/home/docs/checkouts/readthedocs.org/user_builds/advanced-micro-devices-composable-kernel/checkouts/develop/include/ck/tensor_operation/gpu/element/element_wise_operation.hpp Source File

/home/docs/checkouts/readthedocs.org/user_builds/advanced-micro-devices-composable-kernel/checkouts/develop/include/ck/tensor_operation/gpu/element/element_wise_operation.hpp Source File#

Composable Kernel: /home/docs/checkouts/readthedocs.org/user_builds/advanced-micro-devices-composable-kernel/checkouts/develop/include/ck/tensor_operation/gpu/element/element_wise_operation.hpp Source File
Go to the documentation of this file.
 // SPDX-License-Identifier: MIT
 // Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
  
 #pragma once
  
 #include "ck/utility/data_type.hpp"
 #include "ck/utility/math_v2.hpp"
 #include "ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp"
 #include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
 #include "ck/tensor_operation/gpu/element/quantization_operation.hpp"
  
 namespace ck {
 namespace tensor_operation {
 namespace element_wise {
  
 // Need to ensure compiler will fail if there is no matching candidate, instead of compiler
 // siliently do implicit type conversion
 //
 // Example:
 //
 // struct ExampleElementwiseOp
 // {
 //     template<typename Y, typename X>
 //     __host__ __device__ constexpr void
 //     operator()(Y&, const X) const;
 //
 //     template<>
 //     __host__ __device__ constexpr void
 //     operator()<half_t, half_t>(half_t& y, const half_t& x) const
 //     {
 //     }
 // };
  
 struct AddReluAdd
 {
     template <typename Y, typename X0, typename X1, typename X2>
     __host__ __device__ constexpr void operator()(Y&, const X0&, const X1&, const X2&) const;
  
     template <>
     __host__ __device__ constexpr void operator()<half_t, half_t, half_t, half_t>(
         half_t& y, const half_t& x0, const half_t& x1, const half_t& x2) const
     {
         half_t a = x0 + x1;
         half_t b = a > 0 ? a : 0;
         y        = b + x2;
     }
  
     template <>
     __host__ __device__ constexpr void operator()<float, float, float, float>(float& y,
                                                                               const float& x0,
                                                                               const float& x1,
                                                                               const float& x2) const
     {
         float a = x0 + x1;
         float b = a > 0 ? a : 0;
         float c = b + x2;
         y       = c;
     }
  
     template <>
     __host__ __device__ constexpr void operator()<half_t, float, half_t, half_t>(
         half_t& y, const float& x0, const half_t& x1, const half_t& x2) const
     {
         float a = x0 + x1;
         float b = a > 0 ? a : 0;
         float c = b + x2;
         y       = c;
     }
  
     template <>
     __host__ __device__ constexpr void operator()<bhalf_t, float, bhalf_t, bhalf_t>(
         bhalf_t& y, const float& x0, const bhalf_t& x1, const bhalf_t& x2) const
     {
         float a = x0 + x1;
         float b = a > 0 ? a : 0;
         float c = b + x2;
         y       = c;
     }
  
     template <>
     __host__ __device__ constexpr void operator()<int8_t, int8_t, int8_t, int8_t>(
         int8_t& y, const int8_t& x0, const int8_t& x1, const int8_t& x2) const
     {
         int32_t a = x0 + x1;
         int32_t b = a > 0 ? a : 0;
         int32_t c = b + x2;
         y         = c;
     }
  
 #ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
     template <>
     __host__ __device__ constexpr void operator()<int4_t, int8_t, int4_t, int4_t>(
         int4_t& y, const int8_t& x0, const int4_t& x1, const int4_t& x2) const
     {
         int32_t a = x0 + x1;
         int32_t b = a > 0 ? a : 0;
         int32_t c = b + x2;
         y         = c;
     }
 #endif // CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
 };
  
 struct AddHardswishAdd
 {
     template <typename Y, typename X0, typename X1, typename X2>
     __host__ __device__ constexpr void operator()(Y&, const X0&, const X1&, const X2&) const;
  
     template <>
     __host__ __device__ constexpr void operator()<float, float, float, float>(float& y,
                                                                               const float& x0,
                                                                               const float& x1,
                                                                               const float& x2) const
     {
         float a = x0 + x1;
         float b = a + float{3};
         float c = (b > 0) * (b > float{6} ? float{6} : b) * a * float{0.166667};
         float d = c + x2;
         y       = d;
     }
  
     template <>
     __host__ __device__ constexpr void operator()<half_t, half_t, half_t, half_t>(
         half_t& y, const half_t& x0, const half_t& x1, const half_t& x2) const
     {
         float a = x0 + x1;
         float b = a + float{3};
         float c = (b > 0) * (b > float{6} ? float{6} : b) * a * float{0.166667};
         float d = c + x2;
         y       = d;
     }
 };
  
