Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into fix_randomness_in_fusion

Author: huangjiyi

_typos.toml

@@ -10,6 +10,7 @@ grad = "grad"
 arange = "arange"
 ot = 'ot'
 pash = 'pash'
+eles = 'eles'
 
 # These words need to be fixed
 ontext = 'ontext'
@@ -25,7 +26,6 @@ expection = 'expection'
 pacakage = 'pacakage'
 normlized = 'normlized'
 Runing = 'Runing'
-Ags = 'Ags'
 Becasue = 'Becasue'
 craeted = 'craeted'
 oreder = 'oreder'
@@ -154,7 +154,6 @@ CACH = 'CACH'
 endianess = 'endianess'
 VAILD = 'VAILD'
 ues = 'ues'
-algorithem = 'algorithem'
 aer = 'aer'
 elemenents = 'elemenents'
 CANN = 'CANN'
@@ -181,7 +180,6 @@ overrided = 'overrided'
 smll = 'smll'
 outpout = 'outpout'
 staticaly = 'staticaly'
-aranged = 'aranged'
 offets = 'offets'
 olny = 'olny'
 Continer = 'Continer'
@@ -192,7 +190,6 @@ readed = 'readed'
 Opeartion = 'Opeartion'
 shoule = 'shoule'
 inputed = 'inputed'
-arrray = 'arrray'
 positon = 'positon'
 invalide = 'invalide'
 repeatly = 'repeatly'
@@ -255,16 +252,13 @@ defition = 'defition'
 operants = 'operants'
 funcitons = 'funcitons'
 dateset = 'dateset'
-arised = 'arised'
 optimzed = 'optimzed'
 encouter = 'encouter'
-alis = 'alis'
 feeded = 'feeded'
 poped = 'poped'
 parmeter = 'parmeter'
 doens = 'doens'
 cadidate = 'cadidate'
-argumnets = 'argumnets'
 inconsistence = 'inconsistence'
 Caculate = 'Caculate'
 seperator = 'seperator'
@@ -358,7 +352,6 @@ accessable = 'accessable'
 Wheter = 'Wheter'
 processer = 'processer'
 recored = 'recored'
-afer = 'afer'
 wraper = 'wraper'
 Partitial = 'Partitial'
 leafs = 'leafs'
@@ -417,7 +410,6 @@ recieved = 'recieved'
 Hanlder = 'Hanlder'
 EPOCHES = 'EPOCHES'
 sequnce = 'sequnce'
-argmuments = 'argmuments'
 Iteraion = 'Iteraion'
 whill = 'whill'
 tood = 'tood'
@@ -478,7 +470,6 @@ subsituted = 'subsituted'
 automaticly = 'automaticly'
 Minium = 'Minium'
 sequnece = 'sequnece'
-adjancent = 'adjancent'
 payed = 'payed'
 linke = 'linke'
 nagative = 'nagative'
@@ -556,7 +547,6 @@ splited = 'splited'
 instrinsics = 'instrinsics'
 outputing = 'outputing'
 hadler = 'hadler'
-aggragate = 'aggragate'
 qucik = 'qucik'
 alog = 'alog'
 exsit = 'exsit'
@@ -664,7 +654,6 @@ usefull = 'usefull'
 sqaure = 'sqaure'
 adn = 'adn'
 intialize = 'intialize'
-addtional = 'addtional'
 Taget = 'Taget'
 parm = 'parm'
 thrads = 'thrads'
@@ -686,7 +675,6 @@ upsupported = 'upsupported'
 settting = 'settting'
 templat = 'templat'
 priorites = 'priorites'
-admissable = 'admissable'
 optmize = 'optmize'
 bondary = 'bondary'
 Traning = 'Traning'
@@ -732,7 +720,6 @@ mannualy = 'mannualy'
 learing = 'learing'
 noraml = 'noraml'
 padd = 'padd'
-alogrithm = 'alogrithm'
 Requred = 'Requred'
 astroid = 'astroid'
 Fidn = 'Fidn'
@@ -778,14 +765,12 @@ expaned = 'expaned'
 choos = 'choos'
 whos = 'whos'
 architecuture = 'architecuture'
-argumet = 'argumet'
 coule = 'coule'
 instanciate = 'instanciate'
 distrubuted = 'distrubuted'
 Localy = 'Localy'
 PARM = 'PARM'
 thi = 'thi'
-addess = 'addess'
 Oll = 'Oll'
 Auxillary = 'Auxillary'
 Infor = 'Infor'

