feat: Runtime output buffer optimization

Author: keehyuna

core/runtime/TRTEngine.cpp

@@ -319,6 +319,10 @@ int64_t TRTEngine::get_automatic_device_memory_budget() {
   return cuda_engine->getWeightStreamingAutomaticBudget();
 }
 
+void TRTEngine::set_pre_allocated_outputs(bool enable) {
+  use_pre_allocated_outputs = enable;
+}
+
 std::string TRTEngine::to_str() const {
   // clang-format off
   std::stringstream ss;

core/runtime/TRTEngine.h

@@ -88,6 +88,7 @@ struct TRTEngine : torch::CustomClassHolder {
   int64_t get_streamable_device_memory_budget();
   int64_t get_automatic_device_memory_budget();
   std::vector<at::Tensor> infer_outputs(std::vector<std::vector<int64_t>> input_shapes);
+  void set_pre_allocated_outputs(bool enable);
   friend std::ostream& operator<<(std::ostream& os, const TRTEngine& engine);
   static const char BINDING_DELIM = '%';
 
@@ -102,6 +103,9 @@ struct TRTEngine : torch::CustomClassHolder {
   std::vector<at::Tensor> input_buffers = {};
   std::vector<at::Tensor> output_buffers = {};
   std::string shape_key;
+  bool cudagraphs_enabled = false;
+  bool use_pre_allocated_outputs = true;
+  std::vector<at::Tensor> pre_allocated_outputs;
 
   // TODO: Implement a call method
   // c10::List<at::Tensor> Run(c10::List<at::Tensor> inputs);

core/runtime/execute_engine.cpp

@@ -5,6 +5,7 @@
 #include "torch/csrc/jit/runtime/custom_operator.h"
 #include "torch/torch.h"
 
+#include <ATen/record_function.h>
 #include "core/runtime/TRTEngineProfiler.h"
 #include "core/runtime/runtime.h"
 #include "core/util/prelude.h"
@@ -60,9 +61,8 @@ RTDevice select_rt_device(const RTDevice& engine_device, const RTDevice& curr_de
   return new_target_device_opt.value();
 }
 
-bool _cudagraphs_validate_shapes(std::vector<at::Tensor> inputs, c10::intrusive_ptr<TRTEngine> compiled_engine) {
-  // Validate whether the current input shapes to the engine
-  // invalidate the existing cudagraphs object
+bool _validate_shapes(std::vector<at::Tensor> inputs, c10::intrusive_ptr<TRTEngine> compiled_engine) {
+  // Validate whether the current input shapes to the engine has changed
 
   // Populate the shape key for the inputs
   // x: (3, 4), y: (4, 5) --> Key: (3,4)(4,5)
@@ -83,15 +83,32 @@ bool _cudagraphs_validate_shapes(std::vector<at::Tensor> inputs, c10::intrusive_
 
   auto new_shape_key = new_shape_key_ss.str();
 
-  // Compare the shape key to the original key and invalidate shapes if they do not match
+  // Compare the shape key to the original key
   if (new_shape_key != compiled_engine->shape_key) {
-    LOG_DEBUG("Resetting Cudagraph on New Shape Key " << new_shape_key);
+    LOG_DEBUG("Input shape changed " << compiled_engine->shape_key << " -> " << new_shape_key);
     compiled_engine->shape_key = new_shape_key;
-    compiled_engine->cudagraph.reset();
-    return false;
+    return true;
   }
 
-  return true;
+  return false;
+}
+
+std::vector<at::Tensor> create_output_tensors(c10::intrusive_ptr<TRTEngine> compiled_engine) {
+  std::vector<at::Tensor> outputs(compiled_engine->num_io.second);
+  for (auto output_indices : compiled_engine->out_binding_map) {
+    // out_binding_map stores TRT_IDX: PYT_IDX
+    auto pyt_idx = output_indices.second;
+
+    std::string name = compiled_engine->out_binding_names[pyt_idx];
+    auto out_shape = compiled_engine->exec_ctx->getTensorShape(name.c_str());
+    LOG_DEBUG("Output Name: " << name << " Shape: " << out_shape);
+
+    auto dims = core::util::toVec(out_shape);
+    auto type = util::TRTDataTypeToScalarType(compiled_engine->exec_ctx->getEngine().getTensorDataType(name.c_str()));
+    outputs[pyt_idx] = std::move(at::empty(dims, {at::kCUDA}).to(type).contiguous());
+  }
+
+  return outputs;
 }
 
 std::vector<at::Tensor> execute_engine(std::vector<at::Tensor> inputs, c10::intrusive_ptr<TRTEngine> compiled_engine) {
@@ -116,10 +133,15 @@ std::vector<at::Tensor> execute_engine(std::vector<at::Tensor> inputs, c10::intr
     compiled_engine->cudagraph.enable_debug_mode();
   }
 
