[GitHub] [tvm] areusch commented on a change in pull request #7938: Improved MLF to contain workspace info

[GitBox <gi...@apache.org>]

areusch commented on a change in pull request #7938:
URL: https://github.com/apache/tvm/pull/7938#discussion_r625212228



##########
File path: src/relay/backend/graph_executor_codegen.cc
##########
@@ -189,9 +192,109 @@ class GraphExecutorCodegen : public backend::MemoizedExprTranslator<std::vector<
     targets_ = targets;
   }
 
+  /*!
+   * \brief Calculate the storage required to store the type of relay.Expr
+   *
+   * \param func The relay expr for which the storage is calculated
+   */
+  int64_t CalculateRelayExprSizeBytes(const Type& expr_type) {
+    if (expr_type->IsInstance<TupleTypeNode>()) {
+      auto tuple_type = Downcast<TupleType>(expr_type);
+      int64_t size = 0;
+      for (const auto& field : tuple_type->fields) {
+        size += CalculateRelayExprSizeBytes(field);
+      }
+      return size;
+    }
+    auto tensor_type = expr_type.as<TensorTypeNode>();
+    auto shape = tensor_type->shape;
+    int num_of_elements = 1;
+    for (const auto& dim_index_expr : shape) {
+      if (dim_index_expr->IsInstance<IntImmNode>()) {
+        num_of_elements *= dim_index_expr.as<IntImmNode>()->value;
+      } else {
+        // If shape is dynamic, we cannot calculate workspace in compile time.
+        num_of_elements = 0;
+      }
+    }
+    auto element_size = tensor_type->dtype.bytes();
+    return element_size * num_of_elements;
+  }
+
+  /*!
+   * \brief Update the "main" control function's metadata
+   *
+   * \param func The main function that contains calls to relay primitive functions
+   */
+  void UpdateMainWorkspaceSize(const Function& func) {
+    std::unordered_map<int, std::unordered_map<int, int>> sid_workspace;
+    std::unordered_map<int, int> device_workspace;
+    std::unordered_map<int, int> device_io;
+    std::unordered_map<int, int> device_consts;
+
+    for (const auto& kv : storage_device_map_) {
+      auto sids = kv.second[0];
+      auto devices = kv.second[1];
+      CHECK_EQ(sids.size(), devices.size());
+      for (uint32_t i = 0; i < sids.size(); i++) {
+        sid_workspace[devices[i]][sids[i]] = 0;
+        device_io[devices[i]] = 0;
+        device_consts[devices[i]] = 0;
+      }
+    }
+
+    for (const auto& kv : storage_device_map_) {
+      auto size_bytes = CalculateRelayExprSizeBytes(kv.first->checked_type());
+      auto sids = kv.second[0];
+      auto devices = kv.second[1];
+      if (kv.first->IsInstance<ConstantNode>()) {
+        for (const auto& dev : devices) {
+          device_consts[dev] += size_bytes;
+        }
+        continue;
+      } else if (kv.first->IsInstance<VarNode>() || kv.first == func->body) {
+        for (const auto& dev : devices) {
+          device_io[dev] += size_bytes;
+        }
+        continue;
+      }
+      for (uint32_t i = 0; i < sids.size(); i++) {
+        if (size_bytes > sid_workspace[devices[i]][sids[i]]) {

Review comment:
       can you add a brief note why you want the max here?

##########
File path: src/relay/backend/utils.cc
##########
@@ -0,0 +1,44 @@
+/*
+ * Licensed to the Apache Software Foundation (ASF) under one
+ * or more contributor license agreements.  See the NOTICE file
+ * distributed with this work for additional information
+ * regarding copyright ownership.  The ASF licenses this file
+ * to you under the Apache License, Version 2.0 (the
+ * "License"); you may not use this file except in compliance
+ * with the License.  You may obtain a copy of the License at
+ *
+ *   http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+ * KIND, either express or implied.  See the License for the
+ * specific language governing permissions and limitations
+ * under the License.
+ */
+
+/*!
+ * \file relay/backend/util.cc
+ * \brief Relay backend utilities.
+ */
+
+#include "utils.h"
+
+namespace tvm {
+namespace relay {
+namespace backend {
+
+void FunctionInfo::SetWorkspaceSize(Target tgt, tvm::Integer size) {
+  (*this)->workspace_sizes.Set(tgt, size);
+}
+
+TVM_REGISTER_NODE_TYPE(FunctionInfoNode);
+TVM_REGISTER_GLOBAL("relay.backend.FunctionInfo").set_body_typed([]() { return FunctionInfo(); });

Review comment:
       how come we need this?

