Merge pull request #194 from NVIDIA/to_backend_api

Implementation of the PyTorch Backend API

Author: narendasan

.bazelrc

@@ -29,6 +29,7 @@ build --cxxopt='-std=c++14'
 build:python --cxxopt="-D_GLIBCXX_USE_CXX11_ABI=0"
 build:python --linkopt="-D_GLIBCXX_USE_CXX11_ABI=0"
 build:python --define=abi=pre_cxx11_abi
+build:python --define=target_lang=python
 
 build:pre_cxx11_abi --cxxopt="-D_GLIBCXX_USE_CXX11_ABI=0"
 build:pre_cxx11_abi --linkopt="-D_GLIBCXX_USE_CXX11_ABI=0"

.github/pr-labels.yml

@@ -16,8 +16,8 @@
 "component: evaluators":
   - core/conversion/evaluators/**/*
 
-"component: execution":
-  - core/execution/**/*
+"component: runtime":
+  - core/runtime/**/*
 
 "component: lowering":
   - core/lowering/**/*
@@ -32,4 +32,4 @@
 "documentation":
   - docs/**/*
   - docsrc/**/*
- 
+

BUILD

@@ -19,7 +19,7 @@ pkg_tar(
         "//core/conversion/tensorcontainer:include",
         "//core/conversion/evaluators:include",
         "//core/conversion/converters/impl/plugins:include",
-        "//core/execution:include",
+        "//core/runtime:include",
         "//core/lowering:include",
         "//core/lowering/passes:include",
         "//core/util:include",

README.md

@@ -18,7 +18,7 @@ More Information / System Architecture:
 #include "trtorch/trtorch.h"
 
