26 #if defined(__aarch64__) && (defined(FP16_KERNELS) || defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC))
28 #include "../performance_parameters.hpp"
29 #include "../std_transforms_fixed.hpp"
34 void a64_hgemm_asimd_8x24(
const __fp16 *,
const __fp16 *, __fp16 *,
int,
int,
int);
35 void a64_hgemm_asimd_8x24_a55r1(
const __fp16 *,
const __fp16 *, __fp16 *,
int,
int,
int);
36 void a64_hgemm_asimd_8x24_x1(
const __fp16 *,
const __fp16 *, __fp16 *,
int,
int,
int);
42 class cls_a64_hgemm_8x24 {
44 typedef __fp16 operand_type;
45 typedef __fp16 result_type;
47 typedef void (*kern_type)(
const __fp16 *,
const __fp16 *, __fp16 *, int, int, int);
50 static unsigned int out_width() {
54 static unsigned int out_height() {
58 static unsigned int k_unroll() {
63 StdTransformsFixed<operand_type, result_type, 8, 24> transforms = {};
66 static PerformanceParameters get_performance_parameters(
const CPUInfo *
ci) {
69 return { 7.16, 1.14, 0.67 };
72 return { 12.67, 3.98, 1.16 };
77 kern_type kernel = a64_hgemm_asimd_8x24;
82 if (model == CPUModel::A55r1) {
83 kernel = a64_hgemm_asimd_8x24_a55r1;
84 }
else if (model == CPUModel::X1) {
85 kernel = a64_hgemm_asimd_8x24_x1;
92 #endif // __aarch64__ && (FP16_KERNELS || __ARM_FEATURE_FP16_VECTOR_ARITHMETIC)