!2072 optimize the performance of cpu lstm

Merge pull request !2072 from baihuawei/cpulstm

mindspore/ccsrc/kernel/cpu/mkldnn/lstm_cpu_kernel.cc

@@ -22,6 +22,10 @@
 namespace mindspore {
 namespace kernel {
 void LstmCPUKernel::InitKernel(const CNodePtr &kernel_node) {
+#ifdef PLATFORM_86
+  _MM_SET_FLUSH_ZERO_MODE(_MM_FLUSH_ZERO_ON);
+  _MM_SET_DENORMALS_ZERO_MODE(_MM_DENORMALS_ZERO_ON);
+#endif
   MS_EXCEPTION_IF_NULL(kernel_node);
   using tag = dnnl::memory::format_tag;
   using dim = dnnl::memory::dims;

mindspore/ccsrc/kernel/cpu/mkldnn/lstm_cpu_kernel.h

@@ -16,6 +16,12 @@
 
 #ifndef MINDSPORE_CCSRC_KERNEL_CPU_LSTM_CPU_KERNEL_H_
 #define MINDSPORE_CCSRC_KERNEL_CPU_LSTM_CPU_KERNEL_H_
+#if defined(__x86_64__) || defined(__amd64__) || defined(_M_IX86) || defined(_M_X64)
+#define PLATFORM_86
+#endif
+#ifdef PLATFORM_86
+#include <pmmintrin.h>
+#endif
 #include <vector>
 #include <memory>
 #include "kernel/cpu/mkldnn/mkl_cpu_kernel.h"