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_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_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qc8_dwconv_minmax_fp32_ukernel_up16x9__neonv8_mla8_ld64(
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(¶ms->fp32_neonv8.output_zero_point);
34 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->fp32_neonv8.output_min);
35 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->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 int8x8_t vi0x01234567 = vld1_s8(i0); i0 += 8;
93 const int8x8_t vk0x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
94 const int8x8_t vi0x89ABCDEF = vld1_s8(i0); i0 += 8;
95 const int8x8_t vk0x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
96
97 int16x8_t vprod01234567 = vmull_s8(vi0x01234567, vk0x01234567);
98 int16x8_t vprod89ABCDEF = vmull_s8(vi0x89ABCDEF, vk0x89ABCDEF);
99
100 const int8x8_t vi1x01234567 = vld1_s8(i1); i1 += 8;
101 const int8x8_t vk1x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
102 const int8x8_t vi1x89ABCDEF = vld1_s8(i1); i1 += 8;
103 const int8x8_t vk1x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
104
105 vprod01234567 = vmlal_s8(vprod01234567, vi1x01234567, vk1x01234567);
106 vprod89ABCDEF = vmlal_s8(vprod89ABCDEF, vi1x89ABCDEF, vk1x89ABCDEF);
107
108 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
109 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
110 vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
111 vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
112 const int8x8_t vi2x01234567 = vld1_s8(i2); i2 += 8;
113 const int8x8_t vk2x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
114 const int8x8_t vi2x89ABCDEF = vld1_s8(i2); i2 += 8;
115 const int8x8_t vk2x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
116
117 vprod01234567 = vmull_s8(vi2x01234567, vk2x01234567);
118 vprod89ABCDEF = vmull_s8(vi2x89ABCDEF, vk2x89ABCDEF);
119
120 const int8x8_t vi3x01234567 = vld1_s8(i3); i3 += 8;
121 const int8x8_t vk3x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
122 const int8x8_t vi3x89ABCDEF = vld1_s8(i3); i3 += 8;
123 const int8x8_t vk3x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
124
125 vprod01234567 = vmlal_s8(vprod01234567, vi3x01234567, vk3x01234567);
126 vprod89ABCDEF = vmlal_s8(vprod89ABCDEF, vi3x89ABCDEF, vk3x89ABCDEF);
127
128 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
129 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
130 vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
131 vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
132 const int8x8_t vi4x01234567 = vld1_s8(i4); i4 += 8;
133 const int8x8_t vk4x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
134 const int8x8_t vi4x89ABCDEF = vld1_s8(i4); i4 += 8;
135 const int8x8_t vk4x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
136
137 vprod01234567 = vmull_s8(vi4x01234567, vk4x01234567);
138 vprod89ABCDEF = vmull_s8(vi4x89ABCDEF, vk4x89ABCDEF);
139
140 const int8x8_t vi5x01234567 = vld1_s8(i5); i5 += 8;
141 const int8x8_t vk5x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
142 const int8x8_t vi5x89ABCDEF = vld1_s8(i5); i5 += 8;
143 const int8x8_t vk5x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
144
145 vprod01234567 = vmlal_s8(vprod01234567, vi5x01234567, vk5x01234567);
146 vprod89ABCDEF = vmlal_s8(vprod89ABCDEF, vi5x89ABCDEF, vk5x89ABCDEF);
147
148 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
149 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
150 vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
151 vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
152 const int8x8_t vi6x01234567 = vld1_s8(i6); i6 += 8;
153 const int8x8_t vk6x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
154 const int8x8_t vi6x89ABCDEF = vld1_s8(i6); i6 += 8;
155 const int8x8_t vk6x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
