Chris Forbes | cc5697f | 2019-01-30 11:54:08 -0800 | [diff] [blame] | 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 { |
Nicolas Capens | 84c9c45 | 2022-11-18 14:11:05 +0000 | [diff] [blame] | 22 | constexpr uint32_t kExtractCompositeIdInIdx = 0; |
Chris Forbes | cc5697f | 2019-01-30 11:54:08 -0800 | [diff] [blame] | 23 | |
Nicolas Capens | 6cacf18 | 2021-11-30 11:15:46 -0500 | [diff] [blame] | 24 | // Returns a constants with the value NaN of the given type. Only works for |
| 25 | // 32-bit and 64-bit float point types. Returns |nullptr| if an error occurs. |
| 26 | const analysis::Constant* GetNan(const analysis::Type* type, |
| 27 | analysis::ConstantManager* const_mgr) { |
| 28 | const analysis::Float* float_type = type->AsFloat(); |
| 29 | if (float_type == nullptr) { |
| 30 | return nullptr; |
| 31 | } |
| 32 | |
| 33 | switch (float_type->width()) { |
| 34 | case 32: |
| 35 | return const_mgr->GetFloatConst(std::numeric_limits<float>::quiet_NaN()); |
| 36 | case 64: |
| 37 | return const_mgr->GetDoubleConst( |
| 38 | std::numeric_limits<double>::quiet_NaN()); |
| 39 | default: |
| 40 | return nullptr; |
| 41 | } |
| 42 | } |
| 43 | |
| 44 | // Returns a constants with the value INF of the given type. Only works for |
| 45 | // 32-bit and 64-bit float point types. Returns |nullptr| if an error occurs. |
| 46 | const analysis::Constant* GetInf(const analysis::Type* type, |
| 47 | analysis::ConstantManager* const_mgr) { |
| 48 | const analysis::Float* float_type = type->AsFloat(); |
| 49 | if (float_type == nullptr) { |
| 50 | return nullptr; |
| 51 | } |
| 52 | |
| 53 | switch (float_type->width()) { |
| 54 | case 32: |
| 55 | return const_mgr->GetFloatConst(std::numeric_limits<float>::infinity()); |
| 56 | case 64: |
| 57 | return const_mgr->GetDoubleConst(std::numeric_limits<double>::infinity()); |
| 58 | default: |
| 59 | return nullptr; |
| 60 | } |
| 61 | } |
| 62 | |
Chris Forbes | cc5697f | 2019-01-30 11:54:08 -0800 | [diff] [blame] | 63 | // Returns true if |type| is Float or a vector of Float. |
| 64 | bool HasFloatingPoint(const analysis::Type* type) { |
| 65 | if (type->AsFloat()) { |
| 66 | return true; |
| 67 | } else if (const analysis::Vector* vec_type = type->AsVector()) { |
| 68 | return vec_type->element_type()->AsFloat() != nullptr; |
| 69 | } |
| 70 | |
| 71 | return false; |
| 72 | } |
| 73 | |
Nicolas Capens | 6cacf18 | 2021-11-30 11:15:46 -0500 | [diff] [blame] | 74 | // Returns a constants with the value |-val| of the given type. Only works for |
| 75 | // 32-bit and 64-bit float point types. Returns |nullptr| if an error occurs. |
Nicolas Capens | 84c9c45 | 2022-11-18 14:11:05 +0000 | [diff] [blame] | 76 | const analysis::Constant* NegateFPConst(const analysis::Type* result_type, |
Nicolas Capens | 6cacf18 | 2021-11-30 11:15:46 -0500 | [diff] [blame] | 77 | const analysis::Constant* val, |
| 78 | analysis::ConstantManager* const_mgr) { |
| 79 | const analysis::Float* float_type = result_type->AsFloat(); |
| 80 | assert(float_type != nullptr); |
| 81 | if (float_type->width() == 32) { |
| 82 | float fa = val->GetFloat(); |
| 83 | return const_mgr->GetFloatConst(-fa); |
| 84 | } else if (float_type->width() == 64) { |
| 85 | double da = val->GetDouble(); |
| 86 | return const_mgr->GetDoubleConst(-da); |
| 87 | } |
| 88 | return nullptr; |
| 89 | } |
| 90 | |
Chris Forbes | cc5697f | 2019-01-30 11:54:08 -0800 | [diff] [blame] | 91 | // Folds an OpcompositeExtract where input is a composite constant. |
| 92 | ConstantFoldingRule FoldExtractWithConstants() { |
| 93 | return [](IRContext* context, Instruction* inst, |
| 94 | const std::vector<const analysis::Constant*>& constants) |
| 95 | -> const analysis::Constant* { |
| 96 | const analysis::Constant* c = constants[kExtractCompositeIdInIdx]; |
| 97 | if (c == nullptr) { |
| 98 | return nullptr; |
| 99 | } |
| 100 | |
| 101 | for (uint32_t i = 1; i < inst->NumInOperands(); ++i) { |
| 102 | uint32_t element_index = inst->GetSingleWordInOperand(i); |
| 103 | if (c->AsNullConstant()) { |
| 104 | // Return Null for the return type. |
| 105 | analysis::ConstantManager* const_mgr = context->get_constant_mgr(); |
| 106 | analysis::TypeManager* type_mgr = context->get_type_mgr(); |
| 107 | return const_mgr->GetConstant(type_mgr->GetType(inst->type_id()), {}); |
| 108 | } |
| 109 | |
| 110 | auto cc = c->AsCompositeConstant(); |
| 111 | assert(cc != nullptr); |
| 112 | auto components = cc->GetComponents(); |
Ben Clayton | d0f684e | 2019-08-30 22:36:08 +0100 | [diff] [blame] | 113 | // Protect against invalid IR. Refuse to fold if the index is out |
| 114 | // of bounds. |
| 115 | if (element_index >= components.size()) return nullptr; |
Chris Forbes | cc5697f | 2019-01-30 11:54:08 -0800 | [diff] [blame] | 116 | c = components[element_index]; |
| 117 | } |
| 118 | return c; |
| 119 | }; |
| 120 | } |
| 121 | |
Nicolas Capens | 84c9c45 | 2022-11-18 14:11:05 +0000 | [diff] [blame] | 122 | // Folds an OpcompositeInsert where input is a composite constant. |
| 123 | ConstantFoldingRule FoldInsertWithConstants() { |
| 124 | return [](IRContext* context, Instruction* inst, |
| 125 | const std::vector<const analysis::Constant*>& constants) |
| 126 | -> const analysis::Constant* { |
| 127 | analysis::ConstantManager* const_mgr = context->get_constant_mgr(); |
| 128 | const analysis::Constant* object = constants[0]; |
| 129 | const analysis::Constant* composite = constants[1]; |
| 130 | if (object == nullptr || composite == nullptr) { |
| 131 | return nullptr; |
| 132 | } |
| 133 | |
| 134 | // If there is more than 1 index, then each additional constant used by the |
| 135 | // index will need to be recreated to use the inserted object. |
| 136 | std::vector<const analysis::Constant*> chain; |
| 137 | std::vector<const analysis::Constant*> components; |
| 138 | const analysis::Type* type = nullptr; |
Alexis Hetu | 34bbeae | 2022-12-06 09:37:55 -0500 | [diff] [blame] | 139 | const uint32_t final_index = (inst->NumInOperands() - 1); |
Nicolas Capens | 84c9c45 | 2022-11-18 14:11:05 +0000 | [diff] [blame] | 140 | |
Alexis Hetu | 34bbeae | 2022-12-06 09:37:55 -0500 | [diff] [blame] | 141 | // Work down hierarchy of all indexes |
Nicolas Capens | 84c9c45 | 2022-11-18 14:11:05 +0000 | [diff] [blame] | 142 | for (uint32_t i = 2; i < inst->NumInOperands(); ++i) { |
Alexis Hetu | 34bbeae | 2022-12-06 09:37:55 -0500 | [diff] [blame] | 143 | type = composite->type(); |
Alexis Hetu | 1ef51fa | 2022-11-24 09:03:10 -0500 | [diff] [blame] | 144 | |
Alexis Hetu | 34bbeae | 2022-12-06 09:37:55 -0500 | [diff] [blame] | 145 | if (composite->AsNullConstant()) { |
| 146 | // Make new composite so it can be inserted in the index with the |
| 147 | // non-null value |
| 148 | const auto new_composite = const_mgr->GetNullCompositeConstant(type); |
| 149 | // Keep track of any indexes along the way to last index |
| 150 | if (i != final_index) { |
| 151 | chain.push_back(new_composite); |
| 152 | } |
| 153 | components = new_composite->AsCompositeConstant()->GetComponents(); |
| 154 | } else { |
| 155 | // Keep track of any indexes along the way to last index |
| 156 | if (i != final_index) { |
| 157 | chain.push_back(composite); |
| 158 | } |
| 159 | components = composite->AsCompositeConstant()->GetComponents(); |
Nicolas Capens | 84c9c45 | 2022-11-18 14:11:05 +0000 | [diff] [blame] | 160 | } |
| 161 | const uint32_t index = inst->GetSingleWordInOperand(i); |
Nicolas Capens | 84c9c45 | 2022-11-18 14:11:05 +0000 | [diff] [blame] | 162 | composite = components[index]; |
| 163 | } |
| 164 | |
| 165 | // Final index in hierarchy is inserted with new object. |
Alexis Hetu | 34bbeae | 2022-12-06 09:37:55 -0500 | [diff] [blame] | 166 | const uint32_t final_operand = inst->GetSingleWordInOperand(final_index); |
Nicolas Capens | 84c9c45 | 2022-11-18 14:11:05 +0000 | [diff] [blame] | 167 | std::vector<uint32_t> ids; |
| 168 | for (size_t i = 0; i < components.size(); i++) { |
| 169 | const analysis::Constant* constant = |
Alexis Hetu | 34bbeae | 2022-12-06 09:37:55 -0500 | [diff] [blame] | 170 | (i == final_operand) ? object : components[i]; |
Nicolas Capens | 84c9c45 | 2022-11-18 14:11:05 +0000 | [diff] [blame] | 171 | Instruction* member_inst = const_mgr->GetDefiningInstruction(constant); |
| 172 | ids.push_back(member_inst->result_id()); |
| 173 | } |
| 174 | const analysis::Constant* new_constant = const_mgr->GetConstant(type, ids); |
| 175 | |
| 176 | // Work backwards up the chain and replace each index with new constant. |
| 177 | for (size_t i = chain.size(); i > 0; i--) { |
| 178 | // Need to insert any previous instruction into the module first. |
| 179 | // Can't just insert in types_values_begin() because it will move above |
Alexis Hetu | 34bbeae | 2022-12-06 09:37:55 -0500 | [diff] [blame] | 180 | // where the types are declared. |
| 181 | // Can't compare with location of inst because not all new added |
| 182 | // instructions are added to types_values_ |
| 183 | auto iter = context->types_values_end(); |
| 184 | Module::inst_iterator* pos = &iter; |
| 185 | const_mgr->BuildInstructionAndAddToModule(new_constant, pos); |
Nicolas Capens | 84c9c45 | 2022-11-18 14:11:05 +0000 | [diff] [blame] | 186 | |
| 187 | composite = chain[i - 1]; |
| 188 | components = composite->AsCompositeConstant()->GetComponents(); |
Alexis Hetu | 34bbeae | 2022-12-06 09:37:55 -0500 | [diff] [blame] | 189 | type = composite->type(); |
Nicolas Capens | 84c9c45 | 2022-11-18 14:11:05 +0000 | [diff] [blame] | 190 | ids.clear(); |
| 191 | for (size_t k = 0; k < components.size(); k++) { |
| 192 | const uint32_t index = |
| 193 | inst->GetSingleWordInOperand(1 + static_cast<uint32_t>(i)); |
| 194 | const analysis::Constant* constant = |
| 195 | (k == index) ? new_constant : components[k]; |
| 196 | const uint32_t constant_id = |
| 197 | const_mgr->FindDeclaredConstant(constant, 0); |
| 198 | ids.push_back(constant_id); |
| 199 | } |
| 200 | new_constant = const_mgr->GetConstant(type, ids); |
| 201 | } |
| 202 | |
| 203 | // If multiple constants were created, only need to return the top index. |
| 204 | return new_constant; |
| 205 | }; |
| 206 | } |
| 207 | |
Chris Forbes | cc5697f | 2019-01-30 11:54:08 -0800 | [diff] [blame] | 208 | ConstantFoldingRule FoldVectorShuffleWithConstants() { |
| 209 | return [](IRContext* context, Instruction* inst, |
| 210 | const std::vector<const analysis::Constant*>& constants) |
| 211 | -> const analysis::Constant* { |
Nicolas Capens | 84c9c45 | 2022-11-18 14:11:05 +0000 | [diff] [blame] | 212 | assert(inst->opcode() == spv::Op::OpVectorShuffle); |
