Eigen  3.4.90 (git rev 5a9f66fb35d03a4da9ef8976e67a61b30aa16dcf)
 
Loading...
Searching...
No Matches
AVX/TypeCasting.h
1// This file is part of Eigen, a lightweight C++ template library
2// for linear algebra.
3//
4// Copyright (C) 2015 Benoit Steiner <[email protected]>
5//
6// This Source Code Form is subject to the terms of the Mozilla
7// Public License v. 2.0. If a copy of the MPL was not distributed
8// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
9
10#ifndef EIGEN_TYPE_CASTING_AVX_H
11#define EIGEN_TYPE_CASTING_AVX_H
12
13// IWYU pragma: private
14#include "../../InternalHeaderCheck.h"
15
16namespace Eigen {
17
18namespace internal {
19
20#ifndef EIGEN_VECTORIZE_AVX512
21template <>
22struct type_casting_traits<float, bool> : vectorized_type_casting_traits<float, bool> {};
23template <>
24struct type_casting_traits<bool, float> : vectorized_type_casting_traits<bool, float> {};
25
26template <>
27struct type_casting_traits<float, int> : vectorized_type_casting_traits<float, int> {};
28template <>
29struct type_casting_traits<int, float> : vectorized_type_casting_traits<int, float> {};
30
31template <>
32struct type_casting_traits<float, double> : vectorized_type_casting_traits<float, double> {};
33template <>
34struct type_casting_traits<double, float> : vectorized_type_casting_traits<double, float> {};
35
36template <>
37struct type_casting_traits<double, int> : vectorized_type_casting_traits<double, int> {};
38template <>
39struct type_casting_traits<int, double> : vectorized_type_casting_traits<int, double> {};
40
41template <>
42struct type_casting_traits<half, float> : vectorized_type_casting_traits<half, float> {};
43template <>
44struct type_casting_traits<float, half> : vectorized_type_casting_traits<float, half> {};
45
46template <>
47struct type_casting_traits<bfloat16, float> : vectorized_type_casting_traits<bfloat16, float> {};
48template <>
49struct type_casting_traits<float, bfloat16> : vectorized_type_casting_traits<float, bfloat16> {};
50
51#ifdef EIGEN_VECTORIZE_AVX2
52template <>
53struct type_casting_traits<double, int64_t> : vectorized_type_casting_traits<double, int64_t> {};
54template <>
55struct type_casting_traits<int64_t, double> : vectorized_type_casting_traits<int64_t, double> {};
56#endif
57#endif
58
59template <>
60EIGEN_STRONG_INLINE Packet16b pcast<Packet8f, Packet16b>(const Packet8f& a, const Packet8f& b) {
61 __m256 nonzero_a = _mm256_cmp_ps(a, pzero(a), _CMP_NEQ_UQ);
62 __m256 nonzero_b = _mm256_cmp_ps(b, pzero(b), _CMP_NEQ_UQ);
63 constexpr char kFF = '\255';
64#ifndef EIGEN_VECTORIZE_AVX2
65 __m128i shuffle_mask128_a_lo = _mm_set_epi8(kFF, kFF, kFF, kFF, kFF, kFF, kFF, kFF, kFF, kFF, kFF, kFF, 12, 8, 4, 0);
66 __m128i shuffle_mask128_a_hi = _mm_set_epi8(kFF, kFF, kFF, kFF, kFF, kFF, kFF, kFF, 12, 8, 4, 0, kFF, kFF, kFF, kFF);
67 __m128i shuffle_mask128_b_lo = _mm_set_epi8(kFF, kFF, kFF, kFF, 12, 8, 4, 0, kFF, kFF, kFF, kFF, kFF, kFF, kFF, kFF);
68 __m128i shuffle_mask128_b_hi = _mm_set_epi8(12, 8, 4, 0, kFF, kFF, kFF, kFF, kFF, kFF, kFF, kFF, kFF, kFF, kFF, kFF);
