xref: /aosp_15_r20/external/XNNPACK/src/f32-f16-vcvt/gen/vcvt-wasmsimd-x16.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-f16-vcvt/wasmsimd.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 <wasm_simd128.h>
13 
14 #include <xnnpack/common.h>
15 #include <xnnpack/vcvt.h>
16 
17 
xnn_f32_f16_vcvt_ukernel__wasmsimd_x16(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__wasmsimd_x16(
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 v128_t vexp_bias = wasm_v128_load64_splat(params->wasmsimd.exp_bias);
30   const v128_t vscale_to_inf = wasm_v128_load64_splat(params->wasmsimd.scale_to_inf);
31   const v128_t vexpw_max = wasm_v128_load64_splat(params->wasmsimd.expw_max);
32   const v128_t vscale_to_zero = wasm_v128_load64_splat(params->wasmsimd.scale_to_zero);
33   const v128_t vbias_min = wasm_v128_load64_splat(params->wasmsimd.bias_min);
34   const v128_t vmanth_mask = wasm_v128_load64_splat(params->wasmsimd.manth_mask);
35   const v128_t vexph_mask = wasm_v128_load64_splat(params->wasmsimd.exph_mask);
36   const v128_t vnanh = wasm_v128_load64_splat(params->wasmsimd.nanh);
37 
38   uint16_t* o = (uint16_t*) output;
39   for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
40     const v128_t vx0 = wasm_v128_load(input);
41     const v128_t vx1 = wasm_v128_load(input + 4);
42     const v128_t vx2 = wasm_v128_load(input + 8);
43     const v128_t vx3 = wasm_v128_load(input + 12);
44     input += 16;
45 
46     const v128_t vabsx0 = wasm_f32x4_abs(vx0);
47     const v128_t vabsx1 = wasm_f32x4_abs(vx1);
48     const v128_t vabsx2 = wasm_f32x4_abs(vx2);
49     const v128_t vabsx3 = wasm_f32x4_abs(vx3);
50 
51     const v128_t vsignx0 = wasm_v128_xor(vx0, vabsx0);
52     const v128_t vsignx1 = wasm_v128_xor(vx1, vabsx1);
53     const v128_t vsignx2 = wasm_v128_xor(vx2, vabsx2);
54     const v128_t vsignx3 = wasm_v128_xor(vx3, vabsx3);
55 
56     v128_t vbias0 = wasm_i32x4_add(vabsx0, vexp_bias);
57     v128_t vbias1 = wasm_i32x4_add(vabsx1, vexp_bias);
58     v128_t vbias2 = wasm_i32x4_add(vabsx2, vexp_bias);
59     v128_t vbias3 = wasm_i32x4_add(vabsx3, vexp_bias);
60 
61     v128_t vf0 = wasm_f32x4_mul(vabsx0, vscale_to_inf);
62     v128_t vf1 = wasm_f32x4_mul(vabsx1, vscale_to_inf);
63     v128_t vf2 = wasm_f32x4_mul(vabsx2, vscale_to_inf);
64     v128_t vf3 = wasm_f32x4_mul(vabsx3, vscale_to_inf);
65 
66     const v128_t vnanmaskw0 = wasm_i32x4_gt(vabsx0, vexpw_max);
67     const v128_t vnanmaskw1 = wasm_i32x4_gt(vabsx1, vexpw_max);
68     const v128_t vnanmaskw2 = wasm_i32x4_gt(vabsx2, vexpw_max);
69     const v128_t vnanmaskw3 = wasm_i32x4_gt(vabsx3, vexpw_max);
70 
71     vbias0 = wasm_v128_and(vbias0, vexpw_max);
72     vbias1 = wasm_v128_and(vbias1, vexpw_max);
73     vbias2 = wasm_v128_and(vbias2, vexpw_max);
74     vbias3 = wasm_v128_and(vbias3, vexpw_max);
75 
76     vf0 = wasm_f32x4_mul(vf0, vscale_to_zero);
77     vf1 = wasm_f32x4_mul(vf1, vscale_to_zero);
78     vf2 = wasm_f32x4_mul(vf2, vscale_to_zero);
79     vf3 = wasm_f32x4_mul(vf3, vscale_to_zero);
80 
81     const v128_t vnanmaskh0 = wasm_i16x8_narrow_i32x4(vnanmaskw0, vnanmaskw1);
82     const v128_t vnanmaskh1 = wasm_i16x8_narrow_i32x4(vnanmaskw2, vnanmaskw3);
83 
84     const v128_t vsignh0 = wasm_i16x8_narrow_i32x4(vsignx0, vsignx1);
85     const v128_t vsignh1 = wasm_i16x8_narrow_i32x4(vsignx2, vsignx3);
86 
87     vbias0 = wasm_i16x8_max(vbias0, vbias_min);
