xref: /aosp_15_r20/external/XNNPACK/src/f16-dwconv/gen/up8x9-minmax-neonfp16arith-acc2.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
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_up8x9__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_up8x9__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(&params->neon.max));
34   const float16x8_t vmin = vreinterpretq_f16_u16(vld1q_dup_u16(&params->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     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     const __fp16* i4 = (const __fp16*) input[4];
57     assert(i4 != NULL);
58     if XNN_UNPREDICTABLE(i4 != (const __fp16*) zero) {
59       i4 = (const __fp16*) ((uintptr_t) i4 + input_offset);
60     }
61     const __fp16* i5 = (const __fp16*) input[5];
62     assert(i5 != NULL);
63     if XNN_UNPREDICTABLE(i5 != (const __fp16*) zero) {
64       i5 = (const __fp16*) ((uintptr_t) i5 + input_offset);
65     }
66     const __fp16* i6 = (const __fp16*) input[6];
67     assert(i6 != NULL);
68     if XNN_UNPREDICTABLE(i6 != (const __fp16*) zero) {
69       i6 = (const __fp16*) ((uintptr_t) i6 + input_offset);
70     }
71     const __fp16* i7 = (const __fp16*) input[7];
72     assert(i7 != NULL);
73     if XNN_UNPREDICTABLE(i7 != (const __fp16*) zero) {
74       i7 = (const __fp16*) ((uintptr_t) i7 + input_offset);
75     }
76     const __fp16* i8 = (const __fp16*) input[8];
77     assert(i8 != NULL);
78     if XNN_UNPREDICTABLE(i8 != (const __fp16*) zero) {
79       i8 = (const __fp16*) ((uintptr_t) i8 + input_offset);
80     }
81 
82     input = (const void**) ((uintptr_t) input + input_stride);
83 
84     size_t c = channels;
85     const __fp16* w = (const __fp16*) weights;
86     for (; c >= 8; c -= 8) {
87       float16x8_t vacc01234567p0 = vld1q_f16(w); w += 8;
88 
89 
90       const float16x8_t vi0x01234567 = vld1q_f16(i0); i0 += 8;
91       const float16x8_t vk0x01234567 = vld1q_f16(w); w += 8;
92       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi0x01234567, vk0x01234567);
93 
94       const float16x8_t vi1x01234567 = vld1q_f16(i1); i1 += 8;
95       const float16x8_t vk1x01234567 = vld1q_f16(w); w += 8;
96       float16x8_t vacc01234567p1 = vmulq_f16(vi1x01234567, vk1x01234567);
97 
98       const float16x8_t vi2x01234567 = vld1q_f16(i2); i2 += 8;
99       const float16x8_t vk2x01234567 = vld1q_f16(w); w += 8;
100       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi2x01234567, vk2x01234567);
101 
102       const float16x8_t vi3x01234567 = vld1q_f16(i3); i3 += 8;
103       const float16x8_t vk3x01234567 = vld1q_f16(w); w += 8;
104       vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi3x01234567, vk3x01234567);
105 
106       const float16x8_t vi4x01234567 = vld1q_f16(i4); i4 += 8;
107       const float16x8_t vk4x01234567 = vld1q_f16(w); w += 8;
108       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi4x01234567, vk4x01234567);
109 
110       const float16x8_t vi5x01234567 = vld1q_f16(i5); i5 += 8;
111       const float16x8_t vk5x01234567 = vld1q_f16(w); w += 8;
112       vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi5x01234567, vk5x01234567);
113 
114       const float16x8_t vi6x01234567 = vld1q_f16(i6); i6 += 8;
115       const float16x8_t vk6x01234567 = vld1q_f16(w); w += 8;
116       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi6x01234567, vk6x01234567);
117 
118       const float16x8_t vi7x01234567 = vld1q_f16(i7); i7 += 8;
119       const float16x8_t vk7x01234567 = vld1q_f16(w); w += 8;
