1 // Auto-generated file. Do not edit!
2 // Template: src/f32-igemm/wasmsimd-s4.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2020 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/igemm.h>
15
16
xnn_f32_igemm_ukernel_6x8s4__wasmsimd(size_t mr,size_t nc,size_t kc,size_t ks,const float ** restrict a,const float * restrict w,float * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const float * zero,const union xnn_f32_default_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_f32_igemm_ukernel_6x8s4__wasmsimd(
18 size_t mr,
19 size_t nc,
20 size_t kc,
21 size_t ks,
22 const float**restrict a,
23 const float*restrict w,
24 float*restrict c,
25 size_t cm_stride,
26 size_t cn_stride,
27 size_t a_offset,
28 const float* zero,
29 const union xnn_f32_default_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
30 {
31 assert(mr != 0);
32 assert(mr <= 6);
33 assert(nc != 0);
34 assert(kc != 0);
35 assert(kc % sizeof(float) == 0);
36 assert(ks != 0);
37 assert(ks % (6 * sizeof(void*)) == 0);
38 assert(a_offset % sizeof(float) == 0);
39 assert(a != NULL);
40 assert(w != NULL);
41 assert(c != NULL);
42
43 float* c0 = c;
44 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
45 if XNN_UNPREDICTABLE(mr < 2) {
46 c1 = c0;
47 }
48 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
49 if XNN_UNPREDICTABLE(mr <= 2) {
50 c2 = c1;
51 }
52 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
53 if XNN_UNPREDICTABLE(mr < 4) {
54 c3 = c2;
55 }
56 float* c4 = (float*) ((uintptr_t) c3 + cm_stride);
57 if XNN_UNPREDICTABLE(mr <= 4) {
58 c4 = c3;
59 }
60 float* c5 = (float*) ((uintptr_t) c4 + cm_stride);
61 if XNN_UNPREDICTABLE(mr != 6) {
62 c5 = c4;
63 }
64
65 do {
66 v128_t vacc0x0123 = wasm_v128_load(w);
67 v128_t vacc0x4567 = wasm_v128_load(w + 4);
68 v128_t vacc1x0123 = vacc0x0123;
69 v128_t vacc1x4567 = vacc0x4567;
70 v128_t vacc2x0123 = vacc0x0123;
71 v128_t vacc2x4567 = vacc0x4567;
72 v128_t vacc3x0123 = vacc0x0123;
73 v128_t vacc3x4567 = vacc0x4567;
74 v128_t vacc4x0123 = vacc0x0123;
75 v128_t vacc4x4567 = vacc0x4567;
76 v128_t vacc5x0123 = vacc0x0123;
77 v128_t vacc5x4567 = vacc0x4567;
78 w += 8;
79
80 size_t p = ks;
81 do {
82 const float* restrict a0 = a[0];
83 assert(a0 != NULL);
84 if XNN_UNPREDICTABLE(a0 != zero) {
85 a0 = (const float*) ((uintptr_t) a0 + a_offset);
86 }
87 const float* restrict a1 = a[1];
88 assert(a1 != NULL);
89 if XNN_UNPREDICTABLE(a1 != zero) {
90 a1 = (const float*) ((uintptr_t) a1 + a_offset);
91 }
92 const float* restrict a2 = a[2];
93 assert(a2 != NULL);
94 if XNN_UNPREDICTABLE(a2 != zero) {
95 a2 = (const float*) ((uintptr_t) a2 + a_offset);
96 }
97 const float* restrict a3 = a[3];
98 assert(a3 != NULL);
99 if XNN_UNPREDICTABLE(a3 != zero) {
100 a3 = (const float*) ((uintptr_t) a3 + a_offset);
101 }
102 const float* restrict a4 = a[4];
103 assert(a4 != NULL);
104 if XNN_UNPREDICTABLE(a4 != zero) {
105 a4 = (const float*) ((uintptr_t) a4 + a_offset);
106 }
107 const float* restrict a5 = a[5];
108 assert(a5 != NULL);
109 if XNN_UNPREDICTABLE(a5 != zero) {
110 a5 = (const float*) ((uintptr_t) a5 + a_offset);
111 }
112 a += 6;
113
114 size_t k = kc;
115 while (k >= 4 * sizeof(float)) {
116 v128_t va0 = wasm_v128_load(a0);
117 a0 += 4;
118 v128_t va1 = wasm_v128_load(a1);
119 a1 += 4;
120 v128_t va2 = wasm_v128_load(a2);
121 a2 += 4;
122 v128_t va3 = wasm_v128_load(a3);
123 a3 += 4;
124 v128_t va4 = wasm_v128_load(a4);
125 a4 += 4;
126 v128_t va5 = wasm_v128_load(a5);
127 a5 += 4;
128
129
130 const v128_t vb0123c0 = wasm_v128_load(w + 0);
131 const v128_t vb4567c0 = wasm_v128_load(w + 4);
132
133 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c0));