Chris Forbes | cc5697f | 2019-01-30 11:54:08 -0800 | [diff] [blame] | 213 | const analysis::Constant* c1 = constants[0]; |
| 214 | const analysis::Constant* c2 = constants[1]; |
| 215 | if (c1 == nullptr || c2 == nullptr) { |
| 216 | return nullptr; |
| 217 | } |
| 218 | |
| 219 | analysis::ConstantManager* const_mgr = context->get_constant_mgr(); |
| 220 | const analysis::Type* element_type = c1->type()->AsVector()->element_type(); |
| 221 | |
| 222 | std::vector<const analysis::Constant*> c1_components; |
| 223 | if (const analysis::VectorConstant* vec_const = c1->AsVectorConstant()) { |
| 224 | c1_components = vec_const->GetComponents(); |
| 225 | } else { |
| 226 | assert(c1->AsNullConstant()); |
| 227 | const analysis::Constant* element = |
| 228 | const_mgr->GetConstant(element_type, {}); |
| 229 | c1_components.resize(c1->type()->AsVector()->element_count(), element); |
| 230 | } |
| 231 | std::vector<const analysis::Constant*> c2_components; |
| 232 | if (const analysis::VectorConstant* vec_const = c2->AsVectorConstant()) { |
| 233 | c2_components = vec_const->GetComponents(); |
| 234 | } else { |
| 235 | assert(c2->AsNullConstant()); |
| 236 | const analysis::Constant* element = |
| 237 | const_mgr->GetConstant(element_type, {}); |
| 238 | c2_components.resize(c2->type()->AsVector()->element_count(), element); |
| 239 | } |
| 240 | |
| 241 | std::vector<uint32_t> ids; |
| 242 | const uint32_t undef_literal_value = 0xffffffff; |
| 243 | for (uint32_t i = 2; i < inst->NumInOperands(); ++i) { |
| 244 | uint32_t index = inst->GetSingleWordInOperand(i); |
| 245 | if (index == undef_literal_value) { |
| 246 | // Don't fold shuffle with undef literal value. |
| 247 | return nullptr; |
| 248 | } else if (index < c1_components.size()) { |
| 249 | Instruction* member_inst = |
| 250 | const_mgr->GetDefiningInstruction(c1_components[index]); |
| 251 | ids.push_back(member_inst->result_id()); |
| 252 | } else { |
| 253 | Instruction* member_inst = const_mgr->GetDefiningInstruction( |
| 254 | c2_components[index - c1_components.size()]); |
| 255 | ids.push_back(member_inst->result_id()); |
| 256 | } |
| 257 | } |
| 258 | |
| 259 | analysis::TypeManager* type_mgr = context->get_type_mgr(); |
| 260 | return const_mgr->GetConstant(type_mgr->GetType(inst->type_id()), ids); |
| 261 | }; |
| 262 | } |
| 263 | |
| 264 | ConstantFoldingRule FoldVectorTimesScalar() { |
| 265 | return [](IRContext* context, Instruction* inst, |
| 266 | const std::vector<const analysis::Constant*>& constants) |
| 267 | -> const analysis::Constant* { |
Nicolas Capens | 84c9c45 | 2022-11-18 14:11:05 +0000 | [diff] [blame] | 268 | assert(inst->opcode() == spv::Op::OpVectorTimesScalar); |
Chris Forbes | cc5697f | 2019-01-30 11:54:08 -0800 | [diff] [blame] | 269 | analysis::ConstantManager* const_mgr = context->get_constant_mgr(); |
| 270 | analysis::TypeManager* type_mgr = context->get_type_mgr(); |
| 271 | |
| 272 | if (!inst->IsFloatingPointFoldingAllowed()) { |
| 273 | if (HasFloatingPoint(type_mgr->GetType(inst->type_id()))) { |
| 274 | return nullptr; |
| 275 | } |
| 276 | } |
| 277 | |
| 278 | const analysis::Constant* c1 = constants[0]; |
| 279 | const analysis::Constant* c2 = constants[1]; |
| 280 | |
| 281 | if (c1 && c1->IsZero()) { |
| 282 | return c1; |
| 283 | } |
| 284 | |
| 285 | if (c2 && c2->IsZero()) { |
| 286 | // Get or create the NullConstant for this type. |
| 287 | std::vector<uint32_t> ids; |
| 288 | return const_mgr->GetConstant(type_mgr->GetType(inst->type_id()), ids); |
| 289 | } |
| 290 | |
| 291 | if (c1 == nullptr || c2 == nullptr) { |
| 292 | return nullptr; |
| 293 | } |
| 294 | |
| 295 | // Check result type. |
| 296 | const analysis::Type* result_type = type_mgr->GetType(inst->type_id()); |
| 297 | const analysis::Vector* vector_type = result_type->AsVector(); |
| 298 | assert(vector_type != nullptr); |
| 299 | const analysis::Type* element_type = vector_type->element_type(); |
| 300 | assert(element_type != nullptr); |
| 301 | const analysis::Float* float_type = element_type->AsFloat(); |
| 302 | assert(float_type != nullptr); |
| 303 | |
| 304 | // Check types of c1 and c2. |
| 305 | assert(c1->type()->AsVector() == vector_type); |
| 306 | assert(c1->type()->AsVector()->element_type() == element_type && |
| 307 | c2->type() == element_type); |
| 308 | |
| 309 | // Get a float vector that is the result of vector-times-scalar. |
| 310 | std::vector<const analysis::Constant*> c1_components = |
| 311 | c1->GetVectorComponents(const_mgr); |
| 312 | std::vector<uint32_t> ids; |
| 313 | if (float_type->width() == 32) { |
| 314 | float scalar = c2->GetFloat(); |
| 315 | for (uint32_t i = 0; i < c1_components.size(); ++i) { |
| 316 | utils::FloatProxy<float> result(c1_components[i]->GetFloat() * scalar); |
| 317 | std::vector<uint32_t> words = result.GetWords(); |
| 318 | const analysis::Constant* new_elem = |
| 319 | const_mgr->GetConstant(float_type, words); |
| 320 | ids.push_back(const_mgr->GetDefiningInstruction(new_elem)->result_id()); |
| 321 | } |
| 322 | return const_mgr->GetConstant(vector_type, ids); |
| 323 | } else if (float_type->width() == 64) { |
| 324 | double scalar = c2->GetDouble(); |
| 325 | for (uint32_t i = 0; i < c1_components.size(); ++i) { |
| 326 | utils::FloatProxy<double> result(c1_components[i]->GetDouble() * |
| 327 | scalar); |
| 328 | std::vector<uint32_t> words = result.GetWords(); |
| 329 | const analysis::Constant* new_elem = |
| 330 | const_mgr->GetConstant(float_type, words); |
| 331 | ids.push_back(const_mgr->GetDefiningInstruction(new_elem)->result_id()); |
| 332 | } |
| 333 | return const_mgr->GetConstant(vector_type, ids); |
| 334 | } |
| 335 | return nullptr; |
| 336 | }; |
| 337 | } |
| 338 | |
Nicolas Capens | 00a1bcc | 2022-07-29 16:49:40 -0400 | [diff] [blame] | 339 | ConstantFoldingRule FoldVectorTimesMatrix() { |
| 340 | return [](IRContext* context, Instruction* inst, |
| 341 | const std::vector<const analysis::Constant*>& constants) |
| 342 | -> const analysis::Constant* { |
Nicolas Capens | 84c9c45 | 2022-11-18 14:11:05 +0000 | [diff] [blame] | 343 | assert(inst->opcode() == spv::Op::OpVectorTimesMatrix); |
Nicolas Capens | 00a1bcc | 2022-07-29 16:49:40 -0400 | [diff] [blame] | 344 | analysis::ConstantManager* const_mgr = context->get_constant_mgr(); |
| 345 | analysis::TypeManager* type_mgr = context->get_type_mgr(); |
| 346 | |
| 347 | if (!inst->IsFloatingPointFoldingAllowed()) { |
| 348 | if (HasFloatingPoint(type_mgr->GetType(inst->type_id()))) { |
| 349 | return nullptr; |
| 350 | } |
| 351 | } |
| 352 | |
| 353 | const analysis::Constant* c1 = constants[0]; |
| 354 | const analysis::Constant* c2 = constants[1]; |
| 355 | |
| 356 | if (c1 == nullptr || c2 == nullptr) { |
| 357 | return nullptr; |
| 358 | } |
| 359 | |
| 360 | // Check result type. |
| 361 | const analysis::Type* result_type = type_mgr->GetType(inst->type_id()); |
| 362 | const analysis::Vector* vector_type = result_type->AsVector(); |
| 363 | assert(vector_type != nullptr); |
| 364 | const analysis::Type* element_type = vector_type->element_type(); |
| 365 | assert(element_type != nullptr); |
| 366 | const analysis::Float* float_type = element_type->AsFloat(); |
| 367 | assert(float_type != nullptr); |
| 368 | |
| 369 | // Check types of c1 and c2. |
| 370 | assert(c1->type()->AsVector() == vector_type); |
| 371 | assert(c1->type()->AsVector()->element_type() == element_type && |
| 372 | c2->type()->AsMatrix()->element_type() == vector_type); |
| 373 | |
| 374 | // Get a float vector that is the result of vector-times-matrix. |
| 375 | std::vector<const analysis::Constant*> c1_components = |
| 376 | c1->GetVectorComponents(const_mgr); |
| 377 | std::vector<const analysis::Constant*> c2_components = |
| 378 | c2->AsMatrixConstant()->GetComponents(); |
| 379 | uint32_t resultVectorSize = result_type->AsVector()->element_count(); |
| 380 | |
| 381 | std::vector<uint32_t> ids; |
| 382 | |
| 383 | if ((c1 && c1->IsZero()) || (c2 && c2->IsZero())) { |
| 384 | std::vector<uint32_t> words(float_type->width() / 32, 0); |
| 385 | for (uint32_t i = 0; i < resultVectorSize; ++i) { |
| 386 | const analysis::Constant* new_elem = |
| 387 | const_mgr->GetConstant(float_type, words); |
| 388 | ids.push_back(const_mgr->GetDefiningInstruction(new_elem)->result_id()); |
| 389 | } |
| 390 | return const_mgr->GetConstant(vector_type, ids); |
| 391 | } |
| 392 | |
| 393 | if (float_type->width() == 32) { |
| 394 | for (uint32_t i = 0; i < resultVectorSize; ++i) { |
| 395 | float result_scalar = 0.0f; |
| 396 | const analysis::VectorConstant* c2_vec = |
| 397 | c2_components[i]->AsVectorConstant(); |
| 398 | for (uint32_t j = 0; j < c2_vec->GetComponents().size(); ++j) { |
| 399 | float c1_scalar = c1_components[j]->GetFloat(); |
| 400 | float c2_scalar = c2_vec->GetComponents()[j]->GetFloat(); |
| 401 | result_scalar += c1_scalar * c2_scalar; |
| 402 | } |
| 403 | utils::FloatProxy<float> result(result_scalar); |
| 404 | std::vector<uint32_t> words = result.GetWords(); |
| 405 | const analysis::Constant* new_elem = |
| 406 | const_mgr->GetConstant(float_type, words); |
| 407 | ids.push_back(const_mgr->GetDefiningInstruction(new_elem)->result_id()); |
| 408 | } |
| 409 | return const_mgr->GetConstant(vector_type, ids); |
| 410 | } else if (float_type->width() == 64) { |
| 411 | for (uint32_t i = 0; i < c2_components.size(); ++i) { |
| 412 | double result_scalar = 0.0; |
| 413 | const analysis::VectorConstant* c2_vec = |
| 414 | c2_components[i]->AsVectorConstant(); |
| 415 | for (uint32_t j = 0; j < c2_vec->GetComponents().size(); ++j) { |
| 416 | double c1_scalar = c1_components[j]->GetDouble(); |
| 417 | double c2_scalar = c2_vec->GetComponents()[j]->GetDouble(); |
| 418 | result_scalar += c1_scalar * c2_scalar; |
| 419 | } |
| 420 | utils::FloatProxy<double> result(result_scalar); |
| 421 | std::vector<uint32_t> words = result.GetWords(); |
| 422 | const analysis::Constant* new_elem = |
| 423 | const_mgr->GetConstant(float_type, words); |
| 424 | ids.push_back(const_mgr->GetDefiningInstruction(new_elem)->result_id()); |
| 425 | } |
| 426 | return const_mgr->GetConstant(vector_type, ids); |
| 427 | } |
| 428 | return nullptr; |
| 429 | }; |
| 430 | } |
| 431 | |
| 432 | ConstantFoldingRule FoldMatrixTimesVector() { |
| 433 | return [](IRContext* context, Instruction* inst, |
| 434 | const std::vector<const analysis::Constant*>& constants) |
| 435 | -> const analysis::Constant* { |
