Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-3x16c4-minmax-fp32-wasmsdot-u2-acc2.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_acc2(
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 vacc0x1x0123 = wasm_u32x4_const(0, 0, 0, 0);
70 v128_t vacc0x1x4567 = wasm_u32x4_const(0, 0, 0, 0);
71 v128_t vacc0x1x89AB = wasm_u32x4_const(0, 0, 0, 0);
72 v128_t vacc0x1xCDEF = wasm_u32x4_const(0, 0, 0, 0);
73 v128_t vacc1x0x0123= vacc0x0x0123;
74 v128_t vacc1x1x0123= vacc0x1x0123;
75 v128_t vacc1x0x4567= vacc0x0x4567;
76 v128_t vacc1x1x4567= vacc0x1x4567;
77 v128_t vacc1x0x89AB= vacc0x0x89AB;
78 v128_t vacc1x1x89AB= vacc0x1x89AB;
79 v128_t vacc1x0xCDEF= vacc0x0xCDEF;
80 v128_t vacc1x1xCDEF= vacc0x1xCDEF;
81 v128_t vacc2x0x0123= vacc0x0x0123;
82 v128_t vacc2x1x0123= vacc0x1x0123;
83 v128_t vacc2x0x4567= vacc0x0x4567;
84 v128_t vacc2x1x4567= vacc0x1x4567;
85 v128_t vacc2x0x89AB= vacc0x0x89AB;
86 v128_t vacc2x1x89AB= vacc0x1x89AB;
87 v128_t vacc2x0xCDEF= vacc0x0xCDEF;
88 v128_t vacc2x1xCDEF= vacc0x1xCDEF;
89 w = (const int32_t*) w + 16;
90
91 size_t p = ks;
92 do {
93 const int8_t* restrict a0 = a[0];
94 if XNN_UNPREDICTABLE(a0 != zero) {
95 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
96 }
97 const int8_t* restrict a1 = a[1];
98 if XNN_UNPREDICTABLE(a1 != zero) {
99 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
100 }
101 const int8_t* restrict a2 = a[2];
102 if XNN_UNPREDICTABLE(a2 != zero) {
103 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
104 }
105 a += 3;
106
107 size_t k = kc;
108 while (k >= 8 * sizeof(int8_t)) {
109 v128_t va0x0x0123 = wasm_v128_load32_splat(a0);
110 v128_t va0x1x0123 = wasm_v128_load32_splat(a0 + 4);
111 a0 += 8;
112 v128_t va1x0x0123 = wasm_v128_load32_splat(a1);
113 v128_t va1x1x0123 = wasm_v128_load32_splat(a1 + 4);
114 a1 += 8;
115 v128_t va2x0x0123 = wasm_v128_load32_splat(a2);
116 v128_t va2x1x0123 = wasm_v128_load32_splat(a2 + 4);
117 a2 += 8;
118
119
120 const v128_t vb0x0123 = wasm_v128_load((const int8_t*) w);
121 const v128_t vb0x4567 = wasm_v128_load((const int8_t*) w + 16);
122 const v128_t vb0x89AB = wasm_v128_load((const int8_t*) w + 32);
123 const v128_t vb0xCDEF = wasm_v128_load((const int8_t*) w + 48);
124 const v128_t vb1x0123 = wasm_v128_load((const int8_t*) w + 64);
125 const v128_t vb1x4567 = wasm_v128_load((const int8_t*) w + 80);
126 const v128_t vb1x89AB = wasm_v128_load((const int8_t*) w + 96);
127 const v128_t vb1xCDEF = wasm_v128_load((const int8_t*) w + 112);
128
129 vacc0x0x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0x0123, va0x0x0123, vacc0x0x0123);
130 vacc0x0x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0x4567, va0x0x0123, vacc0x0x4567);
131 vacc0x0x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0x89AB, va0x0x0123, vacc0x0x89AB);
132 vacc0x0xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0xCDEF, va0x0x0123, vacc0x0xCDEF);
133 vacc1x0x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0x0123, va1x0x0123, vacc1x0x0123);
