Debian Codegraph @

src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-4x8c8-minmax-wasmsdot-u2.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/c8-wasmdot.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2024 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9
10 #include <assert.h>
11
12 #include <wasm_simd128.h>
13
14 #include "xnnpack/gemm.h"
15 #include "xnnpack/math.h"
16
17
18 void xnn_qd8_f32_qc8w_gemm_minmax_ukernel_4x8c8__wasmsdot_u2(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 const int8_t* restrict a,
23 size_t a_stride,
24 const void* restrict w,
25 float* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
29 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
30 {
31 assert(mr != 0);
32 assert(mr <= 4);
33 assert(nc != 0);
34 assert(kc != 0);
35 assert(kc % sizeof(int8_t) == 0);
36 assert(a != NULL);
37 assert(w != NULL);
38 assert(c != NULL);
39
40 kc = round_up_po2(kc, 8 * sizeof(int8_t));
41 const int8_t* a0 = a;
42 float* c0 = c;
43 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
44 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
45 if XNN_UNPREDICTABLE(mr < 2) {
46 a1 = a0;
47 c1 = c0;
48 }
49 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
50 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
51 if XNN_UNPREDICTABLE(mr <= 2) {
52 a2 = a1;
53 c2 = c1;
54 }
55 const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);
56 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
57 if XNN_UNPREDICTABLE(mr != 4) {
58 a3 = a2;
59 c3 = c2;
60 }
61
62 const v128_t vmin = wasm_v128_load32_splat(&params->scalar.min);
63 const v128_t vmax = wasm_v128_load32_splat(&params->scalar.max);
64 XNN_FORCE_REALIZATION(vmin);
65 XNN_FORCE_REALIZATION(vmax);
66
67 do {
68 const v128_t vinput_zero_point0 = wasm_v128_load32_splat(&quantization_params[0].zero_point);
69 const v128_t vinput_zero_point1 = wasm_v128_load32_splat(&quantization_params[1].zero_point);
70 const v128_t vinput_zero_point2 = wasm_v128_load32_splat(&quantization_params[2].zero_point);
71 const v128_t vinput_zero_point3 = wasm_v128_load32_splat(&quantization_params[3].zero_point);
72 v128_t vksum0123 = wasm_v128_load(w); w = (const int32_t*) w + 4;
73 v128_t vsum0x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point0);
74 v128_t vacc0x01 = wasm_u64x2_extend_low_u32x4(vsum0x0123);
75 v128_t vacc0x23 = wasm_u64x2_extend_high_u32x4(vsum0x0123);
76 v128_t vsum1x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point1);
77 v128_t vacc1x01 = wasm_u64x2_extend_low_u32x4(vsum1x0123);
78 v128_t vacc1x23 = wasm_u64x2_extend_high_u32x4(vsum1x0123);
79 v128_t vsum2x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point2);
80 v128_t vacc2x01 = wasm_u64x2_extend_low_u32x4(vsum2x0123);
81 v128_t vacc2x23 = wasm_u64x2_extend_high_u32x4(vsum2x0123);
82 v128_t vsum3x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point3);
83 v128_t vacc3x01 = wasm_u64x2_extend_low_u32x4(vsum3x0123);
84 v128_t vacc3x23 = wasm_u64x2_extend_high_u32x4(vsum3x0123);
85 v128_t vksum4567 = wasm_v128_load(w); w = (const int32_t*) w + 4;
86 v128_t vsum0x4567 = wasm_i32x4_mul(vksum4567, vinput_zero_point0);
87 v128_t vacc0x45 = wasm_u64x2_extend_low_u32x4(vsum0x4567);
88 v128_t vacc0x67 = wasm_u64x2_extend_high_u32x4(vsum0x4567);
89 v128_t vsum1x4567 = wasm_i32x4_mul(vksum4567, vinput_zero_point1);
90 v128_t vacc1x45 = wasm_u64x2_extend_low_u32x4(vsum1x4567);
91 v128_t vacc1x67 = wasm_u64x2_extend_high_u32x4(vsum1x4567);
92 v128_t vsum2x4567 = wasm_i32x4_mul(vksum4567, vinput_zero_point2);
93 v128_t vacc2x45 = wasm_u64x2_extend_low_u32x4(vsum2x4567);
94 v128_t vacc2x67 = wasm_u64x2_extend_high_u32x4(vsum2x4567);
95 v128_t vsum3x4567 = wasm_i32x4_mul(vksum4567, vinput_zero_point3);
96 v128_t vacc3x45 = wasm_u64x2_extend_low_u32x4(vsum3x4567);
97 v128_t vacc3x67 = wasm_u64x2_extend_high_u32x4(vsum3x4567);
98
99 size_t k = kc;
100 while (k >= 16 * sizeof(int8_t)) {
101 const v128_t va0x01 = wasm_v128_load64_splat(a0);
102 const v128_t va0x23 = wasm_v128_load64_splat((const int8_t*) a0 + 8);
103 a0 += 16;
104 const v128_t va1x01 = wasm_v128_load64_splat(a1);
105 const v128_t va1x23 = wasm_v128_load64_splat((const int8_t*) a1 + 8);
106 a1 += 16;
