solve mklml memory leak

[luotao1]

fix #22827

• related [WIP] Solve mklml memory leakage #23314 (comment)
• use mkl_free_buffers() to solve memory leak
• Inspired by Avoiding Memory Leaks in Intel® MKL. In our application, call the mkl_disable_fast_mm() function or mkl_thread_free_buffers function could not fix the memory leak.

[luotao1]

在windows上test_mkldnn_conv_relu_fuse_pass单测出现MKL_Free_Buffers not found.，但这个单测在Linux CI上能正常运行通过。

[00:25:41]	[Step 5/8] 694/728 Test #705: test_mkldnn_conv_relu_fuse_pass ......................................***Failed    2.39 sec
[00:25:41]	[Step 5/8] I0409 08:25:40.036875 46868 analysis_predictor.cc:136] Profiler is deactivated, and no profiling report will be generated.
[00:25:41]	[Step 5/8] I0409 08:25:40.037909 46868 analysis_predictor.cc:865] MODEL VERSION: 0.0.0
[00:25:41]	[Step 5/8] I0409 08:25:40.037909 46868 analysis_predictor.cc:867] PREDICTOR VERSION: 0.0.0
[00:25:41]	[Step 5/8] W0409 08:25:40.038911 46868 analysis_predictor.cc:880]  - Version incompatible (2) conv2d
[00:25:41]	[Step 5/8] W0409 08:25:40.038911 46868 analysis_predictor.cc:880]  - Version incompatible (1) feed
[00:25:41]	[Step 5/8] W0409 08:25:40.038911 46868 analysis_predictor.cc:880]  - Version incompatible (1) fetch
[00:25:41]	[Step 5/8] W0409 08:25:40.038911 46868 analysis_predictor.cc:880]  - Version incompatible (1) relu
[00:25:41]	[Step 5/8] W0409 08:25:40.038938 46868 analysis_predictor.cc:880]  - Version incompatible (1) scale
[00:25:41]	[Step 5/8] W0409 08:25:40.038946 46868 analysis_predictor.cc:192] WARNING: Results may be DIFF! Please use the corresponding version of the model and prediction library, and do not use the develop branch.
[00:25:41]	[Step 5/8] e[1me[35m--- Running analysis [ir_graph_build_pass]e[0m
[00:25:41]	[Step 5/8] e[1me[35m--- Running analysis [ir_graph_clean_pass]e[0m
[00:25:41]	[Step 5/8] e[1me[35m--- Running analysis [ir_analysis_pass]e[0m
[00:25:41]	[Step 5/8] e[32m--- Running IR pass [simplify_with_basic_ops_pass]e[0m
[00:25:41]	[Step 5/8] e[32m--- Running IR pass [attention_lstm_fuse_pass]e[0m
[00:25:41]	[Step 5/8] e[32m--- Running IR pass [seqconv_eltadd_relu_fuse_pass]e[0m
[00:25:41]	[Step 5/8] e[32m--- Running IR pass [seqpool_cvm_concat_fuse_pass]e[0m
[00:25:41]	[Step 5/8] e[32m--- Running IR pass [fc_lstm_fuse_pass]e[0m
[00:25:41]	[Step 5/8] e[32m--- Running IR pass [mul_lstm_fuse_pass]e[0m
[00:25:41]	[Step 5/8] e[32m--- Running IR pass [fc_gru_fuse_pass]e[0m
[00:25:41]	[Step 5/8] e[32m--- Running IR pass [mul_gru_fuse_pass]e[0m
[00:25:41]	[Step 5/8] e[32m--- Running IR pass [seq_concat_fc_fuse_pass]e[0m
[00:25:41]	[Step 5/8] e[32m--- Running IR pass [fc_fuse_pass]e[0m
[00:25:41]	[Step 5/8] e[32m--- Running IR pass [repeated_fc_relu_fuse_pass]e[0m
[00:25:41]	[Step 5/8] e[32m--- Running IR pass [squared_mat_sub_fuse_pass]e[0m
[00:25:41]	[Step 5/8] e[32m--- Running IR pass [conv_bn_fuse_pass]e[0m
[00:25:41]	[Step 5/8] e[32m--- Running IR pass [conv_eltwiseadd_bn_fuse_pass]e[0m
[00:25:41]	[Step 5/8] e[32m--- Running IR pass [conv_transpose_bn_fuse_pass]e[0m
[00:25:41]	[Step 5/8] e[32m--- Running IR pass [conv_transpose_eltwiseadd_bn_fuse_pass]e[0m
[00:25:41]	[Step 5/8] e[32m--- Running IR pass [is_test_pass]e[0m
[00:25:41]	[Step 5/8] e[32m--- Running IR pass [runtime_context_cache_pass]e[0m
[00:25:41]	[Step 5/8] e[1me[35m--- Running analysis [ir_params_sync_among_devices_pass]e[0m
[00:25:41]	[Step 5/8] e[1me[35m--- Running analysis [adjust_cudnn_workspace_size_pass]e[0m
[00:25:41]	[Step 5/8] e[1me[35m--- Running analysis [inference_op_replace_pass]e[0m
[00:25:41]	[Step 5/8] e[1me[35m--- Running analysis [ir_graph_to_program_pass]e[0m
[00:25:41]	[Step 5/8] test_mkldnn_conv_relu_fuse_pass failed
[00:25:41]	[Step 5/8]  E
[00:25:41]	[Step 5/8] ======================================================================
[00:25:41]	[Step 5/8] ERROR: test_check_output (test_mkldnn_conv_relu_fuse_pass.ConvBnFusePassMKLDNNTest)
[00:25:41]	[Step 5/8] ----------------------------------------------------------------------
[00:25:41]	[Step 5/8] Traceback (most recent call last):
[00:25:41]	[Step 5/8]   File "C:\home\BuildAgent\work\45396e78faf70386\build\python\paddle\fluid\tests\unittests\ir\inference\test_mkldnn_conv_relu_fuse_pass.py", line 41, in test_check_output
[00:25:41]	[Step 5/8]     self.check_output_with_option(use_gpu)
[00:25:41]	[Step 5/8]   File "C:\home\BuildAgent\work\45396e78faf70386\build\python\paddle\fluid\tests\unittests\ir\inference\inference_pass_test.py", line 146, in check_output_with_option
[00:25:41]	[Step 5/8]     self._get_analysis_config(use_gpu=use_gpu))
[00:25:41]	[Step 5/8]   File "C:\home\BuildAgent\work\45396e78faf70386\build\python\paddle\fluid\tests\unittests\ir\inference\inference_pass_test.py", line 83, in _get_analysis_outputs
[00:25:41]	[Step 5/8]     predictor.zero_copy_run()
[00:25:41]	[Step 5/8] RuntimeError: MKL_Free_Buffers not found.
[00:25:41]	[Step 5/8] 
[00:25:41]	[Step 5/8] ----------------------------------------------------------------------