Nicolas Capens | 84c9c45 | 2022-11-18 14:11:05 +0000 | [diff] [blame] | 436 | assert(inst->opcode() == spv::Op::OpMatrixTimesVector); |
Nicolas Capens | 00a1bcc | 2022-07-29 16:49:40 -0400 | [diff] [blame] | 437 | analysis::ConstantManager* const_mgr = context->get_constant_mgr(); |
| 438 | analysis::TypeManager* type_mgr = context->get_type_mgr(); |
| 439 | |
| 440 | if (!inst->IsFloatingPointFoldingAllowed()) { |
| 441 | if (HasFloatingPoint(type_mgr->GetType(inst->type_id()))) { |
| 442 | return nullptr; |
| 443 | } |
| 444 | } |
| 445 | |
| 446 | const analysis::Constant* c1 = constants[0]; |
| 447 | const analysis::Constant* c2 = constants[1]; |
| 448 | |
| 449 | if (c1 == nullptr || c2 == nullptr) { |
| 450 | return nullptr; |
| 451 | } |
| 452 | |
| 453 | // Check result type. |
| 454 | const analysis::Type* result_type = type_mgr->GetType(inst->type_id()); |
| 455 | const analysis::Vector* vector_type = result_type->AsVector(); |
| 456 | assert(vector_type != nullptr); |
| 457 | const analysis::Type* element_type = vector_type->element_type(); |
| 458 | assert(element_type != nullptr); |
| 459 | const analysis::Float* float_type = element_type->AsFloat(); |
| 460 | assert(float_type != nullptr); |
| 461 | |
| 462 | // Check types of c1 and c2. |
| 463 | assert(c1->type()->AsMatrix()->element_type() == vector_type); |
| 464 | assert(c2->type()->AsVector()->element_type() == element_type); |
| 465 | |
| 466 | // Get a float vector that is the result of matrix-times-vector. |
| 467 | std::vector<const analysis::Constant*> c1_components = |
| 468 | c1->AsMatrixConstant()->GetComponents(); |
| 469 | std::vector<const analysis::Constant*> c2_components = |
| 470 | c2->GetVectorComponents(const_mgr); |
| 471 | uint32_t resultVectorSize = result_type->AsVector()->element_count(); |
| 472 | |
| 473 | std::vector<uint32_t> ids; |
| 474 | |
| 475 | if ((c1 && c1->IsZero()) || (c2 && c2->IsZero())) { |
| 476 | std::vector<uint32_t> words(float_type->width() / 32, 0); |
| 477 | for (uint32_t i = 0; i < resultVectorSize; ++i) { |
| 478 | const analysis::Constant* new_elem = |
| 479 | const_mgr->GetConstant(float_type, words); |
| 480 | ids.push_back(const_mgr->GetDefiningInstruction(new_elem)->result_id()); |
| 481 | } |
| 482 | return const_mgr->GetConstant(vector_type, ids); |
| 483 | } |
| 484 | |
| 485 | if (float_type->width() == 32) { |
| 486 | for (uint32_t i = 0; i < resultVectorSize; ++i) { |
| 487 | float result_scalar = 0.0f; |
| 488 | for (uint32_t j = 0; j < c1_components.size(); ++j) { |
| 489 | float c1_scalar = c1_components[j] |
| 490 | ->AsVectorConstant() |
| 491 | ->GetComponents()[i] |
| 492 | ->GetFloat(); |
| 493 | float c2_scalar = c2_components[j]->GetFloat(); |
| 494 | result_scalar += c1_scalar * c2_scalar; |
| 495 | } |
| 496 | utils::FloatProxy<float> result(result_scalar); |
| 497 | std::vector<uint32_t> words = result.GetWords(); |
| 498 | const analysis::Constant* new_elem = |
| 499 | const_mgr->GetConstant(float_type, words); |
| 500 | ids.push_back(const_mgr->GetDefiningInstruction(new_elem)->result_id()); |
| 501 | } |
| 502 | return const_mgr->GetConstant(vector_type, ids); |
| 503 | } else if (float_type->width() == 64) { |
| 504 | for (uint32_t i = 0; i < resultVectorSize; ++i) { |
| 505 | double result_scalar = 0.0; |
| 506 | for (uint32_t j = 0; j < c1_components.size(); ++j) { |
| 507 | double c1_scalar = c1_components[j] |
| 508 | ->AsVectorConstant() |
| 509 | ->GetComponents()[i] |
| 510 | ->GetDouble(); |
| 511 | double c2_scalar = c2_components[j]->GetDouble(); |
| 512 | result_scalar += c1_scalar * c2_scalar; |
| 513 | } |
| 514 | utils::FloatProxy<double> result(result_scalar); |
| 515 | std::vector<uint32_t> words = result.GetWords(); |
| 516 | const analysis::Constant* new_elem = |
| 517 | const_mgr->GetConstant(float_type, words); |
| 518 | ids.push_back(const_mgr->GetDefiningInstruction(new_elem)->result_id()); |
| 519 | } |
| 520 | return const_mgr->GetConstant(vector_type, ids); |
| 521 | } |
| 522 | return nullptr; |
| 523 | }; |
| 524 | } |
| 525 | |
Chris Forbes | cc5697f | 2019-01-30 11:54:08 -0800 | [diff] [blame] | 526 | ConstantFoldingRule FoldCompositeWithConstants() { |
| 527 | // Folds an OpCompositeConstruct where all of the inputs are constants to a |
| 528 | // constant. A new constant is created if necessary. |
| 529 | return [](IRContext* context, Instruction* inst, |
| 530 | const std::vector<const analysis::Constant*>& constants) |
| 531 | -> const analysis::Constant* { |
| 532 | analysis::ConstantManager* const_mgr = context->get_constant_mgr(); |
| 533 | analysis::TypeManager* type_mgr = context->get_type_mgr(); |
| 534 | const analysis::Type* new_type = type_mgr->GetType(inst->type_id()); |
| 535 | Instruction* type_inst = |
| 536 | context->get_def_use_mgr()->GetDef(inst->type_id()); |
| 537 | |
| 538 | std::vector<uint32_t> ids; |
| 539 | for (uint32_t i = 0; i < constants.size(); ++i) { |
| 540 | const analysis::Constant* element_const = constants[i]; |
| 541 | if (element_const == nullptr) { |
| 542 | return nullptr; |
| 543 | } |
| 544 | |
| 545 | uint32_t component_type_id = 0; |
Nicolas Capens | 84c9c45 | 2022-11-18 14:11:05 +0000 | [diff] [blame] | 546 | if (type_inst->opcode() == spv::Op::OpTypeStruct) { |
Chris Forbes | cc5697f | 2019-01-30 11:54:08 -0800 | [diff] [blame] | 547 | component_type_id = type_inst->GetSingleWordInOperand(i); |
Nicolas Capens | 84c9c45 | 2022-11-18 14:11:05 +0000 | [diff] [blame] | 548 | } else if (type_inst->opcode() == spv::Op::OpTypeArray) { |
Chris Forbes | cc5697f | 2019-01-30 11:54:08 -0800 | [diff] [blame] | 549 | component_type_id = type_inst->GetSingleWordInOperand(0); |
| 550 | } |
| 551 | |
| 552 | uint32_t element_id = |
| 553 | const_mgr->FindDeclaredConstant(element_const, component_type_id); |
| 554 | if (element_id == 0) { |
| 555 | return nullptr; |
| 556 | } |
| 557 | ids.push_back(element_id); |
| 558 | } |
| 559 | return const_mgr->GetConstant(new_type, ids); |
| 560 | }; |
| 561 | } |
| 562 | |
| 563 | // The interface for a function that returns the result of applying a scalar |
| 564 | // floating-point binary operation on |a| and |b|. The type of the return value |
| 565 | // will be |type|. The input constants must also be of type |type|. |
| 566 | using UnaryScalarFoldingRule = std::function<const analysis::Constant*( |
| 567 | const analysis::Type* result_type, const analysis::Constant* a, |
| 568 | analysis::ConstantManager*)>; |
| 569 | |
| 570 | // The interface for a function that returns the result of applying a scalar |
| 571 | // floating-point binary operation on |a| and |b|. The type of the return value |
| 572 | // will be |type|. The input constants must also be of type |type|. |
| 573 | using BinaryScalarFoldingRule = std::function<const analysis::Constant*( |
| 574 | const analysis::Type* result_type, const analysis::Constant* a, |
| 575 | const analysis::Constant* b, analysis::ConstantManager*)>; |
| 576 | |
| 577 | // Returns a |ConstantFoldingRule| that folds unary floating point scalar ops |
| 578 | // using |scalar_rule| and unary float point vectors ops by applying |
| 579 | // |scalar_rule| to the elements of the vector. The |ConstantFoldingRule| |
| 580 | // that is returned assumes that |constants| contains 1 entry. If they are |
| 581 | // not |nullptr|, then their type is either |Float| or |Integer| or a |Vector| |
| 582 | // whose element type is |Float| or |Integer|. |
| 583 | ConstantFoldingRule FoldFPUnaryOp(UnaryScalarFoldingRule scalar_rule) { |
| 584 | return [scalar_rule](IRContext* context, Instruction* inst, |
| 585 | const std::vector<const analysis::Constant*>& constants) |
| 586 | -> const analysis::Constant* { |
| 587 | analysis::ConstantManager* const_mgr = context->get_constant_mgr(); |
| 588 | analysis::TypeManager* type_mgr = context->get_type_mgr(); |
| 589 | const analysis::Type* result_type = type_mgr->GetType(inst->type_id()); |
| 590 | const analysis::Vector* vector_type = result_type->AsVector(); |
| 591 | |
| 592 | if (!inst->IsFloatingPointFoldingAllowed()) { |
| 593 | return nullptr; |
| 594 | } |
| 595 | |
Ben Clayton | dc6b76a | 2020-02-24 14:53:40 +0000 | [diff] [blame] | 596 | const analysis::Constant* arg = |
Nicolas Capens | 84c9c45 | 2022-11-18 14:11:05 +0000 | [diff] [blame] | 597 | (inst->opcode() == spv::Op::OpExtInst) ? constants[1] : constants[0]; |
Ben Clayton | dc6b76a | 2020-02-24 14:53:40 +0000 | [diff] [blame] | 598 | |
| 599 | if (arg == nullptr) { |
Chris Forbes | cc5697f | 2019-01-30 11:54:08 -0800 | [diff] [blame] | 600 | return nullptr; |
| 601 | } |
| 602 | |
| 603 | if (vector_type != nullptr) { |
| 604 | std::vector<const analysis::Constant*> a_components; |
| 605 | std::vector<const analysis::Constant*> results_components; |
| 606 | |
Ben Clayton | dc6b76a | 2020-02-24 14:53:40 +0000 | [diff] [blame] | 607 | a_components = arg->GetVectorComponents(const_mgr); |
Chris Forbes | cc5697f | 2019-01-30 11:54:08 -0800 | [diff] [blame] | 608 | |
| 609 | // Fold each component of the vector. |
| 610 | for (uint32_t i = 0; i < a_components.size(); ++i) { |
| 611 | results_components.push_back(scalar_rule(vector_type->element_type(), |
| 612 | a_components[i], const_mgr)); |
| 613 | if (results_components[i] == nullptr) { |
| 614 | return nullptr; |
| 615 | } |
| 616 | } |
| 617 | |
| 618 | // Build the constant object and return it. |
| 619 | std::vector<uint32_t> ids; |
| 620 | for (const analysis::Constant* member : results_components) { |
| 621 | ids.push_back(const_mgr->GetDefiningInstruction(member)->result_id()); |
| 622 | } |
| 623 | return const_mgr->GetConstant(vector_type, ids); |
| 624 | } else { |
Ben Clayton | dc6b76a | 2020-02-24 14:53:40 +0000 | [diff] [blame] | 625 | return scalar_rule(result_type, arg, const_mgr); |
Chris Forbes | cc5697f | 2019-01-30 11:54:08 -0800 | [diff] [blame] | 626 | } |
| 627 | }; |
| 628 | } |
| 629 | |
Ben Clayton | d552f63 | 2019-11-18 11:18:41 +0000 | [diff] [blame] | 630 | // Returns the result of folding the constants in |constants| according the |
| 631 | // |scalar_rule|. If |result_type| is a vector, then |scalar_rule| is applied |
| 632 | // per component. |
| 633 | const analysis::Constant* FoldFPBinaryOp( |
| 634 | BinaryScalarFoldingRule scalar_rule, uint32_t result_type_id, |
| 635 | const std::vector<const analysis::Constant*>& constants, |
| 636 | IRContext* context) { |
| 637 | analysis::ConstantManager* const_mgr = context->get_constant_mgr(); |
| 638 | analysis::TypeManager* type_mgr = context->get_type_mgr(); |