 // C = A * B
 // E = C + D0 + D1
 struct AddAdd
 {
     template <typename E, typename C, typename D0, typename D1>
     __host__ __device__ void operator()(E& e, const C& c, const D0& d0, const D1& d1) const
     {
         // Only support floating so far
         static_assert(is_same<E, half_t>::value || is_same<E, float>::value ||
                           is_same<E, double>::value,
                       "Data type is not supported by this operation!");
  
         static_assert(is_same<C, half_t>::value || is_same<C, float>::value ||
                           is_same<C, double>::value,
                       "Data type is not supported by this operation!");
  
         static_assert(is_same<D0, half_t>::value || is_same<D0, float>::value ||
                           is_same<D0, double>::value,
                       "Data type is not supported by this operation!");
  
         static_assert(is_same<D1, half_t>::value || is_same<D1, float>::value ||
                           is_same<D1, double>::value,
                       "Data type is not supported by this operation!");
  
         const C y = c + type_convert<C>(d0) + type_convert<C>(d1);
         e         = type_convert<E>(y);
     }
 };
  
 // C = A * B
 // E = (C + D0) x D1
 struct AddMultiply
 {
     template <typename E, typename C, typename D0, typename D1>
     __host__ __device__ void operator()(E& e, const C& c, const D0& d0, const D1& d1) const;
  
     template <>
     __host__ __device__ void operator()<half_t, half_t, half_t, half_t>(half_t& e,
                                                                         const half_t& c,
                                                                         const half_t& d0,
                                                                         const half_t& d1) const
     {
         const half_t y = (c + d0) * d1;
         e              = y;
     }
     template <>
     __host__ __device__ void operator()<half_t, float, half_t, half_t>(half_t& e,
                                                                        const float& c,
                                                                        const half_t& d0,
                                                                        const half_t& d1) const
     {
         const half_t y = (type_convert<half_t>(c) + d0) * d1;
         e              = y;
     }
     template <>
     __host__ __device__ void operator()<float, float, half_t, half_t>(float& e,
                                                                       const float& c,
                                                                       const half_t& d0,
                                                                       const half_t& d1) const
     {
         const float y = (c + d0) * d1;
         e             = y;
     }
 };
  
 // C = A * B
 // E = C x D0 + D1
 struct MultiplyAdd
 {
     template <typename E, typename C, typename D0, typename D1>
     __host__ __device__ void operator()(E& e, const C& c, const D0& d0, const D1& d1) const;
  
     template <>
     __host__ __device__ void operator()<half_t, half_t, half_t, half_t>(half_t& e,
                                                                         const half_t& c,
                                                                         const half_t& d0,
                                                                         const half_t& d1) const
     {
         const half_t y = (c * d0) + d1;
         e              = y;
     }
     template <>
     __host__ __device__ void operator()<half_t, float, half_t, half_t>(half_t& e,
                                                                        const float& c,
                                                                        const half_t& d0,
                                                                        const half_t& d1) const
     {
         const half_t y = type_convert<half_t>(c) * d0 + d1;
         e              = y;
     }
     template <>
     __host__ __device__ void operator()<bhalf_t, float, bhalf_t, bhalf_t>(bhalf_t& e,
                                                                           const float& c,
                                                                           const bhalf_t& d0,
                                                                           const bhalf_t& d1) const
     {
         const bhalf_t y = type_convert<bhalf_t>(c) * d0 + d1;
         e               = y;
     }
     template <>
     __host__ __device__ void operator()<float, float, half_t, half_t>(float& e,
                                                                       const float& c,
                                                                       const half_t& d0,
                                                                       const half_t& d1) const
     {
         const float y = c * d0 + d1;
         e             = y;
     }
     template <>
     __host__ __device__ void operator()<half_t, float, float, float>(half_t& e,
                                                                      const float& c,
                                                                      const float& d0,
                                                                      const float& d1) const
     {
         const float y = c * d0 + d1;
         e             = y;
     }
 };
  
