xref: /aosp_15_r20/external/XNNPACK/src/qs8-dwconv/gen/up16x9-minmax-rndnu-neon-mla8-ld64.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_mla8_ld64(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_mla8_ld64(
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 int8x8_t vi0x01234567 = vld1_s8(i0); i0 += 8;
95       const int8x8_t vk0x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
96       const int8x8_t vi0x89ABCDEF = vld1_s8(i0); i0 += 8;
97       const int8x8_t vk0x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
98 
99       int16x8_t vprod01234567 = vmull_s8(vi0x01234567, vk0x01234567);
100       int16x8_t vprod89ABCDEF = vmull_s8(vi0x89ABCDEF, vk0x89ABCDEF);
101 
102       const int8x8_t vi1x01234567 = vld1_s8(i1); i1 += 8;
103       const int8x8_t vk1x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
104       const int8x8_t vi1x89ABCDEF = vld1_s8(i1); i1 += 8;
105       const int8x8_t vk1x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
106 
107       vprod01234567 = vmlal_s8(vprod01234567, vi1x01234567, vk1x01234567);
108       vprod89ABCDEF = vmlal_s8(vprod89ABCDEF, vi1x89ABCDEF, vk1x89ABCDEF);
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 int8x8_t vi2x01234567 = vld1_s8(i2); i2 += 8;
115       const int8x8_t vk2x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
116       const int8x8_t vi2x89ABCDEF = vld1_s8(i2); i2 += 8;
117       const int8x8_t vk2x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
118 
119       vprod01234567 = vmull_s8(vi2x01234567, vk2x01234567);
120       vprod89ABCDEF = vmull_s8(vi2x89ABCDEF, vk2x89ABCDEF);
121 
122       const int8x8_t vi3x01234567 = vld1_s8(i3); i3 += 8;
123       const int8x8_t vk3x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
124       const int8x8_t vi3x89ABCDEF = vld1_s8(i3); i3 += 8;
125       const int8x8_t vk3x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
126 
127       vprod01234567 = vmlal_s8(vprod01234567, vi3x01234567, vk3x01234567);
128       vprod89ABCDEF = vmlal_s8(vprod89ABCDEF, vi3x89ABCDEF, vk3x89ABCDEF);
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 int8x8_t vi4x01234567 = vld1_s8(i4); i4 += 8;
135       const int8x8_t vk4x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
136       const int8x8_t vi4x89ABCDEF = vld1_s8(i4); i4 += 8;
137       const int8x8_t vk4x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
138 
139       vprod01234567 = vmull_s8(vi4x01234567, vk4x01234567);
140       vprod89ABCDEF = vmull_s8(vi4x89ABCDEF, vk4x89ABCDEF);
141 
142       const int8x8_t vi5x01234567 = vld1_s8(i5); i5 += 8;
143       const int8x8_t vk5x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
144       const int8x8_t vi5x89ABCDEF = vld1_s8(i5); i5 += 8;
145       const int8x8_t vk5x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
146 
147       vprod01234567 = vmlal_s8(vprod01234567, vi5x01234567, vk5x01234567);
148       vprod89ABCDEF = vmlal_s8(vprod89ABCDEF, vi5x89ABCDEF, vk5x89ABCDEF);
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 int8x8_t vi6x01234567 = vld1_s8(i6); i6 += 8;
155       const int8x8_t vk6x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
156       const int8x8_t vi6x89ABCDEF = vld1_s8(i6); i6 += 8;
157       const int8x8_t vk6x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
158 
159       vprod01234567 = vmull_s8(vi6x01234567, vk6x01234567);
160       vprod89ABCDEF = vmull_s8(vi6x89ABCDEF, vk6x89ABCDEF);
161 
162       const int8x8_t vi7x01234567 = vld1_s8(i7); i7 += 8;
163       const int8x8_t vk7x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
164       const int8x8_t vi7x89ABCDEF = vld1_s8(i7); i7 += 8;
165       const int8x8_t vk7x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
166 
167       vprod01234567 = vmlal_s8(vprod01234567, vi7x01234567, vk7x01234567);
168       vprod89ABCDEF = vmlal_s8(vprod89ABCDEF, vi7x89ABCDEF, vk7x89ABCDEF);
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 int8x8_t vi8x01234567 = vld1_s8(i8); i8 += 8;
175       const int8x8_t vk8x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
176       const int8x8_t vi8x89ABCDEF = vld1_s8(i8); i8 += 8;
177       const int8x8_t vk8x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
178 
179       vprod01234567 = vmull_s8(vi8x01234567, vk8x01234567);
180       vprod89ABCDEF = vmull_s8(vi8x89ABCDEF, vk8x89ABCDEF);
181 
182       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
183       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
184       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
185       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
186 
187       vacc0123 = vqshlq_s32(vacc0123, vright_pre_shift);
188       vacc4567 = vqshlq_s32(vacc4567, vright_pre_shift);
189       vacc89AB = vqshlq_s32(vacc89AB, vright_pre_shift);
190       vaccCDEF = vqshlq_s32(vaccCDEF, vright_pre_shift);
191 
192       vacc0123 = vqdmulhq_s32(vacc0123, vmultiplier);
193       vacc4567 = vqdmulhq_s32(vacc4567, vmultiplier);
194       vacc89AB = vqdmulhq_s32(vacc89AB, vmultiplier);
195       vaccCDEF = vqdmulhq_s32(vaccCDEF, vmultiplier);
196 
197       vacc0123 = vrshlq_s32(vacc0123, vright_post_shift);
198       vacc4567 = vrshlq_s32(vacc4567, vright_post_shift);
199       vacc89AB = vrshlq_s32(vacc89AB, vright_post_shift);
200       vaccCDEF = vrshlq_s32(vaccCDEF, vright_post_shift);
201 
202 #if XNN_ARCH_ARM64
203       int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
