1 // Auto-generated file. Do not edit!
2 // Template: src/f32-dwconv/up-neon.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2019 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
16
xnn_f32_dwconv_ukernel_up4x9__neonfma(size_t channels,size_t output_width,const float ** input,const float * weights,float * output,size_t input_stride,size_t output_increment,const union xnn_f32_output_params params[restrict static1])17 void xnn_f32_dwconv_ukernel_up4x9__neonfma(
18 size_t channels,
19 size_t output_width,
20 const float** input,
21 const float* weights,
22 float* output,
23 size_t input_stride,
24 size_t output_increment,
25 const union xnn_f32_output_params params[restrict static 1])
26 {
27 assert(channels != 0);
28 assert(output_width != 0);
29
30 const float32x4_t vmax = vld1q_dup_f32(¶ms->scalar.max);
31 const float32x4_t vmin = vld1q_dup_f32(¶ms->scalar.min);
32 do {
33 const float* i0 = input[0];
34 assert(i0 != NULL);
35 const float* i1 = input[1];
36 assert(i1 != NULL);
37 const float* i2 = input[2];
38 assert(i2 != NULL);
39 const float* i3 = input[3];
40 assert(i3 != NULL);
41 const float* i4 = input[4];
42 assert(i4 != NULL);
43 const float* i5 = input[5];
44 assert(i5 != NULL);
45 const float* i6 = input[6];
46 assert(i6 != NULL);
47 const float* i7 = input[7];
48 assert(i7 != NULL);
49 const float* i8 = input[8];
50 assert(i8 != NULL);
51 input = (const float**) ((uintptr_t) input + input_stride);
52
53 size_t c = channels;
54 const float* w = weights;
55 for (; c >= 4; c -= 4) {
56 float32x4_t vacc0123p0 = vld1q_f32(w); w += 4;
57
58
59 const float32x4_t vi0x0123 = vld1q_f32(i0); i0 += 4;
60 const float32x4_t vk0x0123 = vld1q_f32(w); w += 4;
61 vacc0123p0 = vfmaq_f32(vacc0123p0, vi0x0123, vk0x0123);
62
63 const float32x4_t vi1x0123 = vld1q_f32(i1); i1 += 4;
64 const float32x4_t vk1x0123 = vld1q_f32(w); w += 4;
65 vacc0123p0 = vfmaq_f32(vacc0123p0, vi1x0123, vk1x0123);
66
67 const float32x4_t vi2x0123 = vld1q_f32(i2); i2 += 4;
68 const float32x4_t vk2x0123 = vld1q_f32(w); w += 4;
69 vacc0123p0 = vfmaq_f32(vacc0123p0, vi2x0123, vk2x0123);
70
71 const float32x4_t vi3x0123 = vld1q_f32(i3); i3 += 4;
72 const float32x4_t vk3x0123 = vld1q_f32(w); w += 4;
73 vacc0123p0 = vfmaq_f32(vacc0123p0, vi3x0123, vk3x0123);
74
75 const float32x4_t vi4x0123 = vld1q_f32(i4); i4 += 4;
76 const float32x4_t vk4x0123 = vld1q_f32(w); w += 4;
77 vacc0123p0 = vfmaq_f32(vacc0123p0, vi4x0123, vk4x0123);
78
79 const float32x4_t vi5x0123 = vld1q_f32(i5); i5 += 4;
80 const float32x4_t vk5x0123 = vld1q_f32(w); w += 4;
81 vacc0123p0 = vfmaq_f32(vacc0123p0, vi5x0123, vk5x0123);
82
83 const float32x4_t vi6x0123 = vld1q_f32(i6); i6 += 4;
