xref: /aosp_15_r20/external/XNNPACK/src/f32-f16-vcvt/gen/vcvt-neon-x8.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-f16-vcvt/neon.c.in
3 //   Generator: tools/xngen
4 //
5 // Copyright 2021 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/vcvt.h>
16 
17 
xnn_f32_f16_vcvt_ukernel__neon_x8(size_t n,const float * input,void * output,const union xnn_f32_f16_cvt_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f32_f16_vcvt_ukernel__neon_x8(
19     size_t n,
20     const float* input,
21     void* output,
22     const union xnn_f32_f16_cvt_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
23 {
24   assert(n != 0);
25   assert(n % sizeof(float) == 0);
26   assert(input != NULL);
27   assert(output != NULL);
28 
29   const uint32x4_t vexp_bias = vld1q_dup_u32(&params->neon.exp_bias);
30   const float32x4_t vscale_to_inf = vld1q_dup_f32(&params->neon.scale_to_inf);
31   const uint32x4_t vexpw_max = vld1q_dup_u32(&params->neon.expw_max);
32   const float32x4_t vscale_to_zero = vld1q_dup_f32(&params->neon.scale_to_zero);
33   const uint32x4_t vbias_min = vdupq_n_u32(UINT32_C(0x40000000));
34   const uint16x8_t vexph_mask = vdupq_n_u16(UINT16_C(0x7C00));
35   const uint16x8_t vmanth_mask = vdupq_n_u16(UINT16_C(0x0FFF));
36   const uint16x8_t vsignh_mask = vdupq_n_u16(UINT16_C(0x8000));
37   const uint16x8_t vnanh = vdupq_n_u16(UINT16_C(0x7E00));
38 
39   uint16_t* o = (uint16_t*) output;
40   for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
41     const float32x4_t vx0 = vld1q_f32(input); input += 4;
42     const float32x4_t vx1 = vld1q_f32(input); input += 4;
43 
44     const float32x4_t vabsx0 = vabsq_f32(vx0);
45     const float32x4_t vabsx1 = vabsq_f32(vx1);
46 
47     uint32x4_t vbias0 = vaddq_u32(vreinterpretq_u32_f32(vabsx0), vexp_bias);
48     uint32x4_t vbias1 = vaddq_u32(vreinterpretq_u32_f32(vabsx1), vexp_bias);
49 
50     float32x4_t vf0 = vmulq_f32(vabsx0, vscale_to_inf);
51     float32x4_t vf1 = vmulq_f32(vabsx1, vscale_to_inf);
52     const uint32x4_t vnanmaskw0 = vcgtq_u32(vreinterpretq_u32_f32(vabsx0), vexpw_max);
53     const uint32x4_t vnanmaskw1 = vcgtq_u32(vreinterpretq_u32_f32(vabsx1), vexpw_max);
54 
55     vbias0 = vandq_u32(vbias0, vexpw_max);
56     vbias1 = vandq_u32(vbias1, vexpw_max);
57     vf0 = vmulq_f32(vf0, vscale_to_zero);
58     vf1 = vmulq_f32(vf1, vscale_to_zero);
59 
60     const uint16x8_t vnanmaskh0 = vcombine_u16(vmovn_u32(vnanmaskw0), vmovn_u32(vnanmaskw1));
61     vbias0 = vmaxq_u32(vbias0, vbias_min);
62     vbias1 = vmaxq_u32(vbias1, vbias_min);
63 
64     vf0 = vaddq_f32(vf0, vreinterpretq_f32_u32(vbias0));
65     vf1 = vaddq_f32(vf1, vreinterpretq_f32_u32(vbias1));
66 
67     uint16x8_t vexph0 = vcombine_u16(vshrn_n_u32(vreinterpretq_u32_f32(vf0), 13), vshrn_n_u32(vreinterpretq_u32_f32(vf1), 13));
68     uint16x8_t vmanth0 = vcombine_u16(vmovn_u32(vreinterpretq_u32_f32(vf0)), vmovn_u32(vreinterpretq_u32_f32(vf1)));
69     uint16x8_t vsignh0 = vcombine_u16(vshrn_n_u32(vreinterpretq_u32_f32(vx0), 16), vshrn_n_u32(vreinterpretq_u32_f32(vx1), 16));
70 
71     vexph0 = vandq_u16(vexph0, vexph_mask);
