codegen_opencl.cc 6.25 KB
Newer Older
1 2
/*!
 *  Copyright (c) 2017 by Contributors
3
 * \file codegen_opencl.cc
4 5 6 7
 */
#include <tvm/packed_func_ext.h>
#include <vector>
#include <string>
8 9
#include "codegen_opencl.h"
#include "build_common.h"
10
#include "../runtime/thread_storage_scope.h"
11
#include "../runtime/opencl/opencl_module.h"
12 13 14 15

namespace tvm {
namespace codegen {

16 17 18 19
CodeGenOpenCL::CodeGenOpenCL() {
  restrict_keyword_ = "restrict";
}

20 21
void CodeGenOpenCL::InitFuncState(LoweredFunc f) {
  CodeGenC::InitFuncState(f);
22 23 24 25 26
  for (Var arg : f->args) {
    if (arg.type().is_handle()) {
      alloc_storage_scope_[arg.get()] = "global";
    }
  }
27 28 29
}

void CodeGenOpenCL::AddFunction(LoweredFunc f) {
30
  this->stream << "__kernel ";
31
  CodeGenC::AddFunction(f);
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
std::string CodeGenOpenCL::Finish() {
  // inject extension enable pragma for fp16 and fp64
  if (enable_fp16_) {
    decl_stream
        << "#ifdef cl_khr_fp16\n"
           "#pragma OPENCL EXTENSION cl_khr_fp16 : enable\n"
           "#elif defined(cl_amd_fp16)\n"
           "#pragma OPENCL EXTENSION cl_amd_fp16 : enable\n"
           "#else\n"
           "#error \"Half precision floating point not supported"
                    "by OpenCL implementation on your device.\" \n"
           "#endif\n\n";
  }

  if (enable_fp64_) {
    decl_stream
        << "#ifdef cl_khr_fp64\n"
           "#pragma OPENCL EXTENSION cl_khr_fp64 : enable\n"
           "#elif defined(cl_amd_fp64)\n"
           "#pragma OPENCL EXTENSION cl_amd_fp64 : enable\n"
           "#else\n"
           "#error \"Double precision floating point not supported"
                    "by OpenCL implementation on your device.\" \n"
           "#endif\n\n";
  }

  return CodeGenC::Finish();
}

63 64 65 66
void CodeGenOpenCL::BindThreadIndex(const IterVar& iv) {
  CHECK(!var_idmap_.count(iv->var.get()));
  runtime::ThreadScope ts = runtime::ThreadScope::make(iv->thread_tag);
  std::ostringstream os;
67 68
  if (ts.rank == 1) {
    os << "get_local_id(" << ts.dim_index << ")";
69
  } else {
70
    os << "get_group_id(" << ts.dim_index << ")";
71
  }
72 73
  var_idmap_[iv->var.get()] =
      CastFromTo(os.str(), UInt(64), iv->var.type());
74 75
}

76
void CodeGenOpenCL::PrintType(Type t, std::ostream& os) {  // NOLINT(*)
77 78 79 80 81 82
  int lanes = t.lanes();
  if (t.is_handle()) {
    CHECK_EQ(lanes, 1)
        << "do not yet support vector types";
    os << "void*"; return;
  }
83 84 85
  if (t == Bool()) {
    os << "bool"; return;
  }
86 87 88
  bool fail = false;
  if (t.is_float()) {
    switch (t.bits()) {
89 90 91 92
      case 16:
        os << "half";
        enable_fp16_ = true;
        break;
93
      case 32: os << "float"; break;
94 95 96 97
      case 64:
        os << "double";
        enable_fp64_ = true;
        break;
98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142
      default: fail = true; break;
    }
    if (!fail && lanes == 1) return;
    if (!fail && (lanes >= 2 && lanes <= 16)) {
      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 64: os << "long"; break;
      case 1: os << "int"; break;
      default: fail = true; break;
    }
    if (!fail && lanes == 1) return;
    if (!fail && (lanes >= 2 && lanes <= 16)) {
      os << lanes; return;
    }
  }
  LOG(FATAL) << "Cannot convert type " << t << " to OpenCL type";
}

