1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-dwconv2d-chw/5x5s2p2-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/dwconv.h>
15 #include <xnnpack/math.h>
16 
17 
xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc4(size_t input_height,size_t input_width,const float * input,const float * weights,const float * zero,float * output,uint32_t padding_top,const union xnn_f32_chw_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc4(
19     size_t input_height,
20     size_t input_width,
21     const float* input,
22     const float* weights,
23     const float* zero,
24     float* output,
25     uint32_t padding_top,
26     const union xnn_f32_chw_params params[restrict XNN_MIN_ELEMENTS(1)])
27 {
28   assert(input_height != 0);
29   assert(input_width != 0);
30   assert(input_width % sizeof(float) == 0);
31   assert(padding_top >= 1);
32   assert(padding_top <= 2);
33 
34   const uint32x4_t vmask_even = vld1q_u32(params->neon.mask_even);
35   const uint32x4_t vmask_odd = vld1q_u32(params->neon.mask_odd);
36   const float32x4_t vmax = vld1q_dup_f32(&params->neon.max);
37   const float32x4_t vmin = vld1q_dup_f32(&params->neon.min);
38 
39   const float32x4_t vw0123 = vld1q_f32(weights);
40   const float32x4_t vw4567 = vld1q_f32(weights + 4);
41   const float32x4_t vw89AB = vld1q_f32(weights + 8);
42   const float32x4_t vwCDEF = vld1q_f32(weights + 12);
43   const float32x4_t vwGHIJ = vld1q_f32(weights + 16);
44   const float32x4_t vwKLMN = vld1q_f32(weights + 20);
45   const float32x2_t vwOP   = vld1_f32(weights + 24);
46 
47   const uint32_t padding_top_less_1 = padding_top - 1;
48   const size_t input_decrement = round_up_po2(input_width, 8 * sizeof(float));
49 
50   const float* i0 = zero;
51   const float* i1 = (const float*) ((uintptr_t) input - ((-padding_top_less_1) & input_width));
52   const float* i2 = (const float*) ((uintptr_t) i1 + input_width);
53   if XNN_UNPREDICTABLE(padding_top_less_1 != 0) {
54     i1 = zero;
55   }
56   const float* i3 = (const float*) ((uintptr_t) i2 + input_width);
57   const float* i4 = (const float*) ((uintptr_t) i3 + input_width);
58 
59 
60   float* o0 = output;
61 
62   size_t padded_input_height = input_height + (padding_top_less_1 + 1) + 2 /* padding bottom */;
63   size_t output_height = (padded_input_height - 5 /* kernel size */ + 2 /* subsampling */) / 2;
64   do {
65     if XNN_UNPREDICTABLE(padded_input_height < 6) {
66       i3 = zero;
67     }
68     if XNN_UNPREDICTABLE(padded_input_height < 7) {
69       i4 = zero;
70     }
71 
72     float32x4_t vi0x0246 = vmovq_n_f32(0.0f);
73     float32x4_t vi1x0246 = vmovq_n_f32(0.0f);
74     float32x4_t vi2x0246 = vmovq_n_f32(0.0f);
75     float32x4_t vi3x0246 = vmovq_n_f32(0.0f);
