1 // Auto-generated file. Do not edit!
2 // Template: src/f32-prelu/wasmsimd-minmax.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_minmax_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_minmax_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 const v128_t vzero = wasm_i32x4_const_splat(0);
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 >= 16 * sizeof(float); c -= 16 * sizeof(float)) {
49 const v128_t vw0123 = wasm_v128_load(w);
50 const v128_t vw4567 = wasm_v128_load(w + 4);
51 const v128_t vw89AB = wasm_v128_load(w + 8);
52 const v128_t vwCDEF = wasm_v128_load(w + 12);
53 w += 16;
54
55 v128_t vi0x0123 = wasm_v128_load(i0);
56 v128_t vi0x4567 = wasm_v128_load(i0 + 4);
57 v128_t vi0x89AB = wasm_v128_load(i0 + 8);
58 v128_t vi0xCDEF = wasm_v128_load(i0 + 12);
59 i0 += 16;
60 v128_t vi1x0123 = wasm_v128_load(i1);
61 v128_t vi1x4567 = wasm_v128_load(i1 + 4);
62 v128_t vi1x89AB = wasm_v128_load(i1 + 8);
63 v128_t vi1xCDEF = wasm_v128_load(i1 + 12);
64 i1 += 16;
65
66 v128_t vacc0x0123 = wasm_i32x4_max(vi0x0123, vzero);
67 vi0x0123 = wasm_i32x4_min(vi0x0123, vzero);
68 v128_t vacc0x4567 = wasm_i32x4_max(vi0x4567, vzero);
69 vi0x4567 = wasm_i32x4_min(vi0x4567, vzero);
70 v128_t vacc0x89AB = wasm_i32x4_max(vi0x89AB, vzero);
71 vi0x89AB = wasm_i32x4_min(vi0x89AB, vzero);
72 v128_t vacc0xCDEF = wasm_i32x4_max(vi0xCDEF, vzero);
73 vi0xCDEF = wasm_i32x4_min(vi0xCDEF, vzero);
74 v128_t vacc1x0123 = wasm_i32x4_max(vi1x0123, vzero);
75 vi1x0123 = wasm_i32x4_min(vi1x0123, vzero);
76 v128_t vacc1x4567 = wasm_i32x4_max(vi1x4567, vzero);
77 vi1x4567 = wasm_i32x4_min(vi1x4567, vzero);
78 v128_t vacc1x89AB = wasm_i32x4_max(vi1x89AB, vzero);
79 vi1x89AB = wasm_i32x4_min(vi1x89AB, vzero);
80 v128_t vacc1xCDEF = wasm_i32x4_max(vi1xCDEF, vzero);
81 vi1xCDEF = wasm_i32x4_min(vi1xCDEF, vzero);
82
83 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(vi0x0123, vw0123));
84 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(vi0x4567, vw4567));
85 vacc0x89AB = wasm_f32x4_add(vacc0x89AB, wasm_f32x4_mul(vi0x89AB, vw89AB));
86 vacc0xCDEF = wasm_f32x4_add(vacc0xCDEF, wasm_f32x4_mul(vi0xCDEF, vwCDEF));
87 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(vi1x0123, vw0123));
88 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(vi1x4567, vw4567));
89 vacc1x89AB = wasm_f32x4_add(vacc1x89AB, wasm_f32x4_mul(vi1x89AB, vw89AB));
90 vacc1xCDEF = wasm_f32x4_add(vacc1xCDEF, wasm_f32x4_mul(vi1xCDEF, vwCDEF));
91
92 wasm_v128_store(o0, vacc0x0123);
93 wasm_v128_store(o0 + 4, vacc0x4567);
94 wasm_v128_store(o0 + 8, vacc0x89AB);
95 wasm_v128_store(o0 + 12, vacc0xCDEF);
96 o0 += 16;
97 wasm_v128_store(o1, vacc1x0123);
98 wasm_v128_store(o1 + 4, vacc1x4567);
99 wasm_v128_store(o1 + 8, vacc1x89AB);
100 wasm_v128_store(o1 + 12, vacc1xCDEF);
101 o1 += 16;
102 }
103 for (; c >= 4 * sizeof(float); c -= 4 * sizeof(float)) {
104 const v128_t vw0123 = wasm_v128_load(w);
105 w += 4;
106
107 v128_t vi0x0123 = wasm_v128_load(i0);
108 i0 += 4;
109 v128_t vi1x0123 = wasm_v128_load(i1);
110 i1 += 4;
111
112 v128_t vacc0x0123 = wasm_i32x4_max(vi0x0123, vzero);
113 vi0x0123 = wasm_i32x4_min(vi0x0123, vzero);
114 v128_t vacc1x0123 = wasm_i32x4_max(vi1x0123, vzero);
115 vi1x0123 = wasm_i32x4_min(vi1x0123, vzero);
116
117 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(vi0x0123, vw0123));
118 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(vi1x0123, vw0123));
119
120 wasm_v128_store(o0, vacc0x0123);
121 o0 += 4;
122 wasm_v128_store(o1, vacc1x0123);
123 o1 += 4;
124 }
125 if XNN_UNLIKELY(c != 0) {
126 const v128_t vw0123 = wasm_v128_load(w);
127 w = (const float*) ((uintptr_t) w + c);
128
129 v128_t vi0x0123 = wasm_v128_load(i0);
130 i0 = (const float*) ((uintptr_t) i0 + c);
131 v128_t vi1x0123 = wasm_v128_load(i1);
132 i1 = (const float*) ((uintptr_t) i1 + c);
133
134 v128_t vacc0x0123 = wasm_i32x4_max(vi0x0123, vzero);
135 vi0x0123 = wasm_i32x4_min(vi0x0123, vzero);
136 v128_t vacc1x0123 = wasm_i32x4_max(vi1x0123, vzero);
137 vi1x0123 = wasm_i32x4_min(vi1x0123, vzero);
138
139 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(vi0x0123, vw0123));
140 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(vi1x0123, vw0123));
141
142 if (c & (2 * sizeof(float))) {
143 *((double*) o0) = wasm_f64x2_extract_lane(vacc0x0123, 0);
144 *((double*) o1) = wasm_f64x2_extract_lane(vacc1x0123, 0);
145
146 vacc0x0123 = wasm_v32x4_shuffle(vacc0x0123, vacc0x0123, 2, 3, 2, 3);
147 vacc1x0123 = wasm_v32x4_shuffle(vacc1x0123, vacc1x0123, 2, 3, 2, 3);
148
149 o0 += 2;
150 o1 += 2;
151 }
152 if (c & (1 * sizeof(float))) {
153 *o0 = wasm_f32x4_extract_lane(vacc0x0123, 0);
154 *o1 = wasm_f32x4_extract_lane(vacc1x0123, 0);
155
156 o0 += 1;
157 o1 += 1;
158 }
159 }
160 i0 = (const float*) ((uintptr_t) i0 + input_increment);
161 o0 = (float*) ((uintptr_t) o0 + output_increment);
162 i1 = (const float*) ((uintptr_t) i1 + input_increment);
163 o1 = (float*) ((uintptr_t) o1 + output_increment);
164 rows = doz(rows, 2);
165 } while (rows != 0);
166 }
167