• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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 #include <xnnpack/intrinsics-polyfill.h>
16 
17 
xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__neonv8_c32(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)])18 void xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__neonv8_c32(
19     size_t rows,
20     size_t channels,
21     const uint8_t* input,
22     size_t input_stride,
23     const uint8_t* zero,
24     uint8_t* output,
25     const union xnn_qu8_avgpool_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
26 {
27   assert(rows != 0);
28   assert(rows <= 7);
29   assert(channels != 0);
30 
31   const uint8_t* i0 = input;
32   const uint8_t* i1 = (const uint8_t*) ((uintptr_t) i0 + input_stride);
33   if XNN_UNPREDICTABLE(rows < 2) {
34     i1 = zero;
35   }
36   const uint8_t* i2 = (const uint8_t*) ((uintptr_t) i1 + input_stride);
37   if XNN_UNPREDICTABLE(rows <= 2) {
38     i2 = zero;
39   }
40   const uint8_t* i3 = (const uint8_t*) ((uintptr_t) i2 + input_stride);
41   if XNN_UNPREDICTABLE(rows < 4) {
42     i3 = zero;
43   }
44   const uint8_t* i4 = (const uint8_t*) ((uintptr_t) i3 + input_stride);
45   if XNN_UNPREDICTABLE(rows <= 4) {
46     i4 = zero;
47   }
48   const uint8_t* i5 = (const uint8_t*) ((uintptr_t) i4 + input_stride);
49   if XNN_UNPREDICTABLE(rows < 6) {
50     i5 = zero;
51   }
52   const uint8_t* i6 = (const uint8_t*) ((uintptr_t) i5 + input_stride);
53   if XNN_UNPREDICTABLE(rows <= 6) {
54     i6 = zero;
55   }
56 
57   const int32x4_t vinit_bias = vld1q_dup_s32(&params->fp32_neonv8.init_bias);
58   const float32x4_t vscale = vld1q_dup_f32(&params->fp32_neonv8.scale);
59   const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->fp32_neonv8.output_zero_point);
60   const uint8x16_t voutput_min = vld1q_dup_u8(&params->fp32_neonv8.output_min);
61   const uint8x16_t voutput_max = vld1q_dup_u8(&params->fp32_neonv8.output_max);
62   for (; channels >= 32; channels -= 32) {
63     const uint8x8_t vi0x01234567 = vld1_u8(i0); i0 += 8;
64     const uint8x8_t vi0x89ABCDEF = vld1_u8(i0); i0 += 8;
65     const uint8x8_t vi0xGHIJKLMN = vld1_u8(i0); i0 += 8;
66     const uint8x8_t vi0xOPQRSTUV = 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     const uint8x8_t vi1xOPQRSTUV = vld1_u8(i1); i1 += 8;
71 
72     const uint8x8_t vi2x01234567 = vld1_u8(i2); i2 += 8;
73     uint16x8_t vsum01234567 = vaddl_u8(vi0x01234567, vi1x01234567);
74     const uint8x8_t vi2x89ABCDEF = vld1_u8(i2); i2 += 8;
75     uint16x8_t vsum89ABCDEF = vaddl_u8(vi0x89ABCDEF, vi1x89ABCDEF);
76     const uint8x8_t vi2xGHIJKLMN = vld1_u8(i2); i2 += 8;
77     uint16x8_t vsumGHIJKLMN = vaddl_u8(vi0xGHIJKLMN, vi1xGHIJKLMN);
78     const uint8x8_t vi2xOPQRSTUV = vld1_u8(i2); i2 += 8;
79     uint16x8_t vsumOPQRSTUV = vaddl_u8(vi0xOPQRSTUV, vi1xOPQRSTUV);
80 
81     const uint8x8_t vi3x01234567 = vld1_u8(i3); i3 += 8;
82     vsum01234567 = vaddw_u8(vsum01234567, vi2x01234567);
