CMakeLists.txt 15.3 KB
Newer Older
1 2
set(pass_file ${PADDLE_BINARY_DIR}/paddle/fluid/inference/api/paddle_inference_pass.h.tmp)
set(pass_file_final ${PADDLE_BINARY_DIR}/paddle/fluid/inference/api/paddle_inference_pass.h)
L
luotao1 已提交
3
file(WRITE ${pass_file} "// Generated by the paddle/fluid/framework/ir/CMakeLists.txt.  DO NOT EDIT!\n\n")
L
luotao1 已提交
4
file(APPEND ${pass_file} "\#pragma once\n")
L
luotao1 已提交
5
file(APPEND ${pass_file} "\#include \"paddle/fluid/framework/ir/pass.h\"\n")
Y
Yan Chunwei 已提交
6

7
copy_if_different(${pass_file} ${pass_file_final})
8

9 10 11
add_subdirectory(fuse_optimizer_ops_pass)
add_subdirectory(memory_optimize_pass)
add_subdirectory(multi_devices_graph_pass)
12
if(NOT APPLE AND NOT WIN32 AND (WITH_GPU OR WITH_ROCM))
13 14
    add_subdirectory(fusion_group)
endif()
Y
Yan Chunwei 已提交
15 16

# Usage: pass_library(target inference) will append to paddle_inference_pass.h
17
unset(INFER_IR_PASSES CACHE) # clear the global variable
Y
Yan Chunwei 已提交
18
function(pass_library TARGET DEST)
L
luotao1 已提交
19 20
    set(options "")
    set(oneValueArgs "")
21
    set(multiValueArgs SRCS DEPS DIR)
22 23
    set(targetPrefix "")

24 25
    cmake_parse_arguments(pass_library "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
    if(pass_library_DIR)
26
        cc_library(${TARGET} SRCS ${pass_library_DIR}/${TARGET}.cc DEPS graph_pattern_detector pass fuse_pass_base op_version_registry ${pass_library_DEPS})
27
    else()
28
        cc_library(${TARGET} SRCS ${TARGET}.cc DEPS graph_pattern_detector pass fuse_pass_base op_version_registry ${pass_library_DEPS})
29 30
    endif()

Y
Yan Chunwei 已提交
31 32
    # add more DEST here, such as train, dist and collect USE_PASS into a file automatically.
    if (${DEST} STREQUAL "base" OR ${DEST} STREQUAL "inference")
33 34 35
        if(NOT CMAKE_BUILD_TYPE STREQUAL "Release")
            message(STATUS "add pass ${TARGET} ${DEST}")
        endif()
Y
Yan Chunwei 已提交
36
        file(APPEND ${pass_file} "USE_PASS(${TARGET});\n")
37
        set(INFER_IR_PASSES ${INFER_IR_PASSES} ${TARGET} CACHE INTERNAL "")
Y
Yan Chunwei 已提交
38
    endif()
L
luotao1 已提交
39 40
endfunction()

Y
yuyang18 已提交
41
cc_library(node SRCS node.cc DEPS proto_desc)
Y
Yan Chunwei 已提交
42
cc_library(graph SRCS graph.cc DEPS node pretty_log)
X
better  
Xin Pan 已提交
43
cc_library(graph_helper SRCS graph_helper.cc DEPS graph)
X
Xin Pan 已提交
44
cc_library(pass SRCS pass.cc DEPS graph node graph_helper)
45
cc_library(graph_traits SRCS graph_traits.cc DEPS graph)
H
Huihuang Zheng 已提交
46
cc_library(cost_model SRCS cost_model.cc DEPS executor graph profiler proto_desc device_tracer)
H
Huihuang Zheng 已提交
47 48 49 50 51 52 53

SET(GRAPH_PATTERN_DETECTOR_DEPS graph graph_helper graph_traits)
if (WITH_TESTING)
    SET(GRAPH_PATTERN_DETECTOR_DEPS ${GRAPH_PATTERN_DETECTOR_DEPS} gtest)
endif(WITH_TESTING)
cc_library(graph_pattern_detector SRCS graph_pattern_detector.cc DEPS ${GRAPH_PATTERN_DETECTOR_DEPS})

