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
16
xnn_qs8_dwconv_minmax_fp32_ukernel_up8x9__neon_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_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qs8_dwconv_minmax_fp32_ukernel_up8x9__neon_mul16(
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_qs8_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 vscale = vld1q_dup_f32(¶ms->fp32_neon.scale);
33 const float32x4_t vmagic_bias = vld1q_dup_f32(¶ms->fp32_neon.magic_bias);
34 const int32x4_t vmagic_bias_less_output_zero_point = vld1q_dup_s32(¶ms->fp32_neon.magic_bias_less_output_zero_point);
35 const int8x8_t voutput_min = vld1_dup_s8(¶ms->fp32_neon.output_min);
36 const int8x8_t voutput_max = vld1_dup_s8(¶ms->fp32_neon.output_max);
37 do {
38 const int8_t* i0 = input[0];
39 assert(i0 != NULL);
40 if XNN_UNPREDICTABLE(i0 != zero) {
41 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
42 }
43 const int8_t* i1 = input[1];
44 assert(i1 != NULL);
45 if XNN_UNPREDICTABLE(i1 != zero) {
46 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
47 }
48 const int8_t* i2 = input[2];
49 assert(i2 != NULL);
50 if XNN_UNPREDICTABLE(i2 != zero) {
51 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
52 }
53 const int8_t* i3 = input[3];
54 assert(i3 != NULL);
55 if XNN_UNPREDICTABLE(i3 != zero) {
56 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
57 }
58 const int8_t* i4 = input[4];
59 assert(i4 != NULL);
60 if XNN_UNPREDICTABLE(i4 != zero) {
61 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
62 }
63 const int8_t* i5 = input[5];
64 assert(i5 != NULL);
65 if XNN_UNPREDICTABLE(i5 != zero) {
66 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
67 }
68 const int8_t* i6 = input[6];
69 assert(i6 != NULL);
70 if XNN_UNPREDICTABLE(i6 != zero) {
71 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
72 }
73 const int8_t* i7 = input[7];
74 assert(i7 != NULL);
75 if XNN_UNPREDICTABLE(i7 != zero) {
76 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
77 }
78 const int8_t* i8 = input[8];
79 assert(i8 != NULL);
80 if XNN_UNPREDICTABLE(i8 != zero) {
81 i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
82 }
83 input = (const int8_t**) ((uintptr_t) input + input_stride);
84
85 size_t c = channels;
86 const void* w = weights;
87 for (; c >= 8; c -= 8) {
88 int32x4_t vacc0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
89 int32x4_t vacc4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
90
91
92 const int16x8_t vi0x01234567 = vmovl_s8(vld1_s8(i0)); i0 += 8;
93 const int16x8_t vk0x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
94
95 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi0x01234567), vget_low_s16(vk0x01234567));
96 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi0x01234567), vget_high_s16(vk0x01234567));
97
98 const int16x8_t vi1x01234567 = vmovl_s8(vld1_s8(i1)); i1 += 8;
99 const int16x8_t vk1x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
100
101 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi1x01234567), vget_low_s16(vk1x01234567));
102 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi1x01234567), vget_high_s16(vk1x01234567));
103
104 const int16x8_t vi2x01234567 = vmovl_s8(vld1_s8(i2)); i2 += 8;
105 const int16x8_t vk2x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
106
107 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi2x01234567), vget_low_s16(vk2x01234567));
108 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi2x01234567), vget_high_s16(vk2x01234567));
109
110 const int16x8_t vi3x01234567 = vmovl_s8(vld1_s8(i3)); i3 += 8;
111 const int16x8_t vk3x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
112
113 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi3x01234567), vget_low_s16(vk3x01234567));
114 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi3x01234567), vget_high_s16(vk3x01234567));
115
116 const int16x8_t vi4x01234567 = vmovl_s8(vld1_s8(i4)); i4 += 8;
117 const int16x8_t vk4x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
118
119 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi4x01234567), vget_low_s16(vk4x01234567));
120 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi4x01234567), vget_high_s16(vk4x01234567));
121
122 const int16x8_t vi5x01234567 = vmovl_s8(vld1_s8(i5)); i5 += 8;
123 const int16x8_t vk5x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
124
125 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi5x01234567), vget_low_s16(vk5x01234567));
126 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi5x01234567), vget_high_s16(vk5x01234567));
127
128 const int16x8_t vi6x01234567 = vmovl_s8(vld1_s8(i6)); i6 += 8;
129 const int16x8_t vk6x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
130
131 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi6x01234567), vget_low_s16(vk6x01234567));
132 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi6x01234567), vget_high_s16(vk6x01234567));
133
134 const int16x8_t vi7x01234567 = vmovl_s8(vld1_s8(i7)); i7 += 8;
135 const int16x8_t vk7x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
136
137 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi7x01234567), vget_low_s16(vk7x01234567));
138 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi7x01234567), vget_high_s16(vk7x01234567));
139
140 const int16x8_t vi8x01234567 = vmovl_s8(vld1_s8(i8)); i8 += 8;
141 const int16x8_t vk8x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
142
143 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi8x01234567), vget_low_s16(vk8x01234567));
144 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi8x01234567), vget_high_s16(vk8x01234567));
