1 // Auto-generated file. Do not edit!
2 // Template: src/f32-gemm/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/gemm.h>
15
16
xnn_f32_gemm_relu_ukernel_6x8s4__wasmsimd(size_t mr,size_t nc,size_t kc,const float * restrict a,size_t a_stride,const float * restrict w,float * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_f32_relu_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_f32_gemm_relu_ukernel_6x8s4__wasmsimd(
18 size_t mr,
19 size_t nc,
20 size_t kc,
21 const float*restrict a,
22 size_t a_stride,
23 const float*restrict w,
24 float*restrict c,
25 size_t cm_stride,
26 size_t cn_stride,
27 const union xnn_f32_relu_params params[restrict XNN_MIN_ELEMENTS(1)])
28 {
29 assert(mr != 0);
30 assert(mr <= 6);
31 assert(nc != 0);
32 assert(kc != 0);
33 assert(kc % sizeof(float) == 0);
34 assert(a != NULL);
35 assert(w != NULL);
36 assert(c != NULL);
37
38 const float* a0 = a;
39 float* c0 = c;
40 const float* a1 = (const float*) ((uintptr_t) a0 + a_stride);
41 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
42 if XNN_UNPREDICTABLE(mr < 2) {
43 a1 = a0;
44 c1 = c0;
45 }
46 const float* a2 = (const float*) ((uintptr_t) a1 + a_stride);
47 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
48 if XNN_UNPREDICTABLE(mr <= 2) {
49 a2 = a1;
50 c2 = c1;
51 }
52 const float* a3 = (const float*) ((uintptr_t) a2 + a_stride);
53 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
54 if XNN_UNPREDICTABLE(mr < 4) {
55 a3 = a2;
56 c3 = c2;
57 }
58 const float* a4 = (const float*) ((uintptr_t) a3 + a_stride);
59 float* c4 = (float*) ((uintptr_t) c3 + cm_stride);
60 if XNN_UNPREDICTABLE(mr <= 4) {
61 a4 = a3;
62 c4 = c3;
63 }
64 const float* a5 = (const float*) ((uintptr_t) a4 + a_stride);
65 float* c5 = (float*) ((uintptr_t) c4 + cm_stride);
66 if XNN_UNPREDICTABLE(mr != 6) {
67 a5 = a4;
68 c5 = c4;
69 }
70
71 do {
72 v128_t vacc0x0123 = wasm_v128_load(w + 0);
73 v128_t vacc0x4567 = wasm_v128_load(w + 4);
74 v128_t vacc1x0123 = vacc0x0123;
75 v128_t vacc1x4567 = vacc0x4567;
76 v128_t vacc2x0123 = vacc0x0123;
77 v128_t vacc2x4567 = vacc0x4567;
78 v128_t vacc3x0123 = vacc0x0123;
79 v128_t vacc3x4567 = vacc0x4567;
80 v128_t vacc4x0123 = vacc0x0123;
81 v128_t vacc4x4567 = vacc0x4567;
82 v128_t vacc5x0123 = vacc0x0123;
83 v128_t vacc5x4567 = vacc0x4567;
84 w += 8;
85
86 size_t k = kc;
87 while (k >= 4 * sizeof(float)) {
88 v128_t va0 = wasm_v128_load(a0);
89 a0 += 4;
90 v128_t va1 = wasm_v128_load(a1);
91 a1 += 4;
92 v128_t va2 = wasm_v128_load(a2);
93 a2 += 4;
94 v128_t va3 = wasm_v128_load(a3);
95 a3 += 4;
96 v128_t va4 = wasm_v128_load(a4);
97 a4 += 4;
98 v128_t va5 = wasm_v128_load(a5);
99 a5 += 4;
100
101
102 const v128_t vb0123c0 = wasm_v128_load(w + 0);
103 const v128_t vb4567c0 = wasm_v128_load(w + 4);
104
105 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c0));
106 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123c0));
107 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123c0));
108 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3, vb0123c0));
109 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123c0));
110 vacc5x0123 = wasm_f32x4_add(vacc5x0123, wasm_f32x4_mul(va5, vb0123c0));
111 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c0));
112 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1, vb4567c0));
113 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2, vb4567c0));
114 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3, vb4567c0));
115 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4, vb4567c0));
116 vacc5x4567 = wasm_f32x4_add(vacc5x4567, wasm_f32x4_mul(va5, vb4567c0));
117
118 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
