Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-3x16c4-minmax-fp32-wasmsdot-u2.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/MRx16c4-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_qs8_qc8w_igemm_minmax_fp32_ukernel_3x16c4__wasmsdot_u2(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 size_t ks,
23 const int8_t** restrict a,
24 const void* restrict w,
25 int8_t* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 size_t a_offset,
29 const int8_t* zero,
30 const union xnn_qs8_qc8w_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
31 {
32 assert(mr != 0);
33 assert(mr <= 3);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(ks != 0);
37 assert(ks % (3 * sizeof(void*)) == 0);
38 assert(a_offset % sizeof(int8_t) == 0);
39 assert(a != NULL);
40 assert(w != NULL);
41 assert(c != NULL);
42
43 kc = round_up_po2(kc, 4 * sizeof(int8_t));
44 int8_t* c0 = c;
45 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
46 if XNN_UNPREDICTABLE(mr < 2) {
47 c1 = c0;
48 }
49 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
50 if XNN_UNPREDICTABLE(mr <= 2) {
51 c2 = c1;
52 }
53
54 const v128_t vmagic_bias = wasm_f32x4_const_splat(12582912.0f);
55 const int32_t output_min_less_zero_point = (int32_t) params->fp32_scalar.output_min - (int32_t) params->fp32_scalar.output_zero_point;
56 const v128_t vmagic_min = wasm_i32x4_splat((int32_t) float_as_uint32(12582912.0f + output_min_less_zero_point));
57 const v128_t vmagic_bias_less_output_zero_point = wasm_i32x4_splat(INT32_C(0x4B400000) - (int32_t) params->fp32_scalar.output_zero_point);
58 const v128_t voutput_max = wasm_i8x16_splat(params->fp32_scalar.output_max);
59 //XNN_FORCE_REALIZATION(vmagic_bias);
60 //XNN_FORCE_REALIZATION(vmagic_min);
61 //XNN_FORCE_REALIZATION(vmagic_bias_less_output_zero_point);
62 //XNN_FORCE_REALIZATION(voutput_max);
63
64 do {
65 v128_t vacc0x0x0123 = wasm_v128_load(w);
66 v128_t vacc0x0x4567 = wasm_v128_load((const int32_t*) w + 4);
67 v128_t vacc0x0x89AB = wasm_v128_load((const int32_t*) w + 8);
68 v128_t vacc0x0xCDEF = wasm_v128_load((const int32_t*) w + 12);
69 v128_t vacc1x0x0123= vacc0x0x0123;
70 v128_t vacc1x0x4567= vacc0x0x4567;
71 v128_t vacc1x0x89AB= vacc0x0x89AB;
72 v128_t vacc1x0xCDEF= vacc0x0xCDEF;
73 v128_t vacc2x0x0123= vacc0x0x0123;
74 v128_t vacc2x0x4567= vacc0x0x4567;
75 v128_t vacc2x0x89AB= vacc0x0x89AB;
76 v128_t vacc2x0xCDEF= vacc0x0xCDEF;
77 w = (const int32_t*) w + 16;
78
79 size_t p = ks;
80 do {
81 const int8_t* restrict a0 = a[0];
82 if XNN_UNPREDICTABLE(a0 != zero) {
83 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
84 }
85 const int8_t* restrict a1 = a[1];
86 if XNN_UNPREDICTABLE(a1 != zero) {
87 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
88 }
89 const int8_t* restrict a2 = a[2];
90 if XNN_UNPREDICTABLE(a2 != zero) {
91 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
92 }
93 a += 3;
94
95 size_t k = kc;
96 while (k >= 8 * sizeof(int8_t)) {
97 v128_t va0x0x0123 = wasm_v128_load32_splat(a0);
98 v128_t va0x1x0123 = wasm_v128_load32_splat(a0 + 4);
99 a0 += 8;
100 v128_t va1x0x0123 = wasm_v128_load32_splat(a1);
101 v128_t va1x1x0123 = wasm_v128_load32_splat(a1 + 4);
