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_minmax_ukernel_5x8s4__wasmrelaxedsimd_fma(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_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_f32_igemm_minmax_ukernel_5x8s4__wasmrelaxedsimd_fma(
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_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
30 {
31 assert(mr != 0);
32 assert(mr <= 5);
33 assert(nc != 0);
34 assert(kc != 0);
35 assert(kc % sizeof(float) == 0);
36 assert(ks != 0);
37 assert(ks % (5 * 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
61 const v128_t vmin = wasm_v128_load64_splat(params->wasmsimd.min);
62 const v128_t vmax = wasm_v128_load64_splat(params->wasmsimd.max);
63 do {
64 v128_t vacc0x0123 = wasm_v128_load(w);
65 v128_t vacc0x4567 = wasm_v128_load(w + 4);
66 v128_t vacc1x0123 = vacc0x0123;
67 v128_t vacc1x4567 = vacc0x4567;
68 v128_t vacc2x0123 = vacc0x0123;
69 v128_t vacc2x4567 = vacc0x4567;
70 v128_t vacc3x0123 = vacc0x0123;
71 v128_t vacc3x4567 = vacc0x4567;
72 v128_t vacc4x0123 = vacc0x0123;
73 v128_t vacc4x4567 = vacc0x4567;
74 w += 8;
75
76 size_t p = ks;
77 do {
78 const float* restrict a0 = a[0];
79 assert(a0 != NULL);
80 if XNN_UNPREDICTABLE(a0 != zero) {
81 a0 = (const float*) ((uintptr_t) a0 + a_offset);
82 }
83 const float* restrict a1 = a[1];
84 assert(a1 != NULL);
85 if XNN_UNPREDICTABLE(a1 != zero) {
86 a1 = (const float*) ((uintptr_t) a1 + a_offset);
87 }
88 const float* restrict a2 = a[2];
89 assert(a2 != NULL);
90 if XNN_UNPREDICTABLE(a2 != zero) {
91 a2 = (const float*) ((uintptr_t) a2 + a_offset);
92 }
93 const float* restrict a3 = a[3];
94 assert(a3 != NULL);
95 if XNN_UNPREDICTABLE(a3 != zero) {
96 a3 = (const float*) ((uintptr_t) a3 + a_offset);
97 }
98 const float* restrict a4 = a[4];
99 assert(a4 != NULL);
100 if XNN_UNPREDICTABLE(a4 != zero) {
101 a4 = (const float*) ((uintptr_t) a4 + a_offset);
102 }
103 a += 5;
104
105 size_t k = kc;
106 while (k >= 4 * sizeof(float)) {
107 v128_t va0 = wasm_v128_load(a0);
108 a0 += 4;
109 v128_t va1 = wasm_v128_load(a1);
110 a1 += 4;
111 v128_t va2 = wasm_v128_load(a2);
112 a2 += 4;
113 v128_t va3 = wasm_v128_load(a3);
114 a3 += 4;
115 v128_t va4 = wasm_v128_load(a4);
116 a4 += 4;
117
118
119 const v128_t vb0123c0 = wasm_v128_load(w + 0);
120 const v128_t vb4567c0 = wasm_v128_load(w + 4);
121
122 vacc0x0123 = __builtin_wasm_fma_f32x4(vacc0x0123, va0, vb0123c0);
123 vacc1x0123 = __builtin_wasm_fma_f32x4(vacc1x0123, va1, vb0123c0);
124 vacc2x0123 = __builtin_wasm_fma_f32x4(vacc2x0123, va2, vb0123c0);
125 vacc3x0123 = __builtin_wasm_fma_f32x4(vacc3x0123, va3, vb0123c0);
126 vacc4x0123 = __builtin_wasm_fma_f32x4(vacc4x0123, va4, vb0123c0);
127 vacc0x4567 = __builtin_wasm_fma_f32x4(vacc0x4567, va0, vb4567c0);
128 vacc1x4567 = __builtin_wasm_fma_f32x4(vacc1x4567, va1, vb4567c0);
129 vacc2x4567 = __builtin_wasm_fma_f32x4(vacc2x4567, va2, vb4567c0);
130 vacc3x4567 = __builtin_wasm_fma_f32x4(vacc3x4567, va3, vb4567c0);
131 vacc4x4567 = __builtin_wasm_fma_f32x4(vacc4x4567, va4, vb4567c0);
132
133 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
134 va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
135 va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
136 va3 = wasm_v32x4_shuffle(va3, va3, 1, 2, 3, 0);
