/src/spirv-tools/source/opt/upgrade_memory_model.cpp
Line | Count | Source (jump to first uncovered line) |
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 "upgrade_memory_model.h" |
16 | | |
17 | | #include <utility> |
18 | | |
19 | | #include "source/opt/ir_builder.h" |
20 | | #include "source/opt/ir_context.h" |
21 | | #include "source/spirv_constant.h" |
22 | | #include "source/util/make_unique.h" |
23 | | #include "source/util/string_utils.h" |
24 | | |
25 | | namespace spvtools { |
26 | | namespace opt { |
27 | | |
28 | 0 | Pass::Status UpgradeMemoryModel::Process() { |
29 | | // TODO: This pass needs changes to support cooperative matrices. |
30 | 0 | if (context()->get_feature_mgr()->HasCapability( |
31 | 0 | spv::Capability::CooperativeMatrixNV)) { |
32 | 0 | return Pass::Status::SuccessWithoutChange; |
33 | 0 | } |
34 | | |
35 | | // Only update Logical GLSL450 to Logical VulkanKHR. |
36 | 0 | Instruction* memory_model = get_module()->GetMemoryModel(); |
37 | 0 | if (memory_model->GetSingleWordInOperand(0u) != |
38 | 0 | uint32_t(spv::AddressingModel::Logical) || |
39 | 0 | memory_model->GetSingleWordInOperand(1u) != |
40 | 0 | uint32_t(spv::MemoryModel::GLSL450)) { |
41 | 0 | return Pass::Status::SuccessWithoutChange; |
42 | 0 | } |
43 | | |
44 | 0 | UpgradeMemoryModelInstruction(); |
45 | 0 | UpgradeInstructions(); |
46 | 0 | CleanupDecorations(); |
47 | 0 | UpgradeBarriers(); |
48 | 0 | UpgradeMemoryScope(); |
49 | |
|
50 | 0 | return Pass::Status::SuccessWithChange; |
51 | 0 | } |
52 | | |
53 | 0 | void UpgradeMemoryModel::UpgradeMemoryModelInstruction() { |
54 | | // Overall changes necessary: |
55 | | // 1. Add the OpExtension. |
56 | | // 2. Add the OpCapability. |
57 | | // 3. Modify the memory model. |
58 | 0 | Instruction* memory_model = get_module()->GetMemoryModel(); |
59 | 0 | context()->AddCapability(MakeUnique<Instruction>( |
60 | 0 | context(), spv::Op::OpCapability, 0, 0, |
61 | 0 | std::initializer_list<Operand>{ |
62 | 0 | {SPV_OPERAND_TYPE_CAPABILITY, |
63 | 0 | {uint32_t(spv::Capability::VulkanMemoryModelKHR)}}})); |
64 | 0 | const std::string extension = "SPV_KHR_vulkan_memory_model"; |
65 | 0 | std::vector<uint32_t> words = spvtools::utils::MakeVector(extension); |
66 | 0 | context()->AddExtension( |
67 | 0 | MakeUnique<Instruction>(context(), spv::Op::OpExtension, 0, 0, |
68 | 0 | std::initializer_list<Operand>{ |
69 | 0 | {SPV_OPERAND_TYPE_LITERAL_STRING, words}})); |
70 | 0 | memory_model->SetInOperand(1u, {uint32_t(spv::MemoryModel::VulkanKHR)}); |
71 | 0 | } |
72 | | |
73 | 0 | void UpgradeMemoryModel::UpgradeInstructions() { |
74 | | // Coherent and Volatile decorations are deprecated. Remove them and replace |
75 | | // with flags on the memory/image operations. The decorations can occur on |
76 | | // OpVariable, OpFunctionParameter (of pointer type) and OpStructType (member |
77 | | // decoration). Trace from the decoration target(s) to the final memory/image |
78 | | // instructions. Additionally, Workgroup storage class variables and function |
79 | | // parameters are implicitly coherent in GLSL450. |
80 | | |
81 | | // Upgrade modf and frexp first since they generate new stores. |
82 | | // In SPIR-V 1.4 or later, normalize OpCopyMemory* access operands. |
83 | 0 | for (auto& func : *get_module()) { |
84 | 0 | func.ForEachInst([this](Instruction* inst) { |
85 | 0 | if (inst->opcode() == spv::Op::OpExtInst) { |
86 | 0 | auto ext_inst = inst->GetSingleWordInOperand(1u); |
87 | 0 | if (ext_inst == GLSLstd450Modf || ext_inst == GLSLstd450Frexp) { |
88 | 0 | auto import = |
89 | 0 | get_def_use_mgr()->GetDef(inst->GetSingleWordInOperand(0u)); |
90 | 0 | if (import->GetInOperand(0u).AsString() == "GLSL.std.450") { |
91 | 0 | UpgradeExtInst(inst); |
92 | 0 | } |
93 | 0 | } |
94 | 0 | } else if (get_module()->version() >= SPV_SPIRV_VERSION_WORD(1, 4)) { |
95 | 0 | if (inst->opcode() == spv::Op::OpCopyMemory || |
96 | 0 | inst->opcode() == spv::Op::OpCopyMemorySized) { |
97 | 0 | uint32_t start_operand = |
98 | 0 | inst->opcode() == spv::Op::OpCopyMemory ? 2u : 3u; |
99 | 0 | if (inst->NumInOperands() > start_operand) { |
100 | 0 | auto num_access_words = MemoryAccessNumWords( |
101 | 0 | inst->GetSingleWordInOperand(start_operand)); |
102 | 0 | if ((num_access_words + start_operand) == inst->NumInOperands()) { |
