phynexis-deps/armadillo-code
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

add support for cx_fp16

Browse Source
This commit is contained in:
conrad
2025-07-11 23:14:44 +10:00
parent bbe5dcf225
commit c87cf374c4
9 changed files with 116 additions and 35 deletions
Download Patch File Download Diff File Expand all files Collapse all files
include/armadillo_bits/mul_herk.hpp
+20 -1
View File
@@ -484,7 +484,26 @@ class herk
{
herk<do_trans_A, use_alpha, use_beta>::apply_blas_type(C,A,alpha,beta);
}
#if defined(ARMA_HAVE_FP16)
template<typename TA>
arma_inline
static
void
apply
(
Mat< std::complex<fp16> >& C,
const TA& A,
const fp16 alpha = fp16(1),
const fp16 beta = fp16(0)
)
{
herk_emul<do_trans_A, use_alpha, use_beta>::apply(C,A,alpha,beta);
}
#endif
};
include/armadillo_bits/op_dot_bones.hpp
+5 -5
View File
@@ -40,13 +40,13 @@ class op_dot
template<typename eT>
arma_hot inline static typename arma_blas_real_only<eT>::result
direct_dot(const uword n_elem, const eT* const A, const eT* const B);
template<typename eT>
arma_hot inline static typename arma_fp16_only<eT>::result
arma_hot inline static typename arma_blas_cx_only<eT>::result
direct_dot(const uword n_elem, const eT* const A, const eT* const B);
template<typename eT>
arma_hot inline static typename arma_cx_only<eT>::result
arma_hot inline static typename arma_fp16_real_or_cx_only<eT>::result
direct_dot(const uword n_elem, const eT* const A, const eT* const B);
template<typename eT>
@@ -98,7 +98,7 @@ class op_cdot
arma_hot inline static eT direct_cdot(const uword n_elem, const eT* const A, const eT* const B, const typename arma_blas_real_or_cx_only<eT>::result* junk = 0);
template<typename eT>
arma_hot inline static eT direct_cdot(const uword n_elem, const eT* const A, const eT* const B, const typename arma_fp16_only<eT>::result* junk = 0);
arma_hot inline static eT direct_cdot(const uword n_elem, const eT* const A, const eT* const B, const typename arma_fp16_real_or_cx_only<eT>::result* junk = 0);
template<typename T1, typename T2>
arma_hot inline static typename T1::elem_type apply (const T1& X, const T2& Y);
include/armadillo_bits/op_dot_meat.hpp
+14 -16
View File
@@ -98,7 +98,7 @@ op_dot::direct_dot_arma(const uword n_elem, const eT* const A, const eT* const B
//! for two arrays, floating-point version
//! for two arrays, float and double version
template<typename eT>
inline
typename arma_blas_real_only<eT>::result
@@ -129,23 +129,10 @@ op_dot::direct_dot(const uword n_elem, const eT* const A, const eT* const B)
//! for two arrays, fp16 version
template<typename eT>
inline
typename arma_fp16_only<eT>::result
op_dot::direct_dot(const uword n_elem, const eT* const A, const eT* const B)
{
arma_debug_sigprint();
return op_dot::direct_dot_arma(n_elem, A, B);
}
//! for two arrays, complex version
template<typename eT>
inline
typename arma_cx_only<eT>::result
typename arma_blas_cx_only<eT>::result
op_dot::direct_dot(const uword n_elem, const eT* const A, const eT* const B)
{
if(n_elem <= 16u) { return op_dot::direct_dot_arma(n_elem, A, B); }
@@ -171,6 +158,17 @@ op_dot::direct_dot(const uword n_elem, const eT* const A, const eT* const B)
//! for two arrays, fp16 version
template<typename eT>
inline
typename arma_fp16_real_or_cx_only<eT>::result
op_dot::direct_dot(const uword n_elem, const eT* const A, const eT* const B)
{
return op_dot::direct_dot_arma(n_elem, A, B);
}
//! for two arrays, integral version
template<typename eT>
inline
@@ -502,7 +500,7 @@ op_cdot::direct_cdot(const uword n_elem, const eT* const A, const eT* const B, c