| 639 | const analysis::Type* result_type = type_mgr->GetType(result_type_id); |
| 640 | const analysis::Vector* vector_type = result_type->AsVector(); |
| 641 | |
| 642 | if (constants[0] == nullptr || constants[1] == nullptr) { |
| 643 | return nullptr; |
| 644 | } |
| 645 | |
| 646 | if (vector_type != nullptr) { |
| 647 | std::vector<const analysis::Constant*> a_components; |
| 648 | std::vector<const analysis::Constant*> b_components; |
| 649 | std::vector<const analysis::Constant*> results_components; |
| 650 | |
| 651 | a_components = constants[0]->GetVectorComponents(const_mgr); |
| 652 | b_components = constants[1]->GetVectorComponents(const_mgr); |
| 653 | |
| 654 | // Fold each component of the vector. |
| 655 | for (uint32_t i = 0; i < a_components.size(); ++i) { |
| 656 | results_components.push_back(scalar_rule(vector_type->element_type(), |
| 657 | a_components[i], b_components[i], |
| 658 | const_mgr)); |
| 659 | if (results_components[i] == nullptr) { |
| 660 | return nullptr; |
| 661 | } |
| 662 | } |
| 663 | |
| 664 | // Build the constant object and return it. |
| 665 | std::vector<uint32_t> ids; |
| 666 | for (const analysis::Constant* member : results_components) { |
| 667 | ids.push_back(const_mgr->GetDefiningInstruction(member)->result_id()); |
| 668 | } |
| 669 | return const_mgr->GetConstant(vector_type, ids); |
| 670 | } else { |
| 671 | return scalar_rule(result_type, constants[0], constants[1], const_mgr); |
| 672 | } |
| 673 | } |
| 674 | |
Chris Forbes | cc5697f | 2019-01-30 11:54:08 -0800 | [diff] [blame] | 675 | // Returns a |ConstantFoldingRule| that folds floating point scalars using |
| 676 | // |scalar_rule| and vectors of floating point by applying |scalar_rule| to the |
| 677 | // elements of the vector. The |ConstantFoldingRule| that is returned assumes |
| 678 | // that |constants| contains 2 entries. If they are not |nullptr|, then their |
| 679 | // type is either |Float| or a |Vector| whose element type is |Float|. |
| 680 | ConstantFoldingRule FoldFPBinaryOp(BinaryScalarFoldingRule scalar_rule) { |
| 681 | return [scalar_rule](IRContext* context, Instruction* inst, |
| 682 | const std::vector<const analysis::Constant*>& constants) |
| 683 | -> const analysis::Constant* { |
Chris Forbes | cc5697f | 2019-01-30 11:54:08 -0800 | [diff] [blame] | 684 | if (!inst->IsFloatingPointFoldingAllowed()) { |
| 685 | return nullptr; |
| 686 | } |
Nicolas Capens | 84c9c45 | 2022-11-18 14:11:05 +0000 | [diff] [blame] | 687 | if (inst->opcode() == spv::Op::OpExtInst) { |
Ben Clayton | d552f63 | 2019-11-18 11:18:41 +0000 | [diff] [blame] | 688 | return FoldFPBinaryOp(scalar_rule, inst->type_id(), |
| 689 | {constants[1], constants[2]}, context); |
Chris Forbes | cc5697f | 2019-01-30 11:54:08 -0800 | [diff] [blame] | 690 | } |
Ben Clayton | d552f63 | 2019-11-18 11:18:41 +0000 | [diff] [blame] | 691 | return FoldFPBinaryOp(scalar_rule, inst->type_id(), constants, context); |
Chris Forbes | cc5697f | 2019-01-30 11:54:08 -0800 | [diff] [blame] | 692 | }; |
| 693 | } |
| 694 | |
| 695 | // This macro defines a |UnaryScalarFoldingRule| that performs float to |
| 696 | // integer conversion. |
| 697 | // TODO(greg-lunarg): Support for 64-bit integer types. |
| 698 | UnaryScalarFoldingRule FoldFToIOp() { |
| 699 | return [](const analysis::Type* result_type, const analysis::Constant* a, |
| 700 | analysis::ConstantManager* const_mgr) -> const analysis::Constant* { |
| 701 | assert(result_type != nullptr && a != nullptr); |
| 702 | const analysis::Integer* integer_type = result_type->AsInteger(); |
| 703 | const analysis::Float* float_type = a->type()->AsFloat(); |
| 704 | assert(float_type != nullptr); |
| 705 | assert(integer_type != nullptr); |
| 706 | if (integer_type->width() != 32) return nullptr; |
| 707 | if (float_type->width() == 32) { |
| 708 | float fa = a->GetFloat(); |
| 709 | uint32_t result = integer_type->IsSigned() |
| 710 | ? static_cast<uint32_t>(static_cast<int32_t>(fa)) |
| 711 | : static_cast<uint32_t>(fa); |
| 712 | std::vector<uint32_t> words = {result}; |
| 713 | return const_mgr->GetConstant(result_type, words); |
| 714 | } else if (float_type->width() == 64) { |
| 715 | double fa = a->GetDouble(); |
| 716 | uint32_t result = integer_type->IsSigned() |
| 717 | ? static_cast<uint32_t>(static_cast<int32_t>(fa)) |
| 718 | : static_cast<uint32_t>(fa); |
| 719 | std::vector<uint32_t> words = {result}; |
| 720 | return const_mgr->GetConstant(result_type, words); |
| 721 | } |
| 722 | return nullptr; |
| 723 | }; |
| 724 | } |
| 725 | |
| 726 | // This function defines a |UnaryScalarFoldingRule| that performs integer to |
| 727 | // float conversion. |
| 728 | // TODO(greg-lunarg): Support for 64-bit integer types. |
| 729 | UnaryScalarFoldingRule FoldIToFOp() { |
| 730 | return [](const analysis::Type* result_type, const analysis::Constant* a, |
| 731 | analysis::ConstantManager* const_mgr) -> const analysis::Constant* { |
| 732 | assert(result_type != nullptr && a != nullptr); |
| 733 | const analysis::Integer* integer_type = a->type()->AsInteger(); |
| 734 | const analysis::Float* float_type = result_type->AsFloat(); |
| 735 | assert(float_type != nullptr); |
| 736 | assert(integer_type != nullptr); |
| 737 | if (integer_type->width() != 32) return nullptr; |
| 738 | uint32_t ua = a->GetU32(); |
| 739 | if (float_type->width() == 32) { |
| 740 | float result_val = integer_type->IsSigned() |
| 741 | ? static_cast<float>(static_cast<int32_t>(ua)) |
| 742 | : static_cast<float>(ua); |
| 743 | utils::FloatProxy<float> result(result_val); |
| 744 | std::vector<uint32_t> words = {result.data()}; |
| 745 | return const_mgr->GetConstant(result_type, words); |
| 746 | } else if (float_type->width() == 64) { |
| 747 | double result_val = integer_type->IsSigned() |
| 748 | ? static_cast<double>(static_cast<int32_t>(ua)) |
| 749 | : static_cast<double>(ua); |
| 750 | utils::FloatProxy<double> result(result_val); |
| 751 | std::vector<uint32_t> words = result.GetWords(); |
| 752 | return const_mgr->GetConstant(result_type, words); |
| 753 | } |
| 754 | return nullptr; |
| 755 | }; |
| 756 | } |
| 757 | |
Ben Clayton | b73b760 | 2019-07-29 13:56:13 +0100 | [diff] [blame] | 758 | // This defines a |UnaryScalarFoldingRule| that performs |OpQuantizeToF16|. |
| 759 | UnaryScalarFoldingRule FoldQuantizeToF16Scalar() { |
| 760 | return [](const analysis::Type* result_type, const analysis::Constant* a, |
| 761 | analysis::ConstantManager* const_mgr) -> const analysis::Constant* { |
| 762 | assert(result_type != nullptr && a != nullptr); |
| 763 | const analysis::Float* float_type = a->type()->AsFloat(); |
| 764 | assert(float_type != nullptr); |
| 765 | if (float_type->width() != 32) { |
| 766 | return nullptr; |
| 767 | } |
| 768 | |
| 769 | float fa = a->GetFloat(); |
| 770 | utils::HexFloat<utils::FloatProxy<float>> orignal(fa); |
| 771 | utils::HexFloat<utils::FloatProxy<utils::Float16>> quantized(0); |
| 772 | utils::HexFloat<utils::FloatProxy<float>> result(0.0f); |
| 773 | orignal.castTo(quantized, utils::round_direction::kToZero); |
| 774 | quantized.castTo(result, utils::round_direction::kToZero); |
| 775 | std::vector<uint32_t> words = {result.getBits()}; |
| 776 | return const_mgr->GetConstant(result_type, words); |
| 777 | }; |
| 778 | } |
| 779 | |
Chris Forbes | cc5697f | 2019-01-30 11:54:08 -0800 | [diff] [blame] | 780 | // This macro defines a |BinaryScalarFoldingRule| that applies |op|. The |
| 781 | // operator |op| must work for both float and double, and use syntax "f1 op f2". |
Ben Clayton | d552f63 | 2019-11-18 11:18:41 +0000 | [diff] [blame] | 782 | #define FOLD_FPARITH_OP(op) \ |
| 783 | [](const analysis::Type* result_type_in_macro, const analysis::Constant* a, \ |
| 784 | const analysis::Constant* b, \ |
| 785 | analysis::ConstantManager* const_mgr_in_macro) \ |
| 786 | -> const analysis::Constant* { \ |
| 787 | assert(result_type_in_macro != nullptr && a != nullptr && b != nullptr); \ |
| 788 | assert(result_type_in_macro == a->type() && \ |
| 789 | result_type_in_macro == b->type()); \ |
| 790 | const analysis::Float* float_type_in_macro = \ |
| 791 | result_type_in_macro->AsFloat(); \ |
| 792 | assert(float_type_in_macro != nullptr); \ |
| 793 | if (float_type_in_macro->width() == 32) { \ |
| 794 | float fa = a->GetFloat(); \ |
| 795 | float fb = b->GetFloat(); \ |
| 796 | utils::FloatProxy<float> result_in_macro(fa op fb); \ |
| 797 | std::vector<uint32_t> words_in_macro = result_in_macro.GetWords(); \ |
| 798 | return const_mgr_in_macro->GetConstant(result_type_in_macro, \ |
| 799 | words_in_macro); \ |
| 800 | } else if (float_type_in_macro->width() == 64) { \ |
| 801 | double fa = a->GetDouble(); \ |
| 802 | double fb = b->GetDouble(); \ |
| 803 | utils::FloatProxy<double> result_in_macro(fa op fb); \ |
| 804 | std::vector<uint32_t> words_in_macro = result_in_macro.GetWords(); \ |
| 805 | return const_mgr_in_macro->GetConstant(result_type_in_macro, \ |
| 806 | words_in_macro); \ |
| 807 | } \ |
| 808 | return nullptr; \ |
Chris Forbes | cc5697f | 2019-01-30 11:54:08 -0800 | [diff] [blame] | 809 | } |
| 810 | |
| 811 | // Define the folding rule for conversion between floating point and integer |
| 812 | ConstantFoldingRule FoldFToI() { return FoldFPUnaryOp(FoldFToIOp()); } |
| 813 | ConstantFoldingRule FoldIToF() { return FoldFPUnaryOp(FoldIToFOp()); } |
Ben Clayton | b73b760 | 2019-07-29 13:56:13 +0100 | [diff] [blame] | 814 | ConstantFoldingRule FoldQuantizeToF16() { |
| 815 | return FoldFPUnaryOp(FoldQuantizeToF16Scalar()); |
| 816 | } |
Chris Forbes | cc5697f | 2019-01-30 11:54:08 -0800 | [diff] [blame] | 817 | |
| 818 | // Define the folding rules for subtraction, addition, multiplication, and |
| 819 | // division for floating point values. |
| 820 | ConstantFoldingRule FoldFSub() { return FoldFPBinaryOp(FOLD_FPARITH_OP(-)); } |
| 821 | ConstantFoldingRule FoldFAdd() { return FoldFPBinaryOp(FOLD_FPARITH_OP(+)); } |
| 822 | ConstantFoldingRule FoldFMul() { return FoldFPBinaryOp(FOLD_FPARITH_OP(*)); } |
Nicolas Capens | 6cacf18 | 2021-11-30 11:15:46 -0500 | [diff] [blame] | 823 | |
| 824 | // Returns the constant that results from evaluating |numerator| / 0.0. Returns |
sugoi1 | b398bf3 | 2022-02-18 10:27:28 -0500 | [diff] [blame] | 825 | // |nullptr| if the result could not be evaluated. |
Nicolas Capens | 6cacf18 | 2021-11-30 11:15:46 -0500 | [diff] [blame] | 826 | const analysis::Constant* FoldFPScalarDivideByZero( |
| 827 | const analysis::Type* result_type, const analysis::Constant* numerator, |
| 828 | analysis::ConstantManager* const_mgr) { |
