Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-3x16c4-minmax-fp32-wasmusdot-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__wasmusdot_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 const v128_t vsign_mask = wasm_u8x16_const_splat(UINT8_C(0x80));
65 XNN_FORCE_REALIZATION(vsign_mask);
66 do {
67 v128_t vacc0x0x0123 = wasm_v128_load(w);
68 v128_t vacc0x0x4567 = wasm_v128_load((const int32_t*) w + 4);
69 v128_t vacc0x0x89AB = wasm_v128_load((const int32_t*) w + 8);
70 v128_t vacc0x0xCDEF = wasm_v128_load((const int32_t*) w + 12);
71 v128_t vacc0x1x0123 = wasm_u32x4_const(0, 0, 0, 0);
72 v128_t vacc0x1x4567 = wasm_u32x4_const(0, 0, 0, 0);
73 v128_t vacc0x1x89AB = wasm_u32x4_const(0, 0, 0, 0);
74 v128_t vacc0x1xCDEF = wasm_u32x4_const(0, 0, 0, 0);
75 v128_t vacc1x0x0123= vacc0x0x0123;
76 v128_t vacc1x1x0123= vacc0x1x0123;
77 v128_t vacc1x0x4567= vacc0x0x4567;
78 v128_t vacc1x1x4567= vacc0x1x4567;
79 v128_t vacc1x0x89AB= vacc0x0x89AB;
80 v128_t vacc1x1x89AB= vacc0x1x89AB;
81 v128_t vacc1x0xCDEF= vacc0x0xCDEF;
82 v128_t vacc1x1xCDEF= vacc0x1xCDEF;
83 v128_t vacc2x0x0123= vacc0x0x0123;
84 v128_t vacc2x1x0123= vacc0x1x0123;
85 v128_t vacc2x0x4567= vacc0x0x4567;
86 v128_t vacc2x1x4567= vacc0x1x4567;
87 v128_t vacc2x0x89AB= vacc0x0x89AB;
88 v128_t vacc2x1x89AB= vacc0x1x89AB;
89 v128_t vacc2x0xCDEF= vacc0x0xCDEF;
90 v128_t vacc2x1xCDEF= vacc0x1xCDEF;
91 w = (const int32_t*) w + 16;
92
93 size_t p = ks;
94 do {
95 const int8_t* restrict a0 = a[0];
96 if XNN_UNPREDICTABLE(a0 != zero) {
97 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
98 }
99 const int8_t* restrict a1 = a[1];
100 if XNN_UNPREDICTABLE(a1 != zero) {
101 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
102 }
103 const int8_t* restrict a2 = a[2];
104 if XNN_UNPREDICTABLE(a2 != zero) {
105 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
106 }
107 a += 3;
108
109 size_t k = kc;
110 while (k >= 8 * sizeof(int8_t)) {
111 v128_t va0x0x0123 = wasm_v128_load32_splat(a0);
112 v128_t va0x1x0123 = wasm_v128_load32_splat(a0 + 4);
113 a0 += 8;
114 v128_t va1x0x0123 = wasm_v128_load32_splat(a1);
115 v128_t va1x1x0123 = wasm_v128_load32_splat(a1 + 4);
116 a1 += 8;
117 v128_t va2x0x0123 = wasm_v128_load32_splat(a2);
118 v128_t va2x1x0123 = wasm_v128_load32_splat(a2 + 4);
119 a2 += 8;
120
121 va0x0x0123 = wasm_v128_xor(va0x0x0123, vsign_mask);
122 va0x1x0123 = wasm_v128_xor(va0x1x0123, vsign_mask);
123 va1x0x0123 = wasm_v128_xor(va1x0x0123, vsign_mask);
124 va1x1x0123 = wasm_v128_xor(va1x1x0123, vsign_mask);
125 va2x0x0123 = wasm_v128_xor(va2x0x0123, vsign_mask);
126 va2x1x0123 = wasm_v128_xor(va2x1x0123, vsign_mask);
127
128 const v128_t vb0x0123 = wasm_v128_load((const int8_t*) w);
129 const v128_t vb0x4567 = wasm_v128_load((const int8_t*) w + 16);
130 const v128_t vb0x89AB = wasm_v128_load((const int8_t*) w + 32);
131 const v128_t vb0xCDEF = wasm_v128_load((const int8_t*) w + 48);
132 const v128_t vb1x0123 = wasm_v128_load((const int8_t*) w + 64);
133 const v128_t vb1x4567 = wasm_v128_load((const int8_t*) w + 80);
134 const v128_t vb1x89AB = wasm_v128_load((const int8_t*) w + 96);
135 const v128_t vb1xCDEF = wasm_v128_load((const int8_t*) w + 112);
136
