tvm
op.h
Go to the documentation of this file.
1 /*
2  * Licensed to the Apache Software Foundation (ASF) under one
3  * or more contributor license agreements. See the NOTICE file
4  * distributed with this work for additional information
5  * regarding copyright ownership. The ASF licenses this file
6  * to you under the Apache License, Version 2.0 (the
7  * "License"); you may not use this file except in compliance
8  * with the License. You may obtain a copy of the License at
9  *
10  * http://www.apache.org/licenses/LICENSE-2.0
11  *
12  * Unless required by applicable law or agreed to in writing,
13  * software distributed under the License is distributed on an
14  * "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
15  * KIND, either express or implied. See the License for the
16  * specific language governing permissions and limitations
17  * under the License.
18  */
19 
27 // Acknowledgement: Most operator APIs originate from Halide.
28 #ifndef TVM_TIR_OP_H_
29 #define TVM_TIR_OP_H_
30 
31 #include <tvm/ir/expr.h>
32 #include <tvm/ir/op.h>
33 #include <tvm/ir/type.h>
34 #include <tvm/tirx/builtin.h>
35 #include <tvm/tirx/expr.h>
36 #include <tvm/tirx/stmt.h>
37 
38 #include <algorithm>
39 #include <limits>
40 #include <type_traits>
41 
42 namespace tvm {
43 
44 #define TVM_TIR_REGISTER_OP(OpName) \
45  TVM_REGISTER_OP("tirx." OpName).set_attr<TScriptPrinterName>("TScriptPrinterName", OpName)
46 
47 // Most common operators can be overloaded by argument type(PrimExpr).
48 // So we put them under the root namespace.
49 //
50 // We put more developer oriented APIs -- make_const and is_const under tirx
51 // as they are more specific to the tirx namespace.
52 
64 TVM_DLL Type GetType(const PrimExpr& expr);
65 
74 
84 
92 TVM_DLL PrimExpr ret(PrimExpr value, Span span = Span());
93 
100 TVM_DLL PrimExpr thread_return(Span span = Span());
101 
107 TVM_DLL PrimExpr continue_loop(Span span = Span());
108 
114 TVM_DLL PrimExpr break_loop(Span span = Span());
115 
122 TVM_DLL PrimExpr max_value(const DataType& dtype, Span span = Span());
123 
130 TVM_DLL PrimExpr min_value(const DataType& dtype, Span span = Span());
131 
138 TVM_DLL PrimExpr infinity(const DataType& dtype, Span span = Span());
139 
149 TVM_DLL PrimExpr cast(const DataType& t, PrimExpr value, Span span = Span());
159 TVM_DLL PrimExpr reinterpret(const DataType& t, PrimExpr value, Span span = Span());
170 TVM_DLL PrimExpr add(PrimExpr a, PrimExpr b, Span span = Span());
181 TVM_DLL PrimExpr sub(PrimExpr a, PrimExpr b, Span span = Span());
191 TVM_DLL PrimExpr neg(PrimExpr a, Span span = Span());
202 TVM_DLL PrimExpr mul(PrimExpr a, PrimExpr b, Span span = Span());
213 TVM_DLL PrimExpr left_shift(PrimExpr a, PrimExpr b, Span span = Span());
224 TVM_DLL PrimExpr right_shift(PrimExpr a, PrimExpr b, Span span = Span());
235 TVM_DLL PrimExpr greater(PrimExpr a, PrimExpr b, Span span = Span());
257 TVM_DLL PrimExpr less(PrimExpr a, PrimExpr b, Span span = Span());
268 TVM_DLL PrimExpr less_equal(PrimExpr a, PrimExpr b, Span span = Span());
279 TVM_DLL PrimExpr equal(PrimExpr a, PrimExpr b, Span span = Span());
290 TVM_DLL PrimExpr not_equal(PrimExpr a, PrimExpr b, Span span = Span());
300 TVM_DLL PrimExpr logical_and(PrimExpr a, PrimExpr b, Span span = Span());
310 TVM_DLL PrimExpr logical_or(PrimExpr a, PrimExpr b, Span span = Span());
319 TVM_DLL PrimExpr logical_not(PrimExpr a, Span span = Span());
334 TVM_DLL PrimExpr div(PrimExpr a, PrimExpr b, Span span = Span());
347 TVM_DLL PrimExpr truncdiv(PrimExpr a, PrimExpr b, Span span = Span());
360 TVM_DLL PrimExpr truncmod(PrimExpr a, PrimExpr b, Span span = Span());
376 TVM_DLL PrimExpr indexdiv(PrimExpr a, PrimExpr b, Span span = Span());
392 TVM_DLL PrimExpr shapediv(PrimExpr a, PrimExpr b, Span span = Span());
407 TVM_DLL PrimExpr indexmod(PrimExpr a, PrimExpr b, Span span = Span());
418 TVM_DLL PrimExpr floordiv(PrimExpr a, PrimExpr b, Span span = Span());
427 TVM_DLL PrimExpr logaddexp(PrimExpr a, PrimExpr b, Span span = Span());
439 TVM_DLL PrimExpr ceildiv(PrimExpr a, PrimExpr b, Span span = Span());
