xref: /aosp_15_r20/external/XNNPACK/src/qu8-dwconv/gen/up8x9-minmax-rndnu-neon-mul8.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36) 
1 // Auto-generated file. Do not edit!
2 //   Template: src/qu8-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_qu8_dwconv_minmax_rndnu_ukernel_up8x9__neon_mul8(size_t channels,size_t output_width,const uint8_t ** input,const void * weights,uint8_t * output,size_t input_stride,size_t output_increment,size_t input_offset,const uint8_t * zero,const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qu8_dwconv_minmax_rndnu_ukernel_up8x9__neon_mul8(
18     size_t channels,
19     size_t output_width,
20     const uint8_t** input,
21     const void* weights,
22     uint8_t* output,
23     size_t input_stride,
24     size_t output_increment,
25     size_t input_offset,
26     const uint8_t* zero,
27     const union xnn_qu8_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 uint8x8_t vkernel_zero_point = vld1_dup_u8(params->rndnu_neon.kernel_zero_point);
33   const uint16x8_t vkernel_zero_point16 = vmovl_u8(vkernel_zero_point);
34   const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
35   const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
36   const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
37   const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
38   const uint8x8_t voutput_min = vld1_dup_u8(&params->rndnu_neon.output_min);
39   const uint8x8_t voutput_max = vld1_dup_u8(&params->rndnu_neon.output_max);
40   do {
41     const uint8_t* i0 = input[0];
42     assert(i0 != NULL);
43     if XNN_UNPREDICTABLE(i0 != zero) {
44       i0 = (const uint8_t*) ((uintptr_t) i0 + input_offset);
45     }
46     const uint8_t* i1 = input[1];
47     assert(i1 != NULL);
48     if XNN_UNPREDICTABLE(i1 != zero) {
49       i1 = (const uint8_t*) ((uintptr_t) i1 + input_offset);
50     }
51     const uint8_t* i2 = input[2];
52     assert(i2 != NULL);
53     if XNN_UNPREDICTABLE(i2 != zero) {
54       i2 = (const uint8_t*) ((uintptr_t) i2 + input_offset);
55     }
56     const uint8_t* i3 = input[3];
57     assert(i3 != NULL);
58     if XNN_UNPREDICTABLE(i3 != zero) {
59       i3 = (const uint8_t*) ((uintptr_t) i3 + input_offset);
60     }
61     const uint8_t* i4 = input[4];
62     assert(i4 != NULL);
63     if XNN_UNPREDICTABLE(i4 != zero) {
64       i4 = (const uint8_t*) ((uintptr_t) i4 + input_offset);
65     }
66     const uint8_t* i5 = input[5];
67     assert(i5 != NULL);
68     if XNN_UNPREDICTABLE(i5 != zero) {
69       i5 = (const uint8_t*) ((uintptr_t) i5 + input_offset);
70     }
71     const uint8_t* i6 = input[6];
72     assert(i6 != NULL);
73     if XNN_UNPREDICTABLE(i6 != zero) {
74       i6 = (const uint8_t*) ((uintptr_t) i6 + input_offset);
75     }
76     const uint8_t* i7 = input[7];
77     assert(i7 != NULL);
78     if XNN_UNPREDICTABLE(i7 != zero) {
79       i7 = (const uint8_t*) ((uintptr_t) i7 + input_offset);
80     }
81     const uint8_t* i8 = input[8];
82     assert(i8 != NULL);
83     if XNN_UNPREDICTABLE(i8 != zero) {
84       i8 = (const uint8_t*) ((uintptr_t) i8 + input_offset);
85     }
86     input = (const uint8_t**) ((uintptr_t) input + input_stride);
87 
88 
89     size_t c = channels;
90     const void* w = weights;
91     for (; c >= 8; c -= 8) {
92       int32x4_t vacc0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
93       int32x4_t vacc4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
94 
95 
96       const uint8x8_t vi0x01234567 = vld1_u8(i0); i0 += 8;
97       const uint8x8_t vk0x01234567 = vld1_u8(w); w = (const void*) ((const int8_t*) w + 8);
98 
99       uint16x8_t vprod01234567 = vmull_u8(vi0x01234567, vk0x01234567);
100 
101       vacc0123 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc0123), vget_low_u16(vprod01234567)));
102       vacc4567 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc4567), vget_high_u16(vprod01234567)));
103       const uint8x8_t vi1x01234567 = vld1_u8(i1); i1 += 8;
104       const uint8x8_t vk1x01234567 = vld1_u8(w); w = (const void*) ((const int8_t*) w + 8);
105 
106       vprod01234567 = vmull_u8(vi1x01234567, vk1x01234567);
107       uint16x8_t vsum01234567 = vaddl_u8(vi0x01234567, vi1x01234567);
108 
109       vacc0123 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc0123), vget_low_u16(vprod01234567)));
110       vacc4567 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc4567), vget_high_u16(vprod01234567)));
111       const uint8x8_t vi2x01234567 = vld1_u8(i2); i2 += 8;
112       const uint8x8_t vk2x01234567 = vld1_u8(w); w = (const void*) ((const int8_t*) w + 8);
113 
114       vprod01234567 = vmull_u8(vi2x01234567, vk2x01234567);
115       vsum01234567 = vaddw_u8(vsum01234567, vi2x01234567);
116 
117       vacc0123 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc0123), vget_low_u16(vprod01234567)));
118       vacc4567 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc4567), vget_high_u16(vprod01234567)));
119       const uint8x8_t vi3x01234567 = vld1_u8(i3); i3 += 8;
