1+ #include < cstdint>
2+ #include < limits>
3+ #include < type_traits>
4+ #include < cstddef>
5+ #include < algorithm>
6+
7+ #include < arcticdb/util/vector_common.hpp>
8+
9+ namespace arcticdb {
10+
11+ template <typename T>
12+ using vector_type __attribute__ ((vector_size(64 ))) = T;
13+
14+ template <typename T>
15+ class FloatMinFinder {
16+ static_assert (is_supported_float<T>::value, " Type must be float or double"
17+ static_assert (std::is_floating_point_v<T>, " Type must be floating point"
18+
19+ public:
20+ static T find (const T* data, size_t n) {
21+ using vec_t = vector_type<T>;
22+
23+ // Initialize min vector with infinity
24+ vec_t vmin;
25+ for (size_t i = 0 ; i < sizeof (vec_t )/sizeof (T); i++) {
26+ reinterpret_cast <T*>(&vmin)[i] = std::numeric_limits<T>::infinity ();
27+ }
28+
29+ // Process full vectors
30+ const vec_t * vdata = reinterpret_cast <const vec_t *>(data);
31+ const size_t elements_per_vector = sizeof (vec_t ) / sizeof (T);
32+ const size_t vlen = n / elements_per_vector;
33+
34+ // Main SIMD loop
35+ for (size_t i = 0 ; i < vlen; i++) {
36+ vec_t v = vdata[i];
37+ vmin = (v < vmin) ? v : vmin;
38+ }
39+
40+ // Reduce vector to scalar
41+ T min_val = std::numeric_limits<T>::infinity ();
42+ const T* min_arr = reinterpret_cast <const T*>(&vmin);
43+ for (size_t i = 0 ; i < elements_per_vector; i++) {
44+ if (min_arr[i] == min_arr[i]) { // Not NaN
45+ min_val = std::min (min_val, min_arr[i]);
46+ }
47+ }
48+
49+ // Handle remainder
50+ const T* remain = data + (vlen * elements_per_vector);
51+ for (size_t i = 0 ; i < n % elements_per_vector; i++) {
52+ if (remain[i] == remain[i]) { // Not NaN
53+ min_val = std::min (min_val, remain[i]);
54+ }
55+ }
56+
57+ return min_val;
58+ }
59+ };
60+
61+ template <typename T>
62+ class FloatMaxFinder {
63+ static_assert (is_supported_float<T>::value, " Type must be float or double"
64+ static_assert (std::is_floating_point_v<T>, " Type must be floating point"
65+
66+ public:
67+ static T find (const T* data, size_t n) {
68+ using vec_t = vector_type<T>;
69+
70+ // Initialize max vector with negative infinity
71+ vec_t vmax;
72+ for (size_t i = 0 ; i < sizeof (vec_t )/sizeof (T); i++) {
73+ reinterpret_cast <T*>(&vmax)[i] = -std::numeric_limits<T>::infinity ();
74+ }
75+
76+ // Process full vectors
77+ const vec_t * vdata = reinterpret_cast <const vec_t *>(data);
78+ const size_t elements_per_vector = sizeof (vec_t ) / sizeof (T);
79+ const size_t vlen = n / elements_per_vector;
80+
81+ // Main SIMD loop
82+ for (size_t i = 0 ; i < vlen; i++) {
83+ vec_t v = vdata[i];
84+ vmax = (v > vmax) ? v : vmax;
85+ }
86+
87+ // Reduce vector to scalar
88+ T max_val = -std::numeric_limits<T>::infinity ();
89+ const T* max_arr = reinterpret_cast <const T*>(&vmax);
90+ for (size_t i = 0 ; i < elements_per_vector; i++) {
91+ if (max_arr[i] == max_arr[i]) { // Not NaN
92+ max_val = std::max (max_val, max_arr[i]);
93+ }
94+ }
95+
96+ // Handle remainder
97+ const T* remain = data + (vlen * elements_per_vector);
98+ for (size_t i = 0 ; i < n % elements_per_vector; i++) {
99+ if (remain[i] == remain[i]) { // Not NaN
100+ max_val = std::max (max_val, remain[i]);
101+ }
102+ }
103+
104+ return max_val;
105+ }
106+ };
107+
108+ template <typename T>
109+ T find_float_min (const T *data, size_t n) {
110+ return FloatMinFinder<T>::find (data, n);
111+ }
112+
113+ template <typename T>
114+ T find_float_max (const T *data, size_t n) {
115+ return FloatMaxFinder<T>::find (data, n);
116+ }
117+
118+ } // namespace arcticdb
0 commit comments