1 // Auto-generated file. Do not edit!
2 // Template: src/f32-dwconv2d-chw/3x3p1-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_3x3p1__sse_1x4_acc4(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__sse_1x4_acc4(
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
54 float* o0 = output;
55
56 size_t output_height = input_height;
57 do {
58 if XNN_UNPREDICTABLE(output_height < 2) {
59 i2 = zero;
60 }
61
62 // vi0x3012 = ( vi02, vi01, vi{M}0, vi{M}3 )
63 __m128 vi0x3012 = _mm_setzero_ps();
64 // vi1x3012 = ( vi12, vi11, vi{M}0, vi{M}3 )
65 __m128 vi1x3012 = _mm_setzero_ps();
66 // vi2x3012 = ( vi22, vi21, vi{M}0, vi{M}3 )
67 __m128 vi2x3012 = _mm_setzero_ps();
68
69 __m128 vi0x4567 = _mm_loadu_ps(i0);
70 i0 += 4;
71 __m128 vi1x4567 = _mm_loadu_ps(i1);
72 i1 += 4;
73 __m128 vi2x4567 = _mm_loadu_ps(i2);
74 i2 += 4;
75
76 size_t w = input_width;
77 for (; w > 4 * sizeof(float); w -= 4 * sizeof(float)) {
78 // vi0x89AB = ( vi0B, vi0A, vi09, vi08 )
79 const __m128 vi0x89AB = _mm_loadu_ps(i0);
80 i0 += 4;
81 // vi1x89AB = ( vi1B, vi1A, vi19, vi18 )
82 const __m128 vi1x89AB = _mm_loadu_ps(i1);
83 i1 += 4;
84 // vi2x89AB = ( vi2B, vi2A, vi29, vi28 )
85 const __m128 vi2x89AB = _mm_loadu_ps(i2);
86 i2 += 4;
87
88 // vi0x7456 = ( vi06, vi05, vi04, vi07 )
89 const __m128 vi0x7456 = _mm_shuffle_ps(vi0x4567, vi0x4567, _MM_SHUFFLE(2, 1, 0, 3));
90 // vi1x7456 = ( vi16, vi15, vi14, vi17 )
91 const __m128 vi1x7456 = _mm_shuffle_ps(vi1x4567, vi1x4567, _MM_SHUFFLE(2, 1, 0, 3));
92 // vi2x7456 = ( vi26, vi25, vi24, vi27 )
93 const __m128 vi2x7456 = _mm_shuffle_ps(vi2x4567, vi2x4567, _MM_SHUFFLE(2, 1, 0, 3));
94
95 __m128 vo0p0 = _mm_add_ps(vbias, _mm_mul_ps(vi0x4567, vk01));
96 __m128 vo0p1 = _mm_mul_ps(vi1x4567, vk11);
97 __m128 vo0p2 = _mm_mul_ps(vi2x4567, vk21);
98
99 // vi0x3456 = ( vi06, vi05, vi04, vi03 )
100 const __m128 vi0x3456 = _mm_move_ss(vi0x7456, vi0x3012);
101 // vi1x3456 = ( vi16, vi15, vi14, vi13 )
102 const __m128 vi1x3456 = _mm_move_ss(vi1x7456, vi1x3012);
103 // vi2x3456 = ( vi26, vi25, vi24, vi23 )
104 const __m128 vi2x3456 = _mm_move_ss(vi2x7456, vi2x3012);
105
106 __m128 vo0p3 = _mm_mul_ps(vi0x3456, vk00);
107 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x3456, vk10));
108 vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi2x3456, vk20));
109
110 vi0x3012 = vi0x7456;
111 vi1x3012 = vi1x7456;
112 vi2x3012 = vi2x7456;
113
114 // vi0x8567 = ( vi07, vi06, vi05, vi08 )
115 const __m128 vi0x8567 = _mm_move_ss(vi0x4567, vi0x89AB);
116 // vi1x8567 = ( vi17, vi16, vi15, vi18 )
117 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vi1x89AB);
118 // vi2x8567 = ( vi27, vi26, vi25, vi28 )
119 const __m128 vi2x8567 = _mm_move_ss(vi2x4567, vi2x89AB);
120
121 // vi0x5678 = ( vi08, vi07, vi06, vi05 )
122 const __m128 vi0x5678 = _mm_shuffle_ps(vi0x8567, vi0x8567, _MM_SHUFFLE(0, 3, 2, 1));
123 // vi1x5678 = ( vi18, vi17, vi16, vi15 )
124 const __m128 vi1x5678 = _mm_shuffle_ps(vi1x8567, vi1x8567, _MM_SHUFFLE(0, 3, 2, 1));
125 // vi2x5678 = ( vi28, vi27, vi26, vi25 )
126 const __m128 vi2x5678 = _mm_shuffle_ps(vi2x8567, vi2x8567, _MM_SHUFFLE(0, 3, 2, 1));
127
128 vo0p2 = _mm_add_ps(vo0p2, _mm_mul_ps(vi0x5678, vk02));
129 vo0p3 = _mm_add_ps(vo0p3, _mm_mul_ps(vi1x5678, vk12));
130 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x5678, vk22));
131
132 vi0x4567 = vi0x89AB;
133 vi1x4567 = vi1x89AB;
134 vi2x4567 = vi2x89AB;
135
136 vo0p0 = _mm_add_ps(vo0p0, vo0p1);
137 vo0p2 = _mm_add_ps(vo0p2, vo0p3);
138 vo0p0 = _mm_add_ps(vo0p0, vo0p2);
139
140 __m128 vo0 = _mm_max_ps(vo0p0, vmin);
141
142 vo0 = _mm_min_ps(vo0, vmax);
143
144 _mm_storeu_ps(o0, vo0);
145 o0 += 4;
146 }
147 // Always process the last block of 1..4 pixels.