107 const v128_t va2x01 = wasm_v128_load64_splat(a2);
108 const v128_t va2x23 = wasm_v128_load64_splat((const int8_t*) a2 + 8);
109 a2 += 16;
110 const v128_t va3x01 = wasm_v128_load64_splat(a3);
111 const v128_t va3x23 = wasm_v128_load64_splat((const int8_t*) a3 + 8);
112 a3 += 16;
113
114 const v128_t vb01x01 = wasm_v128_load(w); w = (const int8_t*) w + 16;
115 const v128_t vb23x01 = wasm_v128_load(w); w = (const int8_t*) w + 16;
116 const v128_t vb45x01 = wasm_v128_load(w); w = (const int8_t*) w + 16;
117 const v128_t vb67x01 = wasm_v128_load(w); w = (const int8_t*) w + 16;
118 const v128_t vb01x23 = wasm_v128_load(w); w = (const int8_t*) w + 16;
119 const v128_t vb23x23 = wasm_v128_load(w); w = (const int8_t*) w + 16;
120 const v128_t vb45x23 = wasm_v128_load(w); w = (const int8_t*) w + 16;
121 const v128_t vb67x23 = wasm_v128_load(w); w = (const int8_t*) w + 16;
122
123 vacc0x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01x01, va0x01, vacc0x01);
124 vacc0x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23x01, va0x01, vacc0x23);
125 vacc0x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45x01, va0x01, vacc0x45);
126 vacc0x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67x01, va0x01, vacc0x67);
127 vacc1x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01x01, va1x01, vacc1x01);
128 vacc1x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23x01, va1x01, vacc1x23);
129 vacc1x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45x01, va1x01, vacc1x45);
130 vacc1x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67x01, va1x01, vacc1x67);
131 vacc2x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01x01, va2x01, vacc2x01);
132 vacc2x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23x01, va2x01, vacc2x23);
133 vacc2x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45x01, va2x01, vacc2x45);
134 vacc2x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67x01, va2x01, vacc2x67);
135 vacc3x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01x01, va3x01, vacc3x01);
136 vacc3x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23x01, va3x01, vacc3x23);
137 vacc3x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45x01, va3x01, vacc3x45);
138 vacc3x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67x01, va3x01, vacc3x67);
139 vacc0x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01x23, va0x23, vacc0x01);
140 vacc0x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23x23, va0x23, vacc0x23);
141 vacc0x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45x23, va0x23, vacc0x45);
142 vacc0x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67x23, va0x23, vacc0x67);
143 vacc1x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01x23, va1x23, vacc1x01);
144 vacc1x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23x23, va1x23, vacc1x23);
145 vacc1x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45x23, va1x23, vacc1x45);
146 vacc1x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67x23, va1x23, vacc1x67);
147 vacc2x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01x23, va2x23, vacc2x01);
148 vacc2x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23x23, va2x23, vacc2x23);
149 vacc2x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45x23, va2x23, vacc2x45);
150 vacc2x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67x23, va2x23, vacc2x67);
151 vacc3x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01x23, va3x23, vacc3x01);
152 vacc3x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23x23, va3x23, vacc3x23);
153 vacc3x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45x23, va3x23, vacc3x45);
154 vacc3x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67x23, va3x23, vacc3x67);
155
156 k -= 16 * sizeof(int8_t);
157 }
158
159 if (k != 0) {
160 const v128_t va0x01 = wasm_v128_load64_splat(a0);
161 a0 += 8;
162 const v128_t va1x01 = wasm_v128_load64_splat(a1);
163 a1 += 8;
164 const v128_t va2x01 = wasm_v128_load64_splat(a2);
165 a2 += 8;
166 const v128_t va3x01 = wasm_v128_load64_splat(a3);
167 a3 += 8;
168
169 const v128_t vb01 = wasm_v128_load(w); w = (const int8_t*) w + 16;
170 const v128_t vb23 = wasm_v128_load(w); w = (const int8_t*) w + 16;