134 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123c0));
135 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123c0));
136 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3, vb0123c0));
137 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123c0));
138 vacc5x0123 = wasm_f32x4_add(vacc5x0123, wasm_f32x4_mul(va5, vb0123c0));
139 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c0));
140 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1, vb4567c0));
141 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2, vb4567c0));
142 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3, vb4567c0));
143 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4, vb4567c0));
144 vacc5x4567 = wasm_f32x4_add(vacc5x4567, wasm_f32x4_mul(va5, vb4567c0));
145
146 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
147 va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
148 va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
149 va3 = wasm_v32x4_shuffle(va3, va3, 1, 2, 3, 0);
150 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0);
151 va5 = wasm_v32x4_shuffle(va5, va5, 1, 2, 3, 0);
152
153 const v128_t vb0123c1 = wasm_v128_load(w + 8);
154 const v128_t vb4567c1 = wasm_v128_load(w + 12);
155
156 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c1));
157 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123c1));
158 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123c1));
159 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3, vb0123c1));
160 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123c1));
161 vacc5x0123 = wasm_f32x4_add(vacc5x0123, wasm_f32x4_mul(va5, vb0123c1));
162 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c1));
163 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1, vb4567c1));
164 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2, vb4567c1));
165 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3, vb4567c1));
166 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4, vb4567c1));
167 vacc5x4567 = wasm_f32x4_add(vacc5x4567, wasm_f32x4_mul(va5, vb4567c1));
168
169 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
170 va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
171 va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
172 va3 = wasm_v32x4_shuffle(va3, va3, 1, 2, 3, 0);
173 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0);
174 va5 = wasm_v32x4_shuffle(va5, va5, 1, 2, 3, 0);
175
176 const v128_t vb0123c2 = wasm_v128_load(w + 16);
177 const v128_t vb4567c2 = wasm_v128_load(w + 20);
178
179 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c2));
180 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123c2));
181 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123c2));
182 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3, vb0123c2));
183 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123c2));
184 vacc5x0123 = wasm_f32x4_add(vacc5x0123, wasm_f32x4_mul(va5, vb0123c2));
185 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c2));
186 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1, vb4567c2));
187 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2, vb4567c2));
188 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3, vb4567c2));
189 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4, vb4567c2));
190 vacc5x4567 = wasm_f32x4_add(vacc5x4567, wasm_f32x4_mul(va5, vb4567c2));
191
192 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
193 va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
194 va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
195 va3 = wasm_v32x4_shuffle(va3, va3, 1, 2, 3, 0);
196 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0);
