xref: /aosp_15_r20/external/XNNPACK/src/qc8-dwconv/gen/up16x9-minmax-fp32-neonv8-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 #include <xnnpack/intrinsics-polyfill.h>
16 
17 
xnn_qc8_dwconv_minmax_fp32_ukernel_up16x9__neonv8_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_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qc8_dwconv_minmax_fp32_ukernel_up16x9__neonv8_mla8_ld128(
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 int8x16_t voutput_min = vld1q_dup_s8(&params->fp32_neonv8.output_min);
35   const int8x16_t voutput_max = vld1q_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 >= 16; c -= 16) {
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       int32x4_t vacc89AB = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
90       int32x4_t vaccCDEF = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
91 
92       const int8x16_t vi0x0123456789ABCDEF = vld1q_s8(i0); i0 += 16;
93       const int8x16_t vk0x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
94 
95       int16x8_t vprod01234567 = vmull_s8(vget_low_s8(vi0x0123456789ABCDEF), vget_low_s8(vk0x0123456789ABCDEF));
96       int16x8_t vprod89ABCDEF = vmull_s8(vget_high_s8(vi0x0123456789ABCDEF), vget_high_s8(vk0x0123456789ABCDEF));
97 
98       const int8x16_t vi1x0123456789ABCDEF = vld1q_s8(i1); i1 += 16;
99       const int8x16_t vk1x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
100 
101       vprod01234567 = vmlal_s8(vprod01234567, vget_low_s8(vi1x0123456789ABCDEF), vget_low_s8(vk1x0123456789ABCDEF));
102       vprod89ABCDEF = vmlal_s8(vprod89ABCDEF, vget_high_s8(vi1x0123456789ABCDEF), vget_high_s8(vk1x0123456789ABCDEF));
103 
104       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
105       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
106       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
107       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
108       const int8x16_t vi2x0123456789ABCDEF = vld1q_s8(i2); i2 += 16;
109       const int8x16_t vk2x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
110 
111       vprod01234567 = vmull_s8(vget_low_s8(vi2x0123456789ABCDEF), vget_low_s8(vk2x0123456789ABCDEF));
112       vprod89ABCDEF = vmull_s8(vget_high_s8(vi2x0123456789ABCDEF), vget_high_s8(vk2x0123456789ABCDEF));
113 
114       const int8x16_t vi3x0123456789ABCDEF = vld1q_s8(i3); i3 += 16;
115       const int8x16_t vk3x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
116 
117       vprod01234567 = vmlal_s8(vprod01234567, vget_low_s8(vi3x0123456789ABCDEF), vget_low_s8(vk3x0123456789ABCDEF));
118       vprod89ABCDEF = vmlal_s8(vprod89ABCDEF, vget_high_s8(vi3x0123456789ABCDEF), vget_high_s8(vk3x0123456789ABCDEF));
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 vi4x0123456789ABCDEF = vld1q_s8(i4); i4 += 16;
125       const int8x16_t vk4x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
126 
127       vprod01234567 = vmull_s8(vget_low_s8(vi4x0123456789ABCDEF), vget_low_s8(vk4x0123456789ABCDEF));
128       vprod89ABCDEF = vmull_s8(vget_high_s8(vi4x0123456789ABCDEF), vget_high_s8(vk4x0123456789ABCDEF));
129 
130       const int8x16_t vi5x0123456789ABCDEF = vld1q_s8(i5); i5 += 16;
131       const int8x16_t vk5x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
132 
133       vprod01234567 = vmlal_s8(vprod01234567, vget_low_s8(vi5x0123456789ABCDEF), vget_low_s8(vk5x0123456789ABCDEF));
134       vprod89ABCDEF = vmlal_s8(vprod89ABCDEF, vget_high_s8(vi5x0123456789ABCDEF), vget_high_s8(vk5x0123456789ABCDEF));
135 
136       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
137       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
138       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
139       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
140       const int8x16_t vi6x0123456789ABCDEF = vld1q_s8(i6); i6 += 16;
141       const int8x16_t vk6x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
142 
143       vprod01234567 = vmull_s8(vget_low_s8(vi6x0123456789ABCDEF), vget_low_s8(vk6x0123456789ABCDEF));
144       vprod89ABCDEF = vmull_s8(vget_high_s8(vi6x0123456789ABCDEF), vget_high_s8(vk6x0123456789ABCDEF));
145 
146       const int8x16_t vi7x0123456789ABCDEF = vld1q_s8(i7); i7 += 16;
