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_fp32_ukernel_7x__neon_c8(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_fp32_ukernel_7x__neon_c8(
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->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 uint8x8_t voutput_min = vld1_dup_u8(¶ms->fp32_neon.output_min);
61 const uint8x8_t voutput_max = vld1_dup_u8(¶ms->fp32_neon.output_max);
62 for (; channels >= 8; channels -= 8) {
63 const uint8x8_t vi0x01234567 = vld1_u8(i0); i0 += 8;
64 const uint8x8_t vi1x01234567 = vld1_u8(i1); i1 += 8;
65
66 const uint8x8_t vi2x01234567 = vld1_u8(i2); i2 += 8;
67 uint16x8_t vsum01234567 = vaddl_u8(vi0x01234567, vi1x01234567);
68
69 const uint8x8_t vi3x01234567 = vld1_u8(i3); i3 += 8;
70 vsum01234567 = vaddw_u8(vsum01234567, vi2x01234567);
71 const uint8x8_t vi4x01234567 = vld1_u8(i4); i4 += 8;
72 vsum01234567 = vaddw_u8(vsum01234567, vi3x01234567);
73 const uint8x8_t vi5x01234567 = vld1_u8(i5); i5 += 8;
74 vsum01234567 = vaddw_u8(vsum01234567, vi4x01234567);
75 const uint8x8_t vi6x01234567 = vld1_u8(i6); i6 += 8;
76 vsum01234567 = vaddw_u8(vsum01234567, vi5x01234567);
77 vsum01234567 = vaddw_u8(vsum01234567, vi6x01234567);
78
79 int32x4_t vacc0123 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_low_u16(vsum01234567)));
80 int32x4_t vacc4567 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_high_u16(vsum01234567)));
81
82 float32x4_t vfpacc0123 = vcvtq_f32_s32(vacc0123);
83 float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567);
84
85 vfpacc0123 = vmulq_f32(vfpacc0123, vscale);
86 vfpacc4567 = vmulq_f32(vfpacc4567, vscale);
87
88 vacc0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0123, vmagic_bias));
89 vacc4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc4567, vmagic_bias));
90
91 vacc0123 = vqsubq_s32(vacc0123, vmagic_bias_less_output_zero_point);
92 vacc4567 = vqsubq_s32(vacc4567, vmagic_bias_less_output_zero_point);
93
94 #if XNN_ARCH_ARM64
95 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
96 #else // !XNN_ARCH_ARM64
97 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
98 #endif // !XNN_ARCH_ARM64
99
100
101 #if XNN_ARCH_ARM64
102 uint8x8_t vout01234567 = vqmovun_s16(vacc01234567);
103 #else // !XNN_ARCH_ARM64
104 uint8x8_t vout01234567 = vqmovun_s16(vacc01234567);
105 #endif // !XNN_ARCH_ARM64
106
107 vout01234567 = vmax_u8(vout01234567, voutput_min);
108
109 vout01234567 = vmin_u8(vout01234567, voutput_max);
110
111 vst1_u8(output, vout01234567); output += 8;
112 }
113 if XNN_UNLIKELY(channels != 0) {
114 {
115 const uint8x8_t vi0x01234567 = vld1_u8(i0); i0 += 8;
116 const uint8x8_t vi1x01234567 = vld1_u8(i1); i1 += 8;
117 const uint8x8_t vi2x01234567 = vld1_u8(i2); i2 += 8;
118 uint16x8_t vsum01234567 = vaddl_u8(vi0x01234567, vi1x01234567);
119
120 const uint8x8_t vi3x01234567 = vld1_u8(i3); i3 += 8;
121 vsum01234567 = vaddw_u8(vsum01234567, vi2x01234567);
122 const uint8x8_t vi4x01234567 = vld1_u8(i4); i4 += 8;
123 vsum01234567 = vaddw_u8(vsum01234567, vi3x01234567);
124 const uint8x8_t vi5x01234567 = vld1_u8(i5); i5 += 8;
125 vsum01234567 = vaddw_u8(vsum01234567, vi4x01234567);
126 const uint8x8_t vi6x01234567 = vld1_u8(i6); i6 += 8;
127 vsum01234567 = vaddw_u8(vsum01234567, vi5x01234567);
128 vsum01234567 = vaddw_u8(vsum01234567, vi6x01234567);
129
130 int32x4_t vacc0123 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_low_u16(vsum01234567)));
131 int32x4_t vacc4567 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_high_u16(vsum01234567)));
132
133 float32x4_t vfpacc0123 = vcvtq_f32_s32(vacc0123);
134 float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567);
135
136 vfpacc0123 = vmulq_f32(vfpacc0123, vscale);
137 vfpacc4567 = vmulq_f32(vfpacc4567, vscale);
138
139 vacc0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0123, vmagic_bias));
140 vacc4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc4567, vmagic_bias));
141
142 vacc0123 = vqsubq_s32(vacc0123, vmagic_bias_less_output_zero_point);
143 vacc4567 = vqsubq_s32(vacc4567, vmagic_bias_less_output_zero_point);
144
145 #if XNN_ARCH_ARM64
146 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
147 #else
148 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
149 #endif
150
151 uint8x8_t vout01234567 = vqmovun_s16(vacc01234567);
152 vout01234567 = vmax_u8(vout01234567, voutput_min);
153 vout01234567 = vmin_u8(vout01234567, voutput_max);
154
155 if (channels & 4) {
156 vst1_lane_u32((void*) output, vreinterpret_u32_u8(vout01234567), 0); output += 4;
157 vout01234567 = vext_u8(vout01234567, vout01234567, 4);
158 }
159 if (channels & 2) {
160 vst1_lane_u16((void*) output, vreinterpret_u16_u8(vout01234567), 0); output += 2;
161 vout01234567 = vext_u8(vout01234567, vout01234567, 2);
162 }
163 if (channels & 1) {
164 vst1_lane_u8(output, vout01234567, 0);
165 }
166 }
167 }
168 }
169