1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-dwconv2d-chw/3x3s2p1-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_3x3s2p1__neonfma_3x4(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_3x3s2p1__neonfma_3x4(
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 >= 0);
32   assert(padding_top <= 1);
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 float32x2_t vw89 = vld1_f32(weights + 8);
42 
43   const size_t input_decrement = round_down_po2(input_width, 4 /* SIMD output width */ * 2 /* subsampling */ * sizeof(float));
44   const size_t output_width = round_down_po2((input_width + (2 /* padding */ - 3 /* kernel size */ + 2 /* subsampling */) * sizeof(float)) / 2, sizeof(float));
45 
46   const float* i0 = (const float*) ((uintptr_t) input - ((-padding_top) & input_width));
47   const float* i1 = (const float*) ((uintptr_t) i0 + input_width);
48   if XNN_UNPREDICTABLE(padding_top != 0) {
49     i0 = zero;
50   }
51   const float* i2 = (const float*) ((uintptr_t) i1 + input_width);
52   const float* i3 = (const float*) ((uintptr_t) i2 + input_width);
53   const float* i4 = (const float*) ((uintptr_t) i3 + input_width);
54   const float* i5 = (const float*) ((uintptr_t) i4 + input_width);
55   const float* i6 = (const float*) ((uintptr_t) i5 + input_width);
56 
57   float* o0 = output;
58   float* o1 = (float*) ((uintptr_t) o0 + output_width);
59   float* o2 = (float*) ((uintptr_t) o1 + output_width);
60 
61   size_t padded_input_height = input_height + padding_top + 1 /* padding bottom */;
62   size_t output_height = (padded_input_height - 3 /* kernel size */ + 2 /* subsampling */) / 2;
63   do {
64     if XNN_UNPREDICTABLE(padded_input_height < 4) {
65       i2 = zero;
66     }
67     if XNN_UNPREDICTABLE(padded_input_height < 5) {
68       i3 = zero;
69       o1 = o0;
70     }
71     if XNN_UNPREDICTABLE(padded_input_height < 6) {
72       i4 = zero;
73     }
74     if XNN_UNPREDICTABLE(padded_input_height < 7) {
75       i5 = zero;
76       o2 = o1;
77     }
78     if XNN_UNPREDICTABLE(padded_input_height < 8) {
79       i6 = zero;
80     }
81 
82     float32x4_t vi0x1357 = vmovq_n_f32(0.0f);
83     float32x4_t vi1x1357 = vmovq_n_f32(0.0f);
84     float32x4_t vi2x1357 = vmovq_n_f32(0.0f);
85     float32x4_t vi3x1357 = vmovq_n_f32(0.0f);
86     float32x4_t vi4x1357 = vmovq_n_f32(0.0f);
87     float32x4_t vi5x1357 = vmovq_n_f32(0.0f);
88     float32x4_t vi6x1357 = vmovq_n_f32(0.0f);
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       float32x4_t vo1p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0);
94       float32x4_t vo2p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0);
95 
96       const float32x4x2_t vi0x8ACE9BDF = vld2q_f32(i0); i0 += 8;
97       const float32x4x2_t vi1x8ACE9BDF = vld2q_f32(i1); i1 += 8;
98       const float32x4x2_t vi2x8ACE9BDF = vld2q_f32(i2); i2 += 8;
99       const float32x4x2_t vi3x8ACE9BDF = vld2q_f32(i3); i3 += 8;
100       const float32x4x2_t vi4x8ACE9BDF = vld2q_f32(i4); i4 += 8;
101       const float32x4x2_t vi5x8ACE9BDF = vld2q_f32(i5); i5 += 8;
102       const float32x4x2_t vi6x8ACE9BDF = vld2q_f32(i6); i6 += 8;
103 
104       vo0p0 = vfmaq_lane_f32(vo0p0, vi0x8ACE9BDF.val[0], vget_high_f32(vw0123), 0);
105       vo1p0 = vfmaq_lane_f32(vo1p0, vi2x8ACE9BDF.val[0], vget_high_f32(vw0123), 0);