 ...
-auto compile_settings = trtorch::ExtraInfo(dims);
+auto compile_settings = trtorch::CompileSpec(dims);
 // FP16 execution
 compile_settings.op_precision = torch::kFloat;
 // Compile module
@@ -54,7 +54,7 @@ torch.jit.save(trt_ts_module, "trt_torchscript_module.ts")
 ```
 
 > Notes on running in lower precisions:
-> - Set precision with extra_info.op_precision
+> - Set precision with compile_spec.op_precision
 > - The module should be left in FP32 before compilation (FP16 can support half tensor models)
 > - In FP16 only input tensors should be converted to FP16, other precisions use FP32
 

core/BUILD

@@ -7,6 +7,13 @@ config_setting(
     }
 )
 
+config_setting(
+    name = "python_core",
+    values = {
+        "define": "target_lang=python"
+    }
+)
+
 cc_library(
     name = "core",
     hdrs = [
@@ -17,7 +24,7 @@ cc_library(
     ],
     deps = [
         "//core/conversion",
-        "//core/execution",
+        "//core/runtime",
         "//core/lowering",
         "//core/util/logging",
         "@tensorrt//:nvinfer"
@@ -28,11 +35,13 @@ cc_library(
     alwayslink=True,
 )
 
-
 load("@rules_pkg//:pkg.bzl", "pkg_tar")
 
 pkg_tar(
     name = "include",
     package_dir = "core/",
-    srcs = ["compiler.h"],
+    srcs = [
+        "backend.h",
+        "compiler.h",
+    ],
 )

core/compiler.cpp

@@ -20,7 +20,7 @@
 
 #include "core/lowering/lowering.h"
 #include "core/conversion/conversion.h"
-#include "core/execution/execution.h"
+#include "core/runtime/runtime.h"
 
 namespace trtorch {
 namespace core {
@@ -42,15 +42,15 @@ c10::FunctionSchema GenerateGraphSchema(torch::jit::script::Module mod, std::str
 
 
 void AddEngineToGraph(torch::jit::script::Module mod, std::shared_ptr<torch::jit::Graph>& g, std::string& serialized_engine) {
-    auto engine_ptr = c10::make_intrusive<execution::TRTEngine>(mod._ivalue()->name(), serialized_engine);
+    auto engine_ptr = c10::make_intrusive<runtime::TRTEngine>(mod._ivalue()->name(), serialized_engine);
     // Get required metadata about the engine out
     auto num_io = engine_ptr->num_io;
     auto name = engine_ptr->name;
 
     // Add the engine as an attribute of the module, this will let the engine be serialized and deserialized
     mod.register_attribute(
         name,
-        c10::getCustomClassType<c10::intrusive_ptr<execution::TRTEngine>>(),
+        c10::getCustomClassType<c10::intrusive_ptr<runtime::TRTEngine>>(),
         c10::IValue(std::move(engine_ptr)),
         false
     );
@@ -125,7 +125,7 @@ bool CheckMethodOperatorSupport(const torch::jit::script::Module& mod,
 
 std::string ConvertGraphToTRTEngine(const torch::jit::script::Module& mod,
                                     std::string method_name,
-                                    ExtraInfo cfg) {
+                                    CompileSpec cfg) {
 
     // Go through Lowering to simplify graph and extract weight parameters
     auto graph_and_parameters = lowering::Lower(mod, method_name);
@@ -137,12 +137,12 @@ std::string ConvertGraphToTRTEngine(const torch::jit::script::Module& mod,
 
     LOG_INFO(*g << "(CompileGraph)\n");
 
-    auto engine = ConvertBlockToEngine(g->block(), convert_cfg, named_params);
+    auto engine = conversion::ConvertBlockToEngine(g->block(), convert_cfg, named_params);
     return std::move(engine);
 }
 
 torch::jit::script::Module CompileGraph(const torch::jit::script::Module& mod,
-                                        ExtraInfo cfg) {
+                                        CompileSpec cfg) {
     // TODO: Should be doing a functional transform but need PR #31978
     // [jit] More robust mangling
     //torch::jit::script::Module new_mod = mod.clone();

core/compiler.h

@@ -7,18 +7,18 @@
 namespace trtorch {
 namespace core {
 
-struct ExtraInfo {
-    ExtraInfo(std::vector<conversion::InputRange> input_ranges)
+struct CompileSpec {
+    CompileSpec(std::vector<conversion::InputRange> input_ranges)
         : convert_info(std::move(input_ranges)) {}
     conversion::ConversionInfo convert_info;
 };
 
 bool CheckMethodOperatorSupport(const torch::jit::script::Module& mod, std::string method_name);
 
