Update the mainline to include Release 1.7.0 rc (#312)

* Use TVM 1.7

* Use release-1.7.0-rc branch from neo-ai/tvm

* Adding #7027 in TVM Submodule

* Adding #183 in TVM Submodule

Co-authored-by: Trevor Morris <[email protected]>
Co-authored-by: Tushar Dey <[email protected]>

CMakeLists.txt

@@ -1,7 +1,7 @@
 cmake_minimum_required (VERSION 3.6)
 include(cmake/Utils.cmake)
-include(3rdparty/tvm/cmake/util/Util.cmake)
-include(3rdparty/tvm/cmake/util/FindCUDA.cmake)
+include(3rdparty/tvm/cmake/utils/Utils.cmake)
+include(3rdparty/tvm/cmake/utils/FindCUDA.cmake)
 
 # Option for Android on Arm --- has to come before project() function
 option(ANDROID_BUILD "Build for Android target" OFF)
@@ -213,68 +213,14 @@ if(USE_TENSORRT)
     include_directories(${TENSORRT_INCLUDE_DIR})
     list(APPEND DLR_LINKER_LIBS ${TENSORRT_LIB_DIR})
 
-    # NNVM TRT Integration
-    file(GLOB TENSORRT_SRCS ${TVM_SRC}/src/contrib/subgraph/*.cc)
-    list(APPEND DLR_SRC ${TENSORRT_SRCS})
+    # TRT runtime sources
+    file(GLOB RUNTIME_TENSORRT_SRCS ${TVM_SRC}/src/runtime/contrib/tensorrt/*.cc)
+    set_source_files_properties(${RUNTIME_TENSORRT_SRCS} PROPERTIES COMPILE_FLAGS "-Wno-deprecated-declarations")
+    list(APPEND DLR_SRC ${RUNTIME_TENSORRT_SRCS})
+
+    # Set defines
     add_definitions(-DTVM_GRAPH_RUNTIME_TENSORRT)
 
-    # TRT Relay sources
-    file(GLOB TENSORRT_RELAY_SRCS ${TVM_SRC}/src/runtime/contrib/tensorrt/*.cc)
-    list(APPEND DLR_SRC ${TENSORRT_RELAY_SRCS})
-    # Additional TVM compiler sources
-    include_directories("${TVM_SRC}/3rdparty/rang/include")
-    include_directories("${TVM_SRC}/3rdparty/compiler-rt")
-    include_directories("${TVM_SRC}/topi/include")
-    file(GLOB_RECURSE COMPILER_SRCS
-        ${TVM_SRC}/src/auto_scheduler/*.cc
-        ${TVM_SRC}/src/node/*.cc
-        ${TVM_SRC}/src/ir/*.cc
-        ${TVM_SRC}/src/arith/*.cc
-        ${TVM_SRC}/src/te/*.cc
-        ${TVM_SRC}/src/autotvm/*.cc
-        ${TVM_SRC}/src/tir/*.cc
-        ${TVM_SRC}/src/topi/*.cc
-        ${TVM_SRC}/src/driver/*.cc
-        ${TVM_SRC}/src/parser/*.cc
-        ${TVM_SRC}/src/printer/*.cc
-        ${TVM_SRC}/src/support/*.cc
-    )
-    file(GLOB CODEGEN_SRCS
-        ${TVM_SRC}/src/target/*.cc
-        ${TVM_SRC}/src/target/source/*.cc
-    )
-    list(APPEND COMPILER_SRCS ${CODEGEN_SRCS})
-    file(GLOB_RECURSE RELAY_OP_SRCS
-        ${TVM_SRC}/src/relay/op/*.cc
-    )
-    file(GLOB_RECURSE RELAY_PASS_SRCS
-        ${TVM_SRC}/src/relay/analysis/*.cc
-        ${TVM_SRC}/src/relay/transforms/*.cc
-        ${TVM_SRC}/src/relay/quantize/*.cc
-    )
-    file(GLOB RELAY_BACKEND_SRCS
-        ${TVM_SRC}/src/relay/backend/*.cc
-        ${TVM_SRC}/src/relay/backend/vm/*.cc
-    )
-    file(GLOB_RECURSE RELAY_IR_SRCS
-        ${TVM_SRC}/src/relay/ir/*.cc
-    )
-    file(GLOB_RECURSE RELAY_QNN_SRCS
-        ${TVM_SRC}/src/relay/qnn/*.cc
-    )
-    list(APPEND COMPILER_SRCS ${RELAY_OP_SRCS})
-    list(APPEND COMPILER_SRCS ${RELAY_PASS_SRCS})
-    list(APPEND COMPILER_SRCS ${RELAY_BACKEND_SRCS})
-    list(APPEND COMPILER_SRCS ${RELAY_IR_SRCS})
-    list(APPEND COMPILER_SRCS ${RELAY_QNN_SRCS})
-    file(GLOB DATATYPE_SRCS ${TVM_SRC}/src/target/datatype/*.cc)
-    list(APPEND COMPILER_SRCS ${DATATYPE_SRCS})
-    list(APPEND COMPILER_SRCS ${TVM_SRC}/src/target/opt/build_metal_off.cc)
-    list(APPEND COMPILER_SRCS ${TVM_SRC}/src/target/opt/build_sdaccel_off.cc)
-    list(APPEND COMPILER_SRCS ${TVM_SRC}/src/target/opt/build_aocl_off.cc)
-    list(APPEND COMPILER_SRCS ${TVM_SRC}/src/target/opt/build_opencl_off.cc)
-    list(APPEND COMPILER_SRCS ${TVM_SRC}/src/runtime/container.cc)
-    list(APPEND DLR_SRC ${COMPILER_SRCS})
     set(USE_TENSORRT OFF)
 endif()
 if(WITH_HEXAGON)

