Adding upstream version 1:115.7.0.upstream/1%115.7.0upstream

Signed-off-by: Daniel Baumann <daniel.baumann@progress-linux.org>

1 files changed, 419 insertions, 0 deletions

diff --git a/third_party/jpeg-xl/lib/jxl/fast_dct32-inl.h b/third_party/jpeg-xl/lib/jxl/fast_dct32-inl.h
new file mode 100644
index 0000000000..0f3b31cfea
--- /dev/null
+++ b/third_party/jpeg-xl/lib/jxl/fast_dct32-inl.h
@@ -0,0 +1,419 @@
+// 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<32>) { return 1; }
+
+void FastIDCT(FastDCTTag<32>, 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 * 16 + i);
+    int16x8_t v2 = vaddq_s16(v0, v1);
+    int16x8_t v3 = vld1q_s16(in + in_stride * 8 + 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 * 24 + 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 * 4 + 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 * 20 + i);
+    int16x8_t v13 = vld1q_s16(in + in_stride * 12 + i);
+    int16x8_t v14 = vaddq_s16(v12, v13);
+    int16x8_t v15 = vaddq_s16(v11, v14);
+    int16x8_t v16 = vld1q_s16(in + in_stride * 28 + i);
+    int16x8_t v17 = vaddq_s16(v16, v12);
+    int16x8_t v18 = vaddq_s16(v13, v10);
+    int16x8_t v19 = vaddq_s16(v17, v18);
+    int16x8_t v20 = vqrdmulhq_n_s16(v19, 17734);
+    int16x8_t v21 = vqrdmulhq_n_s16(v18, 25080);
+    int16x8_t v22 = vaddq_s16(v20, v21);
+    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 * 2 + 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 * 18 + i);
+    int16x8_t v29 = vld1q_s16(in + in_stride * 14 + i);
+    int16x8_t v30 = vaddq_s16(v28, v29);
+    int16x8_t v31 = vaddq_s16(v27, v30);
+    int16x8_t v32 = vld1q_s16(in + in_stride * 10 + i);
+    int16x8_t v33 = vld1q_s16(in + in_stride * 6 + i);
+    int16x8_t v34 = vaddq_s16(v32, v33);
+    int16x8_t v35 = vqrdmulhq_n_s16(v34, 25080);
+    int16x8_t v36 = vld1q_s16(in + in_stride * 26 + i);
+    int16x8_t v37 = vld1q_s16(in + in_stride * 22 + i);
+    int16x8_t v38 = vaddq_s16(v36, v37);
+    int16x8_t v39 = vaddq_s16(v38, v34);
+    int16x8_t v40 = vqrdmulhq_n_s16(v39, 17734);
+    int16x8_t v41 = vaddq_s16(v35, v40);
+    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(v29, v32);
+    int16x8_t v46 = vaddq_s16(v37, v28);
+    int16x8_t v47 = vaddq_s16(v45, v46);
+    int16x8_t v48 = vaddq_s16(v44, v47);
+    int16x8_t v49 = vqrdmulhq_n_s16(v48, 16705);
+    int16x8_t v50 = vld1q_s16(in + in_stride * 30 + i);
+    int16x8_t v51 = vaddq_s16(v50, v36);
+    int16x8_t v52 = vaddq_s16(v51, v46);
+    int16x8_t v53 = vqrdmulhq_n_s16(v52, 17734);
+    int16x8_t v54 = vaddq_s16(v45, v43);
+    int16x8_t v55_tmp = vqrdmulhq_n_s16(v54, 10045);
+    int16x8_t v55 = vaddq_s16(v55_tmp, v54);