+  bool shape_changed = _validate_shapes(inputs, compiled_engine);
+
   // Whether cudagraphs needs to record the graph on this pass
-  bool need_cudagraphs_record = (CUDAGRAPHS_MODE && (!_cudagraphs_validate_shapes(inputs, compiled_engine)));
+  // Cudagraphs record is required if cudagraphs_enabled is switched to True regardless of shape change
+  bool need_cudagraphs_record =
+      (((!compiled_engine->cudagraphs_enabled) && CUDAGRAPHS_MODE) || (CUDAGRAPHS_MODE && shape_changed));
+  compiled_engine->cudagraphs_enabled = CUDAGRAPHS_MODE;
 
-  if (!CUDAGRAPHS_MODE) {
+  if (!CUDAGRAPHS_MODE || shape_changed) {
     compiled_engine->cudagraph.reset();
   }
 
@@ -180,6 +202,7 @@ std::vector<at::Tensor> execute_engine(std::vector<at::Tensor> inputs, c10::intr
 
   { // Input Setup
     std::unique_ptr<torch::autograd::profiler::RecordProfile> input_profiler_guard;
+    RECORD_FUNCTION("process input", std::vector<c10::IValue>());
     if (compiled_engine->profile_execution) {
       input_profiler_guard =
           std::make_unique<torch::autograd::profiler::RecordProfile>(compiled_engine->input_profile_path);
@@ -261,23 +284,20 @@ std::vector<at::Tensor> execute_engine(std::vector<at::Tensor> inputs, c10::intr
 
   { // Output Setup
     std::unique_ptr<torch::autograd::profiler::RecordProfile> output_profiler_guard;
+    RECORD_FUNCTION("process output", std::vector<c10::IValue>());
     if (compiled_engine->profile_execution) {
       output_profiler_guard =
           std::make_unique<torch::autograd::profiler::RecordProfile>(compiled_engine->output_profile_path);
     }
+    if ((false == compiled_engine->use_pre_allocated_outputs) || shape_changed) {
+      outputs = create_output_tensors(compiled_engine);
+    } else {
+      outputs = compiled_engine->pre_allocated_outputs;
+    }
 
     for (auto output_indices : compiled_engine->out_binding_map) {
-      // out_binding_map stores TRT_IDX: PYT_IDX
       auto pyt_idx = output_indices.second;
-
       std::string name = compiled_engine->out_binding_names[pyt_idx];
-      auto out_shape = compiled_engine->exec_ctx->getTensorShape(name.c_str());
-      LOG_DEBUG("Output Name: " << name << " Shape: " << out_shape);
-
-      auto dims = core::util::toVec(out_shape);
-      auto type = util::TRTDataTypeToScalarType(compiled_engine->exec_ctx->getEngine().getTensorDataType(name.c_str()));
-      outputs[pyt_idx] = std::move(at::empty(dims, {at::kCUDA}).to(type).contiguous());
-
       if (need_cudagraphs_record) {
         // If we are recording the cuda graph then we need to update the persistent output buffer
         compiled_engine->output_buffers[pyt_idx] = std::move(outputs[pyt_idx].clone());
@@ -310,6 +330,7 @@ std::vector<at::Tensor> execute_engine(std::vector<at::Tensor> inputs, c10::intr
   }
 
   { // Engine Execution (execute on engine stream)
+    RECORD_FUNCTION("Trt runtime", std::vector<c10::IValue>());
     c10::cuda::CUDAStreamGuard stream_guard(compiled_engine->engine_stream);
 
     std::unique_ptr<torch::autograd::profiler::RecordProfile> enqueue_profiler_guard;
@@ -344,6 +365,11 @@ std::vector<at::Tensor> execute_engine(std::vector<at::Tensor> inputs, c10::intr
     }
   } // End engine exeuction (resets to caller stream)
 
