1*4bdc9457SAndroid Build Coastguard Worker // Auto-generated file. Do not edit!
2*4bdc9457SAndroid Build Coastguard Worker // Template: src/f16-f32-vcvt/neon-int16.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_int16_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_int16_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 uint16x8_t vsign_mask = vmovq_n_u16(0x8000);
30*4bdc9457SAndroid Build Coastguard Worker const uint16x8_t vexp_offset = vmovq_n_u16(0x7000);
31*4bdc9457SAndroid Build Coastguard Worker const float32x4_t vexp_scale = vld1q_dup_f32(¶ms->neon.exp_scale);
32*4bdc9457SAndroid Build Coastguard Worker const uint32x4_t vmagic_bias = vmovq_n_u32(0x3F000000);
33*4bdc9457SAndroid Build Coastguard Worker const uint16x8_t vdenorm_cutoff = vmovq_n_u16(0x0400);
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 uint16x8_t vsign0 = vandq_u16(vh0, vsign_mask);
41*4bdc9457SAndroid Build Coastguard Worker const uint16x8_t vsign1 = vandq_u16(vh1, vsign_mask);
42*4bdc9457SAndroid Build Coastguard Worker
43*4bdc9457SAndroid Build Coastguard Worker const uint16x8_t vnonsign0 = veorq_u16(vh0, vsign0);
44*4bdc9457SAndroid Build Coastguard Worker const uint16x8_t vnonsign1 = veorq_u16(vh1, vsign1);
45*4bdc9457SAndroid Build Coastguard Worker
46*4bdc9457SAndroid Build Coastguard Worker const uint16x8x2_t vprenorm0 = vzipq_u16(vshlq_n_u16(vnonsign0, 13), vsraq_n_u16(vexp_offset, vnonsign0, 3));
47*4bdc9457SAndroid Build Coastguard Worker const uint16x8x2_t vprenorm1 = vzipq_u16(vshlq_n_u16(vnonsign1, 13), vsraq_n_u16(vexp_offset, vnonsign1, 3));
48*4bdc9457SAndroid Build Coastguard Worker
49*4bdc9457SAndroid Build Coastguard Worker const float32x4_t vnorm0 = vmulq_f32(vreinterpretq_f32_u16(vprenorm0.val[0]), vexp_scale);
50*4bdc9457SAndroid Build Coastguard Worker const float32x4_t vnorm1 = vmulq_f32(vreinterpretq_f32_u16(vprenorm0.val[1]), vexp_scale);
51*4bdc9457SAndroid Build Coastguard Worker const float32x4_t vnorm2 = vmulq_f32(vreinterpretq_f32_u16(vprenorm1.val[0]), vexp_scale);
52*4bdc9457SAndroid Build Coastguard Worker const float32x4_t vnorm3 = vmulq_f32(vreinterpretq_f32_u16(vprenorm1.val[1]), vexp_scale);
53*4bdc9457SAndroid Build Coastguard Worker
54*4bdc9457SAndroid Build Coastguard Worker const float32x4_t vdenorm0 = vsubq_f32(vreinterpretq_f32_u32(vaddw_u16(vmagic_bias, vget_low_u16(vnonsign0))), vreinterpretq_f32_u32(vmagic_bias));
55*4bdc9457SAndroid Build Coastguard Worker const float32x4_t vdenorm1 = vsubq_f32(vreinterpretq_f32_u32(vaddw_u16(vmagic_bias, vget_high_u16(vnonsign0))), vreinterpretq_f32_u32(vmagic_bias));
56*4bdc9457SAndroid Build Coastguard Worker const float32x4_t vdenorm2 = vsubq_f32(vreinterpretq_f32_u32(vaddw_u16(vmagic_bias, vget_low_u16(vnonsign1))), vreinterpretq_f32_u32(vmagic_bias));
57*4bdc9457SAndroid Build Coastguard Worker const float32x4_t vdenorm3 = vsubq_f32(vreinterpretq_f32_u32(vaddw_u16(vmagic_bias, vget_high_u16(vnonsign1))), vreinterpretq_f32_u32(vmagic_bias));
58*4bdc9457SAndroid Build Coastguard Worker
59*4bdc9457SAndroid Build Coastguard Worker const uint16x8_t vmask0 = vcgtq_u16(vnonsign0, vdenorm_cutoff);
60*4bdc9457SAndroid Build Coastguard Worker const uint16x8_t vmask1 = vcgtq_u16(vnonsign1, vdenorm_cutoff);
61*4bdc9457SAndroid Build Coastguard Worker
62*4bdc9457SAndroid Build Coastguard Worker const uint32x4_t vxmask0 = vreinterpretq_u32_s32(vmovl_s16(vreinterpret_s16_u16(vget_low_u16(vmask0))));
63*4bdc9457SAndroid Build Coastguard Worker const uint32x4_t vf0 = vorrq_u32(vshll_n_u16(vget_low_u16(vsign0), 16),
64*4bdc9457SAndroid Build Coastguard Worker vreinterpretq_u32_f32(vbslq_f32(vxmask0, vnorm0, vdenorm0)));
