Debian Codegraph @

src/qs8-dwconv/gen/qs8-dwconv-5f5m5l16c8s8r-minmax-fp32-wasmsimd-mul16-add16.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-dwconv/multipass-wasmsimd-mul16.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/dwconv.h"
15 #include "xnnpack/math.h"
16
17
18 void xnn_qs8_dwconv_minmax_fp32_ukernel_5f5m5l16c8s8r__wasmsimd_mul16_add16(
19 size_t channels,
20 size_t output_width,
21 const int8_t** input,
22 const void* weights,
23 int8_t* output,
24 intptr_t input_stride,
25 size_t output_increment,
26 size_t input_offset,
27 const int8_t* zero,
28 size_t kernel_size,
29 int32_t* buffer,
30 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
31 {
32 assert(channels != 0);
33 assert(output_width != 0);
34 assert(kernel_size > 5);
35
36
37 const v128_t vscale = wasm_v128_load32_splat(&params->fp32_scalar.scale);
38 XNN_FORCE_REALIZATION(vscale);
39
40 const v128_t vmagic_bias = wasm_f32x4_const_splat(12582912.0f);
41 const int32_t output_min_less_zero_point = (int32_t) params->fp32_scalar.output_min - (int32_t) params->fp32_scalar.output_zero_point;
42 const v128_t vmagic_min = wasm_i32x4_splat((int32_t) float_as_uint32(12582912.0f + output_min_less_zero_point));
43 const v128_t vmagic_bias_less_output_zero_point = wasm_i32x4_splat(INT32_C(0x4B400000) - (int32_t) params->fp32_scalar.output_zero_point);
44 const v128_t voutput_max = wasm_i8x16_splat(params->fp32_scalar.output_max);
45 XNN_FORCE_REALIZATION(vmagic_bias);
46 XNN_FORCE_REALIZATION(vmagic_min);
47 XNN_FORCE_REALIZATION(vmagic_bias_less_output_zero_point);
48 XNN_FORCE_REALIZATION(voutput_max);
49
50 do {
51 const void* w = weights;
52
53 // First pass to process 5 inputs.
54 {
55 int32_t* b = buffer;
56 const int8_t* i0 = input[0];
57 assert(i0 != NULL);
58 if XNN_UNPREDICTABLE(i0 != zero) {
59 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
60 }
61 const int8_t* i1 = input[1];
62 assert(i1 != NULL);
63 if XNN_UNPREDICTABLE(i1 != zero) {
64 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
65 }
66 const int8_t* i2 = input[2];
67 assert(i2 != NULL);
68 if XNN_UNPREDICTABLE(i2 != zero) {
69 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
70 }
71 const int8_t* i3 = input[3];
72 assert(i3 != NULL);
73 if XNN_UNPREDICTABLE(i3 != zero) {
74 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
75 }
76 const int8_t* i4 = input[4];
77 assert(i4 != NULL);
78 if XNN_UNPREDICTABLE(i4 != zero) {
79 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
80 }
81 input += 5;
82
83 size_t c = round_up_po2(channels, 8);
84
85 for (; c >= 16; c -= 16) {
86 v128_t vacc0123 = wasm_v128_load(w);
87 v128_t vacc4567 = wasm_v128_load((const void*) ((uintptr_t) w + 4 * sizeof(int32_t)));
88 v128_t vacc89AB = wasm_v128_load((const void*) ((uintptr_t) w + 8 * sizeof(int32_t)));
89 v128_t vaccCDEF = wasm_v128_load((const void*) ((uintptr_t) w + 12 * sizeof(int32_t)));
90
91 const v128_t vi0x01234567 = wasm_i16x8_load8x8(i0);
92 v128_t vk0x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 0 * sizeof(int8_t)));
93 const v128_t vi0x89ABCDEF = wasm_i16x8_load8x8(i0 + 8);
94 v128_t vk0x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 8 * sizeof(int8_t)));
95 i0 += 16;
96
97 v128_t vprod01234567 = wasm_i16x8_mul(vi0x01234567, vk0x01234567);
98 v128_t vprod89ABCDEF = wasm_i16x8_mul(vi0x89ABCDEF, vk0x89ABCDEF);
99
100
101 const v128_t vi1x01234567 = wasm_i16x8_load8x8(i1);
102 v128_t vk1x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 16 * sizeof(int8_t)));
103 const v128_t vi1x89ABCDEF = wasm_i16x8_load8x8(i1 + 8);
104 v128_t vk1x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 24 * sizeof(int8_t)));
105 i1 += 16;
106
107 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi1x01234567, vk1x01234567));
108 vprod89ABCDEF = wasm_i16x8_add(vprod89ABCDEF, wasm_i16x8_mul(vi1x89ABCDEF, vk1x89ABCDEF));
109
110 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
111 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
112 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
113 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
