PaddleOCR-CPU 在飞腾arm版本上编译成C环境后，单页识别在几分钟以上，而且一页内存占用率达2个G

[gaoluhang]

[root@localhost cpp_infer]# build/ppocr --det_model_dir=/opt/ocr/PaddleOCR/inference/ch_PP-OCRv3_det_infer/ --rec_model_dir=/opt/ocr/PaddleOCR/inference/ch_PP-OCRv3_rec_infer/ --image_dir=/opt/ocr/PaddleOCR/doc/imgs/11.jpg
In PP-OCRv3, rec_image_shape parameter defaults to '3, 48, 320',if you are using recognition model with PP-OCRv2 or an older version, please set --rec_image_shape='3,32,320
total images num: 1
--- Running analysis [ir_graph_build_pass]
WARNING: Logging before InitGoogleLogging() is written to STDERR
I0614 20:21:49.510432 378241 executor.cc:187] Old Executor is Running.
--- Running analysis [ir_analysis_pass]
--- Running IR pass [simplify_with_basic_ops_pass]
--- Running IR pass [layer_norm_fuse_pass]
--- Running IR pass [attention_lstm_fuse_pass]
--- Running IR pass [seqconv_eltadd_relu_fuse_pass]
--- Running IR pass [seqpool_cvm_concat_fuse_pass]
--- Running IR pass [mul_lstm_fuse_pass]
--- Running IR pass [fc_gru_fuse_pass]
--- Running IR pass [mul_gru_fuse_pass]
--- Running IR pass [seq_concat_fc_fuse_pass]
--- Running IR pass [gpu_cpu_squeeze2_matmul_fuse_pass]
--- Running IR pass [gpu_cpu_reshape2_matmul_fuse_pass]
--- Running IR pass [gpu_cpu_flatten2_matmul_fuse_pass]
--- Running IR pass [matmul_v2_scale_fuse_pass]
--- Running IR pass [gpu_cpu_map_matmul_v2_to_mul_pass]
--- Running IR pass [gpu_cpu_map_matmul_v2_to_matmul_pass]
--- Running IR pass [matmul_scale_fuse_pass]
--- Running IR pass [gpu_cpu_map_matmul_to_mul_pass]
--- Running IR pass [fc_fuse_pass]
--- Running IR pass [repeated_fc_relu_fuse_pass]
--- Running IR pass [squared_mat_sub_fuse_pass]
--- Running IR pass [conv_bn_fuse_pass]
I0614 20:21:49.694710 378241 fuse_pass_base.cc:59] --- detected 33 subgraphs
--- Running IR pass [conv_eltwiseadd_bn_fuse_pass]
--- Running IR pass [conv_transpose_bn_fuse_pass]
--- Running IR pass [conv_transpose_eltwiseadd_bn_fuse_pass]
I0614 20:21:49.751570 378241 fuse_pass_base.cc:59] --- detected 1 subgraphs
--- Running IR pass [is_test_pass]
--- Running IR pass [constant_folding_pass]
--- Running analysis [save_optimized_model_pass]
--- Running analysis [ir_params_sync_among_devices_pass]
--- Running analysis [adjust_cudnn_workspace_size_pass]
--- Running analysis [inference_op_replace_pass]
--- Running analysis [memory_optimize_pass]
I0614 20:21:49.787482 378241 memory_optimize_pass.cc:118] The persistable params in main graph are : 2.23764MB
I0614 20:21:49.801360 378241 memory_optimize_pass.cc:246] Cluster name : batch_norm_41.tmp_0 size: 1920
I0614 20:21:49.801388 378241 memory_optimize_pass.cc:246] Cluster name : conv2d_221.tmp_0 size: 29491200
I0614 20:21:49.801400 378241 memory_optimize_pass.cc:246] Cluster name : batch_norm_4.tmp_3 size: 29491200
I0614 20:21:49.801411 378241 memory_optimize_pass.cc:246] Cluster name : elementwise_add_1 size: 3686400
I0614 20:21:49.801421 378241 memory_optimize_pass.cc:246] Cluster name : tmp_10 size: 22118400
I0614 20:21:49.801431 378241 memory_optimize_pass.cc:246] Cluster name : x size: 11059200
