1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-dwconv2d-chw/3x3p1-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_3x3p1__neonfma_2x4(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_3x3p1__neonfma_2x4(
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 
33   const uint32x4_t vmask = vld1q_u32(params->neon.mask);
34   const float32x4_t vmax = vld1q_dup_f32(&params->neon.max);
35   const float32x4_t vmin = vld1q_dup_f32(&params->neon.min);
36 
37   const float32x4_t vw0123 = vld1q_f32(weights);
38   const float32x4_t vw4567 = vld1q_f32(weights + 4);
39   const float32x2_t vw89 = vld1_f32(weights + 8);
40 
41   const size_t input_decrement = round_up_po2(input_width, 4 * sizeof(float));
42 
43   const float* i0 = zero;
44   const float* i1 = input;
45   const float* i2 = (const float*) ((uintptr_t) i1 + input_width);
46   const float* i3 = (const float*) ((uintptr_t) i2 + input_width);
47 
48   float* o0 = output;
49   float* o1 = (float*) ((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     float32x4_t vi0x0123 = vmovq_n_f32(0.0f);
62     float32x4_t vi1x0123 = vmovq_n_f32(0.0f);
63     float32x4_t vi2x0123 = vmovq_n_f32(0.0f);
64     float32x4_t vi3x0123 = vmovq_n_f32(0.0f);
65 
66     float32x4_t vi0x4567 = vld1q_f32(i0); i0 += 4;
67     float32x4_t vi1x4567 = vld1q_f32(i1); i1 += 4;
68     float32x4_t vi2x4567 = vld1q_f32(i2); i2 += 4;
69     float32x4_t vi3x4567 = vld1q_f32(i3); i3 += 4;
70 
71     size_t w = input_width;
72     for (; w > 4 * sizeof(float); w -= 4 * sizeof(float)) {
73       float32x4_t vo0p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0);
74       float32x4_t vo1p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0);
75 
76       const float32x4_t vi0x89AB = vld1q_f32(i0); i0 += 4;
77       const float32x4_t vi1x89AB = vld1q_f32(i1); i1 += 4;
78       const float32x4_t vi2x89AB = vld1q_f32(i2); i2 += 4;
79       const float32x4_t vi3x89AB = vld1q_f32(i3); i3 += 4;
80 
81       vo0p0 = vfmaq_lane_f32(vo0p0, vi0x4567, vget_high_f32(vw0123), 0);
82       vo1p0 = vfmaq_lane_f32(vo1p0, vi1x4567, vget_high_f32(vw0123), 0);
83 
84       vo0p0 = vfmaq_lane_f32(vo0p0, vi1x4567, vget_low_f32(vw4567), 1);
85       vo1p0 = vfmaq_lane_f32(vo1p0, vi2x4567, vget_low_f32(vw4567), 1);
86 
87       vo0p0 = vfmaq_lane_f32(vo0p0, vi2x4567, vw89, 0);
88       vo1p0 = vfmaq_lane_f32(vo1p0, vi3x4567, vw89, 0);
89 
90       const float32x4_t vi0x3456 = vextq_f32(vi0x0123, vi0x4567, 3);
91       const float32x4_t vi1x3456 = vextq_f32(vi1x0123, vi1x4567, 3);
92       const float32x4_t vi2x3456 = vextq_f32(vi2x0123, vi2x4567, 3);
93       const float32x4_t vi3x3456 = vextq_f32(vi3x0123, vi3x4567, 3);
94 
95       vo0p0 = vfmaq_lane_f32(vo0p0, vi0x3456, vget_low_f32(vw0123), 1);