| 829 | if (numerator == nullptr) { |
| 830 | return nullptr; |
| 831 | } |
| 832 | |
| 833 | if (numerator->IsZero()) { |
| 834 | return GetNan(result_type, const_mgr); |
| 835 | } |
| 836 | |
| 837 | const analysis::Constant* result = GetInf(result_type, const_mgr); |
| 838 | if (result == nullptr) { |
| 839 | return nullptr; |
| 840 | } |
| 841 | |
| 842 | if (numerator->AsFloatConstant()->GetValueAsDouble() < 0.0) { |
Nicolas Capens | 84c9c45 | 2022-11-18 14:11:05 +0000 | [diff] [blame] | 843 | result = NegateFPConst(result_type, result, const_mgr); |
Nicolas Capens | 6cacf18 | 2021-11-30 11:15:46 -0500 | [diff] [blame] | 844 | } |
| 845 | return result; |
| 846 | } |
| 847 | |
| 848 | // Returns the result of folding |numerator| / |denominator|. Returns |nullptr| |
| 849 | // if it cannot be folded. |
| 850 | const analysis::Constant* FoldScalarFPDivide( |
| 851 | const analysis::Type* result_type, const analysis::Constant* numerator, |
| 852 | const analysis::Constant* denominator, |
| 853 | analysis::ConstantManager* const_mgr) { |
| 854 | if (denominator == nullptr) { |
| 855 | return nullptr; |
| 856 | } |
| 857 | |
| 858 | if (denominator->IsZero()) { |
| 859 | return FoldFPScalarDivideByZero(result_type, numerator, const_mgr); |
| 860 | } |
| 861 | |
| 862 | const analysis::FloatConstant* denominator_float = |
| 863 | denominator->AsFloatConstant(); |
| 864 | if (denominator_float && denominator->GetValueAsDouble() == -0.0) { |
| 865 | const analysis::Constant* result = |
| 866 | FoldFPScalarDivideByZero(result_type, numerator, const_mgr); |
| 867 | if (result != nullptr) |
Nicolas Capens | 84c9c45 | 2022-11-18 14:11:05 +0000 | [diff] [blame] | 868 | result = NegateFPConst(result_type, result, const_mgr); |
Nicolas Capens | 6cacf18 | 2021-11-30 11:15:46 -0500 | [diff] [blame] | 869 | return result; |
| 870 | } else { |
| 871 | return FOLD_FPARITH_OP(/)(result_type, numerator, denominator, const_mgr); |
| 872 | } |
| 873 | } |
| 874 | |
| 875 | // Returns the constant folding rule to fold |OpFDiv| with two constants. |
| 876 | ConstantFoldingRule FoldFDiv() { return FoldFPBinaryOp(FoldScalarFPDivide); } |
Chris Forbes | cc5697f | 2019-01-30 11:54:08 -0800 | [diff] [blame] | 877 | |
| 878 | bool CompareFloatingPoint(bool op_result, bool op_unordered, |
| 879 | bool need_ordered) { |
| 880 | if (need_ordered) { |
| 881 | // operands are ordered and Operand 1 is |op| Operand 2 |
| 882 | return !op_unordered && op_result; |
| 883 | } else { |
| 884 | // operands are unordered or Operand 1 is |op| Operand 2 |
| 885 | return op_unordered || op_result; |
| 886 | } |
| 887 | } |
| 888 | |
| 889 | // This macro defines a |BinaryScalarFoldingRule| that applies |op|. The |
| 890 | // operator |op| must work for both float and double, and use syntax "f1 op f2". |
| 891 | #define FOLD_FPCMP_OP(op, ord) \ |
| 892 | [](const analysis::Type* result_type, const analysis::Constant* a, \ |
| 893 | const analysis::Constant* b, \ |
| 894 | analysis::ConstantManager* const_mgr) -> const analysis::Constant* { \ |
| 895 | assert(result_type != nullptr && a != nullptr && b != nullptr); \ |
| 896 | assert(result_type->AsBool()); \ |
| 897 | assert(a->type() == b->type()); \ |
| 898 | const analysis::Float* float_type = a->type()->AsFloat(); \ |
| 899 | assert(float_type != nullptr); \ |
| 900 | if (float_type->width() == 32) { \ |
| 901 | float fa = a->GetFloat(); \ |
| 902 | float fb = b->GetFloat(); \ |
| 903 | bool result = CompareFloatingPoint( \ |
| 904 | fa op fb, std::isnan(fa) || std::isnan(fb), ord); \ |
| 905 | std::vector<uint32_t> words = {uint32_t(result)}; \ |
| 906 | return const_mgr->GetConstant(result_type, words); \ |
| 907 | } else if (float_type->width() == 64) { \ |
| 908 | double fa = a->GetDouble(); \ |
| 909 | double fb = b->GetDouble(); \ |
| 910 | bool result = CompareFloatingPoint( \ |
| 911 | fa op fb, std::isnan(fa) || std::isnan(fb), ord); \ |
| 912 | std::vector<uint32_t> words = {uint32_t(result)}; \ |
| 913 | return const_mgr->GetConstant(result_type, words); \ |
| 914 | } \ |
| 915 | return nullptr; \ |
| 916 | } |
| 917 | |
| 918 | // Define the folding rules for ordered and unordered comparison for floating |
| 919 | // point values. |
| 920 | ConstantFoldingRule FoldFOrdEqual() { |
| 921 | return FoldFPBinaryOp(FOLD_FPCMP_OP(==, true)); |
| 922 | } |
| 923 | ConstantFoldingRule FoldFUnordEqual() { |
| 924 | return FoldFPBinaryOp(FOLD_FPCMP_OP(==, false)); |
| 925 | } |
| 926 | ConstantFoldingRule FoldFOrdNotEqual() { |
| 927 | return FoldFPBinaryOp(FOLD_FPCMP_OP(!=, true)); |
| 928 | } |
| 929 | ConstantFoldingRule FoldFUnordNotEqual() { |
| 930 | return FoldFPBinaryOp(FOLD_FPCMP_OP(!=, false)); |
| 931 | } |
| 932 | ConstantFoldingRule FoldFOrdLessThan() { |
| 933 | return FoldFPBinaryOp(FOLD_FPCMP_OP(<, true)); |
| 934 | } |
| 935 | ConstantFoldingRule FoldFUnordLessThan() { |
| 936 | return FoldFPBinaryOp(FOLD_FPCMP_OP(<, false)); |
| 937 | } |
| 938 | ConstantFoldingRule FoldFOrdGreaterThan() { |
| 939 | return FoldFPBinaryOp(FOLD_FPCMP_OP(>, true)); |
| 940 | } |
| 941 | ConstantFoldingRule FoldFUnordGreaterThan() { |
| 942 | return FoldFPBinaryOp(FOLD_FPCMP_OP(>, false)); |
| 943 | } |
| 944 | ConstantFoldingRule FoldFOrdLessThanEqual() { |
| 945 | return FoldFPBinaryOp(FOLD_FPCMP_OP(<=, true)); |
| 946 | } |
| 947 | ConstantFoldingRule FoldFUnordLessThanEqual() { |
| 948 | return FoldFPBinaryOp(FOLD_FPCMP_OP(<=, false)); |
| 949 | } |
| 950 | ConstantFoldingRule FoldFOrdGreaterThanEqual() { |
| 951 | return FoldFPBinaryOp(FOLD_FPCMP_OP(>=, true)); |
| 952 | } |
| 953 | ConstantFoldingRule FoldFUnordGreaterThanEqual() { |
| 954 | return FoldFPBinaryOp(FOLD_FPCMP_OP(>=, false)); |
| 955 | } |
| 956 | |
| 957 | // Folds an OpDot where all of the inputs are constants to a |
| 958 | // constant. A new constant is created if necessary. |
| 959 | ConstantFoldingRule FoldOpDotWithConstants() { |
| 960 | return [](IRContext* context, Instruction* inst, |
| 961 | const std::vector<const analysis::Constant*>& constants) |
| 962 | -> const analysis::Constant* { |
| 963 | analysis::ConstantManager* const_mgr = context->get_constant_mgr(); |
| 964 | analysis::TypeManager* type_mgr = context->get_type_mgr(); |
| 965 | const analysis::Type* new_type = type_mgr->GetType(inst->type_id()); |
| 966 | assert(new_type->AsFloat() && "OpDot should have a float return type."); |
| 967 | const analysis::Float* float_type = new_type->AsFloat(); |
| 968 | |
| 969 | if (!inst->IsFloatingPointFoldingAllowed()) { |
| 970 | return nullptr; |
| 971 | } |
| 972 | |
| 973 | // If one of the operands is 0, then the result is 0. |
| 974 | bool has_zero_operand = false; |
| 975 | |
| 976 | for (int i = 0; i < 2; ++i) { |
| 977 | if (constants[i]) { |
| 978 | if (constants[i]->AsNullConstant() || |
| 979 | constants[i]->AsVectorConstant()->IsZero()) { |
| 980 | has_zero_operand = true; |
| 981 | break; |
| 982 | } |
| 983 | } |
| 984 | } |
| 985 | |
| 986 | if (has_zero_operand) { |
| 987 | if (float_type->width() == 32) { |
| 988 | utils::FloatProxy<float> result(0.0f); |
| 989 | std::vector<uint32_t> words = result.GetWords(); |
| 990 | return const_mgr->GetConstant(float_type, words); |
| 991 | } |
| 992 | if (float_type->width() == 64) { |
| 993 | utils::FloatProxy<double> result(0.0); |
| 994 | std::vector<uint32_t> words = result.GetWords(); |
| 995 | return const_mgr->GetConstant(float_type, words); |
| 996 | } |
| 997 | return nullptr; |
| 998 | } |
| 999 | |
| 1000 | if (constants[0] == nullptr || constants[1] == nullptr) { |
| 1001 | return nullptr; |
| 1002 | } |
| 1003 | |
| 1004 | std::vector<const analysis::Constant*> a_components; |
| 1005 | std::vector<const analysis::Constant*> b_components; |
| 1006 | |
| 1007 | a_components = constants[0]->GetVectorComponents(const_mgr); |
| 1008 | b_components = constants[1]->GetVectorComponents(const_mgr); |
| 1009 | |
| 1010 | utils::FloatProxy<double> result(0.0); |
| 1011 | std::vector<uint32_t> words = result.GetWords(); |
| 1012 | const analysis::Constant* result_const = |
| 1013 | const_mgr->GetConstant(float_type, words); |
Ben Clayton | b73b760 | 2019-07-29 13:56:13 +0100 | [diff] [blame] | 1014 | for (uint32_t i = 0; i < a_components.size() && result_const != nullptr; |
| 1015 | ++i) { |
Chris Forbes | cc5697f | 2019-01-30 11:54:08 -0800 | [diff] [blame] | 1016 | if (a_components[i] == nullptr || b_components[i] == nullptr) { |
| 1017 | return nullptr; |
| 1018 | } |
| 1019 | |
| 1020 | const analysis::Constant* component = FOLD_FPARITH_OP(*)( |
| 1021 | new_type, a_components[i], b_components[i], const_mgr); |
Ben Clayton | b73b760 | 2019-07-29 13:56:13 +0100 | [diff] [blame] | 1022 | if (component == nullptr) { |
| 1023 | return nullptr; |
| 1024 | } |
Chris Forbes | cc5697f | 2019-01-30 11:54:08 -0800 | [diff] [blame] | 1025 | result_const = |
| 1026 | FOLD_FPARITH_OP(+)(new_type, result_const, component, const_mgr); |
| 1027 | } |
| 1028 | return result_const; |
| 1029 | }; |
| 1030 | } |
| 1031 | |
| 1032 | // This function defines a |UnaryScalarFoldingRule| that subtracts the constant |
| 1033 | // from zero. |
| 1034 | UnaryScalarFoldingRule FoldFNegateOp() { |
| 1035 | return [](const analysis::Type* result_type, const analysis::Constant* a, |
| 1036 | analysis::ConstantManager* const_mgr) -> const analysis::Constant* { |
| 1037 | assert(result_type != nullptr && a != nullptr); |
| 1038 | assert(result_type == a->type()); |
Nicolas Capens | 84c9c45 | 2022-11-18 14:11:05 +0000 | [diff] [blame] | 1039 | return NegateFPConst(result_type, a, const_mgr); |
Chris Forbes | cc5697f | 2019-01-30 11:54:08 -0800 | [diff] [blame] | 1040 | }; |
| 1041 | } |
| 1042 | |
| 1043 | ConstantFoldingRule FoldFNegate() { return FoldFPUnaryOp(FoldFNegateOp()); } |
| 1044 | |
Nicolas Capens | 84c9c45 | 2022-11-18 14:11:05 +0000 | [diff] [blame] | 1045 | ConstantFoldingRule FoldFClampFeedingCompare(spv::Op cmp_opcode) { |
Chris Forbes | cc5697f | 2019-01-30 11:54:08 -0800 | [diff] [blame] | 1046 | return [cmp_opcode](IRContext* context, Instruction* inst, |
| 1047 | const std::vector<const analysis::Constant*>& constants) |
| 1048 | -> const analysis::Constant* { |
| 1049 | analysis::ConstantManager* const_mgr = context->get_constant_mgr(); |
| 1050 | analysis::DefUseManager* def_use_mgr = context->get_def_use_mgr(); |
| 1051 | |
| 1052 | if (!inst->IsFloatingPointFoldingAllowed()) { |
| 1053 | return nullptr; |
| 1054 | } |
| 1055 | |
| 1056 | uint32_t non_const_idx = (constants[0] ? 1 : 0); |
