xref: /aosp_15_r20/external/XNNPACK/src/f16-f32-vcvt/gen/vcvt-neon-int32-x16.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1*4bdc9457SAndroid Build Coastguard Worker // Auto-generated file. Do not edit!
2*4bdc9457SAndroid Build Coastguard Worker //   Template: src/f16-f32-vcvt/neon-int32.c.in
3*4bdc9457SAndroid Build Coastguard Worker //   Generator: tools/xngen
4*4bdc9457SAndroid Build Coastguard Worker //
5*4bdc9457SAndroid Build Coastguard Worker // Copyright 2021 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 
12*4bdc9457SAndroid Build Coastguard Worker #include <arm_neon.h>
13*4bdc9457SAndroid Build Coastguard Worker 
14*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/common.h>
15*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/vcvt.h>
16*4bdc9457SAndroid Build Coastguard Worker 
17*4bdc9457SAndroid Build Coastguard Worker 
xnn_f16_f32_vcvt_ukernel__neon_int32_x16(size_t n,const void * input,float * output,const union xnn_f16_f32_cvt_params params[restrict XNN_MIN_ELEMENTS (1)])18*4bdc9457SAndroid Build Coastguard Worker void xnn_f16_f32_vcvt_ukernel__neon_int32_x16(
19*4bdc9457SAndroid Build Coastguard Worker     size_t n,
20*4bdc9457SAndroid Build Coastguard Worker     const void* input,
21*4bdc9457SAndroid Build Coastguard Worker     float* output,
22*4bdc9457SAndroid Build Coastguard Worker     const union xnn_f16_f32_cvt_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
23*4bdc9457SAndroid Build Coastguard Worker {
24*4bdc9457SAndroid Build Coastguard Worker   assert(n != 0);
25*4bdc9457SAndroid Build Coastguard Worker   assert(n % sizeof(uint16_t) == 0);
26*4bdc9457SAndroid Build Coastguard Worker   assert(input != NULL);
27*4bdc9457SAndroid Build Coastguard Worker   assert(output != NULL);
28*4bdc9457SAndroid Build Coastguard Worker 
29*4bdc9457SAndroid Build Coastguard Worker   const uint32x4_t vsign_mask = vmovq_n_u32(0x80000000);
30*4bdc9457SAndroid Build Coastguard Worker   const uint32x4_t vexp_offset = vmovq_n_u32(0x70000000);
31*4bdc9457SAndroid Build Coastguard Worker   const float32x4_t vexp_scale = vld1q_dup_f32(&params->neon.exp_scale);
32*4bdc9457SAndroid Build Coastguard Worker   const uint32x4_t vmagic_bias = vmovq_n_u32(0x3F000000);
33*4bdc9457SAndroid Build Coastguard Worker   const uint32x4_t vdenorm_cutoff = vmovq_n_u32(0x04000000);
34*4bdc9457SAndroid Build Coastguard Worker 
35*4bdc9457SAndroid Build Coastguard Worker   const uint16_t* i = (const uint16_t*) input;
36*4bdc9457SAndroid Build Coastguard Worker   for (; n >= 16 * sizeof(uint16_t); n -= 16 * sizeof(uint16_t)) {
37*4bdc9457SAndroid Build Coastguard Worker     const uint16x8_t vh0 = vld1q_u16(i); i += 8;
38*4bdc9457SAndroid Build Coastguard Worker     const uint16x8_t vh1 = vld1q_u16(i); i += 8;
39*4bdc9457SAndroid Build Coastguard Worker 
40*4bdc9457SAndroid Build Coastguard Worker     const uint32x4_t vw0 = vshll_n_u16(vget_low_u16(vh0), 16);
41*4bdc9457SAndroid Build Coastguard Worker     const uint32x4_t vw1 = vshll_n_u16(vget_high_u16(vh0), 16);
42*4bdc9457SAndroid Build Coastguard Worker     const uint32x4_t vw2 = vshll_n_u16(vget_low_u16(vh1), 16);
43*4bdc9457SAndroid Build Coastguard Worker     const uint32x4_t vw3 = vshll_n_u16(vget_high_u16(vh1), 16);
44*4bdc9457SAndroid Build Coastguard Worker 
45*4bdc9457SAndroid Build Coastguard Worker     const uint32x4_t vsign0 = vandq_u32(vw0, vsign_mask);
