xref: /aosp_15_r20/external/XNNPACK/src/f32-vrelu/gen/vrelu-scalar-x2.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-vrelu/scalar.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/vunary.h>
13 #include <xnnpack/common.h>
14 #include <xnnpack/math.h>
15 
xnn_f32_vrelu_ukernel__scalar_x2(size_t n,const float * x_ptr,float * y_ptr,const union xnn_f32_relu_params params[restrict XNN_MIN_ELEMENTS (1)])16 void xnn_f32_vrelu_ukernel__scalar_x2(
17     size_t n,
18     const float* x_ptr,
19     float* y_ptr,
20     const union xnn_f32_relu_params params[restrict XNN_MIN_ELEMENTS(1)])
21 {
22   assert(n != 0);
23   assert(n % sizeof(float) == 0);
24   assert(x_ptr != NULL);
25   assert(y_ptr != NULL);
26 
27   const uint32_t* x = (const uint32_t*)x_ptr;
28   uint32_t* y = (uint32_t*)y_ptr;
29 
30   for (; n >= 2 * sizeof(uint32_t); n -= 2 * sizeof(uint32_t)) {
31     uint32_t vacc0 = x[0];
32     uint32_t vacc1 = x[1];
33     x += 2;
34 
35     vacc0 = ((vacc0 >> 31) - 1) & vacc0;
36     vacc1 = ((vacc1 >> 31) - 1) & vacc1;
37 
38     y[0] = vacc0;
39     y[1] = vacc1;
40     y += 2;
41   }
42   if XNN_UNLIKELY(n != 0) {
43     uint32_t vacc = *x;
44     vacc =  ((vacc >> 31) - 1) & vacc;
45     *y = vacc;
46   }
47 }
48