1 // Auto-generated file. Do not edit!
2 // Template: src/f16-f32-vcvt/sse-int32.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_int32_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_int32_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_int32.sign_mask);
30 const __m128i vexp_offset = _mm_load_si128((const __m128i*) params->sse_int32.exp_offset);
31 const __m128 vexp_scale = _mm_load_ps(params->sse_int32.exp_scale);
32 const __m128i vmagic_bias = _mm_load_si128((const __m128i*) params->sse_int32.magic_bias);
33 const __m128i vdenorm_cutoff = _mm_load_si128((const __m128i*) params->sse_int32.denorm_cutoff);
34
35 const uint16_t* i = (const uint16_t*) input;
36 for (; n >= 24 * sizeof(uint16_t); n -= 24 * sizeof(uint16_t)) {
37 const __m128i vh0 = _mm_loadu_si128((const __m128i*) i);
38 const __m128i vh1 = _mm_loadu_si128((const __m128i*) (i + 8));
39 const __m128i vh2 = _mm_loadu_si128((const __m128i*) (i + 16));
40 i += 24;
41
42 const __m128i vw0 = _mm_unpacklo_epi16(_mm_setzero_si128(), vh0);
43 const __m128i vw1 = _mm_unpackhi_epi16(_mm_setzero_si128(), vh0);
44 const __m128i vw2 = _mm_unpacklo_epi16(_mm_setzero_si128(), vh1);
45 const __m128i vw3 = _mm_unpackhi_epi16(_mm_setzero_si128(), vh1);
46 const __m128i vw4 = _mm_unpacklo_epi16(_mm_setzero_si128(), vh2);
47 const __m128i vw5 = _mm_unpackhi_epi16(_mm_setzero_si128(), vh2);
48
49 const __m128i vsign0 = _mm_and_si128(vw0, vsign_mask);
50 const __m128i vsign1 = _mm_and_si128(vw1, vsign_mask);
51 const __m128i vsign2 = _mm_and_si128(vw2, vsign_mask);
52 const __m128i vsign3 = _mm_and_si128(vw3, vsign_mask);
53 const __m128i vsign4 = _mm_and_si128(vw4, vsign_mask);
54 const __m128i vsign5 = _mm_and_si128(vw5, vsign_mask);
55
56 const __m128i vnonsign0 = _mm_xor_si128(vw0, vsign0);
57 const __m128i vnonsign1 = _mm_xor_si128(vw1, vsign1);
58 const __m128i vnonsign2 = _mm_xor_si128(vw2, vsign2);
59 const __m128i vnonsign3 = _mm_xor_si128(vw3, vsign3);
60 const __m128i vnonsign4 = _mm_xor_si128(vw4, vsign4);
61 const __m128i vnonsign5 = _mm_xor_si128(vw5, vsign5);
62
63 const __m128i vnorm0 = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_add_epi32(_mm_srli_epi32(vnonsign0, 3), vexp_offset)), vexp_scale));
64 const __m128i vnorm1 = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_add_epi32(_mm_srli_epi32(vnonsign1, 3), vexp_offset)), vexp_scale));
65 const __m128i vnorm2 = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_add_epi32(_mm_srli_epi32(vnonsign2, 3), vexp_offset)), vexp_scale));
66 const __m128i vnorm3 = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_add_epi32(_mm_srli_epi32(vnonsign3, 3), vexp_offset)), vexp_scale));
67 const __m128i vnorm4 = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_add_epi32(_mm_srli_epi32(vnonsign4, 3), vexp_offset)), vexp_scale));
68 const __m128i vnorm5 = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_add_epi32(_mm_srli_epi32(vnonsign5, 3), vexp_offset)), vexp_scale));
69