+  // Create output buffer for next execution of graph or trt context.
+  if (compiled_engine->use_pre_allocated_outputs) {
+    compiled_engine->pre_allocated_outputs = create_output_tensors(compiled_engine);
+  }
+
   // Block caller stream until engine execution is complete
   at::cuda::CUDAEvent trt_exec_complete;
   trt_exec_complete.record(compiled_engine->engine_stream);

core/runtime/register_jit_hooks.cpp

@@ -88,6 +88,7 @@ static auto TORCHTRT_UNUSED TRTEngineTSRegistrtion =
         .def("dump_engine_layer_info", &TRTEngine::dump_engine_layer_info)
         .def("get_engine_layer_info", &TRTEngine::get_engine_layer_info)
         .def("infer_outputs", &TRTEngine::infer_outputs)
+        .def("set_pre_allocated_outputs", &TRTEngine::set_pre_allocated_outputs)
         .def_property(
             "device_memory_budget",
             &TRTEngine::get_device_memory_budget,

py/torch_tensorrt/dynamo/runtime/_PythonTorchTensorRTModule.py

@@ -108,6 +108,9 @@ def __init__(
         self.engine = None
         self.weight_name_map = weight_name_map
         self.target_platform = Platform.current_platform()
+        self.cudagraphs_enabled = False
+        self.pre_allocated_outputs: List[torch.Tensor] = []
+        self.use_pre_allocated_outputs = False
 
         if self.serialized_engine is not None and not self.settings.lazy_engine_init:
             self.setup_engine()
@@ -172,7 +175,7 @@ def setup_engine(self) -> None:
             self.engine.get_tensor_shape(input_name) for input_name in self.input_names
         ]
         self.output_dtypes = [
-            dtype._from(self.engine.get_tensor_dtype(output_name))
+            dtype._from(self.engine.get_tensor_dtype(output_name)).to(torch.dtype)
             for output_name in self.output_names
         ]
         self.output_shapes = [
@@ -233,6 +236,19 @@ def __del__(self) -> None:
         if self.cudagraph:
             self.cudagraph.reset()
 
+    def create_output_tensors(self) -> List[torch.Tensor]:
+        # create output tensors
+        outputs: List[torch.Tensor] = []
+
+        for o, _ in enumerate(self.output_names):
+            output = torch.empty(
+                size=self.output_shapes[o],
+                dtype=self.output_dtypes[o],
+                device=torch.cuda.current_device(),
+            )
+            outputs.append(output)
+        return outputs
+
     def forward(self, *inputs: torch.Tensor) -> torch.Tensor | Tuple[torch.Tensor, ...]:
         # Ensure inputs are available in all scopes and cast symbolic integers to Tensors
         contiguous_inputs: List[torch.Tensor] = [
@@ -248,19 +264,25 @@ def forward(self, *inputs: torch.Tensor) -> torch.Tensor | Tuple[torch.Tensor, .
             self._check_initialized()
 
             cudagraphs_enabled = torch_tensorrt.runtime.get_cudagraphs_mode()
-            need_cudagraphs_record = (
-                cudagraphs_enabled and not self.cudagraphs_validate_shapes(inputs)
-            )
+            shape_changed = self.validate_input_shapes(inputs)
+            # Cudagraphs record is required if cudagraphs_enabled is toggled to True regardless of shape change
+            if not self.cudagraphs_enabled and cudagraphs_enabled:
+                need_cudagraphs_record = True
+            else:
+                need_cudagraphs_record = cudagraphs_enabled and shape_changed
+            self.cudagraphs_enabled = cudagraphs_enabled
 
             if need_cudagraphs_record:
+                if self.cudagraph:
+                    self.cudagraph.reset()
                 self._input_buffers = [None] * len(self.input_names)
                 self._output_buffers = [None] * len(self.output_names)
 
             if not cudagraphs_enabled and self.cudagraph:
                 self.cudagraph.reset()
                 self.cudagraph = None
 