119 va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
120 va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
121 va3 = wasm_v32x4_shuffle(va3, va3, 1, 2, 3, 0);
122 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0);
123 va5 = wasm_v32x4_shuffle(va5, va5, 1, 2, 3, 0);
124
125 const v128_t vb0123c1 = wasm_v128_load(w + 8);
126 const v128_t vb4567c1 = wasm_v128_load(w + 12);
127
128 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c1));
129 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123c1));
130 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123c1));
131 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3, vb0123c1));
132 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123c1));
133 vacc5x0123 = wasm_f32x4_add(vacc5x0123, wasm_f32x4_mul(va5, vb0123c1));
134 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c1));
135 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1, vb4567c1));
136 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2, vb4567c1));
137 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3, vb4567c1));
138 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4, vb4567c1));
139 vacc5x4567 = wasm_f32x4_add(vacc5x4567, wasm_f32x4_mul(va5, vb4567c1));
140
141 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
142 va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
143 va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
144 va3 = wasm_v32x4_shuffle(va3, va3, 1, 2, 3, 0);
145 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0);
146 va5 = wasm_v32x4_shuffle(va5, va5, 1, 2, 3, 0);
147
148 const v128_t vb0123c2 = wasm_v128_load(w + 16);
149 const v128_t vb4567c2 = wasm_v128_load(w + 20);
150
151 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c2));
152 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123c2));
153 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123c2));
154 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3, vb0123c2));
155 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123c2));
156 vacc5x0123 = wasm_f32x4_add(vacc5x0123, wasm_f32x4_mul(va5, vb0123c2));
157 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c2));
158 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1, vb4567c2));
159 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2, vb4567c2));
160 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3, vb4567c2));
161 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4, vb4567c2));
162 vacc5x4567 = wasm_f32x4_add(vacc5x4567, wasm_f32x4_mul(va5, vb4567c2));
163
164 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
165 va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
166 va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
167 va3 = wasm_v32x4_shuffle(va3, va3, 1, 2, 3, 0);
168 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0);
169 va5 = wasm_v32x4_shuffle(va5, va5, 1, 2, 3, 0);
170
171 const v128_t vb0123c3 = wasm_v128_load(w + 24);
172 const v128_t vb4567c3 = wasm_v128_load(w + 28);
173
174 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c3));
175 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123c3));
176 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123c3));
177 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3, vb0123c3));
178 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123c3));
179 vacc5x0123 = wasm_f32x4_add(vacc5x0123, wasm_f32x4_mul(va5, vb0123c3));
180 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c3));
181 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1, vb4567c3));
182 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2, vb4567c3));
183 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3, vb4567c3));