197 va5 = wasm_v32x4_shuffle(va5, va5, 1, 2, 3, 0);
198
199 const v128_t vb0123c3 = wasm_v128_load(w + 24);
200 const v128_t vb4567c3 = wasm_v128_load(w + 28);
201
202 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c3));
203 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123c3));
204 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123c3));
205 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3, vb0123c3));
206 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123c3));
207 vacc5x0123 = wasm_f32x4_add(vacc5x0123, wasm_f32x4_mul(va5, vb0123c3));
208 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c3));
209 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1, vb4567c3));
210 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2, vb4567c3));
211 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3, vb4567c3));
212 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4, vb4567c3));
213 vacc5x4567 = wasm_f32x4_add(vacc5x4567, wasm_f32x4_mul(va5, vb4567c3));
214
215
216 w += 32;
217 k -= 4 * sizeof(float);
218 }
219 if XNN_UNLIKELY(k != 0) {
220 v128_t va0 = wasm_v128_load(a0);
221 a0 = (const float*) ((uintptr_t) a0 + k);
222 v128_t va1 = wasm_v128_load(a1);
223 a1 = (const float*) ((uintptr_t) a1 + k);
224 v128_t va2 = wasm_v128_load(a2);
225 a2 = (const float*) ((uintptr_t) a2 + k);
226 v128_t va3 = wasm_v128_load(a3);
227 a3 = (const float*) ((uintptr_t) a3 + k);
228 v128_t va4 = wasm_v128_load(a4);
229 a4 = (const float*) ((uintptr_t) a4 + k);
230 v128_t va5 = wasm_v128_load(a5);
231 a5 = (const float*) ((uintptr_t) a5 + k);
232
233 const v128_t vzero = wasm_f32x4_const_splat(0.0f);
234
235 const v128_t vb0123c0 = wasm_v128_load(w + 0);
236 const v128_t vb4567c0 = wasm_v128_load(w + 4);
237
238 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0));
239 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0));
240 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0));
241 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(wasm_v128_andnot(va3, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0));
242 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(wasm_v128_andnot(va4, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0));
243 vacc5x0123 = wasm_f32x4_add(vacc5x0123, wasm_f32x4_mul(wasm_v128_andnot(va5, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0));
244 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb4567c0, vzero)), vb4567c0));
245 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb4567c0, vzero)), vb4567c0));
246 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb4567c0, vzero)), vb4567c0));
247 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(wasm_v128_andnot(va3, wasm_f32x4_eq(vb4567c0, vzero)), vb4567c0));
248 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(wasm_v128_andnot(va4, wasm_f32x4_eq(vb4567c0, vzero)), vb4567c0));
249 vacc5x4567 = wasm_f32x4_add(vacc5x4567, wasm_f32x4_mul(wasm_v128_andnot(va5, wasm_f32x4_eq(vb4567c0, vzero)), vb4567c0));
250
251 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
252 va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
253 va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
254 va3 = wasm_v32x4_shuffle(va3, va3, 1, 2, 3, 0);
255 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0);
256 va5 = wasm_v32x4_shuffle(va5, va5, 1, 2, 3, 0);
257
258 const v128_t vb0123c1 = wasm_v128_load(w + 8);
259 const v128_t vb4567c1 = wasm_v128_load(w + 12);
260
261 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c1, vzero)), vb0123c1));
262 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb0123c1, vzero)), vb0123c1));
