Debian Codegraph @

src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-4x8c16-minmax-wasmsdot.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/MRx4c16-wasmdot.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2023 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 #include "xnnpack/math.h"
16
17 void xnn_qd8_f32_qc8w_gemm_minmax_ukernel_4x8c16__wasmsdot(
18 size_t mr,
19 size_t nc,
20 size_t kc,
21 const int8_t* restrict a,
22 size_t a_stride,
23 const void* 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 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29 {
30 assert(mr != 0);
31 assert(mr <= 4);
32 assert(nc != 0);
33 assert(kc != 0);
34 assert(kc % sizeof(int8_t) == 0);
35 assert(a != NULL);
36 assert(w != NULL);
37 assert(c != NULL);
38
39 kc = round_up_po2(kc, 16 * sizeof(int8_t));
40 const int8_t* a0 = a;
41 float* c0 = c;
42 const int8_t* a1 = (const int8_t*) ((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 int8_t* a2 = (const int8_t*) ((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 int8_t* a3 = (const int8_t*) ((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
61 const v128_t vmin = wasm_v128_load32_splat(&params->scalar.min);
62 const v128_t vmax = wasm_v128_load32_splat(&params->scalar.max);
63 XNN_FORCE_REALIZATION(vmin);
64 XNN_FORCE_REALIZATION(vmax);
65
66 do {
67 v128_t vksum0 = wasm_v128_load32_zero(w);
68 v128_t vksum1 = wasm_v128_load32_zero((const int32_t*) w + 1);
69 v128_t vksum2 = wasm_v128_load32_zero((const int32_t*) w + 2);
70 v128_t vksum3 = wasm_v128_load32_zero((const int32_t*) w + 3);
71 v128_t vksum4 = wasm_v128_load32_zero((const int32_t*) w + 4);
72 v128_t vksum5 = wasm_v128_load32_zero((const int32_t*) w + 5);
73 v128_t vksum6 = wasm_v128_load32_zero((const int32_t*) w + 6);
74 v128_t vksum7 = wasm_v128_load32_zero((const int32_t*) w + 7);
75 const v128_t vinput_zero_point0 = wasm_v128_load32_splat(&quantization_params[0].zero_point);
76 const v128_t vinput_zero_point1 = wasm_v128_load32_splat(&quantization_params[1].zero_point);
77 const v128_t vinput_zero_point2 = wasm_v128_load32_splat(&quantization_params[2].zero_point);
78 const v128_t vinput_zero_point3 = wasm_v128_load32_splat(&quantization_params[3].zero_point);
79 v128_t vacc0x0 = wasm_i32x4_mul(vksum0, vinput_zero_point0);
80 v128_t vacc0x1 = wasm_i32x4_mul(vksum1, vinput_zero_point0);
81 v128_t vacc0x2 = wasm_i32x4_mul(vksum2, vinput_zero_point0);
82 v128_t vacc0x3 = wasm_i32x4_mul(vksum3, vinput_zero_point0);
83 v128_t vacc0x4 = wasm_i32x4_mul(vksum4, vinput_zero_point0);
84 v128_t vacc0x5 = wasm_i32x4_mul(vksum5, vinput_zero_point0);
85 v128_t vacc0x6 = wasm_i32x4_mul(vksum6, vinput_zero_point0);
86 v128_t vacc0x7 = wasm_i32x4_mul(vksum7, vinput_zero_point0);
87 v128_t vacc1x0 = wasm_i32x4_mul(vksum0, vinput_zero_point1);
88 v128_t vacc1x1 = wasm_i32x4_mul(vksum1, vinput_zero_point1);
89 v128_t vacc1x2 = wasm_i32x4_mul(vksum2, vinput_zero_point1);
90 v128_t vacc1x3 = wasm_i32x4_mul(vksum3, vinput_zero_point1);
91 v128_t vacc1x4 = wasm_i32x4_mul(vksum4, vinput_zero_point1);
92 v128_t vacc1x5 = wasm_i32x4_mul(vksum5, vinput_zero_point1);
93 v128_t vacc1x6 = wasm_i32x4_mul(vksum6, vinput_zero_point1);
94 v128_t vacc1x7 = wasm_i32x4_mul(vksum7, vinput_zero_point1);
95 v128_t vacc2x0 = wasm_i32x4_mul(vksum0, vinput_zero_point2);
96 v128_t vacc2x1 = wasm_i32x4_mul(vksum1, vinput_zero_point2);
97 v128_t vacc2x2 = wasm_i32x4_mul(vksum2, vinput_zero_point2);
98 v128_t vacc2x3 = wasm_i32x4_mul(vksum3, vinput_zero_point2);