184 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4, vb4567c3));
185 vacc5x4567 = wasm_f32x4_add(vacc5x4567, wasm_f32x4_mul(va5, vb4567c3));
186
187
188 w += 32;
189 k -= 4 * sizeof(float);
190 }
191 if XNN_UNLIKELY(k != 0) {
192 v128_t va0 = wasm_v128_load(a0);
193 a0 = (const float*) ((uintptr_t) a0 + k);
194 v128_t va1 = wasm_v128_load(a1);
195 a1 = (const float*) ((uintptr_t) a1 + k);
196 v128_t va2 = wasm_v128_load(a2);
197 a2 = (const float*) ((uintptr_t) a2 + k);
198 v128_t va3 = wasm_v128_load(a3);
199 a3 = (const float*) ((uintptr_t) a3 + k);
200 v128_t va4 = wasm_v128_load(a4);
201 a4 = (const float*) ((uintptr_t) a4 + k);
202 v128_t va5 = wasm_v128_load(a5);
203 a5 = (const float*) ((uintptr_t) a5 + k);
204
205 const v128_t vzero = wasm_f32x4_const_splat(0.0f);
206
207 const v128_t vb0123c0 = wasm_v128_load(w + 0);
208 const v128_t vb4567c0 = wasm_v128_load(w + 4);
209
210 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0));
211 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0));
212 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0));
213 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(wasm_v128_andnot(va3, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0));
214 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(wasm_v128_andnot(va4, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0));
215 vacc5x0123 = wasm_f32x4_add(vacc5x0123, wasm_f32x4_mul(wasm_v128_andnot(va5, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0));
216 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb4567c0, vzero)), vb4567c0));
217 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb4567c0, vzero)), vb4567c0));
218 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb4567c0, vzero)), vb4567c0));
219 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(wasm_v128_andnot(va3, wasm_f32x4_eq(vb4567c0, vzero)), vb4567c0));
220 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(wasm_v128_andnot(va4, wasm_f32x4_eq(vb4567c0, vzero)), vb4567c0));
221 vacc5x4567 = wasm_f32x4_add(vacc5x4567, wasm_f32x4_mul(wasm_v128_andnot(va5, wasm_f32x4_eq(vb4567c0, vzero)), vb4567c0));
222
223 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
224 va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
225 va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
226 va3 = wasm_v32x4_shuffle(va3, va3, 1, 2, 3, 0);
227 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0);
228 va5 = wasm_v32x4_shuffle(va5, va5, 1, 2, 3, 0);
229
230 const v128_t vb0123c1 = wasm_v128_load(w + 8);
231 const v128_t vb4567c1 = wasm_v128_load(w + 12);
232
233 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c1, vzero)), vb0123c1));
234 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb0123c1, vzero)), vb0123c1));
235 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb0123c1, vzero)), vb0123c1));
236 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(wasm_v128_andnot(va3, wasm_f32x4_eq(vb0123c1, vzero)), vb0123c1));
237 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(wasm_v128_andnot(va4, wasm_f32x4_eq(vb0123c1, vzero)), vb0123c1));
238 vacc5x0123 = wasm_f32x4_add(vacc5x0123, wasm_f32x4_mul(wasm_v128_andnot(va5, wasm_f32x4_eq(vb0123c1, vzero)), vb0123c1));
239 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb4567c1, vzero)), vb4567c1));
240 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb4567c1, vzero)), vb4567c1));
241 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb4567c1, vzero)), vb4567c1));
242 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(wasm_v128_andnot(va3, wasm_f32x4_eq(vb4567c1, vzero)), vb4567c1));
