not modify pow_op's original compute logic

Author: betterpig

paddle/fluid/operators/elementwise/elementwise_pow_op.h

@@ -63,8 +63,8 @@ struct PowGradDX {
  HOSTDEVICE T operator()(T x, T y, T out, T dout) const {
 #if defined(__CUDA_ARCH__) || defined(__HIPCC__)
  if (std::is_integral<T>::value) {
- return std::llrint(dout * y * std::pow(static_cast<double>(x),
-  static_cast<double>(y - 1)));
+ return dout * y *
+ std::pow(static_cast<double>(x), static_cast<double>(y - 1));
  }
 #endif
  return dout * y * std::pow(x, y - 1);
@@ -74,19 +74,16 @@ struct PowGradDX {
 template <typename T, typename Enable = void>
 struct PowGradDY {
  HOSTDEVICE T operator()(T x, T y, T out, T dout) const {
+#if defined(__CUDA_ARCH__) || defined(__HIPCC__)
+ if (std::is_integral<T>::value) {
+ return dout * std::log(static_cast<double>(x)) *
+ std::pow(static_cast<double>(x), static_cast<double>(y));
+ }
+#endif
  return dout * std::log(x) * std::pow(x, y);
  }
 };
 
-template <typename T>
-struct PowGradDY<T, typename std::enable_if<std::is_integral<T>::value>::type> {
- HOSTDEVICE T operator()(T x, T y, T out, T dout) const {
- return std::llrint(
- dout * std::log(static_cast<double>(x)) *
- std::pow(static_cast<double>(x), static_cast<double>(y)));
- }
-};
-
 template <typename DeviceContext, typename T>
 class ElementwisePowGradKernel : public ElemwiseGradKernel<T> {
  public:

paddle/scripts/paddle_build.bat

@@ -78,7 +78,6 @@ if not defined PYTHON_ROOT set PYTHON_ROOT=C:\Python37
 if not defined BUILD_DIR set BUILD_DIR=build
 set task_name=%1
 set UPLOAD_TP_FILE=OFF
-set WITH_TPCACHE=OFF
 
 rem ------initialize the python environment------
 set PYTHON_EXECUTABLE=%PYTHON_ROOT%\python.exe
@@ -262,6 +261,7 @@ set ON_INFER=ON
 set WITH_TESTING=ON
 set WITH_TENSORRT=ON
 set WITH_INFERENCE_API_TEST=ON
+set WITH_TPCACHE=OFF
 
 call :cmake || goto cmake_error
 call :build || goto build_error
@@ -333,7 +333,7 @@ echo %task_name%|findstr wincheck_inference >nul && (
 )
 set DISTUTILS_USE_SDK=1
 rem Windows 10 Kit bin dir
-::set PATH=C:\Program Files (x86)\Windows Kits\10\bin\10.0.17763.0\x64;%PATH%
+set PATH=C:\Program Files (x86)\Windows Kits\10\bin\10.0.17763.0\x64;%PATH%
 rem Use 64-bit ToolSet to compile
 set PreferredToolArchitecture=x64
 

python/paddle/fluid/tests/unittests/test_elementwise_pow_op.py

@@ -156,11 +156,11 @@ def setUp(self):
  # dout = 1
  self.grad_res = np.asarray([1, 1, 1])
  # dx = dout * y * pow(x, y-1)
- self.grad_x = (np.rint(self.grad_res * self.y * self.x
- **(self.y - 1))).astype("int")
+ self.grad_x = self.grad_res * self.y * (self.x
+  **(self.y - 1)).astype("int")
  # dy = dout * log(x) * pow(x, y)
- self.grad_y = (np.rint(self.grad_res * np.log(self.x) *
-  (self.x**self.y))).astype("int")
+ self.grad_y = (self.grad_res * np.log(self.x) *
+ (self.x**self.y)).astype("int")
  print(self.grad_res, self.grad_x, self.grad_y)
 
  def test_grad(self):
@@ -176,7 +176,6 @@ def test_grad(self):
  y.stop_gradient = False
  res = x**y
  res.backward()
- print(res.gradient(), x.gradient(), y.gradient())
  self.assertTrue(np.array_equal(res.gradient(), self.grad_res))
  self.assertTrue(np.array_equal(x.gradient(), self.grad_x))
  self.assertTrue(np.array_equal(y.gradient(), self.grad_y))