1 // Auto-generated file. Do not edit!
2 // Template: src/f32-dwconv2d-chw/3x3p1-ssse3.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 <tmmintrin.h>
13
14 #include <xnnpack/dwconv.h>
15 #include <xnnpack/math.h>
16
17
xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_6x4(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_3x3p1__ssse3_6x4(
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 == 1);
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 vk10 = _mm_load1_ps(weights + 4);
42 const __m128 vk11 = _mm_load1_ps(weights + 5);
43 const __m128 vk12 = _mm_load1_ps(weights + 6);
44 const __m128 vk20 = _mm_load1_ps(weights + 7);
45 const __m128 vk21 = _mm_load1_ps(weights + 8);
46 const __m128 vk22 = _mm_load1_ps(weights + 9);
47
48 const size_t input_decrement = round_up_po2(input_width, 4 * sizeof(float));
49
50 const float* i0 = zero;
51 const float* i1 = input;
52 const float* i2 = (const float*) ((uintptr_t) i1 + input_width);
53 const float* i3 = (const float*) ((uintptr_t) i2 + input_width);
54 const float* i4 = (const float*) ((uintptr_t) i3 + input_width);
55 const float* i5 = (const float*) ((uintptr_t) i4 + input_width);
56 const float* i6 = (const float*) ((uintptr_t) i5 + input_width);
57 const float* i7 = (const float*) ((uintptr_t) i6 + input_width);
58
59 float* o0 = output;
60 float* o1 = (float*) ((uintptr_t) o0 + input_width);
61 float* o2 = (float*) ((uintptr_t) o1 + input_width);
62 float* o3 = (float*) ((uintptr_t) o2 + input_width);
63 float* o4 = (float*) ((uintptr_t) o3 + input_width);
64 float* o5 = (float*) ((uintptr_t) o4 + input_width);
65
66 size_t output_height = input_height;
67 do {
68 if XNN_UNPREDICTABLE(output_height < 2) {
69 i2 = zero;
70 o1 = o0;
71 }
72 if XNN_UNPREDICTABLE(output_height < 3) {
73 i3 = zero;
74 o2 = o1;
75 }
76 if XNN_UNPREDICTABLE(output_height < 4) {
77 i4 = zero;
78 o3 = o2;
79 }
80 if XNN_UNPREDICTABLE(output_height < 5) {
81 i5 = zero;
82 o4 = o3;
83 }
84 if XNN_UNPREDICTABLE(output_height < 6) {
85 i6 = zero;
86 o5 = o4;
87 }
88 if XNN_UNPREDICTABLE(output_height < 7) {
89 i7 = zero;
90 }
91
92 __m128 vi0x0123 = _mm_setzero_ps();
93 __m128 vi1x0123 = _mm_setzero_ps();
94 __m128 vi2x0123 = _mm_setzero_ps();
95 __m128 vi3x0123 = _mm_setzero_ps();
96 __m128 vi4x0123 = _mm_setzero_ps();
97 __m128 vi5x0123 = _mm_setzero_ps();
98 __m128 vi6x0123 = _mm_setzero_ps();
99 __m128 vi7x0123 = _mm_setzero_ps();
100
101 __m128 vi0x4567 = _mm_loadu_ps(i0);
102 i0 += 4;
103 __m128 vi1x4567 = _mm_loadu_ps(i1);
104 i1 += 4;
105 __m128 vi2x4567 = _mm_loadu_ps(i2);
106 i2 += 4;
107 __m128 vi3x4567 = _mm_loadu_ps(i3);
108 i3 += 4;
109 __m128 vi4x4567 = _mm_loadu_ps(i4);
110 i4 += 4;
111 __m128 vi5x4567 = _mm_loadu_ps(i5);
112 i5 += 4;
113 __m128 vi6x4567 = _mm_loadu_ps(i6);
114 i6 += 4;
115 __m128 vi7x4567 = _mm_loadu_ps(i7);
116 i7 += 4;
117
118 size_t w = input_width;
119 for (; w > 4 * sizeof(float); w -= 4 * sizeof(float)) {
120 const __m128 vi0x89AB = _mm_loadu_ps(i0);
121 i0 += 4;
122 const __m128 vi1x89AB = _mm_loadu_ps(i1);
123 i1 += 4;
124 const __m128 vi2x89AB = _mm_loadu_ps(i2);
125 i2 += 4;
126 const __m128 vi3x89AB = _mm_loadu_ps(i3);
127 i3 += 4;
128 const __m128 vi4x89AB = _mm_loadu_ps(i4);
129 i4 += 4;
130 const __m128 vi5x89AB = _mm_loadu_ps(i5);
131 i5 += 4;
132 const __m128 vi6x89AB = _mm_loadu_ps(i6);
133 i6 += 4;
134 const __m128 vi7x89AB = _mm_loadu_ps(i7);
135 i7 += 4;
136
137 __m128 vo0p0 = _mm_add_ps(vbias, _mm_mul_ps(vi0x4567, vk01));