102 a1 += 8;
103 v128_t va2x0x0123 = wasm_v128_load32_splat(a2);
104 v128_t va2x1x0123 = wasm_v128_load32_splat(a2 + 4);
105 a2 += 8;
106
107
108 const v128_t vb0x0123 = wasm_v128_load((const int8_t*) w);
109 const v128_t vb0x4567 = wasm_v128_load((const int8_t*) w + 16);
110 const v128_t vb0x89AB = wasm_v128_load((const int8_t*) w + 32);
111 const v128_t vb0xCDEF = wasm_v128_load((const int8_t*) w + 48);
112 const v128_t vb1x0123 = wasm_v128_load((const int8_t*) w + 64);
113 const v128_t vb1x4567 = wasm_v128_load((const int8_t*) w + 80);
114 const v128_t vb1x89AB = wasm_v128_load((const int8_t*) w + 96);
115 const v128_t vb1xCDEF = wasm_v128_load((const int8_t*) w + 112);
116
117 vacc0x0x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0x0123, va0x0x0123, vacc0x0x0123);
118 vacc0x0x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0x4567, va0x0x0123, vacc0x0x4567);
119 vacc0x0x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0x89AB, va0x0x0123, vacc0x0x89AB);
120 vacc0x0xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0xCDEF, va0x0x0123, vacc0x0xCDEF);
121 vacc1x0x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0x0123, va1x0x0123, vacc1x0x0123);
122 vacc1x0x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0x4567, va1x0x0123, vacc1x0x4567);
123 vacc1x0x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0x89AB, va1x0x0123, vacc1x0x89AB);
124 vacc1x0xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0xCDEF, va1x0x0123, vacc1x0xCDEF);
125 vacc2x0x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0x0123, va2x0x0123, vacc2x0x0123);
126 vacc2x0x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0x4567, va2x0x0123, vacc2x0x4567);
127 vacc2x0x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0x89AB, va2x0x0123, vacc2x0x89AB);
128 vacc2x0xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0xCDEF, va2x0x0123, vacc2x0xCDEF);
129 vacc0x0x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1x0123, va0x1x0123, vacc0x0x0123);
130 vacc0x0x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1x4567, va0x1x0123, vacc0x0x4567);
131 vacc0x0x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1x89AB, va0x1x0123, vacc0x0x89AB);
132 vacc0x0xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1xCDEF, va0x1x0123, vacc0x0xCDEF);
133 vacc1x0x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1x0123, va1x1x0123, vacc1x0x0123);
134 vacc1x0x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1x4567, va1x1x0123, vacc1x0x4567);
135 vacc1x0x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1x89AB, va1x1x0123, vacc1x0x89AB);
136 vacc1x0xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1xCDEF, va1x1x0123, vacc1x0xCDEF);
137 vacc2x0x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1x0123, va2x1x0123, vacc2x0x0123);
138 vacc2x0x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1x4567, va2x1x0123, vacc2x0x4567);
139 vacc2x0x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1x89AB, va2x1x0123, vacc2x0x89AB);
140 vacc2x0xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1xCDEF, va2x1x0123, vacc2x0xCDEF);
141
142 w = (const int8_t*) w + 128;
