manupa-arm commented on a change in pull request #7938:
URL: https://github.com/apache/tvm/pull/7938#discussion_r626686759
##########
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:
I did a different change -- kind of unified the inits. See if you agree.
##########
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:
Done
--
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:
[email protected]
