Debian Codegraph @

src/qs8-dwconv/gen/qs8-dwconv-9p8c-minmax-fp32-wasmsimd-mul16-add16.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-dwconv/unipass-wasmsimd-mul16.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/dwconv.h"
15 #include "xnnpack/math.h"
16
17
18 void xnn_qs8_dwconv_minmax_fp32_ukernel_9p8c__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 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29 {
30 assert(channels != 0);
31 assert(output_width != 0);
32
33
34 const v128_t vscale = wasm_v128_load32_splat(&params->fp32_scalar.scale);
35 XNN_FORCE_REALIZATION(vscale);
36 const v128_t vmagic_bias = wasm_f32x4_const_splat(12582912.0f);
37 const int32_t output_min_less_zero_point = (int32_t) params->fp32_scalar.output_min - (int32_t) params->fp32_scalar.output_zero_point;
38 const v128_t vmagic_min = wasm_i32x4_splat((int32_t) float_as_uint32(12582912.0f + output_min_less_zero_point));
39 const v128_t vmagic_bias_less_output_zero_point = wasm_i32x4_splat(INT32_C(0x4B400000) - (int32_t) params->fp32_scalar.output_zero_point);
40 const v128_t voutput_max = wasm_i8x16_splat(params->fp32_scalar.output_max);
41 XNN_FORCE_REALIZATION(vmagic_bias);
42 XNN_FORCE_REALIZATION(vmagic_min);
43 XNN_FORCE_REALIZATION(vmagic_bias_less_output_zero_point);
44 XNN_FORCE_REALIZATION(voutput_max);
45
46 do {
47 const int8_t* i0 = input[0];
48 assert(i0 != NULL);
49 if XNN_UNPREDICTABLE(i0 != zero) {
50 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
51 }
52 const int8_t* i1 = input[1];
53 assert(i1 != NULL);
54 if XNN_UNPREDICTABLE(i1 != zero) {
55 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
56 }
57 const int8_t* i2 = input[2];
58 assert(i2 != NULL);
59 if XNN_UNPREDICTABLE(i2 != zero) {
60 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
61 }
62 const int8_t* i3 = input[3];
63 assert(i3 != NULL);
64 if XNN_UNPREDICTABLE(i3 != zero) {
65 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
66 }
67 const int8_t* i4 = input[4];
68 assert(i4 != NULL);
69 if XNN_UNPREDICTABLE(i4 != zero) {
70 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
71 }
72 const int8_t* i5 = input[5];
73 assert(i5 != NULL);
74 if XNN_UNPREDICTABLE(i5 != zero) {
75 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
76 }
77 const int8_t* i6 = input[6];
78 assert(i6 != NULL);
79 if XNN_UNPREDICTABLE(i6 != zero) {
80 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
81 }
82 const int8_t* i7 = input[7];
83 assert(i7 != NULL);
84 if XNN_UNPREDICTABLE(i7 != zero) {
85 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
86 }
87 const int8_t* i8 = input[8];
88 assert(i8 != NULL);
89 if XNN_UNPREDICTABLE(i8 != zero) {
90 i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
91 }
92 input = (const int8_t**) ((uintptr_t) input + input_stride);
93
94 size_t c = channels;
95 const void* w = weights;
96 for (; c >= 8; c -= 8) {
97 v128_t vacc0123 = wasm_v128_load(w);
98 v128_t vacc4567 = wasm_v128_load((const void*) ((uintptr_t) w + 4 * sizeof(int32_t)));
99
100
101 const v128_t vi0x01234567 = wasm_i16x8_load8x8(i0);
102 const v128_t vk0x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(int8_t)));
103 i0 += 8;
104
105 v128_t vprod01234567 = wasm_i16x8_mul(vi0x01234567, vk0x01234567);
106
107
108 const v128_t vi1x01234567 = wasm_i16x8_load8x8(i1);
109 const v128_t vk1x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(int8_t)));
110 i1 += 8;
111
112 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi1x01234567, vk1x01234567));
113
114 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
115 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
116
117 const v128_t vi2x01234567 = wasm_i16x8_load8x8(i2);
