1 // Auto-generated file. Do not edit!
2 // Template: src/f16-gavgpool/unipass-neonfp16arith.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2022 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_f16_gavgpool_minmax_ukernel_7x__neonfp16arith_c32(size_t rows,size_t channels,const void * input,size_t input_stride,const void * zero,void * output,const union xnn_f16_scaleminmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_f16_gavgpool_minmax_ukernel_7x__neonfp16arith_c32(
18 size_t rows,
19 size_t channels,
20 const void* input,
21 size_t input_stride,
22 const void* zero,
23 void* output,
24 const union xnn_f16_scaleminmax_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 __fp16* i0 = input;
31 const __fp16* i1 = (const __fp16*) ((uintptr_t) i0 + input_stride);
32 if XNN_UNPREDICTABLE(rows < 2) {
33 i1 = (const __fp16*) zero;
34 }
35 const __fp16* i2 = (const __fp16*) ((uintptr_t) i1 + input_stride);
36 if XNN_UNPREDICTABLE(rows <= 2) {
37 i2 = (const __fp16*) zero;
38 }
39 const __fp16* i3 = (const __fp16*) ((uintptr_t) i2 + input_stride);
40 if XNN_UNPREDICTABLE(rows < 4) {
41 i3 = (const __fp16*) zero;
42 }
43 const __fp16* i4 = (const __fp16*) ((uintptr_t) i3 + input_stride);
44 if XNN_UNPREDICTABLE(rows <= 4) {
45 i4 = (const __fp16*) zero;
46 }
47 const __fp16* i5 = (const __fp16*) ((uintptr_t) i4 + input_stride);
48 if XNN_UNPREDICTABLE(rows < 6) {
49 i5 = (const __fp16*) zero;
50 }
51 const __fp16* i6 = (const __fp16*) ((uintptr_t) i5 + input_stride);
52 if XNN_UNPREDICTABLE(rows <= 6) {
53 i6 = (const __fp16*) zero;
54 }
55
56 const float16x8_t vscale = vreinterpretq_f16_u16(vld1q_dup_u16(¶ms->neon.scale));
57 const float16x8_t vmin = vreinterpretq_f16_u16(vld1q_dup_u16(¶ms->neon.min));
58 const float16x8_t vmax = vreinterpretq_f16_u16(vld1q_dup_u16(¶ms->neon.max));
59 for (; channels >= 32; channels -= 32) {
60 const float16x8_t vi0x01234567 = vld1q_f16(i0); i0 += 8;
61 const float16x8_t vi0x89ABCDEF = vld1q_f16(i0); i0 += 8;
62 const float16x8_t vi0xGHIJKLMN = vld1q_f16(i0); i0 += 8;
63 const float16x8_t vi0xOPQRSTUV = vld1q_f16(i0); i0 += 8;
64 const float16x8_t vi1x01234567 = vld1q_f16(i1); i1 += 8;
65 const float16x8_t vi1x89ABCDEF = vld1q_f16(i1); i1 += 8;
66 const float16x8_t vi1xGHIJKLMN = vld1q_f16(i1); i1 += 8;
67 const float16x8_t vi1xOPQRSTUV = vld1q_f16(i1); i1 += 8;
68
69 const float16x8_t vi2x01234567 = vld1q_f16(i2); i2 += 8;
70 float16x8_t vacc01234567 = vaddq_f16(vi0x01234567, vi1x01234567);
71 const float16x8_t vi2x89ABCDEF = vld1q_f16(i2); i2 += 8;
