25 #ifdef ARM_COMPUTE_ENABLE_SVE
27 #include "../std_transforms_sve.hpp"
28 #include "../performance_parameters.hpp"
31 unsigned int, const unsigned int *, \
32 IndirectInputArg<__fp16>, \
35 IndirectOutputArg<__fp16>, \
36 const __fp16 *, Activation, bool
41 void sve_hybrid_fp16_mla_6x4VL( ARGLIST );
42 void sve_hybrid_fp16_mla_6x4VL_a64fx( ARGLIST );
44 class cls_sve_hybrid_fp16_mla_6x4VL
47 typedef __fp16 lhs_operand_type;
48 typedef __fp16 rhs_operand_type;
49 typedef __fp16 result_type;
51 typedef void (*kern_type)( ARGLIST );
54 static constexpr
unsigned int out_height()
59 static unsigned int out_width()
61 return get_vector_length<__fp16>() * 4;
64 static constexpr
unsigned int k_unroll()
69 static constexpr
bool supports_accumulate()
74 StdTransformsSVE<rhs_operand_type, result_type, 6, 4, 1> transforms = {};
76 static inline PerformanceParameters get_performance_parameters(
const CPUInfo *
ci)
78 if (std::is_same<T, __fp16>::value) {
95 kern_type kernel=sve_hybrid_fp16_mla_6x4VL;
96 cls_sve_hybrid_fp16_mla_6x4VL(
const CPUInfo *
ci)
101 case CPUModel::A64FX:
102 kernel=sve_hybrid_fp16_mla_6x4VL_a64fx;
111 #endif // ARM_COMPUTE_ENABLE_SVE