[flang-commits] [flang] [flang][runtime] Fixed performance regression in CopyElement. (PR #102081)
Slava Zakharin via flang-commits
flang-commits at lists.llvm.org
Mon Aug 5 16:46:49 PDT 2024
https://github.com/vzakhari created https://github.com/llvm/llvm-project/pull/102081
Polyhedron/capacita,protein and CPU2000/facerec,wupwise showed up to
60% regression on x86 after #101421. The memcpy loops of the toAt and
fromAt arrays that are run to create the initial work item end up
being encoded as 'rep mov', and they add noticeable overhead
comparing to the total amount of work. 'rep mov' is not the best
choise for small size memcpy (e.g. when the array rank is 1 or 2,
it would be quite slow). Moreover, the rest of the stack related
setup is also noticeable for the simple cases.
I added a shortcut for the simple copy case, and also got rid
of the initial toAt/fromAt copies by allowing the CopyDescriptor
to use the external subscript storages.
>From 54077420b941549b944f0f9282985e6686e6343a Mon Sep 17 00:00:00 2001
From: Slava Zakharin <szakharin at nvidia.com>
Date: Mon, 5 Aug 2024 14:23:23 -0700
Subject: [PATCH] [flang][runtime] Fixed performance regression in CopyElement.
Polyhedron/capacita,protein and CPU2000/facerec,wupwise showed up to
60% regression on x86 after #101421. The memcpy loops of the toAt and
fromAt arrays that are run to create the initial work item end up
being encoded as 'rep mov', and they add noticeable overhead
comparing to the total amount of work. 'rep mov' is not the best
choise for small size memcpy (e.g. when the array rank is 1 or 2,
it would be quite slow). Moreover, the rest of the stack related
setup is also noticeable for the simple cases.
I added a shortcut for the simple copy case, and also got rid
of the initial toAt/fromAt copies by allowing the CopyDescriptor
to use the external subscript storages.
---
flang/runtime/copy.cpp | 49 +++++++++++++++++++++++++++++++++---------
1 file changed, 39 insertions(+), 10 deletions(-)
diff --git a/flang/runtime/copy.cpp b/flang/runtime/copy.cpp
index c2dbbc4a11c06..21525c30b9540 100644
--- a/flang/runtime/copy.cpp
+++ b/flang/runtime/copy.cpp
@@ -23,17 +23,17 @@ using StaticDescTy = StaticDescriptor<maxRank, true, 0>;
// for CopyElement.
struct CopyDescriptor {
// A constructor specifying all members explicitly.
+ // The toAt and fromAt specify subscript storages that might be
+ // external to CopyElement, and cannot be modified.
+ // The copy descriptor only establishes toAtPtr_ and fromAtPtr_
+ // pointers to point to these storages.
RT_API_ATTRS CopyDescriptor(const Descriptor &to, const SubscriptValue toAt[],
const Descriptor &from, const SubscriptValue fromAt[],
std::size_t elements, bool usesStaticDescriptors = false)
: to_(to), from_(from), elements_(elements),
usesStaticDescriptors_(usesStaticDescriptors) {
- for (int dim{0}; dim < to.rank(); ++dim) {
- toAt_[dim] = toAt[dim];
- }
- for (int dim{0}; dim < from.rank(); ++dim) {
- fromAt_[dim] = fromAt[dim];
- }
+ toAtPtr_ = toAt;
+ fromAtPtr_ = fromAt;
}
// The number of elements to copy is initialized from the to descriptor.
// The current element subscripts are initialized from the lower bounds
@@ -46,14 +46,32 @@ struct CopyDescriptor {
from.GetLowerBounds(fromAt_);
}
+ // Increment the toAt_ and fromAt_ subscripts to the next
+ // element.
+ RT_API_ATTRS void incrementSubscripts(Terminator &terminator) {
+ // This method must not be called for copy descriptors
+ // using external non-modifiable subscript storages.
+ RUNTIME_CHECK(terminator, toAt_ == toAtPtr_ && fromAt_ == fromAtPtr_);
+ to_.IncrementSubscripts(toAt_);
+ from_.IncrementSubscripts(fromAt_);
+ }
+
// Descriptor of the destination.
const Descriptor &to_;
// A subscript specifying the current element position to copy to.
SubscriptValue toAt_[maxRank];
+ // A pointer to the storage of the 'to' subscript.
+ // It may point to toAt_ or to an external non-modifiable
+ // subscript storage.
+ const SubscriptValue *toAtPtr_{toAt_};
// Descriptor of the source.
const Descriptor &from_;
// A subscript specifying the current element position to copy from.
SubscriptValue fromAt_[maxRank];
+ // A pointer to the storage of the 'from' subscript.
+ // It may point to fromAt_ or to an external non-modifiable
+ // subscript storage.
+ const SubscriptValue *fromAtPtr_{fromAt_};
// Number of elements left to copy.
std::size_t elements_;
// Must be true, if the to and from descriptors are allocated
@@ -75,6 +93,18 @@ RT_OFFLOAD_API_GROUP_BEGIN
RT_API_ATTRS void CopyElement(const Descriptor &to, const SubscriptValue toAt[],
const Descriptor &from, const SubscriptValue fromAt[],
Terminator &terminator) {
+ if (!to.Addendum()) {
+ // Avoid the overhead of creating the work stacks below
+ // for the simple non-derived type cases, because the overhead
+ // might be noticeable over the total amount of work that
+ // needs to be done for the copy.
+ char *toPtr{to.Element<char>(toAt)};
+ char *fromPtr{from.Element<char>(fromAt)};
+ RUNTIME_CHECK(terminator, to.ElementBytes() == from.ElementBytes());
+ std::memcpy(toPtr, fromPtr, to.ElementBytes());
+ return;
+ }
+
#if !defined(RT_DEVICE_COMPILATION)
constexpr unsigned copyStackReserve{16};
constexpr unsigned descriptorStackReserve{6};
@@ -108,9 +138,9 @@ RT_API_ATTRS void CopyElement(const Descriptor &to, const SubscriptValue toAt[],
continue;
}
const Descriptor &curTo{currentCopy.to_};
- SubscriptValue *curToAt{currentCopy.toAt_};
+ const SubscriptValue *curToAt{currentCopy.toAtPtr_};
const Descriptor &curFrom{currentCopy.from_};
- SubscriptValue *curFromAt{currentCopy.fromAt_};
+ const SubscriptValue *curFromAt{currentCopy.fromAtPtr_};
char *toPtr{curTo.Element<char>(curToAt)};
char *fromPtr{curFrom.Element<char>(curFromAt)};
RUNTIME_CHECK(terminator, curTo.ElementBytes() == curFrom.ElementBytes());
@@ -121,8 +151,7 @@ RT_API_ATTRS void CopyElement(const Descriptor &to, const SubscriptValue toAt[],
std::memcpy(toPtr, fromPtr, curTo.ElementBytes());
--elements;
if (elements != 0) {
- curTo.IncrementSubscripts(curToAt);
- curFrom.IncrementSubscripts(curFromAt);
+ currentCopy.incrementSubscripts(terminator);
}
// Deep copy allocatable and automatic components if any.
More information about the flang-commits
mailing list