Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-4x8c16-minmax-fp32-wasmusdot.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/MRx4c16-wasmdot.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2023 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_4x8c16__wasmusdot(
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 <= 4);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(ks != 0);
37 assert(ks % (4 * 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, 16 * 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 int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
54 if XNN_UNPREDICTABLE(mr != 4) {
55 c3 = c2;
56 }
57
58 const v128_t vmagic_bias = wasm_f32x4_const_splat(12582912.0f);
59 const int32_t output_min_less_zero_point = (int32_t) params->fp32_scalar.output_min - (int32_t) params->fp32_scalar.output_zero_point;
60 const v128_t vmagic_min = wasm_i32x4_splat((int32_t) float_as_uint32(12582912.0f + output_min_less_zero_point));
61 const v128_t vmagic_bias_less_output_zero_point = wasm_i32x4_splat(INT32_C(0x4B400000) - (int32_t) params->fp32_scalar.output_zero_point);
62 const v128_t voutput_max = wasm_i8x16_splat(params->fp32_scalar.output_max);
63 XNN_FORCE_REALIZATION(vmagic_bias);
64 XNN_FORCE_REALIZATION(vmagic_min);
65 XNN_FORCE_REALIZATION(vmagic_bias_less_output_zero_point);
66 XNN_FORCE_REALIZATION(voutput_max);
67
68 const v128_t vsign_mask = wasm_u8x16_const_splat(UINT8_C(0x80));
69 XNN_FORCE_REALIZATION(vsign_mask);
70 do {
71 v128_t vacc0x0 = wasm_v128_load32_zero(w);
72 v128_t vacc0x1 = wasm_v128_load32_zero((const int32_t*) w + 1);
73 v128_t vacc0x2 = wasm_v128_load32_zero((const int32_t*) w + 2);
74 v128_t vacc0x3 = wasm_v128_load32_zero((const int32_t*) w + 3);
75 v128_t vacc0x4 = wasm_v128_load32_zero((const int32_t*) w + 4);
76 v128_t vacc0x5 = wasm_v128_load32_zero((const int32_t*) w + 5);
77 v128_t vacc0x6 = wasm_v128_load32_zero((const int32_t*) w + 6);
78 v128_t vacc0x7 = wasm_v128_load32_zero((const int32_t*) w + 7);
79 v128_t vacc1x0 = vacc0x0;
80 v128_t vacc1x1 = vacc0x1;
81 v128_t vacc1x2 = vacc0x2;
82 v128_t vacc1x3 = vacc0x3;
83 v128_t vacc1x4 = vacc0x4;
84 v128_t vacc1x5 = vacc0x5;
85 v128_t vacc1x6 = vacc0x6;
86 v128_t vacc1x7 = vacc0x7;
87 v128_t vacc2x0 = vacc0x0;
88 v128_t vacc2x1 = vacc0x1;
89 v128_t vacc2x2 = vacc0x2;
90 v128_t vacc2x3 = vacc0x3;
91 v128_t vacc2x4 = vacc0x4;
92 v128_t vacc2x5 = vacc0x5;
93 v128_t vacc2x6 = vacc0x6;
94 v128_t vacc2x7 = vacc0x7;
95 v128_t vacc3x0 = vacc0x0;
96 v128_t vacc3x1 = vacc0x1;
97 v128_t vacc3x2 = vacc0x2;
98 v128_t vacc3x3 = vacc0x3;
99 v128_t vacc3x4 = vacc0x4;
100 v128_t vacc3x5 = vacc0x5;
101 v128_t vacc3x6 = vacc0x6;
102 v128_t vacc3x7 = vacc0x7;
103 w = (const int32_t*) w + 8;
104
105 size_t p = ks;
106 do {
107 const int8_t* restrict a0 = a[0];
108 if XNN_UNPREDICTABLE(a0 != zero) {
109 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
110 }
111 const int8_t* restrict a1 = a[1];
112 if XNN_UNPREDICTABLE(a1 != zero) {
113 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
114 }
115 const int8_t* restrict a2 = a[2];
116 if XNN_UNPREDICTABLE(a2 != zero) {
117 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
118 }
119 const int8_t* restrict a3 = a[3];
120 if XNN_UNPREDICTABLE(a3 != zero) {
121 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
122 }
123 a += 4;
124
125 size_t k = kc;
126 do {
127 const v128_t va0 = wasm_v128_xor(wasm_v128_load(a0), vsign_mask);
128 a0 += 16;
129 const v128_t va1 = wasm_v128_xor(wasm_v128_load(a1), vsign_mask);
130 a1 += 16;
131 const v128_t va2 = wasm_v128_xor(wasm_v128_load(a2), vsign_mask);
132 a2 += 16;
133 const v128_t va3 = wasm_v128_xor(wasm_v128_load(a3), vsign_mask);
134 a3 += 16;
135
136 const v128_t vb0 = wasm_v128_load(w);
137 vacc0x0 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0, va0, vacc0x0);
138 vacc1x0 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0, va1, vacc1x0);
139 vacc2x0 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0, va2, vacc2x0);
140 vacc3x0 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0, va3, vacc3x0);
