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_5x4(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_5x4(
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)])
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
58 float* o0 = output;
59 float* o1 = (float*) ((uintptr_t) o0 + input_width);
60 float* o2 = (float*) ((uintptr_t) o1 + input_width);
61 float* o3 = (float*) ((uintptr_t) o2 + input_width);
62 float* o4 = (float*) ((uintptr_t) o3 + input_width);
63
64 size_t output_height = input_height;
65 do {
66 if XNN_UNPREDICTABLE(output_height < 2) {
67 i2 = zero;
68 o1 = o0;
69 }
70 if XNN_UNPREDICTABLE(output_height < 3) {
71 i3 = zero;
72 o2 = o1;
73 }
74 if XNN_UNPREDICTABLE(output_height < 4) {
75 i4 = zero;
76 o3 = o2;
77 }
78 if XNN_UNPREDICTABLE(output_height < 5) {
79 i5 = zero;
80 o4 = o3;
81 }
82 if XNN_UNPREDICTABLE(output_height < 6) {
83 i6 = zero;
84 }
85
86 __m128 vi0x0123 = _mm_setzero_ps();
87 __m128 vi1x0123 = _mm_setzero_ps();
88 __m128 vi2x0123 = _mm_setzero_ps();
89 __m128 vi3x0123 = _mm_setzero_ps();
90 __m128 vi4x0123 = _mm_setzero_ps();
91 __m128 vi5x0123 = _mm_setzero_ps();
92 __m128 vi6x0123 = _mm_setzero_ps();
93
94 __m128 vi0x4567 = _mm_loadu_ps(i0);
95 i0 += 4;
96 __m128 vi1x4567 = _mm_loadu_ps(i1);
97 i1 += 4;
98 __m128 vi2x4567 = _mm_loadu_ps(i2);
99 i2 += 4;
100 __m128 vi3x4567 = _mm_loadu_ps(i3);
101 i3 += 4;
102 __m128 vi4x4567 = _mm_loadu_ps(i4);
103 i4 += 4;
104 __m128 vi5x4567 = _mm_loadu_ps(i5);
105 i5 += 4;
106 __m128 vi6x4567 = _mm_loadu_ps(i6);
107 i6 += 4;
108
109 size_t w = input_width;
110 for (; w > 4 * sizeof(float); w -= 4 * sizeof(float)) {
111 const __m128 vi0x89AB = _mm_loadu_ps(i0);
112 i0 += 4;
113 const __m128 vi1x89AB = _mm_loadu_ps(i1);
114 i1 += 4;
115 const __m128 vi2x89AB = _mm_loadu_ps(i2);
116 i2 += 4;
117 const __m128 vi3x89AB = _mm_loadu_ps(i3);
118 i3 += 4;
119 const __m128 vi4x89AB = _mm_loadu_ps(i4);
120 i4 += 4;
121 const __m128 vi5x89AB = _mm_loadu_ps(i5);
122 i5 += 4;
123 const __m128 vi6x89AB = _mm_loadu_ps(i6);
124 i6 += 4;
125
126 __m128 vo0p0 = _mm_add_ps(vbias, _mm_mul_ps(vi0x4567, vk01));
127 __m128 vo1p0 = _mm_add_ps(vbias, _mm_mul_ps(vi1x4567, vk01));
128 __m128 vo2p0 = _mm_add_ps(vbias, _mm_mul_ps(vi2x4567, vk01));
129 __m128 vo3p0 = _mm_add_ps(vbias, _mm_mul_ps(vi3x4567, vk01));
130 __m128 vo4p0 = _mm_add_ps(vbias, _mm_mul_ps(vi4x4567, vk01));
131 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x4567, vk11));
132 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x4567, vk11));
133 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi3x4567, vk11));
134 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi4x4567, vk11));
135 vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi5x4567, vk11));
136 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x4567, vk21));
137 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi3x4567, vk21));
138 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi4x4567, vk21));
139 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi5x4567, vk21));
140 vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi6x4567, vk21));
141
142 const __m128 vi0x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi0x4567), _mm_castps_si128(vi0x0123), 12));
143 const __m128 vi1x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi1x4567), _mm_castps_si128(vi1x0123), 12));
144 const __m128 vi2x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi2x4567), _mm_castps_si128(vi2x0123), 12));
145 const __m128 vi3x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi3x4567), _mm_castps_si128(vi3x0123), 12));
146 const __m128 vi4x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi4x4567), _mm_castps_si128(vi4x0123), 12));
147 const __m128 vi5x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi5x4567), _mm_castps_si128(vi5x0123), 12));
148 const __m128 vi6x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi6x4567), _mm_castps_si128(vi6x0123), 12));
149
150 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi0x3456, vk00));
