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_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 union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_f32_gemm_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 union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
28 {
29 assert(mr != 0);
30 assert(mr <= 5);
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
65 const v128_t vmin = wasm_v128_load64_splat(params->wasmsimd.min);
66 const v128_t vmax = wasm_v128_load64_splat(params->wasmsimd.max);
67 do {
68 v128_t vacc0x0123 = wasm_v128_load(w + 0);
69 v128_t vacc0x4567 = wasm_v128_load(w + 4);
70 v128_t vacc1x0123 = vacc0x0123;
71 v128_t vacc1x4567 = vacc0x4567;
72 v128_t vacc2x0123 = vacc0x0123;
73 v128_t vacc2x4567 = vacc0x4567;
74 v128_t vacc3x0123 = vacc0x0123;
75 v128_t vacc3x4567 = vacc0x4567;
76 v128_t vacc4x0123 = vacc0x0123;
77 v128_t vacc4x4567 = vacc0x4567;
78 w += 8;
79
80 size_t k = kc;
81 while (k >= 4 * sizeof(float)) {
82 v128_t va0 = wasm_v128_load(a0);
83 a0 += 4;
84 v128_t va1 = wasm_v128_load(a1);
85 a1 += 4;
86 v128_t va2 = wasm_v128_load(a2);
87 a2 += 4;
88 v128_t va3 = wasm_v128_load(a3);
89 a3 += 4;
90 v128_t va4 = wasm_v128_load(a4);
91 a4 += 4;
92
93
94 const v128_t vb0123c0 = wasm_v128_load(w + 0);
95 const v128_t vb4567c0 = wasm_v128_load(w + 4);
96
97 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c0));
98 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123c0));
99 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123c0));
100 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3, vb0123c0));
101 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123c0));
102 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c0));
103 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1, vb4567c0));
104 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2, vb4567c0));
105 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3, vb4567c0));
106 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4, vb4567c0));
107
108 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
109 va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
110 va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
111 va3 = wasm_v32x4_shuffle(va3, va3, 1, 2, 3, 0);
112 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0);
113
114 const v128_t vb0123c1 = wasm_v128_load(w + 8);
115 const v128_t vb4567c1 = wasm_v128_load(w + 12);
116
117 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c1));
118 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123c1));
119 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123c1));
120 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3, vb0123c1));
121 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123c1));
122 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c1));
123 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1, vb4567c1));
124 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2, vb4567c1));
125 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3, vb4567c1));
126 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4, vb4567c1));
127
128 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
129 va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
130 va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
131 va3 = wasm_v32x4_shuffle(va3, va3, 1, 2, 3, 0);
132 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0);
133
134 const v128_t vb0123c2 = wasm_v128_load(w + 16);
135 const v128_t vb4567c2 = wasm_v128_load(w + 20);
136
137 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c2));
138 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123c2));
139 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123c2));
140 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3, vb0123c2));
141 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123c2));
142 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c2));
143 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1, vb4567c2));
144 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2, vb4567c2));
145 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3, vb4567c2));
146 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4, vb4567c2));
147
148 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
149 va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
150 va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
151 va3 = wasm_v32x4_shuffle(va3, va3, 1, 2, 3, 0);
152 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0);
153
154 const v128_t vb0123c3 = wasm_v128_load(w + 24);
155 const v128_t vb4567c3 = wasm_v128_load(w + 28);
156
157 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c3));
158 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123c3));
159 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123c3));
160 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3, vb0123c3));
161 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123c3));
162 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c3));
163 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1, vb4567c3));
