diff --git a/paddle/fluid/framework/ir/map_matmul_to_mul_pass.cc b/paddle/fluid/framework/ir/map_matmul_to_mul_pass.cc index 865b556f301c0dfa57d05e142d6925ca9ca543f4..734f8957ad09e9aad66f6dcf82a47cf008ad02f9 100644 --- a/paddle/fluid/framework/ir/map_matmul_to_mul_pass.cc +++ b/paddle/fluid/framework/ir/map_matmul_to_mul_pass.cc @@ -169,9 +169,6 @@ Flatten2MatmulFusePass::Flatten2MatmulFusePass() { .AddInput("X") .IsTensor() .End() - .AddInput("Y") - .IsTensor() - .End() .AddOutput("Out") .IsTensor() .End() @@ -179,7 +176,7 @@ Flatten2MatmulFusePass::Flatten2MatmulFusePass() { .IsTensor() .End() .AddAttr("axis") - .IsNumGE(0) + .IsNumEQ(1) .End(); AddOpCompat(OpCompat("mul")) @@ -222,7 +219,7 @@ Squeeze2MatmulFusePass::Squeeze2MatmulFusePass() { .IsBoolEQ(false) .End(); - AddOpCompat(OpCompat("Squeeze2")) + AddOpCompat(OpCompat("squeeze2")) .AddInput("X") .IsTensor() .End() @@ -593,10 +590,10 @@ Reshape2MatmulFusePass::Reshape2MatmulFusePass() { .IsNumLT(1.00001f) .End() .AddAttr("transpose_X") - .IsBoolEQ("False") + .IsBoolEQ(false) .End() .AddAttr("transpose_Y") - .IsBoolEQ("False") + .IsBoolEQ(false) .End(); AddOpCompat(OpCompat("mul")) diff --git a/python/paddle/fluid/tests/unittests/ir/inference/CMakeLists.txt b/python/paddle/fluid/tests/unittests/ir/inference/CMakeLists.txt index c4e5d98b80a8611dea4a54dd926346ca854b11f6..75d182bd01e7492ae7132e699900b006fd19c32a 100755 --- a/python/paddle/fluid/tests/unittests/ir/inference/CMakeLists.txt +++ b/python/paddle/fluid/tests/unittests/ir/inference/CMakeLists.txt @@ -83,9 +83,11 @@ if (WITH_MKLDNN AND TENSORRT_FOUND AND WITH_GPU) set_tests_properties(test_conv_act_mkldnn_fuse_pass PROPERTIES TIMEOUT 120) set_tests_properties(test_conv_elementwise_add2_act_fuse_pass PROPERTIES TIMEOUT 120) set_tests_properties(test_conv_elementwise_add_act_fuse_pass PROPERTIES TIMEOUT 120) - set_tests_properties(test_conv_elementwise_add_act_fuse_pass PROPERTIES TIMEOUT 90) set_tests_properties(test_matmul_scale_fuse_pass PROPERTIES TIMEOUT 60) set_tests_properties(test_matmul_v2_scale_fuse_pass PROPERTIES TIMEOUT 60) + set_tests_properties(test_flatten2_matmul_fuse_pass PROPERTIES TIMEOUT 240) + set_tests_properties(test_squeeze2_matmul_fuse_pass PROPERTIES TIMEOUT 240) + set_tests_properties(test_reshape2_matmul_fuse_pass PROPERTIES TIMEOUT 240) endif() if (WITH_MKLDNN) diff --git a/python/paddle/fluid/tests/unittests/ir/inference/test_flatten2_matmul_fuse_pass.py b/python/paddle/fluid/tests/unittests/ir/inference/test_flatten2_matmul_fuse_pass.py new file mode 100644 index 0000000000000000000000000000000000000000..6cd9ae970bb588a4834ddcae5ed6b1f0c9bfaf92 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/ir/inference/test_flatten2_matmul_fuse_pass.py @@ -0,0 +1,181 @@ +# 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, IgnoreReasons +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, reproduce_failure +import hypothesis.strategies as st + + +class TestFlatten2MatmulFusePass(PassAutoScanTest): + """ + x_var + | + flatten2 + \ + flatten2_out_var y_var + \ / + matmul bias_var + \ / + elementwise_add + """ + + def sample_predictor_configs(self, program_config): + # TRT + # config = self.create_trt_inference_config() + # config.enable_tensorrt_engine( + # max_batch_size=10, + # workspace_size=102400, + # min_subgraph_size=0, + # precision_mode=paddle_infer.PrecisionType.Float32, + # use_static=False, + # use_calib_mode=False) + # yield config, ['mul', 'elementwise_add'], (1e-5, 1e-5) + + # cpu + config = self.create_inference_config(use_gpu=False) + yield config, ["mul", "elementwise_add"], (1e-5, 1e-5) + + # for gpu + config = self.create_inference_config(use_gpu=True) + yield config, ["mul", "elementwise_add"], (1e-5, 1e-5) + + def add_ignore_pass_case(self): + # Here we put some skip rules to avoid known bugs + def teller1(program_config, predictor_config): + if predictor_config.tensorrt_engine_enabled(): + # On 3080, the results of MatMul and Mul are different + # When the input Y is weight + return True + + # On TRT when the input Y is weight, Mul is converted to FC + if "matmul_y" not in program_config.weights \ + or "bias" not in program_config.weights: + return True + + y_shape = list(program_config.weights["matmul_y"].shape) + bias_shape = program_config.weights["bias"].shape + axis = program_config.ops[2].attrs["axis"] + # bias should be [mul_y_shape[-1]] + if axis == 0 or bias_shape[0] != y_shape[1] or len( + bias_shape) != 1: + return True + return False + + self.add_ignore_check_case( + teller1, + IgnoreReasons.PASS_ACCURACY_ERROR, + "The pass error on TRT while shape of bias is not [out_size].", ) + + def sample_program_config(self, draw): + # 1. Generate shape and attr of flatten2 + x_shape = draw( + st.lists( + st.integers( + min_value=1, max_value=10), min_size=4, max_size=4)) + # [a, b, c, d] => [a, b*c*d] + flatten_axis = 1 + flatten_shape = [x_shape[0], x_shape[1] * x_shape[2] * x_shape[3]] + + # 2. Generate attr:transpose_X/transpose_Y/alpha of matmul + alpha = 1.0 + transpose_X = False + transpose_Y = False + + # 3. Generate legal shape of input:Y of matmul + y_shape = draw( + st.lists( + st.integers( + min_value=1, max_value=8), min_size=2, max_size=2)) + y_shape[0] = flatten_shape[1] + + # 4. Generate legal attr:axis of elementwise_add + axis = draw(st.integers(min_value=-1, max_value=1)) + if axis == 0: + bias_shape = [flatten_shape[0], ] + elif axis == 1: + bias_shape = [y_shape[1]] + else: + bias_shape = [flatten_shape[0], y_shape[1]] + if draw(st.booleans()): + bias_shape[1] = 1 + + flatten2_op = OpConfig( + "flatten2", + inputs={"X": ["flatten2_x"], }, + axis=flatten_axis, + outputs={"Out": ["flatten2_out"], + "XShape": ["xshape"]}, ) + matmul_op = OpConfig( + "matmul", + inputs={"X": ["flatten2_out"], + "Y": ["matmul_y"]}, + outputs={"Out": ["matmul_out"]}, + alpha=alpha, + transpose_X=transpose_X, + transpose_Y=transpose_Y, + fused_reshape_X=[], + fused_reshape_Y=[], + fused_transpose_X=[], + fused_transpose_Y=[], + fused_reshape_Out=[], + fused_transpose_Out=[], ) + + add_op = OpConfig( + "elementwise_add", + inputs={"X": ["matmul_out"], + "Y": ["bias"]}, + outputs={"Out": ["add_out"]}, + axis=axis, ) + + ops = [flatten2_op, matmul_op, add_op] + + if draw(st.integers(min_value=1, max_value=10)) <= 8: + program_config = ProgramConfig( + ops=ops, + weights={ + "matmul_y": TensorConfig(shape=y_shape), + "bias": TensorConfig(shape=bias_shape), + }, + inputs={"flatten2_x": TensorConfig(shape=x_shape), }, + outputs=ops[-1].outputs["Out"], ) + else: + program_config = ProgramConfig( + ops=ops, + weights={}, + inputs={ + "flatten2_x": TensorConfig(shape=x_shape), + "matmul_y": TensorConfig(shape=y_shape), + "bias": TensorConfig(shape=bias_shape), + }, + outputs=ops[-1].outputs["Out"], ) + return program_config + + def test(self): + self.run_and_statis( + quant=False, + max_examples=50, + max_duration=1000, + passes=["flatten2_matmul_fuse_pass"]) + + +if __name__ == "__main__": + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/ir/inference/test_reshape2_matmul_fuse_pass.py b/python/paddle/fluid/tests/unittests/ir/inference/test_reshape2_matmul_fuse_pass.py new file mode 100644 index 0000000000000000000000000000000000000000..951ec8e4e8ef42268e480ffe5730e6fa944cbea8 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/ir/inference/test_reshape2_matmul_fuse_pass.py @@ -0,0 +1,179 @@ +# 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, IgnoreReasons +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, reproduce_failure +import hypothesis.strategies as st + + +class TestReshape2MatmulFusePass(PassAutoScanTest): + """ + x_var + | + reshape2 + \ + reshape2_out_var y_var + \ / + matmul bias_var + \ / + elementwise_add + """ + + def sample_predictor_configs(self, program_config): + # TRT + # config = self.create_trt_inference_config() + # config.enable_tensorrt_engine( + # max_batch_size=10, + # workspace_size=102400, + # min_subgraph_size=0, + # precision_mode=paddle_infer.PrecisionType.Float32, + # use_static=False, + # use_calib_mode=False) + # yield config, ['mul', 'elementwise_add'], (1e-5, 1e-5) + + # cpu + config = self.create_inference_config(use_gpu=False) + yield config, ["mul", "elementwise_add"], (1e-5, 1e-5) + + # for gpu + config = self.create_inference_config(use_gpu=True) + yield config, ["mul", "elementwise_add"], (1e-5, 1e-5) + + def add_ignore_pass_case(self): + # Here we put some skip rules to avoid known bugs + def teller1(program_config, predictor_config): + if predictor_config.tensorrt_engine_enabled(): + # On 3080, the results of MatMul and Mul are different + # When the input Y is weight + return True + + # On TRT when the input Y is weight, Mul is converted to FC + if "matmul_y" not in program_config.weights \ + or "bias" not in program_config.weights: + return True + + y_shape = list(program_config.weights["matmul_y"].shape) + bias_shape = program_config.weights["bias"].shape + axis = program_config.ops[2].attrs["axis"] + # bias should be [mul_y_shape[-1]] + if axis == 0 or bias_shape[0] != y_shape[1] or len( + bias_shape) != 1: + return True + return False + + self.add_ignore_check_case( + teller1, + IgnoreReasons.PASS_ACCURACY_ERROR, + "The pass error on TRT while shape of bias is not [out_size].", ) + + def sample_program_config(self, draw): + # 1. Generate shape and attr of reshape2 + reshape = draw( + st.lists( + st.integers( + min_value=1, max_value=10), min_size=2, max_size=2)) + x_shape = reshape + [1, 1] + + # 2. Generate attr:transpose_X/transpose_Y/alpha of matmul + alpha = 1.0 + transpose_X = False + transpose_Y = False + + # 3. Generate legal shape of input:Y of matmul + y_shape = draw( + st.lists( + st.integers( + min_value=1, max_value=8), min_size=2, max_size=2)) + y_shape[0] = x_shape[1] + + # 