summaryrefslogtreecommitdiffstats
path: root/third_party/jpeg-xl/lib/jxl/fast_dct256-inl.h
diff options
context:
space:
mode:
Diffstat (limited to 'third_party/jpeg-xl/lib/jxl/fast_dct256-inl.h')
-rw-r--r--third_party/jpeg-xl/lib/jxl/fast_dct256-inl.h4811
1 files changed, 4811 insertions, 0 deletions
diff --git a/third_party/jpeg-xl/lib/jxl/fast_dct256-inl.h b/third_party/jpeg-xl/lib/jxl/fast_dct256-inl.h
new file mode 100644
index 0000000000..a823440af2
--- /dev/null
+++ b/third_party/jpeg-xl/lib/jxl/fast_dct256-inl.h
@@ -0,0 +1,4811 @@
+// Copyright (c) the JPEG XL Project Authors. All rights reserved.
+//
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+/* This file is automatically generated. Do not modify it directly. */
+#if HWY_TARGET != HWY_NEON
+#error "only include this file from fast_dct-inl.h"
+#endif
+
+constexpr size_t FastIDCTIntegerBits(FastDCTTag<256>) { return 3; }
+
+void FastIDCT(FastDCTTag<256>, const int16_t* in, size_t in_stride,
+ int16_t* out, size_t out_stride, size_t count) {
+ JXL_ASSERT(count % 8 == 0);
+ for (size_t i = 0; i < count; i += 8) {
+ int16x8_t v0 = vld1q_s16(in + in_stride * 0 + i);
+ int16x8_t v1 = vld1q_s16(in + in_stride * 128 + i);
+ int16x8_t v2 = vaddq_s16(v0, v1);
+ int16x8_t v3 = vld1q_s16(in + in_stride * 64 + i);
+ int16x8_t v4_tmp = vqrdmulhq_n_s16(v3, 13573);
+ int16x8_t v4 = vaddq_s16(v4_tmp, v3);
+ int16x8_t v5 = vld1q_s16(in + in_stride * 192 + i);
+ int16x8_t v6 = vaddq_s16(v5, v3);
+ int16x8_t v7 = vaddq_s16(v4, v6);
+ int16x8_t v8 = vqrdmulhq_n_s16(v7, 17734);
+ int16x8_t v9 = vaddq_s16(v2, v8);
+ int16x8_t v10 = vld1q_s16(in + in_stride * 32 + i);
+ int16x8_t v11_tmp = vqrdmulhq_n_s16(v10, 13573);
+ int16x8_t v11 = vaddq_s16(v11_tmp, v10);
+ int16x8_t v12 = vld1q_s16(in + in_stride * 160 + i);
+ int16x8_t v13 = vld1q_s16(in + in_stride * 96 + i);
+ int16x8_t v14 = vaddq_s16(v12, v13);
+ int16x8_t v15 = vaddq_s16(v11, v14);
+ int16x8_t v16 = vaddq_s16(v13, v10);
+ int16x8_t v17_tmp = vqrdmulhq_n_s16(v16, 13573);
+ int16x8_t v17 = vaddq_s16(v17_tmp, v16);
+ int16x8_t v18 = vld1q_s16(in + in_stride * 224 + i);
+ int16x8_t v19 = vaddq_s16(v18, v12);
+ int16x8_t v20 = vaddq_s16(v19, v16);
+ int16x8_t v21 = vaddq_s16(v17, v20);
+ int16x8_t v22 = vqrdmulhq_n_s16(v21, 17734);
+ int16x8_t v23 = vaddq_s16(v15, v22);
+ int16x8_t v24 = vqrdmulhq_n_s16(v23, 16705);
+ int16x8_t v25 = vaddq_s16(v9, v24);
+ int16x8_t v26 = vld1q_s16(in + in_stride * 16 + i);
+ int16x8_t v27_tmp = vqrdmulhq_n_s16(v26, 13573);
+ int16x8_t v27 = vaddq_s16(v27_tmp, v26);
+ int16x8_t v28 = vld1q_s16(in + in_stride * 144 + i);
+ int16x8_t v29 = vld1q_s16(in + in_stride * 112 + i);
+ int16x8_t v30 = vaddq_s16(v28, v29);
+ int16x8_t v31 = vaddq_s16(v27, v30);
+ int16x8_t v32 = vld1q_s16(in + in_stride * 80 + i);
+ int16x8_t v33 = vld1q_s16(in + in_stride * 48 + i);
+ int16x8_t v34 = vaddq_s16(v32, v33);
+ int16x8_t v35_tmp = vqrdmulhq_n_s16(v34, 13573);
+ int16x8_t v35 = vaddq_s16(v35_tmp, v34);
+ int16x8_t v36 = vld1q_s16(in + in_stride * 208 + i);
+ int16x8_t v37 = vld1q_s16(in + in_stride * 176 + i);
+ int16x8_t v38 = vaddq_s16(v36, v37);
+ int16x8_t v39 = vaddq_s16(v38, v34);
+ int16x8_t v40 = vaddq_s16(v35, v39);
+ int16x8_t v41 = vqrdmulhq_n_s16(v40, 17734);
+ int16x8_t v42 = vaddq_s16(v31, v41);
+ int16x8_t v43 = vaddq_s16(v33, v26);
+ int16x8_t v44_tmp = vqrdmulhq_n_s16(v43, 13573);
+ int16x8_t v44 = vaddq_s16(v44_tmp, v43);
+ int16x8_t v45 = vaddq_s16(v37, v28);
+ int16x8_t v46 = vaddq_s16(v29, v32);
+ int16x8_t v47 = vaddq_s16(v45, v46);
+ int16x8_t v48 = vaddq_s16(v44, v47);
+ int16x8_t v49 = vaddq_s16(v46, v43);
+ int16x8_t v50_tmp = vqrdmulhq_n_s16(v49, 13573);
+ int16x8_t v50 = vaddq_s16(v50_tmp, v49);
+ int16x8_t v51 = vld1q_s16(in + in_stride * 240 + i);
+ int16x8_t v52 = vaddq_s16(v51, v36);
+ int16x8_t v53 = vaddq_s16(v52, v45);
+ int16x8_t v54 = vaddq_s16(v53, v49);
+ int16x8_t v55 = vaddq_s16(v50, v54);
+ int16x8_t v56 = vqrdmulhq_n_s16(v55, 17734);
+ int16x8_t v57 = vaddq_s16(v48, v56);
+ int16x8_t v58 = vqrdmulhq_n_s16(v57, 16705);
+ int16x8_t v59 = vaddq_s16(v42, v58);
+ int16x8_t v60 = vqrdmulhq_n_s16(v59, 16463);
+ int16x8_t v61 = vaddq_s16(v25, v60);
+ int16x8_t v62 = vld1q_s16(in + in_stride * 8 + i);
+ int16x8_t v63_tmp = vqrdmulhq_n_s16(v62, 13573);
+ int16x8_t v63 = vaddq_s16(v63_tmp, v62);
+ int16x8_t v64 = vld1q_s16(in + in_stride * 136 + i);
+ int16x8_t v65 = vld1q_s16(in + in_stride * 120 + i);
+ int16x8_t v66 = vaddq_s16(v64, v65);
+ int16x8_t v67 = vaddq_s16(v63, v66);
+ int16x8_t v68 = vld1q_s16(in + in_stride * 72 + i);
+ int16x8_t v69 = vld1q_s16(in + in_stride * 56 + i);
+ int16x8_t v70 = vaddq_s16(v68, v69);
+ int16x8_t v71_tmp = vqrdmulhq_n_s16(v70, 13573);
+ int16x8_t v71 = vaddq_s16(v71_tmp, v70);
+ int16x8_t v72 = vld1q_s16(in + in_stride * 200 + i);
+ int16x8_t v73 = vld1q_s16(in + in_stride * 184 + i);
+ int16x8_t v74 = vaddq_s16(v72, v73);
+ int16x8_t v75 = vaddq_s16(v74, v70);
+ int16x8_t v76 = vaddq_s16(v71, v75);
+ int16x8_t v77 = vqrdmulhq_n_s16(v76, 17734);
+ int16x8_t v78 = vaddq_s16(v67, v77);
+ int16x8_t v79 = vld1q_s16(in + in_stride * 40 + i);
+ int16x8_t v80 = vld1q_s16(in + in_stride * 24 + i);
+ int16x8_t v81 = vaddq_s16(v79, v80);
+ int16x8_t v82_tmp = vqrdmulhq_n_s16(v81, 13573);
+ int16x8_t v82 = vaddq_s16(v82_tmp, v81);
+ int16x8_t v83 = vld1q_s16(in + in_stride * 168 + i);
+ int16x8_t v84 = vld1q_s16(in + in_stride * 152 + i);
+ int16x8_t v85 = vaddq_s16(v83, v84);
+ int16x8_t v86 = vld1q_s16(in + in_stride * 104 + i);
+ int16x8_t v87 = vld1q_s16(in + in_stride * 88 + i);
+ int16x8_t v88 = vaddq_s16(v86, v87);
+ int16x8_t v89 = vaddq_s16(v85, v88);
+ int16x8_t v90 = vaddq_s16(v82, v89);
+ int16x8_t v91 = vaddq_s16(v88, v81);
+ int16x8_t v92_tmp = vqrdmulhq_n_s16(v91, 13573);
+ int16x8_t v92 = vaddq_s16(v92_tmp, v91);
+ int16x8_t v93 = vld1q_s16(in + in_stride * 232 + i);
+ int16x8_t v94 = vld1q_s16(in + in_stride * 216 + i);
+ int16x8_t v95 = vaddq_s16(v93, v94);
+ int16x8_t v96 = vaddq_s16(v95, v85);
+ int16x8_t v97 = vaddq_s16(v96, v91);
+ int16x8_t v98 = vaddq_s16(v92, v97);
+ int16x8_t v99 = vqrdmulhq_n_s16(v98, 17734);
+ int16x8_t v100 = vaddq_s16(v90, v99);
+ int16x8_t v101 = vqrdmulhq_n_s16(v100, 16705);
+ int16x8_t v102 = vaddq_s16(v78, v101);
+ int16x8_t v103 = vaddq_s16(v80, v62);
+ int16x8_t v104_tmp = vqrdmulhq_n_s16(v103, 13573);
+ int16x8_t v104 = vaddq_s16(v104_tmp, v103);
+ int16x8_t v105 = vaddq_s16(v84, v64);
+ int16x8_t v106 = vaddq_s16(v65, v86);
+ int16x8_t v107 = vaddq_s16(v105, v106);
+ int16x8_t v108 = vaddq_s16(v104, v107);
+ int16x8_t v109 = vaddq_s16(v87, v68);
+ int16x8_t v110 = vaddq_s16(v69, v79);
+ int16x8_t v111 = vaddq_s16(v109, v110);
+ int16x8_t v112_tmp = vqrdmulhq_n_s16(v111, 13573);
+ int16x8_t v112 = vaddq_s16(v112_tmp, v111);
+ int16x8_t v113 = vaddq_s16(v94, v72);
+ int16x8_t v114 = vaddq_s16(v73, v83);
+ int16x8_t v115 = vaddq_s16(v113, v114);
+ int16x8_t v116 = vaddq_s16(v115, v111);
+ int16x8_t v117 = vaddq_s16(v112, v116);
+ int16x8_t v118 = vqrdmulhq_n_s16(v117, 17734);
+ int16x8_t v119 = vaddq_s16(v108, v118);
+ int16x8_t v120 = vaddq_s16(v110, v103);
+ int16x8_t v121_tmp = vqrdmulhq_n_s16(v120, 13573);
+ int16x8_t v121 = vaddq_s16(v121_tmp, v120);
+ int16x8_t v122 = vaddq_s16(v114, v105);
+ int16x8_t v123 = vaddq_s16(v106, v109);
+ int16x8_t v124 = vaddq_s16(v122, v123);
+ int16x8_t v125 = vaddq_s16(v121, v124);
+ int16x8_t v126 = vaddq_s16(v123, v120);
+ int16x8_t v127_tmp = vqrdmulhq_n_s16(v126, 13573);
+ int16x8_t v127 = vaddq_s16(v127_tmp, v126);
+ int16x8_t v128 = vld1q_s16(in + in_stride * 248 + i);
+ int16x8_t v129 = vaddq_s16(v128, v93);
+ int16x8_t v130 = vaddq_s16(v129, v113);
+ int16x8_t v131 = vaddq_s16(v130, v122);
+ int16x8_t v132 = vaddq_s16(v131, v126);
+ int16x8_t v133 = vaddq_s16(v127, v132);
+ int16x8_t v134 = vqrdmulhq_n_s16(v133, 17734);
+ int16x8_t v135 = vaddq_s16(v125, v134);
+ int16x8_t v136 = vqrdmulhq_n_s16(v135, 16705);
+ int16x8_t v137 = vaddq_s16(v119, v136);
+ int16x8_t v138 = vqrdmulhq_n_s16(v137, 16463);
+ int16x8_t v139 = vaddq_s16(v102, v138);
+ int16x8_t v140 = vqrdmulhq_n_s16(v139, 16404);
+ int16x8_t v141 = vaddq_s16(v61, v140);
+ int16x8_t v142 = vld1q_s16(in + in_stride * 4 + i);
+ int16x8_t v143_tmp = vqrdmulhq_n_s16(v142, 13573);
+ int16x8_t v143 = vaddq_s16(v143_tmp, v142);
+ int16x8_t v144 = vld1q_s16(in + in_stride * 132 + i);
+ int16x8_t v145 = vld1q_s16(in + in_stride * 124 + i);
+ int16x8_t v146 = vaddq_s16(v144, v145);
+ int16x8_t v147 = vaddq_s16(v143, v146);
+ int16x8_t v148 = vld1q_s16(in + in_stride * 68 + i);
+ int16x8_t v149 = vld1q_s16(in + in_stride * 60 + i);
+ int16x8_t v150 = vaddq_s16(v148, v149);
+ int16x8_t v151_tmp = vqrdmulhq_n_s16(v150, 13573);
+ int16x8_t v151 = vaddq_s16(v151_tmp, v150);
+ int16x8_t v152 = vld1q_s16(in + in_stride * 196 + i);
+ int16x8_t v153 = vld1q_s16(in + in_stride * 188 + i);
+ int16x8_t v154 = vaddq_s16(v152, v153);
+ int16x8_t v155 = vaddq_s16(v154, v150);
+ int16x8_t v156 = vaddq_s16(v151, v155);
+ int16x8_t v157 = vqrdmulhq_n_s16(v156, 17734);
+ int16x8_t v158 = vaddq_s16(v147, v157);
+ int16x8_t v159 = vld1q_s16(in + in_stride * 36 + i);
+ int16x8_t v160 = vld1q_s16(in + in_stride * 28 + i);
+ int16x8_t v161 = vaddq_s16(v159, v160);
+ int16x8_t v162_tmp = vqrdmulhq_n_s16(v161, 13573);
+ int16x8_t v162 = vaddq_s16(v162_tmp, v161);
+ int16x8_t v163 = vld1q_s16(in + in_stride * 164 + i);
+ int16x8_t v164 = vld1q_s16(in + in_stride * 156 + i);
+ int16x8_t v165 = vaddq_s16(v163, v164);
+ int16x8_t v166 = vld1q_s16(in + in_stride * 100 + i);
+ int16x8_t v167 = vld1q_s16(in + in_stride * 92 + i);
+ int16x8_t v168 = vaddq_s16(v166, v167);
+ int16x8_t v169 = vaddq_s16(v165, v168);
+ int16x8_t v170 = vaddq_s16(v162, v169);
+ int16x8_t v171 = vaddq_s16(v168, v161);
+ int16x8_t v172_tmp = vqrdmulhq_n_s16(v171, 13573);
+ int16x8_t v172 = vaddq_s16(v172_tmp, v171);
+ int16x8_t v173 = vld1q_s16(in + in_stride * 228 + i);
+ int16x8_t v174 = vld1q_s16(in + in_stride * 220 + i);
+ int16x8_t v175 = vaddq_s16(v173, v174);
+ int16x8_t v176 = vaddq_s16(v175, v165);
+ int16x8_t v177 = vaddq_s16(v176, v171);
+ int16x8_t v178 = vaddq_s16(v172, v177);
+ int16x8_t v179 = vqrdmulhq_n_s16(v178, 17734);
+ int16x8_t v180 = vaddq_s16(v170, v179);
+ int16x8_t v181 = vqrdmulhq_n_s16(v180, 16705);
+ int16x8_t v182 = vaddq_s16(v158, v181);
+ int16x8_t v183 = vld1q_s16(in + in_stride * 20 + i);
+ int16x8_t v184 = vld1q_s16(in + in_stride * 12 + i);
+ int16x8_t v185 = vaddq_s16(v183, v184);
+ int16x8_t v186_tmp = vqrdmulhq_n_s16(v185, 13573);
+ int16x8_t v186 = vaddq_s16(v186_tmp, v185);
+ int16x8_t v187 = vld1q_s16(in + in_stride * 148 + i);
+ int16x8_t v188 = vld1q_s16(in + in_stride * 140 + i);
+ int16x8_t v189 = vaddq_s16(v187, v188);
+ int16x8_t v190 = vld1q_s16(in + in_stride * 116 + i);
+ int16x8_t v191 = vld1q_s16(in + in_stride * 108 + i);
+ int16x8_t v192 = vaddq_s16(v190, v191);
+ int16x8_t v193 = vaddq_s16(v189, v192);
+ int16x8_t v194 = vaddq_s16(v186, v193);
+ int16x8_t v195 = vld1q_s16(in + in_stride * 84 + i);
+ int16x8_t v196 = vld1q_s16(in + in_stride * 76 + i);
+ int16x8_t v197 = vaddq_s16(v195, v196);
+ int16x8_t v198 = vld1q_s16(in + in_stride * 52 + i);
+ int16x8_t v199 = vld1q_s16(in + in_stride * 44 + i);
+ int16x8_t v200 = vaddq_s16(v198, v199);
+ int16x8_t v201 = vaddq_s16(v197, v200);
+ int16x8_t v202_tmp = vqrdmulhq_n_s16(v201, 13573);
+ int16x8_t v202 = vaddq_s16(v202_tmp, v201);
+ int16x8_t v203 = vld1q_s16(in + in_stride * 212 + i);
+ int16x8_t v204 = vld1q_s16(in + in_stride * 204 + i);
+ int16x8_t v205 = vaddq_s16(v203, v204);
+ int16x8_t v206 = vld1q_s16(in + in_stride * 180 + i);
+ int16x8_t v207 = vld1q_s16(in + in_stride * 172 + i);
+ int16x8_t v208 = vaddq_s16(v206, v207);
+ int16x8_t v209 = vaddq_s16(v205, v208);
+ int16x8_t v210 = vaddq_s16(v209, v201);
+ int16x8_t v211 = vaddq_s16(v202, v210);
+ int16x8_t v212 = vqrdmulhq_n_s16(v211, 17734);
+ int16x8_t v213 = vaddq_s16(v194, v212);
+ int16x8_t v214 = vaddq_s16(v200, v185);
+ int16x8_t v215_tmp = vqrdmulhq_n_s16(v214, 13573);
+ int16x8_t v215 = vaddq_s16(v215_tmp, v214);
+ int16x8_t v216 = vaddq_s16(v208, v189);
+ int16x8_t v217 = vaddq_s16(v192, v197);
+ int16x8_t v218 = vaddq_s16(v216, v217);
+ int16x8_t v219 = vaddq_s16(v215, v218);
+ int16x8_t v220 = vaddq_s16(v217, v214);
+ int16x8_t v221_tmp = vqrdmulhq_n_s16(v220, 13573);
+ int16x8_t v221 = vaddq_s16(v221_tmp, v220);
+ int16x8_t v222 = vld1q_s16(in + in_stride * 244 + i);
+ int16x8_t v223 = vld1q_s16(in + in_stride * 236 + i);
+ int16x8_t v224 = vaddq_s16(v222, v223);
+ int16x8_t v225 = vaddq_s16(v224, v205);
+ int16x8_t v226 = vaddq_s16(v225, v216);
+ int16x8_t v227 = vaddq_s16(v226, v220);
+ int16x8_t v228 = vaddq_s16(v221, v227);
+ int16x8_t v229 = vqrdmulhq_n_s16(v228, 17734);
+ int16x8_t v230 = vaddq_s16(v219, v229);
+ int16x8_t v231 = vqrdmulhq_n_s16(v230, 16705);
+ int16x8_t v232 = vaddq_s16(v213, v231);
+ int16x8_t v233 = vqrdmulhq_n_s16(v232, 16463);
+ int16x8_t v234 = vaddq_s16(v182, v233);
+ int16x8_t v235 = vaddq_s16(v184, v142);
+ int16x8_t v236_tmp = vqrdmulhq_n_s16(v235, 13573);
+ int16x8_t v236 = vaddq_s16(v236_tmp, v235);
+ int16x8_t v237 = vaddq_s16(v188, v144);
+ int16x8_t v238 = vaddq_s16(v145, v190);
+ int16x8_t v239 = vaddq_s16(v237, v238);
+ int16x8_t v240 = vaddq_s16(v236, v239);
+ int16x8_t v241 = vaddq_s16(v196, v148);
+ int16x8_t v242 = vaddq_s16(v149, v198);
+ int16x8_t v243 = vaddq_s16(v241, v242);
+ int16x8_t v244_tmp = vqrdmulhq_n_s16(v243, 13573);
+ int16x8_t v244 = vaddq_s16(v244_tmp, v243);
+ int16x8_t v245 = vaddq_s16(v204, v152);
+ int16x8_t v246 = vaddq_s16(v153, v206);
+ int16x8_t v247 = vaddq_s16(v245, v246);
+ int16x8_t v248 = vaddq_s16(v247, v243);
+ int16x8_t v249 = vaddq_s16(v244, v248);
+ int16x8_t v250 = vqrdmulhq_n_s16(v249, 17734);
+ int16x8_t v251 = vaddq_s16(v240, v250);
+ int16x8_t v252 = vaddq_s16(v199, v159);
+ int16x8_t v253 = vaddq_s16(v160, v183);
+ int16x8_t v254 = vaddq_s16(v252, v253);
+ int16x8_t v255_tmp = vqrdmulhq_n_s16(v254, 13573);
+ int16x8_t v255 = vaddq_s16(v255_tmp, v254);
+ int16x8_t v256 = vaddq_s16(v207, v163);
+ int16x8_t v257 = vaddq_s16(v164, v187);
+ int16x8_t v258 = vaddq_s16(v256, v257);
+ int16x8_t v259 = vaddq_s16(v191, v166);
+ int16x8_t v260 = vaddq_s16(v167, v195);
+ int16x8_t v261 = vaddq_s16(v259, v260);
+ int16x8_t v262 = vaddq_s16(v258, v261);
+ int16x8_t v263 = vaddq_s16(v255, v262);
+ int16x8_t v264 = vaddq_s16(v261, v254);
+ int16x8_t v265_tmp = vqrdmulhq_n_s16(v264, 13573);
+ int16x8_t v265 = vaddq_s16(v265_tmp, v264);
+ int16x8_t v266 = vaddq_s16(v223, v173);
+ int16x8_t v267 = vaddq_s16(v174, v203);
+ int16x8_t v268 = vaddq_s16(v266, v267);
+ int16x8_t v269 = vaddq_s16(v268, v258);
+ int16x8_t v270 = vaddq_s16(v269, v264);
+ int16x8_t v271 = vaddq_s16(v265, v270);
+ int16x8_t v272 = vqrdmulhq_n_s16(v271, 17734);
+ int16x8_t v273 = vaddq_s16(v263, v272);
+ int16x8_t v274 = vqrdmulhq_n_s16(v273, 16705);
+ int16x8_t v275 = vaddq_s16(v251, v274);
+ int16x8_t v276 = vaddq_s16(v253, v235);
+ int16x8_t v277_tmp = vqrdmulhq_n_s16(v276, 13573);
+ int16x8_t v277 = vaddq_s16(v277_tmp, v276);
+ int16x8_t v278 = vaddq_s16(v257, v237);
+ int16x8_t v279 = vaddq_s16(v238, v259);
+ int16x8_t v280 = vaddq_s16(v278, v279);
+ int16x8_t v281 = vaddq_s16(v277, v280);
+ int16x8_t v282 = vaddq_s16(v260, v241);
+ int16x8_t v283 = vaddq_s16(v242, v252);
+ int16x8_t v284 = vaddq_s16(v282, v283);
+ int16x8_t v285_tmp = vqrdmulhq_n_s16(v284, 13573);
+ int16x8_t v285 = vaddq_s16(v285_tmp, v284);
+ int16x8_t v286 = vaddq_s16(v267, v245);
+ int16x8_t v287 = vaddq_s16(v246, v256);
+ int16x8_t v288 = vaddq_s16(v286, v287);
+ int16x8_t v289 = vaddq_s16(v288, v284);
+ int16x8_t v290 = vaddq_s16(v285, v289);
+ int16x8_t v291 = vqrdmulhq_n_s16(v290, 17734);
+ int16x8_t v292 = vaddq_s16(v281, v291);
+ int16x8_t v293 = vaddq_s16(v283, v276);
+ int16x8_t v294_tmp = vqrdmulhq_n_s16(v293, 13573);
+ int16x8_t v294 = vaddq_s16(v294_tmp, v293);
+ int16x8_t v295 = vaddq_s16(v287, v278);
+ int16x8_t v296 = vaddq_s16(v279, v282);
+ int16x8_t v297 = vaddq_s16(v295, v296);
+ int16x8_t v298 = vaddq_s16(v294, v297);
+ int16x8_t v299 = vaddq_s16(v296, v293);
+ int16x8_t v300_tmp = vqrdmulhq_n_s16(v299, 13573);
+ int16x8_t v300 = vaddq_s16(v300_tmp, v299);
+ int16x8_t v301 = vld1q_s16(in + in_stride * 252 + i);
+ int16x8_t v302 = vaddq_s16(v301, v222);
+ int16x8_t v303 = vaddq_s16(v302, v266);
+ int16x8_t v304 = vaddq_s16(v303, v286);
+ int16x8_t v305 = vaddq_s16(v304, v295);
+ int16x8_t v306 = vaddq_s16(v305, v299);
+ int16x8_t v307 = vaddq_s16(v300, v306);
+ int16x8_t v308 = vqrdmulhq_n_s16(v307, 17734);
+ int16x8_t v309 = vaddq_s16(v298, v308);
+ int16x8_t v310 = vqrdmulhq_n_s16(v309, 16705);
+ int16x8_t v311 = vaddq_s16(v292, v310);
+ int16x8_t v312 = vqrdmulhq_n_s16(v311, 16463);
+ int16x8_t v313 = vaddq_s16(v275, v312);
+ int16x8_t v314 = vqrdmulhq_n_s16(v313, 16404);
+ int16x8_t v315 = vaddq_s16(v234, v314);
+ int16x8_t v316 = vqrdmulhq_n_s16(v315, 16389);
+ int16x8_t v317 = vaddq_s16(v141, v316);
+ int16x8_t v318 = vld1q_s16(in + in_stride * 2 + i);
+ int16x8_t v319_tmp = vqrdmulhq_n_s16(v318, 13573);
+ int16x8_t v319 = vaddq_s16(v319_tmp, v318);
+ int16x8_t v320 = vld1q_s16(in + in_stride * 130 + i);
+ int16x8_t v321 = vld1q_s16(in + in_stride * 126 + i);
+ int16x8_t v322 = vaddq_s16(v320, v321);
+ int16x8_t v323 = vaddq_s16(v319, v322);
+ int16x8_t v324 = vld1q_s16(in + in_stride * 66 + i);
+ int16x8_t v325 = vld1q_s16(in + in_stride * 62 + i);
+ int16x8_t v326 = vaddq_s16(v324, v325);
+ int16x8_t v327_tmp = vqrdmulhq_n_s16(v326, 13573);
+ int16x8_t v327 = vaddq_s16(v327_tmp, v326);
+ int16x8_t v328 = vld1q_s16(in + in_stride * 194 + i);
+ int16x8_t v329 = vld1q_s16(in + in_stride * 190 + i);
+ int16x8_t v330 = vaddq_s16(v328, v329);
+ int16x8_t v331 = vaddq_s16(v330, v326);
+ int16x8_t v332 = vaddq_s16(v327, v331);
+ int16x8_t v333 = vqrdmulhq_n_s16(v332, 17734);
+ int16x8_t v334 = vaddq_s16(v323, v333);
+ int16x8_t v335 = vld1q_s16(in + in_stride * 34 + i);
+ int16x8_t v336 = vld1q_s16(in + in_stride * 30 + i);
+ int16x8_t v337 = vaddq_s16(v335, v336);
+ int16x8_t v338_tmp = vqrdmulhq_n_s16(v337, 13573);
+ int16x8_t v338 = vaddq_s16(v338_tmp, v337);
+ int16x8_t v339 = vld1q_s16(in + in_stride * 162 + i);
+ int16x8_t v340 = vld1q_s16(in + in_stride * 158 + i);
+ int16x8_t v341 = vaddq_s16(v339, v340);
+ int16x8_t v342 = vld1q_s16(in + in_stride * 98 + i);
+ int16x8_t v343 = vld1q_s16(in + in_stride * 94 + i);
+ int16x8_t v344 = vaddq_s16(v342, v343);
+ int16x8_t v345 = vaddq_s16(v341, v344);
+ int16x8_t v346 = vaddq_s16(v338, v345);
+ int16x8_t v347 = vaddq_s16(v344, v337);
+ int16x8_t v348_tmp = vqrdmulhq_n_s16(v347, 13573);
+ int16x8_t v348 = vaddq_s16(v348_tmp, v347);
+ int16x8_t v349 = vld1q_s16(in + in_stride * 226 + i);
+ int16x8_t v350 = vld1q_s16(in + in_stride * 222 + i);
+ int16x8_t v351 = vaddq_s16(v349, v350);
+ int16x8_t v352 = vaddq_s16(v351, v341);
+ int16x8_t v353 = vaddq_s16(v352, v347);
+ int16x8_t v354 = vaddq_s16(v348, v353);
+ int16x8_t v355 = vqrdmulhq_n_s16(v354, 17734);
+ int16x8_t v356 = vaddq_s16(v346, v355);
+ int16x8_t v357 = vqrdmulhq_n_s16(v356, 16705);
+ int16x8_t v358 = vaddq_s16(v334, v357);
+ int16x8_t v359 = vld1q_s16(in + in_stride * 18 + i);
+ int16x8_t v360 = vld1q_s16(in + in_stride * 14 + i);
+ int16x8_t v361 = vaddq_s16(v359, v360);
+ int16x8_t v362_tmp = vqrdmulhq_n_s16(v361, 13573);
+ int16x8_t v362 = vaddq_s16(v362_tmp, v361);
+ int16x8_t v363 = vld1q_s16(in + in_stride * 146 + i);
+ int16x8_t v364 = vld1q_s16(in + in_stride * 142 + i);
+ int16x8_t v365 = vaddq_s16(v363, v364);
+ int16x8_t v366 = vld1q_s16(in + in_stride * 114 + i);
+ int16x8_t v367 = vld1q_s16(in + in_stride * 110 + i);
+ int16x8_t v368 = vaddq_s16(v366, v367);
+ int16x8_t v369 = vaddq_s16(v365, v368);
+ int16x8_t v370 = vaddq_s16(v362, v369);
+ int16x8_t v371 = vld1q_s16(in + in_stride * 82 + i);
+ int16x8_t v372 = vld1q_s16(in + in_stride * 78 + i);
+ int16x8_t v373 = vaddq_s16(v371, v372);
+ int16x8_t v374 = vld1q_s16(in + in_stride * 50 + i);
+ int16x8_t v375 = vld1q_s16(in + in_stride * 46 + i);
+ int16x8_t v376 = vaddq_s16(v374, v375);
+ int16x8_t v377 = vaddq_s16(v373, v376);
+ int16x8_t v378_tmp = vqrdmulhq_n_s16(v377, 13573);
+ int16x8_t v378 = vaddq_s16(v378_tmp, v377);
+ int16x8_t v379 = vld1q_s16(in + in_stride * 210 + i);
+ int16x8_t v380 = vld1q_s16(in + in_stride * 206 + i);
+ int16x8_t v381 = vaddq_s16(v379, v380);
+ int16x8_t v382 = vld1q_s16(in + in_stride * 178 + i);
+ int16x8_t v383 = vld1q_s16(in + in_stride * 174 + i);
+ int16x8_t v384 = vaddq_s16(v382, v383);
+ int16x8_t v385 = vaddq_s16(v381, v384);
+ int16x8_t v386 = vaddq_s16(v385, v377);
+ int16x8_t v387 = vaddq_s16(v378, v386);
+ int16x8_t v388 = vqrdmulhq_n_s16(v387, 17734);
+ int16x8_t v389 = vaddq_s16(v370, v388);
+ int16x8_t v390 = vaddq_s16(v376, v361);
+ int16x8_t v391_tmp = vqrdmulhq_n_s16(v390, 13573);
+ int16x8_t v391 = vaddq_s16(v391_tmp, v390);
+ int16x8_t v392 = vaddq_s16(v384, v365);
+ int16x8_t v393 = vaddq_s16(v368, v373);
+ int16x8_t v394 = vaddq_s16(v392, v393);
+ int16x8_t v395 = vaddq_s16(v391, v394);
+ int16x8_t v396 = vaddq_s16(v393, v390);
+ int16x8_t v397_tmp = vqrdmulhq_n_s16(v396, 13573);
+ int16x8_t v397 = vaddq_s16(v397_tmp, v396);
+ int16x8_t v398 = vld1q_s16(in + in_stride * 242 + i);
+ int16x8_t v399 = vld1q_s16(in + in_stride * 238 + i);
+ int16x8_t v400 = vaddq_s16(v398, v399);
+ int16x8_t v401 = vaddq_s16(v400, v381);
+ int16x8_t v402 = vaddq_s16(v401, v392);
+ int16x8_t v403 = vaddq_s16(v402, v396);
+ int16x8_t v404 = vaddq_s16(v397, v403);
+ int16x8_t v405 = vqrdmulhq_n_s16(v404, 17734);
+ int16x8_t v406 = vaddq_s16(v395, v405);
+ int16x8_t v407 = vqrdmulhq_n_s16(v406, 16705);
+ int16x8_t v408 = vaddq_s16(v389, v407);
+ int16x8_t v409 = vqrdmulhq_n_s16(v408, 16463);
+ int16x8_t v410 = vaddq_s16(v358, v409);
+ int16x8_t v411 = vld1q_s16(in + in_stride * 10 + i);
+ int16x8_t v412 = vld1q_s16(in + in_stride * 6 + i);
+ int16x8_t v413 = vaddq_s16(v411, v412);
+ int16x8_t v414_tmp = vqrdmulhq_n_s16(v413, 13573);
+ int16x8_t v414 = vaddq_s16(v414_tmp, v413);
+ int16x8_t v415 = vld1q_s16(in + in_stride * 138 + i);
+ int16x8_t v416 = vld1q_s16(in + in_stride * 134 + i);
+ int16x8_t v417 = vaddq_s16(v415, v416);
+ int16x8_t v418 = vld1q_s16(in + in_stride * 122 + i);
+ int16x8_t v419 = vld1q_s16(in + in_stride * 118 + i);
+ int16x8_t v420 = vaddq_s16(v418, v419);
+ int16x8_t v421 = vaddq_s16(v417, v420);
+ int16x8_t v422 = vaddq_s16(v414, v421);
+ int16x8_t v423 = vld1q_s16(in + in_stride * 74 + i);
+ int16x8_t v424 = vld1q_s16(in + in_stride * 70 + i);
+ int16x8_t v425 = vaddq_s16(v423, v424);
+ int16x8_t v426 = vld1q_s16(in + in_stride * 58 + i);
+ int16x8_t v427 = vld1q_s16(in + in_stride * 54 + i);
+ int16x8_t v428 = vaddq_s16(v426, v427);
+ int16x8_t v429 = vaddq_s16(v425, v428);
+ int16x8_t v430_tmp = vqrdmulhq_n_s16(v429, 13573);
+ int16x8_t v430 = vaddq_s16(v430_tmp, v429);
+ int16x8_t v431 = vld1q_s16(in + in_stride * 202 + i);
+ int16x8_t v432 = vld1q_s16(in + in_stride * 198 + i);
+ int16x8_t v433 = vaddq_s16(v431, v432);
+ int16x8_t v434 = vld1q_s16(in + in_stride * 186 + i);
+ int16x8_t v435 = vld1q_s16(in + in_stride * 182 + i);
+ int16x8_t v436 = vaddq_s16(v434, v435);
+ int16x8_t v437 = vaddq_s16(v433, v436);
+ int16x8_t v438 = vaddq_s16(v437, v429);
+ int16x8_t v439 = vaddq_s16(v430, v438);
+ int16x8_t v440 = vqrdmulhq_n_s16(v439, 17734);
+ int16x8_t v441 = vaddq_s16(v422, v440);
+ int16x8_t v442 = vld1q_s16(in + in_stride * 42 + i);
+ int16x8_t v443 = vld1q_s16(in + in_stride * 38 + i);
+ int16x8_t v444 = vaddq_s16(v442, v443);
+ int16x8_t v445 = vld1q_s16(in + in_stride * 26 + i);
+ int16x8_t v446 = vld1q_s16(in + in_stride * 22 + i);
+ int16x8_t v447 = vaddq_s16(v445, v446);
+ int16x8_t v448 = vaddq_s16(v444, v447);
+ int16x8_t v449_tmp = vqrdmulhq_n_s16(v448, 13573);
+ int16x8_t v449 = vaddq_s16(v449_tmp, v448);
+ int16x8_t v450 = vld1q_s16(in + in_stride * 170 + i);
+ int16x8_t v451 = vld1q_s16(in + in_stride * 166 + i);
+ int16x8_t v452 = vaddq_s16(v450, v451);
+ int16x8_t v453 = vld1q_s16(in + in_stride * 154 + i);
+ int16x8_t v454 = vld1q_s16(in + in_stride * 150 + i);
+ int16x8_t v455 = vaddq_s16(v453, v454);
+ int16x8_t v456 = vaddq_s16(v452, v455);
+ int16x8_t v457 = vld1q_s16(in + in_stride * 106 + i);
+ int16x8_t v458 = vld1q_s16(in + in_stride * 102 + i);
+ int16x8_t v459 = vaddq_s16(v457, v458);
+ int16x8_t v460 = vld1q_s16(in + in_stride * 90 + i);
+ int16x8_t v461 = vld1q_s16(in + in_stride * 86 + i);
+ int16x8_t v462 = vaddq_s16(v460, v461);
+ int16x8_t v463 = vaddq_s16(v459, v462);
+ int16x8_t v464 = vaddq_s16(v456, v463);
+ int16x8_t v465 = vaddq_s16(v449, v464);
+ int16x8_t v466 = vaddq_s16(v463, v448);
+ int16x8_t v467_tmp = vqrdmulhq_n_s16(v466, 13573);
+ int16x8_t v467 = vaddq_s16(v467_tmp, v466);
+ int16x8_t v468 = vld1q_s16(in + in_stride * 234 + i);
+ int16x8_t v469 = vld1q_s16(in + in_stride * 230 + i);
+ int16x8_t v470 = vaddq_s16(v468, v469);
+ int16x8_t v471 = vld1q_s16(in + in_stride * 218 + i);
+ int16x8_t v472 = vld1q_s16(in + in_stride * 214 + i);
+ int16x8_t v473 = vaddq_s16(v471, v472);
+ int16x8_t v474 = vaddq_s16(v470, v473);
+ int16x8_t v475 = vaddq_s16(v474, v456);
+ int16x8_t v476 = vaddq_s16(v475, v466);
+ int16x8_t v477 = vaddq_s16(v467, v476);
+ int16x8_t v478 = vqrdmulhq_n_s16(v477, 17734);
+ int16x8_t v479 = vaddq_s16(v465, v478);
+ int16x8_t v480 = vqrdmulhq_n_s16(v479, 16705);
+ int16x8_t v481 = vaddq_s16(v441, v480);
+ int16x8_t v482 = vaddq_s16(v447, v413);
+ int16x8_t v483_tmp = vqrdmulhq_n_s16(v482, 13573);
+ int16x8_t v483 = vaddq_s16(v483_tmp, v482);
+ int16x8_t v484 = vaddq_s16(v455, v417);
+ int16x8_t v485 = vaddq_s16(v420, v459);
+ int16x8_t v486 = vaddq_s16(v484, v485);
+ int16x8_t v487 = vaddq_s16(v483, v486);
+ int16x8_t v488 = vaddq_s16(v462, v425);
+ int16x8_t v489 = vaddq_s16(v428, v444);
+ int16x8_t v490 = vaddq_s16(v488, v489);
+ int16x8_t v491_tmp = vqrdmulhq_n_s16(v490, 13573);
+ int16x8_t v491 = vaddq_s16(v491_tmp, v490);
+ int16x8_t v492 = vaddq_s16(v473, v433);
+ int16x8_t v493 = vaddq_s16(v436, v452);
+ int16x8_t v494 = vaddq_s16(v492, v493);
+ int16x8_t v495 = vaddq_s16(v494, v490);
+ int16x8_t v496 = vaddq_s16(v491, v495);
+ int16x8_t v497 = vqrdmulhq_n_s16(v496, 17734);
+ int16x8_t v498 = vaddq_s16(v487, v497);
+ int16x8_t v499 = vaddq_s16(v489, v482);
+ int16x8_t v500_tmp = vqrdmulhq_n_s16(v499, 13573);
+ int16x8_t v500 = vaddq_s16(v500_tmp, v499);
+ int16x8_t v501 = vaddq_s16(v493, v484);
+ int16x8_t v502 = vaddq_s16(v485, v488);
+ int16x8_t v503 = vaddq_s16(v501, v502);
+ int16x8_t v504 = vaddq_s16(v500, v503);
+ int16x8_t v505 = vaddq_s16(v502, v499);
+ int16x8_t v506_tmp = vqrdmulhq_n_s16(v505, 13573);
+ int16x8_t v506 = vaddq_s16(v506_tmp, v505);
+ int16x8_t v507 = vld1q_s16(in + in_stride * 250 + i);
+ int16x8_t v508 = vld1q_s16(in + in_stride * 246 + i);
+ int16x8_t v509 = vaddq_s16(v507, v508);
+ int16x8_t v510 = vaddq_s16(v509, v470);
+ int16x8_t v511 = vaddq_s16(v510, v492);
+ int16x8_t v512 = vaddq_s16(v511, v501);
+ int16x8_t v513 = vaddq_s16(v512, v505);
+ int16x8_t v514 = vaddq_s16(v506, v513);
+ int16x8_t v515 = vqrdmulhq_n_s16(v514, 17734);
+ int16x8_t v516 = vaddq_s16(v504, v515);
+ int16x8_t v517 = vqrdmulhq_n_s16(v516, 16705);
+ int16x8_t v518 = vaddq_s16(v498, v517);
+ int16x8_t v519 = vqrdmulhq_n_s16(v518, 16463);
+ int16x8_t v520 = vaddq_s16(v481, v519);
+ int16x8_t v521 = vqrdmulhq_n_s16(v520, 16404);
+ int16x8_t v522 = vaddq_s16(v410, v521);
+ int16x8_t v523 = vaddq_s16(v412, v318);
+ int16x8_t v524_tmp = vqrdmulhq_n_s16(v523, 13573);
+ int16x8_t v524 = vaddq_s16(v524_tmp, v523);
+ int16x8_t v525 = vaddq_s16(v416, v320);
+ int16x8_t v526 = vaddq_s16(v321, v418);
+ int16x8_t v527 = vaddq_s16(v525, v526);
+ int16x8_t v528 = vaddq_s16(v524, v527);
+ int16x8_t v529 = vaddq_s16(v424, v324);
+ int16x8_t v530 = vaddq_s16(v325, v426);
+ int16x8_t v531 = vaddq_s16(v529, v530);
+ int16x8_t v532_tmp = vqrdmulhq_n_s16(v531, 13573);
+ int16x8_t v532 = vaddq_s16(v532_tmp, v531);
+ int16x8_t v533 = vaddq_s16(v432, v328);
+ int16x8_t v534 = vaddq_s16(v329, v434);
+ int16x8_t v535 = vaddq_s16(v533, v534);
+ int16x8_t v536 = vaddq_s16(v535, v531);
+ int16x8_t v537 = vaddq_s16(v532, v536);
+ int16x8_t v538 = vqrdmulhq_n_s16(v537, 17734);
+ int16x8_t v539 = vaddq_s16(v528, v538);
+ int16x8_t v540 = vaddq_s16(v443, v335);
+ int16x8_t v541 = vaddq_s16(v336, v445);