114
115 const v128_t vi2x01234567 = wasm_i16x8_load8x8(i2);
116 v128_t vk2x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 32 * sizeof(int8_t)));
117 const v128_t vi2x89ABCDEF = wasm_i16x8_load8x8(i2 + 8);
118 v128_t vk2x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 40 * sizeof(int8_t)));
119 i2 += 16;
120
121 vprod01234567 = wasm_i16x8_mul(vi2x01234567, vk2x01234567);
122 vprod89ABCDEF = wasm_i16x8_mul(vi2x89ABCDEF, vk2x89ABCDEF);
123
124
125 const v128_t vi3x01234567 = wasm_i16x8_load8x8(i3);
126 v128_t vk3x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 48 * sizeof(int8_t)));
127 const v128_t vi3x89ABCDEF = wasm_i16x8_load8x8(i3 + 8);
128 v128_t vk3x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 56 * sizeof(int8_t)));
129 i3 += 16;
130
131 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi3x01234567, vk3x01234567));
132 vprod89ABCDEF = wasm_i16x8_add(vprod89ABCDEF, wasm_i16x8_mul(vi3x89ABCDEF, vk3x89ABCDEF));
133
134 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
135 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
136 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
137 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
138
139 const v128_t vi4x01234567 = wasm_i16x8_load8x8(i4);
140 v128_t vk4x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 64 * sizeof(int8_t)));
141 const v128_t vi4x89ABCDEF = wasm_i16x8_load8x8(i4 + 8);
142 v128_t vk4x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 72 * sizeof(int8_t)));
143 i4 += 16;
144
145 vprod01234567 = wasm_i16x8_mul(vi4x01234567, vk4x01234567);
146 vprod89ABCDEF = wasm_i16x8_mul(vi4x89ABCDEF, vk4x89ABCDEF);
147
148 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
149 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
150 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
151 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
152
153 w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 80 * sizeof(int8_t));
154
155 wasm_v128_store(b, vacc0123);
156 wasm_v128_store(b + 4, vacc4567);
157 wasm_v128_store(b + 8, vacc89AB);
158 wasm_v128_store(b + 12, vaccCDEF);
159 b += 16;
160 }
161
162 if XNN_UNLIKELY(c != 0) {
163 do {
164 v128_t vacc0123 = wasm_v128_load(w);
165 v128_t vacc4567 = wasm_v128_load((const int32_t*) w + 4);
166
167 const v128_t vi0x01234567 = wasm_i16x8_load8x8(i0);
168 v128_t vk0x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t)));
169 i0 += 8;
170
171 v128_t vprod01234567 = wasm_i16x8_mul(vi0x01234567, vk0x01234567);
172
173
174 const v128_t vi1x01234567 = wasm_i16x8_load8x8(i1);
175 v128_t vk1x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(int8_t)));
176 i1 += 8;
177
178 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi1x01234567, vk1x01234567));
179
180 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
181 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
182
183 const v128_t vi2x01234567 = wasm_i16x8_load8x8(i2);
184 v128_t vk2x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(int8_t)));
185 i2 += 8;
186
187 vprod01234567 = wasm_i16x8_mul(vi2x01234567, vk2x01234567);
188
189
190 const v128_t vi3x01234567 = wasm_i16x8_load8x8(i3);
191 v128_t vk3x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t)));
192 i3 += 8;
193
194 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi3x01234567, vk3x01234567));
195
196 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
197 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
198
199 const v128_t vi4x01234567 = wasm_i16x8_load8x8(i4);
200 v128_t vk4x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(int8_t)));
201 i4 += 8;
202
203 vprod01234567 = wasm_i16x8_mul(vi4x01234567, vk4x01234567);
204
205 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
206 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
207
208
209 w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(int8_t));
210
211 wasm_v128_store(b, vacc0123);
212 wasm_v128_store(b + 4, vacc4567);
213 b += 8;
214 c -= 8;
215 } while (c != 0);
216 }
217 }
218
219
220 // Middle pass to process 5 inputs in each iteration.