84 const float32x4_t vk6x0123 = vld1q_f32(w); w += 4;
85 vacc0123p0 = vfmaq_f32(vacc0123p0, vi6x0123, vk6x0123);
86
87 const float32x4_t vi7x0123 = vld1q_f32(i7); i7 += 4;
88 const float32x4_t vk7x0123 = vld1q_f32(w); w += 4;
89 vacc0123p0 = vfmaq_f32(vacc0123p0, vi7x0123, vk7x0123);
90
91 const float32x4_t vi8x0123 = vld1q_f32(i8); i8 += 4;
92 const float32x4_t vk8x0123 = vld1q_f32(w); w += 4;
93 vacc0123p0 = vfmaq_f32(vacc0123p0, vi8x0123, vk8x0123);
94
95
96 float32x4_t vacc0123 = vmaxq_f32(vacc0123p0, vmin);
97 vacc0123 = vminq_f32(vacc0123, vmax);
98
99 vst1q_f32(output, vacc0123); output += 4;
100 }
101 if XNN_UNLIKELY(c != 0) {
102 float32x4_t vacc0123p0 = vld1q_f32(w); w += 4;
103
104
105 const float32x4_t vi0x0123 = vld1q_f32(i0);
106 const float32x4_t vk0x0123 = vld1q_f32(w); w += 4;
107 vacc0123p0 = vfmaq_f32(vacc0123p0, vi0x0123, vk0x0123);
108
109 const float32x4_t vi1x0123 = vld1q_f32(i1);
110 const float32x4_t vk1x0123 = vld1q_f32(w); w += 4;
111 vacc0123p0 = vfmaq_f32(vacc0123p0, vi1x0123, vk1x0123);
112
113 const float32x4_t vi2x0123 = vld1q_f32(i2);
114 const float32x4_t vk2x0123 = vld1q_f32(w); w += 4;
115 vacc0123p0 = vfmaq_f32(vacc0123p0, vi2x0123, vk2x0123);
116
117 const float32x4_t vi3x0123 = vld1q_f32(i3);
118 const float32x4_t vk3x0123 = vld1q_f32(w); w += 4;
119 vacc0123p0 = vfmaq_f32(vacc0123p0, vi3x0123, vk3x0123);
120
121 const float32x4_t vi4x0123 = vld1q_f32(i4);
122 const float32x4_t vk4x0123 = vld1q_f32(w); w += 4;
123 vacc0123p0 = vfmaq_f32(vacc0123p0, vi4x0123, vk4x0123);
124
125 const float32x4_t vi5x0123 = vld1q_f32(i5);
126 const float32x4_t vk5x0123 = vld1q_f32(w); w += 4;
127 vacc0123p0 = vfmaq_f32(vacc0123p0, vi5x0123, vk5x0123);
128
129 const float32x4_t vi6x0123 = vld1q_f32(i6);
130 const float32x4_t vk6x0123 = vld1q_f32(w); w += 4;
131 vacc0123p0 = vfmaq_f32(vacc0123p0, vi6x0123, vk6x0123);
132
133 const float32x4_t vi7x0123 = vld1q_f32(i7);
134 const float32x4_t vk7x0123 = vld1q_f32(w); w += 4;
135 vacc0123p0 = vfmaq_f32(vacc0123p0, vi7x0123, vk7x0123);
136
137 const float32x4_t vi8x0123 = vld1q_f32(i8);
138 const float32x4_t vk8x0123 = vld1q_f32(w); w += 4;
139 vacc0123p0 = vfmaq_f32(vacc0123p0, vi8x0123, vk8x0123);
140
141
142 float32x4_t vacc0123 = vmaxq_f32(vacc0123p0, vmin);
143 vacc0123 = vminq_f32(vacc0123, vmax);
144
145 float32x2_t vacc01 = vget_low_f32(vacc0123);
146 if (c & 2) {
147 vst1_f32(output, vacc01); output += 2;
148 vacc01 = vget_high_f32(vacc0123);
149 }
150 if (c & 1) {
151 vst1_lane_f32(output, vacc01, 0); output += 1;
152 }
153 }
154
155 output = (float*) ((uintptr_t) output + output_increment);
156 } while (--output_width != 0);
157 }
158