Skip to content

Commit

Permalink
[part 5]change type of function args
Browse files Browse the repository at this point in the history
  • Loading branch information
@zhangting2020
zhangting2020 committed Jan 13, 2022
1 parent 8cc0955 commit 40ccb07
Show file tree
Hide file tree
Showing 5 changed files with 58 additions and 98 deletions.
58 changes: 28 additions & 30 deletions paddle/fluid/operators/elementwise/elementwise_functor.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -54,15 +54,15 @@ using InverseDivFunctor = pten::funcs::InverseDivideFunctor<T>;
// Floor Divide
template <typename T>
struct FloorDivFunctor {
inline HOSTDEVICE T operator()(const T& a, const T& b) const {
inline HOSTDEVICE T operator()(const T a, const T b) const {
PADDLE_ENFORCE(b != 0, DIV_ERROR_INFO);
return static_cast<T>(std::trunc(a / b));
}
};

template <typename T>
struct InverseFloorDivFunctor {
inline HOSTDEVICE T operator()(const T& a, const T& b) const {
inline HOSTDEVICE T operator()(const T a, const T b) const {
PADDLE_ENFORCE(a != 0, DIV_ERROR_INFO);
return static_cast<T>(std::trunc(b / a));
}
Expand All @@ -73,15 +73,15 @@ struct InverseFloorDivFunctor {
// Maximum
template <typename T>
struct MaxFunctor {
inline HOSTDEVICE T operator()(const T& a, const T& b) const {
inline HOSTDEVICE T operator()(const T a, const T b) const {
return a > b ? a : b;
}
};

// Minmum
template <typename T>
struct MinFunctor {
inline HOSTDEVICE T operator()(const T& a, const T& b) const {
inline HOSTDEVICE T operator()(const T a, const T b) const {
return a < b ? a : b;
}
};
Expand Down Expand Up @@ -119,14 +119,14 @@ struct DivGradXYFunctor<Complex<InT>, Complex<OutT>> {
// Float div grad
template <typename T>
struct DivGradXFunctor {
inline HOSTDEVICE T operator()(const T& a, const T& b) const { return a / b; }
inline HOSTDEVICE T operator()(const T a, const T b) const { return a / b; }
};

// Complex div grad
template <typename T>
struct DivGradXFunctor<Complex<T>> {
inline HOSTDEVICE Complex<T> operator()(const Complex<T>& a,
const Complex<T>& b) const {
inline HOSTDEVICE Complex<T> operator()(const Complex<T> a,
const Complex<T> b) const {
Complex<T> b_conj(b.real, -b.imag);
return a / b_conj;
}
Expand All @@ -135,17 +135,17 @@ struct DivGradXFunctor<Complex<T>> {
// Float mul and div
template <typename T>
struct DivGradYFunctor {
inline HOSTDEVICE T operator()(const T& a, const T& b, const T& c) const {
inline HOSTDEVICE T operator()(const T a, const T b, const T c) const {
return -a * b / c;
}
};

// Complex mul and div
template <typename T>
struct DivGradYFunctor<Complex<T>> {
inline HOSTDEVICE Complex<T> operator()(const Complex<T>& a,
const Complex<T>& b,
const Complex<T>& c) const {
inline HOSTDEVICE Complex<T> operator()(const Complex<T> a,
const Complex<T> b,
const Complex<T> c) const {
Complex<T> out_div_c_conj((b / c).real, -(b / c).imag);
return -a * out_div_c_conj;
}
Expand All @@ -154,16 +154,16 @@ struct DivGradYFunctor<Complex<T>> {
// Fmax
template <typename T>
struct FMaxFunctor {
inline HOSTDEVICE T operator()(const T& a, const T& b) const {
inline HOSTDEVICE T operator()(const T a, const T b) const {
return std::fmax(a, b);
}
};

template <>
struct FMaxFunctor<paddle::platform::float16> {
inline HOSTDEVICE paddle::platform::float16 operator()(
const paddle::platform::float16& a,
const paddle::platform::float16& b) const {
const paddle::platform::float16 a,
const paddle::platform::float16 b) const {
float float_a = static_cast<float>(a);
float float_b = static_cast<float>(b);
auto result = std::fmax(float_a, float_b);
Expand All @@ -173,7 +173,7 @@ struct FMaxFunctor<paddle::platform::float16> {

