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_up32x3__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_up32x3__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 >= 32; c -= 32) {
57 float16x8_t vacc01234567p0 = vld1q_f16(w); w += 8;
58 float16x8_t vacc89ABCDEFp0 = vld1q_f16(w); w += 8;
59 float16x8_t vaccGHIJKLMNp0 = vld1q_f16(w); w += 8;
60 float16x8_t vaccOPQRSTUVp0 = vld1q_f16(w); w += 8;
61
62
63 const float16x8_t vi0x01234567 = vld1q_f16(i0); i0 += 8;
64 const float16x8_t vi0x89ABCDEF = vld1q_f16(i0); i0 += 8;
65 const float16x8_t vi0xGHIJKLMN = vld1q_f16(i0); i0 += 8;
66 const float16x8_t vi0xOPQRSTUV = vld1q_f16(i0); i0 += 8;
67 const float16x8_t vk0x01234567 = vld1q_f16(w); w += 8;
68 const float16x8_t vk0x89ABCDEF = vld1q_f16(w); w += 8;
69 const float16x8_t vk0xGHIJKLMN = vld1q_f16(w); w += 8;
70 const float16x8_t vk0xOPQRSTUV = vld1q_f16(w); w += 8;
71 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi0x01234567, vk0x01234567);
72 vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi0x89ABCDEF, vk0x89ABCDEF);
73 vaccGHIJKLMNp0 = vfmaq_f16(vaccGHIJKLMNp0, vi0xGHIJKLMN, vk0xGHIJKLMN);
74 vaccOPQRSTUVp0 = vfmaq_f16(vaccOPQRSTUVp0, vi0xOPQRSTUV, vk0xOPQRSTUV);
75
76 const float16x8_t vi1x01234567 = vld1q_f16(i1); i1 += 8;
77 const float16x8_t vi1x89ABCDEF = vld1q_f16(i1); i1 += 8;
78 const float16x8_t vi1xGHIJKLMN = vld1q_f16(i1); i1 += 8;
79 const float16x8_t vi1xOPQRSTUV = vld1q_f16(i1); i1 += 8;
80 const float16x8_t vk1x01234567 = vld1q_f16(w); w += 8;
81 const float16x8_t vk1x89ABCDEF = vld1q_f16(w); w += 8;
82 const float16x8_t vk1xGHIJKLMN = vld1q_f16(w); w += 8;
83 const float16x8_t vk1xOPQRSTUV = vld1q_f16(w); w += 8;
84 float16x8_t vacc01234567p1 = vmulq_f16(vi1x01234567, vk1x01234567);
85 float16x8_t vacc89ABCDEFp1 = vmulq_f16(vi1x89ABCDEF, vk1x89ABCDEF);
86 float16x8_t vaccGHIJKLMNp1 = vmulq_f16(vi1xGHIJKLMN, vk1xGHIJKLMN);
87 float16x8_t vaccOPQRSTUVp1 = vmulq_f16(vi1xOPQRSTUV, vk1xOPQRSTUV);
88
89 const float16x8_t vi2x01234567 = vld1q_f16(i2); i2 += 8;
90 const float16x8_t vi2x89ABCDEF = vld1q_f16(i2); i2 += 8;
91 const float16x8_t vi2xGHIJKLMN = vld1q_f16(i2); i2 += 8;
92 const float16x8_t vi2xOPQRSTUV = vld1q_f16(i2); i2 += 8;
93 const float16x8_t vk2x01234567 = vld1q_f16(w); w += 8;
94 const float16x8_t vk2x89ABCDEF = vld1q_f16(w); w += 8;
95 const float16x8_t vk2xGHIJKLMN = vld1q_f16(w); w += 8;
96 const float16x8_t vk2xOPQRSTUV = vld1q_f16(w); w += 8;
97 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi2x01234567, vk2x01234567);
98 vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi2x89ABCDEF, vk2x89ABCDEF);
99 vaccGHIJKLMNp0 = vfmaq_f16(vaccGHIJKLMNp0, vi2xGHIJKLMN, vk2xGHIJKLMN);
100 vaccOPQRSTUVp0 = vfmaq_f16(vaccOPQRSTUVp0, vi2xOPQRSTUV, vk2xOPQRSTUV);
101
102 // Add up all accumulators to vacc0123456789ABCDEFGHIJKLMNOPQRSTUVp0
103 vacc01234567p0 = vaddq_f16(vacc01234567p0, vacc01234567p1);
104 vacc89ABCDEFp0 = vaddq_f16(vacc89ABCDEFp0, vacc89ABCDEFp1);
105 vaccGHIJKLMNp0 = vaddq_f16(vaccGHIJKLMNp0, vaccGHIJKLMNp1);
106 vaccOPQRSTUVp0 = vaddq_f16(vaccOPQRSTUVp0, vaccOPQRSTUVp1);
