feat(//core/partitioning): Improved logging and code org for the

segmentation step of partitioning Signed-off-by: Naren Dasan <[email protected]> Signed-off-by: Naren Dasan <[email protected]>
pytorch · Oct 19, 2021 · 8927e77 · 8927e77
1 parent 17e0e8a
commit 8927e77
Show file tree

Hide file tree

Showing 12 changed files with 249 additions and 43 deletions.
diff --git a/core/partitioning/SegmentedBlock.cpp b/core/partitioning/SegmentedBlock.cpp
@@ -4,6 +4,14 @@ namespace trtorch {
 namespace core {
 namespace partitioning {
 
+SegmentedBlock::SegmentedBlock(BlockID id, SegmentedBlockTarget blk_target, const std::vector<torch::jit::Node*>& nodes)
+    : id_(id), target_(blk_target), g_(std::make_shared<torch::jit::Graph>()) {
+  for (auto& node : nodes) {
+    nodes_.push_back(node);
+    appendNode(node);
+  }
+}
+
 SegmentedBlock::SegmentedBlock(SegmentedBlockTarget blk_target, const std::vector<torch::jit::Node*>& nodes)
     : target_(blk_target), g_(std::make_shared<torch::jit::Graph>()) {
   for (auto& node : nodes) {
@@ -62,6 +70,18 @@ torch::jit::Node* SegmentedBlock::cloneNode(torch::jit::Node* node) {
   return new_node;
 }
 
+std::ostream& operator<<(std::ostream& os, const SegmentedBlock& b) {
+  os << "Segment Block @" << b.id_ << ":" << std::endl;
+  os << "    Target: " << b.target_ << std::endl;
+  os << "    Graph: " << *b.g_ << std::endl;
+  return os;
+}
+
+std::ostream& operator<<(std::ostream& os, const SegmentedBlock::SegmentedBlockTarget& t) {
+  os << SegmentedBlock::target_to_str(t) << std::endl;
+  return os;
+}
+
 } // namespace partitioning
 } // namespace core
 } // namespace trtorch
diff --git a/core/partitioning/SegmentedBlock.h b/core/partitioning/SegmentedBlock.h
@@ -1,6 +1,7 @@
 #pragma once
 
 #include <vector>
+#include <ostream>
 
 #include "NvInfer.h"
 #include "core/ir/ir.h"
@@ -18,10 +19,21 @@ struct SegmentedBlock {
     kTensorRT,
   };
 
+  static std::string target_to_str(SegmentedBlockTarget t) {
+    if (t == SegmentedBlockTarget::kTorch) {
+      return "Torch";
+    } else {
+      return "TensorRT";
+    }
+  }
+
+  using BlockID = uint64_t;
+
   SegmentedBlock() = default;
   SegmentedBlock(SegmentedBlockTarget blk_target) : target_(blk_target), g_(std::make_shared<torch::jit::Graph>()) {}
   SegmentedBlock(SegmentedBlockTarget blk_target, const std::vector<torch::jit::Node*>& nodes);
   SegmentedBlock(SegmentedBlockTarget blk_target, std::shared_ptr<torch::jit::Graph> g) : target_(blk_target), g_(g) {}
+  SegmentedBlock(BlockID id, SegmentedBlockTarget blk_target, const std::vector<torch::jit::Node*>& nodes);
 
   torch::jit::Value* getOrAddInputForValue(torch::jit::Value* v);
   torch::jit::Node* cloneNode(torch::jit::Node* node);
@@ -74,7 +86,10 @@ struct SegmentedBlock {
     return target_;
   }
 
+  friend std::ostream& operator<<(std::ostream& os, const SegmentedBlock& b);
+
  private:
+  BlockID id_;
   SegmentedBlockTarget target_;
   std::vector<ir::Input> in_shape_;
   std::vector<torch::jit::Value*> inputs_;
@@ -84,6 +99,8 @@ struct SegmentedBlock {
   std::unordered_map<torch::jit::Value*, torch::jit::Value*> old_to_new_;
 };
 
+std::ostream& operator<<(std::ostream& os, const SegmentedBlock::SegmentedBlockTarget& t);
+
 } // namespace partitioning
 } // namespace core
 } // namespace trtorch
diff --git a/core/partitioning/partitioning.cpp b/core/partitioning/partitioning.cpp
@@ -275,81 +275,120 @@ bool checkLoopEvaluatable(torch::jit::Node* n) {
   return compile_to_trt;
 }
 