137 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0);
138
139 const v128_t vb0123c1 = wasm_v128_load(w + 8);
140 const v128_t vb4567c1 = wasm_v128_load(w + 12);
141
142 vacc0x0123 = __builtin_wasm_fma_f32x4(vacc0x0123, va0, vb0123c1);
143 vacc1x0123 = __builtin_wasm_fma_f32x4(vacc1x0123, va1, vb0123c1);
144 vacc2x0123 = __builtin_wasm_fma_f32x4(vacc2x0123, va2, vb0123c1);
145 vacc3x0123 = __builtin_wasm_fma_f32x4(vacc3x0123, va3, vb0123c1);
146 vacc4x0123 = __builtin_wasm_fma_f32x4(vacc4x0123, va4, vb0123c1);
147 vacc0x4567 = __builtin_wasm_fma_f32x4(vacc0x4567, va0, vb4567c1);
148 vacc1x4567 = __builtin_wasm_fma_f32x4(vacc1x4567, va1, vb4567c1);
149 vacc2x4567 = __builtin_wasm_fma_f32x4(vacc2x4567, va2, vb4567c1);
150 vacc3x4567 = __builtin_wasm_fma_f32x4(vacc3x4567, va3, vb4567c1);
151 vacc4x4567 = __builtin_wasm_fma_f32x4(vacc4x4567, va4, vb4567c1);
152
153 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
154 va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
155 va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
156 va3 = wasm_v32x4_shuffle(va3, va3, 1, 2, 3, 0);
157 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0);
158
159 const v128_t vb0123c2 = wasm_v128_load(w + 16);
160 const v128_t vb4567c2 = wasm_v128_load(w + 20);
161
162 vacc0x0123 = __builtin_wasm_fma_f32x4(vacc0x0123, va0, vb0123c2);
163 vacc1x0123 = __builtin_wasm_fma_f32x4(vacc1x0123, va1, vb0123c2);
164 vacc2x0123 = __builtin_wasm_fma_f32x4(vacc2x0123, va2, vb0123c2);
165 vacc3x0123 = __builtin_wasm_fma_f32x4(vacc3x0123, va3, vb0123c2);
166 vacc4x0123 = __builtin_wasm_fma_f32x4(vacc4x0123, va4, vb0123c2);
167 vacc0x4567 = __builtin_wasm_fma_f32x4(vacc0x4567, va0, vb4567c2);
168 vacc1x4567 = __builtin_wasm_fma_f32x4(vacc1x4567, va1, vb4567c2);
169 vacc2x4567 = __builtin_wasm_fma_f32x4(vacc2x4567, va2, vb4567c2);
170 vacc3x4567 = __builtin_wasm_fma_f32x4(vacc3x4567, va3, vb4567c2);
171 vacc4x4567 = __builtin_wasm_fma_f32x4(vacc4x4567, va4, vb4567c2);
172
173 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
174 va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
175 va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
176 va3 = wasm_v32x4_shuffle(va3, va3, 1, 2, 3, 0);
177 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0);
178
179 const v128_t vb0123c3 = wasm_v128_load(w + 24);
180 const v128_t vb4567c3 = wasm_v128_load(w + 28);
181
182 vacc0x0123 = __builtin_wasm_fma_f32x4(vacc0x0123, va0, vb0123c3);
183 vacc1x0123 = __builtin_wasm_fma_f32x4(vacc1x0123, va1, vb0123c3);
184 vacc2x0123 = __builtin_wasm_fma_f32x4(vacc2x0123, va2, vb0123c3);
185 vacc3x0123 = __builtin_wasm_fma_f32x4(vacc3x0123, va3, vb0123c3);
186 vacc4x0123 = __builtin_wasm_fma_f32x4(vacc4x0123, va4, vb0123c3);
187 vacc0x4567 = __builtin_wasm_fma_f32x4(vacc0x4567, va0, vb4567c3);
188 vacc1x4567 = __builtin_wasm_fma_f32x4(vacc1x4567, va1, vb4567c3);
189 vacc2x4567 = __builtin_wasm_fma_f32x4(vacc2x4567, va2, vb4567c3);
190 vacc3x4567 = __builtin_wasm_fma_f32x4(vacc3x4567, va3, vb4567c3);
191 vacc4x4567 = __builtin_wasm_fma_f32x4(vacc4x4567, va4, vb4567c3);
192
193
194 w += 32;
195 k -= 4 * sizeof(float);
196 }
197 if XNN_UNLIKELY(k != 0) {
198 v128_t va0 = wasm_v128_load(a0);
199 a0 = (const float*) ((uintptr_t) a0 + k);
200 v128_t va1 = wasm_v128_load(a1);