171 const v128_t vb45 = wasm_v128_load(w); w = (const int8_t*) w + 16;
172 const v128_t vb67 = wasm_v128_load(w); w = (const int8_t*) w + 16;
173
174 vacc0x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01, va0x01, vacc0x01);
175 vacc0x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23, va0x01, vacc0x23);
176 vacc0x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45, va0x01, vacc0x45);
177 vacc0x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67, va0x01, vacc0x67);
178 vacc1x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01, va1x01, vacc1x01);
179 vacc1x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23, va1x01, vacc1x23);
180 vacc1x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45, va1x01, vacc1x45);
181 vacc1x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67, va1x01, vacc1x67);
182 vacc2x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01, va2x01, vacc2x01);
183 vacc2x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23, va2x01, vacc2x23);
184 vacc2x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45, va2x01, vacc2x45);
185 vacc2x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67, va2x01, vacc2x67);
186 vacc3x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01, va3x01, vacc3x01);
187 vacc3x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23, va3x01, vacc3x23);
188 vacc3x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45, va3x01, vacc3x45);
189 vacc3x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67, va3x01, vacc3x67);
190
191 k -= 8 * sizeof(int8_t);
192 }
193
194 v128_t vacc0x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x01, vacc0x23, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc0x01, vacc0x23, 1, 3, 5, 7));
195 v128_t vacc0x4567 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x45, vacc0x67, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc0x45, vacc0x67, 1, 3, 5, 7));
196 v128_t vacc1x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x01, vacc1x23, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc1x01, vacc1x23, 1, 3, 5, 7));
197 v128_t vacc1x4567 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x45, vacc1x67, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc1x45, vacc1x67, 1, 3, 5, 7));
198 v128_t vacc2x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc2x01, vacc2x23, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc2x01, vacc2x23, 1, 3, 5, 7));
199 v128_t vacc2x4567 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc2x45, vacc2x67, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc2x45, vacc2x67, 1, 3, 5, 7));
200 v128_t vacc3x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc3x01, vacc3x23, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc3x01, vacc3x23, 1, 3, 5, 7));
201 v128_t vacc3x4567 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc3x45, vacc3x67, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc3x45, vacc3x67, 1, 3, 5, 7));
202
203 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
204 vacc0x4567 = wasm_f32x4_convert_i32x4(vacc0x4567);
205 vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
206 vacc1x4567 = wasm_f32x4_convert_i32x4(vacc1x4567);
207 vacc2x0123 = wasm_f32x4_convert_i32x4(vacc2x0123);
208 vacc2x4567 = wasm_f32x4_convert_i32x4(vacc2x4567);
209 vacc3x0123 = wasm_f32x4_convert_i32x4(vacc3x0123);
210 vacc3x4567 = wasm_f32x4_convert_i32x4(vacc3x4567);
211
212 const v128_t vinput_scale0 = wasm_v128_load32_splat(&quantization_params[0].inv_scale);
213 const v128_t vinput_scale1 = wasm_v128_load32_splat(&quantization_params[1].inv_scale);
214 const v128_t vinput_scale2 = wasm_v128_load32_splat(&quantization_params[2].inv_scale);
215 const v128_t vinput_scale3 = wasm_v128_load32_splat(&quantization_params[3].inv_scale);
216
217 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vinput_scale0);
218 vacc0x4567 = wasm_f32x4_mul(vacc0x4567, vinput_scale0);
219 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vinput_scale1);
220 vacc1x4567 = wasm_f32x4_mul(vacc1x4567, vinput_scale1);
221 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vinput_scale2);
222 vacc2x4567 = wasm_f32x4_mul(vacc2x4567, vinput_scale2);
223 vacc3x0123 = wasm_f32x4_mul(vacc3x0123, vinput_scale3);
224 vacc3x4567 = wasm_f32x4_mul(vacc3x4567, vinput_scale3);
