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_2x4_acc2(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_2x4_acc2(
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
55 float* o0 = output;
56 float* o1 = (float*) ((uintptr_t) o0 + input_width);
57
58 size_t output_height = input_height;
59 do {
60 if XNN_UNPREDICTABLE(output_height < 2) {
61 i2 = zero;
62 o1 = o0;
63 }
64 if XNN_UNPREDICTABLE(output_height < 3) {
65 i3 = zero;
66 }
67
68 __m128 vi0x0123 = _mm_setzero_ps();
69 __m128 vi1x0123 = _mm_setzero_ps();
70 __m128 vi2x0123 = _mm_setzero_ps();
71 __m128 vi3x0123 = _mm_setzero_ps();
72
73 __m128 vi0x4567 = _mm_loadu_ps(i0);
74 i0 += 4;
75 __m128 vi1x4567 = _mm_loadu_ps(i1);
76 i1 += 4;
77 __m128 vi2x4567 = _mm_loadu_ps(i2);
78 i2 += 4;
79 __m128 vi3x4567 = _mm_loadu_ps(i3);
80 i3 += 4;
81
82 size_t w = input_width;
83 for (; w > 4 * sizeof(float); w -= 4 * sizeof(float)) {
84 const __m128 vi0x89AB = _mm_loadu_ps(i0);
85 i0 += 4;
86 const __m128 vi1x89AB = _mm_loadu_ps(i1);
87 i1 += 4;
88 const __m128 vi2x89AB = _mm_loadu_ps(i2);
89 i2 += 4;
90 const __m128 vi3x89AB = _mm_loadu_ps(i3);
91 i3 += 4;
92
93 __m128 vo0p0 = _mm_add_ps(vbias, _mm_mul_ps(vi0x4567, vk01));
94 __m128 vo1p0 = _mm_add_ps(vbias, _mm_mul_ps(vi1x4567, vk01));
95 __m128 vo0p1 = _mm_mul_ps(vi1x4567, vk11);
96 __m128 vo1p1 = _mm_mul_ps(vi2x4567, vk11);
97 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x4567, vk21));
98 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi3x4567, vk21));
99
100 const __m128 vi0x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi0x4567), _mm_castps_si128(vi0x0123), 12));
101 const __m128 vi1x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi1x4567), _mm_castps_si128(vi1x0123), 12));
102 const __m128 vi2x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi2x4567), _mm_castps_si128(vi2x0123), 12));
103 const __m128 vi3x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi3x4567), _mm_castps_si128(vi3x0123), 12));
104
105 vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi0x3456, vk00));
106 vo1p1 = _mm_add_ps(vo1p1, _mm_mul_ps(vi1x3456, vk00));
107 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x3456, vk10));
108 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x3456, vk10));
109 vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi2x3456, vk20));
110 vo1p1 = _mm_add_ps(vo1p1, _mm_mul_ps(vi3x3456, vk20));
111
112 vi0x0123 = vi0x4567;
113 vi1x0123 = vi1x4567;
114 vi2x0123 = vi2x4567;
115 vi3x0123 = vi3x4567;
116
117 const __m128 vi0x5678 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi0x89AB), _mm_castps_si128(vi0x4567), 4));
118 const __m128 vi1x5678 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi1x89AB), _mm_castps_si128(vi1x4567), 4));
119 const __m128 vi2x5678 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi2x89AB), _mm_castps_si128(vi2x4567), 4));
120 const __m128 vi3x5678 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi3x89AB), _mm_castps_si128(vi3x4567), 4));
121
122 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi0x5678, vk02));
123 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi1x5678, vk02));
124 vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi1x5678, vk12));
125 vo1p1 = _mm_add_ps(vo1p1, _mm_mul_ps(vi2x5678, vk12));
126 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x5678, vk22));
127 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi3x5678, vk22));
128
129 vi0x4567 = vi0x89AB;
130 vi1x4567 = vi1x89AB;
131 vi2x4567 = vi2x89AB;
132 vi3x4567 = vi3x89AB;
133
134 vo0p0 = _mm_add_ps(vo0p0, vo0p1);
135 vo1p0 = _mm_add_ps(vo1p0, vo1p1);
136
137 __m128 vo0 = _mm_max_ps(vo0p0, vmin);
138 __m128 vo1 = _mm_max_ps(vo1p0, vmin);
139
140 vo0 = _mm_min_ps(vo0, vmax);
141 vo1 = _mm_min_ps(vo1, vmax);
142
143 _mm_storeu_ps(o1, vo1);
144 o1 += 4;
145 _mm_storeu_ps(o0, vo0);
146 o0 += 4;
147 }
148 // Always process the last block of 1..4 pixels.