103 | | // There is a single memory access operand. Duplicate it to have a |
104 | | // separate operand for both source and target. |
105 | 0 | for (uint32_t i = 0; i < num_access_words; ++i) { |
106 | 0 | auto operand = inst->GetInOperand(start_operand + i); |
107 | 0 | inst->AddOperand(std::move(operand)); |
108 | 0 | } |
109 | 0 | } |
110 | 0 | } else { |
111 | | // Add two memory access operands. |
112 | 0 | inst->AddOperand({SPV_OPERAND_TYPE_MEMORY_ACCESS, |
113 | 0 | {uint32_t(spv::MemoryAccessMask::MaskNone)}}); |
114 | 0 | inst->AddOperand({SPV_OPERAND_TYPE_MEMORY_ACCESS, |
115 | 0 | {uint32_t(spv::MemoryAccessMask::MaskNone)}}); |
116 | 0 | } |
117 | 0 | } |
118 | 0 | } |
119 | 0 | }); |
120 | 0 | } |
121 | |
|
122 | 0 | UpgradeMemoryAndImages(); |
123 | 0 | UpgradeAtomics(); |
124 | 0 | } |
125 | | |
126 | 0 | void UpgradeMemoryModel::UpgradeMemoryAndImages() { |
127 | 0 | for (auto& func : *get_module()) { |
128 | 0 | func.ForEachInst([this](Instruction* inst) { |
129 | 0 | bool is_coherent = false; |
130 | 0 | bool is_volatile = false; |
131 | 0 | bool src_coherent = false; |
132 | 0 | bool src_volatile = false; |
133 | 0 | bool dst_coherent = false; |
134 | 0 | bool dst_volatile = false; |
135 | 0 | uint32_t start_operand = 0u; |
136 | 0 | spv::Scope scope = spv::Scope::QueueFamilyKHR; |
137 | 0 | spv::Scope src_scope = spv::Scope::QueueFamilyKHR; |
138 | 0 | spv::Scope dst_scope = spv::Scope::QueueFamilyKHR; |
139 | 0 | switch (inst->opcode()) { |
140 | 0 | case spv::Op::OpLoad: |
141 | 0 | case spv::Op::OpStore: |
142 | 0 | std::tie(is_coherent, is_volatile, scope) = |
143 | 0 | GetInstructionAttributes(inst->GetSingleWordInOperand(0u)); |
144 | 0 | break; |
145 | 0 | case spv::Op::OpImageRead: |
146 | 0 | case spv::Op::OpImageSparseRead: |
147 | 0 | case spv::Op::OpImageWrite: |
148 | 0 | std::tie(is_coherent, is_volatile, scope) = |
149 | 0 | GetInstructionAttributes(inst->GetSingleWordInOperand(0u)); |
150 | 0 | break; |
151 | 0 | case spv::Op::OpCopyMemory: |
152 | 0 | case spv::Op::OpCopyMemorySized: |
153 | 0 | std::tie(dst_coherent, dst_volatile, dst_scope) = |
154 | 0 | GetInstructionAttributes(inst->GetSingleWordInOperand(0u)); |
155 | 0 | std::tie(src_coherent, src_volatile, src_scope) = |
156 | 0 | GetInstructionAttributes(inst->GetSingleWordInOperand(1u)); |
157 | 0 | break; |
158 | 0 | default: |
159 | 0 | break; |
160 | 0 | } |
161 | | |
162 | 0 | switch (inst->opcode()) { |
163 | 0 | case spv::Op::OpLoad: |
164 | 0 | UpgradeFlags(inst, 1u, is_coherent, is_volatile, kVisibility, |
165 | 0 | kMemory); |
166 | 0 | break; |
167 | 0 | case spv::Op::OpStore: |
168 | 0 | UpgradeFlags(inst, 2u, is_coherent, is_volatile, kAvailability, |
169 | 0 | kMemory); |
170 | 0 | break; |
171 | 0 | case spv::Op::OpCopyMemory: |
172 | 0 | case spv::Op::OpCopyMemorySized: |
173 | 0 | start_operand = inst->opcode() == spv::Op::OpCopyMemory ? 2u : 3u; |
174 | 0 | if (get_module()->version() >= SPV_SPIRV_VERSION_WORD(1, 4)) { |
175 | | // There are guaranteed to be two memory access operands at this |
176 | | // point so treat source and target separately. |
177 | 0 | uint32_t num_access_words = MemoryAccessNumWords( |
178 | 0 | inst->GetSingleWordInOperand(start_operand)); |
179 | 0 | UpgradeFlags(inst, start_operand, dst_coherent, dst_volatile, |
180 | 0 | kAvailability, kMemory); |
181 | 0 | UpgradeFlags(inst, start_operand + num_access_words, src_coherent, |
182 | 0 | src_volatile, kVisibility, kMemory); |
183 | 0 | } else { |
184 | 0 | UpgradeFlags(inst, start_operand, dst_coherent, dst_volatile, |
185 | 0 | kAvailability, kMemory); |
186 | 0 | UpgradeFlags(inst, start_operand, src_coherent, src_volatile, |
187 | 0 | kVisibility, kMemory); |
188 | 0 | } |
189 | 0 | break; |
190 | 0 | case spv::Op::OpImageRead: |
191 | 0 | case spv::Op::OpImageSparseRead: |
192 | 0 | UpgradeFlags(inst, 2u, is_coherent, is_volatile, kVisibility, kImage); |
193 | 0 | break; |
194 | 0 | case spv::Op::OpImageWrite: |
195 | 0 | UpgradeFlags(inst, 3u, is_coherent, is_volatile, kAvailability, |
196 | 0 | kImage); |
197 | 0 | break; |
198 | 0 | default: |
199 | 0 | break; |
200 | 0 | } |
201 | | |
202 | | // |is_coherent| is never used for the same instructions as |
203 | | // |src_coherent| and |dst_coherent|. |
204 | 0 | if (is_coherent) { |
205 | 0 | inst->AddOperand( |
