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_5x8s4__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_5x8s4__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 <= 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 do {
66 v128_t vacc0x0123 = wasm_v128_load(w + 0);
67 v128_t vacc0x4567 = wasm_v128_load(w + 4);
68 v128_t vacc1x0123 = vacc0x0123;
69 v128_t vacc1x4567 = vacc0x4567;
70 v128_t vacc2x0123 = vacc0x0123;
71 v128_t vacc2x4567 = vacc0x4567;
72 v128_t vacc3x0123 = vacc0x0123;
73 v128_t vacc3x4567 = vacc0x4567;
74 v128_t vacc4x0123 = vacc0x0123;
75 v128_t vacc4x4567 = vacc0x4567;
76 w += 8;
77
78 size_t k = kc;
79 while (k >= 4 * sizeof(float)) {
80 v128_t va0 = wasm_v128_load(a0);
81 a0 += 4;
82 v128_t va1 = wasm_v128_load(a1);
83 a1 += 4;
84 v128_t va2 = wasm_v128_load(a2);
85 a2 += 4;
86 v128_t va3 = wasm_v128_load(a3);
87 a3 += 4;
88 v128_t va4 = wasm_v128_load(a4);
89 a4 += 4;
90
91
92 const v128_t vb0123c0 = wasm_v128_load(w + 0);
93 const v128_t vb4567c0 = wasm_v128_load(w + 4);
94
95 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c0));
96 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123c0));
97 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123c0));
98 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3, vb0123c0));
99 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123c0));
100 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c0));
101 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1, vb4567c0));
102 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2, vb4567c0));
103 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3, vb4567c0));
104 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4, vb4567c0));
105
106 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
107 va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
108 va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
109 va3 = wasm_v32x4_shuffle(va3, va3, 1, 2, 3, 0);
110 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0);
111
112 const v128_t vb0123c1 = wasm_v128_load(w + 8);
113 const v128_t vb4567c1 = wasm_v128_load(w + 12);
114
115 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c1));
116 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123c1));
117 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123c1));
118 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3, vb0123c1));
119 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123c1));
120 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c1));
121 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1, vb4567c1));
122 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2, vb4567c1));
123 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3, vb4567c1));
124 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4, vb4567c1));
125
126 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
127 va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
128 va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
129 va3 = wasm_v32x4_shuffle(va3, va3, 1, 2, 3, 0);
130 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0);
131
132 const v128_t vb0123c2 = wasm_v128_load(w + 16);
133 const v128_t vb4567c2 = wasm_v128_load(w + 20);
134
135 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c2));
136 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123c2));
137 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123c2));
138 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3, vb0123c2));
139 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123c2));
140 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c2));
141 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1, vb4567c2));
142 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2, vb4567c2));
143 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3, vb4567c2));
144 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4, vb4567c2));
145
146 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
147 va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
148 va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
149 va3 = wasm_v32x4_shuffle(va3, va3, 1, 2, 3, 0);
150 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0);
151
152 const v128_t vb0123c3 = wasm_v128_load(w + 24);
153 const v128_t vb4567c3 = wasm_v128_load(w + 28);
154
155 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c3));
156 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123c3));
157 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123c3));
158 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3, vb0123c3));
159 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123c3));
