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_relu_ukernel_5x8s4__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_relu_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_f32_igemm_relu_ukernel_5x8s4__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_relu_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 do {
62 v128_t vacc0x0123 = wasm_v128_load(w);
63 v128_t vacc0x4567 = wasm_v128_load(w + 4);
64 v128_t vacc1x0123 = vacc0x0123;
65 v128_t vacc1x4567 = vacc0x4567;
66 v128_t vacc2x0123 = vacc0x0123;
67 v128_t vacc2x4567 = vacc0x4567;
68 v128_t vacc3x0123 = vacc0x0123;
69 v128_t vacc3x4567 = vacc0x4567;
70 v128_t vacc4x0123 = vacc0x0123;
71 v128_t vacc4x4567 = vacc0x4567;
72 w += 8;
73
74 size_t p = ks;
75 do {
76 const float* restrict a0 = a[0];
77 assert(a0 != NULL);
78 if XNN_UNPREDICTABLE(a0 != zero) {
79 a0 = (const float*) ((uintptr_t) a0 + a_offset);
80 }
81 const float* restrict a1 = a[1];
82 assert(a1 != NULL);
83 if XNN_UNPREDICTABLE(a1 != zero) {
84 a1 = (const float*) ((uintptr_t) a1 + a_offset);
85 }
86 const float* restrict a2 = a[2];
87 assert(a2 != NULL);
88 if XNN_UNPREDICTABLE(a2 != zero) {
89 a2 = (const float*) ((uintptr_t) a2 + a_offset);
90 }
91 const float* restrict a3 = a[3];
92 assert(a3 != NULL);
93 if XNN_UNPREDICTABLE(a3 != zero) {
94 a3 = (const float*) ((uintptr_t) a3 + a_offset);
95 }
96 const float* restrict a4 = a[4];
97 assert(a4 != NULL);
98 if XNN_UNPREDICTABLE(a4 != zero) {
99 a4 = (const float*) ((uintptr_t) a4 + a_offset);
100 }
101 a += 5;
102
103 size_t k = kc;
104 while (k >= 4 * sizeof(float)) {
105 v128_t va0 = wasm_v128_load(a0);
106 a0 += 4;
107 v128_t va1 = wasm_v128_load(a1);
108 a1 += 4;
109 v128_t va2 = wasm_v128_load(a2);
110 a2 += 4;
111 v128_t va3 = wasm_v128_load(a3);
112 a3 += 4;
113 v128_t va4 = wasm_v128_load(a4);
114 a4 += 4;
115
116
117 const v128_t vb0123c0 = wasm_v128_load(w + 0);
118 const v128_t vb4567c0 = wasm_v128_load(w + 4);
119
120 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c0));
121 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123c0));
122 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123c0));
123 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3, vb0123c0));
124 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123c0));
125 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c0));
126 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1, vb4567c0));
127 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2, vb4567c0));
128 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3, vb4567c0));
129 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4, vb4567c0));
130
131 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
132 va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
133 va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
134 va3 = wasm_v32x4_shuffle(va3, va3, 1, 2, 3, 0);
135 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0);
136
137 const v128_t vb0123c1 = wasm_v128_load(w + 8);
138 const v128_t vb4567c1 = wasm_v128_load(w + 12);
139
140 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c1));
141 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123c1));
142 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123c1));
143 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3, vb0123c1));
144 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123c1));
145 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c1));
146 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1, vb4567c1));
147 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2, vb4567c1));
148 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3, vb4567c1));
149 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4, vb4567c1));
150
151 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
152 va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
153 va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
154 va3 = wasm_v32x4_shuffle(va3, va3, 1, 2, 3, 0);
155 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0);
156
157 const v128_t vb0123c2 = wasm_v128_load(w + 16);
158 const v128_t vb4567c2 = wasm_v128_load(w + 20);
159
160 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c2));
161 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123c2));
162 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123c2));
163 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3, vb0123c2));
164 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123c2));
165 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c2));
166 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1, vb4567c2));
167 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2, vb4567c2));
168 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3, vb4567c2));
169 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4, vb4567c2));
170
