[CODEGEN/LLVM] Initial support for codegen LLVM.

Makefile

@@ -1,8 +1,8 @@
 ifndef config
 ifneq ("$(wildcard ./config.mk)","")
-	config = config.mk
+	config ?= config.mk
 else
-	config = make/config.mk
+	config ?= make/config.mk
 endif
 endif
 
@@ -19,31 +19,24 @@ SRC = $(wildcard src/*.cc src/*/*.cc src/*/*/*.cc)
 ALL_OBJ = $(patsubst src/%.cc, build/%.o, $(SRC))
 ALL_DEP = $(ALL_OBJ) $(LIB_HALIDE_IR)
 
-ifneq ($(USE_CUDA_PATH), NONE)
-	NVCC=$(USE_CUDA_PATH)/bin/nvcc
-endif
-
 export LDFLAGS = -pthread -lm
-export CFLAGS =  -std=c++11 -Wall -O2\
-	 -Iinclude -Idmlc-core/include -IHalideIR/src  -fPIC
-export FRAMEWORKS=
-
-ifneq ($(ADD_CFLAGS), NONE)
-	CFLAGS += $(ADD_CFLAGS)
-endif
+export CFLAGS =  -std=c++11 -Wall -O2 -fno-rtti\
+	 -Iinclude -Idmlc-core/include -IHalideIR/src  -fPIC -DDMLC_ENABLE_RTTI=0
 
-ifneq ($(ADD_LDFLAGS), NONE)
-	LDFLAGS += $(ADD_LDFLAGS)
+ifdef CUDA_PATH
+	NVCC=$(CUDA_PATH)/bin/nvcc
+	CFLAGS += -I$(CUDA_PATH)/include
+	LDFLAGS += -L$(CUDA_PATH)/lib64
 endif
 
-
 ifeq ($(USE_CUDA), 1)
 	CFLAGS += -DTVM_CUDA_RUNTIME=1
 	LDFLAGS += -lcuda -lcudart -lnvrtc
 else
 	CFLAGS += -DTVM_CUDA_RUNTIME=0
 endif
 
+FRAMEWORKS=
 
 ifeq ($(USE_OPENCL), 1)
 	CFLAGS += -DTVM_OPENCL_RUNTIME=1
@@ -57,6 +50,23 @@ else
 	CFLAGS += -DTVM_OPENCL_RUNTIME=0
 endif
 
+# llvm configuration
+LLVM_CONFIG=llvm-config
+
+ifeq ($(USE_LLVM), 1)
+	LLVM_VERSION=$(shell $(LLVM_CONFIG) --version| cut -b 1,3)
+	LLVM_INCLUDE=$(filter -I%, $(shell $(LLVM_CONFIG) --cxxflags))
+	LDFLAGS += $(shell $(LLVM_CONFIG) --ldflags --libs --system-libs)
+	CFLAGS += $(LLVM_INCLUDE) -DTVM_LLVM_VERSION=$(LLVM_VERSION)
+endif
+
+ifdef $(ADD_CFLAGS)
+	CFLAGS += $(ADD_CFLAGS)
+endif
+
+ifdef $(ADD_LDFLAGS)
+	LDFLAGS += $(ADD_LDFLAGS)
+endif
 
 include tests/cpp/unittest.mk
 

include/tvm/buffer.h

@@ -90,7 +90,7 @@ class BufferNode : public Node {
                      Type dtype);
 
   static constexpr const char* _type_key = "Buffer";
-  TVM_DECLARE_NODE_TYPE_INFO(BufferNode);
+  TVM_DECLARE_NODE_TYPE_INFO(BufferNode, Node);
 };
 
 inline const BufferNode* Buffer::operator->() const {

include/tvm/codegen.h

@@ -31,6 +31,13 @@ PackedFunc BuildStackVM(
     LoweredFunc func,
     const std::unordered_map<LoweredFunc, PackedFunc>& device_funcs);
 
+/*!
+ * \brief Build a LLVM VM function, this is still beta
+ * \param func The LoweredFunc to be build
+ * \return A packed function representing the func.
+ */
+PackedFunc BuildLLVM(LoweredFunc func);
+
 /*!
  * \brief Build a CUDA function with NVRTC
  *

include/tvm/expr.h

@@ -36,7 +36,6 @@ using Halide::Internal::make_zero;
 using Halide::Internal::as_const_int;
 using Halide::Internal::as_const_uint;
 
