Debian Codegraph @

src/qd8-f32-qc8w-igemm/gen/qd8-f32-qc8w-igemm-4x16c4-minmax-wasmusdot-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_qd8_f32_qc8w_igemm_minmax_ukernel_4x16c4__wasmusdot_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 float* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 size_t a_offset,
29 const int8_t* zero,
30 const int8_t* zero_data,
31 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
32 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
33 {
34 assert(mr != 0);
35 assert(mr <= 4);
36 assert(nc != 0);
37 assert(kc != 0);
38 assert(ks != 0);
39 assert(ks % (4 * sizeof(void*)) == 0);
40 assert(a_offset % sizeof(int8_t) == 0);
41 assert(a != NULL);
42 assert(w != NULL);
43 assert(c != NULL);
44
45 kc = round_up_po2(kc, 4 * sizeof(int8_t));
46 float* c0 = c;
47 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
48 if XNN_UNPREDICTABLE(mr < 2) {
49 c1 = c0;
50 }
51 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
52 if XNN_UNPREDICTABLE(mr <= 2) {
53 c2 = c1;
54 }
55 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
56 if XNN_UNPREDICTABLE(mr != 4) {
57 c3 = c2;
58 }
59
60 const v128_t vmin = wasm_v128_load32_splat(&params->scalar.min);
61 const v128_t vmax = wasm_v128_load32_splat(&params->scalar.max);
62 XNN_FORCE_REALIZATION(vmin);
63 XNN_FORCE_REALIZATION(vmax);
64
65 const v128_t vsign_mask = wasm_u8x16_const_splat(UINT8_C(0x80));
66 XNN_FORCE_REALIZATION(vsign_mask);
67 do {
68 v128_t vksum0123 = wasm_v128_load((const int32_t*) w);
69 v128_t vksum4567 = wasm_v128_load((const int32_t*) w + 4);
70 v128_t vksum89AB = wasm_v128_load((const int32_t*) w + 8);
71 v128_t vksumCDEF = wasm_v128_load((const int32_t*) w + 12);
72 const v128_t vinput_zero_point = wasm_i32x4_splat((int32_t) quantization_params->zero_point + 128);
73 v128_t vacc0x0x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point);
74 v128_t vacc0x0x4567 = wasm_i32x4_mul(vksum4567, vinput_zero_point);
75 v128_t vacc0x0x89AB = wasm_i32x4_mul(vksum89AB, vinput_zero_point);
76 v128_t vacc0x0xCDEF = wasm_i32x4_mul(vksumCDEF, vinput_zero_point);
77 v128_t vacc1x0x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point);
78 v128_t vacc1x0x4567 = wasm_i32x4_mul(vksum4567, vinput_zero_point);
79 v128_t vacc1x0x89AB = wasm_i32x4_mul(vksum89AB, vinput_zero_point);
80 v128_t vacc1x0xCDEF = wasm_i32x4_mul(vksumCDEF, vinput_zero_point);
81 v128_t vacc2x0x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point);
82 v128_t vacc2x0x4567 = wasm_i32x4_mul(vksum4567, vinput_zero_point);
83 v128_t vacc2x0x89AB = wasm_i32x4_mul(vksum89AB, vinput_zero_point);
84 v128_t vacc2x0xCDEF = wasm_i32x4_mul(vksumCDEF, vinput_zero_point);
85 v128_t vacc3x0x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point);
86 v128_t vacc3x0x4567 = wasm_i32x4_mul(vksum4567, vinput_zero_point);
87 v128_t vacc3x0x89AB = wasm_i32x4_mul(vksum89AB, vinput_zero_point);