void CodeGenOpenCL::PrintVecAddr(const Variable* buffer, Type t,
                                 Expr base, std::ostream& os) {  // NOLINT(*)
  if (!HandleTypeMatch(buffer, t.element_of())) {
    os << '(';
    auto it = alloc_storage_scope_.find(buffer);
    if (it != alloc_storage_scope_.end()) {
      PrintStorageScope(it->second, os);
    }
    os << ' ';
    PrintType(t.element_of(), os);
    os << "*)";
  }
  os << GetVarID(buffer) << " + ";
  PrintExpr(base, os);
}
143 144
std::string CodeGenOpenCL::GetVecLoad(
    Type t, const Variable* buffer, Expr base) {
145
  std::ostringstream os;
146 147 148
  os << "vload" << t.lanes() << "(0, ";
  PrintVecAddr(buffer, t, base, os);
  os << ")";
149
  return os.str();
150 151 152 153 154 155 156 157 158 159
}

void CodeGenOpenCL::PrintVecStore(const Variable* buffer,
                                  Type t, Expr base,
                                  const std::string& value) {
  this->PrintIndent();
  stream << "vstore" << t.lanes() << "(" << value << ", 0, ";
  PrintVecAddr(buffer, t, base, stream);
  stream << ");\n";
}
160

161 162
void CodeGenOpenCL::PrintStorageSync(const Call* op) {
  const std::string& sync = op->args[0].as<StringImm>()->value;
163
  if (sync == "warp") {
164
    this->PrintIndent();
165
    this->stream << "barrier(CLK_LOCAL_MEM_FENCE);\n";
166
  } else if (sync == "shared") {
167 168 169 170 171 172 173
    this->PrintIndent();
    this->stream << "barrier(CLK_LOCAL_MEM_FENCE);\n";
  } else if (sync == "global") {
    LOG(FATAL) << "not supported";
  }
}

174 175
void CodeGenOpenCL::PrintStorageScope(
    const std::string& scope, std::ostream& os) { // NOLINT(*)
176 177 178
  if (scope == "global") {
    os << "__global";
  } else if (scope == "shared") {
179 180 181 182
    os << "__local";
  }
}

183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198
std::string CodeGenOpenCL::CastFromTo(std::string value, Type from, Type target) {
  if (from == target) return value;
  std::ostringstream os;
  if (target.lanes() == 1) {
    os << "((";
    this->PrintType(target, os);
    os << ")" << value << ")";
  } else {  // convert vector type
    os << "(";
    os << "convert_";
    this->PrintType(target, os);
    os << "(" << value << "))";
  }
  return os.str();
}

199 200
void CodeGenOpenCL::VisitExpr_(const Broadcast* op, std::ostream& os) {   // NOLINT(*)
  std::string v = PrintExpr(op->value);
201
  os << "((";
202 203 204 205 206
  PrintType(op->type, os);
  os << ")(";
  for (int i = 0; i < op->lanes; ++i) {
    if (i != 0) os << ", ";
    os << v;
207
  }
208
  os << "))";
209
}
210 211 212 213 214 215 216 217 218 219 220 221 222 223


runtime::Module BuildOpenCL(Array<LoweredFunc> funcs) {
  using tvm::runtime::Registry;
  bool output_ssa = false;
  CodeGenOpenCL cg;
  cg.Init(output_ssa);
  for (LoweredFunc f : funcs) {
    cg.AddFunction(f);
  }
  std::string code = cg.Finish();
  if (const auto* f = Registry::Get("tvm_callback_opencl_postproc")) {
    code = (*f)(code).operator std::string();
  }
224
  return OpenCLModuleCreate(code, "cl", ExtractFuncInfo(funcs), code);
225 226 227 228 229 230
}

TVM_REGISTER_API("codegen.build_opencl")
.set_body([](TVMArgs args, TVMRetValue* rv) {
    *rv = BuildOpenCL(args[0]);
  });
231 232
}  // namespace codegen
}  // namespace tvm