450 TVM_DLL PrimExpr floormod(PrimExpr a, PrimExpr b, Span span = Span());
461 TVM_DLL PrimExpr max(PrimExpr a, PrimExpr b, Span span = Span());
472 TVM_DLL PrimExpr min(PrimExpr a, PrimExpr b, Span span = Span());
483 TVM_DLL PrimExpr bitwise_and(PrimExpr a, PrimExpr b, Span span = Span());
494 TVM_DLL PrimExpr bitwise_or(PrimExpr a, PrimExpr b, Span span = Span());
505 TVM_DLL PrimExpr bitwise_xor(PrimExpr a, PrimExpr b, Span span = Span());
515 TVM_DLL PrimExpr bitwise_neg(PrimExpr a, Span span = Span());
527 TVM_DLL PrimExpr if_then_else(PrimExpr cond, PrimExpr true_value, PrimExpr false_value,
528  Span span = Span());
535 TVM_DLL PrimExpr likely(PrimExpr cond, Span span = Span());
542 TVM_DLL PrimExpr pow(PrimExpr x, PrimExpr y, Span span = Span());
550 TVM_DLL PrimExpr abs(PrimExpr x, Span span = Span());
557 TVM_DLL PrimExpr isnan(PrimExpr x, Span span = Span());
558 
565 TVM_DLL PrimExpr isfinite(PrimExpr x, Span span = Span());
566 
573 TVM_DLL PrimExpr isinf(PrimExpr x, Span span = Span());
574 
583 TVM_DLL PrimExpr sum(PrimExpr source, ffi::Array<tirx::IterVar> axis,
584  ffi::Array<PrimExpr> init = {}, Span span = Span());
585 
593 TVM_DLL PrimExpr all(PrimExpr source, ffi::Array<tirx::IterVar> axis,
594  ffi::Array<PrimExpr> init = {}, Span span = Span());
595 
604 TVM_DLL PrimExpr any(PrimExpr source, ffi::Array<tirx::IterVar> axis,
605  ffi::Array<PrimExpr> init = {}, Span span = Span());
606 
615 TVM_DLL PrimExpr max(PrimExpr source, ffi::Array<tirx::IterVar> axis,
616  ffi::Array<PrimExpr> init = {}, Span span = Span());
617 
626 TVM_DLL PrimExpr min(PrimExpr source, ffi::Array<tirx::IterVar> axis,
627  ffi::Array<PrimExpr> init = {}, Span span = Span());
628 
637 TVM_DLL PrimExpr prod(PrimExpr source, ffi::Array<tirx::IterVar> axis,
638  ffi::Array<PrimExpr> init = {}, Span span = Span());
639 
646 TVM_DLL PrimExpr floor(PrimExpr x, Span span = Span());
647 
654 TVM_DLL PrimExpr ceil(PrimExpr x, Span span = Span());
655 
666 TVM_DLL PrimExpr round(PrimExpr x, Span span = Span());
667 
677 TVM_DLL PrimExpr nearbyint(PrimExpr x, Span span = Span());
678 
685 TVM_DLL PrimExpr trunc(PrimExpr x, Span span = Span());
686 
695 TVM_DLL PrimExpr LargeUIntImm(DataType dtype, int64_t low, int64_t high, Span span = Span());
696 
718  Span span = Span());
719 
727 TVM_DLL PrimExpr fast_erf_float_expr(PrimExpr arg, int bits);
728 
729 inline void CheckMathUnaryOpInputDType(const char* op_name, DataType dtype) {
730  TVM_FFI_CHECK(dtype.is_float() || dtype.is_bfloat16(), TypeError)
731  << "tirx." << op_name << " only supports floating-point inputs, but got " << dtype;
732 }
733 
734 // Intrinsic operators
735 #define TVM_DECLARE_INTRIN_UNARY_WITH_CHECK(OpName, CheckInputDType) \
736  inline PrimExpr OpName(PrimExpr x, Span span = Span()) { \
737  static const Op& op = Op::Get("tirx." #OpName); \
738  CheckInputDType(#OpName, x.dtype()); \
739  if (x.dtype().is_bfloat16()) { \
740  DataType bf16_dtype = x.dtype(); \
741  DataType fp32_dtype(kDLFloat, 32, bf16_dtype.lanes()); \
742  PrimExpr x_fp32 = tirx::Cast(fp32_dtype, {x}, span); \
743  PrimExpr result_fp32 = tirx::Call(fp32_dtype, op, {x_fp32}, span); \
744  return tirx::Cast(bf16_dtype, {result_fp32}, span); \
745  } else { \
746  return tirx::Call(x.dtype(), op, {x}, span); \
747  } \
748  }
749 
750 #define TVM_DECLARE_INTRIN_UNARY(OpName) \
751  TVM_DECLARE_INTRIN_UNARY_WITH_CHECK(OpName, [](const char*, DataType) {})
752 
753 #define TVM_DECLARE_FLOAT_INTRIN_UNARY(OpName) \
754  TVM_DECLARE_INTRIN_UNARY_WITH_CHECK(OpName, CheckMathUnaryOpInputDType)
755 
781 
782 #define TVM_DECLARE_INTRIN_BINARY(OpName) \
783  inline PrimExpr OpName(PrimExpr x, PrimExpr y, Span span = Span()) { \
784  static const Op& op = Op::Get("tirx." #OpName); \
785  return tirx::Call(x.dtype(), op, {x, y}, span); \
786  }
787 
793 
794 namespace tirx {
795 
802 inline bool IsPointerType(const Type& type, const DataType& element_type) {
803  if (!type.defined()) return false;
804  if (const auto* ptr_type = type.as<PointerTypeNode>()) {
805  if (const auto* prim_type = ptr_type->element_type.as<PrimTypeNode>()) {
