From e4f1b1c5e1fcde4683e2e79076df277412edf224 Mon Sep 17 00:00:00 2001
From: Tao Luo <luotao02@baidu.com>
Date: Fri, 10 Apr 2020 10:59:36 +0800
Subject: [PATCH] solve mklml memory leak (#23557)

---
 .../fluid/inference/api/analysis_predictor.cc | 19 +++++++++++++++++++
 paddle/fluid/platform/dynload/mklml.h         |  1 +
 2 files changed, 20 insertions(+)

diff --git a/paddle/fluid/inference/api/analysis_predictor.cc b/paddle/fluid/inference/api/analysis_predictor.cc
index f5cf7e52040..94b5400a470 100644
--- a/paddle/fluid/inference/api/analysis_predictor.cc
+++ b/paddle/fluid/inference/api/analysis_predictor.cc
@@ -37,6 +37,9 @@
 #include "paddle/fluid/inference/utils/singleton.h"
 #include "paddle/fluid/memory/memcpy.h"
 #include "paddle/fluid/platform/cpu_helper.h"
+#ifdef PADDLE_WITH_MKLML
+#include "paddle/fluid/platform/dynload/mklml.h"
+#endif
 #include "paddle/fluid/platform/gpu_info.h"
 #include "paddle/fluid/platform/place.h"
 #include "paddle/fluid/platform/profiler.h"
@@ -308,6 +311,14 @@ bool AnalysisPredictor::Run(const std::vector<PaddleTensor> &inputs,
   paddle::platform::SetNumThreads(1);
 #ifdef PADDLE_WITH_MKLDNN
   if (config_.use_mkldnn_) MkldnnPostReset();
+#endif
+#if defined(PADDLE_WITH_MKLML) && defined(_LINUX)
+  // Frees unused memory allocated by the Intel® MKL Memory Allocator to
+  // avoid memory leak. See:
+  // https://software.intel.com/en-us/mkl-developer-reference-c-mkl-free-buffers
+  platform::dynload::MKL_Free_Buffers();
+// We don't support windows since MKL_Free_Buffers is not in
+// mklml_win_2019.0.1.20181227.zip. We will upgrade mklml_win version later.
 #endif
   return true;
 }
@@ -648,6 +659,14 @@ bool AnalysisPredictor::ZeroCopyRun() {
   // recover the cpu_math_library_num_threads to 1, in order to avoid thread
   // conflict when integrating it into deployment service.
   paddle::platform::SetNumThreads(1);
+#if defined(PADDLE_WITH_MKLML) && defined(_LINUX)
+  // Frees unused memory allocated by the Intel® MKL Memory Allocator to
+  // avoid memory leak. See:
+  // https://software.intel.com/en-us/mkl-developer-reference-c-mkl-free-buffers
+  platform::dynload::MKL_Free_Buffers();
+// We don't support windows since MKL_Free_Buffers is not in
+// mklml_win_2019.0.1.20181227.zip. We will upgrade mklml_win version later.
+#endif
   return true;
 }
 
diff --git a/paddle/fluid/platform/dynload/mklml.h b/paddle/fluid/platform/dynload/mklml.h
index 839dcd87f57..e5e818f5fba 100644
--- a/paddle/fluid/platform/dynload/mklml.h
+++ b/paddle/fluid/platform/dynload/mklml.h
@@ -89,6 +89,7 @@ extern void* mklml_dso_handle;
   __macro(vdInv);                   \
   __macro(vmsErf);                  \
   __macro(vmdErf);                  \
+  __macro(MKL_Free_Buffers);        \
   __macro(MKL_Set_Num_Threads)
 
 MKLML_ROUTINE_EACH(DECLARE_DYNAMIC_LOAD_MKLML_WRAP);
-- 
GitLab