 struct MultiplyMultiply
 {
     template <typename E, typename C, typename D0, typename D1>
     __host__ __device__ constexpr void
     operator()(E& e, const C& c, const D0& d0, const D1& d1) const;
  
     template <>
     __host__ __device__ constexpr void operator()<ck::half_t, float, float, float>(
         ck::half_t& e, const float& c, const float& d0, const float& d1) const
     {
         const float x0_f = c * d0 * d1;
  
         e = ck::type_convert<ck::half_t>(x0_f);
     }
  
     template <>
     __host__ __device__ constexpr void operator()<ck::bhalf_t, float, float, float>(
         ck::bhalf_t& e, const float& c, const float& d0, const float& d1) const
     {
         const float x0_f = c * d0 * d1;
  
         e = ck::type_convert<ck::bhalf_t>(x0_f);
     }
  
     template <>
     __host__ __device__ constexpr void operator()<ck::half_t, int, ck::half_t, ck::half_t>(
         ck::half_t& e, const int& c, const ck::half_t& d0, const ck::half_t& d1) const
     {
         const float x0_f =
             ck::type_convert<float>(c) * ck::type_convert<float>(d0) * ck::type_convert<float>(d1);
  
         e = ck::type_convert<ck::half_t>(x0_f);
     }
  
     template <>
     __host__ __device__ constexpr void operator()<ck::half_t, int, float, float>(
         ck::half_t& e, const int& c, const float& d0, const float& d1) const
     {
         const float x0_f =
             ck::type_convert<float>(c) * ck::type_convert<float>(d0) * ck::type_convert<float>(d1);
  
         e = ck::type_convert<ck::half_t>(x0_f);
     }
  
     template <>
     __host__ __device__ constexpr void operator()<ck::bhalf_t, int, float, float>(
         ck::bhalf_t& e, const int& c, const float& d0, const float& d1) const
     {
         const float x0_f =
             ck::type_convert<float>(c) * ck::type_convert<float>(d0) * ck::type_convert<float>(d1);
  
         e = ck::type_convert<ck::bhalf_t>(x0_f);
     }
 };
  
 struct MultiplyAddFastGelu
 {
     template <typename E, typename C, typename D0, typename D1>
     __host__ __device__ constexpr void
     operator()(E& e, const C& c, const D0& d0, const D1& d1) const;
  
     template <>
     __host__ __device__ constexpr void operator()<ck::bhalf_t, float, ck::bhalf_t, ck::bhalf_t>(
         ck::bhalf_t& e, const float& c, const ck::bhalf_t& d0, const ck::bhalf_t& d1) const
     {
         const float x0_f = c * ck::type_convert<float>(d0) + ck::type_convert<float>(d1);
  
         float x1_f = 0;
  
         FastGelu{}.template operator()<float, float>(x1_f, x0_f);
  
         e = ck::type_convert<ck::bhalf_t>(x1_f);
     }
 };
  
 // E = FastGelu(C + D0 + D1)
 struct AddAddFastGelu
 {
     template <typename E, typename C, typename D0, typename D1>
     __host__ __device__ constexpr void
     operator()(E& e, const C& c, const D0& d0, const D1& d1) const;
  
     template <>
     __host__ __device__ constexpr void operator()<float, float, float, float>(float& e,
                                                                               const float& c,
                                                                               const float& d0,
                                                                               const float& d1) const
     {
         const float x = c + d0 + d1;
  
         FastGelu{}.template operator()<float, float>(e, x);
     }
  
     template <>
     __host__ __device__ constexpr void operator()<half_t, half_t, half_t, half_t>(
         half_t& e, const half_t& c, const half_t& d0, const half_t& d1) const
     {
         const half_t x = c + d0 + d1;
  
         ck::tensor_operation::element_wise::FastGelu{}.template operator()<half_t, half_t>(e, x);
     }
  
     template <>
     __host__ __device__ constexpr void operator()<half_t, float, half_t, half_t>(
         half_t& e, const float& c, const half_t& d0, const half_t& d1) const
     {
         const float x0_f = c + d0 + d1;
  
         float x1_f = 0;
  
         ck::tensor_operation::element_wise::FastGelu{}.template operator()<float, float>(x1_f,
                                                                                          x0_f);
  
         e = type_convert<half_t>(x1_f);
     }
  
     template <>
     __host__ __device__ constexpr void operator()<bhalf_t, float, bhalf_t, bhalf_t>(
         bhalf_t& e, const float& c, const bhalf_t& d0, const bhalf_t& d1) const
     {
         const float x0_f = c + type_convert<float>(d0) + type_convert<float>(d1);
  