+    int16x8_t v56 = vaddq_s16(v53, v55);
+    int16x8_t v57 = vqrdmulhq_n_s16(v56, 16705);
+    int16x8_t v58 = vaddq_s16(v49, v57);
+    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 * 13 + i);
+    int16x8_t v63 = vld1q_s16(in + in_stride * 11 + i);
+    int16x8_t v64 = vaddq_s16(v62, v63);
+    int16x8_t v65 = vld1q_s16(in + in_stride * 5 + i);
+    int16x8_t v66 = vld1q_s16(in + in_stride * 3 + i);
+    int16x8_t v67 = vaddq_s16(v65, v66);
+    int16x8_t v68 = vaddq_s16(v64, v67);
+    int16x8_t v69_tmp = vqrdmulhq_n_s16(v68, 10045);
+    int16x8_t v69 = vaddq_s16(v69_tmp, v68);
+    int16x8_t v70 = vld1q_s16(in + in_stride * 21 + i);
+    int16x8_t v71 = vld1q_s16(in + in_stride * 19 + i);
+    int16x8_t v72 = vaddq_s16(v70, v71);
+    int16x8_t v73 = vld1q_s16(in + in_stride * 29 + i);
+    int16x8_t v74 = vld1q_s16(in + in_stride * 27 + i);
+    int16x8_t v75 = vaddq_s16(v73, v74);
+    int16x8_t v76 = vaddq_s16(v72, v75);
+    int16x8_t v77 = vqrdmulhq_n_s16(v76, 17734);
+    int16x8_t v78 = vaddq_s16(v69, v77);
+    int16x8_t v79 = vqrdmulhq_n_s16(v78, 16705);
+    int16x8_t v80_tmp = vqrdmulhq_n_s16(v67, 13573);
+    int16x8_t v80 = vaddq_s16(v80_tmp, v67);
+    int16x8_t v81 = vaddq_s16(v64, v72);
+    int16x8_t v82 = vaddq_s16(v80, v81);
+    int16x8_t v83 = vqrdmulhq_n_s16(v82, 16705);
+    int16x8_t v84 = vaddq_s16(v79, v83);
+    int16x8_t v85 = vld1q_s16(in + in_stride * 1 + i);
+    int16x8_t v86_tmp = vqrdmulhq_n_s16(v85, 13573);
+    int16x8_t v86 = vaddq_s16(v86_tmp, v85);
+    int16x8_t v87 = vld1q_s16(in + in_stride * 17 + i);
+    int16x8_t v88 = vld1q_s16(in + in_stride * 15 + i);
+    int16x8_t v89 = vaddq_s16(v87, v88);
+    int16x8_t v90 = vaddq_s16(v86, v89);
+    int16x8_t v91 = vld1q_s16(in + in_stride * 9 + i);
+    int16x8_t v92 = vld1q_s16(in + in_stride * 7 + i);
+    int16x8_t v93 = vaddq_s16(v91, v92);
+    int16x8_t v94 = vqrdmulhq_n_s16(v93, 25080);
+    int16x8_t v95 = vld1q_s16(in + in_stride * 25 + i);
+    int16x8_t v96 = vld1q_s16(in + in_stride * 23 + i);
+    int16x8_t v97 = vaddq_s16(v95, v96);
+    int16x8_t v98 = vaddq_s16(v97, v93);
+    int16x8_t v99 = vqrdmulhq_n_s16(v98, 17734);
+    int16x8_t v100 = vaddq_s16(v94, v99);
+    int16x8_t v101 = vaddq_s16(v90, v100);
+    int16x8_t v102 = vaddq_s16(v84, v101);
+    int16x8_t v103 = vaddq_s16(v92, v65);
+    int16x8_t v104 = vaddq_s16(v66, v85);
+    int16x8_t v105 = vaddq_s16(v103, v104);
+    int16x8_t v106_tmp = vqrdmulhq_n_s16(v105, 13573);
+    int16x8_t v106 = vaddq_s16(v106_tmp, v105);
+    int16x8_t v107 = vaddq_s16(v96, v70);
+    int16x8_t v108 = vaddq_s16(v71, v87);
+    int16x8_t v109 = vaddq_s16(v107, v108);
+    int16x8_t v110 = vaddq_s16(v63, v91);