-std::vector<SegmentedBlock> segment_graph(torch::jit::Block* block, const PartitionInfo& partition_info) {
+bool should_run_in_trt(torch::jit::Node* n, const std::unordered_set<std::string>& torch_ops) {
+  // If the op is not supported by the conversion phase it should run in PyTorch
+  if (!conversion::OpSupported(n)) {
+    LOG_GRAPH("Node not supported by conversion: " << util::node_info(n));
+    return false;
+  }
+
+  // If the user specifies the op to run in Torch it should run in PyTorch
+  if (torch_ops.find(n->kind().toQualString()) != torch_ops.end()) {
+    LOG_GRAPH("Node explicitly set to run in torch: " << util::node_info(n));
+    return false;
+  }
+
+  // If the user specifies the module containing this op to run in torch it should run in PyTorch
+  const auto to_compile_sym = c10::Symbol::attr("to_compile");
+  if (n->hasAttribute(to_compile_sym) && n->i(to_compile_sym) == (int64_t) false) {
+    LOG_GRAPH("Node is within a module set to run in torch: " << util::node_info(n));
+    return false;
+  }
+
+  LOG_GRAPH("Node is going to run in TensorRT: " << util::node_info(n));
+  return true;
+}
+
+void finalize_block(PartitionedGraph& g, SegmentedBlock::SegmentedBlockTarget kind, std::vector<torch::jit::Node*>& nodes) {
+  SegmentedBlock::BlockID b_id= g.size();
+  LOG_DEBUG("Finalizing in progress " << SegmentedBlock::target_to_str(kind) << " block");
+  g.emplace_back(b_id, kind, nodes);
+  nodes.clear();
+  LOG_DEBUG(g.back());
+}
+
+PartitionedGraph segment_graph(torch::jit::Block* block, const PartitionInfo& partition_info) {
   auto min_block_size = partition_info.min_block_size;
-  std::unordered_set<std::string> forced_fallback_operators(
+  std::unordered_set<std::string> forced_fallback_ops(
       partition_info.forced_fallback_operators.begin(), partition_info.forced_fallback_operators.end());
 
   auto nodes = block->nodes();
-  std::vector<SegmentedBlock> segmented_blocks;
+  PartitionedGraph segmented_blocks;
 
   // segment the nodes
-  std::vector<torch::jit::Node*> tensorrt_nodes, pytorch_nodes;
+  std::vector<torch::jit::Node*> in_prog_trt_blk_nodes, in_prog_pyt_blk_nodes;
   for (const auto n : nodes) {
+    // Skip constant nodes as they are resources for both kinds of modules
     if (n->kind() == torch::jit::prim::Constant) {
       continue;
     }
 
-    std::string node_string(n->kind().toQualString());
-    auto has_compile_attribute = n->hasAttribute(c10::Symbol::attr("to_compile"));
-    if (conversion::OpSupported(n) && !forced_fallback_operators.count(node_string) &&
-        (!has_compile_attribute || n->i(c10::Symbol::attr("to_compile")) == (int64_t) true)) {
-      tensorrt_nodes.push_back(n);
-      if (tensorrt_nodes.size() >= min_block_size && !pytorch_nodes.empty()) {
-        segmented_blocks.emplace_back(SegmentedBlock::kTorch, pytorch_nodes);
-        pytorch_nodes.clear();
+    if (should_run_in_trt(n, forced_fallback_ops)) {
+      in_prog_trt_blk_nodes.push_back(n);
+
+      // If there is an active PyTorch block and we have passed the threshold for a valid TRT
+      // block then segment and reset the active PyTorch block
+      if (in_prog_trt_blk_nodes.size() >= min_block_size && !in_prog_pyt_blk_nodes.empty()) {
+        finalize_block(segmented_blocks, SegmentedBlock::kTorch, in_prog_pyt_blk_nodes);
       }
     } else {
-      if (tensorrt_nodes.size() >= min_block_size) {
-        segmented_blocks.emplace_back(SegmentedBlock::kTensorRT, tensorrt_nodes);
+      // If there is an active TRT block that is valid segment and reset the active TRT block
+      // otherwise add it to the active PyTorch block and reset
+      if (in_prog_trt_blk_nodes.size() >= min_block_size) {
+        finalize_block(segmented_blocks, SegmentedBlock::kTensorRT, in_prog_trt_blk_nodes);
       } else {
-        pytorch_nodes.insert(pytorch_nodes.end(), tensorrt_nodes.begin(), tensorrt_nodes.end());
+        LOG_DEBUG("In progress TRT block does not meet minimum block size requirements, therefore folding into in progress PyTorch block");
+        in_prog_pyt_blk_nodes.insert(in_prog_pyt_blk_nodes.end(), in_prog_trt_blk_nodes.begin(), in_prog_trt_blk_nodes.end());
       }
-      tensorrt_nodes.clear();
+      in_prog_trt_blk_nodes.clear();
       // if there is a prim::If then this if node will be encapsulated in a SegmentedBlock
       // we shouldn't inject node for this block in dependency analysis process
       if (n->kind() == torch::jit::prim::If) {
-        if (!pytorch_nodes.empty()) {
-          segmented_blocks.emplace_back(SegmentedBlock::kTorch, pytorch_nodes);
-          pytorch_nodes.clear();
+        LOG_DEBUG("Hit a conditional statement, finializing in progress PYT block and creating a new one for the conditional");
+        if (!in_prog_pyt_blk_nodes.empty()) {
+          finalize_block(segmented_blocks, SegmentedBlock::kTorch, in_prog_pyt_blk_nodes);
         }
-        segmented_blocks.emplace_back(SegmentedBlock::kTorch, std::vector<torch::jit::Node*>{n});
+        auto cond_node = std::vector<torch::jit::Node*>{n};
+        finalize_block(segmented_blocks, SegmentedBlock::kTorch, cond_node);
         continue;
       } else if (n->kind() == torch::jit::prim::Loop) {
-        if (!pytorch_nodes.empty()) {
-          segmented_blocks.emplace_back(SegmentedBlock::kTorch, pytorch_nodes);
-          pytorch_nodes.clear();
+        if (!in_prog_pyt_blk_nodes.empty()) {
+          finalize_block(segmented_blocks, SegmentedBlock::kTorch, in_prog_pyt_blk_nodes);
         }
         if (checkLoopEvaluatable(n)) {
-          tensorrt_nodes.push_back(n);
+          in_prog_trt_blk_nodes.push_back(n);
         } else {
-          segmented_blocks.emplace_back(SegmentedBlock::kTorch, std::vector<torch::jit::Node*>{n});
+          auto loop_node = std::vector<torch::jit::Node*>{n};
+          finalize_block(segmented_blocks, SegmentedBlock::kTorch, loop_node);
         }
         continue;
       }
-      pytorch_nodes.push_back(n);
+      in_prog_pyt_blk_nodes.push_back(n);
     }
   }
 
