xref: /aosp_15_r20/external/XNNPACK/src/f32-f16-vcvt/gen/vcvt-sse2-x32.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-f16-vcvt/sse.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 <emmintrin.h>
13 
14 #include <xnnpack/common.h>
15 #include <xnnpack/unaligned.h>
16 #include <xnnpack/vcvt.h>
17 
18 
xnn_f32_f16_vcvt_ukernel__sse2_x32(size_t n,const float * input,void * output,const union xnn_f32_f16_cvt_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_f32_f16_vcvt_ukernel__sse2_x32(
20     size_t n,
21     const float* input,
22     void* output,
23     const union xnn_f32_f16_cvt_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
24 {
25   assert(n != 0);
26   assert(n % sizeof(float) == 0);
27   assert(input != NULL);
28   assert(output != NULL);
29 
30   const __m128 vnonsign_mask = _mm_load_ps((const float*) params->sse2.nonsign_mask);
31   const __m128i vexp_bias = _mm_load_si128((const __m128i*) params->sse2.exp_bias);
32   const __m128 vscale_to_inf = _mm_load_ps(params->sse2.scale_to_inf);
33   const __m128i vexpw_max = _mm_load_si128((const __m128i*) params->sse2.expw_max);
34   const __m128 vscale_to_zero = _mm_load_ps(params->sse2.scale_to_zero);
35   const __m128i vbias_min = _mm_load_si128((const __m128i*) params->sse2.bias_min);
36   const __m128i vmanth_mask = _mm_load_si128((const __m128i*) params->sse2.manth_mask);
37   const __m128i vexph_mask = _mm_load_si128((const __m128i*) params->sse2.exph_mask);
38   const __m128i vnanh = _mm_load_si128((const __m128i*) params->sse2.nanh);
39 
40   uint16_t* o = (uint16_t*) output;
41   for (; n >= 32 * sizeof(float); n -= 32 * sizeof(float)) {
42     const __m128 vx0 = _mm_loadu_ps(input);
43     const __m128 vx1 = _mm_loadu_ps(input + 4);
44     const __m128 vx2 = _mm_loadu_ps(input + 8);
45     const __m128 vx3 = _mm_loadu_ps(input + 12);
46     const __m128 vx4 = _mm_loadu_ps(input + 16);
47     const __m128 vx5 = _mm_loadu_ps(input + 20);
48     const __m128 vx6 = _mm_loadu_ps(input + 24);
49     const __m128 vx7 = _mm_loadu_ps(input + 28);
50     input += 32;
51 
52     const __m128 vabsx0 = _mm_and_ps(vx0, vnonsign_mask);
53     const __m128 vabsx1 = _mm_and_ps(vx1, vnonsign_mask);
54     const __m128 vabsx2 = _mm_and_ps(vx2, vnonsign_mask);
55     const __m128 vabsx3 = _mm_and_ps(vx3, vnonsign_mask);
56     const __m128 vabsx4 = _mm_and_ps(vx4, vnonsign_mask);
57     const __m128 vabsx5 = _mm_and_ps(vx5, vnonsign_mask);
58     const __m128 vabsx6 = _mm_and_ps(vx6, vnonsign_mask);
59     const __m128 vabsx7 = _mm_and_ps(vx7, vnonsign_mask);
60 
61     const __m128 vsignx0 = _mm_xor_ps(vx0, vabsx0);
62     const __m128 vsignx1 = _mm_xor_ps(vx1, vabsx1);