+    int16x8_t v111 = vaddq_s16(v88, v62);
+    int16x8_t v112 = vaddq_s16(v110, v111);
+    int16x8_t v113 = vaddq_s16(v109, v112);
+    int16x8_t v114 = vaddq_s16(v106, v113);
+    int16x8_t v115 = vqrdmulhq_n_s16(v114, 16705);
+    int16x8_t v116 = vaddq_s16(v112, v105);
+    int16x8_t v117 = vqrdmulhq_n_s16(v116, 25080);
+    int16x8_t v118 = vqrdmulhq_n_s16(v116, 17734);
+    int16x8_t v119 = vaddq_s16(v74, v95);
+    int16x8_t v120 = vld1q_s16(in + in_stride * 31 + i);
+    int16x8_t v121 = vaddq_s16(v120, v73);
+    int16x8_t v122 = vaddq_s16(v119, v121);
+    int16x8_t v123 = vaddq_s16(v122, v109);
+    int16x8_t v124 = vqrdmulhq_n_s16(v123, 17734);
+    int16x8_t v125 = vaddq_s16(v118, v124);
+    int16x8_t v126 = vaddq_s16(v117, v125);
+    int16x8_t v127 = vqrdmulhq_n_s16(v126, 16705);
+    int16x8_t v128 = vaddq_s16(v115, v127);
+    int16x8_t v129 = vqrdmulhq_n_s16(v128, 16463);
+    int16x8_t v130_tmp = vqrdmulhq_n_s16(v104, 13573);
+    int16x8_t v130 = vaddq_s16(v130_tmp, v104);
+    int16x8_t v131 = vaddq_s16(v108, v111);
+    int16x8_t v132 = vaddq_s16(v130, v131);
+    int16x8_t v133 = vaddq_s16(v119, v107);
+    int16x8_t v134 = vqrdmulhq_n_s16(v133, 17734);
+    int16x8_t v135 = vaddq_s16(v110, v103);
+    int16x8_t v136_tmp = vqrdmulhq_n_s16(v135, 10045);
+    int16x8_t v136 = vaddq_s16(v136_tmp, v135);
+    int16x8_t v137 = vaddq_s16(v134, v136);
+    int16x8_t v138 = vaddq_s16(v132, v137);
+    int16x8_t v139 = vqrdmulhq_n_s16(v138, 16463);
+    int16x8_t v140 = vaddq_s16(v129, v139);
+    int16x8_t v141 = vaddq_s16(v102, v140);
+    int16x8_t v142 = vqrdmulhq_n_s16(v141, 16404);
+    int16x8_t v143 = vaddq_s16(v61, v142);
+    int16x8_t v144 = vsubq_s16(v0, v1);
+    int16x8_t v145 = vsubq_s16(v4, v6);
+    int16x8_t v146_tmp = vqrdmulhq_n_s16(v145, 10045);
+    int16x8_t v146 = vaddq_s16(v146_tmp, v145);
+    int16x8_t v147 = vaddq_s16(v144, v146);
+    int16x8_t v148 = vsubq_s16(v11, v14);
+    int16x8_t v149 = vqrdmulhq_n_s16(v18, 17734);
+    int16x8_t v150_tmp = vqrdmulhq_n_s16(v17, 10045);
+    int16x8_t v150 = vaddq_s16(v150_tmp, v17);
+    int16x8_t v151 = vsubq_s16(v149, v150);
+    int16x8_t v152 = vaddq_s16(v148, v151);
+    int16x8_t v153 = vqrdmulhq_n_s16(v152, 19705);
+    int16x8_t v154 = vaddq_s16(v147, v153);
+    int16x8_t v155 = vsubq_s16(v27, v30);
+    int16x8_t v156 = vqrdmulhq_n_s16(v34, 17734);
+    int16x8_t v157_tmp = vqrdmulhq_n_s16(v38, 10045);
+    int16x8_t v157 = vaddq_s16(v157_tmp, v38);
+    int16x8_t v158 = vsubq_s16(v156, v157);
+    int16x8_t v159 = vaddq_s16(v155, v158);
+    int16x8_t v160 = vqrdmulhq_n_s16(v54, 13573);
+    int16x8_t v161 = vsubq_s16(v160, v52);
