1 // Copyright 2021 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 <immintrin.h>
9
10 #include <xnnpack/dwconv.h>
11 #include <xnnpack/math.h>
12
13
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)])14 void xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_2x4_acc2(
15 size_t input_height,
16 size_t input_width,
17 const float* input,
18 const float* weights,
19 const float* zero,
20 float* output,
21 uint32_t padding_top,
22 const union xnn_f32_chw_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
23 {
24 assert(input_height != 0);
25 assert(input_width != 0);
26 assert(input_width % sizeof(float) == 0);
27 assert(padding_top == 1);
28
29 const __m128 vmask = _mm_load_ps((const float*) params->sse.mask);
30 const __m128 vmax = _mm_load_ps(params->sse.max);
31 const __m128 vmin = _mm_load_ps(params->sse.min);
32
33 const __m128 vbias = _mm_load1_ps(weights);
34 const __m128 vk00 = _mm_load1_ps(weights + 1);
35 const __m128 vk01 = _mm_load1_ps(weights + 2);
36 const __m128 vk02 = _mm_load1_ps(weights + 3);
37 const __m128 vk10 = _mm_load1_ps(weights + 4);
38 const __m128 vk11 = _mm_load1_ps(weights + 5);
39 const __m128 vk12 = _mm_load1_ps(weights + 6);
40 const __m128 vk20 = _mm_load1_ps(weights + 7);
41 const __m128 vk21 = _mm_load1_ps(weights + 8);
42 const __m128 vk22 = _mm_load1_ps(weights + 9);
43
44 const size_t input_decrement = round_up_po2(input_width, 4 * sizeof(float));
45
46 const float* i0 = zero;
47 const float* i1 = input;
48 const float* i2 = (const float*) ((uintptr_t) i1 + input_width);
49 const float* i3 = (const float*) ((uintptr_t) i2 + input_width);
50
51 float* o0 = output;
52 float* o1 = (float*) ((uintptr_t) o0 + input_width);
53
54 size_t output_height = input_height;
55 do {
56 if XNN_UNPREDICTABLE(output_height < 2) {
57 i2 = zero;
58 o1 = o0;
59 }
60 if XNN_UNPREDICTABLE(output_height < 3) {
61 i3 = zero;
62 }
63
64 __m128 vi0x0123 = _mm_setzero_ps();
65 __m128 vi1x0123 = _mm_setzero_ps();
66 __m128 vi2x0123 = _mm_setzero_ps();
67 __m128 vi3x0123 = _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 __m128 vi3x4567 = _mm_loadu_ps(i3);
76 i3 += 4;
77
78 size_t w = input_width;
79 for (; w > 4 * sizeof(float); w -= 4 * sizeof(float)) {
80 const __m128 vi0x89AB = _mm_loadu_ps(i0);
81 i0 += 4;
82 const __m128 vi1x89AB = _mm_loadu_ps(i1);
83 i1 += 4;
84 const __m128 vi2x89AB = _mm_loadu_ps(i2);
85 i2 += 4;
86 const __m128 vi3x89AB = _mm_loadu_ps(i3);
87 i3 += 4;
88
89 __m128 vo0p0 = _mm_add_ps(vbias, _mm_mul_ps(vi0x4567, vk01));
90 __m128 vo1p0 = _mm_add_ps(vbias, _mm_mul_ps(vi1x4567, vk01));
91 __m128 vo0p1 = _mm_mul_ps(vi1x4567, vk11);
92 __m128 vo1p1 = _mm_mul_ps(vi2x4567, vk11);
93 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x4567, vk21));
94 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi3x4567, vk21));
95
96 const __m128 vi0x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi0x4567), _mm_castps_si128(vi0x0123), 12));
97 const __m128 vi1x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi1x4567), _mm_castps_si128(vi1x0123), 12));
98 const __m128 vi2x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi2x4567), _mm_castps_si128(vi2x0123), 12));
99 const __m128 vi3x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi3x4567), _mm_castps_si128(vi3x0123), 12));
100
101 vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi0x3456, vk00));
102 vo1p1 = _mm_add_ps(vo1p1, _mm_mul_ps(vi1x3456, vk00));
103 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x3456, vk10));
104 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x3456, vk10));
105 vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi2x3456, vk20));
106 vo1p1 = _mm_add_ps(vo1p1, _mm_mul_ps(vi3x3456, vk20));
107
108 vi0x0123 = vi0x4567;
109 vi1x0123 = vi1x4567;
110 vi2x0123 = vi2x4567;
111 vi3x0123 = vi3x4567;
112
113 const __m128 vi0x5678 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi0x89AB), _mm_castps_si128(vi0x4567), 4));
114 const __m128 vi1x5678 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi1x89AB), _mm_castps_si128(vi1x4567), 4));
115 const __m128 vi2x5678 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi2x89AB), _mm_castps_si128(vi2x4567), 4));
116 const __m128 vi3x5678 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi3x89AB), _mm_castps_si128(vi3x4567), 4));
117
118 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi0x5678, vk02));
119 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi1x5678, vk02));
120 vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi1x5678, vk12));
121 vo1p1 = _mm_add_ps(vo1p1, _mm_mul_ps(vi2x5678, vk12));
122 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x5678, vk22));
123 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi3x5678, vk22));
124
125 vi0x4567 = vi0x89AB;
126 vi1x4567 = vi1x89AB;
127 vi2x4567 = vi2x89AB;
128 vi3x4567 = vi3x89AB;
129
130 vo0p0 = _mm_add_ps(vo0p0, vo0p1);
131 vo1p0 = _mm_add_ps(vo1p0, vo1p1);
132
133 __m128 vo0 = _mm_max_ps(vo0p0, vmin);
134 __m128 vo1 = _mm_max_ps(vo1p0, vmin);
135
136 vo0 = _mm_min_ps(vo0, vmax);
137 vo1 = _mm_min_ps(vo1, vmax);
138
139 _mm_storeu_ps(o1, vo1);
140 o1 += 4;
141 _mm_storeu_ps(o0, vo0);
142 o0 += 4;
143 }
144 // Always process the last block of 1..4 pixels.
