1 // Auto-generated file. Do not edit!
2 // Template: src/f32-dwconv2d-chw/5x5p2-sse.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 <xmmintrin.h>
13
14 #include <xnnpack/dwconv.h>
15 #include <xnnpack/math.h>
16
17
xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4(size_t input_height,size_t input_width,const float * input,const float * weights,const float * zero,float * output,uint32_t padding_top,const union xnn_f32_chw_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4(
19 size_t input_height,
20 size_t input_width,
21 const float* input,
22 const float* weights,
23 const float* zero,
24 float* output,
25 uint32_t padding_top,
26 const union xnn_f32_chw_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
27 {
28 assert(input_height != 0);
29 assert(input_width != 0);
30 assert(input_width % sizeof(float) == 0);
31 assert(padding_top == 2);
32
33 const __m128 vmask = _mm_load_ps((const float*) params->sse.mask);
34 const __m128 vmax = _mm_load_ps(params->sse.max);
35 const __m128 vmin = _mm_load_ps(params->sse.min);
36
37 const __m128 vbias = _mm_load1_ps(weights);
38 const __m128 vk00 = _mm_load1_ps(weights + 1);
39 const __m128 vk01 = _mm_load1_ps(weights + 2);
40 const __m128 vk02 = _mm_load1_ps(weights + 3);
41 const __m128 vk03 = _mm_load1_ps(weights + 4);
42 const __m128 vk04 = _mm_load1_ps(weights + 5);
43 const __m128 vk10 = _mm_load1_ps(weights + 6);
44 const __m128 vk11 = _mm_load1_ps(weights + 7);
45 const __m128 vk12 = _mm_load1_ps(weights + 8);
46 const __m128 vk13 = _mm_load1_ps(weights + 9);
47 const __m128 vk14 = _mm_load1_ps(weights + 10);
48 const __m128 vk20 = _mm_load1_ps(weights + 11);
49 const __m128 vk21 = _mm_load1_ps(weights + 12);
50 const __m128 vk22 = _mm_load1_ps(weights + 13);
51 const __m128 vk23 = _mm_load1_ps(weights + 14);
52 const __m128 vk24 = _mm_load1_ps(weights + 15);
53 const __m128 vk30 = _mm_load1_ps(weights + 16);
54 const __m128 vk31 = _mm_load1_ps(weights + 17);
55 const __m128 vk32 = _mm_load1_ps(weights + 18);
56 const __m128 vk33 = _mm_load1_ps(weights + 19);
57 const __m128 vk34 = _mm_load1_ps(weights + 20);
58 const __m128 vk40 = _mm_load1_ps(weights + 21);
59 const __m128 vk41 = _mm_load1_ps(weights + 22);
60 const __m128 vk42 = _mm_load1_ps(weights + 23);
61 const __m128 vk43 = _mm_load1_ps(weights + 24);
62 const __m128 vk44 = _mm_load1_ps(weights + 25);
63
64 const size_t input_decrement = round_up_po2(input_width, 4 * sizeof(float));
65
66 const float* i0 = zero;
67 const float* i1 = zero;
68 const float* i2 = input;
69 const float* i3 = (const float*) ((uintptr_t) i2 + input_width);
70 const float* i4 = (const float*) ((uintptr_t) i3 + input_width);
71 const float* i5 = (const float*) ((uintptr_t) i4 + input_width);
72 const float* i6 = (const float*) ((uintptr_t) i5 + input_width);
73 const float* i7 = (const float*) ((uintptr_t) i6 + input_width);
74
75 float* o0 = output;
76 float* o1 = (float*) ((uintptr_t) o0 + input_width);
77 float* o2 = (float*) ((uintptr_t) o1 + input_width);
78 float* o3 = (float*) ((uintptr_t) o2 + input_width);
79
80 size_t output_height = input_height;
81 do {
82 if XNN_UNPREDICTABLE(output_height < 2) {
83 i3 = zero;
84 o1 = o0;
85 }
86 if XNN_UNPREDICTABLE(output_height < 3) {
87 i4 = zero;
88 o2 = o1;
89 }
90 if XNN_UNPREDICTABLE(output_height < 4) {
91 i5 = zero;
92 o3 = o2;
93 }
94 if XNN_UNPREDICTABLE(output_height < 5) {
95 i6 = zero;
96 }
97 if XNN_UNPREDICTABLE(output_height < 6) {
98 i7 = zero;
99 }
100
101 __m128 vi0x3012 = _mm_setzero_ps();
102 __m128 vi1x3012 = _mm_setzero_ps();
103 __m128 vi2x3012 = _mm_setzero_ps();
104 __m128 vi3x3012 = _mm_setzero_ps();
105 __m128 vi4x3012 = _mm_setzero_ps();
106 __m128 vi5x3012 = _mm_setzero_ps();
107 __m128 vi6x3012 = _mm_setzero_ps();
108 __m128 vi7x3012 = _mm_setzero_ps();
109
110 __m128 vi0x4567 = _mm_loadu_ps(i0);
111 i0 += 4;
112 __m128 vi1x4567 = _mm_loadu_ps(i1);
113 i1 += 4;
114 __m128 vi2x4567 = _mm_loadu_ps(i2);
115 i2 += 4;
116 __m128 vi3x4567 = _mm_loadu_ps(i3);
117 i3 += 4;
118 __m128 vi4x4567 = _mm_loadu_ps(i4);
119 i4 += 4;
120 __m128 vi5x4567 = _mm_loadu_ps(i5);
121 i5 += 4;
122 __m128 vi6x4567 = _mm_loadu_ps(i6);
123 i6 += 4;
124 __m128 vi7x4567 = _mm_loadu_ps(i7);
125 i7 += 4;
126
127 size_t w = input_width;
128 for (; w > 8 * sizeof(float); w -= 4 * sizeof(float)) {
129 __m128 vo0p0 = _mm_add_ps(vbias, _mm_mul_ps(vi0x4567, vk02));
130 __m128 vo1p0 = _mm_add_ps(vbias, _mm_mul_ps(vi1x4567, vk02));
131 __m128 vo2p0 = _mm_add_ps(vbias, _mm_mul_ps(vi2x4567, vk02));
132 __m128 vo3p0 = _mm_add_ps(vbias, _mm_mul_ps(vi3x4567, vk02));
133 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x4567, vk12));
134 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x4567, vk12));
135 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi3x4567, vk12));
136 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi4x4567, vk12));
137 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x4567, vk22));
138 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi3x4567, vk22));
139 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi4x4567, vk22));
140 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi5x4567, vk22));
141 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi3x4567, vk32));
142 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi4x4567, vk32));
143 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi5x4567, vk32));
144 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi6x4567, vk32));
145 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi4x4567, vk42));
146 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi5x4567, vk42));
