• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Auto-generated file. Do not edit!
2 //   Template: src/f16-dwconv/up-neonfp16arith.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 
16 
xnn_f16_dwconv_minmax_ukernel_up8x4__neonfp16arith(size_t channels,size_t output_width,const void ** input,const void * weights,void * output_ptr,size_t input_stride,size_t output_increment,size_t input_offset,const void * zero,const struct xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_f16_dwconv_minmax_ukernel_up8x4__neonfp16arith(
18     size_t channels,
19     size_t output_width,
20     const void** input,
21     const void* weights,
22     void* output_ptr,
23     size_t input_stride,
24     size_t output_increment,
25     size_t input_offset,
26     const void* zero,
27     const struct xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
28 {
29   assert(channels != 0);
30   assert(output_width != 0);
31 
32   __fp16* output = ( __fp16*) output_ptr;
33   const float16x8_t vmax = vld1q_dup_f16(&params->max);
34   const float16x8_t vmin = vld1q_dup_f16(&params->min);
35   do {
36     const __fp16* i0 = (const __fp16*) input[0];
37     assert(i0 != NULL);
38     if XNN_UNPREDICTABLE(i0 != (const __fp16*) zero) {
39       i0 = (const __fp16*) ((uintptr_t) i0 + input_offset);
40     }
41     const __fp16* i1 = (const __fp16*) input[1];
42     assert(i1 != NULL);
43     if XNN_UNPREDICTABLE(i1 != (const __fp16*) zero) {
44       i1 = (const __fp16*) ((uintptr_t) i1 + input_offset);
45     }
46     const __fp16* i2 = (const __fp16*) input[2];
47     assert(i2 != NULL);
48     if XNN_UNPREDICTABLE(i2 != (const __fp16*) zero) {
49       i2 = (const __fp16*) ((uintptr_t) i2 + input_offset);
50     }
51     const __fp16* i3 = (const __fp16*) input[3];
52     assert(i3 != NULL);
53     if XNN_UNPREDICTABLE(i3 != (const __fp16*) zero) {
54       i3 = (const __fp16*) ((uintptr_t) i3 + input_offset);
55     }
56 
57     input = (const void**) ((uintptr_t) input + input_stride);
58 
59     size_t c = channels;
60     const __fp16* w = (const __fp16*) weights;
61     for (; c >= 8; c -= 8) {
62       float16x8_t vacc01234567p0 = vld1q_f16(w); w += 8;
63 
64 
65       const float16x8_t vi0x01234567 = vld1q_f16(i0); i0 += 8;
66       const float16x8_t vk0x01234567 = vld1q_f16(w); w += 8;
67       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi0x01234567, vk0x01234567);
68 
69       const float16x8_t vi1x01234567 = vld1q_f16(i1); i1 += 8;
70       const float16x8_t vk1x01234567 = vld1q_f16(w); w += 8;
71       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi1x01234567, vk1x01234567);
72 
73       const float16x8_t vi2x01234567 = vld1q_f16(i2); i2 += 8;
74       const float16x8_t vk2x01234567 = vld1q_f16(w); w += 8;
75       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi2x01234567, vk2x01234567);
76 
77       const float16x8_t vi3x01234567 = vld1q_f16(i3); i3 += 8;
78       const float16x8_t vk3x01234567 = vld1q_f16(w); w += 8;
79       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi3x01234567, vk3x01234567);
80 
81 
82       float16x8_t vacc01234567 = vmaxq_f16(vacc01234567p0, vmin);
83       vacc01234567 = vminq_f16(vacc01234567, vmax);
84 
85       vst1q_f16(output, vacc01234567); output += 8;
86     }
87     if XNN_UNLIKELY(c != 0) {
88       float16x8_t vacc01234567p0 = vld1q_f16(w); w += 8;
89 
90 
91       const float16x8_t vi0x01234567 = vld1q_f16(i0);
92       const float16x8_t vk0x01234567 = vld1q_f16(w); w += 8;
93       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi0x01234567, vk0x01234567);
94 
95       const float16x8_t vi1x01234567 = vld1q_f16(i1);
96       const float16x8_t vk1x01234567 = vld1q_f16(w); w += 8;
97       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi1x01234567, vk1x01234567);
98 
99       const float16x8_t vi2x01234567 = vld1q_f16(i2);
100       const float16x8_t vk2x01234567 = vld1q_f16(w); w += 8;
101       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi2x01234567, vk2x01234567);
102 
103       const float16x8_t vi3x01234567 = vld1q_f16(i3);
104       const float16x8_t vk3x01234567 = vld1q_f16(w); w += 8;
105       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi3x01234567, vk3x01234567);
106 
107 
108       float16x8_t vacc01234567 = vmaxq_f16(vacc01234567p0, vmin);
109       vacc01234567 = vminq_f16(vacc01234567, vmax);
110 
111       float16x4_t vacc0123 = vget_low_f16(vacc01234567);
112       if (c & 4) {
113         vst1_f16(output, vacc0123); output += 4;
114         vacc0123 = vget_high_f16(vacc01234567);
115       }
116       if (c & 2) {
117         vst1_lane_u32(__builtin_assume_aligned(output, 1), vreinterpret_u32_f16(vacc0123), 0); output += 2;
118         vacc0123 = vext_f16(vacc0123, vacc0123, 2);
119       }
120       if (c & 1) {
121         vst1_lane_f16(output, vacc0123, 0); output += 1;
122       }
123     }
124 
125     output = (__fp16*) ((uintptr_t) output + output_increment);
126   } while (--output_width != 0);
127 }
128