 std::string ConvertGraphToTRTEngine(const torch::jit::script::Module& mod,
-                                    std::string method_name, ExtraInfo cfg);
+                                    std::string method_name, CompileSpec cfg);
 
-torch::jit::script::Module CompileGraph(const torch::jit::script::Module& module, ExtraInfo cfg);
+torch::jit::script::Module CompileGraph(const torch::jit::script::Module& module, CompileSpec cfg);
 
 } // namespace core
 } // namespace trtorch

core/conversion/conversionctx/ConversionCtx.cpp

@@ -55,7 +55,7 @@ ConversionCtx::ConversionCtx(BuilderSettings build_settings)
             cfg->setFlag(nvinfer1::BuilderFlag::kFP16);
         }
         input_type = nvinfer1::DataType::kFLOAT;
-        TRTORCH_CHECK(settings.calibrator != nullptr, "Requested inference in INT8 but no calibrator provided, set the ptq_calibrator field in the ExtraInfo struct with your calibrator");
+        TRTORCH_CHECK(settings.calibrator != nullptr, "Requested inference in INT8 but no calibrator provided, set the ptq_calibrator field in the CompileSpec struct with your calibrator");
         cfg->setInt8Calibrator(settings.calibrator);
         break;
     case nvinfer1::DataType::kFLOAT:

core/execution/BUILD renamed to core/runtime/BUILD

@@ -8,9 +8,9 @@ config_setting(
 )
 
 cc_library(
-    name = "execution",
+    name = "runtime",
     hdrs = [
-        "execution.h",
+        "runtime.h",
     ],
     srcs = [
         "TRTEngine.cpp",
@@ -30,6 +30,6 @@ load("@rules_pkg//:pkg.bzl", "pkg_tar")
 
 pkg_tar(
     name = "include",
-    package_dir = "core/execution/",
-    srcs = ["execution.h"],
+    package_dir = "core/runtime/",
+    srcs = ["runtime.h"],
 )

core/execution/TRTEngine.cpp renamed to core/runtime/TRTEngine.cpp

@@ -4,11 +4,11 @@
 #include "torch/csrc/jit/frontend/function_schema_parser.h"
 
 #include "core/util/prelude.h"
-#include "core/execution/execution.h"
+#include "core/runtime/runtime.h"
 
 namespace trtorch {
 namespace core {
-namespace execution {
+namespace runtime {
 
 std::string slugify(std::string s) {
     std::replace(s.begin(), s.end(), '.', '_');
@@ -81,6 +81,7 @@ TRTEngine::~TRTEngine() {
 //     return c10::List<at::Tensor>(output_vec);
 // }
 
+namespace {
 static auto TRTORCH_UNUSED TRTEngineTSRegistrtion = torch::class_<TRTEngine>("tensorrt", "Engine")
     .def(torch::init<std::string>())
     // TODO: .def("__call__", &TRTEngine::Run)
@@ -94,7 +95,7 @@ static auto TRTORCH_UNUSED TRTEngineTSRegistrtion = torch::class_<TRTEngine>("te
             return c10::make_intrusive<TRTEngine>(std::move(seralized_engine));
         }
     );
-
-} // namespace execution
+} // namespace
+} // namespace runtime
 } // namespace core
 } // namespace trtorch

core/execution/register_trt_op.cpp renamed to core/runtime/register_trt_op.cpp

@@ -4,11 +4,11 @@
 #include "torch/csrc/jit/runtime/custom_operator.h"
 
 #include "core/util/prelude.h"
-#include "core/execution/execution.h"
+#include "core/runtime/runtime.h"
 
 namespace trtorch {
 namespace core {
-namespace execution {
+namespace runtime {
 
 std::vector<at::Tensor> execute_engine(std::vector<at::Tensor> inputs, c10::intrusive_ptr<TRTEngine> compiled_engine) {
     LOG_DEBUG("Attempting to run engine (ID: " << compiled_engine->name << ")");
@@ -30,7 +30,7 @@ std::vector<at::Tensor> execute_engine(std::vector<at::Tensor> inputs, c10::intr
         gpu_handles.push_back(contig_inputs.back().data_ptr());
     }
 
-    TRTORCH_CHECK(compiled_engine->exec_ctx->allInputDimensionsSpecified(), "Not enough inputs provided (execution.RunCudaEngine)");
+    TRTORCH_CHECK(compiled_engine->exec_ctx->allInputDimensionsSpecified(), "Not enough inputs provided (runtime.RunCudaEngine)");
 
     std::vector<at::Tensor> outputs(compiled_engine->num_io.second);
     for (size_t o = inputs.size(); o < (compiled_engine->num_io.first + compiled_engine->num_io.second); o++) {
@@ -53,6 +53,6 @@ TORCH_LIBRARY(tensorrt, m) {
   m.def("execute_engine", execute_engine);
 }
 
-} // namespace execution
+} // namespace runtime
 } // namespace core
 } // namespace trtorch

core/execution/execution.h renamed to core/runtime/runtime.h

