Support fetch in new ir (#54826)

* add fetch kernel

* support fetch var in new ir

* fix bug

* polish code

* change array equal to np.testing

paddle/fluid/framework/new_executor/interpreter/interpreter_util.cc

@@ -952,10 +952,11 @@ void BuildOpFuncList(
 
     auto op_name = attr_map.at("op_name").dyn_cast<::ir::StrAttribute>().data();
 
-    if (op_name == "pd.fetch" || op_name == "builtin.combine") {
+    if (op_name == "builtin.combine") {
       VLOG(6) << "skip process pd.fetch op";
       continue;
     }
+
     op_func_node.phi_op_name_ = op_name;
 
     ::ir::OpInfo op_info = ctx->GetRegisteredOpInfo(op_name);

paddle/fluid/framework/new_executor/standalone_executor.cc

@@ -68,9 +68,11 @@ StandaloneExecutor::StandaloneExecutor(const platform::Place& place,
 
     if (FLAGS_enable_new_ir_in_executor) {
       VLOG(6) << "begin to translate" << std::endl;
-      auto base_progrm = paddle::TranslateLegacyProgramToProgram(*program);
+      auto base_program = paddle::TranslateLegacyProgramToProgram(*program);
+
       auto kernel_program =
-          paddle::dialect::PdOpLowerToKernelPass(base_progrm.get());
+          paddle::dialect::PdOpLowerToKernelPass(base_program.get());
+
       interpretercores_.emplace_back(std::make_unique<InterpreterCore>(
           place_, std::move(kernel_program), scope_, execution_config));
     } else {

paddle/fluid/ir/dialect/pd_op.yaml

@@ -17,8 +17,20 @@
     data_transform: {}
   attrs:
   - {typename: str, name: name}
-  outputs: []
+  outputs:
+  - {typename: Tensor, name: out, optional: false, intermediate: false}
   no_need_buffer: null
   data_transform: null
+  infer_meta:
+    func: UnchangedInferMeta
+    param: [x]
+  kernel:
+    func: [fetch]
+    param: [x]
+    backend: null
+    layout: null
+    data_type: null
+    dispatch: {fetch: null}
+    force_backend: null
   inplace: null
   backward: null

paddle/fluid/ir/phi_kernel_adaptor/phi_kernel_adaptor.h

@@ -55,15 +55,14 @@ class PhiKernelAdaptor {
   void run(ir::Program* program) {
     auto block = program->block();
     std::unordered_map<ir::Value, std::string> name_map;
-    std::cerr << "run here" << std::endl;
+
     ir::BuildScope(block, scope_, &name_map);
-    std::cerr << "after buid scope" << std::endl;
+
     auto* dev_ctx = phi::DeviceContextPool::Instance().Get(phi::CPUPlace());
     phi::Place cpu_place(phi::AllocationType::CPU);
     for (auto it = block->begin(); it != block->end(); ++it) {
       VLOG(6) << "begin to run op " << (*it)->name();
 
-      std::cerr << (*it)->name() << std::endl;
       auto attr_map = (*it)->attributes();
 
       paddle::dialect::OpYamlInfoInterface op_info_interface =
@@ -91,10 +90,6 @@ class PhiKernelAdaptor {
       }
       auto found_it = phi_kernels.find(kernel_key);
       if (found_it == phi_kernels.end()) {
-        std::cerr << "kernel name " << runtime_info.kernel_func[0] << std::endl;
-        std::cerr << "kernel key " << kernel_key.backend() << "\t"
-                  << kernel_key.dtype() << "\t" << kernel_key.layout()
-                  << std::endl;
         PADDLE_THROW(paddle::platform::errors::NotFound(
             "can not found kerenl for [%s]", (*it)->name()));
       } else {

paddle/fluid/ir/phi_kernel_adaptor/phi_kernel_util.cc

@@ -35,6 +35,7 @@
 #include "paddle/fluid/framework/tensor_ref_array.h"
 #include "paddle/fluid/ir/dialect/kernel_attribute.h"
 #include "paddle/fluid/ir/dialect/pd_attribute.h"
+#include "paddle/phi/core/enforce.h"
 
