Merge pull request #288 from NVIDIA/to_backend_device

Adding the new device API, fixing the a nested dict issue in the existing compile phase, adding new lowering pass for bn

Author: narendasan

.bazelversion

@@ -1 +1 @@
-3.7.0
+4.0.0

WORKSPACE

@@ -79,10 +79,10 @@ http_archive(
 
 http_archive(
     name = "tensorrt",
-    urls = ["https://developer.nvidia.com/compute/machine-learning/tensorrt/secure/7.2.1/tars/TensorRT-7.2.1.6.Ubuntu-18.04.x86_64-gnu.cuda-11.0.cudnn8.0.tar.gz",],
+    urls = ["https://developer.nvidia.com/compute/machine-learning/tensorrt/secure/7.2.2/tars/TensorRT-7.2.2.3.Ubuntu-18.04.x86_64-gnu.cuda-11.0.cudnn8.0.tar.gz",],
     build_file = "@//third_party/tensorrt/archive:BUILD",
-    sha256 = "8def6b03b0c8c3751f560df21b3e99668ae05aab5140b1d38b8e51e4a0ffbbb8",
-    strip_prefix = "TensorRT-7.2.1.6"
+    strip_prefix = "TensorRT-7.2.2.3",
+    sha256 = "b5c325e38e1d92ce1ce92ca8b54ede9c224bf128c9a53eb0b9022f1ee4313ee0"
 )
 
 ####################################################################################

core/lowering/lowering.cpp

@@ -40,6 +40,7 @@ void LowerGraph(std::shared_ptr<torch::jit::Graph>& g) {
   passes::Conv2DToConvolution(g);
   passes::Conv3DToConvolution(g);
   passes::FuseAddMMBranches(g);
+  passes::RemoveBNDimCheck(g);
   torch::jit::EliminateCommonSubexpression(g);
   // torch::jit::UnrollLoops(g);
   torch::jit::EliminateCommonSubexpression(g);

core/lowering/passes/BUILD

@@ -18,6 +18,7 @@ cc_library(
         "exception_elimination.cpp",
         "fuse_addmm_branches.cpp",
         "fuse_flatten_linear.cpp",
+        "remove_bn_dim_check.cpp",
         "remove_contiguous.cpp",
         "remove_dropout.cpp",
         "remove_to.cpp",

core/lowering/passes/passes.h

@@ -12,6 +12,7 @@ void Conv3DToConvolution(std::shared_ptr<torch::jit::Graph>& graph);
 void FuseAddMMBranches(std::shared_ptr<torch::jit::Graph> graph);
 void FuseFlattenLinear(std::shared_ptr<torch::jit::Graph>& graph);
 void EliminateExceptionOrPassPattern(std::shared_ptr<torch::jit::Graph> graph);
+void RemoveBNDimCheck(std::shared_ptr<torch::jit::Graph> graph);
 void RemoveContiguous(std::shared_ptr<torch::jit::Graph>& graph);
 void RemoveDropout(std::shared_ptr<torch::jit::Graph>& graph);
 void RemoveTo(std::shared_ptr<torch::jit::Graph> graph);

core/lowering/passes/remove_bn_dim_check.cpp

@@ -0,0 +1,89 @@
+#include "torch/csrc/jit/ir/alias_analysis.h"
+#include "torch/csrc/jit/jit_log.h"
+#include "torch/csrc/jit/passes/constant_propagation.h"
+#include "torch/csrc/jit/passes/dead_code_elimination.h"
+#include "torch/csrc/jit/passes/guard_elimination.h"
+#include "torch/csrc/jit/passes/peephole.h"
+#include "torch/csrc/jit/runtime/graph_executor.h"
+
+#include "core/util/prelude.h"
+
+#include <vector>
+
+namespace trtorch {
+namespace core {
+namespace lowering {
+namespace passes {
+namespace {
+using namespace torch::jit;
+struct BNDimCheckRemoval {
+  BNDimCheckRemoval(std::shared_ptr<Graph> graph) : graph_(std::move(graph)) {}
+
+  void run() {
+    findBNDimCheckNodes(graph_->block());
+    torch::jit::EliminateDeadCode(graph_);
+    LOG_GRAPH("Post batch norm dim check removal: " << *graph_);
+  }
+
+ private:
+  bool isBNDimCheckNodes(Node* n) {
+    /// Check if this Node hosts a pattern like so:
+    /// %290 : bool = aten::ne(%289, %9)
+    ///   = prim::If(%290)
+    ///     block0():
+    ///       %291 : str = aten::format(%10, %289)
+    ///        = prim::RaiseException(%291)
+    ///       -> ()
+    ///     block1():
+    ///       -> ()
+
+    if (n->blocks().size() != 2) {
+      return false;
+    }
+    auto arm1 = n->blocks()[0];
+    auto arm2 = n->blocks()[1];
+    if (arm1->outputs().size() != 0 || arm2->outputs().size() != 0) {
+      // Make sure that the node doesn't actually produce any Value that are
+      // used by other nodes
+      return false;
+    }
+
+    auto arm1_start = arm1->nodes().begin();
+
+    if ((*arm1_start)->kind() != c10::Symbol::fromQualString("aten::format") &&
+        (*(++arm1_start))->kind() != prim::RaiseException && (*(++arm1_start))->kind() != prim::Return) {
+      // Make sure that block0 is solely just the exception and the return
+      return false;
+    }
+
+    if ((*(arm2->nodes().begin()))->kind() != prim::Return) {
+      // Make sure that block1 is solely the return
+      return false;
+    }
+
+    return true;
+  }
+
+  void findBNDimCheckNodes(Block* b) {
+    for (auto it = b->nodes().begin(); it != b->nodes().end(); it++) {
+      auto n = *it;
+      if (n->kind() == prim::If && isBNDimCheckNodes(n)) {
+        LOG_GRAPH("Found that node " << *n << "  is an batch norm dim check node (EliminateChecks)" << std::endl);
+        it.destroyCurrent();
+      }
+    }
+  }
+
+  std::shared_ptr<Graph> graph_;
+};
+} // namespace
+
+void RemoveBNDimCheck(std::shared_ptr<Graph> graph) {
+  BNDimCheckRemoval bndcr(std::move(graph));
+  bndcr.run();
+}
+
+} // namespace passes
+} // namespace lowering
+} // namespace core
+} // namespace trtorch