225
226 const v128_t vfilter_output_scale0123 = wasm_v128_load(w);
227 w = (const float*) w + 4;
228 const v128_t vfilter_output_scale4567 = wasm_v128_load(w);
229 w = (const float*) w + 4;
230 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vfilter_output_scale0123);
231 vacc0x4567 = wasm_f32x4_mul(vacc0x4567, vfilter_output_scale4567);
232 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vfilter_output_scale0123);
233 vacc1x4567 = wasm_f32x4_mul(vacc1x4567, vfilter_output_scale4567);
234 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vfilter_output_scale0123);
235 vacc2x4567 = wasm_f32x4_mul(vacc2x4567, vfilter_output_scale4567);
236 vacc3x0123 = wasm_f32x4_mul(vacc3x0123, vfilter_output_scale0123);
237 vacc3x4567 = wasm_f32x4_mul(vacc3x4567, vfilter_output_scale4567);
238
239 const v128_t vbias0123 = wasm_v128_load(w);
240 w = (const float*) w + 4;
241 const v128_t vbias4567 = wasm_v128_load(w);
242 w = (const float*) w + 4;
243 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vbias0123);
244 vacc0x4567 = wasm_f32x4_add(vacc0x4567, vbias4567);
245 vacc1x0123 = wasm_f32x4_add(vacc1x0123, vbias0123);
246 vacc1x4567 = wasm_f32x4_add(vacc1x4567, vbias4567);
247 vacc2x0123 = wasm_f32x4_add(vacc2x0123, vbias0123);
248 vacc2x4567 = wasm_f32x4_add(vacc2x4567, vbias4567);
249 vacc3x0123 = wasm_f32x4_add(vacc3x0123, vbias0123);
250 vacc3x4567 = wasm_f32x4_add(vacc3x4567, vbias4567);
251
252 vacc0x0123 = wasm_f32x4_pmax(vacc0x0123, vmin);
253 vacc0x4567 = wasm_f32x4_pmax(vacc0x4567, vmin);
254 vacc1x0123 = wasm_f32x4_pmax(vacc1x0123, vmin);
255 vacc1x4567 = wasm_f32x4_pmax(vacc1x4567, vmin);
256 vacc2x0123 = wasm_f32x4_pmax(vacc2x0123, vmin);
257 vacc2x4567 = wasm_f32x4_pmax(vacc2x4567, vmin);
258 vacc3x0123 = wasm_f32x4_pmax(vacc3x0123, vmin);
259 vacc3x4567 = wasm_f32x4_pmax(vacc3x4567, vmin);
260
261 vacc0x0123 = wasm_f32x4_pmin(vacc0x0123, vmax);
262 vacc0x4567 = wasm_f32x4_pmin(vacc0x4567, vmax);
263 vacc1x0123 = wasm_f32x4_pmin(vacc1x0123, vmax);
264 vacc1x4567 = wasm_f32x4_pmin(vacc1x4567, vmax);
265 vacc2x0123 = wasm_f32x4_pmin(vacc2x0123, vmax);
266 vacc2x4567 = wasm_f32x4_pmin(vacc2x4567, vmax);
267 vacc3x0123 = wasm_f32x4_pmin(vacc3x0123, vmax);
268 vacc3x4567 = wasm_f32x4_pmin(vacc3x4567, vmax);
269
270 if XNN_LIKELY(nc >= 8) {
271 wasm_v128_store(c0, vacc0x0123);
272 wasm_v128_store(c0 + 4, vacc0x4567);
273 wasm_v128_store(c1, vacc1x0123);
274 wasm_v128_store(c1 + 4, vacc1x4567);
275 wasm_v128_store(c2, vacc2x0123);
276 wasm_v128_store(c2 + 4, vacc2x4567);
277 wasm_v128_store(c3, vacc3x0123);
278 wasm_v128_store(c3 + 4, vacc3x4567);
279
280 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
281 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
282 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
283 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
284
285 c0 = (float*) ((uintptr_t) c0 + cn_stride);
286 c1 = (float*) ((uintptr_t) c1 + cn_stride);
287 c2 = (float*) ((uintptr_t) c2 + cn_stride);
288 c3 = (float*) ((uintptr_t) c3 + cn_stride);
289
290 nc -= 8;
291 } else {
292 if (nc & 4) {
293 wasm_v128_store(c0, vacc0x0123);
294 vacc0x0123 = vacc0x4567;
295 c0 += 4;
296 wasm_v128_store(c1, vacc1x0123);
297 vacc1x0123 = vacc1x4567;
298 c1 += 4;
299 wasm_v128_store(c2, vacc2x0123);
300 vacc2x0123 = vacc2x4567;
301 c2 += 4;
302 wasm_v128_store(c3, vacc3x0123);
303 vacc3x0123 = vacc3x4567;
304 c3 += 4;
305 }
306 if (nc & 2) {
307 wasm_v128_store64_lane(c0, vacc0x0123, 0);
308 vacc0x0123 = wasm_v64x2_shuffle(vacc0x0123, vacc0x0123, 1, 1);
309 c0 += 2;
310 wasm_v128_store64_lane(c1, vacc1x0123, 0);
311 vacc1x0123 = wasm_v64x2_shuffle(vacc1x0123, vacc1x0123, 1, 1);
312 c1 += 2;
313 wasm_v128_store64_lane(c2, vacc2x0123, 0);
314 vacc2x0123 = wasm_v64x2_shuffle(vacc2x0123, vacc2x0123, 1, 1);
315 c2 += 2;
316 wasm_v128_store64_lane(c3, vacc3x0123, 0);
317 vacc3x0123 = wasm_v64x2_shuffle(vacc3x0123, vacc3x0123, 1, 1);
318 c3 += 2;
319 }
320 if (nc & 1) {
321 wasm_v128_store32_lane(c0, vacc0x0123, 0);
322 wasm_v128_store32_lane(c1, vacc1x0123, 0);
323 wasm_v128_store32_lane(c2, vacc2x0123, 0);
324 wasm_v128_store32_lane(c3, vacc3x0123, 0);
325 }
326 nc = 0;
327 }
328 } while (nc != 0);
329 }