88 v128_t vacc3x0xCDEF = wasm_i32x4_mul(vksumCDEF, vinput_zero_point);
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 } else {
97 a0 = zero_data;
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 } else {
103 a1 = zero_data;
104 }
105 const int8_t* restrict a2 = a[2];
106 if XNN_UNPREDICTABLE(a2 != zero) {
107 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
108 } else {
109 a2 = zero_data;
110 }
111 const int8_t* restrict a3 = a[3];
112 if XNN_UNPREDICTABLE(a3 != zero) {
113 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
114 } else {
115 a3 = zero_data;
116 }
117 a += 4;
118
119 size_t k = kc;
120 while (k >= 8 * sizeof(int8_t)) {
121 v128_t va0x0x0123 = wasm_v128_load32_splat(a0);
122 v128_t va0x1x0123 = wasm_v128_load32_splat(a0 + 4);
123 a0 += 8;
124 v128_t va1x0x0123 = wasm_v128_load32_splat(a1);
125 v128_t va1x1x0123 = wasm_v128_load32_splat(a1 + 4);
126 a1 += 8;
127 v128_t va2x0x0123 = wasm_v128_load32_splat(a2);
128 v128_t va2x1x0123 = wasm_v128_load32_splat(a2 + 4);
129 a2 += 8;
130 v128_t va3x0x0123 = wasm_v128_load32_splat(a3);
131 v128_t va3x1x0123 = wasm_v128_load32_splat(a3 + 4);
132 a3 += 8;
133
134 va0x0x0123 = wasm_v128_xor(va0x0x0123, vsign_mask);
135 va0x1x0123 = wasm_v128_xor(va0x1x0123, vsign_mask);
136 va1x0x0123 = wasm_v128_xor(va1x0x0123, vsign_mask);
137 va1x1x0123 = wasm_v128_xor(va1x1x0123, vsign_mask);
138 va2x0x0123 = wasm_v128_xor(va2x0x0123, vsign_mask);
139 va2x1x0123 = wasm_v128_xor(va2x1x0123, vsign_mask);
140 va3x0x0123 = wasm_v128_xor(va3x0x0123, vsign_mask);
141 va3x1x0123 = wasm_v128_xor(va3x1x0123, vsign_mask);
142
143 const v128_t vb0x0123 = wasm_v128_load((const int8_t*) w);
144 const v128_t vb0x4567 = wasm_v128_load((const int8_t*) w + 16);
145 const v128_t vb0x89AB = wasm_v128_load((const int8_t*) w + 32);
146 const v128_t vb0xCDEF = wasm_v128_load((const int8_t*) w + 48);
147 const v128_t vb1x0123 = wasm_v128_load((const int8_t*) w + 64);
148 const v128_t vb1x4567 = wasm_v128_load((const int8_t*) w + 80);
149 const v128_t vb1x89AB = wasm_v128_load((const int8_t*) w + 96);
150 const v128_t vb1xCDEF = wasm_v128_load((const int8_t*) w + 112);
151
152 vacc0x0x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0x0123, va0x0x0123, vacc0x0x0123);
153 vacc0x0x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0x4567, va0x0x0123, vacc0x0x4567);
154 vacc0x0x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0x89AB, va0x0x0123, vacc0x0x89AB);
155 vacc0x0xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0xCDEF, va0x0x0123, vacc0x0xCDEF);
156 vacc1x0x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0x0123, va1x0x0123, vacc1x0x0123);
157 vacc1x0x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0x4567, va1x0x0123, vacc1x0x4567);
158 vacc1x0x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0x89AB, va1x0x0123, vacc1x0x89AB);
159 vacc1x0xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0xCDEF, va1x0x0123, vacc1x0xCDEF);
