xref: /aosp_15_r20/external/XNNPACK/src/f32-vsqrt/gen/scalar-sqrt-x4.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1*4bdc9457SAndroid Build Coastguard Worker // Auto-generated file. Do not edit!
2*4bdc9457SAndroid Build Coastguard Worker //   Template: src/f32-vsqrt/scalar-sqrt.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 #include <math.h>
12*4bdc9457SAndroid Build Coastguard Worker 
13*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/common.h>
14*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/vunary.h>
15*4bdc9457SAndroid Build Coastguard Worker 
16*4bdc9457SAndroid Build Coastguard Worker 
xnn_f32_vsqrt_ukernel__scalar_sqrt_x4(size_t n,const float * x,float * y,const union xnn_f32_sqrt_params params[restrict XNN_MIN_ELEMENTS (1)])17*4bdc9457SAndroid Build Coastguard Worker void xnn_f32_vsqrt_ukernel__scalar_sqrt_x4(
18*4bdc9457SAndroid Build Coastguard Worker     size_t n,
19*4bdc9457SAndroid Build Coastguard Worker     const float* x,
20*4bdc9457SAndroid Build Coastguard Worker     float* y,
21*4bdc9457SAndroid Build Coastguard Worker     const union xnn_f32_sqrt_params params[restrict XNN_MIN_ELEMENTS(1)])
22*4bdc9457SAndroid Build Coastguard Worker {
23*4bdc9457SAndroid Build Coastguard Worker   assert(n != 0);
24*4bdc9457SAndroid Build Coastguard Worker   assert(n % sizeof(float) == 0);
25*4bdc9457SAndroid Build Coastguard Worker 
26*4bdc9457SAndroid Build Coastguard Worker   for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
27*4bdc9457SAndroid Build Coastguard Worker     const float vx0 = x[0];
28*4bdc9457SAndroid Build Coastguard Worker     const float vx1 = x[1];
29*4bdc9457SAndroid Build Coastguard Worker     const float vx2 = x[2];
30*4bdc9457SAndroid Build Coastguard Worker     const float vx3 = x[3];
31*4bdc9457SAndroid Build Coastguard Worker     x += 4;
32*4bdc9457SAndroid Build Coastguard Worker 
33*4bdc9457SAndroid Build Coastguard Worker     const float vy0 = sqrtf(vx0);
34*4bdc9457SAndroid Build Coastguard Worker     const float vy1 = sqrtf(vx1);
35*4bdc9457SAndroid Build Coastguard Worker     const float vy2 = sqrtf(vx2);
36*4bdc9457SAndroid Build Coastguard Worker     const float vy3 = sqrtf(vx3);
37*4bdc9457SAndroid Build Coastguard Worker 
38*4bdc9457SAndroid Build Coastguard Worker     y[0] = vy0;
39*4bdc9457SAndroid Build Coastguard Worker     y[1] = vy1;
40*4bdc9457SAndroid Build Coastguard Worker     y[2] = vy2;
41*4bdc9457SAndroid Build Coastguard Worker     y[3] = vy3;
42*4bdc9457SAndroid Build Coastguard Worker     y += 4;
43*4bdc9457SAndroid Build Coastguard Worker   }
44*4bdc9457SAndroid Build Coastguard Worker   if XNN_UNLIKELY(n != 0) {
45*4bdc9457SAndroid Build Coastguard Worker     do {
46*4bdc9457SAndroid Build Coastguard Worker       const float vx = *x++;
47*4bdc9457SAndroid Build Coastguard Worker       const float vy = sqrtf(vx);
48*4bdc9457SAndroid Build Coastguard Worker       *y++ = vy;
49*4bdc9457SAndroid Build Coastguard Worker       n -= sizeof(float);
50*4bdc9457SAndroid Build Coastguard Worker     } while (n != 0);
51*4bdc9457SAndroid Build Coastguard Worker   }
52*4bdc9457SAndroid Build Coastguard Worker }
53