Debian Codegraph @

src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-4x8c8-minmax-wasmusdot-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__wasmusdot_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 const v128_t vsign_mask = wasm_u8x16_const_splat(UINT8_C(0x80));
68 do {
69 const v128_t vinput_zero_point0 = wasm_i32x4_splat((int32_t) quantization_params[0].zero_point + 128);
70 const v128_t vinput_zero_point1 = wasm_i32x4_splat((int32_t) quantization_params[1].zero_point + 128);
71 const v128_t vinput_zero_point2 = wasm_i32x4_splat((int32_t) quantization_params[2].zero_point + 128);
72 const v128_t vinput_zero_point3 = wasm_i32x4_splat((int32_t) quantization_params[3].zero_point + 128);
73 v128_t vksum0123 = wasm_v128_load(w); w = (const int32_t*) w + 4;
74 v128_t vsum0x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point0);
75 v128_t vacc0x01 = wasm_u64x2_extend_low_u32x4(vsum0x0123);
76 v128_t vacc0x23 = wasm_u64x2_extend_high_u32x4(vsum0x0123);
77 v128_t vsum1x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point1);
78 v128_t vacc1x01 = wasm_u64x2_extend_low_u32x4(vsum1x0123);
79 v128_t vacc1x23 = wasm_u64x2_extend_high_u32x4(vsum1x0123);
80 v128_t vsum2x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point2);
81 v128_t vacc2x01 = wasm_u64x2_extend_low_u32x4(vsum2x0123);
82 v128_t vacc2x23 = wasm_u64x2_extend_high_u32x4(vsum2x0123);
83 v128_t vsum3x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point3);
84 v128_t vacc3x01 = wasm_u64x2_extend_low_u32x4(vsum3x0123);
85 v128_t vacc3x23 = wasm_u64x2_extend_high_u32x4(vsum3x0123);
86 v128_t vksum4567 = wasm_v128_load(w); w = (const int32_t*) w + 4;
87 v128_t vsum0x4567 = wasm_i32x4_mul(vksum4567, vinput_zero_point0);
88 v128_t vacc0x45 = wasm_u64x2_extend_low_u32x4(vsum0x4567);
89 v128_t vacc0x67 = wasm_u64x2_extend_high_u32x4(vsum0x4567);
90 v128_t vsum1x4567 = wasm_i32x4_mul(vksum4567, vinput_zero_point1);
91 v128_t vacc1x45 = wasm_u64x2_extend_low_u32x4(vsum1x4567);
92 v128_t vacc1x67 = wasm_u64x2_extend_high_u32x4(vsum1x4567);
93 v128_t vsum2x4567 = wasm_i32x4_mul(vksum4567, vinput_zero_point2);
94 v128_t vacc2x45 = wasm_u64x2_extend_low_u32x4(vsum2x4567);
95 v128_t vacc2x67 = wasm_u64x2_extend_high_u32x4(vsum2x4567);
96 v128_t vsum3x4567 = wasm_i32x4_mul(vksum4567, vinput_zero_point3);
97 v128_t vacc3x45 = wasm_u64x2_extend_low_u32x4(vsum3x4567);
98 v128_t vacc3x67 = wasm_u64x2_extend_high_u32x4(vsum3x4567);
99
100 size_t k = kc;
101 while (k >= 16 * sizeof(int8_t)) {
102 const v128_t va0x01 = wasm_v128_xor(wasm_v128_load64_splat(a0), vsign_mask);
103 const v128_t va0x23 = wasm_v128_xor(wasm_v128_load64_splat((const int8_t*) a0 + 8), vsign_mask);
104 a0 += 16;
105 const v128_t va1x01 = wasm_v128_xor(wasm_v128_load64_splat(a1), vsign_mask);
106 const v128_t va1x23 = wasm_v128_xor(wasm_v128_load64_splat((const int8_t*) a1 + 8), vsign_mask);
107 a1 += 16;
108 const v128_t va2x01 = wasm_v128_xor(wasm_v128_load64_splat(a2), vsign_mask);
109 const v128_t va2x23 = wasm_v128_xor(wasm_v128_load64_splat((const int8_t*) a2 + 8), vsign_mask);
110 a2 += 16;
111 const v128_t va3x01 = wasm_v128_xor(wasm_v128_load64_splat(a3), vsign_mask);
112 const v128_t va3x23 = wasm_v128_xor(wasm_v128_load64_splat((const int8_t*) a3 + 8), vsign_mask);
113 a3 += 16;
114
115 const v128_t vb01x01 = wasm_v128_load(w); w = (const int8_t*) w + 16;
116 const v128_t vb23x01 = wasm_v128_load(w); w = (const int8_t*) w + 16;
117 const v128_t vb45x01 = wasm_v128_load(w); w = (const int8_t*) w + 16;
118 const v128_t vb67x01 = wasm_v128_load(w); w = (const int8_t*) w + 16;
119 const v128_t vb01x23 = wasm_v128_load(w); w = (const int8_t*) w + 16;
120 const v128_t vb23x23 = wasm_v128_load(w); w = (const int8_t*) w + 16;
121 const v128_t vb45x23 = wasm_v128_load(w); w = (const int8_t*) w + 16;
122 const v128_t vb67x23 = wasm_v128_load(w); w = (const int8_t*) w + 16;
123
124 vacc0x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01x01, va0x01, vacc0x01);
