1 // Auto-generated file. Do not edit!
2 // Template: src/f32-vrnd/vrndne-neon.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 <arm_neon.h>
13
14 #include <xnnpack/common.h>
15 #include <xnnpack/math.h>
16 #include <xnnpack/vunary.h>
17
18
xnn_f32_vrndne_ukernel__neon_x4(size_t n,const float * x,float * y,const union xnn_f32_rnd_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_f32_vrndne_ukernel__neon_x4(
20 size_t n,
21 const float* x,
22 float* y,
23 const union xnn_f32_rnd_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
24 {
25 assert(n != 0);
26 assert(n % sizeof(float) == 0);
27
28 const float32x4_t vmagic_number = vreinterpretq_f32_u32(vmovq_n_u32(UINT32_C(0x4B000000)));
29 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
30 const float32x4_t vx0123 = vld1q_f32(x); x += 4;
31
32 const float32x4_t vabsx0123 = vabsq_f32(vx0123);
33 uint32x4_t vrndmask0123 = vcaltq_f32(vmagic_number, vx0123);
34
35 float32x4_t vrndabsx0123 = vaddq_f32(vabsx0123, vmagic_number);
36
37 vrndmask0123 = vorrq_u32(vrndmask0123, vmovq_n_u32(UINT32_C(0x80000000)));
38
39 vrndabsx0123 = vsubq_f32(vrndabsx0123, vmagic_number);
40
41 const float32x4_t vy0123 = vbslq_f32(vrndmask0123, vx0123, vrndabsx0123);
42
43 vst1q_f32(y, vy0123); y += 4;
44 }
45 if XNN_UNLIKELY(n != 0) {
46 const float32x4_t vx = vld1q_f32(x);
47 const float32x4_t vabsx = vabsq_f32(vx);
48 uint32x4_t vrndmask = vcaltq_f32(vmagic_number, vx);
49 float32x4_t vrndabsx = vaddq_f32(vabsx, vmagic_number);
50 vrndmask = vorrq_u32(vrndmask, vmovq_n_u32(UINT32_C(0x80000000)));
51 vrndabsx = vsubq_f32(vrndabsx, vmagic_number);
52 const float32x4_t vy = vbslq_f32(vrndmask, vx, vrndabsx);
53 float32x2_t vy_lo = vget_low_f32(vy);
54 if (n & (2 * sizeof(float))) {
55 vst1_f32(y, vy_lo); y += 2;
56 vy_lo = vget_high_f32(vy);
57 }
58 if (n & (1 * sizeof(float))) {
59 vst1_lane_f32(y, vy_lo, 0);
60 }
61 }
62 }
63