138 __m128 vo1p0 = _mm_add_ps(vbias, _mm_mul_ps(vi1x4567, vk01));
139 __m128 vo2p0 = _mm_add_ps(vbias, _mm_mul_ps(vi2x4567, vk01));
140 __m128 vo3p0 = _mm_add_ps(vbias, _mm_mul_ps(vi3x4567, vk01));
141 __m128 vo4p0 = _mm_add_ps(vbias, _mm_mul_ps(vi4x4567, vk01));
142 __m128 vo5p0 = _mm_add_ps(vbias, _mm_mul_ps(vi5x4567, vk01));
143 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x4567, vk11));
144 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x4567, vk11));
145 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi3x4567, vk11));
146 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi4x4567, vk11));
147 vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi5x4567, vk11));
148 vo5p0 = _mm_add_ps(vo5p0, _mm_mul_ps(vi6x4567, vk11));
149 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x4567, vk21));
150 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi3x4567, vk21));
151 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi4x4567, vk21));
152 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi5x4567, vk21));
153 vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi6x4567, vk21));
154 vo5p0 = _mm_add_ps(vo5p0, _mm_mul_ps(vi7x4567, vk21));
155
156 const __m128 vi0x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi0x4567), _mm_castps_si128(vi0x0123), 12));
157 const __m128 vi1x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi1x4567), _mm_castps_si128(vi1x0123), 12));
158 const __m128 vi2x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi2x4567), _mm_castps_si128(vi2x0123), 12));
159 const __m128 vi3x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi3x4567), _mm_castps_si128(vi3x0123), 12));
160 const __m128 vi4x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi4x4567), _mm_castps_si128(vi4x0123), 12));
161 const __m128 vi5x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi5x4567), _mm_castps_si128(vi5x0123), 12));
162 const __m128 vi6x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi6x4567), _mm_castps_si128(vi6x0123), 12));
163 const __m128 vi7x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi7x4567), _mm_castps_si128(vi7x0123), 12));
164
165 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi0x3456, vk00));
166 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi1x3456, vk00));
167 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi2x3456, vk00));
168 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi3x3456, vk00));
169 vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi4x3456, vk00));
170 vo5p0 = _mm_add_ps(vo5p0, _mm_mul_ps(vi5x3456, vk00));
171 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x3456, vk10));
172 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x3456, vk10));
173 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi3x3456, vk10));
174 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi4x3456, vk10));
175 vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi5x3456, vk10));
176 vo5p0 = _mm_add_ps(vo5p0, _mm_mul_ps(vi6x3456, vk10));
177 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x3456, vk20));
178 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi3x3456, vk20));
179 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi4x3456, vk20));
180 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi5x3456, vk20));
181 vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi6x3456, vk20));
182 vo5p0 = _mm_add_ps(vo5p0, _mm_mul_ps(vi7x3456, vk20));