template<typename eT>
inline
eT
op_cdot::direct_cdot(const uword n_elem, const eT* const A, const eT* const B, const typename arma_fp16_only<eT>::result* junk)
op_cdot::direct_cdot(const uword n_elem, const eT* const A, const eT* const B, const typename arma_fp16_real_or_cx_only<eT>::result* junk)
{
arma_debug_sigprint();
arma_ignore(junk);
include/armadillo_bits/op_norm_bones.hpp
+1 -1
View File
@@ -49,7 +49,7 @@ class op_norm
template<typename eT> inline static typename get_pod_type<eT>::result mat_norm_1(const Mat<eT>& X);
template<typename eT> inline static typename get_pod_type<eT>::result mat_norm_2(const Mat<eT>& X, const typename arma_blas_real_or_cx_only<eT>::result* junk = nullptr);
template<typename eT> inline static typename get_pod_type<eT>::result mat_norm_2(const Mat<eT>& X, const typename arma_fp16_only<eT>::result* junk = nullptr);
template<typename eT> inline static typename get_pod_type<eT>::result mat_norm_2(const Mat<eT>& X, const typename arma_fp16_real_or_cx_only<eT>::result* junk = nullptr);
template<typename eT> inline static typename get_pod_type<eT>::result mat_norm_inf(const Mat<eT>& X);
};
include/armadillo_bits/op_norm_meat.hpp
+1 -1
View File
@@ -954,7 +954,7 @@ op_norm::mat_norm_2(const Mat<eT>& X, const typename arma_blas_real_or_cx_only<e
template<typename eT>
inline
typename get_pod_type<eT>::result
op_norm::mat_norm_2(const Mat<eT>& X, const typename arma_fp16_only<eT>::result* junk)
op_norm::mat_norm_2(const Mat<eT>& X, const typename arma_fp16_real_or_cx_only<eT>::result* junk)
{
arma_debug_sigprint();
arma_ignore(junk);
include/armadillo_bits/restrictors.hpp
+39 -9
View File
@@ -40,11 +40,12 @@ template<> struct arma_scalar_only< ulng_t > { typedef ulng_t result; };
template<> struct arma_scalar_only< slng_t > { typedef slng_t result; };
template<> struct arma_scalar_only< float > { typedef float result; };
template<> struct arma_scalar_only< double > { typedef double result; };
#if defined(ARMA_HAVE_FP16)
template<> struct arma_scalar_only< fp16 > { typedef fp16 result; };
#endif
template<> struct arma_scalar_only< cx_float > { typedef cx_float result; };
template<> struct arma_scalar_only< cx_double > { typedef cx_double result; };
#if defined(ARMA_HAVE_FP16)
template<> struct arma_scalar_only< fp16 > { typedef fp16 result; };
template<> struct arma_scalar_only< cx_fp16 > { typedef cx_fp16 result; };
#endif
@@ -92,11 +93,12 @@ template<> struct arma_signed_only< s64 > { typedef s64 result; };
template<> struct arma_signed_only< slng_t > { typedef slng_t result; };
template<> struct arma_signed_only< float > { typedef float result; };
template<> struct arma_signed_only< double > { typedef double result; };
#if defined(ARMA_HAVE_FP16)
template<> struct arma_signed_only< fp16 > { typedef fp16 result; };
#endif
template<> struct arma_signed_only< cx_float > { typedef cx_float result; };
template<> struct arma_signed_only< cx_double > { typedef cx_double result; };
#if defined(ARMA_HAVE_FP16)
template<> struct arma_signed_only< fp16 > { typedef fp16 result; };
template<> struct arma_signed_only< cx_fp16 > { typedef cx_fp16 result; };
#endif
@@ -129,11 +131,12 @@ template<typename T> struct arma_real_or_cx_only { };
template<> struct arma_real_or_cx_only< float > { typedef float result; };
template<> struct arma_real_or_cx_only< double > { typedef double result; };
#if defined(ARMA_HAVE_FP16)
template<> struct arma_real_or_cx_only< fp16 > { typedef fp16 result; };
#endif