99 v128_t vacc2x4 = wasm_i32x4_mul(vksum4, vinput_zero_point2);
100 v128_t vacc2x5 = wasm_i32x4_mul(vksum5, vinput_zero_point2);
101 v128_t vacc2x6 = wasm_i32x4_mul(vksum6, vinput_zero_point2);
102 v128_t vacc2x7 = wasm_i32x4_mul(vksum7, vinput_zero_point2);
103 v128_t vacc3x0 = wasm_i32x4_mul(vksum0, vinput_zero_point3);
104 v128_t vacc3x1 = wasm_i32x4_mul(vksum1, vinput_zero_point3);
105 v128_t vacc3x2 = wasm_i32x4_mul(vksum2, vinput_zero_point3);
106 v128_t vacc3x3 = wasm_i32x4_mul(vksum3, vinput_zero_point3);
107 v128_t vacc3x4 = wasm_i32x4_mul(vksum4, vinput_zero_point3);
108 v128_t vacc3x5 = wasm_i32x4_mul(vksum5, vinput_zero_point3);
109 v128_t vacc3x6 = wasm_i32x4_mul(vksum6, vinput_zero_point3);
110 v128_t vacc3x7 = wasm_i32x4_mul(vksum7, vinput_zero_point3);
111 w = (const int32_t*) w + 8;
112
113 size_t k = kc;
114 do {
115 const v128_t va0 = wasm_v128_load(a0);
116 a0 += 16;
117 const v128_t va1 = wasm_v128_load(a1);
118 a1 += 16;
119 const v128_t va2 = wasm_v128_load(a2);
120 a2 += 16;
121 const v128_t va3 = wasm_v128_load(a3);
122 a3 += 16;
123
124 const v128_t vb0 = wasm_v128_load(w);
125
126 vacc0x0 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0, va0, vacc0x0);
127 vacc1x0 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0, va1, vacc1x0);
128 vacc2x0 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0, va2, vacc2x0);
129 vacc3x0 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0, va3, vacc3x0);
130 const v128_t vb1 = wasm_v128_load((const int8_t*) w + 16);
131
132 vacc0x1 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1, va0, vacc0x1);
133 vacc1x1 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1, va1, vacc1x1);
134 vacc2x1 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1, va2, vacc2x1);
135 vacc3x1 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1, va3, vacc3x1);
136 const v128_t vb2 = wasm_v128_load((const int8_t*) w + 32);
137
138 vacc0x2 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb2, va0, vacc0x2);
139 vacc1x2 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb2, va1, vacc1x2);
140 vacc2x2 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb2, va2, vacc2x2);
141 vacc3x2 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb2, va3, vacc3x2);
142 const v128_t vb3 = wasm_v128_load((const int8_t*) w + 48);
143
144 vacc0x3 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb3, va0, vacc0x3);
145 vacc1x3 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb3, va1, vacc1x3);
146 vacc2x3 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb3, va2, vacc2x3);
147 vacc3x3 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb3, va3, vacc3x3);
148 const v128_t vb4 = wasm_v128_load((const int8_t*) w + 64);
149
150 vacc0x4 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb4, va0, vacc0x4);
151 vacc1x4 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb4, va1, vacc1x4);
152 vacc2x4 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb4, va2, vacc2x4);
153 vacc3x4 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb4, va3, vacc3x4);
154 const v128_t vb5 = wasm_v128_load((const int8_t*) w + 80);
155
156 vacc0x5 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb5, va0, vacc0x5);
157 vacc1x5 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb5, va1, vacc1x5);
158 vacc2x5 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb5, va2, vacc2x5);
159 vacc3x5 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb5, va3, vacc3x5);
160 const v128_t vb6 = wasm_v128_load((const int8_t*) w + 96);