+ int16x8_t v542 = vaddq_s16(v540, v541);
+ int16x8_t v543_tmp = vqrdmulhq_n_s16(v542, 13573);
+ int16x8_t v543 = vaddq_s16(v543_tmp, v542);
+ int16x8_t v544 = vaddq_s16(v451, v339);
+ int16x8_t v545 = vaddq_s16(v340, v453);
+ int16x8_t v546 = vaddq_s16(v544, v545);
+ int16x8_t v547 = vaddq_s16(v458, v342);
+ int16x8_t v548 = vaddq_s16(v343, v460);
+ int16x8_t v549 = vaddq_s16(v547, v548);
+ int16x8_t v550 = vaddq_s16(v546, v549);
+ int16x8_t v551 = vaddq_s16(v543, v550);
+ int16x8_t v552 = vaddq_s16(v549, v542);
+ int16x8_t v553_tmp = vqrdmulhq_n_s16(v552, 13573);
+ int16x8_t v553 = vaddq_s16(v553_tmp, v552);
+ int16x8_t v554 = vaddq_s16(v469, v349);
+ int16x8_t v555 = vaddq_s16(v350, v471);
+ int16x8_t v556 = vaddq_s16(v554, v555);
+ int16x8_t v557 = vaddq_s16(v556, v546);
+ int16x8_t v558 = vaddq_s16(v557, v552);
+ int16x8_t v559 = vaddq_s16(v553, v558);
+ int16x8_t v560 = vqrdmulhq_n_s16(v559, 17734);
+ int16x8_t v561 = vaddq_s16(v551, v560);
+ int16x8_t v562 = vqrdmulhq_n_s16(v561, 16705);
+ int16x8_t v563 = vaddq_s16(v539, v562);
+ int16x8_t v564 = vaddq_s16(v446, v359);
+ int16x8_t v565 = vaddq_s16(v360, v411);
+ int16x8_t v566 = vaddq_s16(v564, v565);
+ int16x8_t v567_tmp = vqrdmulhq_n_s16(v566, 13573);
+ int16x8_t v567 = vaddq_s16(v567_tmp, v566);
+ int16x8_t v568 = vaddq_s16(v454, v363);
+ int16x8_t v569 = vaddq_s16(v364, v415);
+ int16x8_t v570 = vaddq_s16(v568, v569);
+ int16x8_t v571 = vaddq_s16(v419, v366);
+ int16x8_t v572 = vaddq_s16(v367, v457);
+ int16x8_t v573 = vaddq_s16(v571, v572);
+ int16x8_t v574 = vaddq_s16(v570, v573);
+ int16x8_t v575 = vaddq_s16(v567, v574);
+ int16x8_t v576 = vaddq_s16(v461, v371);
+ int16x8_t v577 = vaddq_s16(v372, v423);
+ int16x8_t v578 = vaddq_s16(v576, v577);
+ int16x8_t v579 = vaddq_s16(v427, v374);
+ int16x8_t v580 = vaddq_s16(v375, v442);
+ int16x8_t v581 = vaddq_s16(v579, v580);
+ int16x8_t v582 = vaddq_s16(v578, v581);
+ int16x8_t v583_tmp = vqrdmulhq_n_s16(v582, 13573);
+ int16x8_t v583 = vaddq_s16(v583_tmp, v582);
+ int16x8_t v584 = vaddq_s16(v472, v379);
+ int16x8_t v585 = vaddq_s16(v380, v431);
+ int16x8_t v586 = vaddq_s16(v584, v585);
+ int16x8_t v587 = vaddq_s16(v435, v382);
+ int16x8_t v588 = vaddq_s16(v383, v450);
+ int16x8_t v589 = vaddq_s16(v587, v588);
+ int16x8_t v590 = vaddq_s16(v586, v589);
+ int16x8_t v591 = vaddq_s16(v590, v582);
+ int16x8_t v592 = vaddq_s16(v583, v591);
+ int16x8_t v593 = vqrdmulhq_n_s16(v592, 17734);
+ int16x8_t v594 = vaddq_s16(v575, v593);
+ int16x8_t v595 = vaddq_s16(v581, v566);
+ int16x8_t v596_tmp = vqrdmulhq_n_s16(v595, 13573);
+ int16x8_t v596 = vaddq_s16(v596_tmp, v595);
+ int16x8_t v597 = vaddq_s16(v589, v570);
+ int16x8_t v598 = vaddq_s16(v573, v578);
+ int16x8_t v599 = vaddq_s16(v597, v598);
+ int16x8_t v600 = vaddq_s16(v596, v599);
+ int16x8_t v601 = vaddq_s16(v598, v595);
+ int16x8_t v602_tmp = vqrdmulhq_n_s16(v601, 13573);
+ int16x8_t v602 = vaddq_s16(v602_tmp, v601);
+ int16x8_t v603 = vaddq_s16(v508, v398);
+ int16x8_t v604 = vaddq_s16(v399, v468);
+ int16x8_t v605 = vaddq_s16(v603, v604);
+ int16x8_t v606 = vaddq_s16(v605, v586);
+ int16x8_t v607 = vaddq_s16(v606, v597);
+ int16x8_t v608 = vaddq_s16(v607, v601);
+ int16x8_t v609 = vaddq_s16(v602, v608);
+ int16x8_t v610 = vqrdmulhq_n_s16(v609, 17734);
+ int16x8_t v611 = vaddq_s16(v600, v610);
+ int16x8_t v612 = vqrdmulhq_n_s16(v611, 16705);
+ int16x8_t v613 = vaddq_s16(v594, v612);
+ int16x8_t v614 = vqrdmulhq_n_s16(v613, 16463);
+ int16x8_t v615 = vaddq_s16(v563, v614);
+ int16x8_t v616 = vaddq_s16(v565, v523);
+ int16x8_t v617_tmp = vqrdmulhq_n_s16(v616, 13573);
+ int16x8_t v617 = vaddq_s16(v617_tmp, v616);
+ int16x8_t v618 = vaddq_s16(v569, v525);
+ int16x8_t v619 = vaddq_s16(v526, v571);
+ int16x8_t v620 = vaddq_s16(v618, v619);
+ int16x8_t v621 = vaddq_s16(v617, v620);
+ int16x8_t v622 = vaddq_s16(v577, v529);
+ int16x8_t v623 = vaddq_s16(v530, v579);
+ int16x8_t v624 = vaddq_s16(v622, v623);
+ int16x8_t v625_tmp = vqrdmulhq_n_s16(v624, 13573);
+ int16x8_t v625 = vaddq_s16(v625_tmp, v624);
+ int16x8_t v626 = vaddq_s16(v585, v533);
+ int16x8_t v627 = vaddq_s16(v534, v587);
+ int16x8_t v628 = vaddq_s16(v626, v627);
+ int16x8_t v629 = vaddq_s16(v628, v624);
+ int16x8_t v630 = vaddq_s16(v625, v629);
+ int16x8_t v631 = vqrdmulhq_n_s16(v630, 17734);
+ int16x8_t v632 = vaddq_s16(v621, v631);
+ int16x8_t v633 = vaddq_s16(v580, v540);
+ int16x8_t v634 = vaddq_s16(v541, v564);
+ int16x8_t v635 = vaddq_s16(v633, v634);
+ int16x8_t v636_tmp = vqrdmulhq_n_s16(v635, 13573);
+ int16x8_t v636 = vaddq_s16(v636_tmp, v635);
+ int16x8_t v637 = vaddq_s16(v588, v544);
+ int16x8_t v638 = vaddq_s16(v545, v568);
+ int16x8_t v639 = vaddq_s16(v637, v638);
+ int16x8_t v640 = vaddq_s16(v572, v547);
+ int16x8_t v641 = vaddq_s16(v548, v576);
+ int16x8_t v642 = vaddq_s16(v640, v641);
+ int16x8_t v643 = vaddq_s16(v639, v642);
+ int16x8_t v644 = vaddq_s16(v636, v643);
+ int16x8_t v645 = vaddq_s16(v642, v635);
+ int16x8_t v646_tmp = vqrdmulhq_n_s16(v645, 13573);
+ int16x8_t v646 = vaddq_s16(v646_tmp, v645);
+ int16x8_t v647 = vaddq_s16(v604, v554);
+ int16x8_t v648 = vaddq_s16(v555, v584);
+ int16x8_t v649 = vaddq_s16(v647, v648);
+ int16x8_t v650 = vaddq_s16(v649, v639);
+ int16x8_t v651 = vaddq_s16(v650, v645);
+ int16x8_t v652 = vaddq_s16(v646, v651);
+ int16x8_t v653 = vqrdmulhq_n_s16(v652, 17734);
+ int16x8_t v654 = vaddq_s16(v644, v653);
+ int16x8_t v655 = vqrdmulhq_n_s16(v654, 16705);
+ int16x8_t v656 = vaddq_s16(v632, v655);
+ int16x8_t v657 = vaddq_s16(v634, v616);
+ int16x8_t v658_tmp = vqrdmulhq_n_s16(v657, 13573);
+ int16x8_t v658 = vaddq_s16(v658_tmp, v657);
+ int16x8_t v659 = vaddq_s16(v638, v618);
+ int16x8_t v660 = vaddq_s16(v619, v640);
+ int16x8_t v661 = vaddq_s16(v659, v660);
+ int16x8_t v662 = vaddq_s16(v658, v661);
+ int16x8_t v663 = vaddq_s16(v641, v622);
+ int16x8_t v664 = vaddq_s16(v623, v633);
+ int16x8_t v665 = vaddq_s16(v663, v664);
+ int16x8_t v666_tmp = vqrdmulhq_n_s16(v665, 13573);
+ int16x8_t v666 = vaddq_s16(v666_tmp, v665);
+ int16x8_t v667 = vaddq_s16(v648, v626);
+ int16x8_t v668 = vaddq_s16(v627, v637);
+ int16x8_t v669 = vaddq_s16(v667, v668);
+ int16x8_t v670 = vaddq_s16(v669, v665);
+ int16x8_t v671 = vaddq_s16(v666, v670);
+ int16x8_t v672 = vqrdmulhq_n_s16(v671, 17734);
+ int16x8_t v673 = vaddq_s16(v662, v672);
+ int16x8_t v674 = vaddq_s16(v664, v657);
+ int16x8_t v675_tmp = vqrdmulhq_n_s16(v674, 13573);
+ int16x8_t v675 = vaddq_s16(v675_tmp, v674);
+ int16x8_t v676 = vaddq_s16(v668, v659);
+ int16x8_t v677 = vaddq_s16(v660, v663);
+ int16x8_t v678 = vaddq_s16(v676, v677);
+ int16x8_t v679 = vaddq_s16(v675, v678);
+ int16x8_t v680 = vaddq_s16(v677, v674);
+ int16x8_t v681_tmp = vqrdmulhq_n_s16(v680, 13573);
+ int16x8_t v681 = vaddq_s16(v681_tmp, v680);
+ int16x8_t v682 = vld1q_s16(in + in_stride * 254 + i);
+ int16x8_t v683 = vaddq_s16(v682, v507);
+ int16x8_t v684 = vaddq_s16(v683, v603);
+ int16x8_t v685 = vaddq_s16(v684, v647);
+ int16x8_t v686 = vaddq_s16(v685, v667);
+ int16x8_t v687 = vaddq_s16(v686, v676);
+ int16x8_t v688 = vaddq_s16(v687, v680);
+ int16x8_t v689 = vaddq_s16(v681, v688);
+ int16x8_t v690 = vqrdmulhq_n_s16(v689, 17734);
+ int16x8_t v691 = vaddq_s16(v679, v690);
+ int16x8_t v692 = vqrdmulhq_n_s16(v691, 16705);
+ int16x8_t v693 = vaddq_s16(v673, v692);
+ int16x8_t v694 = vqrdmulhq_n_s16(v693, 16463);
+ int16x8_t v695 = vaddq_s16(v656, v694);
+ int16x8_t v696 = vqrdmulhq_n_s16(v695, 16404);
+ int16x8_t v697 = vaddq_s16(v615, v696);
+ int16x8_t v698 = vqrdmulhq_n_s16(v697, 16389);
+ int16x8_t v699 = vaddq_s16(v522, v698);
+ int16x8_t v700 = vqrdmulhq_n_s16(v699, 16385);
+ int16x8_t v701 = vaddq_s16(v317, v700);
+ int16x8_t v702 = vld1q_s16(in + in_stride * 1 + i);
+ int16x8_t v703_tmp = vqrdmulhq_n_s16(v702, 13573);
+ int16x8_t v703 = vaddq_s16(v703_tmp, v702);
+ int16x8_t v704 = vld1q_s16(in + in_stride * 129 + i);
+ int16x8_t v705 = vld1q_s16(in + in_stride * 127 + i);
+ int16x8_t v706 = vaddq_s16(v704, v705);
+ int16x8_t v707 = vaddq_s16(v703, v706);
+ int16x8_t v708 = vld1q_s16(in + in_stride * 65 + i);
+ int16x8_t v709 = vld1q_s16(in + in_stride * 63 + i);
+ int16x8_t v710 = vaddq_s16(v708, v709);
+ int16x8_t v711_tmp = vqrdmulhq_n_s16(v710, 13573);
+ int16x8_t v711 = vaddq_s16(v711_tmp, v710);
+ int16x8_t v712 = vld1q_s16(in + in_stride * 193 + i);
+ int16x8_t v713 = vld1q_s16(in + in_stride * 191 + i);
+ int16x8_t v714 = vaddq_s16(v712, v713);
+ int16x8_t v715 = vaddq_s16(v714, v710);
+ int16x8_t v716 = vaddq_s16(v711, v715);
+ int16x8_t v717 = vqrdmulhq_n_s16(v716, 17734);
+ int16x8_t v718 = vaddq_s16(v707, v717);
+ int16x8_t v719 = vld1q_s16(in + in_stride * 33 + i);
+ int16x8_t v720 = vld1q_s16(in + in_stride * 31 + i);
+ int16x8_t v721 = vaddq_s16(v719, v720);
+ int16x8_t v722_tmp = vqrdmulhq_n_s16(v721, 13573);
+ int16x8_t v722 = vaddq_s16(v722_tmp, v721);
+ int16x8_t v723 = vld1q_s16(in + in_stride * 161 + i);
+ int16x8_t v724 = vld1q_s16(in + in_stride * 159 + i);
+ int16x8_t v725 = vaddq_s16(v723, v724);
+ int16x8_t v726 = vld1q_s16(in + in_stride * 97 + i);
+ int16x8_t v727 = vld1q_s16(in + in_stride * 95 + i);
+ int16x8_t v728 = vaddq_s16(v726, v727);
+ int16x8_t v729 = vaddq_s16(v725, v728);
+ int16x8_t v730 = vaddq_s16(v722, v729);
+ int16x8_t v731 = vaddq_s16(v728, v721);
+ int16x8_t v732_tmp = vqrdmulhq_n_s16(v731, 13573);
+ int16x8_t v732 = vaddq_s16(v732_tmp, v731);
+ int16x8_t v733 = vld1q_s16(in + in_stride * 225 + i);
+ int16x8_t v734 = vld1q_s16(in + in_stride * 223 + i);
+ int16x8_t v735 = vaddq_s16(v733, v734);
+ int16x8_t v736 = vaddq_s16(v735, v725);
+ int16x8_t v737 = vaddq_s16(v736, v731);
+ int16x8_t v738 = vaddq_s16(v732, v737);
+ int16x8_t v739 = vqrdmulhq_n_s16(v738, 17734);
+ int16x8_t v740 = vaddq_s16(v730, v739);
+ int16x8_t v741 = vqrdmulhq_n_s16(v740, 16705);
+ int16x8_t v742 = vaddq_s16(v718, v741);
+ int16x8_t v743 = vld1q_s16(in + in_stride * 17 + i);
+ int16x8_t v744 = vld1q_s16(in + in_stride * 15 + i);
+ int16x8_t v745 = vaddq_s16(v743, v744);
+ int16x8_t v746_tmp = vqrdmulhq_n_s16(v745, 13573);
+ int16x8_t v746 = vaddq_s16(v746_tmp, v745);
+ int16x8_t v747 = vld1q_s16(in + in_stride * 145 + i);
+ int16x8_t v748 = vld1q_s16(in + in_stride * 143 + i);
+ int16x8_t v749 = vaddq_s16(v747, v748);
+ int16x8_t v750 = vld1q_s16(in + in_stride * 113 + i);
+ int16x8_t v751 = vld1q_s16(in + in_stride * 111 + i);
+ int16x8_t v752 = vaddq_s16(v750, v751);
+ int16x8_t v753 = vaddq_s16(v749, v752);
+ int16x8_t v754 = vaddq_s16(v746, v753);
+ int16x8_t v755 = vld1q_s16(in + in_stride * 81 + i);
+ int16x8_t v756 = vld1q_s16(in + in_stride * 79 + i);
+ int16x8_t v757 = vaddq_s16(v755, v756);
+ int16x8_t v758 = vld1q_s16(in + in_stride * 49 + i);
+ int16x8_t v759 = vld1q_s16(in + in_stride * 47 + i);
+ int16x8_t v760 = vaddq_s16(v758, v759);
+ int16x8_t v761 = vaddq_s16(v757, v760);
+ int16x8_t v762_tmp = vqrdmulhq_n_s16(v761, 13573);
+ int16x8_t v762 = vaddq_s16(v762_tmp, v761);
+ int16x8_t v763 = vld1q_s16(in + in_stride * 209 + i);
+ int16x8_t v764 = vld1q_s16(in + in_stride * 207 + i);
+ int16x8_t v765 = vaddq_s16(v763, v764);
+ int16x8_t v766 = vld1q_s16(in + in_stride * 177 + i);
+ int16x8_t v767 = vld1q_s16(in + in_stride * 175 + i);
+ int16x8_t v768 = vaddq_s16(v766, v767);
+ int16x8_t v769 = vaddq_s16(v765, v768);
+ int16x8_t v770 = vaddq_s16(v769, v761);
+ int16x8_t v771 = vaddq_s16(v762, v770);
+ int16x8_t v772 = vqrdmulhq_n_s16(v771, 17734);
+ int16x8_t v773 = vaddq_s16(v754, v772);
+ int16x8_t v774 = vaddq_s16(v760, v745);
+ int16x8_t v775_tmp = vqrdmulhq_n_s16(v774, 13573);
+ int16x8_t v775 = vaddq_s16(v775_tmp, v774);
+ int16x8_t v776 = vaddq_s16(v768, v749);
+ int16x8_t v777 = vaddq_s16(v752, v757);
+ int16x8_t v778 = vaddq_s16(v776, v777);
+ int16x8_t v779 = vaddq_s16(v775, v778);
+ int16x8_t v780 = vaddq_s16(v777, v774);
+ int16x8_t v781_tmp = vqrdmulhq_n_s16(v780, 13573);
+ int16x8_t v781 = vaddq_s16(v781_tmp, v780);
+ int16x8_t v782 = vld1q_s16(in + in_stride * 241 + i);
+ int16x8_t v783 = vld1q_s16(in + in_stride * 239 + i);
+ int16x8_t v784 = vaddq_s16(v782, v783);
+ int16x8_t v785 = vaddq_s16(v784, v765);
+ int16x8_t v786 = vaddq_s16(v785, v776);
+ int16x8_t v787 = vaddq_s16(v786, v780);
+ int16x8_t v788 = vaddq_s16(v781, v787);
+ int16x8_t v789 = vqrdmulhq_n_s16(v788, 17734);
+ int16x8_t v790 = vaddq_s16(v779, v789);
+ int16x8_t v791 = vqrdmulhq_n_s16(v790, 16705);
+ int16x8_t v792 = vaddq_s16(v773, v791);
+ int16x8_t v793 = vqrdmulhq_n_s16(v792, 16463);
+ int16x8_t v794 = vaddq_s16(v742, v793);
+ int16x8_t v795 = vld1q_s16(in + in_stride * 9 + i);
+ int16x8_t v796 = vld1q_s16(in + in_stride * 7 + i);
+ int16x8_t v797 = vaddq_s16(v795, v796);
+ int16x8_t v798_tmp = vqrdmulhq_n_s16(v797, 13573);
+ int16x8_t v798 = vaddq_s16(v798_tmp, v797);
+ int16x8_t v799 = vld1q_s16(in + in_stride * 137 + i);
+ int16x8_t v800 = vld1q_s16(in + in_stride * 135 + i);
+ int16x8_t v801 = vaddq_s16(v799, v800);
+ int16x8_t v802 = vld1q_s16(in + in_stride * 121 + i);
+ int16x8_t v803 = vld1q_s16(in + in_stride * 119 + i);
+ int16x8_t v804 = vaddq_s16(v802, v803);
+ int16x8_t v805 = vaddq_s16(v801, v804);
+ int16x8_t v806 = vaddq_s16(v798, v805);
+ int16x8_t v807 = vld1q_s16(in + in_stride * 73 + i);
+ int16x8_t v808 = vld1q_s16(in + in_stride * 71 + i);
+ int16x8_t v809 = vaddq_s16(v807, v808);
+ int16x8_t v810 = vld1q_s16(in + in_stride * 57 + i);
+ int16x8_t v811 = vld1q_s16(in + in_stride * 55 + i);
+ int16x8_t v812 = vaddq_s16(v810, v811);
+ int16x8_t v813 = vaddq_s16(v809, v812);
+ int16x8_t v814_tmp = vqrdmulhq_n_s16(v813, 13573);
+ int16x8_t v814 = vaddq_s16(v814_tmp, v813);
+ int16x8_t v815 = vld1q_s16(in + in_stride * 201 + i);
+ int16x8_t v816 = vld1q_s16(in + in_stride * 199 + i);
+ int16x8_t v817 = vaddq_s16(v815, v816);
+ int16x8_t v818 = vld1q_s16(in + in_stride * 185 + i);
+ int16x8_t v819 = vld1q_s16(in + in_stride * 183 + i);
+ int16x8_t v820 = vaddq_s16(v818, v819);
+ int16x8_t v821 = vaddq_s16(v817, v820);
+ int16x8_t v822 = vaddq_s16(v821, v813);
+ int16x8_t v823 = vaddq_s16(v814, v822);
+ int16x8_t v824 = vqrdmulhq_n_s16(v823, 17734);
+ int16x8_t v825 = vaddq_s16(v806, v824);
+ int16x8_t v826 = vld1q_s16(in + in_stride * 41 + i);
+ int16x8_t v827 = vld1q_s16(in + in_stride * 39 + i);
+ int16x8_t v828 = vaddq_s16(v826, v827);
+ int16x8_t v829 = vld1q_s16(in + in_stride * 25 + i);
+ int16x8_t v830 = vld1q_s16(in + in_stride * 23 + i);
+ int16x8_t v831 = vaddq_s16(v829, v830);
+ int16x8_t v832 = vaddq_s16(v828, v831);
+ int16x8_t v833_tmp = vqrdmulhq_n_s16(v832, 13573);
+ int16x8_t v833 = vaddq_s16(v833_tmp, v832);
+ int16x8_t v834 = vld1q_s16(in + in_stride * 169 + i);
+ int16x8_t v835 = vld1q_s16(in + in_stride * 167 + i);
+ int16x8_t v836 = vaddq_s16(v834, v835);
+ int16x8_t v837 = vld1q_s16(in + in_stride * 153 + i);
+ int16x8_t v838 = vld1q_s16(in + in_stride * 151 + i);
+ int16x8_t v839 = vaddq_s16(v837, v838);
+ int16x8_t v840 = vaddq_s16(v836, v839);
+ int16x8_t v841 = vld1q_s16(in + in_stride * 105 + i);
+ int16x8_t v842 = vld1q_s16(in + in_stride * 103 + i);
+ int16x8_t v843 = vaddq_s16(v841, v842);
+ int16x8_t v844 = vld1q_s16(in + in_stride * 89 + i);
+ int16x8_t v845 = vld1q_s16(in + in_stride * 87 + i);
+ int16x8_t v846 = vaddq_s16(v844, v845);
+ int16x8_t v847 = vaddq_s16(v843, v846);
+ int16x8_t v848 = vaddq_s16(v840, v847);
+ int16x8_t v849 = vaddq_s16(v833, v848);
+ int16x8_t v850 = vaddq_s16(v847, v832);
+ int16x8_t v851_tmp = vqrdmulhq_n_s16(v850, 13573);
+ int16x8_t v851 = vaddq_s16(v851_tmp, v850);
+ int16x8_t v852 = vld1q_s16(in + in_stride * 233 + i);
+ int16x8_t v853 = vld1q_s16(in + in_stride * 231 + i);
+ int16x8_t v854 = vaddq_s16(v852, v853);
+ int16x8_t v855 = vld1q_s16(in + in_stride * 217 + i);
+ int16x8_t v856 = vld1q_s16(in + in_stride * 215 + i);
+ int16x8_t v857 = vaddq_s16(v855, v856);
+ int16x8_t v858 = vaddq_s16(v854, v857);
+ int16x8_t v859 = vaddq_s16(v858, v840);
+ int16x8_t v860 = vaddq_s16(v859, v850);
+ int16x8_t v861 = vaddq_s16(v851, v860);
+ int16x8_t v862 = vqrdmulhq_n_s16(v861, 17734);
+ int16x8_t v863 = vaddq_s16(v849, v862);
+ int16x8_t v864 = vqrdmulhq_n_s16(v863, 16705);
+ int16x8_t v865 = vaddq_s16(v825, v864);
+ int16x8_t v866 = vaddq_s16(v831, v797);
+ int16x8_t v867_tmp = vqrdmulhq_n_s16(v866, 13573);
+ int16x8_t v867 = vaddq_s16(v867_tmp, v866);
+ int16x8_t v868 = vaddq_s16(v839, v801);
+ int16x8_t v869 = vaddq_s16(v804, v843);
+ int16x8_t v870 = vaddq_s16(v868, v869);
+ int16x8_t v871 = vaddq_s16(v867, v870);
+ int16x8_t v872 = vaddq_s16(v846, v809);
+ int16x8_t v873 = vaddq_s16(v812, v828);
+ int16x8_t v874 = vaddq_s16(v872, v873);
+ int16x8_t v875_tmp = vqrdmulhq_n_s16(v874, 13573);
+ int16x8_t v875 = vaddq_s16(v875_tmp, v874);
+ int16x8_t v876 = vaddq_s16(v857, v817);
+ int16x8_t v877 = vaddq_s16(v820, v836);
+ int16x8_t v878 = vaddq_s16(v876, v877);
+ int16x8_t v879 = vaddq_s16(v878, v874);
+ int16x8_t v880 = vaddq_s16(v875, v879);
+ int16x8_t v881 = vqrdmulhq_n_s16(v880, 17734);
+ int16x8_t v882 = vaddq_s16(v871, v881);
+ int16x8_t v883 = vaddq_s16(v873, v866);
+ int16x8_t v884_tmp = vqrdmulhq_n_s16(v883, 13573);
+ int16x8_t v884 = vaddq_s16(v884_tmp, v883);
+ int16x8_t v885 = vaddq_s16(v877, v868);
+ int16x8_t v886 = vaddq_s16(v869, v872);
+ int16x8_t v887 = vaddq_s16(v885, v886);
+ int16x8_t v888 = vaddq_s16(v884, v887);
+ int16x8_t v889 = vaddq_s16(v886, v883);
+ int16x8_t v890_tmp = vqrdmulhq_n_s16(v889, 13573);
+ int16x8_t v890 = vaddq_s16(v890_tmp, v889);
+ int16x8_t v891 = vld1q_s16(in + in_stride * 249 + i);
+ int16x8_t v892 = vld1q_s16(in + in_stride * 247 + i);
+ int16x8_t v893 = vaddq_s16(v891, v892);
+ int16x8_t v894 = vaddq_s16(v893, v854);
+ int16x8_t v895 = vaddq_s16(v894, v876);
+ int16x8_t v896 = vaddq_s16(v895, v885);
+ int16x8_t v897 = vaddq_s16(v896, v889);
+ int16x8_t v898 = vaddq_s16(v890, v897);
+ int16x8_t v899 = vqrdmulhq_n_s16(v898, 17734);
+ int16x8_t v900 = vaddq_s16(v888, v899);
+ int16x8_t v901 = vqrdmulhq_n_s16(v900, 16705);
+ int16x8_t v902 = vaddq_s16(v882, v901);
+ int16x8_t v903 = vqrdmulhq_n_s16(v902, 16463);
+ int16x8_t v904 = vaddq_s16(v865, v903);
+ int16x8_t v905 = vqrdmulhq_n_s16(v904, 16404);
+ int16x8_t v906 = vaddq_s16(v794, v905);
+ int16x8_t v907 = vld1q_s16(in + in_stride * 5 + i);
+ int16x8_t v908 = vld1q_s16(in + in_stride * 3 + i);
+ int16x8_t v909 = vaddq_s16(v907, v908);
+ int16x8_t v910_tmp = vqrdmulhq_n_s16(v909, 13573);
+ int16x8_t v910 = vaddq_s16(v910_tmp, v909);
+ int16x8_t v911 = vld1q_s16(in + in_stride * 133 + i);
+ int16x8_t v912 = vld1q_s16(in + in_stride * 131 + i);
+ int16x8_t v913 = vaddq_s16(v911, v912);
+ int16x8_t v914 = vld1q_s16(in + in_stride * 125 + i);
+ int16x8_t v915 = vld1q_s16(in + in_stride * 123 + i);
+ int16x8_t v916 = vaddq_s16(v914, v915);
+ int16x8_t v917 = vaddq_s16(v913, v916);
+ int16x8_t v918 = vaddq_s16(v910, v917);
+ int16x8_t v919 = vld1q_s16(in + in_stride * 69 + i);
+ int16x8_t v920 = vld1q_s16(in + in_stride * 67 + i);
+ int16x8_t v921 = vaddq_s16(v919, v920);
+ int16x8_t v922 = vld1q_s16(in + in_stride * 61 + i);
+ int16x8_t v923 = vld1q_s16(in + in_stride * 59 + i);
+ int16x8_t v924 = vaddq_s16(v922, v923);
+ int16x8_t v925 = vaddq_s16(v921, v924);
+ int16x8_t v926_tmp = vqrdmulhq_n_s16(v925, 13573);
+ int16x8_t v926 = vaddq_s16(v926_tmp, v925);
+ int16x8_t v927 = vld1q_s16(in + in_stride * 197 + i);
+ int16x8_t v928 = vld1q_s16(in + in_stride * 195 + i);
+ int16x8_t v929 = vaddq_s16(v927, v928);
+ int16x8_t v930 = vld1q_s16(in + in_stride * 189 + i);
+ int16x8_t v931 = vld1q_s16(in + in_stride * 187 + i);
+ int16x8_t v932 = vaddq_s16(v930, v931);
+ int16x8_t v933 = vaddq_s16(v929, v932);
+ int16x8_t v934 = vaddq_s16(v933, v925);
+ int16x8_t v935 = vaddq_s16(v926, v934);
+ int16x8_t v936 = vqrdmulhq_n_s16(v935, 17734);
+ int16x8_t v937 = vaddq_s16(v918, v936);
+ int16x8_t v938 = vld1q_s16(in + in_stride * 37 + i);
+ int16x8_t v939 = vld1q_s16(in + in_stride * 35 + i);
+ int16x8_t v940 = vaddq_s16(v938, v939);
+ int16x8_t v941 = vld1q_s16(in + in_stride * 29 + i);
+ int16x8_t v942 = vld1q_s16(in + in_stride * 27 + i);
+ int16x8_t v943 = vaddq_s16(v941, v942);
+ int16x8_t v944 = vaddq_s16(v940, v943);
+ int16x8_t v945_tmp = vqrdmulhq_n_s16(v944, 13573);
+ int16x8_t v945 = vaddq_s16(v945_tmp, v944);
+ int16x8_t v946 = vld1q_s16(in + in_stride * 165 + i);
+ int16x8_t v947 = vld1q_s16(in + in_stride * 163 + i);
+ int16x8_t v948 = vaddq_s16(v946, v947);
+ int16x8_t v949 = vld1q_s16(in + in_stride * 157 + i);
+ int16x8_t v950 = vld1q_s16(in + in_stride * 155 + i);
+ int16x8_t v951 = vaddq_s16(v949, v950);
+ int16x8_t v952 = vaddq_s16(v948, v951);
+ int16x8_t v953 = vld1q_s16(in + in_stride * 101 + i);
+ int16x8_t v954 = vld1q_s16(in + in_stride * 99 + i);
+ int16x8_t v955 = vaddq_s16(v953, v954);
+ int16x8_t v956 = vld1q_s16(in + in_stride * 93 + i);
+ int16x8_t v957 = vld1q_s16(in + in_stride * 91 + i);
+ int16x8_t v958 = vaddq_s16(v956, v957);
+ int16x8_t v959 = vaddq_s16(v955, v958);
+ int16x8_t v960 = vaddq_s16(v952, v959);
+ int16x8_t v961 = vaddq_s16(v945, v960);
+ int16x8_t v962 = vaddq_s16(v959, v944);
+ int16x8_t v963_tmp = vqrdmulhq_n_s16(v962, 13573);
+ int16x8_t v963 = vaddq_s16(v963_tmp, v962);
+ int16x8_t v964 = vld1q_s16(in + in_stride * 229 + i);
+ int16x8_t v965 = vld1q_s16(in + in_stride * 227 + i);
+ int16x8_t v966 = vaddq_s16(v964, v965);
+ int16x8_t v967 = vld1q_s16(in + in_stride * 221 + i);
+ int16x8_t v968 = vld1q_s16(in + in_stride * 219 + i);
+ int16x8_t v969 = vaddq_s16(v967, v968);
+ int16x8_t v970 = vaddq_s16(v966, v969);
+ int16x8_t v971 = vaddq_s16(v970, v952);
+ int16x8_t v972 = vaddq_s16(v971, v962);
+ int16x8_t v973 = vaddq_s16(v963, v972);
+ int16x8_t v974 = vqrdmulhq_n_s16(v973, 17734);
+ int16x8_t v975 = vaddq_s16(v961, v974);
+ int16x8_t v976 = vqrdmulhq_n_s16(v975, 16705);
+ int16x8_t v977 = vaddq_s16(v937, v976);
+ int16x8_t v978 = vld1q_s16(in + in_stride * 21 + i);
+ int16x8_t v979 = vld1q_s16(in + in_stride * 19 + i);
+ int16x8_t v980 = vaddq_s16(v978, v979);
+ int16x8_t v981 = vld1q_s16(in + in_stride * 13 + i);
+ int16x8_t v982 = vld1q_s16(in + in_stride * 11 + i);
+ int16x8_t v983 = vaddq_s16(v981, v982);
+ int16x8_t v984 = vaddq_s16(v980, v983);
+ int16x8_t v985_tmp = vqrdmulhq_n_s16(v984, 13573);
+ int16x8_t v985 = vaddq_s16(v985_tmp, v984);
+ int16x8_t v986 = vld1q_s16(in + in_stride * 149 + i);
+ int16x8_t v987 = vld1q_s16(in + in_stride * 147 + i);
+ int16x8_t v988 = vaddq_s16(v986, v987);
+ int16x8_t v989 = vld1q_s16(in + in_stride * 141 + i);
+ int16x8_t v990 = vld1q_s16(in + in_stride * 139 + i);
+ int16x8_t v991 = vaddq_s16(v989, v990);
+ int16x8_t v992 = vaddq_s16(v988, v991);
+ int16x8_t v993 = vld1q_s16(in + in_stride * 117 + i);
+ int16x8_t v994 = vld1q_s16(in + in_stride * 115 + i);
+ int16x8_t v995 = vaddq_s16(v993, v994);
+ int16x8_t v996 = vld1q_s16(in + in_stride * 109 + i);
+ int16x8_t v997 = vld1q_s16(in + in_stride * 107 + i);
+ int16x8_t v998 = vaddq_s16(v996, v997);
+ int16x8_t v999 = vaddq_s16(v995, v998);
+ int16x8_t v1000 = vaddq_s16(v992, v999);
+ int16x8_t v1001 = vaddq_s16(v985, v1000);
+ int16x8_t v1002 = vld1q_s16(in + in_stride * 85 + i);
+ int16x8_t v1003 = vld1q_s16(in + in_stride * 83 + i);
+ int16x8_t v1004 = vaddq_s16(v1002, v1003);
+ int16x8_t v1005 = vld1q_s16(in + in_stride * 77 + i);
+ int16x8_t v1006 = vld1q_s16(in + in_stride * 75 + i);
+ int16x8_t v1007 = vaddq_s16(v1005, v1006);
+ int16x8_t v1008 = vaddq_s16(v1004, v1007);
+ int16x8_t v1009 = vld1q_s16(in + in_stride * 53 + i);
+ int16x8_t v1010 = vld1q_s16(in + in_stride * 51 + i);
+ int16x8_t v1011 = vaddq_s16(v1009, v1010);
+ int16x8_t v1012 = vld1q_s16(in + in_stride * 45 + i);
+ int16x8_t v1013 = vld1q_s16(in + in_stride * 43 + i);
+ int16x8_t v1014 = vaddq_s16(v1012, v1013);
+ int16x8_t v1015 = vaddq_s16(v1011, v1014);
+ int16x8_t v1016 = vaddq_s16(v1008, v1015);
+ int16x8_t v1017_tmp = vqrdmulhq_n_s16(v1016, 13573);
+ int16x8_t v1017 = vaddq_s16(v1017_tmp, v1016);
+ int16x8_t v1018 = vld1q_s16(in + in_stride * 213 + i);
+ int16x8_t v1019 = vld1q_s16(in + in_stride * 211 + i);
+ int16x8_t v1020 = vaddq_s16(v1018, v1019);
+ int16x8_t v1021 = vld1q_s16(in + in_stride * 205 + i);
+ int16x8_t v1022 = vld1q_s16(in + in_stride * 203 + i);
+ int16x8_t v1023 = vaddq_s16(v1021, v1022);
+ int16x8_t v1024 = vaddq_s16(v1020, v1023);
+ int16x8_t v1025 = vld1q_s16(in + in_stride * 181 + i);
+ int16x8_t v1026 = vld1q_s16(in + in_stride * 179 + i);
+ int16x8_t v1027 = vaddq_s16(v1025, v1026);
+ int16x8_t v1028 = vld1q_s16(in + in_stride * 173 + i);
+ int16x8_t v1029 = vld1q_s16(in + in_stride * 171 + i);
+ int16x8_t v1030 = vaddq_s16(v1028, v1029);
+ int16x8_t v1031 = vaddq_s16(v1027, v1030);
+ int16x8_t v1032 = vaddq_s16(v1024, v1031);
+ int16x8_t v1033 = vaddq_s16(v1032, v1016);
+ int16x8_t v1034 = vaddq_s16(v1017, v1033);
+ int16x8_t v1035 = vqrdmulhq_n_s16(v1034, 17734);
+ int16x8_t v1036 = vaddq_s16(v1001, v1035);
+ int16x8_t v1037 = vaddq_s16(v1015, v984);
+ int16x8_t v1038_tmp = vqrdmulhq_n_s16(v1037, 13573);
+ int16x8_t v1038 = vaddq_s16(v1038_tmp, v1037);
+ int16x8_t v1039 = vaddq_s16(v1031, v992);
+ int16x8_t v1040 = vaddq_s16(v999, v1008);
+ int16x8_t v1041 = vaddq_s16(v1039, v1040);
+ int16x8_t v1042 = vaddq_s16(v1038, v1041);
+ int16x8_t v1043 = vaddq_s16(v1040, v1037);
+ int16x8_t v1044_tmp = vqrdmulhq_n_s16(v1043, 13573);
+ int16x8_t v1044 = vaddq_s16(v1044_tmp, v1043);
+ int16x8_t v1045 = vld1q_s16(in + in_stride * 245 + i);
+ int16x8_t v1046 = vld1q_s16(in + in_stride * 243 + i);
+ int16x8_t v1047 = vaddq_s16(v1045, v1046);
+ int16x8_t v1048 = vld1q_s16(in + in_stride * 237 + i);
+ int16x8_t v1049 = vld1q_s16(in + in_stride * 235 + i);
+ int16x8_t v1050 = vaddq_s16(v1048, v1049);
+ int16x8_t v1051 = vaddq_s16(v1047, v1050);
+ int16x8_t v1052 = vaddq_s16(v1051, v1024);
+ int16x8_t v1053 = vaddq_s16(v1052, v1039);
+ int16x8_t v1054 = vaddq_s16(v1053, v1043);
+ int16x8_t v1055 = vaddq_s16(v1044, v1054);
+ int16x8_t v1056 = vqrdmulhq_n_s16(v1055, 17734);
+ int16x8_t v1057 = vaddq_s16(v1042, v1056);
+ int16x8_t v1058 = vqrdmulhq_n_s16(v1057, 16705);
+ int16x8_t v1059 = vaddq_s16(v1036, v1058);
+ int16x8_t v1060 = vqrdmulhq_n_s16(v1059, 16463);
+ int16x8_t v1061 = vaddq_s16(v977, v1060);
+ int16x8_t v1062 = vaddq_s16(v983, v909);
+ int16x8_t v1063_tmp = vqrdmulhq_n_s16(v1062, 13573);
+ int16x8_t v1063 = vaddq_s16(v1063_tmp, v1062);
+ int16x8_t v1064 = vaddq_s16(v991, v913);
+ int16x8_t v1065 = vaddq_s16(v916, v995);
+ int16x8_t v1066 = vaddq_s16(v1064, v1065);
+ int16x8_t v1067 = vaddq_s16(v1063, v1066);
+ int16x8_t v1068 = vaddq_s16(v1007, v921);
+ int16x8_t v1069 = vaddq_s16(v924, v1011);
+ int16x8_t v1070 = vaddq_s16(v1068, v1069);
+ int16x8_t v1071_tmp = vqrdmulhq_n_s16(v1070, 13573);
+ int16x8_t v1071 = vaddq_s16(v1071_tmp, v1070);
+ int16x8_t v1072 = vaddq_s16(v1023, v929);
+ int16x8_t v1073 = vaddq_s16(v932, v1027);
+ int16x8_t v1074 = vaddq_s16(v1072, v1073);
+ int16x8_t v1075 = vaddq_s16(v1074, v1070);
+ int16x8_t v1076 = vaddq_s16(v1071, v1075);
+ int16x8_t v1077 = vqrdmulhq_n_s16(v1076, 17734);
+ int16x8_t v1078 = vaddq_s16(v1067, v1077);
+ int16x8_t v1079 = vaddq_s16(v1014, v940);
+ int16x8_t v1080 = vaddq_s16(v943, v980);
+ int16x8_t v1081 = vaddq_s16(v1079, v1080);
+ int16x8_t v1082_tmp = vqrdmulhq_n_s16(v1081, 13573);
+ int16x8_t v1082 = vaddq_s16(v1082_tmp, v1081);
+ int16x8_t v1083 = vaddq_s16(v1030, v948);
+ int16x8_t v1084 = vaddq_s16(v951, v988);
+ int16x8_t v1085 = vaddq_s16(v1083, v1084);
+ int16x8_t v1086 = vaddq_s16(v998, v955);
+ int16x8_t v1087 = vaddq_s16(v958, v1004);
+ int16x8_t v1088 = vaddq_s16(v1086, v1087);
+ int16x8_t v1089 = vaddq_s16(v1085, v1088);
+ int16x8_t v1090 = vaddq_s16(v1082, v1089);
+ int16x8_t v1091 = vaddq_s16(v1088, v1081);
+ int16x8_t v1092_tmp = vqrdmulhq_n_s16(v1091, 13573);
+ int16x8_t v1092 = vaddq_s16(v1092_tmp, v1091);
+ int16x8_t v1093 = vaddq_s16(v1050, v966);
+ int16x8_t v1094 = vaddq_s16(v969, v1020);
+ int16x8_t v1095 = vaddq_s16(v1093, v1094);
+ int16x8_t v1096 = vaddq_s16(v1095, v1085);
+ int16x8_t v1097 = vaddq_s16(v1096, v1091);
+ int16x8_t v1098 = vaddq_s16(v1092, v1097);
+ int16x8_t v1099 = vqrdmulhq_n_s16(v1098, 17734);
+ int16x8_t v1100 = vaddq_s16(v1090, v1099);
+ int16x8_t v1101 = vqrdmulhq_n_s16(v1100, 16705);
+ int16x8_t v1102 = vaddq_s16(v1078, v1101);
+ int16x8_t v1103 = vaddq_s16(v1080, v1062);
+ int16x8_t v1104_tmp = vqrdmulhq_n_s16(v1103, 13573);
+ int16x8_t v1104 = vaddq_s16(v1104_tmp, v1103);
+ int16x8_t v1105 = vaddq_s16(v1084, v1064);
