |
| 992 |
} |
| 993 |
|
| 994 |
// Fold an clamp instruction when all three operands are constant. |
| 995 |
const analysis::Constant* FoldClamp1( |
| 996 |
IRContext* context, Instruction* inst, |
| 997 |
const std::vector<const analysis::Constant*>& constants) { |
| 998 |
assert(inst->opcode() == SpvOpExtInst && |
| 999 |
"Expecting an extended instruction."); |
| 1000 |
assert(inst->GetSingleWordInOperand(0) == |
| 1001 |
context->get_feature_mgr()->GetExtInstImportId_GLSLstd450() && |
| 1002 |
"Expecting a GLSLstd450 extended instruction."); |
| 1003 |
|
| 1004 |
// Make sure all Clamp operands are constants. |
| 1005 |
for (uint32_t i = 1; i < 3; i++) { |
| 1006 |
if (constants[i] == nullptr) { |
| 1007 |
return nullptr; |
| 1008 |
} |
| 1009 |
} |
| 1010 |
|
| 1011 |
const analysis::Constant* temp = FoldFPBinaryOp( |
| 1012 |
FoldMax, inst->type_id(), {constants[1], constants[2]}, context); |
| 1013 |
if (temp == nullptr) { |
| 1014 |
return nullptr; |
| 1015 |
} |
| 1016 |
return FoldFPBinaryOp(FoldMin, inst->type_id(), {temp, constants[3]}, |
| 1017 |
context); |
| 1018 |
} |
| 1019 |
|
| 1020 |
// Fold a clamp instruction when |x >= min_val|. |
| 1021 |
const analysis::Constant* FoldClamp2( |
| 1022 |
IRContext* context, Instruction* inst, |
| 1023 |
const std::vector<const analysis::Constant*>& constants) { |
| 1024 |
assert(inst->opcode() == SpvOpExtInst && |
| 1025 |
"Expecting an extended instruction."); |
| 1026 |
assert(inst->GetSingleWordInOperand(0) == |
| 1027 |
context->get_feature_mgr()->GetExtInstImportId_GLSLstd450() && |
| 1028 |
"Expecting a GLSLstd450 extended instruction."); |
| 1029 |
|
| 1030 |
const analysis::Constant* x = constants[1]; |
| 1031 |
const analysis::Constant* min_val = constants[2]; |
| 1032 |
|
| 1033 |
if (x == nullptr || min_val == nullptr) { |
| 1034 |
return nullptr; |
| 1035 |
} |
| 1036 |
|
| 1037 |
const analysis::Constant* temp = |
| 1038 |
FoldFPBinaryOp(FoldMax, inst->type_id(), {x, min_val}, context); |
| 1039 |
if (temp == min_val) { |
| 1040 |
// We can assume that |min_val| is less than |max_val|. Therefore, if the |
| 1041 |
// result of the max operation is |min_val|, we know the result of the min |
| 1042 |
// operation, even if |max_val| is not a constant. |
| 1043 |
return min_val; |
| 1044 |
} |
| 1045 |
return nullptr; |
| 1046 |
} |
| 1047 |
|
| 1048 |
// Fold a clamp instruction when |x >= max_val|. |
| 1049 |
const analysis::Constant* FoldClamp3( |
| 1050 |
IRContext* context, Instruction* inst, |
| 1051 |
const std::vector<const analysis::Constant*>& constants) { |
| 1052 |
assert(inst->opcode() == SpvOpExtInst && |
| 1053 |
"Expecting an extended instruction."); |
| 1054 |
assert(inst->GetSingleWordInOperand(0) == |
| 1055 |
context->get_feature_mgr()->GetExtInstImportId_GLSLstd450() && |