83     const uint8x8_t vi3x89ABCDEF = vld1_u8(i3); i3 += 8;
84     vsum89ABCDEF = vaddw_u8(vsum89ABCDEF, vi2x89ABCDEF);
85     const uint8x8_t vi3xGHIJKLMN = vld1_u8(i3); i3 += 8;
86     vsumGHIJKLMN = vaddw_u8(vsumGHIJKLMN, vi2xGHIJKLMN);
87     const uint8x8_t vi3xOPQRSTUV = vld1_u8(i3); i3 += 8;
88     vsumOPQRSTUV = vaddw_u8(vsumOPQRSTUV, vi2xOPQRSTUV);
89     const uint8x8_t vi4x01234567 = vld1_u8(i4); i4 += 8;
90     vsum01234567 = vaddw_u8(vsum01234567, vi3x01234567);
91     const uint8x8_t vi4x89ABCDEF = vld1_u8(i4); i4 += 8;
92     vsum89ABCDEF = vaddw_u8(vsum89ABCDEF, vi3x89ABCDEF);
93     const uint8x8_t vi4xGHIJKLMN = vld1_u8(i4); i4 += 8;
94     vsumGHIJKLMN = vaddw_u8(vsumGHIJKLMN, vi3xGHIJKLMN);
95     const uint8x8_t vi4xOPQRSTUV = vld1_u8(i4); i4 += 8;
96     vsumOPQRSTUV = vaddw_u8(vsumOPQRSTUV, vi3xOPQRSTUV);
97     const uint8x8_t vi5x01234567 = vld1_u8(i5); i5 += 8;
98     vsum01234567 = vaddw_u8(vsum01234567, vi4x01234567);
99     const uint8x8_t vi5x89ABCDEF = vld1_u8(i5); i5 += 8;
100     vsum89ABCDEF = vaddw_u8(vsum89ABCDEF, vi4x89ABCDEF);
101     const uint8x8_t vi5xGHIJKLMN = vld1_u8(i5); i5 += 8;
102     vsumGHIJKLMN = vaddw_u8(vsumGHIJKLMN, vi4xGHIJKLMN);
103     const uint8x8_t vi5xOPQRSTUV = vld1_u8(i5); i5 += 8;
104     vsumOPQRSTUV = vaddw_u8(vsumOPQRSTUV, vi4xOPQRSTUV);
105     const uint8x8_t vi6x01234567 = vld1_u8(i6); i6 += 8;
106     vsum01234567 = vaddw_u8(vsum01234567, vi5x01234567);
107     const uint8x8_t vi6x89ABCDEF = vld1_u8(i6); i6 += 8;
108     vsum89ABCDEF = vaddw_u8(vsum89ABCDEF, vi5x89ABCDEF);
109     const uint8x8_t vi6xGHIJKLMN = vld1_u8(i6); i6 += 8;
110     vsumGHIJKLMN = vaddw_u8(vsumGHIJKLMN, vi5xGHIJKLMN);
111     const uint8x8_t vi6xOPQRSTUV = vld1_u8(i6); i6 += 8;
112     vsumOPQRSTUV = vaddw_u8(vsumOPQRSTUV, vi5xOPQRSTUV);
113     vsum01234567 = vaddw_u8(vsum01234567, vi6x01234567);
114     vsum89ABCDEF = vaddw_u8(vsum89ABCDEF, vi6x89ABCDEF);
115     vsumGHIJKLMN = vaddw_u8(vsumGHIJKLMN, vi6xGHIJKLMN);
116     vsumOPQRSTUV = vaddw_u8(vsumOPQRSTUV, vi6xOPQRSTUV);
117 
118     int32x4_t vacc0123 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_low_u16(vsum01234567)));
119     int32x4_t vacc4567 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_high_u16(vsum01234567)));
120     int32x4_t vacc89AB = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_low_u16(vsum89ABCDEF)));
121     int32x4_t vaccCDEF = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_high_u16(vsum89ABCDEF)));
122     int32x4_t vaccGHIJ = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_low_u16(vsumGHIJKLMN)));
123     int32x4_t vaccKLMN = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_high_u16(vsumGHIJKLMN)));
124     int32x4_t vaccOPQR = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_low_u16(vsumOPQRSTUV)));
125     int32x4_t vaccSTUV = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_high_u16(vsumOPQRSTUV)));
126 
127     float32x4_t vfpacc0123 = vcvtq_f32_s32(vacc0123);