| 1057 | uint32_t operand_id = inst->GetSingleWordInOperand(non_const_idx); |
| 1058 | Instruction* operand_inst = def_use_mgr->GetDef(operand_id); |
| 1059 | |
| 1060 | analysis::TypeManager* type_mgr = context->get_type_mgr(); |
| 1061 | const analysis::Type* operand_type = |
| 1062 | type_mgr->GetType(operand_inst->type_id()); |
| 1063 | |
| 1064 | if (!operand_type->AsFloat()) { |
| 1065 | return nullptr; |
| 1066 | } |
| 1067 | |
| 1068 | if (operand_type->AsFloat()->width() != 32 && |
| 1069 | operand_type->AsFloat()->width() != 64) { |
| 1070 | return nullptr; |
| 1071 | } |
| 1072 | |
Nicolas Capens | 84c9c45 | 2022-11-18 14:11:05 +0000 | [diff] [blame] | 1073 | if (operand_inst->opcode() != spv::Op::OpExtInst) { |
Chris Forbes | cc5697f | 2019-01-30 11:54:08 -0800 | [diff] [blame] | 1074 | return nullptr; |
| 1075 | } |
| 1076 | |
| 1077 | if (operand_inst->GetSingleWordInOperand(1) != GLSLstd450FClamp) { |
| 1078 | return nullptr; |
| 1079 | } |
| 1080 | |
| 1081 | if (constants[1] == nullptr && constants[0] == nullptr) { |
| 1082 | return nullptr; |
| 1083 | } |
| 1084 | |
| 1085 | uint32_t max_id = operand_inst->GetSingleWordInOperand(4); |
| 1086 | const analysis::Constant* max_const = |
| 1087 | const_mgr->FindDeclaredConstant(max_id); |
| 1088 | |
| 1089 | uint32_t min_id = operand_inst->GetSingleWordInOperand(3); |
| 1090 | const analysis::Constant* min_const = |
| 1091 | const_mgr->FindDeclaredConstant(min_id); |
| 1092 | |
| 1093 | bool found_result = false; |
| 1094 | bool result = false; |
| 1095 | |
| 1096 | switch (cmp_opcode) { |
Nicolas Capens | 84c9c45 | 2022-11-18 14:11:05 +0000 | [diff] [blame] | 1097 | case spv::Op::OpFOrdLessThan: |
| 1098 | case spv::Op::OpFUnordLessThan: |
| 1099 | case spv::Op::OpFOrdGreaterThanEqual: |
| 1100 | case spv::Op::OpFUnordGreaterThanEqual: |
Chris Forbes | cc5697f | 2019-01-30 11:54:08 -0800 | [diff] [blame] | 1101 | if (constants[0]) { |
| 1102 | if (min_const) { |
| 1103 | if (constants[0]->GetValueAsDouble() < |
| 1104 | min_const->GetValueAsDouble()) { |
| 1105 | found_result = true; |
Nicolas Capens | 84c9c45 | 2022-11-18 14:11:05 +0000 | [diff] [blame] | 1106 | result = (cmp_opcode == spv::Op::OpFOrdLessThan || |
| 1107 | cmp_opcode == spv::Op::OpFUnordLessThan); |
Chris Forbes | cc5697f | 2019-01-30 11:54:08 -0800 | [diff] [blame] | 1108 | } |
| 1109 | } |
| 1110 | if (max_const) { |
| 1111 | if (constants[0]->GetValueAsDouble() >= |
| 1112 | max_const->GetValueAsDouble()) { |
| 1113 | found_result = true; |
Nicolas Capens | 84c9c45 | 2022-11-18 14:11:05 +0000 | [diff] [blame] | 1114 | result = !(cmp_opcode == spv::Op::OpFOrdLessThan || |
| 1115 | cmp_opcode == spv::Op::OpFUnordLessThan); |
Chris Forbes | cc5697f | 2019-01-30 11:54:08 -0800 | [diff] [blame] | 1116 | } |
| 1117 | } |
| 1118 | } |
| 1119 | |
| 1120 | if (constants[1]) { |
| 1121 | if (max_const) { |
| 1122 | if (max_const->GetValueAsDouble() < |
| 1123 | constants[1]->GetValueAsDouble()) { |
| 1124 | found_result = true; |
Nicolas Capens | 84c9c45 | 2022-11-18 14:11:05 +0000 | [diff] [blame] | 1125 | result = (cmp_opcode == spv::Op::OpFOrdLessThan || |
| 1126 | cmp_opcode == spv::Op::OpFUnordLessThan); |
Chris Forbes | cc5697f | 2019-01-30 11:54:08 -0800 | [diff] [blame] | 1127 | } |
| 1128 | } |
| 1129 | |
| 1130 | if (min_const) { |
| 1131 | if (min_const->GetValueAsDouble() >= |
| 1132 | constants[1]->GetValueAsDouble()) { |
| 1133 | found_result = true; |
Nicolas Capens | 84c9c45 | 2022-11-18 14:11:05 +0000 | [diff] [blame] | 1134 | result = !(cmp_opcode == spv::Op::OpFOrdLessThan || |
| 1135 | cmp_opcode == spv::Op::OpFUnordLessThan); |
Chris Forbes | cc5697f | 2019-01-30 11:54:08 -0800 | [diff] [blame] | 1136 | } |
| 1137 | } |
| 1138 | } |
| 1139 | break; |
Nicolas Capens | 84c9c45 | 2022-11-18 14:11:05 +0000 | [diff] [blame] | 1140 | case spv::Op::OpFOrdGreaterThan: |
| 1141 | case spv::Op::OpFUnordGreaterThan: |
| 1142 | case spv::Op::OpFOrdLessThanEqual: |
| 1143 | case spv::Op::OpFUnordLessThanEqual: |
Chris Forbes | cc5697f | 2019-01-30 11:54:08 -0800 | [diff] [blame] | 1144 | if (constants[0]) { |
| 1145 | if (min_const) { |
| 1146 | if (constants[0]->GetValueAsDouble() <= |
| 1147 | min_const->GetValueAsDouble()) { |
| 1148 | found_result = true; |
Nicolas Capens | 84c9c45 | 2022-11-18 14:11:05 +0000 | [diff] [blame] | 1149 | result = (cmp_opcode == spv::Op::OpFOrdLessThanEqual || |
| 1150 | cmp_opcode == spv::Op::OpFUnordLessThanEqual); |
Chris Forbes | cc5697f | 2019-01-30 11:54:08 -0800 | [diff] [blame] | 1151 | } |
| 1152 | } |
| 1153 | if (max_const) { |
| 1154 | if (constants[0]->GetValueAsDouble() > |
| 1155 | max_const->GetValueAsDouble()) { |
| 1156 | found_result = true; |
Nicolas Capens | 84c9c45 | 2022-11-18 14:11:05 +0000 | [diff] [blame] | 1157 | result = !(cmp_opcode == spv::Op::OpFOrdLessThanEqual || |
| 1158 | cmp_opcode == spv::Op::OpFUnordLessThanEqual); |
Chris Forbes | cc5697f | 2019-01-30 11:54:08 -0800 | [diff] [blame] | 1159 | } |
| 1160 | } |
| 1161 | } |
| 1162 | |
| 1163 | if (constants[1]) { |
| 1164 | if (max_const) { |
| 1165 | if (max_const->GetValueAsDouble() <= |
| 1166 | constants[1]->GetValueAsDouble()) { |
| 1167 | found_result = true; |
Nicolas Capens | 84c9c45 | 2022-11-18 14:11:05 +0000 | [diff] [blame] | 1168 | result = (cmp_opcode == spv::Op::OpFOrdLessThanEqual || |
| 1169 | cmp_opcode == spv::Op::OpFUnordLessThanEqual); |
Chris Forbes | cc5697f | 2019-01-30 11:54:08 -0800 | [diff] [blame] | 1170 | } |
| 1171 | } |
| 1172 | |
| 1173 | if (min_const) { |
| 1174 | if (min_const->GetValueAsDouble() > |
| 1175 | constants[1]->GetValueAsDouble()) { |
| 1176 | found_result = true; |
Nicolas Capens | 84c9c45 | 2022-11-18 14:11:05 +0000 | [diff] [blame] | 1177 | result = !(cmp_opcode == spv::Op::OpFOrdLessThanEqual || |
| 1178 | cmp_opcode == spv::Op::OpFUnordLessThanEqual); |
Chris Forbes | cc5697f | 2019-01-30 11:54:08 -0800 | [diff] [blame] | 1179 | } |
| 1180 | } |
| 1181 | } |
| 1182 | break; |
| 1183 | default: |
| 1184 | return nullptr; |
| 1185 | } |
| 1186 | |
| 1187 | if (!found_result) { |
| 1188 | return nullptr; |
| 1189 | } |
| 1190 | |
| 1191 | const analysis::Type* bool_type = |
| 1192 | context->get_type_mgr()->GetType(inst->type_id()); |
| 1193 | const analysis::Constant* result_const = |
| 1194 | const_mgr->GetConstant(bool_type, {static_cast<uint32_t>(result)}); |
| 1195 | assert(result_const); |
| 1196 | return result_const; |
| 1197 | }; |
| 1198 | } |
| 1199 | |
Ben Clayton | d0f684e | 2019-08-30 22:36:08 +0100 | [diff] [blame] | 1200 | ConstantFoldingRule FoldFMix() { |
| 1201 | return [](IRContext* context, Instruction* inst, |
| 1202 | const std::vector<const analysis::Constant*>& constants) |
| 1203 | -> const analysis::Constant* { |
| 1204 | analysis::ConstantManager* const_mgr = context->get_constant_mgr(); |
Nicolas Capens | 84c9c45 | 2022-11-18 14:11:05 +0000 | [diff] [blame] | 1205 | assert(inst->opcode() == spv::Op::OpExtInst && |
Ben Clayton | d0f684e | 2019-08-30 22:36:08 +0100 | [diff] [blame] | 1206 | "Expecting an extended instruction."); |
| 1207 | assert(inst->GetSingleWordInOperand(0) == |
| 1208 | context->get_feature_mgr()->GetExtInstImportId_GLSLstd450() && |
| 1209 | "Expecting a GLSLstd450 extended instruction."); |
| 1210 | assert(inst->GetSingleWordInOperand(1) == GLSLstd450FMix && |
| 1211 | "Expecting and FMix instruction."); |
| 1212 | |
| 1213 | if (!inst->IsFloatingPointFoldingAllowed()) { |
| 1214 | return nullptr; |
| 1215 | } |
| 1216 | |
| 1217 | // Make sure all FMix operands are constants. |
| 1218 | for (uint32_t i = 1; i < 4; i++) { |
| 1219 | if (constants[i] == nullptr) { |
| 1220 | return nullptr; |
| 1221 | } |
| 1222 | } |
| 1223 | |
| 1224 | const analysis::Constant* one; |
Ben Clayton | d552f63 | 2019-11-18 11:18:41 +0000 | [diff] [blame] | 1225 | bool is_vector = false; |
| 1226 | const analysis::Type* result_type = constants[1]->type(); |
| 1227 | const analysis::Type* base_type = result_type; |
| 1228 | if (base_type->AsVector()) { |
| 1229 | is_vector = true; |
| 1230 | base_type = base_type->AsVector()->element_type(); |
| 1231 | } |
| 1232 | assert(base_type->AsFloat() != nullptr && |
| 1233 | "FMix is suppose to act on floats or vectors of floats."); |
| 1234 | |
| 1235 | if (base_type->AsFloat()->width() == 32) { |
| 1236 | one = const_mgr->GetConstant(base_type, |
Ben Clayton | d0f684e | 2019-08-30 22:36:08 +0100 | [diff] [blame] | 1237 | utils::FloatProxy<float>(1.0f).GetWords()); |
| 1238 | } else { |
Ben Clayton | d552f63 | 2019-11-18 11:18:41 +0000 | [diff] [blame] | 1239 | one = const_mgr->GetConstant(base_type, |
Ben Clayton | d0f684e | 2019-08-30 22:36:08 +0100 | [diff] [blame] | 1240 | utils::FloatProxy<double>(1.0).GetWords()); |
| 1241 | } |
| 1242 | |
Ben Clayton | d552f63 | 2019-11-18 11:18:41 +0000 | [diff] [blame] | 1243 | if (is_vector) { |
| 1244 | uint32_t one_id = const_mgr->GetDefiningInstruction(one)->result_id(); |
| 1245 | one = |
| 1246 | const_mgr->GetConstant(result_type, std::vector<uint32_t>(4, one_id)); |
| 1247 | } |
| 1248 | |
| 1249 | const analysis::Constant* temp1 = FoldFPBinaryOp( |
| 1250 | FOLD_FPARITH_OP(-), inst->type_id(), {one, constants[3]}, context); |
Ben Clayton | d0f684e | 2019-08-30 22:36:08 +0100 | [diff] [blame] | 1251 | if (temp1 == nullptr) { |
| 1252 | return nullptr; |
| 1253 | } |
| 1254 | |
Ben Clayton | d552f63 | 2019-11-18 11:18:41 +0000 | [diff] [blame] | 1255 | const analysis::Constant* temp2 = FoldFPBinaryOp( |
| 1256 | FOLD_FPARITH_OP(*), inst->type_id(), {constants[1], temp1}, context); |
Ben Clayton | d0f684e | 2019-08-30 22:36:08 +0100 | [diff] [blame] | 1257 | if (temp2 == nullptr) { |
| 1258 | return nullptr; |
| 1259 | } |
Ben Clayton | d552f63 | 2019-11-18 11:18:41 +0000 | [diff] [blame] | 1260 | const analysis::Constant* temp3 = |
| 1261 | FoldFPBinaryOp(FOLD_FPARITH_OP(*), inst->type_id(), |
| 1262 | {constants[2], constants[3]}, context); |
Ben Clayton | d0f684e | 2019-08-30 22:36:08 +0100 | [diff] [blame] | 1263 | if (temp3 == nullptr) { |
| 1264 | return nullptr; |
| 1265 | } |
Ben Clayton | d552f63 | 2019-11-18 11:18:41 +0000 | [diff] [blame] | 1266 | return FoldFPBinaryOp(FOLD_FPARITH_OP(+), inst->type_id(), {temp2, temp3}, |
| 1267 | context); |
Ben Clayton | d0f684e | 2019-08-30 22:36:08 +0100 | [diff] [blame] | 1268 | }; |
| 1269 | } |
| 1270 | |
Ben Clayton | d552f63 | 2019-11-18 11:18:41 +0000 | [diff] [blame] | 1271 | const analysis::Constant* FoldMin(const analysis::Type* result_type, |
| 1272 | const analysis::Constant* a, |
