xref: /aosp_15_r20/external/XNNPACK/src/qs8-dwconv/gen/up16x9-minmax-rndnu-neon-mul8-ld128.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36) 
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-dwconv/unipass-neon-mul8.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 <arm_neon.h>
13 
14 #include <xnnpack/dwconv.h>
15 
16 
xnn_qs8_dwconv_minmax_rndnu_ukernel_up16x9__neon_mul8_ld128(size_t channels,size_t output_width,const int8_t ** input,const void * weights,int8_t * output,size_t input_stride,size_t output_increment,size_t input_offset,const int8_t * zero,const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qs8_dwconv_minmax_rndnu_ukernel_up16x9__neon_mul8_ld128(
18     size_t channels,
19     size_t output_width,
20     const int8_t** input,
21     const void* weights,
22     int8_t* output,
23     size_t input_stride,
24     size_t output_increment,
25     size_t input_offset,
26     const int8_t* zero,
27     const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
28 {
29   assert(channels != 0);
30   assert(output_width != 0);
31 
32   const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
33   const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
34   const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
35   const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
36   const int8x16_t voutput_min = vld1q_dup_s8(&params->rndnu_neon.output_min);
37   const int8x16_t voutput_max = vld1q_dup_s8(&params->rndnu_neon.output_max);
38   do {
39     const int8_t* i0 = input[0];
40     assert(i0 != NULL);
41     if XNN_UNPREDICTABLE(i0 != zero) {
42       i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
43     }
44     const int8_t* i1 = input[1];
45     assert(i1 != NULL);
46     if XNN_UNPREDICTABLE(i1 != zero) {
47       i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
48     }
49     const int8_t* i2 = input[2];
50     assert(i2 != NULL);
51     if XNN_UNPREDICTABLE(i2 != zero) {
52       i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
53     }
54     const int8_t* i3 = input[3];
55     assert(i3 != NULL);
56     if XNN_UNPREDICTABLE(i3 != zero) {
57       i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
58     }
59     const int8_t* i4 = input[4];
60     assert(i4 != NULL);
61     if XNN_UNPREDICTABLE(i4 != zero) {
62       i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
63     }
64     const int8_t* i5 = input[5];
65     assert(i5 != NULL);
66     if XNN_UNPREDICTABLE(i5 != zero) {
67       i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
68     }
69     const int8_t* i6 = input[6];
70     assert(i6 != NULL);
71     if XNN_UNPREDICTABLE(i6 != zero) {
72       i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
73     }
74     const int8_t* i7 = input[7];
75     assert(i7 != NULL);
76     if XNN_UNPREDICTABLE(i7 != zero) {
77       i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
78     }
79     const int8_t* i8 = input[8];
80     assert(i8 != NULL);
81     if XNN_UNPREDICTABLE(i8 != zero) {
82       i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
83     }
84     input = (const int8_t**) ((uintptr_t) input + input_stride);
85 
86     size_t c = channels;
87     const void* w = weights;
88     for (; c >= 16; c -= 16) {
89       int32x4_t vacc0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
90       int32x4_t vacc4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
91       int32x4_t vacc89AB = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
92       int32x4_t vaccCDEF = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
93 
94       const int8x16_t vi0x0123456789ABCDEF = vld1q_s8(i0); i0 += 16;
95       const int8x16_t vk0x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
96 
97       int16x8_t vprod01234567 = vmull_s8(vget_low_s8(vi0x0123456789ABCDEF), vget_low_s8(vk0x0123456789ABCDEF));
98       int16x8_t vprod89ABCDEF = vmull_s8(vget_high_s8(vi0x0123456789ABCDEF), vget_high_s8(vk0x0123456789ABCDEF));
99 
100       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
101       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
102       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
103       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
104       const int8x16_t vi1x0123456789ABCDEF = vld1q_s8(i1); i1 += 16;
105       const int8x16_t vk1x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
106 
107       vprod01234567 = vmull_s8(vget_low_s8(vi1x0123456789ABCDEF), vget_low_s8(vk1x0123456789ABCDEF));
108       vprod89ABCDEF = vmull_s8(vget_high_s8(vi1x0123456789ABCDEF), vget_high_s8(vk1x0123456789ABCDEF));
109 
110       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
111       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
112       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
113       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
114       const int8x16_t vi2x0123456789ABCDEF = vld1q_s8(i2); i2 += 16;
115       const int8x16_t vk2x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
116 
117       vprod01234567 = vmull_s8(vget_low_s8(vi2x0123456789ABCDEF), vget_low_s8(vk2x0123456789ABCDEF));
118       vprod89ABCDEF = vmull_s8(vget_high_s8(vi2x0123456789ABCDEF), vget_high_s8(vk2x0123456789ABCDEF));
119 
120       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
121       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
122       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
