1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gavgpool/unipass-neon.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/gavgpool.h>
15
16
xnn_qu8_gavgpool_minmax_rndnu_ukernel_7x__neon_c24(size_t rows,size_t channels,const uint8_t * input,size_t input_stride,const uint8_t * zero,uint8_t * output,const union xnn_qu8_avgpool_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qu8_gavgpool_minmax_rndnu_ukernel_7x__neon_c24(
18 size_t rows,
19 size_t channels,
20 const uint8_t* input,
21 size_t input_stride,
22 const uint8_t* zero,
23 uint8_t* output,
24 const union xnn_qu8_avgpool_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
25 {
26 assert(rows != 0);
27 assert(rows <= 7);
28 assert(channels != 0);
29
30 const uint8_t* i0 = input;
31 const uint8_t* i1 = (const uint8_t*) ((uintptr_t) i0 + input_stride);
32 if XNN_UNPREDICTABLE(rows < 2) {
33 i1 = zero;
34 }
35 const uint8_t* i2 = (const uint8_t*) ((uintptr_t) i1 + input_stride);
36 if XNN_UNPREDICTABLE(rows <= 2) {
37 i2 = zero;
38 }
39 const uint8_t* i3 = (const uint8_t*) ((uintptr_t) i2 + input_stride);
40 if XNN_UNPREDICTABLE(rows < 4) {
41 i3 = zero;
42 }
43 const uint8_t* i4 = (const uint8_t*) ((uintptr_t) i3 + input_stride);
44 if XNN_UNPREDICTABLE(rows <= 4) {
45 i4 = zero;
46 }
47 const uint8_t* i5 = (const uint8_t*) ((uintptr_t) i4 + input_stride);
48 if XNN_UNPREDICTABLE(rows < 6) {
49 i5 = zero;
50 }
51 const uint8_t* i6 = (const uint8_t*) ((uintptr_t) i5 + input_stride);
52 if XNN_UNPREDICTABLE(rows <= 6) {
53 i6 = zero;
54 }
55
56 const int32x4_t vinit_bias = vld1q_dup_s32(¶ms->rndnu_neon.init_bias);
57 const int32x4_t vleft_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.left_pre_shift);
58 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
59 const int32x4_t vleft_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.left_post_shift);
60 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
61 const uint8x16_t voutput_min = vld1q_dup_u8(¶ms->rndnu_neon.output_min);
62 const uint8x16_t voutput_max = vld1q_dup_u8(¶ms->rndnu_neon.output_max);
63 for (; channels >= 24; channels -= 24) {
64 const uint8x8_t vi0x01234567 = vld1_u8(i0); i0 += 8;
65 const uint8x8_t vi0x89ABCDEF = vld1_u8(i0); i0 += 8;
66 const uint8x8_t vi0xGHIJKLMN = vld1_u8(i0); i0 += 8;
67 const uint8x8_t vi1x01234567 = vld1_u8(i1); i1 += 8;
68 const uint8x8_t vi1x89ABCDEF = vld1_u8(i1); i1 += 8;
69 const uint8x8_t vi1xGHIJKLMN = vld1_u8(i1); i1 += 8;
70
71 const uint8x8_t vi2x01234567 = vld1_u8(i2); i2 += 8;
72 uint16x8_t vsum01234567 = vaddl_u8(vi0x01234567, vi1x01234567);
73 const uint8x8_t vi2x89ABCDEF = vld1_u8(i2); i2 += 8;
74 uint16x8_t vsum89ABCDEF = vaddl_u8(vi0x89ABCDEF, vi1x89ABCDEF);
75 const uint8x8_t vi2xGHIJKLMN = vld1_u8(i2); i2 += 8;
76 uint16x8_t vsumGHIJKLMN = vaddl_u8(vi0xGHIJKLMN, vi1xGHIJKLMN);
77
78 const uint8x8_t vi3x01234567 = vld1_u8(i3); i3 += 8;
79 vsum01234567 = vaddw_u8(vsum01234567, vi2x01234567);
80 const uint8x8_t vi3x89ABCDEF = vld1_u8(i3); i3 += 8;
81 vsum89ABCDEF = vaddw_u8(vsum89ABCDEF, vi2x89ABCDEF);
82 const uint8x8_t vi3xGHIJKLMN = vld1_u8(i3); i3 += 8;
83 vsumGHIJKLMN = vaddw_u8(vsumGHIJKLMN, vi2xGHIJKLMN);
84 const uint8x8_t vi4x01234567 = vld1_u8(i4); i4 += 8;
85 vsum01234567 = vaddw_u8(vsum01234567, vi3x01234567);