107
108 float16x8_t vacc01234567 = vmaxq_f16(vacc01234567p0, vmin);
109 float16x8_t vacc89ABCDEF = vmaxq_f16(vacc89ABCDEFp0, vmin);
110 float16x8_t vaccGHIJKLMN = vmaxq_f16(vaccGHIJKLMNp0, vmin);
111 float16x8_t vaccOPQRSTUV = vmaxq_f16(vaccOPQRSTUVp0, vmin);
112 vacc01234567 = vminq_f16(vacc01234567, vmax);
113 vacc89ABCDEF = vminq_f16(vacc89ABCDEF, vmax);
114 vaccGHIJKLMN = vminq_f16(vaccGHIJKLMN, vmax);
115 vaccOPQRSTUV = vminq_f16(vaccOPQRSTUV, vmax);
116
117 vst1q_f16(output, vacc01234567); output += 8;
118 vst1q_f16(output, vacc89ABCDEF); output += 8;
119 vst1q_f16(output, vaccGHIJKLMN); output += 8;
120 vst1q_f16(output, vaccOPQRSTUV); output += 8;
121 }
122 for (; c >= 8; c -= 8) {
123 float16x8_t vacc01234567p0 = vld1q_f16(w); w += 8;
124
125
126 const float16x8_t vi0x01234567 = vld1q_f16(i0); i0 += 8;
127 const float16x8_t vk0x01234567 = vld1q_f16(w + 24);
128 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi0x01234567, vk0x01234567);
129
130 const float16x8_t vi1x01234567 = vld1q_f16(i1); i1 += 8;
131 const float16x8_t vk1x01234567 = vld1q_f16(w + 56);
132 float16x8_t vacc01234567p1 = vmulq_f16(vi1x01234567, vk1x01234567);
133
134 const float16x8_t vi2x01234567 = vld1q_f16(i2); i2 += 8;
135 const float16x8_t vk2x01234567 = vld1q_f16(w + 88);
136 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi2x01234567, vk2x01234567);
137
138 // Add up all accumulators to vacc01234567p0
139 vacc01234567p0 = vaddq_f16(vacc01234567p0, vacc01234567p1);
140
141 float16x8_t vacc01234567 = vmaxq_f16(vacc01234567p0, vmin);
142 vacc01234567 = vminq_f16(vacc01234567, vmax);
143
144 vst1q_f16(output, vacc01234567); output += 8;
145 }
146 if XNN_UNLIKELY(c != 0) {
147 float16x8_t vacc01234567p0 = vld1q_f16(w);
148
149
150 const float16x8_t vi0x01234567 = vld1q_f16(i0);
151 const float16x8_t vk0x01234567 = vld1q_f16(w + 32);
152 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi0x01234567, vk0x01234567);
153
154 const float16x8_t vi1x01234567 = vld1q_f16(i1);
155 const float16x8_t vk1x01234567 = vld1q_f16(w + 64);
156 float16x8_t vacc01234567p1 = vmulq_f16(vi1x01234567, vk1x01234567);
157
158 const float16x8_t vi2x01234567 = vld1q_f16(i2);
159 const float16x8_t vk2x01234567 = vld1q_f16(w + 96);
160 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi2x01234567, vk2x01234567);
161
162 // Add up all accumulators to vacc01234567p0
163 vacc01234567p0 = vaddq_f16(vacc01234567p0, vacc01234567p1);
164
165 float16x8_t vacc01234567 = vmaxq_f16(vacc01234567p0, vmin);
166 vacc01234567 = vminq_f16(vacc01234567, vmax);
167
168 float16x4_t vacc0123 = vget_low_f16(vacc01234567);
169 if (c & 4) {
170 vst1_f16(output, vacc0123); output += 4;
171 vacc0123 = vget_high_f16(vacc01234567);
172 }
173 if (c & 2) {
174 vst1_lane_u32((void*) output, vreinterpret_u32_f16(vacc0123), 0); output += 2;
175 vacc0123 = vext_f16(vacc0123, vacc0123, 2);
176 }
177 if (c & 1) {
178 vst1_lane_f16(output, vacc0123, 0); output += 1;
179 }
180 }
181
182 output = (__fp16*) ((uintptr_t) output + output_increment);
183 } while (--output_width != 0);
184 }
185