156
157 vprod01234567 = vmull_s8(vi6x01234567, vk6x01234567);
158 vprod89ABCDEF = vmull_s8(vi6x89ABCDEF, vk6x89ABCDEF);
159
160 const int8x8_t vi7x01234567 = vld1_s8(i7); i7 += 8;
161 const int8x8_t vk7x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
162 const int8x8_t vi7x89ABCDEF = vld1_s8(i7); i7 += 8;
163 const int8x8_t vk7x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
164
165 vprod01234567 = vmlal_s8(vprod01234567, vi7x01234567, vk7x01234567);
166 vprod89ABCDEF = vmlal_s8(vprod89ABCDEF, vi7x89ABCDEF, vk7x89ABCDEF);
167
168 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
169 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
170 vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
171 vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
172 const int8x8_t vi8x01234567 = vld1_s8(i8); i8 += 8;
173 const int8x8_t vk8x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
174 const int8x8_t vi8x89ABCDEF = vld1_s8(i8); i8 += 8;
175 const int8x8_t vk8x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
176
177 vprod01234567 = vmull_s8(vi8x01234567, vk8x01234567);
178 vprod89ABCDEF = vmull_s8(vi8x89ABCDEF, vk8x89ABCDEF);
179
180 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
181 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
182 vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
183 vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
184
185 float32x4_t vfpacc0123 = vcvtq_f32_s32(vacc0123);
186 float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567);
187 float32x4_t vfpacc89AB = vcvtq_f32_s32(vacc89AB);
188 float32x4_t vfpaccCDEF = vcvtq_f32_s32(vaccCDEF);
189
190 const float32x4_t vscale0123 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
191 const float32x4_t vscale4567 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
192 const float32x4_t vscale89AB = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
193 const float32x4_t vscaleCDEF = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
194
195 vfpacc0123 = vmulq_f32(vfpacc0123, vscale0123);
196 vfpacc4567 = vmulq_f32(vfpacc4567, vscale4567);
197 vfpacc89AB = vmulq_f32(vfpacc89AB, vscale89AB);
198 vfpaccCDEF = vmulq_f32(vfpaccCDEF, vscaleCDEF);
199
200 vacc0123 = vcvtnq_s32_f32(vfpacc0123);
201 vacc4567 = vcvtnq_s32_f32(vfpacc4567);
202 vacc89AB = vcvtnq_s32_f32(vfpacc89AB);
203 vaccCDEF = vcvtnq_s32_f32(vfpaccCDEF);
204
205 #if XNN_ARCH_ARM64
206 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
207 int16x8_t vacc89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc89AB), vaccCDEF);
208
209 vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
210 vacc89ABCDEF = vqaddq_s16(vacc89ABCDEF, voutput_zero_point);
211
212 int8x16_t vout0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc01234567), vacc89ABCDEF);
213 #else // !XNN_ARCH_ARM64
214 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
215 int16x8_t vacc89ABCDEF = vcombine_s16(vqmovn_s32(vacc89AB), vqmovn_s32(vaccCDEF));
216
217 vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
218 vacc89ABCDEF = vqaddq_s16(vacc89ABCDEF, voutput_zero_point);
219
220 int8x16_t vout0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc01234567), vqmovn_s16(vacc89ABCDEF));
221 #endif // !XNN_ARCH_ARM64
222
223 vout0123456789ABCDEF = vmaxq_s8(vout0123456789ABCDEF, voutput_min);
224
225 vout0123456789ABCDEF = vminq_s8(vout0123456789ABCDEF, voutput_max);
226
227 vst1q_s8(output, vout0123456789ABCDEF); output += 16;
228 }
229 if XNN_UNLIKELY(c != 0) {
230 const int8_t* k = (const int8_t*) ((const int32_t*) w + 16);
231 do {
232 int32x4_t vacc0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
233 int32x4_t vacc4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
234
235 const int8x8_t vi0x01234567 = vld1_s8(i0); i0 += 8;
236 const int8x8_t vk0x01234567 = vld1_s8(k); k += 8;
237