183
184 vi0x0123 = vi0x4567;
185 vi1x0123 = vi1x4567;
186 vi2x0123 = vi2x4567;
187 vi3x0123 = vi3x4567;
188 vi4x0123 = vi4x4567;
189 vi5x0123 = vi5x4567;
190 vi6x0123 = vi6x4567;
191 vi7x0123 = vi7x4567;
192
193 const __m128 vi0x5678 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi0x89AB), _mm_castps_si128(vi0x4567), 4));
194 const __m128 vi1x5678 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi1x89AB), _mm_castps_si128(vi1x4567), 4));
195 const __m128 vi2x5678 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi2x89AB), _mm_castps_si128(vi2x4567), 4));
196 const __m128 vi3x5678 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi3x89AB), _mm_castps_si128(vi3x4567), 4));
197 const __m128 vi4x5678 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi4x89AB), _mm_castps_si128(vi4x4567), 4));
198 const __m128 vi5x5678 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi5x89AB), _mm_castps_si128(vi5x4567), 4));
199 const __m128 vi6x5678 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi6x89AB), _mm_castps_si128(vi6x4567), 4));
200 const __m128 vi7x5678 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi7x89AB), _mm_castps_si128(vi7x4567), 4));
201
202 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi0x5678, vk02));
203 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi1x5678, vk02));
204 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi2x5678, vk02));
205 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi3x5678, vk02));
206 vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi4x5678, vk02));
207 vo5p0 = _mm_add_ps(vo5p0, _mm_mul_ps(vi5x5678, vk02));
208 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x5678, vk12));
209 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x5678, vk12));
210 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi3x5678, vk12));
211 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi4x5678, vk12));
212 vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi5x5678, vk12));
213 vo5p0 = _mm_add_ps(vo5p0, _mm_mul_ps(vi6x5678, vk12));
214 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x5678, vk22));
215 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi3x5678, vk22));
216 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi4x5678, vk22));
217 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi5x5678, vk22));
218 vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi6x5678, vk22));
219 vo5p0 = _mm_add_ps(vo5p0, _mm_mul_ps(vi7x5678, vk22));
220
221 vi0x4567 = vi0x89AB;
222 vi1x4567 = vi1x89AB;
223 vi2x4567 = vi2x89AB;
224 vi3x4567 = vi3x89AB;
225 vi4x4567 = vi4x89AB;
226 vi5x4567 = vi5x89AB;
227 vi6x4567 = vi6x89AB;
228 vi7x4567 = vi7x89AB;
229
230
231 __m128 vo0 = _mm_max_ps(vo0p0, vmin);
232 __m128 vo1 = _mm_max_ps(vo1p0, vmin);
233 __m128 vo2 = _mm_max_ps(vo2p0, vmin);
234 __m128 vo3 = _mm_max_ps(vo3p0, vmin);
235 __m128 vo4 = _mm_max_ps(vo4p0, vmin);
236 __m128 vo5 = _mm_max_ps(vo5p0, vmin);
237
238 vo0 = _mm_min_ps(vo0, vmax);
239 vo1 = _mm_min_ps(vo1, vmax);
240 vo2 = _mm_min_ps(vo2, vmax);
241 vo3 = _mm_min_ps(vo3, vmax);
242 vo4 = _mm_min_ps(vo4, vmax);
243 vo5 = _mm_min_ps(vo5, vmax);
244
245 _mm_storeu_ps(o5, vo5);
246 o5 += 4;
247 _mm_storeu_ps(o4, vo4);
248 o4 += 4;
249 _mm_storeu_ps(o3, vo3);
250 o3 += 4;
251 _mm_storeu_ps(o2, vo2);
252 o2 += 4;
253 _mm_storeu_ps(o1, vo1);
254 o1 += 4;
255 _mm_storeu_ps(o0, vo0);
256 o0 += 4;
257 }
258 // Always process the last block of 1..4 pixels.