151 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi1x3456, vk00));
152 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi2x3456, vk00));
153 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi3x3456, vk00));
154 vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi4x3456, vk00));
155 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x3456, vk10));
156 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x3456, vk10));
157 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi3x3456, vk10));
158 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi4x3456, vk10));
159 vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi5x3456, vk10));
160 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x3456, vk20));
161 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi3x3456, vk20));
162 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi4x3456, vk20));
163 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi5x3456, vk20));
164 vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi6x3456, vk20));
165
166 vi0x0123 = vi0x4567;
167 vi1x0123 = vi1x4567;
168 vi2x0123 = vi2x4567;
169 vi3x0123 = vi3x4567;
170 vi4x0123 = vi4x4567;
171 vi5x0123 = vi5x4567;
172 vi6x0123 = vi6x4567;
173
174 const __m128 vi0x5678 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi0x89AB), _mm_castps_si128(vi0x4567), 4));
175 const __m128 vi1x5678 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi1x89AB), _mm_castps_si128(vi1x4567), 4));
176 const __m128 vi2x5678 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi2x89AB), _mm_castps_si128(vi2x4567), 4));
177 const __m128 vi3x5678 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi3x89AB), _mm_castps_si128(vi3x4567), 4));
178 const __m128 vi4x5678 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi4x89AB), _mm_castps_si128(vi4x4567), 4));
179 const __m128 vi5x5678 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi5x89AB), _mm_castps_si128(vi5x4567), 4));
180 const __m128 vi6x5678 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi6x89AB), _mm_castps_si128(vi6x4567), 4));
181
182 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi0x5678, vk02));
183 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi1x5678, vk02));
184 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi2x5678, vk02));
185 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi3x5678, vk02));
186 vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi4x5678, vk02));
187 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x5678, vk12));
188 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x5678, vk12));
189 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi3x5678, vk12));
190 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi4x5678, vk12));
191 vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi5x5678, vk12));
192 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x5678, vk22));
193 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi3x5678, vk22));
194 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi4x5678, vk22));
195 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi5x5678, vk22));
196 vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi6x5678, vk22));
197
198 vi0x4567 = vi0x89AB;
199 vi1x4567 = vi1x89AB;
200 vi2x4567 = vi2x89AB;
201 vi3x4567 = vi3x89AB;
202 vi4x4567 = vi4x89AB;
203 vi5x4567 = vi5x89AB;
204 vi6x4567 = vi6x89AB;
205
206
207 __m128 vo0 = _mm_max_ps(vo0p0, vmin);
208 __m128 vo1 = _mm_max_ps(vo1p0, vmin);
209 __m128 vo2 = _mm_max_ps(vo2p0, vmin);
210 __m128 vo3 = _mm_max_ps(vo3p0, vmin);
211 __m128 vo4 = _mm_max_ps(vo4p0, vmin);
212
213 vo0 = _mm_min_ps(vo0, vmax);
214 vo1 = _mm_min_ps(vo1, vmax);
215 vo2 = _mm_min_ps(vo2, vmax);
216 vo3 = _mm_min_ps(vo3, vmax);
217 vo4 = _mm_min_ps(vo4, vmax);
218
219 _mm_storeu_ps(o4, vo4);
220 o4 += 4;
221 _mm_storeu_ps(o3, vo3);
222 o3 += 4;
223 _mm_storeu_ps(o2, vo2);
224 o2 += 4;
225 _mm_storeu_ps(o1, vo1);
226 o1 += 4;
227 _mm_storeu_ps(o0, vo0);
228 o0 += 4;
229 }
230 // Always process the last block of 1..4 pixels.