806  return prim_type->dtype == element_type;
807  }
808  }
809  return false;
810 }
811 
820 template <typename ValueType,
821  typename = typename std::enable_if<std::is_pod<ValueType>::value>::type>
822 inline PrimExpr make_const(DataType t, ValueType value, Span span = Span());
829 inline PrimExpr make_zero(DataType t, Span span = Span());
836 inline PrimExpr const_true(int lanes = 1, Span span = Span()) {
837  return make_const(DataType::Bool(lanes), 1);
838 }
845 inline PrimExpr const_false(int lanes = 1, Span span = Span()) {
846  return make_const(DataType::Bool(lanes), 0);
847 }
854 inline const int64_t* as_const_int(const PrimExpr& x) {
855  if (!x.defined()) return nullptr;
856  if (const tirx::IntImmNode* op = x.as<tirx::IntImmNode>()) {
857  return &(op->value);
858  }
859 
860  return nullptr;
861 }
862 
869 inline bool is_const_int(const PrimExpr& x, int64_t value);
870 
876 inline bool is_no_op(const tirx::Stmt& stmt);
877 
884 inline bool is_one(const PrimExpr& x) { return is_const_int(x, 1); }
885 
892 inline bool is_zero(const PrimExpr& x) { return is_const_int(x, 0); }
893 
899 inline bool is_const_int(const PrimExpr& x);
900 
906 inline bool is_const_number(const PrimExpr& x);
907 
917 template <typename FReduce>
918 inline PrimExpr foldl(FReduce freduce, PrimExpr init_value, const ffi::Array<PrimExpr>& values,
919  Span span = Span()) {
920  for (PrimExpr val : values) {
921  init_value = freduce(init_value, val, span);
922  }
923  return init_value;
924 }
925 
934 TVM_DLL bool is_const_power_of_two_integer(const PrimExpr& x, int* shift);
935 
936 // Implementation details after this
937 inline bool is_const_int(const PrimExpr& x) { return as_const_int(x); }
938 
939 inline bool is_const_number(const PrimExpr& x) {
940  if (x.as<tirx::IntImmNode>()) {
941  return true;
942  } else if (x.as<tirx::FloatImmNode>()) {
943  return true;
944  } else if (const auto* op = x.as<tirx::BroadcastNode>()) {
945  return (op->value->IsInstance<tirx::IntImmNode>() ||
946  op->value->IsInstance<tirx::FloatImmNode>());
947  }
948  return false;
949 }
950 
951 inline bool is_positive_const(const PrimExpr& a) {
952  const int64_t* as_int = as_const_int(a);
953  return as_int && (*as_int > 0);
954 }
955 
956 inline bool is_negative_const(const PrimExpr& a) {
957  const int64_t* as_int = as_const_int(a);
958  return as_int && (*as_int < 0);
959 }
960 
961 inline bool is_const_int(const PrimExpr& x, int64_t value) {
962  const int64_t* as_int = as_const_int(x);
963  return as_int && (*as_int == value);
964 }
965 
966 inline bool is_no_op(const tirx::Stmt& stmt) {
967  if (!stmt.defined()) return true;
968  if (const auto* op = stmt.as<tirx::EvaluateNode>()) {
969  return is_const_int(op->value);
970  }
971  if (const auto* op = stmt.as<tirx::SeqStmtNode>()) {
972  return op->seq.size() == 0;
973  }
974  return false;
975 }
976 
977 template <typename ValueType>
978 inline PrimExpr MakeConstScalar(DataType t, ValueType value, Span span = Span()) {
979  if (t.is_int() || t.is_bool()) return IntImm(t, static_cast<int64_t>(value), span);
980  if (t.is_uint()) {
981  // Use IntImm if it is a small integer
982  uint64_t uval = static_cast<uint64_t>(value);
983  if (value < static_cast<ValueType>(0)) {
984  TVM_FFI_THROW(InternalError) << "cannot make uint from negative value " << value;
985  } else if (uval <= static_cast<uint64_t>(std::numeric_limits<int64_t>::max())) {
986  return IntImm(t, static_cast<int64_t>(value), span);
987  } else {
988  uint64_t mask = (static_cast<uint64_t>(1) << 32U) - 1U;
989  uint64_t low = uval & mask;
990  uint64_t high = uval >> 32U;
991  return LargeUIntImm(t, static_cast<int64_t>(low), static_cast<int64_t>(high), span);
992  }
993  }
994  if (t.is_float() || t.is_bfloat16() || t.is_float8() || t.is_float6() || t.is_float4())
995  return FloatImm(t, static_cast<double>(value), span);
996  // For now, we store const scalar values of custom datatypes within doubles; later, during the
997  // datatypes lowering pass, we will lower the value to its true representation in the format
998  // specified by the datatype.