+ int16x8_t v1106 = vaddq_s16(v1065, v1086);
+ int16x8_t v1107 = vaddq_s16(v1105, v1106);
+ int16x8_t v1108 = vaddq_s16(v1104, v1107);
+ int16x8_t v1109 = vaddq_s16(v1087, v1068);
+ int16x8_t v1110 = vaddq_s16(v1069, v1079);
+ int16x8_t v1111 = vaddq_s16(v1109, v1110);
+ int16x8_t v1112_tmp = vqrdmulhq_n_s16(v1111, 13573);
+ int16x8_t v1112 = vaddq_s16(v1112_tmp, v1111);
+ int16x8_t v1113 = vaddq_s16(v1094, v1072);
+ int16x8_t v1114 = vaddq_s16(v1073, v1083);
+ int16x8_t v1115 = vaddq_s16(v1113, v1114);
+ int16x8_t v1116 = vaddq_s16(v1115, v1111);
+ int16x8_t v1117 = vaddq_s16(v1112, v1116);
+ int16x8_t v1118 = vqrdmulhq_n_s16(v1117, 17734);
+ int16x8_t v1119 = vaddq_s16(v1108, v1118);
+ int16x8_t v1120 = vaddq_s16(v1110, v1103);
+ int16x8_t v1121_tmp = vqrdmulhq_n_s16(v1120, 13573);
+ int16x8_t v1121 = vaddq_s16(v1121_tmp, v1120);
+ int16x8_t v1122 = vaddq_s16(v1114, v1105);
+ int16x8_t v1123 = vaddq_s16(v1106, v1109);
+ int16x8_t v1124 = vaddq_s16(v1122, v1123);
+ int16x8_t v1125 = vaddq_s16(v1121, v1124);
+ int16x8_t v1126 = vaddq_s16(v1123, v1120);
+ int16x8_t v1127_tmp = vqrdmulhq_n_s16(v1126, 13573);
+ int16x8_t v1127 = vaddq_s16(v1127_tmp, v1126);
+ int16x8_t v1128 = vld1q_s16(in + in_stride * 253 + i);
+ int16x8_t v1129 = vld1q_s16(in + in_stride * 251 + i);
+ int16x8_t v1130 = vaddq_s16(v1128, v1129);
+ int16x8_t v1131 = vaddq_s16(v1130, v1047);
+ int16x8_t v1132 = vaddq_s16(v1131, v1093);
+ int16x8_t v1133 = vaddq_s16(v1132, v1113);
+ int16x8_t v1134 = vaddq_s16(v1133, v1122);
+ int16x8_t v1135 = vaddq_s16(v1134, v1126);
+ int16x8_t v1136 = vaddq_s16(v1127, v1135);
+ int16x8_t v1137 = vqrdmulhq_n_s16(v1136, 17734);
+ int16x8_t v1138 = vaddq_s16(v1125, v1137);
+ int16x8_t v1139 = vqrdmulhq_n_s16(v1138, 16705);
+ int16x8_t v1140 = vaddq_s16(v1119, v1139);
+ int16x8_t v1141 = vqrdmulhq_n_s16(v1140, 16463);
+ int16x8_t v1142 = vaddq_s16(v1102, v1141);
+ int16x8_t v1143 = vqrdmulhq_n_s16(v1142, 16404);
+ int16x8_t v1144 = vaddq_s16(v1061, v1143);
+ int16x8_t v1145 = vqrdmulhq_n_s16(v1144, 16389);
+ int16x8_t v1146 = vaddq_s16(v906, v1145);
+ int16x8_t v1147 = vaddq_s16(v908, v702);
+ int16x8_t v1148_tmp = vqrdmulhq_n_s16(v1147, 13573);
+ int16x8_t v1148 = vaddq_s16(v1148_tmp, v1147);
+ int16x8_t v1149 = vaddq_s16(v912, v704);
+ int16x8_t v1150 = vaddq_s16(v705, v914);
+ int16x8_t v1151 = vaddq_s16(v1149, v1150);
+ int16x8_t v1152 = vaddq_s16(v1148, v1151);
+ int16x8_t v1153 = vaddq_s16(v920, v708);
+ int16x8_t v1154 = vaddq_s16(v709, v922);
+ int16x8_t v1155 = vaddq_s16(v1153, v1154);
+ int16x8_t v1156_tmp = vqrdmulhq_n_s16(v1155, 13573);
+ int16x8_t v1156 = vaddq_s16(v1156_tmp, v1155);
+ int16x8_t v1157 = vaddq_s16(v928, v712);
+ int16x8_t v1158 = vaddq_s16(v713, v930);
+ int16x8_t v1159 = vaddq_s16(v1157, v1158);
+ int16x8_t v1160 = vaddq_s16(v1159, v1155);
+ int16x8_t v1161 = vaddq_s16(v1156, v1160);
+ int16x8_t v1162 = vqrdmulhq_n_s16(v1161, 17734);
+ int16x8_t v1163 = vaddq_s16(v1152, v1162);
+ int16x8_t v1164 = vaddq_s16(v939, v719);
+ int16x8_t v1165 = vaddq_s16(v720, v941);
+ int16x8_t v1166 = vaddq_s16(v1164, v1165);
+ int16x8_t v1167_tmp = vqrdmulhq_n_s16(v1166, 13573);
+ int16x8_t v1167 = vaddq_s16(v1167_tmp, v1166);
+ int16x8_t v1168 = vaddq_s16(v947, v723);
+ int16x8_t v1169 = vaddq_s16(v724, v949);
+ int16x8_t v1170 = vaddq_s16(v1168, v1169);
+ int16x8_t v1171 = vaddq_s16(v954, v726);
+ int16x8_t v1172 = vaddq_s16(v727, v956);
+ int16x8_t v1173 = vaddq_s16(v1171, v1172);
+ int16x8_t v1174 = vaddq_s16(v1170, v1173);
+ int16x8_t v1175 = vaddq_s16(v1167, v1174);
+ int16x8_t v1176 = vaddq_s16(v1173, v1166);
+ int16x8_t v1177_tmp = vqrdmulhq_n_s16(v1176, 13573);
+ int16x8_t v1177 = vaddq_s16(v1177_tmp, v1176);
+ int16x8_t v1178 = vaddq_s16(v965, v733);
+ int16x8_t v1179 = vaddq_s16(v734, v967);
+ int16x8_t v1180 = vaddq_s16(v1178, v1179);
+ int16x8_t v1181 = vaddq_s16(v1180, v1170);
+ int16x8_t v1182 = vaddq_s16(v1181, v1176);
+ int16x8_t v1183 = vaddq_s16(v1177, v1182);
+ int16x8_t v1184 = vqrdmulhq_n_s16(v1183, 17734);
+ int16x8_t v1185 = vaddq_s16(v1175, v1184);
+ int16x8_t v1186 = vqrdmulhq_n_s16(v1185, 16705);
+ int16x8_t v1187 = vaddq_s16(v1163, v1186);
+ int16x8_t v1188 = vaddq_s16(v979, v743);
+ int16x8_t v1189 = vaddq_s16(v744, v981);
+ int16x8_t v1190 = vaddq_s16(v1188, v1189);
+ int16x8_t v1191_tmp = vqrdmulhq_n_s16(v1190, 13573);
+ int16x8_t v1191 = vaddq_s16(v1191_tmp, v1190);
+ int16x8_t v1192 = vaddq_s16(v987, v747);
+ int16x8_t v1193 = vaddq_s16(v748, v989);
+ int16x8_t v1194 = vaddq_s16(v1192, v1193);
+ int16x8_t v1195 = vaddq_s16(v994, v750);
+ int16x8_t v1196 = vaddq_s16(v751, v996);
+ int16x8_t v1197 = vaddq_s16(v1195, v1196);
+ int16x8_t v1198 = vaddq_s16(v1194, v1197);
+ int16x8_t v1199 = vaddq_s16(v1191, v1198);
+ int16x8_t v1200 = vaddq_s16(v1003, v755);
+ int16x8_t v1201 = vaddq_s16(v756, v1005);
+ int16x8_t v1202 = vaddq_s16(v1200, v1201);
+ int16x8_t v1203 = vaddq_s16(v1010, v758);
+ int16x8_t v1204 = vaddq_s16(v759, v1012);
+ int16x8_t v1205 = vaddq_s16(v1203, v1204);
+ int16x8_t v1206 = vaddq_s16(v1202, v1205);
+ int16x8_t v1207_tmp = vqrdmulhq_n_s16(v1206, 13573);
+ int16x8_t v1207 = vaddq_s16(v1207_tmp, v1206);
+ int16x8_t v1208 = vaddq_s16(v1019, v763);
+ int16x8_t v1209 = vaddq_s16(v764, v1021);
+ int16x8_t v1210 = vaddq_s16(v1208, v1209);
+ int16x8_t v1211 = vaddq_s16(v1026, v766);
+ int16x8_t v1212 = vaddq_s16(v767, v1028);
+ int16x8_t v1213 = vaddq_s16(v1211, v1212);
+ int16x8_t v1214 = vaddq_s16(v1210, v1213);
+ int16x8_t v1215 = vaddq_s16(v1214, v1206);
+ int16x8_t v1216 = vaddq_s16(v1207, v1215);
+ int16x8_t v1217 = vqrdmulhq_n_s16(v1216, 17734);
+ int16x8_t v1218 = vaddq_s16(v1199, v1217);
+ int16x8_t v1219 = vaddq_s16(v1205, v1190);
+ int16x8_t v1220_tmp = vqrdmulhq_n_s16(v1219, 13573);
+ int16x8_t v1220 = vaddq_s16(v1220_tmp, v1219);
+ int16x8_t v1221 = vaddq_s16(v1213, v1194);
+ int16x8_t v1222 = vaddq_s16(v1197, v1202);
+ int16x8_t v1223 = vaddq_s16(v1221, v1222);
+ int16x8_t v1224 = vaddq_s16(v1220, v1223);
+ int16x8_t v1225 = vaddq_s16(v1222, v1219);
+ int16x8_t v1226_tmp = vqrdmulhq_n_s16(v1225, 13573);
+ int16x8_t v1226 = vaddq_s16(v1226_tmp, v1225);
+ int16x8_t v1227 = vaddq_s16(v1046, v782);
+ int16x8_t v1228 = vaddq_s16(v783, v1048);
+ int16x8_t v1229 = vaddq_s16(v1227, v1228);
+ int16x8_t v1230 = vaddq_s16(v1229, v1210);
+ int16x8_t v1231 = vaddq_s16(v1230, v1221);
+ int16x8_t v1232 = vaddq_s16(v1231, v1225);
+ int16x8_t v1233 = vaddq_s16(v1226, v1232);
+ int16x8_t v1234 = vqrdmulhq_n_s16(v1233, 17734);
+ int16x8_t v1235 = vaddq_s16(v1224, v1234);
+ int16x8_t v1236 = vqrdmulhq_n_s16(v1235, 16705);
+ int16x8_t v1237 = vaddq_s16(v1218, v1236);
+ int16x8_t v1238 = vqrdmulhq_n_s16(v1237, 16463);
+ int16x8_t v1239 = vaddq_s16(v1187, v1238);
+ int16x8_t v1240 = vaddq_s16(v982, v795);
+ int16x8_t v1241 = vaddq_s16(v796, v907);
+ int16x8_t v1242 = vaddq_s16(v1240, v1241);
+ int16x8_t v1243_tmp = vqrdmulhq_n_s16(v1242, 13573);
+ int16x8_t v1243 = vaddq_s16(v1243_tmp, v1242);
+ int16x8_t v1244 = vaddq_s16(v990, v799);
+ int16x8_t v1245 = vaddq_s16(v800, v911);
+ int16x8_t v1246 = vaddq_s16(v1244, v1245);
+ int16x8_t v1247 = vaddq_s16(v915, v802);
+ int16x8_t v1248 = vaddq_s16(v803, v993);
+ int16x8_t v1249 = vaddq_s16(v1247, v1248);
+ int16x8_t v1250 = vaddq_s16(v1246, v1249);
+ int16x8_t v1251 = vaddq_s16(v1243, v1250);
+ int16x8_t v1252 = vaddq_s16(v1006, v807);
+ int16x8_t v1253 = vaddq_s16(v808, v919);
+ int16x8_t v1254 = vaddq_s16(v1252, v1253);
+ int16x8_t v1255 = vaddq_s16(v923, v810);
+ int16x8_t v1256 = vaddq_s16(v811, v1009);
+ int16x8_t v1257 = vaddq_s16(v1255, v1256);
+ int16x8_t v1258 = vaddq_s16(v1254, v1257);
+ int16x8_t v1259_tmp = vqrdmulhq_n_s16(v1258, 13573);
+ int16x8_t v1259 = vaddq_s16(v1259_tmp, v1258);
+ int16x8_t v1260 = vaddq_s16(v1022, v815);
+ int16x8_t v1261 = vaddq_s16(v816, v927);
+ int16x8_t v1262 = vaddq_s16(v1260, v1261);
+ int16x8_t v1263 = vaddq_s16(v931, v818);
+ int16x8_t v1264 = vaddq_s16(v819, v1025);
+ int16x8_t v1265 = vaddq_s16(v1263, v1264);
+ int16x8_t v1266 = vaddq_s16(v1262, v1265);
+ int16x8_t v1267 = vaddq_s16(v1266, v1258);
+ int16x8_t v1268 = vaddq_s16(v1259, v1267);
+ int16x8_t v1269 = vqrdmulhq_n_s16(v1268, 17734);
+ int16x8_t v1270 = vaddq_s16(v1251, v1269);
+ int16x8_t v1271 = vaddq_s16(v1013, v826);
+ int16x8_t v1272 = vaddq_s16(v827, v938);
+ int16x8_t v1273 = vaddq_s16(v1271, v1272);
+ int16x8_t v1274 = vaddq_s16(v942, v829);
+ int16x8_t v1275 = vaddq_s16(v830, v978);
+ int16x8_t v1276 = vaddq_s16(v1274, v1275);
+ int16x8_t v1277 = vaddq_s16(v1273, v1276);
+ int16x8_t v1278_tmp = vqrdmulhq_n_s16(v1277, 13573);
+ int16x8_t v1278 = vaddq_s16(v1278_tmp, v1277);
+ int16x8_t v1279 = vaddq_s16(v1029, v834);
+ int16x8_t v1280 = vaddq_s16(v835, v946);
+ int16x8_t v1281 = vaddq_s16(v1279, v1280);
+ int16x8_t v1282 = vaddq_s16(v950, v837);
+ int16x8_t v1283 = vaddq_s16(v838, v986);
+ int16x8_t v1284 = vaddq_s16(v1282, v1283);
+ int16x8_t v1285 = vaddq_s16(v1281, v1284);
+ int16x8_t v1286 = vaddq_s16(v997, v841);
+ int16x8_t v1287 = vaddq_s16(v842, v953);
+ int16x8_t v1288 = vaddq_s16(v1286, v1287);
+ int16x8_t v1289 = vaddq_s16(v957, v844);
+ int16x8_t v1290 = vaddq_s16(v845, v1002);
+ int16x8_t v1291 = vaddq_s16(v1289, v1290);
+ int16x8_t v1292 = vaddq_s16(v1288, v1291);
+ int16x8_t v1293 = vaddq_s16(v1285, v1292);
+ int16x8_t v1294 = vaddq_s16(v1278, v1293);
+ int16x8_t v1295 = vaddq_s16(v1292, v1277);
+ int16x8_t v1296_tmp = vqrdmulhq_n_s16(v1295, 13573);
+ int16x8_t v1296 = vaddq_s16(v1296_tmp, v1295);
+ int16x8_t v1297 = vaddq_s16(v1049, v852);
+ int16x8_t v1298 = vaddq_s16(v853, v964);
+ int16x8_t v1299 = vaddq_s16(v1297, v1298);
+ int16x8_t v1300 = vaddq_s16(v968, v855);
+ int16x8_t v1301 = vaddq_s16(v856, v1018);
+ int16x8_t v1302 = vaddq_s16(v1300, v1301);
+ int16x8_t v1303 = vaddq_s16(v1299, v1302);
+ int16x8_t v1304 = vaddq_s16(v1303, v1285);
+ int16x8_t v1305 = vaddq_s16(v1304, v1295);
+ int16x8_t v1306 = vaddq_s16(v1296, v1305);
+ int16x8_t v1307 = vqrdmulhq_n_s16(v1306, 17734);
+ int16x8_t v1308 = vaddq_s16(v1294, v1307);
+ int16x8_t v1309 = vqrdmulhq_n_s16(v1308, 16705);
+ int16x8_t v1310 = vaddq_s16(v1270, v1309);
+ int16x8_t v1311 = vaddq_s16(v1276, v1242);
+ int16x8_t v1312_tmp = vqrdmulhq_n_s16(v1311, 13573);
+ int16x8_t v1312 = vaddq_s16(v1312_tmp, v1311);
+ int16x8_t v1313 = vaddq_s16(v1284, v1246);
+ int16x8_t v1314 = vaddq_s16(v1249, v1288);
+ int16x8_t v1315 = vaddq_s16(v1313, v1314);
+ int16x8_t v1316 = vaddq_s16(v1312, v1315);
+ int16x8_t v1317 = vaddq_s16(v1291, v1254);
+ int16x8_t v1318 = vaddq_s16(v1257, v1273);
+ int16x8_t v1319 = vaddq_s16(v1317, v1318);
+ int16x8_t v1320_tmp = vqrdmulhq_n_s16(v1319, 13573);
+ int16x8_t v1320 = vaddq_s16(v1320_tmp, v1319);
+ int16x8_t v1321 = vaddq_s16(v1302, v1262);
+ int16x8_t v1322 = vaddq_s16(v1265, v1281);
+ int16x8_t v1323 = vaddq_s16(v1321, v1322);
+ int16x8_t v1324 = vaddq_s16(v1323, v1319);
+ int16x8_t v1325 = vaddq_s16(v1320, v1324);
+ int16x8_t v1326 = vqrdmulhq_n_s16(v1325, 17734);
+ int16x8_t v1327 = vaddq_s16(v1316, v1326);
+ int16x8_t v1328 = vaddq_s16(v1318, v1311);
+ int16x8_t v1329_tmp = vqrdmulhq_n_s16(v1328, 13573);
+ int16x8_t v1329 = vaddq_s16(v1329_tmp, v1328);
+ int16x8_t v1330 = vaddq_s16(v1322, v1313);
+ int16x8_t v1331 = vaddq_s16(v1314, v1317);
+ int16x8_t v1332 = vaddq_s16(v1330, v1331);
+ int16x8_t v1333 = vaddq_s16(v1329, v1332);
+ int16x8_t v1334 = vaddq_s16(v1331, v1328);
+ int16x8_t v1335_tmp = vqrdmulhq_n_s16(v1334, 13573);
+ int16x8_t v1335 = vaddq_s16(v1335_tmp, v1334);
+ int16x8_t v1336 = vaddq_s16(v1129, v891);
+ int16x8_t v1337 = vaddq_s16(v892, v1045);
+ int16x8_t v1338 = vaddq_s16(v1336, v1337);
+ int16x8_t v1339 = vaddq_s16(v1338, v1299);
+ int16x8_t v1340 = vaddq_s16(v1339, v1321);
+ int16x8_t v1341 = vaddq_s16(v1340, v1330);
+ int16x8_t v1342 = vaddq_s16(v1341, v1334);
+ int16x8_t v1343 = vaddq_s16(v1335, v1342);
+ int16x8_t v1344 = vqrdmulhq_n_s16(v1343, 17734);
+ int16x8_t v1345 = vaddq_s16(v1333, v1344);
+ int16x8_t v1346 = vqrdmulhq_n_s16(v1345, 16705);
+ int16x8_t v1347 = vaddq_s16(v1327, v1346);
+ int16x8_t v1348 = vqrdmulhq_n_s16(v1347, 16463);
+ int16x8_t v1349 = vaddq_s16(v1310, v1348);
+ int16x8_t v1350 = vqrdmulhq_n_s16(v1349, 16404);
+ int16x8_t v1351 = vaddq_s16(v1239, v1350);
+ int16x8_t v1352 = vaddq_s16(v1241, v1147);
+ int16x8_t v1353_tmp = vqrdmulhq_n_s16(v1352, 13573);
+ int16x8_t v1353 = vaddq_s16(v1353_tmp, v1352);
+ int16x8_t v1354 = vaddq_s16(v1245, v1149);
+ int16x8_t v1355 = vaddq_s16(v1150, v1247);
+ int16x8_t v1356 = vaddq_s16(v1354, v1355);
+ int16x8_t v1357 = vaddq_s16(v1353, v1356);
+ int16x8_t v1358 = vaddq_s16(v1253, v1153);
+ int16x8_t v1359 = vaddq_s16(v1154, v1255);
+ int16x8_t v1360 = vaddq_s16(v1358, v1359);
+ int16x8_t v1361_tmp = vqrdmulhq_n_s16(v1360, 13573);
+ int16x8_t v1361 = vaddq_s16(v1361_tmp, v1360);
+ int16x8_t v1362 = vaddq_s16(v1261, v1157);
+ int16x8_t v1363 = vaddq_s16(v1158, v1263);
+ int16x8_t v1364 = vaddq_s16(v1362, v1363);
+ int16x8_t v1365 = vaddq_s16(v1364, v1360);
+ int16x8_t v1366 = vaddq_s16(v1361, v1365);
+ int16x8_t v1367 = vqrdmulhq_n_s16(v1366, 17734);
+ int16x8_t v1368 = vaddq_s16(v1357, v1367);
+ int16x8_t v1369 = vaddq_s16(v1272, v1164);
+ int16x8_t v1370 = vaddq_s16(v1165, v1274);
+ int16x8_t v1371 = vaddq_s16(v1369, v1370);
+ int16x8_t v1372_tmp = vqrdmulhq_n_s16(v1371, 13573);
+ int16x8_t v1372 = vaddq_s16(v1372_tmp, v1371);
+ int16x8_t v1373 = vaddq_s16(v1280, v1168);
+ int16x8_t v1374 = vaddq_s16(v1169, v1282);
+ int16x8_t v1375 = vaddq_s16(v1373, v1374);
+ int16x8_t v1376 = vaddq_s16(v1287, v1171);
+ int16x8_t v1377 = vaddq_s16(v1172, v1289);
+ int16x8_t v1378 = vaddq_s16(v1376, v1377);
+ int16x8_t v1379 = vaddq_s16(v1375, v1378);
+ int16x8_t v1380 = vaddq_s16(v1372, v1379);
+ int16x8_t v1381 = vaddq_s16(v1378, v1371);
+ int16x8_t v1382_tmp = vqrdmulhq_n_s16(v1381, 13573);
+ int16x8_t v1382 = vaddq_s16(v1382_tmp, v1381);
+ int16x8_t v1383 = vaddq_s16(v1298, v1178);
+ int16x8_t v1384 = vaddq_s16(v1179, v1300);
+ int16x8_t v1385 = vaddq_s16(v1383, v1384);
+ int16x8_t v1386 = vaddq_s16(v1385, v1375);
+ int16x8_t v1387 = vaddq_s16(v1386, v1381);
+ int16x8_t v1388 = vaddq_s16(v1382, v1387);
+ int16x8_t v1389 = vqrdmulhq_n_s16(v1388, 17734);
+ int16x8_t v1390 = vaddq_s16(v1380, v1389);
+ int16x8_t v1391 = vqrdmulhq_n_s16(v1390, 16705);
+ int16x8_t v1392 = vaddq_s16(v1368, v1391);
+ int16x8_t v1393 = vaddq_s16(v1275, v1188);
+ int16x8_t v1394 = vaddq_s16(v1189, v1240);
+ int16x8_t v1395 = vaddq_s16(v1393, v1394);
+ int16x8_t v1396_tmp = vqrdmulhq_n_s16(v1395, 13573);
+ int16x8_t v1396 = vaddq_s16(v1396_tmp, v1395);
+ int16x8_t v1397 = vaddq_s16(v1283, v1192);
+ int16x8_t v1398 = vaddq_s16(v1193, v1244);
+ int16x8_t v1399 = vaddq_s16(v1397, v1398);
+ int16x8_t v1400 = vaddq_s16(v1248, v1195);
+ int16x8_t v1401 = vaddq_s16(v1196, v1286);
+ int16x8_t v1402 = vaddq_s16(v1400, v1401);
+ int16x8_t v1403 = vaddq_s16(v1399, v1402);
+ int16x8_t v1404 = vaddq_s16(v1396, v1403);
+ int16x8_t v1405 = vaddq_s16(v1290, v1200);
+ int16x8_t v1406 = vaddq_s16(v1201, v1252);
+ int16x8_t v1407 = vaddq_s16(v1405, v1406);
+ int16x8_t v1408 = vaddq_s16(v1256, v1203);
+ int16x8_t v1409 = vaddq_s16(v1204, v1271);
+ int16x8_t v1410 = vaddq_s16(v1408, v1409);
+ int16x8_t v1411 = vaddq_s16(v1407, v1410);
+ int16x8_t v1412_tmp = vqrdmulhq_n_s16(v1411, 13573);
+ int16x8_t v1412 = vaddq_s16(v1412_tmp, v1411);
+ int16x8_t v1413 = vaddq_s16(v1301, v1208);
+ int16x8_t v1414 = vaddq_s16(v1209, v1260);
+ int16x8_t v1415 = vaddq_s16(v1413, v1414);
+ int16x8_t v1416 = vaddq_s16(v1264, v1211);
+ int16x8_t v1417 = vaddq_s16(v1212, v1279);
+ int16x8_t v1418 = vaddq_s16(v1416, v1417);
+ int16x8_t v1419 = vaddq_s16(v1415, v1418);
+ int16x8_t v1420 = vaddq_s16(v1419, v1411);
+ int16x8_t v1421 = vaddq_s16(v1412, v1420);
+ int16x8_t v1422 = vqrdmulhq_n_s16(v1421, 17734);
+ int16x8_t v1423 = vaddq_s16(v1404, v1422);
+ int16x8_t v1424 = vaddq_s16(v1410, v1395);
+ int16x8_t v1425_tmp = vqrdmulhq_n_s16(v1424, 13573);
+ int16x8_t v1425 = vaddq_s16(v1425_tmp, v1424);
+ int16x8_t v1426 = vaddq_s16(v1418, v1399);
+ int16x8_t v1427 = vaddq_s16(v1402, v1407);
+ int16x8_t v1428 = vaddq_s16(v1426, v1427);
+ int16x8_t v1429 = vaddq_s16(v1425, v1428);
+ int16x8_t v1430 = vaddq_s16(v1427, v1424);
+ int16x8_t v1431_tmp = vqrdmulhq_n_s16(v1430, 13573);
+ int16x8_t v1431 = vaddq_s16(v1431_tmp, v1430);
+ int16x8_t v1432 = vaddq_s16(v1337, v1227);
+ int16x8_t v1433 = vaddq_s16(v1228, v1297);
+ int16x8_t v1434 = vaddq_s16(v1432, v1433);
+ int16x8_t v1435 = vaddq_s16(v1434, v1415);
+ int16x8_t v1436 = vaddq_s16(v1435, v1426);
+ int16x8_t v1437 = vaddq_s16(v1436, v1430);
+ int16x8_t v1438 = vaddq_s16(v1431, v1437);
+ int16x8_t v1439 = vqrdmulhq_n_s16(v1438, 17734);
+ int16x8_t v1440 = vaddq_s16(v1429, v1439);
+ int16x8_t v1441 = vqrdmulhq_n_s16(v1440, 16705);
+ int16x8_t v1442 = vaddq_s16(v1423, v1441);
+ int16x8_t v1443 = vqrdmulhq_n_s16(v1442, 16463);
+ int16x8_t v1444 = vaddq_s16(v1392, v1443);
+ int16x8_t v1445 = vaddq_s16(v1394, v1352);
+ int16x8_t v1446_tmp = vqrdmulhq_n_s16(v1445, 13573);
+ int16x8_t v1446 = vaddq_s16(v1446_tmp, v1445);
+ int16x8_t v1447 = vaddq_s16(v1398, v1354);
+ int16x8_t v1448 = vaddq_s16(v1355, v1400);
+ int16x8_t v1449 = vaddq_s16(v1447, v1448);
+ int16x8_t v1450 = vaddq_s16(v1446, v1449);
+ int16x8_t v1451 = vaddq_s16(v1406, v1358);
+ int16x8_t v1452 = vaddq_s16(v1359, v1408);
+ int16x8_t v1453 = vaddq_s16(v1451, v1452);
+ int16x8_t v1454_tmp = vqrdmulhq_n_s16(v1453, 13573);
+ int16x8_t v1454 = vaddq_s16(v1454_tmp, v1453);
+ int16x8_t v1455 = vaddq_s16(v1414, v1362);
+ int16x8_t v1456 = vaddq_s16(v1363, v1416);
+ int16x8_t v1457 = vaddq_s16(v1455, v1456);
+ int16x8_t v1458 = vaddq_s16(v1457, v1453);
+ int16x8_t v1459 = vaddq_s16(v1454, v1458);
+ int16x8_t v1460 = vqrdmulhq_n_s16(v1459, 17734);
+ int16x8_t v1461 = vaddq_s16(v1450, v1460);
+ int16x8_t v1462 = vaddq_s16(v1409, v1369);
+ int16x8_t v1463 = vaddq_s16(v1370, v1393);
+ int16x8_t v1464 = vaddq_s16(v1462, v1463);
+ int16x8_t v1465_tmp = vqrdmulhq_n_s16(v1464, 13573);
+ int16x8_t v1465 = vaddq_s16(v1465_tmp, v1464);
+ int16x8_t v1466 = vaddq_s16(v1417, v1373);
+ int16x8_t v1467 = vaddq_s16(v1374, v1397);
+ int16x8_t v1468 = vaddq_s16(v1466, v1467);
+ int16x8_t v1469 = vaddq_s16(v1401, v1376);
+ int16x8_t v1470 = vaddq_s16(v1377, v1405);
+ int16x8_t v1471 = vaddq_s16(v1469, v1470);
+ int16x8_t v1472 = vaddq_s16(v1468, v1471);
+ int16x8_t v1473 = vaddq_s16(v1465, v1472);
+ int16x8_t v1474 = vaddq_s16(v1471, v1464);
+ int16x8_t v1475_tmp = vqrdmulhq_n_s16(v1474, 13573);
+ int16x8_t v1475 = vaddq_s16(v1475_tmp, v1474);
+ int16x8_t v1476 = vaddq_s16(v1433, v1383);
+ int16x8_t v1477 = vaddq_s16(v1384, v1413);
+ int16x8_t v1478 = vaddq_s16(v1476, v1477);
+ int16x8_t v1479 = vaddq_s16(v1478, v1468);
+ int16x8_t v1480 = vaddq_s16(v1479, v1474);
+ int16x8_t v1481 = vaddq_s16(v1475, v1480);
+ int16x8_t v1482 = vqrdmulhq_n_s16(v1481, 17734);
+ int16x8_t v1483 = vaddq_s16(v1473, v1482);
+ int16x8_t v1484 = vqrdmulhq_n_s16(v1483, 16705);
+ int16x8_t v1485 = vaddq_s16(v1461, v1484);
+ int16x8_t v1486 = vaddq_s16(v1463, v1445);
+ int16x8_t v1487_tmp = vqrdmulhq_n_s16(v1486, 13573);
+ int16x8_t v1487 = vaddq_s16(v1487_tmp, v1486);
+ int16x8_t v1488 = vaddq_s16(v1467, v1447);
+ int16x8_t v1489 = vaddq_s16(v1448, v1469);
+ int16x8_t v1490 = vaddq_s16(v1488, v1489);
+ int16x8_t v1491 = vaddq_s16(v1487, v1490);
+ int16x8_t v1492 = vaddq_s16(v1470, v1451);
+ int16x8_t v1493 = vaddq_s16(v1452, v1462);
+ int16x8_t v1494 = vaddq_s16(v1492, v1493);
+ int16x8_t v1495_tmp = vqrdmulhq_n_s16(v1494, 13573);
+ int16x8_t v1495 = vaddq_s16(v1495_tmp, v1494);
+ int16x8_t v1496 = vaddq_s16(v1477, v1455);
+ int16x8_t v1497 = vaddq_s16(v1456, v1466);
+ int16x8_t v1498 = vaddq_s16(v1496, v1497);
+ int16x8_t v1499 = vaddq_s16(v1498, v1494);
+ int16x8_t v1500 = vaddq_s16(v1495, v1499);
+ int16x8_t v1501 = vqrdmulhq_n_s16(v1500, 17734);
+ int16x8_t v1502 = vaddq_s16(v1491, v1501);
+ int16x8_t v1503 = vaddq_s16(v1493, v1486);
+ int16x8_t v1504_tmp = vqrdmulhq_n_s16(v1503, 13573);
+ int16x8_t v1504 = vaddq_s16(v1504_tmp, v1503);
+ int16x8_t v1505 = vaddq_s16(v1497, v1488);
+ int16x8_t v1506 = vaddq_s16(v1489, v1492);
+ int16x8_t v1507 = vaddq_s16(v1505, v1506);
+ int16x8_t v1508 = vaddq_s16(v1504, v1507);
+ int16x8_t v1509 = vaddq_s16(v1506, v1503);
+ int16x8_t v1510_tmp = vqrdmulhq_n_s16(v1509, 13573);
+ int16x8_t v1510 = vaddq_s16(v1510_tmp, v1509);
+ int16x8_t v1511 = vld1q_s16(in + in_stride * 255 + i);
+ int16x8_t v1512 = vaddq_s16(v1511, v1128);
+ int16x8_t v1513 = vaddq_s16(v1512, v1336);
+ int16x8_t v1514 = vaddq_s16(v1513, v1432);
+ int16x8_t v1515 = vaddq_s16(v1514, v1476);
+ int16x8_t v1516 = vaddq_s16(v1515, v1496);
+ int16x8_t v1517 = vaddq_s16(v1516, v1505);
+ int16x8_t v1518 = vaddq_s16(v1517, v1509);
+ int16x8_t v1519 = vaddq_s16(v1510, v1518);
+ int16x8_t v1520 = vqrdmulhq_n_s16(v1519, 17734);
+ int16x8_t v1521 = vaddq_s16(v1508, v1520);
+ int16x8_t v1522 = vqrdmulhq_n_s16(v1521, 16705);
+ int16x8_t v1523 = vaddq_s16(v1502, v1522);
+ int16x8_t v1524 = vqrdmulhq_n_s16(v1523, 16463);
+ int16x8_t v1525 = vaddq_s16(v1485, v1524);
+ int16x8_t v1526 = vqrdmulhq_n_s16(v1525, 16404);
+ int16x8_t v1527 = vaddq_s16(v1444, v1526);
+ int16x8_t v1528 = vqrdmulhq_n_s16(v1527, 16389);
+ int16x8_t v1529 = vaddq_s16(v1351, v1528);
+ int16x8_t v1530 = vqrdmulhq_n_s16(v1529, 16385);
+ int16x8_t v1531 = vaddq_s16(v1146, v1530);
+ int16x8_t v1532 = vqrdmulhq_n_s16(v1531, 16384);
+ int16x8_t v1533 = vaddq_s16(v701, v1532);
+ int16x8_t v1534 = vsubq_s16(v0, v1);
+ int16x8_t v1535 = vsubq_s16(v4, v6);
+ int16x8_t v1536_tmp = vqrdmulhq_n_s16(v1535, 10045);
+ int16x8_t v1536 = vaddq_s16(v1536_tmp, v1535);
+ int16x8_t v1537 = vaddq_s16(v1534, v1536);
+ int16x8_t v1538 = vsubq_s16(v11, v14);
+ int16x8_t v1539 = vsubq_s16(v17, v20);
+ int16x8_t v1540_tmp = vqrdmulhq_n_s16(v1539, 10045);
+ int16x8_t v1540 = vaddq_s16(v1540_tmp, v1539);
+ int16x8_t v1541 = vaddq_s16(v1538, v1540);
+ int16x8_t v1542 = vqrdmulhq_n_s16(v1541, 19705);
+ int16x8_t v1543 = vaddq_s16(v1537, v1542);
+ int16x8_t v1544 = vsubq_s16(v27, v30);
+ int16x8_t v1545 = vsubq_s16(v35, v39);
+ int16x8_t v1546_tmp = vqrdmulhq_n_s16(v1545, 10045);
+ int16x8_t v1546 = vaddq_s16(v1546_tmp, v1545);
+ int16x8_t v1547 = vaddq_s16(v1544, v1546);
+ int16x8_t v1548 = vsubq_s16(v44, v47);
+ int16x8_t v1549 = vsubq_s16(v50, v54);
+ int16x8_t v1550_tmp = vqrdmulhq_n_s16(v1549, 10045);
+ int16x8_t v1550 = vaddq_s16(v1550_tmp, v1549);
+ int16x8_t v1551 = vaddq_s16(v1548, v1550);
+ int16x8_t v1552 = vqrdmulhq_n_s16(v1551, 19705);
+ int16x8_t v1553 = vaddq_s16(v1547, v1552);
+ int16x8_t v1554 = vqrdmulhq_n_s16(v1553, 17121);
+ int16x8_t v1555 = vaddq_s16(v1543, v1554);
+ int16x8_t v1556 = vsubq_s16(v63, v66);
+ int16x8_t v1557 = vsubq_s16(v71, v75);
+ int16x8_t v1558_tmp = vqrdmulhq_n_s16(v1557, 10045);
+ int16x8_t v1558 = vaddq_s16(v1558_tmp, v1557);
+ int16x8_t v1559 = vaddq_s16(v1556, v1558);
+ int16x8_t v1560 = vsubq_s16(v82, v89);
+ int16x8_t v1561 = vsubq_s16(v92, v97);
+ int16x8_t v1562_tmp = vqrdmulhq_n_s16(v1561, 10045);
+ int16x8_t v1562 = vaddq_s16(v1562_tmp, v1561);
+ int16x8_t v1563 = vaddq_s16(v1560, v1562);
+ int16x8_t v1564 = vqrdmulhq_n_s16(v1563, 19705);
+ int16x8_t v1565 = vaddq_s16(v1559, v1564);
+ int16x8_t v1566 = vsubq_s16(v104, v107);
+ int16x8_t v1567 = vsubq_s16(v112, v116);
+ int16x8_t v1568_tmp = vqrdmulhq_n_s16(v1567, 10045);
+ int16x8_t v1568 = vaddq_s16(v1568_tmp, v1567);
+ int16x8_t v1569 = vaddq_s16(v1566, v1568);
+ int16x8_t v1570 = vsubq_s16(v121, v124);
+ int16x8_t v1571 = vsubq_s16(v127, v132);
+ int16x8_t v1572_tmp = vqrdmulhq_n_s16(v1571, 10045);
+ int16x8_t v1572 = vaddq_s16(v1572_tmp, v1571);
+ int16x8_t v1573 = vaddq_s16(v1570, v1572);
+ int16x8_t v1574 = vqrdmulhq_n_s16(v1573, 19705);
+ int16x8_t v1575 = vaddq_s16(v1569, v1574);
+ int16x8_t v1576 = vqrdmulhq_n_s16(v1575, 17121);
+ int16x8_t v1577 = vaddq_s16(v1565, v1576);
+ int16x8_t v1578 = vqrdmulhq_n_s16(v1577, 16563);
+ int16x8_t v1579 = vaddq_s16(v1555, v1578);
+ int16x8_t v1580 = vsubq_s16(v143, v146);
+ int16x8_t v1581 = vsubq_s16(v151, v155);
+ int16x8_t v1582_tmp = vqrdmulhq_n_s16(v1581, 10045);
+ int16x8_t v1582 = vaddq_s16(v1582_tmp, v1581);
+ int16x8_t v1583 = vaddq_s16(v1580, v1582);
+ int16x8_t v1584 = vsubq_s16(v162, v169);
+ int16x8_t v1585 = vsubq_s16(v172, v177);
+ int16x8_t v1586_tmp = vqrdmulhq_n_s16(v1585, 10045);
+ int16x8_t v1586 = vaddq_s16(v1586_tmp, v1585);
+ int16x8_t v1587 = vaddq_s16(v1584, v1586);
+ int16x8_t v1588 = vqrdmulhq_n_s16(v1587, 19705);
+ int16x8_t v1589 = vaddq_s16(v1583, v1588);
+ int16x8_t v1590 = vsubq_s16(v186, v193);
+ int16x8_t v1591 = vsubq_s16(v202, v210);
+ int16x8_t v1592_tmp = vqrdmulhq_n_s16(v1591, 10045);
+ int16x8_t v1592 = vaddq_s16(v1592_tmp, v1591);
+ int16x8_t v1593 = vaddq_s16(v1590, v1592);
+ int16x8_t v1594 = vsubq_s16(v215, v218);
+ int16x8_t v1595 = vsubq_s16(v221, v227);
+ int16x8_t v1596_tmp = vqrdmulhq_n_s16(v1595, 10045);
+ int16x8_t v1596 = vaddq_s16(v1596_tmp, v1595);
+ int16x8_t v1597 = vaddq_s16(v1594, v1596);
+ int16x8_t v1598 = vqrdmulhq_n_s16(v1597, 19705);
+ int16x8_t v1599 = vaddq_s16(v1593, v1598);
+ int16x8_t v1600 = vqrdmulhq_n_s16(v1599, 17121);
+ int16x8_t v1601 = vaddq_s16(v1589, v1600);
+ int16x8_t v1602 = vsubq_s16(v236, v239);
+ int16x8_t v1603 = vsubq_s16(v244, v248);
+ int16x8_t v1604_tmp = vqrdmulhq_n_s16(v1603, 10045);
+ int16x8_t v1604 = vaddq_s16(v1604_tmp, v1603);
+ int16x8_t v1605 = vaddq_s16(v1602, v1604);
+ int16x8_t v1606 = vsubq_s16(v255, v262);
+ int16x8_t v1607 = vsubq_s16(v265, v270);
+ int16x8_t v1608_tmp = vqrdmulhq_n_s16(v1607, 10045);
+ int16x8_t v1608 = vaddq_s16(v1608_tmp, v1607);
+ int16x8_t v1609 = vaddq_s16(v1606, v1608);
+ int16x8_t v1610 = vqrdmulhq_n_s16(v1609, 19705);
+ int16x8_t v1611 = vaddq_s16(v1605, v1610);
+ int16x8_t v1612 = vsubq_s16(v277, v280);
+ int16x8_t v1613 = vsubq_s16(v285, v289);
+ int16x8_t v1614_tmp = vqrdmulhq_n_s16(v1613, 10045);
+ int16x8_t v1614 = vaddq_s16(v1614_tmp, v1613);
+ int16x8_t v1615 = vaddq_s16(v1612, v1614);
+ int16x8_t v1616 = vsubq_s16(v294, v297);
+ int16x8_t v1617 = vsubq_s16(v300, v306);
+ int16x8_t v1618_tmp = vqrdmulhq_n_s16(v1617, 10045);
+ int16x8_t v1618 = vaddq_s16(v1618_tmp, v1617);
+ int16x8_t v1619 = vaddq_s16(v1616, v1618);
+ int16x8_t v1620 = vqrdmulhq_n_s16(v1619, 19705);
+ int16x8_t v1621 = vaddq_s16(v1615, v1620);
+ int16x8_t v1622 = vqrdmulhq_n_s16(v1621, 17121);
+ int16x8_t v1623 = vaddq_s16(v1611, v1622);
+ int16x8_t v1624 = vqrdmulhq_n_s16(v1623, 16563);
+ int16x8_t v1625 = vaddq_s16(v1601, v1624);
+ int16x8_t v1626 = vqrdmulhq_n_s16(v1625, 16429);
+ int16x8_t v1627 = vaddq_s16(v1579, v1626);
+ int16x8_t v1628 = vsubq_s16(v319, v322);
+ int16x8_t v1629 = vsubq_s16(v327, v331);
+ int16x8_t v1630_tmp = vqrdmulhq_n_s16(v1629, 10045);
+ int16x8_t v1630 = vaddq_s16(v1630_tmp, v1629);
+ int16x8_t v1631 = vaddq_s16(v1628, v1630);
+ int16x8_t v1632 = vsubq_s16(v338, v345);
+ int16x8_t v1633 = vsubq_s16(v348, v353);
+ int16x8_t v1634_tmp = vqrdmulhq_n_s16(v1633, 10045);
+ int16x8_t v1634 = vaddq_s16(v1634_tmp, v1633);
+ int16x8_t v1635 = vaddq_s16(v1632, v1634);
+ int16x8_t v1636 = vqrdmulhq_n_s16(v1635, 19705);
+ int16x8_t v1637 = vaddq_s16(v1631, v1636);
+ int16x8_t v1638 = vsubq_s16(v362, v369);
+ int16x8_t v1639 = vsubq_s16(v378, v386);
+ int16x8_t v1640_tmp = vqrdmulhq_n_s16(v1639, 10045);
+ int16x8_t v1640 = vaddq_s16(v1640_tmp, v1639);
+ int16x8_t v1641 = vaddq_s16(v1638, v1640);
+ int16x8_t v1642 = vsubq_s16(v391, v394);
+ int16x8_t v1643 = vsubq_s16(v397, v403);
+ int16x8_t v1644_tmp = vqrdmulhq_n_s16(v1643, 10045);
+ int16x8_t v1644 = vaddq_s16(v1644_tmp, v1643);
+ int16x8_t v1645 = vaddq_s16(v1642, v1644);
+ int16x8_t v1646 = vqrdmulhq_n_s16(v1645, 19705);
+ int16x8_t v1647 = vaddq_s16(v1641, v1646);
+ int16x8_t v1648 = vqrdmulhq_n_s16(v1647, 17121);
+ int16x8_t v1649 = vaddq_s16(v1637, v1648);
+ int16x8_t v1650 = vsubq_s16(v414, v421);
+ int16x8_t v1651 = vsubq_s16(v430, v438);
+ int16x8_t v1652_tmp = vqrdmulhq_n_s16(v1651, 10045);
+ int16x8_t v1652 = vaddq_s16(v1652_tmp, v1651);
+ int16x8_t v1653 = vaddq_s16(v1650, v1652);
+ int16x8_t v1654 = vsubq_s16(v449, v464);
+ int16x8_t v1655 = vsubq_s16(v467, v476);
+ int16x8_t v1656_tmp = vqrdmulhq_n_s16(v1655, 10045);
+ int16x8_t v1656 = vaddq_s16(v1656_tmp, v1655);
+ int16x8_t v1657 = vaddq_s16(v1654, v1656);
+ int16x8_t v1658 = vqrdmulhq_n_s16(v1657, 19705);