96       vo1p0 = vfmaq_lane_f32(vo1p0, vi1x3456, vget_low_f32(vw0123), 1);
97 
98       vo0p0 = vfmaq_lane_f32(vo0p0, vi1x3456, vget_low_f32(vw4567), 0);
99       vo1p0 = vfmaq_lane_f32(vo1p0, vi2x3456, vget_low_f32(vw4567), 0);
100 
101       vo0p0 = vfmaq_lane_f32(vo0p0, vi2x3456, vget_high_f32(vw4567), 1);
102       vo1p0 = vfmaq_lane_f32(vo1p0, vi3x3456, vget_high_f32(vw4567), 1);
103 
104       vi0x0123 = vi0x4567;
105       vi1x0123 = vi1x4567;
106       vi2x0123 = vi2x4567;
107       vi3x0123 = vi3x4567;
108 
109       const float32x4_t vi0x5678 = vextq_f32(vi0x4567, vi0x89AB, 1);
110       const float32x4_t vi1x5678 = vextq_f32(vi1x4567, vi1x89AB, 1);
111       const float32x4_t vi2x5678 = vextq_f32(vi2x4567, vi2x89AB, 1);
112       const float32x4_t vi3x5678 = vextq_f32(vi3x4567, vi3x89AB, 1);
113 
114       vo0p0 = vfmaq_lane_f32(vo0p0, vi0x5678, vget_high_f32(vw0123), 1);
115       vo1p0 = vfmaq_lane_f32(vo1p0, vi1x5678, vget_high_f32(vw0123), 1);
116 
117       vo0p0 = vfmaq_lane_f32(vo0p0, vi1x5678, vget_high_f32(vw4567), 0);
118       vo1p0 = vfmaq_lane_f32(vo1p0, vi2x5678, vget_high_f32(vw4567), 0);
119 
120       vo0p0 = vfmaq_lane_f32(vo0p0, vi2x5678, vw89, 1);
121       vo1p0 = vfmaq_lane_f32(vo1p0, vi3x5678, vw89, 1);
122 
123       vi0x4567 = vi0x89AB;
124       vi1x4567 = vi1x89AB;
125       vi2x4567 = vi2x89AB;
126       vi3x4567 = vi3x89AB;
127 
128 
129       float32x4_t vo0 = vmaxq_f32(vo0p0, vmin);
130       float32x4_t vo1 = vmaxq_f32(vo1p0, vmin);
131 
132       vo0 = vminq_f32(vo0, vmax);
133       vo1 = vminq_f32(vo1, vmax);
134 
135       vst1q_f32(o1, vo1); o1 += 4;
136       vst1q_f32(o0, vo0); o0 += 4;
137     }
138     // Always process the last block of 1..4 pixels.
139     assert(w >= 1 * sizeof(float));
140     assert(w <= 4 * sizeof(float));
141     {
142       float32x4_t vo0p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0);
143       float32x4_t vo1p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0);
144 
145       vi0x4567 = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi0x4567)));
146       vi1x4567 = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi1x4567)));
147       vi2x4567 = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi2x4567)));
148       vi3x4567 = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi3x4567)));
149 
150       vo0p0 = vfmaq_lane_f32(vo0p0, vi0x4567, vget_high_f32(vw0123), 0);
151       vo1p0 = vfmaq_lane_f32(vo1p0, vi1x4567, vget_high_f32(vw0123), 0);
152 
153       vo0p0 = vfmaq_lane_f32(vo0p0, vi1x4567, vget_low_f32(vw4567), 1);
154       vo1p0 = vfmaq_lane_f32(vo1p0, vi2x4567, vget_low_f32(vw4567), 1);
155 
156       vo0p0 = vfmaq_lane_f32(vo0p0, vi2x4567, vw89, 0);
157       vo1p0 = vfmaq_lane_f32(vo1p0, vi3x4567, vw89, 0);
158 
159       const float32x4_t vi0x3456 = vextq_f32(vi0x0123, vi0x4567, 3);
