1 // Auto-generated file. Do not edit!
2 // Template: src/f32-prelu/psimd.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 <psimd.h>
13
14 #include <xnnpack/math.h>
15 #include <xnnpack/prelu.h>
16
17
xnn_f32_prelu_ukernel__psimd_2x4(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,const union xnn_f32_output_params params[restrict static1])18 void xnn_f32_prelu_ukernel__psimd_2x4(
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,
26 const union xnn_f32_output_params params[restrict static 1])
27 {
28 assert(rows != 0);
29 assert(channels != 0);
30 assert(channels % sizeof(float) == 0);
31
32 const float* i0 = input;
33 float* o0 = output;
34 const float* i1 = (const float*) ((uintptr_t) i0 + input_stride);
35 float* o1 = (float*) ((uintptr_t) o0 + output_stride);
36 if XNN_UNPREDICTABLE(rows < 2) {
37 i1 = i0;
38 o1 = o0;
39 }
40
41 const size_t input_increment = input_stride * 2 - channels;
42 const size_t output_increment = output_stride * 2 - channels;
43
44 const psimd_f32 vmin = psimd_load_splat_f32(¶ms->scalar.min);
45 const psimd_f32 vmax = psimd_load_splat_f32(¶ms->scalar.max);
46 do {
47 const float* w = weights;
48 size_t c = channels;
49 for (; c >= 4 * sizeof(float); c -= 4 * sizeof(float)) {
50 const psimd_f32 vw0123 = psimd_load_f32(w);
51 w += 4;
52
53 const psimd_f32 vi0x0123 = psimd_load_f32(i0);
54 i0 += 4;
55 const psimd_f32 vi1x0123 = psimd_load_f32(i1);
56 i1 += 4;
57
58 psimd_f32 vacc0x0123 = psimd_mul_f32(vi0x0123, vw0123);
59 psimd_f32 vacc1x0123 = psimd_mul_f32(vi1x0123, vw0123);
60
61 vacc0x0123 = psimd_signblend_f32(vi0x0123, vacc0x0123, vi0x0123);
62 vacc1x0123 = psimd_signblend_f32(vi1x0123, vacc1x0123, vi1x0123);
63
64 vacc0x0123 = psimd_max_f32(vacc0x0123, vmin);
65 vacc1x0123 = psimd_max_f32(vacc1x0123, vmin);
66
67 vacc0x0123 = psimd_min_f32(vacc0x0123, vmax);
68 vacc1x0123 = psimd_min_f32(vacc1x0123, vmax);
69
70 psimd_store_f32(o0, vacc0x0123);
71 o0 += 4;
72 psimd_store_f32(o1, vacc1x0123);
73 o1 += 4;
74 }
75 if XNN_UNLIKELY(c != 0) {
76 const psimd_f32 vw0123 = psimd_load_f32(w);
77 w = (const float*) ((uintptr_t) w + c);
78
79 const psimd_f32 vi0x0123 = psimd_load_f32(i0);
80 i0 = (const float*) ((uintptr_t) i0 + c);
81 const psimd_f32 vi1x0123 = psimd_load_f32(i1);
82 i1 = (const float*) ((uintptr_t) i1 + c);
83
84 psimd_f32 vacc0x0123 = psimd_mul_f32(vi0x0123, vw0123);
85 psimd_f32 vacc1x0123 = psimd_mul_f32(vi1x0123, vw0123);
86
87 vacc0x0123 = psimd_signblend_f32(vi0x0123, vacc0x0123, vi0x0123);
88 vacc1x0123 = psimd_signblend_f32(vi1x0123, vacc1x0123, vi1x0123);
89
90 vacc0x0123 = psimd_max_f32(vacc0x0123, vmin);
91 vacc1x0123 = psimd_max_f32(vacc1x0123, vmin);
92
93 vacc0x0123 = psimd_min_f32(vacc0x0123, vmax);
94 vacc1x0123 = psimd_min_f32(vacc1x0123, vmax);
95
96 if (c & (2 * sizeof(float))) {
97 psimd_store2_f32(o0, vacc0x0123);
98 psimd_store2_f32(o1, vacc1x0123);
99
100 vacc0x0123 = psimd_concat_hi_f32(vacc0x0123, vacc0x0123);
101 vacc1x0123 = psimd_concat_hi_f32(vacc1x0123, vacc1x0123);
102
103 o0 += 2;
104 o1 += 2;
105 }
106 if (c & (1 * sizeof(float))) {
107 psimd_store1_f32(o0, vacc0x0123);
108 psimd_store1_f32(o1, vacc1x0123);
109
110 o0 += 1;
111 o1 += 1;
112 }
113 }
114 i0 = (const float*) ((uintptr_t) i0 + input_increment);
115 o0 = (float*) ((uintptr_t) o0 + output_increment);
116 i1 = (const float*) ((uintptr_t) i1 + input_increment);
117 o1 = (float*) ((uintptr_t) o1 + output_increment);
118 if XNN_UNPREDICTABLE(rows < 4) {
119 i1 = i0;
120 o1 = o0;
121 }
122 rows = doz(rows, 2);
123 } while (rows != 0);
124 }
125