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_up8x9__neon_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)])17 void xnn_qc8_dwconv_minmax_fp32_ukernel_up8x9__neon_mla8_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 int8x8_t voutput_min = vld1_dup_s8(¶ms->fp32_neon.output_min);
35 const int8x8_t voutput_max = vld1_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 >= 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 const int8x8_t vi0x01234567 = vld1_s8(i0); i0 += 8;
91 const int8x8_t vk0x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
92
93 int16x8_t vprod01234567 = vmull_s8(vi0x01234567, vk0x01234567);
94
95 const int8x8_t vi1x01234567 = vld1_s8(i1); i1 += 8;
96 const int8x8_t vk1x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
97
98 vprod01234567 = vmlal_s8(vprod01234567, vi1x01234567, vk1x01234567);
99
100 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
101 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
102 const int8x8_t vi2x01234567 = vld1_s8(i2); i2 += 8;
103 const int8x8_t vk2x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
104
105 vprod01234567 = vmull_s8(vi2x01234567, vk2x01234567);
106
107 const int8x8_t vi3x01234567 = vld1_s8(i3); i3 += 8;
108 const int8x8_t vk3x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
109
110 vprod01234567 = vmlal_s8(vprod01234567, vi3x01234567, vk3x01234567);
111
112 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
113 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
114 const int8x8_t vi4x01234567 = vld1_s8(i4); i4 += 8;
115 const int8x8_t vk4x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
116
117 vprod01234567 = vmull_s8(vi4x01234567, vk4x01234567);
118
119 const int8x8_t vi5x01234567 = vld1_s8(i5); i5 += 8;
120 const int8x8_t vk5x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
121
122 vprod01234567 = vmlal_s8(vprod01234567, vi5x01234567, vk5x01234567);
123
124 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
125 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
126 const int8x8_t vi6x01234567 = vld1_s8(i6); i6 += 8;
127 const int8x8_t vk6x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
128
129 vprod01234567 = vmull_s8(vi6x01234567, vk6x01234567);
130
131 const int8x8_t vi7x01234567 = vld1_s8(i7); i7 += 8;
132 const int8x8_t vk7x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
133
134 vprod01234567 = vmlal_s8(vprod01234567, vi7x01234567, vk7x01234567);
135
136 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
137 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
138 const int8x8_t vi8x01234567 = vld1_s8(i8); i8 += 8;
139 const int8x8_t vk8x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
140
141 vprod01234567 = vmull_s8(vi8x01234567, vk8x01234567);
142
143 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
144 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
145
146 float32x4_t vfpacc0123 = vcvtq_f32_s32(vacc0123);
147 float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567);
148
149 const float32x4_t vscale0123 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
150 const float32x4_t vscale4567 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
151
152 vfpacc0123 = vmulq_f32(vfpacc0123, vscale0123);
153 vfpacc4567 = vmulq_f32(vfpacc4567, vscale4567);
154
155 vacc0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0123, vmagic_bias));
156 vacc4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc4567, vmagic_bias));
157
158 vacc0123 = vqsubq_s32(vacc0123, vmagic_bias_less_output_zero_point);
159 vacc4567 = vqsubq_s32(vacc4567, vmagic_bias_less_output_zero_point);
160
161 #if XNN_ARCH_ARM64
162 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
163
164
165 int8x8_t vout01234567 = vqmovn_s16(vacc01234567);
166 #else // !XNN_ARCH_ARM64
167 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
168
169
170 int8x8_t vout01234567 = vqmovn_s16(vacc01234567);
171 #endif // !XNN_ARCH_ARM64
172
173 vout01234567 = vmax_s8(vout01234567, voutput_min);
174
175 vout01234567 = vmin_s8(vout01234567, voutput_max);
176
177 vst1_s8(output, vout01234567); output += 8;
178 }
179 if XNN_UNLIKELY(c != 0) {
180 {
181 int32x4_t vacc0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
