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_c16(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_c16(
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 >= 16; channels -= 16) {
64 const uint8x8_t vi0x01234567 = vld1_u8(i0); i0 += 8;
65 const uint8x8_t vi0x89ABCDEF = vld1_u8(i0); i0 += 8;
66 const uint8x8_t vi1x01234567 = vld1_u8(i1); i1 += 8;
67 const uint8x8_t vi1x89ABCDEF = vld1_u8(i1); i1 += 8;
68
69 const uint8x8_t vi2x01234567 = vld1_u8(i2); i2 += 8;
70 uint16x8_t vsum01234567 = vaddl_u8(vi0x01234567, vi1x01234567);
71 const uint8x8_t vi2x89ABCDEF = vld1_u8(i2); i2 += 8;
72 uint16x8_t vsum89ABCDEF = vaddl_u8(vi0x89ABCDEF, vi1x89ABCDEF);
73
74 const uint8x8_t vi3x01234567 = vld1_u8(i3); i3 += 8;
75 vsum01234567 = vaddw_u8(vsum01234567, vi2x01234567);
76 const uint8x8_t vi3x89ABCDEF = vld1_u8(i3); i3 += 8;
77 vsum89ABCDEF = vaddw_u8(vsum89ABCDEF, vi2x89ABCDEF);
78 const uint8x8_t vi4x01234567 = vld1_u8(i4); i4 += 8;
79 vsum01234567 = vaddw_u8(vsum01234567, vi3x01234567);
80 const uint8x8_t vi4x89ABCDEF = vld1_u8(i4); i4 += 8;
81 vsum89ABCDEF = vaddw_u8(vsum89ABCDEF, vi3x89ABCDEF);
82 const uint8x8_t vi5x01234567 = vld1_u8(i5); i5 += 8;
83 vsum01234567 = vaddw_u8(vsum01234567, vi4x01234567);
84 const uint8x8_t vi5x89ABCDEF = vld1_u8(i5); i5 += 8;
85 vsum89ABCDEF = vaddw_u8(vsum89ABCDEF, vi4x89ABCDEF);
86 const uint8x8_t vi6x01234567 = vld1_u8(i6); i6 += 8;
87 vsum01234567 = vaddw_u8(vsum01234567, vi5x01234567);
88 const uint8x8_t vi6x89ABCDEF = vld1_u8(i6); i6 += 8;
89 vsum89ABCDEF = vaddw_u8(vsum89ABCDEF, vi5x89ABCDEF);
90 vsum01234567 = vaddw_u8(vsum01234567, vi6x01234567);
91 vsum89ABCDEF = vaddw_u8(vsum89ABCDEF, vi6x89ABCDEF);
92
93 int32x4_t vacc0123 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_low_u16(vsum01234567)));
94 int32x4_t vacc4567 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_high_u16(vsum01234567)));
95 int32x4_t vacc89AB = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_low_u16(vsum89ABCDEF)));
96 int32x4_t vaccCDEF = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_high_u16(vsum89ABCDEF)));
97
98 vacc0123 = vqshlq_s32(vacc0123, vleft_pre_shift);
99 vacc4567 = vqshlq_s32(vacc4567, vleft_pre_shift);
100 vacc89AB = vqshlq_s32(vacc89AB, vleft_pre_shift);
101 vaccCDEF = vqshlq_s32(vaccCDEF, vleft_pre_shift);
102
103 vacc0123 = vqdmulhq_s32(vacc0123, vmultiplier);
104 vacc4567 = vqdmulhq_s32(vacc4567, vmultiplier);
105 vacc89AB = vqdmulhq_s32(vacc89AB, vmultiplier);
106 vaccCDEF = vqdmulhq_s32(vaccCDEF, vmultiplier);
107
108 vacc0123 = vrshlq_s32(vacc0123, vleft_post_shift);
109 vacc4567 = vrshlq_s32(vacc4567, vleft_post_shift);
110 vacc89AB = vrshlq_s32(vacc89AB, vleft_post_shift);
111 vaccCDEF = vrshlq_s32(vaccCDEF, vleft_post_shift);
112
113 #if XNN_ARCH_ARM64
114 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
115 int16x8_t vacc89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc89AB), vaccCDEF);
116 #else // !XNN_ARCH_ARM64
117 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
118 int16x8_t vacc89ABCDEF = vcombine_s16(vqmovn_s32(vacc89AB), vqmovn_s32(vaccCDEF));