+ int16x8_t v1659 = vaddq_s16(v1653, v1658);
+ int16x8_t v1660 = vsubq_s16(v483, v486);
+ int16x8_t v1661 = vsubq_s16(v491, v495);
+ int16x8_t v1662_tmp = vqrdmulhq_n_s16(v1661, 10045);
+ int16x8_t v1662 = vaddq_s16(v1662_tmp, v1661);
+ int16x8_t v1663 = vaddq_s16(v1660, v1662);
+ int16x8_t v1664 = vsubq_s16(v500, v503);
+ int16x8_t v1665 = vsubq_s16(v506, v513);
+ int16x8_t v1666_tmp = vqrdmulhq_n_s16(v1665, 10045);
+ int16x8_t v1666 = vaddq_s16(v1666_tmp, v1665);
+ int16x8_t v1667 = vaddq_s16(v1664, v1666);
+ int16x8_t v1668 = vqrdmulhq_n_s16(v1667, 19705);
+ int16x8_t v1669 = vaddq_s16(v1663, v1668);
+ int16x8_t v1670 = vqrdmulhq_n_s16(v1669, 17121);
+ int16x8_t v1671 = vaddq_s16(v1659, v1670);
+ int16x8_t v1672 = vqrdmulhq_n_s16(v1671, 16563);
+ int16x8_t v1673 = vaddq_s16(v1649, v1672);
+ int16x8_t v1674 = vsubq_s16(v524, v527);
+ int16x8_t v1675 = vsubq_s16(v532, v536);
+ int16x8_t v1676_tmp = vqrdmulhq_n_s16(v1675, 10045);
+ int16x8_t v1676 = vaddq_s16(v1676_tmp, v1675);
+ int16x8_t v1677 = vaddq_s16(v1674, v1676);
+ int16x8_t v1678 = vsubq_s16(v543, v550);
+ int16x8_t v1679 = vsubq_s16(v553, v558);
+ int16x8_t v1680_tmp = vqrdmulhq_n_s16(v1679, 10045);
+ int16x8_t v1680 = vaddq_s16(v1680_tmp, v1679);
+ int16x8_t v1681 = vaddq_s16(v1678, v1680);
+ int16x8_t v1682 = vqrdmulhq_n_s16(v1681, 19705);
+ int16x8_t v1683 = vaddq_s16(v1677, v1682);
+ int16x8_t v1684 = vsubq_s16(v567, v574);
+ int16x8_t v1685 = vsubq_s16(v583, v591);
+ int16x8_t v1686_tmp = vqrdmulhq_n_s16(v1685, 10045);
+ int16x8_t v1686 = vaddq_s16(v1686_tmp, v1685);
+ int16x8_t v1687 = vaddq_s16(v1684, v1686);
+ int16x8_t v1688 = vsubq_s16(v596, v599);
+ int16x8_t v1689 = vsubq_s16(v602, v608);
+ int16x8_t v1690_tmp = vqrdmulhq_n_s16(v1689, 10045);
+ int16x8_t v1690 = vaddq_s16(v1690_tmp, v1689);
+ int16x8_t v1691 = vaddq_s16(v1688, v1690);
+ int16x8_t v1692 = vqrdmulhq_n_s16(v1691, 19705);
+ int16x8_t v1693 = vaddq_s16(v1687, v1692);
+ int16x8_t v1694 = vqrdmulhq_n_s16(v1693, 17121);
+ int16x8_t v1695 = vaddq_s16(v1683, v1694);
+ int16x8_t v1696 = vsubq_s16(v617, v620);
+ int16x8_t v1697 = vsubq_s16(v625, v629);
+ int16x8_t v1698_tmp = vqrdmulhq_n_s16(v1697, 10045);
+ int16x8_t v1698 = vaddq_s16(v1698_tmp, v1697);
+ int16x8_t v1699 = vaddq_s16(v1696, v1698);
+ int16x8_t v1700 = vsubq_s16(v636, v643);
+ int16x8_t v1701 = vsubq_s16(v646, v651);
+ int16x8_t v1702_tmp = vqrdmulhq_n_s16(v1701, 10045);
+ int16x8_t v1702 = vaddq_s16(v1702_tmp, v1701);
+ int16x8_t v1703 = vaddq_s16(v1700, v1702);
+ int16x8_t v1704 = vqrdmulhq_n_s16(v1703, 19705);
+ int16x8_t v1705 = vaddq_s16(v1699, v1704);
+ int16x8_t v1706 = vsubq_s16(v658, v661);
+ int16x8_t v1707 = vsubq_s16(v666, v670);
+ int16x8_t v1708_tmp = vqrdmulhq_n_s16(v1707, 10045);
+ int16x8_t v1708 = vaddq_s16(v1708_tmp, v1707);
+ int16x8_t v1709 = vaddq_s16(v1706, v1708);
+ int16x8_t v1710 = vsubq_s16(v675, v678);
+ int16x8_t v1711 = vsubq_s16(v681, v688);
+ int16x8_t v1712_tmp = vqrdmulhq_n_s16(v1711, 10045);
+ int16x8_t v1712 = vaddq_s16(v1712_tmp, v1711);
+ int16x8_t v1713 = vaddq_s16(v1710, v1712);
+ int16x8_t v1714 = vqrdmulhq_n_s16(v1713, 19705);
+ int16x8_t v1715 = vaddq_s16(v1709, v1714);
+ int16x8_t v1716 = vqrdmulhq_n_s16(v1715, 17121);
+ int16x8_t v1717 = vaddq_s16(v1705, v1716);
+ int16x8_t v1718 = vqrdmulhq_n_s16(v1717, 16563);
+ int16x8_t v1719 = vaddq_s16(v1695, v1718);
+ int16x8_t v1720 = vqrdmulhq_n_s16(v1719, 16429);
+ int16x8_t v1721 = vaddq_s16(v1673, v1720);
+ int16x8_t v1722 = vqrdmulhq_n_s16(v1721, 16395);
+ int16x8_t v1723 = vaddq_s16(v1627, v1722);
+ int16x8_t v1724 = vsubq_s16(v703, v706);
+ int16x8_t v1725 = vsubq_s16(v711, v715);
+ int16x8_t v1726_tmp = vqrdmulhq_n_s16(v1725, 10045);
+ int16x8_t v1726 = vaddq_s16(v1726_tmp, v1725);
+ int16x8_t v1727 = vaddq_s16(v1724, v1726);
+ int16x8_t v1728 = vsubq_s16(v722, v729);
+ int16x8_t v1729 = vsubq_s16(v732, v737);
+ int16x8_t v1730_tmp = vqrdmulhq_n_s16(v1729, 10045);
+ int16x8_t v1730 = vaddq_s16(v1730_tmp, v1729);
+ int16x8_t v1731 = vaddq_s16(v1728, v1730);
+ int16x8_t v1732 = vqrdmulhq_n_s16(v1731, 19705);
+ int16x8_t v1733 = vaddq_s16(v1727, v1732);
+ int16x8_t v1734 = vsubq_s16(v746, v753);
+ int16x8_t v1735 = vsubq_s16(v762, v770);
+ int16x8_t v1736_tmp = vqrdmulhq_n_s16(v1735, 10045);
+ int16x8_t v1736 = vaddq_s16(v1736_tmp, v1735);
+ int16x8_t v1737 = vaddq_s16(v1734, v1736);
+ int16x8_t v1738 = vsubq_s16(v775, v778);
+ int16x8_t v1739 = vsubq_s16(v781, v787);
+ int16x8_t v1740_tmp = vqrdmulhq_n_s16(v1739, 10045);
+ int16x8_t v1740 = vaddq_s16(v1740_tmp, v1739);
+ int16x8_t v1741 = vaddq_s16(v1738, v1740);
+ int16x8_t v1742 = vqrdmulhq_n_s16(v1741, 19705);
+ int16x8_t v1743 = vaddq_s16(v1737, v1742);
+ int16x8_t v1744 = vqrdmulhq_n_s16(v1743, 17121);
+ int16x8_t v1745 = vaddq_s16(v1733, v1744);
+ int16x8_t v1746 = vsubq_s16(v798, v805);
+ int16x8_t v1747 = vsubq_s16(v814, v822);
+ int16x8_t v1748_tmp = vqrdmulhq_n_s16(v1747, 10045);
+ int16x8_t v1748 = vaddq_s16(v1748_tmp, v1747);
+ int16x8_t v1749 = vaddq_s16(v1746, v1748);
+ int16x8_t v1750 = vsubq_s16(v833, v848);
+ int16x8_t v1751 = vsubq_s16(v851, v860);
+ int16x8_t v1752_tmp = vqrdmulhq_n_s16(v1751, 10045);
+ int16x8_t v1752 = vaddq_s16(v1752_tmp, v1751);
+ int16x8_t v1753 = vaddq_s16(v1750, v1752);
+ int16x8_t v1754 = vqrdmulhq_n_s16(v1753, 19705);
+ int16x8_t v1755 = vaddq_s16(v1749, v1754);
+ int16x8_t v1756 = vsubq_s16(v867, v870);
+ int16x8_t v1757 = vsubq_s16(v875, v879);
+ int16x8_t v1758_tmp = vqrdmulhq_n_s16(v1757, 10045);
+ int16x8_t v1758 = vaddq_s16(v1758_tmp, v1757);
+ int16x8_t v1759 = vaddq_s16(v1756, v1758);
+ int16x8_t v1760 = vsubq_s16(v884, v887);
+ int16x8_t v1761 = vsubq_s16(v890, v897);
+ int16x8_t v1762_tmp = vqrdmulhq_n_s16(v1761, 10045);
+ int16x8_t v1762 = vaddq_s16(v1762_tmp, v1761);
+ int16x8_t v1763 = vaddq_s16(v1760, v1762);
+ int16x8_t v1764 = vqrdmulhq_n_s16(v1763, 19705);
+ int16x8_t v1765 = vaddq_s16(v1759, v1764);
+ int16x8_t v1766 = vqrdmulhq_n_s16(v1765, 17121);
+ int16x8_t v1767 = vaddq_s16(v1755, v1766);
+ int16x8_t v1768 = vqrdmulhq_n_s16(v1767, 16563);
+ int16x8_t v1769 = vaddq_s16(v1745, v1768);
+ int16x8_t v1770 = vsubq_s16(v910, v917);
+ int16x8_t v1771 = vsubq_s16(v926, v934);
+ int16x8_t v1772_tmp = vqrdmulhq_n_s16(v1771, 10045);
+ int16x8_t v1772 = vaddq_s16(v1772_tmp, v1771);
+ int16x8_t v1773 = vaddq_s16(v1770, v1772);
+ int16x8_t v1774 = vsubq_s16(v945, v960);
+ int16x8_t v1775 = vsubq_s16(v963, v972);
+ int16x8_t v1776_tmp = vqrdmulhq_n_s16(v1775, 10045);
+ int16x8_t v1776 = vaddq_s16(v1776_tmp, v1775);
+ int16x8_t v1777 = vaddq_s16(v1774, v1776);
+ int16x8_t v1778 = vqrdmulhq_n_s16(v1777, 19705);
+ int16x8_t v1779 = vaddq_s16(v1773, v1778);
+ int16x8_t v1780 = vsubq_s16(v985, v1000);
+ int16x8_t v1781 = vsubq_s16(v1017, v1033);
+ int16x8_t v1782_tmp = vqrdmulhq_n_s16(v1781, 10045);
+ int16x8_t v1782 = vaddq_s16(v1782_tmp, v1781);
+ int16x8_t v1783 = vaddq_s16(v1780, v1782);
+ int16x8_t v1784 = vsubq_s16(v1038, v1041);
+ int16x8_t v1785 = vsubq_s16(v1044, v1054);
+ int16x8_t v1786_tmp = vqrdmulhq_n_s16(v1785, 10045);
+ int16x8_t v1786 = vaddq_s16(v1786_tmp, v1785);
+ int16x8_t v1787 = vaddq_s16(v1784, v1786);
+ int16x8_t v1788 = vqrdmulhq_n_s16(v1787, 19705);
+ int16x8_t v1789 = vaddq_s16(v1783, v1788);
+ int16x8_t v1790 = vqrdmulhq_n_s16(v1789, 17121);
+ int16x8_t v1791 = vaddq_s16(v1779, v1790);
+ int16x8_t v1792 = vsubq_s16(v1063, v1066);
+ int16x8_t v1793 = vsubq_s16(v1071, v1075);
+ int16x8_t v1794_tmp = vqrdmulhq_n_s16(v1793, 10045);
+ int16x8_t v1794 = vaddq_s16(v1794_tmp, v1793);
+ int16x8_t v1795 = vaddq_s16(v1792, v1794);
+ int16x8_t v1796 = vsubq_s16(v1082, v1089);
+ int16x8_t v1797 = vsubq_s16(v1092, v1097);
+ int16x8_t v1798_tmp = vqrdmulhq_n_s16(v1797, 10045);
+ int16x8_t v1798 = vaddq_s16(v1798_tmp, v1797);
+ int16x8_t v1799 = vaddq_s16(v1796, v1798);
+ int16x8_t v1800 = vqrdmulhq_n_s16(v1799, 19705);
+ int16x8_t v1801 = vaddq_s16(v1795, v1800);
+ int16x8_t v1802 = vsubq_s16(v1104, v1107);
+ int16x8_t v1803 = vsubq_s16(v1112, v1116);
+ int16x8_t v1804_tmp = vqrdmulhq_n_s16(v1803, 10045);
+ int16x8_t v1804 = vaddq_s16(v1804_tmp, v1803);
+ int16x8_t v1805 = vaddq_s16(v1802, v1804);
+ int16x8_t v1806 = vsubq_s16(v1121, v1124);
+ int16x8_t v1807 = vsubq_s16(v1127, v1135);
+ int16x8_t v1808_tmp = vqrdmulhq_n_s16(v1807, 10045);
+ int16x8_t v1808 = vaddq_s16(v1808_tmp, v1807);
+ int16x8_t v1809 = vaddq_s16(v1806, v1808);
+ int16x8_t v1810 = vqrdmulhq_n_s16(v1809, 19705);
+ int16x8_t v1811 = vaddq_s16(v1805, v1810);
+ int16x8_t v1812 = vqrdmulhq_n_s16(v1811, 17121);
+ int16x8_t v1813 = vaddq_s16(v1801, v1812);
+ int16x8_t v1814 = vqrdmulhq_n_s16(v1813, 16563);
+ int16x8_t v1815 = vaddq_s16(v1791, v1814);
+ int16x8_t v1816 = vqrdmulhq_n_s16(v1815, 16429);
+ int16x8_t v1817 = vaddq_s16(v1769, v1816);
+ int16x8_t v1818 = vsubq_s16(v1148, v1151);
+ int16x8_t v1819 = vsubq_s16(v1156, v1160);
+ int16x8_t v1820_tmp = vqrdmulhq_n_s16(v1819, 10045);
+ int16x8_t v1820 = vaddq_s16(v1820_tmp, v1819);
+ int16x8_t v1821 = vaddq_s16(v1818, v1820);
+ int16x8_t v1822 = vsubq_s16(v1167, v1174);
+ int16x8_t v1823 = vsubq_s16(v1177, v1182);
+ int16x8_t v1824_tmp = vqrdmulhq_n_s16(v1823, 10045);
+ int16x8_t v1824 = vaddq_s16(v1824_tmp, v1823);
+ int16x8_t v1825 = vaddq_s16(v1822, v1824);
+ int16x8_t v1826 = vqrdmulhq_n_s16(v1825, 19705);
+ int16x8_t v1827 = vaddq_s16(v1821, v1826);
+ int16x8_t v1828 = vsubq_s16(v1191, v1198);
+ int16x8_t v1829 = vsubq_s16(v1207, v1215);
+ int16x8_t v1830_tmp = vqrdmulhq_n_s16(v1829, 10045);
+ int16x8_t v1830 = vaddq_s16(v1830_tmp, v1829);
+ int16x8_t v1831 = vaddq_s16(v1828, v1830);
+ int16x8_t v1832 = vsubq_s16(v1220, v1223);
+ int16x8_t v1833 = vsubq_s16(v1226, v1232);
+ int16x8_t v1834_tmp = vqrdmulhq_n_s16(v1833, 10045);
+ int16x8_t v1834 = vaddq_s16(v1834_tmp, v1833);
+ int16x8_t v1835 = vaddq_s16(v1832, v1834);
+ int16x8_t v1836 = vqrdmulhq_n_s16(v1835, 19705);
+ int16x8_t v1837 = vaddq_s16(v1831, v1836);
+ int16x8_t v1838 = vqrdmulhq_n_s16(v1837, 17121);
+ int16x8_t v1839 = vaddq_s16(v1827, v1838);
+ int16x8_t v1840 = vsubq_s16(v1243, v1250);
+ int16x8_t v1841 = vsubq_s16(v1259, v1267);
+ int16x8_t v1842_tmp = vqrdmulhq_n_s16(v1841, 10045);
+ int16x8_t v1842 = vaddq_s16(v1842_tmp, v1841);
+ int16x8_t v1843 = vaddq_s16(v1840, v1842);
+ int16x8_t v1844 = vsubq_s16(v1278, v1293);
+ int16x8_t v1845 = vsubq_s16(v1296, v1305);
+ int16x8_t v1846_tmp = vqrdmulhq_n_s16(v1845, 10045);
+ int16x8_t v1846 = vaddq_s16(v1846_tmp, v1845);
+ int16x8_t v1847 = vaddq_s16(v1844, v1846);
+ int16x8_t v1848 = vqrdmulhq_n_s16(v1847, 19705);
+ int16x8_t v1849 = vaddq_s16(v1843, v1848);
+ int16x8_t v1850 = vsubq_s16(v1312, v1315);
+ int16x8_t v1851 = vsubq_s16(v1320, v1324);
+ int16x8_t v1852_tmp = vqrdmulhq_n_s16(v1851, 10045);
+ int16x8_t v1852 = vaddq_s16(v1852_tmp, v1851);
+ int16x8_t v1853 = vaddq_s16(v1850, v1852);
+ int16x8_t v1854 = vsubq_s16(v1329, v1332);
+ int16x8_t v1855 = vsubq_s16(v1335, v1342);
+ int16x8_t v1856_tmp = vqrdmulhq_n_s16(v1855, 10045);
+ int16x8_t v1856 = vaddq_s16(v1856_tmp, v1855);
+ int16x8_t v1857 = vaddq_s16(v1854, v1856);
+ int16x8_t v1858 = vqrdmulhq_n_s16(v1857, 19705);
+ int16x8_t v1859 = vaddq_s16(v1853, v1858);
+ int16x8_t v1860 = vqrdmulhq_n_s16(v1859, 17121);
+ int16x8_t v1861 = vaddq_s16(v1849, v1860);
+ int16x8_t v1862 = vqrdmulhq_n_s16(v1861, 16563);
+ int16x8_t v1863 = vaddq_s16(v1839, v1862);
+ int16x8_t v1864 = vsubq_s16(v1353, v1356);
+ int16x8_t v1865 = vsubq_s16(v1361, v1365);
+ int16x8_t v1866_tmp = vqrdmulhq_n_s16(v1865, 10045);
+ int16x8_t v1866 = vaddq_s16(v1866_tmp, v1865);
+ int16x8_t v1867 = vaddq_s16(v1864, v1866);
+ int16x8_t v1868 = vsubq_s16(v1372, v1379);
+ int16x8_t v1869 = vsubq_s16(v1382, v1387);
+ int16x8_t v1870_tmp = vqrdmulhq_n_s16(v1869, 10045);
+ int16x8_t v1870 = vaddq_s16(v1870_tmp, v1869);
+ int16x8_t v1871 = vaddq_s16(v1868, v1870);
+ int16x8_t v1872 = vqrdmulhq_n_s16(v1871, 19705);
+ int16x8_t v1873 = vaddq_s16(v1867, v1872);
+ int16x8_t v1874 = vsubq_s16(v1396, v1403);
+ int16x8_t v1875 = vsubq_s16(v1412, v1420);
+ int16x8_t v1876_tmp = vqrdmulhq_n_s16(v1875, 10045);
+ int16x8_t v1876 = vaddq_s16(v1876_tmp, v1875);
+ int16x8_t v1877 = vaddq_s16(v1874, v1876);
+ int16x8_t v1878 = vsubq_s16(v1425, v1428);
+ int16x8_t v1879 = vsubq_s16(v1431, v1437);
+ int16x8_t v1880_tmp = vqrdmulhq_n_s16(v1879, 10045);
+ int16x8_t v1880 = vaddq_s16(v1880_tmp, v1879);
+ int16x8_t v1881 = vaddq_s16(v1878, v1880);
+ int16x8_t v1882 = vqrdmulhq_n_s16(v1881, 19705);
+ int16x8_t v1883 = vaddq_s16(v1877, v1882);
+ int16x8_t v1884 = vqrdmulhq_n_s16(v1883, 17121);
+ int16x8_t v1885 = vaddq_s16(v1873, v1884);
+ int16x8_t v1886 = vsubq_s16(v1446, v1449);
+ int16x8_t v1887 = vsubq_s16(v1454, v1458);
+ int16x8_t v1888_tmp = vqrdmulhq_n_s16(v1887, 10045);
+ int16x8_t v1888 = vaddq_s16(v1888_tmp, v1887);
+ int16x8_t v1889 = vaddq_s16(v1886, v1888);
+ int16x8_t v1890 = vsubq_s16(v1465, v1472);
+ int16x8_t v1891 = vsubq_s16(v1475, v1480);
+ int16x8_t v1892_tmp = vqrdmulhq_n_s16(v1891, 10045);
+ int16x8_t v1892 = vaddq_s16(v1892_tmp, v1891);
+ int16x8_t v1893 = vaddq_s16(v1890, v1892);
+ int16x8_t v1894 = vqrdmulhq_n_s16(v1893, 19705);
+ int16x8_t v1895 = vaddq_s16(v1889, v1894);
+ int16x8_t v1896 = vsubq_s16(v1487, v1490);
+ int16x8_t v1897 = vsubq_s16(v1495, v1499);
+ int16x8_t v1898_tmp = vqrdmulhq_n_s16(v1897, 10045);
+ int16x8_t v1898 = vaddq_s16(v1898_tmp, v1897);
+ int16x8_t v1899 = vaddq_s16(v1896, v1898);
+ int16x8_t v1900 = vsubq_s16(v1504, v1507);
+ int16x8_t v1901 = vsubq_s16(v1510, v1518);
+ int16x8_t v1902_tmp = vqrdmulhq_n_s16(v1901, 10045);
+ int16x8_t v1902 = vaddq_s16(v1902_tmp, v1901);
+ int16x8_t v1903 = vaddq_s16(v1900, v1902);
+ int16x8_t v1904 = vqrdmulhq_n_s16(v1903, 19705);
+ int16x8_t v1905 = vaddq_s16(v1899, v1904);
+ int16x8_t v1906 = vqrdmulhq_n_s16(v1905, 17121);
+ int16x8_t v1907 = vaddq_s16(v1895, v1906);
+ int16x8_t v1908 = vqrdmulhq_n_s16(v1907, 16563);
+ int16x8_t v1909 = vaddq_s16(v1885, v1908);
+ int16x8_t v1910 = vqrdmulhq_n_s16(v1909, 16429);
+ int16x8_t v1911 = vaddq_s16(v1863, v1910);
+ int16x8_t v1912 = vqrdmulhq_n_s16(v1911, 16395);
+ int16x8_t v1913 = vaddq_s16(v1817, v1912);
+ int16x8_t v1914 = vqrdmulhq_n_s16(v1913, 16387);
+ int16x8_t v1915 = vaddq_s16(v1723, v1914);
+ int16x8_t v1916 = vsubq_s16(v1534, v1536);
+ int16x8_t v1917 = vsubq_s16(v1538, v1540);
+ int16x8_t v1918 = vqrdmulhq_n_s16(v1917, 29490);
+ int16x8_t v1919 = vaddq_s16(v1916, v1918);
+ int16x8_t v1920 = vsubq_s16(v1544, v1546);
+ int16x8_t v1921 = vsubq_s16(v1548, v1550);
+ int16x8_t v1922 = vqrdmulhq_n_s16(v1921, 29490);
+ int16x8_t v1923 = vaddq_s16(v1920, v1922);
+ int16x8_t v1924 = vqrdmulhq_n_s16(v1923, 18578);
+ int16x8_t v1925 = vaddq_s16(v1919, v1924);
+ int16x8_t v1926 = vsubq_s16(v1556, v1558);
+ int16x8_t v1927 = vsubq_s16(v1560, v1562);
+ int16x8_t v1928 = vqrdmulhq_n_s16(v1927, 29490);
+ int16x8_t v1929 = vaddq_s16(v1926, v1928);
+ int16x8_t v1930 = vsubq_s16(v1566, v1568);
+ int16x8_t v1931 = vsubq_s16(v1570, v1572);
+ int16x8_t v1932 = vqrdmulhq_n_s16(v1931, 29490);
+ int16x8_t v1933 = vaddq_s16(v1930, v1932);
+ int16x8_t v1934 = vqrdmulhq_n_s16(v1933, 18578);
+ int16x8_t v1935 = vaddq_s16(v1929, v1934);
+ int16x8_t v1936 = vqrdmulhq_n_s16(v1935, 16890);
+ int16x8_t v1937 = vaddq_s16(v1925, v1936);
+ int16x8_t v1938 = vsubq_s16(v1580, v1582);
+ int16x8_t v1939 = vsubq_s16(v1584, v1586);
+ int16x8_t v1940 = vqrdmulhq_n_s16(v1939, 29490);
+ int16x8_t v1941 = vaddq_s16(v1938, v1940);
+ int16x8_t v1942 = vsubq_s16(v1590, v1592);
+ int16x8_t v1943 = vsubq_s16(v1594, v1596);
+ int16x8_t v1944 = vqrdmulhq_n_s16(v1943, 29490);
+ int16x8_t v1945 = vaddq_s16(v1942, v1944);
+ int16x8_t v1946 = vqrdmulhq_n_s16(v1945, 18578);
+ int16x8_t v1947 = vaddq_s16(v1941, v1946);
+ int16x8_t v1948 = vsubq_s16(v1602, v1604);
+ int16x8_t v1949 = vsubq_s16(v1606, v1608);
+ int16x8_t v1950 = vqrdmulhq_n_s16(v1949, 29490);
+ int16x8_t v1951 = vaddq_s16(v1948, v1950);
+ int16x8_t v1952 = vsubq_s16(v1612, v1614);
+ int16x8_t v1953 = vsubq_s16(v1616, v1618);
+ int16x8_t v1954 = vqrdmulhq_n_s16(v1953, 29490);
+ int16x8_t v1955 = vaddq_s16(v1952, v1954);
+ int16x8_t v1956 = vqrdmulhq_n_s16(v1955, 18578);
+ int16x8_t v1957 = vaddq_s16(v1951, v1956);
+ int16x8_t v1958 = vqrdmulhq_n_s16(v1957, 16890);
+ int16x8_t v1959 = vaddq_s16(v1947, v1958);
+ int16x8_t v1960 = vqrdmulhq_n_s16(v1959, 16508);
+ int16x8_t v1961 = vaddq_s16(v1937, v1960);
+ int16x8_t v1962 = vsubq_s16(v1628, v1630);
+ int16x8_t v1963 = vsubq_s16(v1632, v1634);
+ int16x8_t v1964 = vqrdmulhq_n_s16(v1963, 29490);
+ int16x8_t v1965 = vaddq_s16(v1962, v1964);
+ int16x8_t v1966 = vsubq_s16(v1638, v1640);
+ int16x8_t v1967 = vsubq_s16(v1642, v1644);
+ int16x8_t v1968 = vqrdmulhq_n_s16(v1967, 29490);
+ int16x8_t v1969 = vaddq_s16(v1966, v1968);
+ int16x8_t v1970 = vqrdmulhq_n_s16(v1969, 18578);
+ int16x8_t v1971 = vaddq_s16(v1965, v1970);
+ int16x8_t v1972 = vsubq_s16(v1650, v1652);
+ int16x8_t v1973 = vsubq_s16(v1654, v1656);
+ int16x8_t v1974 = vqrdmulhq_n_s16(v1973, 29490);
+ int16x8_t v1975 = vaddq_s16(v1972, v1974);
+ int16x8_t v1976 = vsubq_s16(v1660, v1662);
+ int16x8_t v1977 = vsubq_s16(v1664, v1666);
+ int16x8_t v1978 = vqrdmulhq_n_s16(v1977, 29490);
+ int16x8_t v1979 = vaddq_s16(v1976, v1978);
+ int16x8_t v1980 = vqrdmulhq_n_s16(v1979, 18578);
+ int16x8_t v1981 = vaddq_s16(v1975, v1980);
+ int16x8_t v1982 = vqrdmulhq_n_s16(v1981, 16890);
+ int16x8_t v1983 = vaddq_s16(v1971, v1982);
+ int16x8_t v1984 = vsubq_s16(v1674, v1676);
+ int16x8_t v1985 = vsubq_s16(v1678, v1680);
+ int16x8_t v1986 = vqrdmulhq_n_s16(v1985, 29490);
+ int16x8_t v1987 = vaddq_s16(v1984, v1986);
+ int16x8_t v1988 = vsubq_s16(v1684, v1686);
+ int16x8_t v1989 = vsubq_s16(v1688, v1690);
+ int16x8_t v1990 = vqrdmulhq_n_s16(v1989, 29490);
+ int16x8_t v1991 = vaddq_s16(v1988, v1990);
+ int16x8_t v1992 = vqrdmulhq_n_s16(v1991, 18578);
+ int16x8_t v1993 = vaddq_s16(v1987, v1992);
+ int16x8_t v1994 = vsubq_s16(v1696, v1698);
+ int16x8_t v1995 = vsubq_s16(v1700, v1702);
+ int16x8_t v1996 = vqrdmulhq_n_s16(v1995, 29490);
+ int16x8_t v1997 = vaddq_s16(v1994, v1996);
+ int16x8_t v1998 = vsubq_s16(v1706, v1708);
+ int16x8_t v1999 = vsubq_s16(v1710, v1712);
+ int16x8_t v2000 = vqrdmulhq_n_s16(v1999, 29490);
+ int16x8_t v2001 = vaddq_s16(v1998, v2000);
+ int16x8_t v2002 = vqrdmulhq_n_s16(v2001, 18578);
+ int16x8_t v2003 = vaddq_s16(v1997, v2002);
+ int16x8_t v2004 = vqrdmulhq_n_s16(v2003, 16890);
+ int16x8_t v2005 = vaddq_s16(v1993, v2004);
+ int16x8_t v2006 = vqrdmulhq_n_s16(v2005, 16508);
+ int16x8_t v2007 = vaddq_s16(v1983, v2006);
+ int16x8_t v2008 = vqrdmulhq_n_s16(v2007, 16415);
+ int16x8_t v2009 = vaddq_s16(v1961, v2008);
+ int16x8_t v2010 = vsubq_s16(v1724, v1726);
+ int16x8_t v2011 = vsubq_s16(v1728, v1730);
+ int16x8_t v2012 = vqrdmulhq_n_s16(v2011, 29490);
+ int16x8_t v2013 = vaddq_s16(v2010, v2012);
+ int16x8_t v2014 = vsubq_s16(v1734, v1736);
+ int16x8_t v2015 = vsubq_s16(v1738, v1740);
+ int16x8_t v2016 = vqrdmulhq_n_s16(v2015, 29490);
+ int16x8_t v2017 = vaddq_s16(v2014, v2016);
+ int16x8_t v2018 = vqrdmulhq_n_s16(v2017, 18578);
+ int16x8_t v2019 = vaddq_s16(v2013, v2018);
+ int16x8_t v2020 = vsubq_s16(v1746, v1748);
+ int16x8_t v2021 = vsubq_s16(v1750, v1752);
+ int16x8_t v2022 = vqrdmulhq_n_s16(v2021, 29490);
+ int16x8_t v2023 = vaddq_s16(v2020, v2022);
+ int16x8_t v2024 = vsubq_s16(v1756, v1758);
+ int16x8_t v2025 = vsubq_s16(v1760, v1762);
+ int16x8_t v2026 = vqrdmulhq_n_s16(v2025, 29490);
+ int16x8_t v2027 = vaddq_s16(v2024, v2026);
+ int16x8_t v2028 = vqrdmulhq_n_s16(v2027, 18578);
+ int16x8_t v2029 = vaddq_s16(v2023, v2028);
+ int16x8_t v2030 = vqrdmulhq_n_s16(v2029, 16890);
+ int16x8_t v2031 = vaddq_s16(v2019, v2030);
+ int16x8_t v2032 = vsubq_s16(v1770, v1772);
+ int16x8_t v2033 = vsubq_s16(v1774, v1776);
+ int16x8_t v2034 = vqrdmulhq_n_s16(v2033, 29490);
+ int16x8_t v2035 = vaddq_s16(v2032, v2034);
+ int16x8_t v2036 = vsubq_s16(v1780, v1782);
+ int16x8_t v2037 = vsubq_s16(v1784, v1786);
+ int16x8_t v2038 = vqrdmulhq_n_s16(v2037, 29490);
+ int16x8_t v2039 = vaddq_s16(v2036, v2038);
+ int16x8_t v2040 = vqrdmulhq_n_s16(v2039, 18578);
+ int16x8_t v2041 = vaddq_s16(v2035, v2040);
+ int16x8_t v2042 = vsubq_s16(v1792, v1794);
+ int16x8_t v2043 = vsubq_s16(v1796, v1798);
+ int16x8_t v2044 = vqrdmulhq_n_s16(v2043, 29490);
+ int16x8_t v2045 = vaddq_s16(v2042, v2044);
+ int16x8_t v2046 = vsubq_s16(v1802, v1804);
+ int16x8_t v2047 = vsubq_s16(v1806, v1808);
+ int16x8_t v2048 = vqrdmulhq_n_s16(v2047, 29490);
+ int16x8_t v2049 = vaddq_s16(v2046, v2048);
+ int16x8_t v2050 = vqrdmulhq_n_s16(v2049, 18578);
+ int16x8_t v2051 = vaddq_s16(v2045, v2050);
+ int16x8_t v2052 = vqrdmulhq_n_s16(v2051, 16890);
+ int16x8_t v2053 = vaddq_s16(v2041, v2052);
+ int16x8_t v2054 = vqrdmulhq_n_s16(v2053, 16508);
+ int16x8_t v2055 = vaddq_s16(v2031, v2054);
+ int16x8_t v2056 = vsubq_s16(v1818, v1820);
+ int16x8_t v2057 = vsubq_s16(v1822, v1824);
+ int16x8_t v2058 = vqrdmulhq_n_s16(v2057, 29490);
+ int16x8_t v2059 = vaddq_s16(v2056, v2058);
+ int16x8_t v2060 = vsubq_s16(v1828, v1830);
+ int16x8_t v2061 = vsubq_s16(v1832, v1834);
+ int16x8_t v2062 = vqrdmulhq_n_s16(v2061, 29490);
+ int16x8_t v2063 = vaddq_s16(v2060, v2062);
+ int16x8_t v2064 = vqrdmulhq_n_s16(v2063, 18578);
+ int16x8_t v2065 = vaddq_s16(v2059, v2064);
+ int16x8_t v2066 = vsubq_s16(v1840, v1842);
+ int16x8_t v2067 = vsubq_s16(v1844, v1846);
+ int16x8_t v2068 = vqrdmulhq_n_s16(v2067, 29490);
+ int16x8_t v2069 = vaddq_s16(v2066, v2068);
+ int16x8_t v2070 = vsubq_s16(v1850, v1852);
+ int16x8_t v2071 = vqrdmulhq_n_s16(v2070, 18578);
+ int16x8_t v2072 = vsubq_s16(v1854, v1856);
+ int16x8_t v2073 = vqrdmulhq_n_s16(v2072, 16719);
+ int16x8_t v2074 = vaddq_s16(v2071, v2073);
+ int16x8_t v2075 = vaddq_s16(v2069, v2074);
+ int16x8_t v2076 = vqrdmulhq_n_s16(v2075, 16890);
+ int16x8_t v2077 = vaddq_s16(v2065, v2076);
+ int16x8_t v2078 = vsubq_s16(v1864, v1866);
+ int16x8_t v2079 = vsubq_s16(v1868, v1870);
+ int16x8_t v2080 = vqrdmulhq_n_s16(v2079, 29490);
+ int16x8_t v2081 = vaddq_s16(v2078, v2080);
+ int16x8_t v2082 = vsubq_s16(v1874, v1876);
+ int16x8_t v2083 = vsubq_s16(v1878, v1880);
+ int16x8_t v2084 = vqrdmulhq_n_s16(v2083, 29490);
+ int16x8_t v2085 = vaddq_s16(v2082, v2084);
+ int16x8_t v2086 = vqrdmulhq_n_s16(v2085, 18578);
+ int16x8_t v2087 = vaddq_s16(v2081, v2086);
+ int16x8_t v2088 = vsubq_s16(v1886, v1888);
+ int16x8_t v2089 = vsubq_s16(v1890, v1892);
+ int16x8_t v2090 = vqrdmulhq_n_s16(v2089, 29490);
+ int16x8_t v2091 = vaddq_s16(v2088, v2090);
+ int16x8_t v2092 = vsubq_s16(v1896, v1898);
+ int16x8_t v2093 = vsubq_s16(v1900, v1902);
+ int16x8_t v2094 = vqrdmulhq_n_s16(v2093, 29490);
+ int16x8_t v2095 = vaddq_s16(v2092, v2094);
+ int16x8_t v2096 = vqrdmulhq_n_s16(v2095, 18578);
+ int16x8_t v2097 = vaddq_s16(v2091, v2096);
+ int16x8_t v2098 = vqrdmulhq_n_s16(v2097, 16890);
+ int16x8_t v2099 = vaddq_s16(v2087, v2098);
+ int16x8_t v2100 = vqrdmulhq_n_s16(v2099, 16508);
+ int16x8_t v2101 = vaddq_s16(v2077, v2100);
+ int16x8_t v2102 = vqrdmulhq_n_s16(v2101, 16415);
+ int16x8_t v2103 = vaddq_s16(v2055, v2102);
+ int16x8_t v2104 = vqrdmulhq_n_s16(v2103, 16392);
+ int16x8_t v2105 = vaddq_s16(v2009, v2104);
+ int16x8_t v2106 = vsubq_s16(v2, v8);
+ int16x8_t v2107 = vsubq_s16(v15, v22);
+ int16x8_t v2108_tmp = vqrdmulhq_n_s16(v2107, 18446);
+ int16x8_t v2108 = vmlaq_n_s16(v2108_tmp, v2107, 2);
+ int16x8_t v2109 = vaddq_s16(v2106, v2108);
+ int16x8_t v2110 = vsubq_s16(v31, v41);
+ int16x8_t v2111 = vsubq_s16(v48, v56);
+ int16x8_t v2112_tmp = vqrdmulhq_n_s16(v2111, 18446);
+ int16x8_t v2112 = vmlaq_n_s16(v2112_tmp, v2111, 2);
+ int16x8_t v2113 = vaddq_s16(v2110, v2112);
+ int16x8_t v2114 = vqrdmulhq_n_s16(v2113, 21195);
+ int16x8_t v2115 = vaddq_s16(v2109, v2114);
+ int16x8_t v2116 = vsubq_s16(v67, v77);
+ int16x8_t v2117 = vsubq_s16(v90, v99);
+ int16x8_t v2118_tmp = vqrdmulhq_n_s16(v2117, 18446);
+ int16x8_t v2118 = vmlaq_n_s16(v2118_tmp, v2117, 2);
+ int16x8_t v2119 = vaddq_s16(v2116, v2118);
+ int16x8_t v2120 = vsubq_s16(v108, v118);
+ int16x8_t v2121 = vsubq_s16(v125, v134);
+ int16x8_t v2122_tmp = vqrdmulhq_n_s16(v2121, 18446);
+ int16x8_t v2122 = vmlaq_n_s16(v2122_tmp, v2121, 2);
+ int16x8_t v2123 = vaddq_s16(v2120, v2122);
+ int16x8_t v2124 = vqrdmulhq_n_s16(v2123, 21195);
+ int16x8_t v2125 = vaddq_s16(v2119, v2124);
+ int16x8_t v2126 = vqrdmulhq_n_s16(v2125, 17401);
+ int16x8_t v2127 = vaddq_s16(v2115, v2126);
+ int16x8_t v2128 = vsubq_s16(v147, v157);
+ int16x8_t v2129 = vsubq_s16(v170, v179);
+ int16x8_t v2130_tmp = vqrdmulhq_n_s16(v2129, 18446);
+ int16x8_t v2130 = vmlaq_n_s16(v2130_tmp, v2129, 2);
+ int16x8_t v2131 = vaddq_s16(v2128, v2130);
+ int16x8_t v2132 = vsubq_s16(v194, v212);
+ int16x8_t v2133 = vsubq_s16(v219, v229);
+ int16x8_t v2134_tmp = vqrdmulhq_n_s16(v2133, 18446);
+ int16x8_t v2134 = vmlaq_n_s16(v2134_tmp, v2133, 2);
+ int16x8_t v2135 = vaddq_s16(v2132, v2134);
+ int16x8_t v2136 = vqrdmulhq_n_s16(v2135, 21195);
+ int16x8_t v2137 = vaddq_s16(v2131, v2136);
+ int16x8_t v2138 = vsubq_s16(v240, v250);
+ int16x8_t v2139 = vsubq_s16(v263, v272);
+ int16x8_t v2140_tmp = vqrdmulhq_n_s16(v2139, 18446);
+ int16x8_t v2140 = vmlaq_n_s16(v2140_tmp, v2139, 2);
+ int16x8_t v2141 = vaddq_s16(v2138, v2140);
+ int16x8_t v2142 = vsubq_s16(v281, v291);
+ int16x8_t v2143 = vsubq_s16(v298, v308);
+ int16x8_t v2144_tmp = vqrdmulhq_n_s16(v2143, 18446);
+ int16x8_t v2144 = vmlaq_n_s16(v2144_tmp, v2143, 2);
+ int16x8_t v2145 = vaddq_s16(v2142, v2144);
+ int16x8_t v2146 = vqrdmulhq_n_s16(v2145, 21195);
+ int16x8_t v2147 = vaddq_s16(v2141, v2146);
+ int16x8_t v2148 = vqrdmulhq_n_s16(v2147, 17401);
+ int16x8_t v2149 = vaddq_s16(v2137, v2148);
+ int16x8_t v2150 = vqrdmulhq_n_s16(v2149, 16629);
+ int16x8_t v2151 = vaddq_s16(v2127, v2150);
+ int16x8_t v2152 = vsubq_s16(v323, v333);
+ int16x8_t v2153 = vsubq_s16(v346, v355);
+ int16x8_t v2154_tmp = vqrdmulhq_n_s16(v2153, 18446);
+ int16x8_t v2154 = vmlaq_n_s16(v2154_tmp, v2153, 2);
+ int16x8_t v2155 = vaddq_s16(v2152, v2154);
+ int16x8_t v2156 = vsubq_s16(v370, v388);
+ int16x8_t v2157 = vsubq_s16(v395, v405);
+ int16x8_t v2158_tmp = vqrdmulhq_n_s16(v2157, 18446);
+ int16x8_t v2158 = vmlaq_n_s16(v2158_tmp, v2157, 2);
+ int16x8_t v2159 = vaddq_s16(v2156, v2158);
+ int16x8_t v2160 = vqrdmulhq_n_s16(v2159, 21195);
+ int16x8_t v2161 = vaddq_s16(v2155, v2160);
+ int16x8_t v2162 = vsubq_s16(v422, v440);
+ int16x8_t v2163 = vsubq_s16(v465, v478);
+ int16x8_t v2164_tmp = vqrdmulhq_n_s16(v2163, 18446);
+ int16x8_t v2164 = vmlaq_n_s16(v2164_tmp, v2163, 2);
+ int16x8_t v2165 = vaddq_s16(v2162, v2164);
+ int16x8_t v2166 = vsubq_s16(v487, v497);
+ int16x8_t v2167 = vsubq_s16(v504, v515);
+ int16x8_t v2168_tmp = vqrdmulhq_n_s16(v2167, 18446);
+ int16x8_t v2168 = vmlaq_n_s16(v2168_tmp, v2167, 2);
+ int16x8_t v2169 = vaddq_s16(v2166, v2168);
+ int16x8_t v2170 = vqrdmulhq_n_s16(v2169, 21195);
+ int16x8_t v2171 = vaddq_s16(v2165, v2170);
+ int16x8_t v2172 = vqrdmulhq_n_s16(v2171, 17401);
+ int16x8_t v2173 = vaddq_s16(v2161, v2172);
+ int16x8_t v2174 = vsubq_s16(v528, v538);
+ int16x8_t v2175 = vsubq_s16(v551, v560);
+ int16x8_t v2176_tmp = vqrdmulhq_n_s16(v2175, 18446);
+ int16x8_t v2176 = vmlaq_n_s16(v2176_tmp, v2175, 2);
+ int16x8_t v2177 = vaddq_s16(v2174, v2176);
+ int16x8_t v2178 = vsubq_s16(v575, v593);
+ int16x8_t v2179 = vsubq_s16(v600, v610);
+ int16x8_t v2180_tmp = vqrdmulhq_n_s16(v2179, 18446);
+ int16x8_t v2180 = vmlaq_n_s16(v2180_tmp, v2179, 2);
+ int16x8_t v2181 = vaddq_s16(v2178, v2180);
+ int16x8_t v2182 = vqrdmulhq_n_s16(v2181, 21195);
+ int16x8_t v2183 = vaddq_s16(v2177, v2182);
+ int16x8_t v2184 = vsubq_s16(v621, v631);
+ int16x8_t v2185 = vsubq_s16(v644, v653);
+ int16x8_t v2186_tmp = vqrdmulhq_n_s16(v2185, 18446);
+ int16x8_t v2186 = vmlaq_n_s16(v2186_tmp, v2185, 2);
+ int16x8_t v2187 = vaddq_s16(v2184, v2186);
+ int16x8_t v2188 = vsubq_s16(v662, v672);
+ int16x8_t v2189 = vsubq_s16(v679, v690);
+ int16x8_t v2190_tmp = vqrdmulhq_n_s16(v2189, 18446);
+ int16x8_t v2190 = vmlaq_n_s16(v2190_tmp, v2189, 2);
+ int16x8_t v2191 = vaddq_s16(v2188, v2190);
+ int16x8_t v2192 = vqrdmulhq_n_s16(v2191, 21195);
+ int16x8_t v2193 = vaddq_s16(v2187, v2192);
+ int16x8_t v2194 = vqrdmulhq_n_s16(v2193, 17401);
+ int16x8_t v2195 = vaddq_s16(v2183, v2194);
+ int16x8_t v2196 = vqrdmulhq_n_s16(v2195, 16629);
+ int16x8_t v2197 = vaddq_s16(v2173, v2196);
+ int16x8_t v2198 = vqrdmulhq_n_s16(v2197, 16445);