137 vacc0x0x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0x0123, va0x0x0123, vacc0x0x0123);
138 vacc0x0x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0x4567, va0x0x0123, vacc0x0x4567);
139 vacc0x0x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0x89AB, va0x0x0123, vacc0x0x89AB);
140 vacc0x0xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0xCDEF, va0x0x0123, vacc0x0xCDEF);
141 vacc1x0x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0x0123, va1x0x0123, vacc1x0x0123);
142 vacc1x0x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0x4567, va1x0x0123, vacc1x0x4567);
143 vacc1x0x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0x89AB, va1x0x0123, vacc1x0x89AB);
144 vacc1x0xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0xCDEF, va1x0x0123, vacc1x0xCDEF);
145 vacc2x0x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0x0123, va2x0x0123, vacc2x0x0123);
146 vacc2x0x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0x4567, va2x0x0123, vacc2x0x4567);
147 vacc2x0x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0x89AB, va2x0x0123, vacc2x0x89AB);
148 vacc2x0xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0xCDEF, va2x0x0123, vacc2x0xCDEF);
149 vacc0x1x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1x0123, va0x1x0123, vacc0x1x0123);
150 vacc0x1x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1x4567, va0x1x0123, vacc0x1x4567);
151 vacc0x1x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1x89AB, va0x1x0123, vacc0x1x89AB);
152 vacc0x1xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1xCDEF, va0x1x0123, vacc0x1xCDEF);
153 vacc1x1x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1x0123, va1x1x0123, vacc1x1x0123);
154 vacc1x1x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1x4567, va1x1x0123, vacc1x1x4567);
155 vacc1x1x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1x89AB, va1x1x0123, vacc1x1x89AB);
156 vacc1x1xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1xCDEF, va1x1x0123, vacc1x1xCDEF);
157 vacc2x1x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1x0123, va2x1x0123, vacc2x1x0123);
158 vacc2x1x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1x4567, va2x1x0123, vacc2x1x4567);
159 vacc2x1x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1x89AB, va2x1x0123, vacc2x1x89AB);
160 vacc2x1xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1xCDEF, va2x1x0123, vacc2x1xCDEF);
161
162 w = (const int8_t*) w + 128;
163 k -= 8 * sizeof(int8_t);
164 }
165 vacc0x0x0123 = wasm_i32x4_add(vacc0x0x0123, vacc0x1x0123);
166 vacc0x0x4567 = wasm_i32x4_add(vacc0x0x4567, vacc0x1x4567);
167 vacc0x0x89AB = wasm_i32x4_add(vacc0x0x89AB, vacc0x1x89AB);
168 vacc0x0xCDEF = wasm_i32x4_add(vacc0x0xCDEF, vacc0x1xCDEF);
169 vacc1x0x0123 = wasm_i32x4_add(vacc1x0x0123, vacc1x1x0123);
170 vacc1x0x4567 = wasm_i32x4_add(vacc1x0x4567, vacc1x1x4567);
171 vacc1x0x89AB = wasm_i32x4_add(vacc1x0x89AB, vacc1x1x89AB);
172 vacc1x0xCDEF = wasm_i32x4_add(vacc1x0xCDEF, vacc1x1xCDEF);
173 vacc2x0x0123 = wasm_i32x4_add(vacc2x0x0123, vacc2x1x0123);
174 vacc2x0x4567 = wasm_i32x4_add(vacc2x0x4567, vacc2x1x4567);
175 vacc2x0x89AB = wasm_i32x4_add(vacc2x0x89AB, vacc2x1x89AB);
