xref: /aosp_15_r20/external/XNNPACK/src/qs8-dwconv/gen/up16x9-minmax-rndnu-neon-mla8-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_mla8_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_mla8_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       const int8x16_t vi1x0123456789ABCDEF = vld1q_s8(i1); i1 += 16;
101       const int8x16_t vk1x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
102 
103       vprod01234567 = vmlal_s8(vprod01234567, vget_low_s8(vi1x0123456789ABCDEF), vget_low_s8(vk1x0123456789ABCDEF));
104       vprod89ABCDEF = vmlal_s8(vprod89ABCDEF, vget_high_s8(vi1x0123456789ABCDEF), vget_high_s8(vk1x0123456789ABCDEF));
105 
106       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
107       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
108       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
109       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
110       const int8x16_t vi2x0123456789ABCDEF = vld1q_s8(i2); i2 += 16;
111       const int8x16_t vk2x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
112 
113       vprod01234567 = vmull_s8(vget_low_s8(vi2x0123456789ABCDEF), vget_low_s8(vk2x0123456789ABCDEF));
114       vprod89ABCDEF = vmull_s8(vget_high_s8(vi2x0123456789ABCDEF), vget_high_s8(vk2x0123456789ABCDEF));
115 
116       const int8x16_t vi3x0123456789ABCDEF = vld1q_s8(i3); i3 += 16;
117       const int8x16_t vk3x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
118 
119       vprod01234567 = vmlal_s8(vprod01234567, vget_low_s8(vi3x0123456789ABCDEF), vget_low_s8(vk3x0123456789ABCDEF));
120       vprod89ABCDEF = vmlal_s8(vprod89ABCDEF, vget_high_s8(vi3x0123456789ABCDEF), vget_high_s8(vk3x0123456789ABCDEF));
121 
122       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
123       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
124       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
125       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
126       const int8x16_t vi4x0123456789ABCDEF = vld1q_s8(i4); i4 += 16;
127       const int8x16_t vk4x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
128 
129       vprod01234567 = vmull_s8(vget_low_s8(vi4x0123456789ABCDEF), vget_low_s8(vk4x0123456789ABCDEF));
130       vprod89ABCDEF = vmull_s8(vget_high_s8(vi4x0123456789ABCDEF), vget_high_s8(vk4x0123456789ABCDEF));
131 
132       const int8x16_t vi5x0123456789ABCDEF = vld1q_s8(i5); i5 += 16;
133       const int8x16_t vk5x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
134 
135       vprod01234567 = vmlal_s8(vprod01234567, vget_low_s8(vi5x0123456789ABCDEF), vget_low_s8(vk5x0123456789ABCDEF));
136       vprod89ABCDEF = vmlal_s8(vprod89ABCDEF, vget_high_s8(vi5x0123456789ABCDEF), vget_high_s8(vk5x0123456789ABCDEF));
137 
138       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
139       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
140       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
141       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
142       const int8x16_t vi6x0123456789ABCDEF = vld1q_s8(i6); i6 += 16;
143       const int8x16_t vk6x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
144 
145       vprod01234567 = vmull_s8(vget_low_s8(vi6x0123456789ABCDEF), vget_low_s8(vk6x0123456789ABCDEF));
146       vprod89ABCDEF = vmull_s8(vget_high_s8(vi6x0123456789ABCDEF), vget_high_s8(vk6x0123456789ABCDEF));
147 
148       const int8x16_t vi7x0123456789ABCDEF = vld1q_s8(i7); i7 += 16;
149       const int8x16_t vk7x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
150 
151       vprod01234567 = vmlal_s8(vprod01234567, vget_low_s8(vi7x0123456789ABCDEF), vget_low_s8(vk7x0123456789ABCDEF));
152       vprod89ABCDEF = vmlal_s8(vprod89ABCDEF, vget_high_s8(vi7x0123456789ABCDEF), vget_high_s8(vk7x0123456789ABCDEF));
153 
154       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
155       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
156       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
157       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
158       const int8x16_t vi8x0123456789ABCDEF = vld1q_s8(i8); i8 += 16;
159       const int8x16_t vk8x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
160 
161       vprod01234567 = vmull_s8(vget_low_s8(vi8x0123456789ABCDEF), vget_low_s8(vk8x0123456789ABCDEF));
162       vprod89ABCDEF = vmull_s8(vget_high_s8(vi8x0123456789ABCDEF), vget_high_s8(vk8x0123456789ABCDEF));
163 
164       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
165       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
166       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
167       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
168 
169       vacc0123 = vqshlq_s32(vacc0123, vright_pre_shift);
170       vacc4567 = vqshlq_s32(vacc4567, vright_pre_shift);
171       vacc89AB = vqshlq_s32(vacc89AB, vright_pre_shift);
172       vaccCDEF = vqshlq_s32(vaccCDEF, vright_pre_shift);
173 
174       vacc0123 = vqdmulhq_s32(vacc0123, vmultiplier);
175       vacc4567 = vqdmulhq_s32(vacc4567, vmultiplier);
176       vacc89AB = vqdmulhq_s32(vacc89AB, vmultiplier);
177       vaccCDEF = vqdmulhq_s32(vaccCDEF, vmultiplier);
178 
179       vacc0123 = vrshlq_s32(vacc0123, vright_post_shift);
180       vacc4567 = vrshlq_s32(vacc4567, vright_post_shift);
181       vacc89AB = vrshlq_s32(vacc89AB, vright_post_shift);
182       vaccCDEF = vrshlq_s32(vaccCDEF, vright_post_shift);
183 
184 #if XNN_ARCH_ARM64
185       int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
186       int16x8_t vacc89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc89AB), vaccCDEF);
187 
