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_up8x4__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_up8x4__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 const float* i3 = input[3];
51 assert(i3 != NULL);
52 if XNN_UNPREDICTABLE(i3 != zero) {
53 i3 = (const float*) ((uintptr_t) i3 + input_offset);
54 }
55
56 input = (const float**) ((uintptr_t) input + input_stride);
57
58 size_t c = channels;
59 const float* w = weights;
60 for (; c >= 8; c -= 8) {
61 float32x4_t vacc0123p0 = vld1q_f32(w); w += 4;
62 float32x4_t vacc4567p0 = vld1q_f32(w); w += 4;
63
64
65 const float32x4_t vi0x0123 = vld1q_f32(i0); i0 += 4;
66 const float32x4_t vi0x4567 = vld1q_f32(i0); i0 += 4;
67 const float32x4_t vk0x0123 = vld1q_f32(w); w += 4;
68 const float32x4_t vk0x4567 = vld1q_f32(w); w += 4;
69 vacc0123p0 = vmlaq_f32(vacc0123p0, vi0x0123, vk0x0123);
70 vacc4567p0 = vmlaq_f32(vacc4567p0, vi0x4567, vk0x4567);
71
72 const float32x4_t vi1x0123 = vld1q_f32(i1); i1 += 4;
73 const float32x4_t vi1x4567 = vld1q_f32(i1); i1 += 4;
74 const float32x4_t vk1x0123 = vld1q_f32(w); w += 4;
75 const float32x4_t vk1x4567 = vld1q_f32(w); w += 4;
76 float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123);
77 float32x4_t vacc4567p1 = vmulq_f32(vi1x4567, vk1x4567);
78
79 const float32x4_t vi2x0123 = vld1q_f32(i2); i2 += 4;
80 const float32x4_t vi2x4567 = vld1q_f32(i2); i2 += 4;
81 const float32x4_t vk2x0123 = vld1q_f32(w); w += 4;
82 const float32x4_t vk2x4567 = vld1q_f32(w); w += 4;
83 vacc0123p0 = vmlaq_f32(vacc0123p0, vi2x0123, vk2x0123);
84 vacc4567p0 = vmlaq_f32(vacc4567p0, vi2x4567, vk2x4567);
85
86 const float32x4_t vi3x0123 = vld1q_f32(i3); i3 += 4;
87 const float32x4_t vi3x4567 = vld1q_f32(i3); i3 += 4;
88 const float32x4_t vk3x0123 = vld1q_f32(w); w += 4;
89 const float32x4_t vk3x4567 = vld1q_f32(w); w += 4;
90 vacc0123p1 = vmlaq_f32(vacc0123p1, vi3x0123, vk3x0123);
91 vacc4567p1 = vmlaq_f32(vacc4567p1, vi3x4567, vk3x4567);
92
93 // Add up all accumulators to vacc01234567p0
94 vacc0123p0 = vaddq_f32(vacc0123p0, vacc0123p1);
95 vacc4567p0 = vaddq_f32(vacc4567p0, vacc4567p1);
96
97 float32x4_t vacc0123 = vmaxq_f32(vacc0123p0, vmin);
98 float32x4_t vacc4567 = vmaxq_f32(vacc4567p0, vmin);
99 vacc0123 = vminq_f32(vacc0123, vmax);
100 vacc4567 = vminq_f32(vacc4567, vmax);
101
102 vst1q_f32(output, vacc0123); output += 4;
103 vst1q_f32(output, vacc4567); output += 4;
104 }
105 for (; c >= 4; c -= 4) {
106 float32x4_t vacc0123p0 = vld1q_f32(w); w += 4;
107
108
109 const float32x4_t vi0x0123 = vld1q_f32(i0); i0 += 4;
110 const float32x4_t vk0x0123 = vld1q_f32(w + 4);
111 vacc0123p0 = vmlaq_f32(vacc0123p0, vi0x0123, vk0x0123);
112
113 const float32x4_t vi1x0123 = vld1q_f32(i1); i1 += 4;
114 const float32x4_t vk1x0123 = vld1q_f32(w + 12);
115 float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123);
116
117 const float32x4_t vi2x0123 = vld1q_f32(i2); i2 += 4;
118 const float32x4_t vk2x0123 = vld1q_f32(w + 20);
119 vacc0123p0 = vmlaq_f32(vacc0123p0, vi2x0123, vk2x0123);
120
121 const float32x4_t vi3x0123 = vld1q_f32(i3); i3 += 4;
122 const float32x4_t vk3x0123 = vld1q_f32(w + 28);
123 vacc0123p1 = vmlaq_f32(vacc0123p1, vi3x0123, vk3x0123);
124
125 // Add up all accumulators to vacc0123p0
126 vacc0123p0 = vaddq_f32(vacc0123p0, vacc0123p1);
127
128 float32x4_t vacc0123 = vmaxq_f32(vacc0123p0, vmin);
129 vacc0123 = vminq_f32(vacc0123, vmax);
130
131 vst1q_f32(output, vacc0123); output += 4;
132 }
133 if XNN_UNLIKELY(c != 0) {
134 float32x4_t vacc0123p0 = vld1q_f32(w);
135
136
137 const float32x4_t vi0x0123 = vld1q_f32(i0);
138 const float32x4_t vk0x0123 = vld1q_f32(w + 8);
139 vacc0123p0 = vmlaq_f32(vacc0123p0, vi0x0123, vk0x0123);
140
141 const float32x4_t vi1x0123 = vld1q_f32(i1);
142 const float32x4_t vk1x0123 = vld1q_f32(w + 16);
143 float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123);
144
145 const float32x4_t vi2x0123 = vld1q_f32(i2);
146 const float32x4_t vk2x0123 = vld1q_f32(w + 24);
147 vacc0123p0 = vmlaq_f32(vacc0123p0, vi2x0123, vk2x0123);
148
149 const float32x4_t vi3x0123 = vld1q_f32(i3);
150 const float32x4_t vk3x0123 = vld1q_f32(w + 32);
151 vacc0123p1 = vmlaq_f32(vacc0123p1, vi3x0123, vk3x0123);
152
153 // Add up all accumulators to vacc0123p0
154 vacc0123p0 = vaddq_f32(vacc0123p0, vacc0123p1);
155
156 float32x4_t vacc0123 = vmaxq_f32(vacc0123p0, vmin);
157 vacc0123 = vminq_f32(vacc0123, vmax);
158
159 float32x2_t vacc01 = vget_low_f32(vacc0123);
160 if (c & 2) {
161 vst1_f32(output, vacc01); output += 2;
162 vacc01 = vget_high_f32(vacc0123);
163 }
164 if (c & 1) {
165 vst1_lane_f32(output, vacc01, 0); output += 1;
166 }
167 }
168
169 output = (float*) ((uintptr_t) output + output_increment);
170 } while (--output_width != 0);
171 }
172