1 // Auto-generated file. Do not edit!
2 // Template: src/f32-dwconv/up-sse.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2019 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
16
xnn_f32_dwconv_minmax_ukernel_up8x4__sse_acc2(size_t channels,size_t output_width,const float ** input,const float * weights,float * output,size_t input_stride,size_t output_increment,size_t input_offset,const float * zero,const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_f32_dwconv_minmax_ukernel_up8x4__sse_acc2(
18 size_t channels,
19 size_t output_width,
20 const float** input,
21 const float* weights,
22 float* output,
23 size_t input_stride,
24 size_t output_increment,
25 size_t input_offset,
26 const float* zero,
27 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
28 {
29 assert(channels != 0);
30 assert(output_width != 0);
31
32 const __m128 vmax = _mm_load_ps(params->sse.max);
33 const __m128 vmin = _mm_load_ps(params->sse.min);
34 do {
35 const float* i0 = input[0];
36 assert(i0 != NULL);
37 if XNN_UNPREDICTABLE(i0 != zero) {
38 i0 = (const float*) ((uintptr_t) i0 + input_offset);
39 }
40 const float* i1 = input[1];
41 assert(i1 != NULL);
42 if XNN_UNPREDICTABLE(i1 != zero) {
43 i1 = (const float*) ((uintptr_t) i1 + input_offset);
44 }
45 const float* i2 = input[2];
46 assert(i2 != NULL);
47 if XNN_UNPREDICTABLE(i2 != zero) {
48 i2 = (const float*) ((uintptr_t) i2 + input_offset);
49 }
50 const float* i3 = input[3];
51 assert(i3 != NULL);
52 if XNN_UNPREDICTABLE(i3 != zero) {
53 i3 = (const float*) ((uintptr_t) i3 + input_offset);
54 }
55 input = (const float**) ((uintptr_t) input + input_stride);
56
57 size_t c = channels;
58 const float* w = weights;
59 for (; c >= 8; c -= 8) {
60 __m128 vacc0123p0 = _mm_load_ps(w);
61 __m128 vacc4567p0 = _mm_load_ps(w + 4);
62
63
64 const __m128 vi0x0123 = _mm_loadu_ps(i0);
65 const __m128 vi0x4567 = _mm_loadu_ps(i0 + 4);
66 i0 += 8;
67
68 const __m128 vk0x0123 = _mm_load_ps(w + 8);
69 const __m128 vk0x4567 = _mm_load_ps(w + 12);
70 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi0x0123, vk0x0123));
71 vacc4567p0 = _mm_add_ps(vacc4567p0, _mm_mul_ps(vi0x4567, vk0x4567));
72
73 const __m128 vi1x0123 = _mm_loadu_ps(i1);
74 const __m128 vi1x4567 = _mm_loadu_ps(i1 + 4);
75 i1 += 8;
76
77 const __m128 vk1x0123 = _mm_load_ps(w + 16);
78 const __m128 vk1x4567 = _mm_load_ps(w + 20);
79 __m128 vacc0123p1 = _mm_mul_ps(vi1x0123, vk1x0123);
80 __m128 vacc4567p1 = _mm_mul_ps(vi1x4567, vk1x4567);
81
82 const __m128 vi2x0123 = _mm_loadu_ps(i2);
83 const __m128 vi2x4567 = _mm_loadu_ps(i2 + 4);
84 i2 += 8;
85
86 const __m128 vk2x0123 = _mm_load_ps(w + 24);
87 const __m128 vk2x4567 = _mm_load_ps(w + 28);
88 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi2x0123, vk2x0123));
89 vacc4567p0 = _mm_add_ps(vacc4567p0, _mm_mul_ps(vi2x4567, vk2x4567));
90
91 const __m128 vi3x0123 = _mm_loadu_ps(i3);
92 const __m128 vi3x4567 = _mm_loadu_ps(i3 + 4);
93 i3 += 8;
94
95 const __m128 vk3x0123 = _mm_load_ps(w + 32);
96 const __m128 vk3x4567 = _mm_load_ps(w + 36);
