lower_tvm_builtin.cc 10.6 KB
Newer Older
1 2
/*!
 *  Copyright (c) 2017 by Contributors
3 4
 *  Lower TVM related buildin intrinsics such as packed call.
 * \file lower_tvm_buildin.cc
5 6 7 8 9 10
 */
#include <tvm/ir.h>
#include <tvm/ir_mutator.h>
#include <tvm/ir_pass.h>
#include <unordered_set>
#include "./ir_util.h"
11
#include "../arithmetic/compute_expr.h"
12

13 14 15 16 17 18 19 20 21 22 23 24 25 26 27
namespace tvm {
namespace ir {

inline Expr ConstInt32(size_t index) {
  CHECK_LE(index, std::numeric_limits<int>::max());
  return make_const(Int(32), static_cast<int>(index));
}

inline Expr StackAlloca(std::string type, size_t num) {
  Array<Expr> args = {StringImm::make(type), ConstInt32(num)};
  return Call::make(Handle(), intrinsic::tvm_stack_alloca, args, Call::Intrinsic);
}

// Calculate the statistics of packed function.
// These information are needed during codegen.
28
class BuiltinLower : public IRMutator {
29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51
 public:
  Stmt Build(Stmt stmt) {
    stack_shape_ = Var("stack_shape", Handle());
    stack_array_ = Var("stack_array", Handle());
    stack_value_ = Var("stack_value", Handle());
    stack_tcode_ = Var("stack_tcode", Handle());
    stmt = this->Mutate(stmt);
    if (max_shape_stack_ != 0) {
      stmt = LetStmt::make(
          stack_shape_, StackAlloca("shape", max_shape_stack_), stmt);
    }
    if (max_array_stack_ != 0) {
      stmt = LetStmt::make(
          stack_array_, StackAlloca("array", max_array_stack_), stmt);
    }
    if (max_arg_stack_ != 0) {
      stmt = LetStmt::make(
          stack_value_, StackAlloca("arg_value", max_arg_stack_), stmt);
      stmt = LetStmt::make(
          stack_tcode_, StackAlloca("arg_tcode", max_arg_stack_), stmt);
    }
    return stmt;
  }
52

53 54 55 56 57 58 59 60 61 62 63
  Stmt Mutate(Stmt stmt) final {
    stmt = IRMutator::Mutate(stmt);
    CHECK_EQ(run_shape_stack_, 0);
    CHECK_EQ(run_array_stack_, 0);
    CHECK_EQ(run_arg_stack_, 0);
    while (prep_seq_.size() != 0) {
      stmt = Block::make(prep_seq_.back(), stmt);
      prep_seq_.pop_back();
    }
    return stmt;
  }
64 65 66 67 68

  Stmt Mutate_(const Allocate* op, const Stmt& s) {
    // Lower allocate to device allocate when needed.
    Stmt stmt = IRMutator::Mutate_(op, s);
    op = stmt.as<Allocate>();
69
    if (op->new_expr.defined()) return stmt;
70 71 72 73 74
    // Get constant allocation bound.
    int64_t dev_type;
    int64_t nbytes = GetVectorBytes(op->type);
    if (device_type_.defined()) {
      if (arith::GetConst(device_type_, &dev_type)) {
75
        if (dev_type == kDLCPU) {
76 77 78 79 80 81 82 83 84 85 86
          int32_t constant_size = op->constant_allocation_size();
          if (constant_size > 0 && constant_size * nbytes < runtime::kMaxStackAlloca) {
            return stmt;
          }
        }
      }
    }
    Expr total_bytes = make_const(op->extents[0].type(), nbytes);
    for (size_t i = 0; i < op->extents.size(); ++i) {
      total_bytes = total_bytes * op->extents[i];
    }
87 88
    CHECK(device_type_.defined()) << "Unknown device type in current IR";
    CHECK(device_id_.defined()) << "Unknown device id in current IR";
89 90 91 92 93 94 95 96 97 98
    Stmt throw_last_error = Evaluate::make(Call::make(Int(32),
                                           intrinsic::tvm_throw_last_error, {},
                                           Call::Intrinsic));

    Stmt body = Block::make(
        IfThenElse::make(Call::make(Bool(1),
                                    intrinsic::tvm_handle_is_null,
                                    {op->buffer_var}, Call::PureIntrinsic),
                         throw_last_error),
        op->body);
99 100 101 102 103 104 105 106 107 108 109 110