125 vacc0x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23x01, va0x01, vacc0x23);
126 vacc0x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45x01, va0x01, vacc0x45);
127 vacc0x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67x01, va0x01, vacc0x67);
128 vacc1x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01x01, va1x01, vacc1x01);
129 vacc1x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23x01, va1x01, vacc1x23);
130 vacc1x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45x01, va1x01, vacc1x45);
131 vacc1x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67x01, va1x01, vacc1x67);
132 vacc2x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01x01, va2x01, vacc2x01);
133 vacc2x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23x01, va2x01, vacc2x23);
134 vacc2x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45x01, va2x01, vacc2x45);
135 vacc2x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67x01, va2x01, vacc2x67);
136 vacc3x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01x01, va3x01, vacc3x01);
137 vacc3x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23x01, va3x01, vacc3x23);
138 vacc3x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45x01, va3x01, vacc3x45);
139 vacc3x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67x01, va3x01, vacc3x67);
140 vacc0x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01x23, va0x23, vacc0x01);
141 vacc0x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23x23, va0x23, vacc0x23);
142 vacc0x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45x23, va0x23, vacc0x45);
143 vacc0x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67x23, va0x23, vacc0x67);
144 vacc1x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01x23, va1x23, vacc1x01);
145 vacc1x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23x23, va1x23, vacc1x23);
146 vacc1x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45x23, va1x23, vacc1x45);
147 vacc1x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67x23, va1x23, vacc1x67);
148 vacc2x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01x23, va2x23, vacc2x01);
149 vacc2x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23x23, va2x23, vacc2x23);
150 vacc2x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45x23, va2x23, vacc2x45);
151 vacc2x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67x23, va2x23, vacc2x67);
152 vacc3x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01x23, va3x23, vacc3x01);
153 vacc3x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23x23, va3x23, vacc3x23);
154 vacc3x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45x23, va3x23, vacc3x45);
155 vacc3x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67x23, va3x23, vacc3x67);
156
157 k -= 16 * sizeof(int8_t);
158 }
159
160 if (k != 0) {
161 const v128_t va0x01 = wasm_v128_xor(wasm_v128_load64_splat(a0), vsign_mask);
162 a0 += 8;
163 const v128_t va1x01 = wasm_v128_xor(wasm_v128_load64_splat(a1), vsign_mask);
164 a1 += 8;
165 const v128_t va2x01 = wasm_v128_xor(wasm_v128_load64_splat(a2), vsign_mask);
166 a2 += 8;
167 const v128_t va3x01 = wasm_v128_xor(wasm_v128_load64_splat(a3), vsign_mask);
168 a3 += 8;
169
170 const v128_t vb01 = wasm_v128_load(w); w = (const int8_t*) w + 16;
171 const v128_t vb23 = wasm_v128_load(w); w = (const int8_t*) w + 16;
172 const v128_t vb45 = wasm_v128_load(w); w = (const int8_t*) w + 16;
173 const v128_t vb67 = wasm_v128_load(w); w = (const int8_t*) w + 16;
174
175 vacc0x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01, va0x01, vacc0x01);
176 vacc0x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23, va0x01, vacc0x23);
177 vacc0x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45, va0x01, vacc0x45);
178 vacc0x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67, va0x01, vacc0x67);
179 vacc1x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01, va1x01, vacc1x01);