201 a1 = (const float*) ((uintptr_t) a1 + k);
202 v128_t va2 = wasm_v128_load(a2);
203 a2 = (const float*) ((uintptr_t) a2 + k);
204 v128_t va3 = wasm_v128_load(a3);
205 a3 = (const float*) ((uintptr_t) a3 + k);
206 v128_t va4 = wasm_v128_load(a4);
207 a4 = (const float*) ((uintptr_t) a4 + k);
208
209 const v128_t vzero = wasm_f32x4_const_splat(0.0f);
210
211 const v128_t vb0123c0 = wasm_v128_load(w + 0);
212 const v128_t vb4567c0 = wasm_v128_load(w + 4);
213
214 vacc0x0123 = __builtin_wasm_fma_f32x4(vacc0x0123, wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0);
215 vacc1x0123 = __builtin_wasm_fma_f32x4(vacc1x0123, wasm_v128_andnot(va1, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0);
216 vacc2x0123 = __builtin_wasm_fma_f32x4(vacc2x0123, wasm_v128_andnot(va2, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0);
217 vacc3x0123 = __builtin_wasm_fma_f32x4(vacc3x0123, wasm_v128_andnot(va3, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0);
218 vacc4x0123 = __builtin_wasm_fma_f32x4(vacc4x0123, wasm_v128_andnot(va4, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0);
219 vacc0x4567 = __builtin_wasm_fma_f32x4(vacc0x4567, wasm_v128_andnot(va0, wasm_f32x4_eq(vb4567c0, vzero)), vb4567c0);
220 vacc1x4567 = __builtin_wasm_fma_f32x4(vacc1x4567, wasm_v128_andnot(va1, wasm_f32x4_eq(vb4567c0, vzero)), vb4567c0);
221 vacc2x4567 = __builtin_wasm_fma_f32x4(vacc2x4567, wasm_v128_andnot(va2, wasm_f32x4_eq(vb4567c0, vzero)), vb4567c0);
222 vacc3x4567 = __builtin_wasm_fma_f32x4(vacc3x4567, wasm_v128_andnot(va3, wasm_f32x4_eq(vb4567c0, vzero)), vb4567c0);
223 vacc4x4567 = __builtin_wasm_fma_f32x4(vacc4x4567, wasm_v128_andnot(va4, wasm_f32x4_eq(vb4567c0, vzero)), vb4567c0);
224
225 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
226 va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
227 va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
228 va3 = wasm_v32x4_shuffle(va3, va3, 1, 2, 3, 0);
229 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0);
230
231 const v128_t vb0123c1 = wasm_v128_load(w + 8);
232 const v128_t vb4567c1 = wasm_v128_load(w + 12);
233
234 vacc0x0123 = __builtin_wasm_fma_f32x4(vacc0x0123, wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c1, vzero)), vb0123c1);
235 vacc1x0123 = __builtin_wasm_fma_f32x4(vacc1x0123, wasm_v128_andnot(va1, wasm_f32x4_eq(vb0123c1, vzero)), vb0123c1);
236 vacc2x0123 = __builtin_wasm_fma_f32x4(vacc2x0123, wasm_v128_andnot(va2, wasm_f32x4_eq(vb0123c1, vzero)), vb0123c1);
237 vacc3x0123 = __builtin_wasm_fma_f32x4(vacc3x0123, wasm_v128_andnot(va3, wasm_f32x4_eq(vb0123c1, vzero)), vb0123c1);
238 vacc4x0123 = __builtin_wasm_fma_f32x4(vacc4x0123, wasm_v128_andnot(va4, wasm_f32x4_eq(vb0123c1, vzero)), vb0123c1);
239 vacc0x4567 = __builtin_wasm_fma_f32x4(vacc0x4567, wasm_v128_andnot(va0, wasm_f32x4_eq(vb4567c1, vzero)), vb4567c1);
240 vacc1x4567 = __builtin_wasm_fma_f32x4(vacc1x4567, wasm_v128_andnot(va1, wasm_f32x4_eq(vb4567c1, vzero)), vb4567c1);
241 vacc2x4567 = __builtin_wasm_fma_f32x4(vacc2x4567, wasm_v128_andnot(va2, wasm_f32x4_eq(vb4567c1, vzero)), vb4567c1);
242 vacc3x4567 = __builtin_wasm_fma_f32x4(vacc3x4567, wasm_v128_andnot(va3, wasm_f32x4_eq(vb4567c1, vzero)), vb4567c1);
243 vacc4x4567 = __builtin_wasm_fma_f32x4(vacc4x4567, wasm_v128_andnot(va4, wasm_f32x4_eq(vb4567c1, vzero)), vb4567c1);