143 k -= 8 * sizeof(int8_t);
144 }
145
146 while (k != 0) {
147 v128_t va0x0123 = wasm_v128_load32_splat(a0);
148 a0 += 4;
149 v128_t va1x0123 = wasm_v128_load32_splat(a1);
150 a1 += 4;
151 v128_t va2x0123 = wasm_v128_load32_splat(a2);
152 a2 += 4;
153
154
155 const v128_t vb0123 = wasm_v128_load((const int8_t*) w);
156 const v128_t vb4567 = wasm_v128_load((const int8_t*) w + 16);
157 const v128_t vb89AB = wasm_v128_load((const int8_t*) w + 32);
158 const v128_t vbCDEF = wasm_v128_load((const int8_t*) w + 48);
159
160 vacc0x0x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0123, va0x0123, vacc0x0x0123);
161 vacc0x0x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb4567, va0x0123, vacc0x0x4567);
162 vacc0x0x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb89AB, va0x0123, vacc0x0x89AB);
163 vacc0x0xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vbCDEF, va0x0123, vacc0x0xCDEF);
164 vacc1x0x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0123, va1x0123, vacc1x0x0123);
165 vacc1x0x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb4567, va1x0123, vacc1x0x4567);
166 vacc1x0x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb89AB, va1x0123, vacc1x0x89AB);
167 vacc1x0xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vbCDEF, va1x0123, vacc1x0xCDEF);
168 vacc2x0x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0123, va2x0123, vacc2x0x0123);
169 vacc2x0x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb4567, va2x0123, vacc2x0x4567);
170 vacc2x0x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb89AB, va2x0123, vacc2x0x89AB);
171 vacc2x0xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vbCDEF, va2x0123, vacc2x0xCDEF);
172
173 w = (const int8_t*) w + 64;
174 k -= 4 * sizeof(int8_t);
175 }
176
177 p -= 3 * sizeof(void*);
178 } while (p != 0);
179
180 vacc0x0x0123 = wasm_f32x4_convert_i32x4(vacc0x0x0123);
181 vacc0x0x4567 = wasm_f32x4_convert_i32x4(vacc0x0x4567);
182 vacc0x0x89AB = wasm_f32x4_convert_i32x4(vacc0x0x89AB);
183 vacc0x0xCDEF = wasm_f32x4_convert_i32x4(vacc0x0xCDEF);
184 vacc1x0x0123 = wasm_f32x4_convert_i32x4(vacc1x0x0123);
185 vacc1x0x4567 = wasm_f32x4_convert_i32x4(vacc1x0x4567);
186 vacc1x0x89AB = wasm_f32x4_convert_i32x4(vacc1x0x89AB);
187 vacc1x0xCDEF = wasm_f32x4_convert_i32x4(vacc1x0xCDEF);
188 vacc2x0x0123 = wasm_f32x4_convert_i32x4(vacc2x0x0123);
189 vacc2x0x4567 = wasm_f32x4_convert_i32x4(vacc2x0x4567);
190 vacc2x0x89AB = wasm_f32x4_convert_i32x4(vacc2x0x89AB);
191 vacc2x0xCDEF = wasm_f32x4_convert_i32x4(vacc2x0xCDEF);
192
193 const v128_t vscale0123 = wasm_v128_load(w);
194 w = (const float*) w + 4;
195 const v128_t vscale4567 = wasm_v128_load(w);
196 w = (const float*) w + 4;
197 const v128_t vscale89AB = wasm_v128_load(w);
198 w = (const float*) w + 4;
199 const v128_t vscaleCDEF = wasm_v128_load(w);
200 w = (const float*) w + 4;
201 vacc0x0x0123 = wasm_f32x4_mul(vacc0x0x0123, vscale0123);
202 vacc0x0x4567 = wasm_f32x4_mul(vacc0x0x4567, vscale4567);
203 vacc0x0x89AB = wasm_f32x4_mul(vacc0x0x89AB, vscale89AB);
204 vacc0x0xCDEF = wasm_f32x4_mul(vacc0x0xCDEF, vscaleCDEF);
205 vacc1x0x0123 = wasm_f32x4_mul(vacc1x0x0123, vscale0123);