I0614 20:21:49.801441 378241 memory_optimize_pass.cc:246] Cluster name : tmp_1 size: 345600
I0614 20:21:49.801452 378241 memory_optimize_pass.cc:246] Cluster name : hardswish_0.tmp_0 size: 7372800
I0614 20:21:49.801462 378241 memory_optimize_pass.cc:246] Cluster name : nearest_interp_v2_5.tmp_0 size: 5529600
--- Running analysis [ir_graph_to_program_pass]
I0614 20:21:50.085887 378241 analysis_predictor.cc:1934] ======= ir optimization completed =======
I0614 20:21:50.118196 378241 naive_executor.cc:203] --- skip [feed], feed -> x
I0614 20:21:50.127347 378241 naive_executor.cc:203] --- skip [sigmoid_0.tmp_0], fetch -> fetch
In PP-OCRv3, default rec_img_h is 48,if you use other model, you should set the param rec_img_h=32
--- Running analysis [ir_graph_build_pass]
--- Running analysis [ir_analysis_pass]
--- Running IR pass [simplify_with_basic_ops_pass]
--- Running IR pass [layer_norm_fuse_pass]
--- Running IR pass [attention_lstm_fuse_pass]
--- Running IR pass [seqconv_eltadd_relu_fuse_pass]
--- Running IR pass [seqpool_cvm_concat_fuse_pass]
--- Running IR pass [mul_lstm_fuse_pass]
--- Running IR pass [fc_gru_fuse_pass]
--- Running IR pass [mul_gru_fuse_pass]
--- Running IR pass [seq_concat_fc_fuse_pass]
--- Running IR pass [gpu_cpu_squeeze2_matmul_fuse_pass]
--- Running IR pass [gpu_cpu_reshape2_matmul_fuse_pass]
--- Running IR pass [gpu_cpu_flatten2_matmul_fuse_pass]
--- Running IR pass [matmul_v2_scale_fuse_pass]
--- Running IR pass [gpu_cpu_map_matmul_v2_to_mul_pass]
I0614 20:21:50.244401 378241 fuse_pass_base.cc:59] --- detected 9 subgraphs
--- Running IR pass [gpu_cpu_map_matmul_v2_to_matmul_pass]
I0614 20:21:50.247087 378241 fuse_pass_base.cc:59] --- detected 4 subgraphs
--- Running IR pass [matmul_scale_fuse_pass]
--- Running IR pass [gpu_cpu_map_matmul_to_mul_pass]
--- Running IR pass [fc_fuse_pass]
I0614 20:21:50.258428 378241 fuse_pass_base.cc:59] --- detected 9 subgraphs
--- Running IR pass [repeated_fc_relu_fuse_pass]
--- Running IR pass [squared_mat_sub_fuse_pass]
--- Running IR pass [conv_bn_fuse_pass]
I0614 20:21:50.313755 378241 fuse_pass_base.cc:59] --- detected 19 subgraphs
--- Running IR pass [conv_eltwiseadd_bn_fuse_pass]
--- Running IR pass [conv_transpose_bn_fuse_pass]
--- Running IR pass [conv_transpose_eltwiseadd_bn_fuse_pass]
--- Running IR pass [is_test_pass]
--- Running IR pass [constant_folding_pass]
--- Running analysis [save_optimized_model_pass]
--- Running analysis [ir_params_sync_among_devices_pass]
--- Running analysis [adjust_cudnn_workspace_size_pass]
--- Running analysis [inference_op_replace_pass]
--- Running analysis [memory_optimize_pass]
I0614 20:21:50.357383 378241 memory_optimize_pass.cc:118] The persistable params in main graph are : 10.0928MB
I0614 20:21:50.363580 378241 memory_optimize_pass.cc:246] Cluster name : matmul_v2_4.tmp_0 size: 32
I0614 20:21:50.363600 378241 memory_optimize_pass.cc:246] Cluster name : tmp_7 size: 480
I0614 20:21:50.363611 378241 memory_optimize_pass.cc:246] Cluster name : batch_norm_39.tmp_4 size: 6144
I0614 20:21:50.363621 378241 memory_optimize_pass.cc:246] Cluster name : linear_43.tmp_1 size: 26500
