Debian Codegraph @

src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-4x16c4-minmax-wasmusdot-u2.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/MRx16c4-wasmdot.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2024 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_4x16c4__wasmusdot_u2(
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, 4 * 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 const v128_t vsign_mask = wasm_u8x16_const_splat(UINT8_C(0x80));
67 XNN_FORCE_REALIZATION(vsign_mask);
68
69 do {
70 v128_t vksum0123 = wasm_v128_load((const int32_t*) w);
71 v128_t vksum4567 = wasm_v128_load((const int32_t*) w + 4);
72 v128_t vksum89AB = wasm_v128_load((const int32_t*) w + 8);
73 v128_t vksumCDEF = wasm_v128_load((const int32_t*) w + 12);
74 const v128_t vinput_zero_point0 = wasm_i32x4_splat((int32_t) quantization_params[0].zero_point + 128);
75 v128_t vacc0x0x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point0);
76 v128_t vacc0x0x4567 = wasm_i32x4_mul(vksum4567, vinput_zero_point0);
77 v128_t vacc0x0x89AB = wasm_i32x4_mul(vksum89AB, vinput_zero_point0);
78 v128_t vacc0x0xCDEF = wasm_i32x4_mul(vksumCDEF, vinput_zero_point0);
79 const v128_t vinput_zero_point1 = wasm_i32x4_splat((int32_t) quantization_params[1].zero_point + 128);
80 v128_t vacc1x0x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point1);
81 v128_t vacc1x0x4567 = wasm_i32x4_mul(vksum4567, vinput_zero_point1);
82 v128_t vacc1x0x89AB = wasm_i32x4_mul(vksum89AB, vinput_zero_point1);
83 v128_t vacc1x0xCDEF = wasm_i32x4_mul(vksumCDEF, vinput_zero_point1);
84 const v128_t vinput_zero_point2 = wasm_i32x4_splat((int32_t) quantization_params[2].zero_point + 128);
85 v128_t vacc2x0x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point2);
86 v128_t vacc2x0x4567 = wasm_i32x4_mul(vksum4567, vinput_zero_point2);
87 v128_t vacc2x0x89AB = wasm_i32x4_mul(vksum89AB, vinput_zero_point2);
88 v128_t vacc2x0xCDEF = wasm_i32x4_mul(vksumCDEF, vinput_zero_point2);
89 const v128_t vinput_zero_point3 = wasm_i32x4_splat((int32_t) quantization_params[3].zero_point + 128);
90 v128_t vacc3x0x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point3);
91 v128_t vacc3x0x4567 = wasm_i32x4_mul(vksum4567, vinput_zero_point3);
92 v128_t vacc3x0x89AB = wasm_i32x4_mul(vksum89AB, vinput_zero_point3);
93 v128_t vacc3x0xCDEF = wasm_i32x4_mul(vksumCDEF, vinput_zero_point3);
94 w = (const int32_t*) w + 16;
95
96 size_t k = kc;
97 while (k >= 8 * sizeof(int8_t)) {
98 v128_t va0x0x0123 = wasm_v128_load32_splat(a0);
99 v128_t va0x1x0123 = wasm_v128_load32_splat(a0 + 4);
100 a0 += 8;
101 v128_t va1x0x0123 = wasm_v128_load32_splat(a1);
102 v128_t va1x1x0123 = wasm_v128_load32_splat(a1 + 4);
103 a1 += 8;
104 v128_t va2x0x0123 = wasm_v128_load32_splat(a2);
105 v128_t va2x1x0123 = wasm_v128_load32_splat(a2 + 4);