180 vacc1x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23, va1x01, vacc1x23);
181 vacc1x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45, va1x01, vacc1x45);
182 vacc1x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67, va1x01, vacc1x67);
183 vacc2x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01, va2x01, vacc2x01);
184 vacc2x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23, va2x01, vacc2x23);
185 vacc2x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45, va2x01, vacc2x45);
186 vacc2x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67, va2x01, vacc2x67);
187 vacc3x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01, va3x01, vacc3x01);
188 vacc3x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23, va3x01, vacc3x23);
189 vacc3x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45, va3x01, vacc3x45);
190 vacc3x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67, va3x01, vacc3x67);
191
192 k -= 8 * sizeof(int8_t);
193 }
194
195 v128_t vacc0x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x01, vacc0x23, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc0x01, vacc0x23, 1, 3, 5, 7));
196 v128_t vacc0x4567 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x45, vacc0x67, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc0x45, vacc0x67, 1, 3, 5, 7));
197 v128_t vacc1x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x01, vacc1x23, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc1x01, vacc1x23, 1, 3, 5, 7));
198 v128_t vacc1x4567 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x45, vacc1x67, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc1x45, vacc1x67, 1, 3, 5, 7));
199 v128_t vacc2x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc2x01, vacc2x23, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc2x01, vacc2x23, 1, 3, 5, 7));
200 v128_t vacc2x4567 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc2x45, vacc2x67, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc2x45, vacc2x67, 1, 3, 5, 7));
201 v128_t vacc3x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc3x01, vacc3x23, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc3x01, vacc3x23, 1, 3, 5, 7));
202 v128_t vacc3x4567 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc3x45, vacc3x67, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc3x45, vacc3x67, 1, 3, 5, 7));
203
204 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
205 vacc0x4567 = wasm_f32x4_convert_i32x4(vacc0x4567);
206 vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
207 vacc1x4567 = wasm_f32x4_convert_i32x4(vacc1x4567);
208 vacc2x0123 = wasm_f32x4_convert_i32x4(vacc2x0123);
209 vacc2x4567 = wasm_f32x4_convert_i32x4(vacc2x4567);
210 vacc3x0123 = wasm_f32x4_convert_i32x4(vacc3x0123);
211 vacc3x4567 = wasm_f32x4_convert_i32x4(vacc3x4567);
212
213 const v128_t vinput_scale0 = wasm_v128_load32_splat(&quantization_params[0].inv_scale);
214 const v128_t vinput_scale1 = wasm_v128_load32_splat(&quantization_params[1].inv_scale);
215 const v128_t vinput_scale2 = wasm_v128_load32_splat(&quantization_params[2].inv_scale);
216 const v128_t vinput_scale3 = wasm_v128_load32_splat(&quantization_params[3].inv_scale);
217
218 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vinput_scale0);
219 vacc0x4567 = wasm_f32x4_mul(vacc0x4567, vinput_scale0);
220 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vinput_scale1);
221 vacc1x4567 = wasm_f32x4_mul(vacc1x4567, vinput_scale1);
222 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vinput_scale2);
223 vacc2x4567 = wasm_f32x4_mul(vacc2x4567, vinput_scale2);
224 vacc3x0123 = wasm_f32x4_mul(vacc3x0123, vinput_scale3);
225 vacc3x4567 = wasm_f32x4_mul(vacc3x4567, vinput_scale3);
226
227 const v128_t vfilter_output_scale0123 = wasm_v128_load(w);
228 w = (const float*) w + 4;
229 const v128_t vfilter_output_scale4567 = wasm_v128_load(w);
230 w = (const float*) w + 4;
231 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vfilter_output_scale0123);
