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_qc8_dwconv_minmax_fp32_ukernel_up16x9__neon_mul8_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)])17 void xnn_qc8_dwconv_minmax_fp32_ukernel_up16x9__neon_mul8_ld64(
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_qc8_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 float32x4_t vmagic_bias = vld1q_dup_f32(¶ms->fp32_neon.magic_bias);
33 const int32x4_t vmagic_bias_less_output_zero_point = vld1q_dup_s32(¶ms->fp32_neon.magic_bias_less_output_zero_point);
34 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->fp32_neon.output_min);
35 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->fp32_neon.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 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
101 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
102 vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
103 vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
104 const int8x8_t vi1x01234567 = vld1_s8(i1); i1 += 8;
105 const int8x8_t vk1x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
106 const int8x8_t vi1x89ABCDEF = vld1_s8(i1); i1 += 8;
107 const int8x8_t vk1x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
108
109 vprod01234567 = vmull_s8(vi1x01234567, vk1x01234567);
110 vprod89ABCDEF = vmull_s8(vi1x89ABCDEF, vk1x89ABCDEF);
111
112 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
113 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
114 vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
115 vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
116 const int8x8_t vi2x01234567 = vld1_s8(i2); i2 += 8;
117 const int8x8_t vk2x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
118 const int8x8_t vi2x89ABCDEF = vld1_s8(i2); i2 += 8;
119 const int8x8_t vk2x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
120
121 vprod01234567 = vmull_s8(vi2x01234567, vk2x01234567);
122 vprod89ABCDEF = vmull_s8(vi2x89ABCDEF, vk2x89ABCDEF);
123
124 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
125 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
126 vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
127 vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
128 const int8x8_t vi3x01234567 = vld1_s8(i3); i3 += 8;
129 const int8x8_t vk3x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
130 const int8x8_t vi3x89ABCDEF = vld1_s8(i3); i3 += 8;
131 const int8x8_t vk3x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
132
133 vprod01234567 = vmull_s8(vi3x01234567, vk3x01234567);
134 vprod89ABCDEF = vmull_s8(vi3x89ABCDEF, vk3x89ABCDEF);
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 int8x8_t vi4x01234567 = vld1_s8(i4); i4 += 8;
141 const int8x8_t vk4x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
142 const int8x8_t vi4x89ABCDEF = vld1_s8(i4); i4 += 8;
143 const int8x8_t vk4x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
144
145 vprod01234567 = vmull_s8(vi4x01234567, vk4x01234567);
146 vprod89ABCDEF = vmull_s8(vi4x89ABCDEF, vk4x89ABCDEF);
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 vi5x01234567 = vld1_s8(i5); i5 += 8;
153 const int8x8_t vk5x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
154 const int8x8_t vi5x89ABCDEF = vld1_s8(i5); i5 += 8;
155 const int8x8_t vk5x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
156
157 vprod01234567 = vmull_s8(vi5x01234567, vk5x01234567);
158 vprod89ABCDEF = vmull_s8(vi5x89ABCDEF, vk5x89ABCDEF);
159
160 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
161 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
162 vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
163 vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
164 const int8x8_t vi6x01234567 = vld1_s8(i6); i6 += 8;
165 const int8x8_t vk6x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
166 const int8x8_t vi6x89ABCDEF = vld1_s8(i6); i6 += 8;
167 const int8x8_t vk6x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
168
169 vprod01234567 = vmull_s8(vi6x01234567, vk6x01234567);
170 vprod89ABCDEF = vmull_s8(vi6x89ABCDEF, vk6x89ABCDEF);
171
172 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
173 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
174 vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