149 assert(w >= 1 * sizeof(float));
150 assert(w <= 4 * sizeof(float));
151 {
152 vi0x4567 = _mm_and_ps(vmask, vi0x4567);
153 vi1x4567 = _mm_and_ps(vmask, vi1x4567);
154 vi2x4567 = _mm_and_ps(vmask, vi2x4567);
155 vi3x4567 = _mm_and_ps(vmask, vi3x4567);
156
157 __m128 vo0p0 = _mm_add_ps(vbias, _mm_mul_ps(vi0x4567, vk01));
158 __m128 vo1p0 = _mm_add_ps(vbias, _mm_mul_ps(vi1x4567, vk01));
159 __m128 vo0p1 = _mm_mul_ps(vi1x4567, vk11);
160 __m128 vo1p1 = _mm_mul_ps(vi2x4567, vk11);
161 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x4567, vk21));
162 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi3x4567, vk21));
163
164 const __m128 vi0x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi0x4567), _mm_castps_si128(vi0x0123), 12));
165 const __m128 vi1x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi1x4567), _mm_castps_si128(vi1x0123), 12));
166 const __m128 vi2x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi2x4567), _mm_castps_si128(vi2x0123), 12));
167 const __m128 vi3x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi3x4567), _mm_castps_si128(vi3x0123), 12));
168
169 vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi0x3456, vk00));
170 vo1p1 = _mm_add_ps(vo1p1, _mm_mul_ps(vi1x3456, vk00));
171 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x3456, vk10));
172 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x3456, vk10));
173 vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi2x3456, vk20));
174 vo1p1 = _mm_add_ps(vo1p1, _mm_mul_ps(vi3x3456, vk20));
175
176 const __m128i vzero = _mm_setzero_si128();
177 const __m128 vi0x5678 = _mm_castsi128_ps(_mm_alignr_epi8(vzero, _mm_castps_si128(vi0x4567), 4));
178 const __m128 vi1x5678 = _mm_castsi128_ps(_mm_alignr_epi8(vzero, _mm_castps_si128(vi1x4567), 4));
179 const __m128 vi2x5678 = _mm_castsi128_ps(_mm_alignr_epi8(vzero, _mm_castps_si128(vi2x4567), 4));
180 const __m128 vi3x5678 = _mm_castsi128_ps(_mm_alignr_epi8(vzero, _mm_castps_si128(vi3x4567), 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 vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi1x5678, vk12));
185 vo1p1 = _mm_add_ps(vo1p1, _mm_mul_ps(vi2x5678, vk12));
186 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x5678, vk22));
187 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi3x5678, vk22));
188
189 vo0p0 = _mm_add_ps(vo0p0, vo0p1);
190 vo1p0 = _mm_add_ps(vo1p0, vo1p1);
191
192 __m128 vo0 = _mm_max_ps(vo0p0, vmin);
193 __m128 vo1 = _mm_max_ps(vo1p0, vmin);
194
195 vo0 = _mm_min_ps(vo0, vmax);
196 vo1 = _mm_min_ps(vo1, vmax);
197
198 if XNN_LIKELY(w == 4 * sizeof(float)) {
199 _mm_storeu_ps(o1, vo1);
200 o1 += 4;
201 _mm_storeu_ps(o0, vo0);
202 o0 += 4;
203 } else {
204 if (w & (2 * sizeof(float))) {
205 _mm_storel_pi((__m64*) o1, vo1);
206 o1 += 2;
207 _mm_storel_pi((__m64*) o0, vo0);
208 o0 += 2;
209
210 vo0 = _mm_movehl_ps(vo0, vo0);
211 vo1 = _mm_movehl_ps(vo1, vo1);
212 }
213 if (w & (1 * sizeof(float))) {
214 _mm_store_ss(o1, vo1);
215 o1 += 1;
216 _mm_store_ss(o0, vo0);
217 o0 += 1;
218 }
219 }
220 }
221
222 i0 = (const float*) ((uintptr_t) i2 - input_decrement);
223 i1 = (const float*) ((uintptr_t) i3 - input_decrement);
224 i2 = (const float*) ((uintptr_t) i1 + input_width);
225 i3 = (const float*) ((uintptr_t) i2 + input_width);
226
227 o0 = o1;
228 o1 = (float*) ((uintptr_t) o0 + input_width);
229
230 output_height = doz(output_height, 2);
231 } while (output_height != 0);
232 }
233