123       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
124       const int8x16_t vi3x0123456789ABCDEF = vld1q_s8(i3); i3 += 16;
125       const int8x16_t vk3x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
126 
127       vprod01234567 = vmull_s8(vget_low_s8(vi3x0123456789ABCDEF), vget_low_s8(vk3x0123456789ABCDEF));
128       vprod89ABCDEF = vmull_s8(vget_high_s8(vi3x0123456789ABCDEF), vget_high_s8(vk3x0123456789ABCDEF));
129 
130       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
131       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
132       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
133       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
134       const int8x16_t vi4x0123456789ABCDEF = vld1q_s8(i4); i4 += 16;
135       const int8x16_t vk4x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
136 
137       vprod01234567 = vmull_s8(vget_low_s8(vi4x0123456789ABCDEF), vget_low_s8(vk4x0123456789ABCDEF));
138       vprod89ABCDEF = vmull_s8(vget_high_s8(vi4x0123456789ABCDEF), vget_high_s8(vk4x0123456789ABCDEF));
139 
140       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
141       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
142       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
143       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
144       const int8x16_t vi5x0123456789ABCDEF = vld1q_s8(i5); i5 += 16;
145       const int8x16_t vk5x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
146 
147       vprod01234567 = vmull_s8(vget_low_s8(vi5x0123456789ABCDEF), vget_low_s8(vk5x0123456789ABCDEF));
148       vprod89ABCDEF = vmull_s8(vget_high_s8(vi5x0123456789ABCDEF), vget_high_s8(vk5x0123456789ABCDEF));
149 
150       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
151       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
152       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
153       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
154       const int8x16_t vi6x0123456789ABCDEF = vld1q_s8(i6); i6 += 16;
155       const int8x16_t vk6x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
156 
157       vprod01234567 = vmull_s8(vget_low_s8(vi6x0123456789ABCDEF), vget_low_s8(vk6x0123456789ABCDEF));
158       vprod89ABCDEF = vmull_s8(vget_high_s8(vi6x0123456789ABCDEF), vget_high_s8(vk6x0123456789ABCDEF));
159 
160       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
161       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
162       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
163       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
164       const int8x16_t vi7x0123456789ABCDEF = vld1q_s8(i7); i7 += 16;
165       const int8x16_t vk7x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
166 
167       vprod01234567 = vmull_s8(vget_low_s8(vi7x0123456789ABCDEF), vget_low_s8(vk7x0123456789ABCDEF));
168       vprod89ABCDEF = vmull_s8(vget_high_s8(vi7x0123456789ABCDEF), vget_high_s8(vk7x0123456789ABCDEF));
169 
170       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
171       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
172       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
173       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
174       const int8x16_t vi8x0123456789ABCDEF = vld1q_s8(i8); i8 += 16;
175       const int8x16_t vk8x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
176 
177       vprod01234567 = vmull_s8(vget_low_s8(vi8x0123456789ABCDEF), vget_low_s8(vk8x0123456789ABCDEF));
178       vprod89ABCDEF = vmull_s8(vget_high_s8(vi8x0123456789ABCDEF), vget_high_s8(vk8x0123456789ABCDEF));
179 
180       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
181       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
182       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
183       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
184 
185       vacc0123 = vqshlq_s32(vacc0123, vright_pre_shift);
186       vacc4567 = vqshlq_s32(vacc4567, vright_pre_shift);
187       vacc89AB = vqshlq_s32(vacc89AB, vright_pre_shift);
188       vaccCDEF = vqshlq_s32(vaccCDEF, vright_pre_shift);
189 
190       vacc0123 = vqdmulhq_s32(vacc0123, vmultiplier);
191       vacc4567 = vqdmulhq_s32(vacc4567, vmultiplier);
192       vacc89AB = vqdmulhq_s32(vacc89AB, vmultiplier);
193       vaccCDEF = vqdmulhq_s32(vaccCDEF, vmultiplier);
194 
195       vacc0123 = vrshlq_s32(vacc0123, vright_post_shift);
196       vacc4567 = vrshlq_s32(vacc4567, vright_post_shift);
197       vacc89AB = vrshlq_s32(vacc89AB, vright_post_shift);
198       vaccCDEF = vrshlq_s32(vaccCDEF, vright_post_shift);
199 
200 #if XNN_ARCH_ARM64
201       int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
202       int16x8_t vacc89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc89AB), vaccCDEF);
203 
204       vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
205       vacc89ABCDEF = vqaddq_s16(vacc89ABCDEF, voutput_zero_point);
206 
207       int8x16_t vout0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc01234567), vacc89ABCDEF);
208 #else  // !XNN_ARCH_ARM64
209       int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
210       int16x8_t vacc89ABCDEF = vcombine_s16(vqmovn_s32(vacc89AB), vqmovn_s32(vaccCDEF));
211 
212       vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
213       vacc89ABCDEF = vqaddq_s16(vacc89ABCDEF, voutput_zero_point);
214 
215       int8x16_t vout0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc01234567), vqmovn_s16(vacc89ABCDEF));