206 | 0 | {SPV_OPERAND_TYPE_SCOPE_ID, {GetScopeConstant(scope)}}); |
207 | 0 | } |
208 | 0 | if (get_module()->version() >= SPV_SPIRV_VERSION_WORD(1, 4)) { |
209 | | // There are two memory access operands. The first is for the target and |
210 | | // the second is for the source. |
211 | 0 | if (dst_coherent || src_coherent) { |
212 | 0 | start_operand = inst->opcode() == spv::Op::OpCopyMemory ? 2u : 3u; |
213 | 0 | std::vector<Operand> new_operands; |
214 | 0 | uint32_t num_access_words = |
215 | 0 | MemoryAccessNumWords(inst->GetSingleWordInOperand(start_operand)); |
216 | | // The flags were already updated so subtract if we're adding a |
217 | | // scope. |
218 | 0 | if (dst_coherent) --num_access_words; |
219 | 0 | for (uint32_t i = 0; i < start_operand + num_access_words; ++i) { |
220 | 0 | new_operands.push_back(inst->GetInOperand(i)); |
221 | 0 | } |
222 | | // Add the target scope if necessary. |
223 | 0 | if (dst_coherent) { |
224 | 0 | new_operands.push_back( |
225 | 0 | {SPV_OPERAND_TYPE_SCOPE_ID, {GetScopeConstant(dst_scope)}}); |
226 | 0 | } |
227 | | // Copy the remaining current operands. |
228 | 0 | for (uint32_t i = start_operand + num_access_words; |
229 | 0 | i < inst->NumInOperands(); ++i) { |
230 | 0 | new_operands.push_back(inst->GetInOperand(i)); |
231 | 0 | } |
232 | | // Add the source scope if necessary. |
233 | 0 | if (src_coherent) { |
234 | 0 | new_operands.push_back( |
235 | 0 | {SPV_OPERAND_TYPE_SCOPE_ID, {GetScopeConstant(src_scope)}}); |
236 | 0 | } |
237 | 0 | inst->SetInOperands(std::move(new_operands)); |
238 | 0 | } |
239 | 0 | } else { |
240 | | // According to SPV_KHR_vulkan_memory_model, if both available and |
241 | | // visible flags are used the first scope operand is for availability |
242 | | // (writes) and the second is for visibility (reads). |
243 | 0 | if (dst_coherent) { |
244 | 0 | inst->AddOperand( |
245 | 0 | {SPV_OPERAND_TYPE_SCOPE_ID, {GetScopeConstant(dst_scope)}}); |
246 | 0 | } |
247 | 0 | if (src_coherent) { |
248 | 0 | inst->AddOperand( |
249 | 0 | {SPV_OPERAND_TYPE_SCOPE_ID, {GetScopeConstant(src_scope)}}); |
250 | 0 | } |
251 | 0 | } |
252 | 0 | }); |
253 | 0 | } |
254 | 0 | } |
255 | | |
256 | 0 | void UpgradeMemoryModel::UpgradeAtomics() { |
257 | 0 | for (auto& func : *get_module()) { |
258 | 0 | func.ForEachInst([this](Instruction* inst) { |
259 | 0 | if (spvOpcodeIsAtomicOp(inst->opcode())) { |
260 | 0 | bool unused_coherent = false; |
261 | 0 | bool is_volatile = false; |
262 | 0 | spv::Scope unused_scope = spv::Scope::QueueFamilyKHR; |
263 | 0 | std::tie(unused_coherent, is_volatile, unused_scope) = |
264 | 0 | GetInstructionAttributes(inst->GetSingleWordInOperand(0)); |
265 | |
|
266 | 0 | UpgradeSemantics(inst, 2u, is_volatile); |
267 | 0 | if (inst->opcode() == spv::Op::OpAtomicCompareExchange || |
268 | 0 | inst->opcode() == spv::Op::OpAtomicCompareExchangeWeak) { |
269 | 0 | UpgradeSemantics(inst, 3u, is_volatile); |
270 | 0 | } |
271 | 0 | } |
272 | 0 | }); |
273 | 0 | } |
274 | 0 | } |
275 | | |
276 | | void UpgradeMemoryModel::UpgradeSemantics(Instruction* inst, |
277 | | uint32_t in_operand, |
278 | 0 | bool is_volatile) { |
279 | 0 | if (!is_volatile) return; |
280 | | |
281 | 0 | uint32_t semantics_id = inst->GetSingleWordInOperand(in_operand); |
282 | 0 | const analysis::Constant* constant = |
283 | 0 | context()->get_constant_mgr()->FindDeclaredConstant(semantics_id); |
284 | 0 | const analysis::Integer* type = constant->type()->AsInteger(); |
285 | 0 | assert(type && type->width() == 32); |
286 | 0 | uint32_t value = 0; |
287 | 0 | if (type->IsSigned()) { |
288 | 0 | value = static_cast<uint32_t>(constant->GetS32()); |
289 | 0 | } else { |
290 | 0 | value = constant->GetU32(); |
291 | 0 | } |
292 | |
|
293 | 0 | value |= uint32_t(spv::MemorySemanticsMask::Volatile); |
294 | 0 | auto new_constant = context()->get_constant_mgr()->GetConstant(type, {value}); |
295 | 0 | auto new_semantics = |
296 | 0 | context()->get_constant_mgr()->GetDefiningInstruction(new_constant); |
297 | 0 | inst->SetInOperand(in_operand, {new_semantics->result_id()}); |
298 | 0 | } |
299 | | |
300 | | std::tuple<bool, bool, spv::Scope> UpgradeMemoryModel::GetInstructionAttributes( |
301 | 0 | uint32_t id) { |
302 | | // |id| is a pointer used in a memory/image instruction. Need to determine if |
303 | | // that pointer points to volatile or coherent memory. Workgroup storage |