175 vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
176 const int8x8_t vi7x01234567 = vld1_s8(i7); i7 += 8;
177 const int8x8_t vk7x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
178 const int8x8_t vi7x89ABCDEF = vld1_s8(i7); i7 += 8;
179 const int8x8_t vk7x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
180
181 vprod01234567 = vmull_s8(vi7x01234567, vk7x01234567);
182 vprod89ABCDEF = vmull_s8(vi7x89ABCDEF, vk7x89ABCDEF);
183
184 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
185 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
186 vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
187 vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
188 const int8x8_t vi8x01234567 = vld1_s8(i8); i8 += 8;
189 const int8x8_t vk8x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
190 const int8x8_t vi8x89ABCDEF = vld1_s8(i8); i8 += 8;
191 const int8x8_t vk8x89ABCDEF = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
192
193 vprod01234567 = vmull_s8(vi8x01234567, vk8x01234567);
194 vprod89ABCDEF = vmull_s8(vi8x89ABCDEF, vk8x89ABCDEF);
195
196 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
197 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
198 vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
199 vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
200
201 float32x4_t vfpacc0123 = vcvtq_f32_s32(vacc0123);
202 float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567);
203 float32x4_t vfpacc89AB = vcvtq_f32_s32(vacc89AB);
204 float32x4_t vfpaccCDEF = vcvtq_f32_s32(vaccCDEF);
205
206 const float32x4_t vscale0123 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
207 const float32x4_t vscale4567 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
208 const float32x4_t vscale89AB = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
209 const float32x4_t vscaleCDEF = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
210
211 vfpacc0123 = vmulq_f32(vfpacc0123, vscale0123);
212 vfpacc4567 = vmulq_f32(vfpacc4567, vscale4567);
213 vfpacc89AB = vmulq_f32(vfpacc89AB, vscale89AB);
214 vfpaccCDEF = vmulq_f32(vfpaccCDEF, vscaleCDEF);
215
216 vacc0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0123, vmagic_bias));
217 vacc4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc4567, vmagic_bias));
218 vacc89AB = vreinterpretq_s32_f32(vaddq_f32(vfpacc89AB, vmagic_bias));
219 vaccCDEF = vreinterpretq_s32_f32(vaddq_f32(vfpaccCDEF, vmagic_bias));
220
221 vacc0123 = vqsubq_s32(vacc0123, vmagic_bias_less_output_zero_point);
222 vacc4567 = vqsubq_s32(vacc4567, vmagic_bias_less_output_zero_point);
223 vacc89AB = vqsubq_s32(vacc89AB, vmagic_bias_less_output_zero_point);
224 vaccCDEF = vqsubq_s32(vaccCDEF, vmagic_bias_less_output_zero_point);
225
226 #if XNN_ARCH_ARM64
227 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
228 int16x8_t vacc89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc89AB), vaccCDEF);
229
230
231 int8x16_t vout0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc01234567), vacc89ABCDEF);
232 #else // !XNN_ARCH_ARM64
233 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
234 int16x8_t vacc89ABCDEF = vcombine_s16(vqmovn_s32(vacc89AB), vqmovn_s32(vaccCDEF));
235
236
237 int8x16_t vout0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc01234567), vqmovn_s16(vacc89ABCDEF));
238 #endif // !XNN_ARCH_ARM64
239
240 vout0123456789ABCDEF = vmaxq_s8(vout0123456789ABCDEF, voutput_min);
241
242 vout0123456789ABCDEF = vminq_s8(vout0123456789ABCDEF, voutput_max);
243
244 vst1q_s8(output, vout0123456789ABCDEF); output += 16;
245 }
246 if XNN_UNLIKELY(c != 0) {
247 const int8_t* k = (const int8_t*) ((const int32_t*) w + 16);
248 do {
249 int32x4_t vacc0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
250 int32x4_t vacc4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
251
252 const int8x8_t vi0x01234567 = vld1_s8(i0); i0 += 8;
253 const int8x8_t vk0x01234567 = vld1_s8(k); k += 8;
254
255 int16x8_t vprod01234567 = vmull_s8(vi0x01234567, vk0x01234567);
256
257 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
258 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
259 const int8x8_t vi1x01234567 = vld1_s8(i1); i1 += 8;
