From 64f7516aee218bf5a079fcdeae8fbec3dcca33fe Mon Sep 17 00:00:00 2001
From: tensor-tang <tangjian03@baidu.com>
Date: Fri, 16 Nov 2018 10:22:40 +0800
Subject: [PATCH] fix lrn on mac (#14426)

* rename and fix blas vsqr

test=develop

* update
---
 paddle/fluid/operators/lrn_op.cc        |  2 +-
 paddle/fluid/operators/math/blas.h      |  6 +++---
 paddle/fluid/operators/math/blas_impl.h | 14 ++++++++------
 3 files changed, 12 insertions(+), 10 deletions(-)

diff --git a/paddle/fluid/operators/lrn_op.cc b/paddle/fluid/operators/lrn_op.cc
index 8994f27086..a3bb2be5c7 100644
--- a/paddle/fluid/operators/lrn_op.cc
+++ b/paddle/fluid/operators/lrn_op.cc
@@ -46,7 +46,7 @@ struct LRNFunctor<platform::CPUDeviceContext, T> {
     int pre_pad = (n - 1) / 2;
     // compute batches one by one
     for (int i = 0; i < N; ++i) {
-      blas.VSQR(fea_size, idata + i * fea_size, sdata + pre_pad * img_size);
+      blas.VSQUARE(fea_size, idata + i * fea_size, sdata + pre_pad * img_size);
       // init the first channel of mid
       for (int c = 0; c < n; ++c) {
         blas.AXPY(img_size, alpha, sdata + c * img_size, mdata + i * fea_size);
diff --git a/paddle/fluid/operators/math/blas.h b/paddle/fluid/operators/math/blas.h
index 5d0d562030..6734df1530 100644
--- a/paddle/fluid/operators/math/blas.h
+++ b/paddle/fluid/operators/math/blas.h
@@ -153,7 +153,7 @@ class Blas {
   void VEXP(int n, const T* x, T* y) const;
 
   template <typename T>
-  void VSQR(int n, const T* x, T* y) const;
+  void VSQUARE(int n, const T* x, T* y) const;
 
   template <typename T>
   void VPOW(int n, const T* x, T alpha, T* y) const;
@@ -245,8 +245,8 @@ class BlasT : private Blas<DeviceContext> {
   }
 
   template <typename... ARGS>
-  void VSQR(ARGS... args) const {
-    Base()->template VSQR<T>(args...);
+  void VSQUARE(ARGS... args) const {
+    Base()->template VSQUARE<T>(args...);
   }
 
   template <typename... ARGS>
diff --git a/paddle/fluid/operators/math/blas_impl.h b/paddle/fluid/operators/math/blas_impl.h
index 59454669be..93bf7c7c88 100644
--- a/paddle/fluid/operators/math/blas_impl.h
+++ b/paddle/fluid/operators/math/blas_impl.h
@@ -105,7 +105,7 @@ struct CBlas<float> {
   }
 
   template <typename... ARGS>
-  static void VSQR(ARGS... args) {
+  static void VSQUARE(ARGS... args) {
     platform::dynload::vsSqr(args...);
   }
 
@@ -195,7 +195,7 @@ struct CBlas<double> {
   }
 
   template <typename... ARGS>
-  static void VSQR(ARGS... args) {
+  static void VSQUARE(ARGS... args) {
     platform::dynload::vdSqr(args...);
   }
 
@@ -262,7 +262,9 @@ struct CBlas<platform::float16> {
   }
   static void VMUL(...) { PADDLE_THROW("float16 VMUL not supported on CPU"); }
   static void VEXP(...) { PADDLE_THROW("float16 VEXP not supported on CPU"); }
-  static void VSQR(...) { PADDLE_THROW("float16 VSQR not supported on CPU"); }
+  static void VSQUARE(...) {
+    PADDLE_THROW("float16 VSQUARE not supported on CPU");
+  }
   static void VPOW(...) { PADDLE_THROW("float16 VPOW not supported on CPU"); }
   static void DOT(...) { PADDLE_THROW("float16 DOT not supported on CPU"); };
   static void SCAL(...) { PADDLE_THROW("float16 SCAL not supported on CPU"); };
@@ -423,12 +425,12 @@ void Blas<platform::CPUDeviceContext>::VEXP(int n, const T *x, T *y) const {
 
 template <>
 template <typename T>
-void Blas<platform::CPUDeviceContext>::VSQR(int n, const T *x, T *y) const {
+void Blas<platform::CPUDeviceContext>::VSQUARE(int n, const T *x, T *y) const {
 #ifdef PADDLE_WITH_MKLML
-  CBlas<T>::VSQR(n, x, y);
+  CBlas<T>::VSQUARE(n, x, y);
 #else
   for (int i = 0; i < n; ++i) {
-    y[i] = std::sqrt(x[i]);
+    y[i] = x[i] * x[i];
   }
 #endif
 }
-- 
GitLab