[GitHub] [tvm] Mousius commented on a change in pull request #9565: [TIR][USMP] Integrating USMP to AoT Executor

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

Mousius commented on a change in pull request #9565:
URL: https://github.com/apache/tvm/pull/9565#discussion_r776250758



##########
File path: src/relay/backend/aot_executor_codegen.cc
##########
@@ -645,32 +632,120 @@ class AOTExecutorCodegen : public MixedModeVisitor {
         // TODO(giuseros): we should allocate this once outside the PrimFunc
         // so we don't pay the price of allocation for every inference
         if (!allocated[sid]) {
-          body = tir::Allocate(sids_table_[sid], DataType::Int(8), {size}, tir::const_true(), body);
+          PointerType ptype = Downcast<PointerType>(sids_table_[sid]->type_annotation);
+          DataType element_type = Downcast<PrimType>(ptype->element_type)->dtype;
+          body = tir::Allocate(sids_table_[sid], element_type, {size}, tir::const_true(), body);
         }
         allocated[sid] = true;
       }
     }
 
-    // Define the attributes
-    body = tir::AttrStmt(PrimExpr(), tvm::tir::attr::device_type, 1, body);
-    body = tir::AttrStmt(PrimExpr(), tvm::tir::attr::device_id, 0, body);
-
     // Define the PrimFunc attributes
     Map<String, ObjectRef> dict_attrs;
     String run_func_name =
         runtime::get_name_mangled(mod_name, runtime::symbol::tvm_run_func_suffix);
     dict_attrs.Set("global_symbol", run_func_name);
     dict_attrs.Set("runner_function", Bool(true));
+    dict_attrs.Set(tvm::attr::kTarget, target_host_);
 
     tir::Stmt device_activations = GenerateAllDeviceHook("Activate");
     tir::Stmt device_deactivations = GenerateAllDeviceHook("Deactivate");
     tir::Stmt final_body = tir::SeqStmt({device_activations, body, device_deactivations});
 
     // Make the PrimFunc
-    return tir::PrimFunc(main_signature_, final_body, VoidType(), Map<tir::Var, tir::Buffer>(),
+    return tir::PrimFunc(main_signature_, final_body, VoidType(), main_buffer_map_,
                          DictAttrs(dict_attrs));
   }
 
+  /*!
+   * brief Access IO vars using the buffer vars and
+   * not the actual var.
+   */
+  tir::Var GetBufferVarForIO(int index) { return main_buffer_map_[main_signature_[index]]->data; }
+
+  /*!
+   * brief Create tir::Var for input/output while updating
+   * the buffer_maps.
+   */
+  void CreateIOVar(const Expr& expr, std::string name) {
+    if (expr->IsInstance<TupleNode>()) {
+      Tuple tuple = Downcast<Tuple>(expr);
+      for (unsigned i = 0; i < tuple->fields.size(); i++) {
+        CreateIOVar(tuple->fields[i], name + std::to_string(i) + "_");
+      }
+    } else {
+      tir::Var var = tir::Var(name, DataType::Handle());
+      main_signature_.push_back(var);
+      auto tensor_type = expr->checked_type().as<TensorTypeNode>();
+      DataType elem_type = tensor_type->dtype;
+      tir::Var buffer_var =
+          tir::Var(name + "_buffer_var", PointerType(PrimType(elem_type), "global"));
+      tir::Buffer buffer = tir::Buffer(buffer_var, elem_type, tensor_type->shape, {}, 0,
+                                       name + "_buffer", 16, 1, tir::BufferType::kDefault);
+      main_buffer_map_.Set(var, buffer);
+    }
+  }
+
+  /*!
+   * brief This function is a wrapper to run memory planning
+   * followed by recording the latest workspaces required.
+   */
+  IRModule PlanMemoryLoweredModule(const IRModule& mod) {
+    transform::PassContext pass_ctx = transform::PassContext::Current();
+    bool enable_usmp = pass_ctx->GetConfig<Bool>("tir.usmp.enable", Bool(false)).value();
+
+    IRModule lowered_mod = mod->ShallowCopy();
+    Executor executor_config = mod->GetAttr<Executor>(tvm::attr::kExecutor).value();
+    Integer workspace_byte_alignment =
+        executor_config->GetAttr<Integer>("workspace-byte-alignment").value_or(16);
+    if (enable_usmp) {

Review comment:
       Would be easier to read these two blocks in their own functions, such as:
   ```c++
       if (enable_usmp) {
          return PlanMemoryLoweredModuleWithUSMP(mod);
       } else {
          return PlanMemoryLoweredModuleWithoutUSMP(mod);
       }
   ```

##########
File path: src/target/source/source_module.cc
##########
@@ -26,6 +26,9 @@
 #include <tvm/runtime/ndarray.h>
 #include <tvm/runtime/packed_func.h>
 #include <tvm/runtime/registry.h>
+#include <tvm/tir/function.h>
+#include <tvm/tir/stmt_functor.h>
+#include <tvm/tir/transform.h>

Review comment:
       It's likely that these are the reason why your Windows CI builds are failing as iirc `source_module.cc` only gets linked with the runtime, not the entirety of TVM?

##########
File path: src/target/source/codegen_source_base.cc
##########
@@ -113,5 +113,15 @@ void CodeGenSourceBase::EndScope(int scope_id) {
   indent_ -= 2;
 }
 
+std::string CodeGenSourceBase::SanitiseName(std::string name) {

Review comment:
       Can we use https://github.com/apache/tvm/blob/main/src/relay/backend/name_transforms.cc#L87-L95 ? 

##########
File path: src/tir/transforms/make_unpacked_api.cc
##########
@@ -57,33 +57,26 @@ PrimFunc MakeUnpackedAPI(PrimFunc&& func) {
   const Stmt nop = Evaluate(0);
   std::vector<Stmt> device_init;
 
-  // Create arg to buffer binder
-  std::unordered_map<const VarNode*, PrimExpr> vmap;
-  ArgBinder binder(&vmap);

Review comment:
       This is a great improvement :smile_cat: 

##########
File path: include/tvm/tir/usmp/algo/algo.h
##########
@@ -0,0 +1,62 @@
+/*
+ * 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 tir/usmp/algo/algo.h

Review comment:
       nit: is there a reason to not just have an `algorithms.h` instead of an `algo.h` nested in an  `algo` folder. `algo` is also unnecessarily ambiguous.

##########
File path: tests/python/relay/aot/test_c_device_api.py
##########
@@ -136,44 +136,29 @@ def test_device_api_hooks_unpacked_api(device_api_main_func):
 
     # Activate Device
     assert (
-        str(main_func.body[0][0].value)
-        == "@tir.call_extern("
-        + '"TVMDeviceEthosUActivate",'
-        + " device_context_ethos_u: handle,"
-        + " dtype=int32)"
+        str(main_func.body[0])
+        == "tir.call_extern(" + '"TVMDeviceEthosUActivate",' + " device_context_ethos_u)\n"
     )
     # Open Device
+    print(str(main_func.body[1]))

Review comment:
       Is this meant to still be here? If so it would make sense to print the entire `body` before any asserts?




-- 
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.

To unsubscribe, e-mail: commits-unsubscribe@tvm.apache.org

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