161
162 vacc0x6 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb6, va0, vacc0x6);
163 vacc1x6 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb6, va1, vacc1x6);
164 vacc2x6 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb6, va2, vacc2x6);
165 vacc3x6 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb6, va3, vacc3x6);
166 const v128_t vb7 = wasm_v128_load((const int8_t*) w + 112);
167
168 vacc0x7 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb7, va0, vacc0x7);
169 vacc1x7 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb7, va1, vacc1x7);
170 vacc2x7 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb7, va2, vacc2x7);
171 vacc3x7 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb7, va3, vacc3x7);
172
173 w = (const int8_t*) w + 128;
174 k -= 16 * sizeof(int8_t);
175 } while (k != 0);
176
177 const v128_t vacc0x02 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x0, vacc0x2, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x0, vacc0x2, 2, 6, 3, 7));
178 const v128_t vacc0x13 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x1, vacc0x3, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x1, vacc0x3, 2, 6, 3, 7));
179 const v128_t vacc0x46 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x4, vacc0x6, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x4, vacc0x6, 2, 6, 3, 7));
180 const v128_t vacc0x57 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x5, vacc0x7, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x5, vacc0x7, 2, 6, 3, 7));
181 const v128_t vacc1x02 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x0, vacc1x2, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x0, vacc1x2, 2, 6, 3, 7));
182 const v128_t vacc1x13 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x1, vacc1x3, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x1, vacc1x3, 2, 6, 3, 7));
183 const v128_t vacc1x46 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x4, vacc1x6, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x4, vacc1x6, 2, 6, 3, 7));
184 const v128_t vacc1x57 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x5, vacc1x7, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x5, vacc1x7, 2, 6, 3, 7));
185 const v128_t vacc2x02 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc2x0, vacc2x2, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc2x0, vacc2x2, 2, 6, 3, 7));
186 const v128_t vacc2x13 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc2x1, vacc2x3, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc2x1, vacc2x3, 2, 6, 3, 7));
187 const v128_t vacc2x46 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc2x4, vacc2x6, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc2x4, vacc2x6, 2, 6, 3, 7));
188 const v128_t vacc2x57 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc2x5, vacc2x7, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc2x5, vacc2x7, 2, 6, 3, 7));
189 const v128_t vacc3x02 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc3x0, vacc3x2, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc3x0, vacc3x2, 2, 6, 3, 7));
190 const v128_t vacc3x13 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc3x1, vacc3x3, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc3x1, vacc3x3, 2, 6, 3, 7));
191 const v128_t vacc3x46 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc3x4, vacc3x6, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc3x4, vacc3x6, 2, 6, 3, 7));
