New relay backend for meta schedule task extraction

Author: masahi

python/tvm/meta_schedule/integration.py

@@ -18,7 +18,10 @@
 from contextlib import contextmanager
 from typing import Callable, Dict, List, Optional, Union
 
-from tvm._ffi import register_object
+import numpy as np
+import tvm.runtime.ndarray as nd
+
+from tvm._ffi import register_object, get_global_func
 from tvm.ir import IRModule, transform
 from tvm.relay import Any
 from tvm.relay import Function as RelayFunc
@@ -230,6 +233,20 @@ def extract_task_from_relay(
         The tasks extracted from this network
     """
 
+    extract_task_func = get_global_func("relay.backend.MetaScheduleExtractTask")
+    assert extract_task_func
+
+    target = Target(target) if isinstance(target, str) else target
+
+    for name, param in params.items():
+        if isinstance(param, np.ndarray):
+            params[name] = nd.array(param)
+
+    with transform.PassContext(opt_level=opt_level):
+        with target:
+            tasks = extract_task_func(mod, target, params)
+            return tasks
+
     @contextmanager
     def _autotvm_silencer():
         from tvm import autotvm  # pylint: disable=import-outside-toplevel

src/relay/backend/metaschedule_task_extraction.cc

@@ -0,0 +1,86 @@
+/*
+ * 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.
+ */
+
+#include <tvm/meta_schedule/integration.h>
+#include <tvm/relay/expr.h>
+#include <tvm/relay/expr_functor.h>
+#include <tvm/relay/function.h>
+#include <tvm/target/target.h>
+
+#include "../../te/operation/create_primfunc.h"
+#include "te_compiler_cache.h"
+#include "utils.h"
+
+namespace tvm {
+namespace relay {
+namespace backend {
+namespace metaschedule {
+
+using meta_schedule::ExtractedTask;
+
+Array<ExtractedTask> ExtractTask(IRModule mod, Target target, Map<String, Constant> params) {
+  if (params.size()) {
+    std::unordered_map<std::string, runtime::NDArray> params_;
+    BaseFunc base_func = mod->Lookup("main");
+    ICHECK(base_func->IsInstance<FunctionNode>());
+    auto f = relay::backend::BindParamsByName(Downcast<Function>(base_func), params_);
+    auto gvar = mod->GetGlobalVar("main");
+    mod->Add(gvar, f);
+  }
+
+  Array<Pass> pass_seqs = relay::backend::GetPassPrefix(/*is_homogenous=*/true, /*is_vm=*/true);
+  pass_seqs.push_back(transform::FuseOps());
+
+  transform::Sequential seq(pass_seqs);
+  auto opt_mod = seq(std::move(mod));
+
+  Array<ExtractedTask> tasks;
+  LOG(INFO) << opt_mod;
+  LOG(INFO) << opt_mod->Lookup("main");
+  PostOrderVisit(opt_mod->Lookup("main"), [target, &tasks](const Expr& exp) {
+    if (exp->IsInstance<FunctionNode>()) {
+      Function relay_func = Downcast<Function>(exp);
+      if (relay_func->HasNonzeroAttr(attr::kPrimitive)) {
+        LOG(INFO) << relay_func;
+        Array<te::Tensor> outputs;
+        std::string fused_name;
+        std::tie(outputs, fused_name) = tec::LowerTECompute(target, relay_func);
+        LOG(INFO) << fused_name;
+        LOG(INFO) << outputs;
+        auto prim_func = tir::CreatePrimFunc(outputs);
+        auto prim_fn_var = GlobalVar(fused_name);
+        auto relay_mod = IRModule({{prim_fn_var, relay_func}});
+        auto tir_mod = IRModule({{prim_fn_var, prim_func}});
+        tasks.push_back(ExtractedTask(prim_fn_var->name_hint, relay_mod, target, {tir_mod}));
+      }
+    }
+  });
+
+  return tasks;
+}
+
+TVM_REGISTER_GLOBAL("relay.backend.MetaScheduleExtractTask")
+    .set_body_typed([](IRModule mod, Target target, Map<String, Constant> params) {
+      return ExtractTask(mod, target, params);
+    });
+
+}  // namespace metaschedule
+}  // namespace backend
+}  // namespace relay
+}  // namespace tvm

src/relay/backend/te_compiler_cache.cc

@@ -754,6 +754,24 @@ CachedFunc ShapeFuncFor(const Function& prim_func, const Target& target,
   return MakeShapeFunc().Create(prim_func, target, renamer);
 }
 
+std::pair<Array<te::Tensor>, std::string> LowerTECompute(Target target, const Function& relay_func,
+                                                         bool return_inputs) {
+  LowerToTECompute lower_te_compute(target);
+  auto outputs = lower_te_compute.Lower(relay_func, [&](std::string name) { return name; });
+  // Following ScheduleBuilder, remove placeholder ops from outputs.
+  tvm::Array<te::Tensor> tensor_outs;
+  for (const auto& tensor : outputs) {
+    if (!tensor->op.as<te::PlaceholderOpNode>()) {
+      tensor_outs.push_back(tensor);
+    }
+  }
+  if (return_inputs) {
+    return std::make_pair(Concat(lower_te_compute.fn_inputs_, tensor_outs),
+                          lower_te_compute.candidate_name_);
+  }
+  return std::make_pair(tensor_outs, lower_te_compute.candidate_name_);
+}
+
 /*!
  * \brief Get unique name from name.
  * \param name The orginal name.

src/relay/backend/te_compiler_cache.h

@@ -204,6 +204,8 @@ class CCacheValue : public ObjectRef {
 
 Array<IndexExpr> GetShape(const Array<IndexExpr>& shape);
 
+std::pair<Array<te::Tensor>, std::string> LowerTECompute(Target target, const Function& relay_func, bool return_inputs=true);
+
 /*!
  * \brief Create schedule for target.
  * \param source_func The primitive function to be lowered.