145
146 float32x4_t vfpacc0123 = vcvtq_f32_s32(vacc0123);
147 float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567);
148
149 vfpacc0123 = vmulq_f32(vfpacc0123, vscale);
150 vfpacc4567 = vmulq_f32(vfpacc4567, vscale);
151
152 vacc0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0123, vmagic_bias));
153 vacc4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc4567, vmagic_bias));
154
155 vacc0123 = vqsubq_s32(vacc0123, vmagic_bias_less_output_zero_point);
156 vacc4567 = vqsubq_s32(vacc4567, vmagic_bias_less_output_zero_point);
157
158 #if XNN_ARCH_ARM64
159 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
160
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
167 int8x8_t vout01234567 = vqmovn_s16(vacc01234567);
168 #endif // !XNN_ARCH_ARM64
169
170 vout01234567 = vmax_s8(vout01234567, voutput_min);
171
172 vout01234567 = vmin_s8(vout01234567, voutput_max);
173
174 vst1_s8(output, vout01234567); output += 8;
175 }
176 if XNN_UNLIKELY(c != 0) {
177 {
178 int32x4_t vacc0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
179 int32x4_t vacc4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
180
181 const int16x8_t vi0x01234567 = vmovl_s8(vld1_s8(i0));
182 const int16x8_t vk0x01234567 = vmovl_s8(vld1_s8(w));
183
184 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi0x01234567), vget_low_s16(vk0x01234567));
185 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi0x01234567), vget_high_s16(vk0x01234567));
186 const int16x8_t vi1x01234567 = vmovl_s8(vld1_s8(i1));
187 const int16x8_t vk1x01234567 = vmovl_s8(vld1_s8((const void*) ((const int8_t*) w + 8)));
188
189 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi1x01234567), vget_low_s16(vk1x01234567));
190 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi1x01234567), vget_high_s16(vk1x01234567));
191 const int16x8_t vi2x01234567 = vmovl_s8(vld1_s8(i2));
192 const int16x8_t vk2x01234567 = vmovl_s8(vld1_s8((const void*) ((const int8_t*) w + 16)));
193
194 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi2x01234567), vget_low_s16(vk2x01234567));
195 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi2x01234567), vget_high_s16(vk2x01234567));
196 const int16x8_t vi3x01234567 = vmovl_s8(vld1_s8(i3));
197 const int16x8_t vk3x01234567 = vmovl_s8(vld1_s8((const void*) ((const int8_t*) w + 24)));
198
199 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi3x01234567), vget_low_s16(vk3x01234567));
200 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi3x01234567), vget_high_s16(vk3x01234567));
201 const int16x8_t vi4x01234567 = vmovl_s8(vld1_s8(i4));
202 const int16x8_t vk4x01234567 = vmovl_s8(vld1_s8((const void*) ((const int8_t*) w + 32)));
203
204 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi4x01234567), vget_low_s16(vk4x01234567));
205 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi4x01234567), vget_high_s16(vk4x01234567));
206 const int16x8_t vi5x01234567 = vmovl_s8(vld1_s8(i5));
207 const int16x8_t vk5x01234567 = vmovl_s8(vld1_s8((const void*) ((const int8_t*) w + 40)));
208
209 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi5x01234567), vget_low_s16(vk5x01234567));
210 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi5x01234567), vget_high_s16(vk5x01234567));
211 const int16x8_t vi6x01234567 = vmovl_s8(vld1_s8(i6));
212 const int16x8_t vk6x01234567 = vmovl_s8(vld1_s8((const void*) ((const int8_t*) w + 48)));
213
214 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi6x01234567), vget_low_s16(vk6x01234567));
215 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi6x01234567), vget_high_s16(vk6x01234567));
216 const int16x8_t vi7x01234567 = vmovl_s8(vld1_s8(i7));
217 const int16x8_t vk7x01234567 = vmovl_s8(vld1_s8((const void*) ((const int8_t*) w + 56)));
218
219 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi7x01234567), vget_low_s16(vk7x01234567));
220 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi7x01234567), vget_high_s16(vk7x01234567));
221 const int16x8_t vi8x01234567 = vmovl_s8(vld1_s8(i8));
222 const int16x8_t vk8x01234567 = vmovl_s8(vld1_s8((const void*) ((const int8_t*) w + 64)));
223
224 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi8x01234567), vget_low_s16(vk8x01234567));
225 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi8x01234567), vget_high_s16(vk8x01234567));
226
227 float32x4_t vfpacc0123 = vcvtq_f32_s32(vacc0123);
228 float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567);
229
230 vfpacc0123 = vmulq_f32(vfpacc0123, vscale);
231 vfpacc4567 = vmulq_f32(vfpacc4567, vscale);
232
233 vacc0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0123, vmagic_bias));
234 vacc4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc4567, vmagic_bias));
235
236 vacc0123 = vqsubq_s32(vacc0123, vmagic_bias_less_output_zero_point);
237 vacc4567 = vqsubq_s32(vacc4567, vmagic_bias_less_output_zero_point);
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
245 int8x8_t vout01234567 = vqmovn_s16(vacc01234567);
246 vout01234567 = vmax_s8(vout01234567, voutput_min);
247 vout01234567 = vmin_s8(vout01234567, voutput_max);
248
249 if (c & 4) {
250 vst1_lane_u32((void*) output, vreinterpret_u32_s8(vout01234567), 0); output += 4;
251 vout01234567 = vext_s8(vout01234567, vout01234567, 4);
252 }
253 if (c & 2) {
254 vst1_lane_u16((void*) output, vreinterpret_u16_s8(vout01234567), 0); output += 2;
255 vout01234567 = vext_s8(vout01234567, vout01234567, 2);
256 }
257 if (c & 1) {
258 vst1_lane_s8(output, vout01234567, 0); output += 1;
259 }
260 }
261 }
262
263 output = (int8_t*) ((uintptr_t) output + output_increment);
264 } while (--output_width != 0);
265 }
266