72 float16x8_t vacc89ABCDEF = vaddq_f16(vi0x89ABCDEF, vi1x89ABCDEF);
73 const float16x8_t vi2xGHIJKLMN = vld1q_f16(i2); i2 += 8;
74 float16x8_t vaccGHIJKLMN = vaddq_f16(vi0xGHIJKLMN, vi1xGHIJKLMN);
75 const float16x8_t vi2xOPQRSTUV = vld1q_f16(i2); i2 += 8;
76 float16x8_t vaccOPQRSTUV = vaddq_f16(vi0xOPQRSTUV, vi1xOPQRSTUV);
77
78 const float16x8_t vi3x01234567 = vld1q_f16(i3); i3 += 8;
79 vacc01234567 = vaddq_f16(vacc01234567, vi2x01234567);
80 const float16x8_t vi3x89ABCDEF = vld1q_f16(i3); i3 += 8;
81 vacc89ABCDEF = vaddq_f16(vacc89ABCDEF, vi2x89ABCDEF);
82 const float16x8_t vi3xGHIJKLMN = vld1q_f16(i3); i3 += 8;
83 vaccGHIJKLMN = vaddq_f16(vaccGHIJKLMN, vi2xGHIJKLMN);
84 const float16x8_t vi3xOPQRSTUV = vld1q_f16(i3); i3 += 8;
85 vaccOPQRSTUV = vaddq_f16(vaccOPQRSTUV, vi2xOPQRSTUV);
86 const float16x8_t vi4x01234567 = vld1q_f16(i4); i4 += 8;
87 vacc01234567 = vaddq_f16(vacc01234567, vi3x01234567);
88 const float16x8_t vi4x89ABCDEF = vld1q_f16(i4); i4 += 8;
89 vacc89ABCDEF = vaddq_f16(vacc89ABCDEF, vi3x89ABCDEF);
90 const float16x8_t vi4xGHIJKLMN = vld1q_f16(i4); i4 += 8;
91 vaccGHIJKLMN = vaddq_f16(vaccGHIJKLMN, vi3xGHIJKLMN);
92 const float16x8_t vi4xOPQRSTUV = vld1q_f16(i4); i4 += 8;
93 vaccOPQRSTUV = vaddq_f16(vaccOPQRSTUV, vi3xOPQRSTUV);
94 const float16x8_t vi5x01234567 = vld1q_f16(i5); i5 += 8;
95 vacc01234567 = vaddq_f16(vacc01234567, vi4x01234567);
96 const float16x8_t vi5x89ABCDEF = vld1q_f16(i5); i5 += 8;
97 vacc89ABCDEF = vaddq_f16(vacc89ABCDEF, vi4x89ABCDEF);
98 const float16x8_t vi5xGHIJKLMN = vld1q_f16(i5); i5 += 8;
99 vaccGHIJKLMN = vaddq_f16(vaccGHIJKLMN, vi4xGHIJKLMN);
100 const float16x8_t vi5xOPQRSTUV = vld1q_f16(i5); i5 += 8;
101 vaccOPQRSTUV = vaddq_f16(vaccOPQRSTUV, vi4xOPQRSTUV);
102 const float16x8_t vi6x01234567 = vld1q_f16(i6); i6 += 8;
103 vacc01234567 = vaddq_f16(vacc01234567, vi5x01234567);
104 const float16x8_t vi6x89ABCDEF = vld1q_f16(i6); i6 += 8;
105 vacc89ABCDEF = vaddq_f16(vacc89ABCDEF, vi5x89ABCDEF);
106 const float16x8_t vi6xGHIJKLMN = vld1q_f16(i6); i6 += 8;
107 vaccGHIJKLMN = vaddq_f16(vaccGHIJKLMN, vi5xGHIJKLMN);
108 const float16x8_t vi6xOPQRSTUV = vld1q_f16(i6); i6 += 8;
109 vaccOPQRSTUV = vaddq_f16(vaccOPQRSTUV, vi5xOPQRSTUV);
110 vacc01234567 = vaddq_f16(vacc01234567, vi6x01234567);
111 vacc89ABCDEF = vaddq_f16(vacc89ABCDEF, vi6x89ABCDEF);
112 vaccGHIJKLMN = vaddq_f16(vaccGHIJKLMN, vi6xGHIJKLMN);
113 vaccOPQRSTUV = vaddq_f16(vaccOPQRSTUV, vi6xOPQRSTUV);
114
115 vacc01234567 = vmulq_f16(vacc01234567, vscale);
