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_up16x3__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_up16x3__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 >= 16; c -= 16) {
56 float32x4_t vacc0123p0 = vld1q_f32(w); w += 4;
57 float32x4_t vacc4567p0 = vld1q_f32(w); w += 4;
58 float32x4_t vacc89ABp0 = vld1q_f32(w); w += 4;
59 float32x4_t vaccCDEFp0 = vld1q_f32(w); w += 4;
60
61
62 const float32x4_t vi0x0123 = vld1q_f32(i0); i0 += 4;
63 const float32x4_t vi0x4567 = vld1q_f32(i0); i0 += 4;
64 const float32x4_t vi0x89AB = vld1q_f32(i0); i0 += 4;
65 const float32x4_t vi0xCDEF = vld1q_f32(i0); i0 += 4;
66 const float32x4_t vk0x0123 = vld1q_f32(w); w += 4;
67 const float32x4_t vk0x4567 = vld1q_f32(w); w += 4;
68 const float32x4_t vk0x89AB = vld1q_f32(w); w += 4;
69 const float32x4_t vk0xCDEF = vld1q_f32(w); w += 4;
70 vacc0123p0 = vmlaq_f32(vacc0123p0, vi0x0123, vk0x0123);
71 vacc4567p0 = vmlaq_f32(vacc4567p0, vi0x4567, vk0x4567);
72 vacc89ABp0 = vmlaq_f32(vacc89ABp0, vi0x89AB, vk0x89AB);
73 vaccCDEFp0 = vmlaq_f32(vaccCDEFp0, vi0xCDEF, vk0xCDEF);
74
75 const float32x4_t vi1x0123 = vld1q_f32(i1); i1 += 4;
76 const float32x4_t vi1x4567 = vld1q_f32(i1); i1 += 4;
77 const float32x4_t vi1x89AB = vld1q_f32(i1); i1 += 4;
78 const float32x4_t vi1xCDEF = vld1q_f32(i1); i1 += 4;
79 const float32x4_t vk1x0123 = vld1q_f32(w); w += 4;
80 const float32x4_t vk1x4567 = vld1q_f32(w); w += 4;
81 const float32x4_t vk1x89AB = vld1q_f32(w); w += 4;
82 const float32x4_t vk1xCDEF = vld1q_f32(w); w += 4;
83 float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123);
84 float32x4_t vacc4567p1 = vmulq_f32(vi1x4567, vk1x4567);
85 float32x4_t vacc89ABp1 = vmulq_f32(vi1x89AB, vk1x89AB);
86 float32x4_t vaccCDEFp1 = vmulq_f32(vi1xCDEF, vk1xCDEF);
87
88 const float32x4_t vi2x0123 = vld1q_f32(i2); i2 += 4;
89 const float32x4_t vi2x4567 = vld1q_f32(i2); i2 += 4;
90 const float32x4_t vi2x89AB = vld1q_f32(i2); i2 += 4;
91 const float32x4_t vi2xCDEF = vld1q_f32(i2); i2 += 4;
92 const float32x4_t vk2x0123 = vld1q_f32(w); w += 4;
93 const float32x4_t vk2x4567 = vld1q_f32(w); w += 4;
94 const float32x4_t vk2x89AB = vld1q_f32(w); w += 4;
95 const float32x4_t vk2xCDEF = vld1q_f32(w); w += 4;
96 vacc0123p0 = vmlaq_f32(vacc0123p0, vi2x0123, vk2x0123);
97 vacc4567p0 = vmlaq_f32(vacc4567p0, vi2x4567, vk2x4567);
98 vacc89ABp0 = vmlaq_f32(vacc89ABp0, vi2x89AB, vk2x89AB);
99 vaccCDEFp0 = vmlaq_f32(vaccCDEFp0, vi2xCDEF, vk2xCDEF);
100
101 // Add up all accumulators to vacc0123456789ABCDEFp0
102 vacc0123p0 = vaddq_f32(vacc0123p0, vacc0123p1);
103 vacc4567p0 = vaddq_f32(vacc4567p0, vacc4567p1);
104 vacc89ABp0 = vaddq_f32(vacc89ABp0, vacc89ABp1);