231 assert(w >= 1 * sizeof(float));
232 assert(w <= 4 * sizeof(float));
233 {
234 vi0x4567 = _mm_and_ps(vmask, vi0x4567);
235 vi1x4567 = _mm_and_ps(vmask, vi1x4567);
236 vi2x4567 = _mm_and_ps(vmask, vi2x4567);
237 vi3x4567 = _mm_and_ps(vmask, vi3x4567);
238 vi4x4567 = _mm_and_ps(vmask, vi4x4567);
239 vi5x4567 = _mm_and_ps(vmask, vi5x4567);
240 vi6x4567 = _mm_and_ps(vmask, vi6x4567);
241
242 __m128 vo0p0 = _mm_add_ps(vbias, _mm_mul_ps(vi0x4567, vk01));
243 __m128 vo1p0 = _mm_add_ps(vbias, _mm_mul_ps(vi1x4567, vk01));
244 __m128 vo2p0 = _mm_add_ps(vbias, _mm_mul_ps(vi2x4567, vk01));
245 __m128 vo3p0 = _mm_add_ps(vbias, _mm_mul_ps(vi3x4567, vk01));
246 __m128 vo4p0 = _mm_add_ps(vbias, _mm_mul_ps(vi4x4567, vk01));
247 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x4567, vk11));
248 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x4567, vk11));
249 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi3x4567, vk11));
250 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi4x4567, vk11));
251 vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi5x4567, vk11));
252 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x4567, vk21));
253 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi3x4567, vk21));
254 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi4x4567, vk21));
255 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi5x4567, vk21));
256 vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi6x4567, vk21));
257
258 const __m128 vi0x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi0x4567), _mm_castps_si128(vi0x0123), 12));
259 const __m128 vi1x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi1x4567), _mm_castps_si128(vi1x0123), 12));
260 const __m128 vi2x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi2x4567), _mm_castps_si128(vi2x0123), 12));
261 const __m128 vi3x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi3x4567), _mm_castps_si128(vi3x0123), 12));
262 const __m128 vi4x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi4x4567), _mm_castps_si128(vi4x0123), 12));
263 const __m128 vi5x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi5x4567), _mm_castps_si128(vi5x0123), 12));
264 const __m128 vi6x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi6x4567), _mm_castps_si128(vi6x0123), 12));
265
266 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi0x3456, vk00));
267 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi1x3456, vk00));
268 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi2x3456, vk00));
269 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi3x3456, vk00));
270 vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi4x3456, vk00));
271 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x3456, vk10));
272 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x3456, vk10));
273 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi3x3456, vk10));
274 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi4x3456, vk10));
275 vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi5x3456, vk10));
276 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x3456, vk20));
277 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi3x3456, vk20));
278 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi4x3456, vk20));
279 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi5x3456, vk20));
280 vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi6x3456, vk20));
281
282 const __m128i vzero = _mm_setzero_si128();
283 const __m128 vi0x5678 = _mm_castsi128_ps(_mm_alignr_epi8(vzero, _mm_castps_si128(vi0x4567), 4));
284 const __m128 vi1x5678 = _mm_castsi128_ps(_mm_alignr_epi8(vzero, _mm_castps_si128(vi1x4567), 4));