+    int16x8_t v162 = vqrdmulhq_n_s16(v161, 25746);
+    int16x8_t v163 = vsubq_s16(v44, v47);
+    int16x8_t v164 = vqrdmulhq_n_s16(v163, 19705);
+    int16x8_t v165 = vaddq_s16(v162, v164);
+    int16x8_t v166 = vaddq_s16(v159, v165);
+    int16x8_t v167 = vqrdmulhq_n_s16(v166, 17121);
+    int16x8_t v168 = vaddq_s16(v154, v167);
+    int16x8_t v169 = vsubq_s16(v86, v89);
+    int16x8_t v170 = vqrdmulhq_n_s16(v93, 17734);
+    int16x8_t v171_tmp = vqrdmulhq_n_s16(v97, 10045);
+    int16x8_t v171 = vaddq_s16(v171_tmp, v97);
+    int16x8_t v172 = vsubq_s16(v170, v171);
+    int16x8_t v173 = vaddq_s16(v169, v172);
+    int16x8_t v174 = vsubq_s16(v80, v81);
+    int16x8_t v175 = vqrdmulhq_n_s16(v174, 19705);
+    int16x8_t v176 = vqrdmulhq_n_s16(v68, 13573);
+    int16x8_t v177 = vsubq_s16(v176, v76);
+    int16x8_t v178 = vqrdmulhq_n_s16(v177, 25746);
+    int16x8_t v179 = vaddq_s16(v175, v178);
+    int16x8_t v180 = vaddq_s16(v173, v179);
+    int16x8_t v181 = vsubq_s16(v130, v131);
+    int16x8_t v182 = vqrdmulhq_n_s16(v135, 13573);
+    int16x8_t v183 = vsubq_s16(v182, v133);
+    int16x8_t v184_tmp = vqrdmulhq_n_s16(v183, 10045);
+    int16x8_t v184 = vaddq_s16(v184_tmp, v183);
+    int16x8_t v185 = vaddq_s16(v181, v184);
+    int16x8_t v186 = vqrdmulhq_n_s16(v185, 17121);
+    int16x8_t v187 = vqrdmulhq_n_s16(v105, 27867);
+    int16x8_t v188 = vqrdmulhq_n_s16(v113, 19705);
+    int16x8_t v189 = vsubq_s16(v187, v188);
+    int16x8_t v190 = vqrdmulhq_n_s16(v116, 13573);
+    int16x8_t v191 = vsubq_s16(v190, v123);
+    int16x8_t v192 = vqrdmulhq_n_s16(v191, 25746);
+    int16x8_t v193 = vaddq_s16(v189, v192);
+    int16x8_t v194 = vqrdmulhq_n_s16(v193, 17121);
+    int16x8_t v195 = vaddq_s16(v186, v194);
+    int16x8_t v196 = vaddq_s16(v180, v195);
+    int16x8_t v197 = vqrdmulhq_n_s16(v196, 16563);
+    int16x8_t v198 = vaddq_s16(v168, v197);
+    int16x8_t v199 = vsubq_s16(v144, v146);
+    int16x8_t v200 = vsubq_s16(v148, v151);
+    int16x8_t v201 = vqrdmulhq_n_s16(v200, 29490);
+    int16x8_t v202 = vaddq_s16(v199, v201);
+    int16x8_t v203 = vsubq_s16(v155, v158);
+    int16x8_t v204 = vqrdmulhq_n_s16(v163, 29490);
+    int16x8_t v205_tmp = vqrdmulhq_n_s16(v161, 5763);
+    int16x8_t v205 = vaddq_s16(v205_tmp, v161);
+    int16x8_t v206 = vsubq_s16(v204, v205);
+    int16x8_t v207 = vaddq_s16(v203, v206);
+    int16x8_t v208 = vqrdmulhq_n_s16(v207, 18578);
+    int16x8_t v209 = vaddq_s16(v202, v208);
+    int16x8_t v210 = vsubq_s16(v169, v172);
+    int16x8_t v211 = vqrdmulhq_n_s16(v174, 29490);
+    int16x8_t v212_tmp = vqrdmulhq_n_s16(v177, 5763);
+    int16x8_t v212 = vaddq_s16(v212_tmp, v177);
+    int16x8_t v213 = vsubq_s16(v211, v212);