134 vacc1x0x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0x4567, va1x0x0123, vacc1x0x4567);
135 vacc1x0x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0x89AB, va1x0x0123, vacc1x0x89AB);
136 vacc1x0xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0xCDEF, va1x0x0123, vacc1x0xCDEF);
137 vacc2x0x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0x0123, va2x0x0123, vacc2x0x0123);
138 vacc2x0x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0x4567, va2x0x0123, vacc2x0x4567);
139 vacc2x0x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0x89AB, va2x0x0123, vacc2x0x89AB);
140 vacc2x0xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0xCDEF, va2x0x0123, vacc2x0xCDEF);
141 vacc0x1x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1x0123, va0x1x0123, vacc0x1x0123);
142 vacc0x1x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1x4567, va0x1x0123, vacc0x1x4567);
143 vacc0x1x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1x89AB, va0x1x0123, vacc0x1x89AB);
144 vacc0x1xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1xCDEF, va0x1x0123, vacc0x1xCDEF);
145 vacc1x1x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1x0123, va1x1x0123, vacc1x1x0123);
146 vacc1x1x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1x4567, va1x1x0123, vacc1x1x4567);
147 vacc1x1x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1x89AB, va1x1x0123, vacc1x1x89AB);
148 vacc1x1xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1xCDEF, va1x1x0123, vacc1x1xCDEF);
149 vacc2x1x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1x0123, va2x1x0123, vacc2x1x0123);
150 vacc2x1x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1x4567, va2x1x0123, vacc2x1x4567);
151 vacc2x1x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1x89AB, va2x1x0123, vacc2x1x89AB);
152 vacc2x1xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1xCDEF, va2x1x0123, vacc2x1xCDEF);
153
154 w = (const int8_t*) w + 128;
155 k -= 8 * sizeof(int8_t);
156 }
157 vacc0x0x0123 = wasm_i32x4_add(vacc0x0x0123, vacc0x1x0123);
158 vacc0x0x4567 = wasm_i32x4_add(vacc0x0x4567, vacc0x1x4567);
159 vacc0x0x89AB = wasm_i32x4_add(vacc0x0x89AB, vacc0x1x89AB);
160 vacc0x0xCDEF = wasm_i32x4_add(vacc0x0xCDEF, vacc0x1xCDEF);
161 vacc1x0x0123 = wasm_i32x4_add(vacc1x0x0123, vacc1x1x0123);
162 vacc1x0x4567 = wasm_i32x4_add(vacc1x0x4567, vacc1x1x4567);
163 vacc1x0x89AB = wasm_i32x4_add(vacc1x0x89AB, vacc1x1x89AB);
164 vacc1x0xCDEF = wasm_i32x4_add(vacc1x0xCDEF, vacc1x1xCDEF);
165 vacc2x0x0123 = wasm_i32x4_add(vacc2x0x0123, vacc2x1x0123);
166 vacc2x0x4567 = wasm_i32x4_add(vacc2x0x4567, vacc2x1x4567);
167 vacc2x0x89AB = wasm_i32x4_add(vacc2x0x89AB, vacc2x1x89AB);
168 vacc2x0xCDEF = wasm_i32x4_add(vacc2x0xCDEF, vacc2x1xCDEF);
169
170 while (k != 0) {
171 v128_t va0x0123 = wasm_v128_load32_splat(a0);
172 a0 += 4;
173 v128_t va1x0123 = wasm_v128_load32_splat(a1);
174 a1 += 4;
175 v128_t va2x0123 = wasm_v128_load32_splat(a2);
176 a2 += 4;
177
178
179 const v128_t vb0123 = wasm_v128_load((const int8_t*) w);