106 a2 += 8;
107 v128_t va3x0x0123 = wasm_v128_load32_splat(a3);
108 v128_t va3x1x0123 = wasm_v128_load32_splat(a3 + 4);
109 a3 += 8;
110
111 va0x0x0123 = wasm_v128_xor(va0x0x0123, vsign_mask);
112 va0x1x0123 = wasm_v128_xor(va0x1x0123, vsign_mask);
113 va1x0x0123 = wasm_v128_xor(va1x0x0123, vsign_mask);
114 va1x1x0123 = wasm_v128_xor(va1x1x0123, vsign_mask);
115 va2x0x0123 = wasm_v128_xor(va2x0x0123, vsign_mask);
116 va2x1x0123 = wasm_v128_xor(va2x1x0123, vsign_mask);
117 va3x0x0123 = wasm_v128_xor(va3x0x0123, vsign_mask);
118 va3x1x0123 = wasm_v128_xor(va3x1x0123, vsign_mask);
119
120 const v128_t vb0x0123 = wasm_v128_load((const int8_t*) w);
121 const v128_t vb0x4567 = wasm_v128_load((const int8_t*) w + 16);
122 const v128_t vb0x89AB = wasm_v128_load((const int8_t*) w + 32);
123 const v128_t vb0xCDEF = wasm_v128_load((const int8_t*) w + 48);
124 const v128_t vb1x0123 = wasm_v128_load((const int8_t*) w + 64);
125 const v128_t vb1x4567 = wasm_v128_load((const int8_t*) w + 80);
126 const v128_t vb1x89AB = wasm_v128_load((const int8_t*) w + 96);
127 const v128_t vb1xCDEF = wasm_v128_load((const int8_t*) w + 112);
128
129 vacc0x0x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0x0123, va0x0x0123, vacc0x0x0123);
130 vacc0x0x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0x4567, va0x0x0123, vacc0x0x4567);
131 vacc0x0x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0x89AB, va0x0x0123, vacc0x0x89AB);
132 vacc0x0xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0xCDEF, va0x0x0123, vacc0x0xCDEF);
133 vacc1x0x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0x0123, va1x0x0123, vacc1x0x0123);
134 vacc1x0x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0x4567, va1x0x0123, vacc1x0x4567);
135 vacc1x0x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0x89AB, va1x0x0123, vacc1x0x89AB);
136 vacc1x0xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0xCDEF, va1x0x0123, vacc1x0xCDEF);
137 vacc2x0x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0x0123, va2x0x0123, vacc2x0x0123);
138 vacc2x0x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0x4567, va2x0x0123, vacc2x0x4567);
139 vacc2x0x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0x89AB, va2x0x0123, vacc2x0x89AB);
140 vacc2x0xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0xCDEF, va2x0x0123, vacc2x0xCDEF);
141 vacc3x0x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0x0123, va3x0x0123, vacc3x0x0123);
142 vacc3x0x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0x4567, va3x0x0123, vacc3x0x4567);
143 vacc3x0x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0x89AB, va3x0x0123, vacc3x0x89AB);
144 vacc3x0xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0xCDEF, va3x0x0123, vacc3x0xCDEF);
145 vacc0x0x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1x0123, va0x1x0123, vacc0x0x0123);