128     float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567);
129     float32x4_t vfpacc89AB = vcvtq_f32_s32(vacc89AB);
130     float32x4_t vfpaccCDEF = vcvtq_f32_s32(vaccCDEF);
131     float32x4_t vfpaccGHIJ = vcvtq_f32_s32(vaccGHIJ);
132     float32x4_t vfpaccKLMN = vcvtq_f32_s32(vaccKLMN);
133     float32x4_t vfpaccOPQR = vcvtq_f32_s32(vaccOPQR);
134     float32x4_t vfpaccSTUV = vcvtq_f32_s32(vaccSTUV);
135 
136     vfpacc0123 = vmulq_f32(vfpacc0123, vscale);
137     vfpacc4567 = vmulq_f32(vfpacc4567, vscale);
138     vfpacc89AB = vmulq_f32(vfpacc89AB, vscale);
139     vfpaccCDEF = vmulq_f32(vfpaccCDEF, vscale);
140     vfpaccGHIJ = vmulq_f32(vfpaccGHIJ, vscale);
141     vfpaccKLMN = vmulq_f32(vfpaccKLMN, vscale);
142     vfpaccOPQR = vmulq_f32(vfpaccOPQR, vscale);
143     vfpaccSTUV = vmulq_f32(vfpaccSTUV, vscale);
144 
145     vacc0123 = vcvtnq_s32_f32(vfpacc0123);
146     vacc4567 = vcvtnq_s32_f32(vfpacc4567);
147     vacc89AB = vcvtnq_s32_f32(vfpacc89AB);
148     vaccCDEF = vcvtnq_s32_f32(vfpaccCDEF);
149     vaccGHIJ = vcvtnq_s32_f32(vfpaccGHIJ);
150     vaccKLMN = vcvtnq_s32_f32(vfpaccKLMN);
151     vaccOPQR = vcvtnq_s32_f32(vfpaccOPQR);
152     vaccSTUV = vcvtnq_s32_f32(vfpaccSTUV);
153 
154     #if XNN_ARCH_ARM64
155       int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
156       int16x8_t vacc89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc89AB), vaccCDEF);
157       int16x8_t vaccGHIJKLMN = vqmovn_high_s32(vqmovn_s32(vaccGHIJ), vaccKLMN);
158       int16x8_t vaccOPQRSTUV = vqmovn_high_s32(vqmovn_s32(vaccOPQR), vaccSTUV);
159     #else  // !XNN_ARCH_ARM64
160       int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
161       int16x8_t vacc89ABCDEF = vcombine_s16(vqmovn_s32(vacc89AB), vqmovn_s32(vaccCDEF));
162       int16x8_t vaccGHIJKLMN = vcombine_s16(vqmovn_s32(vaccGHIJ), vqmovn_s32(vaccKLMN));
163       int16x8_t vaccOPQRSTUV = vcombine_s16(vqmovn_s32(vaccOPQR), vqmovn_s32(vaccSTUV));
164     #endif  // !XNN_ARCH_ARM64
165 
166     vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
167     vacc89ABCDEF = vqaddq_s16(vacc89ABCDEF, voutput_zero_point);
168     vaccGHIJKLMN = vqaddq_s16(vaccGHIJKLMN, voutput_zero_point);
169     vaccOPQRSTUV = vqaddq_s16(vaccOPQRSTUV, voutput_zero_point);
170 
171     #if XNN_ARCH_ARM64
172       uint8x16_t vout0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc01234567), vacc89ABCDEF);
173       uint8x16_t voutGHIJKLMNOPQRSTUV = vqmovun_high_s16(vqmovun_s16(vaccGHIJKLMN), vaccOPQRSTUV);
174     #else  // !XNN_ARCH_ARM64
175       uint8x16_t vout0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc01234567), vqmovun_s16(vacc89ABCDEF));
176       uint8x16_t voutGHIJKLMNOPQRSTUV = vcombine_u8(vqmovun_s16(vaccGHIJKLMN), vqmovun_s16(vaccOPQRSTUV));
177     #endif  // !XNN_ARCH_ARM64
178 
179     vout0123456789ABCDEF = vmaxq_u8(vout0123456789ABCDEF, voutput_min);
180     voutGHIJKLMNOPQRSTUV = vmaxq_u8(voutGHIJKLMNOPQRSTUV, voutput_min);
181 
