1 // Copyright 2019 Google LLC
2 //
3 // This source code is licensed under the BSD-style license found in the
4 // LICENSE file in the root directory of this source tree.
5
6 #include <assert.h>
7
8 #include <xmmintrin.h>
9
10 #include <xnnpack/dwconv.h>
11 #include <xnnpack/math.h>
12
13
xnn_f32_dwconv_spchw_ukernel_3x3p1__sse(size_t m,size_t n,const float * input,const float * weights,float * output,size_t input_tuple_stride,size_t output_tuple_stride,size_t input_width_stride,size_t output_width_stride,const union xnn_f32_spchw_params params[restrict static1])14 void xnn_f32_dwconv_spchw_ukernel_3x3p1__sse(
15 size_t m,
16 size_t n,
17 const float* input,
18 const float* weights,
19 float* output,
20 size_t input_tuple_stride,
21 size_t output_tuple_stride,
22 size_t input_width_stride,
23 size_t output_width_stride,
24 const union xnn_f32_spchw_params params[restrict static 1])
25 {
26 assert(n != 0);
27
28 const __m128 vmask = _mm_load_ps((const float*) params->sse.mask);
29 const __m128 vmax = _mm_load_ps(params->sse.max);
30 const __m128 vmin = _mm_load_ps(params->sse.min);
31
32 const size_t input_width_increment = input_width_stride - round_up_po2(n, 4) / 4 * input_tuple_stride;
33 const size_t output_width_increment = output_width_stride - (n - 1) / 4 * output_tuple_stride;
34
35 // No vertical padding.
36 const float* i0 = input;
37 const float* i1 = (const float*) ((uintptr_t) i0 + input_width_stride);
38 const float* i2 = (const float*) ((uintptr_t) i1 + input_width_stride);
39
40 const __m128 vbias = _mm_load1_ps(weights);
41 const __m128 vk00 = _mm_load1_ps(weights + 1);
42 const __m128 vk01 = _mm_load1_ps(weights + 2);
43 const __m128 vk02 = _mm_load1_ps(weights + 3);
44 const __m128 vk10 = _mm_load1_ps(weights + 4);
45 const __m128 vk11 = _mm_load1_ps(weights + 5);
46 const __m128 vk12 = _mm_load1_ps(weights + 6);
47 const __m128 vk20 = _mm_load1_ps(weights + 7);
48 const __m128 vk21 = _mm_load1_ps(weights + 8);
49 const __m128 vk22 = _mm_load1_ps(weights + 9);
50
51 do {
52 // vi0x3012 = ( vi02, vi01, vi00, vi03 )
53 __m128 vi0x3012 = _mm_setzero_ps();
54 // vi1x3012 = ( vi12, vi11, vi10, vi13 )
55 __m128 vi1x3012 = _mm_setzero_ps();
56 // vi2x3012 = ( vi22, vi21, vi20, vi13 )
57 __m128 vi2x3012 = _mm_setzero_ps();
58 // vi0x4567 = ( vi07, vi06, vi05, vi04 )
59 __m128 vi0x4567 = _mm_loadu_ps(i0);
60 i0 = (const float*) ((uintptr_t) i0 + input_tuple_stride);
61 // vi1x4567 = ( vi17, vi16, vi15, vi14 )
62 __m128 vi1x4567 = _mm_loadu_ps(i1);
63 i1 = (const float*) ((uintptr_t) i1 + input_tuple_stride);
64 // vi2x4567 = ( vi27, vi26, vi25, vi24 )
65 __m128 vi2x4567 = _mm_loadu_ps(i2);
66 i2 = (const float*) ((uintptr_t) i2 + input_tuple_stride);
67
68 size_t k = n;
69 for (; k > 4; k -= 4) {
70 __m128 vo4567p0 = vbias;
71
72 // vi0x89AB = ( vi0B, vi0A, vi09, vi08 )
73 const __m128 vi0x89AB = _mm_loadu_ps(i0);