259 assert(w >= 1 * sizeof(float));
260 assert(w <= 4 * sizeof(float));
261 {
262 vi0x4567 = _mm_and_ps(vmask, vi0x4567);
263 vi1x4567 = _mm_and_ps(vmask, vi1x4567);
264 vi2x4567 = _mm_and_ps(vmask, vi2x4567);
265 vi3x4567 = _mm_and_ps(vmask, vi3x4567);
266 vi4x4567 = _mm_and_ps(vmask, vi4x4567);
267 vi5x4567 = _mm_and_ps(vmask, vi5x4567);
268 vi6x4567 = _mm_and_ps(vmask, vi6x4567);
269 vi7x4567 = _mm_and_ps(vmask, vi7x4567);
270
271 __m128 vo0p0 = _mm_add_ps(vbias, _mm_mul_ps(vi0x4567, vk01));
272 __m128 vo1p0 = _mm_add_ps(vbias, _mm_mul_ps(vi1x4567, vk01));
273 __m128 vo2p0 = _mm_add_ps(vbias, _mm_mul_ps(vi2x4567, vk01));
274 __m128 vo3p0 = _mm_add_ps(vbias, _mm_mul_ps(vi3x4567, vk01));
275 __m128 vo4p0 = _mm_add_ps(vbias, _mm_mul_ps(vi4x4567, vk01));
276 __m128 vo5p0 = _mm_add_ps(vbias, _mm_mul_ps(vi5x4567, vk01));
277 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x4567, vk11));
278 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x4567, vk11));
279 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi3x4567, vk11));
280 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi4x4567, vk11));
281 vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi5x4567, vk11));
282 vo5p0 = _mm_add_ps(vo5p0, _mm_mul_ps(vi6x4567, vk11));
283 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x4567, vk21));
284 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi3x4567, vk21));
285 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi4x4567, vk21));
286 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi5x4567, vk21));
287 vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi6x4567, vk21));
288 vo5p0 = _mm_add_ps(vo5p0, _mm_mul_ps(vi7x4567, vk21));
289
290 const __m128 vi0x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi0x4567), _mm_castps_si128(vi0x0123), 12));
291 const __m128 vi1x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi1x4567), _mm_castps_si128(vi1x0123), 12));
292 const __m128 vi2x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi2x4567), _mm_castps_si128(vi2x0123), 12));
293 const __m128 vi3x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi3x4567), _mm_castps_si128(vi3x0123), 12));
294 const __m128 vi4x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi4x4567), _mm_castps_si128(vi4x0123), 12));
295 const __m128 vi5x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi5x4567), _mm_castps_si128(vi5x0123), 12));
296 const __m128 vi6x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi6x4567), _mm_castps_si128(vi6x0123), 12));
297 const __m128 vi7x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi7x4567), _mm_castps_si128(vi7x0123), 12));
298
299 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi0x3456, vk00));
300 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi1x3456, vk00));
301 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi2x3456, vk00));
302 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi3x3456, vk00));
303 vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi4x3456, vk00));
304 vo5p0 = _mm_add_ps(vo5p0, _mm_mul_ps(vi5x3456, vk00));
305 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x3456, vk10));
306 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x3456, vk10));
307 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi3x3456, vk10));
308 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi4x3456, vk10));
309 vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi5x3456, vk10));
310 vo5p0 = _mm_add_ps(vo5p0, _mm_mul_ps(vi6x3456, vk10));
311 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x3456, vk20));
312 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi3x3456, vk20));
313 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi4x3456, vk20));
314 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi5x3456, vk20));
315 vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi6x3456, vk20));
316 vo5p0 = _mm_add_ps(vo5p0, _mm_mul_ps(vi7x3456, vk20));
317
318 const __m128i vzero = _mm_setzero_si128();
319 const __m128 vi0x5678 = _mm_castsi128_ps(_mm_alignr_epi8(vzero, _mm_castps_si128(vi0x4567), 4));
320 const __m128 vi1x5678 = _mm_castsi128_ps(_mm_alignr_epi8(vzero, _mm_castps_si128(vi1x4567), 4));
321 const __m128 vi2x5678 = _mm_castsi128_ps(_mm_alignr_epi8(vzero, _mm_castps_si128(vi2x4567), 4));
