xref: /aosp_15_r20/external/XNNPACK/src/f32-prelu/gen/wasmsimd-minmax-1x4.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
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_1x4(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_1x4(
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 
34   const size_t input_increment = input_stride * 1 - channels;
35   const size_t output_increment = output_stride * 1 - channels;
36 
37   const v128_t vzero = wasm_i32x4_const_splat(0);
38   do {
39 
40     const float* w = weights;
41     size_t c = channels;
42     for (; c >= 4 * sizeof(float); c -= 4 * sizeof(float)) {
43       const v128_t vw0123 = wasm_v128_load(w);
44       w += 4;
45 
46       v128_t vi0x0123 = wasm_v128_load(i0);
47       i0 += 4;
48 
49       v128_t vacc0x0123 = wasm_i32x4_max(vi0x0123, vzero);
50       vi0x0123 = wasm_i32x4_min(vi0x0123, vzero);
51 
52       vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(vi0x0123, vw0123));
53 
54       wasm_v128_store(o0, vacc0x0123);
55       o0 += 4;
56     }
57     if XNN_UNLIKELY(c != 0) {
58       const v128_t vw0123 = wasm_v128_load(w);
59       w = (const float*) ((uintptr_t) w + c);
60 
61       v128_t vi0x0123 = wasm_v128_load(i0);
62       i0 = (const float*) ((uintptr_t) i0 + c);
63 
64       v128_t vacc0x0123 = wasm_i32x4_max(vi0x0123, vzero);
65       vi0x0123 = wasm_i32x4_min(vi0x0123, vzero);
66 
67       vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(vi0x0123, vw0123));
68 
69       if (c & (2 * sizeof(float))) {
70         *((double*) o0) = wasm_f64x2_extract_lane(vacc0x0123, 0);
71 
72         vacc0x0123 = wasm_v32x4_shuffle(vacc0x0123, vacc0x0123, 2, 3, 2, 3);
73 
74         o0 += 2;
75       }
76       if (c & (1 * sizeof(float))) {
77         *o0 = wasm_f32x4_extract_lane(vacc0x0123, 0);
78 
79         o0 += 1;
80       }
81     }
82     i0 = (const float*) ((uintptr_t) i0 + input_increment);
83     o0 = (float*) ((uintptr_t) o0 + output_increment);
84     rows = doz(rows, 1);
85   } while (rows != 0);
86 }
87