4. Generate legal attr:axis of elementwise_add + axis = draw(st.integers(min_value=-1, max_value=1)) + if axis == 0: + bias_shape = [x_shape[0]] + elif axis == 1: + bias_shape = [y_shape[1]] + else: + bias_shape = [x_shape[0], y_shape[1]] + if draw(st.booleans()): + bias_shape[1] = 1 + + reshape2_op = OpConfig( + "reshape2", + inputs={"X": ["reshape2_x"], }, + shape=reshape, + outputs={"Out": ["reshape2_out"], + "XShape": ["xshape"]}, ) + matmul_op = OpConfig( + "matmul", + inputs={"X": ["reshape2_out"], + "Y": ["matmul_y"]}, + outputs={"Out": ["matmul_out"]}, + alpha=alpha, + transpose_X=transpose_X, + transpose_Y=transpose_Y, + fused_reshape_X=[], + fused_reshape_Y=[], + fused_transpose_X=[], + fused_transpose_Y=[], + fused_reshape_Out=[], + fused_transpose_Out=[], ) + + add_op = OpConfig( + "elementwise_add", + inputs={"X": ["matmul_out"], + "Y": ["bias"]}, + outputs={"Out": ["add_out"]}, + axis=axis, ) + + ops = [reshape2_op, matmul_op, add_op] + + if draw(st.integers(min_value=1, max_value=10)) <= 8: + program_config = ProgramConfig( + ops=ops, + weights={ + "matmul_y": TensorConfig(shape=y_shape), + "bias": TensorConfig(shape=bias_shape), + }, + inputs={"reshape2_x": TensorConfig(shape=x_shape), }, + outputs=ops[-1].outputs["Out"], ) + else: + program_config = ProgramConfig( + ops=ops, + weights={}, + inputs={ + "reshape2_x": TensorConfig(shape=x_shape), + "matmul_y": TensorConfig(shape=y_shape), + "bias": TensorConfig(shape=bias_shape), + }, + outputs=ops[-1].outputs["Out"], ) + return program_config + + def test(self): + self.run_and_statis( + quant=False, + max_examples=50, + max_duration=1000, + passes=["reshape2_matmul_fuse_pass"]) + + +if __name__ == "__main__": + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/ir/inference/test_squeeze2_matmul_fuse_pass.py b/python/paddle/fluid/tests/unittests/ir/inference/test_squeeze2_matmul_fuse_pass.py new file mode 100644 index 0000000000000000000000000000000000000000..605dc4edbe8c60dfc63838eeeb8f6936444e96d1 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/ir/inference/test_squeeze2_matmul_fuse_pass.py @@ -0,0 +1,187 @@ +# 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, IgnoreReasons +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, reproduce_failure +import hypothesis.strategies as st + + +class TestSqueeze2MatmulFusePass(PassAutoScanTest): + """ + x_var + | + squeeze2 + \ + squeeze2_out_var y_var + \ / + matmul bias_var + \ / + elementwise_add + """ + + def sample_predictor_configs(self, program_config): + # cpu + config = self.create_inference_config(use_gpu=False) + yield config, ["mul", "elementwise_add"], (1e-5, 1e-5) + + # for gpu + config = self.create_inference_config(use_gpu=True) + yield config, ["mul", "elementwise_add"], (1e-5, 1e-5) + + # TRT + # config = self.create_trt_inference_config() + # config.enable_tensorrt_engine( + # max_batch_size=10, + # workspace_size=10240, + # min_subgraph_size=0, + # precision_mode=paddle_infer.PrecisionType.Float32, + # use_static=False, + # use_calib_mode=False) + # yield config, ['mul', 'elementwise_add'], (1e-5, 1e-5) + + def add_ignore_pass_case(self): + # Here we put some skip rules to avoid known bugs + def