xref: /aosp_15_r20/external/XNNPACK/src/qc8-dwconv/gen/up8x9-minmax-fp32-neonv8-mul16.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36) 
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-dwconv/unipass-neon-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 <arm_neon.h>
13 
14 #include <xnnpack/dwconv.h>
15 #include <xnnpack/intrinsics-polyfill.h>
16 
17 
xnn_qc8_dwconv_minmax_fp32_ukernel_up8x9__neonv8_mul16(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_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qc8_dwconv_minmax_fp32_ukernel_up8x9__neonv8_mul16(
19     size_t channels,
20     size_t output_width,
21     const int8_t** input,
22     const void* weights,
23     int8_t* output,
24     size_t input_stride,
25     size_t output_increment,
26     size_t input_offset,
27     const int8_t* zero,
28     const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29 {
30   assert(channels != 0);
31   assert(output_width != 0);
32 
33   const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->fp32_neonv8.output_zero_point);
34   const int8x8_t voutput_min = vld1_dup_s8(&params->fp32_neonv8.output_min);
35   const int8x8_t voutput_max = vld1_dup_s8(&params->fp32_neonv8.output_max);
36   do {
37     const int8_t* i0 = input[0];
38     assert(i0 != NULL);
39     if XNN_UNPREDICTABLE(i0 != zero) {
40       i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
41     }
42     const int8_t* i1 = input[1];
43     assert(i1 != NULL);
44     if XNN_UNPREDICTABLE(i1 != zero) {
45       i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
46     }
47     const int8_t* i2 = input[2];
48     assert(i2 != NULL);
49     if XNN_UNPREDICTABLE(i2 != zero) {
50       i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
51     }
52     const int8_t* i3 = input[3];
53     assert(i3 != NULL);
54     if XNN_UNPREDICTABLE(i3 != zero) {
55       i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
56     }
57     const int8_t* i4 = input[4];
58     assert(i4 != NULL);
59     if XNN_UNPREDICTABLE(i4 != zero) {
60       i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
61     }
62     const int8_t* i5 = input[5];
63     assert(i5 != NULL);
64     if XNN_UNPREDICTABLE(i5 != zero) {
65       i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
66     }
67     const int8_t* i6 = input[6];
68     assert(i6 != NULL);
69     if XNN_UNPREDICTABLE(i6 != zero) {
70       i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
71     }
72     const int8_t* i7 = input[7];
73     assert(i7 != NULL);
74     if XNN_UNPREDICTABLE(i7 != zero) {
75       i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
76     }
77     const int8_t* i8 = input[8];
78     assert(i8 != NULL);
79     if XNN_UNPREDICTABLE(i8 != zero) {
80       i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
81     }
82     input = (const int8_t**) ((uintptr_t) input + input_stride);
83 
84     size_t c = channels;
85     const void* w = weights;
86     for (; c >= 8; c -= 8) {
87       int32x4_t vacc0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
88       int32x4_t vacc4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
89 
90 
91       const int16x8_t vi0x01234567 = vmovl_s8(vld1_s8(i0)); i0 += 8;
92       const int16x8_t vk0x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
93 
94       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi0x01234567), vget_low_s16(vk0x01234567));
95       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi0x01234567), vget_high_s16(vk0x01234567));
96 
97       const int16x8_t vi1x01234567 = vmovl_s8(vld1_s8(i1)); i1 += 8;
98       const int16x8_t vk1x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
99 
100       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi1x01234567), vget_low_s16(vk1x01234567));
101       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi1x01234567), vget_high_s16(vk1x01234567));
102 
103       const int16x8_t vi2x01234567 = vmovl_s8(vld1_s8(i2)); i2 += 8;
104       const int16x8_t vk2x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
105 
106       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi2x01234567), vget_low_s16(vk2x01234567));
107       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi2x01234567), vget_high_s16(vk2x01234567));
108 
109       const int16x8_t vi3x01234567 = vmovl_s8(vld1_s8(i3)); i3 += 8;
110       const int16x8_t vk3x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