   // if there is any kTorch nodes left, then either the last nodes are kTorch or last nodes are kTensorRT but num <
   // min_block_size
-  if (!pytorch_nodes.empty()) {
-    pytorch_nodes.insert(pytorch_nodes.end(), tensorrt_nodes.begin(), tensorrt_nodes.end());
-    segmented_blocks.emplace_back(SegmentedBlock::kTorch, pytorch_nodes);
-  } else {
-    segmented_blocks.emplace_back(SegmentedBlock::kTensorRT, tensorrt_nodes);
+  if (in_prog_trt_blk_nodes.size() >= min_block_size) {
+    finalize_block(segmented_blocks, SegmentedBlock::kTensorRT, in_prog_trt_blk_nodes);
+  }
+
+  if (!in_prog_pyt_blk_nodes.empty()) {
+    in_prog_pyt_blk_nodes.insert(in_prog_pyt_blk_nodes.end(), in_prog_trt_blk_nodes.begin(), in_prog_trt_blk_nodes.end());
+    finalize_block(segmented_blocks, SegmentedBlock::kTorch, in_prog_pyt_blk_nodes);
   }
 
   return std::move(segmented_blocks);
 }
 
-std::vector<SegmentedBlock> Partition(
+PartitionedGraph Partition(
     torch::jit::Block* block,
     std::unordered_map<torch::jit::Value*, torch::jit::IValue>& input_ivalues_map,
     const PartitionInfo& partition_info) {
   LOG_DEBUG(partition_info);
   // segment lowering global graph into blocks
-  std::vector<SegmentedBlock> segmented_blocks = segment_graph(block, partition_info);
+  LOG_DEBUG("Parititioning source module into PyTorch and TensorRT sub blocks");
+  PartitionedGraph segmented_blocks = segment_graph(block, partition_info);
 
   // resolve nonTensor inputs/outputs
   resolveNonTensorInputs(segmented_blocks);
@@ -358,11 +397,22 @@ std::vector<SegmentedBlock> Partition(
   registerSegmentsOutputs(segmented_blocks, block);
 
   // run shape analysis on each segmented block
-  runShapeAnalysis(segmented_blocks, input_ivalues_map);
+  runShapeAnalysis(segmented_blocks, input_ivalues_map, at::kFloat);
+
+  LOG_INFO(segmented_blocks);
 
   return segmented_blocks;
 }
 
+std::ostream& operator<<(std::ostream& os, const PartitionedGraph& g) {
+  os << "Partitioned Graph: [";
+  for (auto b : g) {
+    os << b;
+  }
+  os << "]";
+  return os;
+}
+
 } // namespace partitioning
 } // namespace core
 } // namespace trtorch
diff --git a/core/partitioning/partitioning.h b/core/partitioning/partitioning.h
@@ -1,6 +1,7 @@
 #pragma once
 
 #include <vector>
+#include <iostream>
 
 #include "core/ir/ir.h"
 #include "core/partitioning/PartitionInfo.h"
@@ -17,11 +18,13 @@ typedef std::vector<SegmentedBlock> PartitionedGraph;
 