template <>
struct FMaxFunctor<int> {
inline HOSTDEVICE int operator()(const int& a, const int& b) const {
inline HOSTDEVICE int operator()(const int a, const int b) const {
float float_a = static_cast<float>(a);
float float_b = static_cast<float>(b);
auto result = std::fmax(float_a, float_b);
Expand All @@ -183,8 +183,7 @@ struct FMaxFunctor<int> {

template <>
struct FMaxFunctor<int64_t> {
inline HOSTDEVICE int64_t operator()(const int64_t& a,
const int64_t& b) const {
inline HOSTDEVICE int64_t operator()(const int64_t a, const int64_t b) const {
double double_a = static_cast<double>(a);
double double_b = static_cast<double>(b);
auto result = std::fmax(double_a, double_b);
Expand All @@ -195,16 +194,16 @@ struct FMaxFunctor<int64_t> {
// Fmin
template <typename T>
struct FMinFunctor {
inline HOSTDEVICE T operator()(const T& a, const T& b) const {
inline HOSTDEVICE T operator()(const T a, const T b) const {
return std::fmin(a, b);
}
};

template <>
struct FMinFunctor<paddle::platform::float16> {
inline HOSTDEVICE paddle::platform::float16 operator()(
const paddle::platform::float16& a,
const paddle::platform::float16& b) const {
const paddle::platform::float16 a,
const paddle::platform::float16 b) const {
float float_a = static_cast<float>(a);
float float_b = static_cast<float>(b);
auto result = std::fmin(float_a, float_b);
Expand All @@ -214,7 +213,7 @@ struct FMinFunctor<paddle::platform::float16> {

template <>
struct FMinFunctor<int> {
inline HOSTDEVICE int operator()(const int& a, const int& b) const {
inline HOSTDEVICE int operator()(const int a, const int b) const {
float float_a = static_cast<float>(a);
float float_b = static_cast<float>(b);
auto result = std::fmin(float_a, float_b);
Expand All @@ -224,8 +223,7 @@ struct FMinFunctor<int> {

template <>
struct FMinFunctor<int64_t> {
inline HOSTDEVICE int64_t operator()(const int64_t& a,
const int64_t& b) const {
inline HOSTDEVICE int64_t operator()(const int64_t a, const int64_t b) const {
double double_a = static_cast<double>(a);
double double_b = static_cast<double>(b);
auto result = std::fmin(double_a, double_b);
Expand All @@ -235,22 +233,22 @@ struct FMinFunctor<int64_t> {

template <typename T>
struct MulGradFunctor {
inline HOSTDEVICE T operator()(const T& a, const T& b) const { return a * b; }
inline HOSTDEVICE T operator()(const T a, const T b) const { return a * b; }
};
template <typename T>
struct MulGradFunctor<Complex<T>> {
inline HOSTDEVICE Complex<T> operator()(const Complex<T>& a,
const Complex<T>& b) const {
inline HOSTDEVICE Complex<T> operator()(const Complex<T> a,
const Complex<T> b) const {
Complex<T> b_conj(b.real, -b.imag);
return a * b_conj;
}
};

template <typename InT, typename OutT>
struct MulGradXYFunctor {
inline HOSTDEVICE paddle::framework::Array<OutT, 2> operator()(const InT& a,
const InT& b,
const InT& c) {
inline HOSTDEVICE paddle::framework::Array<OutT, 2> operator()(const InT a,
const InT b,
const InT c) {
paddle::framework::Array<OutT, 2> outs;
// dx = dout * y
outs[0] = a * b;
Expand All @@ -263,7 +261,7 @@ struct MulGradXYFunctor {
template <typename InT, typename OutT>
struct MulGradXYFunctor<Complex<InT>, Complex<OutT>> {
inline HOSTDEVICE paddle::framework::Array<Complex<OutT>, 2> operator()(
const Complex<InT>& a, const Complex<InT>& b, const Complex<InT>& c) {
const Complex<InT> a, const Complex<InT> b, const Complex<InT> c) {
paddle::framework::Array<Complex<OutT>, 2> outs;
// dx = dout * y
Complex<InT> b_conj(b.real, -b.imag);
Expand Down
27 changes: 1 addition & 26 deletions paddle/fluid/operators/elementwise/elementwise_mod_op.cu
View file
Original file line number Diff line number Diff line change
Expand Up @@ -20,31 +20,6 @@ namespace plat = paddle::platform;
namespace paddle {
namespace operators {

template <typename T, typename Enable = void>
struct CudaModFunctor {
inline HOSTDEVICE T operator()(const T* args) const {
T res = args[0] % args[1];

