xref: /aosp_15_r20/external/XNNPACK/src/f32-vrnd/gen/vrndd-neon-x8.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-vrnd/vrndd-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_vrndd_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_vrndd_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_OOB_READS
24 {
25   assert(n != 0);
26   assert(n % sizeof(float) == 0);
27 
28   const float32x4_t vintegral_threshold = vreinterpretq_f32_u32(vmovq_n_u32(UINT32_C(0x4B000000)));
29   const uint32x4_t vone = vreinterpretq_u32_f32(vmovq_n_f32(1.0f));
30   for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
31     const float32x4_t vx0123 = vld1q_f32(x); x += 4;
32     const float32x4_t vx4567 = vld1q_f32(x); x += 4;
33 
34     const int32x4_t vintx0123 = vcvtq_s32_f32(vx0123);
35     const int32x4_t vintx4567 = vcvtq_s32_f32(vx4567);
36 
37     uint32x4_t vrndmask0123 = vcaltq_f32(vx0123, vintegral_threshold);
38     uint32x4_t vrndmask4567 = vcaltq_f32(vx4567, vintegral_threshold);
39 
40     const float32x4_t vprerndx0123 = vcvtq_f32_s32(vintx0123);
41     const float32x4_t vprerndx4567 = vcvtq_f32_s32(vintx4567);
42 
43     vrndmask0123 = vbicq_u32(vrndmask0123, vmovq_n_u32(UINT32_C(0x80000000)));
44     vrndmask4567 = vbicq_u32(vrndmask4567, vmovq_n_u32(UINT32_C(0x80000000)));
45 
46     const float32x4_t vrndx0123 = vbslq_f32(vrndmask0123, vprerndx0123, vx0123);
47     const float32x4_t vrndx4567 = vbslq_f32(vrndmask4567, vprerndx4567, vx4567);
48 
49     const uint32x4_t vadjmask0123 = vcgtq_f32(vrndx0123, vx0123);
50     const uint32x4_t vadjmask4567 = vcgtq_f32(vrndx4567, vx4567);
51 
52     const float32x4_t vadjrndx0123 = vreinterpretq_f32_u32(vandq_u32(vadjmask0123, vone));
53     const float32x4_t vadjrndx4567 = vreinterpretq_f32_u32(vandq_u32(vadjmask4567, vone));
54 
55     const float32x4_t vy0123 = vsubq_f32(vrndx0123, vadjrndx0123);
56     const float32x4_t vy4567 = vsubq_f32(vrndx4567, vadjrndx4567);
57 
58     vst1q_f32(y, vy0123); y += 4;
59     vst1q_f32(y, vy4567); y += 4;
60   }
61   for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
62     const float32x4_t vx = vld1q_f32(x); x += 4;
63     const int32x4_t vintx = vcvtq_s32_f32(vx);
64     uint32x4_t vrndmask = vcaltq_f32(vx, vintegral_threshold);
65     const float32x4_t vprerndx = vcvtq_f32_s32(vintx);
66     vrndmask = vbicq_u32(vrndmask, vmovq_n_u32(UINT32_C(0x80000000)));
67     const float32x4_t vrndx = vbslq_f32(vrndmask, vprerndx, vx);
68     const uint32x4_t vadjmask = vcgtq_f32(vrndx, vx);
69     const float32x4_t vadjrndx = vreinterpretq_f32_u32(vandq_u32(vadjmask, vone));
70     const float32x4_t vy = vsubq_f32(vrndx, vadjrndx);
71     vst1q_f32(y, vy); y += 4;
72   }
73   if XNN_UNLIKELY(n != 0) {
74     const float32x4_t vx = vld1q_f32(x);
75     const int32x4_t vintx = vcvtq_s32_f32(vx);
76     uint32x4_t vrndmask = vcaltq_f32(vx, vintegral_threshold);
77     const float32x4_t vprerndx = vcvtq_f32_s32(vintx);
78     vrndmask = vbicq_u32(vrndmask, vmovq_n_u32(UINT32_C(0x80000000)));
79     const float32x4_t vrndx = vbslq_f32(vrndmask, vprerndx, vx);
80     const uint32x4_t vadjmask = vcgtq_f32(vrndx, vx);
81     const float32x4_t vadjrndx = vreinterpretq_f32_u32(vandq_u32(vadjmask, vone));
82     const float32x4_t vy = vsubq_f32(vrndx, vadjrndx);
83     float32x2_t vy_lo = vget_low_f32(vy);
84     if (n & (2 * sizeof(float))) {
85       vst1_f32(y, vy_lo); y += 2;
86       vy_lo = vget_high_f32(vy);
87     }
88     if (n & (1 * sizeof(float))) {
89       vst1_lane_f32(y, vy_lo, 0);
90     }
91   }
92 }
93