| 1273 | const analysis::Constant* b, |
| 1274 | analysis::ConstantManager*) { |
| 1275 | if (const analysis::Integer* int_type = result_type->AsInteger()) { |
| 1276 | if (int_type->width() == 32) { |
| 1277 | if (int_type->IsSigned()) { |
| 1278 | int32_t va = a->GetS32(); |
| 1279 | int32_t vb = b->GetS32(); |
| 1280 | return (va < vb ? a : b); |
| 1281 | } else { |
| 1282 | uint32_t va = a->GetU32(); |
| 1283 | uint32_t vb = b->GetU32(); |
| 1284 | return (va < vb ? a : b); |
| 1285 | } |
| 1286 | } else if (int_type->width() == 64) { |
| 1287 | if (int_type->IsSigned()) { |
| 1288 | int64_t va = a->GetS64(); |
| 1289 | int64_t vb = b->GetS64(); |
| 1290 | return (va < vb ? a : b); |
| 1291 | } else { |
| 1292 | uint64_t va = a->GetU64(); |
| 1293 | uint64_t vb = b->GetU64(); |
| 1294 | return (va < vb ? a : b); |
| 1295 | } |
| 1296 | } |
| 1297 | } else if (const analysis::Float* float_type = result_type->AsFloat()) { |
| 1298 | if (float_type->width() == 32) { |
| 1299 | float va = a->GetFloat(); |
| 1300 | float vb = b->GetFloat(); |
| 1301 | return (va < vb ? a : b); |
| 1302 | } else if (float_type->width() == 64) { |
| 1303 | double va = a->GetDouble(); |
| 1304 | double vb = b->GetDouble(); |
| 1305 | return (va < vb ? a : b); |
| 1306 | } |
| 1307 | } |
| 1308 | return nullptr; |
| 1309 | } |
| 1310 | |
| 1311 | const analysis::Constant* FoldMax(const analysis::Type* result_type, |
| 1312 | const analysis::Constant* a, |
| 1313 | const analysis::Constant* b, |
| 1314 | analysis::ConstantManager*) { |
| 1315 | if (const analysis::Integer* int_type = result_type->AsInteger()) { |
| 1316 | if (int_type->width() == 32) { |
| 1317 | if (int_type->IsSigned()) { |
| 1318 | int32_t va = a->GetS32(); |
| 1319 | int32_t vb = b->GetS32(); |
| 1320 | return (va > vb ? a : b); |
| 1321 | } else { |
| 1322 | uint32_t va = a->GetU32(); |
| 1323 | uint32_t vb = b->GetU32(); |
| 1324 | return (va > vb ? a : b); |
| 1325 | } |
| 1326 | } else if (int_type->width() == 64) { |
| 1327 | if (int_type->IsSigned()) { |
| 1328 | int64_t va = a->GetS64(); |
| 1329 | int64_t vb = b->GetS64(); |
| 1330 | return (va > vb ? a : b); |
| 1331 | } else { |
| 1332 | uint64_t va = a->GetU64(); |
| 1333 | uint64_t vb = b->GetU64(); |
| 1334 | return (va > vb ? a : b); |
| 1335 | } |
| 1336 | } |
| 1337 | } else if (const analysis::Float* float_type = result_type->AsFloat()) { |
| 1338 | if (float_type->width() == 32) { |
| 1339 | float va = a->GetFloat(); |
| 1340 | float vb = b->GetFloat(); |
| 1341 | return (va > vb ? a : b); |
| 1342 | } else if (float_type->width() == 64) { |
| 1343 | double va = a->GetDouble(); |
| 1344 | double vb = b->GetDouble(); |
| 1345 | return (va > vb ? a : b); |
| 1346 | } |
| 1347 | } |
| 1348 | return nullptr; |
| 1349 | } |
| 1350 | |
| 1351 | // Fold an clamp instruction when all three operands are constant. |
| 1352 | const analysis::Constant* FoldClamp1( |
| 1353 | IRContext* context, Instruction* inst, |
| 1354 | const std::vector<const analysis::Constant*>& constants) { |
Nicolas Capens | 84c9c45 | 2022-11-18 14:11:05 +0000 | [diff] [blame] | 1355 | assert(inst->opcode() == spv::Op::OpExtInst && |
Ben Clayton | d552f63 | 2019-11-18 11:18:41 +0000 | [diff] [blame] | 1356 | "Expecting an extended instruction."); |
| 1357 | assert(inst->GetSingleWordInOperand(0) == |
| 1358 | context->get_feature_mgr()->GetExtInstImportId_GLSLstd450() && |
| 1359 | "Expecting a GLSLstd450 extended instruction."); |
| 1360 | |
| 1361 | // Make sure all Clamp operands are constants. |
Alexis Hetu | 00e0af1 | 2021-11-08 08:57:46 -0500 | [diff] [blame] | 1362 | for (uint32_t i = 1; i < 4; i++) { |
Ben Clayton | d552f63 | 2019-11-18 11:18:41 +0000 | [diff] [blame] | 1363 | if (constants[i] == nullptr) { |
| 1364 | return nullptr; |
| 1365 | } |
| 1366 | } |
| 1367 | |
| 1368 | const analysis::Constant* temp = FoldFPBinaryOp( |
| 1369 | FoldMax, inst->type_id(), {constants[1], constants[2]}, context); |
| 1370 | if (temp == nullptr) { |
| 1371 | return nullptr; |
| 1372 | } |
| 1373 | return FoldFPBinaryOp(FoldMin, inst->type_id(), {temp, constants[3]}, |
| 1374 | context); |
| 1375 | } |
| 1376 | |
Alexis Hetu | 00e0af1 | 2021-11-08 08:57:46 -0500 | [diff] [blame] | 1377 | // Fold a clamp instruction when |x <= min_val|. |
Ben Clayton | d552f63 | 2019-11-18 11:18:41 +0000 | [diff] [blame] | 1378 | const analysis::Constant* FoldClamp2( |
| 1379 | IRContext* context, Instruction* inst, |
| 1380 | const std::vector<const analysis::Constant*>& constants) { |
Nicolas Capens | 84c9c45 | 2022-11-18 14:11:05 +0000 | [diff] [blame] | 1381 | assert(inst->opcode() == spv::Op::OpExtInst && |
Ben Clayton | d552f63 | 2019-11-18 11:18:41 +0000 | [diff] [blame] | 1382 | "Expecting an extended instruction."); |
| 1383 | assert(inst->GetSingleWordInOperand(0) == |
| 1384 | context->get_feature_mgr()->GetExtInstImportId_GLSLstd450() && |
| 1385 | "Expecting a GLSLstd450 extended instruction."); |
| 1386 | |
| 1387 | const analysis::Constant* x = constants[1]; |
| 1388 | const analysis::Constant* min_val = constants[2]; |
| 1389 | |
| 1390 | if (x == nullptr || min_val == nullptr) { |
| 1391 | return nullptr; |
| 1392 | } |
| 1393 | |
| 1394 | const analysis::Constant* temp = |
| 1395 | FoldFPBinaryOp(FoldMax, inst->type_id(), {x, min_val}, context); |
| 1396 | if (temp == min_val) { |
| 1397 | // We can assume that |min_val| is less than |max_val|. Therefore, if the |
| 1398 | // result of the max operation is |min_val|, we know the result of the min |
| 1399 | // operation, even if |max_val| is not a constant. |
| 1400 | return min_val; |
| 1401 | } |
| 1402 | return nullptr; |
| 1403 | } |
| 1404 | |
| 1405 | // Fold a clamp instruction when |x >= max_val|. |
| 1406 | const analysis::Constant* FoldClamp3( |
| 1407 | IRContext* context, Instruction* inst, |
| 1408 | const std::vector<const analysis::Constant*>& constants) { |
Nicolas Capens | 84c9c45 | 2022-11-18 14:11:05 +0000 | [diff] [blame] | 1409 | assert(inst->opcode() == spv::Op::OpExtInst && |
Ben Clayton | d552f63 | 2019-11-18 11:18:41 +0000 | [diff] [blame] | 1410 | "Expecting an extended instruction."); |
| 1411 | assert(inst->GetSingleWordInOperand(0) == |
| 1412 | context->get_feature_mgr()->GetExtInstImportId_GLSLstd450() && |
| 1413 | "Expecting a GLSLstd450 extended instruction."); |
| 1414 | |
| 1415 | const analysis::Constant* x = constants[1]; |
| 1416 | const analysis::Constant* max_val = constants[3]; |
| 1417 | |
| 1418 | if (x == nullptr || max_val == nullptr) { |
| 1419 | return nullptr; |
| 1420 | } |
| 1421 | |
| 1422 | const analysis::Constant* temp = |
| 1423 | FoldFPBinaryOp(FoldMin, inst->type_id(), {x, max_val}, context); |
| 1424 | if (temp == max_val) { |
| 1425 | // We can assume that |min_val| is less than |max_val|. Therefore, if the |
| 1426 | // result of the max operation is |min_val|, we know the result of the min |
| 1427 | // operation, even if |max_val| is not a constant. |
| 1428 | return max_val; |
| 1429 | } |
| 1430 | return nullptr; |
| 1431 | } |
| 1432 | |
Ben Clayton | dc6b76a | 2020-02-24 14:53:40 +0000 | [diff] [blame] | 1433 | UnaryScalarFoldingRule FoldFTranscendentalUnary(double (*fp)(double)) { |
| 1434 | return |
| 1435 | [fp](const analysis::Type* result_type, const analysis::Constant* a, |
| 1436 | analysis::ConstantManager* const_mgr) -> const analysis::Constant* { |
| 1437 | assert(result_type != nullptr && a != nullptr); |
| 1438 | const analysis::Float* float_type = a->type()->AsFloat(); |
| 1439 | assert(float_type != nullptr); |
| 1440 | assert(float_type == result_type->AsFloat()); |
| 1441 | if (float_type->width() == 32) { |
| 1442 | float fa = a->GetFloat(); |
| 1443 | float res = static_cast<float>(fp(fa)); |
| 1444 | utils::FloatProxy<float> result(res); |
| 1445 | std::vector<uint32_t> words = result.GetWords(); |
| 1446 | return const_mgr->GetConstant(result_type, words); |
| 1447 | } else if (float_type->width() == 64) { |
| 1448 | double fa = a->GetDouble(); |
| 1449 | double res = fp(fa); |
| 1450 | utils::FloatProxy<double> result(res); |
| 1451 | std::vector<uint32_t> words = result.GetWords(); |
| 1452 | return const_mgr->GetConstant(result_type, words); |
| 1453 | } |
| 1454 | return nullptr; |
| 1455 | }; |
| 1456 | } |
| 1457 | |
| 1458 | BinaryScalarFoldingRule FoldFTranscendentalBinary(double (*fp)(double, |
| 1459 | double)) { |
| 1460 | return |
| 1461 | [fp](const analysis::Type* result_type, const analysis::Constant* a, |
| 1462 | const analysis::Constant* b, |
| 1463 | analysis::ConstantManager* const_mgr) -> const analysis::Constant* { |
| 1464 | assert(result_type != nullptr && a != nullptr); |
| 1465 | const analysis::Float* float_type = a->type()->AsFloat(); |
| 1466 | assert(float_type != nullptr); |
| 1467 | assert(float_type == result_type->AsFloat()); |
| 1468 | assert(float_type == b->type()->AsFloat()); |
| 1469 | if (float_type->width() == 32) { |
| 1470 | float fa = a->GetFloat(); |
| 1471 | float fb = b->GetFloat(); |
| 1472 | float res = static_cast<float>(fp(fa, fb)); |
| 1473 | utils::FloatProxy<float> result(res); |
| 1474 | std::vector<uint32_t> words = result.GetWords(); |
| 1475 | return const_mgr->GetConstant(result_type, words); |
| 1476 | } else if (float_type->width() == 64) { |
| 1477 | double fa = a->GetDouble(); |
| 1478 | double fb = b->GetDouble(); |
| 1479 | double res = fp(fa, fb); |
| 1480 | utils::FloatProxy<double> result(res); |
| 1481 | std::vector<uint32_t> words = result.GetWords(); |
| 1482 | return const_mgr->GetConstant(result_type, words); |
| 1483 | } |
| 1484 | return nullptr; |
| 1485 | }; |
| 1486 | } |
Chris Forbes | cc5697f | 2019-01-30 11:54:08 -0800 | [diff] [blame] | 1487 | } // namespace |
| 1488 | |
Ben Clayton | d0f684e | 2019-08-30 22:36:08 +0100 | [diff] [blame] | 1489 | void ConstantFoldingRules::AddFoldingRules() { |
Chris Forbes | cc5697f | 2019-01-30 11:54:08 -0800 | [diff] [blame] | 1490 | // Add all folding rules to the list for the opcodes to which they apply. |
| 1491 | // Note that the order in which rules are added to the list matters. If a rule |
| 1492 | // applies to the instruction, the rest of the rules will not be attempted. |
| 1493 | // Take that into consideration. |
| 1494 | |
Nicolas Capens | 84c9c45 | 2022-11-18 14:11:05 +0000 | [diff] [blame] | 1495 | rules_[spv::Op::OpCompositeConstruct].push_back(FoldCompositeWithConstants()); |
Chris Forbes | cc5697f | 2019-01-30 11:54:08 -0800 | [diff] [blame] | 1496 | |