76     float32x4_t vi4x0246 = vmovq_n_f32(0.0f);
77 
78     float32x4_t vi0x1357 = vmovq_n_f32(0.0f);
79     float32x4_t vi1x1357 = vmovq_n_f32(0.0f);
80     float32x4_t vi2x1357 = vmovq_n_f32(0.0f);
81     float32x4_t vi3x1357 = vmovq_n_f32(0.0f);
82     float32x4_t vi4x1357 = vmovq_n_f32(0.0f);
83 
84     float32x4x2_t vi0x8ACE9BDF = vld2q_f32(i0); i0 += 8;
85     float32x4x2_t vi1x8ACE9BDF = vld2q_f32(i1); i1 += 8;
86     float32x4x2_t vi2x8ACE9BDF = vld2q_f32(i2); i2 += 8;
87     float32x4x2_t vi3x8ACE9BDF = vld2q_f32(i3); i3 += 8;
88     float32x4x2_t vi4x8ACE9BDF = vld2q_f32(i4); i4 += 8;
89 
90     size_t w = input_width;
91     for (; w > 8 * sizeof(float); w -= 8 * sizeof(float)) {
92       float32x4_t vo0p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0);
93 
94       float32x4_t vo0p1 = vmulq_lane_f32(vi0x8ACE9BDF.val[0], vget_high_f32(vw0123), 1);
95 
96       float32x4_t vo0p2 = vmulq_lane_f32(vi1x8ACE9BDF.val[0], vget_low_f32(vw89AB), 0);
97 
98       float32x4_t vo0p3 = vmulq_lane_f32(vi2x8ACE9BDF.val[0], vget_low_f32(vwCDEF), 1);
99 
100       vo0p1 = vmlaq_lane_f32(vo0p1, vi3x8ACE9BDF.val[0], vget_high_f32(vwGHIJ), 0);
101 
102       vo0p2 = vmlaq_lane_f32(vo0p2, vi4x8ACE9BDF.val[0], vget_high_f32(vwKLMN), 1);
103 
104       vo0p3 = vmlaq_lane_f32(vo0p3, vi0x8ACE9BDF.val[1], vget_low_f32(vw4567), 0);
105 
106       vo0p0 = vmlaq_lane_f32(vo0p0, vi1x8ACE9BDF.val[1], vget_low_f32(vw89AB), 1);
107 
108       vo0p1 = vmlaq_lane_f32(vo0p1, vi2x8ACE9BDF.val[1], vget_high_f32(vwCDEF), 0);
109 
110       vo0p2 = vmlaq_lane_f32(vo0p2, vi3x8ACE9BDF.val[1], vget_high_f32(vwGHIJ), 1);
111 
112       vo0p3 = vmlaq_lane_f32(vo0p3, vi4x8ACE9BDF.val[1], vwOP, 0);
113 
114       const float32x4_t vi0x68AC = vextq_f32(vi0x0246, vi0x8ACE9BDF.val[0], 3);
115       vi0x0246 = vi0x8ACE9BDF.val[0];
116       const float32x4_t vi1x68AC = vextq_f32(vi1x0246, vi1x8ACE9BDF.val[0], 3);
117       vi1x0246 = vi1x8ACE9BDF.val[0];
118       const float32x4_t vi2x68AC = vextq_f32(vi2x0246, vi2x8ACE9BDF.val[0], 3);
119       vi2x0246 = vi2x8ACE9BDF.val[0];
120       const float32x4_t vi3x68AC = vextq_f32(vi3x0246, vi3x8ACE9BDF.val[0], 3);
121       vi3x0246 = vi3x8ACE9BDF.val[0];
122       const float32x4_t vi4x68AC = vextq_f32(vi4x0246, vi4x8ACE9BDF.val[0], 3);
123       vi4x0246 = vi4x8ACE9BDF.val[0];
124 
125       vo0p0 = vmlaq_lane_f32(vo0p0, vi0x68AC, vget_low_f32(vw0123), 1);
126 
127       vo0p1 = vmlaq_lane_f32(vo0p1, vi1x68AC, vget_high_f32(vw4567), 0);
128 
129       vo0p2 = vmlaq_lane_f32(vo0p2, vi2x68AC, vget_high_f32(vw89AB), 1);
130 
131       vo0p3 = vmlaq_lane_f32(vo0p3, vi3x68AC, vget_low_f32(vwGHIJ), 0);
132 
133       vo0p0 = vmlaq_lane_f32(vo0p0, vi4x68AC, vget_low_f32(vwKLMN), 1);
134 