docsrc/tutorials/use_from_pytorch.rst

@@ -32,31 +32,36 @@ at the documentation for the TRTorch ``TensorRTCompileSpec`` API.
 .. code-block:: python
 
     spec = {
-        "forward": trtorch.TensorRTCompileSpec({
-            "input_shapes": [[1, 3, 300, 300]],
-            "op_precision": torch.half,
-            "refit": False,
-            "debug": False,
-            "strict_types": False,
-            "allow_gpu_fallback": True,
-            "device_type": "gpu",
-            "capability": trtorch.EngineCapability.default,
-            "num_min_timing_iters": 2,
-            "num_avg_timing_iters": 1,
-            "max_batch_size": 0,
-        })
-    }
+        "forward":
+            trtorch.TensorRTCompileSpec({
+                "input_shapes": [[1, 3, 300, 300]],
+                "op_precision": torch.half,
+                "refit": False,
+                "debug": False,
+                "strict_types": False,
+                "device": {
+                    "device_type": trtorch.DeviceType.GPU,
+                    "gpu_id": 0,
+                    "dla_core": 0,
+                    "allow_gpu_fallback": True
+                },
+                "capability": trtorch.EngineCapability.default,
+                "num_min_timing_iters": 2,
+                "num_avg_timing_iters": 1,
+                "max_batch_size": 0,
+            })
+        }
 
 Now to compile with TRTorch, provide the target module objects and the spec dictionary to ``torch._C._jit_to_tensorrt``
 
 .. code-block:: python
 
-    trt_model = torch._C._jit_to_tensorrt(script_model._c, spec)
+    trt_model = torch._C._jit_to_backend("tensorrt", script_model, spec)
 
 To run explicitly call the function of the method you want to run (vs. how you can just call on the module itself in standard PyTorch)
 
 .. code-block:: python
 
-    input = torch.randn((1, 3, 300, 300).to("cuda").to(torch.half)
+    input = torch.randn((1, 3, 300, 300)).to("cuda").to(torch.half)
     print(trt_model.forward(input))
 

py/trtorch/_compile_spec.py

@@ -147,10 +147,6 @@ def _parse_compile_spec(compile_spec: Dict[str, Any]) -> trtorch._C.CompileSpec:
         assert isinstance(compile_spec["strict_types"], bool)
         info.strict_types = compile_spec["strict_types"]
 
-    if "allow_gpu_fallback" in compile_spec:
-        assert isinstance(compile_spec["allow_gpu_fallback"], bool)
-        info.allow_gpu_fallback = compile_spec["allow_gpu_fallback"]
-
     if "device" in compile_spec:
         info.device = _parse_device(compile_spec["device"])
 
@@ -177,7 +173,7 @@ def _parse_compile_spec(compile_spec: Dict[str, Any]) -> trtorch._C.CompileSpec:
     return info
 
 
-def TensorRTCompileSpec(compile_spec: Dict[str, Any]):
+def TensorRTCompileSpec(compile_spec: Dict[str, Any]) -> torch.classes.tensorrt.CompileSpec:
     """
     Utility to create a formated spec dictionary for using the PyTorch TensorRT backend
 