304 | | // class is implicitly coherent and cannot be decorated with volatile, so |
305 | | // short circuit that case. |
306 | 0 | Instruction* inst = context()->get_def_use_mgr()->GetDef(id); |
307 | 0 | analysis::Type* type = context()->get_type_mgr()->GetType(inst->type_id()); |
308 | 0 | if (type->AsPointer() && |
309 | 0 | type->AsPointer()->storage_class() == spv::StorageClass::Workgroup) { |
310 | 0 | return std::make_tuple(true, false, spv::Scope::Workgroup); |
311 | 0 | } |
312 | | |
313 | 0 | bool is_coherent = false; |
314 | 0 | bool is_volatile = false; |
315 | 0 | std::unordered_set<uint32_t> visited; |
316 | 0 | std::tie(is_coherent, is_volatile) = |
317 | 0 | TraceInstruction(context()->get_def_use_mgr()->GetDef(id), |
318 | 0 | std::vector<uint32_t>(), &visited); |
319 | |
|
320 | 0 | return std::make_tuple(is_coherent, is_volatile, spv::Scope::QueueFamilyKHR); |
321 | 0 | } |
322 | | |
323 | | std::pair<bool, bool> UpgradeMemoryModel::TraceInstruction( |
324 | | Instruction* inst, std::vector<uint32_t> indices, |
325 | 0 | std::unordered_set<uint32_t>* visited) { |
326 | 0 | auto iter = cache_.find(std::make_pair(inst->result_id(), indices)); |
327 | 0 | if (iter != cache_.end()) { |
328 | 0 | return iter->second; |
329 | 0 | } |
330 | | |
331 | 0 | if (!visited->insert(inst->result_id()).second) { |
332 | 0 | return std::make_pair(false, false); |
333 | 0 | } |
334 | | |
335 | | // Initialize the cache before |indices| is (potentially) modified. |
336 | 0 | auto& cached_result = cache_[std::make_pair(inst->result_id(), indices)]; |
337 | 0 | cached_result.first = false; |
338 | 0 | cached_result.second = false; |
339 | |
|
340 | 0 | bool is_coherent = false; |
341 | 0 | bool is_volatile = false; |
342 | 0 | switch (inst->opcode()) { |
343 | 0 | case spv::Op::OpVariable: |
344 | 0 | case spv::Op::OpFunctionParameter: |
345 | 0 | is_coherent |= HasDecoration(inst, 0, spv::Decoration::Coherent); |
346 | 0 | is_volatile |= HasDecoration(inst, 0, spv::Decoration::Volatile); |
347 | 0 | if (!is_coherent || !is_volatile) { |
348 | 0 | bool type_coherent = false; |
349 | 0 | bool type_volatile = false; |
350 | 0 | std::tie(type_coherent, type_volatile) = |
351 | 0 | CheckType(inst->type_id(), indices); |
352 | 0 | is_coherent |= type_coherent; |
353 | 0 | is_volatile |= type_volatile; |
354 | 0 | } |
355 | 0 | break; |
356 | 0 | case spv::Op::OpAccessChain: |
357 | 0 | case spv::Op::OpInBoundsAccessChain: |
358 | | // Store indices in reverse order. |
359 | 0 | for (uint32_t i = inst->NumInOperands() - 1; i > 0; --i) { |
360 | 0 | indices.push_back(inst->GetSingleWordInOperand(i)); |
361 | 0 | } |
362 | 0 | break; |
363 | 0 | case spv::Op::OpPtrAccessChain: |
364 | | // Store indices in reverse order. Skip the |Element| operand. |
365 | 0 | for (uint32_t i = inst->NumInOperands() - 1; i > 1; --i) { |
366 | 0 | indices.push_back(inst->GetSingleWordInOperand(i)); |
367 | 0 | } |
368 | 0 | break; |
369 | 0 | default: |
370 | 0 | break; |
371 | 0 | } |
372 | | |
373 | | // No point searching further. |
374 | 0 | if (is_coherent && is_volatile) { |
375 | 0 | cached_result.first = true; |
376 | 0 | cached_result.second = true; |
377 | 0 | return std::make_pair(true, true); |
378 | 0 | } |
379 | | |
380 | | // Variables and function parameters are sources. Continue searching until we |
381 | | // reach them. |
382 | 0 | if (inst->opcode() != spv::Op::OpVariable && |
383 | 0 | inst->opcode() != spv::Op::OpFunctionParameter) { |
384 | 0 | inst->ForEachInId([this, &is_coherent, &is_volatile, &indices, |
385 | 0 | &visited](const uint32_t* id_ptr) { |
386 | 0 | Instruction* op_inst = context()->get_def_use_mgr()->GetDef(*id_ptr); |
387 | 0 | const analysis::Type* type = |
388 | 0 | context()->get_type_mgr()->GetType(op_inst->type_id()); |
389 | 0 | if (type && |
390 | 0 | (type->AsPointer() || type->AsImage() || type->AsSampledImage())) { |
391 | 0 | bool operand_coherent = false; |
392 | 0 | bool operand_volatile = false; |
393 | 0 | std::tie(operand_coherent, operand_volatile) = |
394 | 0 | TraceInstruction(op_inst, indices, visited); |
395 | 0 | is_coherent |= operand_coherent; |
396 | 0 | is_volatile |= operand_volatile; |
397 | 0 | } |
398 | 0 | }); |
399 | 0 | } |
400 | |
|
401 | 0 | cached_result.first = is_coherent; |