147 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi6x4567, vk42));
148 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi7x4567, vk42));
149
150 const __m128 vi0x7456 = _mm_shuffle_ps(vi0x4567, vi0x4567, _MM_SHUFFLE(2, 1, 0, 3));
151 const __m128 vi1x7456 = _mm_shuffle_ps(vi1x4567, vi1x4567, _MM_SHUFFLE(2, 1, 0, 3));
152 const __m128 vi2x7456 = _mm_shuffle_ps(vi2x4567, vi2x4567, _MM_SHUFFLE(2, 1, 0, 3));
153 const __m128 vi3x7456 = _mm_shuffle_ps(vi3x4567, vi3x4567, _MM_SHUFFLE(2, 1, 0, 3));
154 const __m128 vi4x7456 = _mm_shuffle_ps(vi4x4567, vi4x4567, _MM_SHUFFLE(2, 1, 0, 3));
155 const __m128 vi5x7456 = _mm_shuffle_ps(vi5x4567, vi5x4567, _MM_SHUFFLE(2, 1, 0, 3));
156 const __m128 vi6x7456 = _mm_shuffle_ps(vi6x4567, vi6x4567, _MM_SHUFFLE(2, 1, 0, 3));
157 const __m128 vi7x7456 = _mm_shuffle_ps(vi7x4567, vi7x4567, _MM_SHUFFLE(2, 1, 0, 3));
158
159 const __m128 vi0x89AB = _mm_loadu_ps(i0);
160 i0 += 4;
161 const __m128 vi1x89AB = _mm_loadu_ps(i1);
162 i1 += 4;
163 const __m128 vi2x89AB = _mm_loadu_ps(i2);
164 i2 += 4;
165 const __m128 vi3x89AB = _mm_loadu_ps(i3);
166 i3 += 4;
167 const __m128 vi4x89AB = _mm_loadu_ps(i4);
168 i4 += 4;
169 const __m128 vi5x89AB = _mm_loadu_ps(i5);
170 i5 += 4;
171 const __m128 vi6x89AB = _mm_loadu_ps(i6);
172 i6 += 4;
173 const __m128 vi7x89AB = _mm_loadu_ps(i7);
174 i7 += 4;
175
176 const __m128 vi0x3456 = _mm_move_ss(vi0x7456, vi0x3012);
177 const __m128 vi1x3456 = _mm_move_ss(vi1x7456, vi1x3012);
178 const __m128 vi2x3456 = _mm_move_ss(vi2x7456, vi2x3012);
179 const __m128 vi3x3456 = _mm_move_ss(vi3x7456, vi3x3012);
180 const __m128 vi4x3456 = _mm_move_ss(vi4x7456, vi4x3012);
181 const __m128 vi5x3456 = _mm_move_ss(vi5x7456, vi5x3012);
182 const __m128 vi6x3456 = _mm_move_ss(vi6x7456, vi6x3012);
183 const __m128 vi7x3456 = _mm_move_ss(vi7x7456, vi7x3012);
184
185 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi0x3456, vk01));
186 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi1x3456, vk01));
187 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi2x3456, vk01));
188 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi3x3456, vk01));
189 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x3456, vk11));
190 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x3456, vk11));
191 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi3x3456, vk11));
192 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi4x3456, vk11));
193 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x3456, vk21));
194 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi3x3456, vk21));
195 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi4x3456, vk21));
196 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi5x3456, vk21));
197 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi3x3456, vk31));
198 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi4x3456, vk31));
199 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi5x3456, vk31));
200 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi6x3456, vk31));
201 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi4x3456, vk41));
202 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi5x3456, vk41));
203 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi6x3456, vk41));
204 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi7x3456, vk41));
205
206 const __m128 vi0x2345 = _mm_shuffle_ps(vi0x3012, vi0x7456, _MM_SHUFFLE(2, 1, 0, 3));
207 vi0x3012 = vi0x7456;
208 const __m128 vi1x2345 = _mm_shuffle_ps(vi1x3012, vi1x7456, _MM_SHUFFLE(2, 1, 0, 3));
209 vi1x3012 = vi1x7456;
210 const __m128 vi2x2345 = _mm_shuffle_ps(vi2x3012, vi2x7456, _MM_SHUFFLE(2, 1, 0, 3));
211 vi2x3012 = vi2x7456;
212 const __m128 vi3x2345 = _mm_shuffle_ps(vi3x3012, vi3x7456, _MM_SHUFFLE(2, 1, 0, 3));
213 vi3x3012 = vi3x7456;
214 const __m128 vi4x2345 = _mm_shuffle_ps(vi4x3012, vi4x7456, _MM_SHUFFLE(2, 1, 0, 3));
215 vi4x3012 = vi4x7456;
216 const __m128 vi5x2345 = _mm_shuffle_ps(vi5x3012, vi5x7456, _MM_SHUFFLE(2, 1, 0, 3));
217 vi5x3012 = vi5x7456;
218 const __m128 vi6x2345 = _mm_shuffle_ps(vi6x3012, vi6x7456, _MM_SHUFFLE(2, 1, 0, 3));
219 vi6x3012 = vi6x7456;
220 const __m128 vi7x2345 = _mm_shuffle_ps(vi7x3012, vi7x7456, _MM_SHUFFLE(2, 1, 0, 3));
221 vi7x3012 = vi7x7456;
222
223 const __m128 vi0x8567 = _mm_move_ss(vi0x4567, vi0x89AB);
224 vi0x4567 = vi0x89AB;
225 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vi1x89AB);
226 vi1x4567 = vi1x89AB;
227 const __m128 vi2x8567 = _mm_move_ss(vi2x4567, vi2x89AB);
228 vi2x4567 = vi2x89AB;
229 const __m128 vi3x8567 = _mm_move_ss(vi3x4567, vi3x89AB);
230 vi3x4567 = vi3x89AB;
231 const __m128 vi4x8567 = _mm_move_ss(vi4x4567, vi4x89AB);
232 vi4x4567 = vi4x89AB;
233 const __m128 vi5x8567 = _mm_move_ss(vi5x4567, vi5x89AB);
234 vi5x4567 = vi5x89AB;
235 const __m128 vi6x8567 = _mm_move_ss(vi6x4567, vi6x89AB);
236 vi6x4567 = vi6x89AB;
237 const __m128 vi7x8567 = _mm_move_ss(vi7x4567, vi7x89AB);
238 vi7x4567 = vi7x89AB;
239
240 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi0x2345, vk00));
241 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi1x2345, vk00));
242 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi2x2345, vk00));
243 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi3x2345, vk00));
244 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x2345, vk10));
245 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x2345, vk10));
246 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi3x2345, vk10));
247 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi4x2345, vk10));
248 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x2345, vk20));
249 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi3x2345, vk20));