Nicolas Capens | 84c9c45 | 2022-11-18 14:11:05 +0000 | [diff] [blame] | 1497 | rules_[spv::Op::OpCompositeExtract].push_back(FoldExtractWithConstants()); |
| 1498 | rules_[spv::Op::OpCompositeInsert].push_back(FoldInsertWithConstants()); |
Chris Forbes | cc5697f | 2019-01-30 11:54:08 -0800 | [diff] [blame] | 1499 | |
Nicolas Capens | 84c9c45 | 2022-11-18 14:11:05 +0000 | [diff] [blame] | 1500 | rules_[spv::Op::OpConvertFToS].push_back(FoldFToI()); |
| 1501 | rules_[spv::Op::OpConvertFToU].push_back(FoldFToI()); |
| 1502 | rules_[spv::Op::OpConvertSToF].push_back(FoldIToF()); |
| 1503 | rules_[spv::Op::OpConvertUToF].push_back(FoldIToF()); |
Chris Forbes | cc5697f | 2019-01-30 11:54:08 -0800 | [diff] [blame] | 1504 | |
Nicolas Capens | 84c9c45 | 2022-11-18 14:11:05 +0000 | [diff] [blame] | 1505 | rules_[spv::Op::OpDot].push_back(FoldOpDotWithConstants()); |
| 1506 | rules_[spv::Op::OpFAdd].push_back(FoldFAdd()); |
| 1507 | rules_[spv::Op::OpFDiv].push_back(FoldFDiv()); |
| 1508 | rules_[spv::Op::OpFMul].push_back(FoldFMul()); |
| 1509 | rules_[spv::Op::OpFSub].push_back(FoldFSub()); |
Chris Forbes | cc5697f | 2019-01-30 11:54:08 -0800 | [diff] [blame] | 1510 | |
Nicolas Capens | 84c9c45 | 2022-11-18 14:11:05 +0000 | [diff] [blame] | 1511 | rules_[spv::Op::OpFOrdEqual].push_back(FoldFOrdEqual()); |
Chris Forbes | cc5697f | 2019-01-30 11:54:08 -0800 | [diff] [blame] | 1512 | |
Nicolas Capens | 84c9c45 | 2022-11-18 14:11:05 +0000 | [diff] [blame] | 1513 | rules_[spv::Op::OpFUnordEqual].push_back(FoldFUnordEqual()); |
Chris Forbes | cc5697f | 2019-01-30 11:54:08 -0800 | [diff] [blame] | 1514 | |
Nicolas Capens | 84c9c45 | 2022-11-18 14:11:05 +0000 | [diff] [blame] | 1515 | rules_[spv::Op::OpFOrdNotEqual].push_back(FoldFOrdNotEqual()); |
Chris Forbes | cc5697f | 2019-01-30 11:54:08 -0800 | [diff] [blame] | 1516 | |
Nicolas Capens | 84c9c45 | 2022-11-18 14:11:05 +0000 | [diff] [blame] | 1517 | rules_[spv::Op::OpFUnordNotEqual].push_back(FoldFUnordNotEqual()); |
Chris Forbes | cc5697f | 2019-01-30 11:54:08 -0800 | [diff] [blame] | 1518 | |
Nicolas Capens | 84c9c45 | 2022-11-18 14:11:05 +0000 | [diff] [blame] | 1519 | rules_[spv::Op::OpFOrdLessThan].push_back(FoldFOrdLessThan()); |
| 1520 | rules_[spv::Op::OpFOrdLessThan].push_back( |
| 1521 | FoldFClampFeedingCompare(spv::Op::OpFOrdLessThan)); |
Chris Forbes | cc5697f | 2019-01-30 11:54:08 -0800 | [diff] [blame] | 1522 | |
Nicolas Capens | 84c9c45 | 2022-11-18 14:11:05 +0000 | [diff] [blame] | 1523 | rules_[spv::Op::OpFUnordLessThan].push_back(FoldFUnordLessThan()); |
| 1524 | rules_[spv::Op::OpFUnordLessThan].push_back( |
| 1525 | FoldFClampFeedingCompare(spv::Op::OpFUnordLessThan)); |
Chris Forbes | cc5697f | 2019-01-30 11:54:08 -0800 | [diff] [blame] | 1526 | |
Nicolas Capens | 84c9c45 | 2022-11-18 14:11:05 +0000 | [diff] [blame] | 1527 | rules_[spv::Op::OpFOrdGreaterThan].push_back(FoldFOrdGreaterThan()); |
| 1528 | rules_[spv::Op::OpFOrdGreaterThan].push_back( |
| 1529 | FoldFClampFeedingCompare(spv::Op::OpFOrdGreaterThan)); |
Chris Forbes | cc5697f | 2019-01-30 11:54:08 -0800 | [diff] [blame] | 1530 | |
Nicolas Capens | 84c9c45 | 2022-11-18 14:11:05 +0000 | [diff] [blame] | 1531 | rules_[spv::Op::OpFUnordGreaterThan].push_back(FoldFUnordGreaterThan()); |
| 1532 | rules_[spv::Op::OpFUnordGreaterThan].push_back( |
| 1533 | FoldFClampFeedingCompare(spv::Op::OpFUnordGreaterThan)); |
Chris Forbes | cc5697f | 2019-01-30 11:54:08 -0800 | [diff] [blame] | 1534 | |
Nicolas Capens | 84c9c45 | 2022-11-18 14:11:05 +0000 | [diff] [blame] | 1535 | rules_[spv::Op::OpFOrdLessThanEqual].push_back(FoldFOrdLessThanEqual()); |
| 1536 | rules_[spv::Op::OpFOrdLessThanEqual].push_back( |
| 1537 | FoldFClampFeedingCompare(spv::Op::OpFOrdLessThanEqual)); |
Chris Forbes | cc5697f | 2019-01-30 11:54:08 -0800 | [diff] [blame] | 1538 | |
Nicolas Capens | 84c9c45 | 2022-11-18 14:11:05 +0000 | [diff] [blame] | 1539 | rules_[spv::Op::OpFUnordLessThanEqual].push_back(FoldFUnordLessThanEqual()); |
| 1540 | rules_[spv::Op::OpFUnordLessThanEqual].push_back( |
| 1541 | FoldFClampFeedingCompare(spv::Op::OpFUnordLessThanEqual)); |
Chris Forbes | cc5697f | 2019-01-30 11:54:08 -0800 | [diff] [blame] | 1542 | |
Nicolas Capens | 84c9c45 | 2022-11-18 14:11:05 +0000 | [diff] [blame] | 1543 | rules_[spv::Op::OpFOrdGreaterThanEqual].push_back(FoldFOrdGreaterThanEqual()); |
| 1544 | rules_[spv::Op::OpFOrdGreaterThanEqual].push_back( |
| 1545 | FoldFClampFeedingCompare(spv::Op::OpFOrdGreaterThanEqual)); |
Chris Forbes | cc5697f | 2019-01-30 11:54:08 -0800 | [diff] [blame] | 1546 | |
Nicolas Capens | 84c9c45 | 2022-11-18 14:11:05 +0000 | [diff] [blame] | 1547 | rules_[spv::Op::OpFUnordGreaterThanEqual].push_back( |
| 1548 | FoldFUnordGreaterThanEqual()); |
| 1549 | rules_[spv::Op::OpFUnordGreaterThanEqual].push_back( |
| 1550 | FoldFClampFeedingCompare(spv::Op::OpFUnordGreaterThanEqual)); |
Chris Forbes | cc5697f | 2019-01-30 11:54:08 -0800 | [diff] [blame] | 1551 | |
Nicolas Capens | 84c9c45 | 2022-11-18 14:11:05 +0000 | [diff] [blame] | 1552 | rules_[spv::Op::OpVectorShuffle].push_back(FoldVectorShuffleWithConstants()); |
| 1553 | rules_[spv::Op::OpVectorTimesScalar].push_back(FoldVectorTimesScalar()); |
| 1554 | rules_[spv::Op::OpVectorTimesMatrix].push_back(FoldVectorTimesMatrix()); |
| 1555 | rules_[spv::Op::OpMatrixTimesVector].push_back(FoldMatrixTimesVector()); |
Chris Forbes | cc5697f | 2019-01-30 11:54:08 -0800 | [diff] [blame] | 1556 | |
Nicolas Capens | 84c9c45 | 2022-11-18 14:11:05 +0000 | [diff] [blame] | 1557 | rules_[spv::Op::OpFNegate].push_back(FoldFNegate()); |
| 1558 | rules_[spv::Op::OpQuantizeToF16].push_back(FoldQuantizeToF16()); |
Ben Clayton | d0f684e | 2019-08-30 22:36:08 +0100 | [diff] [blame] | 1559 | |
| 1560 | // Add rules for GLSLstd450 |
| 1561 | FeatureManager* feature_manager = context_->get_feature_mgr(); |
| 1562 | uint32_t ext_inst_glslstd450_id = |
| 1563 | feature_manager->GetExtInstImportId_GLSLstd450(); |
| 1564 | if (ext_inst_glslstd450_id != 0) { |
| 1565 | ext_rules_[{ext_inst_glslstd450_id, GLSLstd450FMix}].push_back(FoldFMix()); |
Ben Clayton | d552f63 | 2019-11-18 11:18:41 +0000 | [diff] [blame] | 1566 | ext_rules_[{ext_inst_glslstd450_id, GLSLstd450SMin}].push_back( |
| 1567 | FoldFPBinaryOp(FoldMin)); |
| 1568 | ext_rules_[{ext_inst_glslstd450_id, GLSLstd450UMin}].push_back( |
| 1569 | FoldFPBinaryOp(FoldMin)); |
| 1570 | ext_rules_[{ext_inst_glslstd450_id, GLSLstd450FMin}].push_back( |
| 1571 | FoldFPBinaryOp(FoldMin)); |
| 1572 | ext_rules_[{ext_inst_glslstd450_id, GLSLstd450SMax}].push_back( |
| 1573 | FoldFPBinaryOp(FoldMax)); |
| 1574 | ext_rules_[{ext_inst_glslstd450_id, GLSLstd450UMax}].push_back( |
| 1575 | FoldFPBinaryOp(FoldMax)); |
| 1576 | ext_rules_[{ext_inst_glslstd450_id, GLSLstd450FMax}].push_back( |
| 1577 | FoldFPBinaryOp(FoldMax)); |
| 1578 | ext_rules_[{ext_inst_glslstd450_id, GLSLstd450UClamp}].push_back( |
| 1579 | FoldClamp1); |
| 1580 | ext_rules_[{ext_inst_glslstd450_id, GLSLstd450UClamp}].push_back( |
| 1581 | FoldClamp2); |
| 1582 | ext_rules_[{ext_inst_glslstd450_id, GLSLstd450UClamp}].push_back( |
| 1583 | FoldClamp3); |
| 1584 | ext_rules_[{ext_inst_glslstd450_id, GLSLstd450SClamp}].push_back( |
| 1585 | FoldClamp1); |
| 1586 | ext_rules_[{ext_inst_glslstd450_id, GLSLstd450SClamp}].push_back( |
| 1587 | FoldClamp2); |
| 1588 | ext_rules_[{ext_inst_glslstd450_id, GLSLstd450SClamp}].push_back( |
| 1589 | FoldClamp3); |
| 1590 | ext_rules_[{ext_inst_glslstd450_id, GLSLstd450FClamp}].push_back( |
| 1591 | FoldClamp1); |
| 1592 | ext_rules_[{ext_inst_glslstd450_id, GLSLstd450FClamp}].push_back( |
| 1593 | FoldClamp2); |
| 1594 | ext_rules_[{ext_inst_glslstd450_id, GLSLstd450FClamp}].push_back( |
| 1595 | FoldClamp3); |
Ben Clayton | dc6b76a | 2020-02-24 14:53:40 +0000 | [diff] [blame] | 1596 | ext_rules_[{ext_inst_glslstd450_id, GLSLstd450Sin}].push_back( |
| 1597 | FoldFPUnaryOp(FoldFTranscendentalUnary(std::sin))); |
| 1598 | ext_rules_[{ext_inst_glslstd450_id, GLSLstd450Cos}].push_back( |
| 1599 | FoldFPUnaryOp(FoldFTranscendentalUnary(std::cos))); |
| 1600 | ext_rules_[{ext_inst_glslstd450_id, GLSLstd450Tan}].push_back( |
| 1601 | FoldFPUnaryOp(FoldFTranscendentalUnary(std::tan))); |
| 1602 | ext_rules_[{ext_inst_glslstd450_id, GLSLstd450Asin}].push_back( |
| 1603 | FoldFPUnaryOp(FoldFTranscendentalUnary(std::asin))); |
| 1604 | ext_rules_[{ext_inst_glslstd450_id, GLSLstd450Acos}].push_back( |
| 1605 | FoldFPUnaryOp(FoldFTranscendentalUnary(std::acos))); |
| 1606 | ext_rules_[{ext_inst_glslstd450_id, GLSLstd450Atan}].push_back( |
| 1607 | FoldFPUnaryOp(FoldFTranscendentalUnary(std::atan))); |
| 1608 | ext_rules_[{ext_inst_glslstd450_id, GLSLstd450Exp}].push_back( |
| 1609 | FoldFPUnaryOp(FoldFTranscendentalUnary(std::exp))); |
| 1610 | ext_rules_[{ext_inst_glslstd450_id, GLSLstd450Log}].push_back( |
| 1611 | FoldFPUnaryOp(FoldFTranscendentalUnary(std::log))); |
| 1612 | |
| 1613 | #ifdef __ANDROID__ |
sugoi1 | b398bf3 | 2022-02-18 10:27:28 -0500 | [diff] [blame] | 1614 | // Android NDK r15c targeting ABI 15 doesn't have full support for C++11 |
Ben Clayton | dc6b76a | 2020-02-24 14:53:40 +0000 | [diff] [blame] | 1615 | // (no std::exp2/log2). ::exp2 is available from C99 but ::log2 isn't |
| 1616 | // available up until ABI 18 so we use a shim |
| 1617 | auto log2_shim = [](double v) -> double { return log(v) / log(2.0); }; |
| 1618 | ext_rules_[{ext_inst_glslstd450_id, GLSLstd450Exp2}].push_back( |
| 1619 | FoldFPUnaryOp(FoldFTranscendentalUnary(::exp2))); |
| 1620 | ext_rules_[{ext_inst_glslstd450_id, GLSLstd450Log2}].push_back( |
| 1621 | FoldFPUnaryOp(FoldFTranscendentalUnary(log2_shim))); |
| 1622 | #else |
| 1623 | ext_rules_[{ext_inst_glslstd450_id, GLSLstd450Exp2}].push_back( |
| 1624 | FoldFPUnaryOp(FoldFTranscendentalUnary(std::exp2))); |
| 1625 | ext_rules_[{ext_inst_glslstd450_id, GLSLstd450Log2}].push_back( |
| 1626 | FoldFPUnaryOp(FoldFTranscendentalUnary(std::log2))); |
| 1627 | #endif |
| 1628 | |
| 1629 | ext_rules_[{ext_inst_glslstd450_id, GLSLstd450Sqrt}].push_back( |
| 1630 | FoldFPUnaryOp(FoldFTranscendentalUnary(std::sqrt))); |
| 1631 | ext_rules_[{ext_inst_glslstd450_id, GLSLstd450Atan2}].push_back( |
| 1632 | FoldFPBinaryOp(FoldFTranscendentalBinary(std::atan2))); |
| 1633 | ext_rules_[{ext_inst_glslstd450_id, GLSLstd450Pow}].push_back( |
| 1634 | FoldFPBinaryOp(FoldFTranscendentalBinary(std::pow))); |
Ben Clayton | d0f684e | 2019-08-30 22:36:08 +0100 | [diff] [blame] | 1635 | } |
Chris Forbes | cc5697f | 2019-01-30 11:54:08 -0800 | [diff] [blame] | 1636 | } |
| 1637 | } // namespace opt |
| 1638 | } // namespace spvtools |