135       const float32x4_t vi0x79BD = vextq_f32(vi0x1357, vi0x8ACE9BDF.val[1], 3);
136       vi0x1357 = vi0x8ACE9BDF.val[1];
137       const float32x4_t vi1x79BD = vextq_f32(vi1x1357, vi1x8ACE9BDF.val[1], 3);
138       vi1x1357 = vi1x8ACE9BDF.val[1];
139       const float32x4_t vi2x79BD = vextq_f32(vi2x1357, vi2x8ACE9BDF.val[1], 3);
140       vi2x1357 = vi2x8ACE9BDF.val[1];
141       const float32x4_t vi3x79BD = vextq_f32(vi3x1357, vi3x8ACE9BDF.val[1], 3);
142       vi3x1357 = vi3x8ACE9BDF.val[1];
143       const float32x4_t vi4x79BD = vextq_f32(vi4x1357, vi4x8ACE9BDF.val[1], 3);
144       vi4x1357 = vi4x8ACE9BDF.val[1];
145 
146       const float32x4x2_t vi0xGIKMHJLN = vld2q_f32(i0); i0 += 8;
147       const float32x4x2_t vi1xGIKMHJLN = vld2q_f32(i1); i1 += 8;
148       const float32x4x2_t vi2xGIKMHJLN = vld2q_f32(i2); i2 += 8;
149       const float32x4x2_t vi3xGIKMHJLN = vld2q_f32(i3); i3 += 8;
150       const float32x4x2_t vi4xGIKMHJLN = vld2q_f32(i4); i4 += 8;
151 
152       vo0p1 = vmlaq_lane_f32(vo0p1, vi0x79BD, vget_high_f32(vw0123), 0);
153 
154       vo0p2 = vmlaq_lane_f32(vo0p2, vi1x79BD, vget_high_f32(vw4567), 1);
155 
156       vo0p3 = vmlaq_lane_f32(vo0p3, vi2x79BD, vget_low_f32(vwCDEF), 0);
157 
158       vo0p0 = vmlaq_lane_f32(vo0p0, vi3x79BD, vget_low_f32(vwGHIJ), 1);
159 
160       vo0p1 = vmlaq_lane_f32(vo0p1, vi4x79BD, vget_high_f32(vwKLMN), 0);
161 
162       const float32x4_t vi0xACEG = vextq_f32(vi0x8ACE9BDF.val[0], vi0xGIKMHJLN.val[0], 1);
163       vi0x8ACE9BDF = vi0xGIKMHJLN;
164       const float32x4_t vi1xACEG = vextq_f32(vi1x8ACE9BDF.val[0], vi1xGIKMHJLN.val[0], 1);
165       vi1x8ACE9BDF = vi1xGIKMHJLN;
166       const float32x4_t vi2xACEG = vextq_f32(vi2x8ACE9BDF.val[0], vi2xGIKMHJLN.val[0], 1);
167       vi2x8ACE9BDF = vi2xGIKMHJLN;
168       const float32x4_t vi3xACEG = vextq_f32(vi3x8ACE9BDF.val[0], vi3xGIKMHJLN.val[0], 1);
169       vi3x8ACE9BDF = vi3xGIKMHJLN;
170       const float32x4_t vi4xACEG = vextq_f32(vi4x8ACE9BDF.val[0], vi4xGIKMHJLN.val[0], 1);
171       vi4x8ACE9BDF = vi4xGIKMHJLN;
172 
173       vo0p2 = vmlaq_lane_f32(vo0p2, vi0xACEG, vget_low_f32(vw4567), 1);
174 
175       vo0p3 = vmlaq_lane_f32(vo0p3, vi1xACEG, vget_high_f32(vw89AB), 0);
176 
177       vo0p0 = vmlaq_lane_f32(vo0p0, vi2xACEG, vget_high_f32(vwCDEF), 1);
178 
179       vo0p1 = vmlaq_lane_f32(vo0p1, vi3xACEG, vget_low_f32(vwKLMN), 0);
180 
181       vo0p2 = vmlaq_lane_f32(vo0p2, vi4xACEG, vwOP, 1);
182 
183       vo0p0 = vaddq_f32(vo0p0, vo0p1);
184       vo0p2 = vaddq_f32(vo0p2, vo0p3);
185       vo0p0 = vaddq_f32(vo0p0, vo0p2);
186 
187       float32x4_t vo0 = vmaxq_f32(vo0p0, vmin);
188 
189       vo0 = vminq_f32(vo0, vmax);
190 
191       vst1q_f32(o0, vo0); o0 += 4;
192     }
193     // Last block has 1-8 pixels to process.
