1 // Copyright 2019 Google LLC
2 //
3 // This source code is licensed under the BSD-style license found in the
4 // LICENSE file in the root directory of this source tree.
5
6 #include <assert.h>
7
8 #include <xmmintrin.h>
9
10 #include <xnnpack/gavgpool.h>
11
12
xnn_f32_gavgpool_minmax_ukernel_7x__sse_c4(size_t rows,size_t channels,const float * input,size_t input_stride,const float * zero,float * output,const union xnn_f32_scaleminmax_params params[restrict XNN_MIN_ELEMENTS (1)])13 void xnn_f32_gavgpool_minmax_ukernel_7x__sse_c4(
14 size_t rows,
15 size_t channels,
16 const float* input,
17 size_t input_stride,
18 const float* zero,
19 float* output,
20 const union xnn_f32_scaleminmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
21 {
22 assert(rows != 0);
23 assert(rows <= 7);
24 assert(channels != 0);
25
26 const float* i0 = input;
27 const float* i1 = (const float*) ((uintptr_t) i0 + input_stride);
28 if (rows < 2) {
29 i1 = zero;
30 }
31 const float* i2 = (const float*) ((uintptr_t) i1 + input_stride);
32 if (rows <= 2) {
33 i2 = zero;
34 }
35 const float* i3 = (const float*) ((uintptr_t) i2 + input_stride);
36 if (rows < 4) {
37 i3 = zero;
38 }
39 const float* i4 = (const float*) ((uintptr_t) i3 + input_stride);
40 if (rows <= 4) {
41 i4 = zero;
42 }
43 const float* i5 = (const float*) ((uintptr_t) i4 + input_stride);
44 if (rows < 6) {
45 i5 = zero;
46 }
47 const float* i6 = (const float*) ((uintptr_t) i5 + input_stride);
48 if (rows <= 6) {
49 i6 = zero;
50 }
51 const __m128 vscale = _mm_load_ps(params->sse2.scale);
52 const __m128 vmin = _mm_load_ps(params->sse2.min);
53 const __m128 vmax = _mm_load_ps(params->sse2.max);
54
55 while (channels >= 4) {
56 const __m128 vi0 = _mm_loadu_ps(i0);
57 i0 += 4;
58 const __m128 vi1 = _mm_loadu_ps(i1);
59 i1 += 4;
60 const __m128 vi2 = _mm_loadu_ps(i2);
61 i2 += 4;
62 const __m128 vi3 = _mm_loadu_ps(i3);
63 i3 += 4;
64 const __m128 vi4 = _mm_loadu_ps(i4);
65 i4 += 4;
66 const __m128 vi5 = _mm_loadu_ps(i5);
67 i5 += 4;
68 const __m128 vi6 = _mm_loadu_ps(i6);
69 i6 += 4;
70
71 const __m128 vsum01 = _mm_add_ps(vi0, vi1);
72 const __m128 vsum23 = _mm_add_ps(vi2, vi3);
73 const __m128 vsum45 = _mm_add_ps(vi4, vi5);
74
75 const __m128 vsum016 = _mm_add_ps(vsum01, vi6);
76 const __m128 vsum2345 = _mm_add_ps(vsum23, vsum45);
77
78 const __m128 vsum = _mm_add_ps(vsum016, vsum2345);
79
80 __m128 vout = _mm_mul_ps(vsum, vscale);
81 vout = _mm_max_ps(vout, vmin);
82 vout = _mm_min_ps(vout, vmax);
83
84 _mm_storeu_ps(output, vout);
85 output += 4;
86
87 channels -= 4;
88 }
89 if (channels != 0) {
90 const __m128 vi0 = _mm_loadu_ps(i0);
91 const __m128 vi1 = _mm_loadu_ps(i1);
92 const __m128 vi2 = _mm_loadu_ps(i2);
93 const __m128 vi3 = _mm_loadu_ps(i3);
94 const __m128 vi4 = _mm_loadu_ps(i4);
95 const __m128 vi5 = _mm_loadu_ps(i5);
96 const __m128 vi6 = _mm_loadu_ps(i6);
97
98 const __m128 vsum01 = _mm_add_ps(vi0, vi1);
99 const __m128 vsum23 = _mm_add_ps(vi2, vi3);
100 const __m128 vsum45 = _mm_add_ps(vi4, vi5);
101
102 const __m128 vsum016 = _mm_add_ps(vsum01, vi6);
103 const __m128 vsum2345 = _mm_add_ps(vsum23, vsum45);
104
105 const __m128 vsum = _mm_add_ps(vsum016, vsum2345);
106
107 __m128 vout = _mm_mul_ps(vsum, vscale);
108 vout = _mm_max_ps(vout, vmin);
109 vout = _mm_min_ps(vout, vmax);
110
111 if (channels & 2) {
112 _mm_storel_pi((__m64*) output, vout);
113 vout = _mm_movehl_ps(vout, vout);
114 output += 2;
115 }
116 if (channels & 1) {
117 _mm_store_ss(output, vout);
118 }
119 }
120 }
121