paddle/cinn/backends/codegen_gpu_dev.cc

@@ -151,7 +151,7 @@ void CodeGenGpuDev::Visit(const ir::_LoweredFunc_ *op) {
   auto axis_range_assumptions = op->PrepareAxisRangeAssumptions();
   auto alloca_temp_buffers = op->PrepareAllocTempBufferExprs();
   auto temp_buffer_alias = GenerateBufferAliasExprs(op, op->temp_bufs);
-  auto alis_var_exprs = op->CudaAliasVarExprs();
+  auto alias_var_exprs = op->CudaAliasVarExprs();
   auto dealloc_temp_buffers =
       FilterDeallocTempBuffers(op->PrepareDeallocTempBufferExprs());
 
@@ -160,7 +160,7 @@ void CodeGenGpuDev::Visit(const ir::_LoweredFunc_ *op) {
   APPEND_TO_NEW_BODY(axis_range_assumptions)
   APPEND_TO_NEW_BODY(alloca_temp_buffers)
   APPEND_TO_NEW_BODY(temp_buffer_alias)
-  APPEND_TO_NEW_BODY(alis_var_exprs)
+  APPEND_TO_NEW_BODY(alias_var_exprs)
 
   new_body.push_back(op->body);
   APPEND_TO_NEW_BODY(dealloc_temp_buffers);

paddle/cinn/common/object.h

@@ -75,5 +75,42 @@ struct Object {
 using object_ptr = Object*;
 using shared_object = Shared<Object>;
 
+/*
+ * \brief Delete the default copy/move constructor and assign operator
+ * \param TypeName The class typename.
+ */
+#define CINN_DELETE_COPY_MOVE_AND_ASSIGN_FOR_OBJECT_NODE(TypeName) \
+  TypeName(const TypeName& other) = delete;                        \
+  TypeName(TypeName&& other) = delete;                             \
+  TypeName& operator=(const TypeName& other) = delete;             \
+  TypeName& operator=(TypeName&& other) = delete;
+
+/*
+ * \brief Define the default copy/move constructor and assign operator
+ * \param TypeName The class typename.
+ */
+#define CINN_DEFINE_COPY_MOVE_AND_ASSIGN_FOR_OBJECT_REF(TypeName) \
+  TypeName(const TypeName& other) = default;                      \
+  TypeName(TypeName&& other) = default;                           \
+  TypeName& operator=(const TypeName& other) = default;           \
+  TypeName& operator=(TypeName&& other) = default;
+
+/*
+ * \brief Define object reference methods
+ * \param TypeName The object type name
+ * \param ParentType The parent type of the objectref
+ * \param ObjectName The type name of the object
+ */
+#define CINN_DEFINE_OBJECT_REF_METHODS(TypeName, ParentType, ObjectName) \
+  TypeName() = default;                                                  \
+  explicit TypeName(ObjectName* n) : ParentType(n) {}                    \
+  explicit TypeName(const cinn::common::Shared<Object>& ref)             \
+      : ParentType(ref) {}                                               \
+  CINN_DEFINE_COPY_MOVE_AND_ASSIGN_FOR_OBJECT_REF(TypeName);             \
+  const ObjectName* operator->() const {                                 \
+    return static_cast<const ObjectName*>(this->p_);                     \
+  }                                                                      \
+  ObjectName* operator->() { return static_cast<ObjectName*>(this->p_); }
+
 }  // namespace common
 }  // namespace cinn

paddle/cinn/ir/CMakeLists.txt

@@ -19,6 +19,8 @@ gather_srcs(
   intrinsic_ops.cc
   layout.cc
   schedule_block_graph.cc
+  stmt.cc
+  stmt_visitors.cc
   dim.cc)
 
 add_subdirectory(ir_analyzer)

paddle/cinn/ir/ir_base.cc

@@ -66,6 +66,21 @@ std::ostream &operator<<(std::ostream &os, IrNodeTy type) {
   return os;
 }
 