88     vbias1 = wasm_i16x8_max(vbias1, vbias_min);
89     vbias2 = wasm_i16x8_max(vbias2, vbias_min);
90     vbias3 = wasm_i16x8_max(vbias3, vbias_min);
91 
92     vf0 = wasm_f32x4_add(vf0, vbias0);
93     vf1 = wasm_f32x4_add(vf1, vbias1);
94     vf2 = wasm_f32x4_add(vf2, vbias2);
95     vf3 = wasm_f32x4_add(vf3, vbias3);
96 
97     v128_t vexpw0 = wasm_i32x4_shr(vf0, 13);
98     v128_t vexpw1 = wasm_i32x4_shr(vf1, 13);
99     v128_t vexpw2 = wasm_i32x4_shr(vf2, 13);
100     v128_t vexpw3 = wasm_i32x4_shr(vf3, 13);
101 
102     const v128_t vmantw0 = wasm_v128_and(vf0, vmanth_mask);
103     const v128_t vmantw1 = wasm_v128_and(vf1, vmanth_mask);
104     const v128_t vmantw2 = wasm_v128_and(vf2, vmanth_mask);
105     const v128_t vmantw3 = wasm_v128_and(vf3, vmanth_mask);
106 
107     vexpw0 = wasm_v128_and(vexpw0, vexph_mask);
108     vexpw1 = wasm_v128_and(vexpw1, vexph_mask);
109     vexpw2 = wasm_v128_and(vexpw2, vexph_mask);
110     vexpw3 = wasm_v128_and(vexpw3, vexph_mask);
111 
112     const v128_t vnonsignw0 = wasm_i32x4_add(vmantw0, vexpw0);
113     const v128_t vnonsignw1 = wasm_i32x4_add(vmantw1, vexpw1);
114     const v128_t vnonsignw2 = wasm_i32x4_add(vmantw2, vexpw2);
115     const v128_t vnonsignw3 = wasm_i32x4_add(vmantw3, vexpw3);
116 
117     const v128_t vnonsignh0 = wasm_i16x8_narrow_i32x4(vnonsignw0, vnonsignw1);
118     const v128_t vnonsignh1 = wasm_i16x8_narrow_i32x4(vnonsignw2, vnonsignw3);
119 
120     const v128_t vabsh0 = wasm_v128_bitselect(vnanh, vnonsignh0, vnanmaskh0);
121     const v128_t vabsh1 = wasm_v128_bitselect(vnanh, vnonsignh1, vnanmaskh1);
122 
123     const v128_t vh0 = wasm_v128_or(vabsh0, vsignh0);
124     const v128_t vh1 = wasm_v128_or(vabsh1, vsignh1);
125 
126     wasm_v128_store(o, vh0);
127     wasm_v128_store(o + 8, vh1);
128     o += 16;
129   }
130   for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
131     const v128_t vx_lo = wasm_v128_load(input);
132     const v128_t vx_hi = wasm_v128_load(input + 4);
133     input += 8;
134 
135     const v128_t vabsx_lo = wasm_f32x4_abs(vx_lo);
136     const v128_t vabsx_hi = wasm_f32x4_abs(vx_hi);
137 
138     const v128_t vsignx_lo = wasm_v128_xor(vx_lo, vabsx_lo);
139     const v128_t vsignx_hi = wasm_v128_xor(vx_hi, vabsx_hi);
140     v128_t vbias_lo = wasm_i32x4_add(vabsx_lo, vexp_bias);
141     v128_t vbias_hi = wasm_i32x4_add(vabsx_hi, vexp_bias);
142     v128_t vf_lo = wasm_f32x4_mul(vabsx_lo, vscale_to_inf);
143     v128_t vf_hi = wasm_f32x4_mul(vabsx_hi, vscale_to_inf);
144     const v128_t vnanmaskw_lo = wasm_i32x4_gt(vabsx_lo, vexpw_max);
145     const v128_t vnanmaskw_hi = wasm_i32x4_gt(vabsx_hi, vexpw_max);
146 
147     vbias_lo = wasm_v128_and(vbias_lo, vexpw_max);
148     vbias_hi = wasm_v128_and(vbias_hi, vexpw_max);
149     vf_lo = wasm_f32x4_mul(vf_lo, vscale_to_zero);
150     vf_hi = wasm_f32x4_mul(vf_hi, vscale_to_zero);
151     const v128_t vnanmaskh = wasm_i16x8_narrow_i32x4(vnanmaskw_lo, vnanmaskw_hi);
152     const v128_t vsignh = wasm_i16x8_narrow_i32x4(vsignx_lo, vsignx_hi);
153 
154     vbias_lo = wasm_i16x8_max(vbias_lo, vbias_min);
155     vbias_hi = wasm_i16x8_max(vbias_hi, vbias_min);
156 
157     vf_lo = wasm_f32x4_add(vf_lo, vbias_lo);
158     vf_hi = wasm_f32x4_add(vf_hi, vbias_hi);
159 
160     v128_t vexpw_lo = wasm_i32x4_shr(vf_lo, 13);
161     v128_t vexpw_hi = wasm_i32x4_shr(vf_hi, 13);
162     const v128_t vmantw_lo = wasm_v128_and(vf_lo, vmanth_mask);