176 vacc2x0xCDEF = wasm_i32x4_add(vacc2x0xCDEF, vacc2x1xCDEF);
177
178 while (k != 0) {
179 v128_t va0x0123 = wasm_v128_load32_splat(a0);
180 a0 += 4;
181 v128_t va1x0123 = wasm_v128_load32_splat(a1);
182 a1 += 4;
183 v128_t va2x0123 = wasm_v128_load32_splat(a2);
184 a2 += 4;
185
186 va0x0123 = wasm_v128_xor(va0x0123, vsign_mask);
187 va1x0123 = wasm_v128_xor(va1x0123, vsign_mask);
188 va2x0123 = wasm_v128_xor(va2x0123, vsign_mask);
189
190 const v128_t vb0123 = wasm_v128_load((const int8_t*) w);
191 const v128_t vb4567 = wasm_v128_load((const int8_t*) w + 16);
192 const v128_t vb89AB = wasm_v128_load((const int8_t*) w + 32);
193 const v128_t vbCDEF = wasm_v128_load((const int8_t*) w + 48);
194
195 vacc0x0x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0123, va0x0123, vacc0x0x0123);
196 vacc0x0x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb4567, va0x0123, vacc0x0x4567);
197 vacc0x0x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb89AB, va0x0123, vacc0x0x89AB);
198 vacc0x0xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vbCDEF, va0x0123, vacc0x0xCDEF);
199 vacc1x0x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0123, va1x0123, vacc1x0x0123);
200 vacc1x0x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb4567, va1x0123, vacc1x0x4567);
201 vacc1x0x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb89AB, va1x0123, vacc1x0x89AB);
202 vacc1x0xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vbCDEF, va1x0123, vacc1x0xCDEF);
203 vacc2x0x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0123, va2x0123, vacc2x0x0123);
204 vacc2x0x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb4567, va2x0123, vacc2x0x4567);
205 vacc2x0x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb89AB, va2x0123, vacc2x0x89AB);
206 vacc2x0xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vbCDEF, va2x0123, vacc2x0xCDEF);
207
208 w = (const int8_t*) w + 64;
209 k -= 4 * sizeof(int8_t);
210 }
211
212 p -= 3 * sizeof(void*);
213 } while (p != 0);
214
215 vacc0x0x0123 = wasm_f32x4_convert_i32x4(vacc0x0x0123);
216 vacc0x0x4567 = wasm_f32x4_convert_i32x4(vacc0x0x4567);
217 vacc0x0x89AB = wasm_f32x4_convert_i32x4(vacc0x0x89AB);
218 vacc0x0xCDEF = wasm_f32x4_convert_i32x4(vacc0x0xCDEF);
219 vacc1x0x0123 = wasm_f32x4_convert_i32x4(vacc1x0x0123);
220 vacc1x0x4567 = wasm_f32x4_convert_i32x4(vacc1x0x4567);
221 vacc1x0x89AB = wasm_f32x4_convert_i32x4(vacc1x0x89AB);
222 vacc1x0xCDEF = wasm_f32x4_convert_i32x4(vacc1x0xCDEF);
223 vacc2x0x0123 = wasm_f32x4_convert_i32x4(vacc2x0x0123);
224 vacc2x0x4567 = wasm_f32x4_convert_i32x4(vacc2x0x4567);
225 vacc2x0x89AB = wasm_f32x4_convert_i32x4(vacc2x0x89AB);
226 vacc2x0xCDEF = wasm_f32x4_convert_i32x4(vacc2x0xCDEF);
227
228 const v128_t vscale0123 = wasm_v128_load(w);
229 w = (const float*) w + 4;
230 const v128_t vscale4567 = wasm_v128_load(w);
231 w = (const float*) w + 4;
232 const v128_t vscale89AB = wasm_v128_load(w);
233 w = (const float*) w + 4;
234 const v128_t vscaleCDEF = wasm_v128_load(w);
235 w = (const float*) w + 4;
236 vacc0x0x0123 = wasm_f32x4_mul(vacc0x0x0123, vscale0123);
237 vacc0x0x4567 = wasm_f32x4_mul(vacc0x0x4567, vscale4567);
238 vacc0x0x89AB = wasm_f32x4_mul(vacc0x0x89AB, vscale89AB);