146 vacc0x0x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1x4567, va0x1x0123, vacc0x0x4567);
147 vacc0x0x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1x89AB, va0x1x0123, vacc0x0x89AB);
148 vacc0x0xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1xCDEF, va0x1x0123, vacc0x0xCDEF);
149 vacc1x0x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1x0123, va1x1x0123, vacc1x0x0123);
150 vacc1x0x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1x4567, va1x1x0123, vacc1x0x4567);
151 vacc1x0x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1x89AB, va1x1x0123, vacc1x0x89AB);
152 vacc1x0xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1xCDEF, va1x1x0123, vacc1x0xCDEF);
153 vacc2x0x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1x0123, va2x1x0123, vacc2x0x0123);
154 vacc2x0x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1x4567, va2x1x0123, vacc2x0x4567);
155 vacc2x0x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1x89AB, va2x1x0123, vacc2x0x89AB);
156 vacc2x0xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1xCDEF, va2x1x0123, vacc2x0xCDEF);
157 vacc3x0x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1x0123, va3x1x0123, vacc3x0x0123);
158 vacc3x0x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1x4567, va3x1x0123, vacc3x0x4567);
159 vacc3x0x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1x89AB, va3x1x0123, vacc3x0x89AB);
160 vacc3x0xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1xCDEF, va3x1x0123, vacc3x0xCDEF);
161
162 w = (const int8_t*) w + 128;
163 k -= 8 * sizeof(int8_t);
164 }
165
166 while (k != 0) {
167 v128_t va0x0123 = wasm_v128_load32_splat(a0);
168 a0 += 4;
169 v128_t va1x0123 = wasm_v128_load32_splat(a1);
170 a1 += 4;
171 v128_t va2x0123 = wasm_v128_load32_splat(a2);
172 a2 += 4;
173 v128_t va3x0123 = wasm_v128_load32_splat(a3);
174 a3 += 4;
175
176 va0x0123 = wasm_v128_xor(va0x0123, vsign_mask);
177 va1x0123 = wasm_v128_xor(va1x0123, vsign_mask);
178 va2x0123 = wasm_v128_xor(va2x0123, vsign_mask);
179 va3x0123 = wasm_v128_xor(va3x0123, vsign_mask);
180
181 const v128_t vb0123 = wasm_v128_load((const int8_t*) w);
182 const v128_t vb4567 = wasm_v128_load((const int8_t*) w + 16);
183 const v128_t vb89AB = wasm_v128_load((const int8_t*) w + 32);
184 const v128_t vbCDEF = wasm_v128_load((const int8_t*) w + 48);
185
186 vacc0x0x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0123, va0x0123, vacc0x0x0123);
187 vacc0x0x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb4567, va0x0123, vacc0x0x4567);
188 vacc0x0x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb89AB, va0x0123, vacc0x0x89AB);
189 vacc0x0xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vbCDEF, va0x0123, vacc0x0xCDEF);
190 vacc1x0x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0123, va1x0123, vacc1x0x0123);
