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_gemminc_minmax_ukernel_5x8s4__wasmrelaxedsimd(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 float * restrict acc,const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmrelaxedsimd(
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 float*restrict acc,
28 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
29 {
30 assert(mr != 0);
31 assert(mr <= 5);
32 assert(nc != 0);
33 assert(kc != 0);
34 assert(kc % sizeof(float) == 0);
35 assert(a != NULL);
36 assert(w != NULL);
37 assert(c != NULL);
38 assert(acc != NULL);
39
40 const float* a0 = a;
41 float* c0 = c;
42 const float* a1 = (const float*) ((uintptr_t) a0 + a_stride);
43 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
44 if XNN_UNPREDICTABLE(mr < 2) {
45 a1 = a0;
46 c1 = c0;
47 }
48 const float* a2 = (const float*) ((uintptr_t) a1 + a_stride);
49 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
50 if XNN_UNPREDICTABLE(mr <= 2) {
51 a2 = a1;
52 c2 = c1;
53 }
54 const float* a3 = (const float*) ((uintptr_t) a2 + a_stride);
55 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
56 if XNN_UNPREDICTABLE(mr < 4) {
57 a3 = a2;
58 c3 = c2;
59 }
60 const float* a4 = (const float*) ((uintptr_t) a3 + a_stride);
61 float* c4 = (float*) ((uintptr_t) c3 + cm_stride);
62 if XNN_UNPREDICTABLE(mr <= 4) {
63 a4 = a3;
64 c4 = c3;
65 }
66
67 const v128_t vmin = wasm_v128_load64_splat(params->wasmsimd.min);
68 const v128_t vmax = wasm_v128_load64_splat(params->wasmsimd.max);
69 do {
70 v128_t vacc0x0123 = wasm_v128_load(acc + 0);
71 v128_t vacc0x4567 = wasm_v128_load(acc + 4);
72 v128_t vacc1x0123 = wasm_v128_load(acc + 8);
73 v128_t vacc1x4567 = wasm_v128_load(acc + 12);
74 v128_t vacc2x0123 = wasm_v128_load(acc + 16);
75 v128_t vacc2x4567 = wasm_v128_load(acc + 20);
76 v128_t vacc3x0123 = wasm_v128_load(acc + 24);
77 v128_t vacc3x4567 = wasm_v128_load(acc + 28);
78 v128_t vacc4x0123 = wasm_v128_load(acc + 32);
79 v128_t vacc4x4567 = wasm_v128_load(acc + 36);
80 acc += 40;
81
82 size_t k = kc;
83 while (k >= 4 * sizeof(float)) {
84 v128_t va0 = wasm_v128_load(a0);
85 a0 += 4;
86 v128_t va1 = wasm_v128_load(a1);
87 a1 += 4;
88 v128_t va2 = wasm_v128_load(a2);
89 a2 += 4;
90 v128_t va3 = wasm_v128_load(a3);
91 a3 += 4;
92 v128_t va4 = wasm_v128_load(a4);
93 a4 += 4;
94
95
96 const v128_t vb0123c0 = wasm_v128_load(w + 0);
97 const v128_t vb4567c0 = wasm_v128_load(w + 4);
98
99 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c0));
100 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123c0));
101 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123c0));
102 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3, vb0123c0));
103 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123c0));
104 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c0));
105 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1, vb4567c0));
106 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2, vb4567c0));
107 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3, vb4567c0));
108 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4, vb4567c0));
109
110 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
111 va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
112 va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
113 va3 = wasm_v32x4_shuffle(va3, va3, 1, 2, 3, 0);
114 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0);
115
116 const v128_t vb0123c1 = wasm_v128_load(w + 8);
117 const v128_t vb4567c1 = wasm_v128_load(w + 12);
118
119 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c1));
120 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123c1));
121 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123c1));
122 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3, vb0123c1));
123 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123c1));
124 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c1));
125 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1, vb4567c1));
126 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2, vb4567c1));
127 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3, vb4567c1));
128 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4, vb4567c1));
129
130 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
131 va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
132 va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
133 va3 = wasm_v32x4_shuffle(va3, va3, 1, 2, 3, 0);
134 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0);
135
136 const v128_t vb0123c2 = wasm_v128_load(w + 16);
137 const v128_t vb4567c2 = wasm_v128_load(w + 20);
138
139 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c2));
140 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123c2));
141 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123c2));
142 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3, vb0123c2));
143 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123c2));
144 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c2));
145 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1, vb4567c2));
146 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2, vb4567c2));
147 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3, vb4567c2));
148 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4, vb4567c2));
149
150 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
151 va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
152 va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