@@ -8,7 +8,7 @@
 
 namespace trtorch {
 namespace core {
-namespace execution {
+namespace runtime {
 
 using EngineID = int64_t;
 
@@ -35,6 +35,6 @@ struct TRTEngine : torch::CustomClassHolder {
 
 std::vector<at::Tensor> execute_engine(std::vector<at::Tensor> inputs, c10::intrusive_ptr<TRTEngine> compiled_engine);
 
-} // namespace execution
+} // namespace runtime
 } // namespace core
 } // namespace trtorch

cpp/api/BUILD

@@ -9,7 +9,7 @@ cc_library(
         "include/trtorch/ptq.h"
     ],
     srcs = [
-        "src/extra_info.cpp",
+        "src/compile_spec.cpp",
         "src/logging.cpp",
         "src/trtorch.cpp",
         "src/ptq.cpp"

cpp/api/README.md

@@ -31,7 +31,7 @@ namespace trtorch {
  * Settings data structure for TRTorch compilation
  *
  */
-struct TRTORCH_API ExtraInfo {
+struct TRTORCH_API CompileSpec {
     /**
      * @brief A struct to hold an input range (used by TensorRT Optimization profile)
      *
@@ -132,10 +132,10 @@ struct TRTORCH_API ExtraInfo {
         kSAFE_DLA,
     };
 
-    ExtraInfo(std::vector<InputRange> input_ranges)
+    CompileSpec(std::vector<InputRange> input_ranges)
         : input_ranges(std::move(input_ranges)) {}
-    ExtraInfo(std::vector<std::vector<int64_t>> fixed_sizes);
-    ExtraInfo(std::vector<c10::ArrayRef<int64_t>> fixed_sizes);
+    CompileSpec(std::vector<std::vector<int64_t>> fixed_sizes);
+    CompileSpec(std::vector<c10::ArrayRef<int64_t>> fixed_sizes);
 
     // Defaults should reflect TensorRT defaults for BuilderConfig
 
@@ -236,27 +236,27 @@ TRTORCH_API bool CheckMethodOperatorSupport(const torch::jit::script::Module& mo
  * @brief Compile a TorchScript module for NVIDIA GPUs using TensorRT
  *
  * @param module: torch::jit::script::Module - Existing TorchScript module
- * @param info: trtorch::ExtraInfo - Compilation settings
+ * @param info: trtorch::CompileSpec - Compilation settings
  *
  * Takes a existing TorchScript module and a set of settings to configure the compiler
  * and will convert methods to JIT Graphs which call equivalent TensorRT engines
  *
  * Converts specifically the forward method of a TorchScript Module
  */
-TRTORCH_API torch::jit::script::Module CompileGraph(const torch::jit::script::Module& module, ExtraInfo info);
+TRTORCH_API torch::jit::script::Module CompileGraph(const torch::jit::script::Module& module, CompileSpec info);
 
 /**
  * @brief Compile a TorchScript method for NVIDIA GPUs using TensorRT
  *
  * @param module: torch::jit::script::Module - Existing TorchScript module
  * @param method_name: std::string - Name of method to compile
- * @param info: trtorch::ExtraInfo - Compilation settings
+ * @param info: trtorch::CompileSpec - Compilation settings
  *
  * Takes a existing TorchScript module and a set of settings to configure the compiler
  * and will convert selected method to a serialized TensorRT engine which can be run with
  * TensorRT
  */
-TRTORCH_API std::string ConvertGraphToTRTEngine(const torch::jit::script::Module& module, std::string method_name, ExtraInfo info);
+TRTORCH_API std::string ConvertGraphToTRTEngine(const torch::jit::script::Module& module, std::string method_name, CompileSpec info);
 
 namespace ptq {
 /**

cpp/api/include/trtorch/ptq.h

@@ -145,7 +145,7 @@ class Int8Calibrator : Algorithm {
     /**
      * @brief operator to cast to nvinfer1::IInt8Calibrator*
      *
-     * Convience function to convert to a IInt8Calibrator* to easily be assigned to the ptq_calibrator field in ExtraInfo
+     * Convience function to convert to a IInt8Calibrator* to easily be assigned to the ptq_calibrator field in CompileSpec
      *
      * @return nvinfer1::IInt8Calibrator*
      */
@@ -259,7 +259,7 @@ class Int8CacheCalibrator : Algorithm {
     /**
      * @brief operator to cast to nvinfer1::IInt8Calibrator*
      *
-     * Convience function to convert to a IInt8Calibrator* to easily be assigned to the ptq_calibrator field in ExtraInfo
+     * Convience function to convert to a IInt8Calibrator* to easily be assigned to the ptq_calibrator field in CompileSpec
      *
      * @return nvinfer1::IInt8Calibrator*
      */