402 | 0 | cached_result.second = is_volatile; |
403 | 0 | return std::make_pair(is_coherent, is_volatile); |
404 | 0 | } |
405 | | |
406 | | std::pair<bool, bool> UpgradeMemoryModel::CheckType( |
407 | 0 | uint32_t type_id, const std::vector<uint32_t>& indices) { |
408 | 0 | bool is_coherent = false; |
409 | 0 | bool is_volatile = false; |
410 | 0 | Instruction* type_inst = context()->get_def_use_mgr()->GetDef(type_id); |
411 | 0 | assert(type_inst->opcode() == spv::Op::OpTypePointer); |
412 | 0 | Instruction* element_inst = context()->get_def_use_mgr()->GetDef( |
413 | 0 | type_inst->GetSingleWordInOperand(1u)); |
414 | 0 | for (int i = (int)indices.size() - 1; i >= 0; --i) { |
415 | 0 | if (is_coherent && is_volatile) break; |
416 | | |
417 | 0 | if (element_inst->opcode() == spv::Op::OpTypePointer) { |
418 | 0 | element_inst = context()->get_def_use_mgr()->GetDef( |
419 | 0 | element_inst->GetSingleWordInOperand(1u)); |
420 | 0 | } else if (element_inst->opcode() == spv::Op::OpTypeStruct) { |
421 | 0 | uint32_t index = indices.at(i); |
422 | 0 | Instruction* index_inst = context()->get_def_use_mgr()->GetDef(index); |
423 | 0 | assert(index_inst->opcode() == spv::Op::OpConstant); |
424 | 0 | uint64_t value = GetIndexValue(index_inst); |
425 | 0 | is_coherent |= HasDecoration(element_inst, static_cast<uint32_t>(value), |
426 | 0 | spv::Decoration::Coherent); |
427 | 0 | is_volatile |= HasDecoration(element_inst, static_cast<uint32_t>(value), |
428 | 0 | spv::Decoration::Volatile); |
429 | 0 | element_inst = context()->get_def_use_mgr()->GetDef( |
430 | 0 | element_inst->GetSingleWordInOperand(static_cast<uint32_t>(value))); |
431 | 0 | } else { |
432 | 0 | assert(spvOpcodeIsComposite(element_inst->opcode())); |
433 | 0 | element_inst = context()->get_def_use_mgr()->GetDef( |
434 | 0 | element_inst->GetSingleWordInOperand(0u)); |
435 | 0 | } |
436 | 0 | } |
437 | |
|
438 | 0 | if (!is_coherent || !is_volatile) { |
439 | 0 | bool remaining_coherent = false; |
440 | 0 | bool remaining_volatile = false; |
441 | 0 | std::tie(remaining_coherent, remaining_volatile) = |
442 | 0 | CheckAllTypes(element_inst); |
443 | 0 | is_coherent |= remaining_coherent; |
444 | 0 | is_volatile |= remaining_volatile; |
445 | 0 | } |
446 | |
|
447 | 0 | return std::make_pair(is_coherent, is_volatile); |
448 | 0 | } |
449 | | |
450 | | std::pair<bool, bool> UpgradeMemoryModel::CheckAllTypes( |
451 | 0 | const Instruction* inst) { |
452 | 0 | std::unordered_set<const Instruction*> visited; |
453 | 0 | std::vector<const Instruction*> stack; |
454 | 0 | stack.push_back(inst); |
455 | |
|
456 | 0 | bool is_coherent = false; |
457 | 0 | bool is_volatile = false; |
458 | 0 | while (!stack.empty()) { |
459 | 0 | const Instruction* def = stack.back(); |
460 | 0 | stack.pop_back(); |
461 | |
|
462 | 0 | if (!visited.insert(def).second) continue; |
463 | | |
464 | 0 | if (def->opcode() == spv::Op::OpTypeStruct) { |
465 | | // Any member decorated with coherent and/or volatile is enough to have |
466 | | // the related operation be flagged as coherent and/or volatile. |
467 | 0 | is_coherent |= HasDecoration(def, std::numeric_limits<uint32_t>::max(), |
468 | 0 | spv::Decoration::Coherent); |
469 | 0 | is_volatile |= HasDecoration(def, std::numeric_limits<uint32_t>::max(), |
470 | 0 | spv::Decoration::Volatile); |
471 | 0 | if (is_coherent && is_volatile) |
472 | 0 | return std::make_pair(is_coherent, is_volatile); |
473 | | |
474 | | // Check the subtypes. |
475 | 0 | for (uint32_t i = 0; i < def->NumInOperands(); ++i) { |
476 | 0 | stack.push_back(context()->get_def_use_mgr()->GetDef( |
477 | 0 | def->GetSingleWordInOperand(i))); |
478 | 0 | } |
479 | 0 | } else if (spvOpcodeIsComposite(def->opcode())) { |
480 | 0 | stack.push_back(context()->get_def_use_mgr()->GetDef( |
481 | 0 | def->GetSingleWordInOperand(0u))); |
482 | 0 | } else if (def->opcode() == spv::Op::OpTypePointer) { |
483 | 0 | stack.push_back(context()->get_def_use_mgr()->GetDef( |
484 | 0 | def->GetSingleWordInOperand(1u))); |
485 | 0 | } |
486 | 0 | } |
487 | | |
488 | 0 | return std::make_pair(is_coherent, is_volatile); |
489 | 0 | } |
490 | | |
491 | 0 | uint64_t UpgradeMemoryModel::GetIndexValue(Instruction* index_inst) { |
492 | 0 | const analysis::Constant* index_constant = |
493 | 0 | context()->get_constant_mgr()->GetConstantFromInst(index_inst); |
