[feature] arm: speed up fp16 exp_ps floor step on aarch64 (#6659)

Author: crafcat7

src/layer/arm/neon_mathfun_fp16s.h

@@ -197,13 +197,7 @@ static inline float16x4_t exp_ps_f16(float16x4_t x)
 #endif
 
     /* perform a floorf */
-    tmp = vcvt_f16_s16(vcvt_s16_f16(fx));
-
-    /* if greater, substract 1 */
-    uint16x4_t mask = vcgt_f16(tmp, fx);
-    mask = vand_u16(mask, (uint16x4_t)(one));
-
-    fx = vsub_f16(tmp, (float16x4_t)(mask));
+    fx = vrndm_f16(fx);
 
 #if defined(_MSC_VER) && !defined(__clang__)
     tmp = vmul_f16(fx, vcvt_f16_f32(vdupq_n_f32(c_cephes_exp_C1)));
@@ -255,13 +249,7 @@ static inline float16x8_t exp_ps_f16(float16x8_t x)
 #endif
 
     /* perform a floorf */
-    tmp = vcvtq_f16_s16(vcvtq_s16_f16(fx));
-
-    /* if greater, substract 1 */
-    uint16x8_t mask = vcgtq_f16(tmp, fx);
-    mask = vandq_u16(mask, vreinterpretq_u16_f16(one));
-
-    fx = vsubq_f16(tmp, vreinterpretq_f16_u16(mask));
+    fx = vrndmq_f16(fx);
 
 #if defined(_MSC_VER) && !defined(__clang__)
     float16x4_t _c_cephes_exp_C1 = vcvt_f16_f32(vdupq_n_f32(c_cephes_exp_C1));