141 const v128_t vb1 = wasm_v128_load((const int8_t*) w + 16);
142 vacc0x1 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1, va0, vacc0x1);
143 vacc1x1 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1, va1, vacc1x1);
144 vacc2x1 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1, va2, vacc2x1);
145 vacc3x1 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1, va3, vacc3x1);
146 const v128_t vb2 = wasm_v128_load((const int8_t*) w + 32);
147 vacc0x2 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb2, va0, vacc0x2);
148 vacc1x2 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb2, va1, vacc1x2);
149 vacc2x2 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb2, va2, vacc2x2);
150 vacc3x2 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb2, va3, vacc3x2);
151 const v128_t vb3 = wasm_v128_load((const int8_t*) w + 48);
152 vacc0x3 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb3, va0, vacc0x3);
153 vacc1x3 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb3, va1, vacc1x3);
154 vacc2x3 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb3, va2, vacc2x3);
155 vacc3x3 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb3, va3, vacc3x3);
156 const v128_t vb4 = wasm_v128_load((const int8_t*) w + 64);
157 vacc0x4 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb4, va0, vacc0x4);
158 vacc1x4 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb4, va1, vacc1x4);
159 vacc2x4 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb4, va2, vacc2x4);
160 vacc3x4 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb4, va3, vacc3x4);
161 const v128_t vb5 = wasm_v128_load((const int8_t*) w + 80);
162 vacc0x5 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb5, va0, vacc0x5);
163 vacc1x5 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb5, va1, vacc1x5);
164 vacc2x5 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb5, va2, vacc2x5);
165 vacc3x5 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb5, va3, vacc3x5);
166 const v128_t vb6 = wasm_v128_load((const int8_t*) w + 96);
167 vacc0x6 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb6, va0, vacc0x6);
168 vacc1x6 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb6, va1, vacc1x6);
169 vacc2x6 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb6, va2, vacc2x6);
170 vacc3x6 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb6, va3, vacc3x6);
171 const v128_t vb7 = wasm_v128_load((const int8_t*) w + 112);
172 vacc0x7 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb7, va0, vacc0x7);
173 vacc1x7 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb7, va1, vacc1x7);
174 vacc2x7 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb7, va2, vacc2x7);
175 vacc3x7 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb7, va3, vacc3x7);
176
177 w = (const int8_t*) w + 128;
178 k -= 16 * sizeof(int8_t);
179 } while (k != 0);
180 p -= 4 * sizeof(void*);
181 } while (p != 0);
182
183 const v128_t vacc0x02 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x0, vacc0x2, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x0, vacc0x2, 2, 6, 3, 7));
184 const v128_t vacc0x13 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x1, vacc0x3, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x1, vacc0x3, 2, 6, 3, 7));
185 const v128_t vacc0x46 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x4, vacc0x6, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x4, vacc0x6, 2, 6, 3, 7));
186 const v128_t vacc0x57 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x5, vacc0x7, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x5, vacc0x7, 2, 6, 3, 7));
187 const v128_t vacc1x02 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x0, vacc1x2, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x0, vacc1x2, 2, 6, 3, 7));
188 const v128_t vacc1x13 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x1, vacc1x3, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x1, vacc1x3, 2, 6, 3, 7));
189 const v128_t vacc1x46 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x4, vacc1x6, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x4, vacc1x6, 2, 6, 3, 7));