70 const __m128i vdenorm0 = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_or_si128(_mm_srli_epi32(vnonsign0, 16), vmagic_bias)), _mm_castsi128_ps(vmagic_bias)));
71 const __m128i vdenorm1 = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_or_si128(_mm_srli_epi32(vnonsign1, 16), vmagic_bias)), _mm_castsi128_ps(vmagic_bias)));
72 const __m128i vdenorm2 = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_or_si128(_mm_srli_epi32(vnonsign2, 16), vmagic_bias)), _mm_castsi128_ps(vmagic_bias)));
73 const __m128i vdenorm3 = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_or_si128(_mm_srli_epi32(vnonsign3, 16), vmagic_bias)), _mm_castsi128_ps(vmagic_bias)));
74 const __m128i vdenorm4 = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_or_si128(_mm_srli_epi32(vnonsign4, 16), vmagic_bias)), _mm_castsi128_ps(vmagic_bias)));
75 const __m128i vdenorm5 = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_or_si128(_mm_srli_epi32(vnonsign5, 16), vmagic_bias)), _mm_castsi128_ps(vmagic_bias)));
76
77 const __m128i vmask0 = _mm_cmpgt_epi32(vnonsign0, vdenorm_cutoff);
78 const __m128i vmask1 = _mm_cmpgt_epi32(vnonsign1, vdenorm_cutoff);
79 const __m128i vmask2 = _mm_cmpgt_epi32(vnonsign2, vdenorm_cutoff);
80 const __m128i vmask3 = _mm_cmpgt_epi32(vnonsign3, vdenorm_cutoff);
81 const __m128i vmask4 = _mm_cmpgt_epi32(vnonsign4, vdenorm_cutoff);
82 const __m128i vmask5 = _mm_cmpgt_epi32(vnonsign5, vdenorm_cutoff);
83
84 const __m128i vf0 = _mm_or_si128(vsign0, _mm_blendv_epi8(vdenorm0, vnorm0, vmask0));
85 const __m128i vf1 = _mm_or_si128(vsign1, _mm_blendv_epi8(vdenorm1, vnorm1, vmask1));
86 const __m128i vf2 = _mm_or_si128(vsign2, _mm_blendv_epi8(vdenorm2, vnorm2, vmask2));
87 const __m128i vf3 = _mm_or_si128(vsign3, _mm_blendv_epi8(vdenorm3, vnorm3, vmask3));
88 const __m128i vf4 = _mm_or_si128(vsign4, _mm_blendv_epi8(vdenorm4, vnorm4, vmask4));
89 const __m128i vf5 = _mm_or_si128(vsign5, _mm_blendv_epi8(vdenorm5, vnorm5, vmask5));
90
91 _mm_storeu_ps(output, _mm_castsi128_ps(vf0));
92 _mm_storeu_ps(output + 4, _mm_castsi128_ps(vf1));
93 _mm_storeu_ps(output + 8, _mm_castsi128_ps(vf2));
94 _mm_storeu_ps(output + 12, _mm_castsi128_ps(vf3));
95 _mm_storeu_ps(output + 16, _mm_castsi128_ps(vf4));
96 _mm_storeu_ps(output + 20, _mm_castsi128_ps(vf5));
97 output += 24;
98 }
99 for (; n >= 8 * sizeof(uint16_t); n -= 8 * sizeof(uint16_t)) {
100 const __m128i vh = _mm_loadu_si128((const __m128i*) i);
101 i += 8;
102
103 const __m128i vw_lo = _mm_unpacklo_epi16(_mm_setzero_si128(), vh);
104 const __m128i vw_hi = _mm_unpackhi_epi16(_mm_setzero_si128(), vh);
105
106 const __m128i vsign_lo = _mm_and_si128(vw_lo, vsign_mask);
107 const __m128i vsign_hi = _mm_and_si128(vw_hi, vsign_mask);
108
109 const __m128i vnonsign_lo = _mm_xor_si128(vw_lo, vsign_lo);
110 const __m128i vnonsign_hi = _mm_xor_si128(vw_hi, vsign_hi);
111
112 const __m128i vnorm_lo = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_add_epi32(_mm_srli_epi32(vnonsign_lo, 3), vexp_offset)), vexp_scale));
