src/amd/compiler/aco_reduce_assign.cpp

   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_ir.h"
  27 #include "aco_builder.h"
  28
  29 /*
  30  * Insert p_linear_start instructions right before RA to correctly allocate
  31  * temporaries for reductions that have to disrespect EXEC by executing in
  32  * WWM.
  33  */
  34
  35 namespace aco {
  36
  37 void setup_reduce_temp(Program* program)
  38 {
  39    unsigned last_top_level_block_idx = 0;
  40    unsigned maxSize = 0;
  41
  42    std::vector<bool> hasReductions(program->blocks.size());
  43    for (Block& block : program->blocks) {
  44       for (aco_ptr<Instruction>& instr : block.instructions) {
  45          if (instr->format != Format::PSEUDO_REDUCTION)
  46             continue;
  47
  48          maxSize = MAX2(maxSize, instr->operands[0].size());
  49          hasReductions[block.index] = true;
  50       }
  51    }
  52
  53    if (maxSize == 0)
  54       return;
  55
  56    assert(maxSize == 1 || maxSize == 2);
  57    Temp reduceTmp(0, RegClass(RegType::vgpr, maxSize).as_linear());
  58    Temp vtmp(0, RegClass(RegType::vgpr, maxSize).as_linear());
  59    int inserted_at = -1;
  60    int vtmp_inserted_at = -1;
  61    bool reduceTmp_in_loop = false;
  62    bool vtmp_in_loop = false;
  63
  64    for (Block& block : program->blocks) {
  65
  66       /* insert p_end_linear_vgpr after the outermost loop */
  67       if (reduceTmp_in_loop && block.loop_nest_depth == 0) {
  68          assert(inserted_at == (int)last_top_level_block_idx);
  69
  70          aco_ptr<Instruction> end{create_instruction<Instruction>(aco_opcode::p_end_linear_vgpr, Format::PSEUDO, vtmp_in_loop ? 2 : 1, 0)};
  71          end->operands[0] = Operand(reduceTmp);
  72          if (vtmp_in_loop)
  73             end->operands[1] = Operand(vtmp);
  74          /* insert after the phis of the loop exit block */
  75          std::vector<aco_ptr<Instruction>>::iterator it = block.instructions.begin();
  76          while ((*it)->opcode == aco_opcode::p_linear_phi || (*it)->opcode == aco_opcode::p_phi)
  77             ++it;
  78          block.instructions.insert(it, std::move(end));
  79          reduceTmp_in_loop = false;
  80       }
  81
  82       if (block.kind & block_kind_top_level)
  83          last_top_level_block_idx = block.index;
  84
  85       if (!hasReductions[block.index])
  86          continue;
  87
  88       std::vector<aco_ptr<Instruction>>::iterator it;
  89       for (it = block.instructions.begin(); it != block.instructions.end(); ++it) {
  90          Instruction *instr = (*it).get();
  91          if (instr->format != Format::PSEUDO_REDUCTION)
  92             continue;
  93
  94          ReduceOp op = static_cast<Pseudo_reduction_instruction *>(instr)->reduce_op;
  95          reduceTmp_in_loop |= block.loop_nest_depth > 0;
  96
  97          if ((int)last_top_level_block_idx != inserted_at) {
  98             reduceTmp = {program->allocateId(), reduceTmp.regClass()};
  99             aco_ptr<Pseudo_instruction> create{create_instruction<Pseudo_instruction>(aco_opcode::p_start_linear_vgpr, Format::PSEUDO, 0, 1)};
 100             create->definitions[0] = Definition(reduceTmp);
 101             /* find the right place to insert this definition */
 102             if (last_top_level_block_idx == block.index) {
 103                /* insert right before the current instruction */
 104                it = block.instructions.insert(it, std::move(create));
 105                it++;
 106                /* inserted_at is intentionally not updated here, so later blocks
 107                 * would insert at the end instead of using this one. */
 108             } else {
 109                assert(last_top_level_block_idx < block.index);
 110                /* insert before the branch at last top level block */
 111                std::vector<aco_ptr<Instruction>>& instructions = program->blocks[last_top_level_block_idx].instructions;