63     const __m128 vsignx2 = _mm_xor_ps(vx2, vabsx2);
64     const __m128 vsignx3 = _mm_xor_ps(vx3, vabsx3);
65     const __m128 vsignx4 = _mm_xor_ps(vx4, vabsx4);
66     const __m128 vsignx5 = _mm_xor_ps(vx5, vabsx5);
67     const __m128 vsignx6 = _mm_xor_ps(vx6, vabsx6);
68     const __m128 vsignx7 = _mm_xor_ps(vx7, vabsx7);
69 
70     __m128i vbias0 = _mm_add_epi32(_mm_castps_si128(vabsx0), vexp_bias);
71     __m128i vbias1 = _mm_add_epi32(_mm_castps_si128(vabsx1), vexp_bias);
72     __m128i vbias2 = _mm_add_epi32(_mm_castps_si128(vabsx2), vexp_bias);
73     __m128i vbias3 = _mm_add_epi32(_mm_castps_si128(vabsx3), vexp_bias);
74     __m128i vbias4 = _mm_add_epi32(_mm_castps_si128(vabsx4), vexp_bias);
75     __m128i vbias5 = _mm_add_epi32(_mm_castps_si128(vabsx5), vexp_bias);
76     __m128i vbias6 = _mm_add_epi32(_mm_castps_si128(vabsx6), vexp_bias);
77     __m128i vbias7 = _mm_add_epi32(_mm_castps_si128(vabsx7), vexp_bias);
78 
79     __m128 vf0 = _mm_mul_ps(vabsx0, vscale_to_inf);
80     __m128 vf1 = _mm_mul_ps(vabsx1, vscale_to_inf);
81     __m128 vf2 = _mm_mul_ps(vabsx2, vscale_to_inf);
82     __m128 vf3 = _mm_mul_ps(vabsx3, vscale_to_inf);
83     __m128 vf4 = _mm_mul_ps(vabsx4, vscale_to_inf);
84     __m128 vf5 = _mm_mul_ps(vabsx5, vscale_to_inf);
85     __m128 vf6 = _mm_mul_ps(vabsx6, vscale_to_inf);
86     __m128 vf7 = _mm_mul_ps(vabsx7, vscale_to_inf);
87 
88     const __m128i vnanmaskw0 = _mm_cmpgt_epi32(_mm_castps_si128(vabsx0), vexpw_max);
89     const __m128i vnanmaskw1 = _mm_cmpgt_epi32(_mm_castps_si128(vabsx1), vexpw_max);
90     const __m128i vnanmaskw2 = _mm_cmpgt_epi32(_mm_castps_si128(vabsx2), vexpw_max);
91     const __m128i vnanmaskw3 = _mm_cmpgt_epi32(_mm_castps_si128(vabsx3), vexpw_max);
92     const __m128i vnanmaskw4 = _mm_cmpgt_epi32(_mm_castps_si128(vabsx4), vexpw_max);
93     const __m128i vnanmaskw5 = _mm_cmpgt_epi32(_mm_castps_si128(vabsx5), vexpw_max);
94     const __m128i vnanmaskw6 = _mm_cmpgt_epi32(_mm_castps_si128(vabsx6), vexpw_max);
95     const __m128i vnanmaskw7 = _mm_cmpgt_epi32(_mm_castps_si128(vabsx7), vexpw_max);
96 
97     vbias0 = _mm_and_si128(vbias0, vexpw_max);
98     vbias1 = _mm_and_si128(vbias1, vexpw_max);
99     vbias2 = _mm_and_si128(vbias2, vexpw_max);
100     vbias3 = _mm_and_si128(vbias3, vexpw_max);
101     vbias4 = _mm_and_si128(vbias4, vexpw_max);
102     vbias5 = _mm_and_si128(vbias5, vexpw_max);
103     vbias6 = _mm_and_si128(vbias6, vexpw_max);
104     vbias7 = _mm_and_si128(vbias7, vexpw_max);
105 
106     vf0 = _mm_mul_ps(vf0, vscale_to_zero);
107     vf1 = _mm_mul_ps(vf1, vscale_to_zero);
108     vf2 = _mm_mul_ps(vf2, vscale_to_zero);