54
cc_library(op_compat_sensible_pass SRCS op_compat_sensible_pass.cc DEPS graph_pattern_detector op_def_api pass)
55
cc_library(subgraph_detector SRCS subgraph_detector.cc DEPS graph_pattern_detector executor)
56
cc_library(fuse_pass_base SRCS fuse_pass_base.cc DEPS op_compat_sensible_pass)
57
cc_library(placement_pass_base SRCS placement_pass_base.cc DEPS pass)
L
luotao1 已提交
58

59
cc_library(coalesce_grad_tensor_pass SRCS coalesce_grad_tensor_pass.cc DEPS graph graph_helper)
60

Y
Yan Chunwei 已提交
61 62
pass_library(graph_to_program_pass base)
pass_library(graph_viz_pass base)
63
pass_library(lock_free_optimize_pass base DEPS string_helper)
Y
Yan Chunwei 已提交
64 65 66
pass_library(fc_fuse_pass inference)
pass_library(attention_lstm_fuse_pass inference)
pass_library(fc_lstm_fuse_pass inference)
67
pass_library(embedding_fc_lstm_fuse_pass inference)
68
pass_library(fc_gru_fuse_pass inference)
Y
Yan Chunwei 已提交
69
pass_library(seq_concat_fc_fuse_pass inference)
70
pass_library(multi_batch_merge_pass base)
S
Sylwester Fraczek 已提交
71
pass_library(conv_bn_fuse_pass inference)
T
tensor-tang 已提交
72
pass_library(seqconv_eltadd_relu_fuse_pass inference)
T
tensor-tang 已提交
73
pass_library(seqpool_concat_fuse_pass inference)
74
pass_library(seqpool_cvm_concat_fuse_pass inference)
T
tensor-tang 已提交
75
pass_library(repeated_fc_relu_fuse_pass inference)
76
pass_library(squared_mat_sub_fuse_pass inference)
77
pass_library(is_test_pass base)
78 79
pass_library(conv_elementwise_add_act_fuse_pass inference)
pass_library(conv_elementwise_add2_act_fuse_pass inference)
N
nhzlx 已提交
80
pass_library(conv_elementwise_add_fuse_pass inference)
81
pass_library(transpose_flatten_concat_fuse_pass inference)
82
pass_library(identity_scale_op_clean_pass base)
Q
qingqing01 已提交
83
pass_library(sync_batch_norm_pass base)
L
luotao1 已提交
84
pass_library(runtime_context_cache_pass base)
N
nhzlx 已提交
85
pass_library(quant_conv2d_dequant_fuse_pass inference)
86
pass_library(shuffle_channel_detect_pass inference)
87
pass_library(delete_quant_dequant_op_pass inference)
88
pass_library(delete_quant_dequant_filter_op_pass inference)
89 90
pass_library(delete_weight_dequant_linear_op_pass inference)
pass_library(delete_quant_dequant_linear_op_pass inference)
D
denglin-github 已提交
91
pass_library(delete_dropout_op_pass inference)
92
pass_library(simplify_with_basic_ops_pass base)
93
pass_library(fc_elementwise_layernorm_fuse_pass base)
94
pass_library(skip_layernorm_fuse_pass base)
95
pass_library(multihead_matmul_fuse_pass inference)
96 97
pass_library(adaptive_pool2d_convert_global_pass inference)
pass_library(unsqueeze2_eltwise_fuse_pass inference)
98
pass_library(layer_norm_fuse_pass inference)
99
pass_library(add_support_int8_pass inference)
H
heliqi 已提交
100
pass_library(matmul_scale_fuse_pass inference)
101
pass_library(gpu_cpu_map_matmul_to_mul_pass inference)
102
pass_library(mixed_precision_configure_pass inference)
103 104
pass_library(generate_pass DEPS pass_desc_proto)
target_link_libraries(generate_pass pass_desc_proto)
105 106 107

if(WITH_TENSORRT)
    pass_library(trt_map_matmul_to_mul_pass inference)
108 109
    pass_library(preln_embedding_eltwise_layernorm_fuse_pass inference)
    pass_library(preln_skip_layernorm_fuse_pass inference)
110 111
endif()