69 __m128i a_hi = _mm_shuffle_epi8(_mm256_extractf128_si256(_mm256_castps_si256(nonzero_a), 1), shuffle_mask128_a_hi);
70 __m128i a_lo = _mm_shuffle_epi8(_mm256_extractf128_si256(_mm256_castps_si256(nonzero_a), 0), shuffle_mask128_a_lo);
71 __m128i b_hi = _mm_shuffle_epi8(_mm256_extractf128_si256(_mm256_castps_si256(nonzero_b), 1), shuffle_mask128_b_hi);
72 __m128i b_lo = _mm_shuffle_epi8(_mm256_extractf128_si256(_mm256_castps_si256(nonzero_b), 0), shuffle_mask128_b_lo);
73 __m128i merged = _mm_or_si128(_mm_or_si128(b_lo, b_hi), _mm_or_si128(a_lo, a_hi));
74 return _mm_and_si128(merged, _mm_set1_epi8(1));
75#else
76 __m256i a_shuffle_mask = _mm256_set_epi8(kFF, kFF, kFF, kFF, kFF, kFF, kFF, kFF, 12, 8, 4, 0, kFF, kFF, kFF, kFF, kFF,
77 kFF, kFF, kFF, kFF, kFF, kFF, kFF, kFF, kFF, kFF, kFF, 12, 8, 4, 0);
78 __m256i b_shuffle_mask = _mm256_set_epi8(12, 8, 4, 0, kFF, kFF, kFF, kFF, kFF, kFF, kFF, kFF, kFF, kFF, kFF, kFF, kFF,
79 kFF, kFF, kFF, 12, 8, 4, 0, kFF, kFF, kFF, kFF, kFF, kFF, kFF, kFF);
80 __m256i a_shuff = _mm256_shuffle_epi8(_mm256_castps_si256(nonzero_a), a_shuffle_mask);
81 __m256i b_shuff = _mm256_shuffle_epi8(_mm256_castps_si256(nonzero_b), b_shuffle_mask);
82 __m256i a_or_b = _mm256_or_si256(a_shuff, b_shuff);
83 __m256i merged = _mm256_or_si256(a_or_b, _mm256_castsi128_si256(_mm256_extractf128_si256(a_or_b, 1)));
84 return _mm256_castsi256_si128(_mm256_and_si256(merged, _mm256_set1_epi8(1)));
85#endif
86}
87
88template <>
89EIGEN_STRONG_INLINE Packet8f pcast<Packet16b, Packet8f>(const Packet16b& a) {
90 const __m256 cst_one = _mm256_set1_ps(1.0f);
91#ifdef EIGEN_VECTORIZE_AVX2
92 __m256i a_extended = _mm256_cvtepi8_epi32(a);
93 __m256i abcd_efgh = _mm256_cmpeq_epi32(a_extended, _mm256_setzero_si256());
94#else
95 __m128i abcd_efhg_ijkl_mnop = _mm_cmpeq_epi8(a, _mm_setzero_si128());
96 __m128i aabb_ccdd_eeff_gghh = _mm_unpacklo_epi8(abcd_efhg_ijkl_mnop, abcd_efhg_ijkl_mnop);
97 __m128i aaaa_bbbb_cccc_dddd = _mm_unpacklo_epi8(aabb_ccdd_eeff_gghh, aabb_ccdd_eeff_gghh);
98 __m128i eeee_ffff_gggg_hhhh = _mm_unpackhi_epi8(aabb_ccdd_eeff_gghh, aabb_ccdd_eeff_gghh);
99 __m256i abcd_efgh = _mm256_setr_m128i(aaaa_bbbb_cccc_dddd, eeee_ffff_gggg_hhhh);
100#endif
101 __m256 result = _mm256_andnot_ps(_mm256_castsi256_ps(abcd_efgh), cst_one);
102 return result;
103}
104
105template <>
106EIGEN_STRONG_INLINE Packet8i pcast<Packet8f, Packet8i>(const Packet8f& a) {
107 return _mm256_cvttps_epi32(a);
108}
109
110template <>
111EIGEN_STRONG_INLINE Packet8i pcast<Packet4d, Packet8i>(const Packet4d& a, const Packet4d& b) {
112 return _mm256_set_m128i(_mm256_cvttpd_epi32(b), _mm256_cvttpd_epi32(a));
113}
114
115template <>
116EIGEN_STRONG_INLINE Packet4i pcast<Packet4d, Packet4i>(const Packet4d& a) {
117 return _mm256_cvttpd_epi32(a);
118}
119
120template <>
121EIGEN_STRONG_INLINE Packet8f pcast<Packet8i, Packet8f>(const Packet8i& a) {
122 return _mm256_cvtepi32_ps(a);
123}
124
125template <>
126EIGEN_STRONG_INLINE Packet8f pcast<Packet4d, Packet8f>(const Packet4d& a, const Packet4d& b) {
127 return _mm256_set_m128(_mm256_cvtpd_ps(b), _mm256_cvtpd_ps(a));
128}
129
130template <>