221 for (size_t ks = kernel_size - 5; ks > 5; ks -= 5) {
222 int32_t* b = buffer;
223 const int8_t* i0 = input[0];
224 assert(i0 != NULL);
225 if XNN_UNPREDICTABLE(i0 != zero) {
226 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
227 }
228 const int8_t* i1 = input[1];
229 assert(i1 != NULL);
230 if XNN_UNPREDICTABLE(i1 != zero) {
231 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
232 }
233 const int8_t* i2 = input[2];
234 assert(i2 != NULL);
235 if XNN_UNPREDICTABLE(i2 != zero) {
236 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
237 }
238 const int8_t* i3 = input[3];
239 assert(i3 != NULL);
240 if XNN_UNPREDICTABLE(i3 != zero) {
241 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
242 }
243 const int8_t* i4 = input[4];
244 assert(i4 != NULL);
245 if XNN_UNPREDICTABLE(i4 != zero) {
246 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
247 }
248 input += 5;
249
250 size_t c = round_up_po2(channels, 8);
251
252 for (; c >= 16; c -= 16) {
253 v128_t vacc0123 = wasm_v128_load(b);
254 v128_t vacc4567 = wasm_v128_load(b + 4);
255 v128_t vacc89AB = wasm_v128_load(b + 8);
256 v128_t vaccCDEF = wasm_v128_load(b + 12);
257
258 const v128_t vi0x01234567 = wasm_i16x8_load8x8(i0);
259 v128_t vk0x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 0 * sizeof(int8_t)));
260 const v128_t vi0x89ABCDEF = wasm_i16x8_load8x8(i0 + 8);
261 v128_t vk0x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int8_t)));
262 i0 += 16;
263
264 v128_t vprod01234567 = wasm_i16x8_mul(vi0x01234567, vk0x01234567);
265 v128_t vprod89ABCDEF = wasm_i16x8_mul(vi0x89ABCDEF, vk0x89ABCDEF);
266
267
268 const v128_t vi1x01234567 = wasm_i16x8_load8x8(i1);
269 v128_t vk1x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int8_t)));
270 const v128_t vi1x89ABCDEF = wasm_i16x8_load8x8(i1 + 8);
271 v128_t vk1x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int8_t)));
272 i1 += 16;
273
274 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi1x01234567, vk1x01234567));
275 vprod89ABCDEF = wasm_i16x8_add(vprod89ABCDEF, wasm_i16x8_mul(vi1x89ABCDEF, vk1x89ABCDEF));
276
277 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
278 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
279 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
280 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
281
282 const v128_t vi2x01234567 = wasm_i16x8_load8x8(i2);
283 v128_t vk2x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 32 * sizeof(int8_t)));
284 const v128_t vi2x89ABCDEF = wasm_i16x8_load8x8(i2 + 8);
285 v128_t vk2x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 40 * sizeof(int8_t)));
286 i2 += 16;
287
288 vprod01234567 = wasm_i16x8_mul(vi2x01234567, vk2x01234567);
289 vprod89ABCDEF = wasm_i16x8_mul(vi2x89ABCDEF, vk2x89ABCDEF);
290
291
292 const v128_t vi3x01234567 = wasm_i16x8_load8x8(i3);
293 v128_t vk3x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 48 * sizeof(int8_t)));
294 const v128_t vi3x89ABCDEF = wasm_i16x8_load8x8(i3 + 8);
295 v128_t vk3x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 56 * sizeof(int8_t)));
296 i3 += 16;
297
298 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi3x01234567, vk3x01234567));
299 vprod89ABCDEF = wasm_i16x8_add(vprod89ABCDEF, wasm_i16x8_mul(vi3x89ABCDEF, vk3x89ABCDEF));
300
301 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
302 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
303 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
304 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
305
306 const v128_t vi4x01234567 = wasm_i16x8_load8x8(i4);
307 v128_t vk4x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 64 * sizeof(int8_t)));
308 const v128_t vi4x89ABCDEF = wasm_i16x8_load8x8(i4 + 8);
309 v128_t vk4x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 72 * sizeof(int8_t)));
310 i4 += 16;
311