         float x1_f = 0;
  
         ck::tensor_operation::element_wise::FastGelu{}.template operator()<float, float>(x1_f,
                                                                                          x0_f);
  
         e = type_convert<bhalf_t>(x1_f);
     }
  
     template <>
     __host__ __device__ constexpr void operator()<int8_t, int32_t, int8_t, int8_t>(
         int8_t& e, const int32_t& c, const int8_t& d0, const int8_t& d1) const
     {
         const float x0_f =
             type_convert<float>(c) + type_convert<float>(d0) + type_convert<float>(d1);
  
         float x1_f = 0;
  
         ck::tensor_operation::element_wise::FastGelu{}.template operator()<float, float>(x1_f,
                                                                                          x0_f);
  
         e = type_convert<int8_t>(x1_f);
     }
 };
  
 // E = Relu(alpha1 * C + alpha2 * D0 + D1)
 struct ScaleAddScaleAddRelu
 {
  
     ScaleAddScaleAddRelu(const float alpha1 = 1.f, const float alpha2 = 1.f)
         : alpha1_(alpha1), alpha2_(alpha2)
     {
     }
  
     template <typename E, typename C, typename D0, typename D1>
     __host__ __device__ constexpr void
     operator()(E& e, const C& c, const D0& d0, const D1& d1) const;
  
     template <>
     __host__ __device__ constexpr void operator()<float, float, float, float>(float& e,
                                                                               const float& c,
                                                                               const float& d0,
                                                                               const float& d1) const
     {
         const float x = c * alpha1_ + alpha2_ * d0 + d1;
         e             = x > 0 ? x : 0;
     }
  
     template <>
     __host__ __device__ constexpr void operator()<half_t, half_t, half_t, half_t>(
         half_t& e, const half_t& c, const half_t& d0, const half_t& d1) const
     {
         const float x = type_convert<float>(c) * alpha1_ + alpha2_ * type_convert<float>(d0) +
                         type_convert<float>(d1);
  
         float result = 0;
         result       = x > 0 ? x : 0;
  
         e = type_convert<half_t>(result);
     }
  
     template <>
     __host__ __device__ constexpr void operator()<bhalf_t, bhalf_t, bhalf_t, bhalf_t>(
         bhalf_t& e, const bhalf_t& c, const bhalf_t& d0, const bhalf_t& d1) const
     {
         const float x = type_convert<float>(c) * alpha1_ + alpha2_ * type_convert<float>(d0) +
                         type_convert<float>(d1);
  
         float result = 0;
         result       = x > 0 ? x : 0;
  
         e = type_convert<bhalf_t>(result);
     }
  
     template <>
     __host__ __device__ constexpr void operator()<int8_t, int8_t, float, float>(
         int8_t& e, const int8_t& c, const float& d0, const float& d1) const
     {
         const float x = type_convert<float>(c) * alpha1_ + alpha2_ * d0 + d1;
  
         float result = 0;
         result       = x > 0 ? x : 0;
  
         e = type_convert<int8_t>(result);
     }
  
     const float alpha1_;
     const float alpha2_;
 };
  
 struct Normalize
 {
     // FIXME: is double absolutely necessary?
     Normalize(double epsilon = 1e-4) : epsilon_(epsilon) {}
  
     template <typename T1, typename T2, typename T3>
     __host__ __device__ constexpr void operator()(T1& y,
                                                   const T1& x,
                                                   const T2& mean,
                                                   const T2& mean_square,
                                                   const T3& gamma,
                                                   const T3& beta) const;
  
     template <>
     __host__ __device__ constexpr void operator()<half_t, float, half_t>(half_t& y,
                                                                          const half_t& x,
                                                                          const float& mean,
                                                                          const float& mean_square,
                                                                          const half_t& gamma,
                                                                          const half_t& beta) const
     {
         using ck::math::sqrt;
  
         float variance = mean_square - (mean * mean);
  
         float tmp_x     = type_convert<float>(x);
         float tmp_gamma = type_convert<float>(gamma);
         float tmp_beta  = type_convert<float>(beta);
  
         float tmp_y =
             ((tmp_x - mean) / sqrt(variance + type_convert<float>(epsilon_))) * tmp_gamma +
             tmp_beta;
  
         y = type_convert<half_t>(tmp_y);
     };
  