243 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(wasm_v128_andnot(va4, wasm_f32x4_eq(vb4567c1, vzero)), vb4567c1));
244 vacc5x4567 = wasm_f32x4_add(vacc5x4567, wasm_f32x4_mul(wasm_v128_andnot(va5, wasm_f32x4_eq(vb4567c1, vzero)), vb4567c1));
245
246 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
247 va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
248 va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
249 va3 = wasm_v32x4_shuffle(va3, va3, 1, 2, 3, 0);
250 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0);
251 va5 = wasm_v32x4_shuffle(va5, va5, 1, 2, 3, 0);
252
253 const v128_t vb0123c2 = wasm_v128_load(w + 16);
254 const v128_t vb4567c2 = wasm_v128_load(w + 20);
255
256 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c2, vzero)), vb0123c2));
257 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb0123c2, vzero)), vb0123c2));
258 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb0123c2, vzero)), vb0123c2));
259 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(wasm_v128_andnot(va3, wasm_f32x4_eq(vb0123c2, vzero)), vb0123c2));
260 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(wasm_v128_andnot(va4, wasm_f32x4_eq(vb0123c2, vzero)), vb0123c2));
261 vacc5x0123 = wasm_f32x4_add(vacc5x0123, wasm_f32x4_mul(wasm_v128_andnot(va5, wasm_f32x4_eq(vb0123c2, vzero)), vb0123c2));
262 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb4567c2, vzero)), vb4567c2));
263 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb4567c2, vzero)), vb4567c2));
264 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb4567c2, vzero)), vb4567c2));
265 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(wasm_v128_andnot(va3, wasm_f32x4_eq(vb4567c2, vzero)), vb4567c2));
266 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(wasm_v128_andnot(va4, wasm_f32x4_eq(vb4567c2, vzero)), vb4567c2));
267 vacc5x4567 = wasm_f32x4_add(vacc5x4567, wasm_f32x4_mul(wasm_v128_andnot(va5, wasm_f32x4_eq(vb4567c2, vzero)), vb4567c2));
268
269 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
270 va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
271 va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
272 va3 = wasm_v32x4_shuffle(va3, va3, 1, 2, 3, 0);
273 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0);
274 va5 = wasm_v32x4_shuffle(va5, va5, 1, 2, 3, 0);
275
276 const v128_t vb0123c3 = wasm_v128_load(w + 24);
277 const v128_t vb4567c3 = wasm_v128_load(w + 28);
278
279 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c3, vzero)), vb0123c3));
280 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb0123c3, vzero)), vb0123c3));
281 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb0123c3, vzero)), vb0123c3));
282 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(wasm_v128_andnot(va3, wasm_f32x4_eq(vb0123c3, vzero)), vb0123c3));
283 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(wasm_v128_andnot(va4, wasm_f32x4_eq(vb0123c3, vzero)), vb0123c3));
284 vacc5x0123 = wasm_f32x4_add(vacc5x0123, wasm_f32x4_mul(wasm_v128_andnot(va5, wasm_f32x4_eq(vb0123c3, vzero)), vb0123c3));
285 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb4567c3, vzero)), vb4567c3));
286 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb4567c3, vzero)), vb4567c3));
287 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb4567c3, vzero)), vb4567c3));
288 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(wasm_v128_andnot(va3, wasm_f32x4_eq(vb4567c3, vzero)), vb4567c3));
289 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(wasm_v128_andnot(va4, wasm_f32x4_eq(vb4567c3, vzero)), vb4567c3));
290 vacc5x4567 = wasm_f32x4_add(vacc5x4567, wasm_f32x4_mul(wasm_v128_andnot(va5, wasm_f32x4_eq(vb4567c3, vzero)), vb4567c3));