204       int16x8_t vacc89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc89AB), vaccCDEF);
205 
206       vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
207       vacc89ABCDEF = vqaddq_s16(vacc89ABCDEF, voutput_zero_point);
208 
209       int8x16_t vout0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc01234567), vacc89ABCDEF);
210 #else  // !XNN_ARCH_ARM64
211       int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
212       int16x8_t vacc89ABCDEF = vcombine_s16(vqmovn_s32(vacc89AB), vqmovn_s32(vaccCDEF));
213 
214       vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
215       vacc89ABCDEF = vqaddq_s16(vacc89ABCDEF, voutput_zero_point);
216 
217       int8x16_t vout0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc01234567), vqmovn_s16(vacc89ABCDEF));
218 #endif  // !XNN_ARCH_ARM64
219 
220       vout0123456789ABCDEF = vmaxq_s8(vout0123456789ABCDEF, voutput_min);
221 
222       vout0123456789ABCDEF = vminq_s8(vout0123456789ABCDEF, voutput_max);
223 
224       vst1q_s8(output, vout0123456789ABCDEF); output += 16;
225     }
226     if XNN_UNLIKELY(c != 0) {
227       const int8_t* k = (const int8_t*) ((const int32_t*) w + 16);
228       do {
229         int32x4_t vacc0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
230         int32x4_t vacc4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
231 
232         const int8x8_t vi0x01234567 = vld1_s8(i0); i0 += 8;
233         const int8x8_t vk0x01234567 = vld1_s8(k); k += 8;
234 
235         int16x8_t vprod01234567 = vmull_s8(vi0x01234567, vk0x01234567);
236 
237         const int8x8_t vi1x01234567 = vld1_s8(i1); i1 += 8;
238         const int8x8_t vk1x01234567 = vld1_s8((const void*) (k + 8));
239 
240         vprod01234567 = vmlal_s8(vprod01234567, 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         const int8x8_t vi3x01234567 = vld1_s8(i3); i3 += 8;
250         const int8x8_t vk3x01234567 = vld1_s8((const void*) (k + 40));
251 
252         vprod01234567 = vmlal_s8(vprod01234567, vi3x01234567, vk3x01234567);
253 
254         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
255         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
256         const int8x8_t vi4x01234567 = vld1_s8(i4); i4 += 8;
257         const int8x8_t vk4x01234567 = vld1_s8((const void*) (k + 56));
258 
259         vprod01234567 = vmull_s8(vi4x01234567, vk4x01234567);
260 
261         const int8x8_t vi5x01234567 = vld1_s8(i5); i5 += 8;
262         const int8x8_t vk5x01234567 = vld1_s8((const void*) (k + 72));
263 
264         vprod01234567 = vmlal_s8(vprod01234567, vi5x01234567, vk5x01234567);
265 
266         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
267         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
268         const int8x8_t vi6x01234567 = vld1_s8(i6); i6 += 8;
269         const int8x8_t vk6x01234567 = vld1_s8((const void*) (k + 88));
270 
271         vprod01234567 = vmull_s8(vi6x01234567, vk6x01234567);
272 
273         const int8x8_t vi7x01234567 = vld1_s8(i7); i7 += 8;
274         const int8x8_t vk7x01234567 = vld1_s8((const void*) (k + 104));
275 
276         vprod01234567 = vmlal_s8(vprod01234567, vi7x01234567, vk7x01234567);
277 
278         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
279         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
280         const int8x8_t vi8x01234567 = vld1_s8(i8); i8 += 8;
281         const int8x8_t vk8x01234567 = vld1_s8((const void*) (k + 120));
282 
283         vprod01234567 = vmull_s8(vi8x01234567, vk8x01234567);
284 
285         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
286         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
287 
288         vacc0123 = vqshlq_s32(vacc0123, vright_pre_shift);
289         vacc4567 = vqshlq_s32(vacc4567, vright_pre_shift);
290 
291         vacc0123 = vqdmulhq_s32(vacc0123, vmultiplier);
292         vacc4567 = vqdmulhq_s32(vacc4567, vmultiplier);
293 
294         vacc0123 = vrshlq_s32(vacc0123, vright_post_shift);
295         vacc4567 = vrshlq_s32(vacc4567, vright_post_shift);
296 
297 #if XNN_ARCH_ARM64
298         int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
299 #else
300         int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
301 #endif
302         vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
303 
304         int8x8_t vout01234567 = vqmovn_s16(vacc01234567);
305         vout01234567 = vmax_s8(vout01234567, vget_low_s8(voutput_min));
306         vout01234567 = vmin_s8(vout01234567, vget_low_s8(voutput_max));
307 
308         if XNN_LIKELY(c >= 8) {
309           vst1_s8(output, vout01234567); output += 8;
310           c -= 8;
311         } else {
312           if (c & 4) {
313             vst1_lane_u32((void*) output, vreinterpret_u32_s8(vout01234567), 0); output += 4;
314             vout01234567 = vext_s8(vout01234567, vout01234567, 4);
315           }
316           if (c & 2) {
317             vst1_lane_u16((void*) output, vreinterpret_u16_s8(vout01234567), 0); output += 2;
318             vout01234567 = vext_s8(vout01234567, vout01234567, 2);
319           }
320           if (c & 1) {
321             vst1_lane_s8(output, vout01234567, 0); output += 1;
322           }
323           c = 0;
324         }
325       } while (c != 0);
326     }
327 
328     output = (int8_t*) ((uintptr_t) output + output_increment);
329   } while (--output_width != 0);
330 }
331