rooty/dxvk
1
0
Fork 0
mirror of https://github.com/doitsujin/dxvk.git synced 2025-02-27 04:54:15 +01:00
Code Issues Projects Releases Wiki Activity

[dxvk] Track order-invariant access ops in barrier tracker

Browse Source 
This elides barriers between draws or dispatches if we can prove
order-invariance through atomic operations.
This commit is contained in:
Philip Rebohle 2025-02-15 22:16:03 +01:00 committed by Philip Rebohle
parent c475960754
commit dd1ca4ce59
4 changed files with 67 additions and 21 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

43
src/dxvk/dxvk_barrier.cpp
View File

@ -20,20 +20,37 @@ namespace dxvk {
bool DxvkBarrierTracker::findRange(
const DxvkAddressRange& range,
DxvkAccess accessType) const {
DxvkAccess accessType,
DxvkAccessOp accessOp) const {
uint32_t rootIndex = computeRootIndex(range, accessType);
return findNode(range, rootIndex);
uint32_t nodeIndex = findNode(range, rootIndex);
if (likely(!nodeIndex || accessOp == DxvkAccessOp::None))
return nodeIndex;
// If we are checking for a specific order-invariant store
// op, the op must have been the only op used to access the
// resource, and the tracked range must cover the requested
// range in its entirety so we can rule out that other parts
// of the resource have been accessed in a different way.
auto& node = m_nodes[nodeIndex];
return node.payload.accessOps != DxvkAccessOps(accessOp)
|| !node.addressRange.contains(range);
}
void DxvkBarrierTracker::insertRange(
const DxvkAddressRange& range,
DxvkAccess accessType) {
uint32_t rootIndex = computeRootIndex(range, accessType);
DxvkAccess accessType,
DxvkAccessOp accessOp) {
DxvkBarrierPayload payload = { };
payload.accessOps.set(accessOp);
// If we can just insert the node with no conflicts,
// we don't have to do anything.
uint32_t nodeIndex = insertNode(range, rootIndex);
uint32_t rootIndex = computeRootIndex(range, accessType);
uint32_t nodeIndex = insertNode(range, rootIndex, payload);
if (likely(!nodeIndex))
return;
@ -41,6 +58,7 @@ namespace dxvk {
// If there's an existing node and it contains the entire
// range we want to add already, also don't do anything.
auto& node = m_nodes[nodeIndex];
node.payload.accessOps.set(payload.accessOps);
if (node.addressRange.contains(range))
return;
@ -82,12 +100,14 @@ namespace dxvk {
mergedRange.rangeStart = std::min(mergedRange.rangeStart, node.addressRange.rangeStart);
mergedRange.rangeEnd = std::max(mergedRange.rangeEnd, node.addressRange.rangeEnd);
payload.accessOps.set(node.payload.accessOps);
removeNode(nodeIndex, rootIndex);
nodeIndex = findNode(range, rootIndex);
}
insertNode(mergedRange, rootIndex);
insertNode(mergedRange, rootIndex, payload);
}
@ -166,7 +186,8 @@ namespace dxvk {
uint32_t DxvkBarrierTracker::insertNode(
const DxvkAddressRange& range,
uint32_t rootIndex) {
uint32_t rootIndex,
DxvkBarrierPayload payload) {
// Check if the given root is valid at all
uint64_t rootBit = uint64_t(1u) << (rootIndex - 1u);
@ -178,6 +199,7 @@ namespace dxvk {
auto& node = m_nodes[rootIndex];
node.header = 0;
node.addressRange = range;
node.payload = payload;
return 0;
} else {
// Traverse tree and abort if we find any range
@ -209,6 +231,7 @@ namespace dxvk {
node.setRed(true);
node.setParent(parentIndex);
node.addressRange = range;
node.payload = payload;
// Only do the fixup to maintain red-black properties if
// we haven't marked the root node as red in a deletion.
@ -238,6 +261,7 @@ namespace dxvk {
childIndex = m_nodes[childIndex].child(0);
node.addressRange = m_nodes[childIndex].addressRange;
node.payload = m_nodes[childIndex].payload;
removeNode(childIndex, rootIndex);
} else {
// Deletion is expected to be exceptionally rare, to the point of
@ -268,6 +292,7 @@ namespace dxvk {
node.setRed(child.isRed());
node.addressRange = child.addressRange;
node.payload = child.payload;
if (cl) m_nodes[cl].setParent(nodeIndex);
if (cr) m_nodes[cr].setParent(nodeIndex);
@ -378,6 +403,7 @@ namespace dxvk {
node.setChild(1, rr);
std::swap(node.addressRange, m_nodes[r].addressRange);
std::swap(node.payload, m_nodes[r].payload);
}
@ -406,6 +432,7 @@ namespace dxvk {
node.setChild(1, l);
std::swap(node.addressRange, m_nodes[l].addressRange);
std::swap(node.payload, m_nodes[l].payload);
}
@ -498,4 +525,4 @@ namespace dxvk {
flush(list);
}
}
}