291
292
293 w += 32;
294 }
295
296 const v128_t vzero = wasm_i32x4_const_splat(0);
297 vacc0x0123 = wasm_i32x4_max(vacc0x0123, vzero);
298 vacc1x0123 = wasm_i32x4_max(vacc1x0123, vzero);
299 vacc2x0123 = wasm_i32x4_max(vacc2x0123, vzero);
300 vacc3x0123 = wasm_i32x4_max(vacc3x0123, vzero);
301 vacc4x0123 = wasm_i32x4_max(vacc4x0123, vzero);
302 vacc5x0123 = wasm_i32x4_max(vacc5x0123, vzero);
303 vacc0x4567 = wasm_i32x4_max(vacc0x4567, vzero);
304 vacc1x4567 = wasm_i32x4_max(vacc1x4567, vzero);
305 vacc2x4567 = wasm_i32x4_max(vacc2x4567, vzero);
306 vacc3x4567 = wasm_i32x4_max(vacc3x4567, vzero);
307 vacc4x4567 = wasm_i32x4_max(vacc4x4567, vzero);
308 vacc5x4567 = wasm_i32x4_max(vacc5x4567, vzero);
309
310 if XNN_LIKELY(nc >= 8) {
311 wasm_v128_store(c5, vacc5x0123);
312 wasm_v128_store(c5 + 4, vacc5x4567);
313 c5 = (float*) ((uintptr_t) c5 + cn_stride);
314 wasm_v128_store(c4, vacc4x0123);
315 wasm_v128_store(c4 + 4, vacc4x4567);
316 c4 = (float*) ((uintptr_t) c4 + cn_stride);
317 wasm_v128_store(c3, vacc3x0123);
318 wasm_v128_store(c3 + 4, vacc3x4567);
319 c3 = (float*) ((uintptr_t) c3 + cn_stride);
320 wasm_v128_store(c2, vacc2x0123);
321 wasm_v128_store(c2 + 4, vacc2x4567);
322 c2 = (float*) ((uintptr_t) c2 + cn_stride);
323 wasm_v128_store(c1, vacc1x0123);
324 wasm_v128_store(c1 + 4, vacc1x4567);
325 c1 = (float*) ((uintptr_t) c1 + cn_stride);
326 wasm_v128_store(c0, vacc0x0123);
327 wasm_v128_store(c0 + 4, vacc0x4567);
328 c0 = (float*) ((uintptr_t) c0 + cn_stride);
329
330 a5 = (const float*) ((uintptr_t) a5 - kc);
331 a4 = (const float*) ((uintptr_t) a4 - kc);
332 a3 = (const float*) ((uintptr_t) a3 - kc);
333 a2 = (const float*) ((uintptr_t) a2 - kc);
334 a1 = (const float*) ((uintptr_t) a1 - kc);
335 a0 = (const float*) ((uintptr_t) a0 - kc);
336
337 nc -= 8;
338 } else {
339 if (nc & 4) {
340 wasm_v128_store(c5, vacc5x0123);
341 wasm_v128_store(c4, vacc4x0123);
342 wasm_v128_store(c3, vacc3x0123);
343 wasm_v128_store(c2, vacc2x0123);
344 wasm_v128_store(c1, vacc1x0123);
345 wasm_v128_store(c0, vacc0x0123);
346
347 vacc5x0123 = vacc5x4567;
348 vacc4x0123 = vacc4x4567;
349 vacc3x0123 = vacc3x4567;
350 vacc2x0123 = vacc2x4567;
351 vacc1x0123 = vacc1x4567;
352 vacc0x0123 = vacc0x4567;
353
354 c5 += 4;
355 c4 += 4;
356 c3 += 4;
357 c2 += 4;
358 c1 += 4;
359 c0 += 4;
360 }
361 if (nc & 2) {
362 *((double*) c5) = wasm_f64x2_extract_lane(vacc5x0123, 0);
363 *((double*) c4) = wasm_f64x2_extract_lane(vacc4x0123, 0);
364 *((double*) c3) = wasm_f64x2_extract_lane(vacc3x0123, 0);
365 *((double*) c2) = wasm_f64x2_extract_lane(vacc2x0123, 0);
366 *((double*) c1) = wasm_f64x2_extract_lane(vacc1x0123, 0);
367 *((double*) c0) = wasm_f64x2_extract_lane(vacc0x0123, 0);
368
369 vacc5x0123 = wasm_v32x4_shuffle(vacc5x0123, vacc5x0123, 2, 3, 2, 3);
370 vacc4x0123 = wasm_v32x4_shuffle(vacc4x0123, vacc4x0123, 2, 3, 2, 3);
371 vacc3x0123 = wasm_v32x4_shuffle(vacc3x0123, vacc3x0123, 2, 3, 2, 3);
372 vacc2x0123 = wasm_v32x4_shuffle(vacc2x0123, vacc2x0123, 2, 3, 2, 3);
373 vacc1x0123 = wasm_v32x4_shuffle(vacc1x0123, vacc1x0123, 2, 3, 2, 3);
374 vacc0x0123 = wasm_v32x4_shuffle(vacc0x0123, vacc0x0123, 2, 3, 2, 3);
375
376 c5 += 2;
377 c4 += 2;
378 c3 += 2;
379 c2 += 2;
380 c1 += 2;
381 c0 += 2;
382 }
383 if (nc & 1) {
384 *c5 = wasm_f32x4_extract_lane(vacc5x0123, 0);
385 *c4 = wasm_f32x4_extract_lane(vacc4x0123, 0);
386 *c3 = wasm_f32x4_extract_lane(vacc3x0123, 0);
387 *c2 = wasm_f32x4_extract_lane(vacc2x0123, 0);
388 *c1 = wasm_f32x4_extract_lane(vacc1x0123, 0);
389 *c0 = wasm_f32x4_extract_lane(vacc0x0123, 0);
390 }
391
392 nc = 0;
393 }
394 } while (nc != 0);
395 }
396