191 vacc1x0x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb4567, va1x0123, vacc1x0x4567);
192 vacc1x0x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb89AB, va1x0123, vacc1x0x89AB);
193 vacc1x0xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vbCDEF, va1x0123, vacc1x0xCDEF);
194 vacc2x0x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0123, va2x0123, vacc2x0x0123);
195 vacc2x0x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb4567, va2x0123, vacc2x0x4567);
196 vacc2x0x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb89AB, va2x0123, vacc2x0x89AB);
197 vacc2x0xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vbCDEF, va2x0123, vacc2x0xCDEF);
198 vacc3x0x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0123, va3x0123, vacc3x0x0123);
199 vacc3x0x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb4567, va3x0123, vacc3x0x4567);
200 vacc3x0x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb89AB, va3x0123, vacc3x0x89AB);
201 vacc3x0xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vbCDEF, va3x0123, vacc3x0xCDEF);
202
203 w = (const int8_t*) w + 64;
204 k -= 4 * sizeof(int8_t);
205 }
206
207 vacc0x0x0123 = wasm_f32x4_convert_i32x4(vacc0x0x0123);
208 vacc0x0x4567 = wasm_f32x4_convert_i32x4(vacc0x0x4567);
209 vacc0x0x89AB = wasm_f32x4_convert_i32x4(vacc0x0x89AB);
210 vacc0x0xCDEF = wasm_f32x4_convert_i32x4(vacc0x0xCDEF);
211 vacc1x0x0123 = wasm_f32x4_convert_i32x4(vacc1x0x0123);
212 vacc1x0x4567 = wasm_f32x4_convert_i32x4(vacc1x0x4567);
213 vacc1x0x89AB = wasm_f32x4_convert_i32x4(vacc1x0x89AB);
214 vacc1x0xCDEF = wasm_f32x4_convert_i32x4(vacc1x0xCDEF);
215 vacc2x0x0123 = wasm_f32x4_convert_i32x4(vacc2x0x0123);
216 vacc2x0x4567 = wasm_f32x4_convert_i32x4(vacc2x0x4567);
217 vacc2x0x89AB = wasm_f32x4_convert_i32x4(vacc2x0x89AB);
218 vacc2x0xCDEF = wasm_f32x4_convert_i32x4(vacc2x0xCDEF);
219 vacc3x0x0123 = wasm_f32x4_convert_i32x4(vacc3x0x0123);
220 vacc3x0x4567 = wasm_f32x4_convert_i32x4(vacc3x0x4567);
221 vacc3x0x89AB = wasm_f32x4_convert_i32x4(vacc3x0x89AB);
222 vacc3x0xCDEF = wasm_f32x4_convert_i32x4(vacc3x0xCDEF);
223
224 const v128_t vinput_scale0 = wasm_v128_load32_splat(&quantization_params[0].inv_scale);
225 const v128_t vinput_scale1 = wasm_v128_load32_splat(&quantization_params[1].inv_scale);
226 const v128_t vinput_scale2 = wasm_v128_load32_splat(&quantization_params[2].inv_scale);
227 const v128_t vinput_scale3 = wasm_v128_load32_splat(&quantization_params[3].inv_scale);
228
229 vacc0x0x0123 = wasm_f32x4_mul(vacc0x0x0123, vinput_scale0);
230 vacc0x0x4567 = wasm_f32x4_mul(vacc0x0x4567, vinput_scale0);
231 vacc0x0x89AB = wasm_f32x4_mul(vacc0x0x89AB, vinput_scale0);
232 vacc0x0xCDEF = wasm_f32x4_mul(vacc0x0xCDEF, vinput_scale0);
233 vacc1x0x0123 = wasm_f32x4_mul(vacc1x0x0123, vinput_scale1);