999  // TODO(gus) when do we need to start worrying about doubles not being precise enough?
1000  if (static_cast<uint8_t>(t.code()) >= static_cast<uint8_t>(DataType::kCustomBegin)) {
1001  return FloatImm(t, static_cast<double>(value), span);
1002  }
1003  TVM_FFI_THROW(InternalError) << "cannot make const for type " << t;
1004  throw;
1005 }
1006 
1007 template <>
1008 inline PrimExpr MakeConstScalar(DataType t, bool value, Span span) {
1009  return MakeConstScalar(t, static_cast<int>(value), span);
1010 }
1011 
1012 template <typename ValueType, typename>
1013 inline PrimExpr make_const(DataType t, ValueType value, Span span) {
1014  if (t.is_scalar()) {
1015  return MakeConstScalar(t, value, span);
1016  } else {
1017  if (t.is_fixed_length_vector()) {
1018  return tirx::Broadcast(MakeConstScalar(t.element_of(), value, span), t.lanes(), span);
1019  } else {
1020  PrimExpr lanes =
1022  return tirx::Broadcast(MakeConstScalar(t.element_of(), value, span), lanes, span);
1023  }
1024  }
1025 }
1026 
1027 inline PrimExpr make_zero(DataType t, Span span) {
1028  if (t.is_handle()) {
1029  return reinterpret(t, make_const(DataType::UInt(64), 0, span));
1030  }
1031  return make_const(t, 0, span);
1032 }
1033 
1034 } // namespace tirx
1035 
1036 // additional const expression overloading
1037 #define TVM_DEFINE_ASSIGN_OP_OVERLOAD(Name, OpFunc) \
1038  inline PrimExpr Name(PrimExpr& a, PrimExpr b) { \
1039  a = OpFunc(a, b); \
1040  return a; \
1041  }
1042 
1043 #define TVM_DEFINE_BINOP_CONST_VAL_OVERLOAD(Name) \
1044  inline PrimExpr Name(const PrimExpr& a, float b) { return Name(a, PrimExpr(b)); } \
1045  inline PrimExpr Name(float a, const PrimExpr& b) { return Name(PrimExpr(a), b); } \
1046  inline PrimExpr Name(int a, const PrimExpr& b) { \
1047  return Name(tirx::make_const(b.dtype(), a), b); \
1048  } \
1049  inline PrimExpr Name(const PrimExpr& a, int b) { \
1050  return Name(a, tirx::make_const(a.dtype(), b)); \
1051  } \
1052  inline PrimExpr Name(const PrimExpr& a, double b) { \
1053  return Name(a, tirx::make_const(DataType::Float(64), b)); \
1054  }
1055 
1056 #define TVM_DEFINE_BINOP_CONST_VAL_OVERLOAD_SPANNED(Name) \
1057  inline PrimExpr Name(const PrimExpr& a, float b, Span span = Span()) { \
1058  return Name(a, PrimExpr(b), span); \
1059  } \
1060  inline PrimExpr Name(float a, const PrimExpr& b, Span span = Span()) { \
1061  return Name(PrimExpr(a), b, span); \
1062  } \
1063  inline PrimExpr Name(int a, const PrimExpr& b, Span span = Span()) { \
1064  return Name(tirx::make_const(b.dtype(), a), b, span); \
1065  } \
1066  inline PrimExpr Name(const PrimExpr& a, int b, Span span = Span()) { \
1067  return Name(a, tirx::make_const(a.dtype(), b), span); \
1068  } \
1069  inline PrimExpr Name(const PrimExpr& a, double b, Span span = Span()) { \
1070  return Name(a, tirx::make_const(DataType::Float(64), b), span); \
1071  }
1072 
1073 #define TVM_DEFINE_LOGICAL_OP_CONST_VAL_OVERLOAD(Name) \
1074  inline PrimExpr Name(const PrimExpr& a, bool b) { return Name(a, PrimExpr(b)); } \
1075  inline PrimExpr Name(bool a, const PrimExpr& b) { return Name(PrimExpr(a), b); }
1076 
1077 #define TVM_DEFINE_LOGICAL_OP_CONST_VAL_OVERLOAD_SPANNED(Name) \
1078  inline PrimExpr Name(const PrimExpr& a, bool b, Span span = Span()) { \
1079  return Name(a, PrimExpr(b), span); \
1080  } \
1081  inline PrimExpr Name(bool a, const PrimExpr& b, Span span = Span()) { \
1082  return Name(PrimExpr(a), b, span); \
1083  }
1084 
1085 #define TVM_DEFINE_INT_OP_CONST_VAL_OVERLOAD(Name) \
1086  inline PrimExpr Name(const PrimExpr& a, int b) { \
1087  return Name(a, tirx::make_const(a.dtype(), b)); \