160 vacc2x0x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0x0123, va2x0x0123, vacc2x0x0123);
161 vacc2x0x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0x4567, va2x0x0123, vacc2x0x4567);
162 vacc2x0x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0x89AB, va2x0x0123, vacc2x0x89AB);
163 vacc2x0xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0xCDEF, va2x0x0123, vacc2x0xCDEF);
164 vacc3x0x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0x0123, va3x0x0123, vacc3x0x0123);
165 vacc3x0x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0x4567, va3x0x0123, vacc3x0x4567);
166 vacc3x0x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0x89AB, va3x0x0123, vacc3x0x89AB);
167 vacc3x0xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0xCDEF, va3x0x0123, vacc3x0xCDEF);
168 vacc0x0x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1x0123, va0x1x0123, vacc0x0x0123);
169 vacc0x0x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1x4567, va0x1x0123, vacc0x0x4567);
170 vacc0x0x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1x89AB, va0x1x0123, vacc0x0x89AB);
171 vacc0x0xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1xCDEF, va0x1x0123, vacc0x0xCDEF);
172 vacc1x0x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1x0123, va1x1x0123, vacc1x0x0123);
173 vacc1x0x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1x4567, va1x1x0123, vacc1x0x4567);
174 vacc1x0x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1x89AB, va1x1x0123, vacc1x0x89AB);
175 vacc1x0xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1xCDEF, va1x1x0123, vacc1x0xCDEF);
176 vacc2x0x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1x0123, va2x1x0123, vacc2x0x0123);
177 vacc2x0x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1x4567, va2x1x0123, vacc2x0x4567);
178 vacc2x0x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1x89AB, va2x1x0123, vacc2x0x89AB);
179 vacc2x0xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1xCDEF, va2x1x0123, vacc2x0xCDEF);
180 vacc3x0x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1x0123, va3x1x0123, vacc3x0x0123);
181 vacc3x0x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1x4567, va3x1x0123, vacc3x0x4567);
182 vacc3x0x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1x89AB, va3x1x0123, vacc3x0x89AB);
183 vacc3x0xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1xCDEF, va3x1x0123, vacc3x0xCDEF);
184
185 w = (const int8_t*) w + 128;
186 k -= 8 * sizeof(int8_t);
187 }
188
189 while (k != 0) {
190 v128_t va0x0123 = wasm_v128_load32_splat(a0);
191 a0 += 4;
192 v128_t va1x0123 = wasm_v128_load32_splat(a1);
193 a1 += 4;
194 v128_t va2x0123 = wasm_v128_load32_splat(a2);
195 a2 += 4;
196 v128_t va3x0123 = wasm_v128_load32_splat(a3);
197 a3 += 4;
198
199 va0x0123 = wasm_v128_xor(va0x0123, vsign_mask);
200 va1x0123 = wasm_v128_xor(va1x0123, vsign_mask);