+    int16x8_t v214 = vaddq_s16(v210, v213);
+    int16x8_t v215 = vsubq_s16(v181, v184);
+    int16x8_t v216 = vqrdmulhq_n_s16(v215, 18578);
+    int16x8_t v217 = vqrdmulhq_n_s16(v189, 27803);
+    int16x8_t v218 = vqrdmulhq_n_s16(v191, 21845);
+    int16x8_t v219 = vsubq_s16(v217, v218);
+    int16x8_t v220 = vaddq_s16(v216, v219);
+    int16x8_t v221 = vaddq_s16(v214, v220);
+    int16x8_t v222 = vqrdmulhq_n_s16(v221, 16890);
+    int16x8_t v223 = vaddq_s16(v209, v222);
+    int16x8_t v224 = vsubq_s16(v2, v8);
+    int16x8_t v225 = vsubq_s16(v15, v22);
+    int16x8_t v226_tmp = vqrdmulhq_n_s16(v225, 18446);
+    int16x8_t v226 = vmlaq_n_s16(v226_tmp, v225, 2);
+    int16x8_t v227 = vaddq_s16(v224, v226);
+    int16x8_t v228 = vsubq_s16(v31, v41);
+    int16x8_t v229 = vsubq_s16(v48, v56);
+    int16x8_t v230_tmp = vqrdmulhq_n_s16(v229, 18446);
+    int16x8_t v230 = vmlaq_n_s16(v230_tmp, v229, 2);
+    int16x8_t v231 = vaddq_s16(v228, v230);
+    int16x8_t v232 = vqrdmulhq_n_s16(v231, 21195);
+    int16x8_t v233 = vaddq_s16(v227, v232);
+    int16x8_t v234 = vsubq_s16(v82, v78);
+    int16x8_t v235_tmp = vqrdmulhq_n_s16(v234, 18446);
+    int16x8_t v235 = vmlaq_n_s16(v235_tmp, v234, 2);
+    int16x8_t v236 = vsubq_s16(v90, v100);
+    int16x8_t v237 = vaddq_s16(v235, v236);
+    int16x8_t v238 = vsubq_s16(v132, v137);
+    int16x8_t v239 = vsubq_s16(v114, v126);
+    int16x8_t v240_tmp = vqrdmulhq_n_s16(v239, 18446);
+    int16x8_t v240 = vmlaq_n_s16(v240_tmp, v239, 2);
+    int16x8_t v241 = vaddq_s16(v238, v240);
+    int16x8_t v242 = vqrdmulhq_n_s16(v241, 21195);
+    int16x8_t v243 = vaddq_s16(v237, v242);
+    int16x8_t v244 = vqrdmulhq_n_s16(v243, 17401);
+    int16x8_t v245 = vaddq_s16(v233, v244);
+    int16x8_t v246 = vsubq_s16(v228, v230);
+    int16x8_t v247 = vqrdmulhq_n_s16(v246, 25826);
+    int16x8_t v248 = vsubq_s16(v224, v226);
+    int16x8_t v249 = vaddq_s16(v247, v248);
+    int16x8_t v250 = vsubq_s16(v238, v240);
+    int16x8_t v251 = vqrdmulhq_n_s16(v250, 25826);
+    int16x8_t v252 = vsubq_s16(v236, v235);
+    int16x8_t v253 = vaddq_s16(v251, v252);
+    int16x8_t v254 = vqrdmulhq_n_s16(v253, 18124);
+    int16x8_t v255 = vaddq_s16(v249, v254);
+    int16x8_t v256 = vsubq_s16(v199, v201);
+    int16x8_t v257 = vsubq_s16(v203, v206);
+    int16x8_t v258_tmp = vqrdmulhq_n_s16(v257, 1988);
+    int16x8_t v258 = vaddq_s16(v258_tmp, v257);
+    int16x8_t v259 = vaddq_s16(v256, v258);
+    int16x8_t v260 = vsubq_s16(v210, v213);
+    int16x8_t v261_tmp = vqrdmulhq_n_s16(v219, 25030);
+    int16x8_t v261 = vaddq_s16(v261_tmp, v219);
+    int16x8_t v262 = vsubq_s16(v215, v261);
+    int16x8_t v263_tmp = vqrdmulhq_n_s16(v262, 1988);