160 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c3));
161 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1, vb4567c3));
162 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2, vb4567c3));
163 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3, vb4567c3));
164 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4, vb4567c3));
165
166
167 w += 32;
168 k -= 4 * sizeof(float);
169 }
170 if XNN_UNLIKELY(k != 0) {
171 v128_t va0 = wasm_v128_load(a0);
172 a0 = (const float*) ((uintptr_t) a0 + k);
173 v128_t va1 = wasm_v128_load(a1);
174 a1 = (const float*) ((uintptr_t) a1 + k);
175 v128_t va2 = wasm_v128_load(a2);
176 a2 = (const float*) ((uintptr_t) a2 + k);
177 v128_t va3 = wasm_v128_load(a3);
178 a3 = (const float*) ((uintptr_t) a3 + k);
179 v128_t va4 = wasm_v128_load(a4);
180 a4 = (const float*) ((uintptr_t) a4 + k);
181
182 const v128_t vzero = wasm_f32x4_const_splat(0.0f);
183
184 const v128_t vb0123c0 = wasm_v128_load(w + 0);
185 const v128_t vb4567c0 = wasm_v128_load(w + 4);
186
187 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0));
188 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0));
189 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0));
190 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(wasm_v128_andnot(va3, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0));
191 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(wasm_v128_andnot(va4, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0));
192 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb4567c0, vzero)), vb4567c0));
193 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb4567c0, vzero)), vb4567c0));
194 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb4567c0, vzero)), vb4567c0));
195 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(wasm_v128_andnot(va3, wasm_f32x4_eq(vb4567c0, vzero)), vb4567c0));
196 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(wasm_v128_andnot(va4, wasm_f32x4_eq(vb4567c0, vzero)), vb4567c0));
197
198 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
199 va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
200 va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
201 va3 = wasm_v32x4_shuffle(va3, va3, 1, 2, 3, 0);
202 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0);
203
204 const v128_t vb0123c1 = wasm_v128_load(w + 8);
205 const v128_t vb4567c1 = wasm_v128_load(w + 12);
206
207 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c1, vzero)), vb0123c1));
208 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb0123c1, vzero)), vb0123c1));
209 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb0123c1, vzero)), vb0123c1));
210 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(wasm_v128_andnot(va3, wasm_f32x4_eq(vb0123c1, vzero)), vb0123c1));
211 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(wasm_v128_andnot(va4, wasm_f32x4_eq(vb0123c1, vzero)), vb0123c1));
212 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb4567c1, vzero)), vb4567c1));
213 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb4567c1, vzero)), vb4567c1));
214 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb4567c1, vzero)), vb4567c1));
215 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(wasm_v128_andnot(va3, wasm_f32x4_eq(vb4567c1, vzero)), vb4567c1));
216 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(wasm_v128_andnot(va4, wasm_f32x4_eq(vb4567c1, vzero)), vb4567c1));
217
218 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
219 va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
220 va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
221 va3 = wasm_v32x4_shuffle(va3, va3, 1, 2, 3, 0);
222 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0);
223
224 const v128_t vb0123c2 = wasm_v128_load(w + 16);
225 const v128_t vb4567c2 = wasm_v128_load(w + 20);
226
227 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c2, vzero)), vb0123c2));
228 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb0123c2, vzero)), vb0123c2));
229 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb0123c2, vzero)), vb0123c2));
230 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(wasm_v128_andnot(va3, wasm_f32x4_eq(vb0123c2, vzero)), vb0123c2));
231 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(wasm_v128_andnot(va4, wasm_f32x4_eq(vb0123c2, vzero)), vb0123c2));
232 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb4567c2, vzero)), vb4567c2));
233 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb4567c2, vzero)), vb4567c2));
234 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb4567c2, vzero)), vb4567c2));
235 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(wasm_v128_andnot(va3, wasm_f32x4_eq(vb4567c2, vzero)), vb4567c2));
