25 #ifdef ARM_COMPUTE_ENABLE_SVE
27 #include "../std_transforms_sve.hpp"
28 #include "../performance_parameters.hpp"
31 const uint8_t *, const uint8_t *, \
32 uint32_t *, int, int, int
37 void sve_interleaved_u8u32_dot_8x3VL( ARGLIST );
38 void sve_interleaved_u8u32_dot_8x3VL_a64fx( ARGLIST );
40 class cls_sve_interleaved_u8u32_dot_8x3VL
43 typedef uint8_t operand_type;
44 typedef uint32_t 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<uint32_t>() * 3;
59 static constexpr
unsigned int k_unroll()
65 StdTransformsSVE<operand_type, result_type, 8, 3, 4, 1> transforms = {};
66 StdTransformsSVE<operand_type, result_type, 8, 3, 4, 1, true> transforms_quantized = {};
68 static inline PerformanceParameters get_performance_parameters(
const CPUInfo *
ci)
71 if (std::is_same<T, uint32_t>::value) {
74 return { 31.66, 4.11, 7.94 };
76 return { 27.44, 3.41, 2.90 };
78 return { 63.30, 4.97, 11.52 };
80 return { 109.76, 3.88, 6.76 };
85 if (std::is_same<T, uint8_t>::value) {
88 return { 31.67, 4.04, 0.50 };
90 return { 27.45, 1.65, 0.28 };
92 return { 52.24, 7.49, 0.80 };
94 return { 110.18, 2.34, 0.40 };
102 kern_type kernel=sve_interleaved_u8u32_dot_8x3VL;
103 cls_sve_interleaved_u8u32_dot_8x3VL(
const CPUInfo *
ci)
108 case CPUModel::A64FX:
109 kernel=sve_interleaved_u8u32_dot_8x3VL_a64fx;
118 #endif // ARM_COMPUTE_ENABLE_SVE