263 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb0123c1, vzero)), vb0123c1));
264 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(wasm_v128_andnot(va3, wasm_f32x4_eq(vb0123c1, vzero)), vb0123c1));
265 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(wasm_v128_andnot(va4, wasm_f32x4_eq(vb0123c1, vzero)), vb0123c1));
266 vacc5x0123 = wasm_f32x4_add(vacc5x0123, wasm_f32x4_mul(wasm_v128_andnot(va5, wasm_f32x4_eq(vb0123c1, vzero)), vb0123c1));
267 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb4567c1, vzero)), vb4567c1));
268 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb4567c1, vzero)), vb4567c1));
269 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb4567c1, vzero)), vb4567c1));
270 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(wasm_v128_andnot(va3, wasm_f32x4_eq(vb4567c1, vzero)), vb4567c1));
271 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(wasm_v128_andnot(va4, wasm_f32x4_eq(vb4567c1, vzero)), vb4567c1));
272 vacc5x4567 = wasm_f32x4_add(vacc5x4567, wasm_f32x4_mul(wasm_v128_andnot(va5, wasm_f32x4_eq(vb4567c1, vzero)), vb4567c1));
273
274 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
275 va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
276 va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
277 va3 = wasm_v32x4_shuffle(va3, va3, 1, 2, 3, 0);
278 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0);
279 va5 = wasm_v32x4_shuffle(va5, va5, 1, 2, 3, 0);
280
281 const v128_t vb0123c2 = wasm_v128_load(w + 16);
282 const v128_t vb4567c2 = wasm_v128_load(w + 20);
283
284 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c2, vzero)), vb0123c2));
285 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb0123c2, vzero)), vb0123c2));
286 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb0123c2, vzero)), vb0123c2));
287 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(wasm_v128_andnot(va3, wasm_f32x4_eq(vb0123c2, vzero)), vb0123c2));
288 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(wasm_v128_andnot(va4, wasm_f32x4_eq(vb0123c2, vzero)), vb0123c2));
289 vacc5x0123 = wasm_f32x4_add(vacc5x0123, wasm_f32x4_mul(wasm_v128_andnot(va5, wasm_f32x4_eq(vb0123c2, vzero)), vb0123c2));
290 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb4567c2, vzero)), vb4567c2));
291 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb4567c2, vzero)), vb4567c2));
292 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb4567c2, vzero)), vb4567c2));
293 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(wasm_v128_andnot(va3, wasm_f32x4_eq(vb4567c2, vzero)), vb4567c2));
294 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(wasm_v128_andnot(va4, wasm_f32x4_eq(vb4567c2, vzero)), vb4567c2));
295 vacc5x4567 = wasm_f32x4_add(vacc5x4567, wasm_f32x4_mul(wasm_v128_andnot(va5, wasm_f32x4_eq(vb4567c2, vzero)), vb4567c2));
296
297 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
298 va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
299 va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
300 va3 = wasm_v32x4_shuffle(va3, va3, 1, 2, 3, 0);
301 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0);
302 va5 = wasm_v32x4_shuffle(va5, va5, 1, 2, 3, 0);
303
304 const v128_t vb0123c3 = wasm_v128_load(w + 24);
305 const v128_t vb4567c3 = wasm_v128_load(w + 28);
306
307 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c3, vzero)), vb0123c3));
308 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb0123c3, vzero)), vb0123c3));
309 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb0123c3, vzero)), vb0123c3));
310 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(wasm_v128_andnot(va3, wasm_f32x4_eq(vb0123c3, vzero)), vb0123c3));
