27 #include "../std_transforms_fixed.hpp"
28 #include "../performance_parameters.hpp"
31 const int8_t *, const int8_t *, \
32 int32_t *, int, int, int
37 void a64_interleaved_s8s32_mmla_8x12( ARGLIST );
38 void a64_interleaved_s8s32_mmla_8x12_a510( ARGLIST );
40 class cls_a64_interleaved_s8s32_mmla_8x12
43 typedef int8_t operand_type;
44 typedef int32_t 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, 8> transforms = {};
66 StdTransformsFixed<operand_type, result_type, 8, 12, 8, true> transforms_quantized = {};
68 static inline PerformanceParameters get_performance_parameters(
const CPUInfo *
ci)
71 if (std::is_same<T, int32_t>::value) {
74 return { 62.57, 4.08, 8.01 };
76 return { 48.25, 3.53, 3.71 };
78 return { 117.02, 4.98, 10.87 };
83 if (std::is_same<T, int8_t>::value) {
86 return { 62.53, 3.70, 0.50 };
88 return { 48.22, 2.49, 0.29 };
90 return { 75.54, 8.06, 0.63 };
98 kern_type kernel=a64_interleaved_s8s32_mmla_8x12;
99 cls_a64_interleaved_s8s32_mmla_8x12(
const CPUInfo *
ci)
105 kernel=a64_interleaved_s8s32_mmla_8x12_a510;
114 #endif // __aarch64__