 #include "glog/logging.h"
 
@@ -57,18 +58,10 @@ void BuildScope(ir::Block* block,
     if (op_name == "pd.fetch") {
       // fetch is a very special op, with no output
       for (size_t i = 0; i < input_num; ++i) {
-        auto ptr = (*it)->operand(i).source();
-        auto var_name = attr_map.at("name").dyn_cast<ir::StrAttribute>().data();
-
-        PADDLE_ENFORCE_EQ(
-            name_map->count(ptr),
-            true,
-            phi::errors::PreconditionNotMet(
-                "input of fetch op should in name mape, var_name is [%s]",
-                var_name));
-
-        scope->Rename(name_map->at(ptr), var_name);
-        (*name_map)[ptr] = var_name;
+        auto var = scope->Var("fetch");
+        auto fetch_list = var->GetMutable<paddle::framework::FetchList>();
+        // for now only support one fetch
+        fetch_list->resize(1);
       }
       continue;
     }
@@ -245,10 +238,21 @@ void BuildInferMetaContext(
     }
   }
 
-  ir::Value out_ptr = op->result(0);
-  auto name = name_map.at(out_ptr);
-
-  ctx->EmplaceBackOutput(scope->Var(name)->Get<phi::DenseTensor>());
+  // update here, support fetch list for now
+  // [todo update here]
+  if (op->attributes().count("op_name") &&
+      (op->attributes().at("op_name").dyn_cast<ir::StrAttribute>().data() ==
+       "pd.fetch")) {
+    // process fetch op
+    auto fetch_var = scope->Var("fetch");
+    auto* fetch_list = fetch_var->GetMutable<paddle::framework::FetchList>();
+    auto* out_tensor = &(PADDLE_GET(phi::DenseTensor, fetch_list->at(0)));
+    ctx->EmplaceBackOutput(out_tensor);
+  } else {
+    ir::Value out_ptr = op->result(0);
+    auto name = name_map.at(out_ptr);
+    ctx->EmplaceBackOutput(scope->Var(name)->Get<phi::DenseTensor>());
+  }
 }
 
 void BuildPhiKernelContext(
@@ -367,17 +371,26 @@ void BuildPhiKernelContext(
     }
   }
 
-  ir::Value out_ptr = op->result(0);
-  auto name = name_map.at(out_ptr);
-
-  ctx->EmplaceBackOutput(const_cast<phi::DenseTensor*>(
-      &(scope->Var(name)->Get<phi::DenseTensor>())));
-
-  if (output_map != nullptr) {
-    // only deal with single input for now, [todo] need support multi input like
-    // concat
-    size_t tmp_id = std::atol(name.substr(4, 100).c_str());
-    (*output_map)["out"].push_back(tmp_id);
+  if (op->attributes().count("op_name") &&
+      (op->attributes().at("op_name").dyn_cast<ir::StrAttribute>().data() ==
+       "pd.fetch")) {
+    // process fetch op
+    auto fetch_var = scope->Var("fetch");
+    auto* fetch_list = fetch_var->GetMutable<paddle::framework::FetchList>();
+    auto* out_tensor = &(PADDLE_GET(phi::DenseTensor, fetch_list->at(0)));
+    ctx->EmplaceBackOutput(out_tensor);
+  } else {
+    ir::Value out_ptr = op->result(0);
+    auto name = name_map.at(out_ptr);
+    ctx->EmplaceBackOutput(const_cast<phi::DenseTensor*>(
+        &(scope->Var(name)->Get<phi::DenseTensor>())));
+
+    if (output_map != nullptr) {
+      // only deal with single input for now, [todo] need support multi input
+      // like concat
+      size_t tmp_id = std::atol(name.substr(4, 100).c_str());
+      (*output_map)["out"].push_back(tmp_id);
+    }
   }
 }
 

paddle/fluid/ir_adaptor/translator/op_translator.cc

@@ -572,6 +572,7 @@ ir::Operation* FetchOpHandler(ir::IrContext* ctx,
       {"name", ir::StrAttribute::get(ctx, op_desc.InputArgumentNames()[0])},
   };
 