238 int16x8_t vprod01234567 = vmull_s8(vi0x01234567, vk0x01234567);
239
240 const int8x8_t vi1x01234567 = vld1_s8(i1); i1 += 8;
241 const int8x8_t vk1x01234567 = vld1_s8((const void*) (k + 8));
242
243 vprod01234567 = vmlal_s8(vprod01234567, vi1x01234567, vk1x01234567);
244
245 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
246 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
247 const int8x8_t vi2x01234567 = vld1_s8(i2); i2 += 8;
248 const int8x8_t vk2x01234567 = vld1_s8((const void*) (k + 24));
249
250 vprod01234567 = vmull_s8(vi2x01234567, vk2x01234567);
251
252 const int8x8_t vi3x01234567 = vld1_s8(i3); i3 += 8;
253 const int8x8_t vk3x01234567 = vld1_s8((const void*) (k + 40));
254
255 vprod01234567 = vmlal_s8(vprod01234567, vi3x01234567, vk3x01234567);
256
257 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
258 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
259 const int8x8_t vi4x01234567 = vld1_s8(i4); i4 += 8;
260 const int8x8_t vk4x01234567 = vld1_s8((const void*) (k + 56));
261
262 vprod01234567 = vmull_s8(vi4x01234567, vk4x01234567);
263
264 const int8x8_t vi5x01234567 = vld1_s8(i5); i5 += 8;
265 const int8x8_t vk5x01234567 = vld1_s8((const void*) (k + 72));
266
267 vprod01234567 = vmlal_s8(vprod01234567, vi5x01234567, vk5x01234567);
268
269 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
270 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
271 const int8x8_t vi6x01234567 = vld1_s8(i6); i6 += 8;
272 const int8x8_t vk6x01234567 = vld1_s8((const void*) (k + 88));
273
274 vprod01234567 = vmull_s8(vi6x01234567, vk6x01234567);
275
276 const int8x8_t vi7x01234567 = vld1_s8(i7); i7 += 8;
277 const int8x8_t vk7x01234567 = vld1_s8((const void*) (k + 104));
278
279 vprod01234567 = vmlal_s8(vprod01234567, vi7x01234567, vk7x01234567);
280
281 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
282 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
283 const int8x8_t vi8x01234567 = vld1_s8(i8); i8 += 8;
284 const int8x8_t vk8x01234567 = vld1_s8((const void*) (k + 120));
285
286 vprod01234567 = vmull_s8(vi8x01234567, vk8x01234567);
287
288 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
289 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
290
291 float32x4_t vfpacc0123 = vcvtq_f32_s32(vacc0123);
292 float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567);
293
294 const float32x4_t vscale0123 = vld1q_f32((const float*) ((uintptr_t) w + 8 * sizeof(int32_t) + 144 * sizeof(int8_t)));
295 const float32x4_t vscale4567 = vld1q_f32((const float*) ((uintptr_t) w + 8 * sizeof(int32_t) + 144 * sizeof(int8_t) + 4 * sizeof(float)));
296 vfpacc0123 = vmulq_f32(vfpacc0123, vscale0123);
297 vfpacc4567 = vmulq_f32(vfpacc4567, vscale4567);
298
299 vacc0123 = vcvtnq_s32_f32(vfpacc0123);
300 vacc4567 = vcvtnq_s32_f32(vfpacc4567);
301
302 #if XNN_ARCH_ARM64
303 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
304 #else
305 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
306 #endif
307 vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
308
309 int8x8_t vout01234567 = vqmovn_s16(vacc01234567);
310 vout01234567 = vmax_s8(vout01234567, vget_low_s8(voutput_min));
311 vout01234567 = vmin_s8(vout01234567, vget_low_s8(voutput_max));
312
313 if XNN_LIKELY(c >= 8) {
314 vst1_s8(output, vout01234567); output += 8;
315 c -= 8;
316 } else {
317 if (c & 4) {
318 vst1_lane_u32((void*) output, vreinterpret_u32_s8(vout01234567), 0); output += 4;
319 vout01234567 = vext_s8(vout01234567, vout01234567, 4);
320 }
321 if (c & 2) {
322 vst1_lane_u16((void*) output, vreinterpret_u16_s8(vout01234567), 0); output += 2;
323 vout01234567 = vext_s8(vout01234567, vout01234567, 2);
324 }
325 if (c & 1) {
326 vst1_lane_s8(output, vout01234567, 0); output += 1;
327 }
328 c = 0;
329 }
330 } while (c != 0);
331 }
332
333 output = (int8_t*) ((uintptr_t) output + output_increment);
334 } while (--output_width != 0);
335 }
336