145 assert(w >= 1 * sizeof(float));
146 assert(w <= 4 * sizeof(float));
147 {
148 vi0x4567 = _mm_and_ps(vmask, vi0x4567);
149 vi1x4567 = _mm_and_ps(vmask, vi1x4567);
150 vi2x4567 = _mm_and_ps(vmask, vi2x4567);
151 vi3x4567 = _mm_and_ps(vmask, vi3x4567);
152
153 __m128 vo0p0 = _mm_add_ps(vbias, _mm_mul_ps(vi0x4567, vk01));
154 __m128 vo1p0 = _mm_add_ps(vbias, _mm_mul_ps(vi1x4567, vk01));
155 __m128 vo0p1 = _mm_mul_ps(vi1x4567, vk11);
156 __m128 vo1p1 = _mm_mul_ps(vi2x4567, vk11);
157 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x4567, vk21));
158 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi3x4567, vk21));
159
160 const __m128 vi0x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi0x4567), _mm_castps_si128(vi0x0123), 12));
161 const __m128 vi1x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi1x4567), _mm_castps_si128(vi1x0123), 12));
162 const __m128 vi2x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi2x4567), _mm_castps_si128(vi2x0123), 12));
163 const __m128 vi3x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi3x4567), _mm_castps_si128(vi3x0123), 12));
164
165 vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi0x3456, vk00));
166 vo1p1 = _mm_add_ps(vo1p1, _mm_mul_ps(vi1x3456, vk00));
167 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x3456, vk10));
168 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x3456, vk10));
169 vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi2x3456, vk20));
170 vo1p1 = _mm_add_ps(vo1p1, _mm_mul_ps(vi3x3456, vk20));
171
172 const __m128i vzero = _mm_setzero_si128();
173 const __m128 vi0x5678 = _mm_castsi128_ps(_mm_alignr_epi8(vzero, _mm_castps_si128(vi0x4567), 4));
174 const __m128 vi1x5678 = _mm_castsi128_ps(_mm_alignr_epi8(vzero, _mm_castps_si128(vi1x4567), 4));
175 const __m128 vi2x5678 = _mm_castsi128_ps(_mm_alignr_epi8(vzero, _mm_castps_si128(vi2x4567), 4));
176 const __m128 vi3x5678 = _mm_castsi128_ps(_mm_alignr_epi8(vzero, _mm_castps_si128(vi3x4567), 4));
177
178 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi0x5678, vk02));
179 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi1x5678, vk02));
180 vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi1x5678, vk12));
181 vo1p1 = _mm_add_ps(vo1p1, _mm_mul_ps(vi2x5678, vk12));
182 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x5678, vk22));
183 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi3x5678, vk22));
184
185 vo0p0 = _mm_add_ps(vo0p0, vo0p1);
186 vo1p0 = _mm_add_ps(vo1p0, vo1p1);
187
188 __m128 vo0 = _mm_max_ps(vo0p0, vmin);
189 __m128 vo1 = _mm_max_ps(vo1p0, vmin);
190
191 vo0 = _mm_min_ps(vo0, vmax);
192 vo1 = _mm_min_ps(vo1, vmax);
193
194 if XNN_LIKELY(w == 4 * sizeof(float)) {
195 _mm_storeu_ps(o1, vo1);
196 o1 += 4;
197 _mm_storeu_ps(o0, vo0);
198 o0 += 4;
199 } else {
200 if (w & (2 * sizeof(float))) {
201 _mm_storel_pi((__m64*) o1, vo1);
202 o1 += 2;
203 _mm_storel_pi((__m64*) o0, vo0);
204 o0 += 2;
205
206 vo0 = _mm_movehl_ps(vo0, vo0);
207 vo1 = _mm_movehl_ps(vo1, vo1);
208 }
209 if (w & (1 * sizeof(float))) {
210 _mm_store_ss(o1, vo1);
211 o1 += 1;
212 _mm_store_ss(o0, vo0);
213 o0 += 1;
214 }
215 }
216 }
217
218 i0 = (const float*) ((uintptr_t) i2 - input_decrement);
219 i1 = (const float*) ((uintptr_t) i3 - input_decrement);
220 i2 = (const float*) ((uintptr_t) i1 + input_width);
221 i3 = (const float*) ((uintptr_t) i2 + input_width);
222
223 o0 = o1;
224 o1 = (float*) ((uintptr_t) o0 + input_width);
225
226 output_height = doz(output_height, 2);
227 } while (output_height != 0);
228 }
229