239 vacc0x0xCDEF = wasm_f32x4_mul(vacc0x0xCDEF, vscaleCDEF);
240 vacc1x0x0123 = wasm_f32x4_mul(vacc1x0x0123, vscale0123);
241 vacc1x0x4567 = wasm_f32x4_mul(vacc1x0x4567, vscale4567);
242 vacc1x0x89AB = wasm_f32x4_mul(vacc1x0x89AB, vscale89AB);
243 vacc1x0xCDEF = wasm_f32x4_mul(vacc1x0xCDEF, vscaleCDEF);
244 vacc2x0x0123 = wasm_f32x4_mul(vacc2x0x0123, vscale0123);
245 vacc2x0x4567 = wasm_f32x4_mul(vacc2x0x4567, vscale4567);
246 vacc2x0x89AB = wasm_f32x4_mul(vacc2x0x89AB, vscale89AB);
247 vacc2x0xCDEF = wasm_f32x4_mul(vacc2x0xCDEF, vscaleCDEF);
248
249 vacc0x0x0123 = wasm_f32x4_add(vacc0x0x0123, vmagic_bias);
250 vacc0x0x4567 = wasm_f32x4_add(vacc0x0x4567, vmagic_bias);
251 vacc0x0x89AB = wasm_f32x4_add(vacc0x0x89AB, vmagic_bias);
252 vacc0x0xCDEF = wasm_f32x4_add(vacc0x0xCDEF, vmagic_bias);
253 vacc1x0x0123 = wasm_f32x4_add(vacc1x0x0123, vmagic_bias);
254 vacc1x0x4567 = wasm_f32x4_add(vacc1x0x4567, vmagic_bias);
255 vacc1x0x89AB = wasm_f32x4_add(vacc1x0x89AB, vmagic_bias);
256 vacc1x0xCDEF = wasm_f32x4_add(vacc1x0xCDEF, vmagic_bias);
257 vacc2x0x0123 = wasm_f32x4_add(vacc2x0x0123, vmagic_bias);
258 vacc2x0x4567 = wasm_f32x4_add(vacc2x0x4567, vmagic_bias);
259 vacc2x0x89AB = wasm_f32x4_add(vacc2x0x89AB, vmagic_bias);
260 vacc2x0xCDEF = wasm_f32x4_add(vacc2x0xCDEF, vmagic_bias);
261
262 vacc0x0x0123 = wasm_i32x4_max(vacc0x0x0123, vmagic_min);
263 vacc0x0x4567 = wasm_i32x4_max(vacc0x0x4567, vmagic_min);
264 vacc0x0x89AB = wasm_i32x4_max(vacc0x0x89AB, vmagic_min);
265 vacc0x0xCDEF = wasm_i32x4_max(vacc0x0xCDEF, vmagic_min);
266 vacc1x0x0123 = wasm_i32x4_max(vacc1x0x0123, vmagic_min);
267 vacc1x0x4567 = wasm_i32x4_max(vacc1x0x4567, vmagic_min);
268 vacc1x0x89AB = wasm_i32x4_max(vacc1x0x89AB, vmagic_min);
269 vacc1x0xCDEF = wasm_i32x4_max(vacc1x0xCDEF, vmagic_min);
270 vacc2x0x0123 = wasm_i32x4_max(vacc2x0x0123, vmagic_min);
271 vacc2x0x4567 = wasm_i32x4_max(vacc2x0x4567, vmagic_min);
272 vacc2x0x89AB = wasm_i32x4_max(vacc2x0x89AB, vmagic_min);
273 vacc2x0xCDEF = wasm_i32x4_max(vacc2x0xCDEF, vmagic_min);
274
275 vacc0x0x0123 = wasm_i32x4_sub(vacc0x0x0123, vmagic_bias_less_output_zero_point);
276 vacc0x0x4567 = wasm_i32x4_sub(vacc0x0x4567, vmagic_bias_less_output_zero_point);
277 vacc0x0x89AB = wasm_i32x4_sub(vacc0x0x89AB, vmagic_bias_less_output_zero_point);
278 vacc0x0xCDEF = wasm_i32x4_sub(vacc0x0xCDEF, vmagic_bias_less_output_zero_point);
279 vacc1x0x0123 = wasm_i32x4_sub(vacc1x0x0123, vmagic_bias_less_output_zero_point);
280 vacc1x0x4567 = wasm_i32x4_sub(vacc1x0x4567, vmagic_bias_less_output_zero_point);
281 vacc1x0x89AB = wasm_i32x4_sub(vacc1x0x89AB, vmagic_bias_less_output_zero_point);
282 vacc1x0xCDEF = wasm_i32x4_sub(vacc1x0xCDEF, vmagic_bias_less_output_zero_point);
283 vacc2x0x0123 = wasm_i32x4_sub(vacc2x0x0123, vmagic_bias_less_output_zero_point);
284 vacc2x0x4567 = wasm_i32x4_sub(vacc2x0x4567, vmagic_bias_less_output_zero_point);
285 vacc2x0x89AB = wasm_i32x4_sub(vacc2x0x89AB, vmagic_bias_less_output_zero_point);
286 vacc2x0xCDEF = wasm_i32x4_sub(vacc2x0xCDEF, vmagic_bias_less_output_zero_point);
287
288 v128_t vacc0x01234567 = wasm_i16x8_narrow_i32x4(vacc0x0x0123, vacc0x0x4567);