+ int16x8_t v2199 = vaddq_s16(v2151, v2198);
+ int16x8_t v2200 = vsubq_s16(v707, v717);
+ int16x8_t v2201 = vsubq_s16(v730, v739);
+ int16x8_t v2202_tmp = vqrdmulhq_n_s16(v2201, 18446);
+ int16x8_t v2202 = vmlaq_n_s16(v2202_tmp, v2201, 2);
+ int16x8_t v2203 = vaddq_s16(v2200, v2202);
+ int16x8_t v2204 = vsubq_s16(v754, v772);
+ int16x8_t v2205 = vsubq_s16(v779, v789);
+ int16x8_t v2206_tmp = vqrdmulhq_n_s16(v2205, 18446);
+ int16x8_t v2206 = vmlaq_n_s16(v2206_tmp, v2205, 2);
+ int16x8_t v2207 = vaddq_s16(v2204, v2206);
+ int16x8_t v2208 = vqrdmulhq_n_s16(v2207, 21195);
+ int16x8_t v2209 = vaddq_s16(v2203, v2208);
+ int16x8_t v2210 = vsubq_s16(v806, v824);
+ int16x8_t v2211 = vsubq_s16(v849, v862);
+ int16x8_t v2212_tmp = vqrdmulhq_n_s16(v2211, 18446);
+ int16x8_t v2212 = vmlaq_n_s16(v2212_tmp, v2211, 2);
+ int16x8_t v2213 = vaddq_s16(v2210, v2212);
+ int16x8_t v2214 = vsubq_s16(v871, v881);
+ int16x8_t v2215 = vsubq_s16(v888, v899);
+ int16x8_t v2216_tmp = vqrdmulhq_n_s16(v2215, 18446);
+ int16x8_t v2216 = vmlaq_n_s16(v2216_tmp, v2215, 2);
+ int16x8_t v2217 = vaddq_s16(v2214, v2216);
+ int16x8_t v2218 = vqrdmulhq_n_s16(v2217, 21195);
+ int16x8_t v2219 = vaddq_s16(v2213, v2218);
+ int16x8_t v2220 = vqrdmulhq_n_s16(v2219, 17401);
+ int16x8_t v2221 = vaddq_s16(v2209, v2220);
+ int16x8_t v2222 = vsubq_s16(v918, v936);
+ int16x8_t v2223 = vsubq_s16(v961, v974);
+ int16x8_t v2224_tmp = vqrdmulhq_n_s16(v2223, 18446);
+ int16x8_t v2224 = vmlaq_n_s16(v2224_tmp, v2223, 2);
+ int16x8_t v2225 = vaddq_s16(v2222, v2224);
+ int16x8_t v2226 = vsubq_s16(v1001, v1035);
+ int16x8_t v2227 = vsubq_s16(v1042, v1056);
+ int16x8_t v2228_tmp = vqrdmulhq_n_s16(v2227, 18446);
+ int16x8_t v2228 = vmlaq_n_s16(v2228_tmp, v2227, 2);
+ int16x8_t v2229 = vaddq_s16(v2226, v2228);
+ int16x8_t v2230 = vqrdmulhq_n_s16(v2229, 21195);
+ int16x8_t v2231 = vaddq_s16(v2225, v2230);
+ int16x8_t v2232 = vsubq_s16(v1067, v1077);
+ int16x8_t v2233 = vsubq_s16(v1090, v1099);
+ int16x8_t v2234_tmp = vqrdmulhq_n_s16(v2233, 18446);
+ int16x8_t v2234 = vmlaq_n_s16(v2234_tmp, v2233, 2);
+ int16x8_t v2235 = vaddq_s16(v2232, v2234);
+ int16x8_t v2236 = vsubq_s16(v1108, v1118);
+ int16x8_t v2237 = vsubq_s16(v1125, v1137);
+ int16x8_t v2238_tmp = vqrdmulhq_n_s16(v2237, 18446);
+ int16x8_t v2238 = vmlaq_n_s16(v2238_tmp, v2237, 2);
+ int16x8_t v2239 = vaddq_s16(v2236, v2238);
+ int16x8_t v2240 = vqrdmulhq_n_s16(v2239, 21195);
+ int16x8_t v2241 = vaddq_s16(v2235, v2240);
+ int16x8_t v2242 = vqrdmulhq_n_s16(v2241, 17401);
+ int16x8_t v2243 = vaddq_s16(v2231, v2242);
+ int16x8_t v2244 = vqrdmulhq_n_s16(v2243, 16629);
+ int16x8_t v2245 = vaddq_s16(v2221, v2244);
+ int16x8_t v2246 = vsubq_s16(v1152, v1162);
+ int16x8_t v2247 = vsubq_s16(v1175, v1184);
+ int16x8_t v2248_tmp = vqrdmulhq_n_s16(v2247, 18446);
+ int16x8_t v2248 = vmlaq_n_s16(v2248_tmp, v2247, 2);
+ int16x8_t v2249 = vaddq_s16(v2246, v2248);
+ int16x8_t v2250 = vsubq_s16(v1199, v1217);
+ int16x8_t v2251 = vsubq_s16(v1224, v1234);
+ int16x8_t v2252_tmp = vqrdmulhq_n_s16(v2251, 18446);
+ int16x8_t v2252 = vmlaq_n_s16(v2252_tmp, v2251, 2);
+ int16x8_t v2253 = vaddq_s16(v2250, v2252);
+ int16x8_t v2254 = vqrdmulhq_n_s16(v2253, 21195);
+ int16x8_t v2255 = vaddq_s16(v2249, v2254);
+ int16x8_t v2256 = vsubq_s16(v1251, v1269);
+ int16x8_t v2257 = vsubq_s16(v1294, v1307);
+ int16x8_t v2258_tmp = vqrdmulhq_n_s16(v2257, 18446);
+ int16x8_t v2258 = vmlaq_n_s16(v2258_tmp, v2257, 2);
+ int16x8_t v2259 = vaddq_s16(v2256, v2258);
+ int16x8_t v2260 = vsubq_s16(v1316, v1326);
+ int16x8_t v2261 = vsubq_s16(v1333, v1344);
+ int16x8_t v2262_tmp = vqrdmulhq_n_s16(v2261, 18446);
+ int16x8_t v2262 = vmlaq_n_s16(v2262_tmp, v2261, 2);
+ int16x8_t v2263 = vaddq_s16(v2260, v2262);
+ int16x8_t v2264 = vqrdmulhq_n_s16(v2263, 21195);
+ int16x8_t v2265 = vaddq_s16(v2259, v2264);
+ int16x8_t v2266 = vqrdmulhq_n_s16(v2265, 17401);
+ int16x8_t v2267 = vaddq_s16(v2255, v2266);
+ int16x8_t v2268 = vsubq_s16(v1357, v1367);
+ int16x8_t v2269 = vsubq_s16(v1380, v1389);
+ int16x8_t v2270_tmp = vqrdmulhq_n_s16(v2269, 18446);
+ int16x8_t v2270 = vmlaq_n_s16(v2270_tmp, v2269, 2);
+ int16x8_t v2271 = vaddq_s16(v2268, v2270);
+ int16x8_t v2272 = vsubq_s16(v1404, v1422);
+ int16x8_t v2273 = vsubq_s16(v1429, v1439);
+ int16x8_t v2274_tmp = vqrdmulhq_n_s16(v2273, 18446);
+ int16x8_t v2274 = vmlaq_n_s16(v2274_tmp, v2273, 2);
+ int16x8_t v2275 = vaddq_s16(v2272, v2274);
+ int16x8_t v2276 = vqrdmulhq_n_s16(v2275, 21195);
+ int16x8_t v2277 = vaddq_s16(v2271, v2276);
+ int16x8_t v2278 = vsubq_s16(v1450, v1460);
+ int16x8_t v2279 = vsubq_s16(v1473, v1482);
+ int16x8_t v2280_tmp = vqrdmulhq_n_s16(v2279, 18446);
+ int16x8_t v2280 = vmlaq_n_s16(v2280_tmp, v2279, 2);
+ int16x8_t v2281 = vaddq_s16(v2278, v2280);
+ int16x8_t v2282 = vsubq_s16(v1491, v1501);
+ int16x8_t v2283 = vsubq_s16(v1508, v1520);
+ int16x8_t v2284_tmp = vqrdmulhq_n_s16(v2283, 18446);
+ int16x8_t v2284 = vmlaq_n_s16(v2284_tmp, v2283, 2);
+ int16x8_t v2285 = vaddq_s16(v2282, v2284);
+ int16x8_t v2286 = vqrdmulhq_n_s16(v2285, 21195);
+ int16x8_t v2287 = vaddq_s16(v2281, v2286);
+ int16x8_t v2288 = vqrdmulhq_n_s16(v2287, 17401);
+ int16x8_t v2289 = vaddq_s16(v2277, v2288);
+ int16x8_t v2290 = vqrdmulhq_n_s16(v2289, 16629);
+ int16x8_t v2291 = vaddq_s16(v2267, v2290);
+ int16x8_t v2292 = vqrdmulhq_n_s16(v2291, 16445);
+ int16x8_t v2293 = vaddq_s16(v2245, v2292);
+ int16x8_t v2294 = vqrdmulhq_n_s16(v2293, 16399);
+ int16x8_t v2295 = vaddq_s16(v2199, v2294);
+ int16x8_t v2296 = vsubq_s16(v2106, v2108);
+ int16x8_t v2297 = vsubq_s16(v2110, v2112);
+ int16x8_t v2298 = vqrdmulhq_n_s16(v2297, 25826);
+ int16x8_t v2299 = vaddq_s16(v2296, v2298);
+ int16x8_t v2300 = vsubq_s16(v2116, v2118);
+ int16x8_t v2301 = vsubq_s16(v2120, v2122);
+ int16x8_t v2302 = vqrdmulhq_n_s16(v2301, 25826);
+ int16x8_t v2303 = vaddq_s16(v2300, v2302);
+ int16x8_t v2304 = vqrdmulhq_n_s16(v2303, 18124);
+ int16x8_t v2305 = vaddq_s16(v2299, v2304);
+ int16x8_t v2306 = vsubq_s16(v2128, v2130);
+ int16x8_t v2307 = vsubq_s16(v2132, v2134);
+ int16x8_t v2308 = vqrdmulhq_n_s16(v2307, 25826);
+ int16x8_t v2309 = vaddq_s16(v2306, v2308);
+ int16x8_t v2310 = vsubq_s16(v2138, v2140);
+ int16x8_t v2311 = vsubq_s16(v2142, v2144);
+ int16x8_t v2312 = vqrdmulhq_n_s16(v2311, 25826);
+ int16x8_t v2313 = vaddq_s16(v2310, v2312);
+ int16x8_t v2314 = vqrdmulhq_n_s16(v2313, 18124);
+ int16x8_t v2315 = vaddq_s16(v2309, v2314);
+ int16x8_t v2316 = vqrdmulhq_n_s16(v2315, 16792);
+ int16x8_t v2317 = vaddq_s16(v2305, v2316);
+ int16x8_t v2318 = vsubq_s16(v2152, v2154);
+ int16x8_t v2319 = vsubq_s16(v2156, v2158);
+ int16x8_t v2320 = vqrdmulhq_n_s16(v2319, 25826);
+ int16x8_t v2321 = vaddq_s16(v2318, v2320);
+ int16x8_t v2322 = vsubq_s16(v2162, v2164);
+ int16x8_t v2323 = vsubq_s16(v2166, v2168);
+ int16x8_t v2324 = vqrdmulhq_n_s16(v2323, 25826);
+ int16x8_t v2325 = vaddq_s16(v2322, v2324);
+ int16x8_t v2326 = vqrdmulhq_n_s16(v2325, 18124);
+ int16x8_t v2327 = vaddq_s16(v2321, v2326);
+ int16x8_t v2328 = vsubq_s16(v2174, v2176);
+ int16x8_t v2329 = vsubq_s16(v2178, v2180);
+ int16x8_t v2330 = vqrdmulhq_n_s16(v2329, 25826);
+ int16x8_t v2331 = vaddq_s16(v2328, v2330);
+ int16x8_t v2332 = vsubq_s16(v2184, v2186);
+ int16x8_t v2333 = vsubq_s16(v2188, v2190);
+ int16x8_t v2334 = vqrdmulhq_n_s16(v2333, 25826);
+ int16x8_t v2335 = vaddq_s16(v2332, v2334);
+ int16x8_t v2336 = vqrdmulhq_n_s16(v2335, 18124);
+ int16x8_t v2337 = vaddq_s16(v2331, v2336);
+ int16x8_t v2338 = vqrdmulhq_n_s16(v2337, 16792);
+ int16x8_t v2339 = vaddq_s16(v2327, v2338);
+ int16x8_t v2340 = vqrdmulhq_n_s16(v2339, 16484);
+ int16x8_t v2341 = vaddq_s16(v2317, v2340);
+ int16x8_t v2342 = vsubq_s16(v2200, v2202);
+ int16x8_t v2343 = vsubq_s16(v2204, v2206);
+ int16x8_t v2344 = vqrdmulhq_n_s16(v2343, 25826);
+ int16x8_t v2345 = vaddq_s16(v2342, v2344);
+ int16x8_t v2346 = vsubq_s16(v2210, v2212);
+ int16x8_t v2347 = vsubq_s16(v2214, v2216);
+ int16x8_t v2348 = vqrdmulhq_n_s16(v2347, 25826);
+ int16x8_t v2349 = vaddq_s16(v2346, v2348);
+ int16x8_t v2350 = vqrdmulhq_n_s16(v2349, 18124);
+ int16x8_t v2351 = vaddq_s16(v2345, v2350);
+ int16x8_t v2352 = vsubq_s16(v2222, v2224);
+ int16x8_t v2353 = vsubq_s16(v2226, v2228);
+ int16x8_t v2354 = vqrdmulhq_n_s16(v2353, 25826);
+ int16x8_t v2355 = vaddq_s16(v2352, v2354);
+ int16x8_t v2356 = vsubq_s16(v2232, v2234);
+ int16x8_t v2357 = vsubq_s16(v2236, v2238);
+ int16x8_t v2358 = vqrdmulhq_n_s16(v2357, 25826);
+ int16x8_t v2359 = vaddq_s16(v2356, v2358);
+ int16x8_t v2360 = vqrdmulhq_n_s16(v2359, 18124);
+ int16x8_t v2361 = vaddq_s16(v2355, v2360);
+ int16x8_t v2362 = vqrdmulhq_n_s16(v2361, 16792);
+ int16x8_t v2363 = vaddq_s16(v2351, v2362);
+ int16x8_t v2364 = vsubq_s16(v2246, v2248);
+ int16x8_t v2365 = vsubq_s16(v2250, v2252);
+ int16x8_t v2366 = vqrdmulhq_n_s16(v2365, 25826);
+ int16x8_t v2367 = vaddq_s16(v2364, v2366);
+ int16x8_t v2368 = vsubq_s16(v2256, v2258);
+ int16x8_t v2369 = vsubq_s16(v2260, v2262);
+ int16x8_t v2370 = vqrdmulhq_n_s16(v2369, 25826);
+ int16x8_t v2371 = vaddq_s16(v2368, v2370);
+ int16x8_t v2372 = vqrdmulhq_n_s16(v2371, 18124);
+ int16x8_t v2373 = vaddq_s16(v2367, v2372);
+ int16x8_t v2374 = vsubq_s16(v2268, v2270);
+ int16x8_t v2375 = vsubq_s16(v2272, v2274);
+ int16x8_t v2376 = vqrdmulhq_n_s16(v2375, 25826);
+ int16x8_t v2377 = vaddq_s16(v2374, v2376);
+ int16x8_t v2378 = vsubq_s16(v2278, v2280);
+ int16x8_t v2379 = vsubq_s16(v2282, v2284);
+ int16x8_t v2380 = vqrdmulhq_n_s16(v2379, 25826);
+ int16x8_t v2381 = vaddq_s16(v2378, v2380);
+ int16x8_t v2382 = vqrdmulhq_n_s16(v2381, 18124);
+ int16x8_t v2383 = vaddq_s16(v2377, v2382);
+ int16x8_t v2384 = vqrdmulhq_n_s16(v2383, 16792);
+ int16x8_t v2385 = vaddq_s16(v2373, v2384);
+ int16x8_t v2386 = vqrdmulhq_n_s16(v2385, 16484);
+ int16x8_t v2387 = vaddq_s16(v2363, v2386);
+ int16x8_t v2388 = vqrdmulhq_n_s16(v2387, 16409);
+ int16x8_t v2389 = vaddq_s16(v2341, v2388);
+ int16x8_t v2390 = vsubq_s16(v1916, v1918);
+ int16x8_t v2391 = vsubq_s16(v1920, v1922);
+ int16x8_t v2392_tmp = vqrdmulhq_n_s16(v2391, 1988);
+ int16x8_t v2392 = vaddq_s16(v2392_tmp, v2391);
+ int16x8_t v2393 = vaddq_s16(v2390, v2392);
+ int16x8_t v2394 = vsubq_s16(v1926, v1928);
+ int16x8_t v2395 = vsubq_s16(v1930, v1932);
+ int16x8_t v2396_tmp = vqrdmulhq_n_s16(v2395, 1988);
+ int16x8_t v2396 = vaddq_s16(v2396_tmp, v2395);
+ int16x8_t v2397 = vaddq_s16(v2394, v2396);
+ int16x8_t v2398 = vqrdmulhq_n_s16(v2397, 19102);
+ int16x8_t v2399 = vaddq_s16(v2393, v2398);
+ int16x8_t v2400 = vsubq_s16(v1938, v1940);
+ int16x8_t v2401 = vsubq_s16(v1942, v1944);
+ int16x8_t v2402_tmp = vqrdmulhq_n_s16(v2401, 1988);
+ int16x8_t v2402 = vaddq_s16(v2402_tmp, v2401);
+ int16x8_t v2403 = vaddq_s16(v2400, v2402);
+ int16x8_t v2404 = vsubq_s16(v1948, v1950);
+ int16x8_t v2405 = vsubq_s16(v1952, v1954);
+ int16x8_t v2406_tmp = vqrdmulhq_n_s16(v2405, 1988);
+ int16x8_t v2406 = vaddq_s16(v2406_tmp, v2405);
+ int16x8_t v2407 = vaddq_s16(v2404, v2406);
+ int16x8_t v2408 = vqrdmulhq_n_s16(v2407, 19102);
+ int16x8_t v2409 = vaddq_s16(v2403, v2408);
+ int16x8_t v2410 = vqrdmulhq_n_s16(v2409, 17000);
+ int16x8_t v2411 = vaddq_s16(v2399, v2410);
+ int16x8_t v2412 = vsubq_s16(v1962, v1964);
+ int16x8_t v2413 = vsubq_s16(v1966, v1968);
+ int16x8_t v2414_tmp = vqrdmulhq_n_s16(v2413, 1988);
+ int16x8_t v2414 = vaddq_s16(v2414_tmp, v2413);
+ int16x8_t v2415 = vaddq_s16(v2412, v2414);
+ int16x8_t v2416 = vsubq_s16(v1972, v1974);
+ int16x8_t v2417 = vsubq_s16(v1976, v1978);
+ int16x8_t v2418_tmp = vqrdmulhq_n_s16(v2417, 1988);
+ int16x8_t v2418 = vaddq_s16(v2418_tmp, v2417);
+ int16x8_t v2419 = vaddq_s16(v2416, v2418);
+ int16x8_t v2420 = vqrdmulhq_n_s16(v2419, 19102);
+ int16x8_t v2421 = vaddq_s16(v2415, v2420);
+ int16x8_t v2422 = vsubq_s16(v1984, v1986);
+ int16x8_t v2423 = vsubq_s16(v1988, v1990);
+ int16x8_t v2424_tmp = vqrdmulhq_n_s16(v2423, 1988);
+ int16x8_t v2424 = vaddq_s16(v2424_tmp, v2423);
+ int16x8_t v2425 = vaddq_s16(v2422, v2424);
+ int16x8_t v2426 = vsubq_s16(v1994, v1996);
+ int16x8_t v2427 = vsubq_s16(v1998, v2000);
+ int16x8_t v2428_tmp = vqrdmulhq_n_s16(v2427, 1988);
+ int16x8_t v2428 = vaddq_s16(v2428_tmp, v2427);
+ int16x8_t v2429 = vaddq_s16(v2426, v2428);
+ int16x8_t v2430 = vqrdmulhq_n_s16(v2429, 19102);
+ int16x8_t v2431 = vaddq_s16(v2425, v2430);
+ int16x8_t v2432 = vqrdmulhq_n_s16(v2431, 17000);
+ int16x8_t v2433 = vaddq_s16(v2421, v2432);
+ int16x8_t v2434 = vqrdmulhq_n_s16(v2433, 16534);
+ int16x8_t v2435 = vaddq_s16(v2411, v2434);
+ int16x8_t v2436 = vsubq_s16(v2010, v2012);
+ int16x8_t v2437 = vsubq_s16(v2014, v2016);
+ int16x8_t v2438_tmp = vqrdmulhq_n_s16(v2437, 1988);
+ int16x8_t v2438 = vaddq_s16(v2438_tmp, v2437);
+ int16x8_t v2439 = vaddq_s16(v2436, v2438);
+ int16x8_t v2440 = vsubq_s16(v2020, v2022);
+ int16x8_t v2441 = vsubq_s16(v2024, v2026);
+ int16x8_t v2442_tmp = vqrdmulhq_n_s16(v2441, 1988);
+ int16x8_t v2442 = vaddq_s16(v2442_tmp, v2441);
+ int16x8_t v2443 = vaddq_s16(v2440, v2442);
+ int16x8_t v2444 = vqrdmulhq_n_s16(v2443, 19102);
+ int16x8_t v2445 = vaddq_s16(v2439, v2444);
+ int16x8_t v2446 = vsubq_s16(v2032, v2034);
+ int16x8_t v2447 = vsubq_s16(v2036, v2038);
+ int16x8_t v2448_tmp = vqrdmulhq_n_s16(v2447, 1988);
+ int16x8_t v2448 = vaddq_s16(v2448_tmp, v2447);
+ int16x8_t v2449 = vaddq_s16(v2446, v2448);
+ int16x8_t v2450 = vsubq_s16(v2042, v2044);
+ int16x8_t v2451 = vsubq_s16(v2046, v2048);
+ int16x8_t v2452_tmp = vqrdmulhq_n_s16(v2451, 1988);
+ int16x8_t v2452 = vaddq_s16(v2452_tmp, v2451);
+ int16x8_t v2453 = vaddq_s16(v2450, v2452);
+ int16x8_t v2454 = vqrdmulhq_n_s16(v2453, 19102);
+ int16x8_t v2455 = vaddq_s16(v2449, v2454);
+ int16x8_t v2456 = vqrdmulhq_n_s16(v2455, 17000);
+ int16x8_t v2457 = vaddq_s16(v2445, v2456);
+ int16x8_t v2458 = vsubq_s16(v2056, v2058);
+ int16x8_t v2459 = vsubq_s16(v2060, v2062);
+ int16x8_t v2460_tmp = vqrdmulhq_n_s16(v2459, 1988);
+ int16x8_t v2460 = vaddq_s16(v2460_tmp, v2459);
+ int16x8_t v2461 = vaddq_s16(v2458, v2460);
+ int16x8_t v2462 = vsubq_s16(v2066, v2068);
+ int16x8_t v2463 = vqrdmulhq_n_s16(v2072, 29490);
+ int16x8_t v2464 = vsubq_s16(v2070, v2463);
+ int16x8_t v2465_tmp = vqrdmulhq_n_s16(v2464, 1988);
+ int16x8_t v2465 = vaddq_s16(v2465_tmp, v2464);
+ int16x8_t v2466 = vaddq_s16(v2462, v2465);
+ int16x8_t v2467 = vqrdmulhq_n_s16(v2466, 19102);
+ int16x8_t v2468 = vaddq_s16(v2461, v2467);
+ int16x8_t v2469 = vsubq_s16(v2078, v2080);
+ int16x8_t v2470 = vsubq_s16(v2082, v2084);
+ int16x8_t v2471_tmp = vqrdmulhq_n_s16(v2470, 1988);
+ int16x8_t v2471 = vaddq_s16(v2471_tmp, v2470);
+ int16x8_t v2472 = vaddq_s16(v2469, v2471);
+ int16x8_t v2473 = vsubq_s16(v2088, v2090);
+ int16x8_t v2474 = vsubq_s16(v2092, v2094);
+ int16x8_t v2475_tmp = vqrdmulhq_n_s16(v2474, 1988);
+ int16x8_t v2475 = vaddq_s16(v2475_tmp, v2474);
+ int16x8_t v2476 = vaddq_s16(v2473, v2475);
+ int16x8_t v2477 = vqrdmulhq_n_s16(v2476, 19102);
+ int16x8_t v2478 = vaddq_s16(v2472, v2477);
+ int16x8_t v2479 = vqrdmulhq_n_s16(v2478, 17000);
+ int16x8_t v2480 = vaddq_s16(v2468, v2479);
+ int16x8_t v2481 = vqrdmulhq_n_s16(v2480, 16534);
+ int16x8_t v2482 = vaddq_s16(v2457, v2481);
+ int16x8_t v2483 = vqrdmulhq_n_s16(v2482, 16421);
+ int16x8_t v2484 = vaddq_s16(v2435, v2483);
+ int16x8_t v2485 = vsubq_s16(v1537, v1542);
+ int16x8_t v2486 = vsubq_s16(v1547, v1552);
+ int16x8_t v2487_tmp = vqrdmulhq_n_s16(v2486, 23673);
+ int16x8_t v2487 = vaddq_s16(v2487_tmp, v2486);
+ int16x8_t v2488 = vaddq_s16(v2485, v2487);
+ int16x8_t v2489 = vsubq_s16(v1559, v1564);
+ int16x8_t v2490 = vsubq_s16(v1569, v1574);
+ int16x8_t v2491_tmp = vqrdmulhq_n_s16(v2490, 23673);
+ int16x8_t v2491 = vaddq_s16(v2491_tmp, v2490);
+ int16x8_t v2492 = vaddq_s16(v2489, v2491);
+ int16x8_t v2493 = vqrdmulhq_n_s16(v2492, 20398);
+ int16x8_t v2494 = vaddq_s16(v2488, v2493);
+ int16x8_t v2495 = vsubq_s16(v1583, v1588);
+ int16x8_t v2496 = vsubq_s16(v1593, v1598);
+ int16x8_t v2497_tmp = vqrdmulhq_n_s16(v2496, 23673);
+ int16x8_t v2497 = vaddq_s16(v2497_tmp, v2496);
+ int16x8_t v2498 = vaddq_s16(v2495, v2497);
+ int16x8_t v2499 = vsubq_s16(v1605, v1610);
+ int16x8_t v2500 = vsubq_s16(v1615, v1620);
+ int16x8_t v2501_tmp = vqrdmulhq_n_s16(v2500, 23673);
+ int16x8_t v2501 = vaddq_s16(v2501_tmp, v2500);
+ int16x8_t v2502 = vaddq_s16(v2499, v2501);
+ int16x8_t v2503 = vqrdmulhq_n_s16(v2502, 20398);
+ int16x8_t v2504 = vaddq_s16(v2498, v2503);
+ int16x8_t v2505 = vqrdmulhq_n_s16(v2504, 17255);
+ int16x8_t v2506 = vaddq_s16(v2494, v2505);
+ int16x8_t v2507 = vsubq_s16(v1631, v1636);
+ int16x8_t v2508 = vsubq_s16(v1641, v1646);
+ int16x8_t v2509_tmp = vqrdmulhq_n_s16(v2508, 23673);
+ int16x8_t v2509 = vaddq_s16(v2509_tmp, v2508);
+ int16x8_t v2510 = vaddq_s16(v2507, v2509);
+ int16x8_t v2511 = vsubq_s16(v1653, v1658);
+ int16x8_t v2512 = vsubq_s16(v1663, v1668);
+ int16x8_t v2513_tmp = vqrdmulhq_n_s16(v2512, 23673);
+ int16x8_t v2513 = vaddq_s16(v2513_tmp, v2512);
+ int16x8_t v2514 = vaddq_s16(v2511, v2513);
+ int16x8_t v2515 = vqrdmulhq_n_s16(v2514, 20398);
+ int16x8_t v2516 = vaddq_s16(v2510, v2515);
+ int16x8_t v2517 = vsubq_s16(v1677, v1682);
+ int16x8_t v2518 = vsubq_s16(v1687, v1692);
+ int16x8_t v2519_tmp = vqrdmulhq_n_s16(v2518, 23673);
+ int16x8_t v2519 = vaddq_s16(v2519_tmp, v2518);
+ int16x8_t v2520 = vaddq_s16(v2517, v2519);
+ int16x8_t v2521 = vsubq_s16(v1699, v1704);
+ int16x8_t v2522 = vsubq_s16(v1709, v1714);
+ int16x8_t v2523_tmp = vqrdmulhq_n_s16(v2522, 23673);
+ int16x8_t v2523 = vaddq_s16(v2523_tmp, v2522);
+ int16x8_t v2524 = vaddq_s16(v2521, v2523);
+ int16x8_t v2525 = vqrdmulhq_n_s16(v2524, 20398);
+ int16x8_t v2526 = vaddq_s16(v2520, v2525);
+ int16x8_t v2527 = vqrdmulhq_n_s16(v2526, 17255);
+ int16x8_t v2528 = vaddq_s16(v2516, v2527);
+ int16x8_t v2529 = vqrdmulhq_n_s16(v2528, 16595);
+ int16x8_t v2530 = vaddq_s16(v2506, v2529);
+ int16x8_t v2531 = vsubq_s16(v1727, v1732);
+ int16x8_t v2532 = vsubq_s16(v1737, v1742);
+ int16x8_t v2533_tmp = vqrdmulhq_n_s16(v2532, 23673);
+ int16x8_t v2533 = vaddq_s16(v2533_tmp, v2532);
+ int16x8_t v2534 = vaddq_s16(v2531, v2533);
+ int16x8_t v2535 = vsubq_s16(v1749, v1754);
+ int16x8_t v2536 = vsubq_s16(v1759, v1764);
+ int16x8_t v2537_tmp = vqrdmulhq_n_s16(v2536, 23673);
+ int16x8_t v2537 = vaddq_s16(v2537_tmp, v2536);
+ int16x8_t v2538 = vaddq_s16(v2535, v2537);
+ int16x8_t v2539 = vqrdmulhq_n_s16(v2538, 20398);
+ int16x8_t v2540 = vaddq_s16(v2534, v2539);
+ int16x8_t v2541 = vsubq_s16(v1773, v1778);
+ int16x8_t v2542 = vsubq_s16(v1783, v1788);
+ int16x8_t v2543_tmp = vqrdmulhq_n_s16(v2542, 23673);
+ int16x8_t v2543 = vaddq_s16(v2543_tmp, v2542);
+ int16x8_t v2544 = vaddq_s16(v2541, v2543);
+ int16x8_t v2545 = vsubq_s16(v1795, v1800);
+ int16x8_t v2546 = vsubq_s16(v1805, v1810);
+ int16x8_t v2547_tmp = vqrdmulhq_n_s16(v2546, 23673);
+ int16x8_t v2547 = vaddq_s16(v2547_tmp, v2546);
+ int16x8_t v2548 = vaddq_s16(v2545, v2547);
+ int16x8_t v2549 = vqrdmulhq_n_s16(v2548, 20398);
+ int16x8_t v2550 = vaddq_s16(v2544, v2549);
+ int16x8_t v2551 = vqrdmulhq_n_s16(v2550, 17255);
+ int16x8_t v2552 = vaddq_s16(v2540, v2551);
+ int16x8_t v2553 = vsubq_s16(v1821, v1826);
+ int16x8_t v2554 = vsubq_s16(v1831, v1836);
+ int16x8_t v2555_tmp = vqrdmulhq_n_s16(v2554, 23673);
+ int16x8_t v2555 = vaddq_s16(v2555_tmp, v2554);
+ int16x8_t v2556 = vaddq_s16(v2553, v2555);
+ int16x8_t v2557 = vsubq_s16(v1843, v1848);
+ int16x8_t v2558 = vsubq_s16(v1853, v1858);
+ int16x8_t v2559_tmp = vqrdmulhq_n_s16(v2558, 23673);
+ int16x8_t v2559 = vaddq_s16(v2559_tmp, v2558);
+ int16x8_t v2560 = vaddq_s16(v2557, v2559);
+ int16x8_t v2561 = vqrdmulhq_n_s16(v2560, 20398);
+ int16x8_t v2562 = vaddq_s16(v2556, v2561);
+ int16x8_t v2563 = vsubq_s16(v1867, v1872);
+ int16x8_t v2564 = vsubq_s16(v1877, v1882);
+ int16x8_t v2565_tmp = vqrdmulhq_n_s16(v2564, 23673);
+ int16x8_t v2565 = vaddq_s16(v2565_tmp, v2564);
+ int16x8_t v2566 = vaddq_s16(v2563, v2565);
+ int16x8_t v2567 = vsubq_s16(v1889, v1894);
+ int16x8_t v2568 = vsubq_s16(v1899, v1904);
+ int16x8_t v2569_tmp = vqrdmulhq_n_s16(v2568, 23673);
+ int16x8_t v2569 = vaddq_s16(v2569_tmp, v2568);
+ int16x8_t v2570 = vaddq_s16(v2567, v2569);
+ int16x8_t v2571 = vqrdmulhq_n_s16(v2570, 20398);
+ int16x8_t v2572 = vaddq_s16(v2566, v2571);
+ int16x8_t v2573 = vqrdmulhq_n_s16(v2572, 17255);
+ int16x8_t v2574 = vaddq_s16(v2562, v2573);
+ int16x8_t v2575 = vqrdmulhq_n_s16(v2574, 16595);
+ int16x8_t v2576 = vaddq_s16(v2552, v2575);
+ int16x8_t v2577 = vqrdmulhq_n_s16(v2576, 16436);
+ int16x8_t v2578 = vaddq_s16(v2530, v2577);
+ int16x8_t v2579 = vsubq_s16(v9, v24);
+ int16x8_t v2580 = vsubq_s16(v42, v58);
+ int16x8_t v2581_tmp = vqrdmulhq_n_s16(v2580, 3314);
+ int16x8_t v2581 = vmlaq_n_s16(v2581_tmp, v2580, 5);
+ int16x8_t v2582 = vaddq_s16(v2579, v2581);
+ int16x8_t v2583 = vsubq_s16(v78, v101);
+ int16x8_t v2584 = vsubq_s16(v119, v136);
+ int16x8_t v2585_tmp = vqrdmulhq_n_s16(v2584, 3314);
+ int16x8_t v2585 = vmlaq_n_s16(v2585_tmp, v2584, 5);
+ int16x8_t v2586 = vaddq_s16(v2583, v2585);
+ int16x8_t v2587 = vqrdmulhq_n_s16(v2586, 22112);
+ int16x8_t v2588 = vaddq_s16(v2582, v2587);
+ int16x8_t v2589 = vsubq_s16(v158, v181);
+ int16x8_t v2590 = vsubq_s16(v213, v231);
+ int16x8_t v2591_tmp = vqrdmulhq_n_s16(v2590, 3314);
+ int16x8_t v2591 = vmlaq_n_s16(v2591_tmp, v2590, 5);
+ int16x8_t v2592 = vaddq_s16(v2589, v2591);
+ int16x8_t v2593 = vsubq_s16(v251, v274);
+ int16x8_t v2594 = vsubq_s16(v292, v310);
+ int16x8_t v2595_tmp = vqrdmulhq_n_s16(v2594, 3314);
+ int16x8_t v2595 = vmlaq_n_s16(v2595_tmp, v2594, 5);
+ int16x8_t v2596 = vaddq_s16(v2593, v2595);
+ int16x8_t v2597 = vqrdmulhq_n_s16(v2596, 22112);
+ int16x8_t v2598 = vaddq_s16(v2592, v2597);
+ int16x8_t v2599 = vqrdmulhq_n_s16(v2598, 17561);
+ int16x8_t v2600 = vaddq_s16(v2588, v2599);
+ int16x8_t v2601 = vsubq_s16(v334, v357);
+ int16x8_t v2602 = vsubq_s16(v389, v407);
+ int16x8_t v2603_tmp = vqrdmulhq_n_s16(v2602, 3314);
+ int16x8_t v2603 = vmlaq_n_s16(v2603_tmp, v2602, 5);
+ int16x8_t v2604 = vaddq_s16(v2601, v2603);
+ int16x8_t v2605 = vsubq_s16(v441, v480);
+ int16x8_t v2606 = vsubq_s16(v498, v517);
+ int16x8_t v2607_tmp = vqrdmulhq_n_s16(v2606, 3314);
+ int16x8_t v2607 = vmlaq_n_s16(v2607_tmp, v2606, 5);
+ int16x8_t v2608 = vaddq_s16(v2605, v2607);
+ int16x8_t v2609 = vqrdmulhq_n_s16(v2608, 22112);
+ int16x8_t v2610 = vaddq_s16(v2604, v2609);
+ int16x8_t v2611 = vsubq_s16(v539, v562);
+ int16x8_t v2612 = vsubq_s16(v594, v612);
+ int16x8_t v2613_tmp = vqrdmulhq_n_s16(v2612, 3314);
+ int16x8_t v2613 = vmlaq_n_s16(v2613_tmp, v2612, 5);
+ int16x8_t v2614 = vaddq_s16(v2611, v2613);
+ int16x8_t v2615 = vsubq_s16(v632, v655);
+ int16x8_t v2616 = vsubq_s16(v673, v692);
+ int16x8_t v2617_tmp = vqrdmulhq_n_s16(v2616, 3314);
+ int16x8_t v2617 = vmlaq_n_s16(v2617_tmp, v2616, 5);
+ int16x8_t v2618 = vaddq_s16(v2615, v2617);
+ int16x8_t v2619 = vqrdmulhq_n_s16(v2618, 22112);
+ int16x8_t v2620 = vaddq_s16(v2614, v2619);
+ int16x8_t v2621 = vqrdmulhq_n_s16(v2620, 17561);
+ int16x8_t v2622 = vaddq_s16(v2610, v2621);
+ int16x8_t v2623 = vqrdmulhq_n_s16(v2622, 16666);
+ int16x8_t v2624 = vaddq_s16(v2600, v2623);
+ int16x8_t v2625 = vsubq_s16(v718, v741);
+ int16x8_t v2626 = vsubq_s16(v773, v791);
+ int16x8_t v2627_tmp = vqrdmulhq_n_s16(v2626, 3314);
+ int16x8_t v2627 = vmlaq_n_s16(v2627_tmp, v2626, 5);
+ int16x8_t v2628 = vaddq_s16(v2625, v2627);
+ int16x8_t v2629 = vsubq_s16(v825, v864);
+ int16x8_t v2630 = vsubq_s16(v882, v901);
+ int16x8_t v2631_tmp = vqrdmulhq_n_s16(v2630, 3314);
+ int16x8_t v2631 = vmlaq_n_s16(v2631_tmp, v2630, 5);
+ int16x8_t v2632 = vaddq_s16(v2629, v2631);
+ int16x8_t v2633 = vqrdmulhq_n_s16(v2632, 22112);
+ int16x8_t v2634 = vaddq_s16(v2628, v2633);
+ int16x8_t v2635 = vsubq_s16(v937, v976);
+ int16x8_t v2636 = vsubq_s16(v1036, v1058);
+ int16x8_t v2637_tmp = vqrdmulhq_n_s16(v2636, 3314);
+ int16x8_t v2637 = vmlaq_n_s16(v2637_tmp, v2636, 5);
+ int16x8_t v2638 = vaddq_s16(v2635, v2637);
+ int16x8_t v2639 = vsubq_s16(v1078, v1101);
+ int16x8_t v2640 = vsubq_s16(v1119, v1139);
+ int16x8_t v2641_tmp = vqrdmulhq_n_s16(v2640, 3314);
+ int16x8_t v2641 = vmlaq_n_s16(v2641_tmp, v2640, 5);
+ int16x8_t v2642 = vaddq_s16(v2639, v2641);
+ int16x8_t v2643 = vqrdmulhq_n_s16(v2642, 22112);
+ int16x8_t v2644 = vaddq_s16(v2638, v2643);
+ int16x8_t v2645 = vqrdmulhq_n_s16(v2644, 17561);
+ int16x8_t v2646 = vaddq_s16(v2634, v2645);
+ int16x8_t v2647 = vsubq_s16(v1163, v1186);
+ int16x8_t v2648 = vsubq_s16(v1218, v1236);
+ int16x8_t v2649_tmp = vqrdmulhq_n_s16(v2648, 3314);
+ int16x8_t v2649 = vmlaq_n_s16(v2649_tmp, v2648, 5);
+ int16x8_t v2650 = vaddq_s16(v2647, v2649);
+ int16x8_t v2651 = vsubq_s16(v1270, v1309);
+ int16x8_t v2652 = vsubq_s16(v1327, v1346);
+ int16x8_t v2653_tmp = vqrdmulhq_n_s16(v2652, 3314);
+ int16x8_t v2653 = vmlaq_n_s16(v2653_tmp, v2652, 5);
+ int16x8_t v2654 = vaddq_s16(v2651, v2653);
+ int16x8_t v2655 = vqrdmulhq_n_s16(v2654, 22112);
+ int16x8_t v2656 = vaddq_s16(v2650, v2655);
+ int16x8_t v2657 = vsubq_s16(v1368, v1391);
+ int16x8_t v2658 = vsubq_s16(v1423, v1441);
+ int16x8_t v2659_tmp = vqrdmulhq_n_s16(v2658, 3314);
+ int16x8_t v2659 = vmlaq_n_s16(v2659_tmp, v2658, 5);
+ int16x8_t v2660 = vaddq_s16(v2657, v2659);
+ int16x8_t v2661 = vsubq_s16(v1461, v1484);
+ int16x8_t v2662 = vsubq_s16(v1502, v1522);
+ int16x8_t v2663_tmp = vqrdmulhq_n_s16(v2662, 3314);
+ int16x8_t v2663 = vmlaq_n_s16(v2663_tmp, v2662, 5);
+ int16x8_t v2664 = vaddq_s16(v2661, v2663);
+ int16x8_t v2665 = vqrdmulhq_n_s16(v2664, 22112);
+ int16x8_t v2666 = vaddq_s16(v2660, v2665);
+ int16x8_t v2667 = vqrdmulhq_n_s16(v2666, 17561);
+ int16x8_t v2668 = vaddq_s16(v2656, v2667);
+ int16x8_t v2669 = vqrdmulhq_n_s16(v2668, 16666);
+ int16x8_t v2670 = vaddq_s16(v2646, v2669);
+ int16x8_t v2671 = vqrdmulhq_n_s16(v2670, 16454);
+ int16x8_t v2672 = vaddq_s16(v2624, v2671);
+ int16x8_t v2673 = vsubq_s16(v2579, v2581);
+ int16x8_t v2674 = vsubq_s16(v2583, v2585);
+ int16x8_t v2675 = vqrdmulhq_n_s16(v2674, 24397);
+ int16x8_t v2676 = vaddq_s16(v2673, v2675);
+ int16x8_t v2677 = vsubq_s16(v2589, v2591);
+ int16x8_t v2678 = vsubq_s16(v2593, v2595);
+ int16x8_t v2679 = vqrdmulhq_n_s16(v2678, 24397);
+ int16x8_t v2680 = vaddq_s16(v2677, v2679);
+ int16x8_t v2681 = vqrdmulhq_n_s16(v2680, 17921);
+ int16x8_t v2682 = vaddq_s16(v2676, v2681);
+ int16x8_t v2683 = vsubq_s16(v2601, v2603);
+ int16x8_t v2684 = vsubq_s16(v2605, v2607);
+ int16x8_t v2685 = vqrdmulhq_n_s16(v2684, 24397);
+ int16x8_t v2686 = vaddq_s16(v2683, v2685);
+ int16x8_t v2687 = vsubq_s16(v2611, v2613);
+ int16x8_t v2688 = vsubq_s16(v2615, v2617);
+ int16x8_t v2689 = vqrdmulhq_n_s16(v2688, 24397);
+ int16x8_t v2690 = vaddq_s16(v2687, v2689);
+ int16x8_t v2691 = vqrdmulhq_n_s16(v2690, 17921);
+ int16x8_t v2692 = vaddq_s16(v2686, v2691);
+ int16x8_t v2693 = vqrdmulhq_n_s16(v2692, 16747);
+ int16x8_t v2694 = vaddq_s16(v2682, v2693);
+ int16x8_t v2695 = vsubq_s16(v2625, v2627);
+ int16x8_t v2696 = vsubq_s16(v2629, v2631);
+ int16x8_t v2697 = vqrdmulhq_n_s16(v2696, 24397);
+ int16x8_t v2698 = vaddq_s16(v2695, v2697);
+ int16x8_t v2699 = vsubq_s16(v2635, v2637);
+ int16x8_t v2700 = vsubq_s16(v2639, v2641);
+ int16x8_t v2701 = vqrdmulhq_n_s16(v2700, 24397);
+ int16x8_t v2702 = vaddq_s16(v2699, v2701);
+ int16x8_t v2703 = vqrdmulhq_n_s16(v2702, 17921);
+ int16x8_t v2704 = vaddq_s16(v2698, v2703);
+ int16x8_t v2705 = vsubq_s16(v2647, v2649);
+ int16x8_t v2706 = vsubq_s16(v2651, v2653);
+ int16x8_t v2707 = vqrdmulhq_n_s16(v2706, 24397);
+ int16x8_t v2708 = vaddq_s16(v2705, v2707);
+ int16x8_t v2709 = vsubq_s16(v2657, v2659);
+ int16x8_t v2710 = vsubq_s16(v2661, v2663);
+ int16x8_t v2711 = vqrdmulhq_n_s16(v2710, 24397);
+ int16x8_t v2712 = vaddq_s16(v2709, v2711);
+ int16x8_t v2713 = vqrdmulhq_n_s16(v2712, 17921);
+ int16x8_t v2714 = vaddq_s16(v2708, v2713);
+ int16x8_t v2715 = vqrdmulhq_n_s16(v2714, 16747);
+ int16x8_t v2716 = vaddq_s16(v2704, v2715);
+ int16x8_t v2717 = vqrdmulhq_n_s16(v2716, 16474);
+ int16x8_t v2718 = vaddq_s16(v2694, v2717);
+ int16x8_t v2719 = vsubq_s16(v2485, v2487);
+ int16x8_t v2720 = vsubq_s16(v2489, v2491);
+ int16x8_t v2721 = vqrdmulhq_n_s16(v2720, 27504);
+ int16x8_t v2722 = vaddq_s16(v2719, v2721);
+ int16x8_t v2723 = vsubq_s16(v2495, v2497);
+ int16x8_t v2724 = vsubq_s16(v2499, v2501);
+ int16x8_t v2725 = vqrdmulhq_n_s16(v2724, 27504);
+ int16x8_t v2726 = vaddq_s16(v2723, v2725);
+ int16x8_t v2727 = vqrdmulhq_n_s16(v2726, 18343);
+ int16x8_t v2728 = vaddq_s16(v2722, v2727);
+ int16x8_t v2729 = vsubq_s16(v2507, v2509);
