add mul_lstm_fuse_pass ut (#37795)

* add mul_lstm_fuse_pass ut

* update mul_lstm_fuse_pass ut

* update ut

* update ut

* update ut

* add CPU ut cmake setting

* update ut

paddle/fluid/framework/ir/fc_lstm_fuse_pass.cc

@@ -68,13 +68,13 @@ MulLstmFusePass::MulLstmFusePass() {
       .IsType<bool>()
       .End()
       .AddAttr("gate_activation")
-      .IsStringIn({"sigmoid", "tanh", "relu", "identity"})
+      .IsStringIn({"sigmoid"})
       .End()
       .AddAttr("cell_activation")
-      .IsStringIn({"sigmoid", "tanh", "relu", "identity"})
+      .IsStringIn({"tanh", "relu", "identity"})
       .End()
       .AddAttr("candidate_activation")
-      .IsStringIn({"sigmoid", "tanh", "relu", "identity"})
+      .IsStringIn({"tanh", "relu", "identity"})
       .End();
   AddOpCompat(OpCompat("mul"))
       .AddInput("X")

python/paddle/fluid/tests/unittests/ir/inference/CMakeLists.txt

@@ -7,6 +7,13 @@ string(REPLACE ".py" "" TEST_TRT_IR_PASSES "${TEST_TRT_IR_PASSES}")
 file(GLOB TEST_TRT_CONVERTER RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "test_trt_convert_*.py")
 string(REPLACE ".py" "" TEST_TRT_CONVERTER "${TEST_TRT_CONVERTER}")
 
+# Only for cpu(mkl + openblas)
+set(TEST_INFERENCE_CPU_UT "test_mul_lstm_fuse_pass")
+
+foreach(CPU_UT ${TEST_INFERENCE_CPU_UT})
+  list(REMOVE_ITEM TEST_INFERENCE_IR_PASSES ${CPU_UT})
+endforeach()
+
 foreach(TEST_INFERENCE_IR_PASS ${TEST_TRT_IR_PASSES})
   list(REMOVE_ITEM TEST_INFERENCE_IR_PASSES ${TEST_INFERENCE_IR_PASS})
 endforeach()
@@ -53,6 +60,14 @@ if (WITH_MKLDNN AND TENSORRT_FOUND AND WITH_GPU)
   endforeach()
 endif()
 
+if (NOT WITH_MKLDNN AND NOT TENSORRT_FOUND AND NOT WITH_GPU)
+  foreach(target ${TEST_INFERENCE_CPU_UT})
+    py_test_modules(${target} MODULES ${target})
+  endforeach()
+
+set_tests_properties(test_mul_lstm_fuse_pass PROPERTIES TIMEOUT 300)
+endif()
+
 if(WITH_GPU AND TENSORRT_FOUND)
 set_tests_properties(test_trt_subgraph_pass PROPERTIES TIMEOUT 120)
 set_tests_properties(test_trt_activation_pass PROPERTIES TIMEOUT 120)

python/paddle/fluid/tests/unittests/ir/inference/test_mul_lstm_fuse_pass.py

+# Copyright (c) 2021 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.
+
+from auto_scan_test import PassAutoScanTest, SkipReasons
+from program_config import TensorConfig, ProgramConfig, OpConfig
+import numpy as np
+import paddle.inference as paddle_infer
+from functools import partial
+from typing import Optional, List, Callable, Dict, Any, Set
+import unittest
+
+import hypothesis
+from hypothesis import given, settings, seed, example, assume
+import hypothesis.strategies as st
+from functools import reduce
+
+
+class TestMulLstmFusePass(PassAutoScanTest):
+    def is_program_valid(self, program_config: ProgramConfig) -> bool:
+        return True
+
+    def sample_program_config(self, draw):
+        x_col = draw(st.sampled_from([1]))
+        y_col = draw(st.sampled_from([1]))
+        use_peepholes = draw(st.booleans())
+        is_reverse = draw(st.booleans())
+        gate_activation = draw(st.sampled_from(["sigmoid"]))
+        cell_activation = draw(st.sampled_from(["tanh", "relu", "identity"]))
+        candidate_activation = draw(
+            st.sampled_from(["tanh", "relu", "identity"]))
+        batch_size = draw(st.integers(min_value=1, max_value=40))
+
+        def generate_input():
+            shape = [batch_size, 128, 6, 120]
+            return np.full(shape, 0.01).astype(np.float32)
+
+        def generate_weight(shape):
+            return np.full(shape, 0.0001).astype(np.float32)
+
+        im2sequence_op = OpConfig(
+            type="im2sequence",
+            inputs={"X": ["input_data"]},
+            outputs={"Out": ["seq_out"]},
+            attrs={
+                "kernels": [6, 1],
+                "out_stride": [1, 1],
+                "paddings": [0, 0, 0, 0],
+                "strides": [1, 1]
+            })
+
+        mul_op = OpConfig(
+            type="mul",
+            inputs={"X": ["seq_out"],
+                    "Y": ["mul_weight"]},
+            outputs={"Out": ["mul_out"]},
+            attrs={"x_num_col_dims": x_col,
+                   "y_num_col_dims": y_col})
+
+        lstm_op = OpConfig(
+            type="lstm",
+            inputs={
+                "Input": ["mul_out"],
+                "Weight": ["lstm_weight"],
+                "Bias": ["lstm_bias"]
+            },
+            outputs={
+                "Hidden": ["lstm_hidden"],
+                "Cell": ["lstm_cell"],
+                "BatchGate": ["lstm_gate"],
+                "BatchCellPreAct": ["lstm_batch_cell"]
+            },
+            attrs={
+                'use_peepholes': use_peepholes,
+                'is_reverse': is_reverse,
+                'gate_activation': gate_activation,
+                'cell_activation': cell_activation,
+                'candidate_activation': candidate_activation,
+                'is_test': True
+            })
+
+        model_net = [im2sequence_op, mul_op, lstm_op]
+
+        if use_peepholes:
+            lstm_bias_shape = [1, 1050]
+        else:
+            lstm_bias_shape = [1, 600]
+
+        program_config = ProgramConfig(
+            ops=model_net,
+            weights={
+                "mul_weight":
+                TensorConfig(data_gen=partial(generate_weight, [768, 600])),
+                "lstm_weight":
+                TensorConfig(data_gen=partial(generate_weight, [150, 600])),
+                "lstm_bias":
+                TensorConfig(data_gen=partial(generate_weight, lstm_bias_shape))
+            },
+            inputs={
+                "input_data": TensorConfig(data_gen=partial(generate_input)),
+            },
+            outputs=["lstm_hidden"])
+
+        return program_config
+
+    def sample_predictor_configs(self, program_config):
+        config = self.create_inference_config()
+        yield config, ["im2sequence", "fusion_lstm"], (1e-5, 1e-5)
+
+    def test(self):
+        self.run_and_statis(
+            quant=False, max_duration=300, passes=["mul_lstm_fuse_pass"])
+
+
+if __name__ == "__main__":
+    unittest.main()