494 | 0 | assert(index_constant->AsIntConstant()); |
495 | 0 | if (index_constant->type()->AsInteger()->IsSigned()) { |
496 | 0 | if (index_constant->type()->AsInteger()->width() == 32) { |
497 | 0 | return index_constant->GetS32(); |
498 | 0 | } else { |
499 | 0 | return index_constant->GetS64(); |
500 | 0 | } |
501 | 0 | } else { |
502 | 0 | if (index_constant->type()->AsInteger()->width() == 32) { |
503 | 0 | return index_constant->GetU32(); |
504 | 0 | } else { |
505 | 0 | return index_constant->GetU64(); |
506 | 0 | } |
507 | 0 | } |
508 | 0 | } |
509 | | |
510 | | bool UpgradeMemoryModel::HasDecoration(const Instruction* inst, uint32_t value, |
511 | 0 | spv::Decoration decoration) { |
512 | | // If the iteration was terminated early then an appropriate decoration was |
513 | | // found. |
514 | 0 | return !context()->get_decoration_mgr()->WhileEachDecoration( |
515 | 0 | inst->result_id(), (uint32_t)decoration, [value](const Instruction& i) { |
516 | 0 | if (i.opcode() == spv::Op::OpDecorate || |
517 | 0 | i.opcode() == spv::Op::OpDecorateId) { |
518 | 0 | return false; |
519 | 0 | } else if (i.opcode() == spv::Op::OpMemberDecorate) { |
520 | 0 | if (value == i.GetSingleWordInOperand(1u) || |
521 | 0 | value == std::numeric_limits<uint32_t>::max()) |
522 | 0 | return false; |
523 | 0 | } |
524 | | |
525 | 0 | return true; |
526 | 0 | }); |
527 | 0 | } |
528 | | |
529 | | void UpgradeMemoryModel::UpgradeFlags(Instruction* inst, uint32_t in_operand, |
530 | | bool is_coherent, bool is_volatile, |
531 | | OperationType operation_type, |
532 | 0 | InstructionType inst_type) { |
533 | 0 | if (!is_coherent && !is_volatile) return; |
534 | | |
535 | 0 | uint32_t flags = 0; |
536 | 0 | if (inst->NumInOperands() > in_operand) { |
537 | 0 | flags |= inst->GetSingleWordInOperand(in_operand); |
538 | 0 | } |
539 | 0 | if (is_coherent) { |
540 | 0 | if (inst_type == kMemory) { |
541 | 0 | flags |= uint32_t(spv::MemoryAccessMask::NonPrivatePointerKHR); |
542 | 0 | if (operation_type == kVisibility) { |
543 | 0 | flags |= uint32_t(spv::MemoryAccessMask::MakePointerVisibleKHR); |
544 | 0 | } else { |
545 | 0 | flags |= uint32_t(spv::MemoryAccessMask::MakePointerAvailableKHR); |
546 | 0 | } |
547 | 0 | } else { |
548 | 0 | flags |= uint32_t(spv::ImageOperandsMask::NonPrivateTexelKHR); |
549 | 0 | if (operation_type == kVisibility) { |
550 | 0 | flags |= uint32_t(spv::ImageOperandsMask::MakeTexelVisibleKHR); |
551 | 0 | } else { |
552 | 0 | flags |= uint32_t(spv::ImageOperandsMask::MakeTexelAvailableKHR); |
553 | 0 | } |
554 | 0 | } |
555 | 0 | } |
556 | |
|
557 | 0 | if (is_volatile) { |
558 | 0 | if (inst_type == kMemory) { |
559 | 0 | flags |= uint32_t(spv::MemoryAccessMask::Volatile); |
560 | 0 | } else { |
561 | 0 | flags |= uint32_t(spv::ImageOperandsMask::VolatileTexelKHR); |
562 | 0 | } |
563 | 0 | } |
564 | |
|
565 | 0 | if (inst->NumInOperands() > in_operand) { |
566 | 0 | inst->SetInOperand(in_operand, {flags}); |
567 | 0 | } else if (inst_type == kMemory) { |
568 | 0 | inst->AddOperand({SPV_OPERAND_TYPE_OPTIONAL_MEMORY_ACCESS, {flags}}); |
569 | 0 | } else { |
570 | 0 | inst->AddOperand({SPV_OPERAND_TYPE_OPTIONAL_IMAGE, {flags}}); |
571 | 0 | } |
572 | 0 | } |
573 | | |
574 | 0 | uint32_t UpgradeMemoryModel::GetScopeConstant(spv::Scope scope) { |
575 | 0 | analysis::Integer int_ty(32, false); |
576 | 0 | uint32_t int_id = context()->get_type_mgr()->GetTypeInstruction(&int_ty); |
577 | 0 | const analysis::Constant* constant = |
578 | 0 | context()->get_constant_mgr()->GetConstant( |
579 | 0 | context()->get_type_mgr()->GetType(int_id), |
580 | 0 | {static_cast<uint32_t>(scope)}); |
581 | 0 | return context() |
582 | 0 | ->get_constant_mgr() |
583 | 0 | ->GetDefiningInstruction(constant) |
584 | 0 | ->result_id(); |
585 | 0 | } |
586 | | |
587 | 0 | void UpgradeMemoryModel::CleanupDecorations() { |
588 | | // All of the volatile and coherent decorations have been dealt with, so now |
589 | | // we can just remove them. |
590 | 0 | get_module()->ForEachInst([this](Instruction* inst) { |
591 | 0 | if (inst->result_id() != 0) { |
592 | 0 | context()->get_decoration_mgr()->RemoveDecorationsFrom( |
593 | 0 | inst->result_id(), [](const Instruction& dec) { |
594 | 0 | switch (dec.opcode()) { |
595 | 0 | case spv::Op::OpDecorate: |
596 | 0 | case spv::Op::OpDecorateId: |