163     const v128_t vmantw_hi = wasm_v128_and(vf_hi, vmanth_mask);
164 
165     vexpw_lo = wasm_v128_and(vexpw_lo, vexph_mask);
166     vexpw_hi = wasm_v128_and(vexpw_hi, vexph_mask);
167 
168     const v128_t vnonsignw_lo = wasm_i32x4_add(vmantw_lo, vexpw_lo);
169     const v128_t vnonsignw_hi = wasm_i32x4_add(vmantw_hi, vexpw_hi);
170 
171     const v128_t vnonsignh = wasm_i16x8_narrow_i32x4(vnonsignw_lo, vnonsignw_hi);
172 
173     const v128_t vabsh = wasm_v128_bitselect(vnanh, vnonsignh, vnanmaskh);
174 
175     const v128_t vh = wasm_v128_or(vabsh, vsignh);
176 
177     wasm_v128_store(o, vh);
178     o += 8;
179   }
180   if XNN_UNPREDICTABLE(n != 0) {
181     const v128_t vx_lo = wasm_v128_load(input);
182     const float* input_hi = (const float*) ((uintptr_t) input + (n & (4 * sizeof(float))));
183     const v128_t vx_hi = wasm_v128_load(input_hi);
184 
185     const v128_t vabsx_lo = wasm_f32x4_abs(vx_lo);
186     const v128_t vabsx_hi = wasm_f32x4_abs(vx_hi);
187 
188     const v128_t vsignx_lo = wasm_v128_xor(vx_lo, vabsx_lo);
189     const v128_t vsignx_hi = wasm_v128_xor(vx_hi, vabsx_hi);
190     v128_t vbias_lo = wasm_i32x4_add(vabsx_lo, vexp_bias);
191     v128_t vbias_hi = wasm_i32x4_add(vabsx_hi, vexp_bias);
192     v128_t vf_lo = wasm_f32x4_mul(vabsx_lo, vscale_to_inf);
193     v128_t vf_hi = wasm_f32x4_mul(vabsx_hi, vscale_to_inf);
194     const v128_t vnanmaskw_lo = wasm_i32x4_gt(vabsx_lo, vexpw_max);
195     const v128_t vnanmaskw_hi = wasm_i32x4_gt(vabsx_hi, vexpw_max);
196 
197     vbias_lo = wasm_v128_and(vbias_lo, vexpw_max);
198     vbias_hi = wasm_v128_and(vbias_hi, vexpw_max);
199     vf_lo = wasm_f32x4_mul(vf_lo, vscale_to_zero);
200     vf_hi = wasm_f32x4_mul(vf_hi, vscale_to_zero);
201     const v128_t vnanmaskh = wasm_i16x8_narrow_i32x4(vnanmaskw_lo, vnanmaskw_hi);
202     const v128_t vsignh = wasm_i16x8_narrow_i32x4(vsignx_lo, vsignx_hi);
203 
204     vbias_lo = wasm_i16x8_max(vbias_lo, vbias_min);
205     vbias_hi = wasm_i16x8_max(vbias_hi, vbias_min);
206 
207     vf_lo = wasm_f32x4_add(vf_lo, vbias_lo);
208     vf_hi = wasm_f32x4_add(vf_hi, vbias_hi);
209 
210     v128_t vexpw_lo = wasm_i32x4_shr(vf_lo, 13);
211     v128_t vexpw_hi = wasm_i32x4_shr(vf_hi, 13);
212     const v128_t vmantw_lo = wasm_v128_and(vf_lo, vmanth_mask);
213     const v128_t vmantw_hi = wasm_v128_and(vf_hi, vmanth_mask);
214 
215     vexpw_lo = wasm_v128_and(vexpw_lo, vexph_mask);
216     vexpw_hi = wasm_v128_and(vexpw_hi, vexph_mask);
217 
218     const v128_t vnonsignw_lo = wasm_i32x4_add(vmantw_lo, vexpw_lo);
219     const v128_t vnonsignw_hi = wasm_i32x4_add(vmantw_hi, vexpw_hi);
220 
221     const v128_t vnonsignh = wasm_i16x8_narrow_i32x4(vnonsignw_lo, vnonsignw_hi);
222 
223     const v128_t vabsh = wasm_v128_bitselect(vnanh, vnonsignh, vnanmaskh);
224 
225     v128_t vh = wasm_v128_or(vabsh, vsignh);
226 
227     if (n & (4 * sizeof(float))) {
228       *((double*) o) = wasm_f64x2_extract_lane(vh, 0);
229       vh = wasm_v64x2_shuffle(vh, vh, 1, 1);
230       o += 4;
231     }
232     if (n & (2 * sizeof(float))) {
233       *((float*) o) = (float) wasm_f32x4_extract_lane(vh, 0);
234       vh = wasm_i64x2_shr(vh, 32);
235       o += 2;
236     }
237     const uint32_t vh_lo = wasm_i32x4_extract_lane(vh, 0);
238     if (n & (1 * sizeof(float))) {
239       *o = (uint16_t) vh_lo;
240     }
241   }
242 }
243