206 vacc1x0x4567 = wasm_f32x4_mul(vacc1x0x4567, vscale4567);
207 vacc1x0x89AB = wasm_f32x4_mul(vacc1x0x89AB, vscale89AB);
208 vacc1x0xCDEF = wasm_f32x4_mul(vacc1x0xCDEF, vscaleCDEF);
209 vacc2x0x0123 = wasm_f32x4_mul(vacc2x0x0123, vscale0123);
210 vacc2x0x4567 = wasm_f32x4_mul(vacc2x0x4567, vscale4567);
211 vacc2x0x89AB = wasm_f32x4_mul(vacc2x0x89AB, vscale89AB);
212 vacc2x0xCDEF = wasm_f32x4_mul(vacc2x0xCDEF, vscaleCDEF);
213
214 vacc0x0x0123 = wasm_f32x4_add(vacc0x0x0123, vmagic_bias);
215 vacc0x0x4567 = wasm_f32x4_add(vacc0x0x4567, vmagic_bias);
216 vacc0x0x89AB = wasm_f32x4_add(vacc0x0x89AB, vmagic_bias);
217 vacc0x0xCDEF = wasm_f32x4_add(vacc0x0xCDEF, vmagic_bias);
218 vacc1x0x0123 = wasm_f32x4_add(vacc1x0x0123, vmagic_bias);
219 vacc1x0x4567 = wasm_f32x4_add(vacc1x0x4567, vmagic_bias);
220 vacc1x0x89AB = wasm_f32x4_add(vacc1x0x89AB, vmagic_bias);
221 vacc1x0xCDEF = wasm_f32x4_add(vacc1x0xCDEF, vmagic_bias);
222 vacc2x0x0123 = wasm_f32x4_add(vacc2x0x0123, vmagic_bias);
223 vacc2x0x4567 = wasm_f32x4_add(vacc2x0x4567, vmagic_bias);
224 vacc2x0x89AB = wasm_f32x4_add(vacc2x0x89AB, vmagic_bias);
225 vacc2x0xCDEF = wasm_f32x4_add(vacc2x0xCDEF, vmagic_bias);
226
227 vacc0x0x0123 = wasm_i32x4_max(vacc0x0x0123, vmagic_min);
228 vacc0x0x4567 = wasm_i32x4_max(vacc0x0x4567, vmagic_min);
229 vacc0x0x89AB = wasm_i32x4_max(vacc0x0x89AB, vmagic_min);
230 vacc0x0xCDEF = wasm_i32x4_max(vacc0x0xCDEF, vmagic_min);
231 vacc1x0x0123 = wasm_i32x4_max(vacc1x0x0123, vmagic_min);
232 vacc1x0x4567 = wasm_i32x4_max(vacc1x0x4567, vmagic_min);
233 vacc1x0x89AB = wasm_i32x4_max(vacc1x0x89AB, vmagic_min);
234 vacc1x0xCDEF = wasm_i32x4_max(vacc1x0xCDEF, vmagic_min);
235 vacc2x0x0123 = wasm_i32x4_max(vacc2x0x0123, vmagic_min);
236 vacc2x0x4567 = wasm_i32x4_max(vacc2x0x4567, vmagic_min);
237 vacc2x0x89AB = wasm_i32x4_max(vacc2x0x89AB, vmagic_min);
238 vacc2x0xCDEF = wasm_i32x4_max(vacc2x0xCDEF, vmagic_min);
239
240 vacc0x0x0123 = wasm_i32x4_sub(vacc0x0x0123, vmagic_bias_less_output_zero_point);
241 vacc0x0x4567 = wasm_i32x4_sub(vacc0x0x4567, vmagic_bias_less_output_zero_point);
242 vacc0x0x89AB = wasm_i32x4_sub(vacc0x0x89AB, vmagic_bias_less_output_zero_point);
243 vacc0x0xCDEF = wasm_i32x4_sub(vacc0x0xCDEF, vmagic_bias_less_output_zero_point);
244 vacc1x0x0123 = wasm_i32x4_sub(vacc1x0x0123, vmagic_bias_less_output_zero_point);
245 vacc1x0x4567 = wasm_i32x4_sub(vacc1x0x4567, vmagic_bias_less_output_zero_point);
246 vacc1x0x89AB = wasm_i32x4_sub(vacc1x0x89AB, vmagic_bias_less_output_zero_point);
247 vacc1x0xCDEF = wasm_i32x4_sub(vacc1x0xCDEF, vmagic_bias_less_output_zero_point);
248 vacc2x0x0123 = wasm_i32x4_sub(vacc2x0x0123, vmagic_bias_less_output_zero_point);
249 vacc2x0x4567 = wasm_i32x4_sub(vacc2x0x4567, vmagic_bias_less_output_zero_point);
250 vacc2x0x89AB = wasm_i32x4_sub(vacc2x0x89AB, vmagic_bias_less_output_zero_point);
251 vacc2x0xCDEF = wasm_i32x4_sub(vacc2x0xCDEF, vmagic_bias_less_output_zero_point);
252
253 v128_t vacc0x01234567 = wasm_i16x8_narrow_i32x4(vacc0x0x0123, vacc0x0x4567);
254 v128_t vacc0x89ABCDEF = wasm_i16x8_narrow_i32x4(vacc0x0x89AB, vacc0x0xCDEF);