180 const v128_t vb4567 = wasm_v128_load((const int8_t*) w + 16);
181 const v128_t vb89AB = wasm_v128_load((const int8_t*) w + 32);
182 const v128_t vbCDEF = wasm_v128_load((const int8_t*) w + 48);
183
184 vacc0x0x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0123, va0x0123, vacc0x0x0123);
185 vacc0x0x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb4567, va0x0123, vacc0x0x4567);
186 vacc0x0x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb89AB, va0x0123, vacc0x0x89AB);
187 vacc0x0xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vbCDEF, va0x0123, vacc0x0xCDEF);
188 vacc1x0x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0123, va1x0123, vacc1x0x0123);
189 vacc1x0x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb4567, va1x0123, vacc1x0x4567);
190 vacc1x0x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb89AB, va1x0123, vacc1x0x89AB);
191 vacc1x0xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vbCDEF, va1x0123, vacc1x0xCDEF);
192 vacc2x0x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0123, va2x0123, vacc2x0x0123);
193 vacc2x0x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb4567, va2x0123, vacc2x0x4567);
194 vacc2x0x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb89AB, va2x0123, vacc2x0x89AB);
195 vacc2x0xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vbCDEF, va2x0123, vacc2x0xCDEF);
196
197 w = (const int8_t*) w + 64;
198 k -= 4 * sizeof(int8_t);
199 }
200
201 p -= 3 * sizeof(void*);
202 } while (p != 0);
203
204 vacc0x0x0123 = wasm_f32x4_convert_i32x4(vacc0x0x0123);
205 vacc0x0x4567 = wasm_f32x4_convert_i32x4(vacc0x0x4567);
206 vacc0x0x89AB = wasm_f32x4_convert_i32x4(vacc0x0x89AB);
207 vacc0x0xCDEF = wasm_f32x4_convert_i32x4(vacc0x0xCDEF);
208 vacc1x0x0123 = wasm_f32x4_convert_i32x4(vacc1x0x0123);
209 vacc1x0x4567 = wasm_f32x4_convert_i32x4(vacc1x0x4567);
210 vacc1x0x89AB = wasm_f32x4_convert_i32x4(vacc1x0x89AB);
211 vacc1x0xCDEF = wasm_f32x4_convert_i32x4(vacc1x0xCDEF);
212 vacc2x0x0123 = wasm_f32x4_convert_i32x4(vacc2x0x0123);
213 vacc2x0x4567 = wasm_f32x4_convert_i32x4(vacc2x0x4567);
214 vacc2x0x89AB = wasm_f32x4_convert_i32x4(vacc2x0x89AB);
215 vacc2x0xCDEF = wasm_f32x4_convert_i32x4(vacc2x0xCDEF);
216
217 const v128_t vscale0123 = wasm_v128_load(w);
218 w = (const float*) w + 4;
219 const v128_t vscale4567 = wasm_v128_load(w);
220 w = (const float*) w + 4;
221 const v128_t vscale89AB = wasm_v128_load(w);
222 w = (const float*) w + 4;
223 const v128_t vscaleCDEF = wasm_v128_load(w);
224 w = (const float*) w + 4;
225 vacc0x0x0123 = wasm_f32x4_mul(vacc0x0x0123, vscale0123);
226 vacc0x0x4567 = wasm_f32x4_mul(vacc0x0x4567, vscale4567);
227 vacc0x0x89AB = wasm_f32x4_mul(vacc0x0x89AB, vscale89AB);
228 vacc0x0xCDEF = wasm_f32x4_mul(vacc0x0xCDEF, vscaleCDEF);
229 vacc1x0x0123 = wasm_f32x4_mul(vacc1x0x0123, vscale0123);
230 vacc1x0x4567 = wasm_f32x4_mul(vacc1x0x4567, vscale4567);
231 vacc1x0x89AB = wasm_f32x4_mul(vacc1x0x89AB, vscale89AB);
232 vacc1x0xCDEF = wasm_f32x4_mul(vacc1x0xCDEF, vscaleCDEF);
233 vacc2x0x0123 = wasm_f32x4_mul(vacc2x0x0123, vscale0123);