250 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi4x2345, vk20));
251 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi5x2345, vk20));
252 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi3x2345, vk30));
253 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi4x2345, vk30));
254 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi5x2345, vk30));
255 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi6x2345, vk30));
256 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi4x2345, vk40));
257 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi5x2345, vk40));
258 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi6x2345, vk40));
259 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi7x2345, vk40));
260
261 const __m128 vi0x5678 = _mm_shuffle_ps(vi0x8567, vi0x8567, _MM_SHUFFLE(0, 3, 2, 1));
262 const __m128 vi1x5678 = _mm_shuffle_ps(vi1x8567, vi1x8567, _MM_SHUFFLE(0, 3, 2, 1));
263 const __m128 vi2x5678 = _mm_shuffle_ps(vi2x8567, vi2x8567, _MM_SHUFFLE(0, 3, 2, 1));
264 const __m128 vi3x5678 = _mm_shuffle_ps(vi3x8567, vi3x8567, _MM_SHUFFLE(0, 3, 2, 1));
265 const __m128 vi4x5678 = _mm_shuffle_ps(vi4x8567, vi4x8567, _MM_SHUFFLE(0, 3, 2, 1));
266 const __m128 vi5x5678 = _mm_shuffle_ps(vi5x8567, vi5x8567, _MM_SHUFFLE(0, 3, 2, 1));
267 const __m128 vi6x5678 = _mm_shuffle_ps(vi6x8567, vi6x8567, _MM_SHUFFLE(0, 3, 2, 1));
268 const __m128 vi7x5678 = _mm_shuffle_ps(vi7x8567, vi7x8567, _MM_SHUFFLE(0, 3, 2, 1));
269
270 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi0x5678, vk03));
271 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi1x5678, vk03));
272 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi2x5678, vk03));
273 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi3x5678, vk03));
274 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x5678, vk13));
275 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x5678, vk13));
276 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi3x5678, vk13));
277 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi4x5678, vk13));
278 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x5678, vk23));
279 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi3x5678, vk23));
280 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi4x5678, vk23));
281 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi5x5678, vk23));
282 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi3x5678, vk33));
283 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi4x5678, vk33));
284 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi5x5678, vk33));
285 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi6x5678, vk33));
286 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi4x5678, vk43));
287 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi5x5678, vk43));
288 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi6x5678, vk43));
289 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi7x5678, vk43));
290
291 const __m128 vi0x6789 = _mm_shuffle_ps(vi0x5678, vi0x89AB, _MM_SHUFFLE(1, 0, 2, 1));
292 const __m128 vi1x6789 = _mm_shuffle_ps(vi1x5678, vi1x89AB, _MM_SHUFFLE(1, 0, 2, 1));
293 const __m128 vi2x6789 = _mm_shuffle_ps(vi2x5678, vi2x89AB, _MM_SHUFFLE(1, 0, 2, 1));
294 const __m128 vi3x6789 = _mm_shuffle_ps(vi3x5678, vi3x89AB, _MM_SHUFFLE(1, 0, 2, 1));
295 const __m128 vi4x6789 = _mm_shuffle_ps(vi4x5678, vi4x89AB, _MM_SHUFFLE(1, 0, 2, 1));
296 const __m128 vi5x6789 = _mm_shuffle_ps(vi5x5678, vi5x89AB, _MM_SHUFFLE(1, 0, 2, 1));
297 const __m128 vi6x6789 = _mm_shuffle_ps(vi6x5678, vi6x89AB, _MM_SHUFFLE(1, 0, 2, 1));
298 const __m128 vi7x6789 = _mm_shuffle_ps(vi7x5678, vi7x89AB, _MM_SHUFFLE(1, 0, 2, 1));
299
300 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi0x6789, vk04));
301 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi1x6789, vk04));
302 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi2x6789, vk04));
303 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi3x6789, vk04));
304 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x6789, vk14));
305 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x6789, vk14));
306 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi3x6789, vk14));
307 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi4x6789, vk14));
308 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x6789, vk24));
309 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi3x6789, vk24));
310 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi4x6789, vk24));
311 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi5x6789, vk24));
312 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi3x6789, vk34));
313 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi4x6789, vk34));
314 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi5x6789, vk34));
315 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi6x6789, vk34));
316 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi4x6789, vk44));
317 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi5x6789, vk44));
318 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi6x6789, vk44));
319 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi7x6789, vk44));
320
321
322 __m128 vo0 = _mm_max_ps(vo0p0, vmin);
323 __m128 vo1 = _mm_max_ps(vo1p0, vmin);
324 __m128 vo2 = _mm_max_ps(vo2p0, vmin);
325 __m128 vo3 = _mm_max_ps(vo3p0, vmin);
326
327 vo0 = _mm_min_ps(vo0, vmax);
328 vo1 = _mm_min_ps(vo1, vmax);
329 vo2 = _mm_min_ps(vo2, vmax);
330 vo3 = _mm_min_ps(vo3, vmax);
331
332 _mm_storeu_ps(o3, vo3);
333 o3 += 4;
334 _mm_storeu_ps(o2, vo2);
335 o2 += 4;
336 _mm_storeu_ps(o1, vo1);
337 o1 += 4;
338 _mm_storeu_ps(o0, vo0);
339 o0 += 4;
340 }
341 // Always process the last block of 5..8 pixels.