153 va3 = wasm_v32x4_shuffle(va3, va3, 1, 2, 3, 0);
154 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0);
155
156 const v128_t vb0123c3 = wasm_v128_load(w + 24);
157 const v128_t vb4567c3 = wasm_v128_load(w + 28);
158
159 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c3));
160 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123c3));
161 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123c3));
162 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3, vb0123c3));
163 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123c3));
164 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c3));
165 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1, vb4567c3));
166 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2, vb4567c3));
167 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3, vb4567c3));
168 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4, vb4567c3));
169
170
171 w += 32;
172 k -= 4 * sizeof(float);
173 }
174 if XNN_UNLIKELY(k != 0) {
175 v128_t va0 = wasm_v128_load(a0);
176 a0 = (const float*) ((uintptr_t) a0 + k);
177 v128_t va1 = wasm_v128_load(a1);
178 a1 = (const float*) ((uintptr_t) a1 + k);
179 v128_t va2 = wasm_v128_load(a2);
180 a2 = (const float*) ((uintptr_t) a2 + k);
181 v128_t va3 = wasm_v128_load(a3);
182 a3 = (const float*) ((uintptr_t) a3 + k);
183 v128_t va4 = wasm_v128_load(a4);
184 a4 = (const float*) ((uintptr_t) a4 + k);
185
186 const v128_t vzero = wasm_f32x4_const_splat(0.0f);
187
188 const v128_t vb0123c0 = wasm_v128_load(w + 0);
189 const v128_t vb4567c0 = wasm_v128_load(w + 4);
190
191 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0));
192 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0));
193 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0));
194 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(wasm_v128_andnot(va3, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0));
195 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(wasm_v128_andnot(va4, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0));
196 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb4567c0, vzero)), vb4567c0));
197 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb4567c0, vzero)), vb4567c0));
198 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb4567c0, vzero)), vb4567c0));
199 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(wasm_v128_andnot(va3, wasm_f32x4_eq(vb4567c0, vzero)), vb4567c0));
200 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(wasm_v128_andnot(va4, wasm_f32x4_eq(vb4567c0, vzero)), vb4567c0));
201
202 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
203 va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
204 va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
205 va3 = wasm_v32x4_shuffle(va3, va3, 1, 2, 3, 0);
206 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0);
207
208 const v128_t vb0123c1 = wasm_v128_load(w + 8);
209 const v128_t vb4567c1 = wasm_v128_load(w + 12);
210
211 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c1, vzero)), vb0123c1));
212 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb0123c1, vzero)), vb0123c1));
213 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb0123c1, vzero)), vb0123c1));
214 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(wasm_v128_andnot(va3, wasm_f32x4_eq(vb0123c1, vzero)), vb0123c1));
215 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(wasm_v128_andnot(va4, wasm_f32x4_eq(vb0123c1, vzero)), vb0123c1));
216 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb4567c1, vzero)), vb4567c1));
217 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb4567c1, vzero)), vb4567c1));
218 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb4567c1, vzero)), vb4567c1));
219 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(wasm_v128_andnot(va3, wasm_f32x4_eq(vb4567c1, vzero)), vb4567c1));
220 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(wasm_v128_andnot(va4, wasm_f32x4_eq(vb4567c1, vzero)), vb4567c1));
221
222 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
223 va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
224 va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
225 va3 = wasm_v32x4_shuffle(va3, va3, 1, 2, 3, 0);
226 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0);
227
228 const v128_t vb0123c2 = wasm_v128_load(w + 16);
229 const v128_t vb4567c2 = wasm_v128_load(w + 20);
230
231 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c2, vzero)), vb0123c2));
232 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb0123c2, vzero)), vb0123c2));
233 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb0123c2, vzero)), vb0123c2));
234 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(wasm_v128_andnot(va3, wasm_f32x4_eq(vb0123c2, vzero)), vb0123c2));
235 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(wasm_v128_andnot(va4, wasm_f32x4_eq(vb0123c2, vzero)), vb0123c2));
236 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb4567c2, vzero)), vb4567c2));
237 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb4567c2, vzero)), vb4567c2));
238 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb4567c2, vzero)), vb4567c2));
239 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(wasm_v128_andnot(va3, wasm_f32x4_eq(vb4567c2, vzero)), vb4567c2));
240 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(wasm_v128_andnot(va4, wasm_f32x4_eq(vb4567c2, vzero)), vb4567c2));
241
