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_x8(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_x8(
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 >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
30 const float32x4_t vx0123 = vld1q_f32(x); x += 4;
31 const float32x4_t vx4567 = vld1q_f32(x); x += 4;
32
33 const float32x4_t vabsx0123 = vabsq_f32(vx0123);
34 uint32x4_t vrndmask0123 = vcaltq_f32(vmagic_number, vx0123);
35 const float32x4_t vabsx4567 = vabsq_f32(vx4567);
36 uint32x4_t vrndmask4567 = vcaltq_f32(vmagic_number, vx4567);
37
38 float32x4_t vrndabsx0123 = vaddq_f32(vabsx0123, vmagic_number);
39 float32x4_t vrndabsx4567 = vaddq_f32(vabsx4567, vmagic_number);
40
41 vrndmask0123 = vorrq_u32(vrndmask0123, vmovq_n_u32(UINT32_C(0x80000000)));
42 vrndmask4567 = vorrq_u32(vrndmask4567, vmovq_n_u32(UINT32_C(0x80000000)));
43
44 vrndabsx0123 = vsubq_f32(vrndabsx0123, vmagic_number);
45 vrndabsx4567 = vsubq_f32(vrndabsx4567, vmagic_number);
46
47 const float32x4_t vy0123 = vbslq_f32(vrndmask0123, vx0123, vrndabsx0123);
48 const float32x4_t vy4567 = vbslq_f32(vrndmask4567, vx4567, vrndabsx4567);
49
50 vst1q_f32(y, vy0123); y += 4;
51 vst1q_f32(y, vy4567); y += 4;
52 }
53 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
54 const float32x4_t vx = vld1q_f32(x); x += 4;
55 const float32x4_t vabsx = vabsq_f32(vx);
56 uint32x4_t vrndmask = vcaltq_f32(vmagic_number, vx);
57 float32x4_t vrndabsx = vaddq_f32(vabsx, vmagic_number);
58 vrndmask = vorrq_u32(vrndmask, vmovq_n_u32(UINT32_C(0x80000000)));
59 vrndabsx = vsubq_f32(vrndabsx, vmagic_number);
60 const float32x4_t vy = vbslq_f32(vrndmask, vx, vrndabsx);
61 vst1q_f32(y, vy); y += 4;
62 }
63 if XNN_UNLIKELY(n != 0) {
64 const float32x4_t vx = vld1q_f32(x);
65 const float32x4_t vabsx = vabsq_f32(vx);
66 uint32x4_t vrndmask = vcaltq_f32(vmagic_number, vx);
67 float32x4_t vrndabsx = vaddq_f32(vabsx, vmagic_number);
68 vrndmask = vorrq_u32(vrndmask, vmovq_n_u32(UINT32_C(0x80000000)));
69 vrndabsx = vsubq_f32(vrndabsx, vmagic_number);
70 const float32x4_t vy = vbslq_f32(vrndmask, vx, vrndabsx);
71 float32x2_t vy_lo = vget_low_f32(vy);
72 if (n & (2 * sizeof(float))) {
73 vst1_f32(y, vy_lo); y += 2;
74 vy_lo = vget_high_f32(vy);
75 }
76 if (n & (1 * sizeof(float))) {
77 vst1_lane_f32(y, vy_lo, 0);
78 }
79 }
80 }
81