342 if XNN_LIKELY(w > 4 * sizeof(float)) {
343 __m128 vo0p0 = _mm_add_ps(vbias, _mm_mul_ps(vi0x4567, vk02));
344 __m128 vo1p0 = _mm_add_ps(vbias, _mm_mul_ps(vi1x4567, vk02));
345 __m128 vo2p0 = _mm_add_ps(vbias, _mm_mul_ps(vi2x4567, vk02));
346 __m128 vo3p0 = _mm_add_ps(vbias, _mm_mul_ps(vi3x4567, vk02));
347 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x4567, vk12));
348 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x4567, vk12));
349 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi3x4567, vk12));
350 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi4x4567, vk12));
351 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x4567, vk22));
352 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi3x4567, vk22));
353 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi4x4567, vk22));
354 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi5x4567, vk22));
355 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi3x4567, vk32));
356 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi4x4567, vk32));
357 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi5x4567, vk32));
358 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi6x4567, vk32));
359 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi4x4567, vk42));
360 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi5x4567, vk42));
361 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi6x4567, vk42));
362 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi7x4567, vk42));
363
364 const __m128 vi0x7456 = _mm_shuffle_ps(vi0x4567, vi0x4567, _MM_SHUFFLE(2, 1, 0, 3));
365 const __m128 vi1x7456 = _mm_shuffle_ps(vi1x4567, vi1x4567, _MM_SHUFFLE(2, 1, 0, 3));
366 const __m128 vi2x7456 = _mm_shuffle_ps(vi2x4567, vi2x4567, _MM_SHUFFLE(2, 1, 0, 3));
367 const __m128 vi3x7456 = _mm_shuffle_ps(vi3x4567, vi3x4567, _MM_SHUFFLE(2, 1, 0, 3));
368 const __m128 vi4x7456 = _mm_shuffle_ps(vi4x4567, vi4x4567, _MM_SHUFFLE(2, 1, 0, 3));
369 const __m128 vi5x7456 = _mm_shuffle_ps(vi5x4567, vi5x4567, _MM_SHUFFLE(2, 1, 0, 3));
370 const __m128 vi6x7456 = _mm_shuffle_ps(vi6x4567, vi6x4567, _MM_SHUFFLE(2, 1, 0, 3));
371 const __m128 vi7x7456 = _mm_shuffle_ps(vi7x4567, vi7x4567, _MM_SHUFFLE(2, 1, 0, 3));
372
373 const __m128 vi0x89AB = _mm_and_ps(_mm_loadu_ps(i0), vmask);
374 i0 += 4;
375 const __m128 vi1x89AB = _mm_and_ps(_mm_loadu_ps(i1), vmask);
376 i1 += 4;
377 const __m128 vi2x89AB = _mm_and_ps(_mm_loadu_ps(i2), vmask);
378 i2 += 4;
379 const __m128 vi3x89AB = _mm_and_ps(_mm_loadu_ps(i3), vmask);
380 i3 += 4;
381 const __m128 vi4x89AB = _mm_and_ps(_mm_loadu_ps(i4), vmask);
382 i4 += 4;
383 const __m128 vi5x89AB = _mm_and_ps(_mm_loadu_ps(i5), vmask);
384 i5 += 4;
385 const __m128 vi6x89AB = _mm_and_ps(_mm_loadu_ps(i6), vmask);
386 i6 += 4;
387 const __m128 vi7x89AB = _mm_and_ps(_mm_loadu_ps(i7), vmask);
388 i7 += 4;
389
390 const __m128 vi0x3456 = _mm_move_ss(vi0x7456, vi0x3012);
391 const __m128 vi1x3456 = _mm_move_ss(vi1x7456, vi1x3012);
392 const __m128 vi2x3456 = _mm_move_ss(vi2x7456, vi2x3012);
393 const __m128 vi3x3456 = _mm_move_ss(vi3x7456, vi3x3012);
394 const __m128 vi4x3456 = _mm_move_ss(vi4x7456, vi4x3012);
395 const __m128 vi5x3456 = _mm_move_ss(vi5x7456, vi5x3012);
396 const __m128 vi6x3456 = _mm_move_ss(vi6x7456, vi6x3012);
397 const __m128 vi7x3456 = _mm_move_ss(vi7x7456, vi7x3012);
398
399 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi0x3456, vk01));
400 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi1x3456, vk01));
401 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi2x3456, vk01));
402 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi3x3456, vk01));
403 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x3456, vk11));
404 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x3456, vk11));
405 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi3x3456, vk11));
406 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi4x3456, vk11));
407 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x3456, vk21));
408 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi3x3456, vk21));
409 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi4x3456, vk21));
410 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi5x3456, vk21));
411 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi3x3456, vk31));
412 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi4x3456, vk31));
413 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi5x3456, vk31));
414 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi6x3456, vk31));
415 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi4x3456, vk41));
416 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi5x3456, vk41));
417 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi6x3456, vk41));
418 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi7x3456, vk41));
419
420 const __m128 vi0x2345 = _mm_shuffle_ps(vi0x3012, vi0x7456, _MM_SHUFFLE(2, 1, 0, 3));
421 vi0x3012 = vi0x7456;
422 const __m128 vi1x2345 = _mm_shuffle_ps(vi1x3012, vi1x7456, _MM_SHUFFLE(2, 1, 0, 3));
423 vi1x3012 = vi1x7456;
424 const __m128 vi2x2345 = _mm_shuffle_ps(vi2x3012, vi2x7456, _MM_SHUFFLE(2, 1, 0, 3));
425 vi2x3012 = vi2x7456;
426 const __m128 vi3x2345 = _mm_shuffle_ps(vi3x3012, vi3x7456, _MM_SHUFFLE(2, 1, 0, 3));
427 vi3x3012 = vi3x7456;