194     assert(w <= 8 * sizeof(float));
195     assert(w >= 1 * sizeof(float));
196     {
197       float32x4_t vo0p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0);
198 
199       const float32x4_t vi0x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi0x8ACE9BDF.val[0])));
200       const float32x4_t vi1x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi1x8ACE9BDF.val[0])));
201       const float32x4_t vi2x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi2x8ACE9BDF.val[0])));
202       const float32x4_t vi3x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi3x8ACE9BDF.val[0])));
203       const float32x4_t vi4x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi4x8ACE9BDF.val[0])));
204 
205       const float32x4_t vi0x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi0x8ACE9BDF.val[1])));
206       const float32x4_t vi1x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi1x8ACE9BDF.val[1])));
207       const float32x4_t vi2x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi2x8ACE9BDF.val[1])));
208       const float32x4_t vi3x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi3x8ACE9BDF.val[1])));
209       const float32x4_t vi4x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi4x8ACE9BDF.val[1])));
210 
211       float32x4_t vo0p1 = vmulq_lane_f32(vi0x8ACE, vget_high_f32(vw0123), 1);
212 
213       float32x4_t vo0p2 = vmulq_lane_f32(vi1x8ACE, vget_low_f32(vw89AB), 0);
214 
215       float32x4_t vo0p3 = vmulq_lane_f32(vi2x8ACE, vget_low_f32(vwCDEF), 1);
216 
217       vo0p1 = vmlaq_lane_f32(vo0p1, vi3x8ACE, vget_high_f32(vwGHIJ), 0);
218 
219       vo0p2 = vmlaq_lane_f32(vo0p2, vi4x8ACE, vget_high_f32(vwKLMN), 1);
220 
221       vo0p3 = vmlaq_lane_f32(vo0p3, vi0x9BDF, vget_low_f32(vw4567), 0);
222 
223       vo0p0 = vmlaq_lane_f32(vo0p0, vi1x9BDF, vget_low_f32(vw89AB), 1);
224 
225       vo0p1 = vmlaq_lane_f32(vo0p1, vi2x9BDF, vget_high_f32(vwCDEF), 0);
226 
227       vo0p2 = vmlaq_lane_f32(vo0p2, vi3x9BDF, vget_high_f32(vwGHIJ), 1);
228 
229       vo0p3 = vmlaq_lane_f32(vo0p3, vi4x9BDF, vwOP, 0);
230 
231       const float32x4_t vi0x68AC = vextq_f32(vi0x0246, vi0x8ACE, 3);
232       const float32x4_t vi1x68AC = vextq_f32(vi1x0246, vi1x8ACE, 3);
233       const float32x4_t vi2x68AC = vextq_f32(vi2x0246, vi2x8ACE, 3);
234       const float32x4_t vi3x68AC = vextq_f32(vi3x0246, vi3x8ACE, 3);
235       const float32x4_t vi4x68AC = vextq_f32(vi4x0246, vi4x8ACE, 3);
236 
237       vo0p0 = vmlaq_lane_f32(vo0p0, vi0x68AC, vget_low_f32(vw0123), 1);
238 
239       vo0p1 = vmlaq_lane_f32(vo0p1, vi1x68AC, vget_high_f32(vw4567), 0);
240 
241       vo0p2 = vmlaq_lane_f32(vo0p2, vi2x68AC, vget_high_f32(vw89AB), 1);
242 
243       vo0p3 = vmlaq_lane_f32(vo0p3, vi3x68AC, vget_low_f32(vwGHIJ), 0);
244 
