109 #define vtrn256_128q(a, b)                   \  argument
111 	float32x4_t vtrn128_temp = a.val[1]; \
112 	a.val[1] = b.val[0];                 \
116 #define vtrn128_64q(a, b)           \  argument
119 	ab = vget_low_f32(a);	    \
121 	cd = vget_high_f32(a);	    \
123 	a = vcombine_f32(ab, ef);   \
127 #define vtrn256_64q(a, b)                  \  argument
131 	a_0 = vget_low_f32(a.val[0]);      \
132 	a_1 = vget_high_f32(a.val[0]);     \
133 	a_2 = vget_low_f32(a.val[1]);      \
134 	a_3 = vget_high_f32(a.val[1]);     \
139 	a.val[0] = vcombine_f32(a_0, b_0); \
140 	a.val[1] = vcombine_f32(a_2, b_2); \
145 #define vtrn128_32q(a, b)                               \  argument
147 	float32x4x2_t vtrn32_tmp = vtrnq_f32((a), (b)); \
148 	(a) = vtrn32_tmp.val[0];                        \
152 #define vtrn256_32q(a, b)               \  argument
154 	float32x4x2_t vtrn32_tmp_1 = vtrnq_f32((a.val[0]), (b.val[0])); \
155 	float32x4x2_t vtrn32_tmp_2 = vtrnq_f32((a.val[1]), (b.val[1])); \
156 	a.val[0] = vtrn32_tmp_1.val[0]; \
157 	a.val[1] = vtrn32_tmp_2.val[0]; \
162 #define vminmaxq(a, b)                    \  argument
164 	float32x4_t minmax_tmp = (a);     \
165 	(a) = vminq_f32((a), (b));        \
169 #define vminmax256q(a, b)                         \  argument
171 	float32x4x2_t minmax256_tmp = (a);        \
172 	a.val[0] = vminq_f32(a.val[0], b.val[0]); \
173 	a.val[1] = vminq_f32(a.val[1], b.val[1]); \
178 #define vrev128q_f32(a) \  argument
179         vcombine_f32(vrev64_f32(vget_high_f32(a)), vrev64_f32(vget_low_f32(a)))
181 #define vrev256q_f32(a)     \  argument
183 …oat32x4_t rev_tmp = vcombine_f32(vrev64_f32(vget_high_f32(a.val[0])), vrev64_f32(vget_low_f32(a.va…
184 …a.val[0] = vcombine_f32(vrev64_f32(vget_high_f32(a.val[1])), vrev64_f32(vget_low_f32(a.val[1]))); …
185 	a.val[1] = rev_tmp; \
188 #define vldrev128q_f32(a, p) \  argument
190 	a = vld1q_f32(p);    \
191 	a = vrev128q_f32(a); \