147       const int8x16_t vk7x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
148 
149       vprod01234567 = vmlal_s8(vprod01234567, vget_low_s8(vi7x0123456789ABCDEF), vget_low_s8(vk7x0123456789ABCDEF));
150       vprod89ABCDEF = vmlal_s8(vprod89ABCDEF, vget_high_s8(vi7x0123456789ABCDEF), vget_high_s8(vk7x0123456789ABCDEF));
151 
152       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
153       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
154       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
155       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
156       const int8x16_t vi8x0123456789ABCDEF = vld1q_s8(i8); i8 += 16;
157       const int8x16_t vk8x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
158 
159       vprod01234567 = vmull_s8(vget_low_s8(vi8x0123456789ABCDEF), vget_low_s8(vk8x0123456789ABCDEF));
160       vprod89ABCDEF = vmull_s8(vget_high_s8(vi8x0123456789ABCDEF), vget_high_s8(vk8x0123456789ABCDEF));
161 
162       vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
163       vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
164       vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
165       vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
166 
167       float32x4_t vfpacc0123 = vcvtq_f32_s32(vacc0123);
168       float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567);
169       float32x4_t vfpacc89AB = vcvtq_f32_s32(vacc89AB);
170       float32x4_t vfpaccCDEF = vcvtq_f32_s32(vaccCDEF);
171 
172       const float32x4_t vscale0123 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
173       const float32x4_t vscale4567 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
174       const float32x4_t vscale89AB = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
175       const float32x4_t vscaleCDEF = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
176 
177       vfpacc0123 = vmulq_f32(vfpacc0123, vscale0123);
178       vfpacc4567 = vmulq_f32(vfpacc4567, vscale4567);
179       vfpacc89AB = vmulq_f32(vfpacc89AB, vscale89AB);
180       vfpaccCDEF = vmulq_f32(vfpaccCDEF, vscaleCDEF);
181 
182       vacc0123 = vcvtnq_s32_f32(vfpacc0123);
183       vacc4567 = vcvtnq_s32_f32(vfpacc4567);
184       vacc89AB = vcvtnq_s32_f32(vfpacc89AB);
185       vaccCDEF = vcvtnq_s32_f32(vfpaccCDEF);
186 
187 #if XNN_ARCH_ARM64
188       int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
189       int16x8_t vacc89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc89AB), vaccCDEF);
190 
191       vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
192       vacc89ABCDEF = vqaddq_s16(vacc89ABCDEF, voutput_zero_point);
193 
194       int8x16_t vout0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc01234567), vacc89ABCDEF);
195 #else  // !XNN_ARCH_ARM64
196       int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
197       int16x8_t vacc89ABCDEF = vcombine_s16(vqmovn_s32(vacc89AB), vqmovn_s32(vaccCDEF));
198 
199       vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
200       vacc89ABCDEF = vqaddq_s16(vacc89ABCDEF, voutput_zero_point);
201 
202       int8x16_t vout0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc01234567), vqmovn_s16(vacc89ABCDEF));
203 #endif  // !XNN_ARCH_ARM64
204 
205       vout0123456789ABCDEF = vmaxq_s8(vout0123456789ABCDEF, voutput_min);
206 
207       vout0123456789ABCDEF = vminq_s8(vout0123456789ABCDEF, voutput_max);
208 
209       vst1q_s8(output, vout0123456789ABCDEF); output += 16;
210     }
211     if XNN_UNLIKELY(c != 0) {
212       const int8_t* k = (const int8_t*) ((const int32_t*) w + 16);
213       do {
214         int32x4_t vacc0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
215         int32x4_t vacc4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
216 
217         const int8x8_t vi0x01234567 = vld1_s8(i0); i0 += 8;
218         const int8x8_t vk0x01234567 = vld1_s8(k); k += 8;
219 
220         int16x8_t vprod01234567 = vmull_s8(vi0x01234567, vk0x01234567);
221 
222         const int8x8_t vi1x01234567 = vld1_s8(i1); i1 += 8;
223         const int8x8_t vk1x01234567 = vld1_s8((const void*) (k + 8));
224 
225         vprod01234567 = vmlal_s8(vprod01234567, vi1x01234567, vk1x01234567);