111 
112       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi3x01234567), vget_low_s16(vk3x01234567));
113       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi3x01234567), vget_high_s16(vk3x01234567));
114 
115       const int16x8_t vi4x01234567 = vmovl_s8(vld1_s8(i4)); i4 += 8;
116       const int16x8_t vk4x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
117 
118       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi4x01234567), vget_low_s16(vk4x01234567));
119       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi4x01234567), vget_high_s16(vk4x01234567));
120 
121       const int16x8_t vi5x01234567 = vmovl_s8(vld1_s8(i5)); i5 += 8;
122       const int16x8_t vk5x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
123 
124       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi5x01234567), vget_low_s16(vk5x01234567));
125       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi5x01234567), vget_high_s16(vk5x01234567));
126 
127       const int16x8_t vi6x01234567 = vmovl_s8(vld1_s8(i6)); i6 += 8;
128       const int16x8_t vk6x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
129 
130       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi6x01234567), vget_low_s16(vk6x01234567));
131       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi6x01234567), vget_high_s16(vk6x01234567));
132 
133       const int16x8_t vi7x01234567 = vmovl_s8(vld1_s8(i7)); i7 += 8;
134       const int16x8_t vk7x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
135 
136       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi7x01234567), vget_low_s16(vk7x01234567));
137       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi7x01234567), vget_high_s16(vk7x01234567));
138 
139       const int16x8_t vi8x01234567 = vmovl_s8(vld1_s8(i8)); i8 += 8;
140       const int16x8_t vk8x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
141 
142       vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi8x01234567), vget_low_s16(vk8x01234567));
143       vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi8x01234567), vget_high_s16(vk8x01234567));
144 
145       float32x4_t vfpacc0123 = vcvtq_f32_s32(vacc0123);
146       float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567);
147 
148       const float32x4_t vscale0123 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
149       const float32x4_t vscale4567 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
150 
151       vfpacc0123 = vmulq_f32(vfpacc0123, vscale0123);
152       vfpacc4567 = vmulq_f32(vfpacc4567, vscale4567);
153 
154       vacc0123 = vcvtnq_s32_f32(vfpacc0123);
155       vacc4567 = vcvtnq_s32_f32(vfpacc4567);
156 
157 #if XNN_ARCH_ARM64
158       int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
159 
160       vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
161 
162       int8x8_t vout01234567 = vqmovn_s16(vacc01234567);
163 #else  // !XNN_ARCH_ARM64
164       int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
165 
166       vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
167 
168       int8x8_t vout01234567 = vqmovn_s16(vacc01234567);
169 #endif  // !XNN_ARCH_ARM64
170 
171       vout01234567 = vmax_s8(vout01234567, voutput_min);
172 
173       vout01234567 = vmin_s8(vout01234567, voutput_max);
174 
175       vst1_s8(output, vout01234567); output += 8;
176     }
177     if XNN_UNLIKELY(c != 0) {
178       {
179         int32x4_t vacc0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
180         int32x4_t vacc4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
181 
182         const int16x8_t vi0x01234567 = vmovl_s8(vld1_s8(i0));
183         const int16x8_t vk0x01234567 = vmovl_s8(vld1_s8(w));
184 
185         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi0x01234567), vget_low_s16(vk0x01234567));
186         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi0x01234567), vget_high_s16(vk0x01234567));
187         const int16x8_t vi1x01234567 = vmovl_s8(vld1_s8(i1));
188         const int16x8_t vk1x01234567 = vmovl_s8(vld1_s8((const void*) ((const int8_t*) w + 8)));
189 
190         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi1x01234567), vget_low_s16(vk1x01234567));
191         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi1x01234567), vget_high_s16(vk1x01234567));
192         const int16x8_t vi2x01234567 = vmovl_s8(vld1_s8(i2));