// Accoding to #PR26732: in dividen % divsor
// remainder shall have the same sign as divsor.
if ((res != 0) && ((args[1] ^ res) < 0)) res += args[1];
return res;
}
};

template <typename T>
struct CudaModFunctor<
T, typename std::enable_if_t<std::is_floating_point<T>::value>> {
inline HOSTDEVICE T operator()(const T* args) const {
T res = fmod(args[0], args[1]);

// Accoding to #PR26732: in dividen % divsor
// remainder shall have the same sign as divsor.
if ((res != 0) && ((res < 0) != (args[1] < 0))) res += args[1];
return res;
}
};

template <typename T>
class ElementwiseModKernel<platform::CUDADeviceContext, T>
: public framework::OpKernel<T> {
Expand All @@ -56,7 +31,7 @@ class ElementwiseModKernel<platform::CUDADeviceContext, T>
ctx.template device_context<platform::CUDADeviceContext>();
int axis = PackTensorsIntoVector<T>(ctx, &ins, &outs);
LaunchElementwiseCudaKernel<ElementwiseType::kBinary, T, T>(
cuda_ctx, ins, &outs, axis, CudaModFunctor<T>());
cuda_ctx, ins, &outs, axis, ModFunctor<T>());
}
};

Expand Down
31 changes: 19 additions & 12 deletions paddle/fluid/operators/elementwise/elementwise_mod_op.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -19,29 +19,36 @@ limitations under the License. */
namespace paddle {
namespace operators {

template <typename T>
template <typename T, typename Enable = void>
struct ModFunctor {
inline HOSTDEVICE T operator()(T a, T b) const {
inline HOSTDEVICE T operator()(const T a, const T b) const {
T res = a % b;
if ((res != 0) && ((res < 0) != (b < 0))) res += b;

// Accoding to #PR26732: in dividen % divsor
// remainder shall have the same sign as divsor.
if ((res != 0) && ((b ^ res) < 0)) res += b;
return res;
}
};

template <typename T>
struct InverseModFunctor {
inline HOSTDEVICE T operator()(T a, T b) const {
T res = b % a;
if ((res != 0) && ((res < 0) != (a < 0))) res += a;
struct ModFunctor<T,
typename std::enable_if_t<std::is_floating_point<T>::value>> {
inline HOSTDEVICE T operator()(const T a, const T b) const {
T res = fmod(a, b);

// Accoding to #PR26732: in dividen % divsor
// remainder shall have the same sign as divsor.
if ((res != 0) && ((res < 0) != (b < 0))) res += b;
return res;
}
};

template <typename T>
struct ModFunctorFP {
struct InverseModFunctor {
inline HOSTDEVICE T operator()(T a, T b) const {
T res = fmod(a, b);
if ((res != 0) && ((b < 0) != (res < 0))) res += b;
T res = b % a;
if ((res != 0) && ((res < 0) != (a < 0))) res += a;
return res;
}
};
Expand Down Expand Up @@ -79,8 +86,8 @@ void elementwise_mod_fp(const framework::ExecutionContext &ctx,
auto x_dims = x->dims();
auto y_dims = y->dims();
if (x_dims.size() >= y_dims.size()) {
ElementwiseComputeEx<ModFunctorFP<T>, DeviceContext, T>(
ctx, x, y, axis, ModFunctorFP<T>(), z);
ElementwiseComputeEx<ModFunctor<T>, DeviceContext, T>(ctx, x, y, axis,
ModFunctor<T>(), z);
} else {
ElementwiseComputeEx<InverseModFunctorFP<T>, DeviceContext, T>(
ctx, x, y, axis, InverseModFunctorFP<T>(), z);
Expand Down
22 changes: 1 addition & 21 deletions paddle/fluid/operators/elementwise/elementwise_pow_op.cu
View file
Original file line number Diff line number Diff line change
Expand Up @@ -16,26 +16,6 @@ namespace ops = paddle::operators;
namespace paddle {
namespace operators {