597 | 0 | if (spv::Decoration(dec.GetSingleWordInOperand(1u)) == |
598 | 0 | spv::Decoration::Coherent || |
599 | 0 | spv::Decoration(dec.GetSingleWordInOperand(1u)) == |
600 | 0 | spv::Decoration::Volatile) |
601 | 0 | return true; |
602 | 0 | break; |
603 | 0 | case spv::Op::OpMemberDecorate: |
604 | 0 | if (spv::Decoration(dec.GetSingleWordInOperand(2u)) == |
605 | 0 | spv::Decoration::Coherent || |
606 | 0 | spv::Decoration(dec.GetSingleWordInOperand(2u)) == |
607 | 0 | spv::Decoration::Volatile) |
608 | 0 | return true; |
609 | 0 | break; |
610 | 0 | default: |
611 | 0 | break; |
612 | 0 | } |
613 | 0 | return false; |
614 | 0 | }); |
615 | 0 | } |
616 | 0 | }); |
617 | 0 | } |
618 | | |
619 | 0 | void UpgradeMemoryModel::UpgradeBarriers() { |
620 | 0 | std::vector<Instruction*> barriers; |
621 | | // Collects all the control barriers in |function|. Returns true if the |
622 | | // function operates on the Output storage class. |
623 | 0 | ProcessFunction CollectBarriers = [this, &barriers](Function* function) { |
624 | 0 | bool operates_on_output = false; |
625 | 0 | for (auto& block : *function) { |
626 | 0 | block.ForEachInst([this, &barriers, |
627 | 0 | &operates_on_output](Instruction* inst) { |
628 | 0 | if (inst->opcode() == spv::Op::OpControlBarrier) { |
629 | 0 | barriers.push_back(inst); |
630 | 0 | } else if (!operates_on_output) { |
631 | | // This instruction operates on output storage class if it is a |
632 | | // pointer to output type or any input operand is a pointer to output |
633 | | // type. |
634 | 0 | analysis::Type* type = |
635 | 0 | context()->get_type_mgr()->GetType(inst->type_id()); |
636 | 0 | if (type && type->AsPointer() && |
637 | 0 | type->AsPointer()->storage_class() == spv::StorageClass::Output) { |
638 | 0 | operates_on_output = true; |
639 | 0 | return; |
640 | 0 | } |
641 | 0 | inst->ForEachInId([this, &operates_on_output](uint32_t* id_ptr) { |
642 | 0 | Instruction* op_inst = |
643 | 0 | context()->get_def_use_mgr()->GetDef(*id_ptr); |
644 | 0 | analysis::Type* op_type = |
645 | 0 | context()->get_type_mgr()->GetType(op_inst->type_id()); |
646 | 0 | if (op_type && op_type->AsPointer() && |
647 | 0 | op_type->AsPointer()->storage_class() == |
648 | 0 | spv::StorageClass::Output) |
649 | 0 | operates_on_output = true; |
650 | 0 | }); |
651 | 0 | } |
652 | 0 | }); |
653 | 0 | } |
654 | 0 | return operates_on_output; |
655 | 0 | }; |
656 | |
|
657 | 0 | std::queue<uint32_t> roots; |
658 | 0 | for (auto& e : get_module()->entry_points()) |
659 | 0 | if (spv::ExecutionModel(e.GetSingleWordInOperand(0u)) == |
660 | 0 | spv::ExecutionModel::TessellationControl) { |
661 | 0 | roots.push(e.GetSingleWordInOperand(1u)); |
662 | 0 | if (context()->ProcessCallTreeFromRoots(CollectBarriers, &roots)) { |
663 | 0 | for (auto barrier : barriers) { |
664 | | // Add OutputMemoryKHR to the semantics of the barriers. |
665 | 0 | uint32_t semantics_id = barrier->GetSingleWordInOperand(2u); |
666 | 0 | Instruction* semantics_inst = |
667 | 0 | context()->get_def_use_mgr()->GetDef(semantics_id); |
668 | 0 | analysis::Type* semantics_type = |
669 | 0 | context()->get_type_mgr()->GetType(semantics_inst->type_id()); |
670 | 0 | uint64_t semantics_value = GetIndexValue(semantics_inst); |
671 | 0 | const analysis::Constant* constant = |
672 | 0 | context()->get_constant_mgr()->GetConstant( |
673 | 0 | semantics_type, |
674 | 0 | {static_cast<uint32_t>(semantics_value) | |
675 | 0 | uint32_t(spv::MemorySemanticsMask::OutputMemoryKHR)}); |
676 | 0 | barrier->SetInOperand(2u, {context() |
677 | 0 | ->get_constant_mgr() |
678 | 0 | ->GetDefiningInstruction(constant) |
679 | 0 | ->result_id()}); |
680 | 0 | } |
681 | 0 | } |
682 | 0 | barriers.clear(); |
683 | 0 | } |
684 | 0 | } |
685 | | |
686 | 0 | void UpgradeMemoryModel::UpgradeMemoryScope() { |
687 | 0 | get_module()->ForEachInst([this](Instruction* inst) { |
688 | | // Don't need to handle all the operations that take a scope. |
689 | | // * Group operations can only be subgroup |
690 | | // * Non-uniform can only be workgroup or subgroup |
691 | | // * Named barriers are not supported by Vulkan |
692 | | // * Workgroup ops (e.g. async_copy) have at most workgroup scope. |
693 | 0 | if (spvOpcodeIsAtomicOp(inst->opcode())) { |
694 | 0 | if (IsDeviceScope(inst->GetSingleWordInOperand(1))) { |
