1 // Auto-generated file. Do not edit!
2 // Template: src/f16-dwconv2d-chw/3x3p1-neonfp16arith.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/dwconv.h>
15 #include <xnnpack/math.h>
16
17
xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_1x8_acc3(size_t input_height,size_t input_width,const void * input,const void * weights,const void * zero,void * output,uint32_t padding_top,const union xnn_f16_chw_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_1x8_acc3(
19 size_t input_height,
20 size_t input_width,
21 const void* input,
22 const void* weights,
23 const void* zero,
24 void* output,
25 uint32_t padding_top,
26 const union xnn_f16_chw_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
27 {
28 assert(input_height != 0);
29 assert(input_width != 0);
30 assert(input_width % sizeof(__fp16) == 0);
31 assert(padding_top == 1);
32
33 const uint16x8_t vmask = vld1q_u16(params->neonfp16arith.maskx8);
34 const float16x8_t vmax = vld1q_dup_f16(¶ms->neonfp16arith.max);
35 const float16x8_t vmin = vld1q_dup_f16(¶ms->neonfp16arith.min);
36
37 const __fp16* w0 = (const __fp16*)weights;
38 const float16x8_t vw01234567 = vld1q_f16(w0);
39 const float16x4_t vw89 = vreinterpret_f16_u32(vld1_lane_u32((const void*)(w0 + 8), vmov_n_u32(0), 0));
40
41 const size_t input_decrement = round_up_po2(input_width, 8 * sizeof(__fp16));
42
43 const __fp16* i0 = zero;
44 const __fp16* i1 = input;
45 const __fp16* i2 = (const __fp16*) ((uintptr_t) i1 + input_width);
46
47 __fp16* o0 = output;
48
49 size_t output_height = input_height;
50 do {
51 if XNN_UNPREDICTABLE(output_height < 2) {
52 i2 = zero;
53 }
54
55 float16x8_t vi0x01234567 = vmovq_n_f16(0);
56 float16x8_t vi1x01234567 = vmovq_n_f16(0);
57 float16x8_t vi2x01234567 = vmovq_n_f16(0);
58
59 float16x8_t vi0x89ABCDEF = vld1q_f16(i0); i0 += 8;
60 float16x8_t vi1x89ABCDEF = vld1q_f16(i1); i1 += 8;
61 float16x8_t vi2x89ABCDEF = vld1q_f16(i2); i2 += 8;
62
63 size_t w = input_width;
64 for (; w > 8 * sizeof(__fp16); w -= 8 * sizeof(__fp16)) {
65 float16x8_t vo0p0 = vdupq_lane_f16(vget_low_f16(vw01234567), 0);
66
67 const float16x8_t vi0xGHIJKLMN = vld1q_f16(i0); i0 += 8;
68 const float16x8_t vi1xGHIJKLMN = vld1q_f16(i1); i1 += 8;
69 const float16x8_t vi2xGHIJKLMN = vld1q_f16(i2); i2 += 8;
70
71 // Center column
72 vo0p0 = vfmaq_lane_f16(vo0p0, vi0x89ABCDEF, vget_low_f16(vw01234567), 2);
73
74 float16x8_t vo0p1 = vmulq_lane_f16(vi1x89ABCDEF, vget_high_f16(vw01234567), 1);
75
76 float16x8_t vo0p2 = vmulq_lane_f16(vi2x89ABCDEF, vw89, 0);
77
78 // Left column
79 const float16x8_t vi0x789ABCDE = vextq_f16(vi0x01234567, vi0x89ABCDEF, 7);
80 const float16x8_t vi1x789ABCDE = vextq_f16(vi1x01234567, vi1x89ABCDEF, 7);
81 const float16x8_t vi2x789ABCDE = vextq_f16(vi2x01234567, vi2x89ABCDEF, 7);
82
83 vo0p0 = vfmaq_lane_f16(vo0p0, vi0x789ABCDE, vget_low_f16(vw01234567), 1);
84
85 vo0p1 = vfmaq_lane_f16(vo0p1, vi1x789ABCDE, vget_high_f16(vw01234567), 0);
86
87 vo0p2 = vfmaq_lane_f16(vo0p2, vi2x789ABCDE, vget_high_f16(vw01234567), 3);
88
89 vi0x01234567 = vi0x89ABCDEF;
90 vi1x01234567 = vi1x89ABCDEF;
91 vi2x01234567 = vi2x89ABCDEF;
92
93 // Right column
94 const float16x8_t vi0x9ABCDEFG = vextq_f16(vi0x89ABCDEF, vi0xGHIJKLMN, 1);
95 const float16x8_t vi1x9ABCDEFG = vextq_f16(vi1x89ABCDEF, vi1xGHIJKLMN, 1);
96 const float16x8_t vi2x9ABCDEFG = vextq_f16(vi2x89ABCDEF, vi2xGHIJKLMN, 1);
97
98 vo0p0 = vfmaq_lane_f16(vo0p0, vi0x9ABCDEFG, vget_low_f16(vw01234567), 3);
99
100 vo0p1 = vfmaq_lane_f16(vo0p1, vi1x9ABCDEFG, vget_high_f16(vw01234567), 2);
101
102 vo0p2 = vfmaq_lane_f16(vo0p2, vi2x9ABCDEFG, vw89, 1);
103
104 vi0x89ABCDEF = vi0xGHIJKLMN;
105 vi1x89ABCDEF = vi1xGHIJKLMN;
106 vi2x89ABCDEF = vi2xGHIJKLMN;
107
108 vo0p0 = vaddq_f16(vo0p0, vo0p1);
109 vo0p0 = vaddq_f16(vo0p0, vo0p2);
110
111 float16x8_t vo0 = vmaxq_f16(vo0p0, vmin);
112
113 vo0 = vminq_f16(vo0, vmax);
114
115 vst1q_f16(o0, vo0); o0 += 8;
116 }
117 // Always process the last block of 1..8 pixels.