template<> struct arma_real_or_cx_only< cx_float > { typedef cx_float result; };
template<> struct arma_real_or_cx_only< cx_double > { typedef cx_double result; };
#if defined(ARMA_HAVE_FP16)
template<> struct arma_real_or_cx_only< fp16 > { typedef fp16 result; };
template<> struct arma_real_or_cx_only< cx_fp16 > { typedef cx_fp16 result; };
#endif
template<typename T> struct arma_blas_real_or_cx_only { };
@@ -145,10 +148,37 @@ template<> struct arma_blas_real_or_cx_only< cx_double > { typedef cx_double res
template<typename T> struct arma_fp16_real_or_cx_only { };
#if defined(ARMA_HAVE_FP16)
template<> struct arma_fp16_real_or_cx_only< fp16 > { typedef fp16 result; };
template<> struct arma_fp16_real_or_cx_only< cx_fp16 > { typedef cx_fp16 result; };
#endif
template<typename T> struct arma_cx_only { };
template<> struct arma_cx_only< cx_float > { typedef cx_float result; };
template<> struct arma_cx_only< cx_double > { typedef cx_double result; };
#if defined(ARMA_HAVE_FP16)
template<> struct arma_cx_only< cx_fp16 > { typedef cx_fp16 result; };
#endif
template<typename T> struct arma_blas_cx_only { };
template<> struct arma_blas_real_only< cx_float > { typedef cx_float result; };
template<> struct arma_blas_real_only< cx_double > { typedef cx_double result; };
template<typename T> struct arma_cx_fp16_only { };
#if defined(ARMA_HAVE_FP16)
template<> struct arma_cx_fp16_only< cx_fp16 > { typedef cx_fp16 result; };
#endif
include/armadillo_bits/traits.hpp
+23 -1
View File
@@ -1119,6 +1119,26 @@ struct is_cx_double< std::complex<double> >
template<typename T1>
struct is_cx_fp16
{
static constexpr bool value = false;
static constexpr bool yes = false;
static constexpr bool no = true;
};
#ifdef ARMA_HAVE_FP16
template<>
struct is_cx_fp16< std::complex<fp16> >
{
static constexpr bool value = true;
static constexpr bool yes = true;
static constexpr bool no = false;
};
#endif
template<typename T1>
struct is_supported_elem_type
{
@@ -1137,7 +1157,8 @@ struct is_supported_elem_type
is_double<T1>::value ||
is_fp16<T1>::value ||
is_cx_float<T1>::value ||
is_cx_double<T1>::value;
is_cx_double<T1>::value ||
is_cx_fp16<T1>::value;
};
@@ -1194,6 +1215,7 @@ template<> struct is_non_integral< std::complex<double> > { static constexpr boo
#if defined(ARMA_HAVE_FP16)
template<> struct is_non_integral< fp16 > { static constexpr bool value = true; };
template<> struct is_non_integral< std::complex<fp16> > { static constexpr bool value = true; };
#endif
include/armadillo_bits/typedef_elem.hpp
+2 -1
View File
@@ -154,7 +154,8 @@ typedef void* void_ptr;
#if defined(ARMA_FORCE_USE_FP16) || defined(ARMA_GOOD_FP16)
#define ARMA_HAVE_FP16
typedef std::float16_t fp16;
typedef std::float16_t fp16;
typedef std::complex<std::float16_t> cx_fp16;
#endif
#if defined(ARMA_FORCE_USE_FP16) && !defined(ARMA_GOOD_FP16)
include/armadillo_bits/typedef_mat.hpp
+11
View File
@@ -104,6 +104,12 @@ typedef Col <fp16> fp16_vec;
typedef Col <fp16> fp16_colvec;
typedef Row <fp16> fp16_rowvec;
typedef Cube<fp16> fp16_cube;
typedef Mat <cx_fp16> cx_fp16_mat;
typedef Col <cx_fp16> cx_fp16_vec;
typedef Col <cx_fp16> cx_fp16_colvec;
typedef Row <cx_fp16> cx_fp16_rowvec;
typedef Cube<cx_fp16> cx_fp16_cube;
#endif
@@ -153,6 +159,11 @@ typedef SpMat <fp16> sp_fp16_mat;
typedef SpCol <fp16> sp_fp16_vec;
typedef SpCol <fp16> sp_fp16_colvec;
typedef SpRow <fp16> sp_fp16_rowvec;
typedef SpMat <cx_fp16> sp_cx_fp16_mat;
typedef SpCol <cx_fp16> sp_cx_fp16_vec;
typedef SpCol <cx_fp16> sp_cx_fp16_colvec;
typedef SpRow <cx_fp16> sp_cx_fp16_rowvec;
#endif