120       const uint8x8_t vk3x01234567 = vld1_u8(w); w = (const void*) ((const int8_t*) w + 8);
121 
122       vprod01234567 = vmull_u8(vi3x01234567, vk3x01234567);
123       vsum01234567 = vaddw_u8(vsum01234567, vi3x01234567);
124 
125       vacc0123 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc0123), vget_low_u16(vprod01234567)));
126       vacc4567 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc4567), vget_high_u16(vprod01234567)));
127       const uint8x8_t vi4x01234567 = vld1_u8(i4); i4 += 8;
128       const uint8x8_t vk4x01234567 = vld1_u8(w); w = (const void*) ((const int8_t*) w + 8);
129 
130       vprod01234567 = vmull_u8(vi4x01234567, vk4x01234567);
131       vsum01234567 = vaddw_u8(vsum01234567, vi4x01234567);
132 
133       vacc0123 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc0123), vget_low_u16(vprod01234567)));
134       vacc4567 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc4567), vget_high_u16(vprod01234567)));
135       const uint8x8_t vi5x01234567 = vld1_u8(i5); i5 += 8;
136       const uint8x8_t vk5x01234567 = vld1_u8(w); w = (const void*) ((const int8_t*) w + 8);
137 
138       vprod01234567 = vmull_u8(vi5x01234567, vk5x01234567);
139       vsum01234567 = vaddw_u8(vsum01234567, vi5x01234567);
140 
141       vacc0123 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc0123), vget_low_u16(vprod01234567)));
142       vacc4567 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc4567), vget_high_u16(vprod01234567)));
143       const uint8x8_t vi6x01234567 = vld1_u8(i6); i6 += 8;
144       const uint8x8_t vk6x01234567 = vld1_u8(w); w = (const void*) ((const int8_t*) w + 8);
145 
146       vprod01234567 = vmull_u8(vi6x01234567, vk6x01234567);
147       vsum01234567 = vaddw_u8(vsum01234567, vi6x01234567);
148 
149       vacc0123 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc0123), vget_low_u16(vprod01234567)));
150       vacc4567 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc4567), vget_high_u16(vprod01234567)));
151       const uint8x8_t vi7x01234567 = vld1_u8(i7); i7 += 8;
152       const uint8x8_t vk7x01234567 = vld1_u8(w); w = (const void*) ((const int8_t*) w + 8);
153 
154       vprod01234567 = vmull_u8(vi7x01234567, vk7x01234567);
155       vsum01234567 = vaddw_u8(vsum01234567, vi7x01234567);
156 
157       vacc0123 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc0123), vget_low_u16(vprod01234567)));
158       vacc4567 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc4567), vget_high_u16(vprod01234567)));
159       const uint8x8_t vi8x01234567 = vld1_u8(i8); i8 += 8;
160       const uint8x8_t vk8x01234567 = vld1_u8(w); w = (const void*) ((const int8_t*) w + 8);
161 
162       vprod01234567 = vmull_u8(vi8x01234567, vk8x01234567);
163       vsum01234567 = vaddw_u8(vsum01234567, vi8x01234567);
164 
165       vacc0123 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc0123), vget_low_u16(vprod01234567)));
166       vacc4567 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vacc4567), vget_high_u16(vprod01234567)));
167 
168       vacc0123 = vreinterpretq_s32_u32(vmlsl_u16(vreinterpretq_u32_s32(vacc0123), vget_low_u16(vsum01234567), vget_low_u16(vkernel_zero_point16)));
169       vacc4567 = vreinterpretq_s32_u32(vmlsl_u16(vreinterpretq_u32_s32(vacc4567), vget_high_u16(vsum01234567), vget_high_u16(vkernel_zero_point16)));
170 
171       vacc0123 = vshlq_s32(vacc0123, vright_pre_shift);
172       vacc4567 = vshlq_s32(vacc4567, vright_pre_shift);
173 
174       vacc0123 = vqdmulhq_s32(vacc0123, vmultiplier);
175       vacc4567 = vqdmulhq_s32(vacc4567, vmultiplier);
176 
177       vacc0123 = vrshlq_s32(vacc0123, vright_post_shift);
178       vacc4567 = vrshlq_s32(vacc4567, vright_post_shift);
179 
180 #if XNN_ARCH_ARM64
181       const int16x8_t vacc01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567), voutput_zero_point);
182 
183       uint8x8_t vout01234567 = vqmovun_s16(vacc01234567);
184 #else
185       const int16x8_t vacc01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)), voutput_zero_point);
186 
187       uint8x8_t vout01234567 = vqmovun_s16(vacc01234567);
188 #endif
189 
190       vout01234567 = vmax_u8(vout01234567, voutput_min);
191 
192       vout01234567 = vmin_u8(vout01234567, voutput_max);
193 
194       vst1_u8(output, vout01234567); output += 8;
195     }
196     if XNN_UNLIKELY(c != 0) {
197       {
198         int32x4_t vacc0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
199         int32x4_t vacc4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
200 
201         const int16x8_t vi0x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i0)));
202         const int16x8_t vk0x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point));
203 
204         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi0x01234567), vget_low_s16(vk0x01234567));
205         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi0x01234567), vget_high_s16(vk0x01234567));
206         const int16x8_t vi1x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i1)));