65*4bdc9457SAndroid Build Coastguard Worker const uint32x4_t vxmask2 = vreinterpretq_u32_s32(vmovl_s16(vreinterpret_s16_u16(vget_low_u16(vmask1))));
66*4bdc9457SAndroid Build Coastguard Worker const uint32x4_t vf2 = vorrq_u32(vshll_n_u16(vget_low_u16(vsign1), 16),
67*4bdc9457SAndroid Build Coastguard Worker vreinterpretq_u32_f32(vbslq_f32(vxmask2, vnorm2, vdenorm2)));
68*4bdc9457SAndroid Build Coastguard Worker
69*4bdc9457SAndroid Build Coastguard Worker const uint32x4_t vxmask1 = vreinterpretq_u32_s32(vmovl_s16(vreinterpret_s16_u16(vget_high_u16(vmask0))));
70*4bdc9457SAndroid Build Coastguard Worker const uint32x4_t vf1 = vorrq_u32(vshll_n_u16(vget_high_u16(vsign0), 16),
71*4bdc9457SAndroid Build Coastguard Worker vreinterpretq_u32_f32(vbslq_f32(vxmask1, vnorm1, vdenorm1)));
72*4bdc9457SAndroid Build Coastguard Worker const uint32x4_t vxmask3 = vreinterpretq_u32_s32(vmovl_s16(vreinterpret_s16_u16(vget_high_u16(vmask1))));
73*4bdc9457SAndroid Build Coastguard Worker const uint32x4_t vf3 = vorrq_u32(vshll_n_u16(vget_high_u16(vsign1), 16),
74*4bdc9457SAndroid Build Coastguard Worker vreinterpretq_u32_f32(vbslq_f32(vxmask3, vnorm3, vdenorm3)));
75*4bdc9457SAndroid Build Coastguard Worker
76*4bdc9457SAndroid Build Coastguard Worker vst1q_f32(output, vreinterpretq_f32_u32(vf0)); output += 4;
77*4bdc9457SAndroid Build Coastguard Worker vst1q_f32(output, vreinterpretq_f32_u32(vf1)); output += 4;
78*4bdc9457SAndroid Build Coastguard Worker vst1q_f32(output, vreinterpretq_f32_u32(vf2)); output += 4;
79*4bdc9457SAndroid Build Coastguard Worker vst1q_f32(output, vreinterpretq_f32_u32(vf3)); output += 4;
80*4bdc9457SAndroid Build Coastguard Worker }
81*4bdc9457SAndroid Build Coastguard Worker for (; n >= 8 * sizeof(uint16_t); n -= 8 * sizeof(uint16_t)) {
82*4bdc9457SAndroid Build Coastguard Worker const uint16x8_t vh = vld1q_u16(i); i += 8;
83*4bdc9457SAndroid Build Coastguard Worker
84*4bdc9457SAndroid Build Coastguard Worker const uint16x8_t vsign = vandq_u16(vh, vsign_mask);
85*4bdc9457SAndroid Build Coastguard Worker
86*4bdc9457SAndroid Build Coastguard Worker const uint16x8_t vnonsign = veorq_u16(vh, vsign);
87*4bdc9457SAndroid Build Coastguard Worker
88*4bdc9457SAndroid Build Coastguard Worker const uint16x8x2_t vprenorm = vzipq_u16(vshlq_n_u16(vnonsign, 13), vsraq_n_u16(vexp_offset, vnonsign, 3));
89*4bdc9457SAndroid Build Coastguard Worker const float32x4_t vnorm_lo = vmulq_f32(vreinterpretq_f32_u16(vprenorm.val[0]), vexp_scale);
90*4bdc9457SAndroid Build Coastguard Worker const float32x4_t vnorm_hi = vmulq_f32(vreinterpretq_f32_u16(vprenorm.val[1]), vexp_scale);
91*4bdc9457SAndroid Build Coastguard Worker
92*4bdc9457SAndroid Build Coastguard Worker const float32x4_t vdenorm_lo = vsubq_f32(vreinterpretq_f32_u32(vaddw_u16(vmagic_bias, vget_low_u16(vnonsign))), vreinterpretq_f32_u32(vmagic_bias));
93*4bdc9457SAndroid Build Coastguard Worker const float32x4_t vdenorm_hi = vsubq_f32(vreinterpretq_f32_u32(vaddw_u16(vmagic_bias, vget_high_u16(vnonsign))), vreinterpretq_f32_u32(vmagic_bias));
94*4bdc9457SAndroid Build Coastguard Worker
95*4bdc9457SAndroid Build Coastguard Worker const uint16x8_t vmask = vcgtq_u16(vnonsign, vdenorm_cutoff);
96*4bdc9457SAndroid Build Coastguard Worker
97*4bdc9457SAndroid Build Coastguard Worker const uint32x4_t vxmask_lo = vreinterpretq_u32_s32(vmovl_s16(vreinterpret_s16_u16(vget_low_u16(vmask))));
98*4bdc9457SAndroid Build Coastguard Worker const uint32x4_t vf_lo = vorrq_u32(vshll_n_u16(vget_low_u16(vsign), 16),