+ int16x8_t v2730 = vsubq_s16(v2511, v2513);
+ int16x8_t v2731 = vqrdmulhq_n_s16(v2730, 27504);
+ int16x8_t v2732 = vaddq_s16(v2729, v2731);
+ int16x8_t v2733 = vsubq_s16(v2517, v2519);
+ int16x8_t v2734 = vsubq_s16(v2521, v2523);
+ int16x8_t v2735 = vqrdmulhq_n_s16(v2734, 27504);
+ int16x8_t v2736 = vaddq_s16(v2733, v2735);
+ int16x8_t v2737 = vqrdmulhq_n_s16(v2736, 18343);
+ int16x8_t v2738 = vaddq_s16(v2732, v2737);
+ int16x8_t v2739 = vqrdmulhq_n_s16(v2738, 16840);
+ int16x8_t v2740 = vaddq_s16(v2728, v2739);
+ int16x8_t v2741 = vsubq_s16(v2531, v2533);
+ int16x8_t v2742 = vsubq_s16(v2535, v2537);
+ int16x8_t v2743 = vqrdmulhq_n_s16(v2742, 27504);
+ int16x8_t v2744 = vaddq_s16(v2741, v2743);
+ int16x8_t v2745 = vsubq_s16(v2541, v2543);
+ int16x8_t v2746 = vsubq_s16(v2545, v2547);
+ int16x8_t v2747 = vqrdmulhq_n_s16(v2746, 27504);
+ int16x8_t v2748 = vaddq_s16(v2745, v2747);
+ int16x8_t v2749 = vqrdmulhq_n_s16(v2748, 18343);
+ int16x8_t v2750 = vaddq_s16(v2744, v2749);
+ int16x8_t v2751 = vsubq_s16(v2553, v2555);
+ int16x8_t v2752 = vsubq_s16(v2557, v2559);
+ int16x8_t v2753 = vqrdmulhq_n_s16(v2752, 27504);
+ int16x8_t v2754 = vaddq_s16(v2751, v2753);
+ int16x8_t v2755 = vsubq_s16(v2563, v2565);
+ int16x8_t v2756 = vsubq_s16(v2567, v2569);
+ int16x8_t v2757 = vqrdmulhq_n_s16(v2756, 27504);
+ int16x8_t v2758 = vaddq_s16(v2755, v2757);
+ int16x8_t v2759 = vqrdmulhq_n_s16(v2758, 18343);
+ int16x8_t v2760 = vaddq_s16(v2754, v2759);
+ int16x8_t v2761 = vqrdmulhq_n_s16(v2760, 16840);
+ int16x8_t v2762 = vaddq_s16(v2750, v2761);
+ int16x8_t v2763 = vqrdmulhq_n_s16(v2762, 16496);
+ int16x8_t v2764 = vaddq_s16(v2740, v2763);
+ int16x8_t v2765 = vsubq_s16(v2390, v2392);
+ int16x8_t v2766 = vsubq_s16(v2394, v2396);
+ int16x8_t v2767 = vqrdmulhq_n_s16(v2766, 31869);
+ int16x8_t v2768 = vaddq_s16(v2765, v2767);
+ int16x8_t v2769 = vsubq_s16(v2400, v2402);
+ int16x8_t v2770 = vsubq_s16(v2404, v2406);
+ int16x8_t v2771 = vqrdmulhq_n_s16(v2770, 31869);
+ int16x8_t v2772 = vaddq_s16(v2769, v2771);
+ int16x8_t v2773 = vqrdmulhq_n_s16(v2772, 18830);
+ int16x8_t v2774 = vaddq_s16(v2768, v2773);
+ int16x8_t v2775 = vsubq_s16(v2412, v2414);
+ int16x8_t v2776 = vsubq_s16(v2416, v2418);
+ int16x8_t v2777 = vqrdmulhq_n_s16(v2776, 31869);
+ int16x8_t v2778 = vaddq_s16(v2775, v2777);
+ int16x8_t v2779 = vsubq_s16(v2422, v2424);
+ int16x8_t v2780 = vsubq_s16(v2426, v2428);
+ int16x8_t v2781 = vqrdmulhq_n_s16(v2780, 31869);
+ int16x8_t v2782 = vaddq_s16(v2779, v2781);
+ int16x8_t v2783 = vqrdmulhq_n_s16(v2782, 18830);
+ int16x8_t v2784 = vaddq_s16(v2778, v2783);
+ int16x8_t v2785 = vqrdmulhq_n_s16(v2784, 16944);
+ int16x8_t v2786 = vaddq_s16(v2774, v2785);
+ int16x8_t v2787 = vsubq_s16(v2436, v2438);
+ int16x8_t v2788 = vsubq_s16(v2440, v2442);
+ int16x8_t v2789 = vqrdmulhq_n_s16(v2788, 31869);
+ int16x8_t v2790 = vaddq_s16(v2787, v2789);
+ int16x8_t v2791 = vsubq_s16(v2446, v2448);
+ int16x8_t v2792 = vsubq_s16(v2450, v2452);
+ int16x8_t v2793 = vqrdmulhq_n_s16(v2792, 31869);
+ int16x8_t v2794 = vaddq_s16(v2791, v2793);
+ int16x8_t v2795 = vqrdmulhq_n_s16(v2794, 18830);
+ int16x8_t v2796 = vaddq_s16(v2790, v2795);
+ int16x8_t v2797 = vsubq_s16(v2458, v2460);
+ int16x8_t v2798 = vsubq_s16(v2462, v2465);
+ int16x8_t v2799 = vqrdmulhq_n_s16(v2798, 31869);
+ int16x8_t v2800 = vaddq_s16(v2797, v2799);
+ int16x8_t v2801 = vsubq_s16(v2469, v2471);
+ int16x8_t v2802 = vsubq_s16(v2473, v2475);
+ int16x8_t v2803 = vqrdmulhq_n_s16(v2802, 31869);
+ int16x8_t v2804 = vaddq_s16(v2801, v2803);
+ int16x8_t v2805 = vqrdmulhq_n_s16(v2804, 18830);
+ int16x8_t v2806 = vaddq_s16(v2800, v2805);
+ int16x8_t v2807 = vqrdmulhq_n_s16(v2806, 16944);
+ int16x8_t v2808 = vaddq_s16(v2796, v2807);
+ int16x8_t v2809 = vqrdmulhq_n_s16(v2808, 16521);
+ int16x8_t v2810 = vaddq_s16(v2786, v2809);
+ int16x8_t v2811 = vsubq_s16(v2296, v2298);
+ int16x8_t v2812 = vsubq_s16(v2300, v2302);
+ int16x8_t v2813_tmp = vqrdmulhq_n_s16(v2812, 5552);
+ int16x8_t v2813 = vaddq_s16(v2813_tmp, v2812);
+ int16x8_t v2814 = vaddq_s16(v2811, v2813);
+ int16x8_t v2815 = vsubq_s16(v2306, v2308);
+ int16x8_t v2816 = vsubq_s16(v2310, v2312);
+ int16x8_t v2817_tmp = vqrdmulhq_n_s16(v2816, 5552);
+ int16x8_t v2817 = vaddq_s16(v2817_tmp, v2816);
+ int16x8_t v2818 = vaddq_s16(v2815, v2817);
+ int16x8_t v2819 = vqrdmulhq_n_s16(v2818, 19393);
+ int16x8_t v2820 = vaddq_s16(v2814, v2819);
+ int16x8_t v2821 = vsubq_s16(v2318, v2320);
+ int16x8_t v2822 = vsubq_s16(v2322, v2324);
+ int16x8_t v2823_tmp = vqrdmulhq_n_s16(v2822, 5552);
+ int16x8_t v2823 = vaddq_s16(v2823_tmp, v2822);
+ int16x8_t v2824 = vaddq_s16(v2821, v2823);
+ int16x8_t v2825 = vsubq_s16(v2328, v2330);
+ int16x8_t v2826 = vsubq_s16(v2332, v2334);
+ int16x8_t v2827_tmp = vqrdmulhq_n_s16(v2826, 5552);
+ int16x8_t v2827 = vaddq_s16(v2827_tmp, v2826);
+ int16x8_t v2828 = vaddq_s16(v2825, v2827);
+ int16x8_t v2829 = vqrdmulhq_n_s16(v2828, 19393);
+ int16x8_t v2830 = vaddq_s16(v2824, v2829);
+ int16x8_t v2831 = vqrdmulhq_n_s16(v2830, 17059);
+ int16x8_t v2832 = vaddq_s16(v2820, v2831);
+ int16x8_t v2833 = vsubq_s16(v2342, v2344);
+ int16x8_t v2834 = vsubq_s16(v2346, v2348);
+ int16x8_t v2835_tmp = vqrdmulhq_n_s16(v2834, 5552);
+ int16x8_t v2835 = vaddq_s16(v2835_tmp, v2834);
+ int16x8_t v2836 = vaddq_s16(v2833, v2835);
+ int16x8_t v2837 = vsubq_s16(v2352, v2354);
+ int16x8_t v2838 = vsubq_s16(v2356, v2358);
+ int16x8_t v2839_tmp = vqrdmulhq_n_s16(v2838, 5552);
+ int16x8_t v2839 = vaddq_s16(v2839_tmp, v2838);
+ int16x8_t v2840 = vaddq_s16(v2837, v2839);
+ int16x8_t v2841 = vqrdmulhq_n_s16(v2840, 19393);
+ int16x8_t v2842 = vaddq_s16(v2836, v2841);
+ int16x8_t v2843 = vsubq_s16(v2364, v2366);
+ int16x8_t v2844 = vsubq_s16(v2368, v2370);
+ int16x8_t v2845_tmp = vqrdmulhq_n_s16(v2844, 5552);
+ int16x8_t v2845 = vaddq_s16(v2845_tmp, v2844);
+ int16x8_t v2846 = vaddq_s16(v2843, v2845);
+ int16x8_t v2847 = vsubq_s16(v2374, v2376);
+ int16x8_t v2848 = vsubq_s16(v2378, v2380);
+ int16x8_t v2849_tmp = vqrdmulhq_n_s16(v2848, 5552);
+ int16x8_t v2849 = vaddq_s16(v2849_tmp, v2848);
+ int16x8_t v2850 = vaddq_s16(v2847, v2849);
+ int16x8_t v2851 = vqrdmulhq_n_s16(v2850, 19393);
+ int16x8_t v2852 = vaddq_s16(v2846, v2851);
+ int16x8_t v2853 = vqrdmulhq_n_s16(v2852, 17059);
+ int16x8_t v2854 = vaddq_s16(v2842, v2853);
+ int16x8_t v2855 = vqrdmulhq_n_s16(v2854, 16549);
+ int16x8_t v2856 = vaddq_s16(v2832, v2855);
+ int16x8_t v2857 = vsubq_s16(v2109, v2114);
+ int16x8_t v2858 = vsubq_s16(v2119, v2124);
+ int16x8_t v2859_tmp = vqrdmulhq_n_s16(v2858, 15865);
+ int16x8_t v2859 = vaddq_s16(v2859_tmp, v2858);
+ int16x8_t v2860 = vaddq_s16(v2857, v2859);
+ int16x8_t v2861 = vsubq_s16(v2131, v2136);
+ int16x8_t v2862 = vsubq_s16(v2141, v2146);
+ int16x8_t v2863_tmp = vqrdmulhq_n_s16(v2862, 15865);
+ int16x8_t v2863 = vaddq_s16(v2863_tmp, v2862);
+ int16x8_t v2864 = vaddq_s16(v2861, v2863);
+ int16x8_t v2865 = vqrdmulhq_n_s16(v2864, 20040);
+ int16x8_t v2866 = vaddq_s16(v2860, v2865);
+ int16x8_t v2867 = vsubq_s16(v2155, v2160);
+ int16x8_t v2868 = vsubq_s16(v2165, v2170);
+ int16x8_t v2869_tmp = vqrdmulhq_n_s16(v2868, 15865);
+ int16x8_t v2869 = vaddq_s16(v2869_tmp, v2868);
+ int16x8_t v2870 = vaddq_s16(v2867, v2869);
+ int16x8_t v2871 = vsubq_s16(v2177, v2182);
+ int16x8_t v2872 = vsubq_s16(v2187, v2192);
+ int16x8_t v2873_tmp = vqrdmulhq_n_s16(v2872, 15865);
+ int16x8_t v2873 = vaddq_s16(v2873_tmp, v2872);
+ int16x8_t v2874 = vaddq_s16(v2871, v2873);
+ int16x8_t v2875 = vqrdmulhq_n_s16(v2874, 20040);
+ int16x8_t v2876 = vaddq_s16(v2870, v2875);
+ int16x8_t v2877 = vqrdmulhq_n_s16(v2876, 17187);
+ int16x8_t v2878 = vaddq_s16(v2866, v2877);
+ int16x8_t v2879 = vsubq_s16(v2203, v2208);
+ int16x8_t v2880 = vsubq_s16(v2213, v2218);
+ int16x8_t v2881_tmp = vqrdmulhq_n_s16(v2880, 15865);
+ int16x8_t v2881 = vaddq_s16(v2881_tmp, v2880);
+ int16x8_t v2882 = vaddq_s16(v2879, v2881);
+ int16x8_t v2883 = vsubq_s16(v2225, v2230);
+ int16x8_t v2884 = vsubq_s16(v2235, v2240);
+ int16x8_t v2885_tmp = vqrdmulhq_n_s16(v2884, 15865);
+ int16x8_t v2885 = vaddq_s16(v2885_tmp, v2884);
+ int16x8_t v2886 = vaddq_s16(v2883, v2885);
+ int16x8_t v2887 = vqrdmulhq_n_s16(v2886, 20040);
+ int16x8_t v2888 = vaddq_s16(v2882, v2887);
+ int16x8_t v2889 = vsubq_s16(v2249, v2254);
+ int16x8_t v2890 = vsubq_s16(v2259, v2264);
+ int16x8_t v2891_tmp = vqrdmulhq_n_s16(v2890, 15865);
+ int16x8_t v2891 = vaddq_s16(v2891_tmp, v2890);
+ int16x8_t v2892 = vaddq_s16(v2889, v2891);
+ int16x8_t v2893 = vsubq_s16(v2271, v2276);
+ int16x8_t v2894 = vsubq_s16(v2281, v2286);
+ int16x8_t v2895_tmp = vqrdmulhq_n_s16(v2894, 15865);
+ int16x8_t v2895 = vaddq_s16(v2895_tmp, v2894);
+ int16x8_t v2896 = vaddq_s16(v2893, v2895);
+ int16x8_t v2897 = vqrdmulhq_n_s16(v2896, 20040);
+ int16x8_t v2898 = vaddq_s16(v2892, v2897);
+ int16x8_t v2899 = vqrdmulhq_n_s16(v2898, 17187);
+ int16x8_t v2900 = vaddq_s16(v2888, v2899);
+ int16x8_t v2901 = vqrdmulhq_n_s16(v2900, 16579);
+ int16x8_t v2902 = vaddq_s16(v2878, v2901);
+ int16x8_t v2903 = vsubq_s16(v1919, v1924);
+ int16x8_t v2904 = vsubq_s16(v1929, v1934);
+ int16x8_t v2905_tmp = vqrdmulhq_n_s16(v2904, 1893);
+ int16x8_t v2905 = vmlaq_n_s16(v2905_tmp, v2904, 2);
+ int16x8_t v2906 = vaddq_s16(v2903, v2905);
+ int16x8_t v2907 = vsubq_s16(v1941, v1946);
+ int16x8_t v2908 = vsubq_s16(v1951, v1956);
+ int16x8_t v2909_tmp = vqrdmulhq_n_s16(v2908, 1893);
+ int16x8_t v2909 = vmlaq_n_s16(v2909_tmp, v2908, 2);
+ int16x8_t v2910 = vaddq_s16(v2907, v2909);
+ int16x8_t v2911 = vqrdmulhq_n_s16(v2910, 20783);
+ int16x8_t v2912 = vaddq_s16(v2906, v2911);
+ int16x8_t v2913 = vsubq_s16(v1965, v1970);
+ int16x8_t v2914 = vsubq_s16(v1975, v1980);
+ int16x8_t v2915_tmp = vqrdmulhq_n_s16(v2914, 1893);
+ int16x8_t v2915 = vmlaq_n_s16(v2915_tmp, v2914, 2);
+ int16x8_t v2916 = vaddq_s16(v2913, v2915);
+ int16x8_t v2917 = vsubq_s16(v1987, v1992);
+ int16x8_t v2918 = vsubq_s16(v1997, v2002);
+ int16x8_t v2919_tmp = vqrdmulhq_n_s16(v2918, 1893);
+ int16x8_t v2919 = vmlaq_n_s16(v2919_tmp, v2918, 2);
+ int16x8_t v2920 = vaddq_s16(v2917, v2919);
+ int16x8_t v2921 = vqrdmulhq_n_s16(v2920, 20783);
+ int16x8_t v2922 = vaddq_s16(v2916, v2921);
+ int16x8_t v2923 = vqrdmulhq_n_s16(v2922, 17326);
+ int16x8_t v2924 = vaddq_s16(v2912, v2923);
+ int16x8_t v2925 = vsubq_s16(v2013, v2018);
+ int16x8_t v2926 = vsubq_s16(v2023, v2028);
+ int16x8_t v2927_tmp = vqrdmulhq_n_s16(v2926, 1893);
+ int16x8_t v2927 = vmlaq_n_s16(v2927_tmp, v2926, 2);
+ int16x8_t v2928 = vaddq_s16(v2925, v2927);
+ int16x8_t v2929 = vsubq_s16(v2035, v2040);
+ int16x8_t v2930 = vsubq_s16(v2045, v2050);
+ int16x8_t v2931_tmp = vqrdmulhq_n_s16(v2930, 1893);
+ int16x8_t v2931 = vmlaq_n_s16(v2931_tmp, v2930, 2);
+ int16x8_t v2932 = vaddq_s16(v2929, v2931);
+ int16x8_t v2933 = vqrdmulhq_n_s16(v2932, 20783);
+ int16x8_t v2934 = vaddq_s16(v2928, v2933);
+ int16x8_t v2935 = vsubq_s16(v2059, v2064);
+ int16x8_t v2936 = vsubq_s16(v2069, v2074);
+ int16x8_t v2937_tmp = vqrdmulhq_n_s16(v2936, 1893);
+ int16x8_t v2937 = vmlaq_n_s16(v2937_tmp, v2936, 2);
+ int16x8_t v2938 = vaddq_s16(v2935, v2937);
+ int16x8_t v2939 = vsubq_s16(v2081, v2086);
+ int16x8_t v2940 = vsubq_s16(v2091, v2096);
+ int16x8_t v2941_tmp = vqrdmulhq_n_s16(v2940, 1893);
+ int16x8_t v2941 = vmlaq_n_s16(v2941_tmp, v2940, 2);
+ int16x8_t v2942 = vaddq_s16(v2939, v2941);
+ int16x8_t v2943 = vqrdmulhq_n_s16(v2942, 20783);
+ int16x8_t v2944 = vaddq_s16(v2938, v2943);
+ int16x8_t v2945 = vqrdmulhq_n_s16(v2944, 17326);
+ int16x8_t v2946 = vaddq_s16(v2934, v2945);
+ int16x8_t v2947 = vqrdmulhq_n_s16(v2946, 16611);
+ int16x8_t v2948 = vaddq_s16(v2924, v2947);
+ int16x8_t v2949 = vsubq_s16(v1543, v1554);
+ int16x8_t v2950 = vsubq_s16(v1565, v1576);
+ int16x8_t v2951_tmp = vqrdmulhq_n_s16(v2950, 13357);
+ int16x8_t v2951 = vmlaq_n_s16(v2951_tmp, v2950, 3);
+ int16x8_t v2952 = vaddq_s16(v2949, v2951);
+ int16x8_t v2953 = vsubq_s16(v1589, v1600);
+ int16x8_t v2954 = vsubq_s16(v1611, v1622);
+ int16x8_t v2955_tmp = vqrdmulhq_n_s16(v2954, 13357);
+ int16x8_t v2955 = vmlaq_n_s16(v2955_tmp, v2954, 3);
+ int16x8_t v2956 = vaddq_s16(v2953, v2955);
+ int16x8_t v2957 = vqrdmulhq_n_s16(v2956, 21637);
+ int16x8_t v2958 = vaddq_s16(v2952, v2957);
+ int16x8_t v2959 = vsubq_s16(v1637, v1648);
+ int16x8_t v2960 = vsubq_s16(v1659, v1670);
+ int16x8_t v2961_tmp = vqrdmulhq_n_s16(v2960, 13357);
+ int16x8_t v2961 = vmlaq_n_s16(v2961_tmp, v2960, 3);
+ int16x8_t v2962 = vaddq_s16(v2959, v2961);
+ int16x8_t v2963 = vsubq_s16(v1683, v1694);
+ int16x8_t v2964 = vsubq_s16(v1705, v1716);
+ int16x8_t v2965_tmp = vqrdmulhq_n_s16(v2964, 13357);
+ int16x8_t v2965 = vmlaq_n_s16(v2965_tmp, v2964, 3);
+ int16x8_t v2966 = vaddq_s16(v2963, v2965);
+ int16x8_t v2967 = vqrdmulhq_n_s16(v2966, 21637);
+ int16x8_t v2968 = vaddq_s16(v2962, v2967);
+ int16x8_t v2969 = vqrdmulhq_n_s16(v2968, 17479);
+ int16x8_t v2970 = vaddq_s16(v2958, v2969);
+ int16x8_t v2971 = vsubq_s16(v1733, v1744);
+ int16x8_t v2972 = vsubq_s16(v1755, v1766);
+ int16x8_t v2973_tmp = vqrdmulhq_n_s16(v2972, 13357);
+ int16x8_t v2973 = vmlaq_n_s16(v2973_tmp, v2972, 3);
+ int16x8_t v2974 = vaddq_s16(v2971, v2973);
+ int16x8_t v2975 = vsubq_s16(v1779, v1790);
+ int16x8_t v2976 = vsubq_s16(v1801, v1812);
+ int16x8_t v2977_tmp = vqrdmulhq_n_s16(v2976, 13357);
+ int16x8_t v2977 = vmlaq_n_s16(v2977_tmp, v2976, 3);
+ int16x8_t v2978 = vaddq_s16(v2975, v2977);
+ int16x8_t v2979 = vqrdmulhq_n_s16(v2978, 21637);
+ int16x8_t v2980 = vaddq_s16(v2974, v2979);
+ int16x8_t v2981 = vsubq_s16(v1827, v1838);
+ int16x8_t v2982 = vsubq_s16(v1849, v1860);
+ int16x8_t v2983_tmp = vqrdmulhq_n_s16(v2982, 13357);
+ int16x8_t v2983 = vmlaq_n_s16(v2983_tmp, v2982, 3);
+ int16x8_t v2984 = vaddq_s16(v2981, v2983);
+ int16x8_t v2985 = vsubq_s16(v1873, v1884);
+ int16x8_t v2986 = vsubq_s16(v1895, v1906);
+ int16x8_t v2987_tmp = vqrdmulhq_n_s16(v2986, 13357);
+ int16x8_t v2987 = vmlaq_n_s16(v2987_tmp, v2986, 3);
+ int16x8_t v2988 = vaddq_s16(v2985, v2987);
+ int16x8_t v2989 = vqrdmulhq_n_s16(v2988, 21637);
+ int16x8_t v2990 = vaddq_s16(v2984, v2989);
+ int16x8_t v2991 = vqrdmulhq_n_s16(v2990, 17479);
+ int16x8_t v2992 = vaddq_s16(v2980, v2991);
+ int16x8_t v2993 = vqrdmulhq_n_s16(v2992, 16647);
+ int16x8_t v2994 = vaddq_s16(v2970, v2993);
+ int16x8_t v2995 = vsubq_s16(v25, v60);
+ int16x8_t v2996 = vsubq_s16(v102, v138);
+ int16x8_t v2997_tmp = vqrdmulhq_n_s16(v2996, 6226);
+ int16x8_t v2997 = vmlaq_n_s16(v2997_tmp, v2996, 10);
+ int16x8_t v2998 = vaddq_s16(v2995, v2997);
+ int16x8_t v2999 = vsubq_s16(v182, v233);
+ int16x8_t v3000 = vsubq_s16(v275, v312);
+ int16x8_t v3001_tmp = vqrdmulhq_n_s16(v3000, 6226);
+ int16x8_t v3001 = vmlaq_n_s16(v3001_tmp, v3000, 10);
+ int16x8_t v3002 = vaddq_s16(v2999, v3001);
+ int16x8_t v3003 = vqrdmulhq_n_s16(v3002, 22622);
+ int16x8_t v3004 = vaddq_s16(v2998, v3003);
+ int16x8_t v3005 = vsubq_s16(v358, v409);
+ int16x8_t v3006 = vsubq_s16(v481, v519);
+ int16x8_t v3007_tmp = vqrdmulhq_n_s16(v3006, 6226);
+ int16x8_t v3007 = vmlaq_n_s16(v3007_tmp, v3006, 10);
+ int16x8_t v3008 = vaddq_s16(v3005, v3007);
+ int16x8_t v3009 = vsubq_s16(v563, v614);
+ int16x8_t v3010 = vsubq_s16(v656, v694);
+ int16x8_t v3011_tmp = vqrdmulhq_n_s16(v3010, 6226);
+ int16x8_t v3011 = vmlaq_n_s16(v3011_tmp, v3010, 10);
+ int16x8_t v3012 = vaddq_s16(v3009, v3011);
+ int16x8_t v3013 = vqrdmulhq_n_s16(v3012, 22622);
+ int16x8_t v3014 = vaddq_s16(v3008, v3013);
+ int16x8_t v3015 = vqrdmulhq_n_s16(v3014, 17646);
+ int16x8_t v3016 = vaddq_s16(v3004, v3015);
+ int16x8_t v3017 = vsubq_s16(v742, v793);
+ int16x8_t v3018 = vsubq_s16(v865, v903);
+ int16x8_t v3019_tmp = vqrdmulhq_n_s16(v3018, 6226);
+ int16x8_t v3019 = vmlaq_n_s16(v3019_tmp, v3018, 10);
+ int16x8_t v3020 = vaddq_s16(v3017, v3019);
+ int16x8_t v3021 = vsubq_s16(v977, v1060);
+ int16x8_t v3022 = vsubq_s16(v1102, v1141);
+ int16x8_t v3023_tmp = vqrdmulhq_n_s16(v3022, 6226);
+ int16x8_t v3023 = vmlaq_n_s16(v3023_tmp, v3022, 10);
+ int16x8_t v3024 = vaddq_s16(v3021, v3023);
+ int16x8_t v3025 = vqrdmulhq_n_s16(v3024, 22622);
+ int16x8_t v3026 = vaddq_s16(v3020, v3025);
+ int16x8_t v3027 = vsubq_s16(v1187, v1238);
+ int16x8_t v3028 = vsubq_s16(v1310, v1348);
+ int16x8_t v3029_tmp = vqrdmulhq_n_s16(v3028, 6226);
+ int16x8_t v3029 = vmlaq_n_s16(v3029_tmp, v3028, 10);
+ int16x8_t v3030 = vaddq_s16(v3027, v3029);
+ int16x8_t v3031 = vsubq_s16(v1392, v1443);
+ int16x8_t v3032 = vsubq_s16(v1485, v1524);
+ int16x8_t v3033_tmp = vqrdmulhq_n_s16(v3032, 6226);
+ int16x8_t v3033 = vmlaq_n_s16(v3033_tmp, v3032, 10);
+ int16x8_t v3034 = vaddq_s16(v3031, v3033);
+ int16x8_t v3035 = vqrdmulhq_n_s16(v3034, 22622);
+ int16x8_t v3036 = vaddq_s16(v3030, v3035);
+ int16x8_t v3037 = vqrdmulhq_n_s16(v3036, 17646);
+ int16x8_t v3038 = vaddq_s16(v3026, v3037);
+ int16x8_t v3039 = vqrdmulhq_n_s16(v3038, 16685);
+ int16x8_t v3040 = vaddq_s16(v3016, v3039);
+ int16x8_t v3041 = vsubq_s16(v2995, v2997);
+ int16x8_t v3042 = vsubq_s16(v2999, v3001);
+ int16x8_t v3043 = vqrdmulhq_n_s16(v3042, 23761);
+ int16x8_t v3044 = vaddq_s16(v3041, v3043);
+ int16x8_t v3045 = vsubq_s16(v3005, v3007);
+ int16x8_t v3046 = vsubq_s16(v3009, v3011);
+ int16x8_t v3047 = vqrdmulhq_n_s16(v3046, 23761);
+ int16x8_t v3048 = vaddq_s16(v3045, v3047);
+ int16x8_t v3049 = vqrdmulhq_n_s16(v3048, 17826);
+ int16x8_t v3050 = vaddq_s16(v3044, v3049);
+ int16x8_t v3051 = vsubq_s16(v3017, v3019);
+ int16x8_t v3052 = vsubq_s16(v3021, v3023);
+ int16x8_t v3053 = vqrdmulhq_n_s16(v3052, 23761);
+ int16x8_t v3054 = vaddq_s16(v3051, v3053);
+ int16x8_t v3055 = vsubq_s16(v3027, v3029);
+ int16x8_t v3056 = vsubq_s16(v3031, v3033);
+ int16x8_t v3057 = vqrdmulhq_n_s16(v3056, 23761);
+ int16x8_t v3058 = vaddq_s16(v3055, v3057);
+ int16x8_t v3059 = vqrdmulhq_n_s16(v3058, 17826);
+ int16x8_t v3060 = vaddq_s16(v3054, v3059);
+ int16x8_t v3061 = vqrdmulhq_n_s16(v3060, 16726);
+ int16x8_t v3062 = vaddq_s16(v3050, v3061);
+ int16x8_t v3063 = vsubq_s16(v2949, v2951);
+ int16x8_t v3064 = vsubq_s16(v2953, v2955);
+ int16x8_t v3065 = vqrdmulhq_n_s16(v3064, 25084);
+ int16x8_t v3066 = vaddq_s16(v3063, v3065);
+ int16x8_t v3067 = vsubq_s16(v2959, v2961);
+ int16x8_t v3068 = vsubq_s16(v2963, v2965);
+ int16x8_t v3069 = vqrdmulhq_n_s16(v3068, 25084);
+ int16x8_t v3070 = vaddq_s16(v3067, v3069);
+ int16x8_t v3071 = vqrdmulhq_n_s16(v3070, 18021);
+ int16x8_t v3072 = vaddq_s16(v3066, v3071);
+ int16x8_t v3073 = vsubq_s16(v2971, v2973);
+ int16x8_t v3074 = vsubq_s16(v2975, v2977);
+ int16x8_t v3075 = vqrdmulhq_n_s16(v3074, 25084);
+ int16x8_t v3076 = vaddq_s16(v3073, v3075);
+ int16x8_t v3077 = vsubq_s16(v2981, v2983);
+ int16x8_t v3078 = vsubq_s16(v2985, v2987);
+ int16x8_t v3079 = vqrdmulhq_n_s16(v3078, 25084);
+ int16x8_t v3080 = vaddq_s16(v3077, v3079);
+ int16x8_t v3081 = vqrdmulhq_n_s16(v3080, 18021);
+ int16x8_t v3082 = vaddq_s16(v3076, v3081);
+ int16x8_t v3083 = vqrdmulhq_n_s16(v3082, 16769);
+ int16x8_t v3084 = vaddq_s16(v3072, v3083);
+ int16x8_t v3085 = vsubq_s16(v2903, v2905);
+ int16x8_t v3086 = vsubq_s16(v2907, v2909);
+ int16x8_t v3087 = vqrdmulhq_n_s16(v3086, 26631);
+ int16x8_t v3088 = vaddq_s16(v3085, v3087);
+ int16x8_t v3089 = vsubq_s16(v2913, v2915);
+ int16x8_t v3090 = vsubq_s16(v2917, v2919);
+ int16x8_t v3091 = vqrdmulhq_n_s16(v3090, 26631);
+ int16x8_t v3092 = vaddq_s16(v3089, v3091);
+ int16x8_t v3093 = vqrdmulhq_n_s16(v3092, 18231);
+ int16x8_t v3094 = vaddq_s16(v3088, v3093);
+ int16x8_t v3095 = vsubq_s16(v2925, v2927);
+ int16x8_t v3096 = vsubq_s16(v2929, v2931);
+ int16x8_t v3097 = vqrdmulhq_n_s16(v3096, 26631);
+ int16x8_t v3098 = vaddq_s16(v3095, v3097);
+ int16x8_t v3099 = vsubq_s16(v2935, v2937);
+ int16x8_t v3100 = vsubq_s16(v2939, v2941);
+ int16x8_t v3101 = vqrdmulhq_n_s16(v3100, 26631);
+ int16x8_t v3102 = vaddq_s16(v3099, v3101);
+ int16x8_t v3103 = vqrdmulhq_n_s16(v3102, 18231);
+ int16x8_t v3104 = vaddq_s16(v3098, v3103);
+ int16x8_t v3105 = vqrdmulhq_n_s16(v3104, 16815);
+ int16x8_t v3106 = vaddq_s16(v3094, v3105);
+ int16x8_t v3107 = vsubq_s16(v2857, v2859);
+ int16x8_t v3108 = vsubq_s16(v2861, v2863);
+ int16x8_t v3109 = vqrdmulhq_n_s16(v3108, 28454);
+ int16x8_t v3110 = vaddq_s16(v3107, v3109);
+ int16x8_t v3111 = vsubq_s16(v2867, v2869);
+ int16x8_t v3112 = vsubq_s16(v2871, v2873);
+ int16x8_t v3113 = vqrdmulhq_n_s16(v3112, 28454);
+ int16x8_t v3114 = vaddq_s16(v3111, v3113);
+ int16x8_t v3115 = vqrdmulhq_n_s16(v3114, 18458);
+ int16x8_t v3116 = vaddq_s16(v3110, v3115);
+ int16x8_t v3117 = vsubq_s16(v2879, v2881);
+ int16x8_t v3118 = vsubq_s16(v2883, v2885);
+ int16x8_t v3119 = vqrdmulhq_n_s16(v3118, 28454);
+ int16x8_t v3120 = vaddq_s16(v3117, v3119);
+ int16x8_t v3121 = vsubq_s16(v2889, v2891);
+ int16x8_t v3122 = vsubq_s16(v2893, v2895);
+ int16x8_t v3123 = vqrdmulhq_n_s16(v3122, 28454);
+ int16x8_t v3124 = vaddq_s16(v3121, v3123);
+ int16x8_t v3125 = vqrdmulhq_n_s16(v3124, 18458);
+ int16x8_t v3126 = vaddq_s16(v3120, v3125);
+ int16x8_t v3127 = vqrdmulhq_n_s16(v3126, 16865);
+ int16x8_t v3128 = vaddq_s16(v3116, v3127);
+ int16x8_t v3129 = vsubq_s16(v2811, v2813);
+ int16x8_t v3130 = vsubq_s16(v2815, v2817);
+ int16x8_t v3131 = vqrdmulhq_n_s16(v3130, 30624);
+ int16x8_t v3132 = vaddq_s16(v3129, v3131);
+ int16x8_t v3133 = vsubq_s16(v2821, v2823);
+ int16x8_t v3134 = vsubq_s16(v2825, v2827);
+ int16x8_t v3135 = vqrdmulhq_n_s16(v3134, 30624);
+ int16x8_t v3136 = vaddq_s16(v3133, v3135);
+ int16x8_t v3137 = vqrdmulhq_n_s16(v3136, 18702);
+ int16x8_t v3138 = vaddq_s16(v3132, v3137);
+ int16x8_t v3139 = vsubq_s16(v2833, v2835);
+ int16x8_t v3140 = vsubq_s16(v2837, v2839);
+ int16x8_t v3141 = vqrdmulhq_n_s16(v3140, 30624);
+ int16x8_t v3142 = vaddq_s16(v3139, v3141);
+ int16x8_t v3143 = vsubq_s16(v2843, v2845);
+ int16x8_t v3144 = vsubq_s16(v2847, v2849);
+ int16x8_t v3145 = vqrdmulhq_n_s16(v3144, 30624);
+ int16x8_t v3146 = vaddq_s16(v3143, v3145);
+ int16x8_t v3147 = vqrdmulhq_n_s16(v3146, 18702);
+ int16x8_t v3148 = vaddq_s16(v3142, v3147);
+ int16x8_t v3149 = vqrdmulhq_n_s16(v3148, 16916);
+ int16x8_t v3150 = vaddq_s16(v3138, v3149);
+ int16x8_t v3151 = vsubq_s16(v2765, v2767);
+ int16x8_t v3152 = vsubq_s16(v2769, v2771);
+ int16x8_t v3153_tmp = vqrdmulhq_n_s16(v3152, 472);
+ int16x8_t v3153 = vaddq_s16(v3153_tmp, v3152);
+ int16x8_t v3154 = vaddq_s16(v3151, v3153);
+ int16x8_t v3155 = vsubq_s16(v2775, v2777);
+ int16x8_t v3156 = vsubq_s16(v2779, v2781);
+ int16x8_t v3157_tmp = vqrdmulhq_n_s16(v3156, 472);
+ int16x8_t v3157 = vaddq_s16(v3157_tmp, v3156);
+ int16x8_t v3158 = vaddq_s16(v3155, v3157);
+ int16x8_t v3159 = vqrdmulhq_n_s16(v3158, 18964);
+ int16x8_t v3160 = vaddq_s16(v3154, v3159);
+ int16x8_t v3161 = vsubq_s16(v2787, v2789);
+ int16x8_t v3162 = vsubq_s16(v2791, v2793);
+ int16x8_t v3163_tmp = vqrdmulhq_n_s16(v3162, 472);
+ int16x8_t v3163 = vaddq_s16(v3163_tmp, v3162);
+ int16x8_t v3164 = vaddq_s16(v3161, v3163);
+ int16x8_t v3165 = vsubq_s16(v2797, v2799);
+ int16x8_t v3166 = vsubq_s16(v2801, v2803);
+ int16x8_t v3167_tmp = vqrdmulhq_n_s16(v3166, 472);
+ int16x8_t v3167 = vaddq_s16(v3167_tmp, v3166);
+ int16x8_t v3168 = vaddq_s16(v3165, v3167);
+ int16x8_t v3169 = vqrdmulhq_n_s16(v3168, 18964);
+ int16x8_t v3170 = vaddq_s16(v3164, v3169);
+ int16x8_t v3171 = vqrdmulhq_n_s16(v3170, 16971);
+ int16x8_t v3172 = vaddq_s16(v3160, v3171);
+ int16x8_t v3173 = vsubq_s16(v2719, v2721);
+ int16x8_t v3174 = vsubq_s16(v2723, v2725);
+ int16x8_t v3175_tmp = vqrdmulhq_n_s16(v3174, 3672);
+ int16x8_t v3175 = vaddq_s16(v3175_tmp, v3174);
+ int16x8_t v3176 = vaddq_s16(v3173, v3175);
+ int16x8_t v3177 = vsubq_s16(v2729, v2731);
+ int16x8_t v3178 = vsubq_s16(v2733, v2735);
+ int16x8_t v3179_tmp = vqrdmulhq_n_s16(v3178, 3672);
+ int16x8_t v3179 = vaddq_s16(v3179_tmp, v3178);
+ int16x8_t v3180 = vaddq_s16(v3177, v3179);
+ int16x8_t v3181 = vqrdmulhq_n_s16(v3180, 19245);
+ int16x8_t v3182 = vaddq_s16(v3176, v3181);
+ int16x8_t v3183 = vsubq_s16(v2741, v2743);
+ int16x8_t v3184 = vsubq_s16(v2745, v2747);
+ int16x8_t v3185_tmp = vqrdmulhq_n_s16(v3184, 3672);
+ int16x8_t v3185 = vaddq_s16(v3185_tmp, v3184);
+ int16x8_t v3186 = vaddq_s16(v3183, v3185);
+ int16x8_t v3187 = vsubq_s16(v2751, v2753);
+ int16x8_t v3188 = vsubq_s16(v2755, v2757);
+ int16x8_t v3189_tmp = vqrdmulhq_n_s16(v3188, 3672);
+ int16x8_t v3189 = vaddq_s16(v3189_tmp, v3188);
+ int16x8_t v3190 = vaddq_s16(v3187, v3189);
+ int16x8_t v3191 = vqrdmulhq_n_s16(v3190, 19245);
+ int16x8_t v3192 = vaddq_s16(v3186, v3191);
+ int16x8_t v3193 = vqrdmulhq_n_s16(v3192, 17029);
+ int16x8_t v3194 = vaddq_s16(v3182, v3193);
+ int16x8_t v3195 = vsubq_s16(v2673, v2675);
+ int16x8_t v3196 = vsubq_s16(v2677, v2679);
+ int16x8_t v3197_tmp = vqrdmulhq_n_s16(v3196, 7662);
+ int16x8_t v3197 = vaddq_s16(v3197_tmp, v3196);
+ int16x8_t v3198 = vaddq_s16(v3195, v3197);
+ int16x8_t v3199 = vsubq_s16(v2683, v2685);
+ int16x8_t v3200 = vsubq_s16(v2687, v2689);
+ int16x8_t v3201_tmp = vqrdmulhq_n_s16(v3200, 7662);
+ int16x8_t v3201 = vaddq_s16(v3201_tmp, v3200);
+ int16x8_t v3202 = vaddq_s16(v3199, v3201);
+ int16x8_t v3203 = vqrdmulhq_n_s16(v3202, 19546);
+ int16x8_t v3204 = vaddq_s16(v3198, v3203);
+ int16x8_t v3205 = vsubq_s16(v2695, v2697);
+ int16x8_t v3206 = vsubq_s16(v2699, v2701);
+ int16x8_t v3207_tmp = vqrdmulhq_n_s16(v3206, 7662);
+ int16x8_t v3207 = vaddq_s16(v3207_tmp, v3206);
+ int16x8_t v3208 = vaddq_s16(v3205, v3207);
+ int16x8_t v3209 = vsubq_s16(v2705, v2707);
+ int16x8_t v3210 = vsubq_s16(v2709, v2711);
+ int16x8_t v3211_tmp = vqrdmulhq_n_s16(v3210, 7662);
+ int16x8_t v3211 = vaddq_s16(v3211_tmp, v3210);
+ int16x8_t v3212 = vaddq_s16(v3209, v3211);
+ int16x8_t v3213 = vqrdmulhq_n_s16(v3212, 19546);
+ int16x8_t v3214 = vaddq_s16(v3208, v3213);
+ int16x8_t v3215 = vqrdmulhq_n_s16(v3214, 17090);
+ int16x8_t v3216 = vaddq_s16(v3204, v3215);
+ int16x8_t v3217 = vsubq_s16(v2582, v2587);
+ int16x8_t v3218 = vsubq_s16(v2592, v2597);
+ int16x8_t v3219_tmp = vqrdmulhq_n_s16(v3218, 12756);
+ int16x8_t v3219 = vaddq_s16(v3219_tmp, v3218);
+ int16x8_t v3220 = vaddq_s16(v3217, v3219);
+ int16x8_t v3221 = vsubq_s16(v2604, v2609);
+ int16x8_t v3222 = vsubq_s16(v2614, v2619);
+ int16x8_t v3223_tmp = vqrdmulhq_n_s16(v3222, 12756);
+ int16x8_t v3223 = vaddq_s16(v3223_tmp, v3222);
+ int16x8_t v3224 = vaddq_s16(v3221, v3223);
+ int16x8_t v3225 = vqrdmulhq_n_s16(v3224, 19869);
+ int16x8_t v3226 = vaddq_s16(v3220, v3225);
+ int16x8_t v3227 = vsubq_s16(v2628, v2633);
+ int16x8_t v3228 = vsubq_s16(v2638, v2643);
+ int16x8_t v3229_tmp = vqrdmulhq_n_s16(v3228, 12756);
+ int16x8_t v3229 = vaddq_s16(v3229_tmp, v3228);
+ int16x8_t v3230 = vaddq_s16(v3227, v3229);
+ int16x8_t v3231 = vsubq_s16(v2650, v2655);
+ int16x8_t v3232 = vsubq_s16(v2660, v2665);
+ int16x8_t v3233_tmp = vqrdmulhq_n_s16(v3232, 12756);
+ int16x8_t v3233 = vaddq_s16(v3233_tmp, v3232);
+ int16x8_t v3234 = vaddq_s16(v3231, v3233);
+ int16x8_t v3235 = vqrdmulhq_n_s16(v3234, 19869);
+ int16x8_t v3236 = vaddq_s16(v3230, v3235);
+ int16x8_t v3237 = vqrdmulhq_n_s16(v3236, 17153);
+ int16x8_t v3238 = vaddq_s16(v3226, v3237);
+ int16x8_t v3239 = vsubq_s16(v2488, v2493);
+ int16x8_t v3240 = vsubq_s16(v2498, v2503);
+ int16x8_t v3241_tmp = vqrdmulhq_n_s16(v3240, 19463);
+ int16x8_t v3241 = vaddq_s16(v3241_tmp, v3240);
+ int16x8_t v3242 = vaddq_s16(v3239, v3241);
+ int16x8_t v3243 = vsubq_s16(v2510, v2515);
+ int16x8_t v3244 = vsubq_s16(v2520, v2525);
+ int16x8_t v3245_tmp = vqrdmulhq_n_s16(v3244, 19463);
+ int16x8_t v3245 = vaddq_s16(v3245_tmp, v3244);
+ int16x8_t v3246 = vaddq_s16(v3243, v3245);
+ int16x8_t v3247 = vqrdmulhq_n_s16(v3246, 20216);
+ int16x8_t v3248 = vaddq_s16(v3242, v3247);
+ int16x8_t v3249 = vsubq_s16(v2534, v2539);
+ int16x8_t v3250 = vsubq_s16(v2544, v2549);
+ int16x8_t v3251_tmp = vqrdmulhq_n_s16(v3250, 19463);
+ int16x8_t v3251 = vaddq_s16(v3251_tmp, v3250);
+ int16x8_t v3252 = vaddq_s16(v3249, v3251);
+ int16x8_t v3253 = vsubq_s16(v2556, v2561);
+ int16x8_t v3254 = vsubq_s16(v2566, v2571);
+ int16x8_t v3255_tmp = vqrdmulhq_n_s16(v3254, 19463);
+ int16x8_t v3255 = vaddq_s16(v3255_tmp, v3254);
+ int16x8_t v3256 = vaddq_s16(v3253, v3255);
+ int16x8_t v3257 = vqrdmulhq_n_s16(v3256, 20216);
+ int16x8_t v3258 = vaddq_s16(v3252, v3257);
+ int16x8_t v3259 = vqrdmulhq_n_s16(v3258, 17220);
+ int16x8_t v3260 = vaddq_s16(v3248, v3259);
+ int16x8_t v3261 = vsubq_s16(v2393, v2398);
+ int16x8_t v3262 = vsubq_s16(v2403, v2408);
+ int16x8_t v3263_tmp = vqrdmulhq_n_s16(v3262, 28661);
+ int16x8_t v3263 = vaddq_s16(v3263_tmp, v3262);
+ int16x8_t v3264 = vaddq_s16(v3261, v3263);
