1 | // Copyright (c) 2019 Valve Corporation |
2 | // Copyright (c) 2019 LunarG Inc. |
3 | // |
4 | // Licensed under the Apache License, Version 2.0 (the "License"); |
5 | // you may not use this file except in compliance with the License. |
6 | // You may obtain a copy of the License at |
7 | // |
8 | // http://www.apache.org/licenses/LICENSE-2.0 |
9 | // |
10 | // Unless required by applicable law or agreed to in writing, software |
11 | // distributed under the License is distributed on an "AS IS" BASIS, |
12 | // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
13 | // See the License for the specific language governing permissions and |
14 | // limitations under the License. |
15 | |
16 | #ifndef LIBSPIRV_OPT_CONVERT_TO_HALF_PASS_H_ |
17 | #define LIBSPIRV_OPT_CONVERT_TO_HALF_PASS_H_ |
18 | |
19 | #include "source/opt/ir_builder.h" |
20 | #include "source/opt/pass.h" |
21 | |
22 | namespace spvtools { |
23 | namespace opt { |
24 | |
25 | class ConvertToHalfPass : public Pass { |
26 | public: |
27 | ConvertToHalfPass() : Pass() {} |
28 | |
29 | ~ConvertToHalfPass() override = default; |
30 | |
31 | IRContext::Analysis GetPreservedAnalyses() override { |
32 | return IRContext::kAnalysisDefUse | IRContext::kAnalysisInstrToBlockMapping; |
33 | } |
34 | |
35 | // See optimizer.hpp for pass user documentation. |
36 | Status Process() override; |
37 | |
38 | const char* name() const override { return "convert-to-half-pass" ; } |
39 | |
40 | private: |
41 | // Return true if |inst| is an arithmetic, composite or phi op that can be |
42 | // of type float16 |
43 | bool IsArithmetic(Instruction* inst); |
44 | |
45 | // Return true if |inst| returns scalar, vector or matrix type with base |
46 | // float and |width| |
47 | bool IsFloat(Instruction* inst, uint32_t width); |
48 | |
49 | // Return true if |inst| is decorated with RelaxedPrecision |
50 | bool IsDecoratedRelaxed(Instruction* inst); |
51 | |
52 | // Return true if |id| has been added to the relaxed id set |
53 | bool IsRelaxed(uint32_t id); |
54 | |
55 | // Add |id| to the relaxed id set |
56 | void AddRelaxed(uint32_t id); |
57 | |
58 | // Return type id for float with |width| |
59 | analysis::Type* FloatScalarType(uint32_t width); |
60 | |
61 | // Return type id for vector of length |vlen| of float of |width| |
62 | analysis::Type* FloatVectorType(uint32_t v_len, uint32_t width); |
63 | |
64 | // Return type id for matrix of |v_cnt| vectors of length identical to |
65 | // |vty_id| of float of |width| |
66 | analysis::Type* FloatMatrixType(uint32_t v_cnt, uint32_t vty_id, |
67 | uint32_t width); |
68 | |
69 | // Return equivalent to float type |ty_id| with |width| |
70 | uint32_t EquivFloatTypeId(uint32_t ty_id, uint32_t width); |
71 | |
72 | // Append instructions to builder to convert value |*val_idp| to type |
73 | // |ty_id| but with |width|. Set |*val_idp| to the new id. |
74 | void GenConvert(uint32_t* val_idp, uint32_t width, Instruction* inst); |
75 | |
76 | // Remove RelaxedPrecision decoration of |id|. |
77 | void RemoveRelaxedDecoration(uint32_t id); |
78 | |
79 | // Add |inst| to relaxed instruction set if warranted. Specifically, if |
80 | // it is float32 and either decorated relaxed or a composite or phi |
81 | // instruction where all operands are relaxed or all uses are relaxed. |
82 | bool CloseRelaxInst(Instruction* inst); |
83 | |
84 | // If |inst| is an arithmetic, phi, extract or convert instruction of float32 |
85 | // base type and decorated with RelaxedPrecision, change it to the equivalent |
86 | // float16 based type instruction. Specifically, insert instructions to |
87 | // convert all operands to float16 (if needed) and change its type to the |
88 | // equivalent float16 type. Otherwise, insert instructions to convert its |
89 | // operands back to their original types, if needed. |
90 | bool GenHalfInst(Instruction* inst); |
91 | |
92 | // Gen code for relaxed arithmetic |inst| |
93 | bool GenHalfArith(Instruction* inst); |
94 | |
95 | // Gen code for relaxed phi |inst| |
96 | bool ProcessPhi(Instruction* inst); |
97 | |
98 | // Gen code for relaxed convert |inst| |
99 | bool ProcessConvert(Instruction* inst); |
100 | |
101 | // Gen code for image reference |inst| |
102 | bool ProcessImageRef(Instruction* inst); |
103 | |
104 | // Process default non-relaxed |inst| |
105 | bool ProcessDefault(Instruction* inst); |
106 | |
107 | // If |inst| is an FConvert of a matrix type, decompose it to a series |
108 | // of vector extracts, converts and inserts into an Undef. These are |
109 | // generated by GenHalfInst because they are easier to manipulate, but are |
110 | // invalid so we need to clean them up. |
111 | bool MatConvertCleanup(Instruction* inst); |
112 | |
113 | // Call GenHalfInst on every instruction in |func|. |
114 | // If code is generated for an instruction, replace the instruction |
115 | // with the new instructions that are generated. |
116 | bool ProcessFunction(Function* func); |
117 | |
118 | Pass::Status ProcessImpl(); |
119 | |
120 | // Initialize state for converting to half |
121 | void Initialize(); |
122 | |
123 | // Set of core operations to be processed |
124 | std::unordered_set<uint32_t> target_ops_core_; |
125 | |
126 | // Set of 450 extension operations to be processed |
127 | std::unordered_set<uint32_t> target_ops_450_; |
128 | |
129 | // Set of sample operations |
130 | std::unordered_set<uint32_t> image_ops_; |
131 | |
132 | // Set of dref sample operations |
133 | std::unordered_set<uint32_t> dref_image_ops_; |
134 | |
135 | // Set of dref sample operations |
136 | std::unordered_set<uint32_t> closure_ops_; |
137 | |
138 | // Set of ids of all relaxed instructions |
139 | std::unordered_set<uint32_t> relaxed_ids_set_; |
140 | |
141 | // Ids of all converted instructions |
142 | std::unordered_set<uint32_t> converted_ids_; |
143 | }; |
144 | |
145 | } // namespace opt |
146 | } // namespace spvtools |
147 | |
148 | #endif // LIBSPIRV_OPT_CONVERT_TO_HALF_PASS_H_ |
149 | |