rooty/dxvk
1
0
Fork 0
mirror of https://github.com/doitsujin/dxvk.git synced 2024-12-02 10:24:12 +01:00
Code Issues Projects Releases Wiki Activity

[dxbc] Only use atomic append/consume optimization in compute shaders

Browse Source 
- For fragment shaders, this isn't safe since ballots include helper invocations
- For vertex shaders, if drivers don't support subgroup operations in those
  stages, we don't want it to affect the performance of compute shadres.
This commit is contained in:
Philip Rebohle 2019-07-26 14:07:39 +02:00
parent dfe2922136
commit 6ab074c95b
No known key found for this signature in database
GPG Key ID: C8CC613427A31C99
2 changed files with 8 additions and 5 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

9
src/dxbc/dxbc_compiler.cpp
View File

@ -2411,6 +2411,11 @@ namespace dxvk {
m_module.opLabel(cond.labelIf);
// Only use subgroup ops on compute to avoid having to
// deal with helper invocations or hardware limitations
bool useSubgroupOps = m_moduleInfo.options.useSubgroupOpsForAtomicCounters
&& m_programInfo.type() == DxbcProgramType::ComputeShader;
// In case we have subgroup ops enabled, we need to
// count the number of active lanes, the lane index,
// and we need to perform the atomic op conditionally
@ -2419,7 +2424,7 @@ namespace dxvk {
DxbcConditional elect;
if (m_moduleInfo.options.useSubgroupOpsForAtomicCounters) {
if (useSubgroupOps) {
m_module.enableCapability(spv::CapabilityGroupNonUniform);
m_module.enableCapability(spv::CapabilityGroupNonUniformBallot);
@ -2504,7 +2509,7 @@ namespace dxvk {
// If we're using subgroup ops, we have to broadcast
// the result of the atomic op and compute the index
if (m_moduleInfo.options.useSubgroupOpsForAtomicCounters) {
if (useSubgroupOps) {
m_module.opBranch(elect.labelEnd);
m_module.opLabel (elect.labelEnd);

4
src/dxbc/dxbc_options.cpp
View File

@ -15,14 +15,12 @@ namespace dxvk {
const DxvkDeviceFeatures& devFeatures = device->features();
const DxvkDeviceInfo& devInfo = adapter->devicePropertiesExt();
const VkShaderStageFlags allShaderStages = device->getShaderPipelineStages();
useDepthClipWorkaround
= !devFeatures.extDepthClipEnable.depthClipEnable;
useStorageImageReadWithoutFormat
= devFeatures.core.features.shaderStorageImageReadWithoutFormat;
useSubgroupOpsForAtomicCounters
= (devInfo.coreSubgroup.supportedStages & allShaderStages) == allShaderStages
= (devInfo.coreSubgroup.supportedStages & VK_SHADER_STAGE_COMPUTE_BIT)
&& (devInfo.coreSubgroup.supportedOperations & VK_SUBGROUP_FEATURE_BALLOT_BIT);
useDemoteToHelperInvocation
= (devFeatures.extShaderDemoteToHelperInvocation.shaderDemoteToHelperInvocation);