428 const __m128 vi4x2345 = _mm_shuffle_ps(vi4x3012, vi4x7456, _MM_SHUFFLE(2, 1, 0, 3));
429 vi4x3012 = vi4x7456;
430 const __m128 vi5x2345 = _mm_shuffle_ps(vi5x3012, vi5x7456, _MM_SHUFFLE(2, 1, 0, 3));
431 vi5x3012 = vi5x7456;
432 const __m128 vi6x2345 = _mm_shuffle_ps(vi6x3012, vi6x7456, _MM_SHUFFLE(2, 1, 0, 3));
433 vi6x3012 = vi6x7456;
434 const __m128 vi7x2345 = _mm_shuffle_ps(vi7x3012, vi7x7456, _MM_SHUFFLE(2, 1, 0, 3));
435 vi7x3012 = vi7x7456;
436
437 const __m128 vi0x8567 = _mm_move_ss(vi0x4567, vi0x89AB);
438 vi0x4567 = vi0x89AB;
439 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vi1x89AB);
440 vi1x4567 = vi1x89AB;
441 const __m128 vi2x8567 = _mm_move_ss(vi2x4567, vi2x89AB);
442 vi2x4567 = vi2x89AB;
443 const __m128 vi3x8567 = _mm_move_ss(vi3x4567, vi3x89AB);
444 vi3x4567 = vi3x89AB;
445 const __m128 vi4x8567 = _mm_move_ss(vi4x4567, vi4x89AB);
446 vi4x4567 = vi4x89AB;
447 const __m128 vi5x8567 = _mm_move_ss(vi5x4567, vi5x89AB);
448 vi5x4567 = vi5x89AB;
449 const __m128 vi6x8567 = _mm_move_ss(vi6x4567, vi6x89AB);
450 vi6x4567 = vi6x89AB;
451 const __m128 vi7x8567 = _mm_move_ss(vi7x4567, vi7x89AB);
452 vi7x4567 = vi7x89AB;
453
454 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi0x2345, vk00));
455 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi1x2345, vk00));
456 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi2x2345, vk00));
457 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi3x2345, vk00));
458 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x2345, vk10));
459 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x2345, vk10));
460 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi3x2345, vk10));
461 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi4x2345, vk10));
462 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x2345, vk20));
463 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi3x2345, vk20));
464 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi4x2345, vk20));
465 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi5x2345, vk20));
466 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi3x2345, vk30));
467 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi4x2345, vk30));
468 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi5x2345, vk30));
469 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi6x2345, vk30));
470 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi4x2345, vk40));
471 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi5x2345, vk40));
472 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi6x2345, vk40));
473 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi7x2345, vk40));
474
475 const __m128 vi0x5678 = _mm_shuffle_ps(vi0x8567, vi0x8567, _MM_SHUFFLE(0, 3, 2, 1));
476 const __m128 vi1x5678 = _mm_shuffle_ps(vi1x8567, vi1x8567, _MM_SHUFFLE(0, 3, 2, 1));
477 const __m128 vi2x5678 = _mm_shuffle_ps(vi2x8567, vi2x8567, _MM_SHUFFLE(0, 3, 2, 1));
478 const __m128 vi3x5678 = _mm_shuffle_ps(vi3x8567, vi3x8567, _MM_SHUFFLE(0, 3, 2, 1));
479 const __m128 vi4x5678 = _mm_shuffle_ps(vi4x8567, vi4x8567, _MM_SHUFFLE(0, 3, 2, 1));
480 const __m128 vi5x5678 = _mm_shuffle_ps(vi5x8567, vi5x8567, _MM_SHUFFLE(0, 3, 2, 1));
481 const __m128 vi6x5678 = _mm_shuffle_ps(vi6x8567, vi6x8567, _MM_SHUFFLE(0, 3, 2, 1));
482 const __m128 vi7x5678 = _mm_shuffle_ps(vi7x8567, vi7x8567, _MM_SHUFFLE(0, 3, 2, 1));
483
484 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi0x5678, vk03));
485 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi1x5678, vk03));
486 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi2x5678, vk03));
487 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi3x5678, vk03));
488 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x5678, vk13));
489 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x5678, vk13));
490 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi3x5678, vk13));
491 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi4x5678, vk13));
492 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x5678, vk23));
493 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi3x5678, vk23));
494 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi4x5678, vk23));
495 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi5x5678, vk23));
496 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi3x5678, vk33));
497 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi4x5678, vk33));
498 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi5x5678, vk33));
499 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi6x5678, vk33));
500 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi4x5678, vk43));
501 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi5x5678, vk43));
502 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi6x5678, vk43));
503 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi7x5678, vk43));
504
505 const __m128 vi0x6789 = _mm_shuffle_ps(vi0x5678, vi0x89AB, _MM_SHUFFLE(1, 0, 2, 1));
506 const __m128 vi1x6789 = _mm_shuffle_ps(vi1x5678, vi1x89AB, _MM_SHUFFLE(1, 0, 2, 1));
507 const __m128 vi2x6789 = _mm_shuffle_ps(vi2x5678, vi2x89AB, _MM_SHUFFLE(1, 0, 2, 1));
508 const __m128 vi3x6789 = _mm_shuffle_ps(vi3x5678, vi3x89AB, _MM_SHUFFLE(1, 0, 2, 1));
509 const __m128 vi4x6789 = _mm_shuffle_ps(vi4x5678, vi4x89AB, _MM_SHUFFLE(1, 0, 2, 1));
