xref: /aosp_15_r20/external/XNNPACK/src/f32-vrelu/gen/vrelu-scalar-x2.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1*4bdc9457SAndroid Build Coastguard Worker // Auto-generated file. Do not edit!
2*4bdc9457SAndroid Build Coastguard Worker //   Template: src/f32-vrelu/scalar.c.in
3*4bdc9457SAndroid Build Coastguard Worker //   Generator: tools/xngen
4*4bdc9457SAndroid Build Coastguard Worker //
5*4bdc9457SAndroid Build Coastguard Worker // Copyright 2020 Google LLC
6*4bdc9457SAndroid Build Coastguard Worker //
7*4bdc9457SAndroid Build Coastguard Worker // This source code is licensed under the BSD-style license found in the
8*4bdc9457SAndroid Build Coastguard Worker // LICENSE file in the root directory of this source tree.
9*4bdc9457SAndroid Build Coastguard Worker 
10*4bdc9457SAndroid Build Coastguard Worker #include <assert.h>
11*4bdc9457SAndroid Build Coastguard Worker 
12*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/vunary.h>
13*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/common.h>
14*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/math.h>
15*4bdc9457SAndroid Build Coastguard Worker 
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*4bdc9457SAndroid Build Coastguard Worker void xnn_f32_vrelu_ukernel__scalar_x2(
17*4bdc9457SAndroid Build Coastguard Worker     size_t n,
18*4bdc9457SAndroid Build Coastguard Worker     const float* x_ptr,
19*4bdc9457SAndroid Build Coastguard Worker     float* y_ptr,
20*4bdc9457SAndroid Build Coastguard Worker     const union xnn_f32_relu_params params[restrict XNN_MIN_ELEMENTS(1)])
21*4bdc9457SAndroid Build Coastguard Worker {
22*4bdc9457SAndroid Build Coastguard Worker   assert(n != 0);
23*4bdc9457SAndroid Build Coastguard Worker   assert(n % sizeof(float) == 0);
24*4bdc9457SAndroid Build Coastguard Worker   assert(x_ptr != NULL);
25*4bdc9457SAndroid Build Coastguard Worker   assert(y_ptr != NULL);
26*4bdc9457SAndroid Build Coastguard Worker 
27*4bdc9457SAndroid Build Coastguard Worker   const uint32_t* x = (const uint32_t*)x_ptr;
28*4bdc9457SAndroid Build Coastguard Worker   uint32_t* y = (uint32_t*)y_ptr;
29*4bdc9457SAndroid Build Coastguard Worker 
30*4bdc9457SAndroid Build Coastguard Worker   for (; n >= 2 * sizeof(uint32_t); n -= 2 * sizeof(uint32_t)) {
31*4bdc9457SAndroid Build Coastguard Worker     uint32_t vacc0 = x[0];
32*4bdc9457SAndroid Build Coastguard Worker     uint32_t vacc1 = x[1];
33*4bdc9457SAndroid Build Coastguard Worker     x += 2;
34*4bdc9457SAndroid Build Coastguard Worker 
35*4bdc9457SAndroid Build Coastguard Worker     vacc0 = ((vacc0 >> 31) - 1) & vacc0;
36*4bdc9457SAndroid Build Coastguard Worker     vacc1 = ((vacc1 >> 31) - 1) & vacc1;
37*4bdc9457SAndroid Build Coastguard Worker 
38*4bdc9457SAndroid Build Coastguard Worker     y[0] = vacc0;
39*4bdc9457SAndroid Build Coastguard Worker     y[1] = vacc1;
40*4bdc9457SAndroid Build Coastguard Worker     y += 2;
41*4bdc9457SAndroid Build Coastguard Worker   }
42*4bdc9457SAndroid Build Coastguard Worker   if XNN_UNLIKELY(n != 0) {
43*4bdc9457SAndroid Build Coastguard Worker     uint32_t vacc = *x;
44*4bdc9457SAndroid Build Coastguard Worker     vacc =  ((vacc >> 31) - 1) & vacc;
45*4bdc9457SAndroid Build Coastguard Worker     *y = vacc;
46*4bdc9457SAndroid Build Coastguard Worker   }
47*4bdc9457SAndroid Build Coastguard Worker }
48