xref: /aosp_15_r20/external/XNNPACK/src/f32-prelu/gen/wasm-2x4.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-prelu/wasm.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 <xnnpack/math.h>
13 #include <xnnpack/prelu.h>
14 
15 
xnn_f32_prelu_ukernel__wasm_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)16 void xnn_f32_prelu_ukernel__wasm_2x4(
17     size_t rows,
18     size_t channels,
19     const float*restrict input,
20     size_t input_stride,
21     const float*restrict weights,
22     float*restrict output,
23     size_t output_stride)
24 {
25   assert(rows != 0);
26   assert(channels != 0);
27   assert(channels % sizeof(float) == 0);
28 
29   const float* i0 = input;
30   float* o0 = output;
31   const float* i1 = (const float*) ((uintptr_t) i0 + input_stride);
32   float* o1 = (float*) ((uintptr_t) o0 + output_stride);
33 
34   const size_t input_increment = input_stride * 2 - channels;
35   const size_t output_increment = output_stride * 2 - channels;
36 
37   const float vzero = 0.0f;
38   do {
39     if XNN_UNPREDICTABLE(rows < 2) {
40       i1 = i0;
41       o1 = o0;
42     }
43 
44     const float* w = weights;
45     size_t c = channels;
46     for (; c >= 4 * sizeof(float); c -= 4 * sizeof(float)) {
47       const float vw0 = w[0];
48       const float vw1 = w[1];
49       const float vw2 = w[2];
50       const float vw3 = w[3];
51 
52       float vi0x0 = i0[0];
53       float vi0x1 = i0[1];
54       float vi0x2 = i0[2];
55       float vi0x3 = i0[3];
56       i0 += 4;
57       float vi1x0 = i1[0];
58       float vi1x1 = i1[1];
59       float vi1x2 = i1[2];
60       float vi1x3 = i1[3];
61       i1 += 4;
62 
63       float vacc0x0 = __builtin_wasm_max_f32(vi0x0, vzero);
64       vi0x0 = __builtin_wasm_min_f32(vi0x0, vzero);
65       float vacc0x1 = __builtin_wasm_max_f32(vi0x1, vzero);
66       vi0x1 = __builtin_wasm_min_f32(vi0x1, vzero);
67       float vacc0x2 = __builtin_wasm_max_f32(vi0x2, vzero);
68       vi0x2 = __builtin_wasm_min_f32(vi0x2, vzero);
69       float vacc0x3 = __builtin_wasm_max_f32(vi0x3, vzero);
70       vi0x3 = __builtin_wasm_min_f32(vi0x3, vzero);
71       float vacc1x0 = __builtin_wasm_max_f32(vi1x0, vzero);
72       vi1x0 = __builtin_wasm_min_f32(vi1x0, vzero);
73       float vacc1x1 = __builtin_wasm_max_f32(vi1x1, vzero);
74       vi1x1 = __builtin_wasm_min_f32(vi1x1, vzero);
75       float vacc1x2 = __builtin_wasm_max_f32(vi1x2, vzero);
76       vi1x2 = __builtin_wasm_min_f32(vi1x2, vzero);
77       float vacc1x3 = __builtin_wasm_max_f32(vi1x3, vzero);
78       vi1x3 = __builtin_wasm_min_f32(vi1x3, vzero);
79 
80       vacc0x0 += vi0x0 * vw0;
81       vacc0x1 += vi0x1 * vw1;
82       vacc0x2 += vi0x2 * vw2;
83       vacc0x3 += vi0x3 * vw3;
84       vacc1x0 += vi1x0 * vw0;
85       vacc1x1 += vi1x1 * vw1;
86       vacc1x2 += vi1x2 * vw2;
87       vacc1x3 += vi1x3 * vw3;
88 
89       o0[0] = vacc0x0;
90       o0[1] = vacc0x1;
91       o0[2] = vacc0x2;
92       o0[3] = vacc0x3;
93       o0 += 4;
94       o1[0] = vacc1x0;
95       o1[1] = vacc1x1;
96       o1[2] = vacc1x2;
97       o1[3] = vacc1x3;
98       o1 += 4;
99 
100       w += 4;
101     }
102     for (; c != 0; c -= sizeof(float)) {
103       const float vw = *w++;
104 
105       float vi0 = *i0++;
106       float vi1 = *i1++;
107 
108       float vacc0 = __builtin_wasm_max_f32(vi0, vzero);
109       vi0 = __builtin_wasm_min_f32(vi0, vzero);
110       float vacc1 = __builtin_wasm_max_f32(vi1, vzero);
111       vi1 = __builtin_wasm_min_f32(vi1, vzero);
112 
113       vacc0 += vi0 * vw;
114       vacc1 += vi1 * vw;
115 
116       *o0++ = vacc0;
117       *o1++ = vacc1;
118     }
119     i0 = (const float*) ((uintptr_t) i0 + input_increment);
120     o0 = (float*) ((uintptr_t) o0 + output_increment);
121     i1 = (const float*) ((uintptr_t) i1 + input_increment);
122     o1 = (float*) ((uintptr_t) o1 + output_increment);
123     rows = doz(rows, 2);
124   } while (rows != 0);
125 }
126