xref: /aosp_15_r20/external/XNNPACK/src/f16-dwconv/gen/up32x4-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_up32x4__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_up32x4__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 
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 >= 32; c -= 32) {
62       float16x8_t vacc01234567p0 = vld1q_f16(w); w += 8;
63       float16x8_t vacc89ABCDEFp0 = vld1q_f16(w); w += 8;
64       float16x8_t vaccGHIJKLMNp0 = vld1q_f16(w); w += 8;
65       float16x8_t vaccOPQRSTUVp0 = vld1q_f16(w); w += 8;
66 
67 
68       const float16x8_t vi0x01234567 = vld1q_f16(i0); i0 += 8;
69       const float16x8_t vi0x89ABCDEF = vld1q_f16(i0); i0 += 8;
70       const float16x8_t vi0xGHIJKLMN = vld1q_f16(i0); i0 += 8;
71       const float16x8_t vi0xOPQRSTUV = vld1q_f16(i0); i0 += 8;
72       const float16x8_t vk0x01234567 = vld1q_f16(w); w += 8;
73       const float16x8_t vk0x89ABCDEF = vld1q_f16(w); w += 8;
74       const float16x8_t vk0xGHIJKLMN = vld1q_f16(w); w += 8;
75       const float16x8_t vk0xOPQRSTUV = vld1q_f16(w); w += 8;
76       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi0x01234567, vk0x01234567);
77       vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi0x89ABCDEF, vk0x89ABCDEF);
78       vaccGHIJKLMNp0 = vfmaq_f16(vaccGHIJKLMNp0, vi0xGHIJKLMN, vk0xGHIJKLMN);
79       vaccOPQRSTUVp0 = vfmaq_f16(vaccOPQRSTUVp0, vi0xOPQRSTUV, vk0xOPQRSTUV);
80 
81       const float16x8_t vi1x01234567 = vld1q_f16(i1); i1 += 8;
82       const float16x8_t vi1x89ABCDEF = vld1q_f16(i1); i1 += 8;
83       const float16x8_t vi1xGHIJKLMN = vld1q_f16(i1); i1 += 8;
84       const float16x8_t vi1xOPQRSTUV = vld1q_f16(i1); i1 += 8;
85       const float16x8_t vk1x01234567 = vld1q_f16(w); w += 8;
86       const float16x8_t vk1x89ABCDEF = vld1q_f16(w); w += 8;
87       const float16x8_t vk1xGHIJKLMN = vld1q_f16(w); w += 8;
88       const float16x8_t vk1xOPQRSTUV = vld1q_f16(w); w += 8;
89       float16x8_t vacc01234567p1 = vmulq_f16(vi1x01234567, vk1x01234567);
90       float16x8_t vacc89ABCDEFp1 = vmulq_f16(vi1x89ABCDEF, vk1x89ABCDEF);
91       float16x8_t vaccGHIJKLMNp1 = vmulq_f16(vi1xGHIJKLMN, vk1xGHIJKLMN);
92       float16x8_t vaccOPQRSTUVp1 = vmulq_f16(vi1xOPQRSTUV, vk1xOPQRSTUV);
93 
94       const float16x8_t vi2x01234567 = vld1q_f16(i2); i2 += 8;
95       const float16x8_t vi2x89ABCDEF = vld1q_f16(i2); i2 += 8;
96       const float16x8_t vi2xGHIJKLMN = vld1q_f16(i2); i2 += 8;
97       const float16x8_t vi2xOPQRSTUV = vld1q_f16(i2); i2 += 8;
98       const float16x8_t vk2x01234567 = vld1q_f16(w); w += 8;
99       const float16x8_t vk2x89ABCDEF = vld1q_f16(w); w += 8;
100       const float16x8_t vk2xGHIJKLMN = vld1q_f16(w); w += 8;
101       const float16x8_t vk2xOPQRSTUV = vld1q_f16(w); w += 8;
102       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi2x01234567, vk2x01234567);
103       vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi2x89ABCDEF, vk2x89ABCDEF);
104       vaccGHIJKLMNp0 = vfmaq_f16(vaccGHIJKLMNp0, vi2xGHIJKLMN, vk2xGHIJKLMN);
105       vaccOPQRSTUVp0 = vfmaq_f16(vaccOPQRSTUVp0, vi2xOPQRSTUV, vk2xOPQRSTUV);
106 
107       const float16x8_t vi3x01234567 = vld1q_f16(i3); i3 += 8;
108       const float16x8_t vi3x89ABCDEF = vld1q_f16(i3); i3 += 8;
109       const float16x8_t vi3xGHIJKLMN = vld1q_f16(i3); i3 += 8;
110       const float16x8_t vi3xOPQRSTUV = vld1q_f16(i3); i3 += 8;
111       const float16x8_t vk3x01234567 = vld1q_f16(w); w += 8;
112       const float16x8_t vk3x89ABCDEF = vld1q_f16(w); w += 8;
113       const float16x8_t vk3xGHIJKLMN = vld1q_f16(w); w += 8;
114       const float16x8_t vk3xOPQRSTUV = vld1q_f16(w); w += 8;
115       vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi3x01234567, vk3x01234567);
116       vacc89ABCDEFp1 = vfmaq_f16(vacc89ABCDEFp1, vi3x89ABCDEF, vk3x89ABCDEF);
117       vaccGHIJKLMNp1 = vfmaq_f16(vaccGHIJKLMNp1, vi3xGHIJKLMN, vk3xGHIJKLMN);
118       vaccOPQRSTUVp1 = vfmaq_f16(vaccOPQRSTUVp1, vi3xOPQRSTUV, vk3xOPQRSTUV);
119 
120       // Add up all accumulators to vacc0123456789ABCDEFGHIJKLMNOPQRSTUVp0
121       vacc01234567p0 = vaddq_f16(vacc01234567p0, vacc01234567p1);