510 const __m128 vi5x6789 = _mm_shuffle_ps(vi5x5678, vi5x89AB, _MM_SHUFFLE(1, 0, 2, 1));
511 const __m128 vi6x6789 = _mm_shuffle_ps(vi6x5678, vi6x89AB, _MM_SHUFFLE(1, 0, 2, 1));
512 const __m128 vi7x6789 = _mm_shuffle_ps(vi7x5678, vi7x89AB, _MM_SHUFFLE(1, 0, 2, 1));
513
514 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi0x6789, vk04));
515 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi1x6789, vk04));
516 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi2x6789, vk04));
517 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi3x6789, vk04));
518 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x6789, vk14));
519 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x6789, vk14));
520 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi3x6789, vk14));
521 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi4x6789, vk14));
522 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x6789, vk24));
523 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi3x6789, vk24));
524 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi4x6789, vk24));
525 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi5x6789, vk24));
526 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi3x6789, vk34));
527 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi4x6789, vk34));
528 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi5x6789, vk34));
529 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi6x6789, vk34));
530 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi4x6789, vk44));
531 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi5x6789, vk44));
532 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi6x6789, vk44));
533 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi7x6789, vk44));
534
535
536 __m128 vo0 = _mm_max_ps(vo0p0, vmin);
537 __m128 vo1 = _mm_max_ps(vo1p0, vmin);
538 __m128 vo2 = _mm_max_ps(vo2p0, vmin);
539 __m128 vo3 = _mm_max_ps(vo3p0, vmin);
540
541 vo0 = _mm_min_ps(vo0, vmax);
542 vo1 = _mm_min_ps(vo1, vmax);
543 vo2 = _mm_min_ps(vo2, vmax);
544 vo3 = _mm_min_ps(vo3, vmax);
545
546 _mm_storeu_ps(o3, vo3);
547 o3 += 4;
548 _mm_storeu_ps(o2, vo2);
549 o2 += 4;
550 _mm_storeu_ps(o1, vo1);
551 o1 += 4;
552 _mm_storeu_ps(o0, vo0);
553 o0 += 4;
554
555 w -= 4 * sizeof(float);
556 }
557 assert(w >= 1 * sizeof(float));
558 assert(w <= 4 * sizeof(float));
559 {
560 vi0x4567 = _mm_and_ps(vi0x4567, vmask);
561 vi1x4567 = _mm_and_ps(vi1x4567, vmask);
562 vi2x4567 = _mm_and_ps(vi2x4567, vmask);
563 vi3x4567 = _mm_and_ps(vi3x4567, vmask);
564 vi4x4567 = _mm_and_ps(vi4x4567, vmask);
565 vi5x4567 = _mm_and_ps(vi5x4567, vmask);
566 vi6x4567 = _mm_and_ps(vi6x4567, vmask);
567 vi7x4567 = _mm_and_ps(vi7x4567, vmask);
568
569 __m128 vo0p0 = _mm_add_ps(vbias, _mm_mul_ps(vi0x4567, vk02));
570 __m128 vo1p0 = _mm_add_ps(vbias, _mm_mul_ps(vi1x4567, vk02));
571 __m128 vo2p0 = _mm_add_ps(vbias, _mm_mul_ps(vi2x4567, vk02));
572 __m128 vo3p0 = _mm_add_ps(vbias, _mm_mul_ps(vi3x4567, vk02));
573 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x4567, vk12));
574 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x4567, vk12));
575 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi3x4567, vk12));
576 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi4x4567, vk12));
577 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x4567, vk22));
578 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi3x4567, vk22));
579 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi4x4567, vk22));
580 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi5x4567, vk22));
581 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi3x4567, vk32));
582 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi4x4567, vk32));
583 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi5x4567, vk32));
584 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi6x4567, vk32));
585 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi4x4567, vk42));
586 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi5x4567, vk42));
587 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi6x4567, vk42));
588 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi7x4567, vk42));
589
590 const __m128 vi0x7456 = _mm_shuffle_ps(vi0x4567, vi0x4567, _MM_SHUFFLE(2, 1, 0, 3));
591 const __m128 vi1x7456 = _mm_shuffle_ps(vi1x4567, vi1x4567, _MM_SHUFFLE(2, 1, 0, 3));
592 const __m128 vi2x7456 = _mm_shuffle_ps(vi2x4567, vi2x4567, _MM_SHUFFLE(2, 1, 0, 3));
593 const __m128 vi3x7456 = _mm_shuffle_ps(vi3x4567, vi3x4567, _MM_SHUFFLE(2, 1, 0, 3));
594 const __m128 vi4x7456 = _mm_shuffle_ps(vi4x4567, vi4x4567, _MM_SHUFFLE(2, 1, 0, 3));
595 const __m128 vi5x7456 = _mm_shuffle_ps(vi5x4567, vi5x4567, _MM_SHUFFLE(2, 1, 0, 3));
596 const __m128 vi6x7456 = _mm_shuffle_ps(vi6x4567, vi6x4567, _MM_SHUFFLE(2, 1, 0, 3));
597 const __m128 vi7x7456 = _mm_shuffle_ps(vi7x4567, vi7x4567, _MM_SHUFFLE(2, 1, 0, 3));
598
599 const __m128 vi0x3456 = _mm_move_ss(vi0x7456, vi0x3012);
600 const __m128 vi1x3456 = _mm_move_ss(vi1x7456, vi1x3012);
601 const __m128 vi2x3456 = _mm_move_ss(vi2x7456, vi2x3012);
602 const __m128 vi3x3456 = _mm_move_ss(vi3x7456, vi3x3012);
603 const __m128 vi4x3456 = _mm_move_ss(vi4x7456, vi4x3012);
604 const __m128 vi5x3456 = _mm_move_ss(vi5x7456, vi5x3012);