131EIGEN_STRONG_INLINE Packet4f pcast<Packet4d, Packet4f>(const Packet4d& a) {
132 return _mm256_cvtpd_ps(a);
133}
134
135template <>
136EIGEN_STRONG_INLINE Packet4d pcast<Packet8i, Packet4d>(const Packet8i& a) {
137 return _mm256_cvtepi32_pd(_mm256_castsi256_si128(a));
138}
139
140template <>
141EIGEN_STRONG_INLINE Packet4d pcast<Packet4i, Packet4d>(const Packet4i& a) {
142 return _mm256_cvtepi32_pd(a);
143}
144
145template <>
146EIGEN_STRONG_INLINE Packet4d pcast<Packet8f, Packet4d>(const Packet8f& a) {
147 return _mm256_cvtps_pd(_mm256_castps256_ps128(a));
148}
149
150template <>
151EIGEN_STRONG_INLINE Packet4d pcast<Packet4f, Packet4d>(const Packet4f& a) {
152 return _mm256_cvtps_pd(a);
153}
154
155template <>
156EIGEN_STRONG_INLINE Packet8i preinterpret<Packet8i, Packet8f>(const Packet8f& a) {
157 return _mm256_castps_si256(a);
158}
159
160template <>
161EIGEN_STRONG_INLINE Packet8f preinterpret<Packet8f, Packet8i>(const Packet8i& a) {
162 return _mm256_castsi256_ps(a);
163}
164
165template <>
166EIGEN_STRONG_INLINE Packet8ui preinterpret<Packet8ui, Packet8i>(const Packet8i& a) {
167 return Packet8ui(a);
168}
169
170template <>
171EIGEN_STRONG_INLINE Packet8i preinterpret<Packet8i, Packet8ui>(const Packet8ui& a) {
172 return Packet8i(a);
173}
174
175// truncation operations
176
177template <>
178EIGEN_STRONG_INLINE Packet4f preinterpret<Packet4f, Packet8f>(const Packet8f& a) {
179 return _mm256_castps256_ps128(a);
180}
181
182template <>
183EIGEN_STRONG_INLINE Packet2d preinterpret<Packet2d, Packet4d>(const Packet4d& a) {
184 return _mm256_castpd256_pd128(a);
185}
186
187template <>
188EIGEN_STRONG_INLINE Packet4i preinterpret<Packet4i, Packet8i>(const Packet8i& a) {
189 return _mm256_castsi256_si128(a);
190}
191
192template <>
193EIGEN_STRONG_INLINE Packet4ui preinterpret<Packet4ui, Packet8ui>(const Packet8ui& a) {
194 return _mm256_castsi256_si128(a);
195}
196
197#ifdef EIGEN_VECTORIZE_AVX2
198template <>
199EIGEN_STRONG_INLINE Packet4l pcast<Packet4d, Packet4l>(const Packet4d& a) {
200#if defined(EIGEN_VECTORIZE_AVX512DQ) && defined(EIGEN_VECTORIZE_AVS512VL)
201 return _mm256_cvttpd_epi64(a);
202#else
203
204 // if 'a' exceeds the numerical limits of int64_t, the behavior is undefined
205
206 // e <= 0 corresponds to |a| < 1, which should result in zero. incidentally, intel intrinsics with shift arguments
207 // greater than or equal to 64 produce zero. furthermore, negative shifts appear to be interpreted as large positive
208 // shifts (two's complement), which also result in zero. therefore, e does not need to be clamped to [0, 64)
209
210 constexpr int kTotalBits = sizeof(double) * CHAR_BIT, kMantissaBits = std::numeric_limits<double>::digits - 1,
211 kExponentBits = kTotalBits - kMantissaBits - 1, kBias = (1 << (kExponentBits - 1)) - 1;
212
213 const __m256i cst_one = _mm256_set1_epi64x(1);
214 const __m256i cst_total_bits = _mm256_set1_epi64x(kTotalBits);
215 const __m256i cst_bias = _mm256_set1_epi64x(kBias);
216
217 __m256i a_bits = _mm256_castpd_si256(a);
218 // shift left by 1 to clear the sign bit, and shift right by kMantissaBits + 1 to recover biased exponent
