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_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,const union xnn_f32_output_params params[restrict static1])17 void xnn_f32_dwconv_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 const union xnn_f32_output_params params[restrict static 1])
26 {
27 assert(channels != 0);
28 assert(output_width != 0);
29
30 const __m128 vmax = _mm_load_ps(params->sse.max);
31 const __m128 vmin = _mm_load_ps(params->sse.min);
32 do {
33 const float* i0 = input[0];
34 assert(i0 != NULL);
35 const float* i1 = input[1];
36 assert(i1 != NULL);
37 const float* i2 = input[2];
38 assert(i2 != NULL);
39 const float* i3 = input[3];
40 assert(i3 != NULL);
41 input = (const float**) ((uintptr_t) input + input_stride);
42
43 size_t c = channels;
44 const float* w = weights;
45 for (; c >= 8; c -= 8) {
46 __m128 vacc0123p0 = _mm_load_ps(w);
47 __m128 vacc4567p0 = _mm_load_ps(w + 4);
48
49
50 const __m128 vi0x0123 = _mm_loadu_ps(i0);
51 const __m128 vi0x4567 = _mm_loadu_ps(i0 + 4);
52 i0 += 8;
53
54 const __m128 vk0x0123 = _mm_load_ps(w + 8);
55 const __m128 vk0x4567 = _mm_load_ps(w + 12);
56 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi0x0123, vk0x0123));
57 vacc4567p0 = _mm_add_ps(vacc4567p0, _mm_mul_ps(vi0x4567, vk0x4567));
58
59 const __m128 vi1x0123 = _mm_loadu_ps(i1);
60 const __m128 vi1x4567 = _mm_loadu_ps(i1 + 4);
61 i1 += 8;
62
63 const __m128 vk1x0123 = _mm_load_ps(w + 16);
64 const __m128 vk1x4567 = _mm_load_ps(w + 20);
65 __m128 vacc0123p1 = _mm_mul_ps(vi1x0123, vk1x0123);
66 __m128 vacc4567p1 = _mm_mul_ps(vi1x4567, vk1x4567);
67
68 const __m128 vi2x0123 = _mm_loadu_ps(i2);
69 const __m128 vi2x4567 = _mm_loadu_ps(i2 + 4);
70 i2 += 8;
71
72 const __m128 vk2x0123 = _mm_load_ps(w + 24);
73 const __m128 vk2x4567 = _mm_load_ps(w + 28);
74 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi2x0123, vk2x0123));
75 vacc4567p0 = _mm_add_ps(vacc4567p0, _mm_mul_ps(vi2x4567, vk2x4567));
76
77 const __m128 vi3x0123 = _mm_loadu_ps(i3);
78 const __m128 vi3x4567 = _mm_loadu_ps(i3 + 4);
79 i3 += 8;
80
81 const __m128 vk3x0123 = _mm_load_ps(w + 32);
82 const __m128 vk3x4567 = _mm_load_ps(w + 36);
83 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi3x0123, vk3x0123));
84 vacc4567p1 = _mm_add_ps(vacc4567p1, _mm_mul_ps(vi3x4567, vk3x4567));
85
86 w += 40;
87
88 // Add up all accumulators to vacc01234567p0
89 vacc0123p0 = _mm_add_ps(vacc0123p0, vacc0123p1);
90 vacc4567p0 = _mm_add_ps(vacc4567p0, vacc4567p1);
91
92 __m128 vacc0123 = _mm_max_ps(vacc0123p0, vmin);
93 __m128 vacc4567 = _mm_max_ps(vacc4567p0, vmin);
94 vacc0123 = _mm_min_ps(vacc0123, vmax);
95 vacc4567 = _mm_min_ps(vacc4567, vmax);
96
97 _mm_storeu_ps(output, vacc0123);
98 _mm_storeu_ps(output + 4, vacc4567);
99 output += 8;
100 }
101 for (; c >= 4; c -= 4) {
102 __m128 vacc0123p0 = _mm_load_ps(w);
103
104 const __m128 vi0x0123 = _mm_loadu_ps(i0);
105 i0 += 4;
106
107 const __m128 vk0x0123 = _mm_load_ps(w + 8);
108 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi0x0123, vk0x0123));
109
110 const __m128 vi1x0123 = _mm_loadu_ps(i1);
111 i1 += 4;
112
113 const __m128 vk1x0123 = _mm_load_ps(w + 16);
114 __m128 vacc0123p1 = _mm_mul_ps(vi1x0123, vk1x0123);
115
116 const __m128 vi2x0123 = _mm_loadu_ps(i2);
117 i2 += 4;
118
119 const __m128 vk2x0123 = _mm_load_ps(w + 24);
120 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi2x0123, vk2x0123));
121
122 const __m128 vi3x0123 = _mm_loadu_ps(i3);
123 i3 += 4;
124
125 const __m128 vk3x0123 = _mm_load_ps(w + 32);
126 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi3x0123, vk3x0123));
127
128 w += 4;
129
130 // Add up all accumulators to vacc01234567p0
131 vacc0123p0 = _mm_add_ps(vacc0123p0, vacc0123p1);
132
133 __m128 vacc0123 = _mm_max_ps(vacc0123p0, vmin);
134 vacc0123 = _mm_min_ps(vacc0123, vmax);
135
136 _mm_storeu_ps(output, vacc0123);
137 output += 4;
138 }
139 if XNN_UNLIKELY(c != 0) {
140 __m128 vacc0123p0 = _mm_load_ps(w);
141
142 const __m128 vi0x0123 = _mm_loadu_ps(i0);
143 const __m128 vk0x0123 = _mm_load_ps(w + 8);
144 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi0x0123, vk0x0123));
145
146 const __m128 vi1x0123 = _mm_loadu_ps(i1);
147 const __m128 vk1x0123 = _mm_load_ps(w + 16);
148 __m128 vacc0123p1 = _mm_mul_ps(vi1x0123, vk1x0123);
149
150 const __m128 vi2x0123 = _mm_loadu_ps(i2);
151 const __m128 vk2x0123 = _mm_load_ps(w + 24);
152 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi2x0123, vk2x0123));
153
154 const __m128 vi3x0123 = _mm_loadu_ps(i3);
155 const __m128 vk3x0123 = _mm_load_ps(w + 32);
156 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi3x0123, vk3x0123));
157
158 // Add up all accumulators to vacc01234567p0
159 vacc0123p0 = _mm_add_ps(vacc0123p0, vacc0123p1);
160
161 __m128 vacc0123 = _mm_max_ps(vacc0123p0, vmin);
162 vacc0123 = _mm_min_ps(vacc0123, vmax);
163
164 if (c & 2) {
165 _mm_storel_pi((__m64*) output, vacc0123);
166 vacc0123 = _mm_movehl_ps(vacc0123, vacc0123);
167 output += 2;
168 }
169 if (c & 1) {
170 _mm_store_ss(output, vacc0123);
171 output += 1;
172 }
173 }
174
175 output = (float*) ((uintptr_t) output + output_increment);
176 } while (--output_width != 0);
177 }
178