171 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
172 va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
173 va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
174 va3 = wasm_v32x4_shuffle(va3, va3, 1, 2, 3, 0);
175 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0);
176
177 const v128_t vb0123c3 = wasm_v128_load(w + 24);
178 const v128_t vb4567c3 = wasm_v128_load(w + 28);
179
180 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c3));
181 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123c3));
182 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123c3));
183 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3, vb0123c3));
184 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123c3));
185 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c3));
186 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1, vb4567c3));
187 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2, vb4567c3));
188 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3, vb4567c3));
189 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4, vb4567c3));
190
191
192 w += 32;
193 k -= 4 * sizeof(float);
194 }
195 if XNN_UNLIKELY(k != 0) {
196 v128_t va0 = wasm_v128_load(a0);
197 a0 = (const float*) ((uintptr_t) a0 + k);
198 v128_t va1 = wasm_v128_load(a1);
199 a1 = (const float*) ((uintptr_t) a1 + k);
200 v128_t va2 = wasm_v128_load(a2);
201 a2 = (const float*) ((uintptr_t) a2 + k);
202 v128_t va3 = wasm_v128_load(a3);
203 a3 = (const float*) ((uintptr_t) a3 + k);
204 v128_t va4 = wasm_v128_load(a4);
205 a4 = (const float*) ((uintptr_t) a4 + k);
206
207 const v128_t vzero = wasm_f32x4_const_splat(0.0f);
208
209 const v128_t vb0123c0 = wasm_v128_load(w + 0);
210 const v128_t vb4567c0 = wasm_v128_load(w + 4);
211
212 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0));
213 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0));
214 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0));
215 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(wasm_v128_andnot(va3, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0));
216 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(wasm_v128_andnot(va4, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0));
217 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb4567c0, vzero)), vb4567c0));
218 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb4567c0, vzero)), vb4567c0));
219 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb4567c0, vzero)), vb4567c0));
220 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(wasm_v128_andnot(va3, wasm_f32x4_eq(vb4567c0, vzero)), vb4567c0));
221 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(wasm_v128_andnot(va4, 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
229 const v128_t vb0123c1 = wasm_v128_load(w + 8);
230 const v128_t vb4567c1 = wasm_v128_load(w + 12);
231
232 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c1, vzero)), vb0123c1));
233 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb0123c1, vzero)), vb0123c1));
234 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb0123c1, vzero)), vb0123c1));
235 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(wasm_v128_andnot(va3, wasm_f32x4_eq(vb0123c1, vzero)), vb0123c1));
236 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(wasm_v128_andnot(va4, wasm_f32x4_eq(vb0123c1, vzero)), vb0123c1));
237 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb4567c1, vzero)), vb4567c1));
238 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb4567c1, vzero)), vb4567c1));
239 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb4567c1, vzero)), vb4567c1));
240 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(wasm_v128_andnot(va3, wasm_f32x4_eq(vb4567c1, vzero)), vb4567c1));
241 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(wasm_v128_andnot(va4, wasm_f32x4_eq(vb4567c1, vzero)), vb4567c1));
242
243 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
244 va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
245 va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
246 va3 = wasm_v32x4_shuffle(va3, va3, 1, 2, 3, 0);
247 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0);
248
249 const v128_t vb0123c2 = wasm_v128_load(w + 16);
250 const v128_t vb4567c2 = wasm_v128_load(w + 20);
251
252 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c2, vzero)), vb0123c2));
253 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb0123c2, vzero)), vb0123c2));
254 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb0123c2, vzero)), vb0123c2));
255 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(wasm_v128_andnot(va3, wasm_f32x4_eq(vb0123c2, vzero)), vb0123c2));
256 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(wasm_v128_andnot(va4, wasm_f32x4_eq(vb0123c2, vzero)), vb0123c2));
257 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb4567c2, vzero)), vb4567c2));
