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_2x8_acc2(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_2x8_acc2(
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 const __fp16* i3 = (const __fp16*) ((uintptr_t) i2 + input_width);
47
48 __fp16* o0 = output;
49 __fp16* o1 = (__fp16*) ((uintptr_t) o0 + input_width);
50
51 size_t output_height = input_height;
52 do {
53 if XNN_UNPREDICTABLE(output_height < 2) {
54 i2 = zero;
55 o1 = o0;
56 }
57 if XNN_UNPREDICTABLE(output_height < 3) {
58 i3 = zero;
59 }
60
61 float16x8_t vi0x01234567 = vmovq_n_f16(0);
62 float16x8_t vi1x01234567 = vmovq_n_f16(0);
63 float16x8_t vi2x01234567 = vmovq_n_f16(0);
64 float16x8_t vi3x01234567 = vmovq_n_f16(0);
65
66 float16x8_t vi0x89ABCDEF = vld1q_f16(i0); i0 += 8;
67 float16x8_t vi1x89ABCDEF = vld1q_f16(i1); i1 += 8;
68 float16x8_t vi2x89ABCDEF = vld1q_f16(i2); i2 += 8;
69 float16x8_t vi3x89ABCDEF = vld1q_f16(i3); i3 += 8;
70
71 size_t w = input_width;
72 for (; w > 8 * sizeof(__fp16); w -= 8 * sizeof(__fp16)) {
73 float16x8_t vo0p0 = vdupq_lane_f16(vget_low_f16(vw01234567), 0);
74 float16x8_t vo1p0 = vdupq_lane_f16(vget_low_f16(vw01234567), 0);
75
76 const float16x8_t vi0xGHIJKLMN = vld1q_f16(i0); i0 += 8;
77 const float16x8_t vi1xGHIJKLMN = vld1q_f16(i1); i1 += 8;
78 const float16x8_t vi2xGHIJKLMN = vld1q_f16(i2); i2 += 8;
79 const float16x8_t vi3xGHIJKLMN = vld1q_f16(i3); i3 += 8;
80
81 // Center column
82 vo0p0 = vfmaq_lane_f16(vo0p0, vi0x89ABCDEF, vget_low_f16(vw01234567), 2);
83 vo1p0 = vfmaq_lane_f16(vo1p0, vi1x89ABCDEF, vget_low_f16(vw01234567), 2);
84
85 float16x8_t vo0p1 = vmulq_lane_f16(vi1x89ABCDEF, vget_high_f16(vw01234567), 1);
86 float16x8_t vo1p1 = vmulq_lane_f16(vi2x89ABCDEF, vget_high_f16(vw01234567), 1);
87
88 vo0p0 = vfmaq_lane_f16(vo0p0, vi2x89ABCDEF, vw89, 0);
89 vo1p0 = vfmaq_lane_f16(vo1p0, vi3x89ABCDEF, vw89, 0);
90
91 // Left column
92 const float16x8_t vi0x789ABCDE = vextq_f16(vi0x01234567, vi0x89ABCDEF, 7);
93 const float16x8_t vi1x789ABCDE = vextq_f16(vi1x01234567, vi1x89ABCDEF, 7);
94 const float16x8_t vi2x789ABCDE = vextq_f16(vi2x01234567, vi2x89ABCDEF, 7);
95 const float16x8_t vi3x789ABCDE = vextq_f16(vi3x01234567, vi3x89ABCDEF, 7);
96
97 vo0p1 = vfmaq_lane_f16(vo0p1, vi0x789ABCDE, vget_low_f16(vw01234567), 1);
98 vo1p1 = vfmaq_lane_f16(vo1p1, vi1x789ABCDE, vget_low_f16(vw01234567), 1);
99
100 vo0p0 = vfmaq_lane_f16(vo0p0, vi1x789ABCDE, vget_high_f16(vw01234567), 0);
101 vo1p0 = vfmaq_lane_f16(vo1p0, vi2x789ABCDE, vget_high_f16(vw01234567), 0);
102
103 vo0p1 = vfmaq_lane_f16(vo0p1, vi2x789ABCDE, vget_high_f16(vw01234567), 3);
104 vo1p1 = vfmaq_lane_f16(vo1p1, vi3x789ABCDE, vget_high_f16(vw01234567), 3);
105
106 vi0x01234567 = vi0x89ABCDEF;
107 vi1x01234567 = vi1x89ABCDEF;