216 #endif  // !XNN_ARCH_ARM64
217 
218       vout0123456789ABCDEF = vmaxq_s8(vout0123456789ABCDEF, voutput_min);
219 
220       vout0123456789ABCDEF = vminq_s8(vout0123456789ABCDEF, voutput_max);
221 
222       vst1q_s8(output, vout0123456789ABCDEF); output += 16;
223     }
224     if XNN_UNLIKELY(c != 0) {
225       const int8_t* k = (const int8_t*) ((const int32_t*) w + 16);
226       do {
227         int32x4_t vacc0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
228         int32x4_t vacc4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
229 
230         const int8x8_t vi0x01234567 = vld1_s8(i0); i0 += 8;
231         const int8x8_t vk0x01234567 = vld1_s8(k); k += 8;
232 
233         int16x8_t vprod01234567 = vmull_s8(vi0x01234567, vk0x01234567);
234 
235         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
236         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
237         const int8x8_t vi1x01234567 = vld1_s8(i1); i1 += 8;
238         const int8x8_t vk1x01234567 = vld1_s8((const void*) (k + 8));
239 
240         vprod01234567 = vmull_s8(vi1x01234567, vk1x01234567);
241 
242         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
243         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
244         const int8x8_t vi2x01234567 = vld1_s8(i2); i2 += 8;
245         const int8x8_t vk2x01234567 = vld1_s8((const void*) (k + 24));
246 
247         vprod01234567 = vmull_s8(vi2x01234567, vk2x01234567);
248 
249         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
250         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
251         const int8x8_t vi3x01234567 = vld1_s8(i3); i3 += 8;
252         const int8x8_t vk3x01234567 = vld1_s8((const void*) (k + 40));
253 
254         vprod01234567 = vmull_s8(vi3x01234567, vk3x01234567);
255 
256         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
257         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
258         const int8x8_t vi4x01234567 = vld1_s8(i4); i4 += 8;
259         const int8x8_t vk4x01234567 = vld1_s8((const void*) (k + 56));
260 
261         vprod01234567 = vmull_s8(vi4x01234567, vk4x01234567);
262 
263         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
264         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
265         const int8x8_t vi5x01234567 = vld1_s8(i5); i5 += 8;
266         const int8x8_t vk5x01234567 = vld1_s8((const void*) (k + 72));
267 
268         vprod01234567 = vmull_s8(vi5x01234567, vk5x01234567);
269 
270         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
271         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
272         const int8x8_t vi6x01234567 = vld1_s8(i6); i6 += 8;
273         const int8x8_t vk6x01234567 = vld1_s8((const void*) (k + 88));
274 
275         vprod01234567 = vmull_s8(vi6x01234567, vk6x01234567);
276 
277         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
278         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
279         const int8x8_t vi7x01234567 = vld1_s8(i7); i7 += 8;
280         const int8x8_t vk7x01234567 = vld1_s8((const void*) (k + 104));
281 
282         vprod01234567 = vmull_s8(vi7x01234567, vk7x01234567);
283 
284         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
285         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
286         const int8x8_t vi8x01234567 = vld1_s8(i8); i8 += 8;
287         const int8x8_t vk8x01234567 = vld1_s8((const void*) (k + 120));
288 
289         vprod01234567 = vmull_s8(vi8x01234567, vk8x01234567);
290 
291         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
292         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
293 
294         vacc0123 = vqshlq_s32(vacc0123, vright_pre_shift);
295         vacc4567 = vqshlq_s32(vacc4567, vright_pre_shift);
296 
297         vacc0123 = vqdmulhq_s32(vacc0123, vmultiplier);
298         vacc4567 = vqdmulhq_s32(vacc4567, vmultiplier);
299 
300         vacc0123 = vrshlq_s32(vacc0123, vright_post_shift);
301         vacc4567 = vrshlq_s32(vacc4567, vright_post_shift);
302 
303 #if XNN_ARCH_ARM64
304         int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
305 #else
306         int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
307 #endif
308         vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
309 
310         int8x8_t vout01234567 = vqmovn_s16(vacc01234567);
311         vout01234567 = vmax_s8(vout01234567, vget_low_s8(voutput_min));
312         vout01234567 = vmin_s8(vout01234567, vget_low_s8(voutput_max));
313 
314         if XNN_LIKELY(c >= 8) {
315           vst1_s8(output, vout01234567); output += 8;
316           c -= 8;
317         } else {
318           if (c & 4) {
319             vst1_lane_u32((void*) output, vreinterpret_u32_s8(vout01234567), 0); output += 4;
320             vout01234567 = vext_s8(vout01234567, vout01234567, 4);
321           }
322           if (c & 2) {
323             vst1_lane_u16((void*) output, vreinterpret_u16_s8(vout01234567), 0); output += 2;
324             vout01234567 = vext_s8(vout01234567, vout01234567, 2);
325           }
326           if (c & 1) {
327             vst1_lane_s8(output, vout01234567, 0); output += 1;
328           }
329           c = 0;
330         }
331       } while (c != 0);
332     }
333 
334     output = (int8_t*) ((uintptr_t) output + output_increment);
335   } while (--output_width != 0);
336 }
337