112
if(WITH_GPU OR WITH_ROCM)
113
    pass_library(cudnn_placement_pass base DEPS placement_pass_base)
114
    pass_library(embedding_eltwise_layernorm_fuse_pass inference)
115
endif()
116

W
Wojciech Uss 已提交
117
if(WITH_MKLDNN)
118
    pass_library(mkldnn_placement_pass base DEPS placement_pass_base DIR mkldnn)
119
    pass_library(mkldnn_inplace_pass inference DEPS mkldnn_placement_pass op_registry elementwise_add_op gelu_op activation_op softmax_op softmax DIR mkldnn)
120
    pass_library(depthwise_conv_mkldnn_pass base DIR mkldnn)
121
    pass_library(conv_affine_channel_mkldnn_fuse_pass inference DIR mkldnn)
122 123 124 125
    pass_library(conv_bias_mkldnn_fuse_pass inference DIR mkldnn)
    pass_library(conv_activation_mkldnn_fuse_pass inference DIR mkldnn)
    pass_library(conv_concat_relu_mkldnn_fuse_pass inference DIR mkldnn)
    pass_library(conv_elementwise_add_mkldnn_fuse_pass inference DIR mkldnn)
126
    pass_library(fc_elementwise_add_mkldnn_fuse_pass inference DIR mkldnn)
127
    pass_library(scale_matmul_fuse_pass inference DIR mkldnn)
128 129
    pass_library(cpu_bfloat16_placement_pass inference DIR mkldnn)
    pass_library(cpu_bfloat16_pass inference DIR mkldnn)
130
    pass_library(fc_mkldnn_pass inference DIR mkldnn)
131
    pass_library(interpolate_mkldnn_pass inference DIR mkldnn)
132
    pass_library(softplus_activation_mkldnn_fuse_pass inference DIR mkldnn)
133
    pass_library(shuffle_channel_mkldnn_detect_pass inference DIR mkldnn)
134
    pass_library(fc_act_mkldnn_fuse_pass inference DIR mkldnn)
135
    pass_library(elt_act_mkldnn_fuse_pass inference DIR mkldnn)
136 137 138
    pass_library(cpu_quantize_placement_pass base DIR mkldnn)
    pass_library(cpu_quantize_pass inference DIR mkldnn)
    pass_library(cpu_quantize_squash_pass inference DIR mkldnn)
139
    pass_library(reshape_transpose_matmul_mkldnn_fuse_pass inference DIR mkldnn)
140
    pass_library(reshape_transpose_matmul_v2_mkldnn_fuse_pass inference DIR mkldnn)
141
    pass_library(matmul_transpose_reshape_fuse_pass inference DIR mkldnn)
142
    pass_library(matmul_v2_transpose_reshape_fuse_pass inference DIR mkldnn)
143
    pass_library(batch_norm_act_fuse_pass inference DIR mkldnn)
144
    pass_library(multi_gru_fuse_pass inference DIR mkldnn)
145
    pass_library(multi_gru_seq_fuse_pass inference DIR mkldnn)
B
baoachun 已提交
146
    pass_library(quant_dequant_mkldnn_pass inference DIR mkldnn)
147
    pass_library(compute_propagate_scales_mkldnn_pass inference DIR mkldnn)
W
Wojciech Uss 已提交
148
endif()
149

