GoaLitiuM/FlaxEngine
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Add explicit GPU resource transitions, memory and compute barriers

Browse Source
This commit is contained in:
Wojtek Figat
2025-08-09 23:57:43 +02:00
parent 3907bc4957
commit b5a431d2f5
12 changed files with 353 additions and 48 deletions
Download Patch File Download Diff File Expand all files Collapse all files

109
Source/Engine/GraphicsDevice/Vulkan/GPUContextVulkan.cpp
View File

@@ -78,13 +78,14 @@ const Char* ToString(VkImageLayout layout)
void PipelineBarrierVulkan::Execute(const CmdBufferVulkan* cmdBuffer)
{
ASSERT(cmdBuffer->IsOutsideRenderPass());
vkCmdPipelineBarrier(cmdBuffer->GetHandle(), SourceStage, DestStage, 0, 0, nullptr, BufferBarriers.Count(), BufferBarriers.Get(), ImageBarriers.Count(), ImageBarriers.Get());
vkCmdPipelineBarrier(cmdBuffer->GetHandle(), SourceStage, DestStage, 0, MemoryBarriers.Count(), MemoryBarriers.Get(), BufferBarriers.Count(), BufferBarriers.Get(), ImageBarriers.Count(), ImageBarriers.Get());
// Reset
SourceStage = 0;
DestStage = 0;
ImageBarriers.Clear();
BufferBarriers.Clear();
MemoryBarriers.Clear();
#if VK_ENABLE_BARRIERS_DEBUG
ImageBarriersDebug.Clear();
#endif
@@ -153,12 +154,7 @@ void GPUContextVulkan::AddImageBarrier(VkImage image, VkImageLayout srcLayout, V
#if VK_ENABLE_BARRIERS_BATCHING
// Auto-flush on overflow
if (_barriers.IsFull())
{
const auto cmdBuffer = _cmdBufferManager->GetCmdBuffer();
if (cmdBuffer->IsInsideRenderPass())
EndRenderPass();
_barriers.Execute(cmdBuffer);
}
FlushBarriers();
#endif
// Insert barrier
@@ -190,10 +186,7 @@ void GPUContextVulkan::AddImageBarrier(VkImage image, VkImageLayout srcLayout, V
#if !VK_ENABLE_BARRIERS_BATCHING
// Auto-flush without batching
const auto cmdBuffer = _cmdBufferManager->GetCmdBuffer();
if (cmdBuffer->IsInsideRenderPass())
EndRenderPass();
_barriers.Execute(cmdBuffer);
FlushBarriers();
#endif
}
@@ -315,12 +308,7 @@ void GPUContextVulkan::AddBufferBarrier(GPUBufferVulkan* buffer, VkAccessFlags d
#if VK_ENABLE_BARRIERS_BATCHING
// Auto-flush on overflow
if (_barriers.IsFull())
{
const auto cmdBuffer = _cmdBufferManager->GetCmdBuffer();
if (cmdBuffer->IsInsideRenderPass())
EndRenderPass();
_barriers.Execute(cmdBuffer);
}
FlushBarriers();
#endif
// Insert barrier
@@ -339,13 +327,38 @@ void GPUContextVulkan::AddBufferBarrier(GPUBufferVulkan* buffer, VkAccessFlags d
#if !VK_ENABLE_BARRIERS_BATCHING
// Auto-flush without batching
const auto cmdBuffer = _cmdBufferManager->GetCmdBuffer();
if (cmdBuffer->IsInsideRenderPass())
EndRenderPass();
_barriers.Execute(cmdBuffer);
FlushBarriers();
#endif
}
void GPUContextVulkan::AddMemoryBarrier()
{
#if VK_ENABLE_BARRIERS_BATCHING
// Auto-flush on overflow
if (_barriers.IsFull())
FlushBarriers();
#endif
// Insert barrier
VkMemoryBarrier& memoryBarrier = _barriers.MemoryBarriers.AddOne();
RenderToolsVulkan::ZeroStruct(memoryBarrier, VK_STRUCTURE_TYPE_MEMORY_BARRIER);
memoryBarrier.srcAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT;
memoryBarrier.dstAccessMask = VK_ACCESS_MEMORY_READ_BIT | VK_ACCESS_MEMORY_WRITE_BIT;
_barriers.SourceStage |= VK_PIPELINE_STAGE_TRANSFER_BIT;
_barriers.DestStage |= VK_PIPELINE_STAGE_ALL_COMMANDS_BIT;
#if !VK_ENABLE_BARRIERS_BATCHING
// Auto-flush without batching
FlushBarriers();
#endif
}
void GPUContextVulkan::AddUABarrier()
{
_barriers.SourceStage |= VK_PIPELINE_STAGE_BOTTOM_OF_PIPE_BIT;
_barriers.DestStage |= VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT;
}
void GPUContextVulkan::FlushBarriers()
{
#if VK_ENABLE_BARRIERS_BATCHING
@@ -475,7 +488,7 @@ void GPUContextVulkan::EndRenderPass()
cmdBuffer->EndRenderPass();
_renderPass = nullptr;
// Place a barrier between RenderPasses, so that color / depth outputs can be read in subsequent passes
// Place a barrier between RenderPasses, so that color/depth outputs can be read in subsequent passes
// TODO: remove it in future and use proper barriers without whole pipeline stalls
vkCmdPipelineBarrier(cmdBuffer->GetHandle(), VK_PIPELINE_STAGE_BOTTOM_OF_PIPE_BIT, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT, 0, 0, nullptr, 0, nullptr, 0, nullptr);