46*4bdc9457SAndroid Build Coastguard Worker     const uint32x4_t vsign1 = vandq_u32(vw1, vsign_mask);
47*4bdc9457SAndroid Build Coastguard Worker     const uint32x4_t vsign2 = vandq_u32(vw2, vsign_mask);
48*4bdc9457SAndroid Build Coastguard Worker     const uint32x4_t vsign3 = vandq_u32(vw3, vsign_mask);
49*4bdc9457SAndroid Build Coastguard Worker 
50*4bdc9457SAndroid Build Coastguard Worker     const uint32x4_t vnonsign0 = veorq_u32(vw0, vsign0);
51*4bdc9457SAndroid Build Coastguard Worker     const uint32x4_t vnonsign1 = veorq_u32(vw1, vsign1);
52*4bdc9457SAndroid Build Coastguard Worker     const uint32x4_t vnonsign2 = veorq_u32(vw2, vsign2);
53*4bdc9457SAndroid Build Coastguard Worker     const uint32x4_t vnonsign3 = veorq_u32(vw3, vsign3);
54*4bdc9457SAndroid Build Coastguard Worker 
55*4bdc9457SAndroid Build Coastguard Worker     const float32x4_t vnorm0 = vmulq_f32(vreinterpretq_f32_u32(vsraq_n_u32(vexp_offset, vnonsign0, 3)), vexp_scale);
56*4bdc9457SAndroid Build Coastguard Worker     const float32x4_t vnorm1 = vmulq_f32(vreinterpretq_f32_u32(vsraq_n_u32(vexp_offset, vnonsign1, 3)), vexp_scale);
57*4bdc9457SAndroid Build Coastguard Worker     const float32x4_t vnorm2 = vmulq_f32(vreinterpretq_f32_u32(vsraq_n_u32(vexp_offset, vnonsign2, 3)), vexp_scale);
58*4bdc9457SAndroid Build Coastguard Worker     const float32x4_t vnorm3 = vmulq_f32(vreinterpretq_f32_u32(vsraq_n_u32(vexp_offset, vnonsign3, 3)), vexp_scale);
59*4bdc9457SAndroid Build Coastguard Worker 
60*4bdc9457SAndroid Build Coastguard Worker     const float32x4_t vdenorm0 = vsubq_f32(vreinterpretq_f32_u32(vsriq_n_u32(vmagic_bias, vnonsign0, 16)), vreinterpretq_f32_u32(vmagic_bias));
61*4bdc9457SAndroid Build Coastguard Worker     const float32x4_t vdenorm1 = vsubq_f32(vreinterpretq_f32_u32(vsriq_n_u32(vmagic_bias, vnonsign1, 16)), vreinterpretq_f32_u32(vmagic_bias));
62*4bdc9457SAndroid Build Coastguard Worker     const float32x4_t vdenorm2 = vsubq_f32(vreinterpretq_f32_u32(vsriq_n_u32(vmagic_bias, vnonsign2, 16)), vreinterpretq_f32_u32(vmagic_bias));
63*4bdc9457SAndroid Build Coastguard Worker     const float32x4_t vdenorm3 = vsubq_f32(vreinterpretq_f32_u32(vsriq_n_u32(vmagic_bias, vnonsign3, 16)), vreinterpretq_f32_u32(vmagic_bias));
64*4bdc9457SAndroid Build Coastguard Worker 
65*4bdc9457SAndroid Build Coastguard Worker     const uint32x4_t vxmask0 = vcgtq_u32(vnonsign0, vdenorm_cutoff);
66*4bdc9457SAndroid Build Coastguard Worker     const uint32x4_t vxmask1 = vcgtq_u32(vnonsign1, vdenorm_cutoff);
67*4bdc9457SAndroid Build Coastguard Worker     const uint32x4_t vxmask2 = vcgtq_u32(vnonsign2, vdenorm_cutoff);
68*4bdc9457SAndroid Build Coastguard Worker     const uint32x4_t vxmask3 = vcgtq_u32(vnonsign3, vdenorm_cutoff);
69*4bdc9457SAndroid Build Coastguard Worker 
70*4bdc9457SAndroid Build Coastguard Worker     const uint32x4_t vf0 = vorrq_u32(vsign0, vreinterpretq_u32_f32(vbslq_f32(vxmask0, vnorm0, vdenorm0)));
71*4bdc9457SAndroid Build Coastguard Worker     const uint32x4_t vf1 = vorrq_u32(vsign1, vreinterpretq_u32_f32(vbslq_f32(vxmask1, vnorm1, vdenorm1)));
72*4bdc9457SAndroid Build Coastguard Worker     const uint32x4_t vf2 = vorrq_u32(vsign2, vreinterpretq_u32_f32(vbslq_f32(vxmask2, vnorm2, vdenorm2)));
73*4bdc9457SAndroid Build Coastguard Worker     const uint32x4_t vf3 = vorrq_u32(vsign3, vreinterpretq_u32_f32(vbslq_f32(vxmask3, vnorm3, vdenorm3)));