605 const __m128 vi6x3456 = _mm_move_ss(vi6x7456, vi6x3012);
606 const __m128 vi7x3456 = _mm_move_ss(vi7x7456, vi7x3012);
607
608 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi0x3456, vk01));
609 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi1x3456, vk01));
610 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi2x3456, vk01));
611 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi3x3456, vk01));
612 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x3456, vk11));
613 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x3456, vk11));
614 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi3x3456, vk11));
615 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi4x3456, vk11));
616 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x3456, vk21));
617 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi3x3456, vk21));
618 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi4x3456, vk21));
619 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi5x3456, vk21));
620 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi3x3456, vk31));
621 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi4x3456, vk31));
622 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi5x3456, vk31));
623 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi6x3456, vk31));
624 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi4x3456, vk41));
625 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi5x3456, vk41));
626 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi6x3456, vk41));
627 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi7x3456, vk41));
628
629 const __m128 vi0x2345 = _mm_shuffle_ps(vi0x3012, vi0x7456, _MM_SHUFFLE(2, 1, 0, 3));
630 const __m128 vi1x2345 = _mm_shuffle_ps(vi1x3012, vi1x7456, _MM_SHUFFLE(2, 1, 0, 3));
631 const __m128 vi2x2345 = _mm_shuffle_ps(vi2x3012, vi2x7456, _MM_SHUFFLE(2, 1, 0, 3));
632 const __m128 vi3x2345 = _mm_shuffle_ps(vi3x3012, vi3x7456, _MM_SHUFFLE(2, 1, 0, 3));
633 const __m128 vi4x2345 = _mm_shuffle_ps(vi4x3012, vi4x7456, _MM_SHUFFLE(2, 1, 0, 3));
634 const __m128 vi5x2345 = _mm_shuffle_ps(vi5x3012, vi5x7456, _MM_SHUFFLE(2, 1, 0, 3));
635 const __m128 vi6x2345 = _mm_shuffle_ps(vi6x3012, vi6x7456, _MM_SHUFFLE(2, 1, 0, 3));
636 const __m128 vi7x2345 = _mm_shuffle_ps(vi7x3012, vi7x7456, _MM_SHUFFLE(2, 1, 0, 3));
637
638 const __m128 vzero = _mm_setzero_ps();
639 const __m128 vi0x8567 = _mm_move_ss(vi0x4567, vzero);
640 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vzero);
641 const __m128 vi2x8567 = _mm_move_ss(vi2x4567, vzero);
642 const __m128 vi3x8567 = _mm_move_ss(vi3x4567, vzero);
643 const __m128 vi4x8567 = _mm_move_ss(vi4x4567, vzero);
644 const __m128 vi5x8567 = _mm_move_ss(vi5x4567, vzero);
645 const __m128 vi6x8567 = _mm_move_ss(vi6x4567, vzero);
646 const __m128 vi7x8567 = _mm_move_ss(vi7x4567, vzero);
647
648 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi0x2345, vk00));
649 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi1x2345, vk00));
650 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi2x2345, vk00));
651 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi3x2345, vk00));
652 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x2345, vk10));
653 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x2345, vk10));
654 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi3x2345, vk10));
655 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi4x2345, vk10));
656 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x2345, vk20));
657 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi3x2345, vk20));
658 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi4x2345, vk20));
659 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi5x2345, vk20));
660 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi3x2345, vk30));
661 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi4x2345, vk30));
662 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi5x2345, vk30));
663 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi6x2345, vk30));
664 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi4x2345, vk40));
665 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi5x2345, vk40));
666 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi6x2345, vk40));
667 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi7x2345, vk40));
668
669 const __m128 vi0x5678 = _mm_shuffle_ps(vi0x8567, vi0x8567, _MM_SHUFFLE(0, 3, 2, 1));
670 const __m128 vi1x5678 = _mm_shuffle_ps(vi1x8567, vi1x8567, _MM_SHUFFLE(0, 3, 2, 1));
671 const __m128 vi2x5678 = _mm_shuffle_ps(vi2x8567, vi2x8567, _MM_SHUFFLE(0, 3, 2, 1));
672 const __m128 vi3x5678 = _mm_shuffle_ps(vi3x8567, vi3x8567, _MM_SHUFFLE(0, 3, 2, 1));
673 const __m128 vi4x5678 = _mm_shuffle_ps(vi4x8567, vi4x8567, _MM_SHUFFLE(0, 3, 2, 1));
674 const __m128 vi5x5678 = _mm_shuffle_ps(vi5x8567, vi5x8567, _MM_SHUFFLE(0, 3, 2, 1));
675 const __m128 vi6x5678 = _mm_shuffle_ps(vi6x8567, vi6x8567, _MM_SHUFFLE(0, 3, 2, 1));
676 const __m128 vi7x5678 = _mm_shuffle_ps(vi7x8567, vi7x8567, _MM_SHUFFLE(0, 3, 2, 1));
677
678 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi0x5678, vk03));
679 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi1x5678, vk03));
680 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi2x5678, vk03));