244
245 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
246 va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
247 va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
248 va3 = wasm_v32x4_shuffle(va3, va3, 1, 2, 3, 0);
249 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0);
250
251 const v128_t vb0123c2 = wasm_v128_load(w + 16);
252 const v128_t vb4567c2 = wasm_v128_load(w + 20);
253
254 vacc0x0123 = __builtin_wasm_fma_f32x4(vacc0x0123, wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c2, vzero)), vb0123c2);
255 vacc1x0123 = __builtin_wasm_fma_f32x4(vacc1x0123, wasm_v128_andnot(va1, wasm_f32x4_eq(vb0123c2, vzero)), vb0123c2);
256 vacc2x0123 = __builtin_wasm_fma_f32x4(vacc2x0123, wasm_v128_andnot(va2, wasm_f32x4_eq(vb0123c2, vzero)), vb0123c2);
257 vacc3x0123 = __builtin_wasm_fma_f32x4(vacc3x0123, wasm_v128_andnot(va3, wasm_f32x4_eq(vb0123c2, vzero)), vb0123c2);
258 vacc4x0123 = __builtin_wasm_fma_f32x4(vacc4x0123, wasm_v128_andnot(va4, wasm_f32x4_eq(vb0123c2, vzero)), vb0123c2);
259 vacc0x4567 = __builtin_wasm_fma_f32x4(vacc0x4567, wasm_v128_andnot(va0, wasm_f32x4_eq(vb4567c2, vzero)), vb4567c2);
260 vacc1x4567 = __builtin_wasm_fma_f32x4(vacc1x4567, wasm_v128_andnot(va1, wasm_f32x4_eq(vb4567c2, vzero)), vb4567c2);
261 vacc2x4567 = __builtin_wasm_fma_f32x4(vacc2x4567, wasm_v128_andnot(va2, wasm_f32x4_eq(vb4567c2, vzero)), vb4567c2);
262 vacc3x4567 = __builtin_wasm_fma_f32x4(vacc3x4567, wasm_v128_andnot(va3, wasm_f32x4_eq(vb4567c2, vzero)), vb4567c2);
263 vacc4x4567 = __builtin_wasm_fma_f32x4(vacc4x4567, wasm_v128_andnot(va4, wasm_f32x4_eq(vb4567c2, vzero)), vb4567c2);
264
265 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
266 va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
267 va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
268 va3 = wasm_v32x4_shuffle(va3, va3, 1, 2, 3, 0);
269 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0);
270
271 const v128_t vb0123c3 = wasm_v128_load(w + 24);
272 const v128_t vb4567c3 = wasm_v128_load(w + 28);
273
274 vacc0x0123 = __builtin_wasm_fma_f32x4(vacc0x0123, wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c3, vzero)), vb0123c3);
275 vacc1x0123 = __builtin_wasm_fma_f32x4(vacc1x0123, wasm_v128_andnot(va1, wasm_f32x4_eq(vb0123c3, vzero)), vb0123c3);
276 vacc2x0123 = __builtin_wasm_fma_f32x4(vacc2x0123, wasm_v128_andnot(va2, wasm_f32x4_eq(vb0123c3, vzero)), vb0123c3);
277 vacc3x0123 = __builtin_wasm_fma_f32x4(vacc3x0123, wasm_v128_andnot(va3, wasm_f32x4_eq(vb0123c3, vzero)), vb0123c3);
278 vacc4x0123 = __builtin_wasm_fma_f32x4(vacc4x0123, wasm_v128_andnot(va4, wasm_f32x4_eq(vb0123c3, vzero)), vb0123c3);
279 vacc0x4567 = __builtin_wasm_fma_f32x4(vacc0x4567, wasm_v128_andnot(va0, wasm_f32x4_eq(vb4567c3, vzero)), vb4567c3);
280 vacc1x4567 = __builtin_wasm_fma_f32x4(vacc1x4567, wasm_v128_andnot(va1, wasm_f32x4_eq(vb4567c3, vzero)), vb4567c3);
281 vacc2x4567 = __builtin_wasm_fma_f32x4(vacc2x4567, wasm_v128_andnot(va2, wasm_f32x4_eq(vb4567c3, vzero)), vb4567c3);
282 vacc3x4567 = __builtin_wasm_fma_f32x4(vacc3x4567, wasm_v128_andnot(va3, wasm_f32x4_eq(vb4567c3, vzero)), vb4567c3);
283 vacc4x4567 = __builtin_wasm_fma_f32x4(vacc4x4567, wasm_v128_andnot(va4, wasm_f32x4_eq(vb4567c3, vzero)), vb4567c3);
284
285
286 w += 32;
287 }
288 p -= 5 * sizeof(void*);