201 va2x0123 = wasm_v128_xor(va2x0123, vsign_mask);
202 va3x0123 = wasm_v128_xor(va3x0123, vsign_mask);
203
204 const v128_t vb0123 = wasm_v128_load((const int8_t*) w);
205 const v128_t vb4567 = wasm_v128_load((const int8_t*) w + 16);
206 const v128_t vb89AB = wasm_v128_load((const int8_t*) w + 32);
207 const v128_t vbCDEF = wasm_v128_load((const int8_t*) w + 48);
208
209 vacc0x0x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0123, va0x0123, vacc0x0x0123);
210 vacc0x0x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb4567, va0x0123, vacc0x0x4567);
211 vacc0x0x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb89AB, va0x0123, vacc0x0x89AB);
212 vacc0x0xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vbCDEF, va0x0123, vacc0x0xCDEF);
213 vacc1x0x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0123, va1x0123, vacc1x0x0123);
214 vacc1x0x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb4567, va1x0123, vacc1x0x4567);
215 vacc1x0x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb89AB, va1x0123, vacc1x0x89AB);
216 vacc1x0xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vbCDEF, va1x0123, vacc1x0xCDEF);
217 vacc2x0x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0123, va2x0123, vacc2x0x0123);
218 vacc2x0x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb4567, va2x0123, vacc2x0x4567);
219 vacc2x0x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb89AB, va2x0123, vacc2x0x89AB);
220 vacc2x0xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vbCDEF, va2x0123, vacc2x0xCDEF);
221 vacc3x0x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0123, va3x0123, vacc3x0x0123);
222 vacc3x0x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb4567, va3x0123, vacc3x0x4567);
223 vacc3x0x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb89AB, va3x0123, vacc3x0x89AB);
224 vacc3x0xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vbCDEF, va3x0123, vacc3x0xCDEF);
225
226 w = (const int8_t*) w + 64;
227 k -= 4 * sizeof(int8_t);
228 }
229
230 p -= 4 * sizeof(void*);
231 } while (p != 0);
232
233 vacc0x0x0123 = wasm_f32x4_convert_i32x4(vacc0x0x0123);
234 vacc0x0x4567 = wasm_f32x4_convert_i32x4(vacc0x0x4567);
235 vacc0x0x89AB = wasm_f32x4_convert_i32x4(vacc0x0x89AB);
236 vacc0x0xCDEF = wasm_f32x4_convert_i32x4(vacc0x0xCDEF);
237 vacc1x0x0123 = wasm_f32x4_convert_i32x4(vacc1x0x0123);
238 vacc1x0x4567 = wasm_f32x4_convert_i32x4(vacc1x0x4567);
239 vacc1x0x89AB = wasm_f32x4_convert_i32x4(vacc1x0x89AB);
240 vacc1x0xCDEF = wasm_f32x4_convert_i32x4(vacc1x0xCDEF);
241 vacc2x0x0123 = wasm_f32x4_convert_i32x4(vacc2x0x0123);
242 vacc2x0x4567 = wasm_f32x4_convert_i32x4(vacc2x0x4567);
243 vacc2x0x89AB = wasm_f32x4_convert_i32x4(vacc2x0x89AB);
244 vacc2x0xCDEF = wasm_f32x4_convert_i32x4(vacc2x0xCDEF);
245 vacc3x0x0123 = wasm_f32x4_convert_i32x4(vacc3x0x0123);
246 vacc3x0x4567 = wasm_f32x4_convert_i32x4(vacc3x0x4567);
247 vacc3x0x89AB = wasm_f32x4_convert_i32x4(vacc3x0x89AB);