24
src/dxvk/dxvk_barrier.h
View File

@ -42,6 +42,14 @@ namespace dxvk {
};
/**
* \brief Barrier node payload
*/
struct DxvkBarrierPayload {
DxvkAccessOps accessOps = 0u;
};
/**
* \brief Barrier tree node
*
@ -62,6 +70,9 @@ namespace dxvk {
// Address range of the node
DxvkAddressRange addressRange = { };
// Node payload
DxvkBarrierPayload payload = { };
void setRed(bool red) {
header &= ~uint64_t(1u);
header |= uint64_t(red);
@ -117,21 +128,25 @@ namespace dxvk {
*
* \param [in] range Resource range
* \param [in] accessType Access type
* \param [in] accessOp Access operation
* \returns \c true if the range has a pending access
*/
bool findRange(
const DxvkAddressRange& range,
DxvkAccess accessType) const;
DxvkAccess accessType,
DxvkAccessOp accessOp) const;
/**
* \brief Inserts address range for a given access type
*
* \param [in] range Resource range
* \param [in] accessType Access type
* \param [in] accessOp Access operation
*/
void insertRange(
const DxvkAddressRange& range,
DxvkAccess accessType);
DxvkAccess accessType,
DxvkAccessOp accessOp);
/**
* \brief Clears the entire structure
@ -166,7 +181,8 @@ namespace dxvk {
uint32_t insertNode(
const DxvkAddressRange& range,
uint32_t rootIndex);
uint32_t rootIndex,
DxvkBarrierPayload payload);
void removeNode(
uint32_t nodeIndex,
@ -285,4 +301,4 @@ namespace dxvk {
};
}
}

18
src/dxvk/dxvk_context.cpp
View File

@ -7796,9 +7796,9 @@ namespace dxvk {
+ (subresources.baseArrayLayer + subresources.layerCount - 1u);
if (hasWrite)
m_barrierTracker.insertRange(range, DxvkAccess::Write);
m_barrierTracker.insertRange(range, DxvkAccess::Write, accessOp);
if (hasRead)
m_barrierTracker.insertRange(range, DxvkAccess::Read);
m_barrierTracker.insertRange(range, DxvkAccess::Read, accessOp);
} else {
DxvkAddressRange range;
range.resource = image.getResourceId();
@ -7808,9 +7808,9 @@ namespace dxvk {
range.rangeEnd = range.rangeStart + subresources.layerCount - 1u;
if (hasWrite)
m_barrierTracker.insertRange(range, DxvkAccess::Write);
m_barrierTracker.insertRange(range, DxvkAccess::Write, accessOp);
if (hasRead)
m_barrierTracker.insertRange(range, DxvkAccess::Read);
m_barrierTracker.insertRange(range, DxvkAccess::Read, accessOp);
}
}
}
@ -7863,9 +7863,9 @@ namespace dxvk {
range.rangeEnd = offset + size - 1;
if (srcAccess & vk::AccessWriteMask)
m_barrierTracker.insertRange(range, DxvkAccess::Write);
m_barrierTracker.insertRange(range, DxvkAccess::Write, accessOp);
if (srcAccess & vk::AccessReadMask)
m_barrierTracker.insertRange(range, DxvkAccess::Read);
m_barrierTracker.insertRange(range, DxvkAccess::Read, accessOp);
}
}
@ -8037,7 +8037,7 @@ namespace dxvk {
range.rangeStart = offset;
range.rangeEnd = offset + size - 1;
return m_barrierTracker.findRange(range, access);
return m_barrierTracker.findRange(range, access, accessOp);
}
@ -8071,7 +8071,7 @@ namespace dxvk {
// Probe all subresources first, only check individual mip levels
// if there are overlaps and if we are checking a subset of array
// layers of multiple mips.
bool dirty = m_barrierTracker.findRange(range, access);
bool dirty = m_barrierTracker.findRange(range, access, accessOp);
if (!dirty || subresources.levelCount == 1u || subresources.layerCount == layerCount)
return dirty;
@ -8080,7 +8080,7 @@ namespace dxvk {
range.rangeStart = i * layerCount + subresources.baseArrayLayer;
range.rangeEnd = range.rangeStart + subresources.layerCount - 1u;
dirty = m_barrierTracker.findRange(range, access);
dirty = m_barrierTracker.findRange(range, access, accessOp);
}
return dirty;

3
src/dxvk/dxvk_pipelayout.h
View File

@ -29,6 +29,9 @@ namespace dxvk {
UMax = 8,
};
using DxvkAccessOps = Flags<DxvkAccessOp>;
/**
* \brief Descriptor set indices
*/