148 assert(w >= 1 * sizeof(float));
149 assert(w <= 4 * sizeof(float));
150 {
151 vi0x4567 = _mm_and_ps(vmask, vi0x4567);
152 vi1x4567 = _mm_and_ps(vmask, vi1x4567);
153 vi2x4567 = _mm_and_ps(vmask, vi2x4567);
154
155 // vi0x7456 = ( vi06, vi05, vi04, vi07 )
156 const __m128 vi0x7456 = _mm_shuffle_ps(vi0x4567, vi0x4567, _MM_SHUFFLE(2, 1, 0, 3));
157 // vi1x7456 = ( vi16, vi15, vi14, vi17 )
158 const __m128 vi1x7456 = _mm_shuffle_ps(vi1x4567, vi1x4567, _MM_SHUFFLE(2, 1, 0, 3));
159 // vi2x7456 = ( vi26, vi25, vi24, vi27 )
160 const __m128 vi2x7456 = _mm_shuffle_ps(vi2x4567, vi2x4567, _MM_SHUFFLE(2, 1, 0, 3));
161
162 __m128 vo0p0 = _mm_add_ps(vbias, _mm_mul_ps(vi0x4567, vk01));
163 __m128 vo0p1 = _mm_mul_ps(vi1x4567, vk11);
164 __m128 vo0p2 = _mm_mul_ps(vi2x4567, vk21);
165
166 // vi0x3456 = ( vi06, vi05, vi04, vi03 )
167 const __m128 vi0x3456 = _mm_move_ss(vi0x7456, vi0x3012);
168 // vi1x3456 = ( vi16, vi15, vi14, vi13 )
169 const __m128 vi1x3456 = _mm_move_ss(vi1x7456, vi1x3012);
170 // vi2x3456 = ( vi26, vi25, vi24, vi23 )
171 const __m128 vi2x3456 = _mm_move_ss(vi2x7456, vi2x3012);
172
173 __m128 vo0p3 = _mm_mul_ps(vi0x3456, vk00);
174 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x3456, vk10));
175 vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi2x3456, vk20));
176
177 const __m128 vzero = _mm_setzero_ps();
178 // vi0x8567 = ( vi07, vi06, vi05, 0.0 )
179 const __m128 vi0x8567 = _mm_move_ss(vi0x4567, vzero);
180 // vi1x8567 = ( vi17, vi16, vi15, 0.0 )
181 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vzero);
182 // vi2x8567 = ( vi27, vi26, vi25, 0.0 )
183 const __m128 vi2x8567 = _mm_move_ss(vi2x4567, vzero);
184
185 // vi0x5678 = ( vi08, vi07, vi06, vi05 )
186 const __m128 vi0x5678 = _mm_shuffle_ps(vi0x8567, vi0x8567, _MM_SHUFFLE(0, 3, 2, 1));
187 // vi1x5678 = ( vi18, vi17, vi16, vi15 )
188 const __m128 vi1x5678 = _mm_shuffle_ps(vi1x8567, vi1x8567, _MM_SHUFFLE(0, 3, 2, 1));
189 // vi2x5678 = ( vi28, vi27, vi26, vi25 )
190 const __m128 vi2x5678 = _mm_shuffle_ps(vi2x8567, vi2x8567, _MM_SHUFFLE(0, 3, 2, 1));
191
192 vo0p2 = _mm_add_ps(vo0p2, _mm_mul_ps(vi0x5678, vk02));
193 vo0p3 = _mm_add_ps(vo0p3, _mm_mul_ps(vi1x5678, vk12));
194 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x5678, vk22));
195
196 vo0p0 = _mm_add_ps(vo0p0, vo0p1);
197 vo0p2 = _mm_add_ps(vo0p2, vo0p3);
198 vo0p0 = _mm_add_ps(vo0p0, vo0p2);
199
200 __m128 vo0 = _mm_max_ps(vo0p0, vmin);
201
202 vo0 = _mm_min_ps(vo0, vmax);
203
204 if XNN_LIKELY(w == 4 * sizeof(float)) {
205 _mm_storeu_ps(o0, vo0);
206 o0 += 4;
207 } else {
208 if (w & (2 * sizeof(float))) {
209 _mm_storel_pi((__m64*) o0, vo0);
210 o0 += 2;
211
212 vo0 = _mm_movehl_ps(vo0, vo0);
213 }
214 if (w & (1 * sizeof(float))) {
215 _mm_store_ss(o0, vo0);
216 o0 += 1;
217 }
218 }
219 }
220
221 i0 = (const float*) ((uintptr_t) i1 - input_decrement);
222 i1 = (const float*) ((uintptr_t) i2 - input_decrement);
223 i2 = (const float*) ((uintptr_t) i1 + input_width);
224
225
226 } while (--output_height != 0);
227 }
228