118 const v128_t vk2x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(int8_t)));
119 i2 += 8;
120
121 vprod01234567 = wasm_i16x8_mul(vi2x01234567, vk2x01234567);
122
123
124 const v128_t vi3x01234567 = wasm_i16x8_load8x8(i3);
125 const v128_t vk3x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t)));
126 i3 += 8;
127
128 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi3x01234567, vk3x01234567));
129
130 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
131 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
132
133 const v128_t vi4x01234567 = wasm_i16x8_load8x8(i4);
134 const v128_t vk4x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(int8_t)));
135 i4 += 8;
136
137 vprod01234567 = wasm_i16x8_mul(vi4x01234567, vk4x01234567);
138
139
140 const v128_t vi5x01234567 = wasm_i16x8_load8x8(i5);
141 const v128_t vk5x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(int8_t)));
142 i5 += 8;
143
144 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi5x01234567, vk5x01234567));
145
146 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
147 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
148
149 const v128_t vi6x01234567 = wasm_i16x8_load8x8(i6);
150 const v128_t vk6x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(int8_t)));
151 i6 += 8;
152
153 vprod01234567 = wasm_i16x8_mul(vi6x01234567, vk6x01234567);
154
155
156 const v128_t vi7x01234567 = wasm_i16x8_load8x8(i7);
157 const v128_t vk7x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(int8_t)));
158 i7 += 8;
159
160 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi7x01234567, vk7x01234567));
161
162 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
163 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
164
165 const v128_t vi8x01234567 = wasm_i16x8_load8x8(i8);
166 const v128_t vk8x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(int8_t)));
167 i8 += 8;
168
169 vprod01234567 = wasm_i16x8_mul(vi8x01234567, vk8x01234567);
170
171 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
172 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
173
174
175 w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 72 * sizeof(int8_t));
176
177 vacc0123 = wasm_f32x4_convert_i32x4(vacc0123);
178 vacc4567 = wasm_f32x4_convert_i32x4(vacc4567);
179
180 vacc0123 = wasm_f32x4_mul(vacc0123, vscale);
181 vacc4567 = wasm_f32x4_mul(vacc4567, vscale);
182
183 vacc0123 = wasm_f32x4_add(vacc0123, vmagic_bias);
184 vacc4567 = wasm_f32x4_add(vacc4567, vmagic_bias);
185
186 vacc0123 = wasm_i32x4_max(vacc0123, vmagic_min);
187 vacc4567 = wasm_i32x4_max(vacc4567, vmagic_min);
188
189 vacc0123 = wasm_i32x4_sub(vacc0123, vmagic_bias_less_output_zero_point);
190 vacc4567 = wasm_i32x4_sub(vacc4567, vmagic_bias_less_output_zero_point);
191
192 v128_t vout01234567 = wasm_i16x8_narrow_i32x4(vacc0123, vacc4567);
193
194 v128_t vout0123456701234567 = wasm_i8x16_narrow_i16x8(vout01234567, vout01234567);
195
196 vout0123456701234567 = wasm_i8x16_min(vout0123456701234567, voutput_max);
197
198 wasm_v128_store64_lane(output, vout0123456701234567, 0);
199 output += 8;
200 }
201 if XNN_UNLIKELY(c != 0) {
202 {
203 v128_t vacc0123 = wasm_v128_load(w);
204 v128_t vacc4567 = wasm_v128_load((const void*) ((uintptr_t) w + 4 * sizeof(int32_t)));
205
206
207 const v128_t vi0x01234567 = wasm_i16x8_load8x8(i0);
208 const v128_t vk0x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(int8_t)));
209
210 v128_t vprod01234567 = wasm_i16x8_mul(vi0x01234567, vk0x01234567);
211
212
213 const v128_t vi1x01234567 = wasm_i16x8_load8x8(i1);
214 const v128_t vk1x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(int8_t)));
215
