25 #ifdef ARM_COMPUTE_ENABLE_SVE
27 #include "../std_transforms_sve.hpp"
28 #include "../bfloat.hpp"
29 #include "../performance_parameters.hpp"
32 const bfloat16 *, const bfloat16 *, \
33 float *, int, int, int
38 void sve_interleaved_bf16fp32_mmla_8x3VL( ARGLIST );
40 class cls_sve_interleaved_bf16fp32_mmla_8x3VL
44 typedef float result_type;
46 typedef void (*kern_type)( ARGLIST );
49 static constexpr
unsigned int out_height()
54 static unsigned int out_width()
56 return get_vector_length<float>() * 3;
59 static constexpr
unsigned int k_unroll()
65 StdTransformsSVE<operand_type, result_type, 8, 6, 4, 2> transforms = {};
66 StdTransformsSVE<operand_type, result_type, 8, 6, 4, 2, true> transforms_quantized = {};
68 static inline PerformanceParameters get_performance_parameters(
const CPUInfo *
ci)
71 if (std::is_same<T, bfloat16>::value) {
74 return { 31.41, 4.30, 7.14 };
76 return { 7.78, 4.01, 2.43 };
78 return { 62.50, 5.09, 11.32 };
83 if (std::is_same<T, float>::value) {
86 return { 30.86, 2.36, 5.28 };
88 return { 7.75, 2.47, 2.39 };
90 return { 47.63, 5.11, 6.80 };
98 kern_type kernel=sve_interleaved_bf16fp32_mmla_8x3VL;
99 cls_sve_interleaved_bf16fp32_mmla_8x3VL(
const CPUInfo *)
107 #endif // ARM_COMPUTE_ENABLE_SVE