106       vo2p0 = vfmaq_lane_f32(vo2p0, vi4x8ACE9BDF.val[0], vget_high_f32(vw0123), 0);
107 
108       vo0p0 = vfmaq_lane_f32(vo0p0, vi1x8ACE9BDF.val[0], vget_low_f32(vw4567), 1);
109       vo1p0 = vfmaq_lane_f32(vo1p0, vi3x8ACE9BDF.val[0], vget_low_f32(vw4567), 1);
110       vo2p0 = vfmaq_lane_f32(vo2p0, vi5x8ACE9BDF.val[0], vget_low_f32(vw4567), 1);
111 
112       vo0p0 = vfmaq_lane_f32(vo0p0, vi2x8ACE9BDF.val[0], vw89, 0);
113       vo1p0 = vfmaq_lane_f32(vo1p0, vi4x8ACE9BDF.val[0], vw89, 0);
114       vo2p0 = vfmaq_lane_f32(vo2p0, vi6x8ACE9BDF.val[0], vw89, 0);
115 
116       const float32x4_t vi0x7BDF = vextq_f32(vi0x1357, vi0x8ACE9BDF.val[1], 3);
117       vi0x1357 = vi0x8ACE9BDF.val[1];
118       const float32x4_t vi1x7BDF = vextq_f32(vi1x1357, vi1x8ACE9BDF.val[1], 3);
119       vi1x1357 = vi1x8ACE9BDF.val[1];
120       const float32x4_t vi2x7BDF = vextq_f32(vi2x1357, vi2x8ACE9BDF.val[1], 3);
121       vi2x1357 = vi2x8ACE9BDF.val[1];
122       const float32x4_t vi3x7BDF = vextq_f32(vi3x1357, vi3x8ACE9BDF.val[1], 3);
123       vi3x1357 = vi3x8ACE9BDF.val[1];
124       const float32x4_t vi4x7BDF = vextq_f32(vi4x1357, vi4x8ACE9BDF.val[1], 3);
125       vi4x1357 = vi4x8ACE9BDF.val[1];
126       const float32x4_t vi5x7BDF = vextq_f32(vi5x1357, vi5x8ACE9BDF.val[1], 3);
127       vi5x1357 = vi5x8ACE9BDF.val[1];
128       const float32x4_t vi6x7BDF = vextq_f32(vi6x1357, vi6x8ACE9BDF.val[1], 3);
129       vi6x1357 = vi6x8ACE9BDF.val[1];
130 
131       vo0p0 = vfmaq_lane_f32(vo0p0, vi0x7BDF, vget_low_f32(vw0123), 1);
132       vo1p0 = vfmaq_lane_f32(vo1p0, vi2x7BDF, vget_low_f32(vw0123), 1);
133       vo2p0 = vfmaq_lane_f32(vo2p0, vi4x7BDF, vget_low_f32(vw0123), 1);
134 
135       vo0p0 = vfmaq_lane_f32(vo0p0, vi1x7BDF, vget_low_f32(vw4567), 0);
136       vo1p0 = vfmaq_lane_f32(vo1p0, vi3x7BDF, vget_low_f32(vw4567), 0);
137       vo2p0 = vfmaq_lane_f32(vo2p0, vi5x7BDF, vget_low_f32(vw4567), 0);
138 
139       vo0p0 = vfmaq_lane_f32(vo0p0, vi2x7BDF, vget_high_f32(vw4567), 1);
140       vo1p0 = vfmaq_lane_f32(vo1p0, vi4x7BDF, vget_high_f32(vw4567), 1);
141       vo2p0 = vfmaq_lane_f32(vo2p0, vi6x7BDF, vget_high_f32(vw4567), 1);
142 
143       vo0p0 = vfmaq_lane_f32(vo0p0, vi0x8ACE9BDF.val[1], vget_high_f32(vw0123), 1);
144       vo1p0 = vfmaq_lane_f32(vo1p0, vi2x8ACE9BDF.val[1], vget_high_f32(vw0123), 1);
145       vo2p0 = vfmaq_lane_f32(vo2p0, vi4x8ACE9BDF.val[1], vget_high_f32(vw0123), 1);
146 
147       vo0p0 = vfmaq_lane_f32(vo0p0, vi1x8ACE9BDF.val[1], vget_high_f32(vw4567), 0);
148       vo1p0 = vfmaq_lane_f32(vo1p0, vi3x8ACE9BDF.val[1], vget_high_f32(vw4567), 0);
149       vo2p0 = vfmaq_lane_f32(vo2p0, vi5x8ACE9BDF.val[1], vget_high_f32(vw4567), 0);
150 
151       vo0p0 = vfmaq_lane_f32(vo0p0, vi2x8ACE9BDF.val[1], vw89, 1);
152       vo1p0 = vfmaq_lane_f32(vo1p0, vi4x8ACE9BDF.val[1], vw89, 1);
153       vo2p0 = vfmaq_lane_f32(vo2p0, vi6x8ACE9BDF.val[1], vw89, 1);
154 
155 
156       float32x4_t vo0 = vmaxq_f32(vo0p0, vmin);
157       float32x4_t vo1 = vmaxq_f32(vo1p0, vmin);
158       float32x4_t vo2 = vmaxq_f32(vo2p0, vmin);
159 
160       vo0 = vminq_f32(vo0, vmax);
161       vo1 = vminq_f32(vo1, vmax);
162       vo2 = vminq_f32(vo2, vmax);
163 
164       vst1q_f32(o2, vo2); o2 += 4;
165       vst1q_f32(o1, vo1); o1 += 4;
166       vst1q_f32(o0, vo0); o0 += 4;
167     }
168     // Last block has 0-7 pixels to process.
