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_3x4(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_3x4(
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
56 float* o0 = output;
57 float* o1 = (float*) ((uintptr_t) o0 + input_width);
58 float* o2 = (float*) ((uintptr_t) o1 + input_width);
59
60 size_t output_height = input_height;
61 do {
62 if XNN_UNPREDICTABLE(output_height < 2) {
63 i2 = zero;
64 o1 = o0;
65 }
66 if XNN_UNPREDICTABLE(output_height < 3) {
67 i3 = zero;
68 o2 = o1;
69 }
70 if XNN_UNPREDICTABLE(output_height < 4) {
71 i4 = zero;
72 }
73
74 __m128 vi0x0123 = _mm_setzero_ps();
75 __m128 vi1x0123 = _mm_setzero_ps();
76 __m128 vi2x0123 = _mm_setzero_ps();
77 __m128 vi3x0123 = _mm_setzero_ps();
78 __m128 vi4x0123 = _mm_setzero_ps();
79
80 __m128 vi0x4567 = _mm_loadu_ps(i0);
81 i0 += 4;
82 __m128 vi1x4567 = _mm_loadu_ps(i1);
83 i1 += 4;
84 __m128 vi2x4567 = _mm_loadu_ps(i2);
85 i2 += 4;
86 __m128 vi3x4567 = _mm_loadu_ps(i3);
87 i3 += 4;
88 __m128 vi4x4567 = _mm_loadu_ps(i4);
89 i4 += 4;
90
91 size_t w = input_width;
92 for (; w > 4 * sizeof(float); w -= 4 * sizeof(float)) {
93 const __m128 vi0x89AB = _mm_loadu_ps(i0);
94 i0 += 4;
95 const __m128 vi1x89AB = _mm_loadu_ps(i1);
96 i1 += 4;
97 const __m128 vi2x89AB = _mm_loadu_ps(i2);
98 i2 += 4;
99 const __m128 vi3x89AB = _mm_loadu_ps(i3);
100 i3 += 4;
101 const __m128 vi4x89AB = _mm_loadu_ps(i4);
102 i4 += 4;
103
104 __m128 vo0p0 = _mm_add_ps(vbias, _mm_mul_ps(vi0x4567, vk01));
105 __m128 vo1p0 = _mm_add_ps(vbias, _mm_mul_ps(vi1x4567, vk01));
106 __m128 vo2p0 = _mm_add_ps(vbias, _mm_mul_ps(vi2x4567, vk01));
107 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x4567, vk11));
108 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x4567, vk11));
109 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi3x4567, vk11));
110 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x4567, vk21));
111 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi3x4567, vk21));
112 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi4x4567, vk21));
113
114 const __m128 vi0x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi0x4567), _mm_castps_si128(vi0x0123), 12));
115 const __m128 vi1x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi1x4567), _mm_castps_si128(vi1x0123), 12));
116 const __m128 vi2x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi2x4567), _mm_castps_si128(vi2x0123), 12));
117 const __m128 vi3x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi3x4567), _mm_castps_si128(vi3x0123), 12));
118 const __m128 vi4x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi4x4567), _mm_castps_si128(vi4x0123), 12));
119
120 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi0x3456, vk00));
121 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi1x3456, vk00));
122 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi2x3456, vk00));
123 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x3456, vk10));
124 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x3456, vk10));
125 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi3x3456, vk10));
126 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x3456, vk20));
127 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi3x3456, vk20));
128 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi4x3456, vk20));
129
130 vi0x0123 = vi0x4567;
131 vi1x0123 = vi1x4567;
132 vi2x0123 = vi2x4567;
133 vi3x0123 = vi3x4567;
134 vi4x0123 = vi4x4567;
135
136 const __m128 vi0x5678 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi0x89AB), _mm_castps_si128(vi0x4567), 4));
137 const __m128 vi1x5678 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi1x89AB), _mm_castps_si128(vi1x4567), 4));
138 const __m128 vi2x5678 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi2x89AB), _mm_castps_si128(vi2x4567), 4));
139 const __m128 vi3x5678 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi3x89AB), _mm_castps_si128(vi3x4567), 4));
140 const __m128 vi4x5678 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi4x89AB), _mm_castps_si128(vi4x4567), 4));
141
142 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi0x5678, vk02));
143 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi1x5678, vk02));
144 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi2x5678, vk02));
145 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x5678, vk12));
146 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x5678, vk12));
147 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi3x5678, vk12));
148 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x5678, vk22));
149 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi3x5678, vk22));
150 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi4x5678, vk22));
151
152 vi0x4567 = vi0x89AB;
153 vi1x4567 = vi1x89AB;
154 vi2x4567 = vi2x89AB;
155 vi3x4567 = vi3x89AB;
156 vi4x4567 = vi4x89AB;
157
158
159 __m128 vo0 = _mm_max_ps(vo0p0, vmin);
160 __m128 vo1 = _mm_max_ps(vo1p0, vmin);
161 __m128 vo2 = _mm_max_ps(vo2p0, vmin);
162
163 vo0 = _mm_min_ps(vo0, vmax);
164 vo1 = _mm_min_ps(vo1, vmax);
165 vo2 = _mm_min_ps(vo2, vmax);
166
167 _mm_storeu_ps(o2, vo2);
168 o2 += 4;
169 _mm_storeu_ps(o1, vo1);
170 o1 += 4;
171 _mm_storeu_ps(o0, vo0);
172 o0 += 4;
173 }
174 // Always process the last block of 1..4 pixels.