164 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2, vb4567c3));
165 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3, vb4567c3));
166 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4, vb4567c3));
167
168
169 w += 32;
170 k -= 4 * sizeof(float);
171 }
172 if XNN_UNLIKELY(k != 0) {
173 v128_t va0 = wasm_v128_load(a0);
174 a0 = (const float*) ((uintptr_t) a0 + k);
175 v128_t va1 = wasm_v128_load(a1);
176 a1 = (const float*) ((uintptr_t) a1 + k);
177 v128_t va2 = wasm_v128_load(a2);
178 a2 = (const float*) ((uintptr_t) a2 + k);
179 v128_t va3 = wasm_v128_load(a3);
180 a3 = (const float*) ((uintptr_t) a3 + k);
181 v128_t va4 = wasm_v128_load(a4);
182 a4 = (const float*) ((uintptr_t) a4 + k);
183
184 const v128_t vzero = wasm_f32x4_const_splat(0.0f);
185
186 const v128_t vb0123c0 = wasm_v128_load(w + 0);
187 const v128_t vb4567c0 = wasm_v128_load(w + 4);
188
189 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0));
190 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0));
191 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0));
192 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(wasm_v128_andnot(va3, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0));
193 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(wasm_v128_andnot(va4, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0));
194 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb4567c0, vzero)), vb4567c0));
195 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb4567c0, vzero)), vb4567c0));
196 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb4567c0, vzero)), vb4567c0));
197 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(wasm_v128_andnot(va3, wasm_f32x4_eq(vb4567c0, vzero)), vb4567c0));
198 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(wasm_v128_andnot(va4, wasm_f32x4_eq(vb4567c0, vzero)), vb4567c0));
199
200 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
201 va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
202 va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
203 va3 = wasm_v32x4_shuffle(va3, va3, 1, 2, 3, 0);
204 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0);
205
206 const v128_t vb0123c1 = wasm_v128_load(w + 8);
207 const v128_t vb4567c1 = wasm_v128_load(w + 12);
208
209 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c1, vzero)), vb0123c1));
210 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb0123c1, vzero)), vb0123c1));
211 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb0123c1, vzero)), vb0123c1));
212 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(wasm_v128_andnot(va3, wasm_f32x4_eq(vb0123c1, vzero)), vb0123c1));
213 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(wasm_v128_andnot(va4, wasm_f32x4_eq(vb0123c1, vzero)), vb0123c1));
214 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb4567c1, vzero)), vb4567c1));
215 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb4567c1, vzero)), vb4567c1));
216 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb4567c1, vzero)), vb4567c1));
217 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(wasm_v128_andnot(va3, wasm_f32x4_eq(vb4567c1, vzero)), vb4567c1));
218 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(wasm_v128_andnot(va4, wasm_f32x4_eq(vb4567c1, vzero)), vb4567c1));
219
220 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
221 va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
222 va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
223 va3 = wasm_v32x4_shuffle(va3, va3, 1, 2, 3, 0);
224 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0);
225
226 const v128_t vb0123c2 = wasm_v128_load(w + 16);
227 const v128_t vb4567c2 = wasm_v128_load(w + 20);
228
229 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c2, vzero)), vb0123c2));
230 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb0123c2, vzero)), vb0123c2));
231 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb0123c2, vzero)), vb0123c2));
232 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(wasm_v128_andnot(va3, wasm_f32x4_eq(vb0123c2, vzero)), vb0123c2));
233 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(wasm_v128_andnot(va4, wasm_f32x4_eq(vb0123c2, vzero)), vb0123c2));
234 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb4567c2, vzero)), vb4567c2));
235 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb4567c2, vzero)), vb4567c2));
236 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb4567c2, vzero)), vb4567c2));
237 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(wasm_v128_andnot(va3, wasm_f32x4_eq(vb4567c2, vzero)), vb4567c2));
238 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(wasm_v128_andnot(va4, wasm_f32x4_eq(vb4567c2, vzero)), vb4567c2));
239
240 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
241 va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
242 va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
243 va3 = wasm_v32x4_shuffle(va3, va3, 1, 2, 3, 0);
244 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0);
245
246 const v128_t vb0123c3 = wasm_v128_load(w + 24);
247 const v128_t vb4567c3 = wasm_v128_load(w + 28);