74*4bdc9457SAndroid Build Coastguard Worker 
75*4bdc9457SAndroid Build Coastguard Worker     vst1q_f32(output, vreinterpretq_f32_u32(vf0)); output += 4;
76*4bdc9457SAndroid Build Coastguard Worker     vst1q_f32(output, vreinterpretq_f32_u32(vf1)); output += 4;
77*4bdc9457SAndroid Build Coastguard Worker     vst1q_f32(output, vreinterpretq_f32_u32(vf2)); output += 4;
78*4bdc9457SAndroid Build Coastguard Worker     vst1q_f32(output, vreinterpretq_f32_u32(vf3)); output += 4;
79*4bdc9457SAndroid Build Coastguard Worker   }
80*4bdc9457SAndroid Build Coastguard Worker   for (; n >= 8 * sizeof(uint16_t); n -= 8 * sizeof(uint16_t)) {
81*4bdc9457SAndroid Build Coastguard Worker     const uint16x8_t vh = vld1q_u16(i); i += 8;
82*4bdc9457SAndroid Build Coastguard Worker 
83*4bdc9457SAndroid Build Coastguard Worker     const uint32x4_t vw_lo = vshll_n_u16(vget_low_u16(vh), 16);
84*4bdc9457SAndroid Build Coastguard Worker     const uint32x4_t vw_hi = vshll_n_u16(vget_high_u16(vh), 16);
85*4bdc9457SAndroid Build Coastguard Worker 
86*4bdc9457SAndroid Build Coastguard Worker     const uint32x4_t vsign_lo = vandq_u32(vw_lo, vsign_mask);
87*4bdc9457SAndroid Build Coastguard Worker     const uint32x4_t vsign_hi = vandq_u32(vw_hi, vsign_mask);
88*4bdc9457SAndroid Build Coastguard Worker 
89*4bdc9457SAndroid Build Coastguard Worker     const uint32x4_t vnonsign_lo = veorq_u32(vw_lo, vsign_lo);
90*4bdc9457SAndroid Build Coastguard Worker     const uint32x4_t vnonsign_hi = veorq_u32(vw_hi, vsign_hi);
91*4bdc9457SAndroid Build Coastguard Worker 
92*4bdc9457SAndroid Build Coastguard Worker     const float32x4_t vnorm_lo = vmulq_f32(vreinterpretq_f32_u32(vsraq_n_u32(vexp_offset, vnonsign_lo, 3)), vexp_scale);
93*4bdc9457SAndroid Build Coastguard Worker     const float32x4_t vnorm_hi = vmulq_f32(vreinterpretq_f32_u32(vsraq_n_u32(vexp_offset, vnonsign_hi, 3)), vexp_scale);
94*4bdc9457SAndroid Build Coastguard Worker 
95*4bdc9457SAndroid Build Coastguard Worker     const float32x4_t vdenorm_lo = vsubq_f32(vreinterpretq_f32_u32(vsriq_n_u32(vmagic_bias, vnonsign_lo, 16)), vreinterpretq_f32_u32(vmagic_bias));
96*4bdc9457SAndroid Build Coastguard Worker     const float32x4_t vdenorm_hi = vsubq_f32(vreinterpretq_f32_u32(vsriq_n_u32(vmagic_bias, vnonsign_hi, 16)), vreinterpretq_f32_u32(vmagic_bias));
97*4bdc9457SAndroid Build Coastguard Worker 
98*4bdc9457SAndroid Build Coastguard Worker     const uint32x4_t vxmask_lo = vcgtq_u32(vnonsign_lo, vdenorm_cutoff);
99*4bdc9457SAndroid Build Coastguard Worker     const uint32x4_t vf_lo = vorrq_u32(vsign_lo, vreinterpretq_u32_f32(vbslq_f32(vxmask_lo, vnorm_lo, vdenorm_lo)));
100*4bdc9457SAndroid Build Coastguard Worker 
101*4bdc9457SAndroid Build Coastguard Worker     const uint32x4_t vxmask_hi = vcgtq_u32(vnonsign_hi, vdenorm_cutoff);
102*4bdc9457SAndroid Build Coastguard Worker     const uint32x4_t vf_hi = vorrq_u32(vsign_hi, vreinterpretq_u32_f32(vbslq_f32(vxmask_hi, vnorm_hi, vdenorm_hi)));
103*4bdc9457SAndroid Build Coastguard Worker 
104*4bdc9457SAndroid Build Coastguard Worker     vst1q_f32(output, vreinterpretq_f32_u32(vf_lo)); output += 4;
105*4bdc9457SAndroid Build Coastguard Worker     vst1q_f32(output, vreinterpretq_f32_u32(vf_hi)); output += 4;
106*4bdc9457SAndroid Build Coastguard Worker   }
107*4bdc9457SAndroid Build Coastguard Worker   if XNN_UNPREDICTABLE(n != 0) {