A
Allen Guo 已提交
150 151 152 153 154 155 156 157
if(WITH_IPU)
    pass_library(forward_graph_extract_pass base DIR ipu)
    pass_library(optimizer_extract_pass base DIR ipu)
    pass_library(optimizer_state_align_pass base DIR ipu)
    pass_library(ipu_graph_builder_pass base DIR ipu)
    pass_library(ipu_runtime_replacer_pass base DIR ipu)
    pass_library(inference_process_pass base DIR ipu)
    pass_library(inference_postprocess_pass base DIR ipu)
158
    pass_library(popart_canonicalization_pass base DIR ipu)
A
Allen Guo 已提交
159 160 161 162 163 164
    pass_library(ipu_inplace_pass base DIR ipu)
    pass_library(infer_shape_pass base DIR ipu)
    pass_library(delete_scale_op_pass base DIR ipu)
    pass_library(avg_shard_pass base DIR ipu)
endif()

Z
Zhen Wang 已提交
165
cc_library(fuse_bn_act_pass SRCS fuse_bn_act_pass.cc DEPS pass graph_pattern_detector )
Z
Zhang Ting 已提交
166
cc_library(fuse_bn_add_act_pass SRCS fuse_bn_add_act_pass.cc DEPS pass graph_pattern_detector )
C
chengduo 已提交
167
cc_library(fuse_elewise_add_act_pass SRCS fuse_elewise_add_act_pass.cc DEPS pass graph_pattern_detector )
168
cc_library(fuse_gemm_epilogue_pass SRCS fuse_gemm_epilogue_pass.cc DEPS pass graph_pattern_detector )
169
cc_library(fuse_relu_depthwise_conv_pass SRCS fuse_relu_depthwise_conv_pass.cc DEPS pass graph_pattern_detector )
C
chengduo 已提交
170

L
luotao1 已提交
171
set(GLOB_PASS_LIB ${PASS_LIBRARY} CACHE INTERNAL "Global PASS library")
X
Xin Pan 已提交
172

173
cc_library(pass_builder SRCS pass_builder.cc DEPS pass)
174
cc_library(pass_test_util SRCS pass_test_util.cc DEPS graph pass)
175

