1 // Auto-generated file. Do not edit!
2 // Template: src/f16-f32-vcvt/sse-int16.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 <smmintrin.h>
13
14 #include <xnnpack/common.h>
15 #include <xnnpack/vcvt.h>
16
17
xnn_f16_f32_vcvt_ukernel__sse41_int16_x24(size_t n,const void * input,float * output,const union xnn_f16_f32_cvt_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f16_f32_vcvt_ukernel__sse41_int16_x24(
19 size_t n,
20 const void* input,
21 float* output,
22 const union xnn_f16_f32_cvt_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
23 {
24 assert(n != 0);
25 assert(n % sizeof(uint16_t) == 0);
26 assert(input != NULL);
27 assert(output != NULL);
28
29 const __m128i vsign_mask = _mm_load_si128((const __m128i*) params->sse_int16.sign_mask);
30 const __m128i vexp_offset = _mm_load_si128((const __m128i*) params->sse_int16.exp_offset);
31 const __m128 vexp_scale = _mm_load_ps(params->sse_int16.exp_scale);
32 const __m128i vmagic_mask = _mm_load_si128((const __m128i*) params->sse_int16.magic_mask);
33 const __m128 vmagic_bias = _mm_load_ps(params->sse_int16.magic_bias);
34 const __m128i vdenorm_cutoff = _mm_load_si128((const __m128i*) params->sse_int16.denorm_cutoff);
35
36 const uint16_t* i = (const uint16_t*) input;
37 for (; n >= 24 * sizeof(uint16_t); n -= 24 * sizeof(uint16_t)) {
38 const __m128i vh0 = _mm_loadu_si128((const __m128i*) i);
39 const __m128i vh1 = _mm_loadu_si128((const __m128i*) (i + 8));
40 const __m128i vh2 = _mm_loadu_si128((const __m128i*) (i + 16));
41 i += 24;
42
43 const __m128i vsign0 = _mm_and_si128(vh0, vsign_mask);
44 const __m128i vsign1 = _mm_and_si128(vh1, vsign_mask);
45 const __m128i vsign2 = _mm_and_si128(vh2, vsign_mask);
46
47 const __m128i vnonsign0 = _mm_xor_si128(vh0, vsign0);
48 const __m128i vnonsign1 = _mm_xor_si128(vh1, vsign1);
49 const __m128i vnonsign2 = _mm_xor_si128(vh2, vsign2);
50
51 const __m128i vprenorm0 = _mm_slli_epi16(vnonsign0, 13);
52 const __m128i vprenorm1 = _mm_add_epi16(_mm_srli_epi16(vnonsign0, 3), vexp_offset);
53 const __m128i vprenorm2 = _mm_slli_epi16(vnonsign1, 13);
54 const __m128i vprenorm3 = _mm_add_epi16(_mm_srli_epi16(vnonsign1, 3), vexp_offset);
55 const __m128i vprenorm4 = _mm_slli_epi16(vnonsign2, 13);
56 const __m128i vprenorm5 = _mm_add_epi16(_mm_srli_epi16(vnonsign2, 3), vexp_offset);
57
58 const __m128i vnorm0 = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vprenorm0, vprenorm1)), vexp_scale));
59 const __m128i vnorm1 = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_unpackhi_epi16(vprenorm0, vprenorm1)), vexp_scale));
60 const __m128i vnorm2 = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vprenorm2, vprenorm3)), vexp_scale));
61 const __m128i vnorm3 = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_unpackhi_epi16(vprenorm2, vprenorm3)), vexp_scale));
62 const __m128i vnorm4 = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vprenorm4, vprenorm5)), vexp_scale));
63 const __m128i vnorm5 = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_unpackhi_epi16(vprenorm4, vprenorm5)), vexp_scale));
64
65 const __m128i vdenorm0 = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vnonsign0, vmagic_mask)), vmagic_bias));
66 const __m128i vdenorm1 = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_unpackhi_epi16(vnonsign0, vmagic_mask)), vmagic_bias));
67 const __m128i vdenorm2 = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vnonsign1, vmagic_mask)), vmagic_bias));
68 const __m128i vdenorm3 = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_unpackhi_epi16(vnonsign1, vmagic_mask)), vmagic_bias));
69 const __m128i vdenorm4 = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vnonsign2, vmagic_mask)), vmagic_bias));
70 const __m128i vdenorm5 = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_unpackhi_epi16(vnonsign2, vmagic_mask)), vmagic_bias));
71
72 const __m128i vmask0 = _mm_cmpgt_epi16(vnonsign0, vdenorm_cutoff);
73 const __m128i vmask1 = _mm_cmpgt_epi16(vnonsign1, vdenorm_cutoff);
74 const __m128i vmask2 = _mm_cmpgt_epi16(vnonsign2, vdenorm_cutoff);
75
76 const __m128i vf0 = _mm_or_si128(_mm_unpacklo_epi16(_mm_setzero_si128(), vsign0),
77 _mm_blendv_epi8(vdenorm0, vnorm0, _mm_cvtepi16_epi32(vmask0)));
78 const __m128i vf1 = _mm_or_si128(_mm_unpackhi_epi16(_mm_setzero_si128(), vsign0),
