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 #include <xnnpack/math.h>
12
13
xnn_f32_gavgpool_cw_ukernel__sse_x4(size_t elements,size_t channels,const float * input,float * output,const union xnn_f32_gavgpool_params params[restrict XNN_MIN_ELEMENTS (1)])14 void xnn_f32_gavgpool_cw_ukernel__sse_x4(
15 size_t elements,
16 size_t channels,
17 const float* input,
18 float* output,
19 const union xnn_f32_gavgpool_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
20 {
21 assert(elements != 0);
22 assert(elements % sizeof(float) == 0);
23 assert(channels != 0);
24
25 const float* i0 = input;
26 const float* i1 = (const float*) ((uintptr_t) i0 + elements);
27 const float* i2 = (const float*) ((uintptr_t) i1 + elements);
28 const float* i3 = (const float*) ((uintptr_t) i2 + elements);
29
30 const __m128 vmask = _mm_load_ps((const float*) params->sse.mask);
31 const __m128 vmultiplier = _mm_load_ps(params->sse.multiplier);
32 const __m128 voutput_min = _mm_load_ps(params->sse.output_min);
33 const __m128 voutput_max = _mm_load_ps(params->sse.output_max);
34
35 while (channels >= 4) {
36 __m128 vsum0 = _mm_setzero_ps();
37 __m128 vsum1 = _mm_setzero_ps();
38 __m128 vsum2 = _mm_setzero_ps();
39 __m128 vsum3 = _mm_setzero_ps();
40 size_t n = elements;
41 while (n >= 4 * sizeof(float)) {
42 const __m128 vi0 = _mm_loadu_ps(i0);
43 i0 += 4;
44 const __m128 vi1 = _mm_loadu_ps(i1);
45 i1 += 4;
46 const __m128 vi2 = _mm_loadu_ps(i2);
47 i2 += 4;
48 const __m128 vi3 = _mm_loadu_ps(i3);
49 i3 += 4;
50
51 vsum0 = _mm_add_ps(vsum0, vi0);
52 vsum1 = _mm_add_ps(vsum1, vi1);
53 vsum2 = _mm_add_ps(vsum2, vi2);
54 vsum3 = _mm_add_ps(vsum3, vi3);
55 n -= 4 * sizeof(float);
56 }
57
58 if XNN_UNLIKELY(n != 0) {
59 const __m128 vi0 = _mm_and_ps(_mm_loadu_ps(i0), vmask);
60 i0 = (const float*) ((uintptr_t) i0 + n);
61 const __m128 vi1 = _mm_and_ps(_mm_loadu_ps(i1), vmask);
62 i1 = (const float*) ((uintptr_t) i1 + n);
63 const __m128 vi2 = _mm_and_ps(_mm_loadu_ps(i2), vmask);
64 i2 = (const float*) ((uintptr_t) i2 + n);
65 const __m128 vi3 = _mm_and_ps(_mm_loadu_ps(i3), vmask);
66 i3 = (const float*) ((uintptr_t) i3 + n);
67
68 vsum0 = _mm_add_ps(vsum0, vi0);
69 vsum1 = _mm_add_ps(vsum1, vi1);
70 vsum2 = _mm_add_ps(vsum2, vi2);
71 vsum3 = _mm_add_ps(vsum3, vi3);
72 }
73
74 // Having exactly 4 rows makes this work out nicely as we end up with
75 // the 4 totals in 4 different lanes of the same vector.
76 const __m128 vsum01 = _mm_add_ps(_mm_unpacklo_ps(vsum0, vsum1), _mm_unpackhi_ps(vsum0, vsum1));
77 const __m128 vsum23 = _mm_add_ps(_mm_unpacklo_ps(vsum2, vsum3), _mm_unpackhi_ps(vsum2, vsum3));
78 const __m128 vsum = _mm_add_ps(_mm_movelh_ps(vsum01, vsum23), _mm_movehl_ps(vsum23, vsum01));
79 __m128 vout = _mm_mul_ps(vsum, vmultiplier);
80
81 vout = _mm_max_ps(vout, voutput_min);
82 vout = _mm_min_ps(vout, voutput_max);
83
84 _mm_storeu_ps(output, vout);
85 output += 4;
86 i0 = i3;
87 i1 = (const float*) ((uintptr_t) i0 + elements);
88 i2 = (const float*) ((uintptr_t) i1 + elements);
89 i3 = (const float*) ((uintptr_t) i2 + elements);
90 channels -= 4;
91 }
92
93 while (channels != 0) {
94 __m128 vsum = _mm_setzero_ps();
95 size_t n = elements;
96 while (n >= 4 * sizeof(float)) {
97 const __m128 vi0 = _mm_loadu_ps(i0);
98 i0 += 4;
99 vsum = _mm_add_ps(vsum, vi0);
100 n -= 4 * sizeof(float);
101 }
102
103 if XNN_UNLIKELY(n != 0) {
104 __m128 vi0 = _mm_and_ps(_mm_loadu_ps(i0), vmask);
105 i0 = (const float*) ((uintptr_t) i0 + n);
106 vsum = _mm_add_ps(vsum, vi0);
107 }
108
109 vsum = _mm_add_ps(vsum, _mm_movehl_ps(vsum, vsum));
110 vsum = _mm_add_ss(vsum, _mm_shuffle_ps(vsum, vsum, _MM_SHUFFLE(3, 2, 1, 1)));
111
112 __m128 vout = _mm_mul_ss(vsum, vmultiplier);
113
114 vout = _mm_max_ss(vout, voutput_min);
115 vout = _mm_min_ss(vout, voutput_max);
116
117 _mm_store_ss(output, vout);
118 output += 1;
119 channels -= 1;
120 }
121 }
122