1 // Auto-generated file. Do not edit!
2 // Template: src/f32-dwconv/up-neon.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2019 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_f32_dwconv_minmax_ukernel_up8x9__neon_acc2(size_t channels,size_t output_width,const float ** input,const float * weights,float * output,size_t input_stride,size_t output_increment,size_t input_offset,const float * zero,const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2(
18 size_t channels,
19 size_t output_width,
20 const float** input,
21 const float* weights,
22 float* output,
23 size_t input_stride,
24 size_t output_increment,
25 size_t input_offset,
26 const float* zero,
27 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
28 {
29 assert(channels != 0);
30 assert(output_width != 0);
31
32 const float32x4_t vmax = vld1q_dup_f32(¶ms->scalar.max);
33 const float32x4_t vmin = vld1q_dup_f32(¶ms->scalar.min);
34 do {
35 const float* i0 = input[0];
36 assert(i0 != NULL);
37 if XNN_UNPREDICTABLE(i0 != zero) {
38 i0 = (const float*) ((uintptr_t) i0 + input_offset);
39 }
40 const float* i1 = input[1];
41 assert(i1 != NULL);
42 if XNN_UNPREDICTABLE(i1 != zero) {
43 i1 = (const float*) ((uintptr_t) i1 + input_offset);
44 }
45 const float* i2 = input[2];
46 assert(i2 != NULL);
47 if XNN_UNPREDICTABLE(i2 != zero) {
48 i2 = (const float*) ((uintptr_t) i2 + input_offset);
49 }
50 const float* i3 = input[3];
51 assert(i3 != NULL);
52 if XNN_UNPREDICTABLE(i3 != zero) {
53 i3 = (const float*) ((uintptr_t) i3 + input_offset);
54 }
55 const float* i4 = input[4];
56 assert(i4 != NULL);
57 if XNN_UNPREDICTABLE(i4 != zero) {
58 i4 = (const float*) ((uintptr_t) i4 + input_offset);
59 }
60 const float* i5 = input[5];
61 assert(i5 != NULL);
62 if XNN_UNPREDICTABLE(i5 != zero) {
63 i5 = (const float*) ((uintptr_t) i5 + input_offset);
64 }
65 const float* i6 = input[6];
66 assert(i6 != NULL);
67 if XNN_UNPREDICTABLE(i6 != zero) {
68 i6 = (const float*) ((uintptr_t) i6 + input_offset);
69 }
70 const float* i7 = input[7];
71 assert(i7 != NULL);
72 if XNN_UNPREDICTABLE(i7 != zero) {
73 i7 = (const float*) ((uintptr_t) i7 + input_offset);
74 }
75 const float* i8 = input[8];
76 assert(i8 != NULL);
77 if XNN_UNPREDICTABLE(i8 != zero) {
78 i8 = (const float*) ((uintptr_t) i8 + input_offset);
79 }
80
81 input = (const float**) ((uintptr_t) input + input_stride);
82
83 size_t c = channels;
84 const float* w = weights;
85 for (; c >= 8; c -= 8) {
86 float32x4_t vacc0123p0 = vld1q_f32(w); w += 4;
87 float32x4_t vacc4567p0 = vld1q_f32(w); w += 4;
88
89
90 const float32x4_t vi0x0123 = vld1q_f32(i0); i0 += 4;
91 const float32x4_t vi0x4567 = vld1q_f32(i0); i0 += 4;
92 const float32x4_t vk0x0123 = vld1q_f32(w); w += 4;
93 const float32x4_t vk0x4567 = vld1q_f32(w); w += 4;
94 vacc0123p0 = vmlaq_f32(vacc0123p0, vi0x0123, vk0x0123);
95 vacc4567p0 = vmlaq_f32(vacc4567p0, vi0x4567, vk0x4567);
96
97 const float32x4_t vi1x0123 = vld1q_f32(i1); i1 += 4;
98 const float32x4_t vi1x4567 = vld1q_f32(i1); i1 += 4;
99 const float32x4_t vk1x0123 = vld1q_f32(w); w += 4;
100 const float32x4_t vk1x4567 = vld1q_f32(w); w += 4;
101 float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123);
102 float32x4_t vacc4567p1 = vmulq_f32(vi1x4567, vk1x4567);
103
104 const float32x4_t vi2x0123 = vld1q_f32(i2); i2 += 4;
105 const float32x4_t vi2x4567 = vld1q_f32(i2); i2 += 4;