109     vf3 = _mm_mul_ps(vf3, vscale_to_zero);
110     vf4 = _mm_mul_ps(vf4, vscale_to_zero);
111     vf5 = _mm_mul_ps(vf5, vscale_to_zero);
112     vf6 = _mm_mul_ps(vf6, vscale_to_zero);
113     vf7 = _mm_mul_ps(vf7, vscale_to_zero);
114 
115     const __m128i vnanmaskh0 = _mm_packs_epi32(vnanmaskw0, vnanmaskw1);
116     const __m128i vnanmaskh1 = _mm_packs_epi32(vnanmaskw2, vnanmaskw3);
117     const __m128i vnanmaskh2 = _mm_packs_epi32(vnanmaskw4, vnanmaskw5);
118     const __m128i vnanmaskh3 = _mm_packs_epi32(vnanmaskw6, vnanmaskw7);
119 
120     const __m128i vsignh0 = _mm_packs_epi32(_mm_castps_si128(vsignx0), _mm_castps_si128(vsignx1));
121     const __m128i vsignh1 = _mm_packs_epi32(_mm_castps_si128(vsignx2), _mm_castps_si128(vsignx3));
122     const __m128i vsignh2 = _mm_packs_epi32(_mm_castps_si128(vsignx4), _mm_castps_si128(vsignx5));
123     const __m128i vsignh3 = _mm_packs_epi32(_mm_castps_si128(vsignx6), _mm_castps_si128(vsignx7));
124 
125     vbias0 = _mm_max_epi16(vbias0, vbias_min);
126     vbias1 = _mm_max_epi16(vbias1, vbias_min);
127     vbias2 = _mm_max_epi16(vbias2, vbias_min);
128     vbias3 = _mm_max_epi16(vbias3, vbias_min);
129     vbias4 = _mm_max_epi16(vbias4, vbias_min);
130     vbias5 = _mm_max_epi16(vbias5, vbias_min);
131     vbias6 = _mm_max_epi16(vbias6, vbias_min);
132     vbias7 = _mm_max_epi16(vbias7, vbias_min);
133 
134     __m128i vh0 = _mm_and_si128(vnanh, vnanmaskh0);
135     __m128i vh1 = _mm_and_si128(vnanh, vnanmaskh1);
136     __m128i vh2 = _mm_and_si128(vnanh, vnanmaskh2);
137     __m128i vh3 = _mm_and_si128(vnanh, vnanmaskh3);
138 
139     vf0 = _mm_add_ps(vf0, _mm_castsi128_ps(vbias0));
140     vf1 = _mm_add_ps(vf1, _mm_castsi128_ps(vbias1));
141     vf2 = _mm_add_ps(vf2, _mm_castsi128_ps(vbias2));
142     vf3 = _mm_add_ps(vf3, _mm_castsi128_ps(vbias3));
143     vf4 = _mm_add_ps(vf4, _mm_castsi128_ps(vbias4));
144     vf5 = _mm_add_ps(vf5, _mm_castsi128_ps(vbias5));
145     vf6 = _mm_add_ps(vf6, _mm_castsi128_ps(vbias6));
146     vf7 = _mm_add_ps(vf7, _mm_castsi128_ps(vbias7));
147 
148     vh0 = _mm_or_si128(vh0, vsignh0);
149     vh1 = _mm_or_si128(vh1, vsignh1);
150     vh2 = _mm_or_si128(vh2, vsignh2);
151     vh3 = _mm_or_si128(vh3, vsignh3);
152 
153     __m128i vexpw0 = _mm_srli_epi32(_mm_castps_si128(vf0), 13);
154     __m128i vexpw1 = _mm_srli_epi32(_mm_castps_si128(vf1), 13);
155     __m128i vexpw2 = _mm_srli_epi32(_mm_castps_si128(vf2), 13);
156     __m128i vexpw3 = _mm_srli_epi32(_mm_castps_si128(vf3), 13);
157     __m128i vexpw4 = _mm_srli_epi32(_mm_castps_si128(vf4), 13);
158     __m128i vexpw5 = _mm_srli_epi32(_mm_castps_si128(vf5), 13);