190 const v128_t vacc1x57 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x5, vacc1x7, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x5, vacc1x7, 2, 6, 3, 7));
191 const v128_t vacc2x02 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc2x0, vacc2x2, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc2x0, vacc2x2, 2, 6, 3, 7));
192 const v128_t vacc2x13 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc2x1, vacc2x3, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc2x1, vacc2x3, 2, 6, 3, 7));
193 const v128_t vacc2x46 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc2x4, vacc2x6, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc2x4, vacc2x6, 2, 6, 3, 7));
194 const v128_t vacc2x57 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc2x5, vacc2x7, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc2x5, vacc2x7, 2, 6, 3, 7));
195 const v128_t vacc3x02 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc3x0, vacc3x2, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc3x0, vacc3x2, 2, 6, 3, 7));
196 const v128_t vacc3x13 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc3x1, vacc3x3, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc3x1, vacc3x3, 2, 6, 3, 7));
197 const v128_t vacc3x46 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc3x4, vacc3x6, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc3x4, vacc3x6, 2, 6, 3, 7));
198 const v128_t vacc3x57 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc3x5, vacc3x7, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc3x5, vacc3x7, 2, 6, 3, 7));
199
200 v128_t vacc0x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x02, vacc0x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x02, vacc0x13, 2, 6, 3, 7));
201 v128_t vacc0x4567 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x46, vacc0x57, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x46, vacc0x57, 2, 6, 3, 7));
202 v128_t vacc1x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x02, vacc1x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x02, vacc1x13, 2, 6, 3, 7));
203 v128_t vacc1x4567 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x46, vacc1x57, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x46, vacc1x57, 2, 6, 3, 7));
204 v128_t vacc2x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc2x02, vacc2x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc2x02, vacc2x13, 2, 6, 3, 7));
205 v128_t vacc2x4567 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc2x46, vacc2x57, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc2x46, vacc2x57, 2, 6, 3, 7));
206 v128_t vacc3x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc3x02, vacc3x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc3x02, vacc3x13, 2, 6, 3, 7));
207 v128_t vacc3x4567 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc3x46, vacc3x57, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc3x46, vacc3x57, 2, 6, 3, 7));
208
209 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
210 vacc0x4567 = wasm_f32x4_convert_i32x4(vacc0x4567);
211 vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
212 vacc1x4567 = wasm_f32x4_convert_i32x4(vacc1x4567);
213 vacc2x0123 = wasm_f32x4_convert_i32x4(vacc2x0123);
214 vacc2x4567 = wasm_f32x4_convert_i32x4(vacc2x4567);
215 vacc3x0123 = wasm_f32x4_convert_i32x4(vacc3x0123);
216 vacc3x4567 = wasm_f32x4_convert_i32x4(vacc3x4567);
217
218 const v128_t vscale0123 = wasm_v128_load(w);
219 w = (const float*) w + 4;
220 const v128_t vscale4567 = wasm_v128_load(w);
221 w = (const float*) w + 4;
222 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vscale0123);
223 vacc0x4567 = wasm_f32x4_mul(vacc0x4567, vscale4567);
224 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vscale0123);
225 vacc1x4567 = wasm_f32x4_mul(vacc1x4567, vscale4567);
226 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vscale0123);
227 vacc2x4567 = wasm_f32x4_mul(vacc2x4567, vscale4567);
228 vacc3x0123 = wasm_f32x4_mul(vacc3x0123, vscale0123);
229 vacc3x4567 = wasm_f32x4_mul(vacc3x4567, vscale4567);
230
231 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vmagic_bias);