+ int16x8_t v3265 = vsubq_s16(v2415, v2420);
+ int16x8_t v3266 = vsubq_s16(v2425, v2430);
+ int16x8_t v3267_tmp = vqrdmulhq_n_s16(v3266, 28661);
+ int16x8_t v3267 = vaddq_s16(v3267_tmp, v3266);
+ int16x8_t v3268 = vaddq_s16(v3265, v3267);
+ int16x8_t v3269 = vqrdmulhq_n_s16(v3268, 20587);
+ int16x8_t v3270 = vaddq_s16(v3264, v3269);
+ int16x8_t v3271 = vsubq_s16(v2439, v2444);
+ int16x8_t v3272 = vsubq_s16(v2449, v2454);
+ int16x8_t v3273_tmp = vqrdmulhq_n_s16(v3272, 28661);
+ int16x8_t v3273 = vaddq_s16(v3273_tmp, v3272);
+ int16x8_t v3274 = vaddq_s16(v3271, v3273);
+ int16x8_t v3275 = vsubq_s16(v2461, v2467);
+ int16x8_t v3276 = vsubq_s16(v2472, v2477);
+ int16x8_t v3277_tmp = vqrdmulhq_n_s16(v3276, 28661);
+ int16x8_t v3277 = vaddq_s16(v3277_tmp, v3276);
+ int16x8_t v3278 = vaddq_s16(v3275, v3277);
+ int16x8_t v3279 = vqrdmulhq_n_s16(v3278, 20587);
+ int16x8_t v3280 = vaddq_s16(v3274, v3279);
+ int16x8_t v3281 = vqrdmulhq_n_s16(v3280, 17290);
+ int16x8_t v3282 = vaddq_s16(v3270, v3281);
+ int16x8_t v3283 = vsubq_s16(v2299, v2304);
+ int16x8_t v3284 = vsubq_s16(v2309, v2314);
+ int16x8_t v3285_tmp = vqrdmulhq_n_s16(v3284, 9242);
+ int16x8_t v3285 = vmlaq_n_s16(v3285_tmp, v3284, 2);
+ int16x8_t v3286 = vaddq_s16(v3283, v3285);
+ int16x8_t v3287 = vsubq_s16(v2321, v2326);
+ int16x8_t v3288 = vsubq_s16(v2331, v2336);
+ int16x8_t v3289_tmp = vqrdmulhq_n_s16(v3288, 9242);
+ int16x8_t v3289 = vmlaq_n_s16(v3289_tmp, v3288, 2);
+ int16x8_t v3290 = vaddq_s16(v3287, v3289);
+ int16x8_t v3291 = vqrdmulhq_n_s16(v3290, 20985);
+ int16x8_t v3292 = vaddq_s16(v3286, v3291);
+ int16x8_t v3293 = vsubq_s16(v2345, v2350);
+ int16x8_t v3294 = vsubq_s16(v2355, v2360);
+ int16x8_t v3295_tmp = vqrdmulhq_n_s16(v3294, 9242);
+ int16x8_t v3295 = vmlaq_n_s16(v3295_tmp, v3294, 2);
+ int16x8_t v3296 = vaddq_s16(v3293, v3295);
+ int16x8_t v3297 = vsubq_s16(v2367, v2372);
+ int16x8_t v3298 = vsubq_s16(v2377, v2382);
+ int16x8_t v3299_tmp = vqrdmulhq_n_s16(v3298, 9242);
+ int16x8_t v3299 = vmlaq_n_s16(v3299_tmp, v3298, 2);
+ int16x8_t v3300 = vaddq_s16(v3297, v3299);
+ int16x8_t v3301 = vqrdmulhq_n_s16(v3300, 20985);
+ int16x8_t v3302 = vaddq_s16(v3296, v3301);
+ int16x8_t v3303 = vqrdmulhq_n_s16(v3302, 17363);
+ int16x8_t v3304 = vaddq_s16(v3292, v3303);
+ int16x8_t v3305 = vsubq_s16(v2115, v2126);
+ int16x8_t v3306 = vsubq_s16(v2137, v2148);
+ int16x8_t v3307_tmp = vqrdmulhq_n_s16(v3306, 30298);
+ int16x8_t v3307 = vmlaq_n_s16(v3307_tmp, v3306, 2);
+ int16x8_t v3308 = vaddq_s16(v3305, v3307);
+ int16x8_t v3309 = vsubq_s16(v2161, v2172);
+ int16x8_t v3310 = vsubq_s16(v2183, v2194);
+ int16x8_t v3311_tmp = vqrdmulhq_n_s16(v3310, 30298);
+ int16x8_t v3311 = vmlaq_n_s16(v3311_tmp, v3310, 2);
+ int16x8_t v3312 = vaddq_s16(v3309, v3311);
+ int16x8_t v3313 = vqrdmulhq_n_s16(v3312, 21412);
+ int16x8_t v3314 = vaddq_s16(v3308, v3313);
+ int16x8_t v3315 = vsubq_s16(v2209, v2220);
+ int16x8_t v3316 = vsubq_s16(v2231, v2242);
+ int16x8_t v3317_tmp = vqrdmulhq_n_s16(v3316, 30298);
+ int16x8_t v3317 = vmlaq_n_s16(v3317_tmp, v3316, 2);
+ int16x8_t v3318 = vaddq_s16(v3315, v3317);
+ int16x8_t v3319 = vsubq_s16(v2255, v2266);
+ int16x8_t v3320 = vsubq_s16(v2277, v2288);
+ int16x8_t v3321_tmp = vqrdmulhq_n_s16(v3320, 30298);
+ int16x8_t v3321 = vmlaq_n_s16(v3321_tmp, v3320, 2);
+ int16x8_t v3322 = vaddq_s16(v3319, v3321);
+ int16x8_t v3323 = vqrdmulhq_n_s16(v3322, 21412);
+ int16x8_t v3324 = vaddq_s16(v3318, v3323);
+ int16x8_t v3325 = vqrdmulhq_n_s16(v3324, 17440);
+ int16x8_t v3326 = vaddq_s16(v3314, v3325);
+ int16x8_t v3327 = vsubq_s16(v1925, v1936);
+ int16x8_t v3328 = vsubq_s16(v1947, v1958);
+ int16x8_t v3329_tmp = vqrdmulhq_n_s16(v3328, 2773);
+ int16x8_t v3329 = vmlaq_n_s16(v3329_tmp, v3328, 4);
+ int16x8_t v3330 = vaddq_s16(v3327, v3329);
+ int16x8_t v3331 = vsubq_s16(v1971, v1982);
+ int16x8_t v3332 = vsubq_s16(v1993, v2004);
+ int16x8_t v3333_tmp = vqrdmulhq_n_s16(v3332, 2773);
+ int16x8_t v3333 = vmlaq_n_s16(v3333_tmp, v3332, 4);
+ int16x8_t v3334 = vaddq_s16(v3331, v3333);
+ int16x8_t v3335 = vqrdmulhq_n_s16(v3334, 21871);
+ int16x8_t v3336 = vaddq_s16(v3330, v3335);
+ int16x8_t v3337 = vsubq_s16(v2019, v2030);
+ int16x8_t v3338 = vsubq_s16(v2041, v2052);
+ int16x8_t v3339_tmp = vqrdmulhq_n_s16(v3338, 2773);
+ int16x8_t v3339 = vmlaq_n_s16(v3339_tmp, v3338, 4);
+ int16x8_t v3340 = vaddq_s16(v3337, v3339);
+ int16x8_t v3341 = vsubq_s16(v2065, v2076);
+ int16x8_t v3342 = vsubq_s16(v2087, v2098);
+ int16x8_t v3343_tmp = vqrdmulhq_n_s16(v3342, 2773);
+ int16x8_t v3343 = vmlaq_n_s16(v3343_tmp, v3342, 4);
+ int16x8_t v3344 = vaddq_s16(v3341, v3343);
+ int16x8_t v3345 = vqrdmulhq_n_s16(v3344, 21871);
+ int16x8_t v3346 = vaddq_s16(v3340, v3345);
+ int16x8_t v3347 = vqrdmulhq_n_s16(v3346, 17520);
+ int16x8_t v3348 = vaddq_s16(v3336, v3347);
+ int16x8_t v3349 = vsubq_s16(v1555, v1578);
+ int16x8_t v3350 = vsubq_s16(v1601, v1624);
+ int16x8_t v3351_tmp = vqrdmulhq_n_s16(v3350, 26108);
+ int16x8_t v3351 = vmlaq_n_s16(v3351_tmp, v3350, 6);
+ int16x8_t v3352 = vaddq_s16(v3349, v3351);
+ int16x8_t v3353 = vsubq_s16(v1649, v1672);
+ int16x8_t v3354 = vsubq_s16(v1695, v1718);
+ int16x8_t v3355_tmp = vqrdmulhq_n_s16(v3354, 26108);
+ int16x8_t v3355 = vmlaq_n_s16(v3355_tmp, v3354, 6);
+ int16x8_t v3356 = vaddq_s16(v3353, v3355);
+ int16x8_t v3357 = vqrdmulhq_n_s16(v3356, 22363);
+ int16x8_t v3358 = vaddq_s16(v3352, v3357);
+ int16x8_t v3359 = vsubq_s16(v1745, v1768);
+ int16x8_t v3360 = vsubq_s16(v1791, v1814);
+ int16x8_t v3361_tmp = vqrdmulhq_n_s16(v3360, 26108);
+ int16x8_t v3361 = vmlaq_n_s16(v3361_tmp, v3360, 6);
+ int16x8_t v3362 = vaddq_s16(v3359, v3361);
+ int16x8_t v3363 = vsubq_s16(v1839, v1862);
+ int16x8_t v3364 = vsubq_s16(v1885, v1908);
+ int16x8_t v3365_tmp = vqrdmulhq_n_s16(v3364, 26108);
+ int16x8_t v3365 = vmlaq_n_s16(v3365_tmp, v3364, 6);
+ int16x8_t v3366 = vaddq_s16(v3363, v3365);
+ int16x8_t v3367 = vqrdmulhq_n_s16(v3366, 22363);
+ int16x8_t v3368 = vaddq_s16(v3362, v3367);
+ int16x8_t v3369 = vqrdmulhq_n_s16(v3368, 17603);
+ int16x8_t v3370 = vaddq_s16(v3358, v3369);
+ int16x8_t v3371 = vsubq_s16(v61, v140);
+ int16x8_t v3372 = vsubq_s16(v234, v314);
+ int16x8_t v3373_tmp = vqrdmulhq_n_s16(v3372, 12251);
+ int16x8_t v3373 = vmlaq_n_s16(v3373_tmp, v3372, 20);
+ int16x8_t v3374 = vaddq_s16(v3371, v3373);
+ int16x8_t v3375 = vsubq_s16(v410, v521);
+ int16x8_t v3376 = vsubq_s16(v615, v696);
+ int16x8_t v3377_tmp = vqrdmulhq_n_s16(v3376, 12251);
+ int16x8_t v3377 = vmlaq_n_s16(v3377_tmp, v3376, 20);
+ int16x8_t v3378 = vaddq_s16(v3375, v3377);
+ int16x8_t v3379 = vqrdmulhq_n_s16(v3378, 22891);
+ int16x8_t v3380 = vaddq_s16(v3374, v3379);
+ int16x8_t v3381 = vsubq_s16(v794, v905);
+ int16x8_t v3382 = vsubq_s16(v1061, v1143);
+ int16x8_t v3383_tmp = vqrdmulhq_n_s16(v3382, 12251);
+ int16x8_t v3383 = vmlaq_n_s16(v3383_tmp, v3382, 20);
+ int16x8_t v3384 = vaddq_s16(v3381, v3383);
+ int16x8_t v3385 = vsubq_s16(v1239, v1350);
+ int16x8_t v3386 = vsubq_s16(v1444, v1526);
+ int16x8_t v3387_tmp = vqrdmulhq_n_s16(v3386, 12251);
+ int16x8_t v3387 = vmlaq_n_s16(v3387_tmp, v3386, 20);
+ int16x8_t v3388 = vaddq_s16(v3385, v3387);
+ int16x8_t v3389 = vqrdmulhq_n_s16(v3388, 22891);
+ int16x8_t v3390 = vaddq_s16(v3384, v3389);
+ int16x8_t v3391 = vqrdmulhq_n_s16(v3390, 17689);
+ int16x8_t v3392 = vaddq_s16(v3380, v3391);
+ int16x8_t v3393 = vsubq_s16(v3371, v3373);
+ int16x8_t v3394 = vsubq_s16(v3375, v3377);
+ int16x8_t v3395 = vqrdmulhq_n_s16(v3394, 23460);
+ int16x8_t v3396 = vaddq_s16(v3393, v3395);
+ int16x8_t v3397 = vsubq_s16(v3381, v3383);
+ int16x8_t v3398 = vsubq_s16(v3385, v3387);
+ int16x8_t v3399 = vqrdmulhq_n_s16(v3398, 23460);
+ int16x8_t v3400 = vaddq_s16(v3397, v3399);
+ int16x8_t v3401 = vqrdmulhq_n_s16(v3400, 17779);
+ int16x8_t v3402 = vaddq_s16(v3396, v3401);
+ int16x8_t v3403 = vsubq_s16(v3349, v3351);
+ int16x8_t v3404 = vsubq_s16(v3353, v3355);
+ int16x8_t v3405 = vqrdmulhq_n_s16(v3404, 24073);
+ int16x8_t v3406 = vaddq_s16(v3403, v3405);
+ int16x8_t v3407 = vsubq_s16(v3359, v3361);
+ int16x8_t v3408 = vsubq_s16(v3363, v3365);
+ int16x8_t v3409 = vqrdmulhq_n_s16(v3408, 24073);
+ int16x8_t v3410 = vaddq_s16(v3407, v3409);
+ int16x8_t v3411 = vqrdmulhq_n_s16(v3410, 17873);
+ int16x8_t v3412 = vaddq_s16(v3406, v3411);
+ int16x8_t v3413 = vsubq_s16(v3327, v3329);
+ int16x8_t v3414 = vsubq_s16(v3331, v3333);
+ int16x8_t v3415 = vqrdmulhq_n_s16(v3414, 24734);
+ int16x8_t v3416 = vaddq_s16(v3413, v3415);
+ int16x8_t v3417 = vsubq_s16(v3337, v3339);
+ int16x8_t v3418 = vsubq_s16(v3341, v3343);
+ int16x8_t v3419 = vqrdmulhq_n_s16(v3418, 24734);
+ int16x8_t v3420 = vaddq_s16(v3417, v3419);
+ int16x8_t v3421 = vqrdmulhq_n_s16(v3420, 17971);
+ int16x8_t v3422 = vaddq_s16(v3416, v3421);
+ int16x8_t v3423 = vsubq_s16(v3305, v3307);
+ int16x8_t v3424 = vsubq_s16(v3309, v3311);
+ int16x8_t v3425 = vqrdmulhq_n_s16(v3424, 25448);
+ int16x8_t v3426 = vaddq_s16(v3423, v3425);
+ int16x8_t v3427 = vsubq_s16(v3315, v3317);
+ int16x8_t v3428 = vsubq_s16(v3319, v3321);
+ int16x8_t v3429 = vqrdmulhq_n_s16(v3428, 25448);
+ int16x8_t v3430 = vaddq_s16(v3427, v3429);
+ int16x8_t v3431 = vqrdmulhq_n_s16(v3430, 18072);
+ int16x8_t v3432 = vaddq_s16(v3426, v3431);
+ int16x8_t v3433 = vsubq_s16(v3283, v3285);
+ int16x8_t v3434 = vsubq_s16(v3287, v3289);
+ int16x8_t v3435 = vqrdmulhq_n_s16(v3434, 26220);
+ int16x8_t v3436 = vaddq_s16(v3433, v3435);
+ int16x8_t v3437 = vsubq_s16(v3293, v3295);
+ int16x8_t v3438 = vsubq_s16(v3297, v3299);
+ int16x8_t v3439 = vqrdmulhq_n_s16(v3438, 26220);
+ int16x8_t v3440 = vaddq_s16(v3437, v3439);
+ int16x8_t v3441 = vqrdmulhq_n_s16(v3440, 18177);
+ int16x8_t v3442 = vaddq_s16(v3436, v3441);
+ int16x8_t v3443 = vsubq_s16(v3261, v3263);
+ int16x8_t v3444 = vsubq_s16(v3265, v3267);
+ int16x8_t v3445 = vqrdmulhq_n_s16(v3444, 27058);
+ int16x8_t v3446 = vaddq_s16(v3443, v3445);
+ int16x8_t v3447 = vsubq_s16(v3271, v3273);
+ int16x8_t v3448 = vsubq_s16(v3275, v3277);
+ int16x8_t v3449 = vqrdmulhq_n_s16(v3448, 27058);
+ int16x8_t v3450 = vaddq_s16(v3447, v3449);
+ int16x8_t v3451 = vqrdmulhq_n_s16(v3450, 18286);
+ int16x8_t v3452 = vaddq_s16(v3446, v3451);
+ int16x8_t v3453 = vsubq_s16(v3239, v3241);
+ int16x8_t v3454 = vsubq_s16(v3243, v3245);
+ int16x8_t v3455 = vqrdmulhq_n_s16(v3454, 27969);
+ int16x8_t v3456 = vaddq_s16(v3453, v3455);
+ int16x8_t v3457 = vsubq_s16(v3249, v3251);
+ int16x8_t v3458 = vsubq_s16(v3253, v3255);
+ int16x8_t v3459 = vqrdmulhq_n_s16(v3458, 27969);
+ int16x8_t v3460 = vaddq_s16(v3457, v3459);
+ int16x8_t v3461 = vqrdmulhq_n_s16(v3460, 18400);
+ int16x8_t v3462 = vaddq_s16(v3456, v3461);
+ int16x8_t v3463 = vsubq_s16(v3217, v3219);
+ int16x8_t v3464 = vsubq_s16(v3221, v3223);
+ int16x8_t v3465 = vqrdmulhq_n_s16(v3464, 28961);
+ int16x8_t v3466 = vaddq_s16(v3463, v3465);
+ int16x8_t v3467 = vsubq_s16(v3227, v3229);
+ int16x8_t v3468 = vsubq_s16(v3231, v3233);
+ int16x8_t v3469 = vqrdmulhq_n_s16(v3468, 28961);
+ int16x8_t v3470 = vaddq_s16(v3467, v3469);
+ int16x8_t v3471 = vqrdmulhq_n_s16(v3470, 18517);
+ int16x8_t v3472 = vaddq_s16(v3466, v3471);
+ int16x8_t v3473 = vsubq_s16(v3195, v3197);
+ int16x8_t v3474 = vsubq_s16(v3199, v3201);
+ int16x8_t v3475 = vqrdmulhq_n_s16(v3474, 30044);
+ int16x8_t v3476 = vaddq_s16(v3473, v3475);
+ int16x8_t v3477 = vsubq_s16(v3205, v3207);
+ int16x8_t v3478 = vsubq_s16(v3209, v3211);
+ int16x8_t v3479 = vqrdmulhq_n_s16(v3478, 30044);
+ int16x8_t v3480 = vaddq_s16(v3477, v3479);
+ int16x8_t v3481 = vqrdmulhq_n_s16(v3480, 18639);
+ int16x8_t v3482 = vaddq_s16(v3476, v3481);
+ int16x8_t v3483 = vsubq_s16(v3173, v3175);
+ int16x8_t v3484 = vsubq_s16(v3177, v3179);
+ int16x8_t v3485 = vqrdmulhq_n_s16(v3484, 31232);
+ int16x8_t v3486 = vaddq_s16(v3483, v3485);
+ int16x8_t v3487 = vsubq_s16(v3183, v3185);
+ int16x8_t v3488 = vsubq_s16(v3187, v3189);
+ int16x8_t v3489 = vqrdmulhq_n_s16(v3488, 31232);
+ int16x8_t v3490 = vaddq_s16(v3487, v3489);
+ int16x8_t v3491 = vqrdmulhq_n_s16(v3490, 18765);
+ int16x8_t v3492 = vaddq_s16(v3486, v3491);
+ int16x8_t v3493 = vsubq_s16(v3151, v3153);
+ int16x8_t v3494 = vsubq_s16(v3155, v3157);
+ int16x8_t v3495 = vqrdmulhq_n_s16(v3494, 32538);
+ int16x8_t v3496 = vaddq_s16(v3493, v3495);
+ int16x8_t v3497 = vsubq_s16(v3161, v3163);
+ int16x8_t v3498 = vsubq_s16(v3165, v3167);
+ int16x8_t v3499 = vqrdmulhq_n_s16(v3498, 32538);
+ int16x8_t v3500 = vaddq_s16(v3497, v3499);
+ int16x8_t v3501 = vqrdmulhq_n_s16(v3500, 18896);
+ int16x8_t v3502 = vaddq_s16(v3496, v3501);
+ int16x8_t v3503 = vsubq_s16(v3129, v3131);
+ int16x8_t v3504 = vsubq_s16(v3133, v3135);
+ int16x8_t v3505_tmp = vqrdmulhq_n_s16(v3504, 1211);
+ int16x8_t v3505 = vaddq_s16(v3505_tmp, v3504);
+ int16x8_t v3506 = vaddq_s16(v3503, v3505);
+ int16x8_t v3507 = vsubq_s16(v3139, v3141);
+ int16x8_t v3508 = vsubq_s16(v3143, v3145);
+ int16x8_t v3509_tmp = vqrdmulhq_n_s16(v3508, 1211);
+ int16x8_t v3509 = vaddq_s16(v3509_tmp, v3508);
+ int16x8_t v3510 = vaddq_s16(v3507, v3509);
+ int16x8_t v3511 = vqrdmulhq_n_s16(v3510, 19032);
+ int16x8_t v3512 = vaddq_s16(v3506, v3511);
+ int16x8_t v3513 = vsubq_s16(v3107, v3109);
+ int16x8_t v3514 = vsubq_s16(v3111, v3113);
+ int16x8_t v3515_tmp = vqrdmulhq_n_s16(v3514, 2808);
+ int16x8_t v3515 = vaddq_s16(v3515_tmp, v3514);
+ int16x8_t v3516 = vaddq_s16(v3513, v3515);
+ int16x8_t v3517 = vsubq_s16(v3117, v3119);
+ int16x8_t v3518 = vsubq_s16(v3121, v3123);
+ int16x8_t v3519_tmp = vqrdmulhq_n_s16(v3518, 2808);
+ int16x8_t v3519 = vaddq_s16(v3519_tmp, v3518);
+ int16x8_t v3520 = vaddq_s16(v3517, v3519);
+ int16x8_t v3521 = vqrdmulhq_n_s16(v3520, 19172);
+ int16x8_t v3522 = vaddq_s16(v3516, v3521);
+ int16x8_t v3523 = vsubq_s16(v3085, v3087);
+ int16x8_t v3524 = vsubq_s16(v3089, v3091);
+ int16x8_t v3525_tmp = vqrdmulhq_n_s16(v3524, 4586);
+ int16x8_t v3525 = vaddq_s16(v3525_tmp, v3524);
+ int16x8_t v3526 = vaddq_s16(v3523, v3525);
+ int16x8_t v3527 = vsubq_s16(v3095, v3097);
+ int16x8_t v3528 = vsubq_s16(v3099, v3101);
+ int16x8_t v3529_tmp = vqrdmulhq_n_s16(v3528, 4586);
+ int16x8_t v3529 = vaddq_s16(v3529_tmp, v3528);
+ int16x8_t v3530 = vaddq_s16(v3527, v3529);
+ int16x8_t v3531 = vqrdmulhq_n_s16(v3530, 19318);
+ int16x8_t v3532 = vaddq_s16(v3526, v3531);
+ int16x8_t v3533 = vsubq_s16(v3063, v3065);
+ int16x8_t v3534 = vsubq_s16(v3067, v3069);
+ int16x8_t v3535_tmp = vqrdmulhq_n_s16(v3534, 6576);
+ int16x8_t v3535 = vaddq_s16(v3535_tmp, v3534);
+ int16x8_t v3536 = vaddq_s16(v3533, v3535);
+ int16x8_t v3537 = vsubq_s16(v3073, v3075);
+ int16x8_t v3538 = vsubq_s16(v3077, v3079);
+ int16x8_t v3539_tmp = vqrdmulhq_n_s16(v3538, 6576);
+ int16x8_t v3539 = vaddq_s16(v3539_tmp, v3538);
+ int16x8_t v3540 = vaddq_s16(v3537, v3539);
+ int16x8_t v3541 = vqrdmulhq_n_s16(v3540, 19469);
+ int16x8_t v3542 = vaddq_s16(v3536, v3541);
+ int16x8_t v3543 = vsubq_s16(v3041, v3043);
+ int16x8_t v3544 = vsubq_s16(v3045, v3047);
+ int16x8_t v3545_tmp = vqrdmulhq_n_s16(v3544, 8817);
+ int16x8_t v3545 = vaddq_s16(v3545_tmp, v3544);
+ int16x8_t v3546 = vaddq_s16(v3543, v3545);
+ int16x8_t v3547 = vsubq_s16(v3051, v3053);
+ int16x8_t v3548 = vsubq_s16(v3055, v3057);
+ int16x8_t v3549_tmp = vqrdmulhq_n_s16(v3548, 8817);
+ int16x8_t v3549 = vaddq_s16(v3549_tmp, v3548);
+ int16x8_t v3550 = vaddq_s16(v3547, v3549);
+ int16x8_t v3551 = vqrdmulhq_n_s16(v3550, 19625);
+ int16x8_t v3552 = vaddq_s16(v3546, v3551);
+ int16x8_t v3553 = vsubq_s16(v2998, v3003);
+ int16x8_t v3554 = vsubq_s16(v3008, v3013);
+ int16x8_t v3555_tmp = vqrdmulhq_n_s16(v3554, 11356);
+ int16x8_t v3555 = vaddq_s16(v3555_tmp, v3554);
+ int16x8_t v3556 = vaddq_s16(v3553, v3555);
+ int16x8_t v3557 = vsubq_s16(v3020, v3025);
+ int16x8_t v3558 = vsubq_s16(v3030, v3035);
+ int16x8_t v3559_tmp = vqrdmulhq_n_s16(v3558, 11356);
+ int16x8_t v3559 = vaddq_s16(v3559_tmp, v3558);
+ int16x8_t v3560 = vaddq_s16(v3557, v3559);
+ int16x8_t v3561 = vqrdmulhq_n_s16(v3560, 19786);
+ int16x8_t v3562 = vaddq_s16(v3556, v3561);
+ int16x8_t v3563 = vsubq_s16(v2952, v2957);
+ int16x8_t v3564 = vsubq_s16(v2962, v2967);
+ int16x8_t v3565_tmp = vqrdmulhq_n_s16(v3564, 14256);
+ int16x8_t v3565 = vaddq_s16(v3565_tmp, v3564);
+ int16x8_t v3566 = vaddq_s16(v3563, v3565);
+ int16x8_t v3567 = vsubq_s16(v2974, v2979);
+ int16x8_t v3568 = vsubq_s16(v2984, v2989);
+ int16x8_t v3569_tmp = vqrdmulhq_n_s16(v3568, 14256);
+ int16x8_t v3569 = vaddq_s16(v3569_tmp, v3568);
+ int16x8_t v3570 = vaddq_s16(v3567, v3569);
+ int16x8_t v3571 = vqrdmulhq_n_s16(v3570, 19954);
+ int16x8_t v3572 = vaddq_s16(v3566, v3571);
+ int16x8_t v3573 = vsubq_s16(v2906, v2911);
+ int16x8_t v3574 = vsubq_s16(v2916, v2921);
+ int16x8_t v3575_tmp = vqrdmulhq_n_s16(v3574, 17596);
+ int16x8_t v3575 = vaddq_s16(v3575_tmp, v3574);
+ int16x8_t v3576 = vaddq_s16(v3573, v3575);
+ int16x8_t v3577 = vsubq_s16(v2928, v2933);
+ int16x8_t v3578 = vsubq_s16(v2938, v2943);
+ int16x8_t v3579_tmp = vqrdmulhq_n_s16(v3578, 17596);
+ int16x8_t v3579 = vaddq_s16(v3579_tmp, v3578);
+ int16x8_t v3580 = vaddq_s16(v3577, v3579);
+ int16x8_t v3581 = vqrdmulhq_n_s16(v3580, 20127);
+ int16x8_t v3582 = vaddq_s16(v3576, v3581);
+ int16x8_t v3583 = vsubq_s16(v2860, v2865);
+ int16x8_t v3584 = vsubq_s16(v2870, v2875);
+ int16x8_t v3585_tmp = vqrdmulhq_n_s16(v3584, 21483);
+ int16x8_t v3585 = vaddq_s16(v3585_tmp, v3584);
+ int16x8_t v3586 = vaddq_s16(v3583, v3585);
+ int16x8_t v3587 = vsubq_s16(v2882, v2887);
+ int16x8_t v3588 = vsubq_s16(v2892, v2897);
+ int16x8_t v3589_tmp = vqrdmulhq_n_s16(v3588, 21483);
+ int16x8_t v3589 = vaddq_s16(v3589_tmp, v3588);
+ int16x8_t v3590 = vaddq_s16(v3587, v3589);
+ int16x8_t v3591 = vqrdmulhq_n_s16(v3590, 20306);
+ int16x8_t v3592 = vaddq_s16(v3586, v3591);
+ int16x8_t v3593 = vsubq_s16(v2814, v2819);
+ int16x8_t v3594 = vsubq_s16(v2824, v2829);
+ int16x8_t v3595_tmp = vqrdmulhq_n_s16(v3594, 26057);
+ int16x8_t v3595 = vaddq_s16(v3595_tmp, v3594);
+ int16x8_t v3596 = vaddq_s16(v3593, v3595);
+ int16x8_t v3597 = vsubq_s16(v2836, v2841);
+ int16x8_t v3598 = vsubq_s16(v2846, v2851);
+ int16x8_t v3599_tmp = vqrdmulhq_n_s16(v3598, 26057);
+ int16x8_t v3599 = vaddq_s16(v3599_tmp, v3598);
+ int16x8_t v3600 = vaddq_s16(v3597, v3599);
+ int16x8_t v3601 = vqrdmulhq_n_s16(v3600, 20492);
+ int16x8_t v3602 = vaddq_s16(v3596, v3601);
+ int16x8_t v3603 = vsubq_s16(v2768, v2773);
+ int16x8_t v3604 = vsubq_s16(v2778, v2783);
+ int16x8_t v3605_tmp = vqrdmulhq_n_s16(v3604, 31517);
+ int16x8_t v3605 = vaddq_s16(v3605_tmp, v3604);
+ int16x8_t v3606 = vaddq_s16(v3603, v3605);
+ int16x8_t v3607 = vsubq_s16(v2790, v2795);
+ int16x8_t v3608 = vsubq_s16(v2800, v2805);
+ int16x8_t v3609_tmp = vqrdmulhq_n_s16(v3608, 31517);
+ int16x8_t v3609 = vaddq_s16(v3609_tmp, v3608);
+ int16x8_t v3610 = vaddq_s16(v3607, v3609);
+ int16x8_t v3611 = vqrdmulhq_n_s16(v3610, 20684);
+ int16x8_t v3612 = vaddq_s16(v3606, v3611);
+ int16x8_t v3613 = vsubq_s16(v2722, v2727);
+ int16x8_t v3614 = vsubq_s16(v2732, v2737);
+ int16x8_t v3615_tmp = vqrdmulhq_n_s16(v3614, 5373);
+ int16x8_t v3615 = vmlaq_n_s16(v3615_tmp, v3614, 2);
+ int16x8_t v3616 = vaddq_s16(v3613, v3615);
+ int16x8_t v3617 = vsubq_s16(v2744, v2749);
+ int16x8_t v3618 = vsubq_s16(v2754, v2759);
+ int16x8_t v3619_tmp = vqrdmulhq_n_s16(v3618, 5373);
+ int16x8_t v3619 = vmlaq_n_s16(v3619_tmp, v3618, 2);
+ int16x8_t v3620 = vaddq_s16(v3617, v3619);
+ int16x8_t v3621 = vqrdmulhq_n_s16(v3620, 20883);
+ int16x8_t v3622 = vaddq_s16(v3616, v3621);
+ int16x8_t v3623 = vsubq_s16(v2676, v2681);
+ int16x8_t v3624 = vsubq_s16(v2686, v2691);
+ int16x8_t v3625_tmp = vqrdmulhq_n_s16(v3624, 13571);
+ int16x8_t v3625 = vmlaq_n_s16(v3625_tmp, v3624, 2);
+ int16x8_t v3626 = vaddq_s16(v3623, v3625);
+ int16x8_t v3627 = vsubq_s16(v2698, v2703);
+ int16x8_t v3628 = vsubq_s16(v2708, v2713);
+ int16x8_t v3629_tmp = vqrdmulhq_n_s16(v3628, 13571);
+ int16x8_t v3629 = vmlaq_n_s16(v3629_tmp, v3628, 2);
+ int16x8_t v3630 = vaddq_s16(v3627, v3629);
+ int16x8_t v3631 = vqrdmulhq_n_s16(v3630, 21089);
+ int16x8_t v3632 = vaddq_s16(v3626, v3631);
+ int16x8_t v3633 = vsubq_s16(v2588, v2599);
+ int16x8_t v3634 = vsubq_s16(v2610, v2621);
+ int16x8_t v3635_tmp = vqrdmulhq_n_s16(v3634, 23975);
+ int16x8_t v3635 = vmlaq_n_s16(v3635_tmp, v3634, 2);
+ int16x8_t v3636 = vaddq_s16(v3633, v3635);
+ int16x8_t v3637 = vsubq_s16(v2634, v2645);
+ int16x8_t v3638 = vsubq_s16(v2656, v2667);
+ int16x8_t v3639_tmp = vqrdmulhq_n_s16(v3638, 23975);
+ int16x8_t v3639 = vmlaq_n_s16(v3639_tmp, v3638, 2);
+ int16x8_t v3640 = vaddq_s16(v3637, v3639);
+ int16x8_t v3641 = vqrdmulhq_n_s16(v3640, 21303);
+ int16x8_t v3642 = vaddq_s16(v3636, v3641);
+ int16x8_t v3643 = vsubq_s16(v2494, v2505);
+ int16x8_t v3644 = vsubq_s16(v2516, v2527);
+ int16x8_t v3645_tmp = vqrdmulhq_n_s16(v3644, 4832);
+ int16x8_t v3645 = vmlaq_n_s16(v3645_tmp, v3644, 3);
+ int16x8_t v3646 = vaddq_s16(v3643, v3645);
+ int16x8_t v3647 = vsubq_s16(v2540, v2551);
+ int16x8_t v3648 = vsubq_s16(v2562, v2573);
+ int16x8_t v3649_tmp = vqrdmulhq_n_s16(v3648, 4832);
+ int16x8_t v3649 = vmlaq_n_s16(v3649_tmp, v3648, 3);
+ int16x8_t v3650 = vaddq_s16(v3647, v3649);
+ int16x8_t v3651 = vqrdmulhq_n_s16(v3650, 21524);
+ int16x8_t v3652 = vaddq_s16(v3646, v3651);
+ int16x8_t v3653 = vsubq_s16(v2399, v2410);
+ int16x8_t v3654 = vsubq_s16(v2421, v2432);
+ int16x8_t v3655_tmp = vqrdmulhq_n_s16(v3654, 23437);
+ int16x8_t v3655 = vmlaq_n_s16(v3655_tmp, v3654, 3);
+ int16x8_t v3656 = vaddq_s16(v3653, v3655);
+ int16x8_t v3657 = vsubq_s16(v2445, v2456);
+ int16x8_t v3658 = vsubq_s16(v2468, v2479);
+ int16x8_t v3659_tmp = vqrdmulhq_n_s16(v3658, 23437);
+ int16x8_t v3659 = vmlaq_n_s16(v3659_tmp, v3658, 3);
+ int16x8_t v3660 = vaddq_s16(v3657, v3659);
+ int16x8_t v3661 = vqrdmulhq_n_s16(v3660, 21753);
+ int16x8_t v3662 = vaddq_s16(v3656, v3661);
+ int16x8_t v3663 = vsubq_s16(v2305, v2316);
+ int16x8_t v3664 = vsubq_s16(v2327, v2338);
+ int16x8_t v3665_tmp = vqrdmulhq_n_s16(v3664, 17573);
+ int16x8_t v3665 = vmlaq_n_s16(v3665_tmp, v3664, 4);
+ int16x8_t v3666 = vaddq_s16(v3663, v3665);
+ int16x8_t v3667 = vsubq_s16(v2351, v2362);
+ int16x8_t v3668 = vsubq_s16(v2373, v2384);
+ int16x8_t v3669_tmp = vqrdmulhq_n_s16(v3668, 17573);
+ int16x8_t v3669 = vmlaq_n_s16(v3669_tmp, v3668, 4);
+ int16x8_t v3670 = vaddq_s16(v3667, v3669);
+ int16x8_t v3671 = vqrdmulhq_n_s16(v3670, 21990);
+ int16x8_t v3672 = vaddq_s16(v3666, v3671);
+ int16x8_t v3673 = vsubq_s16(v2127, v2150);
+ int16x8_t v3674 = vsubq_s16(v2173, v2196);
+ int16x8_t v3675_tmp = vqrdmulhq_n_s16(v3674, 27122);
+ int16x8_t v3675 = vmlaq_n_s16(v3675_tmp, v3674, 5);
+ int16x8_t v3676 = vaddq_s16(v3673, v3675);
+ int16x8_t v3677 = vsubq_s16(v2221, v2244);
+ int16x8_t v3678 = vsubq_s16(v2267, v2290);
+ int16x8_t v3679_tmp = vqrdmulhq_n_s16(v3678, 27122);
+ int16x8_t v3679 = vmlaq_n_s16(v3679_tmp, v3678, 5);
+ int16x8_t v3680 = vaddq_s16(v3677, v3679);
+ int16x8_t v3681 = vqrdmulhq_n_s16(v3680, 22236);
+ int16x8_t v3682 = vaddq_s16(v3676, v3681);
+ int16x8_t v3683 = vsubq_s16(v1937, v1960);
+ int16x8_t v3684 = vsubq_s16(v1983, v2006);
+ int16x8_t v3685_tmp = vqrdmulhq_n_s16(v3684, 5041);
+ int16x8_t v3685 = vmlaq_n_s16(v3685_tmp, v3684, 8);
+ int16x8_t v3686 = vaddq_s16(v3683, v3685);
+ int16x8_t v3687 = vsubq_s16(v2031, v2054);
+ int16x8_t v3688 = vsubq_s16(v2077, v2100);
+ int16x8_t v3689_tmp = vqrdmulhq_n_s16(v3688, 5041);
+ int16x8_t v3689 = vmlaq_n_s16(v3689_tmp, v3688, 8);
+ int16x8_t v3690 = vaddq_s16(v3687, v3689);
+ int16x8_t v3691 = vqrdmulhq_n_s16(v3690, 22491);
+ int16x8_t v3692 = vaddq_s16(v3686, v3691);
+ int16x8_t v3693 = vsubq_s16(v1579, v1626);
+ int16x8_t v3694 = vsubq_s16(v1673, v1720);
+ int16x8_t v3695_tmp = vqrdmulhq_n_s16(v3694, 19146);
+ int16x8_t v3695 = vmlaq_n_s16(v3695_tmp, v3694, 13);
+ int16x8_t v3696 = vaddq_s16(v3693, v3695);
+ int16x8_t v3697 = vsubq_s16(v1769, v1816);
+ int16x8_t v3698 = vsubq_s16(v1863, v1910);
+ int16x8_t v3699_tmp = vqrdmulhq_n_s16(v3698, 19146);
+ int16x8_t v3699 = vmlaq_n_s16(v3699_tmp, v3698, 13);
+ int16x8_t v3700 = vaddq_s16(v3697, v3699);
+ int16x8_t v3701 = vqrdmulhq_n_s16(v3700, 22755);
+ int16x8_t v3702 = vaddq_s16(v3696, v3701);
+ int16x8_t v3703 = vsubq_s16(v141, v316);
+ int16x8_t v3704 = vsubq_s16(v522, v698);
+ int16x8_t v3705_tmp = vqrdmulhq_n_s16(v3704, 24402);
+ int16x8_t v3705 = vmlaq_n_s16(v3705_tmp, v3704, 40);
+ int16x8_t v3706 = vaddq_s16(v3703, v3705);
+ int16x8_t v3707 = vsubq_s16(v906, v1145);
+ int16x8_t v3708 = vsubq_s16(v1351, v1528);
+ int16x8_t v3709_tmp = vqrdmulhq_n_s16(v3708, 24402);
+ int16x8_t v3709 = vmlaq_n_s16(v3709_tmp, v3708, 40);
+ int16x8_t v3710 = vaddq_s16(v3707, v3709);
+ int16x8_t v3711 = vqrdmulhq_n_s16(v3710, 23030);
+ int16x8_t v3712 = vaddq_s16(v3706, v3711);
+ int16x8_t v3713 = vsubq_s16(v3703, v3705);
+ int16x8_t v3714 = vsubq_s16(v3707, v3709);
+ int16x8_t v3715 = vqrdmulhq_n_s16(v3714, 23314);
+ int16x8_t v3716 = vaddq_s16(v3713, v3715);
+ int16x8_t v3717 = vsubq_s16(v3693, v3695);
+ int16x8_t v3718 = vsubq_s16(v3697, v3699);
+ int16x8_t v3719 = vqrdmulhq_n_s16(v3718, 23609);
+ int16x8_t v3720 = vaddq_s16(v3717, v3719);
+ int16x8_t v3721 = vsubq_s16(v3683, v3685);
+ int16x8_t v3722 = vsubq_s16(v3687, v3689);
+ int16x8_t v3723 = vqrdmulhq_n_s16(v3722, 23915);
+ int16x8_t v3724 = vaddq_s16(v3721, v3723);
+ int16x8_t v3725 = vsubq_s16(v3673, v3675);
+ int16x8_t v3726 = vsubq_s16(v3677, v3679);
+ int16x8_t v3727 = vqrdmulhq_n_s16(v3726, 24233);
+ int16x8_t v3728 = vaddq_s16(v3725, v3727);
+ int16x8_t v3729 = vsubq_s16(v3663, v3665);
+ int16x8_t v3730 = vsubq_s16(v3667, v3669);
+ int16x8_t v3731 = vqrdmulhq_n_s16(v3730, 24564);
+ int16x8_t v3732 = vaddq_s16(v3729, v3731);
+ int16x8_t v3733 = vsubq_s16(v3653, v3655);
+ int16x8_t v3734 = vsubq_s16(v3657, v3659);
+ int16x8_t v3735 = vqrdmulhq_n_s16(v3734, 24907);
+ int16x8_t v3736 = vaddq_s16(v3733, v3735);
+ int16x8_t v3737 = vsubq_s16(v3643, v3645);
+ int16x8_t v3738 = vsubq_s16(v3647, v3649);
+ int16x8_t v3739 = vqrdmulhq_n_s16(v3738, 25264);
+ int16x8_t v3740 = vaddq_s16(v3737, v3739);
+ int16x8_t v3741 = vsubq_s16(v3633, v3635);
+ int16x8_t v3742 = vsubq_s16(v3637, v3639);
+ int16x8_t v3743 = vqrdmulhq_n_s16(v3742, 25635);
+ int16x8_t v3744 = vaddq_s16(v3741, v3743);
+ int16x8_t v3745 = vsubq_s16(v3623, v3625);
+ int16x8_t v3746 = vsubq_s16(v3627, v3629);
+ int16x8_t v3747 = vqrdmulhq_n_s16(v3746, 26021);
+ int16x8_t v3748 = vaddq_s16(v3745, v3747);
+ int16x8_t v3749 = vsubq_s16(v3613, v3615);
+ int16x8_t v3750 = vsubq_s16(v3617, v3619);
+ int16x8_t v3751 = vqrdmulhq_n_s16(v3750, 26423);
+ int16x8_t v3752 = vaddq_s16(v3749, v3751);
+ int16x8_t v3753 = vsubq_s16(v3603, v3605);
+ int16x8_t v3754 = vsubq_s16(v3607, v3609);
+ int16x8_t v3755 = vqrdmulhq_n_s16(v3754, 26842);
+ int16x8_t v3756 = vaddq_s16(v3753, v3755);
+ int16x8_t v3757 = vsubq_s16(v3593, v3595);
+ int16x8_t v3758 = vsubq_s16(v3597, v3599);
+ int16x8_t v3759 = vqrdmulhq_n_s16(v3758, 27279);
+ int16x8_t v3760 = vaddq_s16(v3757, v3759);
+ int16x8_t v3761 = vsubq_s16(v3583, v3585);
+ int16x8_t v3762 = vsubq_s16(v3587, v3589);
+ int16x8_t v3763 = vqrdmulhq_n_s16(v3762, 27734);
+ int16x8_t v3764 = vaddq_s16(v3761, v3763);
+ int16x8_t v3765 = vsubq_s16(v3573, v3575);
+ int16x8_t v3766 = vsubq_s16(v3577, v3579);
+ int16x8_t v3767 = vqrdmulhq_n_s16(v3766, 28209);
+ int16x8_t v3768 = vaddq_s16(v3765, v3767);
+ int16x8_t v3769 = vsubq_s16(v3563, v3565);
+ int16x8_t v3770 = vsubq_s16(v3567, v3569);
+ int16x8_t v3771 = vqrdmulhq_n_s16(v3770, 28705);
+ int16x8_t v3772 = vaddq_s16(v3769, v3771);
+ int16x8_t v3773 = vsubq_s16(v3553, v3555);
+ int16x8_t v3774 = vsubq_s16(v3557, v3559);
+ int16x8_t v3775 = vqrdmulhq_n_s16(v3774, 29223);
+ int16x8_t v3776 = vaddq_s16(v3773, v3775);
+ int16x8_t v3777 = vsubq_s16(v3543, v3545);
+ int16x8_t v3778 = vsubq_s16(v3547, v3549);
+ int16x8_t v3779 = vqrdmulhq_n_s16(v3778, 29764);
+ int16x8_t v3780 = vaddq_s16(v3777, v3779);
+ int16x8_t v3781 = vsubq_s16(v3533, v3535);
+ int16x8_t v3782 = vsubq_s16(v3537, v3539);
+ int16x8_t v3783 = vqrdmulhq_n_s16(v3782, 30331);
+ int16x8_t v3784 = vaddq_s16(v3781, v3783);
+ int16x8_t v3785 = vsubq_s16(v3523, v3525);
+ int16x8_t v3786 = vsubq_s16(v3527, v3529);
+ int16x8_t v3787 = vqrdmulhq_n_s16(v3786, 30925);
+ int16x8_t v3788 = vaddq_s16(v3785, v3787);
+ int16x8_t v3789 = vsubq_s16(v3513, v3515);
+ int16x8_t v3790 = vsubq_s16(v3517, v3519);
+ int16x8_t v3791 = vqrdmulhq_n_s16(v3790, 31547);
+ int16x8_t v3792 = vaddq_s16(v3789, v3791);
+ int16x8_t v3793 = vsubq_s16(v3503, v3505);
+ int16x8_t v3794 = vsubq_s16(v3507, v3509);