255 v128_t vacc1x01234567 = wasm_i16x8_narrow_i32x4(vacc1x0x0123, vacc1x0x4567);
256 v128_t vacc1x89ABCDEF = wasm_i16x8_narrow_i32x4(vacc1x0x89AB, vacc1x0xCDEF);
257 v128_t vacc2x01234567 = wasm_i16x8_narrow_i32x4(vacc2x0x0123, vacc2x0x4567);
258 v128_t vacc2x89ABCDEF = wasm_i16x8_narrow_i32x4(vacc2x0x89AB, vacc2x0xCDEF);
259
260 vacc0x01234567 = wasm_i8x16_narrow_i16x8(vacc0x01234567, vacc0x01234567);
261 vacc0x89ABCDEF = wasm_i8x16_narrow_i16x8(vacc0x89ABCDEF, vacc0x89ABCDEF);
262 vacc1x01234567 = wasm_i8x16_narrow_i16x8(vacc1x01234567, vacc1x01234567);
263 vacc1x89ABCDEF = wasm_i8x16_narrow_i16x8(vacc1x89ABCDEF, vacc1x89ABCDEF);
264 vacc2x01234567 = wasm_i8x16_narrow_i16x8(vacc2x01234567, vacc2x01234567);
265 vacc2x89ABCDEF = wasm_i8x16_narrow_i16x8(vacc2x89ABCDEF, vacc2x89ABCDEF);
266
267 vacc0x01234567 = wasm_i8x16_min(vacc0x01234567, voutput_max);
268 vacc0x89ABCDEF = wasm_i8x16_min(vacc0x89ABCDEF, voutput_max);
269 vacc1x01234567 = wasm_i8x16_min(vacc1x01234567, voutput_max);
270 vacc1x89ABCDEF = wasm_i8x16_min(vacc1x89ABCDEF, voutput_max);
271 vacc2x01234567 = wasm_i8x16_min(vacc2x01234567, voutput_max);
272 vacc2x89ABCDEF = wasm_i8x16_min(vacc2x89ABCDEF, voutput_max);
273
274 if (nc >= 16) {
275 wasm_v128_store64_lane(c2, vacc2x01234567, 0);
276 wasm_v128_store64_lane(c2 + 8, vacc2x89ABCDEF, 0);
277 wasm_v128_store64_lane(c1, vacc1x01234567, 0);
278 wasm_v128_store64_lane(c1 + 8, vacc1x89ABCDEF, 0);
279 wasm_v128_store64_lane(c0, vacc0x01234567, 0);
280 wasm_v128_store64_lane(c0 + 8, vacc0x89ABCDEF, 0);
281
282 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
283 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
284 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
285
286 a = (const int8_t**restrict) ((uintptr_t) a - ks);
287
288 nc -= 16;
289 } else {
290 if (nc & 8) {
291 wasm_v128_store64_lane(c2, vacc2x01234567, 0);
292 c2 += 8;
293 wasm_v128_store64_lane(c1, vacc1x01234567, 0);
294 c1 += 8;
295 wasm_v128_store64_lane(c0, vacc0x01234567, 0);
296 c0 += 8;
297
298 vacc0x01234567 = vacc0x89ABCDEF;
299 vacc1x01234567 = vacc1x89ABCDEF;
300 vacc2x01234567 = vacc2x89ABCDEF;
301 }
302 if (nc & 4) {
303 wasm_v128_store32_lane(c2, vacc2x01234567, 0);
304 c2 += 4;
305 wasm_v128_store32_lane(c1, vacc1x01234567, 0);
306 c1 += 4;
307 wasm_v128_store32_lane(c0, vacc0x01234567, 0);
308 c0 += 4;
309
310 vacc0x01234567 = wasm_u64x2_shr(vacc0x01234567, 32);
311 vacc1x01234567 = wasm_u64x2_shr(vacc1x01234567, 32);
312 vacc2x01234567 = wasm_u64x2_shr(vacc2x01234567, 32);
313 }
314 if (nc & 2) {
315 wasm_v128_store16_lane(c2, vacc2x01234567, 0);
316 c2 += 2;
317 wasm_v128_store16_lane(c1, vacc1x01234567, 0);
318 c1 += 2;
319 wasm_v128_store16_lane(c0, vacc0x01234567, 0);
320 c0 += 2;
321
322 vacc0x01234567 = wasm_u32x4_shr(vacc0x01234567, 16);
323 vacc1x01234567 = wasm_u32x4_shr(vacc1x01234567, 16);
324 vacc2x01234567 = wasm_u32x4_shr(vacc2x01234567, 16);
325 }
326 if (nc & 1) {
327 wasm_v128_store8_lane(c2, vacc2x01234567, 0);
328 wasm_v128_store8_lane(c1, vacc1x01234567, 0);
329 wasm_v128_store8_lane(c0, vacc0x01234567, 0);
330 }
331
332 nc = 0;
333 }
334 } while (nc != 0);
335 }