template <typename T, typename Enable = void>
struct CudaPowFunctor {
inline HOSTDEVICE T operator()(const T args[]) const {
return std::pow(args[0], args[1]);
}
};

template <typename T>
struct CudaPowFunctor<
T, typename std::enable_if<std::is_integral<T>::value>::type> {
// On CUDAPlace, std::pow(3, 1) calls pow(float, float), and
// it will return a float number like 2.99... , which floor to 2
// when cast to int by default and it is wrong.
// Use llrint to cast it to the nearest integer, which is 3.
inline HOSTDEVICE T operator()(const T args[]) const {
return std::llrint(
std::pow(static_cast<double>(args[0]), static_cast<double>(args[1])));
}
};

template <typename T>
class ElementwisePowKernel<platform::CUDADeviceContext, T>
: public framework::OpKernel<T> {
Expand All @@ -48,7 +28,7 @@ class ElementwisePowKernel<platform::CUDADeviceContext, T>

int axis = PackTensorsIntoVector<T>(ctx, &ins, &outs);
LaunchElementwiseCudaKernel<ElementwiseType::kBinary, T, T>(
cuda_ctx, ins, &outs, axis, CudaPowFunctor<T>());
cuda_ctx, ins, &outs, axis, PowFunctor<T>());
}
};

Expand Down
18 changes: 9 additions & 9 deletions paddle/pten/kernels/funcs/elementwise_functor.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -26,31 +26,31 @@ namespace funcs {
// Add
template <typename T>
struct AddFunctor {
inline HOSTDEVICE T operator()(const T& a, const T& b) const { return a + b; }
inline HOSTDEVICE T operator()(const T a, const T b) const { return a + b; }
};
template <typename T>
struct InverseAddFunctor {
inline HOSTDEVICE T operator()(const T& a, const T& b) const { return b + a; }
inline HOSTDEVICE T operator()(const T a, const T b) const { return b + a; }
};

// Subtract
template <typename T>
struct SubtractFunctor {
inline HOSTDEVICE T operator()(const T& a, const T& b) const { return a - b; }
inline HOSTDEVICE T operator()(const T a, const T b) const { return a - b; }
};
template <typename T>
struct InverseSubtractFunctor {
inline HOSTDEVICE T operator()(const T& a, const T& b) const { return b - a; }
inline HOSTDEVICE T operator()(const T a, const T b) const { return b - a; }
};

// Multiply
template <typename T>
struct MultiplyFunctor {
inline HOSTDEVICE T operator()(const T& a, const T& b) const { return a * b; }
inline HOSTDEVICE T operator()(const T a, const T b) const { return a * b; }
};
template <typename T>
struct InverseMultiplyFunctor {
inline HOSTDEVICE T operator()(const T& a, const T& b) const { return b * a; }
inline HOSTDEVICE T operator()(const T a, const T b) const { return b * a; }
};

// Divide
Expand All @@ -60,14 +60,14 @@ struct InverseMultiplyFunctor {

template <typename T, typename Enable = void>
struct DivideFunctor {
inline HOSTDEVICE T operator()(const T& a, const T& b) const { return a / b; }
inline HOSTDEVICE T operator()(const T a, const T b) const { return a / b; }
};

template <typename T>
struct DivideFunctor<
T,
typename std::enable_if<std::is_integral<T>::value>::type> {
inline HOSTDEVICE T operator()(const T& a, const T& b) const {
inline HOSTDEVICE T operator()(const T a, const T b) const {
// For int32/int64, need to check whether the divison is zero.
PADDLE_ENFORCE(b != 0, DIV_ERROR_INFO);
return a / b;
Expand All @@ -76,7 +76,7 @@ struct DivideFunctor<

template <typename T, typename Enable = void>
struct InverseDivideFunctor {
inline HOSTDEVICE T operator()(const T& a, const T& b) const { return b / a; }
inline HOSTDEVICE T operator()(const T a, const T b) const { return b / a; }
};

} // namespace funcs
Expand Down

1 comment on commit 40ccb07

@paddle-bot-old
Copy link

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

✅Congratulation! Your pull request passed all required CI. You could ask reviewer(s) to approve and merge. 🎉

Please sign in to comment.