248 vacc3x0xCDEF = wasm_f32x4_convert_i32x4(vacc3x0xCDEF);
249
250 const v128_t vinput_scale = wasm_v128_load32_splat(&quantization_params->inv_scale);
251
252 vacc0x0x0123 = wasm_f32x4_mul(vacc0x0x0123, vinput_scale);
253 vacc0x0x4567 = wasm_f32x4_mul(vacc0x0x4567, vinput_scale);
254 vacc0x0x89AB = wasm_f32x4_mul(vacc0x0x89AB, vinput_scale);
255 vacc0x0xCDEF = wasm_f32x4_mul(vacc0x0xCDEF, vinput_scale);
256 vacc1x0x0123 = wasm_f32x4_mul(vacc1x0x0123, vinput_scale);
257 vacc1x0x4567 = wasm_f32x4_mul(vacc1x0x4567, vinput_scale);
258 vacc1x0x89AB = wasm_f32x4_mul(vacc1x0x89AB, vinput_scale);
259 vacc1x0xCDEF = wasm_f32x4_mul(vacc1x0xCDEF, vinput_scale);
260 vacc2x0x0123 = wasm_f32x4_mul(vacc2x0x0123, vinput_scale);
261 vacc2x0x4567 = wasm_f32x4_mul(vacc2x0x4567, vinput_scale);
262 vacc2x0x89AB = wasm_f32x4_mul(vacc2x0x89AB, vinput_scale);
263 vacc2x0xCDEF = wasm_f32x4_mul(vacc2x0xCDEF, vinput_scale);
264 vacc3x0x0123 = wasm_f32x4_mul(vacc3x0x0123, vinput_scale);
265 vacc3x0x4567 = wasm_f32x4_mul(vacc3x0x4567, vinput_scale);
266 vacc3x0x89AB = wasm_f32x4_mul(vacc3x0x89AB, vinput_scale);
267 vacc3x0xCDEF = wasm_f32x4_mul(vacc3x0xCDEF, vinput_scale);
268
269 const v128_t vfilter_output_scale0123 = wasm_v128_load(w);
270 w = (const float*) w + 4;
271 const v128_t vfilter_output_scale4567 = wasm_v128_load(w);
272 w = (const float*) w + 4;
273 const v128_t vfilter_output_scale89AB = wasm_v128_load(w);
274 w = (const float*) w + 4;
275 const v128_t vfilter_output_scaleCDEF = wasm_v128_load(w);
276 w = (const float*) w + 4;
277 vacc0x0x0123 = wasm_f32x4_mul(vacc0x0x0123, vfilter_output_scale0123);
278 vacc0x0x4567 = wasm_f32x4_mul(vacc0x0x4567, vfilter_output_scale4567);
279 vacc0x0x89AB = wasm_f32x4_mul(vacc0x0x89AB, vfilter_output_scale89AB);
280 vacc0x0xCDEF = wasm_f32x4_mul(vacc0x0xCDEF, vfilter_output_scaleCDEF);
281 vacc1x0x0123 = wasm_f32x4_mul(vacc1x0x0123, vfilter_output_scale0123);
282 vacc1x0x4567 = wasm_f32x4_mul(vacc1x0x4567, vfilter_output_scale4567);
283 vacc1x0x89AB = wasm_f32x4_mul(vacc1x0x89AB, vfilter_output_scale89AB);
284 vacc1x0xCDEF = wasm_f32x4_mul(vacc1x0xCDEF, vfilter_output_scaleCDEF);
285 vacc2x0x0123 = wasm_f32x4_mul(vacc2x0x0123, vfilter_output_scale0123);
286 vacc2x0x4567 = wasm_f32x4_mul(vacc2x0x4567, vfilter_output_scale4567);
287 vacc2x0x89AB = wasm_f32x4_mul(vacc2x0x89AB, vfilter_output_scale89AB);
288 vacc2x0xCDEF = wasm_f32x4_mul(vacc2x0xCDEF, vfilter_output_scaleCDEF);
289 vacc3x0x0123 = wasm_f32x4_mul(vacc3x0x0123, vfilter_output_scale0123);
290 vacc3x0x4567 = wasm_f32x4_mul(vacc3x0x4567, vfilter_output_scale4567);
291 vacc3x0x89AB = wasm_f32x4_mul(vacc3x0x89AB, vfilter_output_scale89AB);
292 vacc3x0xCDEF = wasm_f32x4_mul(vacc3x0xCDEF, vfilter_output_scaleCDEF);
293
294 const v128_t vbias0123 = wasm_v128_load(w);
295 w = (const float*) w + 4;
296 const v128_t vbias4567 = wasm_v128_load(w);
297 w = (const float*) w + 4;