119 #endif // !XNN_ARCH_ARM64
120
121 vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
122 vacc89ABCDEF = vqaddq_s16(vacc89ABCDEF, voutput_zero_point);
123
124 #if XNN_ARCH_ARM64
125 uint8x16_t vout0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc01234567), vacc89ABCDEF);
126 #else // !XNN_ARCH_ARM64
127 uint8x16_t vout0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc01234567), vqmovun_s16(vacc89ABCDEF));
128 #endif // !XNN_ARCH_ARM64
129
130 vout0123456789ABCDEF = vmaxq_u8(vout0123456789ABCDEF, voutput_min);
131
132 vout0123456789ABCDEF = vminq_u8(vout0123456789ABCDEF, voutput_max);
133
134 vst1q_u8(output, vout0123456789ABCDEF); output += 16;
135 }
136 if XNN_UNLIKELY(channels != 0) {
137 do {
138 const uint8x8_t vi0x01234567 = vld1_u8(i0); i0 += 8;
139 const uint8x8_t vi1x01234567 = vld1_u8(i1); i1 += 8;
140 const uint8x8_t vi2x01234567 = vld1_u8(i2); i2 += 8;
141 uint16x8_t vsum01234567 = vaddl_u8(vi0x01234567, vi1x01234567);
142
143 const uint8x8_t vi3x01234567 = vld1_u8(i3); i3 += 8;
144 vsum01234567 = vaddw_u8(vsum01234567, vi2x01234567);
145 const uint8x8_t vi4x01234567 = vld1_u8(i4); i4 += 8;
146 vsum01234567 = vaddw_u8(vsum01234567, vi3x01234567);
147 const uint8x8_t vi5x01234567 = vld1_u8(i5); i5 += 8;
148 vsum01234567 = vaddw_u8(vsum01234567, vi4x01234567);
149 const uint8x8_t vi6x01234567 = vld1_u8(i6); i6 += 8;
150 vsum01234567 = vaddw_u8(vsum01234567, vi5x01234567);
151 vsum01234567 = vaddw_u8(vsum01234567, vi6x01234567);
152
153 int32x4_t vacc0123 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_low_u16(vsum01234567)));
154 int32x4_t vacc4567 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_high_u16(vsum01234567)));
155
156 vacc0123 = vqshlq_s32(vacc0123, vleft_pre_shift);
157 vacc4567 = vqshlq_s32(vacc4567, vleft_pre_shift);
158
159 vacc0123 = vqdmulhq_s32(vacc0123, vmultiplier);
160 vacc4567 = vqdmulhq_s32(vacc4567, vmultiplier);
161
162 vacc0123 = vrshlq_s32(vacc0123, vleft_post_shift);
163 vacc4567 = vrshlq_s32(vacc4567, vleft_post_shift);
164
165 #if XNN_ARCH_ARM64
166 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
167 #else
168 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
169 #endif
170 vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
171
172 uint8x8_t vout01234567 = vqmovun_s16(vacc01234567);
173 vout01234567 = vmax_u8(vout01234567, vget_low_u8(voutput_min));
174 vout01234567 = vmin_u8(vout01234567, vget_low_u8(voutput_max));
175
176 if XNN_LIKELY(channels >= 8) {
177 vst1_u8(output, vout01234567); output += 8;
178 channels -= 8;
179 } else {
180 if (channels & 4) {
181 vst1_lane_u32((void*) output, vreinterpret_u32_u8(vout01234567), 0); output += 4;
182 vout01234567 = vext_u8(vout01234567, vout01234567, 4);
183 }
184 if (channels & 2) {
185 vst1_lane_u16((void*) output, vreinterpret_u16_u8(vout01234567), 0); output += 2;
186 vout01234567 = vext_u8(vout01234567, vout01234567, 2);
187 }
188 if (channels & 1) {
189 vst1_lane_u8(output, vout01234567, 0); output += 1;
190 }
191 channels = 0;
192 }
193 } while (channels != 0);
194 }
195 }
196