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_up8x3__neonfp16arith_acc2(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 union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_f16_dwconv_minmax_ukernel_up8x3__neonfp16arith_acc2(
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 union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
28 {
29 assert(channels != 0);
30 assert(output_width != 0);
31
32 __fp16* output = (__fp16*) output_ptr;
33 const float16x8_t vmax = vreinterpretq_f16_u16(vld1q_dup_u16(¶ms->neon.max));
34 const float16x8_t vmin = vreinterpretq_f16_u16(vld1q_dup_u16(¶ms->neon.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
52 input = (const void**) ((uintptr_t) input + input_stride);
53
54 size_t c = channels;
55 const __fp16* w = (const __fp16*) weights;
56 for (; c >= 8; c -= 8) {
57 float16x8_t vacc01234567p0 = vld1q_f16(w); w += 8;
58
59
60 const float16x8_t vi0x01234567 = vld1q_f16(i0); i0 += 8;
61 const float16x8_t vk0x01234567 = vld1q_f16(w); w += 8;
62 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi0x01234567, vk0x01234567);
63
64 const float16x8_t vi1x01234567 = vld1q_f16(i1); i1 += 8;
65 const float16x8_t vk1x01234567 = vld1q_f16(w); w += 8;
66 float16x8_t vacc01234567p1 = vmulq_f16(vi1x01234567, vk1x01234567);
67
68 const float16x8_t vi2x01234567 = vld1q_f16(i2); i2 += 8;
69 const float16x8_t vk2x01234567 = vld1q_f16(w); w += 8;
70 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi2x01234567, vk2x01234567);
71
72 // Add up all accumulators to vacc01234567p0
73 vacc01234567p0 = vaddq_f16(vacc01234567p0, vacc01234567p1);
74
75 float16x8_t vacc01234567 = vmaxq_f16(vacc01234567p0, vmin);
76 vacc01234567 = vminq_f16(vacc01234567, vmax);
77
78 vst1q_f16(output, vacc01234567); output += 8;
79 }
80 if XNN_UNLIKELY(c != 0) {
81 float16x8_t vacc01234567p0 = vld1q_f16(w); w += 8;
82
83
84 const float16x8_t vi0x01234567 = vld1q_f16(i0);
85 const float16x8_t vk0x01234567 = vld1q_f16(w); w += 8;
86 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi0x01234567, vk0x01234567);
87
88 const float16x8_t vi1x01234567 = vld1q_f16(i1);
89 const float16x8_t vk1x01234567 = vld1q_f16(w); w += 8;
90 float16x8_t vacc01234567p1 = vmulq_f16(vi1x01234567, vk1x01234567);
91
92 const float16x8_t vi2x01234567 = vld1q_f16(i2);
93 const float16x8_t vk2x01234567 = vld1q_f16(w); w += 8;
94 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi2x01234567, vk2x01234567);
95
96 // Add up all accumulators to vacc01234567p0
97 vacc01234567p0 = vaddq_f16(vacc01234567p0, vacc01234567p1);
98
99 float16x8_t vacc01234567 = vmaxq_f16(vacc01234567p0, vmin);
100 vacc01234567 = vminq_f16(vacc01234567, vmax);
101
102 float16x4_t vacc0123 = vget_low_f16(vacc01234567);
103 if (c & 4) {
104 vst1_f16(output, vacc0123); output += 4;
105 vacc0123 = vget_high_f16(vacc01234567);
106 }
107 if (c & 2) {
108 vst1_lane_u32((void*) output, vreinterpret_u32_f16(vacc0123), 0); output += 2;
109 vacc0123 = vext_f16(vacc0123, vacc0123, 2);
110 }
111 if (c & 1) {
112 vst1_lane_f16(output, vacc0123, 0); output += 1;
113 }
114 }
115
116 output = (__fp16*) ((uintptr_t) output + output_increment);
117 } while (--output_width != 0);
118 }
119