    Stmt alloca = LetStmt::make(
        op->buffer_var,
        Call::make(op->buffer_var.type(),
                   "TVMBackendAllocWorkspace",
                   {cast(Int(32), device_type_),
                    cast(Int(32), device_id_),
                    cast(UInt(64), total_bytes),
                    IntImm::make(Int(32), op->type.code()),
                    IntImm::make(Int(32), op->type.bits())},
                   Call::Extern),
        body);
111 112 113 114 115 116 117 118

    Expr free_op = Call::make(Int(32),
                              "TVMBackendFreeWorkspace",
                              {cast(Int(32), device_type_),
                                    cast(Int(32), device_id_),
                                    op->buffer_var},
                              Call::Extern);
    Stmt free_stmt = IfThenElse::make(free_op != make_zero(Int(32)), throw_last_error);
119 120 121 122 123 124
    body = Block::make(alloca, free_stmt);
    body = AttrStmt::make(
        op->buffer_var, attr::storage_alignment,
        make_const(Int(32), runtime::kTempAllocaAlignment),
        body);
    return body;
125 126
  }

127 128 129 130 131 132 133 134 135 136 137 138 139 140
  Stmt Mutate_(const AttrStmt* op, const Stmt &s) final {
    if (op->attr_key == attr::device_context_id) {
      CHECK(!device_id_.defined());
      device_id_ = op->value;
      return Mutate(op->body);
    } else if (op->attr_key == attr::device_context_type) {
      CHECK(!device_type_.defined());
      device_type_ = op->value;
      return Mutate(op->body);
    } else {
      return IRMutator::Mutate_(op, s);
    }
  }
  Expr Mutate_(const Call* op, const Expr &e) final {
141
    if (op->is_intrinsic(intrinsic::tvm_call_packed)) {
142 143 144 145 146
      return MakeCallPacked(op, e);
    } else if (op->is_intrinsic(intrinsic::tvm_stack_make_shape)) {
      return MakeShape(op, e);
    } else if (op->is_intrinsic(intrinsic::tvm_stack_make_array)) {
      return MakeArray(op, e);
147 148
    } else if (op->is_intrinsic(intrinsic::tvm_context_id)) {
      return make_zero(op->type);
149 150 151 152 153 154 155 156 157 158 159 160
    } else {
      return IRMutator::Mutate_(op, e);
    }
  }
  // call shape
  Expr MakeShape(const Call* op, const Expr& e) {
    size_t stack_begin = run_shape_stack_;
    run_shape_stack_ += op->args.size();
    Expr expr = IRMutator::Mutate_(op, e);
    op = expr.as<Call>();
    for (size_t i = 0; i < op->args.size(); ++i) {
      prep_seq_.emplace_back(
161
          Store::make(stack_shape_, cast(Int(64), op->args[i]),
162
                      ConstInt32(stack_begin +i), const_true(1)));
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
    }
    return AddressOffset(stack_shape_, Int(64), stack_begin);
  }
  // make array
  Expr MakeArray(const Call* op, const Expr& e) {
    size_t idx = run_array_stack_;
    run_array_stack_ += 1;
    Expr expr = IRMutator::Mutate_(op, e);
    op = expr.as<Call>();
    prep_seq_.emplace_back(
        TVMStructSet(stack_array_, idx, intrinsic::kArrData, op->args[0]));
    prep_seq_.emplace_back(
        TVMStructSet(stack_array_, idx, intrinsic::kArrShape, op->args[1]));
    Expr strides = op->args[2];
    if (!strides.defined() || is_zero(strides)) {
      strides = make_zero(Handle());
    }
    prep_seq_.emplace_back(
        TVMStructSet(stack_array_, idx, intrinsic::kArrStrides, strides));
    prep_seq_.emplace_back(
        TVMStructSet(stack_array_, idx, intrinsic::kArrNDim, op->args[3]));
    Type dtype = op->args[4].type();
    prep_seq_.emplace_back(
        TVMStructSet(stack_array_, idx, intrinsic::kArrTypeCode,
                     make_const(UInt(8), static_cast<int>(dtype.code()))));
    prep_seq_.emplace_back(
        TVMStructSet(stack_array_, idx, intrinsic::kArrTypeBits,
                     make_const(UInt(8), dtype.bits())));
    prep_seq_.emplace_back(
        TVMStructSet(stack_array_, idx, intrinsic::kArrTypeLanes,