-
 inline Type TVMType2Type(TVMType t) {
   return Type(static_cast<halide_type_code_t>(t.code), t.bits, t.lanes);
 }
@@ -182,7 +181,7 @@ class IterVarNode : public Node {
   static IterVar make(Range dom, Var var, std::string thread_tag);
 
   static constexpr const char* _type_key = "IterVar";
-  TVM_DECLARE_NODE_TYPE_INFO(IterVarNode);
+  TVM_DECLARE_NODE_TYPE_INFO(IterVarNode, Node);
 };
 
 // inline implementations

include/tvm/ir.h

@@ -200,6 +200,8 @@ using Halide::Internal::Realize;
 using Halide::Internal::Block;
 using Halide::Internal::IfThenElse;
 using Halide::Internal::Evaluate;
+// ir functions
+using Halide::Internal::is_const_power_of_two_integer;
 
 }  // namespace ir
 }  // namespace tvm

include/tvm/lowered_func.h

@@ -92,7 +92,7 @@ class LoweredFuncNode : public FunctionBaseNode {
   }
 
   static constexpr const char* _type_key = "LoweredFunc";
-  TVM_DECLARE_NODE_TYPE_INFO(LoweredFuncNode);
+  TVM_DECLARE_NODE_TYPE_INFO(LoweredFuncNode, Node);
 };
 
 // Implementations of inline functions

include/tvm/operation.h

@@ -39,7 +39,7 @@ class PlaceholderOpNode : public OperationNode {
                         Type dtype);
 
   static constexpr const char* _type_key = "PlaceholderOp";
-  TVM_DECLARE_NODE_TYPE_INFO(PlaceholderOpNode);
+  TVM_DECLARE_NODE_TYPE_INFO(PlaceholderOpNode, OperationNode);
 };
 
 /*!
@@ -74,7 +74,7 @@ class ComputeOpNode : public OperationNode {
                         Expr body);
 
   static constexpr const char* _type_key = "ComputeOp";
-  TVM_DECLARE_NODE_TYPE_INFO(ComputeOpNode);
+  TVM_DECLARE_NODE_TYPE_INFO(ComputeOpNode, OperationNode);
 };
 
 /*!
@@ -123,7 +123,7 @@ class ScanOpNode : public OperationNode {
                         Array<Tensor> state_placeholder);
 
   static constexpr const char* _type_key = "ScanOp";
-  TVM_DECLARE_NODE_TYPE_INFO(ScanOpNode);
+  TVM_DECLARE_NODE_TYPE_INFO(ScanOpNode, OperationNode);
 };
 
 

include/tvm/packed_func_ext.h

@@ -33,7 +33,7 @@ struct NodeTypeChecker {
     // It can be turned off, but will make non strict checking.
     // TODO(tqchen) possibly find alternative to turn of RTTI
     using ContainerType = typename T::ContainerType;
-    return (dynamic_cast<ContainerType*>(sptr) != nullptr);
+    return sptr->derived_from<ContainerType>();
   }
   static inline void PrintName(std::ostringstream& os) { // NOLINT(*)
     using ContainerType = typename T::ContainerType;

include/tvm/runtime/c_runtime_api.h

@@ -152,6 +152,13 @@ typedef void* TVMRetValueHandle;
 /*! \brief the array handle */
 typedef TVMArray* TVMArrayHandle;
 
+/*!
+ * \brief Used for implementing C API function.
+ *  Set last error message before return.
+ * \param msg The error message to be set.
+ */
+TVM_DLL void TVMAPISetLastError(const char* msg);
+
 /*!
  * \brief return str message of the last error
  *  all function in this file will return 0 when success
@@ -287,10 +294,10 @@ TVM_DLL int TVMCFuncSetReturn(TVMRetValueHandle ret,
  * \param num_args Number of arguments.
  * \param ret The return value handle.
  * \param resource_handle The handle additional resouce handle from fron-end.
- *
+ * \return 0 if success, -1 if failure happens, set error via TVMAPISetLastError.
  * \sa TVMCFuncSetReturn
  */
-typedef void (*TVMPackedCFunc)(
+typedef int (*TVMPackedCFunc)(
     TVMValue* args,
     int* type_codes,
     int num_args,

include/tvm/schedule.h

@@ -331,7 +331,7 @@ class StageNode : public Node {
   }
 
   static constexpr const char* _type_key = "Stage";
-  TVM_DECLARE_NODE_TYPE_INFO(StageNode);
+  TVM_DECLARE_NODE_TYPE_INFO(StageNode, Node);
 };
 