234 vacc1x0x4567 = wasm_f32x4_mul(vacc1x0x4567, vinput_scale1);
235 vacc1x0x89AB = wasm_f32x4_mul(vacc1x0x89AB, vinput_scale1);
236 vacc1x0xCDEF = wasm_f32x4_mul(vacc1x0xCDEF, vinput_scale1);
237 vacc2x0x0123 = wasm_f32x4_mul(vacc2x0x0123, vinput_scale2);
238 vacc2x0x4567 = wasm_f32x4_mul(vacc2x0x4567, vinput_scale2);
239 vacc2x0x89AB = wasm_f32x4_mul(vacc2x0x89AB, vinput_scale2);
240 vacc2x0xCDEF = wasm_f32x4_mul(vacc2x0xCDEF, vinput_scale2);
241 vacc3x0x0123 = wasm_f32x4_mul(vacc3x0x0123, vinput_scale3);
242 vacc3x0x4567 = wasm_f32x4_mul(vacc3x0x4567, vinput_scale3);
243 vacc3x0x89AB = wasm_f32x4_mul(vacc3x0x89AB, vinput_scale3);
244 vacc3x0xCDEF = wasm_f32x4_mul(vacc3x0xCDEF, vinput_scale3);
245
246 const v128_t vfilter_output_scale0123 = wasm_v128_load(w);
247 w = (const float*) w + 4;
248 const v128_t vfilter_output_scale4567 = wasm_v128_load(w);
249 w = (const float*) w + 4;
250 const v128_t vfilter_output_scale89AB = wasm_v128_load(w);
251 w = (const float*) w + 4;
252 const v128_t vfilter_output_scaleCDEF = wasm_v128_load(w);
253 w = (const float*) w + 4;
254 vacc0x0x0123 = wasm_f32x4_mul(vacc0x0x0123, vfilter_output_scale0123);
255 vacc0x0x4567 = wasm_f32x4_mul(vacc0x0x4567, vfilter_output_scale4567);
256 vacc0x0x89AB = wasm_f32x4_mul(vacc0x0x89AB, vfilter_output_scale89AB);
257 vacc0x0xCDEF = wasm_f32x4_mul(vacc0x0xCDEF, vfilter_output_scaleCDEF);
258 vacc1x0x0123 = wasm_f32x4_mul(vacc1x0x0123, vfilter_output_scale0123);
259 vacc1x0x4567 = wasm_f32x4_mul(vacc1x0x4567, vfilter_output_scale4567);
260 vacc1x0x89AB = wasm_f32x4_mul(vacc1x0x89AB, vfilter_output_scale89AB);
261 vacc1x0xCDEF = wasm_f32x4_mul(vacc1x0xCDEF, vfilter_output_scaleCDEF);
262 vacc2x0x0123 = wasm_f32x4_mul(vacc2x0x0123, vfilter_output_scale0123);
263 vacc2x0x4567 = wasm_f32x4_mul(vacc2x0x4567, vfilter_output_scale4567);
264 vacc2x0x89AB = wasm_f32x4_mul(vacc2x0x89AB, vfilter_output_scale89AB);
265 vacc2x0xCDEF = wasm_f32x4_mul(vacc2x0xCDEF, vfilter_output_scaleCDEF);
266 vacc3x0x0123 = wasm_f32x4_mul(vacc3x0x0123, vfilter_output_scale0123);
267 vacc3x0x4567 = wasm_f32x4_mul(vacc3x0x4567, vfilter_output_scale4567);
268 vacc3x0x89AB = wasm_f32x4_mul(vacc3x0x89AB, vfilter_output_scale89AB);
269 vacc3x0xCDEF = wasm_f32x4_mul(vacc3x0xCDEF, vfilter_output_scaleCDEF);
270
271 const v128_t vbias0123 = wasm_v128_load(w);
272 w = (const float*) w + 4;
273 const v128_t vbias4567 = wasm_v128_load(w);
274 w = (const float*) w + 4;
275 const v128_t vbias89AB = wasm_v128_load(w);
276 w = (const float*) w + 4;
277 const v128_t vbiasCDEF = wasm_v128_load(w);
278 w = (const float*) w + 4;
279 vacc0x0x0123 = wasm_f32x4_add(vacc0x0x0123, vbias0123);
280 vacc0x0x4567 = wasm_f32x4_add(vacc0x0x4567, vbias4567);
281 vacc0x0x89AB = wasm_f32x4_add(vacc0x0x89AB, vbias89AB);