219 __m256i biased_e = _mm256_srli_epi64(_mm256_slli_epi64(a_bits, 1), kMantissaBits + 1);
220 __m256i e = _mm256_sub_epi64(biased_e, cst_bias);
221
222 // shift to the left by kExponentBits + 1 to clear the sign and exponent bits
223 __m256i shifted_mantissa = _mm256_slli_epi64(a_bits, kExponentBits + 1);
224 // shift to the right by kTotalBits - e to convert the significand to an integer
225 __m256i result_significand = _mm256_srlv_epi64(shifted_mantissa, _mm256_sub_epi64(cst_total_bits, e));
226
227 // add the implied bit
228 __m256i result_exponent = _mm256_sllv_epi64(cst_one, e);
229 // e <= 0 is interpreted as a large positive shift (2's complement), which also conveniently results in zero
230 __m256i result = _mm256_add_epi64(result_significand, result_exponent);
231 // handle negative arguments
232 __m256i sign_mask = _mm256_cmpgt_epi64(_mm256_setzero_si256(), a_bits);
233 result = _mm256_sub_epi64(_mm256_xor_si256(result, sign_mask), sign_mask);
234 return result;
235#endif
236}
237
238template <>
239EIGEN_STRONG_INLINE Packet4d pcast<Packet4l, Packet4d>(const Packet4l& a) {
240#if defined(EIGEN_VECTORIZE_AVX512DQ) && defined(EIGEN_VECTORIZE_AVS512VL)
241 return _mm256_cvtepi64_pd(a);
242#else
243 EIGEN_ALIGN16 int64_t aux[4];
244 pstore(aux, a);
245 return _mm256_set_pd(static_cast<double>(aux[3]), static_cast<double>(aux[2]), static_cast<double>(aux[1]),
246 static_cast<double>(aux[0]));
247#endif
248}
249
250template <>
251EIGEN_STRONG_INLINE Packet4d pcast<Packet2l, Packet4d>(const Packet2l& a, const Packet2l& b) {
252 return _mm256_set_m128d((pcast<Packet2l, Packet2d>(b)), (pcast<Packet2l, Packet2d>(a)));
253}
254
255template <>
256EIGEN_STRONG_INLINE Packet4ul preinterpret<Packet4ul, Packet4l>(const Packet4l& a) {
257 return Packet4ul(a);
258}
259
260template <>
261EIGEN_STRONG_INLINE Packet4l preinterpret<Packet4l, Packet4ul>(const Packet4ul& a) {
262 return Packet4l(a);
263}
264
265template <>
266EIGEN_STRONG_INLINE Packet4l preinterpret<Packet4l, Packet4d>(const Packet4d& a) {
267 return _mm256_castpd_si256(a);
268}
269
270template <>
271EIGEN_STRONG_INLINE Packet4d preinterpret<Packet4d, Packet4l>(const Packet4l& a) {
272 return _mm256_castsi256_pd(a);
273}
274
275// truncation operations
276template <>
277EIGEN_STRONG_INLINE Packet2l preinterpret<Packet2l, Packet4l>(const Packet4l& a) {
278 return _mm256_castsi256_si128(a);
279}
280#endif
281
282template <>
283EIGEN_STRONG_INLINE Packet8f pcast<Packet8h, Packet8f>(const Packet8h& a) {
284 return half2float(a);
285}
286
287template <>
288EIGEN_STRONG_INLINE Packet8f pcast<Packet8bf, Packet8f>(const Packet8bf& a) {
289 return Bf16ToF32(a);
290}
291
292template <>
293EIGEN_STRONG_INLINE Packet8h pcast<Packet8f, Packet8h>(const Packet8f& a) {
294 return float2half(a);
295}
296
297template <>
298EIGEN_STRONG_INLINE Packet8bf pcast<Packet8f, Packet8bf>(const Packet8f& a) {
299 return F32ToBf16(a);
300}
301
302} // end namespace internal
303
304} // end namespace Eigen
305
306#endif // EIGEN_TYPE_CASTING_AVX_H
Eigen
Namespace containing all symbols from the Eigen library.
Definition Core:137