mesa/src/amd/compiler/aco_reduce_assign.cpp

/*
 * Copyright © 2018 Valve Corporation
 * Copyright © 2018 Google
 *
 * SPDX-License-Identifier: MIT
 */

#include "aco_builder.h"
#include "aco_ir.h"

#include <vector>

/*
 * Insert p_linear_start instructions right before RA to correctly allocate
 * temporaries for reductions that have to disrespect EXEC by executing in
 * WWM.
 */

namespace aco {

void
setup_reduce_temp(Program* program)
{
   unsigned last_top_level_block_idx = 0;
   unsigned maxSize = 0;

   std::vector<bool> hasReductions(program->blocks.size());
   for (Block& block : program->blocks) {
      for (aco_ptr<Instruction>& instr : block.instructions) {
         if (instr->opcode == aco_opcode::p_interp_gfx11 ||
             instr->opcode == aco_opcode::p_bpermute_permlane) {
            maxSize = MAX2(maxSize, 1);
            hasReductions[block.index] = true;
         } else if (instr->format == Format::PSEUDO_REDUCTION) {
            maxSize = MAX2(maxSize, instr->operands[0].size());
            hasReductions[block.index] = true;
         }
      }
   }

   if (maxSize == 0)
      return;

   assert(maxSize == 1 || maxSize == 2);
   Temp reduceTmp(0, RegClass(RegType::vgpr, maxSize).as_linear());
   Temp vtmp(0, RegClass(RegType::vgpr, maxSize).as_linear());
   int inserted_at = -1;
   int vtmp_inserted_at = -1;

   for (Block& block : program->blocks) {

      if (block.kind & block_kind_top_level) {
         last_top_level_block_idx = block.index;

         /* TODO: this could be improved in this case:
          *    start_linear_vgpr
          *    if (...) {
          *       use_linear_vgpr
          *    }
          *    end_linear_vgpr
          * Here, the linear vgpr is used before any phi copies, so this isn't necessary.
          */
         if (inserted_at >= 0) {
            aco_ptr<Instruction> end{create_instruction(
               aco_opcode::p_end_linear_vgpr, Format::PSEUDO, vtmp_inserted_at >= 0 ? 2 : 1, 0)};
            end->operands[0] = Operand(reduceTmp);
            end->operands[0].setLateKill(true);
            if (vtmp_inserted_at >= 0) {
               end->operands[1] = Operand(vtmp);
               end->operands[1].setLateKill(true);
            }
            /* insert after the phis of the block */
            std::vector<aco_ptr<Instruction>>::iterator it = block.instructions.begin();
            while ((*it)->opcode == aco_opcode::p_linear_phi || (*it)->opcode == aco_opcode::p_phi)
               ++it;
            block.instructions.insert(it, std::move(end));
            inserted_at = vtmp_inserted_at = -1;
         }
      }

      if (!hasReductions[block.index])
         continue;

      std::vector<aco_ptr<Instruction>>::iterator it;
      for (it = block.instructions.begin(); it != block.instructions.end(); ++it) {
         Instruction* instr = (*it).get();
         if (instr->format != Format::PSEUDO_REDUCTION &&
             instr->opcode != aco_opcode::p_interp_gfx11 &&
             instr->opcode != aco_opcode::p_bpermute_permlane)
            continue;

         if ((int)last_top_level_block_idx != inserted_at) {
            reduceTmp = program->allocateTmp(reduceTmp.regClass());
            aco_ptr<Instruction> create{
               create_instruction(aco_opcode::p_start_linear_vgpr, Format::PSEUDO, 0, 1)};
            create->definitions[0] = Definition(reduceTmp);
            /* find the right place to insert this definition */
            if (last_top_level_block_idx == block.index) {
               /* insert right before the current instruction */
               it = block.instructions.insert(it, std::move(create));
               it++;
               /* inserted_at is intentionally not updated here, so later blocks
                * would insert at the end instead of using this one. */
            } else {
               assert(last_top_level_block_idx < block.index);
               /* insert after p_logical_end of the last top-level block */
               std::vector<aco_ptr<Instruction>>& instructions =
                  program->blocks[last_top_level_block_idx].instructions;
               auto insert_point =
                  std::find_if(instructions.rbegin(), instructions.rend(),
                               [](const auto& iter) {
                                  return iter->opcode == aco_opcode::p_logical_end;
                               })
                     .base();
               instructions.insert(insert_point, std::move(create));
               inserted_at = last_top_level_block_idx;
            }
         }

         /* same as before, except for the vector temporary instead of the reduce temporary */
         bool need_vtmp = false;
         if (instr->isReduction()) {
            ReduceOp op = instr->reduction().reduce_op;
            unsigned cluster_size = instr->reduction().cluster_size;
            need_vtmp = op == imul32 || op == fadd64 || op == fmul64 || op == fmin64 ||
                        op == fmax64 || op == umin64 || op == umax64 || op == imin64 ||
                        op == imax64 || op == imul64;
            bool gfx10_need_vtmp = op == imul8 || op == imax8 || op == imin8 || op == umin8 ||
                                   op == imul16 || op == imax16 || op == imin16 || op == umin16 ||
                                   op == iadd64;

            if (program->gfx_level >= GFX10 && cluster_size == 64)
               need_vtmp = true;
            if (program->gfx_level >= GFX10 && gfx10_need_vtmp)
               need_vtmp = true;
            if (program->gfx_level <= GFX7)
               need_vtmp = true;

            need_vtmp |= cluster_size == 32;
         }

         if (need_vtmp && (int)last_top_level_block_idx != vtmp_inserted_at) {
            vtmp = program->allocateTmp(vtmp.regClass());
            aco_ptr<Instruction> create{
               create_instruction(aco_opcode::p_start_linear_vgpr, Format::PSEUDO, 0, 1)};
            create->definitions[0] = Definition(vtmp);
            if (last_top_level_block_idx == block.index) {
               it = block.instructions.insert(it, std::move(create));
               it++;
            } else {
               assert(last_top_level_block_idx < block.index);
               std::vector<aco_ptr<Instruction>>& instructions =
                  program->blocks[last_top_level_block_idx].instructions;
               auto insert_point =
                  std::find_if(instructions.rbegin(), instructions.rend(),
                               [](const auto& iter) {
                                  return iter->opcode == aco_opcode::p_logical_end;
                               })
                     .base();
               instructions.insert(insert_point, std::move(create));
               vtmp_inserted_at = last_top_level_block_idx;
            }
         }

         if (instr->isReduction()) {
            instr->operands[1] = Operand(reduceTmp);
            instr->operands[1].setLateKill(true);
            if (need_vtmp) {
               instr->operands[2] = Operand(vtmp);
               instr->operands[2].setLateKill(true);
            }
         } else {
            assert(instr->opcode == aco_opcode::p_interp_gfx11 ||
                   instr->opcode == aco_opcode::p_bpermute_permlane);
            instr->operands[0] = Operand(reduceTmp);
            instr->operands[0].setLateKill(true);
         }
      }
   }
}

}; // namespace aco