74 i0 = (const float*) ((uintptr_t) i0 + input_tuple_stride);
75 // vi1x89AB = ( vi1B, vi0A, vi09, vi08 )
76 const __m128 vi1x89AB = _mm_loadu_ps(i1);
77 i1 = (const float*) ((uintptr_t) i1 + input_tuple_stride);
78 // vi2x89AB = ( vi2B, vi0A, vi09, vi08 )
79 const __m128 vi2x89AB = _mm_loadu_ps(i2);
80 i2 = (const float*) ((uintptr_t) i2 + input_tuple_stride);
81
82 // vi0x7456 = ( vi06, vi05, vi04, vi07 )
83 const __m128 vi0x7456 = _mm_shuffle_ps(vi0x4567, vi0x4567, _MM_SHUFFLE(2, 1, 0, 3));
84 // vi1x7456 = ( vi16, vi15, vi14, vi17 )
85 const __m128 vi1x7456 = _mm_shuffle_ps(vi1x4567, vi1x4567, _MM_SHUFFLE(2, 1, 0, 3));
86 // vi2x7456 = ( vi26, vi25, vi24, vi27 )
87 const __m128 vi2x7456 = _mm_shuffle_ps(vi2x4567, vi2x4567, _MM_SHUFFLE(2, 1, 0, 3));
88
89 vo4567p0 = _mm_add_ps(vo4567p0, _mm_mul_ps(vi0x4567, vk01));
90 __m128 vo4567p1 = _mm_mul_ps(vi1x4567, vk11);
91 __m128 vo4567p2 = _mm_mul_ps(vi2x4567, vk21);
92
93 // vi0x3456 = ( vi06, vi05, vi04, vi03 )
94 const __m128 vi0x3456 = _mm_move_ss(vi0x7456, vi0x3012);
95 // vi1x3456 = ( vi16, vi15, vi14, vi13 )
96 const __m128 vi1x3456 = _mm_move_ss(vi1x7456, vi1x3012);
97 // vi2x3456 = ( vi26, vi25, vi24, vi23 )
98 const __m128 vi2x3456 = _mm_move_ss(vi2x7456, vi2x3012);
99
100 vo4567p0 = _mm_add_ps(vo4567p0, _mm_mul_ps(vi0x3456, vk00));
101 vo4567p1 = _mm_add_ps(vo4567p1, _mm_mul_ps(vi1x3456, vk10));
102 vo4567p2 = _mm_add_ps(vo4567p2, _mm_mul_ps(vi2x3456, vk20));
103
104 vi0x3012 = vi0x7456;
105 vi1x3012 = vi1x7456;
106 vi2x3012 = vi2x7456;
107
108 // vi0x8567 = ( vi07, vi06, vi05, vi08 )
109 const __m128 vi0x8567 = _mm_move_ss(vi0x4567, vi0x89AB);
110 // vi1x8567 = ( vi17, vi16, vi15, vi18 )
111 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vi1x89AB);
112 // vi2x8567 = ( vi27, vi26, vi25, vi28 )
113 const __m128 vi2x8567 = _mm_move_ss(vi2x4567, vi2x89AB);
114
115 // vi0x5678 = ( vi08, vi07, vi06, vi05 )
116 const __m128 vi0x5678 = _mm_shuffle_ps(vi0x8567, vi0x8567, _MM_SHUFFLE(0, 3, 2, 1));
117 // vi1x5678 = ( vi18, vi17, vi16, vi15 )
118 const __m128 vi1x5678 = _mm_shuffle_ps(vi1x8567, vi1x8567, _MM_SHUFFLE(0, 3, 2, 1));
119 // vi2x5678 = ( vi28, vi27, vi26, vi25 )
120 const __m128 vi2x5678 = _mm_shuffle_ps(vi2x8567, vi2x8567, _MM_SHUFFLE(0, 3, 2, 1));
121
122 vo4567p0 = _mm_add_ps(vo4567p0, _mm_mul_ps(vi0x5678, vk02));
123 vo4567p1 = _mm_add_ps(vo4567p1, _mm_mul_ps(vi1x5678, vk12));
124 vo4567p2 = _mm_add_ps(vo4567p2, _mm_mul_ps(vi2x5678, vk22));
125
126 vi0x4567 = vi0x89AB;
127 vi1x4567 = vi1x89AB;
128 vi2x4567 = vi2x89AB;
129
130 __m128 vo = _mm_add_ps(vo4567p0, vo4567p1);
131 vo = _mm_add_ps(vo, vo4567p2);
132
133 vo = _mm_max_ps(vo, vmin);
134 vo = _mm_min_ps(vo, vmax);
135
136 _mm_storeu_ps(output, vo);
137 output = (float*) ((uintptr_t) output + output_tuple_stride);
138 }
139 // Always process the last block of 1..4 pixels.