282 vacc0x0xCDEF = wasm_f32x4_add(vacc0x0xCDEF, vbiasCDEF);
283 vacc1x0x0123 = wasm_f32x4_add(vacc1x0x0123, vbias0123);
284 vacc1x0x4567 = wasm_f32x4_add(vacc1x0x4567, vbias4567);
285 vacc1x0x89AB = wasm_f32x4_add(vacc1x0x89AB, vbias89AB);
286 vacc1x0xCDEF = wasm_f32x4_add(vacc1x0xCDEF, vbiasCDEF);
287 vacc2x0x0123 = wasm_f32x4_add(vacc2x0x0123, vbias0123);
288 vacc2x0x4567 = wasm_f32x4_add(vacc2x0x4567, vbias4567);
289 vacc2x0x89AB = wasm_f32x4_add(vacc2x0x89AB, vbias89AB);
290 vacc2x0xCDEF = wasm_f32x4_add(vacc2x0xCDEF, vbiasCDEF);
291 vacc3x0x0123 = wasm_f32x4_add(vacc3x0x0123, vbias0123);
292 vacc3x0x4567 = wasm_f32x4_add(vacc3x0x4567, vbias4567);
293 vacc3x0x89AB = wasm_f32x4_add(vacc3x0x89AB, vbias89AB);
294 vacc3x0xCDEF = wasm_f32x4_add(vacc3x0xCDEF, vbiasCDEF);
295
296 vacc0x0x0123 = wasm_f32x4_pmax(vacc0x0x0123, vmin);
297 vacc0x0x4567 = wasm_f32x4_pmax(vacc0x0x4567, vmin);
298 vacc0x0x89AB = wasm_f32x4_pmax(vacc0x0x89AB, vmin);
299 vacc0x0xCDEF = wasm_f32x4_pmax(vacc0x0xCDEF, vmin);
300 vacc1x0x0123 = wasm_f32x4_pmax(vacc1x0x0123, vmin);
301 vacc1x0x4567 = wasm_f32x4_pmax(vacc1x0x4567, vmin);
302 vacc1x0x89AB = wasm_f32x4_pmax(vacc1x0x89AB, vmin);
303 vacc1x0xCDEF = wasm_f32x4_pmax(vacc1x0xCDEF, vmin);
304 vacc2x0x0123 = wasm_f32x4_pmax(vacc2x0x0123, vmin);
305 vacc2x0x4567 = wasm_f32x4_pmax(vacc2x0x4567, vmin);
306 vacc2x0x89AB = wasm_f32x4_pmax(vacc2x0x89AB, vmin);
307 vacc2x0xCDEF = wasm_f32x4_pmax(vacc2x0xCDEF, vmin);
308 vacc3x0x0123 = wasm_f32x4_pmax(vacc3x0x0123, vmin);
309 vacc3x0x4567 = wasm_f32x4_pmax(vacc3x0x4567, vmin);
310 vacc3x0x89AB = wasm_f32x4_pmax(vacc3x0x89AB, vmin);
311 vacc3x0xCDEF = wasm_f32x4_pmax(vacc3x0xCDEF, vmin);
312
313 vacc0x0x0123 = wasm_f32x4_pmin(vacc0x0x0123, vmax);
314 vacc0x0x4567 = wasm_f32x4_pmin(vacc0x0x4567, vmax);
315 vacc0x0x89AB = wasm_f32x4_pmin(vacc0x0x89AB, vmax);
316 vacc0x0xCDEF = wasm_f32x4_pmin(vacc0x0xCDEF, vmax);
317 vacc1x0x0123 = wasm_f32x4_pmin(vacc1x0x0123, vmax);
318 vacc1x0x4567 = wasm_f32x4_pmin(vacc1x0x4567, vmax);
319 vacc1x0x89AB = wasm_f32x4_pmin(vacc1x0x89AB, vmax);
320 vacc1x0xCDEF = wasm_f32x4_pmin(vacc1x0xCDEF, vmax);
321 vacc2x0x0123 = wasm_f32x4_pmin(vacc2x0x0123, vmax);
322 vacc2x0x4567 = wasm_f32x4_pmin(vacc2x0x4567, vmax);
323 vacc2x0x89AB = wasm_f32x4_pmin(vacc2x0x89AB, vmax);
324 vacc2x0xCDEF = wasm_f32x4_pmin(vacc2x0xCDEF, vmax);
325 vacc3x0x0123 = wasm_f32x4_pmin(vacc3x0x0123, vmax);
326 vacc3x0x4567 = wasm_f32x4_pmin(vacc3x0x4567, vmax);
327 vacc3x0x89AB = wasm_f32x4_pmin(vacc3x0x89AB, vmax);
328 vacc3x0xCDEF = wasm_f32x4_pmin(vacc3x0xCDEF, vmax);
329
330 if XNN_LIKELY(nc >= 16) {
331 wasm_v128_store(c0, vacc0x0x0123);
332 wasm_v128_store(c0 + 4, vacc0x0x4567);
333 wasm_v128_store(c0 + 8, vacc0x0x89AB);
334 wasm_v128_store(c0 + 12, vacc0x0xCDEF);
335 wasm_v128_store(c1, vacc1x0x0123);
