• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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_up4x4__sse(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_up4x4__sse(
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_DISABLE_TSAN
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 >= 4; c -= 4) {
60       __m128 vacc0123p0 = _mm_load_ps(w);
61 
62 
63       const __m128 vi0x0123 = _mm_loadu_ps(i0);
64       i0 += 4;
65 
66       const __m128 vk0x0123 = _mm_load_ps(w + 4);
67       vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi0x0123, vk0x0123));
68 
69       const __m128 vi1x0123 = _mm_loadu_ps(i1);
70       i1 += 4;
71 
72       const __m128 vk1x0123 = _mm_load_ps(w + 8);
73       vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi1x0123, vk1x0123));
74 
75       const __m128 vi2x0123 = _mm_loadu_ps(i2);
76       i2 += 4;
77 
78       const __m128 vk2x0123 = _mm_load_ps(w + 12);
79       vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi2x0123, vk2x0123));
80 
81       const __m128 vi3x0123 = _mm_loadu_ps(i3);
82       i3 += 4;
83 
84       const __m128 vk3x0123 = _mm_load_ps(w + 16);
85       vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi3x0123, vk3x0123));
86 
87       w += 20;
88 
89 
90       __m128 vacc0123 = _mm_max_ps(vacc0123p0, vmin);
91       vacc0123 = _mm_min_ps(vacc0123, vmax);
92 
93       _mm_storeu_ps(output, vacc0123);
94       output += 4;
95     }
96     if XNN_UNLIKELY(c != 0) {
97       __m128 vacc0123p0 = _mm_load_ps(w);
98 
99       const __m128 vi0x0123 = _mm_loadu_ps(i0);
100       const __m128 vk0x0123 = _mm_load_ps(w + 4);
101       vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi0x0123, vk0x0123));
102 
103       const __m128 vi1x0123 = _mm_loadu_ps(i1);
104       const __m128 vk1x0123 = _mm_load_ps(w + 8);
105       vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi1x0123, vk1x0123));
106 
107       const __m128 vi2x0123 = _mm_loadu_ps(i2);
108       const __m128 vk2x0123 = _mm_load_ps(w + 12);
109       vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi2x0123, vk2x0123));
110 
111       const __m128 vi3x0123 = _mm_loadu_ps(i3);
112       const __m128 vk3x0123 = _mm_load_ps(w + 16);
113       vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi3x0123, vk3x0123));
114 
115 
116       __m128 vacc0123 = _mm_max_ps(vacc0123p0, vmin);
117       vacc0123 = _mm_min_ps(vacc0123, vmax);
118 
119       if (c & 2) {
120         _mm_storel_pi((__m64*) output, vacc0123);
121         vacc0123 = _mm_movehl_ps(vacc0123, vacc0123);
122         output += 2;
123       }
124       if (c & 1) {
125         _mm_store_ss(output, vacc0123);
126         output += 1;
127       }
128     }
129 
130     output = (float*) ((uintptr_t) output + output_increment);
131   } while (--output_width != 0);
132 }
133