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_mmla_8x3VL( ARGLIST );
39 class cls_sve_interleaved_u8u32_mmla_8x3VL
42 typedef uint8_t operand_type;
43 typedef uint32_t result_type;
45 typedef void (*kern_type)( ARGLIST );
48 static constexpr
unsigned int out_height()
53 static unsigned int out_width()
55 return get_vector_length<uint32_t>() * 3;
58 static constexpr
unsigned int k_unroll()
64 StdTransformsSVE<operand_type, result_type, 8, 6, 8, 2> transforms = {};
65 StdTransformsSVE<operand_type, result_type, 8, 6, 8, 2, true> transforms_quantized = {};
67 static inline PerformanceParameters get_performance_parameters(
const CPUInfo *
ci)
70 if (std::is_same<T, uint32_t>::value) {
73 return { 61.97, 4.11, 7.93 };
75 return { 43.18, 3.57, 2.89 };
77 return { 123.47, 5.03, 11.76 };
82 if (std::is_same<T, uint8_t>::value) {
85 return { 62.00, 4.08, 0.51 };
87 return { 38.02, 1.85, 0.28 };
89 return { 95.28, 7.99, 0.79 };
97 kern_type kernel=sve_interleaved_u8u32_mmla_8x3VL;
98 cls_sve_interleaved_u8u32_mmla_8x3VL(
const CPUInfo *)
106 #endif // ARM_COMPUTE_ENABLE_SVE