79 _mm_blendv_epi8(vdenorm1, vnorm1, _mm_unpackhi_epi16(vmask0, vmask0)));
80 const __m128i vf2 = _mm_or_si128(_mm_unpacklo_epi16(_mm_setzero_si128(), vsign1),
81 _mm_blendv_epi8(vdenorm2, vnorm2, _mm_cvtepi16_epi32(vmask1)));
82 const __m128i vf3 = _mm_or_si128(_mm_unpackhi_epi16(_mm_setzero_si128(), vsign1),
83 _mm_blendv_epi8(vdenorm3, vnorm3, _mm_unpackhi_epi16(vmask1, vmask1)));
84 const __m128i vf4 = _mm_or_si128(_mm_unpacklo_epi16(_mm_setzero_si128(), vsign2),
85 _mm_blendv_epi8(vdenorm4, vnorm4, _mm_cvtepi16_epi32(vmask2)));
86 const __m128i vf5 = _mm_or_si128(_mm_unpackhi_epi16(_mm_setzero_si128(), vsign2),
87 _mm_blendv_epi8(vdenorm5, vnorm5, _mm_unpackhi_epi16(vmask2, vmask2)));
88
89 _mm_storeu_ps(output, _mm_castsi128_ps(vf0));
90 _mm_storeu_ps(output + 4, _mm_castsi128_ps(vf1));
91 _mm_storeu_ps(output + 8, _mm_castsi128_ps(vf2));
92 _mm_storeu_ps(output + 12, _mm_castsi128_ps(vf3));
93 _mm_storeu_ps(output + 16, _mm_castsi128_ps(vf4));
94 _mm_storeu_ps(output + 20, _mm_castsi128_ps(vf5));
95 output += 24;
96 }
97 for (; n >= 8 * sizeof(uint16_t); n -= 8 * sizeof(uint16_t)) {
98 const __m128i vh = _mm_loadu_si128((const __m128i*) i);
99 i += 8;
100
101 const __m128i vsign = _mm_and_si128(vh, vsign_mask);
102
103 const __m128i vnonsign = _mm_xor_si128(vh, vsign);
104
105 const __m128i vprenorm_lo = _mm_slli_epi16(vnonsign, 13);
106 const __m128i vprenorm_hi = _mm_add_epi16(_mm_srli_epi16(vnonsign, 3), vexp_offset);
107
108 const __m128i vnorm_lo = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vprenorm_lo, vprenorm_hi)), vexp_scale));
109 const __m128i vnorm_hi = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_unpackhi_epi16(vprenorm_lo, vprenorm_hi)), vexp_scale));
110
111 const __m128i vdenorm_lo = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vnonsign, vmagic_mask)), vmagic_bias));
112 const __m128i vdenorm_hi = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_unpackhi_epi16(vnonsign, vmagic_mask)), vmagic_bias));
113
114 const __m128i vmask = _mm_cmpgt_epi16(vnonsign, vdenorm_cutoff);
115
116 const __m128i vf_lo = _mm_or_si128(_mm_unpacklo_epi16(_mm_setzero_si128(), vsign),
117 _mm_blendv_epi8(vdenorm_lo, vnorm_lo, _mm_cvtepi16_epi32(vmask)));
118
119 const __m128i vf_hi = _mm_or_si128(_mm_unpackhi_epi16(_mm_setzero_si128(), vsign),
120 _mm_blendv_epi8(vdenorm_hi, vnorm_hi, _mm_unpackhi_epi16(vmask, vmask)));
121
122 _mm_storeu_ps(output, _mm_castsi128_ps(vf_lo));
123 _mm_storeu_ps(output + 4, _mm_castsi128_ps(vf_hi));
124 output += 8;
125 }
126 if XNN_UNPREDICTABLE(n != 0) {
127 const __m128i vh = _mm_loadu_si128((const __m128i*) i);
128
129 const __m128i vsign = _mm_and_si128(vh, vsign_mask);
130
131 const __m128i vnonsign = _mm_xor_si128(vh, vsign);
132
133 const __m128i vprenorm_lo = _mm_slli_epi16(vnonsign, 13);
134 const __m128i vprenorm_hi = _mm_add_epi16(_mm_srli_epi16(vnonsign, 3), vexp_offset);
135
136 const __m128i vnorm_lo = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vprenorm_lo, vprenorm_hi)), vexp_scale));
137 const __m128i vnorm_hi = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_unpackhi_epi16(vprenorm_lo, vprenorm_hi)), vexp_scale));
138
139 const __m128i vdenorm_lo = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vnonsign, vmagic_mask)), vmagic_bias));
140 const __m128i vdenorm_hi = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_unpackhi_epi16(vnonsign, vmagic_mask)), vmagic_bias));
141
142 const __m128i vmask = _mm_cmpgt_epi16(vnonsign, vdenorm_cutoff);
143
144 __m128i vf = _mm_or_si128(_mm_unpacklo_epi16(_mm_setzero_si128(), vsign),
145 _mm_blendv_epi8(vdenorm_lo, vnorm_lo, _mm_cvtepi16_epi32(vmask)));
146
147 if (n & (4 * sizeof(uint16_t))) {
148 _mm_storeu_ps(output, _mm_castsi128_ps(vf));
149 output += 4;
150
151 vf = _mm_or_si128(_mm_unpackhi_epi16(_mm_setzero_si128(), vsign),
152 _mm_blendv_epi8(vdenorm_hi, vnorm_hi, _mm_unpackhi_epi16(vmask, vmask)));
153 }
154 if (n & (2 * sizeof(uint16_t))) {
155 _mm_storel_pi((__m64*) output, _mm_castsi128_ps(vf));
156 output += 2;
157
158 vf = _mm_castps_si128(_mm_movehl_ps(_mm_castsi128_ps(vf), _mm_castsi128_ps(vf)));
159 }
160 if (n & (1 * sizeof(uint16_t))) {
161 _mm_store_ss(output, _mm_castsi128_ps(vf));
162 }
163 }
164 }
165