106 const float32x4_t vk2x0123 = vld1q_f32(w); w += 4;
107 const float32x4_t vk2x4567 = vld1q_f32(w); w += 4;
108 vacc0123p0 = vmlaq_f32(vacc0123p0, vi2x0123, vk2x0123);
109 vacc4567p0 = vmlaq_f32(vacc4567p0, vi2x4567, vk2x4567);
110
111 const float32x4_t vi3x0123 = vld1q_f32(i3); i3 += 4;
112 const float32x4_t vi3x4567 = vld1q_f32(i3); i3 += 4;
113 const float32x4_t vk3x0123 = vld1q_f32(w); w += 4;
114 const float32x4_t vk3x4567 = vld1q_f32(w); w += 4;
115 vacc0123p1 = vmlaq_f32(vacc0123p1, vi3x0123, vk3x0123);
116 vacc4567p1 = vmlaq_f32(vacc4567p1, vi3x4567, vk3x4567);
117
118 const float32x4_t vi4x0123 = vld1q_f32(i4); i4 += 4;
119 const float32x4_t vi4x4567 = vld1q_f32(i4); i4 += 4;
120 const float32x4_t vk4x0123 = vld1q_f32(w); w += 4;
121 const float32x4_t vk4x4567 = vld1q_f32(w); w += 4;
122 vacc0123p0 = vmlaq_f32(vacc0123p0, vi4x0123, vk4x0123);
123 vacc4567p0 = vmlaq_f32(vacc4567p0, vi4x4567, vk4x4567);
124
125 const float32x4_t vi5x0123 = vld1q_f32(i5); i5 += 4;
126 const float32x4_t vi5x4567 = vld1q_f32(i5); i5 += 4;
127 const float32x4_t vk5x0123 = vld1q_f32(w); w += 4;
128 const float32x4_t vk5x4567 = vld1q_f32(w); w += 4;
129 vacc0123p1 = vmlaq_f32(vacc0123p1, vi5x0123, vk5x0123);
130 vacc4567p1 = vmlaq_f32(vacc4567p1, vi5x4567, vk5x4567);
131
132 const float32x4_t vi6x0123 = vld1q_f32(i6); i6 += 4;
133 const float32x4_t vi6x4567 = vld1q_f32(i6); i6 += 4;
134 const float32x4_t vk6x0123 = vld1q_f32(w); w += 4;
135 const float32x4_t vk6x4567 = vld1q_f32(w); w += 4;
136 vacc0123p0 = vmlaq_f32(vacc0123p0, vi6x0123, vk6x0123);
137 vacc4567p0 = vmlaq_f32(vacc4567p0, vi6x4567, vk6x4567);
138
139 const float32x4_t vi7x0123 = vld1q_f32(i7); i7 += 4;
140 const float32x4_t vi7x4567 = vld1q_f32(i7); i7 += 4;
141 const float32x4_t vk7x0123 = vld1q_f32(w); w += 4;
142 const float32x4_t vk7x4567 = vld1q_f32(w); w += 4;
143 vacc0123p1 = vmlaq_f32(vacc0123p1, vi7x0123, vk7x0123);
144 vacc4567p1 = vmlaq_f32(vacc4567p1, vi7x4567, vk7x4567);
145
146 const float32x4_t vi8x0123 = vld1q_f32(i8); i8 += 4;
147 const float32x4_t vi8x4567 = vld1q_f32(i8); i8 += 4;
148 const float32x4_t vk8x0123 = vld1q_f32(w); w += 4;
149 const float32x4_t vk8x4567 = vld1q_f32(w); w += 4;
150 vacc0123p0 = vmlaq_f32(vacc0123p0, vi8x0123, vk8x0123);
151 vacc4567p0 = vmlaq_f32(vacc4567p0, vi8x4567, vk8x4567);
152
153 // Add up all accumulators to vacc01234567p0
154 vacc0123p0 = vaddq_f32(vacc0123p0, vacc0123p1);
155 vacc4567p0 = vaddq_f32(vacc4567p0, vacc4567p1);
156
157 float32x4_t vacc0123 = vmaxq_f32(vacc0123p0, vmin);
158 float32x4_t vacc4567 = vmaxq_f32(vacc4567p0, vmin);
159 vacc0123 = vminq_f32(vacc0123, vmax);
160 vacc4567 = vminq_f32(vacc4567, vmax);
161
162 vst1q_f32(output, vacc0123); output += 4;
163 vst1q_f32(output, vacc4567); output += 4;
164 }
165 for (; c >= 4; c -= 4) {
166 float32x4_t vacc0123p0 = vld1q_f32(w); w += 4;
167
168
169 const float32x4_t vi0x0123 = vld1q_f32(i0); i0 += 4;
170 const float32x4_t vk0x0123 = vld1q_f32(w + 4);
171 vacc0123p0 = vmlaq_f32(vacc0123p0, vi0x0123, vk0x0123);
172
173 const float32x4_t vi1x0123 = vld1q_f32(i1); i1 += 4;
174 const float32x4_t vk1x0123 = vld1q_f32(w + 12);
175 float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123);
176
177 const float32x4_t vi2x0123 = vld1q_f32(i2); i2 += 4;
178 const float32x4_t vk2x0123 = vld1q_f32(w + 20);
179 vacc0123p0 = vmlaq_f32(vacc0123p0, vi2x0123, vk2x0123);
180