193         const int16x8_t vk2x01234567 = vmovl_s8(vld1_s8((const void*) ((const int8_t*) w + 16)));
194 
195         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi2x01234567), vget_low_s16(vk2x01234567));
196         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi2x01234567), vget_high_s16(vk2x01234567));
197         const int16x8_t vi3x01234567 = vmovl_s8(vld1_s8(i3));
198         const int16x8_t vk3x01234567 = vmovl_s8(vld1_s8((const void*) ((const int8_t*) w + 24)));
199 
200         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi3x01234567), vget_low_s16(vk3x01234567));
201         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi3x01234567), vget_high_s16(vk3x01234567));
202         const int16x8_t vi4x01234567 = vmovl_s8(vld1_s8(i4));
203         const int16x8_t vk4x01234567 = vmovl_s8(vld1_s8((const void*) ((const int8_t*) w + 32)));
204 
205         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi4x01234567), vget_low_s16(vk4x01234567));
206         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi4x01234567), vget_high_s16(vk4x01234567));
207         const int16x8_t vi5x01234567 = vmovl_s8(vld1_s8(i5));
208         const int16x8_t vk5x01234567 = vmovl_s8(vld1_s8((const void*) ((const int8_t*) w + 40)));
209 
210         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi5x01234567), vget_low_s16(vk5x01234567));
211         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi5x01234567), vget_high_s16(vk5x01234567));
212         const int16x8_t vi6x01234567 = vmovl_s8(vld1_s8(i6));
213         const int16x8_t vk6x01234567 = vmovl_s8(vld1_s8((const void*) ((const int8_t*) w + 48)));
214 
215         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi6x01234567), vget_low_s16(vk6x01234567));
216         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi6x01234567), vget_high_s16(vk6x01234567));
217         const int16x8_t vi7x01234567 = vmovl_s8(vld1_s8(i7));
218         const int16x8_t vk7x01234567 = vmovl_s8(vld1_s8((const void*) ((const int8_t*) w + 56)));
219 
220         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi7x01234567), vget_low_s16(vk7x01234567));
221         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi7x01234567), vget_high_s16(vk7x01234567));
222         const int16x8_t vi8x01234567 = vmovl_s8(vld1_s8(i8));
223         const int16x8_t vk8x01234567 = vmovl_s8(vld1_s8((const void*) ((const int8_t*) w + 64)));
224 
225         vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi8x01234567), vget_low_s16(vk8x01234567));
226         vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi8x01234567), vget_high_s16(vk8x01234567));
227 
228         float32x4_t vfpacc0123 = vcvtq_f32_s32(vacc0123);
229         float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567);
230 
231         const float32x4_t vscale0123 = vld1q_f32((const float*) ((uintptr_t) w + 0 * sizeof(int32_t) + 72 * sizeof(int8_t)));
232         const float32x4_t vscale4567 = vld1q_f32((const float*) ((uintptr_t) w + 0 * sizeof(int32_t) + 72 * sizeof(int8_t) + 4 * sizeof(float)));
233         vfpacc0123 = vmulq_f32(vfpacc0123, vscale0123);
234         vfpacc4567 = vmulq_f32(vfpacc4567, vscale4567);
235 
236         vacc0123 = vcvtnq_s32_f32(vfpacc0123);
237         vacc4567 = vcvtnq_s32_f32(vfpacc4567);
238 
239 #if XNN_ARCH_ARM64
240         int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
241 #else
242         int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
243 #endif
244         vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
245 
246         int8x8_t vout01234567 = vqmovn_s16(vacc01234567);
247         vout01234567 = vmax_s8(vout01234567, voutput_min);
248         vout01234567 = vmin_s8(vout01234567, voutput_max);
249 
250         if (c & 4) {
251           vst1_lane_u32((void*) output, vreinterpret_u32_s8(vout01234567), 0); output += 4;
252           vout01234567 = vext_s8(vout01234567, vout01234567, 4);
253         }
254         if (c & 2) {
255           vst1_lane_u16((void*) output, vreinterpret_u16_s8(vout01234567), 0); output += 2;
256           vout01234567 = vext_s8(vout01234567, vout01234567, 2);
257         }
258         if (c & 1) {
259           vst1_lane_s8(output, vout01234567, 0); output += 1;
260         }
261       }
262     }
263 
264     output = (int8_t*) ((uintptr_t) output + output_increment);
265   } while (--output_width != 0);
266 }
267