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_minmax_ukernel_up8x3__neon_acc2(size_t channels,size_t output_width,const float ** input,const float * weights,float * output,size_t input_stride,size_t output_increment,size_t input_offset,const float * zero,const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_f32_dwconv_minmax_ukernel_up8x3__neon_acc2(
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 size_t input_offset,
26 const float* zero,
27 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
28 {
29 assert(channels != 0);
30 assert(output_width != 0);
31
32 const float32x4_t vmax = vld1q_dup_f32(¶ms->scalar.max);
33 const float32x4_t vmin = vld1q_dup_f32(¶ms->scalar.min);
34 do {
35 const float* i0 = input[0];
36 assert(i0 != NULL);
37 if XNN_UNPREDICTABLE(i0 != zero) {
38 i0 = (const float*) ((uintptr_t) i0 + input_offset);
39 }
40 const float* i1 = input[1];
41 assert(i1 != NULL);
42 if XNN_UNPREDICTABLE(i1 != zero) {
43 i1 = (const float*) ((uintptr_t) i1 + input_offset);
44 }
45 const float* i2 = input[2];
46 assert(i2 != NULL);
47 if XNN_UNPREDICTABLE(i2 != zero) {
48 i2 = (const float*) ((uintptr_t) i2 + input_offset);
49 }
50
51 input = (const float**) ((uintptr_t) input + input_stride);
52
53 size_t c = channels;
54 const float* w = weights;
55 for (; c >= 8; c -= 8) {
56 float32x4_t vacc0123p0 = vld1q_f32(w); w += 4;
57 float32x4_t vacc4567p0 = vld1q_f32(w); w += 4;
58
59
60 const float32x4_t vi0x0123 = vld1q_f32(i0); i0 += 4;
61 const float32x4_t vi0x4567 = vld1q_f32(i0); i0 += 4;
62 const float32x4_t vk0x0123 = vld1q_f32(w); w += 4;
63 const float32x4_t vk0x4567 = vld1q_f32(w); w += 4;
64 vacc0123p0 = vmlaq_f32(vacc0123p0, vi0x0123, vk0x0123);
65 vacc4567p0 = vmlaq_f32(vacc4567p0, vi0x4567, vk0x4567);
66
67 const float32x4_t vi1x0123 = vld1q_f32(i1); i1 += 4;
68 const float32x4_t vi1x4567 = vld1q_f32(i1); i1 += 4;
69 const float32x4_t vk1x0123 = vld1q_f32(w); w += 4;
70 const float32x4_t vk1x4567 = vld1q_f32(w); w += 4;
71 float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123);
72 float32x4_t vacc4567p1 = vmulq_f32(vi1x4567, vk1x4567);
73
74 const float32x4_t vi2x0123 = vld1q_f32(i2); i2 += 4;
75 const float32x4_t vi2x4567 = vld1q_f32(i2); i2 += 4;
76 const float32x4_t vk2x0123 = vld1q_f32(w); w += 4;
77 const float32x4_t vk2x4567 = vld1q_f32(w); w += 4;
78 vacc0123p0 = vmlaq_f32(vacc0123p0, vi2x0123, vk2x0123);
79 vacc4567p0 = vmlaq_f32(vacc4567p0, vi2x4567, vk2x4567);
80
81 // Add up all accumulators to vacc01234567p0
82 vacc0123p0 = vaddq_f32(vacc0123p0, vacc0123p1);
83 vacc4567p0 = vaddq_f32(vacc4567p0, vacc4567p1);
84
85 float32x4_t vacc0123 = vmaxq_f32(vacc0123p0, vmin);
86 float32x4_t vacc4567 = vmaxq_f32(vacc4567p0, vmin);
87 vacc0123 = vminq_f32(vacc0123, vmax);
88 vacc4567 = vminq_f32(vacc4567, vmax);
89
90 vst1q_f32(output, vacc0123); output += 4;
91 vst1q_f32(output, vacc4567); output += 4;
92 }
93 for (; c >= 4; c -= 4) {
94 float32x4_t vacc0123p0 = vld1q_f32(w); w += 4;
95
96
97 const float32x4_t vi0x0123 = vld1q_f32(i0); i0 += 4;
98 const float32x4_t vk0x0123 = vld1q_f32(w + 4);
99 vacc0123p0 = vmlaq_f32(vacc0123p0, vi0x0123, vk0x0123);
100
101 const float32x4_t vi1x0123 = vld1q_f32(i1); i1 += 4;
102 const float32x4_t vk1x0123 = vld1q_f32(w + 12);
103 float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123);
104
105 const float32x4_t vi2x0123 = vld1q_f32(i2); i2 += 4;
106 const float32x4_t vk2x0123 = vld1q_f32(w + 20);
107 vacc0123p0 = vmlaq_f32(vacc0123p0, vi2x0123, vk2x0123);
108
109 // Add up all accumulators to vacc0123p0
110 vacc0123p0 = vaddq_f32(vacc0123p0, vacc0123p1);
111
112 float32x4_t vacc0123 = vmaxq_f32(vacc0123p0, vmin);
113 vacc0123 = vminq_f32(vacc0123, vmax);
114
115 vst1q_f32(output, vacc0123); output += 4;
116 }
117 if XNN_UNLIKELY(c != 0) {
118 float32x4_t vacc0123p0 = vld1q_f32(w);
119
120
121 const float32x4_t vi0x0123 = vld1q_f32(i0);
122 const float32x4_t vk0x0123 = vld1q_f32(w + 8);
123 vacc0123p0 = vmlaq_f32(vacc0123p0, vi0x0123, vk0x0123);
124
125 const float32x4_t vi1x0123 = vld1q_f32(i1);
126 const float32x4_t vk1x0123 = vld1q_f32(w + 16);
127 float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123);
128
129 const float32x4_t vi2x0123 = vld1q_f32(i2);
130 const float32x4_t vk2x0123 = vld1q_f32(w + 24);
131 vacc0123p0 = vmlaq_f32(vacc0123p0, vi2x0123, vk2x0123);
132
133 // Add up all accumulators to vacc0123p0
134 vacc0123p0 = vaddq_f32(vacc0123p0, vacc0123p1);
135
136 float32x4_t vacc0123 = vmaxq_f32(vacc0123p0, vmin);
137 vacc0123 = vminq_f32(vacc0123, vmax);
138
139 float32x2_t vacc01 = vget_low_f32(vacc0123);
140 if (c & 2) {
141 vst1_f32(output, vacc01); output += 2;
142 vacc01 = vget_high_f32(vacc0123);
143 }
144 if (c & 1) {
145 vst1_lane_f32(output, vacc01, 0); output += 1;
146 }
147 }
148
149 output = (float*) ((uintptr_t) output + output_increment);
150 } while (--output_width != 0);
151 }
152