236 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(wasm_v128_andnot(va4, wasm_f32x4_eq(vb4567c2, vzero)), vb4567c2));
237
238 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
239 va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
240 va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
241 va3 = wasm_v32x4_shuffle(va3, va3, 1, 2, 3, 0);
242 va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0);
243
244 const v128_t vb0123c3 = wasm_v128_load(w + 24);
245 const v128_t vb4567c3 = wasm_v128_load(w + 28);
246
247 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c3, vzero)), vb0123c3));
248 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb0123c3, vzero)), vb0123c3));
249 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb0123c3, vzero)), vb0123c3));
250 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(wasm_v128_andnot(va3, wasm_f32x4_eq(vb0123c3, vzero)), vb0123c3));
251 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(wasm_v128_andnot(va4, wasm_f32x4_eq(vb0123c3, vzero)), vb0123c3));
252 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb4567c3, vzero)), vb4567c3));
253 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb4567c3, vzero)), vb4567c3));
254 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb4567c3, vzero)), vb4567c3));
255 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(wasm_v128_andnot(va3, wasm_f32x4_eq(vb4567c3, vzero)), vb4567c3));
256 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(wasm_v128_andnot(va4, wasm_f32x4_eq(vb4567c3, vzero)), vb4567c3));
257
258
259 w += 32;
260 }
261
262 const v128_t vzero = wasm_i32x4_const_splat(0);
263 vacc0x0123 = wasm_i32x4_max(vacc0x0123, vzero);
264 vacc1x0123 = wasm_i32x4_max(vacc1x0123, vzero);
265 vacc2x0123 = wasm_i32x4_max(vacc2x0123, vzero);
266 vacc3x0123 = wasm_i32x4_max(vacc3x0123, vzero);
267 vacc4x0123 = wasm_i32x4_max(vacc4x0123, vzero);
268 vacc0x4567 = wasm_i32x4_max(vacc0x4567, vzero);
269 vacc1x4567 = wasm_i32x4_max(vacc1x4567, vzero);
270 vacc2x4567 = wasm_i32x4_max(vacc2x4567, vzero);
271 vacc3x4567 = wasm_i32x4_max(vacc3x4567, vzero);
272 vacc4x4567 = wasm_i32x4_max(vacc4x4567, vzero);
273
274 if XNN_LIKELY(nc >= 8) {
275 wasm_v128_store(c4, vacc4x0123);
276 wasm_v128_store(c4 + 4, vacc4x4567);
277 c4 = (float*) ((uintptr_t) c4 + cn_stride);
278 wasm_v128_store(c3, vacc3x0123);
279 wasm_v128_store(c3 + 4, vacc3x4567);
280 c3 = (float*) ((uintptr_t) c3 + cn_stride);
281 wasm_v128_store(c2, vacc2x0123);
282 wasm_v128_store(c2 + 4, vacc2x4567);
283 c2 = (float*) ((uintptr_t) c2 + cn_stride);
284 wasm_v128_store(c1, vacc1x0123);
285 wasm_v128_store(c1 + 4, vacc1x4567);
286 c1 = (float*) ((uintptr_t) c1 + cn_stride);
287 wasm_v128_store(c0, vacc0x0123);
288 wasm_v128_store(c0 + 4, vacc0x4567);
289 c0 = (float*) ((uintptr_t) c0 + cn_stride);
290
291 a4 = (const float*) ((uintptr_t) a4 - kc);
292 a3 = (const float*) ((uintptr_t) a3 - kc);
293 a2 = (const float*) ((uintptr_t) a2 - kc);
294 a1 = (const float*) ((uintptr_t) a1 - kc);
295 a0 = (const float*) ((uintptr_t) a0 - kc);
296
297 nc -= 8;
298 } else {
299 if (nc & 4) {
300 wasm_v128_store(c4, vacc4x0123);
301 wasm_v128_store(c3, vacc3x0123);
302 wasm_v128_store(c2, vacc2x0123);
303 wasm_v128_store(c1, vacc1x0123);
304 wasm_v128_store(c0, vacc0x0123);
305
306 vacc4x0123 = vacc4x4567;
307 vacc3x0123 = vacc3x4567;
308 vacc2x0123 = vacc2x4567;
309 vacc1x0123 = vacc1x4567;
310 vacc0x0123 = vacc0x4567;
311
312 c4 += 4;
313 c3 += 4;
314 c2 += 4;
315 c1 += 4;
316 c0 += 4;
317 }
318 if (nc & 2) {
319 *((double*) c4) = wasm_f64x2_extract_lane(vacc4x0123, 0);
320 *((double*) c3) = wasm_f64x2_extract_lane(vacc3x0123, 0);
321 *((double*) c2) = wasm_f64x2_extract_lane(vacc2x0123, 0);
322 *((double*) c1) = wasm_f64x2_extract_lane(vacc1x0123, 0);
323 *((double*) c0) = wasm_f64x2_extract_lane(vacc0x0123, 0);
324
325 vacc4x0123 = wasm_v32x4_shuffle(vacc4x0123, vacc4x0123, 2, 3, 2, 3);
326 vacc3x0123 = wasm_v32x4_shuffle(vacc3x0123, vacc3x0123, 2, 3, 2, 3);
327 vacc2x0123 = wasm_v32x4_shuffle(vacc2x0123, vacc2x0123, 2, 3, 2, 3);
328 vacc1x0123 = wasm_v32x4_shuffle(vacc1x0123, vacc1x0123, 2, 3, 2, 3);
329 vacc0x0123 = wasm_v32x4_shuffle(vacc0x0123, vacc0x0123, 2, 3, 2, 3);
330
331 c4 += 2;
332 c3 += 2;
333 c2 += 2;
334 c1 += 2;
335 c0 += 2;
336 }
337 if (nc & 1) {
338 *c4 = wasm_f32x4_extract_lane(vacc4x0123, 0);
339 *c3 = wasm_f32x4_extract_lane(vacc3x0123, 0);
340 *c2 = wasm_f32x4_extract_lane(vacc2x0123, 0);
341 *c1 = wasm_f32x4_extract_lane(vacc1x0123, 0);
342 *c0 = wasm_f32x4_extract_lane(vacc0x0123, 0);
343 }
344
345 nc = 0;
346 }
347 } while (nc != 0);
348 }
349