248
249 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c3, vzero)), vb0123c3));
250 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb0123c3, vzero)), vb0123c3));
251 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb0123c3, vzero)), vb0123c3));
252 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(wasm_v128_andnot(va3, wasm_f32x4_eq(vb0123c3, vzero)), vb0123c3));
253 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(wasm_v128_andnot(va4, wasm_f32x4_eq(vb0123c3, vzero)), vb0123c3));
254 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb4567c3, vzero)), vb4567c3));
255 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb4567c3, vzero)), vb4567c3));
256 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb4567c3, vzero)), vb4567c3));
257 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(wasm_v128_andnot(va3, wasm_f32x4_eq(vb4567c3, vzero)), vb4567c3));
258 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(wasm_v128_andnot(va4, wasm_f32x4_eq(vb4567c3, vzero)), vb4567c3));
259
260
261 w += 32;
262 }
263
264 vacc0x0123 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc0x0123);
265 vacc1x0123 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc1x0123);
266 vacc2x0123 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc2x0123);
267 vacc3x0123 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc3x0123);
268 vacc4x0123 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc4x0123);
269 vacc0x4567 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc0x4567);
270 vacc1x4567 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc1x4567);
271 vacc2x4567 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc2x4567);
272 vacc3x4567 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc3x4567);
273 vacc4x4567 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc4x4567);
274
275 vacc0x0123 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc0x0123);
276 vacc1x0123 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc1x0123);
277 vacc2x0123 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc2x0123);
278 vacc3x0123 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc3x0123);
279 vacc4x0123 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc4x0123);
280 vacc0x4567 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc0x4567);
281 vacc1x4567 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc1x4567);
282 vacc2x4567 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc2x4567);
283 vacc3x4567 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc3x4567);
284 vacc4x4567 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc4x4567);
285
286 if XNN_LIKELY(nc >= 8) {
287 wasm_v128_store(c4, vacc4x0123);
288 wasm_v128_store(c4 + 4, vacc4x4567);
289 c4 = (float*) ((uintptr_t) c4 + cn_stride);
290 wasm_v128_store(c3, vacc3x0123);
291 wasm_v128_store(c3 + 4, vacc3x4567);
292 c3 = (float*) ((uintptr_t) c3 + cn_stride);
293 wasm_v128_store(c2, vacc2x0123);
294 wasm_v128_store(c2 + 4, vacc2x4567);
295 c2 = (float*) ((uintptr_t) c2 + cn_stride);
296 wasm_v128_store(c1, vacc1x0123);
297 wasm_v128_store(c1 + 4, vacc1x4567);
298 c1 = (float*) ((uintptr_t) c1 + cn_stride);
299 wasm_v128_store(c0, vacc0x0123);
300 wasm_v128_store(c0 + 4, vacc0x4567);
301 c0 = (float*) ((uintptr_t) c0 + cn_stride);
302
303 a4 = (const float*) ((uintptr_t) a4 - kc);
304 a3 = (const float*) ((uintptr_t) a3 - kc);
305 a2 = (const float*) ((uintptr_t) a2 - kc);
306 a1 = (const float*) ((uintptr_t) a1 - kc);
307 a0 = (const float*) ((uintptr_t) a0 - kc);
308
309 nc -= 8;
310 } else {
311 if (nc & 4) {
312 wasm_v128_store(c4, vacc4x0123);
313 wasm_v128_store(c3, vacc3x0123);
314 wasm_v128_store(c2, vacc2x0123);
315 wasm_v128_store(c1, vacc1x0123);
316 wasm_v128_store(c0, vacc0x0123);
317
318 vacc4x0123 = vacc4x4567;
319 vacc3x0123 = vacc3x4567;
320 vacc2x0123 = vacc2x4567;
321 vacc1x0123 = vacc1x4567;
322 vacc0x0123 = vacc0x4567;
323
324 c4 += 4;
325 c3 += 4;
326 c2 += 4;
327 c1 += 4;
328 c0 += 4;
329 }
330 if (nc & 2) {
331 *((double*) c4) = wasm_f64x2_extract_lane(vacc4x0123, 0);
332 *((double*) c3) = wasm_f64x2_extract_lane(vacc3x0123, 0);
333 *((double*) c2) = wasm_f64x2_extract_lane(vacc2x0123, 0);
334 *((double*) c1) = wasm_f64x2_extract_lane(vacc1x0123, 0);
335 *((double*) c0) = wasm_f64x2_extract_lane(vacc0x0123, 0);
336
337 vacc4x0123 = wasm_v32x4_shuffle(vacc4x0123, vacc4x0123, 2, 3, 2, 3);
338 vacc3x0123 = wasm_v32x4_shuffle(vacc3x0123, vacc3x0123, 2, 3, 2, 3);
339 vacc2x0123 = wasm_v32x4_shuffle(vacc2x0123, vacc2x0123, 2, 3, 2, 3);
340 vacc1x0123 = wasm_v32x4_shuffle(vacc1x0123, vacc1x0123, 2, 3, 2, 3);
341 vacc0x0123 = wasm_v32x4_shuffle(vacc0x0123, vacc0x0123, 2, 3, 2, 3);
342
343 c4 += 2;
344 c3 += 2;
345 c2 += 2;
346 c1 += 2;
347 c0 += 2;
348 }
349 if (nc & 1) {
350 *c4 = wasm_f32x4_extract_lane(vacc4x0123, 0);
351 *c3 = wasm_f32x4_extract_lane(vacc3x0123, 0);
352 *c2 = wasm_f32x4_extract_lane(vacc2x0123, 0);
353 *c1 = wasm_f32x4_extract_lane(vacc1x0123, 0);
354 *c0 = wasm_f32x4_extract_lane(vacc0x0123, 0);
355 }
356
357 nc = 0;
358 }
359 } while (nc != 0);
360 }
361