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_up4x4__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_up4x4__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 >= 4; c -= 4) {
46 __m128 vacc0123p0 = _mm_load_ps(w);
47
48
49 const __m128 vi0x0123 = _mm_loadu_ps(i0);
50 i0 += 4;
51
52 const __m128 vk0x0123 = _mm_load_ps(w + 4);
53 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi0x0123, vk0x0123));
54
55 const __m128 vi1x0123 = _mm_loadu_ps(i1);
56 i1 += 4;
57
58 const __m128 vk1x0123 = _mm_load_ps(w + 8);
59 __m128 vacc0123p1 = _mm_mul_ps(vi1x0123, vk1x0123);
60
61 const __m128 vi2x0123 = _mm_loadu_ps(i2);
62 i2 += 4;
63
64 const __m128 vk2x0123 = _mm_load_ps(w + 12);
65 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi2x0123, vk2x0123));
66
67 const __m128 vi3x0123 = _mm_loadu_ps(i3);
68 i3 += 4;
69
70 const __m128 vk3x0123 = _mm_load_ps(w + 16);
71 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi3x0123, vk3x0123));
72
73 w += 20;
74
75 // Add up all accumulators to vacc0123p0
76 vacc0123p0 = _mm_add_ps(vacc0123p0, vacc0123p1);
77
78 __m128 vacc0123 = _mm_max_ps(vacc0123p0, vmin);
79 vacc0123 = _mm_min_ps(vacc0123, vmax);
80
81 _mm_storeu_ps(output, vacc0123);
82 output += 4;
83 }
84 if XNN_UNLIKELY(c != 0) {
85 __m128 vacc0123p0 = _mm_load_ps(w);
86
87 const __m128 vi0x0123 = _mm_loadu_ps(i0);
88 const __m128 vk0x0123 = _mm_load_ps(w + 4);
89 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi0x0123, vk0x0123));
90
91 const __m128 vi1x0123 = _mm_loadu_ps(i1);
92 const __m128 vk1x0123 = _mm_load_ps(w + 8);
93 __m128 vacc0123p1 = _mm_mul_ps(vi1x0123, vk1x0123);
94
95 const __m128 vi2x0123 = _mm_loadu_ps(i2);
96 const __m128 vk2x0123 = _mm_load_ps(w + 12);
97 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi2x0123, vk2x0123));
98
99 const __m128 vi3x0123 = _mm_loadu_ps(i3);
100 const __m128 vk3x0123 = _mm_load_ps(w + 16);
101 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi3x0123, vk3x0123));
102
103 // Add up all accumulators to vacc0123p0
104 vacc0123p0 = _mm_add_ps(vacc0123p0, vacc0123p1);
105
106 __m128 vacc0123 = _mm_max_ps(vacc0123p0, vmin);
107 vacc0123 = _mm_min_ps(vacc0123, vmax);
108
109 if (c & 2) {
110 _mm_storel_pi((__m64*) output, vacc0123);
111 vacc0123 = _mm_movehl_ps(vacc0123, vacc0123);
112 output += 2;
113 }
114 if (c & 1) {
115 _mm_store_ss(output, vacc0123);
116 output += 1;
117 }
118 }
119
120 output = (float*) ((uintptr_t) output + output_increment);
121 } while (--output_width != 0);
122 }
123