122       vacc89ABCDEFp0 = vaddq_f16(vacc89ABCDEFp0, vacc89ABCDEFp1);
123       vaccGHIJKLMNp0 = vaddq_f16(vaccGHIJKLMNp0, vaccGHIJKLMNp1);
124       vaccOPQRSTUVp0 = vaddq_f16(vaccOPQRSTUVp0, vaccOPQRSTUVp1);
125 
126       float16x8_t vacc01234567 = vmaxq_f16(vacc01234567p0, vmin);
127       float16x8_t vacc89ABCDEF = vmaxq_f16(vacc89ABCDEFp0, vmin);
128       float16x8_t vaccGHIJKLMN = vmaxq_f16(vaccGHIJKLMNp0, vmin);
129       float16x8_t vaccOPQRSTUV = vmaxq_f16(vaccOPQRSTUVp0, vmin);
130       vacc01234567 = vminq_f16(vacc01234567, vmax);
131       vacc89ABCDEF = vminq_f16(vacc89ABCDEF, vmax);
132       vaccGHIJKLMN = vminq_f16(vaccGHIJKLMN, vmax);
133       vaccOPQRSTUV = vminq_f16(vaccOPQRSTUV, vmax);
134 
135       vst1q_f16(output, vacc01234567); output += 8;
136       vst1q_f16(output, vacc89ABCDEF); output += 8;
137       vst1q_f16(output, vaccGHIJKLMN); output += 8;
138       vst1q_f16(output, vaccOPQRSTUV); output += 8;
139     }
140     for (; c >= 8; c -= 8) {
141       float16x8_t vacc01234567p0 = vld1q_f16(w); w += 8;
142 
143 
144       const float16x8_t vi0x01234567 = vld1q_f16(i0); i0 += 8;
145       const float16x8_t vk0x01234567 = vld1q_f16(w + 24);
146       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi0x01234567, vk0x01234567);
147 
148       const float16x8_t vi1x01234567 = vld1q_f16(i1); i1 += 8;
149       const float16x8_t vk1x01234567 = vld1q_f16(w + 56);
150       float16x8_t vacc01234567p1 = vmulq_f16(vi1x01234567, vk1x01234567);
151 
152       const float16x8_t vi2x01234567 = vld1q_f16(i2); i2 += 8;
153       const float16x8_t vk2x01234567 = vld1q_f16(w + 88);
154       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi2x01234567, vk2x01234567);
155 
156       const float16x8_t vi3x01234567 = vld1q_f16(i3); i3 += 8;
157       const float16x8_t vk3x01234567 = vld1q_f16(w + 120);
158       vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi3x01234567, vk3x01234567);
159 
160       // Add up all accumulators to vacc01234567p0
161       vacc01234567p0 = vaddq_f16(vacc01234567p0, vacc01234567p1);
162 
163       float16x8_t vacc01234567 = vmaxq_f16(vacc01234567p0, vmin);
164       vacc01234567 = vminq_f16(vacc01234567, vmax);
165 
166       vst1q_f16(output, vacc01234567); output += 8;
167     }
168     if XNN_UNLIKELY(c != 0) {
169       float16x8_t vacc01234567p0 = vld1q_f16(w);
170 
171 
172       const float16x8_t vi0x01234567 = vld1q_f16(i0);
173       const float16x8_t vk0x01234567 = vld1q_f16(w + 32);
174       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi0x01234567, vk0x01234567);
175 
176       const float16x8_t vi1x01234567 = vld1q_f16(i1);
177       const float16x8_t vk1x01234567 = vld1q_f16(w + 64);
178       float16x8_t vacc01234567p1 = vmulq_f16(vi1x01234567, vk1x01234567);
179 
180       const float16x8_t vi2x01234567 = vld1q_f16(i2);
181       const float16x8_t vk2x01234567 = vld1q_f16(w + 96);
182       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi2x01234567, vk2x01234567);
183 
184       const float16x8_t vi3x01234567 = vld1q_f16(i3);
185       const float16x8_t vk3x01234567 = vld1q_f16(w + 128);
186       vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi3x01234567, vk3x01234567);
187 
188       // Add up all accumulators to vacc01234567p0
189       vacc01234567p0 = vaddq_f16(vacc01234567p0, vacc01234567p1);
190 
191       float16x8_t vacc01234567 = vmaxq_f16(vacc01234567p0, vmin);
192       vacc01234567 = vminq_f16(vacc01234567, vmax);
193 
194       float16x4_t vacc0123 = vget_low_f16(vacc01234567);
195       if (c & 4) {
196         vst1_f16(output, vacc0123); output += 4;
197         vacc0123 = vget_high_f16(vacc01234567);
198       }
199       if (c & 2) {
200         vst1_lane_u32((void*) output, vreinterpret_u32_f16(vacc0123), 0); output += 2;
201         vacc0123 = vext_f16(vacc0123, vacc0123, 2);
202       }
203       if (c & 1) {
204         vst1_lane_f16(output, vacc0123, 0); output += 1;
205       }
206     }
207 
208     output = (__fp16*) ((uintptr_t) output + output_increment);
209   } while (--output_width != 0);
210 }
211