108*4bdc9457SAndroid Build Coastguard Worker     const uint16x8_t vh = vld1q_u16(i); i += 8;
109*4bdc9457SAndroid Build Coastguard Worker 
110*4bdc9457SAndroid Build Coastguard Worker     const uint32x4_t vw_lo = vshll_n_u16(vget_low_u16(vh), 16);
111*4bdc9457SAndroid Build Coastguard Worker     const uint32x4_t vw_hi = vshll_n_u16(vget_high_u16(vh), 16);
112*4bdc9457SAndroid Build Coastguard Worker 
113*4bdc9457SAndroid Build Coastguard Worker     const uint32x4_t vsign_lo = vandq_u32(vw_lo, vsign_mask);
114*4bdc9457SAndroid Build Coastguard Worker     const uint32x4_t vsign_hi = vandq_u32(vw_hi, vsign_mask);
115*4bdc9457SAndroid Build Coastguard Worker 
116*4bdc9457SAndroid Build Coastguard Worker     const uint32x4_t vnonsign_lo = veorq_u32(vw_lo, vsign_lo);
117*4bdc9457SAndroid Build Coastguard Worker     const uint32x4_t vnonsign_hi = veorq_u32(vw_hi, vsign_hi);
118*4bdc9457SAndroid Build Coastguard Worker 
119*4bdc9457SAndroid Build Coastguard Worker     const float32x4_t vnorm_lo = vmulq_f32(vreinterpretq_f32_u32(vsraq_n_u32(vexp_offset, vnonsign_lo, 3)), vexp_scale);
120*4bdc9457SAndroid Build Coastguard Worker     const float32x4_t vnorm_hi = vmulq_f32(vreinterpretq_f32_u32(vsraq_n_u32(vexp_offset, vnonsign_hi, 3)), vexp_scale);
121*4bdc9457SAndroid Build Coastguard Worker 
122*4bdc9457SAndroid Build Coastguard Worker     const float32x4_t vdenorm_lo = vsubq_f32(vreinterpretq_f32_u32(vsriq_n_u32(vmagic_bias, vnonsign_lo, 16)), vreinterpretq_f32_u32(vmagic_bias));
123*4bdc9457SAndroid Build Coastguard Worker     const float32x4_t vdenorm_hi = vsubq_f32(vreinterpretq_f32_u32(vsriq_n_u32(vmagic_bias, vnonsign_hi, 16)), vreinterpretq_f32_u32(vmagic_bias));
124*4bdc9457SAndroid Build Coastguard Worker 
125*4bdc9457SAndroid Build Coastguard Worker     const uint32x4_t vxmask_lo = vcgtq_u32(vnonsign_lo, vdenorm_cutoff);
126*4bdc9457SAndroid Build Coastguard Worker     uint32x4_t vf = vorrq_u32(vsign_lo, vreinterpretq_u32_f32(vbslq_f32(vxmask_lo, vnorm_lo, vdenorm_lo)));
127*4bdc9457SAndroid Build Coastguard Worker 
128*4bdc9457SAndroid Build Coastguard Worker     if (n & (4 * sizeof(uint16_t))) {
129*4bdc9457SAndroid Build Coastguard Worker       vst1q_f32(output, vreinterpretq_f32_u32(vf)); output += 4;
130*4bdc9457SAndroid Build Coastguard Worker 
131*4bdc9457SAndroid Build Coastguard Worker       const uint32x4_t vxmask_hi = vcgtq_u32(vnonsign_hi, vdenorm_cutoff);
132*4bdc9457SAndroid Build Coastguard Worker       vf = vorrq_u32(vsign_hi, vreinterpretq_u32_f32(vbslq_f32(vxmask_hi, vnorm_hi, vdenorm_hi)));
133*4bdc9457SAndroid Build Coastguard Worker     }
134*4bdc9457SAndroid Build Coastguard Worker     uint32x2_t vf_lo = vget_low_u32(vf);
135*4bdc9457SAndroid Build Coastguard Worker     if (n & (2 * sizeof(uint16_t))) {
136*4bdc9457SAndroid Build Coastguard Worker       vst1_f32(output, vreinterpret_f32_u32(vf_lo)); output += 2;
137*4bdc9457SAndroid Build Coastguard Worker       vf_lo = vget_high_u32(vf);
138*4bdc9457SAndroid Build Coastguard Worker     }
139*4bdc9457SAndroid Build Coastguard Worker     if (n & (1 * sizeof(uint16_t))) {
140*4bdc9457SAndroid Build Coastguard Worker       vst1_lane_f32(output, vreinterpret_f32_u32(vf_lo), 0);
141*4bdc9457SAndroid Build Coastguard Worker     }
142*4bdc9457SAndroid Build Coastguard Worker   }
143*4bdc9457SAndroid Build Coastguard Worker }
144