169     assert(w < 8 * sizeof(float));
170     if XNN_LIKELY(w != 0) {
171       float32x4_t vo0p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0);
172       float32x4_t vo1p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0);
173       float32x4_t vo2p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0);
174 
175       const float32x4x2_t vi0x8ACE9BDF = vld2q_f32(i0);
176       const float32x4x2_t vi1x8ACE9BDF = vld2q_f32(i1);
177       const float32x4x2_t vi2x8ACE9BDF = vld2q_f32(i2);
178       const float32x4x2_t vi3x8ACE9BDF = vld2q_f32(i3);
179       const float32x4x2_t vi4x8ACE9BDF = vld2q_f32(i4);
180       const float32x4x2_t vi5x8ACE9BDF = vld2q_f32(i5);
181       const float32x4x2_t vi6x8ACE9BDF = vld2q_f32(i6);
182 
183       const float32x4_t vi0x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi0x8ACE9BDF.val[0])));
184       const float32x4_t vi0x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd,  vreinterpretq_u32_f32(vi0x8ACE9BDF.val[1])));
185       const float32x4_t vi1x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi1x8ACE9BDF.val[0])));
186       const float32x4_t vi1x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd,  vreinterpretq_u32_f32(vi1x8ACE9BDF.val[1])));
187       const float32x4_t vi2x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi2x8ACE9BDF.val[0])));
188       const float32x4_t vi2x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd,  vreinterpretq_u32_f32(vi2x8ACE9BDF.val[1])));
189       const float32x4_t vi3x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi3x8ACE9BDF.val[0])));
190       const float32x4_t vi3x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd,  vreinterpretq_u32_f32(vi3x8ACE9BDF.val[1])));
191       const float32x4_t vi4x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi4x8ACE9BDF.val[0])));
192       const float32x4_t vi4x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd,  vreinterpretq_u32_f32(vi4x8ACE9BDF.val[1])));
193       const float32x4_t vi5x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi5x8ACE9BDF.val[0])));
194       const float32x4_t vi5x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd,  vreinterpretq_u32_f32(vi5x8ACE9BDF.val[1])));
195       const float32x4_t vi6x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi6x8ACE9BDF.val[0])));
196       const float32x4_t vi6x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd,  vreinterpretq_u32_f32(vi6x8ACE9BDF.val[1])));
197 
198       vo0p0 = vfmaq_lane_f32(vo0p0, vi0x8ACE, vget_high_f32(vw0123), 0);
199       vo1p0 = vfmaq_lane_f32(vo1p0, vi2x8ACE, vget_high_f32(vw0123), 0);
200       vo2p0 = vfmaq_lane_f32(vo2p0, vi4x8ACE, vget_high_f32(vw0123), 0);
201 
202       vo0p0 = vfmaq_lane_f32(vo0p0, vi1x8ACE, vget_low_f32(vw4567), 1);
203       vo1p0 = vfmaq_lane_f32(vo1p0, vi3x8ACE, vget_low_f32(vw4567), 1);
204       vo2p0 = vfmaq_lane_f32(vo2p0, vi5x8ACE, vget_low_f32(vw4567), 1);
205 
206       vo0p0 = vfmaq_lane_f32(vo0p0, vi2x8ACE, vw89, 0);
207       vo1p0 = vfmaq_lane_f32(vo1p0, vi4x8ACE, vw89, 0);
208       vo2p0 = vfmaq_lane_f32(vo2p0, vi6x8ACE, vw89, 0);
209 
210       const float32x4_t vi0x7BDF = vextq_f32(vi0x1357, vi0x9BDF, 3);
211       const float32x4_t vi1x7BDF = vextq_f32(vi1x1357, vi1x9BDF, 3);
212       const float32x4_t vi2x7BDF = vextq_f32(vi2x1357, vi2x9BDF, 3);
213       const float32x4_t vi3x7BDF = vextq_f32(vi3x1357, vi3x9BDF, 3);
214       const float32x4_t vi4x7BDF = vextq_f32(vi4x1357, vi4x9BDF, 3);