207         const int16x8_t vk1x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) ((const uint8_t*) w + 8)), vkernel_zero_point));
208 
209         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi1x01234567), vget_low_s16(vk1x01234567));
210         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi1x01234567), vget_high_s16(vk1x01234567));
211         const int16x8_t vi2x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i2)));
212         const int16x8_t vk2x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) ((const uint8_t*) w + 16)), vkernel_zero_point));
213 
214         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi2x01234567), vget_low_s16(vk2x01234567));
215         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi2x01234567), vget_high_s16(vk2x01234567));
216         const int16x8_t vi3x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i3)));
217         const int16x8_t vk3x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) ((const uint8_t*) w + 24)), vkernel_zero_point));
218 
219         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi3x01234567), vget_low_s16(vk3x01234567));
220         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi3x01234567), vget_high_s16(vk3x01234567));
221         const int16x8_t vi4x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i4)));
222         const int16x8_t vk4x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) ((const uint8_t*) w + 32)), vkernel_zero_point));
223 
224         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi4x01234567), vget_low_s16(vk4x01234567));
225         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi4x01234567), vget_high_s16(vk4x01234567));
226         const int16x8_t vi5x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i5)));
227         const int16x8_t vk5x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) ((const uint8_t*) w + 40)), vkernel_zero_point));
228 
229         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi5x01234567), vget_low_s16(vk5x01234567));
230         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi5x01234567), vget_high_s16(vk5x01234567));
231         const int16x8_t vi6x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i6)));
232         const int16x8_t vk6x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) ((const uint8_t*) w + 48)), vkernel_zero_point));
233 
234         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi6x01234567), vget_low_s16(vk6x01234567));
235         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi6x01234567), vget_high_s16(vk6x01234567));
236         const int16x8_t vi7x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i7)));
237         const int16x8_t vk7x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) ((const uint8_t*) w + 56)), vkernel_zero_point));
238 
239         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi7x01234567), vget_low_s16(vk7x01234567));
240         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi7x01234567), vget_high_s16(vk7x01234567));
241         const int16x8_t vi8x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i8)));
242         const int16x8_t vk8x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) ((const uint8_t*) w + 64)), vkernel_zero_point));
243 
244         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi8x01234567), vget_low_s16(vk8x01234567));
245         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi8x01234567), vget_high_s16(vk8x01234567));
246 
247         vacc0123 = vrshlq_s32(vacc0123, vright_pre_shift);
248         vacc4567 = vrshlq_s32(vacc4567, vright_pre_shift);
249 
250         vacc0123 = vqdmulhq_s32(vacc0123, vmultiplier);
251         vacc4567 = vqdmulhq_s32(vacc4567, vmultiplier);
252 
253         vacc0123 = vrshlq_s32(vacc0123, vright_post_shift);
254         vacc4567 = vrshlq_s32(vacc4567, vright_post_shift);
255 
256 #if XNN_ARCH_ARM64
257         const int16x8_t vacc01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567), voutput_zero_point);
258         uint8x8_t vout01234567 = vqmovun_s16(vacc01234567);
259 #else
260         const int16x8_t vacc01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)), voutput_zero_point);
261         uint8x8_t vout01234567 = vqmovun_s16(vacc01234567);
262 #endif
263 
264         vout01234567 = vmax_u8(vout01234567, voutput_min);
265         vout01234567 = vmin_u8(vout01234567, voutput_max);
266 
267         if (c & 4) {
268           vst1_lane_u32((void*) output, vreinterpret_u32_u8(vout01234567), 0); output += 4;
269           vout01234567 = vext_u8(vout01234567, vout01234567, 4);
270         }
271         if (c & 2) {
272           vst1_lane_u16((void*) output, vreinterpret_u16_u8(vout01234567), 0); output += 2;
273           vout01234567 = vext_u8(vout01234567, vout01234567, 2);
274         }
275         if (c & 1) {
276           vst1_lane_u8(output, vout01234567, 0); output += 1;
277         }
278       }
279     }
280 
281     output = (uint8_t*) ((uintptr_t) output + output_increment);
282   } while (--output_width != 0);
283 }
284