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_up8x3__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_up8x3__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 input = (const float**) ((uintptr_t) input + input_stride);
51
52 size_t c = channels;
53 const float* w = weights;
54 for (; c >= 8; c -= 8) {
55 __m128 vacc0123p0 = _mm_load_ps(w);
56 __m128 vacc4567p0 = _mm_load_ps(w + 4);
57
58
59 const __m128 vi0x0123 = _mm_loadu_ps(i0);
60 const __m128 vi0x4567 = _mm_loadu_ps(i0 + 4);
61 i0 += 8;
62
63 const __m128 vk0x0123 = _mm_load_ps(w + 8);
64 const __m128 vk0x4567 = _mm_load_ps(w + 12);
65 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi0x0123, vk0x0123));
66 vacc4567p0 = _mm_add_ps(vacc4567p0, _mm_mul_ps(vi0x4567, vk0x4567));
67
68 const __m128 vi1x0123 = _mm_loadu_ps(i1);
69 const __m128 vi1x4567 = _mm_loadu_ps(i1 + 4);
70 i1 += 8;
71
72 const __m128 vk1x0123 = _mm_load_ps(w + 16);
73 const __m128 vk1x4567 = _mm_load_ps(w + 20);
74 __m128 vacc0123p1 = _mm_mul_ps(vi1x0123, vk1x0123);
75 __m128 vacc4567p1 = _mm_mul_ps(vi1x4567, vk1x4567);
76
77 const __m128 vi2x0123 = _mm_loadu_ps(i2);
78 const __m128 vi2x4567 = _mm_loadu_ps(i2 + 4);
79 i2 += 8;
80
81 const __m128 vk2x0123 = _mm_load_ps(w + 24);
82 const __m128 vk2x4567 = _mm_load_ps(w + 28);
83 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi2x0123, vk2x0123));
84 vacc4567p0 = _mm_add_ps(vacc4567p0, _mm_mul_ps(vi2x4567, vk2x4567));
85
86 w += 32;
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 w += 4;
123
124 // Add up all accumulators to vacc01234567p0
125 vacc0123p0 = _mm_add_ps(vacc0123p0, vacc0123p1);
126
127 __m128 vacc0123 = _mm_max_ps(vacc0123p0, vmin);
128 vacc0123 = _mm_min_ps(vacc0123, vmax);
129
130 _mm_storeu_ps(output, vacc0123);
131 output += 4;
132 }
133 if XNN_UNLIKELY(c != 0) {
134 __m128 vacc0123p0 = _mm_load_ps(w);
135
136 const __m128 vi0x0123 = _mm_loadu_ps(i0);
137 const __m128 vk0x0123 = _mm_load_ps(w + 8);
138 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi0x0123, vk0x0123));
139
140 const __m128 vi1x0123 = _mm_loadu_ps(i1);
141 const __m128 vk1x0123 = _mm_load_ps(w + 16);
142 __m128 vacc0123p1 = _mm_mul_ps(vi1x0123, vk1x0123);
143
144 const __m128 vi2x0123 = _mm_loadu_ps(i2);
145 const __m128 vk2x0123 = _mm_load_ps(w + 24);
146 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi2x0123, vk2x0123));
147
148 // Add up all accumulators to vacc01234567p0
149 vacc0123p0 = _mm_add_ps(vacc0123p0, vacc0123p1);
150
151 __m128 vacc0123 = _mm_max_ps(vacc0123p0, vmin);
152 vacc0123 = _mm_min_ps(vacc0123, vmax);
153
154 if (c & 2) {
155 _mm_storel_pi((__m64*) output, vacc0123);
156 vacc0123 = _mm_movehl_ps(vacc0123, vacc0123);
157 output += 2;
158 }
159 if (c & 1) {
160 _mm_store_ss(output, vacc0123);
161 output += 1;
162 }
163 }
164
165 output = (float*) ((uintptr_t) output + output_increment);
166 } while (--output_width != 0);
167 }
168