+    int16x8_t v263 = vaddq_s16(v263_tmp, v262);
+    int16x8_t v264 = vaddq_s16(v260, v263);
+    int16x8_t v265 = vqrdmulhq_n_s16(v264, 19102);
+    int16x8_t v266 = vaddq_s16(v259, v265);
+    int16x8_t v267 = vsubq_s16(v147, v153);
+    int16x8_t v268 = vsubq_s16(v159, v165);
+    int16x8_t v269_tmp = vqrdmulhq_n_s16(v268, 23673);
+    int16x8_t v269 = vaddq_s16(v269_tmp, v268);
+    int16x8_t v270 = vaddq_s16(v267, v269);
+    int16x8_t v271 = vsubq_s16(v173, v179);
+    int16x8_t v272 = vsubq_s16(v185, v193);
+    int16x8_t v273_tmp = vqrdmulhq_n_s16(v272, 23673);
+    int16x8_t v273 = vaddq_s16(v273_tmp, v272);
+    int16x8_t v274 = vaddq_s16(v271, v273);
+    int16x8_t v275 = vqrdmulhq_n_s16(v274, 20398);
+    int16x8_t v276 = vaddq_s16(v270, v275);
+    int16x8_t v277 = vsubq_s16(v9, v24);
+    int16x8_t v278 = vsubq_s16(v42, v58);
+    int16x8_t v279_tmp = vqrdmulhq_n_s16(v278, 3314);
+    int16x8_t v279 = vmlaq_n_s16(v279_tmp, v278, 5);
+    int16x8_t v280 = vaddq_s16(v277, v279);
+    int16x8_t v281 = vsubq_s16(v138, v128);
+    int16x8_t v282_tmp = vqrdmulhq_n_s16(v281, 3314);
+    int16x8_t v282 = vmlaq_n_s16(v282_tmp, v281, 5);
+    int16x8_t v283 = vsubq_s16(v101, v84);
+    int16x8_t v284 = vaddq_s16(v282, v283);
+    int16x8_t v285 = vqrdmulhq_n_s16(v284, 22112);
+    int16x8_t v286 = vaddq_s16(v280, v285);
+    int16x8_t v287 = vsubq_s16(v277, v279);
+    int16x8_t v288 = vsubq_s16(v283, v282);
+    int16x8_t v289 = vqrdmulhq_n_s16(v288, 24397);
+    int16x8_t v290 = vaddq_s16(v287, v289);
+    int16x8_t v291 = vsubq_s16(v267, v269);
+    int16x8_t v292 = vsubq_s16(v271, v273);
+    int16x8_t v293 = vqrdmulhq_n_s16(v292, 27504);
+    int16x8_t v294 = vaddq_s16(v291, v293);
+    int16x8_t v295 = vsubq_s16(v260, v263);
+    int16x8_t v296 = vqrdmulhq_n_s16(v295, 31869);
+    int16x8_t v297 = vsubq_s16(v256, v258);
+    int16x8_t v298 = vaddq_s16(v296, v297);
+    int16x8_t v299 = vsubq_s16(v248, v247);
+    int16x8_t v300 = vsubq_s16(v252, v251);
+    int16x8_t v301_tmp = vqrdmulhq_n_s16(v300, 5552);
+    int16x8_t v301 = vaddq_s16(v301_tmp, v300);
+    int16x8_t v302 = vaddq_s16(v299, v301);
+    int16x8_t v303 = vsubq_s16(v227, v232);
+    int16x8_t v304 = vsubq_s16(v237, v242);
+    int16x8_t v305_tmp = vqrdmulhq_n_s16(v304, 15865);
+    int16x8_t v305 = vaddq_s16(v305_tmp, v304);
+    int16x8_t v306 = vaddq_s16(v303, v305);
+    int16x8_t v307 = vsubq_s16(v202, v208);
+    int16x8_t v308 = vsubq_s16(v214, v220);
+    int16x8_t v309_tmp = vqrdmulhq_n_s16(v308, 1893);
+    int16x8_t v309 = vmlaq_n_s16(v309_tmp, v308, 2);
+    int16x8_t v310 = vaddq_s16(v307, v309);
+    int16x8_t v311 = vsubq_s16(v154, v167);
+    int16x8_t v312 = vsubq_s16(v180, v195);