336 wasm_v128_store(c1 + 4, vacc1x0x4567);
337 wasm_v128_store(c1 + 8, vacc1x0x89AB);
338 wasm_v128_store(c1 + 12, vacc1x0xCDEF);
339 wasm_v128_store(c2, vacc2x0x0123);
340 wasm_v128_store(c2 + 4, vacc2x0x4567);
341 wasm_v128_store(c2 + 8, vacc2x0x89AB);
342 wasm_v128_store(c2 + 12, vacc2x0xCDEF);
343 wasm_v128_store(c3, vacc3x0x0123);
344 wasm_v128_store(c3 + 4, vacc3x0x4567);
345 wasm_v128_store(c3 + 8, vacc3x0x89AB);
346 wasm_v128_store(c3 + 12, vacc3x0xCDEF);
347
348 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
349 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
350 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
351 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
352
353 c0 = (float*) ((uintptr_t) c0 + cn_stride);
354 c1 = (float*) ((uintptr_t) c1 + cn_stride);
355 c2 = (float*) ((uintptr_t) c2 + cn_stride);
356 c3 = (float*) ((uintptr_t) c3 + cn_stride);
357
358 nc -= 16;
359 } else {
360 if (nc & 8) {
361 wasm_v128_store(c0, vacc0x0x0123);
362 vacc0x0x0123 = vacc0x0x89AB;
363 c0 += 4;
364 wasm_v128_store(c0, vacc0x0x4567);
365 vacc0x0x4567 = vacc0x0xCDEF;
366 c0 += 4;
367 wasm_v128_store(c1, vacc1x0x0123);
368 vacc1x0x0123 = vacc1x0x89AB;
369 c1 += 4;
370 wasm_v128_store(c1, vacc1x0x4567);
371 vacc1x0x4567 = vacc1x0xCDEF;
372 c1 += 4;
373 wasm_v128_store(c2, vacc2x0x0123);
374 vacc2x0x0123 = vacc2x0x89AB;
375 c2 += 4;
376 wasm_v128_store(c2, vacc2x0x4567);
377 vacc2x0x4567 = vacc2x0xCDEF;
378 c2 += 4;
379 wasm_v128_store(c3, vacc3x0x0123);
380 vacc3x0x0123 = vacc3x0x89AB;
381 c3 += 4;
382 wasm_v128_store(c3, vacc3x0x4567);
383 vacc3x0x4567 = vacc3x0xCDEF;
384 c3 += 4;
385 }
386 if (nc & 4) {
387 wasm_v128_store(c0, vacc0x0x0123);
388 vacc0x0x0123 = vacc0x0x4567;
389 c0 += 4;
390 wasm_v128_store(c1, vacc1x0x0123);
391 vacc1x0x0123 = vacc1x0x4567;
392 c1 += 4;
393 wasm_v128_store(c2, vacc2x0x0123);
394 vacc2x0x0123 = vacc2x0x4567;
395 c2 += 4;
396 wasm_v128_store(c3, vacc3x0x0123);
397 vacc3x0x0123 = vacc3x0x4567;
398 c3 += 4;
399 }
400 if (nc & 2) {
401 wasm_v128_store64_lane(c0, vacc0x0x0123, 0);
402 vacc0x0x0123 = wasm_v64x2_shuffle(vacc0x0x0123, vacc0x0x0123, 1, 1);
403 c0 += 2;
404 wasm_v128_store64_lane(c1, vacc1x0x0123, 0);
405 vacc1x0x0123 = wasm_v64x2_shuffle(vacc1x0x0123, vacc1x0x0123, 1, 1);
406 c1 += 2;
407 wasm_v128_store64_lane(c2, vacc2x0x0123, 0);
408 vacc2x0x0123 = wasm_v64x2_shuffle(vacc2x0x0123, vacc2x0x0123, 1, 1);
409 c2 += 2;
410 wasm_v128_store64_lane(c3, vacc3x0x0123, 0);
411 vacc3x0x0123 = wasm_v64x2_shuffle(vacc3x0x0123, vacc3x0x0123, 1, 1);
412 c3 += 2;
413 }
414 if (nc & 1) {
415 wasm_v128_store32_lane(c0, vacc0x0x0123, 0);
416 wasm_v128_store32_lane(c1, vacc1x0x0123, 0);
417 wasm_v128_store32_lane(c2, vacc2x0x0123, 0);
418 wasm_v128_store32_lane(c3, vacc3x0x0123, 0);
419 }
420 nc = 0;
421 }
422 } while (nc != 0);
423 }