teller1(program_config, predictor_config): + if predictor_config.tensorrt_engine_enabled(): + # On 3080, the results of MatMul and Mul are different + # When the input Y is weight + return True + + # On TRT when the input Y is weight, Mul is converted to FC + predictor_config.exp_disable_tensorrt_ops(["elementwise_add"]) + if "matmul_y" not in program_config.weights \ + or "bias" not in program_config.weights: + return True + + y_shape = list(program_config.weights["matmul_y"].shape) + bias_shape = program_config.weights["bias"].shape + axis = program_config.ops[2].attrs["axis"] + # bias should be [mul_y_shape[-1]] + if axis == 0 or bias_shape[0] != y_shape[1] or len( + bias_shape) != 1: + return True + return False + + self.add_ignore_check_case( + teller1, + IgnoreReasons.PASS_ACCURACY_ERROR, + "The pass error on TRT while shape of bias is not [out_size].", ) + + def sample_program_config(self, draw): + # 1. Generate shape of input:X of squeeze2 + x_shape = draw( + st.lists( + st.integers( + min_value=1, max_value=8), min_size=2, max_size=2)) + # axes of squeeze2 == [2, 3] + x_shape += [1, 1] + axes = [2, 3] + + # 2. Generate attr:transpose_X/transpose_Y/alpha of matmul + alpha = 1.0 + transpose_X = False + transpose_Y = False + + # 3. Generate legal shape of input:Y of matmul + y_shape = draw( + st.lists( + st.integers( + min_value=1, max_value=8), min_size=2, max_size=2)) + y_shape[0] = x_shape[1] + + # 4. Generate legal attr:axis of elementwise_add + axis = draw(st.integers(min_value=-1, max_value=1)) + if axis == 0 or axis == -1: + bias_shape = [x_shape[0], y_shape[1]] + else: + bias_shape = [y_shape[1], ] + if draw(st.booleans()): + bias_shape[-1] = 1 + if len(bias_shape) == 2 and draw(st.booleans()): + bias_shape[0] = 1 + + axis = 0 + bias_shape = [2, ] + x_shape = [2, 1, 1, 1] + y_shape = [1, 2] + + squeeze2_op = OpConfig( + "squeeze2", + inputs={"X": ["squeeze2_x"], }, + axes=axes, + outputs={"Out": ["squeeze2_out"], + "XShape": ["xshape"]}, ) + matmul_op = OpConfig( + "matmul", + inputs={"X": ["squeeze2_out"], + "Y": ["matmul_y"]}, + outputs={"Out": ["matmul_out"]}, + alpha=alpha, + transpose_X=transpose_X, + transpose_Y=transpose_Y, + fused_reshape_X=[], + fused_reshape_Y=[], + fused_transpose_X=[], + fused_transpose_Y=[], + fused_reshape_Out=[], + fused_transpose_Out=[], ) + + add_op = OpConfig( + "elementwise_add", + inputs={"X": ["matmul_out"], + "Y": ["bias"]}, + outputs={"Out": ["add_out"]}, + axis=axis, ) + + ops = [squeeze2_op, matmul_op, add_op] + + if draw(st.integers(min_value=1, max_value=10)) <= 8: + program_config = ProgramConfig( + ops=ops, + weights={ + "matmul_y": TensorConfig(shape=y_shape), + "bias": TensorConfig(shape=bias_shape), + }, + inputs={"squeeze2_x": TensorConfig(shape=x_shape), }, + outputs=ops[-1].outputs["Out"], ) + else: + program_config = ProgramConfig( + ops=ops, + weights={}, + inputs={ + "squeeze2_x": TensorConfig(shape=x_shape), + "matmul_y": TensorConfig(shape=y_shape), + "bias": TensorConfig(shape=bias_shape), + }, + outputs=ops[-1].outputs["Out"], ) + return program_config + + def test(self): + self.run_and_statis( + quant=False, + max_examples=50, + max_duration=1000, + passes=["squeeze2_matmul_fuse_pass"]) + + +if __name__ == "__main__": + unittest.main()