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_qs8_gavgpool_minmax_fp32_ukernel_7x__neon_c24(size_t rows,size_t channels,const int8_t * input,size_t input_stride,const int8_t * zero,int8_t * output,const union xnn_qs8_avgpool_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__neon_c24(
18 size_t rows,
19 size_t channels,
20 const int8_t* input,
21 size_t input_stride,
22 const int8_t* zero,
23 int8_t* output,
24 const union xnn_qs8_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 int8_t* i0 = input;
31 const int8_t* i1 = (const int8_t*) ((uintptr_t) i0 + input_stride);
32 if XNN_UNPREDICTABLE(rows < 2) {
33 i1 = zero;
34 }
35 const int8_t* i2 = (const int8_t*) ((uintptr_t) i1 + input_stride);
36 if XNN_UNPREDICTABLE(rows <= 2) {
37 i2 = zero;
38 }
39 const int8_t* i3 = (const int8_t*) ((uintptr_t) i2 + input_stride);
40 if XNN_UNPREDICTABLE(rows < 4) {
41 i3 = zero;
42 }
43 const int8_t* i4 = (const int8_t*) ((uintptr_t) i3 + input_stride);
44 if XNN_UNPREDICTABLE(rows <= 4) {
45 i4 = zero;
46 }
47 const int8_t* i5 = (const int8_t*) ((uintptr_t) i4 + input_stride);
48 if XNN_UNPREDICTABLE(rows < 6) {
49 i5 = zero;
50 }
51 const int8_t* i6 = (const int8_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->fp32_neon.init_bias);
57 const float32x4_t vscale = vld1q_dup_f32(¶ms->fp32_neon.scale);
58 const float32x4_t vmagic_bias = vld1q_dup_f32(¶ms->fp32_neon.magic_bias);
59 const int32x4_t vmagic_bias_less_output_zero_point = vld1q_dup_s32(¶ms->fp32_neon.magic_bias_less_output_zero_point);
60 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->fp32_neon.output_min);
61 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->fp32_neon.output_max);
62 for (; channels >= 24; channels -= 24) {
63 const int8x8_t vi0x01234567 = vld1_s8(i0); i0 += 8;
64 const int8x8_t vi0x89ABCDEF = vld1_s8(i0); i0 += 8;
65 const int8x8_t vi0xGHIJKLMN = vld1_s8(i0); i0 += 8;
66 const int8x8_t vi1x01234567 = vld1_s8(i1); i1 += 8;
67 const int8x8_t vi1x89ABCDEF = vld1_s8(i1); i1 += 8;
68 const int8x8_t vi1xGHIJKLMN = vld1_s8(i1); i1 += 8;
69
70 const int8x8_t vi2x01234567 = vld1_s8(i2); i2 += 8;
71 int16x8_t vsum01234567 = vaddl_s8(vi0x01234567, vi1x01234567);
72 const int8x8_t vi2x89ABCDEF = vld1_s8(i2); i2 += 8;
73 int16x8_t vsum89ABCDEF = vaddl_s8(vi0x89ABCDEF, vi1x89ABCDEF);
74 const int8x8_t vi2xGHIJKLMN = vld1_s8(i2); i2 += 8;
75 int16x8_t vsumGHIJKLMN = vaddl_s8(vi0xGHIJKLMN, vi1xGHIJKLMN);
76
77 const int8x8_t vi3x01234567 = vld1_s8(i3); i3 += 8;
78 vsum01234567 = vaddw_s8(vsum01234567, vi2x01234567);
79 const int8x8_t vi3x89ABCDEF = vld1_s8(i3); i3 += 8;
80 vsum89ABCDEF = vaddw_s8(vsum89ABCDEF, vi2x89ABCDEF);
81 const int8x8_t vi3xGHIJKLMN = vld1_s8(i3); i3 += 8;
82 vsumGHIJKLMN = vaddw_s8(vsumGHIJKLMN, vi2xGHIJKLMN);
83 const int8x8_t vi4x01234567 = vld1_s8(i4); i4 += 8;
84 vsum01234567 = vaddw_s8(vsum01234567, vi3x01234567);
85 const int8x8_t vi4x89ABCDEF = vld1_s8(i4); i4 += 8;
86 vsum89ABCDEF = vaddw_s8(vsum89ABCDEF, vi3x89ABCDEF);
87 const int8x8_t vi4xGHIJKLMN = vld1_s8(i4); i4 += 8;
88 vsumGHIJKLMN = vaddw_s8(vsumGHIJKLMN, vi3xGHIJKLMN);
89 const int8x8_t vi5x01234567 = vld1_s8(i5); i5 += 8;
90 vsum01234567 = vaddw_s8(vsum01234567, vi4x01234567);
91 const int8x8_t vi5x89ABCDEF = vld1_s8(i5); i5 += 8;