+ int16x8_t v3795 = vqrdmulhq_n_s16(v3794, 32199);
+ int16x8_t v3796 = vaddq_s16(v3793, v3795);
+ int16x8_t v3797 = vsubq_s16(v3493, v3495);
+ int16x8_t v3798 = vsubq_s16(v3497, v3499);
+ int16x8_t v3799_tmp = vqrdmulhq_n_s16(v3798, 117);
+ int16x8_t v3799 = vaddq_s16(v3799_tmp, v3798);
+ int16x8_t v3800 = vaddq_s16(v3797, v3799);
+ int16x8_t v3801 = vsubq_s16(v3483, v3485);
+ int16x8_t v3802 = vsubq_s16(v3487, v3489);
+ int16x8_t v3803_tmp = vqrdmulhq_n_s16(v3802, 837);
+ int16x8_t v3803 = vaddq_s16(v3803_tmp, v3802);
+ int16x8_t v3804 = vaddq_s16(v3801, v3803);
+ int16x8_t v3805 = vsubq_s16(v3473, v3475);
+ int16x8_t v3806 = vsubq_s16(v3477, v3479);
+ int16x8_t v3807_tmp = vqrdmulhq_n_s16(v3806, 1594);
+ int16x8_t v3807 = vaddq_s16(v3807_tmp, v3806);
+ int16x8_t v3808 = vaddq_s16(v3805, v3807);
+ int16x8_t v3809 = vsubq_s16(v3463, v3465);
+ int16x8_t v3810 = vsubq_s16(v3467, v3469);
+ int16x8_t v3811_tmp = vqrdmulhq_n_s16(v3810, 2393);
+ int16x8_t v3811 = vaddq_s16(v3811_tmp, v3810);
+ int16x8_t v3812 = vaddq_s16(v3809, v3811);
+ int16x8_t v3813 = vsubq_s16(v3453, v3455);
+ int16x8_t v3814 = vsubq_s16(v3457, v3459);
+ int16x8_t v3815_tmp = vqrdmulhq_n_s16(v3814, 3234);
+ int16x8_t v3815 = vaddq_s16(v3815_tmp, v3814);
+ int16x8_t v3816 = vaddq_s16(v3813, v3815);
+ int16x8_t v3817 = vsubq_s16(v3443, v3445);
+ int16x8_t v3818 = vsubq_s16(v3447, v3449);
+ int16x8_t v3819_tmp = vqrdmulhq_n_s16(v3818, 4123);
+ int16x8_t v3819 = vaddq_s16(v3819_tmp, v3818);
+ int16x8_t v3820 = vaddq_s16(v3817, v3819);
+ int16x8_t v3821 = vsubq_s16(v3433, v3435);
+ int16x8_t v3822 = vsubq_s16(v3437, v3439);
+ int16x8_t v3823_tmp = vqrdmulhq_n_s16(v3822, 5062);
+ int16x8_t v3823 = vaddq_s16(v3823_tmp, v3822);
+ int16x8_t v3824 = vaddq_s16(v3821, v3823);
+ int16x8_t v3825 = vsubq_s16(v3423, v3425);
+ int16x8_t v3826 = vsubq_s16(v3427, v3429);
+ int16x8_t v3827_tmp = vqrdmulhq_n_s16(v3826, 6057);
+ int16x8_t v3827 = vaddq_s16(v3827_tmp, v3826);
+ int16x8_t v3828 = vaddq_s16(v3825, v3827);
+ int16x8_t v3829 = vsubq_s16(v3413, v3415);
+ int16x8_t v3830 = vsubq_s16(v3417, v3419);
+ int16x8_t v3831_tmp = vqrdmulhq_n_s16(v3830, 7111);
+ int16x8_t v3831 = vaddq_s16(v3831_tmp, v3830);
+ int16x8_t v3832 = vaddq_s16(v3829, v3831);
+ int16x8_t v3833 = vsubq_s16(v3403, v3405);
+ int16x8_t v3834 = vsubq_s16(v3407, v3409);
+ int16x8_t v3835_tmp = vqrdmulhq_n_s16(v3834, 8231);
+ int16x8_t v3835 = vaddq_s16(v3835_tmp, v3834);
+ int16x8_t v3836 = vaddq_s16(v3833, v3835);
+ int16x8_t v3837 = vsubq_s16(v3393, v3395);
+ int16x8_t v3838 = vsubq_s16(v3397, v3399);
+ int16x8_t v3839_tmp = vqrdmulhq_n_s16(v3838, 9421);
+ int16x8_t v3839 = vaddq_s16(v3839_tmp, v3838);
+ int16x8_t v3840 = vaddq_s16(v3837, v3839);
+ int16x8_t v3841 = vsubq_s16(v3374, v3379);
+ int16x8_t v3842 = vsubq_s16(v3384, v3389);
+ int16x8_t v3843_tmp = vqrdmulhq_n_s16(v3842, 10690);
+ int16x8_t v3843 = vaddq_s16(v3843_tmp, v3842);
+ int16x8_t v3844 = vaddq_s16(v3841, v3843);
+ int16x8_t v3845 = vsubq_s16(v3352, v3357);
+ int16x8_t v3846 = vsubq_s16(v3362, v3367);
+ int16x8_t v3847_tmp = vqrdmulhq_n_s16(v3846, 12044);
+ int16x8_t v3847 = vaddq_s16(v3847_tmp, v3846);
+ int16x8_t v3848 = vaddq_s16(v3845, v3847);
+ int16x8_t v3849 = vsubq_s16(v3330, v3335);
+ int16x8_t v3850 = vsubq_s16(v3340, v3345);
+ int16x8_t v3851_tmp = vqrdmulhq_n_s16(v3850, 13493);
+ int16x8_t v3851 = vaddq_s16(v3851_tmp, v3850);
+ int16x8_t v3852 = vaddq_s16(v3849, v3851);
+ int16x8_t v3853 = vsubq_s16(v3308, v3313);
+ int16x8_t v3854 = vsubq_s16(v3318, v3323);
+ int16x8_t v3855_tmp = vqrdmulhq_n_s16(v3854, 15046);
+ int16x8_t v3855 = vaddq_s16(v3855_tmp, v3854);
+ int16x8_t v3856 = vaddq_s16(v3853, v3855);
+ int16x8_t v3857 = vsubq_s16(v3286, v3291);
+ int16x8_t v3858 = vsubq_s16(v3296, v3301);
+ int16x8_t v3859_tmp = vqrdmulhq_n_s16(v3858, 16715);
+ int16x8_t v3859 = vaddq_s16(v3859_tmp, v3858);
+ int16x8_t v3860 = vaddq_s16(v3857, v3859);
+ int16x8_t v3861 = vsubq_s16(v3264, v3269);
+ int16x8_t v3862 = vsubq_s16(v3274, v3279);
+ int16x8_t v3863_tmp = vqrdmulhq_n_s16(v3862, 18512);
+ int16x8_t v3863 = vaddq_s16(v3863_tmp, v3862);
+ int16x8_t v3864 = vaddq_s16(v3861, v3863);
+ int16x8_t v3865 = vsubq_s16(v3242, v3247);
+ int16x8_t v3866 = vsubq_s16(v3252, v3257);
+ int16x8_t v3867_tmp = vqrdmulhq_n_s16(v3866, 20453);
+ int16x8_t v3867 = vaddq_s16(v3867_tmp, v3866);
+ int16x8_t v3868 = vaddq_s16(v3865, v3867);
+ int16x8_t v3869 = vsubq_s16(v3220, v3225);
+ int16x8_t v3870 = vsubq_s16(v3230, v3235);
+ int16x8_t v3871_tmp = vqrdmulhq_n_s16(v3870, 22555);
+ int16x8_t v3871 = vaddq_s16(v3871_tmp, v3870);
+ int16x8_t v3872 = vaddq_s16(v3869, v3871);
+ int16x8_t v3873 = vsubq_s16(v3198, v3203);
+ int16x8_t v3874 = vsubq_s16(v3208, v3213);
+ int16x8_t v3875_tmp = vqrdmulhq_n_s16(v3874, 24839);
+ int16x8_t v3875 = vaddq_s16(v3875_tmp, v3874);
+ int16x8_t v3876 = vaddq_s16(v3873, v3875);
+ int16x8_t v3877 = vsubq_s16(v3176, v3181);
+ int16x8_t v3878 = vsubq_s16(v3186, v3191);
+ int16x8_t v3879_tmp = vqrdmulhq_n_s16(v3878, 27330);
+ int16x8_t v3879 = vaddq_s16(v3879_tmp, v3878);
+ int16x8_t v3880 = vaddq_s16(v3877, v3879);
+ int16x8_t v3881 = vsubq_s16(v3154, v3159);
+ int16x8_t v3882 = vsubq_s16(v3164, v3169);
+ int16x8_t v3883_tmp = vqrdmulhq_n_s16(v3882, 30056);
+ int16x8_t v3883 = vaddq_s16(v3883_tmp, v3882);
+ int16x8_t v3884 = vaddq_s16(v3881, v3883);
+ int16x8_t v3885 = vsubq_s16(v3132, v3137);
+ int16x8_t v3886 = vsubq_s16(v3142, v3147);
+ int16x8_t v3887_tmp = vqrdmulhq_n_s16(v3886, 282);
+ int16x8_t v3887 = vmlaq_n_s16(v3887_tmp, v3886, 2);
+ int16x8_t v3888 = vaddq_s16(v3885, v3887);
+ int16x8_t v3889 = vsubq_s16(v3110, v3115);
+ int16x8_t v3890 = vsubq_s16(v3120, v3125);
+ int16x8_t v3891_tmp = vqrdmulhq_n_s16(v3890, 3588);
+ int16x8_t v3891 = vmlaq_n_s16(v3891_tmp, v3890, 2);
+ int16x8_t v3892 = vaddq_s16(v3889, v3891);
+ int16x8_t v3893 = vsubq_s16(v3088, v3093);
+ int16x8_t v3894 = vsubq_s16(v3098, v3103);
+ int16x8_t v3895_tmp = vqrdmulhq_n_s16(v3894, 7255);
+ int16x8_t v3895 = vmlaq_n_s16(v3895_tmp, v3894, 2);
+ int16x8_t v3896 = vaddq_s16(v3893, v3895);
+ int16x8_t v3897 = vsubq_s16(v3066, v3071);
+ int16x8_t v3898 = vsubq_s16(v3076, v3081);
+ int16x8_t v3899_tmp = vqrdmulhq_n_s16(v3898, 11344);
+ int16x8_t v3899 = vmlaq_n_s16(v3899_tmp, v3898, 2);
+ int16x8_t v3900 = vaddq_s16(v3897, v3899);
+ int16x8_t v3901 = vsubq_s16(v3044, v3049);
+ int16x8_t v3902 = vsubq_s16(v3054, v3059);
+ int16x8_t v3903_tmp = vqrdmulhq_n_s16(v3902, 15934);
+ int16x8_t v3903 = vmlaq_n_s16(v3903_tmp, v3902, 2);
+ int16x8_t v3904 = vaddq_s16(v3901, v3903);
+ int16x8_t v3905 = vsubq_s16(v3004, v3015);
+ int16x8_t v3906 = vsubq_s16(v3026, v3037);
+ int16x8_t v3907_tmp = vqrdmulhq_n_s16(v3906, 21120);
+ int16x8_t v3907 = vmlaq_n_s16(v3907_tmp, v3906, 2);
+ int16x8_t v3908 = vaddq_s16(v3905, v3907);
+ int16x8_t v3909 = vsubq_s16(v2958, v2969);
+ int16x8_t v3910 = vsubq_s16(v2980, v2991);
+ int16x8_t v3911_tmp = vqrdmulhq_n_s16(v3910, 27027);
+ int16x8_t v3911 = vmlaq_n_s16(v3911_tmp, v3910, 2);
+ int16x8_t v3912 = vaddq_s16(v3909, v3911);
+ int16x8_t v3913 = vsubq_s16(v2912, v2923);
+ int16x8_t v3914 = vsubq_s16(v2934, v2945);
+ int16x8_t v3915_tmp = vqrdmulhq_n_s16(v3914, 1045);
+ int16x8_t v3915 = vmlaq_n_s16(v3915_tmp, v3914, 3);
+ int16x8_t v3916 = vaddq_s16(v3913, v3915);
+ int16x8_t v3917 = vsubq_s16(v2866, v2877);
+ int16x8_t v3918 = vsubq_s16(v2888, v2899);
+ int16x8_t v3919_tmp = vqrdmulhq_n_s16(v3918, 8923);
+ int16x8_t v3919 = vmlaq_n_s16(v3919_tmp, v3918, 3);
+ int16x8_t v3920 = vaddq_s16(v3917, v3919);
+ int16x8_t v3921 = vsubq_s16(v2820, v2831);
+ int16x8_t v3922 = vsubq_s16(v2842, v2853);
+ int16x8_t v3923_tmp = vqrdmulhq_n_s16(v3922, 18177);
+ int16x8_t v3923 = vmlaq_n_s16(v3923_tmp, v3922, 3);
+ int16x8_t v3924 = vaddq_s16(v3921, v3923);
+ int16x8_t v3925 = vsubq_s16(v2774, v2785);
+ int16x8_t v3926 = vsubq_s16(v2796, v2807);
+ int16x8_t v3927_tmp = vqrdmulhq_n_s16(v3926, 29200);
+ int16x8_t v3927 = vmlaq_n_s16(v3927_tmp, v3926, 3);
+ int16x8_t v3928 = vaddq_s16(v3925, v3927);
+ int16x8_t v3929 = vsubq_s16(v2728, v2739);
+ int16x8_t v3930 = vsubq_s16(v2750, v2761);
+ int16x8_t v3931_tmp = vqrdmulhq_n_s16(v3930, 9782);
+ int16x8_t v3931 = vmlaq_n_s16(v3931_tmp, v3930, 4);
+ int16x8_t v3932 = vaddq_s16(v3929, v3931);
+ int16x8_t v3933 = vsubq_s16(v2682, v2693);
+ int16x8_t v3934 = vsubq_s16(v2704, v2715);
+ int16x8_t v3935_tmp = vqrdmulhq_n_s16(v3934, 26282);
+ int16x8_t v3935 = vmlaq_n_s16(v3935_tmp, v3934, 4);
+ int16x8_t v3936 = vaddq_s16(v3933, v3935);
+ int16x8_t v3937 = vsubq_s16(v2600, v2623);
+ int16x8_t v3938 = vsubq_s16(v2646, v2669);
+ int16x8_t v3939_tmp = vqrdmulhq_n_s16(v3938, 14423);
+ int16x8_t v3939 = vmlaq_n_s16(v3939_tmp, v3938, 5);
+ int16x8_t v3940 = vaddq_s16(v3937, v3939);
+ int16x8_t v3941 = vsubq_s16(v2506, v2529);
+ int16x8_t v3942 = vsubq_s16(v2552, v2575);
+ int16x8_t v3943_tmp = vqrdmulhq_n_s16(v3942, 9008);
+ int16x8_t v3943 = vmlaq_n_s16(v3943_tmp, v3942, 6);
+ int16x8_t v3944 = vaddq_s16(v3941, v3943);
+ int16x8_t v3945 = vsubq_s16(v2411, v2434);
+ int16x8_t v3946 = vsubq_s16(v2457, v2481);
+ int16x8_t v3947_tmp = vqrdmulhq_n_s16(v3946, 13552);
+ int16x8_t v3947 = vmlaq_n_s16(v3947_tmp, v3946, 7);
+ int16x8_t v3948 = vaddq_s16(v3945, v3947);
+ int16x8_t v3949 = vsubq_s16(v2317, v2340);
+ int16x8_t v3950 = vsubq_s16(v2363, v2386);
+ int16x8_t v3951_tmp = vqrdmulhq_n_s16(v3950, 1925);
+ int16x8_t v3951 = vmlaq_n_s16(v3951_tmp, v3950, 9);
+ int16x8_t v3952 = vaddq_s16(v3949, v3951);
+ int16x8_t v3953 = vsubq_s16(v2151, v2198);
+ int16x8_t v3954 = vsubq_s16(v2245, v2292);
+ int16x8_t v3955_tmp = vqrdmulhq_n_s16(v3954, 21123);
+ int16x8_t v3955 = vmlaq_n_s16(v3955_tmp, v3954, 11);
+ int16x8_t v3956 = vaddq_s16(v3953, v3955);
+ int16x8_t v3957 = vsubq_s16(v1961, v2008);
+ int16x8_t v3958 = vsubq_s16(v2055, v2102);
+ int16x8_t v3959_tmp = vqrdmulhq_n_s16(v3958, 9831);
+ int16x8_t v3959 = vmlaq_n_s16(v3959_tmp, v3958, 16);
+ int16x8_t v3960 = vaddq_s16(v3957, v3959);
+ int16x8_t v3961 = vsubq_s16(v1627, v1722);
+ int16x8_t v3962 = vsubq_s16(v1817, v1912);
+ int16x8_t v3963_tmp = vqrdmulhq_n_s16(v3962, 5373);
+ int16x8_t v3963 = vmlaq_n_s16(v3963_tmp, v3962, 27);
+ int16x8_t v3964 = vaddq_s16(v3961, v3963);
+ int16x8_t v3965 = vsubq_s16(v317, v700);
+ int16x8_t v3966 = vsubq_s16(v1146, v1530);
+ int16x8_t v3967_tmp = vqrdmulhq_n_s16(v3966, 15986);
+ int16x8_t v3967 = vmlaq_n_s16(v3967_tmp, v3966, 81);
+ int16x8_t v3968 = vaddq_s16(v3965, v3967);
+ int16x8_t v3969 = vsubq_s16(v3965, v3967);
+ int16x8_t v3970 = vsubq_s16(v3961, v3963);
+ int16x8_t v3971 = vsubq_s16(v3957, v3959);
+ int16x8_t v3972 = vsubq_s16(v3953, v3955);
+ int16x8_t v3973 = vsubq_s16(v3949, v3951);
+ int16x8_t v3974 = vsubq_s16(v3945, v3947);
+ int16x8_t v3975 = vsubq_s16(v3941, v3943);
+ int16x8_t v3976 = vsubq_s16(v3937, v3939);
+ int16x8_t v3977 = vsubq_s16(v3933, v3935);
+ int16x8_t v3978 = vsubq_s16(v3929, v3931);
+ int16x8_t v3979 = vsubq_s16(v3925, v3927);
+ int16x8_t v3980 = vsubq_s16(v3921, v3923);
+ int16x8_t v3981 = vsubq_s16(v3917, v3919);
+ int16x8_t v3982 = vsubq_s16(v3913, v3915);
+ int16x8_t v3983 = vsubq_s16(v3909, v3911);
+ int16x8_t v3984 = vsubq_s16(v3905, v3907);
+ int16x8_t v3985 = vsubq_s16(v3901, v3903);
+ int16x8_t v3986 = vsubq_s16(v3897, v3899);
+ int16x8_t v3987 = vsubq_s16(v3893, v3895);
+ int16x8_t v3988 = vsubq_s16(v3889, v3891);
+ int16x8_t v3989 = vsubq_s16(v3885, v3887);
+ int16x8_t v3990 = vsubq_s16(v3881, v3883);
+ int16x8_t v3991 = vsubq_s16(v3877, v3879);
+ int16x8_t v3992 = vsubq_s16(v3873, v3875);
+ int16x8_t v3993 = vsubq_s16(v3869, v3871);
+ int16x8_t v3994 = vsubq_s16(v3865, v3867);
+ int16x8_t v3995 = vsubq_s16(v3861, v3863);
+ int16x8_t v3996 = vsubq_s16(v3857, v3859);
+ int16x8_t v3997 = vsubq_s16(v3853, v3855);
+ int16x8_t v3998 = vsubq_s16(v3849, v3851);
+ int16x8_t v3999 = vsubq_s16(v3845, v3847);
+ int16x8_t v4000 = vsubq_s16(v3841, v3843);
+ int16x8_t v4001 = vsubq_s16(v3837, v3839);
+ int16x8_t v4002 = vsubq_s16(v3833, v3835);
+ int16x8_t v4003 = vsubq_s16(v3829, v3831);
+ int16x8_t v4004 = vsubq_s16(v3825, v3827);
+ int16x8_t v4005 = vsubq_s16(v3821, v3823);
+ int16x8_t v4006 = vsubq_s16(v3817, v3819);
+ int16x8_t v4007 = vsubq_s16(v3813, v3815);
+ int16x8_t v4008 = vsubq_s16(v3809, v3811);
+ int16x8_t v4009 = vsubq_s16(v3805, v3807);
+ int16x8_t v4010 = vsubq_s16(v3801, v3803);
+ int16x8_t v4011 = vsubq_s16(v3797, v3799);
+ int16x8_t v4012 = vsubq_s16(v3793, v3795);
+ int16x8_t v4013 = vsubq_s16(v3789, v3791);
+ int16x8_t v4014 = vsubq_s16(v3785, v3787);
+ int16x8_t v4015 = vsubq_s16(v3781, v3783);
+ int16x8_t v4016 = vsubq_s16(v3777, v3779);
+ int16x8_t v4017 = vsubq_s16(v3773, v3775);
+ int16x8_t v4018 = vsubq_s16(v3769, v3771);
+ int16x8_t v4019 = vsubq_s16(v3765, v3767);
+ int16x8_t v4020 = vsubq_s16(v3761, v3763);
+ int16x8_t v4021 = vsubq_s16(v3757, v3759);
+ int16x8_t v4022 = vsubq_s16(v3753, v3755);
+ int16x8_t v4023 = vsubq_s16(v3749, v3751);
+ int16x8_t v4024 = vsubq_s16(v3745, v3747);
+ int16x8_t v4025 = vsubq_s16(v3741, v3743);
+ int16x8_t v4026 = vsubq_s16(v3737, v3739);
+ int16x8_t v4027 = vsubq_s16(v3733, v3735);
+ int16x8_t v4028 = vsubq_s16(v3729, v3731);
+ int16x8_t v4029 = vsubq_s16(v3725, v3727);
+ int16x8_t v4030 = vsubq_s16(v3721, v3723);
+ int16x8_t v4031 = vsubq_s16(v3717, v3719);
+ int16x8_t v4032 = vsubq_s16(v3713, v3715);
+ int16x8_t v4033 = vsubq_s16(v3706, v3711);
+ int16x8_t v4034 = vsubq_s16(v3696, v3701);
+ int16x8_t v4035 = vsubq_s16(v3686, v3691);
+ int16x8_t v4036 = vsubq_s16(v3676, v3681);
+ int16x8_t v4037 = vsubq_s16(v3666, v3671);
+ int16x8_t v4038 = vsubq_s16(v3656, v3661);
+ int16x8_t v4039 = vsubq_s16(v3646, v3651);
+ int16x8_t v4040 = vsubq_s16(v3636, v3641);
+ int16x8_t v4041 = vsubq_s16(v3626, v3631);
+ int16x8_t v4042 = vsubq_s16(v3616, v3621);
+ int16x8_t v4043 = vsubq_s16(v3606, v3611);
+ int16x8_t v4044 = vsubq_s16(v3596, v3601);
+ int16x8_t v4045 = vsubq_s16(v3586, v3591);
+ int16x8_t v4046 = vsubq_s16(v3576, v3581);
+ int16x8_t v4047 = vsubq_s16(v3566, v3571);
+ int16x8_t v4048 = vsubq_s16(v3556, v3561);
+ int16x8_t v4049 = vsubq_s16(v3546, v3551);
+ int16x8_t v4050 = vsubq_s16(v3536, v3541);
+ int16x8_t v4051 = vsubq_s16(v3526, v3531);
+ int16x8_t v4052 = vsubq_s16(v3516, v3521);
+ int16x8_t v4053 = vsubq_s16(v3506, v3511);
+ int16x8_t v4054 = vsubq_s16(v3496, v3501);
+ int16x8_t v4055 = vsubq_s16(v3486, v3491);
+ int16x8_t v4056 = vsubq_s16(v3476, v3481);
+ int16x8_t v4057 = vsubq_s16(v3466, v3471);
+ int16x8_t v4058 = vsubq_s16(v3456, v3461);
+ int16x8_t v4059 = vsubq_s16(v3446, v3451);
+ int16x8_t v4060 = vsubq_s16(v3436, v3441);
+ int16x8_t v4061 = vsubq_s16(v3426, v3431);
+ int16x8_t v4062 = vsubq_s16(v3416, v3421);
+ int16x8_t v4063 = vsubq_s16(v3406, v3411);
+ int16x8_t v4064 = vsubq_s16(v3396, v3401);
+ int16x8_t v4065 = vsubq_s16(v3380, v3391);
+ int16x8_t v4066 = vsubq_s16(v3358, v3369);
+ int16x8_t v4067 = vsubq_s16(v3336, v3347);
+ int16x8_t v4068 = vsubq_s16(v3314, v3325);
+ int16x8_t v4069 = vsubq_s16(v3292, v3303);
+ int16x8_t v4070 = vsubq_s16(v3270, v3281);
+ int16x8_t v4071 = vsubq_s16(v3248, v3259);
+ int16x8_t v4072 = vsubq_s16(v3226, v3237);
+ int16x8_t v4073 = vsubq_s16(v3204, v3215);
+ int16x8_t v4074 = vsubq_s16(v3182, v3193);
+ int16x8_t v4075 = vsubq_s16(v3160, v3171);
+ int16x8_t v4076 = vsubq_s16(v3138, v3149);
+ int16x8_t v4077 = vsubq_s16(v3116, v3127);
+ int16x8_t v4078 = vsubq_s16(v3094, v3105);
+ int16x8_t v4079 = vsubq_s16(v3072, v3083);
+ int16x8_t v4080 = vsubq_s16(v3050, v3061);
+ int16x8_t v4081 = vsubq_s16(v3016, v3039);
+ int16x8_t v4082 = vsubq_s16(v2970, v2993);
+ int16x8_t v4083 = vsubq_s16(v2924, v2947);
+ int16x8_t v4084 = vsubq_s16(v2878, v2901);
+ int16x8_t v4085 = vsubq_s16(v2832, v2855);
+ int16x8_t v4086 = vsubq_s16(v2786, v2809);
+ int16x8_t v4087 = vsubq_s16(v2740, v2763);
+ int16x8_t v4088 = vsubq_s16(v2694, v2717);
+ int16x8_t v4089 = vsubq_s16(v2624, v2671);
+ int16x8_t v4090 = vsubq_s16(v2530, v2577);
+ int16x8_t v4091 = vsubq_s16(v2435, v2483);
+ int16x8_t v4092 = vsubq_s16(v2341, v2388);
+ int16x8_t v4093 = vsubq_s16(v2199, v2294);
+ int16x8_t v4094 = vsubq_s16(v2009, v2104);
+ int16x8_t v4095 = vsubq_s16(v1723, v1914);
+ int16x8_t v4096 = vsubq_s16(v701, v1532);
+ vst1q_s16(out + out_stride * 0 + i, v1533);
+ vst1q_s16(out + out_stride * 1 + i, v1915);
+ vst1q_s16(out + out_stride * 2 + i, v2105);
+ vst1q_s16(out + out_stride * 3 + i, v2295);
+ vst1q_s16(out + out_stride * 4 + i, v2389);
+ vst1q_s16(out + out_stride * 5 + i, v2484);
+ vst1q_s16(out + out_stride * 6 + i, v2578);
+ vst1q_s16(out + out_stride * 7 + i, v2672);
+ vst1q_s16(out + out_stride * 8 + i, v2718);
+ vst1q_s16(out + out_stride * 9 + i, v2764);
+ vst1q_s16(out + out_stride * 10 + i, v2810);
+ vst1q_s16(out + out_stride * 11 + i, v2856);
+ vst1q_s16(out + out_stride * 12 + i, v2902);
+ vst1q_s16(out + out_stride * 13 + i, v2948);
+ vst1q_s16(out + out_stride * 14 + i, v2994);
+ vst1q_s16(out + out_stride * 15 + i, v3040);
+ vst1q_s16(out + out_stride * 16 + i, v3062);
+ vst1q_s16(out + out_stride * 17 + i, v3084);
+ vst1q_s16(out + out_stride * 18 + i, v3106);
+ vst1q_s16(out + out_stride * 19 + i, v3128);
+ vst1q_s16(out + out_stride * 20 + i, v3150);
+ vst1q_s16(out + out_stride * 21 + i, v3172);
+ vst1q_s16(out + out_stride * 22 + i, v3194);
+ vst1q_s16(out + out_stride * 23 + i, v3216);
+ vst1q_s16(out + out_stride * 24 + i, v3238);
+ vst1q_s16(out + out_stride * 25 + i, v3260);
+ vst1q_s16(out + out_stride * 26 + i, v3282);
+ vst1q_s16(out + out_stride * 27 + i, v3304);
+ vst1q_s16(out + out_stride * 28 + i, v3326);
+ vst1q_s16(out + out_stride * 29 + i, v3348);
+ vst1q_s16(out + out_stride * 30 + i, v3370);
+ vst1q_s16(out + out_stride * 31 + i, v3392);
+ vst1q_s16(out + out_stride * 32 + i, v3402);
+ vst1q_s16(out + out_stride * 33 + i, v3412);
+ vst1q_s16(out + out_stride * 34 + i, v3422);
+ vst1q_s16(out + out_stride * 35 + i, v3432);
+ vst1q_s16(out + out_stride * 36 + i, v3442);
+ vst1q_s16(out + out_stride * 37 + i, v3452);
+ vst1q_s16(out + out_stride * 38 + i, v3462);
+ vst1q_s16(out + out_stride * 39 + i, v3472);
+ vst1q_s16(out + out_stride * 40 + i, v3482);
+ vst1q_s16(out + out_stride * 41 + i, v3492);
+ vst1q_s16(out + out_stride * 42 + i, v3502);
+ vst1q_s16(out + out_stride * 43 + i, v3512);
+ vst1q_s16(out + out_stride * 44 + i, v3522);
+ vst1q_s16(out + out_stride * 45 + i, v3532);
+ vst1q_s16(out + out_stride * 46 + i, v3542);
+ vst1q_s16(out + out_stride * 47 + i, v3552);
+ vst1q_s16(out + out_stride * 48 + i, v3562);
+ vst1q_s16(out + out_stride * 49 + i, v3572);
+ vst1q_s16(out + out_stride * 50 + i, v3582);
+ vst1q_s16(out + out_stride * 51 + i, v3592);
+ vst1q_s16(out + out_stride * 52 + i, v3602);
+ vst1q_s16(out + out_stride * 53 + i, v3612);
+ vst1q_s16(out + out_stride * 54 + i, v3622);
+ vst1q_s16(out + out_stride * 55 + i, v3632);
+ vst1q_s16(out + out_stride * 56 + i, v3642);
+ vst1q_s16(out + out_stride * 57 + i, v3652);
+ vst1q_s16(out + out_stride * 58 + i, v3662);
+ vst1q_s16(out + out_stride * 59 + i, v3672);
+ vst1q_s16(out + out_stride * 60 + i, v3682);
+ vst1q_s16(out + out_stride * 61 + i, v3692);
+ vst1q_s16(out + out_stride * 62 + i, v3702);
+ vst1q_s16(out + out_stride * 63 + i, v3712);
+ vst1q_s16(out + out_stride * 64 + i, v3716);
+ vst1q_s16(out + out_stride * 65 + i, v3720);
+ vst1q_s16(out + out_stride * 66 + i, v3724);
+ vst1q_s16(out + out_stride * 67 + i, v3728);
+ vst1q_s16(out + out_stride * 68 + i, v3732);
+ vst1q_s16(out + out_stride * 69 + i, v3736);
+ vst1q_s16(out + out_stride * 70 + i, v3740);
+ vst1q_s16(out + out_stride * 71 + i, v3744);
+ vst1q_s16(out + out_stride * 72 + i, v3748);
+ vst1q_s16(out + out_stride * 73 + i, v3752);
+ vst1q_s16(out + out_stride * 74 + i, v3756);
+ vst1q_s16(out + out_stride * 75 + i, v3760);
+ vst1q_s16(out + out_stride * 76 + i, v3764);
+ vst1q_s16(out + out_stride * 77 + i, v3768);
+ vst1q_s16(out + out_stride * 78 + i, v3772);
+ vst1q_s16(out + out_stride * 79 + i, v3776);
+ vst1q_s16(out + out_stride * 80 + i, v3780);
+ vst1q_s16(out + out_stride * 81 + i, v3784);
+ vst1q_s16(out + out_stride * 82 + i, v3788);
+ vst1q_s16(out + out_stride * 83 + i, v3792);
+ vst1q_s16(out + out_stride * 84 + i, v3796);
+ vst1q_s16(out + out_stride * 85 + i, v3800);
+ vst1q_s16(out + out_stride * 86 + i, v3804);
+ vst1q_s16(out + out_stride * 87 + i, v3808);
+ vst1q_s16(out + out_stride * 88 + i, v3812);
+ vst1q_s16(out + out_stride * 89 + i, v3816);
+ vst1q_s16(out + out_stride * 90 + i, v3820);
+ vst1q_s16(out + out_stride * 91 + i, v3824);
+ vst1q_s16(out + out_stride * 92 + i, v3828);
+ vst1q_s16(out + out_stride * 93 + i, v3832);
+ vst1q_s16(out + out_stride * 94 + i, v3836);
+ vst1q_s16(out + out_stride * 95 + i, v3840);
+ vst1q_s16(out + out_stride * 96 + i, v3844);
+ vst1q_s16(out + out_stride * 97 + i, v3848);
+ vst1q_s16(out + out_stride * 98 + i, v3852);
+ vst1q_s16(out + out_stride * 99 + i, v3856);
+ vst1q_s16(out + out_stride * 100 + i, v3860);
+ vst1q_s16(out + out_stride * 101 + i, v3864);
+ vst1q_s16(out + out_stride * 102 + i, v3868);
+ vst1q_s16(out + out_stride * 103 + i, v3872);
+ vst1q_s16(out + out_stride * 104 + i, v3876);
+ vst1q_s16(out + out_stride * 105 + i, v3880);
+ vst1q_s16(out + out_stride * 106 + i, v3884);
+ vst1q_s16(out + out_stride * 107 + i, v3888);
+ vst1q_s16(out + out_stride * 108 + i, v3892);
+ vst1q_s16(out + out_stride * 109 + i, v3896);
+ vst1q_s16(out + out_stride * 110 + i, v3900);
+ vst1q_s16(out + out_stride * 111 + i, v3904);
+ vst1q_s16(out + out_stride * 112 + i, v3908);
+ vst1q_s16(out + out_stride * 113 + i, v3912);
+ vst1q_s16(out + out_stride * 114 + i, v3916);
+ vst1q_s16(out + out_stride * 115 + i, v3920);
+ vst1q_s16(out + out_stride * 116 + i, v3924);
+ vst1q_s16(out + out_stride * 117 + i, v3928);
+ vst1q_s16(out + out_stride * 118 + i, v3932);
+ vst1q_s16(out + out_stride * 119 + i, v3936);
+ vst1q_s16(out + out_stride * 120 + i, v3940);
+ vst1q_s16(out + out_stride * 121 + i, v3944);
+ vst1q_s16(out + out_stride * 122 + i, v3948);
+ vst1q_s16(out + out_stride * 123 + i, v3952);
+ vst1q_s16(out + out_stride * 124 + i, v3956);
+ vst1q_s16(out + out_stride * 125 + i, v3960);
+ vst1q_s16(out + out_stride * 126 + i, v3964);
+ vst1q_s16(out + out_stride * 127 + i, v3968);
+ vst1q_s16(out + out_stride * 128 + i, v3969);
+ vst1q_s16(out + out_stride * 129 + i, v3970);
+ vst1q_s16(out + out_stride * 130 + i, v3971);
+ vst1q_s16(out + out_stride * 131 + i, v3972);
+ vst1q_s16(out + out_stride * 132 + i, v3973);
+ vst1q_s16(out + out_stride * 133 + i, v3974);
+ vst1q_s16(out + out_stride * 134 + i, v3975);
+ vst1q_s16(out + out_stride * 135 + i, v3976);
+ vst1q_s16(out + out_stride * 136 + i, v3977);
+ vst1q_s16(out + out_stride * 137 + i, v3978);
+ vst1q_s16(out + out_stride * 138 + i, v3979);
+ vst1q_s16(out + out_stride * 139 + i, v3980);
+ vst1q_s16(out + out_stride * 140 + i, v3981);
+ vst1q_s16(out + out_stride * 141 + i, v3982);
+ vst1q_s16(out + out_stride * 142 + i, v3983);
+ vst1q_s16(out + out_stride * 143 + i, v3984);
+ vst1q_s16(out + out_stride * 144 + i, v3985);
+ vst1q_s16(out + out_stride * 145 + i, v3986);
+ vst1q_s16(out + out_stride * 146 + i, v3987);
+ vst1q_s16(out + out_stride * 147 + i, v3988);
+ vst1q_s16(out + out_stride * 148 + i, v3989);
+ vst1q_s16(out + out_stride * 149 + i, v3990);
+ vst1q_s16(out + out_stride * 150 + i, v3991);
+ vst1q_s16(out + out_stride * 151 + i, v3992);
+ vst1q_s16(out + out_stride * 152 + i, v3993);
+ vst1q_s16(out + out_stride * 153 + i, v3994);
+ vst1q_s16(out + out_stride * 154 + i, v3995);
+ vst1q_s16(out + out_stride * 155 + i, v3996);
+ vst1q_s16(out + out_stride * 156 + i, v3997);
+ vst1q_s16(out + out_stride * 157 + i, v3998);
+ vst1q_s16(out + out_stride * 158 + i, v3999);
+ vst1q_s16(out + out_stride * 159 + i, v4000);
+ vst1q_s16(out + out_stride * 160 + i, v4001);
+ vst1q_s16(out + out_stride * 161 + i, v4002);
+ vst1q_s16(out + out_stride * 162 + i, v4003);
+ vst1q_s16(out + out_stride * 163 + i, v4004);
+ vst1q_s16(out + out_stride * 164 + i, v4005);
+ vst1q_s16(out + out_stride * 165 + i, v4006);
+ vst1q_s16(out + out_stride * 166 + i, v4007);
+ vst1q_s16(out + out_stride * 167 + i, v4008);
+ vst1q_s16(out + out_stride * 168 + i, v4009);
+ vst1q_s16(out + out_stride * 169 + i, v4010);
+ vst1q_s16(out + out_stride * 170 + i, v4011);
+ vst1q_s16(out + out_stride * 171 + i, v4012);
+ vst1q_s16(out + out_stride * 172 + i, v4013);
+ vst1q_s16(out + out_stride * 173 + i, v4014);
+ vst1q_s16(out + out_stride * 174 + i, v4015);
+ vst1q_s16(out + out_stride * 175 + i, v4016);
+ vst1q_s16(out + out_stride * 176 + i, v4017);
+ vst1q_s16(out + out_stride * 177 + i, v4018);
+ vst1q_s16(out + out_stride * 178 + i, v4019);
+ vst1q_s16(out + out_stride * 179 + i, v4020);
+ vst1q_s16(out + out_stride * 180 + i, v4021);
+ vst1q_s16(out + out_stride * 181 + i, v4022);
+ vst1q_s16(out + out_stride * 182 + i, v4023);
+ vst1q_s16(out + out_stride * 183 + i, v4024);
+ vst1q_s16(out + out_stride * 184 + i, v4025);
+ vst1q_s16(out + out_stride * 185 + i, v4026);
+ vst1q_s16(out + out_stride * 186 + i, v4027);
+ vst1q_s16(out + out_stride * 187 + i, v4028);
+ vst1q_s16(out + out_stride * 188 + i, v4029);
+ vst1q_s16(out + out_stride * 189 + i, v4030);
+ vst1q_s16(out + out_stride * 190 + i, v4031);
+ vst1q_s16(out + out_stride * 191 + i, v4032);
+ vst1q_s16(out + out_stride * 192 + i, v4033);
+ vst1q_s16(out + out_stride * 193 + i, v4034);
+ vst1q_s16(out + out_stride * 194 + i, v4035);
+ vst1q_s16(out + out_stride * 195 + i, v4036);
+ vst1q_s16(out + out_stride * 196 + i, v4037);
+ vst1q_s16(out + out_stride * 197 + i, v4038);
+ vst1q_s16(out + out_stride * 198 + i, v4039);
+ vst1q_s16(out + out_stride * 199 + i, v4040);
+ vst1q_s16(out + out_stride * 200 + i, v4041);
+ vst1q_s16(out + out_stride * 201 + i, v4042);
+ vst1q_s16(out + out_stride * 202 + i, v4043);
+ vst1q_s16(out + out_stride * 203 + i, v4044);
+ vst1q_s16(out + out_stride * 204 + i, v4045);
+ vst1q_s16(out + out_stride * 205 + i, v4046);
+ vst1q_s16(out + out_stride * 206 + i, v4047);
+ vst1q_s16(out + out_stride * 207 + i, v4048);
+ vst1q_s16(out + out_stride * 208 + i, v4049);
+ vst1q_s16(out + out_stride * 209 + i, v4050);
+ vst1q_s16(out + out_stride * 210 + i, v4051);
+ vst1q_s16(out + out_stride * 211 + i, v4052);
+ vst1q_s16(out + out_stride * 212 + i, v4053);
+ vst1q_s16(out + out_stride * 213 + i, v4054);
+ vst1q_s16(out + out_stride * 214 + i, v4055);
+ vst1q_s16(out + out_stride * 215 + i, v4056);
+ vst1q_s16(out + out_stride * 216 + i, v4057);
+ vst1q_s16(out + out_stride * 217 + i, v4058);
+ vst1q_s16(out + out_stride * 218 + i, v4059);
+ vst1q_s16(out + out_stride * 219 + i, v4060);
+ vst1q_s16(out + out_stride * 220 + i, v4061);
+ vst1q_s16(out + out_stride * 221 + i, v4062);
+ vst1q_s16(out + out_stride * 222 + i, v4063);
+ vst1q_s16(out + out_stride * 223 + i, v4064);
+ vst1q_s16(out + out_stride * 224 + i, v4065);
+ vst1q_s16(out + out_stride * 225 + i, v4066);
+ vst1q_s16(out + out_stride * 226 + i, v4067);
+ vst1q_s16(out + out_stride * 227 + i, v4068);
+ vst1q_s16(out + out_stride * 228 + i, v4069);
+ vst1q_s16(out + out_stride * 229 + i, v4070);
+ vst1q_s16(out + out_stride * 230 + i, v4071);
+ vst1q_s16(out + out_stride * 231 + i, v4072);
+ vst1q_s16(out + out_stride * 232 + i, v4073);
+ vst1q_s16(out + out_stride * 233 + i, v4074);
+ vst1q_s16(out + out_stride * 234 + i, v4075);
+ vst1q_s16(out + out_stride * 235 + i, v4076);
+ vst1q_s16(out + out_stride * 236 + i, v4077);
+ vst1q_s16(out + out_stride * 237 + i, v4078);
+ vst1q_s16(out + out_stride * 238 + i, v4079);
+ vst1q_s16(out + out_stride * 239 + i, v4080);
+ vst1q_s16(out + out_stride * 240 + i, v4081);
+ vst1q_s16(out + out_stride * 241 + i, v4082);
+ vst1q_s16(out + out_stride * 242 + i, v4083);
+ vst1q_s16(out + out_stride * 243 + i, v4084);
+ vst1q_s16(out + out_stride * 244 + i, v4085);
+ vst1q_s16(out + out_stride * 245 + i, v4086);
+ vst1q_s16(out + out_stride * 246 + i, v4087);
+ vst1q_s16(out + out_stride * 247 + i, v4088);
+ vst1q_s16(out + out_stride * 248 + i, v4089);
+ vst1q_s16(out + out_stride * 249 + i, v4090);
+ vst1q_s16(out + out_stride * 250 + i, v4091);
+ vst1q_s16(out + out_stride * 251 + i, v4092);
+ vst1q_s16(out + out_stride * 252 + i, v4093);
+ vst1q_s16(out + out_stride * 253 + i, v4094);
+ vst1q_s16(out + out_stride * 254 + i, v4095);
+ vst1q_s16(out + out_stride * 255 + i, v4096);
+ }
+}