 /*! \brief node container for schedule */
@@ -354,7 +354,7 @@ class ScheduleNode : public Node {
   }
 
   static constexpr const char* _type_key = "Schedule";
-  TVM_DECLARE_NODE_TYPE_INFO(ScheduleNode);
+  TVM_DECLARE_NODE_TYPE_INFO(ScheduleNode, Node);
 };
 
 /*! \brief node container for IterVar attr */
@@ -368,11 +368,14 @@ class IterVarAttrNode : public Node {
   }
 
   static constexpr const char* _type_key = "IterVarAttr";
-  TVM_DECLARE_NODE_TYPE_INFO(IterVarAttrNode);
+  TVM_DECLARE_NODE_TYPE_INFO(IterVarAttrNode, Node);
 };
 
 /*! \brief base node of iteration var */
 class IterVarRelationNode : public Node {
+ public:
+  static constexpr const char* _type_key = "IterVarRelation";
+  TVM_DECLARE_BASE_NODE_INFO(IterVarRelationNode, Node);
 };
 
 /*!
@@ -402,7 +405,7 @@ class SplitNode : public IterVarRelationNode {
       IterVar inner, Expr factor);
 
   static constexpr const char* _type_key = "Split";
-  TVM_DECLARE_NODE_TYPE_INFO(SplitNode);
+  TVM_DECLARE_NODE_TYPE_INFO(SplitNode, IterVarRelationNode);
 };
 
 /*!
@@ -427,7 +430,7 @@ class FuseNode : public IterVarRelationNode {
       IterVar outer, IterVar inner, IterVar fused);
 
   static constexpr const char* _type_key = "Fuse";
-  TVM_DECLARE_NODE_TYPE_INFO(FuseNode);
+  TVM_DECLARE_NODE_TYPE_INFO(FuseNode, IterVarRelationNode);
 };
 
 /*!
@@ -450,7 +453,7 @@ class RebaseNode : public IterVarRelationNode {
   static IterVarRelation make(IterVar parent, IterVar rebased);
 
   static constexpr const char* _type_key = "Rebase";
-  TVM_DECLARE_NODE_TYPE_INFO(RebaseNode);
+  TVM_DECLARE_NODE_TYPE_INFO(RebaseNode, IterVarRelationNode);
 };
 
 

include/tvm/tensor.h

@@ -153,7 +153,7 @@ class TensorNode : public Node {
                      int value_index);
 
   static constexpr const char* _type_key = "Tensor";
-  TVM_DECLARE_NODE_TYPE_INFO(TensorNode);
+  TVM_DECLARE_NODE_TYPE_INFO(TensorNode, Node);
 };
 
 /*!
@@ -167,8 +167,6 @@ class OperationNode : public FunctionBaseNode {
   const std::string& func_name() const final {
     return name;
   }
-  /*! \return number of outputs of this op */
-  virtual int num_outputs() const = 0;
   /*! \return the list of iteration variable at root */
   virtual Array<IterVar> root_iter_vars() const = 0;
   /*! \return type of i-th output */
@@ -177,6 +175,8 @@ class OperationNode : public FunctionBaseNode {
   virtual Array<Expr> output_shape(size_t i) const = 0;
 
   static constexpr const char* _type_key = "Operation";
+
+  TVM_DECLARE_BASE_NODE_INFO(OperationNode, Node);
 };
 
 // Implementations of inline functions

make/config.mk

@@ -40,10 +40,11 @@ USE_CUDA = 1
 # whether use OpenCL during compile
 USE_OPENCL = 0
 