192 const v128_t vacc3x57 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc3x5, vacc3x7, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc3x5, vacc3x7, 2, 6, 3, 7));
193
194 v128_t vacc0x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x02, vacc0x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x02, vacc0x13, 2, 6, 3, 7));
195 v128_t vacc0x4567 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x46, vacc0x57, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x46, vacc0x57, 2, 6, 3, 7));
196 v128_t vacc1x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x02, vacc1x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x02, vacc1x13, 2, 6, 3, 7));
197 v128_t vacc1x4567 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x46, vacc1x57, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x46, vacc1x57, 2, 6, 3, 7));
198 v128_t vacc2x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc2x02, vacc2x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc2x02, vacc2x13, 2, 6, 3, 7));
199 v128_t vacc2x4567 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc2x46, vacc2x57, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc2x46, vacc2x57, 2, 6, 3, 7));
200 v128_t vacc3x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc3x02, vacc3x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc3x02, vacc3x13, 2, 6, 3, 7));
201 v128_t vacc3x4567 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc3x46, vacc3x57, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc3x46, vacc3x57, 2, 6, 3, 7));
202
203 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
204 vacc0x4567 = wasm_f32x4_convert_i32x4(vacc0x4567);
205 vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
206 vacc1x4567 = wasm_f32x4_convert_i32x4(vacc1x4567);
207 vacc2x0123 = wasm_f32x4_convert_i32x4(vacc2x0123);
208 vacc2x4567 = wasm_f32x4_convert_i32x4(vacc2x4567);
209 vacc3x0123 = wasm_f32x4_convert_i32x4(vacc3x0123);
210 vacc3x4567 = wasm_f32x4_convert_i32x4(vacc3x4567);
211
212 const v128_t vinput_scale0 = wasm_v128_load32_splat(&quantization_params[0].inv_scale);
213 const v128_t vinput_scale1 = wasm_v128_load32_splat(&quantization_params[1].inv_scale);
214 const v128_t vinput_scale2 = wasm_v128_load32_splat(&quantization_params[2].inv_scale);
215 const v128_t vinput_scale3 = wasm_v128_load32_splat(&quantization_params[3].inv_scale);
216
217 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vinput_scale0);
218 vacc0x4567 = wasm_f32x4_mul(vacc0x4567, vinput_scale0);
219 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vinput_scale1);
220 vacc1x4567 = wasm_f32x4_mul(vacc1x4567, vinput_scale1);
221 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vinput_scale2);
222 vacc2x4567 = wasm_f32x4_mul(vacc2x4567, vinput_scale2);
223 vacc3x0123 = wasm_f32x4_mul(vacc3x0123, vinput_scale3);
224 vacc3x4567 = wasm_f32x4_mul(vacc3x4567, vinput_scale3);
225
226 const v128_t vfilter_output_scale0123 = wasm_v128_load(w);
227 w = (const float*) w + 4;
228 const v128_t vfilter_output_scale4567 = wasm_v128_load(w);
229 w = (const float*) w + 4;
230 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vfilter_output_scale0123);
231 vacc0x4567 = wasm_f32x4_mul(vacc0x4567, vfilter_output_scale4567);
232 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vfilter_output_scale0123);
233 vacc1x4567 = wasm_f32x4_mul(vacc1x4567, vfilter_output_scale4567);
234 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vfilter_output_scale0123);
235 vacc2x4567 = wasm_f32x4_mul(vacc2x4567, vfilter_output_scale4567);