108 vi2x01234567 = vi2x89ABCDEF;
109 vi3x01234567 = vi3x89ABCDEF;
110
111 // Right column
112 const float16x8_t vi0x9ABCDEFG = vextq_f16(vi0x89ABCDEF, vi0xGHIJKLMN, 1);
113 const float16x8_t vi1x9ABCDEFG = vextq_f16(vi1x89ABCDEF, vi1xGHIJKLMN, 1);
114 const float16x8_t vi2x9ABCDEFG = vextq_f16(vi2x89ABCDEF, vi2xGHIJKLMN, 1);
115 const float16x8_t vi3x9ABCDEFG = vextq_f16(vi3x89ABCDEF, vi3xGHIJKLMN, 1);
116
117 vo0p0 = vfmaq_lane_f16(vo0p0, vi0x9ABCDEFG, vget_low_f16(vw01234567), 3);
118 vo1p0 = vfmaq_lane_f16(vo1p0, vi1x9ABCDEFG, vget_low_f16(vw01234567), 3);
119
120 vo0p1 = vfmaq_lane_f16(vo0p1, vi1x9ABCDEFG, vget_high_f16(vw01234567), 2);
121 vo1p1 = vfmaq_lane_f16(vo1p1, vi2x9ABCDEFG, vget_high_f16(vw01234567), 2);
122
123 vo0p0 = vfmaq_lane_f16(vo0p0, vi2x9ABCDEFG, vw89, 1);
124 vo1p0 = vfmaq_lane_f16(vo1p0, vi3x9ABCDEFG, vw89, 1);
125
126 vi0x89ABCDEF = vi0xGHIJKLMN;
127 vi1x89ABCDEF = vi1xGHIJKLMN;
128 vi2x89ABCDEF = vi2xGHIJKLMN;
129 vi3x89ABCDEF = vi3xGHIJKLMN;
130
131 vo0p0 = vaddq_f16(vo0p0, vo0p1);
132 vo1p0 = vaddq_f16(vo1p0, vo1p1);
133
134 float16x8_t vo0 = vmaxq_f16(vo0p0, vmin);
135 float16x8_t vo1 = vmaxq_f16(vo1p0, vmin);
136
137 vo0 = vminq_f16(vo0, vmax);
138 vo1 = vminq_f16(vo1, vmax);
139
140 vst1q_f16(o1, vo1); o1 += 8;
141 vst1q_f16(o0, vo0); o0 += 8;
142 }
143 // Always process the last block of 1..8 pixels.
144 assert(w >= 1 * sizeof(__fp16));
145 assert(w <= 8 * sizeof(__fp16));
146 {
147 float16x8_t vo0p0 = vdupq_lane_f16(vget_low_f16(vw01234567), 0);
148 float16x8_t vo1p0 = vdupq_lane_f16(vget_low_f16(vw01234567), 0);
149
150 vi0x89ABCDEF = vreinterpretq_f16_u16(vandq_u16(vmask, vreinterpretq_u16_f16(vi0x89ABCDEF)));
151 vi1x89ABCDEF = vreinterpretq_f16_u16(vandq_u16(vmask, vreinterpretq_u16_f16(vi1x89ABCDEF)));
152 vi2x89ABCDEF = vreinterpretq_f16_u16(vandq_u16(vmask, vreinterpretq_u16_f16(vi2x89ABCDEF)));
153 vi3x89ABCDEF = vreinterpretq_f16_u16(vandq_u16(vmask, vreinterpretq_u16_f16(vi3x89ABCDEF)));
154
155 // Center column
156 vo0p0 = vfmaq_lane_f16(vo0p0, vi0x89ABCDEF, vget_low_f16(vw01234567), 2);
157 vo1p0 = vfmaq_lane_f16(vo1p0, vi1x89ABCDEF, vget_low_f16(vw01234567), 2);
158
159 float16x8_t vo0p1 = vmulq_lane_f16(vi1x89ABCDEF, vget_high_f16(vw01234567), 1);
160 float16x8_t vo1p1 = vmulq_lane_f16(vi2x89ABCDEF, vget_high_f16(vw01234567), 1);
161
162 vo0p0 = vfmaq_lane_f16(vo0p0, vi2x89ABCDEF, vw89, 0);
163 vo1p0 = vfmaq_lane_f16(vo1p0, vi3x89ABCDEF, vw89, 0);
164
165 // Left column
166 const float16x8_t vi0x789ABCDE = vextq_f16(vi0x01234567, vi0x89ABCDEF, 7);
167 const float16x8_t vi1x789ABCDE = vextq_f16(vi1x01234567, vi1x89ABCDEF, 7);
168 const float16x8_t vi2x789ABCDE = vextq_f16(vi2x01234567, vi2x89ABCDEF, 7);
169 const float16x8_t vi3x789ABCDE = vextq_f16(vi3x01234567, vi3x89ABCDEF, 7);