298 const v128_t vbias89AB = wasm_v128_load(w);
299 w = (const float*) w + 4;
300 const v128_t vbiasCDEF = wasm_v128_load(w);
301 w = (const float*) w + 4;
302 vacc0x0x0123 = wasm_f32x4_add(vacc0x0x0123, vbias0123);
303 vacc0x0x4567 = wasm_f32x4_add(vacc0x0x4567, vbias4567);
304 vacc0x0x89AB = wasm_f32x4_add(vacc0x0x89AB, vbias89AB);
305 vacc0x0xCDEF = wasm_f32x4_add(vacc0x0xCDEF, vbiasCDEF);
306 vacc1x0x0123 = wasm_f32x4_add(vacc1x0x0123, vbias0123);
307 vacc1x0x4567 = wasm_f32x4_add(vacc1x0x4567, vbias4567);
308 vacc1x0x89AB = wasm_f32x4_add(vacc1x0x89AB, vbias89AB);
309 vacc1x0xCDEF = wasm_f32x4_add(vacc1x0xCDEF, vbiasCDEF);
310 vacc2x0x0123 = wasm_f32x4_add(vacc2x0x0123, vbias0123);
311 vacc2x0x4567 = wasm_f32x4_add(vacc2x0x4567, vbias4567);
312 vacc2x0x89AB = wasm_f32x4_add(vacc2x0x89AB, vbias89AB);
313 vacc2x0xCDEF = wasm_f32x4_add(vacc2x0xCDEF, vbiasCDEF);
314 vacc3x0x0123 = wasm_f32x4_add(vacc3x0x0123, vbias0123);
315 vacc3x0x4567 = wasm_f32x4_add(vacc3x0x4567, vbias4567);
316 vacc3x0x89AB = wasm_f32x4_add(vacc3x0x89AB, vbias89AB);
317 vacc3x0xCDEF = wasm_f32x4_add(vacc3x0xCDEF, vbiasCDEF);
318
319 vacc0x0x0123 = wasm_f32x4_pmax(vacc0x0x0123, vmin);
320 vacc0x0x4567 = wasm_f32x4_pmax(vacc0x0x4567, vmin);
321 vacc0x0x89AB = wasm_f32x4_pmax(vacc0x0x89AB, vmin);
322 vacc0x0xCDEF = wasm_f32x4_pmax(vacc0x0xCDEF, vmin);
323 vacc1x0x0123 = wasm_f32x4_pmax(vacc1x0x0123, vmin);
324 vacc1x0x4567 = wasm_f32x4_pmax(vacc1x0x4567, vmin);
325 vacc1x0x89AB = wasm_f32x4_pmax(vacc1x0x89AB, vmin);
326 vacc1x0xCDEF = wasm_f32x4_pmax(vacc1x0xCDEF, vmin);
327 vacc2x0x0123 = wasm_f32x4_pmax(vacc2x0x0123, vmin);
328 vacc2x0x4567 = wasm_f32x4_pmax(vacc2x0x4567, vmin);
329 vacc2x0x89AB = wasm_f32x4_pmax(vacc2x0x89AB, vmin);
330 vacc2x0xCDEF = wasm_f32x4_pmax(vacc2x0xCDEF, vmin);
331 vacc3x0x0123 = wasm_f32x4_pmax(vacc3x0x0123, vmin);
332 vacc3x0x4567 = wasm_f32x4_pmax(vacc3x0x4567, vmin);
333 vacc3x0x89AB = wasm_f32x4_pmax(vacc3x0x89AB, vmin);
334 vacc3x0xCDEF = wasm_f32x4_pmax(vacc3x0xCDEF, vmin);
335
336 vacc0x0x0123 = wasm_f32x4_pmin(vacc0x0x0123, vmax);
337 vacc0x0x4567 = wasm_f32x4_pmin(vacc0x0x4567, vmax);
338 vacc0x0x89AB = wasm_f32x4_pmin(vacc0x0x89AB, vmax);
339 vacc0x0xCDEF = wasm_f32x4_pmin(vacc0x0xCDEF, vmax);
340 vacc1x0x0123 = wasm_f32x4_pmin(vacc1x0x0123, vmax);
341 vacc1x0x4567 = wasm_f32x4_pmin(vacc1x0x4567, vmax);
342 vacc1x0x89AB = wasm_f32x4_pmin(vacc1x0x89AB, vmax);
343 vacc1x0xCDEF = wasm_f32x4_pmin(vacc1x0xCDEF, vmax);
344 vacc2x0x0123 = wasm_f32x4_pmin(vacc2x0x0123, vmax);
345 vacc2x0x4567 = wasm_f32x4_pmin(vacc2x0x4567, vmax);
346 vacc2x0x89AB = wasm_f32x4_pmin(vacc2x0x89AB, vmax);
347 vacc2x0xCDEF = wasm_f32x4_pmin(vacc2x0xCDEF, vmax);
348 vacc3x0x0123 = wasm_f32x4_pmin(vacc3x0x0123, vmax);
349 vacc3x0x4567 = wasm_f32x4_pmin(vacc3x0x4567, vmax);
350 vacc3x0x89AB = wasm_f32x4_pmin(vacc3x0x89AB, vmax);
351 vacc3x0xCDEF = wasm_f32x4_pmin(vacc3x0xCDEF, vmax);