下载 https://paddlepaddledeps.bj.bcebos.com/mklml_win_2019.0.1.20181227.zip windows安装包后，nm mklml.lib，其中只有MKL_free，没有MKL_Free_Buffers

$ nm mklml.lib | grep MKL
00000000 T MKL_Comatcopy
00000000 T __imp_MKL_Comatcopy
00000000 T MKL_Domain_Get_Max_Threads
00000000 T __imp_MKL_Domain_Get_Max_Threads
00000000 T MKL_Domain_Set_Num_Threads
00000000 T __imp_MKL_Domain_Set_Num_Threads
00000000 T MKL_Domatcopy
00000000 T __imp_MKL_Domatcopy
00000000 T MKL_Get_Dynamic
00000000 T __imp_MKL_Get_Dynamic
00000000 T MKL_Get_Max_Threads
00000000 T __imp_MKL_Get_Max_Threads
00000000 T MKL_Get_Version
00000000 T __imp_MKL_Get_Version
00000000 T MKL_Get_Version_String
00000000 T __imp_MKL_Get_Version_String
00000000 T MKL_Set_Dynamic
00000000 T __imp_MKL_Set_Dynamic
00000000 T MKL_Set_Num_Threads
00000000 T __imp_MKL_Set_Num_Threads
00000000 T MKL_Set_Num_Threads_Local
00000000 T __imp_MKL_Set_Num_Threads_Local
00000000 T MKL_Somatcopy
00000000 T __imp_MKL_Somatcopy
00000000 T MKL_Zomatcopy
00000000 T __imp_MKL_Zomatcopy
00000000 T MKL_free
00000000 T __imp_MKL_free
00000000 T MKL_malloc
00000000 T __imp_MKL_malloc

@yinghu5 请问windows下是要用MKL_free来代替么？

[luotao1]

已和 @yinghu5 确认目前repo中集成的windows版本mklml.lib中没有打包进MKL_Free_Buffers函数

这个PR会暂时不修windows上的内存泄漏，后续通过升级mklml版本来解决。

[jiweibo]

已在linux平台，用户模型下测试，合入此pr后修复了由于mkl导致的内存泄露问题。

[linzhec]

windows版本的内存泄漏解决了吗，测试发现当前版本仍然存在