232 vacc0x4567 = wasm_f32x4_add(vacc0x4567, vmagic_bias);
233 vacc1x0123 = wasm_f32x4_add(vacc1x0123, vmagic_bias);
234 vacc1x4567 = wasm_f32x4_add(vacc1x4567, vmagic_bias);
235 vacc2x0123 = wasm_f32x4_add(vacc2x0123, vmagic_bias);
236 vacc2x4567 = wasm_f32x4_add(vacc2x4567, vmagic_bias);
237 vacc3x0123 = wasm_f32x4_add(vacc3x0123, vmagic_bias);
238 vacc3x4567 = wasm_f32x4_add(vacc3x4567, vmagic_bias);
239
240 vacc0x0123 = wasm_i32x4_max(vacc0x0123, vmagic_min);
241 vacc0x4567 = wasm_i32x4_max(vacc0x4567, vmagic_min);
242 vacc1x0123 = wasm_i32x4_max(vacc1x0123, vmagic_min);
243 vacc1x4567 = wasm_i32x4_max(vacc1x4567, vmagic_min);
244 vacc2x0123 = wasm_i32x4_max(vacc2x0123, vmagic_min);
245 vacc2x4567 = wasm_i32x4_max(vacc2x4567, vmagic_min);
246 vacc3x0123 = wasm_i32x4_max(vacc3x0123, vmagic_min);
247 vacc3x4567 = wasm_i32x4_max(vacc3x4567, vmagic_min);
248
249 vacc0x0123 = wasm_i32x4_sub(vacc0x0123, vmagic_bias_less_output_zero_point);
250 vacc0x4567 = wasm_i32x4_sub(vacc0x4567, vmagic_bias_less_output_zero_point);
251 vacc1x0123 = wasm_i32x4_sub(vacc1x0123, vmagic_bias_less_output_zero_point);
252 vacc1x4567 = wasm_i32x4_sub(vacc1x4567, vmagic_bias_less_output_zero_point);
253 vacc2x0123 = wasm_i32x4_sub(vacc2x0123, vmagic_bias_less_output_zero_point);
254 vacc2x4567 = wasm_i32x4_sub(vacc2x4567, vmagic_bias_less_output_zero_point);
255 vacc3x0123 = wasm_i32x4_sub(vacc3x0123, vmagic_bias_less_output_zero_point);
256 vacc3x4567 = wasm_i32x4_sub(vacc3x4567, vmagic_bias_less_output_zero_point);
257
258 v128_t vacc0x01234567 = wasm_i16x8_narrow_i32x4(vacc0x0123, vacc0x4567);
259 v128_t vacc1x01234567 = wasm_i16x8_narrow_i32x4(vacc1x0123, vacc1x4567);
260 v128_t vacc2x01234567 = wasm_i16x8_narrow_i32x4(vacc2x0123, vacc2x4567);
261 v128_t vacc3x01234567 = wasm_i16x8_narrow_i32x4(vacc3x0123, vacc3x4567);
262
263 v128_t vout0x01234567_1x01234567 = wasm_i8x16_narrow_i16x8(vacc0x01234567, vacc1x01234567);
264 v128_t vout2x01234567_3x01234567 = wasm_i8x16_narrow_i16x8(vacc2x01234567, vacc3x01234567);
265
266 vout0x01234567_1x01234567 = wasm_i8x16_min(vout0x01234567_1x01234567, voutput_max);
267 vout2x01234567_3x01234567 = wasm_i8x16_min(vout2x01234567_3x01234567, voutput_max);
268
269 if (nc >= 8) {
270 wasm_v128_store64_lane(c3, vout2x01234567_3x01234567, 1);
271 wasm_v128_store64_lane(c2, vout2x01234567_3x01234567, 0);
272 wasm_v128_store64_lane(c1, vout0x01234567_1x01234567, 1);
273 wasm_v128_store64_lane(c0, vout0x01234567_1x01234567, 0);
274
275 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
276 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
277 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
278 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
279
280 a = (const int8_t**restrict) ((uintptr_t) a - ks);
281
282 nc -= 8;
283 } else {
284 if (nc & 4) {
285 wasm_v128_store32_lane(c3, vout2x01234567_3x01234567, 2);
286 c3 += 4;
287 wasm_v128_store32_lane(c2, vout2x01234567_3x01234567, 0);
288 c2 += 4;
289 wasm_v128_store32_lane(c1, vout0x01234567_1x01234567, 2);
290 c1 += 4;
291 wasm_v128_store32_lane(c0, vout0x01234567_1x01234567, 0);
292 c0 += 4;
293
294 vout0x01234567_1x01234567 = wasm_u64x2_shr(vout0x01234567_1x01234567, 32);
295 vout2x01234567_3x01234567 = wasm_u64x2_shr(vout2x01234567_3x01234567, 32);
296 }
297 if (nc & 2) {
298 wasm_v128_store16_lane(c3, vout2x01234567_3x01234567, 4);
299 c3 += 2;
300 wasm_v128_store16_lane(c2, vout2x01234567_3x01234567, 0);
301 c2 += 2;
302 wasm_v128_store16_lane(c1, vout0x01234567_1x01234567, 4);
303 c1 += 2;
304 wasm_v128_store16_lane(c0, vout0x01234567_1x01234567, 0);
305 c0 += 2;
306
307 vout0x01234567_1x01234567 = wasm_u32x4_shr(vout0x01234567_1x01234567, 16);
308 vout2x01234567_3x01234567 = wasm_u32x4_shr(vout2x01234567_3x01234567, 16);
309 }
310 if (nc & 1) {
311 wasm_v128_store8_lane(c3, vout2x01234567_3x01234567, 8);
312 wasm_v128_store8_lane(c2, vout2x01234567_3x01234567, 0);
313 wasm_v128_store8_lane(c1, vout0x01234567_1x01234567, 8);
314 wasm_v128_store8_lane(c0, vout0x01234567_1x01234567, 0);
315 }
316
317 nc = 0;
318 }
319 } while (nc != 0);
320 }