}
@@ -1155,8 +1168,8 @@ void GPUContextVulkan::Dispatch(GPUShaderProgramCS* shader, uint32 threadGroupCo
RENDER_STAT_DISPATCH_CALL();
// Place a barrier between dispatches, so that UAVs can be read+write in subsequent passes
// TODO: optimize it by moving inputs/outputs into higher-layer so eg. Global SDF can manually optimize it
vkCmdPipelineBarrier(cmdBuffer->GetHandle(), VK_PIPELINE_STAGE_BOTTOM_OF_PIPE_BIT, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT, 0, 0, nullptr, 0, nullptr, 0, nullptr);
if (_pass == 0)
AddUABarrier();
#if VK_ENABLE_BARRIERS_DEBUG
LOG(Warning, "Dispatch");
@@ -1191,8 +1204,8 @@ void GPUContextVulkan::DispatchIndirect(GPUShaderProgramCS* shader, GPUBuffer* b
RENDER_STAT_DISPATCH_CALL();
// Place a barrier between dispatches, so that UAVs can be read+write in subsequent passes
// TODO: optimize it by moving inputs/outputs into higher-layer so eg. Global SDF can manually optimize it
vkCmdPipelineBarrier(cmdBuffer->GetHandle(), VK_PIPELINE_STAGE_BOTTOM_OF_PIPE_BIT, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT, 0, 0, nullptr, 0, nullptr, 0, nullptr);
if (_pass == 0)
AddUABarrier();
#if VK_ENABLE_BARRIERS_DEBUG
LOG(Warning, "DispatchIndirect");
@@ -1351,18 +1364,14 @@ void GPUContextVulkan::UpdateBuffer(GPUBuffer* buffer, const void* data, uint32
const auto bufferVulkan = static_cast<GPUBufferVulkan*>(buffer);
// Memory transfer barrier
// TODO: batch pipeline barriers
const VkMemoryBarrier barrierBefore = { VK_STRUCTURE_TYPE_MEMORY_BARRIER, nullptr, VK_ACCESS_MEMORY_WRITE_BIT, VK_ACCESS_MEMORY_READ_BIT };
vkCmdPipelineBarrier(cmdBuffer->GetHandle(), VK_PIPELINE_STAGE_TRANSFER_BIT, VK_PIPELINE_STAGE_TRANSFER_BIT, 0, 1, &barrierBefore, 0, nullptr, 0, nullptr);
// Transition resource
AddBufferBarrier(bufferVulkan, VK_ACCESS_TRANSFER_WRITE_BIT);
FlushBarriers();
// Use direct update for small buffers
const uint32 alignedSize = Math::AlignUp<uint32>(size, 4);
if (size <= 4 * 1024 && alignedSize <= buffer->GetSize())
{
//AddBufferBarrier(bufferVulkan, VK_ACCESS_TRANSFER_WRITE_BIT);
//FlushBarriers();
vkCmdUpdateBuffer(cmdBuffer->GetHandle(), bufferVulkan->GetHandle(), offset, alignedSize, data);
}
else
@@ -1379,10 +1388,9 @@ void GPUContextVulkan::UpdateBuffer(GPUBuffer* buffer, const void* data, uint32
_device->StagingManager.ReleaseBuffer(cmdBuffer, staging);
}
// Memory transfer barrier
// TODO: batch pipeline barriers
const VkMemoryBarrier barrierAfter = { VK_STRUCTURE_TYPE_MEMORY_BARRIER, nullptr, VK_ACCESS_TRANSFER_WRITE_BIT, VK_ACCESS_MEMORY_READ_BIT | VK_ACCESS_MEMORY_WRITE_BIT };
vkCmdPipelineBarrier(cmdBuffer->GetHandle(), VK_PIPELINE_STAGE_TRANSFER_BIT, VK_PIPELINE_STAGE_ALL_COMMANDS_BIT, 0, 1, &barrierAfter, 0, nullptr, 0, nullptr);
// Memory transfer barrier to ensure buffer is ready to read (eg. by Draw or Dispatch)
if (_pass == 0)
AddMemoryBarrier();
}
void GPUContextVulkan::CopyBuffer(GPUBuffer* dstBuffer, GPUBuffer* srcBuffer, uint32 size, uint32 dstOffset, uint32 srcOffset)
@@ -1407,6 +1415,10 @@ void GPUContextVulkan::CopyBuffer(GPUBuffer* dstBuffer, GPUBuffer* srcBuffer, ui
bufferCopy.dstOffset = dstOffset;
bufferCopy.size = size;
vkCmdCopyBuffer(cmdBuffer->GetHandle(), srcBufferVulkan->GetHandle(), dstBufferVulkan->GetHandle(), 1, &bufferCopy);
// Memory transfer barrier to ensure buffer is ready to read (eg. by Draw or Dispatch)
if (_pass == 0)
AddMemoryBarrier();
}
void GPUContextVulkan::UpdateTexture(GPUTexture* texture, int32 arrayIndex, int32 mipIndex, const void* data, uint32 rowPitch, uint32 slicePitch)
@@ -1816,4 +1828,27 @@ void GPUContextVulkan::CopySubresource(GPUResource* dstResource, uint32 dstSubre
}
}
void GPUContextVulkan::Transition(GPUResource* resource, GPUResourceAccess access)
{
if (auto buffer = dynamic_cast<GPUBufferVulkan*>(resource))
{
AddBufferBarrier(buffer, RenderToolsVulkan::GetAccess(access));
}
else if (auto texture = dynamic_cast<GPUTextureVulkan*>(resource))
{
AddImageBarrier(texture, RenderToolsVulkan::GetImageLayout(access));
}
}
void GPUContextVulkan::MemoryBarrier()
{
AddMemoryBarrier();
}
void GPUContextVulkan::OverlapUA(bool end)
{
if (end)
AddUABarrier();
}
#endif