1088  } \
1089  inline PrimExpr Name(int a, const PrimExpr& b) { return Name(tirx::make_const(b.dtype(), a), b); }
1090 
1091 #define TVM_DEFINE_INT_OP_CONST_VAL_OVERLOAD_SPANNED(Name) \
1092  inline PrimExpr Name(const PrimExpr& a, int b, Span span = Span()) { \
1093  return Name(a, tirx::make_const(a.dtype(), b), span); \
1094  } \
1095  inline PrimExpr Name(int a, const PrimExpr& b, Span span = Span()) { \
1096  return Name(tirx::make_const(b.dtype(), a), b, span); \
1097  }
1098 
1099 TVM_DEFINE_ASSIGN_OP_OVERLOAD(operator+=, operator+);
1100 TVM_DEFINE_ASSIGN_OP_OVERLOAD(operator-=, operator-);
1101 TVM_DEFINE_ASSIGN_OP_OVERLOAD(operator*=, operator*);
1105 TVM_DEFINE_BINOP_CONST_VAL_OVERLOAD(operator>); // NOLINT(*)
1107 TVM_DEFINE_BINOP_CONST_VAL_OVERLOAD(operator<); // NOLINT(*)
1119 // integer related ops
1132 TVM_DEFINE_INT_OP_CONST_VAL_OVERLOAD(operator>>); // NOLINT(*)
1133 TVM_DEFINE_INT_OP_CONST_VAL_OVERLOAD(operator<<); // NOLINT(*)
1137 // logical ops
1142 
1148 template <typename TA>
1149 inline void DivAmbiguityError(const TA& a) {
1150  constexpr bool div_ambiguity = !std::is_class<TA>::value;
1151  static_assert(div_ambiguity,
1152  "TVM supports multiple types of integer divisions, "
1153  "please call div, indexdiv/indexmod, "
1154  "floordiv/floormod or truncdiv/truncmod directly "
1155  "to avoid ambiguity in the code. "
1156  "Checkout these functions in tirx/op.h.");
1157 }
1158 
1159 // The following code are not intended to be used in the codebase.
1160 // Instead, they generate clear compiler errors that ask developers
1161 // to use the specific division function.
1162 // The second template argument is necessary to make sure the
1163 // code compiles lazily by the compiler during invocation.
1164 template <typename TB>
1165 inline PrimExpr operator/(const PrimExpr& a, const TB& b) {
1166  DivAmbiguityError(a);
1167  return a;
1168 }
1169 
1170 template <typename TB>
1171 inline PrimExpr operator/=(const PrimExpr& a, const TB& b) {
1172  DivAmbiguityError(a);
1173  return a;
1174 }
1175 
1176 template <typename TB>
1177 inline PrimExpr operator%(const PrimExpr& a, const TB& b) {
1178  DivAmbiguityError(a);
1179  return a;
1180 }
1181 } // namespace tvm
1182 #endif // TVM_TIR_OP_H_
Constant floating point literals in the program.
Definition: expr.h:529
Managed reference class to FloatImmNode.
Definition: expr.h:546
Constant integer literals in the program.
Definition: expr.h:494
int64_t value
the Internal value.
Definition: expr.h:497
Managed reference class to IntImmNode.
Definition: expr.h:511
Low-level raw pointer type.
Definition: type.h:152
Reference to PrimExprNode.
Definition: expr.h:126
Primitive data types used in the low-level IR.
Definition: type.h:112
Definition: source_map.h:111
Managed reference to TypeNode.
Definition: type.h:99
Runtime primitive data type.
Definition: data_type.h:47
bool is_handle() const
Definition: data_type.h:198
bool is_uint() const
Definition: data_type.h:196
bool is_float6() const
Definition: data_type.h:159
DataType element_of() const
Get the scalar version of the type.
Definition: data_type.h:240
@ kCustomBegin
Definition: data_type.h:75
bool is_bool() const
Definition: data_type.h:143
bool is_int() const
Definition: data_type.h:194
int code() const
Definition: data_type.h:114
int lanes() const
Definition: data_type.h:120
static DataType Bool(int lanes=1, bool is_scalable=false)
Construct a bool type.
Definition: data_type.h:387
int vscale_factor() const
Definition: data_type.h:129
bool is_fixed_length_vector() const
Definition: data_type.h:205
static DataType Int(int bits, int lanes=1)
Construct an int type.