92 vsum89ABCDEF = vaddw_s8(vsum89ABCDEF, vi4x89ABCDEF);
93 const int8x8_t vi5xGHIJKLMN = vld1_s8(i5); i5 += 8;
94 vsumGHIJKLMN = vaddw_s8(vsumGHIJKLMN, vi4xGHIJKLMN);
95 const int8x8_t vi6x01234567 = vld1_s8(i6); i6 += 8;
96 vsum01234567 = vaddw_s8(vsum01234567, vi5x01234567);
97 const int8x8_t vi6x89ABCDEF = vld1_s8(i6); i6 += 8;
98 vsum89ABCDEF = vaddw_s8(vsum89ABCDEF, vi5x89ABCDEF);
99 const int8x8_t vi6xGHIJKLMN = vld1_s8(i6); i6 += 8;
100 vsumGHIJKLMN = vaddw_s8(vsumGHIJKLMN, vi5xGHIJKLMN);
101 vsum01234567 = vaddw_s8(vsum01234567, vi6x01234567);
102 vsum89ABCDEF = vaddw_s8(vsum89ABCDEF, vi6x89ABCDEF);
103 vsumGHIJKLMN = vaddw_s8(vsumGHIJKLMN, vi6xGHIJKLMN);
104
105 int32x4_t vacc0123 = vaddw_s16(vinit_bias, vget_low_s16(vsum01234567));
106 int32x4_t vacc4567 = vaddw_s16(vinit_bias, vget_high_s16(vsum01234567));
107 int32x4_t vacc89AB = vaddw_s16(vinit_bias, vget_low_s16(vsum89ABCDEF));
108 int32x4_t vaccCDEF = vaddw_s16(vinit_bias, vget_high_s16(vsum89ABCDEF));
109 int32x4_t vaccGHIJ = vaddw_s16(vinit_bias, vget_low_s16(vsumGHIJKLMN));
110 int32x4_t vaccKLMN = vaddw_s16(vinit_bias, vget_high_s16(vsumGHIJKLMN));
111
112 float32x4_t vfpacc0123 = vcvtq_f32_s32(vacc0123);
113 float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567);
114 float32x4_t vfpacc89AB = vcvtq_f32_s32(vacc89AB);
115 float32x4_t vfpaccCDEF = vcvtq_f32_s32(vaccCDEF);
116 float32x4_t vfpaccGHIJ = vcvtq_f32_s32(vaccGHIJ);
117 float32x4_t vfpaccKLMN = vcvtq_f32_s32(vaccKLMN);
118
119 vfpacc0123 = vmulq_f32(vfpacc0123, vscale);
120 vfpacc4567 = vmulq_f32(vfpacc4567, vscale);
121 vfpacc89AB = vmulq_f32(vfpacc89AB, vscale);
122 vfpaccCDEF = vmulq_f32(vfpaccCDEF, vscale);
123 vfpaccGHIJ = vmulq_f32(vfpaccGHIJ, vscale);
124 vfpaccKLMN = vmulq_f32(vfpaccKLMN, vscale);
125
126 vacc0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0123, vmagic_bias));
127 vacc4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc4567, vmagic_bias));
128 vacc89AB = vreinterpretq_s32_f32(vaddq_f32(vfpacc89AB, vmagic_bias));
129 vaccCDEF = vreinterpretq_s32_f32(vaddq_f32(vfpaccCDEF, vmagic_bias));
130 vaccGHIJ = vreinterpretq_s32_f32(vaddq_f32(vfpaccGHIJ, vmagic_bias));
131 vaccKLMN = vreinterpretq_s32_f32(vaddq_f32(vfpaccKLMN, vmagic_bias));
132
133 vacc0123 = vqsubq_s32(vacc0123, vmagic_bias_less_output_zero_point);
134 vacc4567 = vqsubq_s32(vacc4567, vmagic_bias_less_output_zero_point);
135 vacc89AB = vqsubq_s32(vacc89AB, vmagic_bias_less_output_zero_point);
136 vaccCDEF = vqsubq_s32(vaccCDEF, vmagic_bias_less_output_zero_point);
137 vaccGHIJ = vqsubq_s32(vaccGHIJ, vmagic_bias_less_output_zero_point);
138 vaccKLMN = vqsubq_s32(vaccKLMN, vmagic_bias_less_output_zero_point);
139
140 #if XNN_ARCH_ARM64
141 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
142 int16x8_t vacc89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc89AB), vaccCDEF);
143 int16x8_t vaccGHIJKLMN = vqmovn_high_s32(vqmovn_s32(vaccGHIJ), vaccKLMN);
144 #else // !XNN_ARCH_ARM64
145 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
146 int16x8_t vacc89ABCDEF = vcombine_s16(vqmovn_s32(vacc89AB), vqmovn_s32(vaccCDEF));
147 int16x8_t vaccGHIJKLMN = vcombine_s16(vqmovn_s32(vaccGHIJ), vqmovn_s32(vaccKLMN));
148 #endif // !XNN_ARCH_ARM64
149
150
151 #if XNN_ARCH_ARM64