X
Xin Pan 已提交
176
cc_test(node_test SRCS node_test.cc DEPS node)
X
Xin Pan 已提交
177 178 179
cc_test(pass_test SRCS pass_test.cc DEPS graph pass graph_helper)
cc_test(graph_test SRCS graph_test.cc DEPS graph graph_helper op_registry)
cc_test(graph_helper_test SRCS graph_helper_test.cc DEPS graph graph_helper op_registry)
X
Xin Pan 已提交
180
cc_test(graph_to_program_pass_test SRCS graph_to_program_pass_test.cc DEPS graph_to_program_pass)
H
Huihuang Zheng 已提交
181
cc_test(cost_model_test SRCS cost_model_test.cc DEPS cost_model op_registry)
182
cc_test(test_graph_pattern_detector SRCS graph_pattern_detector_tester.cc DEPS graph_pattern_detector)
183
cc_test(test_op_compat_sensible_pass SRCS op_compat_sensible_pass_tester.cc DEPS op_compat_sensible_pass)
Y
YUNSHEN XIE 已提交
184 185 186
cc_test(test_fc_fuse_pass_cc SRCS fc_fuse_pass_tester.cc DEPS fc_fuse_pass framework_proto)
cc_test(test_fc_lstm_fuse_pass_cc SRCS fc_lstm_fuse_pass_tester.cc DEPS fc_lstm_fuse_pass framework_proto)
cc_test(test_fc_gru_fuse_pass_cc SRCS fc_gru_fuse_pass_tester.cc DEPS fc_gru_fuse_pass framework_proto)
187
cc_test(test_seqpool_concat_fuse_pass SRCS seqpool_concat_fuse_pass_tester.cc DEPS seqpool_concat_fuse_pass framework_proto)
188
cc_test(test_seqpool_cvm_concat_fuse_pass SRCS seqpool_cvm_concat_fuse_pass_tester.cc DEPS seqpool_cvm_concat_fuse_pass framework_proto)
Y
YUNSHEN XIE 已提交
189
cc_test(test_repeated_fc_relu_fuse_pass_cc SRCS repeated_fc_relu_fuse_pass_tester.cc DEPS repeated_fc_relu_fuse_pass framework_proto)
190
cc_test(test_is_test_pass SRCS is_test_pass_tester.cc DEPS is_test_pass)
191
cc_test(test_simplify_with_basic_ops_pass SRCS simplify_with_basic_ops_pass_tester.cc DEPS simplify_with_basic_ops_pass)
192
cc_test(test_fc_elementwise_layernorm_fuse_pass_cc SRCS fc_elementwise_layernorm_fuse_pass_tester.cc DEPS fc_elementwise_layernorm_fuse_pass)
193
cc_test(test_skip_layernorm_fuse_pass SRCS skip_layernorm_fuse_pass_tester.cc DEPS skip_layernorm_fuse_pass)
194
cc_test(test_multihead_matmul_fuse_pass SRCS multihead_matmul_fuse_pass_tester.cc DEPS multihead_matmul_fuse_pass)
Y
YUNSHEN XIE 已提交
195
cc_test(test_conv_bn_fuse_pass_cc SRCS conv_bn_fuse_pass_tester.cc DEPS conv_bn_fuse_pass)
196
cc_test(test_adaptive_pool2d_convert_global_pass SRCS adaptive_pool2d_convert_global_pass_tester.cc DEPS adaptive_pool2d_convert_global_pass)
197
cc_test(test_unsqueeze2_eltwise_fuse_pass_cc SRCS unsqueeze2_eltwise_fuse_pass_tester.cc DEPS unsqueeze2_eltwise_fuse_pass)
198
cc_test(test_generate_pass_cc SRCS generate_pass_tester.cc DEPS generate_pass pass_desc_proto)
199
if(WITH_GPU OR WITH_ROCM)
200
    cc_test(test_embedding_eltwise_layernorm_fuse_pass SRCS embedding_eltwise_layernorm_fuse_pass_tester.cc DEPS embedding_eltwise_layernorm_fuse_pass)
201 202
    cc_test(test_cudnn_placement_pass SRCS cudnn_placement_pass_tester.cc DEPS cudnn_placement_pass)
endif()
Q
qingqing01 已提交
203 204 205
if(NOT WIN32)
    cc_test(test_sync_batch_norm_pass SRCS sync_batch_norm_pass_tester.cc DEPS sync_batch_norm_pass)
endif()
T
Tao Luo 已提交
206
if (WITH_MKLDNN)
207
    cc_test(test_depthwise_conv_mkldnn_pass SRCS mkldnn/depthwise_conv_mkldnn_pass_tester.cc DEPS depthwise_conv_mkldnn_pass)
208
    cc_test(test_conv_bias_mkldnn_fuse_pass_cc SRCS mkldnn/conv_bias_mkldnn_fuse_pass_tester.cc DEPS conv_bias_mkldnn_fuse_pass naive_executor)
209
    cc_test(test_conv_activation_mkldnn_fuse_pass SRCS mkldnn/conv_activation_mkldnn_fuse_pass_tester.cc DEPS conv_activation_mkldnn_fuse_pass)