159     __m128i vexpw6 = _mm_srli_epi32(_mm_castps_si128(vf6), 13);
160     __m128i vexpw7 = _mm_srli_epi32(_mm_castps_si128(vf7), 13);
161 
162     const __m128i vmantw0 = _mm_and_si128(_mm_castps_si128(vf0), vmanth_mask);
163     const __m128i vmantw1 = _mm_and_si128(_mm_castps_si128(vf1), vmanth_mask);
164     const __m128i vmantw2 = _mm_and_si128(_mm_castps_si128(vf2), vmanth_mask);
165     const __m128i vmantw3 = _mm_and_si128(_mm_castps_si128(vf3), vmanth_mask);
166     const __m128i vmantw4 = _mm_and_si128(_mm_castps_si128(vf4), vmanth_mask);
167     const __m128i vmantw5 = _mm_and_si128(_mm_castps_si128(vf5), vmanth_mask);
168     const __m128i vmantw6 = _mm_and_si128(_mm_castps_si128(vf6), vmanth_mask);
169     const __m128i vmantw7 = _mm_and_si128(_mm_castps_si128(vf7), vmanth_mask);
170 
171     vexpw0 = _mm_and_si128(vexpw0, vexph_mask);
172     vexpw1 = _mm_and_si128(vexpw1, vexph_mask);
173     vexpw2 = _mm_and_si128(vexpw2, vexph_mask);
174     vexpw3 = _mm_and_si128(vexpw3, vexph_mask);
175     vexpw4 = _mm_and_si128(vexpw4, vexph_mask);
176     vexpw5 = _mm_and_si128(vexpw5, vexph_mask);
177     vexpw6 = _mm_and_si128(vexpw6, vexph_mask);
178     vexpw7 = _mm_and_si128(vexpw7, vexph_mask);
179 
180     const __m128i vnonsignw0 = _mm_add_epi32(vmantw0, vexpw0);
181     const __m128i vnonsignw1 = _mm_add_epi32(vmantw1, vexpw1);
182     const __m128i vnonsignw2 = _mm_add_epi32(vmantw2, vexpw2);
183     const __m128i vnonsignw3 = _mm_add_epi32(vmantw3, vexpw3);
184     const __m128i vnonsignw4 = _mm_add_epi32(vmantw4, vexpw4);
185     const __m128i vnonsignw5 = _mm_add_epi32(vmantw5, vexpw5);
186     const __m128i vnonsignw6 = _mm_add_epi32(vmantw6, vexpw6);
187     const __m128i vnonsignw7 = _mm_add_epi32(vmantw7, vexpw7);
188 
189     const __m128i vnonsignh0 = _mm_packs_epi32(vnonsignw0, vnonsignw1);
190     const __m128i vnonsignh1 = _mm_packs_epi32(vnonsignw2, vnonsignw3);
191     const __m128i vnonsignh2 = _mm_packs_epi32(vnonsignw4, vnonsignw5);
192     const __m128i vnonsignh3 = _mm_packs_epi32(vnonsignw6, vnonsignw7);
193 
194     vh0 = _mm_or_si128(vh0, _mm_andnot_si128(vnanmaskh0, vnonsignh0));
195     vh1 = _mm_or_si128(vh1, _mm_andnot_si128(vnanmaskh1, vnonsignh1));
196     vh2 = _mm_or_si128(vh2, _mm_andnot_si128(vnanmaskh2, vnonsignh2));
197     vh3 = _mm_or_si128(vh3, _mm_andnot_si128(vnanmaskh3, vnonsignh3));
198 
199     _mm_storeu_si128((__m128i*) o, vh0);
200     _mm_storeu_si128((__m128i*) (o + 8), vh1);
201     _mm_storeu_si128((__m128i*) (o + 16), vh2);
202     _mm_storeu_si128((__m128i*) (o + 24), vh3);
203     o += 32;
204   }
205   for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