234 vacc2x0x4567 = wasm_f32x4_mul(vacc2x0x4567, vscale4567);
235 vacc2x0x89AB = wasm_f32x4_mul(vacc2x0x89AB, vscale89AB);
236 vacc2x0xCDEF = wasm_f32x4_mul(vacc2x0xCDEF, vscaleCDEF);
237
238 vacc0x0x0123 = wasm_f32x4_add(vacc0x0x0123, vmagic_bias);
239 vacc0x0x4567 = wasm_f32x4_add(vacc0x0x4567, vmagic_bias);
240 vacc0x0x89AB = wasm_f32x4_add(vacc0x0x89AB, vmagic_bias);
241 vacc0x0xCDEF = wasm_f32x4_add(vacc0x0xCDEF, vmagic_bias);
242 vacc1x0x0123 = wasm_f32x4_add(vacc1x0x0123, vmagic_bias);
243 vacc1x0x4567 = wasm_f32x4_add(vacc1x0x4567, vmagic_bias);
244 vacc1x0x89AB = wasm_f32x4_add(vacc1x0x89AB, vmagic_bias);
245 vacc1x0xCDEF = wasm_f32x4_add(vacc1x0xCDEF, vmagic_bias);
246 vacc2x0x0123 = wasm_f32x4_add(vacc2x0x0123, vmagic_bias);
247 vacc2x0x4567 = wasm_f32x4_add(vacc2x0x4567, vmagic_bias);
248 vacc2x0x89AB = wasm_f32x4_add(vacc2x0x89AB, vmagic_bias);
249 vacc2x0xCDEF = wasm_f32x4_add(vacc2x0xCDEF, vmagic_bias);
250
251 vacc0x0x0123 = wasm_i32x4_max(vacc0x0x0123, vmagic_min);
252 vacc0x0x4567 = wasm_i32x4_max(vacc0x0x4567, vmagic_min);
253 vacc0x0x89AB = wasm_i32x4_max(vacc0x0x89AB, vmagic_min);
254 vacc0x0xCDEF = wasm_i32x4_max(vacc0x0xCDEF, vmagic_min);
255 vacc1x0x0123 = wasm_i32x4_max(vacc1x0x0123, vmagic_min);
256 vacc1x0x4567 = wasm_i32x4_max(vacc1x0x4567, vmagic_min);
257 vacc1x0x89AB = wasm_i32x4_max(vacc1x0x89AB, vmagic_min);
258 vacc1x0xCDEF = wasm_i32x4_max(vacc1x0xCDEF, vmagic_min);
259 vacc2x0x0123 = wasm_i32x4_max(vacc2x0x0123, vmagic_min);
260 vacc2x0x4567 = wasm_i32x4_max(vacc2x0x4567, vmagic_min);
261 vacc2x0x89AB = wasm_i32x4_max(vacc2x0x89AB, vmagic_min);
262 vacc2x0xCDEF = wasm_i32x4_max(vacc2x0xCDEF, vmagic_min);
263
264 vacc0x0x0123 = wasm_i32x4_sub(vacc0x0x0123, vmagic_bias_less_output_zero_point);
265 vacc0x0x4567 = wasm_i32x4_sub(vacc0x0x4567, vmagic_bias_less_output_zero_point);
266 vacc0x0x89AB = wasm_i32x4_sub(vacc0x0x89AB, vmagic_bias_less_output_zero_point);
267 vacc0x0xCDEF = wasm_i32x4_sub(vacc0x0xCDEF, vmagic_bias_less_output_zero_point);
268 vacc1x0x0123 = wasm_i32x4_sub(vacc1x0x0123, vmagic_bias_less_output_zero_point);
269 vacc1x0x4567 = wasm_i32x4_sub(vacc1x0x4567, vmagic_bias_less_output_zero_point);
270 vacc1x0x89AB = wasm_i32x4_sub(vacc1x0x89AB, vmagic_bias_less_output_zero_point);
271 vacc1x0xCDEF = wasm_i32x4_sub(vacc1x0xCDEF, vmagic_bias_less_output_zero_point);
272 vacc2x0x0123 = wasm_i32x4_sub(vacc2x0x0123, vmagic_bias_less_output_zero_point);
273 vacc2x0x4567 = wasm_i32x4_sub(vacc2x0x4567, vmagic_bias_less_output_zero_point);
274 vacc2x0x89AB = wasm_i32x4_sub(vacc2x0x89AB, vmagic_bias_less_output_zero_point);
275 vacc2x0xCDEF = wasm_i32x4_sub(vacc2x0xCDEF, vmagic_bias_less_output_zero_point);
276
277 v128_t vacc0x01234567 = wasm_i16x8_narrow_i32x4(vacc0x0x0123, vacc0x0x4567);
278 v128_t vacc0x89ABCDEF = wasm_i16x8_narrow_i32x4(vacc0x0x89AB, vacc0x0xCDEF);
279 v128_t vacc1x01234567 = wasm_i16x8_narrow_i32x4(vacc1x0x0123, vacc1x0x4567);
280 v128_t vacc1x89ABCDEF = wasm_i16x8_narrow_i32x4(vacc1x0x89AB, vacc1x0xCDEF);
