xref: /aosp_15_r20/external/XNNPACK/src/qu8-gavgpool/gen/7x-minmax-rndnu-neon-c32.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36) 
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_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)])17 void xnn_qu8_gavgpool_minmax_rndnu_ukernel_7x__neon_c32(
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(&params->rndnu_neon.init_bias);
57   const int32x4_t vleft_pre_shift = vld1q_dup_s32(&params->rndnu_neon.left_pre_shift);
58   const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
59   const int32x4_t vleft_post_shift = vld1q_dup_s32(&params->rndnu_neon.left_post_shift);
60   const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
61   const uint8x16_t voutput_min = vld1q_dup_u8(&params->rndnu_neon.output_min);
62   const uint8x16_t voutput_max = vld1q_dup_u8(&params->rndnu_neon.output_max);
63   for (; channels >= 32; channels -= 32) {
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 vi0xOPQRSTUV = vld1_u8(i0); i0 += 8;
68     const uint8x8_t vi1x01234567 = vld1_u8(i1); i1 += 8;
69     const uint8x8_t vi1x89ABCDEF = vld1_u8(i1); i1 += 8;
70     const uint8x8_t vi1xGHIJKLMN = vld1_u8(i1); i1 += 8;
71     const uint8x8_t vi1xOPQRSTUV = vld1_u8(i1); i1 += 8;
72 
73     const uint8x8_t vi2x01234567 = vld1_u8(i2); i2 += 8;
74     uint16x8_t vsum01234567 = vaddl_u8(vi0x01234567, vi1x01234567);
75     const uint8x8_t vi2x89ABCDEF = vld1_u8(i2); i2 += 8;
76     uint16x8_t vsum89ABCDEF = vaddl_u8(vi0x89ABCDEF, vi1x89ABCDEF);
77     const uint8x8_t vi2xGHIJKLMN = vld1_u8(i2); i2 += 8;
78     uint16x8_t vsumGHIJKLMN = vaddl_u8(vi0xGHIJKLMN, vi1xGHIJKLMN);
79     const uint8x8_t vi2xOPQRSTUV = vld1_u8(i2); i2 += 8;
80     uint16x8_t vsumOPQRSTUV = vaddl_u8(vi0xOPQRSTUV, vi1xOPQRSTUV);
81 
82     const uint8x8_t vi3x01234567 = vld1_u8(i3); i3 += 8;
83     vsum01234567 = vaddw_u8(vsum01234567, vi2x01234567);
84     const uint8x8_t vi3x89ABCDEF = vld1_u8(i3); i3 += 8;
85     vsum89ABCDEF = vaddw_u8(vsum89ABCDEF, vi2x89ABCDEF);
86     const uint8x8_t vi3xGHIJKLMN = vld1_u8(i3); i3 += 8;
87     vsumGHIJKLMN = vaddw_u8(vsumGHIJKLMN, vi2xGHIJKLMN);
88     const uint8x8_t vi3xOPQRSTUV = vld1_u8(i3); i3 += 8;
89     vsumOPQRSTUV = vaddw_u8(vsumOPQRSTUV, vi2xOPQRSTUV);
90     const uint8x8_t vi4x01234567 = vld1_u8(i4); i4 += 8;
91     vsum01234567 = vaddw_u8(vsum01234567, vi3x01234567);
92     const uint8x8_t vi4x89ABCDEF = vld1_u8(i4); i4 += 8;
93     vsum89ABCDEF = vaddw_u8(vsum89ABCDEF, vi3x89ABCDEF);
94     const uint8x8_t vi4xGHIJKLMN = vld1_u8(i4); i4 += 8;
95     vsumGHIJKLMN = vaddw_u8(vsumGHIJKLMN, vi3xGHIJKLMN);
96     const uint8x8_t vi4xOPQRSTUV = vld1_u8(i4); i4 += 8;
97     vsumOPQRSTUV = vaddw_u8(vsumOPQRSTUV, vi3xOPQRSTUV);
98     const uint8x8_t vi5x01234567 = vld1_u8(i5); i5 += 8;
99     vsum01234567 = vaddw_u8(vsum01234567, vi4x01234567);
100     const uint8x8_t vi5x89ABCDEF = vld1_u8(i5); i5 += 8;
101     vsum89ABCDEF = vaddw_u8(vsum89ABCDEF, vi4x89ABCDEF);
102     const uint8x8_t vi5xGHIJKLMN = vld1_u8(i5); i5 += 8;
103     vsumGHIJKLMN = vaddw_u8(vsumGHIJKLMN, vi4xGHIJKLMN);
104     const uint8x8_t vi5xOPQRSTUV = vld1_u8(i5); i5 += 8;
105     vsumOPQRSTUV = vaddw_u8(vsumOPQRSTUV, vi4xOPQRSTUV);
106     const uint8x8_t vi6x01234567 = vld1_u8(i6); i6 += 8;
107     vsum01234567 = vaddw_u8(vsum01234567, vi5x01234567);
108     const uint8x8_t vi6x89ABCDEF = vld1_u8(i6); i6 += 8;
109     vsum89ABCDEF = vaddw_u8(vsum89ABCDEF, vi5x89ABCDEF);
110     const uint8x8_t vi6xGHIJKLMN = vld1_u8(i6); i6 += 8;
111     vsumGHIJKLMN = vaddw_u8(vsumGHIJKLMN, vi5xGHIJKLMN);
112     const uint8x8_t vi6xOPQRSTUV = vld1_u8(i6); i6 += 8;