140 assert(k >= 1);
141 assert(k <= 4);
142 {
143 __m128 vo4567p0 = vbias;
144
145 vi0x4567 = _mm_and_ps(vmask, vi0x4567);
146 vi1x4567 = _mm_and_ps(vmask, vi1x4567);
147 vi2x4567 = _mm_and_ps(vmask, vi2x4567);
148
149 // vi0x7456 = ( vi06, vi05, vi04, vi07 )
150 const __m128 vi0x7456 = _mm_shuffle_ps(vi0x4567, vi0x4567, _MM_SHUFFLE(2, 1, 0, 3));
151 // vi1x7456 = ( vi16, vi15, vi14, vi17 )
152 const __m128 vi1x7456 = _mm_shuffle_ps(vi1x4567, vi1x4567, _MM_SHUFFLE(2, 1, 0, 3));
153 // vi2x7456 = ( vi26, vi25, vi24, vi27 )
154 const __m128 vi2x7456 = _mm_shuffle_ps(vi2x4567, vi2x4567, _MM_SHUFFLE(2, 1, 0, 3));
155
156 vo4567p0 = _mm_add_ps(vo4567p0, _mm_mul_ps(vi0x4567, vk01));
157 __m128 vo4567p1 = _mm_mul_ps(vi1x4567, vk11);
158 __m128 vo4567p2 = _mm_mul_ps(vi2x4567, vk21);
159
160 // vi0x3456 = ( vi06, vi05, vi04, vi03 )
161 const __m128 vi0x3456 = _mm_move_ss(vi0x7456, vi0x3012);
162 // vi1x3456 = ( vi16, vi15, vi14, vi13 )
163 const __m128 vi1x3456 = _mm_move_ss(vi1x7456, vi1x3012);
164 // vi2x3456 = ( vi26, vi25, vi24, vi23 )
165 const __m128 vi2x3456 = _mm_move_ss(vi2x7456, vi2x3012);
166
167 vo4567p0 = _mm_add_ps(vo4567p0, _mm_mul_ps(vi0x3456, vk00));
168 vo4567p1 = _mm_add_ps(vo4567p1, _mm_mul_ps(vi1x3456, vk10));
169 vo4567p2 = _mm_add_ps(vo4567p2, _mm_mul_ps(vi2x3456, vk20));
170
171 const __m128 vzero = _mm_setzero_ps();
172 // vi0x8567 = ( vi07, vi06, vi05, 0.0 )
173 const __m128 vi0x8567 = _mm_move_ss(vi0x4567, vzero);
174 // vi1x8567 = ( vi17, vi16, vi15, 0.0 )
175 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vzero);
176 // vi2x8567 = ( vi27, vi26, vi25, 0.0 )
177 const __m128 vi2x8567 = _mm_move_ss(vi2x4567, vzero);
178
179 // vi0x5678 = ( vi08, vi07, vi06, vi05 )
180 const __m128 vi0x5678 = _mm_shuffle_ps(vi0x8567, vi0x8567, _MM_SHUFFLE(0, 3, 2, 1));
181 // vi1x5678 = ( vi18, vi17, vi16, vi15 )
182 const __m128 vi1x5678 = _mm_shuffle_ps(vi1x8567, vi1x8567, _MM_SHUFFLE(0, 3, 2, 1));
183 // vi2x5678 = ( vi28, vi27, vi26, vi25 )
184 const __m128 vi2x5678 = _mm_shuffle_ps(vi2x8567, vi2x8567, _MM_SHUFFLE(0, 3, 2, 1));
185
186 vo4567p0 = _mm_add_ps(vo4567p0, _mm_mul_ps(vi0x5678, vk02));
187 vo4567p1 = _mm_add_ps(vo4567p1, _mm_mul_ps(vi1x5678, vk12));
188 vo4567p2 = _mm_add_ps(vo4567p2, _mm_mul_ps(vi2x5678, vk22));
189
190 __m128 vo = _mm_add_ps(vo4567p0, vo4567p1);
191 vo = _mm_add_ps(vo, vo4567p2);
192
193 vo = _mm_max_ps(vo, vmin);
194 vo = _mm_min_ps(vo, vmax);
195
196 if XNN_LIKELY(k == 4) {
197 _mm_storeu_ps(output, vo);
198 } else {
199 float* output_lo = output;
200 if (k & 2) {
201 _mm_storel_pi((__m64*) output_lo, vo);
202 output_lo += 2;
203 vo = _mm_movehl_ps(vo, vo);
204 }
205 if (k & 1) {
206 _mm_store_ss(output_lo, vo);
207 }
208 }
209 }
210
211 i0 = (const float*) ((uintptr_t) i0 + input_width_increment);
212 i1 = (const float*) ((uintptr_t) i1 + input_width_increment);
213 i2 = (const float*) ((uintptr_t) i2 + input_width_increment);
214 output = (float*) ((uintptr_t) output + output_width_increment);
215 } while (--m != 0);
216 }
217