188       vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
189       vacc89ABCDEF = vqaddq_s16(vacc89ABCDEF, voutput_zero_point);
190 
191       int8x16_t vout0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc01234567), vacc89ABCDEF);
192 #else  // !XNN_ARCH_ARM64
193       int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
194       int16x8_t vacc89ABCDEF = vcombine_s16(vqmovn_s32(vacc89AB), vqmovn_s32(vaccCDEF));
195 
196       vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
197       vacc89ABCDEF = vqaddq_s16(vacc89ABCDEF, voutput_zero_point);
198 
199       int8x16_t vout0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc01234567), vqmovn_s16(vacc89ABCDEF));
200 #endif  // !XNN_ARCH_ARM64
201 
202       vout0123456789ABCDEF = vmaxq_s8(vout0123456789ABCDEF, voutput_min);
203 
204       vout0123456789ABCDEF = vminq_s8(vout0123456789ABCDEF, voutput_max);
205 
206       vst1q_s8(output, vout0123456789ABCDEF); output += 16;
207     }
208     if XNN_UNLIKELY(c != 0) {
209       const int8_t* k = (const int8_t*) ((const int32_t*) w + 16);
210       do {
211         int32x4_t vacc0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
212         int32x4_t vacc4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
213 
214         const int8x8_t vi0x01234567 = vld1_s8(i0); i0 += 8;
215         const int8x8_t vk0x01234567 = vld1_s8(k); k += 8;
216 
217         int16x8_t vprod01234567 = vmull_s8(vi0x01234567, vk0x01234567);
218 
219         const int8x8_t vi1x01234567 = vld1_s8(i1); i1 += 8;
220         const int8x8_t vk1x01234567 = vld1_s8((const void*) (k + 8));
221 
222         vprod01234567 = vmlal_s8(vprod01234567, vi1x01234567, vk1x01234567);
223 
224         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
225         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
226         const int8x8_t vi2x01234567 = vld1_s8(i2); i2 += 8;
227         const int8x8_t vk2x01234567 = vld1_s8((const void*) (k + 24));
228 
229         vprod01234567 = vmull_s8(vi2x01234567, vk2x01234567);
230 
231         const int8x8_t vi3x01234567 = vld1_s8(i3); i3 += 8;
232         const int8x8_t vk3x01234567 = vld1_s8((const void*) (k + 40));
233 
234         vprod01234567 = vmlal_s8(vprod01234567, vi3x01234567, vk3x01234567);
235 
236         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
237         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
238         const int8x8_t vi4x01234567 = vld1_s8(i4); i4 += 8;
239         const int8x8_t vk4x01234567 = vld1_s8((const void*) (k + 56));
240 
241         vprod01234567 = vmull_s8(vi4x01234567, vk4x01234567);
242 
243         const int8x8_t vi5x01234567 = vld1_s8(i5); i5 += 8;
244         const int8x8_t vk5x01234567 = vld1_s8((const void*) (k + 72));
245 
246         vprod01234567 = vmlal_s8(vprod01234567, vi5x01234567, vk5x01234567);
247 
248         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
249         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
250         const int8x8_t vi6x01234567 = vld1_s8(i6); i6 += 8;
251         const int8x8_t vk6x01234567 = vld1_s8((const void*) (k + 88));
252 
253         vprod01234567 = vmull_s8(vi6x01234567, vk6x01234567);
254 
255         const int8x8_t vi7x01234567 = vld1_s8(i7); i7 += 8;
256         const int8x8_t vk7x01234567 = vld1_s8((const void*) (k + 104));
257 
258         vprod01234567 = vmlal_s8(vprod01234567, vi7x01234567, vk7x01234567);
259 
260         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
261         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
262         const int8x8_t vi8x01234567 = vld1_s8(i8); i8 += 8;
263         const int8x8_t vk8x01234567 = vld1_s8((const void*) (k + 120));
264 
265         vprod01234567 = vmull_s8(vi8x01234567, vk8x01234567);
266 
267         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
268         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
269 
270         vacc0123 = vqshlq_s32(vacc0123, vright_pre_shift);
271         vacc4567 = vqshlq_s32(vacc4567, vright_pre_shift);
272 
273         vacc0123 = vqdmulhq_s32(vacc0123, vmultiplier);
274         vacc4567 = vqdmulhq_s32(vacc4567, vmultiplier);
275 
276         vacc0123 = vrshlq_s32(vacc0123, vright_post_shift);
277         vacc4567 = vrshlq_s32(vacc4567, vright_post_shift);
278 
279 #if XNN_ARCH_ARM64
280         int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
281 #else
282         int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
283 #endif
284         vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
285 
286         int8x8_t vout01234567 = vqmovn_s16(vacc01234567);
287         vout01234567 = vmax_s8(vout01234567, vget_low_s8(voutput_min));
288         vout01234567 = vmin_s8(vout01234567, vget_low_s8(voutput_max));
289 
290         if XNN_LIKELY(c >= 8) {
291           vst1_s8(output, vout01234567); output += 8;
292           c -= 8;
293         } else {
294           if (c & 4) {
295             vst1_lane_u32((void*) output, vreinterpret_u32_s8(vout01234567), 0); output += 4;
296             vout01234567 = vext_s8(vout01234567, vout01234567, 4);
297           }
298           if (c & 2) {
299             vst1_lane_u16((void*) output, vreinterpret_u16_s8(vout01234567), 0); output += 2;
300             vout01234567 = vext_s8(vout01234567, vout01234567, 2);
301           }
302           if (c & 1) {
303             vst1_lane_s8(output, vout01234567, 0); output += 1;
304           }
305           c = 0;
306         }
307       } while (c != 0);
308     }
309 
310     output = (int8_t*) ((uintptr_t) output + output_increment);
311   } while (--output_width != 0);
312 }
313