##########
File path: tests/python/unittest/test_micro_model_library_format.py
##########
@@ -167,11 +199,68 @@ def @main(%a : Tensor[(1, 2), uint8], %b : Tensor[(1, 2), float32], %c : Tensor[
             assert "p0" in params
 
 
+@tvm.testing.requires_micro
+def test_export_model_library_format_workspace():
+    with utils.TempDirectory.set_keep_for_debug(True):

Review comment:
       nit: remove

##########
File path: src/relay/backend/graph_executor_codegen.cc
##########
@@ -189,9 +192,109 @@ class GraphExecutorCodegen : public backend::MemoizedExprTranslator<std::vector<
     targets_ = targets;
   }
 
+  /*!
+   * \brief Calculate the storage required to store the type of relay.Expr
+   *
+   * \param func The relay expr for which the storage is calculated
+   */
+  int64_t CalculateRelayExprSizeBytes(const Type& expr_type) {
+    if (expr_type->IsInstance<TupleTypeNode>()) {
+      auto tuple_type = Downcast<TupleType>(expr_type);
+      int64_t size = 0;
+      for (const auto& field : tuple_type->fields) {
+        size += CalculateRelayExprSizeBytes(field);
+      }
+      return size;
+    }
+    auto tensor_type = expr_type.as<TensorTypeNode>();
+    auto shape = tensor_type->shape;
+    int num_of_elements = 1;
+    for (const auto& dim_index_expr : shape) {
+      if (dim_index_expr->IsInstance<IntImmNode>()) {
+        num_of_elements *= dim_index_expr.as<IntImmNode>()->value;
+      } else {
+        // If shape is dynamic, we cannot calculate workspace in compile time.
+        num_of_elements = 0;
+      }
+    }
+    auto element_size = tensor_type->dtype.bytes();
+    return element_size * num_of_elements;
+  }
+
+  /*!
+   * \brief Update the "main" control function's metadata
+   *
+   * \param func The main function that contains calls to relay primitive functions
+   */
+  void UpdateMainWorkspaceSize(const Function& func) {
+    std::unordered_map<int, std::unordered_map<int, int>> sid_workspace;

Review comment:
       can you add a one-liner comment to each of these explaining key and value?

##########
File path: src/relay/backend/graph_executor_codegen.cc
##########
@@ -551,10 +713,14 @@ class GraphExecutorCodegen : public backend::MemoizedExprTranslator<std::vector<
   Map<Expr, Array<IntegerArray>> storage_device_map_;
   /*! \brief lowered funcs */
   std::unordered_map<std::string, IRModule> lowered_funcs_;
+  /*! \brief lowered funcs */
+  Map<String, FunctionInfo> function_metadata_;
   /*! \brief name map */
   std::unordered_map<std::string, size_t> name_map_;
   /*! \brief compile engine */
   CompileEngine compile_engine_;
+  /*! \brief main function name */
+  const String kMainFuncStr = "main_func";

Review comment:
       i wonder if this should be `mod_name`, as passed to GraphExecutorFactory?

##########
File path: src/relay/backend/graph_executor_codegen.cc
##########
@@ -189,9 +192,109 @@ class GraphExecutorCodegen : public backend::MemoizedExprTranslator<std::vector<
     targets_ = targets;
   }
 
+  /*!
+   * \brief Calculate the storage required to store the type of relay.Expr
+   *
+   * \param func The relay expr for which the storage is calculated
+   */
+  int64_t CalculateRelayExprSizeBytes(const Type& expr_type) {
+    if (expr_type->IsInstance<TupleTypeNode>()) {
+      auto tuple_type = Downcast<TupleType>(expr_type);
+      int64_t size = 0;
+      for (const auto& field : tuple_type->fields) {
+        size += CalculateRelayExprSizeBytes(field);
+      }
+      return size;
+    }
+    auto tensor_type = expr_type.as<TensorTypeNode>();
+    auto shape = tensor_type->shape;
+    int num_of_elements = 1;
+    for (const auto& dim_index_expr : shape) {
+      if (dim_index_expr->IsInstance<IntImmNode>()) {
+        num_of_elements *= dim_index_expr.as<IntImmNode>()->value;
+      } else {
+        // If shape is dynamic, we cannot calculate workspace in compile time.
+        num_of_elements = 0;
+      }
+    }
+    auto element_size = tensor_type->dtype.bytes();
+    return element_size * num_of_elements;
+  }
+
+  /*!
+   * \brief Update the "main" control function's metadata
+   *
+   * \param func The main function that contains calls to relay primitive functions
+   */
+  void UpdateMainWorkspaceSize(const Function& func) {
+    std::unordered_map<int, std::unordered_map<int, int>> sid_workspace;
+    std::unordered_map<int, int> device_workspace;

Review comment:
       would suggest to move this to line 268, where it's built up, so that it's clear it's not needed til then.




-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

For queries about this service, please contact Infrastructure at:
users@infra.apache.org