| 1056 |
"Expecting a GLSLstd450 extended instruction."); |
| 1057 |
|
| 1058 |
const analysis::Constant* x = constants[1]; |
| 1059 |
const analysis::Constant* max_val = constants[3]; |
| 1060 |
|
| 1061 |
if (x == nullptr || max_val == nullptr) { |
| 1062 |
return nullptr; |
| 1063 |
} |
| 1064 |
|
| 1065 |
const analysis::Constant* temp = |
| 1066 |
FoldFPBinaryOp(FoldMin, inst->type_id(), {x, max_val}, context); |
| 1067 |
if (temp == max_val) { |
| 1068 |
// We can assume that |min_val| is less than |max_val|. Therefore, if the |
| 1069 |
// result of the max operation is |min_val|, we know the result of the min |
| 1070 |
// operation, even if |max_val| is not a constant. |
| 1071 |
return max_val; |
| 1072 |
} |
| 1073 |
return nullptr; |
| 1074 |
} |
| 1075 |
|
| 1076 |
UnaryScalarFoldingRule FoldFTranscendentalUnary(double (*fp)(double)) { |
| 1077 |
return |
| 1078 |
[fp](const analysis::Type* result_type, const analysis::Constant* a, |
| 1079 |
analysis::ConstantManager* const_mgr) -> const analysis::Constant* { |
| 1080 |
assert(result_type != nullptr && a != nullptr); |
| 1081 |
const analysis::Float* float_type = a->type()->AsFloat(); |
| 1082 |
assert(float_type != nullptr); |
| 1083 |
assert(float_type == result_type->AsFloat()); |
| 1084 |
if (float_type->width() == 32) { |
| 1085 |
float fa = a->GetFloat(); |
| 1086 |
float res = static_cast<float>(fp(fa)); |
| 1087 |
utils::FloatProxy<float> result(res); |
| 1088 |
std::vector<uint32_t> words = result.GetWords(); |
| 1089 |
return const_mgr->GetConstant(result_type, words); |
| 1090 |
} else if (float_type->width() == 64) { |
| 1091 |
double fa = a->GetDouble(); |
| 1092 |
double res = fp(fa); |
| 1093 |
utils::FloatProxy<double> result(res); |
| 1094 |
std::vector<uint32_t> words = result.GetWords(); |
| 1095 |
return const_mgr->GetConstant(result_type, words); |
| 1096 |
} |
| 1097 |
return nullptr; |
| 1098 |
}; |
| 1099 |
} |
| 1100 |
|
| 1101 |
BinaryScalarFoldingRule FoldFTranscendentalBinary(double (*fp)(double, |
| 1102 |
double)) { |
| 1103 |
return |
| 1104 |
[fp](const analysis::Type* result_type, const analysis::Constant* a, |
| 1105 |
const analysis::Constant* b, |
| 1106 |
analysis::ConstantManager* const_mgr) -> const analysis::Constant* { |
| 1107 |
assert(result_type != nullptr && a != nullptr); |
| 1108 |
const analysis::Float* float_type = a->type()->AsFloat(); |
| 1109 |
assert(float_type != nullptr); |
| 1110 |
assert(float_type == result_type->AsFloat()); |
| 1111 |
assert(float_type == b->type()->AsFloat()); |
| 1112 |
if (float_type->width() == 32) { |
| 1113 |
float fa = a->GetFloat(); |
| 1114 |
float fb = b->GetFloat(); |
| 1115 |
float res = static_cast<float>(fp(fa, fb)); |
| 1116 |
utils::FloatProxy<float> result(res); |
| 1117 |