72     vmanth0 = vandq_u16(vmanth0, vmanth_mask);
73     vsignh0 = vandq_u16(vsignh0, vsignh_mask);
74 
75     uint16x8_t vh0 = vaddq_u16(vmanth0, vexph0);
76 
77     vh0 = vbslq_u16(vnanmaskh0, vnanh, vh0);
78 
79     vh0 = vorrq_u16(vh0, vsignh0);
80 
81     vst1q_u16(o, vh0); o += 8;
82   }
83   for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
84     const float32x4_t vx = vld1q_f32(input); input += 4;
85 
86     const float32x4_t vabsx = vabsq_f32(vx);
87 
88     uint32x4_t vbias = vaddq_u32(vreinterpretq_u32_f32(vabsx), vexp_bias);
89 
90     float32x4_t vf = vmulq_f32(vabsx, vscale_to_inf);
91     const uint32x4_t vnanmaskw = vcgtq_u32(vreinterpretq_u32_f32(vabsx), vexpw_max);
92 
93     vbias = vandq_u32(vbias, vexpw_max);
94     vf = vmulq_f32(vf, vscale_to_zero);
95 
96     const uint16x4_t vnanmaskh = vmovn_u32(vnanmaskw);
97     vbias = vmaxq_u32(vbias, vbias_min);
98 
99     vf = vaddq_f32(vf, vreinterpretq_f32_u32(vbias));
100 
101     uint16x4_t vexph = vshrn_n_u32(vreinterpretq_u32_f32(vf), 13);
102     uint16x4_t vmanth = vmovn_u32(vreinterpretq_u32_f32(vf));
103     uint16x4_t vsignh = vshrn_n_u32(vreinterpretq_u32_f32(vx), 16);
104 
105     vexph = vand_u16(vexph, vget_low_u16(vexph_mask));
106     vmanth = vand_u16(vmanth, vget_low_u16(vmanth_mask));
107     vsignh = vand_u16(vsignh, vget_low_u16(vsignh_mask));
108 
109     uint16x4_t vh = vadd_u16(vmanth, vexph);
110 
111     vh = vbsl_u16(vnanmaskh, vget_low_u16(vnanh), vh);
112 
113     vh = vorr_u16(vh, vsignh);
114 
115     vst1_u16(o, vh); o += 4;
116   }
117   if XNN_UNLIKELY(n != 0) {
118     assert(n % sizeof(float) == 0);
119     assert(n >= 1 * sizeof(float));
120     assert(n <= 3 * sizeof(float));
121     const float32x4_t vx = vld1q_f32(input);
122 
123     const float32x4_t vabsx = vabsq_f32(vx);
124 
125     uint32x4_t vbias = vaddq_u32(vreinterpretq_u32_f32(vabsx), vexp_bias);
126 
127     float32x4_t vf = vmulq_f32(vabsx, vscale_to_inf);
128     const uint32x4_t vnanmaskw = vcgtq_u32(vreinterpretq_u32_f32(vabsx), vexpw_max);
129 
130     vbias = vandq_u32(vbias, vexpw_max);
131     vf = vmulq_f32(vf, vscale_to_zero);
132 
133     const uint16x4_t vnanmaskh = vmovn_u32(vnanmaskw);
134     vbias = vmaxq_u32(vbias, vbias_min);
135 
136     vf = vaddq_f32(vf, vreinterpretq_f32_u32(vbias));
137 
138     uint16x4_t vexph = vshrn_n_u32(vreinterpretq_u32_f32(vf), 13);
139     uint16x4_t vmanth = vmovn_u32(vreinterpretq_u32_f32(vf));
140     uint16x4_t vsignh = vshrn_n_u32(vreinterpretq_u32_f32(vx), 16);
141 
142     vexph = vand_u16(vexph, vget_low_u16(vexph_mask));
143     vmanth = vand_u16(vmanth, vget_low_u16(vmanth_mask));
144     vsignh = vand_u16(vsignh, vget_low_u16(vsignh_mask));
145 
146     uint16x4_t vh = vadd_u16(vmanth, vexph);
147 
148     vh = vbsl_u16(vnanmaskh, vget_low_u16(vnanh), vh);
149 
150     vh = vorr_u16(vh, vsignh);
151 
152     if (n & (2 * sizeof(float))) {
153       vst1_lane_u32((void*) o, vreinterpret_u32_u16(vh), 0); o += 2;
154       vh = vext_u16(vh, vh, 2);
155     }
156     if (n & (1 * sizeof(float))) {
157       vst1_lane_u16(o, vh, 0);
158     }
159   }
160 }
161