+  op_output_types.push_back(op_inputs[0].type());
   ir::Operation* operation =
       ir::Operation::Create(op_inputs, attribute_map, op_output_types, op_info);
   program->block()->push_back(operation);

paddle/phi/kernels/cpu/fetch_kernel.cc

+// Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed 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 "paddle/phi/kernels/fetch_kernel.h"
+
+#include "paddle/phi/backends/cpu/cpu_context.h"
+#include "paddle/phi/core/kernel_registry.h"
+
+namespace phi {
+
+template <typename T, typename Context>
+void FetchKernel(const Context& dev_ctx,
+                 const DenseTensor& x,
+                 DenseTensor* out) {
+  phi::Copy(dev_ctx, x, phi::CPUPlace(), true, out);
+}
+}  // namespace phi
+PD_REGISTER_KERNEL(fetch,
+                   CPU,
+                   ALL_LAYOUT,
+                   phi::FetchKernel,
+                   float,
+                   double,
+                   int,
+                   int64_t,
+                   phi::dtype::complex<float>,
+                   phi::dtype::complex<double>,
+                   bool) {}

paddle/phi/kernels/fetch_kernel.h

+// Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed 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.
+
+#pragma once
+
+#include "paddle/phi/core/dense_tensor.h"
+
+namespace phi {
+
+/**
+ * @brief This kernel is used to fetch tensor from scope
+ * @param  ctx     device context
+ * @param  x       the input tensor of fetch
+ * @param  out     the output tensor of fetch
+ */
+template <typename T, typename Context>
+void FetchKernel(const Context& ctx, const DenseTensor& x, DenseTensor* out);
+
+}  // namespace phi

test/ir/new_ir/test_standalone_new_ir.py

@@ -22,30 +22,22 @@ import paddle
 paddle.enable_static()
 
 
-# class TestNewIr(unittest.TestCase):
-#     def test_with_new_ir(self):
-#         place = paddle.CPUPlace()
-#         exe = paddle.static.Executor(place)
+class TestNewIr(unittest.TestCase):
+    def test_with_new_ir(self):
+        place = paddle.CPUPlace()
+        exe = paddle.static.Executor(place)
 
-#         x = paddle.ones([2, 2], dtype="float32")
-#         y = paddle.ones([2, 2], dtype="float32")
+        x = paddle.ones([2, 2], dtype="float32")
+        y = paddle.ones([2, 2], dtype="float32")
 
-#         z = x + y
-#         out = exe.run(
-#             paddle.static.default_main_program(), {}, fetch_list=[z.name]
-#         )
+        z = x + y
+        out = exe.run(
+            paddle.static.default_main_program(), {}, fetch_list=[z.name]
+        )
 
-#         gold_res = np.ones([2, 2], dtype="float32") * 2
+        gold_res = np.ones([2, 2], dtype="float32") * 2
 
-#         self.assertEqual(
-#             np.array_equal(
-#                 np.array(
-#                     paddle.static.global_scope().find_var(z.name).get_tensor()
-#                 ),
-#                 gold_res,
-#             ),
-#             True,
-#         )
+        np.testing.assert_array_equal(out[0], gold_res)
 
 
 class TestCombineOp(unittest.TestCase):
@@ -63,15 +55,7 @@ class TestCombineOp(unittest.TestCase):
 
         gold_res = np.ones([2, 2], dtype="float32") * 2
 
-        self.assertEqual(
-            np.array_equal(
-                np.array(
-                    paddle.static.global_scope().find_var(z.name).get_tensor()
-                ),
-                gold_res,
-            ),
-            True,
-        )
+        np.testing.assert_array_equal(out[0], gold_res)
 
 
 if __name__ == "__main__":