116 vacc89ABCDEF = vmulq_f16(vacc89ABCDEF, vscale);
117 vaccGHIJKLMN = vmulq_f16(vaccGHIJKLMN, vscale);
118 vaccOPQRSTUV = vmulq_f16(vaccOPQRSTUV, vscale);
119
120 vacc01234567 = vmaxq_f16(vacc01234567, vmin);
121 vacc89ABCDEF = vmaxq_f16(vacc89ABCDEF, vmin);
122 vaccGHIJKLMN = vmaxq_f16(vaccGHIJKLMN, vmin);
123 vaccOPQRSTUV = vmaxq_f16(vaccOPQRSTUV, vmin);
124
125 vacc01234567 = vminq_f16(vacc01234567, vmax);
126 vacc89ABCDEF = vminq_f16(vacc89ABCDEF, vmax);
127 vaccGHIJKLMN = vminq_f16(vaccGHIJKLMN, vmax);
128 vaccOPQRSTUV = vminq_f16(vaccOPQRSTUV, vmax);
129
130 vst1q_f16(output, vacc01234567); output = (__fp16*) output + 8;
131 vst1q_f16(output, vacc89ABCDEF); output = (__fp16*) output + 8;
132 vst1q_f16(output, vaccGHIJKLMN); output = (__fp16*) output + 8;
133 vst1q_f16(output, vaccOPQRSTUV); output = (__fp16*) output + 8;
134 }
135 if XNN_UNLIKELY(channels != 0) {
136 do {
137 const float16x8_t vi0x01234567 = vld1q_f16(i0); i0 += 8;
138 const float16x8_t vi1x01234567 = vld1q_f16(i1); i1 += 8;
139
140 const float16x8_t vi2x01234567 = vld1q_f16(i2); i2 += 8;
141 float16x8_t vacc01234567 = vaddq_f16(vi0x01234567, vi1x01234567);
142
143 const float16x8_t vi3x01234567 = vld1q_f16(i3); i3 += 8;
144 vacc01234567 = vaddq_f16(vacc01234567, vi2x01234567);
145 const float16x8_t vi4x01234567 = vld1q_f16(i4); i4 += 8;
146 vacc01234567 = vaddq_f16(vacc01234567, vi3x01234567);
147 const float16x8_t vi5x01234567 = vld1q_f16(i5); i5 += 8;
148 vacc01234567 = vaddq_f16(vacc01234567, vi4x01234567);
149 const float16x8_t vi6x01234567 = vld1q_f16(i6); i6 += 8;
150 vacc01234567 = vaddq_f16(vacc01234567, vi5x01234567);
151 vacc01234567 = vaddq_f16(vacc01234567, vi6x01234567);
152
153 vacc01234567 = vmulq_f16(vacc01234567, vscale);
154 vacc01234567 = vmaxq_f16(vacc01234567, vmin);
155 vacc01234567 = vminq_f16(vacc01234567, vmax);
156
157 if XNN_LIKELY(channels >= 8) {
158 vst1q_f16(output, vacc01234567); output = (__fp16*) output + 8;
159 channels -= 8;
160 } else {
161 float16x4_t vacc0123 = vget_low_f16(vacc01234567);
162 if (channels & 4) {
163 vst1_f16(output, vacc0123); output = (__fp16*) output + 4;
164 vacc0123 = vget_high_f16(vacc01234567);
165 }
166 if (channels & 2) {
167 vst1_lane_u32(output, vreinterpret_u32_f16(vacc0123), 0); output = (__fp16*) output + 2;
168 vacc0123 = vext_f16(vacc0123, vacc0123, 2);
169 }
170 if (channels & 1) {
171 vst1_lane_f16(output, vacc0123, 0); output = (__fp16*) output + 1;
172 }
173 channels = 0;
174 }
175 } while (channels != 0);
176 }
177 }
178