rooty/dxvk
1
0
Fork 0
mirror of https://github.com/doitsujin/dxvk.git synced 2025-04-05 16:40:17 +02:00
Code Issues Projects Releases Wiki Activity

[dxbc] Optimize write pattern for TGSM initialization

Browse Source 
Reduces bank conflicts.
This commit is contained in:
Philip Rebohle 2022-09-12 22:33:40 +02:00
parent 05f36fd97b
commit af208ceb9e
1 changed files with 7 additions and 8 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

15
src/dxbc/dxbc_compiler.cpp
View File

@ -6012,9 +6012,7 @@ namespace dxvk {
bool hasTgsm = false; bool hasTgsm = false;
SpirvMemoryOperands memoryOperands; SpirvMemoryOperands memoryOperands;
memoryOperands.flags = spv::MemoryAccessNonPrivatePointerMask memoryOperands.flags = spv::MemoryAccessNonPrivatePointerMask;
| spv::MemoryAccessMakePointerAvailableMask;
memoryOperands.makeAvailable = m_module.constu32(spv::ScopeWorkgroup);
for (uint32_t i = 0; i < m_gRegs.size(); i++) { for (uint32_t i = 0; i < m_gRegs.size(); i++) {
if (!m_gRegs[i].varId) if (!m_gRegs[i].varId)
@ -6045,13 +6043,12 @@ namespace dxvk {
uint32_t threadId = m_module.opLoad( uint32_t threadId = m_module.opLoad(
intTypeId, m_cs.builtinLocalInvocationIndex); intTypeId, m_cs.builtinLocalInvocationIndex);
uint32_t strideId = m_module.constu32(numElementsPerThread); uint32_t strideId = m_module.constu32(numThreads);
uint32_t zeroId = m_module.constu32(0); uint32_t zeroId = m_module.constu32(0);
for (uint32_t e = 0; e < numElementsPerThread; e++) { for (uint32_t e = 0; e < numElementsPerThread; e++) {
uint32_t ofsId = m_module.opIAdd(intTypeId, uint32_t ofsId = m_module.opIAdd(intTypeId, threadId,
m_module.opIMul(intTypeId, strideId, threadId), m_module.opIMul(intTypeId, strideId, m_module.constu32(e)));
m_module.constu32(e));
uint32_t ptrId = m_module.opAccessChain( uint32_t ptrId = m_module.opAccessChain(
ptrTypeId, m_gRegs[i].varId, 1, &ofsId); ptrTypeId, m_gRegs[i].varId, 1, &ofsId);
@ -6094,7 +6091,9 @@ namespace dxvk {
m_module.constu32(spv::ScopeWorkgroup), m_module.constu32(spv::ScopeWorkgroup),
m_module.constu32(spv::ScopeWorkgroup), m_module.constu32(spv::ScopeWorkgroup),
m_module.constu32(spv::MemorySemanticsWorkgroupMemoryMask m_module.constu32(spv::MemorySemanticsWorkgroupMemoryMask
| spv::MemorySemanticsAcquireReleaseMask)); | spv::MemorySemanticsAcquireReleaseMask
| spv::MemorySemanticsMakeAvailableMask
| spv::MemorySemanticsMakeVisibleMask));
} }
} }