113     vsumOPQRSTUV = vaddw_u8(vsumOPQRSTUV, vi5xOPQRSTUV);
114     vsum01234567 = vaddw_u8(vsum01234567, vi6x01234567);
115     vsum89ABCDEF = vaddw_u8(vsum89ABCDEF, vi6x89ABCDEF);
116     vsumGHIJKLMN = vaddw_u8(vsumGHIJKLMN, vi6xGHIJKLMN);
117     vsumOPQRSTUV = vaddw_u8(vsumOPQRSTUV, vi6xOPQRSTUV);
118 
119     int32x4_t vacc0123 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_low_u16(vsum01234567)));
120     int32x4_t vacc4567 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_high_u16(vsum01234567)));
121     int32x4_t vacc89AB = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_low_u16(vsum89ABCDEF)));
122     int32x4_t vaccCDEF = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_high_u16(vsum89ABCDEF)));
123     int32x4_t vaccGHIJ = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_low_u16(vsumGHIJKLMN)));
124     int32x4_t vaccKLMN = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_high_u16(vsumGHIJKLMN)));
125     int32x4_t vaccOPQR = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_low_u16(vsumOPQRSTUV)));
126     int32x4_t vaccSTUV = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_high_u16(vsumOPQRSTUV)));
127 
128     vacc0123 = vqshlq_s32(vacc0123, vleft_pre_shift);
129     vacc4567 = vqshlq_s32(vacc4567, vleft_pre_shift);
130     vacc89AB = vqshlq_s32(vacc89AB, vleft_pre_shift);
131     vaccCDEF = vqshlq_s32(vaccCDEF, vleft_pre_shift);
132     vaccGHIJ = vqshlq_s32(vaccGHIJ, vleft_pre_shift);
133     vaccKLMN = vqshlq_s32(vaccKLMN, vleft_pre_shift);
134     vaccOPQR = vqshlq_s32(vaccOPQR, vleft_pre_shift);
135     vaccSTUV = vqshlq_s32(vaccSTUV, vleft_pre_shift);
136 
137     vacc0123 = vqdmulhq_s32(vacc0123, vmultiplier);
138     vacc4567 = vqdmulhq_s32(vacc4567, vmultiplier);
139     vacc89AB = vqdmulhq_s32(vacc89AB, vmultiplier);
140     vaccCDEF = vqdmulhq_s32(vaccCDEF, vmultiplier);
141     vaccGHIJ = vqdmulhq_s32(vaccGHIJ, vmultiplier);
142     vaccKLMN = vqdmulhq_s32(vaccKLMN, vmultiplier);
143     vaccOPQR = vqdmulhq_s32(vaccOPQR, vmultiplier);
144     vaccSTUV = vqdmulhq_s32(vaccSTUV, vmultiplier);
145 
146     vacc0123 = vrshlq_s32(vacc0123, vleft_post_shift);
147     vacc4567 = vrshlq_s32(vacc4567, vleft_post_shift);
148     vacc89AB = vrshlq_s32(vacc89AB, vleft_post_shift);
149     vaccCDEF = vrshlq_s32(vaccCDEF, vleft_post_shift);
150     vaccGHIJ = vrshlq_s32(vaccGHIJ, vleft_post_shift);
151     vaccKLMN = vrshlq_s32(vaccKLMN, vleft_post_shift);
152     vaccOPQR = vrshlq_s32(vaccOPQR, vleft_post_shift);
153     vaccSTUV = vrshlq_s32(vaccSTUV, vleft_post_shift);
154 
155     #if XNN_ARCH_ARM64
156       int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
157       int16x8_t vacc89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc89AB), vaccCDEF);
158       int16x8_t vaccGHIJKLMN = vqmovn_high_s32(vqmovn_s32(vaccGHIJ), vaccKLMN);
159       int16x8_t vaccOPQRSTUV = vqmovn_high_s32(vqmovn_s32(vaccOPQR), vaccSTUV);
160     #else  // !XNN_ARCH_ARM64
161       int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
162       int16x8_t vacc89ABCDEF = vcombine_s16(vqmovn_s32(vacc89AB), vqmovn_s32(vaccCDEF));
163       int16x8_t vaccGHIJKLMN = vcombine_s16(vqmovn_s32(vaccGHIJ), vqmovn_s32(vaccKLMN));
164       int16x8_t vaccOPQRSTUV = vcombine_s16(vqmovn_s32(vaccOPQR), vqmovn_s32(vaccSTUV));
165     #endif  // !XNN_ARCH_ARM64
166 
167     vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
168     vacc89ABCDEF = vqaddq_s16(vacc89ABCDEF, voutput_zero_point);
169     vaccGHIJKLMN = vqaddq_s16(vaccGHIJKLMN, voutput_zero_point);
170     vaccOPQRSTUV = vqaddq_s16(vaccOPQRSTUV, voutput_zero_point);
171 
172     #if XNN_ARCH_ARM64