I0614 20:21:50.363631 378241 memory_optimize_pass.cc:246] Cluster name : batch_norm_52.tmp_1 size: 2048
I0614 20:21:50.363641 378241 memory_optimize_pass.cc:246] Cluster name : batch_norm_42.tmp_3 size: 6144
I0614 20:21:50.363651 378241 memory_optimize_pass.cc:246] Cluster name : x size: 576
I0614 20:21:50.363660 378241 memory_optimize_pass.cc:246] Cluster name : transpose_9.tmp_0 size: 480
--- Running analysis [ir_graph_to_program_pass]
I0614 20:21:50.605667 378241 analysis_predictor.cc:1934] ======= ir optimization completed =======
I0614 20:21:50.607167 378241 naive_executor.cc:203] --- skip [feed], feed -> x
I0614 20:21:50.614113 378241 naive_executor.cc:203] --- skip [softmax_5.tmp_0], fetch -> fetch
predict img: /opt/ocr/PaddleOCR/doc/imgs/11.jpg
0 det boxes: [[27,37],[301,39],[301,72],[27,70]] rec text: 纯臻营养护发素 rec score: 0.927906
1 det boxes: [[25,81],[171,83],[171,103],[25,101]] rec text: 产品信息/参数 rec score: 0.851073
2 det boxes: [[28,115],[330,115],[330,131],[28,131]] rec text: 45元每公斤，100公斤起订） rec score: 0.90874
3 det boxes: [[396,122],[423,113],[430,132],[403,142]] rec text: rec score: 0
4 det boxes: [[27,144],[282,144],[282,164],[27,164]] rec text: 每瓶22元，1000瓶起订） rec score: 0.972308
5 det boxes: [[406,156],[428,156],[428,170],[406,170]] rec text: rec score: 0
6 det boxes: [[26,178],[298,178],[298,195],[26,195]] rec text: 【品牌】：代加工方式/OEMODM rec score: 0.904677
7 det boxes: [[26,209],[232,209],[232,226],[26,226]] rec text: 【品名】：纯臻营养护发素 rec score: 0.883685
8 det boxes: [[26,241],[241,241],[241,257],[26,257]] rec text: 【产品编号】：YM-X-3011 rec score: 0.855335
9 det boxes: [[413,236],[429,236],[429,302],[413,302]] rec text: ODMOEM rec score: 0.855649
10 det boxes: [[23,270],[179,269],[179,289],[23,290]] rec text: 【净含量】：220ml rec score: 0.88486
11 det boxes: [[26,303],[251,303],[251,320],[26,320]] rec text: 【适用人群】：适合所有肤质 rec score: 0.900196
12 det boxes: [[26,334],[342,334],[342,351],[26,351]] rec text: 【主要成分】：鲸蜡硬脂醇、燕麦β-葡聚 rec score: 0.880816
13 det boxes: [[27,366],[281,366],[281,382],[27,382]] rec text: 糖、椰油酰胺丙基甜菜碱、泛醒 rec score: 0.875617
14 det boxes: [[368,369],[475,369],[475,385],[368,385]] rec text: （成品包材） rec score: 0.971406
15 det boxes: [[26,397],[361,397],[361,414],[26,414]] rec text: 【主要功能】：可紧致头发磷层，从而达到 rec score: 0.848575
16 det boxes: [[28,428],[372,428],[372,445],[28,445]] rec text: 即时持久改善头发光泽的效果，给干燥的头 rec score: 0.948419
17 det boxes: [[26,458],[134,458],[134,478],[26,478]] rec text: 发足够的滋养 rec score: 0.864855
The detection visualized image saved in ./output//11.jpg

[gaoluhang]

Paddle编译过程：

cd Paddle
git checkout develop
mkdir build && cd build
ulimit -n 4096
cmake .. -DPY_VERSION=3.12.4 -DPYTHON_EXECUTABLE=which python -DWITH_ARM=ON -DWITH_TESTING=OFF -DCMAKE_BUILD_TYPE=Release -DON_INFER=ON -DWITH_XBYAK=OFF -DWITH_GPU=OFF
make TARGET=ARMV8 -j
make inference_lib_dist

然后修改
PaddleOCR/deploy/cpp_infer/tools/build.sh，将-DWITH_MKL改为OFF
随后tools/build.sh编译

[GreatV]

看着像paddle inference的问题，建议在paddle框架的仓库 https://github.com/PaddlePaddle/Paddle 下提一个issue。或者更换一个推理后端，如onnx。