281 v128_t vacc2x01234567 = wasm_i16x8_narrow_i32x4(vacc2x0x0123, vacc2x0x4567);
282 v128_t vacc2x89ABCDEF = wasm_i16x8_narrow_i32x4(vacc2x0x89AB, vacc2x0xCDEF);
283
284 vacc0x01234567 = wasm_i8x16_narrow_i16x8(vacc0x01234567, vacc0x01234567);
285 vacc0x89ABCDEF = wasm_i8x16_narrow_i16x8(vacc0x89ABCDEF, vacc0x89ABCDEF);
286 vacc1x01234567 = wasm_i8x16_narrow_i16x8(vacc1x01234567, vacc1x01234567);
287 vacc1x89ABCDEF = wasm_i8x16_narrow_i16x8(vacc1x89ABCDEF, vacc1x89ABCDEF);
288 vacc2x01234567 = wasm_i8x16_narrow_i16x8(vacc2x01234567, vacc2x01234567);
289 vacc2x89ABCDEF = wasm_i8x16_narrow_i16x8(vacc2x89ABCDEF, vacc2x89ABCDEF);
290
291 vacc0x01234567 = wasm_i8x16_min(vacc0x01234567, voutput_max);
292 vacc0x89ABCDEF = wasm_i8x16_min(vacc0x89ABCDEF, voutput_max);
293 vacc1x01234567 = wasm_i8x16_min(vacc1x01234567, voutput_max);
294 vacc1x89ABCDEF = wasm_i8x16_min(vacc1x89ABCDEF, voutput_max);
295 vacc2x01234567 = wasm_i8x16_min(vacc2x01234567, voutput_max);
296 vacc2x89ABCDEF = wasm_i8x16_min(vacc2x89ABCDEF, voutput_max);
297
298 if (nc >= 16) {
299 wasm_v128_store64_lane(c2, vacc2x01234567, 0);
300 wasm_v128_store64_lane(c2 + 8, vacc2x89ABCDEF, 0);
301 wasm_v128_store64_lane(c1, vacc1x01234567, 0);
302 wasm_v128_store64_lane(c1 + 8, vacc1x89ABCDEF, 0);
303 wasm_v128_store64_lane(c0, vacc0x01234567, 0);
304 wasm_v128_store64_lane(c0 + 8, vacc0x89ABCDEF, 0);
305
306 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
307 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
308 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
309
310 a = (const int8_t**restrict) ((uintptr_t) a - ks);
311
312 nc -= 16;
313 } else {
314 if (nc & 8) {
315 wasm_v128_store64_lane(c2, vacc2x01234567, 0);
316 c2 += 8;
317 wasm_v128_store64_lane(c1, vacc1x01234567, 0);
318 c1 += 8;
319 wasm_v128_store64_lane(c0, vacc0x01234567, 0);
320 c0 += 8;
321
322 vacc0x01234567 = vacc0x89ABCDEF;
323 vacc1x01234567 = vacc1x89ABCDEF;
324 vacc2x01234567 = vacc2x89ABCDEF;
325 }
326 if (nc & 4) {
327 wasm_v128_store32_lane(c2, vacc2x01234567, 0);
328 c2 += 4;
329 wasm_v128_store32_lane(c1, vacc1x01234567, 0);
330 c1 += 4;
331 wasm_v128_store32_lane(c0, vacc0x01234567, 0);
332 c0 += 4;
333
334 vacc0x01234567 = wasm_u64x2_shr(vacc0x01234567, 32);
335 vacc1x01234567 = wasm_u64x2_shr(vacc1x01234567, 32);
336 vacc2x01234567 = wasm_u64x2_shr(vacc2x01234567, 32);
337 }
338 if (nc & 2) {
339 wasm_v128_store16_lane(c2, vacc2x01234567, 0);
340 c2 += 2;
341 wasm_v128_store16_lane(c1, vacc1x01234567, 0);
342 c1 += 2;
343 wasm_v128_store16_lane(c0, vacc0x01234567, 0);
344 c0 += 2;
345
346 vacc0x01234567 = wasm_u32x4_shr(vacc0x01234567, 16);
347 vacc1x01234567 = wasm_u32x4_shr(vacc1x01234567, 16);
348 vacc2x01234567 = wasm_u32x4_shr(vacc2x01234567, 16);
349 }
350 if (nc & 1) {
351 wasm_v128_store8_lane(c2, vacc2x01234567, 0);
352 wasm_v128_store8_lane(c1, vacc1x01234567, 0);
353 wasm_v128_store8_lane(c0, vacc0x01234567, 0);
354 }
355
356 nc = 0;
357 }
358 } while (nc != 0);
359 }