-            # If in safe mode, check at each iteration for for whether a switch is required
+            # If in safe mode, check at each iteration for whether a switch is required
             if (
                 torch_tensorrt.runtime._multi_device_safe_mode._PY_RT_MULTI_DEVICE_SAFE_MODE
             ):
@@ -351,14 +373,14 @@ def forward(self, *inputs: torch.Tensor) -> torch.Tensor | Tuple[torch.Tensor, .
                             self.context.set_tensor_address(
                                 input_name, contiguous_inputs[i].data_ptr()
                             )
-
-                # Check if input shapes can be inferred.
-                uninferred_input_names = self.context.infer_shapes()
-                if uninferred_input_names:
-                    logger.warning(
-                        f"The shapes of the inputs: {uninferred_input_names} cannot be inferred and could lead to undefined behavior. \
-                                This could happen if the input tensor addresses/shapes haven't been configured correctly"
-                    )
+                if shape_changed:
+                    # Check if input shapes can be inferred.
+                    uninferred_input_names = self.context.infer_shapes()
+                    if uninferred_input_names:
+                        logger.warning(
+                            f"The shapes of the inputs: {uninferred_input_names} cannot be inferred and could lead to undefined behavior. \
+                                    This could happen if the input tensor addresses/shapes haven't been configured correctly"
+                        )
 
             with (
                 torch.autograd.profiler.record_function(
@@ -367,24 +389,20 @@ def forward(self, *inputs: torch.Tensor) -> torch.Tensor | Tuple[torch.Tensor, .
                 if self.profiling_enabled
                 else nullcontext()
             ):
-                # create output tensors
-                outputs: List[torch.Tensor] = []
-
-                for o, output_name in enumerate(self.output_names):
-                    shape = tuple(self.context.get_tensor_shape(output_name))
-
-                    if DYNAMIC_DIM in shape:
+                if not self.use_pre_allocated_outputs or shape_changed:
+                    self.output_shapes = [
+                        tuple(self.context.get_tensor_shape(output_name))
+                        for output_name in self.output_names
+                    ]
+                    if DYNAMIC_DIM in self.output_shapes:
                         raise ValueError(
                             "Encountered dynamic output shapes during runtime. This could mean the network has data-dependent output shapes which is not currently supported."
                         )
+                    outputs = self.create_output_tensors()
+                else:
+                    outputs = self.pre_allocated_outputs
 
-                    output = torch.empty(
-                        size=shape,
-                        dtype=self.output_dtypes[o].to(torch.dtype),
-                        device=torch.cuda.current_device(),
-                    )
-
-                    outputs.append(output)
+                for o, output_name in enumerate(self.output_names):
 
                     if need_cudagraphs_record:
                         self._output_buffers[o] = outputs[o].clone()
@@ -445,6 +463,9 @@ def forward(self, *inputs: torch.Tensor) -> torch.Tensor | Tuple[torch.Tensor, .
 
                 self._caller_stream.wait_stream(self._engine_stream)
 
+            if self.use_pre_allocated_outputs:
+                self.pre_allocated_outputs = self.create_output_tensors()
+
             if cudagraphs_enabled:
                 for idx, o in enumerate(outputs):
                     o.copy_(self._output_buffers[idx])
@@ -486,10 +507,9 @@ def get_layer_info(self) -> str:
         )
         return engine_json
 
-    def cudagraphs_validate_shapes(self, inputs: Sequence[torch.Tensor]) -> bool:
+    def validate_input_shapes(self, inputs: Sequence[torch.Tensor]) -> bool:
         """
-        Validates the input shapes of the forward function
-        versus the version currently active for the
+        Validates the input shapes of the forward function has changed
         """
         # Representation of input shapes to a given model
         # Shapes are concatenated as so:
@@ -499,10 +519,8 @@ def cudagraphs_validate_shapes(self, inputs: Sequence[torch.Tensor]) -> bool:
         # If the new shape key differs from the existing one,
         # invalidate the old shape key and remove the CUDAGraph
         if new_shape_key != self.shape_key:
-            logger.debug(f"Resetting Cudagraph on new shape key {new_shape_key}")
+            logger.debug(f"Input shape changed {self.shape_key} -> {new_shape_key}")
             self.shape_key = new_shape_key
-            if self.cudagraph:
-                self.cudagraph.reset()
-            return False
+            return True
 
-        return True
+        return False

py/torch_tensorrt/dynamo/runtime/_TorchTensorRTModule.py

@@ -208,6 +208,7 @@ def setup_engine(self) -> None:
         if self.engine is not None:
             return
         self.engine = torch.classes.tensorrt.Engine(self._pack_engine_info())
+        self.engine.set_pre_allocated_outputs(True)
 
     def encode_metadata(self, metadata: Any) -> str:
         metadata = copy.deepcopy(metadata)