105 vaccCDEFp0 = vaddq_f32(vaccCDEFp0, vaccCDEFp1);
106
107 float32x4_t vacc0123 = vmaxq_f32(vacc0123p0, vmin);
108 float32x4_t vacc4567 = vmaxq_f32(vacc4567p0, vmin);
109 float32x4_t vacc89AB = vmaxq_f32(vacc89ABp0, vmin);
110 float32x4_t vaccCDEF = vmaxq_f32(vaccCDEFp0, vmin);
111 vacc0123 = vminq_f32(vacc0123, vmax);
112 vacc4567 = vminq_f32(vacc4567, vmax);
113 vacc89AB = vminq_f32(vacc89AB, vmax);
114 vaccCDEF = vminq_f32(vaccCDEF, vmax);
115
116 vst1q_f32(output, vacc0123); output += 4;
117 vst1q_f32(output, vacc4567); output += 4;
118 vst1q_f32(output, vacc89AB); output += 4;
119 vst1q_f32(output, vaccCDEF); output += 4;
120 }
121 for (; c >= 4; c -= 4) {
122 float32x4_t vacc0123p0 = vld1q_f32(w); w += 4;
123
124
125 const float32x4_t vi0x0123 = vld1q_f32(i0); i0 += 4;
126 const float32x4_t vk0x0123 = vld1q_f32(w + 12);
127 vacc0123p0 = vmlaq_f32(vacc0123p0, vi0x0123, vk0x0123);
128
129 const float32x4_t vi1x0123 = vld1q_f32(i1); i1 += 4;
130 const float32x4_t vk1x0123 = vld1q_f32(w + 28);
131 float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123);
132
133 const float32x4_t vi2x0123 = vld1q_f32(i2); i2 += 4;
134 const float32x4_t vk2x0123 = vld1q_f32(w + 44);
135 vacc0123p0 = vmlaq_f32(vacc0123p0, vi2x0123, vk2x0123);
136
137 // Add up all accumulators to vacc0123p0
138 vacc0123p0 = vaddq_f32(vacc0123p0, vacc0123p1);
139
140 float32x4_t vacc0123 = vmaxq_f32(vacc0123p0, vmin);
141 vacc0123 = vminq_f32(vacc0123, vmax);
142
143 vst1q_f32(output, vacc0123); output += 4;
144 }
145 if XNN_UNLIKELY(c != 0) {
146 float32x4_t vacc0123p0 = vld1q_f32(w);
147
148
149 const float32x4_t vi0x0123 = vld1q_f32(i0);
150 const float32x4_t vk0x0123 = vld1q_f32(w + 16);
151 vacc0123p0 = vmlaq_f32(vacc0123p0, vi0x0123, vk0x0123);
152
153 const float32x4_t vi1x0123 = vld1q_f32(i1);
154 const float32x4_t vk1x0123 = vld1q_f32(w + 32);
155 float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123);
156
157 const float32x4_t vi2x0123 = vld1q_f32(i2);
158 const float32x4_t vk2x0123 = vld1q_f32(w + 48);
159 vacc0123p0 = vmlaq_f32(vacc0123p0, vi2x0123, vk2x0123);
160
161 // Add up all accumulators to vacc0123p0
162 vacc0123p0 = vaddq_f32(vacc0123p0, vacc0123p1);
163
164 float32x4_t vacc0123 = vmaxq_f32(vacc0123p0, vmin);
165 vacc0123 = vminq_f32(vacc0123, vmax);
166
167 float32x2_t vacc01 = vget_low_f32(vacc0123);
168 if (c & 2) {
169 vst1_f32(output, vacc01); output += 2;
170 vacc01 = vget_high_f32(vacc0123);
171 }
172 if (c & 1) {
173 vst1_lane_f32(output, vacc01, 0); output += 1;
174 }
175 }
176
177 output = (float*) ((uintptr_t) output + output_increment);
178 } while (--output_width != 0);
179 }
180