1 | /* Copyright (C) 2017 Povilas Kanapickas <povilas@radix.lt> |
2 | |
3 | Distributed under the Boost Software License, Version 1.0. |
4 | (See accompanying file LICENSE_1_0.txt or copy at |
5 | http://www.boost.org/LICENSE_1_0.txt) |
6 | */ |
7 | |
8 | #ifndef LIBSIMDPP_DETAIL_FOR_EACH_H |
9 | #define LIBSIMDPP_DETAIL_FOR_EACH_H |
10 | |
11 | #ifndef LIBSIMDPP_SIMD_H |
12 | #error "This file must be included through simd.h" |
13 | #endif |
14 | |
15 | #include <simdpp/types.h> |
16 | #include <simdpp/core/extract.h> |
17 | #include <simdpp/detail/mem_block.h> |
18 | #include <type_traits> |
19 | |
20 | namespace simdpp { |
21 | namespace SIMDPP_ARCH_NAMESPACE { |
22 | namespace detail { |
23 | |
24 | #if SIMDPP_USE_SSE2 || SIMDPP_USE_NEON || SIMDPP_USE_MSA |
25 | template<class V, class F> SIMDPP_INL |
26 | void foreach_impl(std::integral_constant<unsigned, 2>, |
27 | const V& v, F function) |
28 | { |
29 | function(extract<0>(v)); |
30 | function(extract<1>(v)); |
31 | } |
32 | |
33 | template<class V, class F> SIMDPP_INL |
34 | void foreach_impl(std::integral_constant<unsigned, 4>, |
35 | const V& v, F function) |
36 | { |
37 | function(extract<0>(v)); |
38 | function(extract<1>(v)); |
39 | function(extract<2>(v)); |
40 | function(extract<3>(v)); |
41 | } |
42 | |
43 | template<unsigned N, class V, class F> SIMDPP_INL |
44 | void foreach_impl(std::integral_constant<unsigned, N>, |
45 | const V& v, F function) |
46 | { |
47 | // When we're operating on more than 4-5 elements it makes sense to move |
48 | // the vector to memory and load data from there. This has higher latency, |
49 | // but this is masked by extracting the first several elements directly |
50 | // from the SIMD register set. For the rest of elements it's very likely |
51 | // that loading through memory has higher throughput. |
52 | // |
53 | // Recent x86 (since Sandy Bridge) and NEON (since Cortex A73) processors |
54 | // are able to sustain more than one load memory access per cycle. |
55 | // All x86 processors (at least up to Skylake, newer not checked) are only |
56 | // able to sustain single cross domain data access instruction per cycle. |
57 | |
58 | // TODO: needs tuning on ARM and MIPS |
59 | function(extract<0>(v)); |
60 | function(extract<1>(v)); |
61 | mem_block<V> mem(v); |
62 | for (unsigned i = 2; i < N; ++i) |
63 | function(mem[i]); |
64 | } |
65 | #else |
66 | template<unsigned N, class V, class F> SIMDPP_INL |
67 | void foreach_impl(std::integral_constant<unsigned, N>, |
68 | const V& v, F function) |
69 | { |
70 | mem_block<V> mem(v); |
71 | for (unsigned i = 0; i < N; ++i) |
72 | function(mem[i]); |
73 | } |
74 | #endif |
75 | |
76 | template<unsigned N, class V, class F> SIMDPP_INL |
77 | void for_each(const any_vec<N, V>& v, F function) |
78 | { |
79 | using size_tag = std::integral_constant<unsigned, V::base_vector_type::length>; |
80 | for (unsigned i = 0; i < V::vec_length; ++i) |
81 | foreach_impl(size_tag(), v.wrapped().vec(i), function); |
82 | } |
83 | |
84 | |
85 | } // namespace detail |
86 | } // namespace SIMDPP_ARCH_NAMESPACE |
87 | } // namespace simdpp |
88 | |
89 | #endif |
90 | |