                     make_const(UInt(16), dtype.lanes())));
194 195 196 197 198 199
    // set byte offset
    int data_bytes = GetVectorBytes(dtype);
    Expr byte_offset = op->args[5];
    if (!is_zero(byte_offset)) {
      byte_offset = byte_offset * make_const(byte_offset.type(), data_bytes);
    }
200 201
    prep_seq_.emplace_back(
        TVMStructSet(stack_array_, idx, intrinsic::kArrByteOffset,
202
                     cast(UInt(64), byte_offset)));
203 204 205 206
    CHECK(device_type_.defined()) << "Unknown device type in current IR";
    CHECK(device_id_.defined()) << "Unknown device id in current IR";
    prep_seq_.emplace_back(
        TVMStructSet(stack_array_, idx, intrinsic::kArrDeviceId,
207
                     cast(Int(32), device_id_)));
208 209
    prep_seq_.emplace_back(
        TVMStructSet(stack_array_, idx, intrinsic::kArrDeviceType,
210
                     cast(Int(32), device_type_)));
211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237
    return TVMStructGet(Handle(), stack_array_, idx, intrinsic::kArrAddr);
  }
  // call packled.
  Expr MakeCallPacked(const Call* op, const Expr& e) {
    size_t restore_shape_stack = run_shape_stack_;
    size_t restore_array_stack = run_array_stack_;
    size_t arg_stack_begin = run_arg_stack_;
    run_arg_stack_ += op->args.size();
    // Specially handle the buffer packed intrinsic
    Expr expr = IRMutator::Mutate_(op, e);
    op = expr.as<Call>();
    for (size_t i = 1; i < op->args.size(); ++i) {
      Expr stack_index = ConstInt32(arg_stack_begin + i - 1);
      Expr arg = op->args[i];
      Type t = arg.type();
      Type api_type = APIType(t);
      if (t != api_type) {
        arg = Cast::make(api_type, arg);
      }
      prep_seq_.emplace_back(TVMStructSet(
          stack_value_, static_cast<int>(arg_stack_begin + i - 1),
          intrinsic::kTVMValueContent, arg));
      int arg_tcode = api_type.code();
      if (IsArrayHandle(arg)) arg_tcode = kArrayHandle;
      prep_seq_.emplace_back(
          Store::make(stack_tcode_,
                      ConstInt32(arg_tcode),
238
                      stack_index, const_true(1)));
239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289
    }
    // UPDATE stack value
    max_arg_stack_ = std::max(run_arg_stack_, max_arg_stack_);
    max_shape_stack_ = std::max(run_shape_stack_, max_shape_stack_);
    max_array_stack_ = std::max(run_array_stack_, max_array_stack_);
    run_shape_stack_ = restore_shape_stack;
    run_array_stack_ = restore_array_stack;
    run_arg_stack_ = arg_stack_begin;
    Array<Expr> packed_args = {
      op->args[0],
      stack_value_,
      stack_tcode_,
      ConstInt32(arg_stack_begin),
      ConstInt32(arg_stack_begin + op->args.size() - 1)
    };
    return Call::make(
        Int(32), intrinsic::tvm_call_packed_lowered,
        packed_args, Call::Intrinsic);
  }

 private:
  bool IsArrayHandle(const Expr& arg) {
    // specially set array handle.
    if (const Call* buf = arg.as<Call>()) {
      if (buf->is_intrinsic(intrinsic::tvm_struct_get) &&
          buf->args[2].as<IntImm>()->value == intrinsic::kArrAddr) {
        return true;
      }
    }
    return false;
  }

  // The prepration sequence to be emitted.
  std::vector<Stmt> prep_seq_;
  Expr device_type_;
  Expr device_id_;
  // Var handle for each stack.
  Var stack_shape_;
  Var stack_array_;
  Var stack_tcode_;
  Var stack_value_;
  // The running statistics
  uint64_t run_shape_stack_{0};
  uint64_t run_array_stack_{0};
  uint64_t run_arg_stack_{0};
  // statistics of stacks
  uint64_t max_shape_stack_{0};
  uint64_t max_array_stack_{0};
  uint64_t max_arg_stack_{0};
};

290
LoweredFunc LowerTVMBuiltin(LoweredFunc f) {
291
  auto n = std::make_shared<LoweredFuncNode>(*f.operator->());
292
  n->body = BuiltinLower().Build(n->body);
293 294 295 296 297
  return LoweredFunc(n);
}

}  // namespace ir
}  // namespace tvm