216 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi1x01234567, vk1x01234567));
217
218 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
219 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
220
221 const v128_t vi2x01234567 = wasm_i16x8_load8x8(i2);
222 const v128_t vk2x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(int8_t)));
223
224 vprod01234567 = wasm_i16x8_mul(vi2x01234567, vk2x01234567);
225
226
227 const v128_t vi3x01234567 = wasm_i16x8_load8x8(i3);
228 const v128_t vk3x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t)));
229
230 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi3x01234567, vk3x01234567));
231
232 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
233 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
234
235 const v128_t vi4x01234567 = wasm_i16x8_load8x8(i4);
236 const v128_t vk4x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(int8_t)));
237
238 vprod01234567 = wasm_i16x8_mul(vi4x01234567, vk4x01234567);
239
240
241 const v128_t vi5x01234567 = wasm_i16x8_load8x8(i5);
242 const v128_t vk5x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(int8_t)));
243
244 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi5x01234567, vk5x01234567));
245
246 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
247 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
248
249 const v128_t vi6x01234567 = wasm_i16x8_load8x8(i6);
250 const v128_t vk6x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(int8_t)));
251
252 vprod01234567 = wasm_i16x8_mul(vi6x01234567, vk6x01234567);
253
254
255 const v128_t vi7x01234567 = wasm_i16x8_load8x8(i7);
256 const v128_t vk7x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(int8_t)));
257
258 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi7x01234567, vk7x01234567));
259
260 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
261 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
262
263 const v128_t vi8x01234567 = wasm_i16x8_load8x8(i8);
264 const v128_t vk8x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(int8_t)));
265
266 vprod01234567 = wasm_i16x8_mul(vi8x01234567, vk8x01234567);
267
268 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
269 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
270
271
272
273 vacc0123 = wasm_f32x4_convert_i32x4(vacc0123);
274 vacc4567 = wasm_f32x4_convert_i32x4(vacc4567);
275
276 vacc0123 = wasm_f32x4_mul(vacc0123, vscale);
277 vacc4567 = wasm_f32x4_mul(vacc4567, vscale);
278
279 vacc0123 = wasm_f32x4_add(vacc0123, vmagic_bias);
280 vacc4567 = wasm_f32x4_add(vacc4567, vmagic_bias);
281
282 vacc0123 = wasm_i32x4_max(vacc0123, vmagic_min);
283 vacc4567 = wasm_i32x4_max(vacc4567, vmagic_min);
284
285 vacc0123 = wasm_i32x4_sub(vacc0123, vmagic_bias_less_output_zero_point);
286 vacc4567 = wasm_i32x4_sub(vacc4567, vmagic_bias_less_output_zero_point);
287
288 v128_t vout01234567 = wasm_i16x8_narrow_i32x4(vacc0123, vacc4567);
289 v128_t vout0123456701234567 = wasm_i8x16_narrow_i16x8(vout01234567, vout01234567);
290
291 vout0123456701234567 = wasm_i8x16_min(vout0123456701234567, voutput_max);
292
293
294 if (c & 4) {
295 wasm_v128_store32_lane(output, vout0123456701234567, 0);
296 vout0123456701234567 = wasm_u64x2_shr(vout0123456701234567, 32);
297 output += 4;
298 }
299 if (c & 2) {
300 wasm_v128_store16_lane(output, vout0123456701234567, 0);
301 vout0123456701234567 = wasm_u32x4_shr(vout0123456701234567, 16);
302 output += 2;
303 }
304 if (c & 1) {
305 wasm_v128_store8_lane(output, vout0123456701234567, 0);
306 output += 1;
307 }
308 }
309 }
310
311 output = (int8_t*) ((uintptr_t) output + output_increment);
312 } while (--output_width != 0);
313 }