236 vacc3x0123 = wasm_f32x4_mul(vacc3x0123, vfilter_output_scale0123);
237 vacc3x4567 = wasm_f32x4_mul(vacc3x4567, vfilter_output_scale4567);
238
239 const v128_t vbias0123 = wasm_v128_load(w);
240 w = (const float*) w + 4;
241 const v128_t vbias4567 = wasm_v128_load(w);
242 w = (const float*) w + 4;
243 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vbias0123);
244 vacc0x4567 = wasm_f32x4_add(vacc0x4567, vbias4567);
245 vacc1x0123 = wasm_f32x4_add(vacc1x0123, vbias0123);
246 vacc1x4567 = wasm_f32x4_add(vacc1x4567, vbias4567);
247 vacc2x0123 = wasm_f32x4_add(vacc2x0123, vbias0123);
248 vacc2x4567 = wasm_f32x4_add(vacc2x4567, vbias4567);
249 vacc3x0123 = wasm_f32x4_add(vacc3x0123, vbias0123);
250 vacc3x4567 = wasm_f32x4_add(vacc3x4567, vbias4567);
251
252 vacc0x0123 = wasm_f32x4_pmax(vacc0x0123, vmin);
253 vacc0x4567 = wasm_f32x4_pmax(vacc0x4567, vmin);
254 vacc1x0123 = wasm_f32x4_pmax(vacc1x0123, vmin);
255 vacc1x4567 = wasm_f32x4_pmax(vacc1x4567, vmin);
256 vacc2x0123 = wasm_f32x4_pmax(vacc2x0123, vmin);
257 vacc2x4567 = wasm_f32x4_pmax(vacc2x4567, vmin);
258 vacc3x0123 = wasm_f32x4_pmax(vacc3x0123, vmin);
259 vacc3x4567 = wasm_f32x4_pmax(vacc3x4567, vmin);
260
261 vacc0x0123 = wasm_f32x4_pmin(vacc0x0123, vmax);
262 vacc0x4567 = wasm_f32x4_pmin(vacc0x4567, vmax);
263 vacc1x0123 = wasm_f32x4_pmin(vacc1x0123, vmax);
264 vacc1x4567 = wasm_f32x4_pmin(vacc1x4567, vmax);
265 vacc2x0123 = wasm_f32x4_pmin(vacc2x0123, vmax);
266 vacc2x4567 = wasm_f32x4_pmin(vacc2x4567, vmax);
267 vacc3x0123 = wasm_f32x4_pmin(vacc3x0123, vmax);
268 vacc3x4567 = wasm_f32x4_pmin(vacc3x4567, vmax);
269
270 if XNN_LIKELY(nc >= 8) {
271 wasm_v128_store(c0, vacc0x0123);
272 wasm_v128_store(c0 + 4, vacc0x4567);
273 wasm_v128_store(c1, vacc1x0123);
274 wasm_v128_store(c1 + 4, vacc1x4567);
275 wasm_v128_store(c2, vacc2x0123);
276 wasm_v128_store(c2 + 4, vacc2x4567);
277 wasm_v128_store(c3, vacc3x0123);
278 wasm_v128_store(c3 + 4, vacc3x4567);
279
280 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
281 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
282 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
283 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
284
285 c0 = (float*) ((uintptr_t) c0 + cn_stride);
286 c1 = (float*) ((uintptr_t) c1 + cn_stride);
287 c2 = (float*) ((uintptr_t) c2 + cn_stride);
288 c3 = (float*) ((uintptr_t) c3 + cn_stride);
289
290 nc -= 8;
291 } else {
292 if (nc & 4) {
293 wasm_v128_store(c0, vacc0x0123);
294 vacc0x0123 = vacc0x4567;
295 c0 += 4;
296 wasm_v128_store(c1, vacc1x0123);
297 vacc1x0123 = vacc1x4567;
298 c1 += 4;
299 wasm_v128_store(c2, vacc2x0123);
300 vacc2x0123 = vacc2x4567;
301 c2 += 4;
302 wasm_v128_store(c3, vacc3x0123);
303 vacc3x0123 = vacc3x4567;
304 c3 += 4;
305 }
306 if (nc & 2) {
307 wasm_v128_store64_lane(c0, vacc0x0123, 0);
308 vacc0x0123 = wasm_v64x2_shuffle(vacc0x0123, vacc0x0123, 1, 1);
309 c0 += 2;
310 wasm_v128_store64_lane(c1, vacc1x0123, 0);
311 vacc1x0123 = wasm_v64x2_shuffle(vacc1x0123, vacc1x0123, 1, 1);
312 c1 += 2;
313 wasm_v128_store64_lane(c2, vacc2x0123, 0);
314 vacc2x0123 = wasm_v64x2_shuffle(vacc2x0123, vacc2x0123, 1, 1);
315 c2 += 2;
316 wasm_v128_store64_lane(c3, vacc3x0123, 0);
317 vacc3x0123 = wasm_v64x2_shuffle(vacc3x0123, vacc3x0123, 1, 1);
318 c3 += 2;
319 }
320 if (nc & 1) {
321 wasm_v128_store32_lane(c0, vacc0x0123, 0);
322 wasm_v128_store32_lane(c1, vacc1x0123, 0);
323 wasm_v128_store32_lane(c2, vacc2x0123, 0);
324 wasm_v128_store32_lane(c3, vacc3x0123, 0);
325 }
326 nc = 0;
327 }
328 } while (nc != 0);
329 }