+    int16x8_t v313_tmp = vqrdmulhq_n_s16(v312, 13357);
+    int16x8_t v313 = vmlaq_n_s16(v313_tmp, v312, 3);
+    int16x8_t v314 = vaddq_s16(v311, v313);
+    int16x8_t v315 = vsubq_s16(v102, v140);
+    int16x8_t v316_tmp = vqrdmulhq_n_s16(v315, 6226);
+    int16x8_t v316 = vmlaq_n_s16(v316_tmp, v315, 10);
+    int16x8_t v317 = vsubq_s16(v25, v60);
+    int16x8_t v318 = vaddq_s16(v316, v317);
+    int16x8_t v319 = vsubq_s16(v317, v316);
+    int16x8_t v320 = vsubq_s16(v311, v313);
+    int16x8_t v321 = vsubq_s16(v307, v309);
+    int16x8_t v322 = vsubq_s16(v303, v305);
+    int16x8_t v323 = vsubq_s16(v299, v301);
+    int16x8_t v324 = vsubq_s16(v297, v296);
+    int16x8_t v325 = vsubq_s16(v291, v293);
+    int16x8_t v326 = vsubq_s16(v287, v289);
+    int16x8_t v327 = vsubq_s16(v280, v285);
+    int16x8_t v328 = vsubq_s16(v270, v275);
+    int16x8_t v329 = vsubq_s16(v259, v265);
+    int16x8_t v330 = vsubq_s16(v249, v254);
+    int16x8_t v331 = vsubq_s16(v233, v244);
+    int16x8_t v332 = vsubq_s16(v209, v222);
+    int16x8_t v333 = vsubq_s16(v168, v197);
+    int16x8_t v334 = vsubq_s16(v61, v142);
+    vst1q_s16(out + out_stride * 0 + i, v143);
+    vst1q_s16(out + out_stride * 1 + i, v198);
+    vst1q_s16(out + out_stride * 2 + i, v223);
+    vst1q_s16(out + out_stride * 3 + i, v245);
+    vst1q_s16(out + out_stride * 4 + i, v255);
+    vst1q_s16(out + out_stride * 5 + i, v266);
+    vst1q_s16(out + out_stride * 6 + i, v276);
+    vst1q_s16(out + out_stride * 7 + i, v286);
+    vst1q_s16(out + out_stride * 8 + i, v290);
+    vst1q_s16(out + out_stride * 9 + i, v294);
+    vst1q_s16(out + out_stride * 10 + i, v298);
+    vst1q_s16(out + out_stride * 11 + i, v302);
+    vst1q_s16(out + out_stride * 12 + i, v306);
+    vst1q_s16(out + out_stride * 13 + i, v310);
+    vst1q_s16(out + out_stride * 14 + i, v314);
+    vst1q_s16(out + out_stride * 15 + i, v318);
+    vst1q_s16(out + out_stride * 16 + i, v319);
+    vst1q_s16(out + out_stride * 17 + i, v320);
+    vst1q_s16(out + out_stride * 18 + i, v321);
+    vst1q_s16(out + out_stride * 19 + i, v322);
+    vst1q_s16(out + out_stride * 20 + i, v323);
+    vst1q_s16(out + out_stride * 21 + i, v324);
+    vst1q_s16(out + out_stride * 22 + i, v325);
+    vst1q_s16(out + out_stride * 23 + i, v326);
+    vst1q_s16(out + out_stride * 24 + i, v327);
+    vst1q_s16(out + out_stride * 25 + i, v328);
+    vst1q_s16(out + out_stride * 26 + i, v329);
+    vst1q_s16(out + out_stride * 27 + i, v330);
+    vst1q_s16(out + out_stride * 28 + i, v331);
+    vst1q_s16(out + out_stride * 29 + i, v332);
+    vst1q_s16(out + out_stride * 30 + i, v333);
+    vst1q_s16(out + out_stride * 31 + i, v334);
+  }
+}