322 const __m128 vi3x5678 = _mm_castsi128_ps(_mm_alignr_epi8(vzero, _mm_castps_si128(vi3x4567), 4));
323 const __m128 vi4x5678 = _mm_castsi128_ps(_mm_alignr_epi8(vzero, _mm_castps_si128(vi4x4567), 4));
324 const __m128 vi5x5678 = _mm_castsi128_ps(_mm_alignr_epi8(vzero, _mm_castps_si128(vi5x4567), 4));
325 const __m128 vi6x5678 = _mm_castsi128_ps(_mm_alignr_epi8(vzero, _mm_castps_si128(vi6x4567), 4));
326 const __m128 vi7x5678 = _mm_castsi128_ps(_mm_alignr_epi8(vzero, _mm_castps_si128(vi7x4567), 4));
327
328 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi0x5678, vk02));
329 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi1x5678, vk02));
330 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi2x5678, vk02));
331 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi3x5678, vk02));
332 vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi4x5678, vk02));
333 vo5p0 = _mm_add_ps(vo5p0, _mm_mul_ps(vi5x5678, vk02));
334 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x5678, vk12));
335 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x5678, vk12));
336 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi3x5678, vk12));
337 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi4x5678, vk12));
338 vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi5x5678, vk12));
339 vo5p0 = _mm_add_ps(vo5p0, _mm_mul_ps(vi6x5678, vk12));
340 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x5678, vk22));
341 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi3x5678, vk22));
342 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi4x5678, vk22));
343 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi5x5678, vk22));
344 vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi6x5678, vk22));
345 vo5p0 = _mm_add_ps(vo5p0, _mm_mul_ps(vi7x5678, vk22));
346
347
348 __m128 vo0 = _mm_max_ps(vo0p0, vmin);
349 __m128 vo1 = _mm_max_ps(vo1p0, vmin);
350 __m128 vo2 = _mm_max_ps(vo2p0, vmin);
351 __m128 vo3 = _mm_max_ps(vo3p0, vmin);
352 __m128 vo4 = _mm_max_ps(vo4p0, vmin);
353 __m128 vo5 = _mm_max_ps(vo5p0, vmin);
354
355 vo0 = _mm_min_ps(vo0, vmax);
356 vo1 = _mm_min_ps(vo1, vmax);
357 vo2 = _mm_min_ps(vo2, vmax);
358 vo3 = _mm_min_ps(vo3, vmax);
359 vo4 = _mm_min_ps(vo4, vmax);
360 vo5 = _mm_min_ps(vo5, vmax);
361
362 if XNN_LIKELY(w == 4 * sizeof(float)) {
363 _mm_storeu_ps(o5, vo5);
364 o5 += 4;
365 _mm_storeu_ps(o4, vo4);
366 o4 += 4;
367 _mm_storeu_ps(o3, vo3);
368 o3 += 4;
369 _mm_storeu_ps(o2, vo2);
370 o2 += 4;
371 _mm_storeu_ps(o1, vo1);
372 o1 += 4;
373 _mm_storeu_ps(o0, vo0);
374 o0 += 4;
375 } else {
376 if (w & (2 * sizeof(float))) {
377 _mm_storel_pi((__m64*) o5, vo5);
378 o5 += 2;
379 _mm_storel_pi((__m64*) o4, vo4);
380 o4 += 2;
381 _mm_storel_pi((__m64*) o3, vo3);
382 o3 += 2;
383 _mm_storel_pi((__m64*) o2, vo2);
384 o2 += 2;
385 _mm_storel_pi((__m64*) o1, vo1);
386 o1 += 2;
387 _mm_storel_pi((__m64*) o0, vo0);
388 o0 += 2;
389
390 vo0 = _mm_movehl_ps(vo0, vo0);
391 vo1 = _mm_movehl_ps(vo1, vo1);
392 vo2 = _mm_movehl_ps(vo2, vo2);
393 vo3 = _mm_movehl_ps(vo3, vo3);
394 vo4 = _mm_movehl_ps(vo4, vo4);
395 vo5 = _mm_movehl_ps(vo5, vo5);
396 }
397 if (w & (1 * sizeof(float))) {
398 _mm_store_ss(o5, vo5);
399 o5 += 1;
400 _mm_store_ss(o4, vo4);
401 o4 += 1;
402 _mm_store_ss(o3, vo3);
403 o3 += 1;
404 _mm_store_ss(o2, vo2);
405 o2 += 1;
406 _mm_store_ss(o1, vo1);
407 o1 += 1;
408 _mm_store_ss(o0, vo0);
409 o0 += 1;
410 }
411 }
412 }
413
414 i0 = (const float*) ((uintptr_t) i6 - input_decrement);
415 i1 = (const float*) ((uintptr_t) i7 - input_decrement);
416 i2 = (const float*) ((uintptr_t) i1 + input_width);
417 i3 = (const float*) ((uintptr_t) i2 + input_width);
418 i4 = (const float*) ((uintptr_t) i3 + input_width);
419 i5 = (const float*) ((uintptr_t) i4 + input_width);
420 i6 = (const float*) ((uintptr_t) i5 + input_width);
421 i7 = (const float*) ((uintptr_t) i6 + input_width);
422
423 o0 = o5;
424 o1 = (float*) ((uintptr_t) o0 + input_width);
425 o2 = (float*) ((uintptr_t) o1 + input_width);
426 o3 = (float*) ((uintptr_t) o2 + input_width);
427 o4 = (float*) ((uintptr_t) o3 + input_width);
428 o5 = (float*) ((uintptr_t) o4 + input_width);
429
430 output_height = doz(output_height, 6);
431 } while (output_height != 0);
432 }
433