1 // Auto-generated file. Do not edit!
2 // Template: src/f32-prelu/wasmsimd-bitselect.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2020 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 <wasm_simd128.h>
13
14 #include <xnnpack/math.h>
15 #include <xnnpack/prelu.h>
16
17
xnn_f32_prelu_ukernel__wasmsimd_bitselect_2x16(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__wasmsimd_bitselect_2x16(
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 do {
40 if XNN_UNPREDICTABLE(rows < 2) {
41 i1 = i0;
42 o1 = o0;
43 }
44
45 const float* w = weights;
46 size_t c = channels;
47 for (; c >= 16 * sizeof(float); c -= 16 * sizeof(float)) {
48 const v128_t vw0123 = wasm_v128_load(w);
49 const v128_t vw4567 = wasm_v128_load(w + 4);
50 const v128_t vw89AB = wasm_v128_load(w + 8);
51 const v128_t vwCDEF = wasm_v128_load(w + 12);
52 w += 16;
53
54 const v128_t vi0x0123 = wasm_v128_load(i0);
55 const v128_t vi0x4567 = wasm_v128_load(i0 + 4);
56 const v128_t vi0x89AB = wasm_v128_load(i0 + 8);
57 const v128_t vi0xCDEF = wasm_v128_load(i0 + 12);
58 i0 += 16;
59 const v128_t vi1x0123 = wasm_v128_load(i1);
60 const v128_t vi1x4567 = wasm_v128_load(i1 + 4);
61 const v128_t vi1x89AB = wasm_v128_load(i1 + 8);
62 const v128_t vi1xCDEF = wasm_v128_load(i1 + 12);
63 i1 += 16;
64
65 v128_t vacc0x0123 = wasm_f32x4_mul(vi0x0123, vw0123);
66 const v128_t vmask0x0123 = wasm_i32x4_shr(vi0x0123, 31);
67 v128_t vacc0x4567 = wasm_f32x4_mul(vi0x4567, vw4567);
68 const v128_t vmask0x4567 = wasm_i32x4_shr(vi0x4567, 31);
69 v128_t vacc0x89AB = wasm_f32x4_mul(vi0x89AB, vw89AB);
70 const v128_t vmask0x89AB = wasm_i32x4_shr(vi0x89AB, 31);
71 v128_t vacc0xCDEF = wasm_f32x4_mul(vi0xCDEF, vwCDEF);
72 const v128_t vmask0xCDEF = wasm_i32x4_shr(vi0xCDEF, 31);
73 v128_t vacc1x0123 = wasm_f32x4_mul(vi1x0123, vw0123);
74 const v128_t vmask1x0123 = wasm_i32x4_shr(vi1x0123, 31);
75 v128_t vacc1x4567 = wasm_f32x4_mul(vi1x4567, vw4567);
76 const v128_t vmask1x4567 = wasm_i32x4_shr(vi1x4567, 31);
77 v128_t vacc1x89AB = wasm_f32x4_mul(vi1x89AB, vw89AB);
78 const v128_t vmask1x89AB = wasm_i32x4_shr(vi1x89AB, 31);
79 v128_t vacc1xCDEF = wasm_f32x4_mul(vi1xCDEF, vwCDEF);
80 const v128_t vmask1xCDEF = wasm_i32x4_shr(vi1xCDEF, 31);
81
82 vacc0x0123 = wasm_v128_bitselect(vacc0x0123, vi0x0123, vmask0x0123);
83 vacc0x4567 = wasm_v128_bitselect(vacc0x4567, vi0x4567, vmask0x4567);
84 vacc0x89AB = wasm_v128_bitselect(vacc0x89AB, vi0x89AB, vmask0x89AB);
85 vacc0xCDEF = wasm_v128_bitselect(vacc0xCDEF, vi0xCDEF, vmask0xCDEF);
86 vacc1x0123 = wasm_v128_bitselect(vacc1x0123, vi1x0123, vmask1x0123);
87 vacc1x4567 = wasm_v128_bitselect(vacc1x4567, vi1x4567, vmask1x4567);
88 vacc1x89AB = wasm_v128_bitselect(vacc1x89AB, vi1x89AB, vmask1x89AB);