260 const int8x8_t vk1x01234567 = vld1_s8((const void*) (k + 8));
261
262 vprod01234567 = vmull_s8(vi1x01234567, vk1x01234567);
263
264 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
265 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
266 const int8x8_t vi2x01234567 = vld1_s8(i2); i2 += 8;
267 const int8x8_t vk2x01234567 = vld1_s8((const void*) (k + 24));
268
269 vprod01234567 = vmull_s8(vi2x01234567, vk2x01234567);
270
271 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
272 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
273 const int8x8_t vi3x01234567 = vld1_s8(i3); i3 += 8;
274 const int8x8_t vk3x01234567 = vld1_s8((const void*) (k + 40));
275
276 vprod01234567 = vmull_s8(vi3x01234567, vk3x01234567);
277
278 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
279 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
280 const int8x8_t vi4x01234567 = vld1_s8(i4); i4 += 8;
281 const int8x8_t vk4x01234567 = vld1_s8((const void*) (k + 56));
282
283 vprod01234567 = vmull_s8(vi4x01234567, vk4x01234567);
284
285 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
286 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
287 const int8x8_t vi5x01234567 = vld1_s8(i5); i5 += 8;
288 const int8x8_t vk5x01234567 = vld1_s8((const void*) (k + 72));
289
290 vprod01234567 = vmull_s8(vi5x01234567, vk5x01234567);
291
292 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
293 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
294 const int8x8_t vi6x01234567 = vld1_s8(i6); i6 += 8;
295 const int8x8_t vk6x01234567 = vld1_s8((const void*) (k + 88));
296
297 vprod01234567 = vmull_s8(vi6x01234567, vk6x01234567);
298
299 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
300 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
301 const int8x8_t vi7x01234567 = vld1_s8(i7); i7 += 8;
302 const int8x8_t vk7x01234567 = vld1_s8((const void*) (k + 104));
303
304 vprod01234567 = vmull_s8(vi7x01234567, vk7x01234567);
305
306 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
307 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
308 const int8x8_t vi8x01234567 = vld1_s8(i8); i8 += 8;
309 const int8x8_t vk8x01234567 = vld1_s8((const void*) (k + 120));
310
311 vprod01234567 = vmull_s8(vi8x01234567, vk8x01234567);
312
313 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
314 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
315
316 float32x4_t vfpacc0123 = vcvtq_f32_s32(vacc0123);
317 float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567);
318
319 const float32x4_t vscale0123 = vld1q_f32((const float*) ((uintptr_t) w + 8 * sizeof(int32_t) + 144 * sizeof(int8_t)));
320 const float32x4_t vscale4567 = vld1q_f32((const float*) ((uintptr_t) w + 8 * sizeof(int32_t) + 144 * sizeof(int8_t) + 4 * sizeof(float)));
321 vfpacc0123 = vmulq_f32(vfpacc0123, vscale0123);
322 vfpacc4567 = vmulq_f32(vfpacc4567, vscale4567);
323
324 vacc0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0123, vmagic_bias));
325 vacc4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc4567, vmagic_bias));
326
327 vacc0123 = vqsubq_s32(vacc0123, vmagic_bias_less_output_zero_point);
328 vacc4567 = vqsubq_s32(vacc4567, vmagic_bias_less_output_zero_point);
329
330 #if XNN_ARCH_ARM64
331 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
332 #else
333 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
334 #endif
335
336 int8x8_t vout01234567 = vqmovn_s16(vacc01234567);
337 vout01234567 = vmax_s8(vout01234567, vget_low_s8(voutput_min));
338 vout01234567 = vmin_s8(vout01234567, vget_low_s8(voutput_max));
339
340 if XNN_LIKELY(c >= 8) {
341 vst1_s8(output, vout01234567); output += 8;
342 c -= 8;
343 } else {
344 if (c & 4) {
345 vst1_lane_u32((void*) output, vreinterpret_u32_s8(vout01234567), 0); output += 4;
346 vout01234567 = vext_s8(vout01234567, vout01234567, 4);
347 }
348 if (c & 2) {
349 vst1_lane_u16((void*) output, vreinterpret_u16_s8(vout01234567), 0); output += 2;
350 vout01234567 = vext_s8(vout01234567, vout01234567, 2);
351 }
352 if (c & 1) {
353 vst1_lane_s8(output, vout01234567, 0); output += 1;
354 }
355 c = 0;
356 }
357 } while (c != 0);
358 }
359
360 output = (int8_t*) ((uintptr_t) output + output_increment);
361 } while (--output_width != 0);
362 }
363