232 vacc0x4567 = wasm_f32x4_mul(vacc0x4567, vfilter_output_scale4567);
233 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vfilter_output_scale0123);
234 vacc1x4567 = wasm_f32x4_mul(vacc1x4567, vfilter_output_scale4567);
235 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vfilter_output_scale0123);
236 vacc2x4567 = wasm_f32x4_mul(vacc2x4567, vfilter_output_scale4567);
237 vacc3x0123 = wasm_f32x4_mul(vacc3x0123, vfilter_output_scale0123);
238 vacc3x4567 = wasm_f32x4_mul(vacc3x4567, vfilter_output_scale4567);
239
240 const v128_t vbias0123 = wasm_v128_load(w);
241 w = (const float*) w + 4;
242 const v128_t vbias4567 = wasm_v128_load(w);
243 w = (const float*) w + 4;
244 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vbias0123);
245 vacc0x4567 = wasm_f32x4_add(vacc0x4567, vbias4567);
246 vacc1x0123 = wasm_f32x4_add(vacc1x0123, vbias0123);
247 vacc1x4567 = wasm_f32x4_add(vacc1x4567, vbias4567);
248 vacc2x0123 = wasm_f32x4_add(vacc2x0123, vbias0123);
249 vacc2x4567 = wasm_f32x4_add(vacc2x4567, vbias4567);
250 vacc3x0123 = wasm_f32x4_add(vacc3x0123, vbias0123);
251 vacc3x4567 = wasm_f32x4_add(vacc3x4567, vbias4567);
252
253 vacc0x0123 = wasm_f32x4_pmax(vacc0x0123, vmin);
254 vacc0x4567 = wasm_f32x4_pmax(vacc0x4567, vmin);
255 vacc1x0123 = wasm_f32x4_pmax(vacc1x0123, vmin);
256 vacc1x4567 = wasm_f32x4_pmax(vacc1x4567, vmin);
257 vacc2x0123 = wasm_f32x4_pmax(vacc2x0123, vmin);
258 vacc2x4567 = wasm_f32x4_pmax(vacc2x4567, vmin);
259 vacc3x0123 = wasm_f32x4_pmax(vacc3x0123, vmin);
260 vacc3x4567 = wasm_f32x4_pmax(vacc3x4567, vmin);
261
262 vacc0x0123 = wasm_f32x4_pmin(vacc0x0123, vmax);
263 vacc0x4567 = wasm_f32x4_pmin(vacc0x4567, vmax);
264 vacc1x0123 = wasm_f32x4_pmin(vacc1x0123, vmax);
265 vacc1x4567 = wasm_f32x4_pmin(vacc1x4567, vmax);
266 vacc2x0123 = wasm_f32x4_pmin(vacc2x0123, vmax);
267 vacc2x4567 = wasm_f32x4_pmin(vacc2x4567, vmax);
268 vacc3x0123 = wasm_f32x4_pmin(vacc3x0123, vmax);
269 vacc3x4567 = wasm_f32x4_pmin(vacc3x4567, vmax);
270
271 if XNN_LIKELY(nc >= 8) {
272 wasm_v128_store(c0, vacc0x0123);
273 wasm_v128_store(c0 + 4, vacc0x4567);
274 wasm_v128_store(c1, vacc1x0123);
275 wasm_v128_store(c1 + 4, vacc1x4567);
276 wasm_v128_store(c2, vacc2x0123);
277 wasm_v128_store(c2 + 4, vacc2x4567);
278 wasm_v128_store(c3, vacc3x0123);
279 wasm_v128_store(c3 + 4, vacc3x4567);
280
281 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
282 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
283 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
284 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
285
286 c0 = (float*) ((uintptr_t) c0 + cn_stride);
287 c1 = (float*) ((uintptr_t) c1 + cn_stride);
288 c2 = (float*) ((uintptr_t) c2 + cn_stride);
289 c3 = (float*) ((uintptr_t) c3 + cn_stride);
290
291 nc -= 8;
292 } else {
293 if (nc & 4) {
294 wasm_v128_store(c0, vacc0x0123);
295 vacc0x0123 = vacc0x4567;
296 c0 += 4;
297 wasm_v128_store(c1, vacc1x0123);
298 vacc1x0123 = vacc1x4567;
299 c1 += 4;
300 wasm_v128_store(c2, vacc2x0123);
301 vacc2x0123 = vacc2x4567;
302 c2 += 4;
303 wasm_v128_store(c3, vacc3x0123);
304 vacc3x0123 = vacc3x4567;
305 c3 += 4;
306 }
307 if (nc & 2) {
308 wasm_v128_store64_lane(c0, vacc0x0123, 0);
309 vacc0x0123 = wasm_v64x2_shuffle(vacc0x0123, vacc0x0123, 1, 1);
310 c0 += 2;
311 wasm_v128_store64_lane(c1, vacc1x0123, 0);
312 vacc1x0123 = wasm_v64x2_shuffle(vacc1x0123, vacc1x0123, 1, 1);
313 c1 += 2;
314 wasm_v128_store64_lane(c2, vacc2x0123, 0);
315 vacc2x0123 = wasm_v64x2_shuffle(vacc2x0123, vacc2x0123, 1, 1);
316 c2 += 2;
317 wasm_v128_store64_lane(c3, vacc3x0123, 0);
318 vacc3x0123 = wasm_v64x2_shuffle(vacc3x0123, vacc3x0123, 1, 1);
319 c3 += 2;
320 }
321 if (nc & 1) {
322 wasm_v128_store32_lane(c0, vacc0x0123, 0);
323 wasm_v128_store32_lane(c1, vacc1x0123, 0);
324 wasm_v128_store32_lane(c2, vacc2x0123, 0);
325 wasm_v128_store32_lane(c3, vacc3x0123, 0);
326 }
327 nc = 0;
328 }
329 } while (nc != 0);
330 }