245       vo0p0 = vmlaq_lane_f32(vo0p0, vi4x68AC, vget_low_f32(vwKLMN), 1);
246 
247       const float32x4_t vi0x79BD = vextq_f32(vi0x1357, vi0x9BDF, 3);
248       const float32x4_t vi1x79BD = vextq_f32(vi1x1357, vi1x9BDF, 3);
249       const float32x4_t vi2x79BD = vextq_f32(vi2x1357, vi2x9BDF, 3);
250       const float32x4_t vi3x79BD = vextq_f32(vi3x1357, vi3x9BDF, 3);
251       const float32x4_t vi4x79BD = vextq_f32(vi4x1357, vi4x9BDF, 3);
252 
253       vo0p1 = vmlaq_lane_f32(vo0p1, vi0x79BD, vget_high_f32(vw0123), 0);
254 
255       vo0p2 = vmlaq_lane_f32(vo0p2, vi1x79BD, vget_high_f32(vw4567), 1);
256 
257       vo0p3 = vmlaq_lane_f32(vo0p3, vi2x79BD, vget_low_f32(vwCDEF), 0);
258 
259       vo0p0 = vmlaq_lane_f32(vo0p0, vi3x79BD, vget_low_f32(vwGHIJ), 1);
260 
261       vo0p1 = vmlaq_lane_f32(vo0p1, vi4x79BD, vget_high_f32(vwKLMN), 0);
262 
263       const float32x4_t vzero = vmovq_n_f32(0.0f);
264       const float32x4_t vi0xACEG = vextq_f32(vi0x8ACE, vzero, 1);
265       const float32x4_t vi1xACEG = vextq_f32(vi1x8ACE, vzero, 1);
266       const float32x4_t vi2xACEG = vextq_f32(vi2x8ACE, vzero, 1);
267       const float32x4_t vi3xACEG = vextq_f32(vi3x8ACE, vzero, 1);
268       const float32x4_t vi4xACEG = vextq_f32(vi4x8ACE, vzero, 1);
269 
270       vo0p2 = vmlaq_lane_f32(vo0p2, vi0xACEG, vget_low_f32(vw4567), 1);
271 
272       vo0p3 = vmlaq_lane_f32(vo0p3, vi1xACEG, vget_high_f32(vw89AB), 0);
273 
274       vo0p0 = vmlaq_lane_f32(vo0p0, vi2xACEG, vget_high_f32(vwCDEF), 1);
275 
276       vo0p1 = vmlaq_lane_f32(vo0p1, vi3xACEG, vget_low_f32(vwKLMN), 0);
277 
278       vo0p2 = vmlaq_lane_f32(vo0p2, vi4xACEG, vwOP, 1);
279 
280       vo0p0 = vaddq_f32(vo0p0, vo0p1);
281       vo0p2 = vaddq_f32(vo0p2, vo0p3);
282       vo0p0 = vaddq_f32(vo0p0, vo0p2);
283 
284       float32x4_t vo0 = vmaxq_f32(vo0p0, vmin);
285 
286       vo0 = vminq_f32(vo0, vmax);
287 
288       size_t w_tmp = (w + 1 * sizeof(float)) / (2 * sizeof(float));
289       if XNN_LIKELY(w_tmp >= 4) {
290         vst1q_f32(o0, vo0); o0 += 4;
291       } else {
292         float32x2_t vo0_lo = vget_low_f32(vo0);
293         if (w_tmp & 2) {
294           vst1_f32(o0, vo0_lo); o0 += 2;
295 
296           vo0_lo = vget_high_f32(vo0);
297         }
298         if (w_tmp & 1) {
299           vst1_lane_f32(o0, vo0_lo, 0); o0 += 1;
300         }
301       }
302     }
303 
304     i0 = (const float*) ((uintptr_t) i2 - input_decrement);
305     i1 = (const float*) ((uintptr_t) i3 - input_decrement);
306     i2 = (const float*) ((uintptr_t) i4 - input_decrement);
307     i3 = (const float*) ((uintptr_t) i2 + input_width);
308     i4 = (const float*) ((uintptr_t) i3 + input_width);
309 
310 
311     output_height -= 1;
312     padded_input_height -= 2;
313   } while (output_height != 0);
314 }
315