182 int32x4_t vacc4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
183
184 const int8x8_t vi0x01234567 = vld1_s8(i0);
185 const int8x8_t vk0x01234567 = vld1_s8(w);
186
187 int16x8_t vprod01234567 = vmull_s8(vi0x01234567, vk0x01234567);
188
189 const int8x8_t vi1x01234567 = vld1_s8(i1);
190 const int8x8_t vk1x01234567 = vld1_s8((const void*) ((const int8_t*) w + 8));
191
192 vprod01234567 = vmlal_s8(vprod01234567, vi1x01234567, vk1x01234567);
193
194 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
195 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
196 const int8x8_t vi2x01234567 = vld1_s8(i2);
197 const int8x8_t vk2x01234567 = vld1_s8((const void*) ((const int8_t*) w + 16));
198
199 vprod01234567 = vmull_s8(vi2x01234567, vk2x01234567);
200
201 const int8x8_t vi3x01234567 = vld1_s8(i3);
202 const int8x8_t vk3x01234567 = vld1_s8((const void*) ((const int8_t*) w + 24));
203
204 vprod01234567 = vmlal_s8(vprod01234567, vi3x01234567, vk3x01234567);
205
206 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
207 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
208 const int8x8_t vi4x01234567 = vld1_s8(i4);
209 const int8x8_t vk4x01234567 = vld1_s8((const void*) ((const int8_t*) w + 32));
210
211 vprod01234567 = vmull_s8(vi4x01234567, vk4x01234567);
212
213 const int8x8_t vi5x01234567 = vld1_s8(i5);
214 const int8x8_t vk5x01234567 = vld1_s8((const void*) ((const int8_t*) w + 40));
215
216 vprod01234567 = vmlal_s8(vprod01234567, vi5x01234567, vk5x01234567);
217
218 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
219 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
220 const int8x8_t vi6x01234567 = vld1_s8(i6);
221 const int8x8_t vk6x01234567 = vld1_s8((const void*) ((const int8_t*) w + 48));
222
223 vprod01234567 = vmull_s8(vi6x01234567, vk6x01234567);
224
225 const int8x8_t vi7x01234567 = vld1_s8(i7);
226 const int8x8_t vk7x01234567 = vld1_s8((const void*) ((const int8_t*) w + 56));
227
228 vprod01234567 = vmlal_s8(vprod01234567, vi7x01234567, vk7x01234567);
229
230 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
231 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
232 const int8x8_t vi8x01234567 = vld1_s8(i8);
233 const int8x8_t vk8x01234567 = vld1_s8((const void*) ((const int8_t*) w + 64));
234
235 vprod01234567 = vmull_s8(vi8x01234567, vk8x01234567);
236
237 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
238 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
239
240 float32x4_t vfpacc0123 = vcvtq_f32_s32(vacc0123);
241 float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567);
242
243 const float32x4_t vscale0123 = vld1q_f32((const float*) ((uintptr_t) w + 0 * sizeof(int32_t) + 72 * sizeof(int8_t)));
244 const float32x4_t vscale4567 = vld1q_f32((const float*) ((uintptr_t) w + 0 * sizeof(int32_t) + 72 * sizeof(int8_t) + 4 * sizeof(float)));
245 vfpacc0123 = vmulq_f32(vfpacc0123, vscale0123);
246 vfpacc4567 = vmulq_f32(vfpacc4567, vscale4567);
247
248 vacc0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0123, vmagic_bias));
249 vacc4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc4567, vmagic_bias));
250
251 vacc0123 = vqsubq_s32(vacc0123, vmagic_bias_less_output_zero_point);
252 vacc4567 = vqsubq_s32(vacc4567, vmagic_bias_less_output_zero_point);
253
254 #if XNN_ARCH_ARM64
255 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
256 #else
257 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
258 #endif
259
260 int8x8_t vout01234567 = vqmovn_s16(vacc01234567);
261 vout01234567 = vmax_s8(vout01234567, voutput_min);
262 vout01234567 = vmin_s8(vout01234567, voutput_max);
263
264 if (c & 4) {
265 vst1_lane_u32((void*) output, vreinterpret_u32_s8(vout01234567), 0); output += 4;
266 vout01234567 = vext_s8(vout01234567, vout01234567, 4);
267 }
268 if (c & 2) {
269 vst1_lane_u16((void*) output, vreinterpret_u16_s8(vout01234567), 0); output += 2;
270 vout01234567 = vext_s8(vout01234567, vout01234567, 2);
271 }
272 if (c & 1) {
273 vst1_lane_s8(output, vout01234567, 0); output += 1;
274 }
275 }
276 }
277
278 output = (int8_t*) ((uintptr_t) output + output_increment);
279 } while (--output_width != 0);
280 }
281