 PartitionedGraph segment_graph(torch::jit::Block* block, const PartitionInfo& partition_info);
 
-std::vector<SegmentedBlock> Partition(
+PartitionedGraph Partition(
     torch::jit::Block* block,
     std::unordered_map<torch::jit::Value*, torch::jit::IValue>& input_ivalues_map,
     const PartitionInfo& partition_info);
 
+std::ostream& operator<<(std::ostream& os, const PartitionedGraph& g);
+
 } // namespace partitioning
 } // namespace core
 } // namespace trtorch
diff --git a/core/util/BUILD b/core/util/BUILD
@@ -27,6 +27,9 @@ cc_library(
     hdrs = [
         "jit_util.h",
     ],
+    srcs = [
+        "jit_util.cpp"
+    ],
     deps = select({
         ":use_pre_cxx11_abi": ["@libtorch_pre_cxx11_abi//:libtorch"],
         "//conditions:default": ["@libtorch//:libtorch"],

diff --git a/core/util/jit_util.cpp b/core/util/jit_util.cpp
@@ -0,0 +1,69 @@
+#include "core/util/jit_util.h"
+
+namespace trtorch {
+namespace core {
+namespace util {
+
+c10::optional<at::ScalarType> getBlockFirstCalcDType(const std::shared_ptr<torch::jit::Block>& b) {
+  auto ns = b->nodes();
+
+  c10::optional<at::ScalarType> dtype = {};
+
+  // For each node check the inputs to find a prim:Constant, which will provide a static tensor.
+  // Use that tensor to determine operating dtype for the first calculation in the block
+  for (auto n : ns) {
+    if (n->kind() == torch::jit::prim::Constant) {
+      // Not really helpful to evaluate typing for constants
+      continue;
+    }
+
+    auto ins = n->inputs();
+    auto outs = n->outputs();
+
+    bool outputs_tensor = false;
+    for (auto o : outs) {
+      if (o->type() == c10::TensorType::get()) {
+        outputs_tensor = true;
+      }
+    }
+
+    if (outputs_tensor) {
+      // If all input tensors are block inputs then this node will not give us useful type info so move to the next one
+      std::unordered_set<torch::jit::Value*> node_input_set = {ins.begin(), ins.end()};
+
+      bool all_n_ins_are_b_ins = true;
+      for (auto b_in : b->inputs()) {
+        if (node_input_set.find(b_in) == node_input_set.end()) {
+          all_n_ins_are_b_ins = false;
+        }
+      }
+
+      if (all_n_ins_are_b_ins) {
+        continue;
+      }
+
+
+      // If node outputs a Tensor it might be a result of tensor calcuation so check to see
+      // if any inputs to the calculation can give us hints
+      c10::optional<torch::jit::Node*> const_tensor_n = {};
+
+      // Backtrace to constants which will immediately give us the Tensor type if possible
+      for (auto in : ins) {
+        if (in->type() == c10::TensorType::get()) {
+          if (in->node()->kind() == torch::jit::prim::Constant) {
+            auto const_ival = in->node()->get(c10::Symbol::attr("value"));
+            dtype = {const_ival.value().toTensor().scalar_type()};
+            goto exit_first_calc_dtype;
+          }
+        }
+      }
+    }
+  }
+
+exit_first_calc_dtype:
+  return dtype;
+}
+
+} // namespace util
+} // namespace core
+} // namespace trtorch
diff --git a/core/util/jit_util.h b/core/util/jit_util.h
@@ -52,6 +52,8 @@ inline std::string GetPyTorchSourceCode(const torch::jit::Node* n) {
   return source_code;
 }
 
+c10::optional<at::ScalarType> getBlockFirstCalcDType(const std::shared_ptr<torch::jit::Block>& b);
+
 } // namespace util
 } // namespace core
 } // namespace trtorch
diff --git a/core/util/logging/TRTorchLogger.cpp b/core/util/logging/TRTorchLogger.cpp
@@ -125,7 +125,7 @@ namespace {
 
 TRTorchLogger& get_global_logger() {
 #ifndef NDEBUG
-  static TRTorchLogger global_logger("[TRTorch - Debug Build] - ", LogLevel::kDEBUG, true);
+  static TRTorchLogger global_logger("[TRTorch - Debug Build] - ", LogLevel::kGRAPH, true);
 #else
   static TRTorchLogger global_logger("[TRTorch] - ", LogLevel::kERROR, false);
 #endif

diff --git a/tests/core/partitioning/BUILD b/tests/core/partitioning/BUILD
@@ -81,7 +81,7 @@ cc_test(
 )
 
 test_suite(
-    name = "partitioning_test",
+    name = "partitioning_tests",
     tests = [
         ":test_segmentation",
         ":test_shape_analysis",