118 assert(w >= 1 * sizeof(__fp16));
119 assert(w <= 8 * sizeof(__fp16));
120 {
121 float16x8_t vo0p0 = vdupq_lane_f16(vget_low_f16(vw01234567), 0);
122
123 vi0x89ABCDEF = vreinterpretq_f16_u16(vandq_u16(vmask, vreinterpretq_u16_f16(vi0x89ABCDEF)));
124 vi1x89ABCDEF = vreinterpretq_f16_u16(vandq_u16(vmask, vreinterpretq_u16_f16(vi1x89ABCDEF)));
125 vi2x89ABCDEF = vreinterpretq_f16_u16(vandq_u16(vmask, vreinterpretq_u16_f16(vi2x89ABCDEF)));
126
127 // Center column
128 vo0p0 = vfmaq_lane_f16(vo0p0, vi0x89ABCDEF, vget_low_f16(vw01234567), 2);
129
130 float16x8_t vo0p1 = vmulq_lane_f16(vi1x89ABCDEF, vget_high_f16(vw01234567), 1);
131
132 float16x8_t vo0p2 = vmulq_lane_f16(vi2x89ABCDEF, vw89, 0);
133
134 // Left column
135 const float16x8_t vi0x789ABCDE = vextq_f16(vi0x01234567, vi0x89ABCDEF, 7);
136 const float16x8_t vi1x789ABCDE = vextq_f16(vi1x01234567, vi1x89ABCDEF, 7);
137 const float16x8_t vi2x789ABCDE = vextq_f16(vi2x01234567, vi2x89ABCDEF, 7);
138
139 vo0p0 = vfmaq_lane_f16(vo0p0, vi0x789ABCDE, vget_low_f16(vw01234567), 1);
140
141 vo0p1 = vfmaq_lane_f16(vo0p1, vi1x789ABCDE, vget_high_f16(vw01234567), 0);
142
143 vo0p2 = vfmaq_lane_f16(vo0p2, vi2x789ABCDE, vget_high_f16(vw01234567), 3);
144
145 // Right column
146 const float16x8_t vzero = vmovq_n_f16(0);
147 const float16x8_t vi0x9ABCDEFG = vextq_f16(vi0x89ABCDEF, vzero, 1);
148 const float16x8_t vi1x9ABCDEFG = vextq_f16(vi1x89ABCDEF, vzero, 1);
149 const float16x8_t vi2x9ABCDEFG = vextq_f16(vi2x89ABCDEF, vzero, 1);
150
151 vo0p0 = vfmaq_lane_f16(vo0p0, vi0x9ABCDEFG, vget_low_f16(vw01234567), 3);
152
153 vo0p1 = vfmaq_lane_f16(vo0p1, vi1x9ABCDEFG, vget_high_f16(vw01234567), 2);
154
155 vo0p2 = vfmaq_lane_f16(vo0p2, vi2x9ABCDEFG, vw89, 1);
156
157 vo0p0 = vaddq_f16(vo0p0, vo0p1);
158 vo0p0 = vaddq_f16(vo0p0, vo0p2);
159
160 float16x8_t vo0 = vmaxq_f16(vo0p0, vmin);
161
162 vo0 = vminq_f16(vo0, vmax);
163
164 if XNN_LIKELY(w == 8 * sizeof(__fp16)) {
165 vst1q_f16(o0, vo0); o0 += 8;
166 } else {
167 float16x4_t vo0_lo = vget_low_f16(vo0);
168
169 if (w & (4 * sizeof(__fp16))) {
170 vst1_f16(o0, vo0_lo); o0 += 4;
171
172 vo0_lo = vget_high_f16(vo0);
173 }
174 if (w & (2 * sizeof(__fp16))) {
175 vst1_lane_u32((void*) o0, vreinterpret_u32_f16(vo0_lo), 0); o0 += 2;
176
177 vo0_lo = vext_f16(vo0_lo, vo0_lo, 2);
178 }
179 if (w & (1 * sizeof(__fp16))) {
180 vst1_lane_f16(o0, vo0_lo, 0); o0 += 1;
181 }
182 }
183 }
184
185 i0 = (const __fp16*) ((uintptr_t) i1 - input_decrement);
186 i1 = (const __fp16*) ((uintptr_t) i2 - input_decrement);
187 i2 = (const __fp16*) ((uintptr_t) i1 + input_width);
188
189
190 } while (--output_height != 0);
191 }
192