1 // Auto-generated file. Do not edit!
2 // Template: src/f32-prelu/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/math.h>
15 #include <xnnpack/prelu.h>
16
17
xnn_f32_prelu_ukernel__sse_2x8(size_t rows,size_t channels,const float * restrict input,size_t input_stride,const float * restrict weights,float * restrict output,size_t output_stride)18 void xnn_f32_prelu_ukernel__sse_2x8(
19 size_t rows,
20 size_t channels,
21 const float*restrict input,
22 size_t input_stride,
23 const float*restrict weights,
24 float*restrict output,
25 size_t output_stride) XNN_OOB_READS
26 {
27 assert(rows != 0);
28 assert(channels != 0);
29 assert(channels % sizeof(float) == 0);
30
31 const float* i0 = input;
32 float* o0 = output;
33 const float* i1 = (const float*) ((uintptr_t) i0 + input_stride);
34 float* o1 = (float*) ((uintptr_t) o0 + output_stride);
35
36 const size_t input_increment = input_stride * 2 - channels;
37 const size_t output_increment = output_stride * 2 - channels;
38
39 const __m128 vzero = _mm_setzero_ps();
40 do {
41 if XNN_UNPREDICTABLE(rows < 2) {
42 i1 = i0;
43 o1 = o0;
44 }
45
46 const float* w = weights;
47 size_t c = channels;
48 for (; c >= 8 * sizeof(float); c -= 8 * sizeof(float)) {
49 const __m128 vw0123 = _mm_load_ps(w);
50 const __m128 vw4567 = _mm_load_ps(w + 4);
51 w += 8;
52
53 __m128 vi0x0123 = _mm_loadu_ps(i0);
54 __m128 vi0x4567 = _mm_loadu_ps(i0 + 4);
55 i0 += 8;
56 __m128 vi1x0123 = _mm_loadu_ps(i1);
57 __m128 vi1x4567 = _mm_loadu_ps(i1 + 4);
58 i1 += 8;
59
60 __m128 vacc0x0123 = _mm_max_ps(_mm_setzero_ps(), vi0x0123);
61 vi0x0123 = _mm_min_ps(vi0x0123, vzero);
62 __m128 vacc0x4567 = _mm_max_ps(_mm_setzero_ps(), vi0x4567);
63 vi0x4567 = _mm_min_ps(vi0x4567, vzero);
64 __m128 vacc1x0123 = _mm_max_ps(_mm_setzero_ps(), vi1x0123);
65 vi1x0123 = _mm_min_ps(vi1x0123, vzero);
66 __m128 vacc1x4567 = _mm_max_ps(_mm_setzero_ps(), vi1x4567);
67 vi1x4567 = _mm_min_ps(vi1x4567, vzero);
68
69 vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(vi0x0123, vw0123));
70 vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(vi0x4567, vw4567));
71 vacc1x0123 = _mm_add_ps(vacc1x0123, _mm_mul_ps(vi1x0123, vw0123));
72 vacc1x4567 = _mm_add_ps(vacc1x4567, _mm_mul_ps(vi1x4567, vw4567));
73
74 _mm_storeu_ps(o0, vacc0x0123);
75 _mm_storeu_ps(o0 + 4, vacc0x4567);
76 o0 += 8;
77 _mm_storeu_ps(o1, vacc1x0123);
78 _mm_storeu_ps(o1 + 4, vacc1x4567);
79 o1 += 8;
80 }
81 for (; c >= 4 * sizeof(float); c -= 4 * sizeof(float)) {
82 const __m128 vw0123 = _mm_load_ps(w);
83 w += 4;
84
85 __m128 vi0x0123 = _mm_loadu_ps(i0);
86 i0 += 4;
87 __m128 vi1x0123 = _mm_loadu_ps(i1);
88 i1 += 4;
89
90 __m128 vacc0x0123 = _mm_max_ps(_mm_setzero_ps(), vi0x0123);
91 vi0x0123 = _mm_min_ps(vi0x0123, vzero);
92 __m128 vacc1x0123 = _mm_max_ps(_mm_setzero_ps(), vi1x0123);
93 vi1x0123 = _mm_min_ps(vi1x0123, vzero);
94
95 vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(vi0x0123, vw0123));
96 vacc1x0123 = _mm_add_ps(vacc1x0123, _mm_mul_ps(vi1x0123, vw0123));
97
98 _mm_storeu_ps(o0, vacc0x0123);
99 o0 += 4;
100 _mm_storeu_ps(o1, vacc1x0123);
101 o1 += 4;
102 }
103 if XNN_UNLIKELY(c != 0) {
104 const __m128 vw0123 = _mm_load_ps(w);
105 w = (const float*) ((uintptr_t) w + c);
106
107 __m128 vi0x0123 = _mm_loadu_ps(i0);
108 i0 = (const float*) ((uintptr_t) i0 + c);
109 __m128 vi1x0123 = _mm_loadu_ps(i1);
110 i1 = (const float*) ((uintptr_t) i1 + c);
111
112 __m128 vacc0x0123 = _mm_max_ps(_mm_setzero_ps(), vi0x0123);
113 vi0x0123 = _mm_min_ps(vi0x0123, vzero);
114 __m128 vacc1x0123 = _mm_max_ps(_mm_setzero_ps(), vi1x0123);
115 vi1x0123 = _mm_min_ps(vi1x0123, vzero);
116
117 vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(vi0x0123, vw0123));
118 vacc1x0123 = _mm_add_ps(vacc1x0123, _mm_mul_ps(vi1x0123, vw0123));
119
120 if (c & (2 * sizeof(float))) {
121 _mm_storel_pi((__m64*) o0, vacc0x0123);
122 _mm_storel_pi((__m64*) o1, vacc1x0123);
123
124 vacc0x0123 = _mm_movehl_ps(vacc0x0123, vacc0x0123);
125 vacc1x0123 = _mm_movehl_ps(vacc1x0123, vacc1x0123);
126
127 o0 += 2;
128 o1 += 2;
129 }
130 if (c & (1 * sizeof(float))) {
131 _mm_store_ss(o0, vacc0x0123);
132 _mm_store_ss(o1, vacc1x0123);
133
134 o0 += 1;
135 o1 += 1;
136 }
137 }
138 i0 = (const float*) ((uintptr_t) i0 + input_increment);
139 o0 = (float*) ((uintptr_t) o0 + output_increment);
140 i1 = (const float*) ((uintptr_t) i1 + input_increment);
141 o1 = (float*) ((uintptr_t) o1 + output_increment);
142 rows = doz(rows, 2);
143 } while (rows != 0);
144 }
145