312 vprod01234567 = wasm_i16x8_mul(vi4x01234567, vk4x01234567);
313 vprod89ABCDEF = wasm_i16x8_mul(vi4x89ABCDEF, vk4x89ABCDEF);
314
315 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
316 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
317 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
318 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
319
320 w = (const void*) ((uintptr_t) w + 80 * sizeof(int8_t));
321
322 wasm_v128_store(b, vacc0123);
323 wasm_v128_store(b + 4, vacc4567);
324 wasm_v128_store(b + 8, vacc89AB);
325 wasm_v128_store(b + 12, vaccCDEF);
326 b += 16;
327 }
328
329 if XNN_UNLIKELY(c != 0) {
330 do {
331 v128_t vacc0123 = wasm_v128_load(b);
332 v128_t vacc4567 = wasm_v128_load(b + 4);
333
334 const v128_t vi0x01234567 = wasm_i16x8_load8x8(i0);
335 v128_t vk0x01234567 = wasm_i16x8_load8x8(w);
336 i0 += 8;
337
338 v128_t vprod01234567 = wasm_i16x8_mul(vi0x01234567, vk0x01234567);
339
340
341 const v128_t vi1x01234567 = wasm_i16x8_load8x8(i1);
342 v128_t vk1x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int8_t)));
343 i1 += 8;
344
345 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi1x01234567, vk1x01234567));
346
347 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
348 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
349
350 const v128_t vi2x01234567 = wasm_i16x8_load8x8(i2);
351 v128_t vk2x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int8_t)));
352 i2 += 8;
353
354 vprod01234567 = wasm_i16x8_mul(vi2x01234567, vk2x01234567);
355
356
357 const v128_t vi3x01234567 = wasm_i16x8_load8x8(i3);
358 v128_t vk3x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int8_t)));
359 i3 += 8;
360
361 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi3x01234567, vk3x01234567));
362
363 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
364 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
365
366 const v128_t vi4x01234567 = wasm_i16x8_load8x8(i4);
367 v128_t vk4x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 32 * sizeof(int8_t)));
368 i4 += 8;
369
370 vprod01234567 = wasm_i16x8_mul(vi4x01234567, vk4x01234567);
371
372 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
373 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
374
375
376 w = (const void*) ((uintptr_t) w + 40 * sizeof(int8_t));
377
378 wasm_v128_store(b, vacc0123);
379 wasm_v128_store(b + 4, vacc4567);
380 b += 8;
381 c -= 8;
382 } while (c != 0);
383 }
384 }
385
386 // Last pass to process up to 5 inputs.
387 {
388 const int32_t* b = buffer;
389 const int8_t* i0 = input[0];
390 assert(i0 != NULL);
391 if XNN_UNPREDICTABLE(i0 != zero) {
392 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
393 }
394 const int8_t* i1 = input[1];
395 assert(i1 != NULL);
396 if XNN_UNPREDICTABLE(i1 != zero) {
397 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
398 }
399 const int8_t* i2 = input[2];
400 assert(i2 != NULL);
401 if XNN_UNPREDICTABLE(i2 != zero) {
402 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
403 }
404 const int8_t* i3 = input[3];
405 assert(i3 != NULL);
406 if XNN_UNPREDICTABLE(i3 != zero) {
407 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
408 }
409 const int8_t* i4 = input[4];
410 assert(i4 != NULL);
411 if XNN_UNPREDICTABLE(i4 != zero) {
412 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
413 }
414
415 size_t c = channels;
416
417 for (; c >= 16; c -= 16) {
418 v128_t vacc0123 = wasm_v128_load(b);
419 v128_t vacc4567 = wasm_v128_load(b + 4);
420 v128_t vacc89AB = wasm_v128_load(b + 8);
421 v128_t vaccCDEF = wasm_v128_load(b + 12);
422 b += 16;
423
424 const v128_t vi0x01234567 = wasm_i16x8_load8x8(i0);
425 v128_t vk0x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 0 * sizeof(int8_t)));
426 const v128_t vi0x89ABCDEF = wasm_i16x8_load8x8(i0 + 8);
427 v128_t vk0x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int8_t)));
428 i0 += 16;