210
    cc_test(test_conv_concat_relu_mkldnn_fuse_pass SRCS mkldnn/conv_concat_relu_mkldnn_fuse_pass_tester.cc DEPS conv_concat_relu_mkldnn_fuse_pass)
211
    cc_test(test_conv_elementwise_add_mkldnn_fuse_pass SRCS mkldnn/conv_elementwise_add_mkldnn_fuse_pass_tester.cc DEPS conv_elementwise_add_mkldnn_fuse_pass pass_test_util)
212
    cc_test(test_fc_elementwise_add_mkldnn_fuse_pass SRCS mkldnn/fc_elementwise_add_mkldnn_fuse_pass_tester.cc DEPS fc_elementwise_add_mkldnn_fuse_pass pass_test_util)
213 214
    cc_test(test_fc_act_mkldnn_fuse_pass SRCS mkldnn/fc_act_mkldnn_fuse_pass_tester.cc DEPS fc_act_mkldnn_fuse_pass pass_test_util)
    cc_test(test_batch_norm_act_fuse_pass SRCS mkldnn/batch_norm_act_fuse_pass_tester.cc DEPS batch_norm_act_fuse_pass pass_test_util)
215
    set(TEST_CONV_BN_PASS_DEPS conv_bn_fuse_pass graph_to_program_pass conv_op conv_transpose_op math_function im2col vol2col batch_norm_op gelu_op activation_op elementwise_add_op concat_and_split naive_executor device_context eigen_function)
216
if (WITH_GPU OR WITH_ROCM)
217 218 219
    set(TEST_CONV_BN_PASS_DEPS ${TEST_CONV_BN_PASS_DEPS} depthwise_conv)
endif()
    cc_test(test_conv_batch_norm_mkldnn_fuse_pass SRCS mkldnn/mkldnn_conv_bn_fuse_pass_tester.cc DEPS ${TEST_CONV_BN_PASS_DEPS})
220
    cc_test(test_scale_matmul_fuse_pass SRCS mkldnn/scale_matmul_fuse_pass_tester.cc DEPS scale_matmul_fuse_pass)
221
    cc_test(test_mkldnn_placement_pass SRCS mkldnn/mkldnn_placement_pass_tester.cc DEPS mkldnn_placement_pass)
222
    cc_test(test_mkldnn_inplace_pass SRCS mkldnn/mkldnn_inplace_pass_tester.cc DEPS mkldnn_inplace_pass)
B
baoachun 已提交
223
    cc_test(test_compute_propagate_scales_mkldnn_pass SRCS mkldnn/compute_propagate_scales_mkldnn_pass_tester.cc DEPS compute_propagate_scales_mkldnn_pass naive_executor)
224 225 226
    cc_test(test_cpu_quantize_placement_pass SRCS mkldnn/cpu_quantize_placement_pass_tester.cc DEPS cpu_quantize_placement_pass)
    cc_test(test_cpu_quantize_pass SRCS mkldnn/cpu_quantize_pass_tester.cc DEPS cpu_quantize_pass naive_executor)
    cc_test(test_cpu_quantize_squash_pass SRCS mkldnn/cpu_quantize_squash_pass_tester.cc DEPS cpu_quantize_squash_pass naive_executor)
227
    cc_test(test_reshape_transpose_matmul_mkldnn_fuse_pass SRCS mkldnn/reshape_transpose_matmul_mkldnn_fuse_pass_tester.cc DEPS reshape_transpose_matmul_mkldnn_fuse_pass reshape_transpose_matmul_v2_mkldnn_fuse_pass)
228
    cc_test(test_matmul_transpose_reshape_fuse_pass SRCS mkldnn/matmul_transpose_reshape_fuse_pass_tester.cc DEPS matmul_transpose_reshape_fuse_pass matmul_v2_transpose_reshape_fuse_pass)
229
    cc_test(test_shuffle_channel_mkldnn_detect_pass SRCS mkldnn/shuffle_channel_mkldnn_detect_pass_tester.cc DEPS shuffle_channel_mkldnn_detect_pass)
230 231
    cc_test(test_cpu_bfloat16_placement_pass SRCS mkldnn/cpu_bfloat16_placement_pass_tester.cc DEPS cpu_bfloat16_placement_pass)
    cc_test(test_cpu_bfloat16_pass SRCS mkldnn/cpu_bfloat16_pass_tester.cc DEPS cpu_bfloat16_pass)
232
    cc_test(test_multi_gru_fuse_pass SRCS mkldnn/multi_gru_fuse_pass_tester.cc DEPS multi_gru_fuse_pass)
233
    cc_test(test_multi_gru_seq_fuse_pass SRCS mkldnn/multi_gru_seq_fuse_pass_tester.cc DEPS multi_gru_seq_fuse_pass)
234 235
    set(TEST_FC_RNN_PASS_DEPS fc_gru_fuse_pass fc_lstm_fuse_pass mkldnn_placement_pass)
    cc_test(test_fc_rnn_mkldnn_fuse_pass SRCS mkldnn/mkldnn_fc_rnn_fuse_pass_tester.cc DEPS ${TEST_FC_RNN_PASS_DEPS})
T
Tao Luo 已提交
236
endif ()