258 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb4567c2, vzero)), vb4567c2));
259 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb4567c2, vzero)), vb4567c2));
260 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(wasm_v128_andnot(va3, wasm_f32x4_eq(vb4567c2, vzero)), vb4567c2));
261 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(wasm_v128_andnot(va4, wasm_f32x4_eq(vb4567c2, vzero)), vb4567c2));
262
263 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
264 va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
265 va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
266 va3 = wasm_v32x4_shuffle(va3, va3, 1, 2, 3, 0);
267 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0);
268
269 const v128_t vb0123c3 = wasm_v128_load(w + 24);
270 const v128_t vb4567c3 = wasm_v128_load(w + 28);
271
272 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c3, vzero)), vb0123c3));
273 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb0123c3, vzero)), vb0123c3));
274 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb0123c3, vzero)), vb0123c3));
275 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(wasm_v128_andnot(va3, wasm_f32x4_eq(vb0123c3, vzero)), vb0123c3));
276 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(wasm_v128_andnot(va4, wasm_f32x4_eq(vb0123c3, vzero)), vb0123c3));
277 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb4567c3, vzero)), vb4567c3));
278 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb4567c3, vzero)), vb4567c3));
279 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb4567c3, vzero)), vb4567c3));
280 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(wasm_v128_andnot(va3, wasm_f32x4_eq(vb4567c3, vzero)), vb4567c3));
281 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(wasm_v128_andnot(va4, wasm_f32x4_eq(vb4567c3, vzero)), vb4567c3));
282
283
284 w += 32;
285 }
286 p -= 5 * sizeof(void*);
287 } while (p != 0);
288
289 const v128_t vzero = wasm_i32x4_const_splat(0);
290 vacc0x0123 = wasm_i32x4_max(vacc0x0123, vzero);
291 vacc1x0123 = wasm_i32x4_max(vacc1x0123, vzero);
292 vacc2x0123 = wasm_i32x4_max(vacc2x0123, vzero);
293 vacc3x0123 = wasm_i32x4_max(vacc3x0123, vzero);
294 vacc4x0123 = wasm_i32x4_max(vacc4x0123, vzero);
295 vacc0x4567 = wasm_i32x4_max(vacc0x4567, vzero);
296 vacc1x4567 = wasm_i32x4_max(vacc1x4567, vzero);
297 vacc2x4567 = wasm_i32x4_max(vacc2x4567, vzero);
298 vacc3x4567 = wasm_i32x4_max(vacc3x4567, vzero);
299 vacc4x4567 = wasm_i32x4_max(vacc4x4567, vzero);
300
301 if XNN_LIKELY(nc >= 8) {
302 wasm_v128_store(c4, vacc4x0123);
303 wasm_v128_store(c4 + 4, vacc4x4567);
304 c4 = (float*) ((uintptr_t) c4 + cn_stride);
305 wasm_v128_store(c3, vacc3x0123);
306 wasm_v128_store(c3 + 4, vacc3x4567);
307 c3 = (float*) ((uintptr_t) c3 + cn_stride);
308 wasm_v128_store(c2, vacc2x0123);
309 wasm_v128_store(c2 + 4, vacc2x4567);
310 c2 = (float*) ((uintptr_t) c2 + cn_stride);
311 wasm_v128_store(c1, vacc1x0123);
312 wasm_v128_store(c1 + 4, vacc1x4567);
313 c1 = (float*) ((uintptr_t) c1 + cn_stride);
314 wasm_v128_store(c0, vacc0x0123);
315 wasm_v128_store(c0 + 4, vacc0x4567);
316 c0 = (float*) ((uintptr_t) c0 + cn_stride);
317
318 a = (const float**restrict) ((uintptr_t) a - ks);
319 nc -= 8;
320 } else {
321 if (nc & 4) {
322 wasm_v128_store(c4, vacc4x0123);
323 wasm_v128_store(c3, vacc3x0123);
324 wasm_v128_store(c2, vacc2x0123);
325 wasm_v128_store(c1, vacc1x0123);
326 wasm_v128_store(c0, vacc0x0123);
327
328 vacc4x0123 = vacc4x4567;
329 vacc3x0123 = vacc3x4567;
330 vacc2x0123 = vacc2x4567;
331 vacc1x0123 = vacc1x4567;
332 vacc0x0123 = vacc0x4567;
333
334 c4 += 4;
335 c3 += 4;
336 c2 += 4;
337 c1 += 4;
338 c0 += 4;
339 }
340 if (nc & 2) {
341 *((double*) c4) = wasm_f64x2_extract_lane(vacc4x0123, 0);
342 *((double*) c3) = wasm_f64x2_extract_lane(vacc3x0123, 0);
343 *((double*) c2) = wasm_f64x2_extract_lane(vacc2x0123, 0);
344 *((double*) c1) = wasm_f64x2_extract_lane(vacc1x0123, 0);
345 *((double*) c0) = wasm_f64x2_extract_lane(vacc0x0123, 0);
346
347 vacc4x0123 = wasm_v32x4_shuffle(vacc4x0123, vacc4x0123, 2, 3, 2, 3);
348 vacc3x0123 = wasm_v32x4_shuffle(vacc3x0123, vacc3x0123, 2, 3, 2, 3);
349 vacc2x0123 = wasm_v32x4_shuffle(vacc2x0123, vacc2x0123, 2, 3, 2, 3);
350 vacc1x0123 = wasm_v32x4_shuffle(vacc1x0123, vacc1x0123, 2, 3, 2, 3);
351 vacc0x0123 = wasm_v32x4_shuffle(vacc0x0123, vacc0x0123, 2, 3, 2, 3);
352
353 c4 += 2;
354 c3 += 2;
355 c2 += 2;
356 c1 += 2;
357 c0 += 2;
358 }
359 if (nc & 1) {
360 *c4 = wasm_f32x4_extract_lane(vacc4x0123, 0);
361 *c3 = wasm_f32x4_extract_lane(vacc3x0123, 0);
362 *c2 = wasm_f32x4_extract_lane(vacc2x0123, 0);
363 *c1 = wasm_f32x4_extract_lane(vacc1x0123, 0);
364 *c0 = wasm_f32x4_extract_lane(vacc0x0123, 0);
365 }
366
367 nc = 0;
368 }
369 } while (nc != 0);
370 }
371