429
430 v128_t vprod01234567 = wasm_i16x8_mul(vi0x01234567, vk0x01234567);
431 v128_t vprod89ABCDEF = wasm_i16x8_mul(vi0x89ABCDEF, vk0x89ABCDEF);
432
433
434 const v128_t vi1x01234567 = wasm_i16x8_load8x8(i1);
435 v128_t vk1x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int8_t)));
436 const v128_t vi1x89ABCDEF = wasm_i16x8_load8x8(i1 + 8);
437 v128_t vk1x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int8_t)));
438 i1 += 16;
439
440 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi1x01234567, vk1x01234567));
441 vprod89ABCDEF = wasm_i16x8_add(vprod89ABCDEF, wasm_i16x8_mul(vi1x89ABCDEF, vk1x89ABCDEF));
442
443 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
444 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
445 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
446 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
447
448 const v128_t vi2x01234567 = wasm_i16x8_load8x8(i2);
449 v128_t vk2x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 32 * sizeof(int8_t)));
450 const v128_t vi2x89ABCDEF = wasm_i16x8_load8x8(i2 + 8);
451 v128_t vk2x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 40 * sizeof(int8_t)));
452 i2 += 16;
453
454 vprod01234567 = wasm_i16x8_mul(vi2x01234567, vk2x01234567);
455 vprod89ABCDEF = wasm_i16x8_mul(vi2x89ABCDEF, vk2x89ABCDEF);
456
457
458 const v128_t vi3x01234567 = wasm_i16x8_load8x8(i3);
459 v128_t vk3x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 48 * sizeof(int8_t)));
460 const v128_t vi3x89ABCDEF = wasm_i16x8_load8x8(i3 + 8);
461 v128_t vk3x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 56 * sizeof(int8_t)));
462 i3 += 16;
463
464 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi3x01234567, vk3x01234567));
465 vprod89ABCDEF = wasm_i16x8_add(vprod89ABCDEF, wasm_i16x8_mul(vi3x89ABCDEF, vk3x89ABCDEF));
466
467 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
468 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
469 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
470 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
471
472 const v128_t vi4x01234567 = wasm_i16x8_load8x8(i4);
473 v128_t vk4x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 64 * sizeof(int8_t)));
474 const v128_t vi4x89ABCDEF = wasm_i16x8_load8x8(i4 + 8);
475 v128_t vk4x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 72 * sizeof(int8_t)));
476 i4 += 16;
477
478 vprod01234567 = wasm_i16x8_mul(vi4x01234567, vk4x01234567);
479 vprod89ABCDEF = wasm_i16x8_mul(vi4x89ABCDEF, vk4x89ABCDEF);
480
481 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
482 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
483 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
484 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
485
486
487 w = (const void*) ((uintptr_t) w + 80 * sizeof(int8_t));
488
489 vacc0123 = wasm_f32x4_convert_i32x4(vacc0123);
490 vacc4567 = wasm_f32x4_convert_i32x4(vacc4567);
491 vacc89AB = wasm_f32x4_convert_i32x4(vacc89AB);
492 vaccCDEF = wasm_f32x4_convert_i32x4(vaccCDEF);
493
494 vacc0123 = wasm_f32x4_mul(vacc0123, vscale);
495 vacc4567 = wasm_f32x4_mul(vacc4567, vscale);
496 vacc89AB = wasm_f32x4_mul(vacc89AB, vscale);
497 vaccCDEF = wasm_f32x4_mul(vaccCDEF, vscale);
498
499 vacc0123 = wasm_f32x4_add(vacc0123, vmagic_bias);
500 vacc4567 = wasm_f32x4_add(vacc4567, vmagic_bias);
501 vacc89AB = wasm_f32x4_add(vacc89AB, vmagic_bias);
502 vaccCDEF = wasm_f32x4_add(vaccCDEF, vmagic_bias);
503
504 vacc0123 = wasm_i32x4_max(vacc0123, vmagic_min);
505 vacc4567 = wasm_i32x4_max(vacc4567, vmagic_min);
506 vacc89AB = wasm_i32x4_max(vacc89AB, vmagic_min);
507 vaccCDEF = wasm_i32x4_max(vaccCDEF, vmagic_min);
508
509 vacc0123 = wasm_i32x4_sub(vacc0123, vmagic_bias_less_output_zero_point);
510 vacc4567 = wasm_i32x4_sub(vacc4567, vmagic_bias_less_output_zero_point);
511 vacc89AB = wasm_i32x4_sub(vacc89AB, vmagic_bias_less_output_zero_point);