181 const float32x4_t vi3x0123 = vld1q_f32(i3); i3 += 4;
182 const float32x4_t vk3x0123 = vld1q_f32(w + 28);
183 vacc0123p1 = vmlaq_f32(vacc0123p1, vi3x0123, vk3x0123);
184
185 const float32x4_t vi4x0123 = vld1q_f32(i4); i4 += 4;
186 const float32x4_t vk4x0123 = vld1q_f32(w + 36);
187 vacc0123p0 = vmlaq_f32(vacc0123p0, vi4x0123, vk4x0123);
188
189 const float32x4_t vi5x0123 = vld1q_f32(i5); i5 += 4;
190 const float32x4_t vk5x0123 = vld1q_f32(w + 44);
191 vacc0123p1 = vmlaq_f32(vacc0123p1, vi5x0123, vk5x0123);
192
193 const float32x4_t vi6x0123 = vld1q_f32(i6); i6 += 4;
194 const float32x4_t vk6x0123 = vld1q_f32(w + 52);
195 vacc0123p0 = vmlaq_f32(vacc0123p0, vi6x0123, vk6x0123);
196
197 const float32x4_t vi7x0123 = vld1q_f32(i7); i7 += 4;
198 const float32x4_t vk7x0123 = vld1q_f32(w + 60);
199 vacc0123p1 = vmlaq_f32(vacc0123p1, vi7x0123, vk7x0123);
200
201 const float32x4_t vi8x0123 = vld1q_f32(i8); i8 += 4;
202 const float32x4_t vk8x0123 = vld1q_f32(w + 68);
203 vacc0123p0 = vmlaq_f32(vacc0123p0, vi8x0123, vk8x0123);
204
205 // Add up all accumulators to vacc0123p0
206 vacc0123p0 = vaddq_f32(vacc0123p0, vacc0123p1);
207
208 float32x4_t vacc0123 = vmaxq_f32(vacc0123p0, vmin);
209 vacc0123 = vminq_f32(vacc0123, vmax);
210
211 vst1q_f32(output, vacc0123); output += 4;
212 }
213 if XNN_UNLIKELY(c != 0) {
214 float32x4_t vacc0123p0 = vld1q_f32(w);
215
216
217 const float32x4_t vi0x0123 = vld1q_f32(i0);
218 const float32x4_t vk0x0123 = vld1q_f32(w + 8);
219 vacc0123p0 = vmlaq_f32(vacc0123p0, vi0x0123, vk0x0123);
220
221 const float32x4_t vi1x0123 = vld1q_f32(i1);
222 const float32x4_t vk1x0123 = vld1q_f32(w + 16);
223 float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123);
224
225 const float32x4_t vi2x0123 = vld1q_f32(i2);
226 const float32x4_t vk2x0123 = vld1q_f32(w + 24);
227 vacc0123p0 = vmlaq_f32(vacc0123p0, vi2x0123, vk2x0123);
228
229 const float32x4_t vi3x0123 = vld1q_f32(i3);
230 const float32x4_t vk3x0123 = vld1q_f32(w + 32);
231 vacc0123p1 = vmlaq_f32(vacc0123p1, vi3x0123, vk3x0123);
232
233 const float32x4_t vi4x0123 = vld1q_f32(i4);
234 const float32x4_t vk4x0123 = vld1q_f32(w + 40);
235 vacc0123p0 = vmlaq_f32(vacc0123p0, vi4x0123, vk4x0123);
236
237 const float32x4_t vi5x0123 = vld1q_f32(i5);
238 const float32x4_t vk5x0123 = vld1q_f32(w + 48);
239 vacc0123p1 = vmlaq_f32(vacc0123p1, vi5x0123, vk5x0123);
240
241 const float32x4_t vi6x0123 = vld1q_f32(i6);
242 const float32x4_t vk6x0123 = vld1q_f32(w + 56);
243 vacc0123p0 = vmlaq_f32(vacc0123p0, vi6x0123, vk6x0123);
244
245 const float32x4_t vi7x0123 = vld1q_f32(i7);
246 const float32x4_t vk7x0123 = vld1q_f32(w + 64);
247 vacc0123p1 = vmlaq_f32(vacc0123p1, vi7x0123, vk7x0123);
248
249 const float32x4_t vi8x0123 = vld1q_f32(i8);
250 const float32x4_t vk8x0123 = vld1q_f32(w + 72);
251 vacc0123p0 = vmlaq_f32(vacc0123p0, vi8x0123, vk8x0123);
252
253 // Add up all accumulators to vacc0123p0
254 vacc0123p0 = vaddq_f32(vacc0123p0, vacc0123p1);
255
256 float32x4_t vacc0123 = vmaxq_f32(vacc0123p0, vmin);
257 vacc0123 = vminq_f32(vacc0123, vmax);
258
259 float32x2_t vacc01 = vget_low_f32(vacc0123);
260 if (c & 2) {
261 vst1_f32(output, vacc01); output += 2;
262 vacc01 = vget_high_f32(vacc0123);
263 }
264 if (c & 1) {
265 vst1_lane_f32(output, vacc01, 0); output += 1;
266 }
267 }
268
269 output = (float*) ((uintptr_t) output + output_increment);
270 } while (--output_width != 0);
271 }
272