182     vout0123456789ABCDEF = vminq_u8(vout0123456789ABCDEF, voutput_max);
183     voutGHIJKLMNOPQRSTUV = vminq_u8(voutGHIJKLMNOPQRSTUV, voutput_max);
184 
185     vst1q_u8(output, vout0123456789ABCDEF); output += 16;
186     vst1q_u8(output, voutGHIJKLMNOPQRSTUV); output += 16;
187   }
188   if XNN_UNLIKELY(channels != 0) {
189     do {
190       const uint8x8_t vi0x01234567 = vld1_u8(i0); i0 += 8;
191       const uint8x8_t vi1x01234567 = vld1_u8(i1); i1 += 8;
192       const uint8x8_t vi2x01234567 = vld1_u8(i2); i2 += 8;
193       uint16x8_t vsum01234567 = vaddl_u8(vi0x01234567, vi1x01234567);
194 
195       const uint8x8_t vi3x01234567 = vld1_u8(i3); i3 += 8;
196       vsum01234567 = vaddw_u8(vsum01234567, vi2x01234567);
197       const uint8x8_t vi4x01234567 = vld1_u8(i4); i4 += 8;
198       vsum01234567 = vaddw_u8(vsum01234567, vi3x01234567);
199       const uint8x8_t vi5x01234567 = vld1_u8(i5); i5 += 8;
200       vsum01234567 = vaddw_u8(vsum01234567, vi4x01234567);
201       const uint8x8_t vi6x01234567 = vld1_u8(i6); i6 += 8;
202       vsum01234567 = vaddw_u8(vsum01234567, vi5x01234567);
203       vsum01234567 = vaddw_u8(vsum01234567, vi6x01234567);
204 
205       int32x4_t vacc0123 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_low_u16(vsum01234567)));
206       int32x4_t vacc4567 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_high_u16(vsum01234567)));
207 
208       float32x4_t vfpacc0123 = vcvtq_f32_s32(vacc0123);
209       float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567);
210 
211       vfpacc0123 = vmulq_f32(vfpacc0123, vscale);
212       vfpacc4567 = vmulq_f32(vfpacc4567, vscale);
213 
214       vacc0123 = vcvtnq_s32_f32(vfpacc0123);
215       vacc4567 = vcvtnq_s32_f32(vfpacc4567);
216 
217       #if XNN_ARCH_ARM64
218         int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
219       #else
220         int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
221       #endif
222       vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
223 
224       uint8x8_t vout01234567 = vqmovun_s16(vacc01234567);
225       vout01234567 = vmax_u8(vout01234567, vget_low_u8(voutput_min));
226       vout01234567 = vmin_u8(vout01234567, vget_low_u8(voutput_max));
227 
228       if XNN_LIKELY(channels >= 8) {
229         vst1_u8(output, vout01234567); output += 8;
230         channels -= 8;
231       } else {
232         if (channels & 4) {
233           vst1_lane_u32((void*) output, vreinterpret_u32_u8(vout01234567), 0); output += 4;
234           vout01234567 = vext_u8(vout01234567, vout01234567, 4);
235         }
236         if (channels & 2) {
237           vst1_lane_u16((void*) output, vreinterpret_u16_u8(vout01234567), 0); output += 2;
238           vout01234567 = vext_u8(vout01234567, vout01234567, 2);
239         }
240         if (channels & 1) {
241           vst1_lane_u8(output, vout01234567, 0); output += 1;
242         }
243         channels = 0;
244       }
245     } while (channels != 0);
246   }
247 }
248