     template <>
     __host__ __device__ constexpr void operator()<float, float, float>(float& y,
                                                                        const float& x,
                                                                        const float& mean,
                                                                        const float& mean_square,
                                                                        const float& gamma,
                                                                        const float& beta) const
     {
         using ck::math::sqrt;
  
         float variance = mean_square - (mean * mean);
         y = ((x - mean) / sqrt(variance + type_convert<float>(epsilon_))) * gamma + beta;
     };
  
     template <>
     __host__ __device__ constexpr void operator()<double, double, double>(double& y,
                                                                           const double& x,
                                                                           const double& mean,
                                                                           const double& mean_square,
                                                                           const double& gamma,
                                                                           const double& beta) const
     {
         using ck::math::sqrt;
  
         double variance = mean_square - (mean * mean);
         y               = ((x - mean) / sqrt(variance + epsilon_)) * gamma + beta;
     };
  
     // FIXME: is double absolutely necessary?
     double epsilon_;
 };
  
 // used by BatchNorm inference
 // y = gamma * (x-mean) / sqrt(epsilon+variance) + beta
 // The data type of mean and variance is used as AccDataType
 struct NormalizeInInfer
 {
     NormalizeInInfer(double epsilon = 1e-4) : epsilon_(epsilon) {}
  
     template <typename T1, typename T2, typename T3, typename T4>
     __host__ __device__ constexpr void operator()(T1& y,
                                                   const T1& x,
                                                   const T2& mean,
                                                   const T2& variance,
                                                   const T3& gamma,
                                                   const T4& beta) const
     {
         static_assert(is_same<T2, float>::value || is_same<T2, double>::value,
                       "Data type is not supported by this operation!");
  
         using ck::type_convert;
         using ck::math::sqrt;
  
         T2 tmp_x, tmp_y;
  
         tmp_x = type_convert<T2>(x);
  
         tmp_y = ((tmp_x - mean) / sqrt(variance + type_convert<T2>(epsilon_))) *
                     type_convert<T2>(gamma) +
                 type_convert<T2>(beta);
         y = type_convert<T1>(tmp_y);
     };
  
     double epsilon_;
 };
  
 // used by Conv+Bias+BatchNorm+Clamp inference
 struct BiasNormalizeInInferClamp
 {
     BiasNormalizeInInferClamp(float floor   = 0.f,
                               float ceil    = NumericLimits<float>::Max(),
                               float epsilon = 1e-4)
         : clamp_(floor, ceil), epsilon_(epsilon)
     {
     }
  
     template <typename T>
     __host__ __device__ constexpr void operator()(T& y,
                                                   const T& x,
                                                   const T& bias,
                                                   const T& mean,
                                                   const T& variance,
                                                   const T& gamma,
                                                   const T& beta) const
     {
         using ck::type_convert;
         using ck::math::sqrt;
  
         float tmp_x = type_convert<float>(x) + type_convert<float>(bias);
  
         float tmp_y =
             ((tmp_x - type_convert<float>(mean)) / sqrt(type_convert<float>(variance) + epsilon_)) *
                 type_convert<float>(gamma) +
             type_convert<float>(beta);
         clamp_(tmp_y, tmp_y);
         y = type_convert<T>(tmp_y);
     };
  
     template <>
     __host__ __device__ constexpr void operator()(float& y,
                                                   const float& x,
                                                   const float& bias,
                                                   const float& mean,
                                                   const float& variance,
                                                   const float& gamma,
                                                   const float& beta) const
     {
         using ck::type_convert;
         using ck::math::sqrt;
  
         float tmp_y = (((x + bias) - mean) / sqrt(variance + epsilon_)) * gamma + beta;
         clamp_(y, tmp_y);
     };
  
     Clamp clamp_;
     float epsilon_;
 };
  
 template <typename Y, typename X>
 struct UnaryTypeConvert;
  
 template <>
 struct UnaryTypeConvert<float, ck::bhalf_t>
 {
     __host__ __device__ void operator()(float& y, ck::bhalf_t& x) const
     {
         y = ck::type_convert<float, ck::bhalf_t>(x);
     }
 };
  
 template <>
 struct UnaryTypeConvert<ck::bhalf_t, float>
 {
     __host__ __device__ void operator()(ck::bhalf_t& y, float& x) const
     {
         y = ck::type_convert<ck::bhalf_t, float>(x);
     }
 };
  
 } // namespace element_wise
 } // namespace tensor_operation
 } // namespace ck