std::vector<uint32_t> words = result.GetWords(); |
| 1118 |
return const_mgr->GetConstant(result_type, words); |
| 1119 |
} else if (float_type->width() == 64) { |
| 1120 |
double fa = a->GetDouble(); |
| 1121 |
double fb = b->GetDouble(); |
| 1122 |
double res = fp(fa, fb); |
| 1123 |
utils::FloatProxy<double> result(res); |
| 1124 |
std::vector<uint32_t> words = result.GetWords(); |
| 1125 |
return const_mgr->GetConstant(result_type, words); |
| 1126 |
} |
| 1127 |
return nullptr; |
| 1128 |
}; |
| 1129 |
} |
| 1130 |
} // namespace |
| 1131 |
|
| 1132 |
void ConstantFoldingRules::AddFoldingRules() { |
| 1133 |
// Add all folding rules to the list for the opcodes to which they apply. |
| 1134 |
// Note that the order in which rules are added to the list matters. If a rule |
| 1135 |
// applies to the instruction, the rest of the rules will not be attempted. |
| 1136 |
// Take that into consideration. |
| 1137 |
|
| 1138 |
rules_[SpvOpCompositeConstruct].push_back(FoldCompositeWithConstants()); |
| 1139 |
|
| 1140 |
rules_[SpvOpCompositeExtract].push_back(FoldExtractWithConstants()); |
| 1141 |
|
| 1142 |
rules_[SpvOpConvertFToS].push_back(FoldFToI()); |
| 1143 |
rules_[SpvOpConvertFToU].push_back(FoldFToI()); |
| 1144 |
rules_[SpvOpConvertSToF].push_back(FoldIToF()); |
| 1145 |
rules_[SpvOpConvertUToF].push_back(FoldIToF()); |
| 1146 |
|
| 1147 |
rules_[SpvOpDot].push_back(FoldOpDotWithConstants()); |
| 1148 |
rules_[SpvOpFAdd].push_back(FoldFAdd()); |
| 1149 |
rules_[SpvOpFDiv].push_back(FoldFDiv()); |
| 1150 |
rules_[SpvOpFMul].push_back(FoldFMul()); |
| 1151 |
rules_[SpvOpFSub].push_back(FoldFSub()); |
| 1152 |
|
| 1153 |
rules_[SpvOpFOrdEqual].push_back(FoldFOrdEqual()); |
| 1154 |
|
| 1155 |
rules_[SpvOpFUnordEqual].push_back(FoldFUnordEqual()); |
| 1156 |
|
| 1157 |
rules_[SpvOpFOrdNotEqual].push_back(FoldFOrdNotEqual()); |
| 1158 |
|
| 1159 |
rules_[SpvOpFUnordNotEqual].push_back(FoldFUnordNotEqual()); |
| 1160 |
|
| 1161 |
rules_[SpvOpFOrdLessThan].push_back(FoldFOrdLessThan()); |
| 1162 |
rules_[SpvOpFOrdLessThan].push_back( |
| 1163 |
FoldFClampFeedingCompare(SpvOpFOrdLessThan)); |
| 1164 |
|
| 1165 |
rules_[SpvOpFUnordLessThan].push_back(FoldFUnordLessThan()); |
| 1166 |
rules_[SpvOpFUnordLessThan].push_back( |
| 1167 |
FoldFClampFeedingCompare(SpvOpFUnordLessThan)); |
| 1168 |
|
| 1169 |
rules_[SpvOpFOrdGreaterThan].push_back(FoldFOrdGreaterThan()); |
| 1170 |
rules_[SpvOpFOrdGreaterThan].push_back( |
| 1171 |
FoldFClampFeedingCompare(SpvOpFOrdGreaterThan)); |
| 1172 |
|
| 1173 |
rules_[SpvOpFUnordGreaterThan].push_back(FoldFUnordGreaterThan()); |
| 1174 |
rules_[SpvOpFUnordGreaterThan].push_back( |
| 1175 |
FoldFClampFeedingCompare(SpvOpFUnordGreaterThan)); |
| 1176 |
|
| 1177 |
rules_[SpvOpFOrdLessThanEqual].push_back(FoldFOrdLessThanEqual()); |