173       uint8x16_t vout0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc01234567), vacc89ABCDEF);
174       uint8x16_t voutGHIJKLMNOPQRSTUV = vqmovun_high_s16(vqmovun_s16(vaccGHIJKLMN), vaccOPQRSTUV);
175     #else  // !XNN_ARCH_ARM64
176       uint8x16_t vout0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc01234567), vqmovun_s16(vacc89ABCDEF));
177       uint8x16_t voutGHIJKLMNOPQRSTUV = vcombine_u8(vqmovun_s16(vaccGHIJKLMN), vqmovun_s16(vaccOPQRSTUV));
178     #endif  // !XNN_ARCH_ARM64
179 
180     vout0123456789ABCDEF = vmaxq_u8(vout0123456789ABCDEF, voutput_min);
181     voutGHIJKLMNOPQRSTUV = vmaxq_u8(voutGHIJKLMNOPQRSTUV, voutput_min);
182 
183     vout0123456789ABCDEF = vminq_u8(vout0123456789ABCDEF, voutput_max);
184     voutGHIJKLMNOPQRSTUV = vminq_u8(voutGHIJKLMNOPQRSTUV, voutput_max);
185 
186     vst1q_u8(output, vout0123456789ABCDEF); output += 16;
187     vst1q_u8(output, voutGHIJKLMNOPQRSTUV); output += 16;
188   }
189   if XNN_UNLIKELY(channels != 0) {
190     do {
191       const uint8x8_t vi0x01234567 = vld1_u8(i0); i0 += 8;
192       const uint8x8_t vi1x01234567 = vld1_u8(i1); i1 += 8;
193       const uint8x8_t vi2x01234567 = vld1_u8(i2); i2 += 8;
194       uint16x8_t vsum01234567 = vaddl_u8(vi0x01234567, vi1x01234567);
195 
196       const uint8x8_t vi3x01234567 = vld1_u8(i3); i3 += 8;
197       vsum01234567 = vaddw_u8(vsum01234567, vi2x01234567);
198       const uint8x8_t vi4x01234567 = vld1_u8(i4); i4 += 8;
199       vsum01234567 = vaddw_u8(vsum01234567, vi3x01234567);
200       const uint8x8_t vi5x01234567 = vld1_u8(i5); i5 += 8;
201       vsum01234567 = vaddw_u8(vsum01234567, vi4x01234567);
202       const uint8x8_t vi6x01234567 = vld1_u8(i6); i6 += 8;
203       vsum01234567 = vaddw_u8(vsum01234567, vi5x01234567);
204       vsum01234567 = vaddw_u8(vsum01234567, vi6x01234567);
205 
206       int32x4_t vacc0123 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_low_u16(vsum01234567)));
207       int32x4_t vacc4567 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_high_u16(vsum01234567)));
208 
209       vacc0123 = vqshlq_s32(vacc0123, vleft_pre_shift);
210       vacc4567 = vqshlq_s32(vacc4567, vleft_pre_shift);
211 
212       vacc0123 = vqdmulhq_s32(vacc0123, vmultiplier);
213       vacc4567 = vqdmulhq_s32(vacc4567, vmultiplier);
214 
215       vacc0123 = vrshlq_s32(vacc0123, vleft_post_shift);
216       vacc4567 = vrshlq_s32(vacc4567, vleft_post_shift);
217 
218       #if XNN_ARCH_ARM64
219         int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
220       #else
221         int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
222       #endif
223       vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
224 
225       uint8x8_t vout01234567 = vqmovun_s16(vacc01234567);
226       vout01234567 = vmax_u8(vout01234567, vget_low_u8(voutput_min));
227       vout01234567 = vmin_u8(vout01234567, vget_low_u8(voutput_max));
228 
229       if XNN_LIKELY(channels >= 8) {
230         vst1_u8(output, vout01234567); output += 8;
231         channels -= 8;
232       } else {
233         if (channels & 4) {
234           vst1_lane_u32((void*) output, vreinterpret_u32_u8(vout01234567), 0); output += 4;
235           vout01234567 = vext_u8(vout01234567, vout01234567, 4);
236         }
237         if (channels & 2) {
238           vst1_lane_u16((void*) output, vreinterpret_u16_u8(vout01234567), 0); output += 2;
239           vout01234567 = vext_u8(vout01234567, vout01234567, 2);
240         }
241         if (channels & 1) {
242           vst1_lane_u8(output, vout01234567, 0); output += 1;
243         }
244         channels = 0;
245       }
246     } while (channels != 0);
247   }
248 }
249