120       vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi7x01234567, vk7x01234567);
121 
122       const float16x8_t vi8x01234567 = vld1q_f16(i8); i8 += 8;
123       const float16x8_t vk8x01234567 = vld1q_f16(w); w += 8;
124       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi8x01234567, vk8x01234567);
125 
126       // Add up all accumulators to vacc01234567p0
127       vacc01234567p0 = vaddq_f16(vacc01234567p0, vacc01234567p1);
128 
129       float16x8_t vacc01234567 = vmaxq_f16(vacc01234567p0, vmin);
130       vacc01234567 = vminq_f16(vacc01234567, vmax);
131 
132       vst1q_f16(output, vacc01234567); output += 8;
133     }
134     if XNN_UNLIKELY(c != 0) {
135       float16x8_t vacc01234567p0 = vld1q_f16(w); w += 8;
136 
137 
138       const float16x8_t vi0x01234567 = vld1q_f16(i0);
139       const float16x8_t vk0x01234567 = vld1q_f16(w); w += 8;
140       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi0x01234567, vk0x01234567);
141 
142       const float16x8_t vi1x01234567 = vld1q_f16(i1);
143       const float16x8_t vk1x01234567 = vld1q_f16(w); w += 8;
144       float16x8_t vacc01234567p1 = vmulq_f16(vi1x01234567, vk1x01234567);
145 
146       const float16x8_t vi2x01234567 = vld1q_f16(i2);
147       const float16x8_t vk2x01234567 = vld1q_f16(w); w += 8;
148       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi2x01234567, vk2x01234567);
149 
150       const float16x8_t vi3x01234567 = vld1q_f16(i3);
151       const float16x8_t vk3x01234567 = vld1q_f16(w); w += 8;
152       vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi3x01234567, vk3x01234567);
153 
154       const float16x8_t vi4x01234567 = vld1q_f16(i4);
155       const float16x8_t vk4x01234567 = vld1q_f16(w); w += 8;
156       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi4x01234567, vk4x01234567);
157 
158       const float16x8_t vi5x01234567 = vld1q_f16(i5);
159       const float16x8_t vk5x01234567 = vld1q_f16(w); w += 8;
160       vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi5x01234567, vk5x01234567);
161 
162       const float16x8_t vi6x01234567 = vld1q_f16(i6);
163       const float16x8_t vk6x01234567 = vld1q_f16(w); w += 8;
164       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi6x01234567, vk6x01234567);
165 
166       const float16x8_t vi7x01234567 = vld1q_f16(i7);
167       const float16x8_t vk7x01234567 = vld1q_f16(w); w += 8;
168       vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi7x01234567, vk7x01234567);
169 
170       const float16x8_t vi8x01234567 = vld1q_f16(i8);
171       const float16x8_t vk8x01234567 = vld1q_f16(w); w += 8;
172       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi8x01234567, vk8x01234567);
173 
174       // Add up all accumulators to vacc01234567p0
175       vacc01234567p0 = vaddq_f16(vacc01234567p0, vacc01234567p1);
176 
177       float16x8_t vacc01234567 = vmaxq_f16(vacc01234567p0, vmin);
178       vacc01234567 = vminq_f16(vacc01234567, vmax);
179 
180       float16x4_t vacc0123 = vget_low_f16(vacc01234567);
181       if (c & 4) {
182         vst1_f16(output, vacc0123); output += 4;
183         vacc0123 = vget_high_f16(vacc01234567);
184       }
185       if (c & 2) {
186         vst1_lane_u32((void*) output, vreinterpret_u32_f16(vacc0123), 0); output += 2;
187         vacc0123 = vext_f16(vacc0123, vacc0123, 2);
188       }
189       if (c & 1) {
190         vst1_lane_f16(output, vacc0123, 0); output += 1;
191       }
192     }
193 
194     output = (__fp16*) ((uintptr_t) output + output_increment);
195   } while (--output_width != 0);
196 }
197