215       const float32x4_t vi5x7BDF = vextq_f32(vi5x1357, vi5x9BDF, 3);
216       const float32x4_t vi6x7BDF = vextq_f32(vi6x1357, vi6x9BDF, 3);
217 
218       vo0p0 = vfmaq_lane_f32(vo0p0, vi0x7BDF, vget_low_f32(vw0123), 1);
219       vo1p0 = vfmaq_lane_f32(vo1p0, vi2x7BDF, vget_low_f32(vw0123), 1);
220       vo2p0 = vfmaq_lane_f32(vo2p0, vi4x7BDF, vget_low_f32(vw0123), 1);
221 
222       vo0p0 = vfmaq_lane_f32(vo0p0, vi1x7BDF, vget_low_f32(vw4567), 0);
223       vo1p0 = vfmaq_lane_f32(vo1p0, vi3x7BDF, vget_low_f32(vw4567), 0);
224       vo2p0 = vfmaq_lane_f32(vo2p0, vi5x7BDF, vget_low_f32(vw4567), 0);
225 
226       vo0p0 = vfmaq_lane_f32(vo0p0, vi2x7BDF, vget_high_f32(vw4567), 1);
227       vo1p0 = vfmaq_lane_f32(vo1p0, vi4x7BDF, vget_high_f32(vw4567), 1);
228       vo2p0 = vfmaq_lane_f32(vo2p0, vi6x7BDF, vget_high_f32(vw4567), 1);
229 
230       vo0p0 = vfmaq_lane_f32(vo0p0, vi0x9BDF, vget_high_f32(vw0123), 1);
231       vo1p0 = vfmaq_lane_f32(vo1p0, vi2x9BDF, vget_high_f32(vw0123), 1);
232       vo2p0 = vfmaq_lane_f32(vo2p0, vi4x9BDF, vget_high_f32(vw0123), 1);
233 
234       vo0p0 = vfmaq_lane_f32(vo0p0, vi1x9BDF, vget_high_f32(vw4567), 0);
235       vo1p0 = vfmaq_lane_f32(vo1p0, vi3x9BDF, vget_high_f32(vw4567), 0);
236       vo2p0 = vfmaq_lane_f32(vo2p0, vi5x9BDF, vget_high_f32(vw4567), 0);
237 
238       vo0p0 = vfmaq_lane_f32(vo0p0, vi2x9BDF, vw89, 1);
239       vo1p0 = vfmaq_lane_f32(vo1p0, vi4x9BDF, vw89, 1);
240       vo2p0 = vfmaq_lane_f32(vo2p0, vi6x9BDF, vw89, 1);
241 
242 
243       float32x4_t vo0 = vmaxq_f32(vo0p0, vmin);
244       float32x4_t vo1 = vmaxq_f32(vo1p0, vmin);
245       float32x4_t vo2 = vmaxq_f32(vo2p0, vmin);
246 
247       vo0 = vminq_f32(vo0, vmax);
248       vo1 = vminq_f32(vo1, vmax);
249       vo2 = vminq_f32(vo2, vmax);
250 
251       w += 1 * sizeof(float);
252       if (w & (8 * sizeof(float))) {
253         vst1q_f32(o2, vo2); o2 += 4;
254         vst1q_f32(o1, vo1); o1 += 4;
255         vst1q_f32(o0, vo0); o0 += 4;
256       } else {
257         float32x2_t vo0_lo = vget_low_f32(vo0);
258         float32x2_t vo1_lo = vget_low_f32(vo1);
259         float32x2_t vo2_lo = vget_low_f32(vo2);
260         if (w & (4 * sizeof(float))) {
261           vst1_f32(o2, vo2_lo); o2 += 2;
262           vst1_f32(o1, vo1_lo); o1 += 2;
263           vst1_f32(o0, vo0_lo); o0 += 2;
264 
265           vo0_lo = vget_high_f32(vo0);
266           vo1_lo = vget_high_f32(vo1);
267           vo2_lo = vget_high_f32(vo2);
268         }
269         if (w & (2 * sizeof(float))) {
270           vst1_lane_f32(o2, vo2_lo, 0); o2 += 1;
271           vst1_lane_f32(o1, vo1_lo, 0); o1 += 1;
272           vst1_lane_f32(o0, vo0_lo, 0); o0 += 1;
273         }
274       }
275     }
276 
277     i0 = (const float*) ((uintptr_t) i6 - input_decrement);
278     i1 = (const float*) ((uintptr_t) i0 + input_width);
279     i2 = (const float*) ((uintptr_t) i1 + input_width);
280     i3 = (const float*) ((uintptr_t) i2 + input_width);
281     i4 = (const float*) ((uintptr_t) i3 + input_width);
282     i5 = (const float*) ((uintptr_t) i4 + input_width);
283     i6 = (const float*) ((uintptr_t) i5 + input_width);
284 
285     o0 = o2;
286     o1 = (float*) ((uintptr_t) o0 + output_width);
287     o2 = (float*) ((uintptr_t) o1 + output_width);
288 
289     output_height = doz(output_height, 3);
290     padded_input_height = doz(padded_input_height, 6);
291   } while (output_height != 0);
292 }
293