1 /*
2  * Copyright © 2018 Valve Corporation
3  * Copyright © 2018 Google
4  *
5  * Permission is hereby granted, free of charge, to any person obtaining a
6  * copy of this software and associated documentation files (the "Software"),
7  * to deal in the Software without restriction, including without limitation
8  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
9  * and/or sell copies of the Software, and to permit persons to whom the
10  * Software is furnished to do so, subject to the following conditions:
11  *
12  * The above copyright notice and this permission notice (including the next
13  * paragraph) shall be included in all copies or substantial portions of the
14  * Software.
15  *
16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
19  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
21  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
22  * IN THE SOFTWARE.
23  *
24  */
25 
26 #include "aco_builder.h"
27 #include "aco_ir.h"
28 
29 #include <vector>
30 
31 /*
32  * Insert p_linear_start instructions right before RA to correctly allocate
33  * temporaries for reductions that have to disrespect EXEC by executing in
34  * WWM.
35  */
36 
37 namespace aco {
38 
39 void
setup_reduce_temp(Program * program)40 setup_reduce_temp(Program* program)
41 {
42    unsigned last_top_level_block_idx = 0;
43    unsigned maxSize = 0;
44 
45    std::vector<bool> hasReductions(program->blocks.size());
46    for (Block& block : program->blocks) {
47       for (aco_ptr<Instruction>& instr : block.instructions) {
48          if (instr->format != Format::PSEUDO_REDUCTION)
49             continue;
50 
51          maxSize = MAX2(maxSize, instr->operands[0].size());
52          hasReductions[block.index] = true;
53       }
54    }
55 
56    if (maxSize == 0)
57       return;
58 
59    assert(maxSize == 1 || maxSize == 2);
60    Temp reduceTmp(0, RegClass(RegType::vgpr, maxSize).as_linear());
61    Temp vtmp(0, RegClass(RegType::vgpr, maxSize).as_linear());
62    int inserted_at = -1;
63    int vtmp_inserted_at = -1;
64    bool reduceTmp_in_loop = false;
65    bool vtmp_in_loop = false;
66 
67    for (Block& block : program->blocks) {
68 
69       /* insert p_end_linear_vgpr after the outermost loop */
70       if (reduceTmp_in_loop && block.loop_nest_depth == 0) {
71          assert(inserted_at == (int)last_top_level_block_idx);
72 
73          aco_ptr<Instruction> end{create_instruction<Instruction>(
74             aco_opcode::p_end_linear_vgpr, Format::PSEUDO, vtmp_in_loop ? 2 : 1, 0)};
75          end->operands[0] = Operand(reduceTmp);
76          if (vtmp_in_loop)
77             end->operands[1] = Operand(vtmp);
78          /* insert after the phis of the loop exit block */
79          std::vector<aco_ptr<Instruction>>::iterator it = block.instructions.begin();
80          while ((*it)->opcode == aco_opcode::p_linear_phi || (*it)->opcode == aco_opcode::p_phi)
81             ++it;
82          block.instructions.insert(it, std::move(end));
83          reduceTmp_in_loop = false;
84       }
85 
86       if (block.kind & block_kind_top_level)
87          last_top_level_block_idx = block.index;
88 
89       if (!hasReductions[block.index])
90          continue;
91 
92       std::vector<aco_ptr<Instruction>>::iterator it;
93       for (it = block.instructions.begin(); it != block.instructions.end(); ++it) {
94          Instruction* instr = (*it).get();
95          if (instr->format != Format::PSEUDO_REDUCTION)
96             continue;
97 
98          ReduceOp op = instr->reduction().reduce_op;
99          reduceTmp_in_loop |= block.loop_nest_depth > 0;
100 
101          if ((int)last_top_level_block_idx != inserted_at) {
102             reduceTmp = program->allocateTmp(reduceTmp.regClass());
103             aco_ptr<Pseudo_instruction> create{create_instruction<Pseudo_instruction>(
104                aco_opcode::p_start_linear_vgpr, Format::PSEUDO, 0, 1)};
105             create->definitions[0] = Definition(reduceTmp);
106             /* find the right place to insert this definition */
107             if (last_top_level_block_idx == block.index) {
108                /* insert right before the current instruction */
109                it = block.instructions.insert(it, std::move(create));
110                it++;
111                /* inserted_at is intentionally not updated here, so later blocks
112                 * would insert at the end instead of using this one. */
113             } else {
114                assert(last_top_level_block_idx < block.index);
115                /* insert before the branch at last top level block */
116                std::vector<aco_ptr<Instruction>>& instructions =
117                   program->blocks[last_top_level_block_idx].instructions;
118                instructions.insert(std::next(instructions.begin(), instructions.size() - 1),
119                                    std::move(create));
120                inserted_at = last_top_level_block_idx;
121             }
122          }
123 
124          /* same as before, except for the vector temporary instead of the reduce temporary */
125          unsigned cluster_size = instr->reduction().cluster_size;
126          bool need_vtmp = op == imul32 || op == fadd64 || op == fmul64 || op == fmin64 ||
127                           op == fmax64 || op == umin64 || op == umax64 || op == imin64 ||
128                           op == imax64 || op == imul64;
129          bool gfx10_need_vtmp = op == imul8 || op == imax8 || op == imin8 || op == umin8 ||
130                                 op == imul16 || op == imax16 || op == imin16 || op == umin16 ||
131                                 op == iadd64;
132 
133          if (program->chip_class >= GFX10 && cluster_size == 64)
134             need_vtmp = true;
135          if (program->chip_class >= GFX10 && gfx10_need_vtmp)
136             need_vtmp = true;
137          if (program->chip_class <= GFX7)
138             need_vtmp = true;
139 
140          need_vtmp |= cluster_size == 32;
141 
142          vtmp_in_loop |= need_vtmp && block.loop_nest_depth > 0;
143          if (need_vtmp && (int)last_top_level_block_idx != vtmp_inserted_at) {
144             vtmp = program->allocateTmp(vtmp.regClass());
145             aco_ptr<Pseudo_instruction> create{create_instruction<Pseudo_instruction>(
146                aco_opcode::p_start_linear_vgpr, Format::PSEUDO, 0, 1)};
147             create->definitions[0] = Definition(vtmp);
148             if (last_top_level_block_idx == block.index) {
149                it = block.instructions.insert(it, std::move(create));
150                it++;
151             } else {
152                assert(last_top_level_block_idx < block.index);
153                std::vector<aco_ptr<Instruction>>& instructions =
154                   program->blocks[last_top_level_block_idx].instructions;
155                instructions.insert(std::next(instructions.begin(), instructions.size() - 1),
156                                    std::move(create));
157                vtmp_inserted_at = last_top_level_block_idx;
158             }
159          }
160 
161          instr->operands[1] = Operand(reduceTmp);
162          if (need_vtmp)
163             instr->operands[2] = Operand(vtmp);
164       }
165    }
166 }
167 
168 }; // namespace aco
169