86 const uint8x8_t vi4x89ABCDEF = vld1_u8(i4); i4 += 8;
87 vsum89ABCDEF = vaddw_u8(vsum89ABCDEF, vi3x89ABCDEF);
88 const uint8x8_t vi4xGHIJKLMN = vld1_u8(i4); i4 += 8;
89 vsumGHIJKLMN = vaddw_u8(vsumGHIJKLMN, vi3xGHIJKLMN);
90 const uint8x8_t vi5x01234567 = vld1_u8(i5); i5 += 8;
91 vsum01234567 = vaddw_u8(vsum01234567, vi4x01234567);
92 const uint8x8_t vi5x89ABCDEF = vld1_u8(i5); i5 += 8;
93 vsum89ABCDEF = vaddw_u8(vsum89ABCDEF, vi4x89ABCDEF);
94 const uint8x8_t vi5xGHIJKLMN = vld1_u8(i5); i5 += 8;
95 vsumGHIJKLMN = vaddw_u8(vsumGHIJKLMN, vi4xGHIJKLMN);
96 const uint8x8_t vi6x01234567 = vld1_u8(i6); i6 += 8;
97 vsum01234567 = vaddw_u8(vsum01234567, vi5x01234567);
98 const uint8x8_t vi6x89ABCDEF = vld1_u8(i6); i6 += 8;
99 vsum89ABCDEF = vaddw_u8(vsum89ABCDEF, vi5x89ABCDEF);
100 const uint8x8_t vi6xGHIJKLMN = vld1_u8(i6); i6 += 8;
101 vsumGHIJKLMN = vaddw_u8(vsumGHIJKLMN, vi5xGHIJKLMN);
102 vsum01234567 = vaddw_u8(vsum01234567, vi6x01234567);
103 vsum89ABCDEF = vaddw_u8(vsum89ABCDEF, vi6x89ABCDEF);
104 vsumGHIJKLMN = vaddw_u8(vsumGHIJKLMN, vi6xGHIJKLMN);
105
106 int32x4_t vacc0123 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_low_u16(vsum01234567)));
107 int32x4_t vacc4567 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_high_u16(vsum01234567)));
108 int32x4_t vacc89AB = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_low_u16(vsum89ABCDEF)));
109 int32x4_t vaccCDEF = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_high_u16(vsum89ABCDEF)));
110 int32x4_t vaccGHIJ = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_low_u16(vsumGHIJKLMN)));
111 int32x4_t vaccKLMN = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_high_u16(vsumGHIJKLMN)));
112
113 vacc0123 = vqshlq_s32(vacc0123, vleft_pre_shift);
114 vacc4567 = vqshlq_s32(vacc4567, vleft_pre_shift);
115 vacc89AB = vqshlq_s32(vacc89AB, vleft_pre_shift);
116 vaccCDEF = vqshlq_s32(vaccCDEF, vleft_pre_shift);
117 vaccGHIJ = vqshlq_s32(vaccGHIJ, vleft_pre_shift);
118 vaccKLMN = vqshlq_s32(vaccKLMN, vleft_pre_shift);
119
120 vacc0123 = vqdmulhq_s32(vacc0123, vmultiplier);
121 vacc4567 = vqdmulhq_s32(vacc4567, vmultiplier);
122 vacc89AB = vqdmulhq_s32(vacc89AB, vmultiplier);
123 vaccCDEF = vqdmulhq_s32(vaccCDEF, vmultiplier);
124 vaccGHIJ = vqdmulhq_s32(vaccGHIJ, vmultiplier);
125 vaccKLMN = vqdmulhq_s32(vaccKLMN, vmultiplier);
126
127 vacc0123 = vrshlq_s32(vacc0123, vleft_post_shift);
128 vacc4567 = vrshlq_s32(vacc4567, vleft_post_shift);
129 vacc89AB = vrshlq_s32(vacc89AB, vleft_post_shift);
130 vaccCDEF = vrshlq_s32(vaccCDEF, vleft_post_shift);
131 vaccGHIJ = vrshlq_s32(vaccGHIJ, vleft_post_shift);
132 vaccKLMN = vrshlq_s32(vaccKLMN, vleft_post_shift);
133
134 #if XNN_ARCH_ARM64
135 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
136 int16x8_t vacc89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc89AB), vaccCDEF);
137 int16x8_t vaccGHIJKLMN = vqmovn_high_s32(vqmovn_s32(vaccGHIJ), vaccKLMN);
138 #else // !XNN_ARCH_ARM64
139 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
140 int16x8_t vacc89ABCDEF = vcombine_s16(vqmovn_s32(vacc89AB), vqmovn_s32(vaccCDEF));
141 int16x8_t vaccGHIJKLMN = vcombine_s16(vqmovn_s32(vaccGHIJ), vqmovn_s32(vaccKLMN));
142 #endif // !XNN_ARCH_ARM64
143
144 vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
145 vacc89ABCDEF = vqaddq_s16(vacc89ABCDEF, voutput_zero_point);