 112                instructions.insert(std::next(instructions.begin(), instructions.size() - 1), std::move(create));
 113                inserted_at = last_top_level_block_idx;
 114             }
 115          }
 116
 117          if (op == gfx10_wave64_bpermute) {
 118             instr->operands[1] = Operand(reduceTmp);
 119             continue;
 120          }
 121
 122          /* same as before, except for the vector temporary instead of the reduce temporary */
 123          unsigned cluster_size = static_cast<Pseudo_reduction_instruction *>(instr)->cluster_size;
 124          bool need_vtmp = op == imul32 || op == fadd64 || op == fmul64 ||
 125                           op == fmin64 || op == fmax64 || op == umin64 ||
 126                           op == umax64 || op == imin64 || op == imax64 ||
 127                           op == imul64;
 128          bool gfx10_need_vtmp = op == imul8 || op == imax8 || op == imin8 || op == umin8 ||
 129                                 op == imul16 || op == imax16 || op == imin16 || op == umin16 ||
 130                                 op == iadd64;
 131
 132          if (program->chip_class >= GFX10 && cluster_size == 64)
 133             need_vtmp = true;
 134          if (program->chip_class >= GFX10 && gfx10_need_vtmp)
 135             need_vtmp = true;
 136          if (program->chip_class <= GFX7)
 137             need_vtmp = true;
 138
 139          need_vtmp |= cluster_size == 32;
 140
 141          vtmp_in_loop |= need_vtmp && block.loop_nest_depth > 0;
 142          if (need_vtmp && (int)last_top_level_block_idx != vtmp_inserted_at) {
 143             vtmp = {program->allocateId(), vtmp.regClass()};
 144             aco_ptr<Pseudo_instruction> create{create_instruction<Pseudo_instruction>(aco_opcode::p_start_linear_vgpr, Format::PSEUDO, 0, 1)};
 145             create->definitions[0] = Definition(vtmp);
 146             if (last_top_level_block_idx == block.index) {
 147                it = block.instructions.insert(it, std::move(create));
 148                it++;
 149             } else {
 150                assert(last_top_level_block_idx < block.index);
 151                std::vector<aco_ptr<Instruction>>& instructions = program->blocks[last_top_level_block_idx].instructions;
 152                instructions.insert(std::next(instructions.begin(), instructions.size() - 1), std::move(create));
 153                vtmp_inserted_at = last_top_level_block_idx;
 154             }
 155          }
 156
 157          instr->operands[1] = Operand(reduceTmp);
 158          if (need_vtmp)
 159             instr->operands[2] = Operand(vtmp);
 160
 161          /* scalar temporary */
 162          Builder bld(program);
 163          instr->definitions[1] = bld.def(s2);
 164
 165          /* scalar identity temporary */
 166          bool need_sitmp = (program->chip_class <= GFX7 || program->chip_class >= GFX10) && instr->opcode != aco_opcode::p_reduce;
 167          if (instr->opcode == aco_opcode::p_exclusive_scan) {
 168             need_sitmp |=
 169                (op == imin8 || op == imin16 || op == imin32 || op == imin64 ||
 170                 op == imax8 || op == imax16 || op == imax32 || op == imax64 ||
 171                 op == fmin16 || op == fmin32 || op == fmin64 ||
 172                 op == fmax16 || op == fmax32 || op == fmax64 ||
 173                 op == fmul16 || op == fmul64);
 174          }
 175          if (need_sitmp) {
 176             instr->definitions[2] = bld.def(RegClass(RegType::sgpr, instr->operands[0].size()));
 177          }
 178
 179          /* vcc clobber */
 180          bool clobber_vcc = false;
 181          if ((op == iadd32 || op == imul64) && program->chip_class < GFX9)
 182             clobber_vcc = true;
 183          if (op == iadd64 || op == umin64 || op == umax64 || op == imin64 || op == imax64)
 184             clobber_vcc = true;
 185
 186          if (clobber_vcc)
 187             instr->definitions[4] = Definition(vcc, bld.lm);
 188       }
 189    }
 190 }
 191
 192 };
 193