175 assert(w >= 1 * sizeof(float));
176 assert(w <= 4 * sizeof(float));
177 {
178 vi0x4567 = _mm_and_ps(vmask, vi0x4567);
179 vi1x4567 = _mm_and_ps(vmask, vi1x4567);
180 vi2x4567 = _mm_and_ps(vmask, vi2x4567);
181 vi3x4567 = _mm_and_ps(vmask, vi3x4567);
182 vi4x4567 = _mm_and_ps(vmask, vi4x4567);
183
184 __m128 vo0p0 = _mm_add_ps(vbias, _mm_mul_ps(vi0x4567, vk01));
185 __m128 vo1p0 = _mm_add_ps(vbias, _mm_mul_ps(vi1x4567, vk01));
186 __m128 vo2p0 = _mm_add_ps(vbias, _mm_mul_ps(vi2x4567, vk01));
187 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x4567, vk11));
188 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x4567, vk11));
189 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi3x4567, vk11));
190 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x4567, vk21));
191 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi3x4567, vk21));
192 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi4x4567, vk21));
193
194 const __m128 vi0x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi0x4567), _mm_castps_si128(vi0x0123), 12));
195 const __m128 vi1x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi1x4567), _mm_castps_si128(vi1x0123), 12));
196 const __m128 vi2x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi2x4567), _mm_castps_si128(vi2x0123), 12));
197 const __m128 vi3x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi3x4567), _mm_castps_si128(vi3x0123), 12));
198 const __m128 vi4x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi4x4567), _mm_castps_si128(vi4x0123), 12));
199
200 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi0x3456, vk00));
201 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi1x3456, vk00));
202 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi2x3456, vk00));
203 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x3456, vk10));
204 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x3456, vk10));
205 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi3x3456, vk10));
206 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x3456, vk20));
207 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi3x3456, vk20));
208 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi4x3456, vk20));
209
210 const __m128i vzero = _mm_setzero_si128();
211 const __m128 vi0x5678 = _mm_castsi128_ps(_mm_alignr_epi8(vzero, _mm_castps_si128(vi0x4567), 4));
212 const __m128 vi1x5678 = _mm_castsi128_ps(_mm_alignr_epi8(vzero, _mm_castps_si128(vi1x4567), 4));
213 const __m128 vi2x5678 = _mm_castsi128_ps(_mm_alignr_epi8(vzero, _mm_castps_si128(vi2x4567), 4));
214 const __m128 vi3x5678 = _mm_castsi128_ps(_mm_alignr_epi8(vzero, _mm_castps_si128(vi3x4567), 4));
215 const __m128 vi4x5678 = _mm_castsi128_ps(_mm_alignr_epi8(vzero, _mm_castps_si128(vi4x4567), 4));
216
217 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi0x5678, vk02));
218 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi1x5678, vk02));
219 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi2x5678, vk02));
220 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x5678, vk12));
221 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x5678, vk12));
222 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi3x5678, vk12));
223 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x5678, vk22));
224 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi3x5678, vk22));
225 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi4x5678, vk22));
226
227
228 __m128 vo0 = _mm_max_ps(vo0p0, vmin);
229 __m128 vo1 = _mm_max_ps(vo1p0, vmin);
230 __m128 vo2 = _mm_max_ps(vo2p0, vmin);
231
232 vo0 = _mm_min_ps(vo0, vmax);
233 vo1 = _mm_min_ps(vo1, vmax);
234 vo2 = _mm_min_ps(vo2, vmax);
235
236 if XNN_LIKELY(w == 4 * sizeof(float)) {
237 _mm_storeu_ps(o2, vo2);
238 o2 += 4;
239 _mm_storeu_ps(o1, vo1);
240 o1 += 4;
241 _mm_storeu_ps(o0, vo0);
242 o0 += 4;
243 } else {
244 if (w & (2 * sizeof(float))) {
245 _mm_storel_pi((__m64*) o2, vo2);
246 o2 += 2;
247 _mm_storel_pi((__m64*) o1, vo1);
248 o1 += 2;
249 _mm_storel_pi((__m64*) o0, vo0);
250 o0 += 2;
251
252 vo0 = _mm_movehl_ps(vo0, vo0);
253 vo1 = _mm_movehl_ps(vo1, vo1);
254 vo2 = _mm_movehl_ps(vo2, vo2);
255 }
256 if (w & (1 * sizeof(float))) {
257 _mm_store_ss(o2, vo2);
258 o2 += 1;
259 _mm_store_ss(o1, vo1);
260 o1 += 1;
261 _mm_store_ss(o0, vo0);
262 o0 += 1;
263 }
264 }
265 }
266
267 i0 = (const float*) ((uintptr_t) i3 - input_decrement);
268 i1 = (const float*) ((uintptr_t) i4 - input_decrement);
269 i2 = (const float*) ((uintptr_t) i1 + input_width);
270 i3 = (const float*) ((uintptr_t) i2 + input_width);
271 i4 = (const float*) ((uintptr_t) i3 + input_width);
272
273 o0 = o2;
274 o1 = (float*) ((uintptr_t) o0 + input_width);
275 o2 = (float*) ((uintptr_t) o1 + input_width);
276
277 output_height = doz(output_height, 3);
278 } while (output_height != 0);
279 }
280