152 int8x16_t vout0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc01234567), vacc89ABCDEF);
153 int8x8_t voutGHIJKLMN = vqmovn_s16(vaccGHIJKLMN);
154 #else // !XNN_ARCH_ARM64
155 int8x16_t vout0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc01234567), vqmovn_s16(vacc89ABCDEF));
156 int8x8_t voutGHIJKLMN = vqmovn_s16(vaccGHIJKLMN);
157 #endif // !XNN_ARCH_ARM64
158
159 vout0123456789ABCDEF = vmaxq_s8(vout0123456789ABCDEF, voutput_min);
160 voutGHIJKLMN = vmax_s8(voutGHIJKLMN, vget_low_s8(voutput_min));
161
162 vout0123456789ABCDEF = vminq_s8(vout0123456789ABCDEF, voutput_max);
163 voutGHIJKLMN = vmin_s8(voutGHIJKLMN, vget_low_s8(voutput_max));
164
165 vst1q_s8(output, vout0123456789ABCDEF); output += 16;
166 vst1_s8(output, voutGHIJKLMN); output += 8;
167 }
168 if XNN_UNLIKELY(channels != 0) {
169 do {
170 const int8x8_t vi0x01234567 = vld1_s8(i0); i0 += 8;
171 const int8x8_t vi1x01234567 = vld1_s8(i1); i1 += 8;
172 const int8x8_t vi2x01234567 = vld1_s8(i2); i2 += 8;
173 int16x8_t vsum01234567 = vaddl_s8(vi0x01234567, vi1x01234567);
174
175 const int8x8_t vi3x01234567 = vld1_s8(i3); i3 += 8;
176 vsum01234567 = vaddw_s8(vsum01234567, vi2x01234567);
177 const int8x8_t vi4x01234567 = vld1_s8(i4); i4 += 8;
178 vsum01234567 = vaddw_s8(vsum01234567, vi3x01234567);
179 const int8x8_t vi5x01234567 = vld1_s8(i5); i5 += 8;
180 vsum01234567 = vaddw_s8(vsum01234567, vi4x01234567);
181 const int8x8_t vi6x01234567 = vld1_s8(i6); i6 += 8;
182 vsum01234567 = vaddw_s8(vsum01234567, vi5x01234567);
183 vsum01234567 = vaddw_s8(vsum01234567, vi6x01234567);
184
185 int32x4_t vacc0123 = vaddw_s16(vinit_bias, vget_low_s16(vsum01234567));
186 int32x4_t vacc4567 = vaddw_s16(vinit_bias, vget_high_s16(vsum01234567));
187
188 float32x4_t vfpacc0123 = vcvtq_f32_s32(vacc0123);
189 float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567);
190
191 vfpacc0123 = vmulq_f32(vfpacc0123, vscale);
192 vfpacc4567 = vmulq_f32(vfpacc4567, vscale);
193
194 vacc0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0123, vmagic_bias));
195 vacc4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc4567, vmagic_bias));
196
197 vacc0123 = vqsubq_s32(vacc0123, vmagic_bias_less_output_zero_point);
198 vacc4567 = vqsubq_s32(vacc4567, vmagic_bias_less_output_zero_point);
199
200 #if XNN_ARCH_ARM64
201 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
202 #else
203 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
204 #endif
205
206 int8x8_t vout01234567 = vqmovn_s16(vacc01234567);
207 vout01234567 = vmax_s8(vout01234567, vget_low_s8(voutput_min));
208 vout01234567 = vmin_s8(vout01234567, vget_low_s8(voutput_max));
209
210 if XNN_LIKELY(channels >= 8) {
211 vst1_s8(output, vout01234567); output += 8;
212 channels -= 8;
213 } else {
214 if (channels & 4) {
215 vst1_lane_u32((void*) output, vreinterpret_u32_s8(vout01234567), 0); output += 4;
216 vout01234567 = vext_s8(vout01234567, vout01234567, 4);
217 }
218 if (channels & 2) {
219 vst1_lane_u16((void*) output, vreinterpret_u16_s8(vout01234567), 0); output += 2;
220 vout01234567 = vext_s8(vout01234567, vout01234567, 2);
221 }
222 if (channels & 1) {
223 vst1_lane_s8(output, vout01234567, 0); output += 1;
224 }
225 channels = 0;
226 }
227 } while (channels != 0);
228 }
229 }
230