242 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
243 va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
244 va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
245 va3 = wasm_v32x4_shuffle(va3, va3, 1, 2, 3, 0);
246 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0);
247
248 const v128_t vb0123c3 = wasm_v128_load(w + 24);
249 const v128_t vb4567c3 = wasm_v128_load(w + 28);
250
251 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c3, vzero)), vb0123c3));
252 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb0123c3, vzero)), vb0123c3));
253 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb0123c3, vzero)), vb0123c3));
254 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(wasm_v128_andnot(va3, wasm_f32x4_eq(vb0123c3, vzero)), vb0123c3));
255 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(wasm_v128_andnot(va4, wasm_f32x4_eq(vb0123c3, vzero)), vb0123c3));
256 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb4567c3, vzero)), vb4567c3));
257 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb4567c3, vzero)), vb4567c3));
258 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb4567c3, vzero)), vb4567c3));
259 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(wasm_v128_andnot(va3, wasm_f32x4_eq(vb4567c3, vzero)), vb4567c3));
260 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(wasm_v128_andnot(va4, wasm_f32x4_eq(vb4567c3, vzero)), vb4567c3));
261
262
263 w += 32;
264 }
265
266 vacc0x0123 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc0x0123);
267 vacc1x0123 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc1x0123);
268 vacc2x0123 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc2x0123);
269 vacc3x0123 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc3x0123);
270 vacc4x0123 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc4x0123);
271 vacc0x4567 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc0x4567);
272 vacc1x4567 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc1x4567);
273 vacc2x4567 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc2x4567);
274 vacc3x4567 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc3x4567);
275 vacc4x4567 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc4x4567);
276
277 vacc0x0123 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc0x0123);
278 vacc1x0123 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc1x0123);
279 vacc2x0123 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc2x0123);
280 vacc3x0123 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc3x0123);
281 vacc4x0123 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc4x0123);
282 vacc0x4567 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc0x4567);
283 vacc1x4567 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc1x4567);
284 vacc2x4567 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc2x4567);
285 vacc3x4567 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc3x4567);
286 vacc4x4567 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc4x4567);
287
288 if XNN_LIKELY(nc >= 8) {
289 wasm_v128_store(c4, vacc4x0123);
290 wasm_v128_store(c4 + 4, vacc4x4567);
291 c4 = (float*) ((uintptr_t) c4 + cn_stride);
292 wasm_v128_store(c3, vacc3x0123);
293 wasm_v128_store(c3 + 4, vacc3x4567);
294 c3 = (float*) ((uintptr_t) c3 + cn_stride);
295 wasm_v128_store(c2, vacc2x0123);
296 wasm_v128_store(c2 + 4, vacc2x4567);
297 c2 = (float*) ((uintptr_t) c2 + cn_stride);
298 wasm_v128_store(c1, vacc1x0123);
299 wasm_v128_store(c1 + 4, vacc1x4567);
300 c1 = (float*) ((uintptr_t) c1 + cn_stride);
301 wasm_v128_store(c0, vacc0x0123);
302 wasm_v128_store(c0 + 4, vacc0x4567);
303 c0 = (float*) ((uintptr_t) c0 + cn_stride);
304
305 a4 = (const float*) ((uintptr_t) a4 - kc);
306 a3 = (const float*) ((uintptr_t) a3 - kc);
307 a2 = (const float*) ((uintptr_t) a2 - kc);
308 a1 = (const float*) ((uintptr_t) a1 - kc);
309 a0 = (const float*) ((uintptr_t) a0 - kc);
310
311 nc -= 8;
312 } else {
313 if (nc & 4) {
314 wasm_v128_store(c4, vacc4x0123);
315 wasm_v128_store(c3, vacc3x0123);
316 wasm_v128_store(c2, vacc2x0123);
317 wasm_v128_store(c1, vacc1x0123);
318 wasm_v128_store(c0, vacc0x0123);
319
320 vacc4x0123 = vacc4x4567;
321 vacc3x0123 = vacc3x4567;
322 vacc2x0123 = vacc2x4567;
323 vacc1x0123 = vacc1x4567;
324 vacc0x0123 = vacc0x4567;
325
326 c4 += 4;
327 c3 += 4;
328 c2 += 4;
329 c1 += 4;
330 c0 += 4;
331 }
332 if (nc & 2) {
333 *((double*) c4) = wasm_f64x2_extract_lane(vacc4x0123, 0);
334 *((double*) c3) = wasm_f64x2_extract_lane(vacc3x0123, 0);
335 *((double*) c2) = wasm_f64x2_extract_lane(vacc2x0123, 0);
336 *((double*) c1) = wasm_f64x2_extract_lane(vacc1x0123, 0);
337 *((double*) c0) = wasm_f64x2_extract_lane(vacc0x0123, 0);
338
339 vacc4x0123 = wasm_v32x4_shuffle(vacc4x0123, vacc4x0123, 2, 3, 2, 3);
340 vacc3x0123 = wasm_v32x4_shuffle(vacc3x0123, vacc3x0123, 2, 3, 2, 3);
341 vacc2x0123 = wasm_v32x4_shuffle(vacc2x0123, vacc2x0123, 2, 3, 2, 3);
342 vacc1x0123 = wasm_v32x4_shuffle(vacc1x0123, vacc1x0123, 2, 3, 2, 3);
343 vacc0x0123 = wasm_v32x4_shuffle(vacc0x0123, vacc0x0123, 2, 3, 2, 3);
344
345 c4 += 2;
346 c3 += 2;
347 c2 += 2;
348 c1 += 2;
349 c0 += 2;
350 }
351 if (nc & 1) {
352 *c4 = wasm_f32x4_extract_lane(vacc4x0123, 0);
353 *c3 = wasm_f32x4_extract_lane(vacc3x0123, 0);
354 *c2 = wasm_f32x4_extract_lane(vacc2x0123, 0);
355 *c1 = wasm_f32x4_extract_lane(vacc1x0123, 0);
356 *c0 = wasm_f32x4_extract_lane(vacc0x0123, 0);
357 }
358
359 nc = 0;
360 }
361 } while (nc != 0);
362 }
363