codegen_metal.cc 6.17 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128
/*!
 *  Copyright (c) 2017 by Contributors
 * \file codegen_metal.cc
 */
#include <tvm/runtime/config.h>
#include <tvm/packed_func_ext.h>
#include <vector>
#include <string>
#include "./codegen_metal.h"
#include "../runtime/thread_storage_scope.h"

namespace tvm {
namespace codegen {

void CodeGenMetal::InitFuncState(LoweredFunc f) {
  CodeGenC::InitFuncState(f);
  // analyze the data;
  for (Var arg : f->args) {
    if (arg.type().is_handle()) {
      alloc_storage_scope_[arg.get()] = "global";
    }
  }
}

CodeGenMetal::CodeGenMetal() {
  decl_stream << "#include <metal_stdlib>\n";
  decl_stream << "using namespace metal;\n\n";
  decl_stream << "union __TVMArgUnion {\n"
              << " int v_int;\n"
              << "};\n\n";
}

void CodeGenMetal::AddFunction(LoweredFunc f) {
  // clear previous generated state.
  this->InitFuncState(f);
  // skip the first underscore, so SSA variable starts from _1
  GetUniqueName("_");
  // add to alloc buffer type.
  for (const auto & kv : f->handle_data_type) {
    RegisterHandleType(kv.first.get(), kv.second.type());
  }
  // Function header.
  this->stream << "kernel void " << f->name << "(\n";
  // Buffer arguments
  size_t num_buffer = 0;
  for (size_t i = 0; i < f->args.size(); ++i, ++num_buffer) {
    Var v = f->args[i];
    if (!v.type().is_handle())  break;
    stream << "  ";
    std::string vid = AllocVarID(v.get());
    auto it = alloc_storage_scope_.find(v.get());
    CHECK(it != alloc_storage_scope_.end());
    PrintStorageScope(it->second, stream);
    stream << ' ';
    if (handle_data_type_.count(v.get())) {
      PrintType(handle_data_type_.at(v.get()), stream);
      stream << "*";
    } else {
      PrintType(v.type(), stream);
    }
    stream << ' ' << vid
           << " [[ buffer(" << i << ") ]],\n";
  }
  // Setup normal arguments.
  size_t nargs = f->args.size() - num_buffer;
  std::string varg = GetUniqueName("arg");
  if (nargs != 0) {
    std::string arg_buf_type = f->name + "_args_t";
    stream << "  constant " << arg_buf_type << "& " << varg
           << " [[ buffer(" << num_buffer << ") ]],\n";
    // declare the struct
    decl_stream << "struct " << arg_buf_type << " {\n";
    for (size_t i = num_buffer; i < f->args.size(); ++i) {
      Var v = f->args[i];
      CHECK(!v.type().is_handle());
      std::string vid = AllocVarID(v.get());
      std::ostringstream vref;
      if (v.type().bits() == 32) {
        decl_stream << "  ";
        PrintType(v.type(), decl_stream);
        decl_stream << " " << vid << ";\n";
        vref << varg << "." << vid;
      } else {
        // For non 32bit type, ref through arg union.
        decl_stream << "  __TVMArgUnion " << vid << ";\n";
        vref << varg << "." << vid << ".v_";
        PrintType(v.type(), vref);
      }
      var_idmap_[v.get()] = vref.str();
    }
    decl_stream << "};\n\n";
  }
  // Setup the thread group info.
  CHECK_EQ(GetUniqueName("threadIdx"), "threadIdx");
  CHECK_EQ(GetUniqueName("blockIdx"), "blockIdx");
  int work_dim = 0;
  for (IterVar iv : f->thread_axis) {
    runtime::ThreadScope scope = runtime::ThreadScope::make(iv->thread_tag);
    work_dim = std::max(work_dim, scope.dim_index + 1);
  }
  if (work_dim != 0) {
    // use ushort by default for now
    stream << "  ";
    PrintType(UInt(16, work_dim), stream);
    stream << " blockIdx [[threadgroup_position_in_grid]],\n";
    stream << "  ";
    PrintType(UInt(16, work_dim), stream);
    stream << " threadIdx [[thread_position_in_threadgroup]]\n";
  }
  // bind thread axis
  for (IterVar iv : f->thread_axis) {
    CHECK(!var_idmap_.count(iv->var.get()));
    if (work_dim <= 1) {
      var_idmap_[iv->var.get()] =
          iv->thread_tag.substr(0, iv->thread_tag.length() - 2);
    } else {
      var_idmap_[iv->var.get()] = iv->thread_tag;
    }
  }
  // the function scope.
  stream << ") {\n";
  int func_scope = this->BeginScope();
  this->PrintStmt(f->body);
  this->EndScope(func_scope);
  this->PrintIndent();
  this->stream << "}\n\n";
}

129 130 131 132 133 134
void CodeGenMetal::BindThreadIndex(const IterVar& iv) {
  CHECK(!var_idmap_.count(iv->var.get()));
  var_idmap_[iv->var.get()] =
      CastFromTo(iv->thread_tag, UInt(16), iv->var.type());
}

135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209
void CodeGenMetal::PrintType(Type t, std::ostream& os) const {  // NOLINT(*)
  int lanes = t.lanes();
  if (t.is_handle()) {
    CHECK_EQ(lanes, 1)
        << "do not yet support vector types";
    os << "void*"; return;
  }
  bool fail = false;
  if (t.is_float()) {
    switch (t.bits()) {
      case 16: os << "half"; break;
      case 32: os << "float"; break;
      default: fail = true; break;
    }
    if (!fail && lanes == 1) return;
    if (!fail && (lanes >= 2 && lanes <= 4)) {
      os << lanes; return;
    }
  } else if (t.is_uint() || t.is_int()) {
    if (t.is_uint()) {
      os << 'u';
    }
    if (t.bits() == 8 && t.lanes() == 4) {
      // directly 4 8 bit int in integer.
      os << "int"; return;
    }
    switch (t.bits()) {
      case 8: os << "char"; break;
      case 16: os << "short"; break;
      case 32: os << "int"; break;
      case 1: os << "bool"; break;
      default: fail = true; break;
    }
    if (!fail && lanes == 1) return;
    if (!fail && (lanes >= 2 && lanes <= 4)) {
      os << lanes; return;
    }
  }
  LOG(FATAL) << "Cannot convert type " << t << " to Metal type";
}

void CodeGenMetal::PrintStorageSync(const Call* op) {
  const std::string& sync = op->args[0].as<StringImm>()->value;
  if (sync == "warp") {
    this->PrintIndent();
    this->stream << "simdgroup_barrier(mem_flags::mem_threadgroup);\n";
  } else if (sync == "shared") {
    this->PrintIndent();
    this->stream << "threadgroup_barrier(mem_flags::mem_threadgroup);\n";
  } else if (sync == "global") {
    LOG(FATAL) << "global barrier not supported";
  }
}

void CodeGenMetal::PrintStorageScope(
    const std::string& scope, std::ostream& os) { // NOLINT(*)
  if (scope == "global") {
    os << "device";
  } else if (scope == "shared") {
    os << "threadgroup";
  }
}

void CodeGenMetal::VisitExpr_(const Broadcast* op, std::ostream& os) {   // NOLINT(*)
  std::string v = PrintExpr(op->value);
  PrintType(op->type, os);
  os << "(";
  for (int i = 0; i < op->lanes; ++i) {
    if (i != 0) os << ", ";
    os << v;
  }
  os << ')';
}
}  // namespace codegen
}  // namespace tvm