@@ -199,10 +195,10 @@ def TensorRTCompileSpec(compile_spec: Dict[str, Any]):
                             } # Dynamic input shape for input #2
                         ],
                         "device": {
-                        "device_type": torch.device("cuda"), # Type of device to run engine on (for DLA use trtorch.DeviceType.DLA)
-                        "gpu_id": 0, # Target gpu id to run engine (Use Xavier as gpu id for DLA)
-                        "dla_core": 0, # (DLA only) Target dla core id to run engine
-                        "allow_gpu_fallback": false, # (DLA only) Allow layers unsupported on DLA to run on GPU
+                            "device_type": torch.device("cuda"), # Type of device to run engine on (for DLA use trtorch.DeviceType.DLA)
+                            "gpu_id": 0, # Target gpu id to run engine (Use Xavier as gpu id for DLA)
+                            "dla_core": 0, # (DLA only) Target dla core id to run engine
+                            "allow_gpu_fallback": false, # (DLA only) Allow layers unsupported on DLA to run on GPU
                         },
                         "op_precision": torch.half, # Operating precision set to FP16
                         "refit": False, # enable refit
@@ -235,14 +231,13 @@ def TensorRTCompileSpec(compile_spec: Dict[str, Any]):
         ir.set_max(i.max)
         backend_spec.append_input_range(ir)
 
-    for i in parsed_spec.device:
-        ir = torch.classes.tensorrt.Device()
-        ir.set_device_type(i.device_type)
-        ir.set_gpu_id(i.gpu_id)
-        ir.set_dla_core(i.dla_core)
-        ir.set_allow_gpu_fallback(i.allow_gpu_fallback)
-        backend_spec.set_device(ir)
+    d = torch.classes.tensorrt.Device()
+    d.set_device_type(int(parsed_spec.device.device_type))
+    d.set_gpu_id(parsed_spec.device.gpu_id)
+    d.set_dla_core(parsed_spec.device.dla_core)
+    d.set_allow_gpu_fallback(parsed_spec.device.allow_gpu_fallback)
 
+    backend_spec.set_device(d)
     backend_spec.set_op_precision(int(parsed_spec.op_precision))
     backend_spec.set_refit(parsed_spec.refit)
     backend_spec.set_debug(parsed_spec.debug)

py/trtorch/csrc/register_tensorrt_classes.cpp

@@ -3,22 +3,32 @@
 namespace trtorch {
 namespace backend {
 namespace {
-void RegisterTRTCompileSpec() {
+
 #define ADD_FIELD_GET_SET_REGISTRATION(registry, class_name, field_name) \
   (registry).def("set_" #field_name, &class_name::set_##field_name);     \
   (registry).def("get_" #field_name, &class_name::get_##field_name);
 
+void RegisterTRTCompileSpec() {
   static auto TRTORCH_UNUSED TRTInputRangeTSRegistration =
       torch::class_<trtorch::pyapi::InputRange>("tensorrt", "InputRange").def(torch::init<>());
 
   ADD_FIELD_GET_SET_REGISTRATION(TRTInputRangeTSRegistration, trtorch::pyapi::InputRange, min);
   ADD_FIELD_GET_SET_REGISTRATION(TRTInputRangeTSRegistration, trtorch::pyapi::InputRange, opt);
   ADD_FIELD_GET_SET_REGISTRATION(TRTInputRangeTSRegistration, trtorch::pyapi::InputRange, max);
 
+  static auto TRTORCH_UNUSED TRTDeviceTSRegistration =
+      torch::class_<trtorch::pyapi::Device>("tensorrt", "Device").def(torch::init<>());
+
+  ADD_FIELD_GET_SET_REGISTRATION(TRTDeviceTSRegistration, trtorch::pyapi::Device, device_type);
+  ADD_FIELD_GET_SET_REGISTRATION(TRTDeviceTSRegistration, trtorch::pyapi::Device, gpu_id);
+  ADD_FIELD_GET_SET_REGISTRATION(TRTDeviceTSRegistration, trtorch::pyapi::Device, dla_core);
+  ADD_FIELD_GET_SET_REGISTRATION(TRTDeviceTSRegistration, trtorch::pyapi::Device, allow_gpu_fallback);
+
   static auto TRTORCH_UNUSED TRTCompileSpecTSRegistration =
       torch::class_<trtorch::pyapi::CompileSpec>("tensorrt", "CompileSpec")
           .def(torch::init<>())
           .def("append_input_range", &trtorch::pyapi::CompileSpec::appendInputRange)
+          .def("set_device", &trtorch::pyapi::CompileSpec::setDeviceIntrusive)
           .def("__str__", &trtorch::pyapi::CompileSpec::stringify);
 
   ADD_FIELD_GET_SET_REGISTRATION(TRTCompileSpecTSRegistration, trtorch::pyapi::CompileSpec, op_precision);

py/trtorch/csrc/tensorrt_backend.cpp

@@ -46,7 +46,7 @@ c10::impl::GenericDict TensorRTBackend::compile(c10::IValue processed_mod, c10::
     auto method = mod.get_method(method_name);
     auto g = method.graph();
 
-    auto raw_spec = it->value().toGenericDict().at(it->key()).toCustomClass<trtorch::pyapi::CompileSpec>();
+    auto raw_spec = it->value().toCustomClass<trtorch::pyapi::CompileSpec>();
     LOG_DEBUG(raw_spec->stringify());
     auto cfg = raw_spec->toInternalCompileSpec();
     auto convert_cfg = std::move(cfg.convert_info);