| 1178 |
rules_[SpvOpFOrdLessThanEqual].push_back( |
| 1179 |
FoldFClampFeedingCompare(SpvOpFOrdLessThanEqual)); |
| 1180 |
|
| 1181 |
rules_[SpvOpFUnordLessThanEqual].push_back(FoldFUnordLessThanEqual()); |
| 1182 |
rules_[SpvOpFUnordLessThanEqual].push_back( |
| 1183 |
FoldFClampFeedingCompare(SpvOpFUnordLessThanEqual)); |
| 1184 |
|
| 1185 |
rules_[SpvOpFOrdGreaterThanEqual].push_back(FoldFOrdGreaterThanEqual()); |
| 1186 |
rules_[SpvOpFOrdGreaterThanEqual].push_back( |
| 1187 |
FoldFClampFeedingCompare(SpvOpFOrdGreaterThanEqual)); |
| 1188 |
|
| 1189 |
rules_[SpvOpFUnordGreaterThanEqual].push_back(FoldFUnordGreaterThanEqual()); |
| 1190 |
rules_[SpvOpFUnordGreaterThanEqual].push_back( |
| 1191 |
FoldFClampFeedingCompare(SpvOpFUnordGreaterThanEqual)); |
| 1192 |
|
| 1193 |
rules_[SpvOpVectorShuffle].push_back(FoldVectorShuffleWithConstants()); |
| 1194 |
rules_[SpvOpVectorTimesScalar].push_back(FoldVectorTimesScalar()); |
| 1195 |
|
| 1196 |
rules_[SpvOpFNegate].push_back(FoldFNegate()); |
| 1197 |
rules_[SpvOpQuantizeToF16].push_back(FoldQuantizeToF16()); |
| 1198 |
|
| 1199 |
// Add rules for GLSLstd450 |
| 1200 |
FeatureManager* feature_manager = context_->get_feature_mgr(); |
| 1201 |
uint32_t ext_inst_glslstd450_id = |
| 1202 |
feature_manager->GetExtInstImportId_GLSLstd450(); |
| 1203 |
if (ext_inst_glslstd450_id != 0) { |
| 1204 |
ext_rules_[{ext_inst_glslstd450_id, GLSLstd450FMix}].push_back(FoldFMix()); |
| 1205 |
ext_rules_[{ext_inst_glslstd450_id, GLSLstd450SMin}].push_back( |
| 1206 |
FoldFPBinaryOp(FoldMin)); |
| 1207 |
ext_rules_[{ext_inst_glslstd450_id, GLSLstd450UMin}].push_back( |
| 1208 |
FoldFPBinaryOp(FoldMin)); |
| 1209 |
ext_rules_[{ext_inst_glslstd450_id, GLSLstd450FMin}].push_back( |
| 1210 |
FoldFPBinaryOp(FoldMin)); |
| 1211 |
ext_rules_[{ext_inst_glslstd450_id, GLSLstd450SMax}].push_back( |
| 1212 |
FoldFPBinaryOp(FoldMax)); |
| 1213 |
ext_rules_[{ext_inst_glslstd450_id, GLSLstd450UMax}].push_back( |
| 1214 |
FoldFPBinaryOp(FoldMax)); |
| 1215 |
ext_rules_[{ext_inst_glslstd450_id, GLSLstd450FMax}].push_back( |
| 1216 |
FoldFPBinaryOp(FoldMax)); |
| 1217 |
ext_rules_[{ext_inst_glslstd450_id, GLSLstd450UClamp}].push_back( |
| 1218 |
FoldClamp1); |
| 1219 |
ext_rules_[{ext_inst_glslstd450_id, GLSLstd450UClamp}].push_back( |
| 1220 |
FoldClamp2); |
| 1221 |
ext_rules_[{ext_inst_glslstd450_id, GLSLstd450UClamp}].push_back( |
| 1222 |
FoldClamp3); |
| 1223 |
ext_rules_[{ext_inst_glslstd450_id, GLSLstd450SClamp}].push_back( |
| 1224 |
FoldClamp1); |
| 1225 |
ext_rules_[{ext_inst_glslstd450_id, GLSLstd450SClamp}].push_back( |
| 1226 |
FoldClamp2); |
| 1227 |
ext_rules_[{ext_inst_glslstd450_id, GLSLstd450SClamp}].push_back( |
| 1228 |
FoldClamp3); |
| 1229 |
ext_rules_[{ext_inst_glslstd450_id, GLSLstd450FClamp}].push_back( |
