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_4x4(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_4x4(
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 const float* i2 = (const float*) ((uintptr_t) i1 + input_stride);
36 float* o2 = (float*) ((uintptr_t) o1 + output_stride);
37 const float* i3 = (const float*) ((uintptr_t) i2 + input_stride);
38 float* o3 = (float*) ((uintptr_t) o2 + output_stride);
39
40 const size_t input_increment = input_stride * 4 - channels;
41 const size_t output_increment = output_stride * 4 - channels;
42
43 const v128_t vzero = wasm_i32x4_const_splat(0);
44 do {
45 if XNN_UNPREDICTABLE(rows < 2) {
46 i1 = i0;
47 o1 = o0;
48 }
49 if XNN_UNPREDICTABLE(rows <= 2) {
50 i2 = i1;
51 o2 = o1;
52 }
53 if XNN_UNPREDICTABLE(rows < 4) {
54 i3 = i2;
55 o3 = o2;
56 }
57
58 const float* w = weights;
59 size_t c = channels;
60 for (; c >= 4 * sizeof(float); c -= 4 * sizeof(float)) {
61 const v128_t vw0123 = wasm_v128_load(w);
62 w += 4;
63
64 v128_t vi0x0123 = wasm_v128_load(i0);
65 i0 += 4;
66 v128_t vi1x0123 = wasm_v128_load(i1);
67 i1 += 4;
68 v128_t vi2x0123 = wasm_v128_load(i2);
69 i2 += 4;
70 v128_t vi3x0123 = wasm_v128_load(i3);
71 i3 += 4;
72
73 v128_t vacc0x0123 = wasm_i32x4_max(vi0x0123, vzero);
74 vi0x0123 = wasm_i32x4_min(vi0x0123, vzero);
75 v128_t vacc1x0123 = wasm_i32x4_max(vi1x0123, vzero);
76 vi1x0123 = wasm_i32x4_min(vi1x0123, vzero);
77 v128_t vacc2x0123 = wasm_i32x4_max(vi2x0123, vzero);
78 vi2x0123 = wasm_i32x4_min(vi2x0123, vzero);
79 v128_t vacc3x0123 = wasm_i32x4_max(vi3x0123, vzero);
80 vi3x0123 = wasm_i32x4_min(vi3x0123, vzero);
81
82 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(vi0x0123, vw0123));
83 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(vi1x0123, vw0123));
84 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(vi2x0123, vw0123));
85 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(vi3x0123, vw0123));
86
87 wasm_v128_store(o0, vacc0x0123);
88 o0 += 4;
89 wasm_v128_store(o1, vacc1x0123);
90 o1 += 4;
91 wasm_v128_store(o2, vacc2x0123);
92 o2 += 4;
93 wasm_v128_store(o3, vacc3x0123);
94 o3 += 4;
95 }
96 if XNN_UNLIKELY(c != 0) {
97 const v128_t vw0123 = wasm_v128_load(w);
98 w = (const float*) ((uintptr_t) w + c);
99
100 v128_t vi0x0123 = wasm_v128_load(i0);
101 i0 = (const float*) ((uintptr_t) i0 + c);
102 v128_t vi1x0123 = wasm_v128_load(i1);
103 i1 = (const float*) ((uintptr_t) i1 + c);
104 v128_t vi2x0123 = wasm_v128_load(i2);
105 i2 = (const float*) ((uintptr_t) i2 + c);
106 v128_t vi3x0123 = wasm_v128_load(i3);
107 i3 = (const float*) ((uintptr_t) i3 + c);
108
109 v128_t vacc0x0123 = wasm_i32x4_max(vi0x0123, vzero);
110 vi0x0123 = wasm_i32x4_min(vi0x0123, vzero);
111 v128_t vacc1x0123 = wasm_i32x4_max(vi1x0123, vzero);
112 vi1x0123 = wasm_i32x4_min(vi1x0123, vzero);
113 v128_t vacc2x0123 = wasm_i32x4_max(vi2x0123, vzero);
114 vi2x0123 = wasm_i32x4_min(vi2x0123, vzero);
115 v128_t vacc3x0123 = wasm_i32x4_max(vi3x0123, vzero);
116 vi3x0123 = wasm_i32x4_min(vi3x0123, vzero);
117
118 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(vi0x0123, vw0123));
119 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(vi1x0123, vw0123));
120 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(vi2x0123, vw0123));
121 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(vi3x0123, vw0123));
122
123 if (c & (2 * sizeof(float))) {
124 *((double*) o0) = wasm_f64x2_extract_lane(vacc0x0123, 0);
125 *((double*) o1) = wasm_f64x2_extract_lane(vacc1x0123, 0);
126 *((double*) o2) = wasm_f64x2_extract_lane(vacc2x0123, 0);
127 *((double*) o3) = wasm_f64x2_extract_lane(vacc3x0123, 0);
128
129 vacc0x0123 = wasm_v32x4_shuffle(vacc0x0123, vacc0x0123, 2, 3, 2, 3);
130 vacc1x0123 = wasm_v32x4_shuffle(vacc1x0123, vacc1x0123, 2, 3, 2, 3);
131 vacc2x0123 = wasm_v32x4_shuffle(vacc2x0123, vacc2x0123, 2, 3, 2, 3);
132 vacc3x0123 = wasm_v32x4_shuffle(vacc3x0123, vacc3x0123, 2, 3, 2, 3);
133
134 o0 += 2;
135 o1 += 2;
136 o2 += 2;
137 o3 += 2;
138 }
139 if (c & (1 * sizeof(float))) {
140 *o0 = wasm_f32x4_extract_lane(vacc0x0123, 0);
141 *o1 = wasm_f32x4_extract_lane(vacc1x0123, 0);
142 *o2 = wasm_f32x4_extract_lane(vacc2x0123, 0);
143 *o3 = wasm_f32x4_extract_lane(vacc3x0123, 0);
144
145 o0 += 1;
146 o1 += 1;
147 o2 += 1;
148 o3 += 1;
149 }
150 }
151 i0 = (const float*) ((uintptr_t) i0 + input_increment);
152 o0 = (float*) ((uintptr_t) o0 + output_increment);
153 i1 = (const float*) ((uintptr_t) i1 + input_increment);
154 o1 = (float*) ((uintptr_t) o1 + output_increment);
155 i2 = (const float*) ((uintptr_t) i2 + input_increment);
156 o2 = (float*) ((uintptr_t) o2 + output_increment);
157 i3 = (const float*) ((uintptr_t) i3 + input_increment);
158 o3 = (float*) ((uintptr_t) o3 + output_increment);
159 rows = doz(rows, 4);
160 } while (rows != 0);
161 }
162