311 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(wasm_v128_andnot(va4, wasm_f32x4_eq(vb0123c3, vzero)), vb0123c3));
312 vacc5x0123 = wasm_f32x4_add(vacc5x0123, wasm_f32x4_mul(wasm_v128_andnot(va5, wasm_f32x4_eq(vb0123c3, vzero)), vb0123c3));
313 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb4567c3, vzero)), vb4567c3));
314 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb4567c3, vzero)), vb4567c3));
315 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb4567c3, vzero)), vb4567c3));
316 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(wasm_v128_andnot(va3, wasm_f32x4_eq(vb4567c3, vzero)), vb4567c3));
317 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(wasm_v128_andnot(va4, wasm_f32x4_eq(vb4567c3, vzero)), vb4567c3));
318 vacc5x4567 = wasm_f32x4_add(vacc5x4567, wasm_f32x4_mul(wasm_v128_andnot(va5, wasm_f32x4_eq(vb4567c3, vzero)), vb4567c3));
319
320
321 w += 32;
322 }
323 p -= 6 * sizeof(void*);
324 } while (p != 0);
325
326
327 if XNN_LIKELY(nc >= 8) {
328 wasm_v128_store(c5, vacc5x0123);
329 wasm_v128_store(c5 + 4, vacc5x4567);
330 c5 = (float*) ((uintptr_t) c5 + cn_stride);
331 wasm_v128_store(c4, vacc4x0123);
332 wasm_v128_store(c4 + 4, vacc4x4567);
333 c4 = (float*) ((uintptr_t) c4 + cn_stride);
334 wasm_v128_store(c3, vacc3x0123);
335 wasm_v128_store(c3 + 4, vacc3x4567);
336 c3 = (float*) ((uintptr_t) c3 + cn_stride);
337 wasm_v128_store(c2, vacc2x0123);
338 wasm_v128_store(c2 + 4, vacc2x4567);
339 c2 = (float*) ((uintptr_t) c2 + cn_stride);
340 wasm_v128_store(c1, vacc1x0123);
341 wasm_v128_store(c1 + 4, vacc1x4567);
342 c1 = (float*) ((uintptr_t) c1 + cn_stride);
343 wasm_v128_store(c0, vacc0x0123);
344 wasm_v128_store(c0 + 4, vacc0x4567);
345 c0 = (float*) ((uintptr_t) c0 + cn_stride);
346
347 a = (const float**restrict) ((uintptr_t) a - ks);
348 nc -= 8;
349 } else {
350 if (nc & 4) {
351 wasm_v128_store(c5, vacc5x0123);
352 wasm_v128_store(c4, vacc4x0123);
353 wasm_v128_store(c3, vacc3x0123);
354 wasm_v128_store(c2, vacc2x0123);
355 wasm_v128_store(c1, vacc1x0123);
356 wasm_v128_store(c0, vacc0x0123);
357
358 vacc5x0123 = vacc5x4567;
359 vacc4x0123 = vacc4x4567;
360 vacc3x0123 = vacc3x4567;
361 vacc2x0123 = vacc2x4567;
362 vacc1x0123 = vacc1x4567;
363 vacc0x0123 = vacc0x4567;
364
365 c5 += 4;
366 c4 += 4;
367 c3 += 4;
368 c2 += 4;
369 c1 += 4;
370 c0 += 4;
371 }
372 if (nc & 2) {
373 *((double*) c5) = wasm_f64x2_extract_lane(vacc5x0123, 0);
374 *((double*) c4) = wasm_f64x2_extract_lane(vacc4x0123, 0);
375 *((double*) c3) = wasm_f64x2_extract_lane(vacc3x0123, 0);
376 *((double*) c2) = wasm_f64x2_extract_lane(vacc2x0123, 0);
377 *((double*) c1) = wasm_f64x2_extract_lane(vacc1x0123, 0);
378 *((double*) c0) = wasm_f64x2_extract_lane(vacc0x0123, 0);
379
380 vacc5x0123 = wasm_v32x4_shuffle(vacc5x0123, vacc5x0123, 2, 3, 2, 3);
381 vacc4x0123 = wasm_v32x4_shuffle(vacc4x0123, vacc4x0123, 2, 3, 2, 3);
382 vacc3x0123 = wasm_v32x4_shuffle(vacc3x0123, vacc3x0123, 2, 3, 2, 3);
383 vacc2x0123 = wasm_v32x4_shuffle(vacc2x0123, vacc2x0123, 2, 3, 2, 3);
384 vacc1x0123 = wasm_v32x4_shuffle(vacc1x0123, vacc1x0123, 2, 3, 2, 3);
385 vacc0x0123 = wasm_v32x4_shuffle(vacc0x0123, vacc0x0123, 2, 3, 2, 3);
386
387 c5 += 2;
388 c4 += 2;
389 c3 += 2;
390 c2 += 2;
391 c1 += 2;
392 c0 += 2;
393 }
394 if (nc & 1) {
395 *c5 = wasm_f32x4_extract_lane(vacc5x0123, 0);
396 *c4 = wasm_f32x4_extract_lane(vacc4x0123, 0);
397 *c3 = wasm_f32x4_extract_lane(vacc3x0123, 0);
398 *c2 = wasm_f32x4_extract_lane(vacc2x0123, 0);
399 *c1 = wasm_f32x4_extract_lane(vacc1x0123, 0);
400 *c0 = wasm_f32x4_extract_lane(vacc0x0123, 0);
401 }
402
403 nc = 0;
404 }
405 } while (nc != 0);
406 }
407