89 vacc1xCDEF = wasm_v128_bitselect(vacc1xCDEF, vi1xCDEF, vmask1xCDEF);
90
91 wasm_v128_store(o0, vacc0x0123);
92 wasm_v128_store(o0 + 4, vacc0x4567);
93 wasm_v128_store(o0 + 8, vacc0x89AB);
94 wasm_v128_store(o0 + 12, vacc0xCDEF);
95 o0 += 16;
96 wasm_v128_store(o1, vacc1x0123);
97 wasm_v128_store(o1 + 4, vacc1x4567);
98 wasm_v128_store(o1 + 8, vacc1x89AB);
99 wasm_v128_store(o1 + 12, vacc1xCDEF);
100 o1 += 16;
101 }
102 for (; c >= 4 * sizeof(float); c -= 4 * sizeof(float)) {
103 const v128_t vw0123 = wasm_v128_load(w);
104 w += 4;
105
106 const v128_t vi0x0123 = wasm_v128_load(i0);
107 i0 += 4;
108 const v128_t vi1x0123 = wasm_v128_load(i1);
109 i1 += 4;
110
111 v128_t vacc0x0123 = wasm_f32x4_mul(vi0x0123, vw0123);
112 const v128_t vmask0x0123 = wasm_i32x4_shr(vi0x0123, 31);
113 v128_t vacc1x0123 = wasm_f32x4_mul(vi1x0123, vw0123);
114 const v128_t vmask1x0123 = wasm_i32x4_shr(vi1x0123, 31);
115
116 vacc0x0123 = wasm_v128_bitselect(vacc0x0123, vi0x0123, vmask0x0123);
117 vacc1x0123 = wasm_v128_bitselect(vacc1x0123, vi1x0123, vmask1x0123);
118
119 wasm_v128_store(o0, vacc0x0123);
120 o0 += 4;
121 wasm_v128_store(o1, vacc1x0123);
122 o1 += 4;
123 }
124 if XNN_UNLIKELY(c != 0) {
125 const v128_t vw0123 = wasm_v128_load(w);
126 w = (const float*) ((uintptr_t) w + c);
127
128 const v128_t vi0x0123 = wasm_v128_load(i0);
129 i0 = (const float*) ((uintptr_t) i0 + c);
130 const v128_t vi1x0123 = wasm_v128_load(i1);
131 i1 = (const float*) ((uintptr_t) i1 + c);
132
133 v128_t vacc0x0123 = wasm_f32x4_mul(vi0x0123, vw0123);
134 const v128_t vmask0x0123 = wasm_i32x4_shr(vi0x0123, 31);
135 v128_t vacc1x0123 = wasm_f32x4_mul(vi1x0123, vw0123);
136 const v128_t vmask1x0123 = wasm_i32x4_shr(vi1x0123, 31);
137
138 vacc0x0123 = wasm_v128_bitselect(vacc0x0123, vi0x0123, vmask0x0123);
139 vacc1x0123 = wasm_v128_bitselect(vacc1x0123, vi1x0123, vmask1x0123);
140
141 if (c & (2 * sizeof(float))) {
142 *((double*) o0) = wasm_f64x2_extract_lane(vacc0x0123, 0);
143 *((double*) o1) = wasm_f64x2_extract_lane(vacc1x0123, 0);
144
145 vacc0x0123 = wasm_v32x4_shuffle(vacc0x0123, vacc0x0123, 2, 3, 2, 3);
146 vacc1x0123 = wasm_v32x4_shuffle(vacc1x0123, vacc1x0123, 2, 3, 2, 3);
147
148 o0 += 2;
149 o1 += 2;
150 }
151 if (c & (1 * sizeof(float))) {
152 *o0 = wasm_f32x4_extract_lane(vacc0x0123, 0);
153 *o1 = wasm_f32x4_extract_lane(vacc1x0123, 0);
154
155 o0 += 1;
156 o1 += 1;
157 }
158 }
159 i0 = (const float*) ((uintptr_t) i0 + input_increment);
160 o0 = (float*) ((uintptr_t) o0 + output_increment);
161 i1 = (const float*) ((uintptr_t) i1 + input_increment);
162 o1 = (float*) ((uintptr_t) o1 + output_increment);
163 rows = doz(rows, 2);
164 } while (rows != 0);
165 }
166