Definition: data_type.h:278
bool is_scalar() const
Definition: data_type.h:141
bool is_float8() const
Definition: data_type.h:151
bool is_bfloat16() const
Definition: data_type.h:192
bool is_float4() const
Definition: data_type.h:164
static DataType UInt(int bits, int lanes=1, bool is_scalable=false)
Construct an uint type.
Definition: data_type.h:286
bool is_float() const
Definition: data_type.h:147
Create a vector where all the elements are value.
Definition: expr.h:657
Managed reference to BroadcastNode.
Definition: expr.h:677
Managed reference to CallNode.
Definition: expr.h:744
Evaluates an expression. This is mostly used for putting a Call node into Stmt.
Definition: stmt.h:336
Managed reference to MulNode.
Definition: expr.h:169
The container of seq statement. Represent a sequence of statements.
Definition: stmt.h:311
Container of all statements.
Definition: stmt.h:65
Base expr nodes in TVM.
Primitive operators(builtin intrinsics) and registry for them.
IR/AST nodes for the unified type system in TVM.
const Op & vscale()
Get the target's vscale value. It will be lowered to llvm.vscale intrinsic (https://llvm....
bool is_const_number(const PrimExpr &x)
Check whether x is an integer/float constant.
Definition: op.h:939
bool is_zero(const PrimExpr &x)
Check whether x is a constant integer 0.
Definition: op.h:892
bool is_const_power_of_two_integer(const PrimExpr &x, int *shift)
Check whether x is a constant power of two If x is power of two, write the power to the shift.
bool IsPointerType(const Type &type, const DataType &element_type)
Check if type is a pointer to a runtime element type.
Definition: op.h:802
bool is_positive_const(const PrimExpr &a)
Definition: op.h:951
PrimExpr make_const(DataType t, ValueType value, Span span=Span())
Make a const value with certain data type.
Definition: op.h:1013
PrimExpr MakeConstScalar(DataType t, ValueType value, Span span=Span())
Definition: op.h:978
PrimExpr const_false(int lanes=1, Span span=Span())
Make a constant false expression.
Definition: op.h:845
PrimExpr make_zero(DataType t, Span span=Span())
Make a const zero expr.
Definition: op.h:1027
PrimExpr const_true(int lanes=1, Span span=Span())
Make a constant true expression.
Definition: op.h:836
bool is_negative_const(const PrimExpr &a)
Definition: op.h:956
bool is_const_int(const PrimExpr &x, int64_t value)
Check whether x is a constant integer expression.
Definition: op.h:961
bool is_one(const PrimExpr &x)
Check whether x is a constant integer 1.
Definition: op.h:884
bool is_no_op(const tirx::Stmt &stmt)
Check whether stmt is nop.
Definition: op.h:966
PrimExpr foldl(FReduce freduce, PrimExpr init_value, const ffi::Array< PrimExpr > &values, Span span=Span())
Left fold.
Definition: op.h:918
const int64_t * as_const_int(const PrimExpr &x)
Get x as constant int expression.
Definition: op.h:854
std::function< PrimExpr(PrimExpr source, const ffi::Array< IterVar > &axis, ffi::Array< PrimExpr > init, Span span)> FReduce
The operation to use for CommReduce.
Definition: reduction.h:47
An object that builds and maintains block scope and StmtSref mapping for Dependence analysis.
Definition: analyzer.h:37
runtime::DataType GetRuntimeDataType(const Type &type)
Get the implied DataType for storing values with type during runtime.
PrimExpr isfinite(PrimExpr x, Span span=Span())
Check if x is finite.
PrimExpr ceildiv(PrimExpr a, PrimExpr b, Span span=Span())
compute ceil(a / b)
PrimExpr ret(PrimExpr value, Span span=Span())
Return the value.
PrimExpr max(PrimExpr a, PrimExpr b, Span span=Span())
take maximum of two values
PrimExpr tanh(PrimExpr x, Span span=Span())
Definition: op.h:760
PrimExpr erf(PrimExpr x, Span span=Span())
Definition: op.h:759
PrimExpr shapediv(PrimExpr a, PrimExpr b, Span span=Span())
compute ceil(a / b) where a and b are non-negative.
PrimExpr log10(PrimExpr x, Span span=Span())
Definition: op.h:766
PrimExpr div(PrimExpr a, PrimExpr b, Span span=Span())
compute division in C semantics.
PrimExpr operator/(PrimExpr a, PrimExpr b)
division operator
PrimExpr equal(PrimExpr a, PrimExpr b, Span span=Span())
equal
PrimExpr truncmod(PrimExpr a, PrimExpr b, Span span=Span())
compute the remainder of truncdiv
PrimExpr logical_and(PrimExpr a, PrimExpr b, Span span=Span())
and
PrimExpr hypot(PrimExpr x, PrimExpr y, Span span=Span())
Definition: op.h:791
PrimExpr log1p(PrimExpr x, Span span=Span())
Definition: op.h:767
void DivAmbiguityError(const TA &a)
Helper function to raise a compiler error about division ambiguity.