99*4bdc9457SAndroid Build Coastguard Worker 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 = vreinterpretq_u32_s32(vmovl_s16(vreinterpret_s16_u16(vget_high_u16(vmask))));
102*4bdc9457SAndroid Build Coastguard Worker const uint32x4_t vf_hi = vorrq_u32(vshll_n_u16(vget_high_u16(vsign), 16),
103*4bdc9457SAndroid Build Coastguard Worker vreinterpretq_u32_f32(vbslq_f32(vxmask_hi, vnorm_hi, vdenorm_hi)));
104*4bdc9457SAndroid Build Coastguard Worker
105*4bdc9457SAndroid Build Coastguard Worker vst1q_f32(output, vreinterpretq_f32_u32(vf_lo)); output += 4;
106*4bdc9457SAndroid Build Coastguard Worker vst1q_f32(output, vreinterpretq_f32_u32(vf_hi)); output += 4;
107*4bdc9457SAndroid Build Coastguard Worker }
108*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(n != 0) {
109*4bdc9457SAndroid Build Coastguard Worker const uint16x8_t vh = vld1q_u16(i); i += 8;
110*4bdc9457SAndroid Build Coastguard Worker
111*4bdc9457SAndroid Build Coastguard Worker const uint16x8_t vsign = vandq_u16(vh, vsign_mask);
112*4bdc9457SAndroid Build Coastguard Worker
113*4bdc9457SAndroid Build Coastguard Worker const uint16x8_t vnonsign = veorq_u16(vh, vsign);
114*4bdc9457SAndroid Build Coastguard Worker
115*4bdc9457SAndroid Build Coastguard Worker const uint16x8x2_t vprenorm = vzipq_u16(vshlq_n_u16(vnonsign, 13), vsraq_n_u16(vexp_offset, vnonsign, 3));
116*4bdc9457SAndroid Build Coastguard Worker const float32x4_t vnorm_lo = vmulq_f32(vreinterpretq_f32_u16(vprenorm.val[0]), vexp_scale);
117*4bdc9457SAndroid Build Coastguard Worker const float32x4_t vnorm_hi = vmulq_f32(vreinterpretq_f32_u16(vprenorm.val[1]), vexp_scale);
118*4bdc9457SAndroid Build Coastguard Worker
119*4bdc9457SAndroid Build Coastguard Worker const float32x4_t vdenorm_lo = vsubq_f32(vreinterpretq_f32_u32(vaddw_u16(vmagic_bias, vget_low_u16(vnonsign))), vreinterpretq_f32_u32(vmagic_bias));
120*4bdc9457SAndroid Build Coastguard Worker const float32x4_t vdenorm_hi = vsubq_f32(vreinterpretq_f32_u32(vaddw_u16(vmagic_bias, vget_high_u16(vnonsign))), vreinterpretq_f32_u32(vmagic_bias));
121*4bdc9457SAndroid Build Coastguard Worker
122*4bdc9457SAndroid Build Coastguard Worker const uint16x8_t vmask = vcgtq_u16(vnonsign, vdenorm_cutoff);
123*4bdc9457SAndroid Build Coastguard Worker
124*4bdc9457SAndroid Build Coastguard Worker const uint32x4_t vxmask_lo = vreinterpretq_u32_s32(vmovl_s16(vreinterpret_s16_u16(vget_low_u16(vmask))));
125*4bdc9457SAndroid Build Coastguard Worker uint32x4_t vf = vorrq_u32(vshll_n_u16(vget_low_u16(vsign), 16),
126*4bdc9457SAndroid Build Coastguard Worker 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 = vreinterpretq_u32_s32(vmovl_s16(vreinterpret_s16_u16(vget_high_u16(vmask))));
132*4bdc9457SAndroid Build Coastguard Worker vf = vorrq_u32(vshll_n_u16(vget_high_u16(vsign), 16),
133*4bdc9457SAndroid Build Coastguard Worker vreinterpretq_u32_f32(vbslq_f32(vxmask_hi, vnorm_hi, vdenorm_hi)));
134*4bdc9457SAndroid Build Coastguard Worker }
135*4bdc9457SAndroid Build Coastguard Worker uint32x2_t vf_lo = vget_low_u32(vf);
136*4bdc9457SAndroid Build Coastguard Worker if (n & (2 * sizeof(uint16_t))) {
137*4bdc9457SAndroid Build Coastguard Worker vst1_f32(output, vreinterpret_f32_u32(vf_lo)); output += 2;
138*4bdc9457SAndroid Build Coastguard Worker vf_lo = vget_high_u32(vf);
139*4bdc9457SAndroid Build Coastguard Worker }
140*4bdc9457SAndroid Build Coastguard Worker if (n & (1 * sizeof(uint16_t))) {
141*4bdc9457SAndroid Build Coastguard Worker vst1_lane_f32(output, vreinterpret_f32_u32(vf_lo), 0);
142*4bdc9457SAndroid Build Coastguard Worker }
143*4bdc9457SAndroid Build Coastguard Worker }
144*4bdc9457SAndroid Build Coastguard Worker }
145