695 | 0 | inst->SetInOperand(1, {GetScopeConstant(spv::Scope::QueueFamilyKHR)}); |
696 | 0 | } |
697 | 0 | } else if (inst->opcode() == spv::Op::OpControlBarrier) { |
698 | 0 | if (IsDeviceScope(inst->GetSingleWordInOperand(1))) { |
699 | 0 | inst->SetInOperand(1, {GetScopeConstant(spv::Scope::QueueFamilyKHR)}); |
700 | 0 | } |
701 | 0 | } else if (inst->opcode() == spv::Op::OpMemoryBarrier) { |
702 | 0 | if (IsDeviceScope(inst->GetSingleWordInOperand(0))) { |
703 | 0 | inst->SetInOperand(0, {GetScopeConstant(spv::Scope::QueueFamilyKHR)}); |
704 | 0 | } |
705 | 0 | } |
706 | 0 | }); |
707 | 0 | } |
708 | | |
709 | 0 | bool UpgradeMemoryModel::IsDeviceScope(uint32_t scope_id) { |
710 | 0 | const analysis::Constant* constant = |
711 | 0 | context()->get_constant_mgr()->FindDeclaredConstant(scope_id); |
712 | 0 | assert(constant && "Memory scope must be a constant"); |
713 | | |
714 | 0 | const analysis::Integer* type = constant->type()->AsInteger(); |
715 | 0 | assert(type); |
716 | 0 | assert(type->width() == 32 || type->width() == 64); |
717 | 0 | if (type->width() == 32) { |
718 | 0 | if (type->IsSigned()) |
719 | 0 | return static_cast<spv::Scope>(constant->GetS32()) == spv::Scope::Device; |
720 | 0 | else |
721 | 0 | return static_cast<spv::Scope>(constant->GetU32()) == spv::Scope::Device; |
722 | 0 | } else { |
723 | 0 | if (type->IsSigned()) |
724 | 0 | return static_cast<spv::Scope>(constant->GetS64()) == spv::Scope::Device; |
725 | 0 | else |
726 | 0 | return static_cast<spv::Scope>(constant->GetU64()) == spv::Scope::Device; |
727 | 0 | } |
728 | | |
729 | 0 | assert(false); |
730 | 0 | return false; |
731 | 0 | } |
732 | | |
733 | 0 | void UpgradeMemoryModel::UpgradeExtInst(Instruction* ext_inst) { |
734 | 0 | const bool is_modf = ext_inst->GetSingleWordInOperand(1u) == GLSLstd450Modf; |
735 | 0 | auto ptr_id = ext_inst->GetSingleWordInOperand(3u); |
736 | 0 | auto ptr_type_id = get_def_use_mgr()->GetDef(ptr_id)->type_id(); |
737 | 0 | auto pointee_type_id = |
738 | 0 | get_def_use_mgr()->GetDef(ptr_type_id)->GetSingleWordInOperand(1u); |
739 | 0 | auto element_type_id = ext_inst->type_id(); |
740 | 0 | std::vector<const analysis::Type*> element_types(2); |
741 | 0 | element_types[0] = context()->get_type_mgr()->GetType(element_type_id); |
742 | 0 | element_types[1] = context()->get_type_mgr()->GetType(pointee_type_id); |
743 | 0 | analysis::Struct struct_type(element_types); |
744 | 0 | uint32_t struct_id = |
745 | 0 | context()->get_type_mgr()->GetTypeInstruction(&struct_type); |
746 | | // Change the operation |
747 | 0 | GLSLstd450 new_op = is_modf ? GLSLstd450ModfStruct : GLSLstd450FrexpStruct; |
748 | 0 | ext_inst->SetOperand(3u, {static_cast<uint32_t>(new_op)}); |
749 | | // Remove the pointer argument |
750 | 0 | ext_inst->RemoveOperand(5u); |
751 | | // Set the type id to the new struct. |
752 | 0 | ext_inst->SetResultType(struct_id); |
753 | | |
754 | | // The result is now a struct of the original result. The zero'th element is |
755 | | // old result and should replace the old result. The one'th element needs to |
756 | | // be stored via a new instruction. |
757 | 0 | auto where = ext_inst->NextNode(); |
758 | 0 | InstructionBuilder builder( |
759 | 0 | context(), where, |
760 | 0 | IRContext::kAnalysisDefUse | IRContext::kAnalysisInstrToBlockMapping); |
761 | 0 | auto extract_0 = |
762 | 0 | builder.AddCompositeExtract(element_type_id, ext_inst->result_id(), {0}); |
763 | 0 | context()->ReplaceAllUsesWith(ext_inst->result_id(), extract_0->result_id()); |
764 | | // The extract's input was just changed to itself, so fix that. |
765 | 0 | extract_0->SetInOperand(0u, {ext_inst->result_id()}); |
766 | 0 | auto extract_1 = |
767 | 0 | builder.AddCompositeExtract(pointee_type_id, ext_inst->result_id(), {1}); |
768 | 0 | builder.AddStore(ptr_id, extract_1->result_id()); |
769 | 0 | } |
770 | | |
771 | 0 | uint32_t UpgradeMemoryModel::MemoryAccessNumWords(uint32_t mask) { |
772 | 0 | uint32_t result = 1; |
773 | 0 | if (mask & uint32_t(spv::MemoryAccessMask::Aligned)) ++result; |
774 | 0 | if (mask & uint32_t(spv::MemoryAccessMask::MakePointerAvailableKHR)) ++result; |
775 | 0 | if (mask & uint32_t(spv::MemoryAccessMask::MakePointerVisibleKHR)) ++result; |
776 | 0 | return result; |
777 | 0 | } |
778 | | |
779 | | } // namespace opt |
780 | | } // namespace spvtools |