27 #include "../std_transforms_fixed.hpp"
28 #include "../performance_parameters.hpp"
31 unsigned int, const unsigned int *, \
32 IndirectInputArg<uint8_t>, \
35 IndirectOutputArg<uint32_t>, \
36 const uint32_t *, Activation, bool
41 void a64_hybrid_u8u32_dot_6x16( ARGLIST );
42 void a64_hybrid_u8u32_dot_6x16_a55( ARGLIST );
44 class cls_a64_hybrid_u8u32_dot_6x16
47 typedef uint8_t lhs_operand_type;
48 typedef uint8_t rhs_operand_type;
49 typedef uint32_t result_type;
51 typedef void (*kern_type)( ARGLIST );
54 static constexpr
unsigned int out_height()
59 static unsigned int out_width()
64 static constexpr
unsigned int k_unroll()
69 static constexpr
bool supports_accumulate()
74 StdTransformsFixed<rhs_operand_type, result_type, 6, 16, 4> transforms = {};
76 static inline PerformanceParameters get_performance_parameters(
const CPUInfo *
ci)
78 if (std::is_same<T, uint32_t>::value) {
91 if (std::is_same<T, uint8_t>::value) {
94 return { 9.5238, 2.0799, 0.2279 };
96 return { 29.6736, 11.4025, 0.5591 };
98 return { 16.65, 3.92, 0.48 };
100 return { 42.62, 16.32, 0.83 };
108 kern_type kernel=a64_hybrid_u8u32_dot_6x16;
109 cls_a64_hybrid_u8u32_dot_6x16(
const CPUInfo *
ci)
114 case CPUModel::A55r1:
115 kernel=a64_hybrid_u8u32_dot_6x16_a55;
124 #endif // __aarch64__