97 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi3x0123, vk3x0123));
98 vacc4567p1 = _mm_add_ps(vacc4567p1, _mm_mul_ps(vi3x4567, vk3x4567));
99
100 w += 40;
101
102 // Add up all accumulators to vacc01234567p0
103 vacc0123p0 = _mm_add_ps(vacc0123p0, vacc0123p1);
104 vacc4567p0 = _mm_add_ps(vacc4567p0, vacc4567p1);
105
106 __m128 vacc0123 = _mm_max_ps(vacc0123p0, vmin);
107 __m128 vacc4567 = _mm_max_ps(vacc4567p0, vmin);
108 vacc0123 = _mm_min_ps(vacc0123, vmax);
109 vacc4567 = _mm_min_ps(vacc4567, vmax);
110
111 _mm_storeu_ps(output, vacc0123);
112 _mm_storeu_ps(output + 4, vacc4567);
113 output += 8;
114 }
115 for (; c >= 4; c -= 4) {
116 __m128 vacc0123p0 = _mm_load_ps(w);
117
118 const __m128 vi0x0123 = _mm_loadu_ps(i0);
119 i0 += 4;
120
121 const __m128 vk0x0123 = _mm_load_ps(w + 8);
122 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi0x0123, vk0x0123));
123
124 const __m128 vi1x0123 = _mm_loadu_ps(i1);
125 i1 += 4;
126
127 const __m128 vk1x0123 = _mm_load_ps(w + 16);
128 __m128 vacc0123p1 = _mm_mul_ps(vi1x0123, vk1x0123);
129
130 const __m128 vi2x0123 = _mm_loadu_ps(i2);
131 i2 += 4;
132
133 const __m128 vk2x0123 = _mm_load_ps(w + 24);
134 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi2x0123, vk2x0123));
135
136 const __m128 vi3x0123 = _mm_loadu_ps(i3);
137 i3 += 4;
138
139 const __m128 vk3x0123 = _mm_load_ps(w + 32);
140 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi3x0123, vk3x0123));
141
142 w += 4;
143
144 // Add up all accumulators to vacc01234567p0
145 vacc0123p0 = _mm_add_ps(vacc0123p0, vacc0123p1);
146
147 __m128 vacc0123 = _mm_max_ps(vacc0123p0, vmin);
148 vacc0123 = _mm_min_ps(vacc0123, vmax);
149
150 _mm_storeu_ps(output, vacc0123);
151 output += 4;
152 }
153 if XNN_UNLIKELY(c != 0) {
154 __m128 vacc0123p0 = _mm_load_ps(w);
155
156 const __m128 vi0x0123 = _mm_loadu_ps(i0);
157 const __m128 vk0x0123 = _mm_load_ps(w + 8);
158 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi0x0123, vk0x0123));
159
160 const __m128 vi1x0123 = _mm_loadu_ps(i1);
161 const __m128 vk1x0123 = _mm_load_ps(w + 16);
162 __m128 vacc0123p1 = _mm_mul_ps(vi1x0123, vk1x0123);
163
164 const __m128 vi2x0123 = _mm_loadu_ps(i2);
165 const __m128 vk2x0123 = _mm_load_ps(w + 24);
166 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi2x0123, vk2x0123));
167
168 const __m128 vi3x0123 = _mm_loadu_ps(i3);
169 const __m128 vk3x0123 = _mm_load_ps(w + 32);
170 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi3x0123, vk3x0123));
171
172 // Add up all accumulators to vacc01234567p0
173 vacc0123p0 = _mm_add_ps(vacc0123p0, vacc0123p1);
174
175 __m128 vacc0123 = _mm_max_ps(vacc0123p0, vmin);
176 vacc0123 = _mm_min_ps(vacc0123, vmax);
177
178 if (c & 2) {
179 _mm_storel_pi((__m64*) output, vacc0123);
180 vacc0123 = _mm_movehl_ps(vacc0123, vacc0123);
181 output += 2;
182 }
183 if (c & 1) {
184 _mm_store_ss(output, vacc0123);
185 output += 1;
186 }
187 }
188
189 output = (float*) ((uintptr_t) output + output_increment);
190 } while (--output_width != 0);
191 }
192