leavelet/cutlass
1
0
Fork 0
Code Issues Pull Requests Actions 27 Packages Projects Releases Wiki Activity

Updates for CUTLASS 3.5.0 (#1468)

Browse Source
This commit is contained in:
Vijay Thakkar
2024-04-11 21:33:40 -04:00
committed by GitHub
parent a40e08e9d5
commit 7d49e6c7e2
171 changed files with 7526 additions and 1888 deletions
Download Patch File Download Diff File Expand all files Collapse all files
include/cute/int_tuple.hpp
+15 -4
View File
@@ -325,10 +325,21 @@ CUTE_HOST_DEVICE constexpr
auto
ceil_div(IntTupleA const& a, IntTupleB const& b)
{
if constexpr (is_tuple<IntTupleA>::value && is_tuple<IntTupleB>::value) {
static_assert(tuple_size<IntTupleA>::value >= tuple_size<IntTupleB>::value, "Mismatched ranks");
constexpr int R = tuple_size<IntTupleA>::value; // Missing ranks in TupleB are implicitly 1
return transform(a, append<R>(b,Int<1>{}), [](auto const& x, auto const& y) { return ceil_div(x,y); });
if constexpr (is_tuple<IntTupleA>::value) {
if constexpr (is_tuple<IntTupleB>::value) { // tuple tuple
static_assert(tuple_size<IntTupleA>::value >= tuple_size<IntTupleB>::value, "Mismatched ranks");
constexpr int R = tuple_size<IntTupleA>::value; // Missing ranks in TupleB are implicitly 1
return transform(a, append<R>(b,Int<1>{}), [](auto const& x, auto const& y) { return ceil_div(x,y); });
} else { // tuple int
auto const [result, rest] = fold(a, cute::make_tuple(cute::make_tuple(), b),
[] (auto const& init, auto const& ai) {
return cute::make_tuple(append(get<0>(init), ceil_div(ai, get<1>(init))), ceil_div(get<1>(init), ai));
});
return result;
}
} else
if constexpr (is_tuple<IntTupleB>::value) { // int tuple
return ceil_div(a, product(b));
} else {
return (a + b - Int<1>{}) / b;
}