206     const __m128 vx_lo = _mm_loadu_ps(input);
207     const __m128 vx_hi = _mm_loadu_ps(input + 4);
208     input += 8;
209 
210     const __m128 vabsx_lo = _mm_and_ps(vx_lo, vnonsign_mask);
211     const __m128 vabsx_hi = _mm_and_ps(vx_hi, vnonsign_mask);
212 
213     const __m128 vsignx_lo = _mm_xor_ps(vx_lo, vabsx_lo);
214     const __m128 vsignx_hi = _mm_xor_ps(vx_hi, vabsx_hi);
215     __m128i vbias_lo = _mm_add_epi32(_mm_castps_si128(vabsx_lo), vexp_bias);
216     __m128i vbias_hi = _mm_add_epi32(_mm_castps_si128(vabsx_hi), vexp_bias);
217     __m128 vf_lo = _mm_mul_ps(vabsx_lo, vscale_to_inf);
218     __m128 vf_hi = _mm_mul_ps(vabsx_hi, vscale_to_inf);
219     const __m128i vnanmaskw_lo = _mm_cmpgt_epi32(_mm_castps_si128(vabsx_lo), vexpw_max);
220     const __m128i vnanmaskw_hi = _mm_cmpgt_epi32(_mm_castps_si128(vabsx_hi), vexpw_max);
221 
222     vbias_lo = _mm_and_si128(vbias_lo, vexpw_max);
223     vbias_hi = _mm_and_si128(vbias_hi, vexpw_max);
224     vf_lo = _mm_mul_ps(vf_lo, vscale_to_zero);
225     vf_hi = _mm_mul_ps(vf_hi, vscale_to_zero);
226     const __m128i vnanmaskh = _mm_packs_epi32(vnanmaskw_lo, vnanmaskw_hi);
227     const __m128i vsignh = _mm_packs_epi32(_mm_castps_si128(vsignx_lo), _mm_castps_si128(vsignx_hi));
228 
229     vbias_lo = _mm_max_epi16(vbias_lo, vbias_min);
230     vbias_hi = _mm_max_epi16(vbias_hi, vbias_min);
231     __m128i vh = _mm_and_si128(vnanh, vnanmaskh);
232 
233     vf_lo = _mm_add_ps(vf_lo, _mm_castsi128_ps(vbias_lo));
234     vf_hi = _mm_add_ps(vf_hi, _mm_castsi128_ps(vbias_hi));
235     vh = _mm_or_si128(vh, vsignh);
236 
237     __m128i vexpw_lo = _mm_srli_epi32(_mm_castps_si128(vf_lo), 13);
238     __m128i vexpw_hi = _mm_srli_epi32(_mm_castps_si128(vf_hi), 13);
239     const __m128i vmantw_lo = _mm_and_si128(_mm_castps_si128(vf_lo), vmanth_mask);
240     const __m128i vmantw_hi = _mm_and_si128(_mm_castps_si128(vf_hi), vmanth_mask);
241 
242     vexpw_lo = _mm_and_si128(vexpw_lo, vexph_mask);
243     vexpw_hi = _mm_and_si128(vexpw_hi, vexph_mask);
244 
245     const __m128i vnonsignw_lo = _mm_add_epi32(vmantw_lo, vexpw_lo);
246     const __m128i vnonsignw_hi = _mm_add_epi32(vmantw_hi, vexpw_hi);
247 
248     const __m128i vnonsignh = _mm_packs_epi32(vnonsignw_lo, vnonsignw_hi);
249 
250     vh = _mm_or_si128(vh, _mm_andnot_si128(vnanmaskh, vnonsignh));
251 
252     _mm_storeu_si128((__m128i*) o, vh);
253     o += 8;
254   }
255   if XNN_UNPREDICTABLE(n != 0) {
256     const __m128 vx_lo = _mm_loadu_ps(input);
257     const float* input_hi = (const float*) ((uintptr_t) input + (n & (4 * sizeof(float))));
258     const __m128 vx_hi = _mm_loadu_ps(input_hi);
259 