+std::ostream &operator<<(std::ostream &os, StmtNodeTy type) {
+  switch (type) {
+#define __m(t__)                         \
+  case StmtNodeTy::t__:                  \
+    os << "<stmt node: " << #t__ << ">"; \
+    break;
+
+    NODETY_FORALL_STMT(__m)
+#undef __m
+    default:
+      PADDLE_THROW(
+          ::common::errors::InvalidArgument("unknown StmtNodeTy found"));
+  }
+}
+
 Expr Zero(const Type &type) {
   if (type.is_bfloat16()) return Expr(bfloat16(0.f));
   if (type.is_float16()) return Expr(float16(0.f));
@@ -330,15 +345,18 @@ void IrNode::replace(Expr old_op, Expr new_op) {
 }
 
 void IrNode::convert_int32_to_int64() {
-  if (type_ != Int(64))
-    if (type_ != Int(32))
+  if (type_ != Int(64) && type_ != UInt(64))
+    if (type_ != Int(32) && type_ != UInt(32))
       PADDLE_ENFORCE_EQ(type_.is_unk(),
                         true,
                         ::common::errors::InvalidArgument(
                             "Current only support convert int32_t "
                             "to int64_t, but get type is: %s",
                             type_));
-  type_ = Int(64);
+
+  if (type_ == Int(32)) type_ = Int(64);
+  if (type_ == UInt(32)) type_ = UInt(64);
+
   for (Expr &operand : operands) {
     operand->convert_int32_to_int64();
   }

paddle/cinn/ir/ir_base.h

@@ -116,6 +116,7 @@ class Dim;
 #define NODETY_CONTROL_OP_FOR_INTRINSIC(macro__) \
   macro__(IntrinsicOp)                      \
 
+// TODO(Hongqing-work): change NODETY_FORALL to NODETY_FORALL_EXPR
 #define NODETY_FORALL(__m)              \
   NODETY_PRIMITIVE_TYPE_FOR_EACH(__m)   \
   NODETY_OP_FOR_EACH(__m)               \
@@ -126,6 +127,16 @@ class Dim;
   NODETY_PRIMITIVE_TYPE_FOR_EACH(__m)                    \
   NODETY_OP_FOR_EACH(__m)                                \
   NODETY_CONTROL_OP_FOR_EACH(__m)
+
+#define NODETY_FORALL_STMT(macro__) \
+  macro__(Let)                      \
+  macro__(Store)                    \
+  macro__(Alloc)                    \
+  macro__(Free)                     \
+  macro__(IfThenElse)               \
+  macro__(For)                      \
+  macro__(Schedule)                 \
+  macro__(Evaluate)
 // clang-format on
 
 //! Define IrNodeTy
@@ -143,6 +154,13 @@ enum class IrNodeTy {
 #undef __m
 // @}
 
+//! Define StmtNodeTy
+// @{
+#define __m(x__) x__,
+enum class StmtNodeTy { kUnk = -1, NODETY_FORALL_STMT(__m) };
+#undef __m
+// @}
+
 //! String representations for IrNodeTy.
 // @{
 #define __m(x__) #x__,
@@ -152,6 +170,7 @@ const std::vector<std::string> kIrNodeTyReprs(
 // @}
 
 std::ostream& operator<<(std::ostream& os, IrNodeTy type);
+std::ostream& operator<<(std::ostream& os, StmtNodeTy type);
 
 struct Expr;