512 vaccCDEF = wasm_i32x4_sub(vaccCDEF, vmagic_bias_less_output_zero_point);
513
514 v128_t vout01234567 = wasm_i16x8_narrow_i32x4(vacc0123, vacc4567);
515 v128_t vout89ABCDEF = wasm_i16x8_narrow_i32x4(vacc89AB, vaccCDEF);
516
517 v128_t vout0123456789ABCDEF = wasm_i8x16_narrow_i16x8(vout01234567, vout89ABCDEF);
518
519 vout0123456789ABCDEF = wasm_i8x16_min(vout0123456789ABCDEF, voutput_max);
520
521 wasm_v128_store(output, vout0123456789ABCDEF);
522 output += 16;
523 }
524
525 if XNN_UNLIKELY(c != 0) {
526 do {
527 v128_t vacc0123 = wasm_v128_load(b);
528 v128_t vacc4567 = wasm_v128_load(b + 4);
529 b += 8;
530
531 const v128_t vi0x01234567 = wasm_i16x8_load8x8(i0);
532 v128_t vk0x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w));
533 i0 += 8;
534
535 v128_t vprod01234567 = wasm_i16x8_mul(vi0x01234567, vk0x01234567);
536
537
538 const v128_t vi1x01234567 = wasm_i16x8_load8x8(i1);
539 v128_t vk1x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int8_t)));
540 i1 += 8;
541
542 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi1x01234567, vk1x01234567));
543
544 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
545 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
546
547 const v128_t vi2x01234567 = wasm_i16x8_load8x8(i2);
548 v128_t vk2x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int8_t)));
549 i2 += 8;
550
551 vprod01234567 = wasm_i16x8_mul(vi2x01234567, vk2x01234567);
552
553
554 const v128_t vi3x01234567 = wasm_i16x8_load8x8(i3);
555 v128_t vk3x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int8_t)));
556 i3 += 8;
557
558 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi3x01234567, vk3x01234567));
559
560 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
561 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
562
563 const v128_t vi4x01234567 = wasm_i16x8_load8x8(i4);
564 v128_t vk4x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 32 * sizeof(int8_t)));
565 i4 += 8;
566
567 vprod01234567 = wasm_i16x8_mul(vi4x01234567, vk4x01234567);
568
569 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
570 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
571
572
573 vacc0123 = wasm_f32x4_convert_i32x4(vacc0123);
574 vacc4567 = wasm_f32x4_convert_i32x4(vacc4567);
575
576 vacc0123 = wasm_f32x4_mul(vacc0123, vscale);
577 vacc4567 = wasm_f32x4_mul(vacc4567, vscale);
578
579 w = (void*) ((uintptr_t) w + 40);
580
581 vacc0123 = wasm_f32x4_add(vacc0123, vmagic_bias);
582 vacc4567 = wasm_f32x4_add(vacc4567, vmagic_bias);
583
584 vacc0123 = wasm_i32x4_max(vacc0123, vmagic_min);
585 vacc4567 = wasm_i32x4_max(vacc4567, vmagic_min);
586
587 vacc0123 = wasm_i32x4_sub(vacc0123, vmagic_bias_less_output_zero_point);
588 vacc4567 = wasm_i32x4_sub(vacc4567, vmagic_bias_less_output_zero_point);
589
590 v128_t vout01234567 = wasm_i16x8_narrow_i32x4(vacc0123, vacc4567);
591 v128_t vout0123456701234567 = wasm_i8x16_narrow_i16x8(vout01234567, vout01234567);
592
593 vout0123456701234567 = wasm_i8x16_min(vout0123456701234567, voutput_max);
594
595 if XNN_LIKELY(c >= 8) {
596 wasm_v128_store64_lane(output, vout0123456701234567, 0);
597 output += 8;
598 c -= 8;
599 } else {
600 if (c & 4) {
601 wasm_v128_store32_lane(output, vout0123456701234567, 0);
602 vout0123456701234567 = wasm_u64x2_shr(vout0123456701234567, 32);
603 output += 4;
604 }
605 if (c & 2) {
606 wasm_v128_store16_lane(output, vout0123456701234567, 0);
607 vout0123456701234567 = wasm_u32x4_shr(vout0123456701234567, 16);
608 output += 2;
609 }
610 if (c & 1) {
611 wasm_v128_store8_lane(output, vout0123456701234567, 0);
612 output += 1;
613 }
614 c = 0;
615 }
616 } while (c != 0);
617 }
618 }
619
620 input = (const int8_t**) ((uintptr_t) input + input_stride);
621 output = (int8_t*) ((uintptr_t) output + output_increment);
622 } while (--output_width != 0);
623 }