| 1230 |
FoldClamp1); |
| 1231 |
ext_rules_[{ext_inst_glslstd450_id, GLSLstd450FClamp}].push_back( |
| 1232 |
FoldClamp2); |
| 1233 |
ext_rules_[{ext_inst_glslstd450_id, GLSLstd450FClamp}].push_back( |
| 1234 |
FoldClamp3); |
| 1235 |
ext_rules_[{ext_inst_glslstd450_id, GLSLstd450Sin}].push_back( |
| 1236 |
FoldFPUnaryOp(FoldFTranscendentalUnary(std::sin))); |
| 1237 |
ext_rules_[{ext_inst_glslstd450_id, GLSLstd450Cos}].push_back( |
| 1238 |
FoldFPUnaryOp(FoldFTranscendentalUnary(std::cos))); |
| 1239 |
ext_rules_[{ext_inst_glslstd450_id, GLSLstd450Tan}].push_back( |
| 1240 |
FoldFPUnaryOp(FoldFTranscendentalUnary(std::tan))); |
| 1241 |
ext_rules_[{ext_inst_glslstd450_id, GLSLstd450Asin}].push_back( |
| 1242 |
FoldFPUnaryOp(FoldFTranscendentalUnary(std::asin))); |
| 1243 |
ext_rules_[{ext_inst_glslstd450_id, GLSLstd450Acos}].push_back( |
| 1244 |
FoldFPUnaryOp(FoldFTranscendentalUnary(std::acos))); |
| 1245 |
ext_rules_[{ext_inst_glslstd450_id, GLSLstd450Atan}].push_back( |
| 1246 |
FoldFPUnaryOp(FoldFTranscendentalUnary(std::atan))); |
| 1247 |
ext_rules_[{ext_inst_glslstd450_id, GLSLstd450Exp}].push_back( |
| 1248 |
FoldFPUnaryOp(FoldFTranscendentalUnary(std::exp))); |
| 1249 |
ext_rules_[{ext_inst_glslstd450_id, GLSLstd450Log}].push_back( |
| 1250 |
FoldFPUnaryOp(FoldFTranscendentalUnary(std::log))); |
| 1251 |
|
| 1252 |
#ifdef __ANDROID__ |
| 1253 |
// Android NDK r15c tageting ABI 15 doesn't have full support for C++11 |
| 1254 |
// (no std::exp2/log2). ::exp2 is available from C99 but ::log2 isn't |
| 1255 |
// available up until ABI 18 so we use a shim |
| 1256 |
auto log2_shim = [](double v) -> double { return log(v) / log(2.0); }; |
| 1257 |
ext_rules_[{ext_inst_glslstd450_id, GLSLstd450Exp2}].push_back( |
| 1258 |
FoldFPUnaryOp(FoldFTranscendentalUnary(::exp2))); |
| 1259 |
ext_rules_[{ext_inst_glslstd450_id, GLSLstd450Log2}].push_back( |
| 1260 |
FoldFPUnaryOp(FoldFTranscendentalUnary(log2_shim))); |
| 1261 |
#else |
| 1262 |
ext_rules_[{ext_inst_glslstd450_id, GLSLstd450Exp2}].push_back( |
| 1263 |
FoldFPUnaryOp(FoldFTranscendentalUnary(std::exp2))); |
| 1264 |
ext_rules_[{ext_inst_glslstd450_id, GLSLstd450Log2}].push_back( |
| 1265 |
FoldFPUnaryOp(FoldFTranscendentalUnary(std::log2))); |
| 1266 |
#endif |
| 1267 |
|
| 1268 |
ext_rules_[{ext_inst_glslstd450_id, GLSLstd450Sqrt}].push_back( |
| 1269 |
FoldFPUnaryOp(FoldFTranscendentalUnary(std::sqrt))); |
| 1270 |
ext_rules_[{ext_inst_glslstd450_id, GLSLstd450Atan2}].push_back( |
| 1271 |
FoldFPBinaryOp(FoldFTranscendentalBinary(std::atan2))); |
| 1272 |
ext_rules_[{ext_inst_glslstd450_id, GLSLstd450Pow}].push_back( |
| 1273 |
FoldFPBinaryOp(FoldFTranscendentalBinary(std::pow))); |
| 1274 |
} |
| 1275 |
} |
| 1276 |
} // namespace opt |
| 1277 |
} // namespace spvtools |