160       const float32x4_t vi1x3456 = vextq_f32(vi1x0123, vi1x4567, 3);
161       const float32x4_t vi2x3456 = vextq_f32(vi2x0123, vi2x4567, 3);
162       const float32x4_t vi3x3456 = vextq_f32(vi3x0123, vi3x4567, 3);
163 
164       vo0p0 = vfmaq_lane_f32(vo0p0, vi0x3456, vget_low_f32(vw0123), 1);
165       vo1p0 = vfmaq_lane_f32(vo1p0, vi1x3456, vget_low_f32(vw0123), 1);
166 
167       vo0p0 = vfmaq_lane_f32(vo0p0, vi1x3456, vget_low_f32(vw4567), 0);
168       vo1p0 = vfmaq_lane_f32(vo1p0, vi2x3456, vget_low_f32(vw4567), 0);
169 
170       vo0p0 = vfmaq_lane_f32(vo0p0, vi2x3456, vget_high_f32(vw4567), 1);
171       vo1p0 = vfmaq_lane_f32(vo1p0, vi3x3456, vget_high_f32(vw4567), 1);
172 
173       const float32x4_t vzero = vmovq_n_f32(0.0f);
174       const float32x4_t vi0x5678 = vextq_f32(vi0x4567, vzero, 1);
175       const float32x4_t vi1x5678 = vextq_f32(vi1x4567, vzero, 1);
176       const float32x4_t vi2x5678 = vextq_f32(vi2x4567, vzero, 1);
177       const float32x4_t vi3x5678 = vextq_f32(vi3x4567, vzero, 1);
178 
179       vo0p0 = vfmaq_lane_f32(vo0p0, vi0x5678, vget_high_f32(vw0123), 1);
180       vo1p0 = vfmaq_lane_f32(vo1p0, vi1x5678, vget_high_f32(vw0123), 1);
181 
182       vo0p0 = vfmaq_lane_f32(vo0p0, vi1x5678, vget_high_f32(vw4567), 0);
183       vo1p0 = vfmaq_lane_f32(vo1p0, vi2x5678, vget_high_f32(vw4567), 0);
184 
185       vo0p0 = vfmaq_lane_f32(vo0p0, vi2x5678, vw89, 1);
186       vo1p0 = vfmaq_lane_f32(vo1p0, vi3x5678, vw89, 1);
187 
188 
189       float32x4_t vo0 = vmaxq_f32(vo0p0, vmin);
190       float32x4_t vo1 = vmaxq_f32(vo1p0, vmin);
191 
192       vo0 = vminq_f32(vo0, vmax);
193       vo1 = vminq_f32(vo1, vmax);
194 
195       if XNN_LIKELY(w == 4 * sizeof(float)) {
196         vst1q_f32(o1, vo1); o1 += 4;
197         vst1q_f32(o0, vo0); o0 += 4;
198       } else {
199         float32x2_t vo0_lo = vget_low_f32(vo0);
200         float32x2_t vo1_lo = vget_low_f32(vo1);
201         if (w & (2 * sizeof(float))) {
202           vst1_f32(o1, vo1_lo); o1 += 2;
203           vst1_f32(o0, vo0_lo); o0 += 2;
204 
205           vo0_lo = vget_high_f32(vo0);
206           vo1_lo = vget_high_f32(vo1);
207         }
208         if (w & (1 * sizeof(float))) {
209           vst1_lane_f32(o1, vo1_lo, 0); o1 += 1;
210           vst1_lane_f32(o0, vo0_lo, 0); o0 += 1;
211         }
212       }
213     }
214 
215     i0 = (const float*) ((uintptr_t) i2 - input_decrement);
216     i1 = (const float*) ((uintptr_t) i3 - input_decrement);
217     i2 = (const float*) ((uintptr_t) i1 + input_width);
218     i3 = (const float*) ((uintptr_t) i2 + input_width);
219 
220     o0 = o1;
221     o1 = (float*) ((uintptr_t) o0 + input_width);
222 
223     output_height = doz(output_height, 2);
224   } while (output_height != 0);
225 }
226