include/dlr.h

@@ -18,7 +18,7 @@ extern "C" {  // Open extern "C" block
 /*! \brief major version */
 #define DLR_MAJOR 1
 /*! \brief minor version */
-#define DLR_MINOR 6
+#define DLR_MINOR 7
 /*! \brief patch version */
 #define DLR_PATCH 0
 /*! \brief DLR version */

python/dlr/metadata.py

@@ -1,3 +1,3 @@
 NAME = ['DLRModel']
 
-VERSION = "1.6.0"
+VERSION = "1.7.0"

src/dlr_relayvm.cc

@@ -86,8 +86,16 @@ void RelayVMModel::SetupVMModule(const std::vector<DLRModelElem>& model_elems) {
   vm_module_ = std::make_shared<tvm::runtime::Module>(tvm::runtime::Module(vm));
 
   tvm::runtime::PackedFunc init = vm_module_->GetFunction("init");
-  init(static_cast<int>(ctx_.device_type), ctx_.device_id,
-       static_cast<int>(tvm::runtime::vm::AllocatorType::kPooled));
+  if (ctx_.device_type == DLDeviceType::kDLCPU) {
+    init(static_cast<int>(ctx_.device_type), ctx_.device_id,
+         static_cast<int>(tvm::runtime::vm::AllocatorType::kPooled));
+  } else {
+    // CPU context also must be initialized because input/output data comes from CPU.
+    init(static_cast<int>(ctx_.device_type), ctx_.device_id,
+         static_cast<int>(tvm::runtime::vm::AllocatorType::kPooled),
+         static_cast<int>(DLDeviceType::kDLCPU), 0,
+         static_cast<int>(tvm::runtime::vm::AllocatorType::kPooled));
+  }
 }
 
 void RelayVMModel::FetchInputNodesData() {
@@ -268,7 +276,7 @@ void RelayVMModel::SetInput(const char* name, const int64_t* shape, const void*
   DLDataType dtype = GetInputDLDataType(index);
   DLTensor input_tensor;
   input_tensor.data = const_cast<void*>(input);
-  input_tensor.ctx = ctx_;
+  input_tensor.ctx = DLContext{DLDeviceType::kDLCPU, 0};
   input_tensor.ndim = dim;
   input_tensor.shape = const_cast<int64_t*>(shape);
   input_tensor.strides = nullptr;
@@ -357,7 +365,7 @@ void RelayVMModel::GetOutput(int index, void* output) {
   }
   DLTensor output_tensor;
   output_tensor.data = output;
-  output_tensor.ctx = ctx_;
+  output_tensor.ctx = DLContext{DLDeviceType::kDLCPU, 0};
   output_tensor.ndim = out_array->ndim;
   output_tensor.dtype = out_array->dtype;
   output_tensor.shape = out_array->shape;

src/dlr_tvm.cc

@@ -97,7 +97,7 @@ void TVMModel::SetupTVMModule(const std::vector<DLRModelElem>& model_elems) {
   module = tvm::runtime::Module::LoadFromFile(model_lib_path);
 
   tvm_graph_runtime_ = tvm::runtime::make_object<tvm::runtime::GraphRuntime>();
-  tvm_graph_runtime_->Init(graph_str, module, {ctx_});
+  tvm_graph_runtime_->Init(graph_str, module, {ctx_}, nullptr);
   dmlc::MemoryFixedSizeStream strm(const_cast<char*>(params_data), params_size);
   tvm_graph_runtime_->LoadParams(&strm);