Definition: op.h:1149
PrimExpr likely(PrimExpr cond, Span span=Span())
Mark condition as likely.
PrimExpr reinterpret(const DataType &t, PrimExpr value, Span span=Span())
perform reinterpret cast value to type.
PrimExpr atan2(PrimExpr x, PrimExpr y, Span span=Span())
Definition: op.h:788
PrimExpr if_then_else(PrimExpr cond, PrimExpr true_value, PrimExpr false_value, Span span=Span())
Conditional expression.
PrimExpr min_value(const DataType &dtype, Span span=Span())
PrimExpr bitwise_neg(PrimExpr a, Span span=Span())
take bitwise negation of two values
PrimExpr cosh(PrimExpr x, Span span=Span())
Definition: op.h:771
PrimExpr logical_or(PrimExpr a, PrimExpr b, Span span=Span())
or
PrimExpr thread_return(Span span=Span())
Return from a thread.
PrimExpr atan(PrimExpr x, Span span=Span())
Definition: op.h:776
Type GetType(const PrimExpr &expr)
Get the type of the expression under the unified type system.
PrimExpr isnan(PrimExpr x, Span span=Span())
Check if x is NaN.
PrimExpr cast(const DataType &t, PrimExpr value, Span span=Span())
cast value to type.
PrimExpr max_value(const DataType &dtype, Span span=Span())
PrimExpr exp2(PrimExpr x, Span span=Span())
Definition: op.h:757
PrimExpr rsqrt(PrimExpr x, Span span=Span())
Definition: op.h:763
PrimExpr operator/=(const PrimExpr &a, const TB &b)
Definition: op.h:1171
PrimExpr asinh(PrimExpr x, Span span=Span())
Definition: op.h:778
PrimExpr less(PrimExpr a, PrimExpr b, Span span=Span())
less
PrimExpr sin(PrimExpr x, Span span=Span())
Definition: op.h:772
PrimExpr trunc(PrimExpr x, Span span=Span())
Calculate trunc(x)
PrimExpr round(PrimExpr x, Span span=Span())
Round x to the nearest integer, ties to even.
Type GetTypeFromRuntimeDataType(const DataType &dtype)
Get the type corresponding to DataType.
PrimExpr neg(PrimExpr a, Span span=Span())
negation.
PrimExpr ceil(PrimExpr x, Span span=Span())
Calculate ceil(x)
PrimExpr pow(PrimExpr x, PrimExpr y, Span span=Span())
Calculate power(x, y)
PrimExpr logical_not(PrimExpr a, Span span=Span())
not
PrimExpr exp10(PrimExpr x, Span span=Span())
Definition: op.h:758
PrimExpr copysign(PrimExpr x, PrimExpr y, Span span=Span())
Definition: op.h:790
PrimExpr bitwise_xor(PrimExpr a, PrimExpr b, Span span=Span())
take bitwise xor of two values
PrimExpr less_equal(PrimExpr a, PrimExpr b, Span span=Span())
less_equal
PrimExpr any(PrimExpr source, ffi::Array< tirx::IterVar > axis, ffi::Array< PrimExpr > init={}, Span span=Span())
logical Or of source expression over axis
PrimExpr greater(PrimExpr a, PrimExpr b, Span span=Span())
greater
PrimExpr exp(PrimExpr x, Span span=Span())
Definition: op.h:756
PrimExpr logaddexp(PrimExpr a, PrimExpr b, Span span=Span())
Compute log(exp(a) + exp(b)).
PrimExpr floormod(PrimExpr a, PrimExpr b, Span span=Span())
compute the remainder of floordiv
PrimExpr infinity(const DataType &dtype, Span span=Span())
PrimExpr sub(PrimExpr a, PrimExpr b, Span span=Span())
subtraction operator
PrimExpr all(PrimExpr source, ffi::Array< tirx::IterVar > axis, ffi::Array< PrimExpr > init={}, Span span=Span())
logical And of source expression over axis
PrimExpr indexdiv(PrimExpr a, PrimExpr b, Span span=Span())
compute floor(a / b) where a and b are non-negative.
PrimExpr nextafter(PrimExpr x, PrimExpr y, Span span=Span())
Definition: op.h:789
PrimExpr LargeUIntImm(DataType dtype, int64_t low, int64_t high, Span span=Span())
Construct a large uint constant by its low 32 bits and high 32bits.