146 vaccGHIJKLMN = vqaddq_s16(vaccGHIJKLMN, voutput_zero_point);
147
148 #if XNN_ARCH_ARM64
149 uint8x16_t vout0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc01234567), vacc89ABCDEF);
150 uint8x8_t voutGHIJKLMN = vqmovun_s16(vaccGHIJKLMN);
151 #else // !XNN_ARCH_ARM64
152 uint8x16_t vout0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc01234567), vqmovun_s16(vacc89ABCDEF));
153 uint8x8_t voutGHIJKLMN = vqmovun_s16(vaccGHIJKLMN);
154 #endif // !XNN_ARCH_ARM64
155
156 vout0123456789ABCDEF = vmaxq_u8(vout0123456789ABCDEF, voutput_min);
157 voutGHIJKLMN = vmax_u8(voutGHIJKLMN, vget_low_u8(voutput_min));
158
159 vout0123456789ABCDEF = vminq_u8(vout0123456789ABCDEF, voutput_max);
160 voutGHIJKLMN = vmin_u8(voutGHIJKLMN, vget_low_u8(voutput_max));
161
162 vst1q_u8(output, vout0123456789ABCDEF); output += 16;
163 vst1_u8(output, voutGHIJKLMN); output += 8;
164 }
165 if XNN_UNLIKELY(channels != 0) {
166 do {
167 const uint8x8_t vi0x01234567 = vld1_u8(i0); i0 += 8;
168 const uint8x8_t vi1x01234567 = vld1_u8(i1); i1 += 8;
169 const uint8x8_t vi2x01234567 = vld1_u8(i2); i2 += 8;
170 uint16x8_t vsum01234567 = vaddl_u8(vi0x01234567, vi1x01234567);
171
172 const uint8x8_t vi3x01234567 = vld1_u8(i3); i3 += 8;
173 vsum01234567 = vaddw_u8(vsum01234567, vi2x01234567);
174 const uint8x8_t vi4x01234567 = vld1_u8(i4); i4 += 8;
175 vsum01234567 = vaddw_u8(vsum01234567, vi3x01234567);
176 const uint8x8_t vi5x01234567 = vld1_u8(i5); i5 += 8;
177 vsum01234567 = vaddw_u8(vsum01234567, vi4x01234567);
178 const uint8x8_t vi6x01234567 = vld1_u8(i6); i6 += 8;
179 vsum01234567 = vaddw_u8(vsum01234567, vi5x01234567);
180 vsum01234567 = vaddw_u8(vsum01234567, vi6x01234567);
181
182 int32x4_t vacc0123 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_low_u16(vsum01234567)));
183 int32x4_t vacc4567 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_high_u16(vsum01234567)));
184
185 vacc0123 = vqshlq_s32(vacc0123, vleft_pre_shift);
186 vacc4567 = vqshlq_s32(vacc4567, vleft_pre_shift);
187
188 vacc0123 = vqdmulhq_s32(vacc0123, vmultiplier);
189 vacc4567 = vqdmulhq_s32(vacc4567, vmultiplier);
190
191 vacc0123 = vrshlq_s32(vacc0123, vleft_post_shift);
192 vacc4567 = vrshlq_s32(vacc4567, vleft_post_shift);
193
194 #if XNN_ARCH_ARM64
195 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
196 #else
197 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
198 #endif
199 vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
200
201 uint8x8_t vout01234567 = vqmovun_s16(vacc01234567);
202 vout01234567 = vmax_u8(vout01234567, vget_low_u8(voutput_min));
203 vout01234567 = vmin_u8(vout01234567, vget_low_u8(voutput_max));
204
205 if XNN_LIKELY(channels >= 8) {
206 vst1_u8(output, vout01234567); output += 8;
207 channels -= 8;
208 } else {
209 if (channels & 4) {
210 vst1_lane_u32((void*) output, vreinterpret_u32_u8(vout01234567), 0); output += 4;
211 vout01234567 = vext_u8(vout01234567, vout01234567, 4);
212 }
213 if (channels & 2) {
214 vst1_lane_u16((void*) output, vreinterpret_u16_u8(vout01234567), 0); output += 2;
215 vout01234567 = vext_u8(vout01234567, vout01234567, 2);
216 }
217 if (channels & 1) {
218 vst1_lane_u8(output, vout01234567, 0); output += 1;
219 }
220 channels = 0;
221 }
222 } while (channels != 0);
223 }
224 }
225