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

releaase 2.11 (#703)

Browse Source
This commit is contained in:
Aditya Atluri
2022-11-19 06:02:15 -08:00
committed by GitHub
parent 3c90f6aea6
commit c975e2ccbb
329 changed files with 47332 additions and 10607 deletions
Download Patch File Download Diff File Expand all files Collapse all files

21
include/cutlass/transform/pitch_linear_thread_map.h
View File

@@ -88,20 +88,15 @@ struct PitchLinearStripminedThreadMap {
static_assert(!(Shape::kContiguous % kElementsPerAccess), "");
static_assert(!((Shape::kContiguous * Shape::kStrided) % (kThreads * kElementsPerAccess)),
"Shape must be divisible thread count.");
/// Shape of the tile in units of vectors
using ShapeVec = layout::PitchLinearShape<
Shape::kContiguous / kElementsPerAccess,
Shape::kStrided
>;
static_assert(
(Threads < ShapeVec::kContiguous && !(ShapeVec::kContiguous % kThreads)) ||
(!(kThreads % ShapeVec::kContiguous) && !(ShapeVec::kStrided % (kThreads / ShapeVec::kContiguous))),
"Shape must be divisible by number of iterations of each thread."
);
static_assert((Threads < ShapeVec::kContiguous && !(ShapeVec::kContiguous % kThreads)) ||
(!(kThreads % ShapeVec::kContiguous)),
"Shape must be divisible by number of iterations of each thread.");
};
/// Number of iterations by each thread
@@ -112,11 +107,12 @@ struct PitchLinearStripminedThreadMap {
// Redo the comparison here to work around divide by zero compiler
// error. The compiler evaluates both path of platform::conditional.
(Threads >= Detail::ShapeVec::kContiguous
? Detail::ShapeVec::kStrided /
? (Detail::ShapeVec::kStrided + (kThreads / Detail::ShapeVec::kContiguous - 1)) /
(kThreads / Detail::ShapeVec::kContiguous)
: 0)>,
layout::PitchLinearShape<Detail::ShapeVec::kContiguous / kThreads,
Detail::ShapeVec::kStrided>>::type;
/// Interval between accesses along each dimension of the tensor's logical coordinate space
/// (in units of Elements)
@@ -132,6 +128,13 @@ struct PitchLinearStripminedThreadMap {
>
>::type;
/// Shape of the tile in units of vectors
using StorageShape = typename platform::conditional<
Threads >= Detail::ShapeVec::kContiguous,
layout::PitchLinearShape<Shape::kContiguous,
Iterations::kStrided*(kThreads / Detail::ShapeVec::kContiguous)>,
layout::PitchLinearShape<Shape::kContiguous, Shape::kStrided>>::type;
/// Maps thread ID to a coordinate offset within the tensor's logical coordinate space
/// (in units of Elements)
CUTLASS_HOST_DEVICE