289 } while (p != 0);
290
291 vacc0x0123 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc0x0123);
292 vacc1x0123 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc1x0123);
293 vacc2x0123 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc2x0123);
294 vacc3x0123 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc3x0123);
295 vacc4x0123 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc4x0123);
296 vacc0x4567 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc0x4567);
297 vacc1x4567 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc1x4567);
298 vacc2x4567 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc2x4567);
299 vacc3x4567 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc3x4567);
300 vacc4x4567 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc4x4567);
301
302 vacc0x0123 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc0x0123);
303 vacc1x0123 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc1x0123);
304 vacc2x0123 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc2x0123);
305 vacc3x0123 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc3x0123);
306 vacc4x0123 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc4x0123);
307 vacc0x4567 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc0x4567);
308 vacc1x4567 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc1x4567);
309 vacc2x4567 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc2x4567);
310 vacc3x4567 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc3x4567);
311 vacc4x4567 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc4x4567);
312
313 if XNN_LIKELY(nc >= 8) {
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 a = (const float**restrict) ((uintptr_t) a - ks);
331 nc -= 8;
332 } else {
333 if (nc & 4) {
334 wasm_v128_store(c4, vacc4x0123);
335 wasm_v128_store(c3, vacc3x0123);
336 wasm_v128_store(c2, vacc2x0123);
337 wasm_v128_store(c1, vacc1x0123);
338 wasm_v128_store(c0, vacc0x0123);
339
340 vacc4x0123 = vacc4x4567;
341 vacc3x0123 = vacc3x4567;
342 vacc2x0123 = vacc2x4567;
343 vacc1x0123 = vacc1x4567;
344 vacc0x0123 = vacc0x4567;
345
346 c4 += 4;
347 c3 += 4;
348 c2 += 4;
349 c1 += 4;
350 c0 += 4;
351 }
352 if (nc & 2) {
353 *((double*) c4) = wasm_f64x2_extract_lane(vacc4x0123, 0);
354 *((double*) c3) = wasm_f64x2_extract_lane(vacc3x0123, 0);
355 *((double*) c2) = wasm_f64x2_extract_lane(vacc2x0123, 0);
356 *((double*) c1) = wasm_f64x2_extract_lane(vacc1x0123, 0);
357 *((double*) c0) = wasm_f64x2_extract_lane(vacc0x0123, 0);
358
359 vacc4x0123 = wasm_v32x4_shuffle(vacc4x0123, vacc4x0123, 2, 3, 2, 3);
360 vacc3x0123 = wasm_v32x4_shuffle(vacc3x0123, vacc3x0123, 2, 3, 2, 3);
361 vacc2x0123 = wasm_v32x4_shuffle(vacc2x0123, vacc2x0123, 2, 3, 2, 3);
362 vacc1x0123 = wasm_v32x4_shuffle(vacc1x0123, vacc1x0123, 2, 3, 2, 3);
363 vacc0x0123 = wasm_v32x4_shuffle(vacc0x0123, vacc0x0123, 2, 3, 2, 3);
364
365 c4 += 2;
366 c3 += 2;
367 c2 += 2;
368 c1 += 2;
369 c0 += 2;
370 }
371 if (nc & 1) {
372 *c4 = wasm_f32x4_extract_lane(vacc4x0123, 0);
373 *c3 = wasm_f32x4_extract_lane(vacc3x0123, 0);
374 *c2 = wasm_f32x4_extract_lane(vacc2x0123, 0);
375 *c1 = wasm_f32x4_extract_lane(vacc1x0123, 0);
376 *c0 = wasm_f32x4_extract_lane(vacc0x0123, 0);
377 }
378
379 nc = 0;
380 }
381 } while (nc != 0);
382 }
383