260     const __m128 vabsx_lo = _mm_and_ps(vx_lo, vnonsign_mask);
261     const __m128 vabsx_hi = _mm_and_ps(vx_hi, vnonsign_mask);
262 
263     const __m128 vsignx_lo = _mm_xor_ps(vx_lo, vabsx_lo);
264     const __m128 vsignx_hi = _mm_xor_ps(vx_hi, vabsx_hi);
265     __m128i vbias_lo = _mm_add_epi32(_mm_castps_si128(vabsx_lo), vexp_bias);
266     __m128i vbias_hi = _mm_add_epi32(_mm_castps_si128(vabsx_hi), vexp_bias);
267     __m128 vf_lo = _mm_mul_ps(vabsx_lo, vscale_to_inf);
268     __m128 vf_hi = _mm_mul_ps(vabsx_hi, vscale_to_inf);
269     const __m128i vnanmaskw_lo = _mm_cmpgt_epi32(_mm_castps_si128(vabsx_lo), vexpw_max);
270     const __m128i vnanmaskw_hi = _mm_cmpgt_epi32(_mm_castps_si128(vabsx_hi), vexpw_max);
271 
272     vbias_lo = _mm_and_si128(vbias_lo, vexpw_max);
273     vbias_hi = _mm_and_si128(vbias_hi, vexpw_max);
274     vf_lo = _mm_mul_ps(vf_lo, vscale_to_zero);
275     vf_hi = _mm_mul_ps(vf_hi, vscale_to_zero);
276     const __m128i vnanmaskh = _mm_packs_epi32(vnanmaskw_lo, vnanmaskw_hi);
277     const __m128i vsignh = _mm_packs_epi32(_mm_castps_si128(vsignx_lo), _mm_castps_si128(vsignx_hi));
278 
279     vbias_lo = _mm_max_epi16(vbias_lo, vbias_min);
280     vbias_hi = _mm_max_epi16(vbias_hi, vbias_min);
281     __m128i vh = _mm_and_si128(vnanh, vnanmaskh);
282 
283     vf_lo = _mm_add_ps(vf_lo, _mm_castsi128_ps(vbias_lo));
284     vf_hi = _mm_add_ps(vf_hi, _mm_castsi128_ps(vbias_hi));
285     vh = _mm_or_si128(vh, vsignh);
286 
287     __m128i vexpw_lo = _mm_srli_epi32(_mm_castps_si128(vf_lo), 13);
288     __m128i vexpw_hi = _mm_srli_epi32(_mm_castps_si128(vf_hi), 13);
289     const __m128i vmantw_lo = _mm_and_si128(_mm_castps_si128(vf_lo), vmanth_mask);
290     const __m128i vmantw_hi = _mm_and_si128(_mm_castps_si128(vf_hi), vmanth_mask);
291 
292     vexpw_lo = _mm_and_si128(vexpw_lo, vexph_mask);
293     vexpw_hi = _mm_and_si128(vexpw_hi, vexph_mask);
294 
295     const __m128i vnonsignw_lo = _mm_add_epi32(vmantw_lo, vexpw_lo);
296     const __m128i vnonsignw_hi = _mm_add_epi32(vmantw_hi, vexpw_hi);
297 
298     const __m128i vnonsignh = _mm_packs_epi32(vnonsignw_lo, vnonsignw_hi);
299 
300     vh = _mm_or_si128(vh, _mm_andnot_si128(vnanmaskh, vnonsignh));
301 
302     if (n & (4 * sizeof(float))) {
303       _mm_storel_epi64((__m128i*) o, vh);
304       vh = _mm_unpackhi_epi64(vh, vh);
305       o += 4;
306     }
307     if (n & (2 * sizeof(float))) {
308       unaligned_store_u32(o, (uint32_t) _mm_cvtsi128_si32(vh));
309       vh = _mm_srli_epi64(vh, 32);
310       o += 2;
311     }
312     if (n & (1 * sizeof(float))) {
313       *o = (uint16_t) _mm_cvtsi128_si32(vh);
314     }
315   }
316 }
317