289 v128_t vacc0x89ABCDEF = wasm_i16x8_narrow_i32x4(vacc0x0x89AB, vacc0x0xCDEF);
290 v128_t vacc1x01234567 = wasm_i16x8_narrow_i32x4(vacc1x0x0123, vacc1x0x4567);
291 v128_t vacc1x89ABCDEF = wasm_i16x8_narrow_i32x4(vacc1x0x89AB, vacc1x0xCDEF);
292 v128_t vacc2x01234567 = wasm_i16x8_narrow_i32x4(vacc2x0x0123, vacc2x0x4567);
293 v128_t vacc2x89ABCDEF = wasm_i16x8_narrow_i32x4(vacc2x0x89AB, vacc2x0xCDEF);
294
295 vacc0x01234567 = wasm_i8x16_narrow_i16x8(vacc0x01234567, vacc0x01234567);
296 vacc0x89ABCDEF = wasm_i8x16_narrow_i16x8(vacc0x89ABCDEF, vacc0x89ABCDEF);
297 vacc1x01234567 = wasm_i8x16_narrow_i16x8(vacc1x01234567, vacc1x01234567);
298 vacc1x89ABCDEF = wasm_i8x16_narrow_i16x8(vacc1x89ABCDEF, vacc1x89ABCDEF);
299 vacc2x01234567 = wasm_i8x16_narrow_i16x8(vacc2x01234567, vacc2x01234567);
300 vacc2x89ABCDEF = wasm_i8x16_narrow_i16x8(vacc2x89ABCDEF, vacc2x89ABCDEF);
301
302 vacc0x01234567 = wasm_i8x16_min(vacc0x01234567, voutput_max);
303 vacc0x89ABCDEF = wasm_i8x16_min(vacc0x89ABCDEF, voutput_max);
304 vacc1x01234567 = wasm_i8x16_min(vacc1x01234567, voutput_max);
305 vacc1x89ABCDEF = wasm_i8x16_min(vacc1x89ABCDEF, voutput_max);
306 vacc2x01234567 = wasm_i8x16_min(vacc2x01234567, voutput_max);
307 vacc2x89ABCDEF = wasm_i8x16_min(vacc2x89ABCDEF, voutput_max);
308
309 if (nc >= 16) {
310 wasm_v128_store64_lane(c2, vacc2x01234567, 0);
311 wasm_v128_store64_lane(c2 + 8, vacc2x89ABCDEF, 0);
312 wasm_v128_store64_lane(c1, vacc1x01234567, 0);
313 wasm_v128_store64_lane(c1 + 8, vacc1x89ABCDEF, 0);
314 wasm_v128_store64_lane(c0, vacc0x01234567, 0);
315 wasm_v128_store64_lane(c0 + 8, vacc0x89ABCDEF, 0);
316
317 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
318 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
319 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
320
321 a = (const int8_t**restrict) ((uintptr_t) a - ks);
322
323 nc -= 16;
324 } else {
325 if (nc & 8) {
326 wasm_v128_store64_lane(c2, vacc2x01234567, 0);
327 c2 += 8;
328 wasm_v128_store64_lane(c1, vacc1x01234567, 0);
329 c1 += 8;
330 wasm_v128_store64_lane(c0, vacc0x01234567, 0);
331 c0 += 8;
332
333 vacc0x01234567 = vacc0x89ABCDEF;
334 vacc1x01234567 = vacc1x89ABCDEF;
335 vacc2x01234567 = vacc2x89ABCDEF;
336 }
337 if (nc & 4) {
338 wasm_v128_store32_lane(c2, vacc2x01234567, 0);
339 c2 += 4;
340 wasm_v128_store32_lane(c1, vacc1x01234567, 0);
341 c1 += 4;
342 wasm_v128_store32_lane(c0, vacc0x01234567, 0);
343 c0 += 4;
344
345 vacc0x01234567 = wasm_u64x2_shr(vacc0x01234567, 32);
346 vacc1x01234567 = wasm_u64x2_shr(vacc1x01234567, 32);
347 vacc2x01234567 = wasm_u64x2_shr(vacc2x01234567, 32);
348 }
349 if (nc & 2) {
350 wasm_v128_store16_lane(c2, vacc2x01234567, 0);
351 c2 += 2;
352 wasm_v128_store16_lane(c1, vacc1x01234567, 0);
353 c1 += 2;
354 wasm_v128_store16_lane(c0, vacc0x01234567, 0);
355 c0 += 2;
356
357 vacc0x01234567 = wasm_u32x4_shr(vacc0x01234567, 16);
358 vacc1x01234567 = wasm_u32x4_shr(vacc1x01234567, 16);
359 vacc2x01234567 = wasm_u32x4_shr(vacc2x01234567, 16);
360 }
361 if (nc & 1) {
362 wasm_v128_store8_lane(c2, vacc2x01234567, 0);
363 wasm_v128_store8_lane(c1, vacc1x01234567, 0);
364 wasm_v128_store8_lane(c0, vacc0x01234567, 0);
365 }
366
367 nc = 0;
368 }
369 } while (nc != 0);
370 }