226 
227         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
228         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
229         const int8x8_t vi2x01234567 = vld1_s8(i2); i2 += 8;
230         const int8x8_t vk2x01234567 = vld1_s8((const void*) (k + 24));
231 
232         vprod01234567 = vmull_s8(vi2x01234567, vk2x01234567);
233 
234         const int8x8_t vi3x01234567 = vld1_s8(i3); i3 += 8;
235         const int8x8_t vk3x01234567 = vld1_s8((const void*) (k + 40));
236 
237         vprod01234567 = vmlal_s8(vprod01234567, vi3x01234567, vk3x01234567);
238 
239         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
240         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
241         const int8x8_t vi4x01234567 = vld1_s8(i4); i4 += 8;
242         const int8x8_t vk4x01234567 = vld1_s8((const void*) (k + 56));
243 
244         vprod01234567 = vmull_s8(vi4x01234567, vk4x01234567);
245 
246         const int8x8_t vi5x01234567 = vld1_s8(i5); i5 += 8;
247         const int8x8_t vk5x01234567 = vld1_s8((const void*) (k + 72));
248 
249         vprod01234567 = vmlal_s8(vprod01234567, vi5x01234567, vk5x01234567);
250 
251         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
252         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
253         const int8x8_t vi6x01234567 = vld1_s8(i6); i6 += 8;
254         const int8x8_t vk6x01234567 = vld1_s8((const void*) (k + 88));
255 
256         vprod01234567 = vmull_s8(vi6x01234567, vk6x01234567);
257 
258         const int8x8_t vi7x01234567 = vld1_s8(i7); i7 += 8;
259         const int8x8_t vk7x01234567 = vld1_s8((const void*) (k + 104));
260 
261         vprod01234567 = vmlal_s8(vprod01234567, vi7x01234567, vk7x01234567);
262 
263         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
264         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
265         const int8x8_t vi8x01234567 = vld1_s8(i8); i8 += 8;
266         const int8x8_t vk8x01234567 = vld1_s8((const void*) (k + 120));
267 
268         vprod01234567 = vmull_s8(vi8x01234567, vk8x01234567);
269 
270         vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
271         vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
272 
273         float32x4_t vfpacc0123 = vcvtq_f32_s32(vacc0123);
274         float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567);
275 
276         const float32x4_t vscale0123 = vld1q_f32((const float*) ((uintptr_t) w + 8 * sizeof(int32_t) + 144 * sizeof(int8_t)));
277         const float32x4_t vscale4567 = vld1q_f32((const float*) ((uintptr_t) w + 8 * sizeof(int32_t) + 144 * sizeof(int8_t) + 4 * sizeof(float)));
278         vfpacc0123 = vmulq_f32(vfpacc0123, vscale0123);
279         vfpacc4567 = vmulq_f32(vfpacc4567, vscale4567);
280 
281         vacc0123 = vcvtnq_s32_f32(vfpacc0123);
282         vacc4567 = vcvtnq_s32_f32(vfpacc4567);
283 
284 #if XNN_ARCH_ARM64
285         int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
286 #else
287         int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
288 #endif
289         vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
290 
291         int8x8_t vout01234567 = vqmovn_s16(vacc01234567);
292         vout01234567 = vmax_s8(vout01234567, vget_low_s8(voutput_min));
293         vout01234567 = vmin_s8(vout01234567, vget_low_s8(voutput_max));
294 
295         if XNN_LIKELY(c >= 8) {
296           vst1_s8(output, vout01234567); output += 8;
297           c -= 8;
298         } else {
299           if (c & 4) {
300             vst1_lane_u32((void*) output, vreinterpret_u32_s8(vout01234567), 0); output += 4;
301             vout01234567 = vext_s8(vout01234567, vout01234567, 4);
302           }
303           if (c & 2) {
304             vst1_lane_u16((void*) output, vreinterpret_u16_s8(vout01234567), 0); output += 2;
305             vout01234567 = vext_s8(vout01234567, vout01234567, 2);
306           }
307           if (c & 1) {
308             vst1_lane_s8(output, vout01234567, 0); output += 1;
309           }
310           c = 0;
311         }
312       } while (c != 0);
313     }
314 
315     output = (int8_t*) ((uintptr_t) output + output_increment);
316   } while (--output_width != 0);
317 }
318