681 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi3x5678, vk03));
682 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x5678, vk13));
683 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x5678, vk13));
684 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi3x5678, vk13));
685 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi4x5678, vk13));
686 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x5678, vk23));
687 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi3x5678, vk23));
688 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi4x5678, vk23));
689 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi5x5678, vk23));
690 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi3x5678, vk33));
691 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi4x5678, vk33));
692 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi5x5678, vk33));
693 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi6x5678, vk33));
694 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi4x5678, vk43));
695 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi5x5678, vk43));
696 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi6x5678, vk43));
697 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi7x5678, vk43));
698
699 const __m128 vi0x6789 = _mm_shuffle_ps(vi0x5678, vzero, _MM_SHUFFLE(1, 0, 2, 1));
700 const __m128 vi1x6789 = _mm_shuffle_ps(vi1x5678, vzero, _MM_SHUFFLE(1, 0, 2, 1));
701 const __m128 vi2x6789 = _mm_shuffle_ps(vi2x5678, vzero, _MM_SHUFFLE(1, 0, 2, 1));
702 const __m128 vi3x6789 = _mm_shuffle_ps(vi3x5678, vzero, _MM_SHUFFLE(1, 0, 2, 1));
703 const __m128 vi4x6789 = _mm_shuffle_ps(vi4x5678, vzero, _MM_SHUFFLE(1, 0, 2, 1));
704 const __m128 vi5x6789 = _mm_shuffle_ps(vi5x5678, vzero, _MM_SHUFFLE(1, 0, 2, 1));
705 const __m128 vi6x6789 = _mm_shuffle_ps(vi6x5678, vzero, _MM_SHUFFLE(1, 0, 2, 1));
706 const __m128 vi7x6789 = _mm_shuffle_ps(vi7x5678, vzero, _MM_SHUFFLE(1, 0, 2, 1));
707
708 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi0x6789, vk04));
709 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi1x6789, vk04));
710 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi2x6789, vk04));
711 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi3x6789, vk04));
712 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x6789, vk14));
713 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x6789, vk14));
714 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi3x6789, vk14));
715 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi4x6789, vk14));
716 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x6789, vk24));
717 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi3x6789, vk24));
718 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi4x6789, vk24));
719 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi5x6789, vk24));
720 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi3x6789, vk34));
721 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi4x6789, vk34));
722 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi5x6789, vk34));
723 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi6x6789, vk34));
724 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi4x6789, vk44));
725 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi5x6789, vk44));
726 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi6x6789, vk44));
727 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi7x6789, vk44));
728
729
730 __m128 vo0 = _mm_max_ps(vo0p0, vmin);
731 __m128 vo1 = _mm_max_ps(vo1p0, vmin);
732 __m128 vo2 = _mm_max_ps(vo2p0, vmin);
733 __m128 vo3 = _mm_max_ps(vo3p0, vmin);
734
735 vo0 = _mm_min_ps(vo0, vmax);
736 vo1 = _mm_min_ps(vo1, vmax);
737 vo2 = _mm_min_ps(vo2, vmax);
738 vo3 = _mm_min_ps(vo3, vmax);
739
740 if XNN_LIKELY(w & (4 * sizeof(float))) {
741 _mm_storeu_ps(o3, vo3);
742 o3 += 4;
743 _mm_storeu_ps(o2, vo2);
744 o2 += 4;
745 _mm_storeu_ps(o1, vo1);
746 o1 += 4;
747 _mm_storeu_ps(o0, vo0);
748 o0 += 4;
749 } else {
750 if (w & (2 * sizeof(float))) {
751 _mm_storel_pi((__m64*) o3, vo3);
752 o3 += 2;
753 _mm_storel_pi((__m64*) o2, vo2);
754 o2 += 2;
755 _mm_storel_pi((__m64*) o1, vo1);
756 o1 += 2;
757 _mm_storel_pi((__m64*) o0, vo0);
758 o0 += 2;
759
760 vo0 = _mm_movehl_ps(vo0, vo0);
761 vo1 = _mm_movehl_ps(vo1, vo1);
762 vo2 = _mm_movehl_ps(vo2, vo2);
763 vo3 = _mm_movehl_ps(vo3, vo3);
764 }
765 if (w & (1 * sizeof(float))) {
766 _mm_store_ss(o3, vo3);
767 o3 += 1;
768 _mm_store_ss(o2, vo2);
769 o2 += 1;
770 _mm_store_ss(o1, vo1);
771 o1 += 1;
772 _mm_store_ss(o0, vo0);
773 o0 += 1;
774 }
775 }
776 }
777
778 i0 = (const float*) ((uintptr_t) i4 - input_decrement);
779 i1 = (const float*) ((uintptr_t) i5 - input_decrement);
780 i2 = (const float*) ((uintptr_t) i1 + input_width);
781 i3 = (const float*) ((uintptr_t) i2 + input_width);
782 i4 = (const float*) ((uintptr_t) i3 + input_width);
783 i5 = (const float*) ((uintptr_t) i4 + input_width);
784 i6 = (const float*) ((uintptr_t) i5 + input_width);
785 i7 = (const float*) ((uintptr_t) i6 + input_width);
786
787 o0 = o3;
788 o1 = (float*) ((uintptr_t) o0 + input_width);
789 o2 = (float*) ((uintptr_t) o1 + input_width);
790 o3 = (float*) ((uintptr_t) o2 + input_width);
791
792 output_height = doz(output_height, 4);
793 } while (output_height != 0);
794 }
795