src/backends/cpu/arithmetics.hpp - master - vuron/adept

src/backends/cpu/arithmetics.hpp
109 строк3 KB
kolkir
Refactor simd algos to use map reduce
16 мар 2025, 23:09
16 мар 2025, 23:095f536e1
// should be included in only ones section after arithmetics-inl.hpp
#pragma once

#include <hwy/highway.h>
#include <adept/types.hpp>

#if HWY_ONCE

namespace adept::detail {
// callers for SIMD function platforms overloads

struct add {
  template <typename DataType>
  static void apply(DataType* x_array, const DataType* y_array, index_t size, bool aligned) {
    HWY_EXPORT_AND_DYNAMIC_DISPATCH_T(add_vectors<DataType>)
    (x_array, y_array, size, aligned);
  }

  template <typename DataType>
  static void apply(DataType* x_array, float scalar, index_t size, bool aligned) {
    HWY_EXPORT_AND_DYNAMIC_DISPATCH_T(add_vector_scalar<DataType>)
    (x_array, scalar, size, aligned);
  }
};

struct sub {
  template <typename DataType>
  static void apply(DataType* x_array, const DataType* y_array, index_t size, bool aligned) {
    HWY_EXPORT_AND_DYNAMIC_DISPATCH_T(sub_vectors<DataType>)
    (x_array, y_array, size, aligned);
  }

  template <typename DataType>
  static void apply(DataType* x_array, float32_t scalar, index_t size, bool aligned) {
    HWY_EXPORT_AND_DYNAMIC_DISPATCH_T(sub_vector_scalar<DataType>)
    (x_array, scalar, size, aligned);
  }
};

struct mul {
  template <typename DataType>
  static void apply(DataType* x_array, const DataType* y_array, index_t size, bool aligned) {
    HWY_EXPORT_AND_DYNAMIC_DISPATCH_T(mul_vectors<DataType>)
    (x_array, y_array, size, aligned);
  }

  template <typename DataType>
  static void apply(DataType* x_array, float32_t scalar, index_t size, bool aligned) {
    HWY_EXPORT_AND_DYNAMIC_DISPATCH_T(mul_vector_scalar<DataType>)
    (x_array, scalar, size, aligned);
  }
};

struct div {
  template <typename DataType>
  static void apply(DataType* x_array, const DataType* y_array, index_t size, bool aligned) {
    HWY_EXPORT_AND_DYNAMIC_DISPATCH_T(div_vectors<DataType>)
    (x_array, y_array, size, aligned);
  }

  template <typename DataType>
  static void apply(DataType* x_array, float32_t scalar, index_t size, bool aligned) {
    HWY_EXPORT_AND_DYNAMIC_DISPATCH_T(div_vector_scalar<DataType>)
    (x_array, scalar, size, aligned);
  }
};

struct neg {
  template <typename DataType>
  static void apply(DataType* x_array, index_t size, bool aligned) {
    HWY_EXPORT_AND_DYNAMIC_DISPATCH_T(vector_neg<DataType>)
    (x_array, size, aligned);
  }
};

struct exp {
  template <typename DataType>
  static void apply(DataType* x_array, index_t size, bool aligned) {
    HWY_EXPORT_AND_DYNAMIC_DISPATCH_T(vector_exp<DataType>)
    (x_array, size, aligned);
  }
};

struct sqrt {
  template <typename DataType>
  static void apply(DataType* x_array, index_t size, bool aligned) {
    HWY_EXPORT_AND_DYNAMIC_DISPATCH_T(vector_sqrt<DataType>)
    (x_array, size, aligned);
  }
};

struct max {
  template <typename DataType>
  static void apply(const DataType* x_array, DataType& out, index_t size, bool aligned) {
    HWY_EXPORT_AND_DYNAMIC_DISPATCH_T(vector_max<DataType>)
    (x_array, out, size, aligned);
  }
};

struct sum {
  template <typename DataType>
  static void apply(const DataType* x_array, DataType& out, index_t size, bool aligned) {
    HWY_EXPORT_AND_DYNAMIC_DISPATCH_T(vector_sum<DataType>)
    (x_array, out, size, aligned);
  }
};

}  // namespace adept::detail
#endif  // HWY_ONCE