170
171 vo0p1 = vfmaq_lane_f16(vo0p1, vi0x789ABCDE, vget_low_f16(vw01234567), 1);
172 vo1p1 = vfmaq_lane_f16(vo1p1, vi1x789ABCDE, vget_low_f16(vw01234567), 1);
173
174 vo0p0 = vfmaq_lane_f16(vo0p0, vi1x789ABCDE, vget_high_f16(vw01234567), 0);
175 vo1p0 = vfmaq_lane_f16(vo1p0, vi2x789ABCDE, vget_high_f16(vw01234567), 0);
176
177 vo0p1 = vfmaq_lane_f16(vo0p1, vi2x789ABCDE, vget_high_f16(vw01234567), 3);
178 vo1p1 = vfmaq_lane_f16(vo1p1, vi3x789ABCDE, vget_high_f16(vw01234567), 3);
179
180 // Right column
181 const float16x8_t vzero = vmovq_n_f16(0);
182 const float16x8_t vi0x9ABCDEFG = vextq_f16(vi0x89ABCDEF, vzero, 1);
183 const float16x8_t vi1x9ABCDEFG = vextq_f16(vi1x89ABCDEF, vzero, 1);
184 const float16x8_t vi2x9ABCDEFG = vextq_f16(vi2x89ABCDEF, vzero, 1);
185 const float16x8_t vi3x9ABCDEFG = vextq_f16(vi3x89ABCDEF, vzero, 1);
186
187 vo0p0 = vfmaq_lane_f16(vo0p0, vi0x9ABCDEFG, vget_low_f16(vw01234567), 3);
188 vo1p0 = vfmaq_lane_f16(vo1p0, vi1x9ABCDEFG, vget_low_f16(vw01234567), 3);
189
190 vo0p1 = vfmaq_lane_f16(vo0p1, vi1x9ABCDEFG, vget_high_f16(vw01234567), 2);
191 vo1p1 = vfmaq_lane_f16(vo1p1, vi2x9ABCDEFG, vget_high_f16(vw01234567), 2);
192
193 vo0p0 = vfmaq_lane_f16(vo0p0, vi2x9ABCDEFG, vw89, 1);
194 vo1p0 = vfmaq_lane_f16(vo1p0, vi3x9ABCDEFG, vw89, 1);
195
196 vo0p0 = vaddq_f16(vo0p0, vo0p1);
197 vo1p0 = vaddq_f16(vo1p0, vo1p1);
198
199 float16x8_t vo0 = vmaxq_f16(vo0p0, vmin);
200 float16x8_t vo1 = vmaxq_f16(vo1p0, vmin);
201
202 vo0 = vminq_f16(vo0, vmax);
203 vo1 = vminq_f16(vo1, vmax);
204
205 if XNN_LIKELY(w == 8 * sizeof(__fp16)) {
206 vst1q_f16(o1, vo1); o1 += 8;
207 vst1q_f16(o0, vo0); o0 += 8;
208 } else {
209 float16x4_t vo1_lo = vget_low_f16(vo1);
210 float16x4_t vo0_lo = vget_low_f16(vo0);
211
212 if (w & (4 * sizeof(__fp16))) {
213 vst1_f16(o1, vo1_lo); o1 += 4;
214 vst1_f16(o0, vo0_lo); o0 += 4;
215
216 vo1_lo = vget_high_f16(vo1);
217 vo0_lo = vget_high_f16(vo0);
218 }
219 if (w & (2 * sizeof(__fp16))) {
220 vst1_lane_u32((void*) o1, vreinterpret_u32_f16(vo1_lo), 0); o1 += 2;
221 vst1_lane_u32((void*) o0, vreinterpret_u32_f16(vo0_lo), 0); o0 += 2;
222
223 vo0_lo = vext_f16(vo0_lo, vo0_lo, 2);
224 vo1_lo = vext_f16(vo1_lo, vo1_lo, 2);
225 }
226 if (w & (1 * sizeof(__fp16))) {
227 vst1_lane_f16(o1, vo1_lo, 0); o1 += 1;
228 vst1_lane_f16(o0, vo0_lo, 0); o0 += 1;
229 }
230 }
231 }
232
233 i0 = (const __fp16*) ((uintptr_t) i2 - input_decrement);
234 i1 = (const __fp16*) ((uintptr_t) i3 - input_decrement);
235 i2 = (const __fp16*) ((uintptr_t) i1 + input_width);
236 i3 = (const __fp16*) ((uintptr_t) i2 + input_width);
237
238 o0 = o1;
239 o1 = (__fp16*) ((uintptr_t) o0 + input_width);
240
241 output_height = doz(output_height, 2);
242 } while (output_height != 0);
243 }
244