352
353 if XNN_LIKELY(nc >= 16) {
354 wasm_v128_store(c3, vacc3x0x0123);
355 wasm_v128_store(c3 + 4, vacc3x0x4567);
356 wasm_v128_store(c3 + 8, vacc3x0x89AB);
357 wasm_v128_store(c3 + 12, vacc3x0xCDEF);
358 wasm_v128_store(c2, vacc2x0x0123);
359 wasm_v128_store(c2 + 4, vacc2x0x4567);
360 wasm_v128_store(c2 + 8, vacc2x0x89AB);
361 wasm_v128_store(c2 + 12, vacc2x0xCDEF);
362 wasm_v128_store(c1, vacc1x0x0123);
363 wasm_v128_store(c1 + 4, vacc1x0x4567);
364 wasm_v128_store(c1 + 8, vacc1x0x89AB);
365 wasm_v128_store(c1 + 12, vacc1x0xCDEF);
366 wasm_v128_store(c0, vacc0x0x0123);
367 wasm_v128_store(c0 + 4, vacc0x0x4567);
368 wasm_v128_store(c0 + 8, vacc0x0x89AB);
369 wasm_v128_store(c0 + 12, vacc0x0xCDEF);
370
371 c0 = (float*) ((uintptr_t) c0 + cn_stride);
372 c1 = (float*) ((uintptr_t) c1 + cn_stride);
373 c2 = (float*) ((uintptr_t) c2 + cn_stride);
374 c3 = (float*) ((uintptr_t) c3 + cn_stride);
375
376 a = (const int8_t**restrict) ((uintptr_t) a - ks);
377
378 nc -= 16;
379 } else {
380 if (nc & 8) {
381 wasm_v128_store(c3, vacc3x0x0123);
382 vacc3x0x0123 = vacc3x0x89AB;
383 c3 += 4;
384 wasm_v128_store(c3, vacc3x0x4567);
385 vacc3x0x4567 = vacc3x0xCDEF;
386 c3 += 4;
387 wasm_v128_store(c2, vacc2x0x0123);
388 vacc2x0x0123 = vacc2x0x89AB;
389 c2 += 4;
390 wasm_v128_store(c2, vacc2x0x4567);
391 vacc2x0x4567 = vacc2x0xCDEF;
392 c2 += 4;
393 wasm_v128_store(c1, vacc1x0x0123);
394 vacc1x0x0123 = vacc1x0x89AB;
395 c1 += 4;
396 wasm_v128_store(c1, vacc1x0x4567);
397 vacc1x0x4567 = vacc1x0xCDEF;
398 c1 += 4;
399 wasm_v128_store(c0, vacc0x0x0123);
400 vacc0x0x0123 = vacc0x0x89AB;
401 c0 += 4;
402 wasm_v128_store(c0, vacc0x0x4567);
403 vacc0x0x4567 = vacc0x0xCDEF;
404 c0 += 4;
405 }
406 if (nc & 4) {
407 wasm_v128_store(c3, vacc3x0x0123);
408 vacc3x0x0123 = vacc3x0x4567;
409 c3 += 4;
410 wasm_v128_store(c2, vacc2x0x0123);
411 vacc2x0x0123 = vacc2x0x4567;
412 c2 += 4;
413 wasm_v128_store(c1, vacc1x0x0123);
414 vacc1x0x0123 = vacc1x0x4567;
415 c1 += 4;
416 wasm_v128_store(c0, vacc0x0x0123);
417 vacc0x0x0123 = vacc0x0x4567;
418 c0 += 4;
419 }
420 if (nc & 2) {
421 wasm_v128_store64_lane(c3, vacc3x0x0123, 0);
422 vacc3x0x0123 = wasm_v64x2_shuffle(vacc3x0x0123, vacc3x0x0123, 1, 1);
423 c3 += 2;
424 wasm_v128_store64_lane(c2, vacc2x0x0123, 0);
425 vacc2x0x0123 = wasm_v64x2_shuffle(vacc2x0x0123, vacc2x0x0123, 1, 1);
426 c2 += 2;
427 wasm_v128_store64_lane(c1, vacc1x0x0123, 0);
428 vacc1x0x0123 = wasm_v64x2_shuffle(vacc1x0x0123, vacc1x0x0123, 1, 1);
429 c1 += 2;
430 wasm_v128_store64_lane(c0, vacc0x0x0123, 0);
431 vacc0x0x0123 = wasm_v64x2_shuffle(vacc0x0x0123, vacc0x0x0123, 1, 1);
432 c0 += 2;
433 }
434 if (nc & 1) {
435 wasm_v128_store32_lane(c3, vacc3x0x0123, 0);
436 wasm_v128_store32_lane(c2, vacc2x0x0123, 0);
437 wasm_v128_store32_lane(c1, vacc1x0x0123, 0);
438 wasm_v128_store32_lane(c0, vacc0x0x0123, 0);
439 }
440 nc = 0;
441 }
442 } while (nc != 0);
443 }