285 const __m128 vi2x5678 = _mm_castsi128_ps(_mm_alignr_epi8(vzero, _mm_castps_si128(vi2x4567), 4));
286 const __m128 vi3x5678 = _mm_castsi128_ps(_mm_alignr_epi8(vzero, _mm_castps_si128(vi3x4567), 4));
287 const __m128 vi4x5678 = _mm_castsi128_ps(_mm_alignr_epi8(vzero, _mm_castps_si128(vi4x4567), 4));
288 const __m128 vi5x5678 = _mm_castsi128_ps(_mm_alignr_epi8(vzero, _mm_castps_si128(vi5x4567), 4));
289 const __m128 vi6x5678 = _mm_castsi128_ps(_mm_alignr_epi8(vzero, _mm_castps_si128(vi6x4567), 4));
290
291 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi0x5678, vk02));
292 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi1x5678, vk02));
293 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi2x5678, vk02));
294 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi3x5678, vk02));
295 vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi4x5678, vk02));
296 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x5678, vk12));
297 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x5678, vk12));
298 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi3x5678, vk12));
299 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi4x5678, vk12));
300 vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi5x5678, vk12));
301 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x5678, vk22));
302 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi3x5678, vk22));
303 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi4x5678, vk22));
304 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi5x5678, vk22));
305 vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi6x5678, vk22));
306
307
308 __m128 vo0 = _mm_max_ps(vo0p0, vmin);
309 __m128 vo1 = _mm_max_ps(vo1p0, vmin);
310 __m128 vo2 = _mm_max_ps(vo2p0, vmin);
311 __m128 vo3 = _mm_max_ps(vo3p0, vmin);
312 __m128 vo4 = _mm_max_ps(vo4p0, vmin);
313
314 vo0 = _mm_min_ps(vo0, vmax);
315 vo1 = _mm_min_ps(vo1, vmax);
316 vo2 = _mm_min_ps(vo2, vmax);
317 vo3 = _mm_min_ps(vo3, vmax);
318 vo4 = _mm_min_ps(vo4, vmax);
319
320 if XNN_LIKELY(w == 4 * sizeof(float)) {
321 _mm_storeu_ps(o4, vo4);
322 o4 += 4;
323 _mm_storeu_ps(o3, vo3);
324 o3 += 4;
325 _mm_storeu_ps(o2, vo2);
326 o2 += 4;
327 _mm_storeu_ps(o1, vo1);
328 o1 += 4;
329 _mm_storeu_ps(o0, vo0);
330 o0 += 4;
331 } else {
332 if (w & (2 * sizeof(float))) {
333 _mm_storel_pi((__m64*) o4, vo4);
334 o4 += 2;
335 _mm_storel_pi((__m64*) o3, vo3);
336 o3 += 2;
337 _mm_storel_pi((__m64*) o2, vo2);
338 o2 += 2;
339 _mm_storel_pi((__m64*) o1, vo1);
340 o1 += 2;
341 _mm_storel_pi((__m64*) o0, vo0);
342 o0 += 2;
343
344 vo0 = _mm_movehl_ps(vo0, vo0);
345 vo1 = _mm_movehl_ps(vo1, vo1);
346 vo2 = _mm_movehl_ps(vo2, vo2);
347 vo3 = _mm_movehl_ps(vo3, vo3);
348 vo4 = _mm_movehl_ps(vo4, vo4);
349 }
350 if (w & (1 * sizeof(float))) {
351 _mm_store_ss(o4, vo4);
352 o4 += 1;
353 _mm_store_ss(o3, vo3);
354 o3 += 1;
355 _mm_store_ss(o2, vo2);
356 o2 += 1;
357 _mm_store_ss(o1, vo1);
358 o1 += 1;
359 _mm_store_ss(o0, vo0);
360 o0 += 1;
361 }
362 }
363 }
364
365 i0 = (const float*) ((uintptr_t) i5 - input_decrement);
366 i1 = (const float*) ((uintptr_t) i6 - input_decrement);
367 i2 = (const float*) ((uintptr_t) i1 + input_width);
368 i3 = (const float*) ((uintptr_t) i2 + input_width);
369 i4 = (const float*) ((uintptr_t) i3 + input_width);
370 i5 = (const float*) ((uintptr_t) i4 + input_width);
371 i6 = (const float*) ((uintptr_t) i5 + input_width);
372
373 o0 = o4;
374 o1 = (float*) ((uintptr_t) o0 + input_width);
375 o2 = (float*) ((uintptr_t) o1 + input_width);
376 o3 = (float*) ((uintptr_t) o2 + input_width);
377 o4 = (float*) ((uintptr_t) o3 + input_width);
378
379 output_height = doz(output_height, 5);
380 } while (output_height != 0);
381 }
382