113 const __m128i vnorm_hi = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_add_epi32(_mm_srli_epi32(vnonsign_hi, 3), vexp_offset)), vexp_scale));
114
115 const __m128i vdenorm_lo = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_or_si128(_mm_srli_epi32(vnonsign_lo, 16), vmagic_bias)), _mm_castsi128_ps(vmagic_bias)));
116 const __m128i vdenorm_hi = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_or_si128(_mm_srli_epi32(vnonsign_hi, 16), vmagic_bias)), _mm_castsi128_ps(vmagic_bias)));
117
118 const __m128i vmask_lo = _mm_cmpgt_epi32(vnonsign_lo, vdenorm_cutoff);
119 const __m128i vf_lo = _mm_or_si128(vsign_lo, _mm_blendv_epi8(vdenorm_lo, vnorm_lo, vmask_lo));
120
121 const __m128i vmask_hi = _mm_cmpgt_epi32(vnonsign_hi, vdenorm_cutoff);
122 const __m128i vf_hi = _mm_or_si128(vsign_hi, _mm_blendv_epi8(vdenorm_hi, vnorm_hi, vmask_hi));
123
124 _mm_storeu_ps(output, _mm_castsi128_ps(vf_lo));
125 _mm_storeu_ps(output + 4, _mm_castsi128_ps(vf_hi));
126 output += 8;
127 }
128 if XNN_UNPREDICTABLE(n != 0) {
129 const __m128i vh = _mm_loadu_si128((const __m128i*) i);
130
131 const __m128i vw_lo = _mm_unpacklo_epi16(_mm_setzero_si128(), vh);
132 const __m128i vw_hi = _mm_unpackhi_epi16(_mm_setzero_si128(), vh);
133
134 const __m128i vsign_lo = _mm_and_si128(vw_lo, vsign_mask);
135 const __m128i vsign_hi = _mm_and_si128(vw_hi, vsign_mask);
136
137 const __m128i vnonsign_lo = _mm_xor_si128(vw_lo, vsign_lo);
138 const __m128i vnonsign_hi = _mm_xor_si128(vw_hi, vsign_hi);
139
140 const __m128i vnorm_lo = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_add_epi32(_mm_srli_epi32(vnonsign_lo, 3), vexp_offset)), vexp_scale));
141 const __m128i vnorm_hi = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_add_epi32(_mm_srli_epi32(vnonsign_hi, 3), vexp_offset)), vexp_scale));
142
143 const __m128i vdenorm_lo = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_or_si128(_mm_srli_epi32(vnonsign_lo, 16), vmagic_bias)), _mm_castsi128_ps(vmagic_bias)));
144 const __m128i vdenorm_hi = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_or_si128(_mm_srli_epi32(vnonsign_hi, 16), vmagic_bias)), _mm_castsi128_ps(vmagic_bias)));
145
146 const __m128i vmask_lo = _mm_cmpgt_epi32(vnonsign_lo, vdenorm_cutoff);
147 __m128i vf = _mm_or_si128(vsign_lo, _mm_blendv_epi8(vdenorm_lo, vnorm_lo, vmask_lo));
148
149 if (n & (4 * sizeof(uint16_t))) {
150 _mm_storeu_ps(output, _mm_castsi128_ps(vf));
151 output += 4;
152
153 const __m128i vmask_hi = _mm_cmpgt_epi32(vnonsign_hi, vdenorm_cutoff);
154 vf = _mm_or_si128(vsign_hi, _mm_blendv_epi8(vdenorm_hi, vnorm_hi, vmask_hi));
155 }
156 if (n & (2 * sizeof(uint16_t))) {
157 _mm_storel_pi((__m64*) output, _mm_castsi128_ps(vf));
158 output += 2;
159
160 vf = _mm_castps_si128(_mm_movehl_ps(_mm_castsi128_ps(vf), _mm_castsi128_ps(vf)));
161 }
162 if (n & (1 * sizeof(uint16_t))) {
163 _mm_store_ss(output, _mm_castsi128_ps(vf));
164 }
165 }
166 }
167