PrimExpr asin(PrimExpr x, Span span=Span())
Definition: op.h:774
PrimExpr prod(PrimExpr source, ffi::Array< tirx::IterVar > axis, ffi::Array< PrimExpr > init={}, Span span=Span())
product of source expression over axis
PrimExpr sigmoid(PrimExpr x, Span span=Span())
Definition: op.h:761
PrimExpr max(const PrimExpr &a, double b, Span span=Span())
Definition: op.h:1109
PrimExpr acos(PrimExpr x, Span span=Span())
Definition: op.h:775
PrimExpr mul(PrimExpr a, PrimExpr b, Span span=Span())
multiplication operator
PrimExpr min(PrimExpr a, PrimExpr b, Span span=Span())
take minimum of two values
void CheckMathUnaryOpInputDType(const char *op_name, DataType dtype)
Definition: op.h:729
PrimExpr sum(PrimExpr source, ffi::Array< tirx::IterVar > axis, ffi::Array< PrimExpr > init={}, Span span=Span())
sum of source expression over axis
PrimExpr floor(PrimExpr x, Span span=Span())
Calculate floor(x)
PrimExpr greater_equal(PrimExpr a, PrimExpr b, Span span=Span())
greater_equal
PrimExpr operator%(const PrimExpr &a, const TB &b)
Definition: op.h:1177
PrimExpr abs(PrimExpr x, Span span=Span())
Calculate absolute value of x.
PrimExpr atanh(PrimExpr x, Span span=Span())
Definition: op.h:779
PrimExpr sqrt(PrimExpr x, Span span=Span())
Definition: op.h:762
PrimExpr isinf(PrimExpr x, Span span=Span())
Check if x is infinite.
PrimExpr continue_loop(Span span=Span())
Continue current loop.
PrimExpr log2(PrimExpr x, Span span=Span())
Definition: op.h:765
PrimExpr not_equal(PrimExpr a, PrimExpr b, Span span=Span())
not_equal
PrimExpr ldexp(PrimExpr x, PrimExpr y, Span span=Span())
Definition: op.h:792
PrimExpr truncdiv(PrimExpr a, PrimExpr b, Span span=Span())
compute trunc(a / b)
PrimExpr q_multiply_shift(PrimExpr x, PrimExpr y, PrimExpr q, PrimExpr s, Span span=Span())
Execute a multiplication between two Q-numbers x and y followed by a right shift s....
PrimExpr popcount(PrimExpr x, Span span=Span())
Definition: op.h:768
PrimExpr bitwise_and(PrimExpr a, PrimExpr b, Span span=Span())
take bitwise and of two values
PrimExpr left_shift(PrimExpr a, PrimExpr b, Span span=Span())
left shift operator
PrimExpr sinh(PrimExpr x, Span span=Span())
Definition: op.h:773
PrimExpr indexmod(PrimExpr a, PrimExpr b, Span span=Span())
compute the remainder floor(a / b) where a and b are non-negative.
PrimExpr break_loop(Span span=Span())
Break current loop.
PrimExpr add(PrimExpr a, PrimExpr b, Span span=Span())
add operator
PrimExpr log(PrimExpr x, Span span=Span())
Definition: op.h:764
PrimExpr nearbyint(PrimExpr x, Span span=Span())
Round x to the nearest integer, ties to even.
PrimExpr right_shift(PrimExpr a, PrimExpr b, Span span=Span())
right shift operator
PrimExpr bitwise_or(PrimExpr a, PrimExpr b, Span span=Span())
take bitwise or of two values
PrimExpr clz(PrimExpr x, Span span=Span())
Definition: op.h:780
PrimExpr floordiv(PrimExpr a, PrimExpr b, Span span=Span())
compute floor(a / b)
PrimExpr acosh(PrimExpr x, Span span=Span())
Definition: op.h:777
PrimExpr tan(PrimExpr x, Span span=Span())
Definition: op.h:769
PrimExpr cos(PrimExpr x, Span span=Span())
Definition: op.h:770
PrimExpr fast_erf_float_expr(PrimExpr arg, int bits)
Fast_erf_float expression from Eigen.
TIR builtin intrinsics.
TIR expressions.
#define TVM_DECLARE_INTRIN_UNARY(OpName)
Definition: op.h:750
#define TVM_DEFINE_INT_OP_CONST_VAL_OVERLOAD(Name)
Definition: op.h:1085
#define TVM_DEFINE_LOGICAL_OP_CONST_VAL_OVERLOAD_SPANNED(Name)
Definition: op.h:1077
#define TVM_DEFINE_ASSIGN_OP_OVERLOAD(Name, OpFunc)
Definition: op.h:1037
#define TVM_DECLARE_FLOAT_INTRIN_UNARY(OpName)
Definition: op.h:753
#define TVM_DEFINE_BINOP_CONST_VAL_OVERLOAD_SPANNED(Name)
Definition: op.h:1056
#define TVM_DEFINE_BINOP_CONST_VAL_OVERLOAD(Name)
Definition: op.h:1043
#define TVM_DEFINE_LOGICAL_OP_CONST_VAL_OVERLOAD(Name)
Definition: op.h:1073
#define TVM_DECLARE_INTRIN_BINARY(OpName)
Definition: op.h:782
#define TVM_DEFINE_INT_OP_CONST_VAL_OVERLOAD_SPANNED(Name)
Definition: op.h:1091
TIR statements.