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 <smmintrin.h>
13
14 #include <xnnpack/math.h>
15 #include <xnnpack/prelu.h>
16
17
xnn_f32_prelu_ukernel__sse41_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__sse41_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_DISABLE_TSAN
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 if XNN_UNPREDICTABLE(rows < 2) {
36 i1 = i0;
37 o1 = o0;
38 }
39
40 const size_t input_increment = input_stride * 2 - channels;
41 const size_t output_increment = output_stride * 2 - channels;
42
43 do {
44 const float* w = weights;
45 size_t c = channels;
46 for (; c >= 8 * sizeof(float); c -= 8 * sizeof(float)) {
47 const __m128 vw0123 = _mm_load_ps(w);
48 const __m128 vw4567 = _mm_load_ps(w + 4);
49 w += 8;
50
51 const __m128 vi0x0123 = _mm_loadu_ps(i0);
52 const __m128 vi0x4567 = _mm_loadu_ps(i0 + 4);
53 i0 += 8;
54 const __m128 vi1x0123 = _mm_loadu_ps(i1);
55 const __m128 vi1x4567 = _mm_loadu_ps(i1 + 4);
56 i1 += 8;
57
58 const __m128 vprod0x0123 = _mm_mul_ps(vi0x0123, vw0123);
59 const __m128 vprod0x4567 = _mm_mul_ps(vi0x4567, vw4567);
60 const __m128 vprod1x0123 = _mm_mul_ps(vi1x0123, vw0123);
61 const __m128 vprod1x4567 = _mm_mul_ps(vi1x4567, vw4567);
62
63 const __m128 vacc0x0123 = _mm_blendv_ps(vi0x0123, vprod0x0123, vi0x0123);
64 const __m128 vacc0x4567 = _mm_blendv_ps(vi0x4567, vprod0x4567, vi0x4567);
65 const __m128 vacc1x0123 = _mm_blendv_ps(vi1x0123, vprod1x0123, vi1x0123);
66 const __m128 vacc1x4567 = _mm_blendv_ps(vi1x4567, vprod1x4567, vi1x4567);
67
68 _mm_storeu_ps(o0, vacc0x0123);
69 _mm_storeu_ps(o0 + 4, vacc0x4567);
70 o0 += 8;
71 _mm_storeu_ps(o1, vacc1x0123);
72 _mm_storeu_ps(o1 + 4, vacc1x4567);
73 o1 += 8;
74 }
75 for (; c >= 4 * sizeof(float); c -= 4 * sizeof(float)) {
76 const __m128 vw0123 = _mm_load_ps(w);
77 w += 4;
78
79 const __m128 vi0x0123 = _mm_loadu_ps(i0);
80 i0 += 4;
81 const __m128 vi1x0123 = _mm_loadu_ps(i1);
82 i1 += 4;
83
84 const __m128 vprod0x0123 = _mm_mul_ps(vi0x0123, vw0123);
85 const __m128 vprod1x0123 = _mm_mul_ps(vi1x0123, vw0123);
86
87 __m128 vacc0x0123 = _mm_blendv_ps(vi0x0123, vprod0x0123, vi0x0123);
88 __m128 vacc1x0123 = _mm_blendv_ps(vi1x0123, vprod1x0123, vi1x0123);
89
90 _mm_storeu_ps(o0, vacc0x0123);
91 o0 += 4;
92 _mm_storeu_ps(o1, vacc1x0123);
93 o1 += 4;
94 }
95 if XNN_UNLIKELY(c != 0) {
96 const __m128 vw0123 = _mm_load_ps(w);
97 w = (const float*) ((uintptr_t) w + c);
98
99 const __m128 vi0x0123 = _mm_loadu_ps(i0);
100 i0 = (const float*) ((uintptr_t) i0 + c);
101 const __m128 vi1x0123 = _mm_loadu_ps(i1);
102 i1 = (const float*) ((uintptr_t) i1 + c);
103
104 const __m128 vprod0x0123 = _mm_mul_ps(vi0x0123, vw0123);
105 const __m128 vprod1x0123 = _mm_mul_ps(vi1x0123, vw0123);
106
107 __m128 vacc0x0123 = _mm_blendv_ps(vi0x0123, vprod0x0123, vi0x0123);
108 __m128 vacc1x0123 = _mm_blendv_ps(vi1x0123, vprod1x0123, vi1x0123);
109
110 if (c & (2 * sizeof(float))) {
111 _mm_storel_pi((__m64*) o0, vacc0x0123);
112 _mm_storel_pi((__m64*) o1, vacc1x0123);
113
114 vacc0x0123 = _mm_movehl_ps(vacc0x0123, vacc0x0123);
115 vacc1x0123 = _mm_movehl_ps(vacc1x0123, vacc1x0123);
116
117 o0 += 2;
118 o1 += 2;
119 }
120 if (c & (1 * sizeof(float))) {
121 _mm_store_ss(o0, vacc0x0123);
122 _mm_store_ss(o1, vacc1x0123);
123
124 o0 += 1;
125 o1 += 1;
126 }
127 }
128 i0 = (const float*) ((uintptr_t) i0 + input_increment);
129 o0 = (float*) ((uintptr_t) o0 + output_increment);
130 i1 = (const float*) ((uintptr_t) i1 + input_increment);
131 o1 = (float*) ((uintptr_t) o1 + output_increment);
132 if XNN_UNPREDICTABLE(rows < 4) {
133 i1 = i0;
134 o1 = o0;
135 }
136 rows = doz(rows, 2);
137 } while (rows != 0);
138 }
139