-# add the path to CUDA library to link and compile flag
-# if you have already add them to environment variable, leave it as NONE
-# USE_CUDA_PATH = /usr/local/cuda
-USE_CUDA_PATH = NONE
+# whether build with LLVM support
+# This requires llvm-config to be in your PATH
+# Requires LLVM version >= 4.0
+USE_LLVM = 0
 
-# whether use cuda runtime compiling for writing kernels in native language (i.e. Python)
-USE_NVRTC = 0
+# add the path to CUDA library to link and compile flag
+# if you have already add them to environment variable.
+# CUDA_PATH = /usr/local/cuda

python/tvm/_ctypes/_function.py

@@ -56,6 +56,7 @@ def cfun(args, type_codes, num_args, ret, _):
             check_call(_LIB.TVMCFuncSetReturn(ret, values[0], ctypes.c_int(tcodes[0])))
             _ = temp_args
             _ = rv
+        return 0
 
     handle = FunctionHandle()
     f = TVMPackedCFunc(cfun)

python/tvm/_ctypes/_types.py

@@ -96,7 +96,7 @@ class TVMByteArray(ctypes.Structure):
 
 
 TVMPackedCFunc = ctypes.CFUNCTYPE(
-    None,
+    ctypes.c_int,
     ctypes.POINTER(TVMValue),
     ctypes.POINTER(ctypes.c_int),
     ctypes.c_int,

src/README.md

@@ -6,3 +6,4 @@
 - arithmetic Arithmetic expression and set simplification
 - pass The optimization pass on the IR structure
 - runtime Minimum runtime related codes.
+- codegen The code generator

src/api/api_codegen.cc

@@ -37,6 +37,11 @@ TVM_REGISTER_API(_codegen_BuildStackVM)
                         std::unordered_map<LoweredFunc, PackedFunc>());
   });
 
+TVM_REGISTER_API(_codegen_BuildLLVM)
+.set_body([](TVMArgs args, TVMRetValue *ret) {
+    *ret = BuildLLVM(args[0]);
+  });
+
 TVM_REGISTER_API(_codegen_BuildNVRTC)
 .set_body([](TVMArgs args, TVMRetValue *ret) {
     *ret = BuildNVRTC(args[0], args[1]);

src/arithmetic/int_set.h

@@ -20,7 +20,7 @@ enum SignType {
 };
 
 // internal node container of int set.
-class IntSetNode;
+struct IntSetNode;
 
 /*!
  * \brief Integer set class, represent a set of integers in one dimension.
@@ -104,6 +104,8 @@ class IntSet : public NodeRef {
  * \brief Base class of all IntSet containers.
  */
 struct IntSetNode : public Node {
+  static constexpr const char* _type_key = "IntSet";
+  TVM_DECLARE_BASE_NODE_INFO(IntSetNode, Node);
 };
 
 using ExprIntSetMap = std::unordered_map<Expr, IntSet,

src/arithmetic/int_set_internal.h

@@ -35,7 +35,7 @@ struct IntervalSet : public IntSetNode {
   }
 
   static constexpr const char* _type_key = "IntervalSet";
-  TVM_DECLARE_NODE_TYPE_INFO(IntervalSet);
+  TVM_DECLARE_NODE_TYPE_INFO(IntervalSet, IntSetNode);
 };
 
 /*!
@@ -51,7 +51,7 @@ struct StrideSet : public IntSetNode {
   Array<Expr> strides;
 
   static constexpr const char* _type_key = "StrideSet";
-  TVM_DECLARE_NODE_TYPE_INFO(StrideSet);
+  TVM_DECLARE_NODE_TYPE_INFO(StrideSet, IntSetNode);
 };
 
 }  // namespace arith

src/codegen/codegen_stack_vm.cc

@@ -272,9 +272,6 @@ inline void PushBinary(StackVM::OpCode op_int64,
   }
 }
 
-
-
-
 inline void PushCast(Type dst,
                      Type src,
                      CodeGenStackVM* p) {
@@ -496,7 +493,5 @@ TVM_STATIC_IR_FUNCTOR(CodeGenStackVM, vtable)
 .set_dispatch<Call>([](const Call *op, CodeGenStackVM* p) {
     p->Push_(op);
   });
-
-
 }  // namespace codegen
 }  // namespace tvm