27 #include "../std_transforms_fixed.hpp"
28 #include "../bfloat.hpp"
29 #include "../performance_parameters.hpp"
32 const bfloat16 *, const bfloat16 *, \
33 float *, int, int, int
38 void a64_interleaved_bf16fp32_dot_8x12( ARGLIST );
40 class cls_a64_interleaved_bf16fp32_dot_8x12
44 typedef float result_type;
46 typedef void (*kern_type)( ARGLIST );
49 static constexpr
unsigned int out_height()
54 static unsigned int out_width()
59 static constexpr
unsigned int k_unroll()
65 StdTransformsFixed<operand_type, result_type, 8, 12, 2> transforms = {};
66 StdTransformsFixed<operand_type, result_type, 8, 12, 2, true> transforms_quantized = {};
68 static inline PerformanceParameters get_performance_parameters(
const CPUInfo *
ci)
71 if (std::is_same<T, bfloat16>::value) {
74 return { 15.93, 4.16, 7.19 };
76 return { 20.88, 5.10, 6.57 };
78 return { 7.77, 3.69, 3.02 };
86 kern_type kernel=a64_interleaved_bf16fp32_dot_8x12;
87 cls_a64_interleaved_bf16fp32_dot_8x12(
const CPUInfo *)