MemoryAllocation WrappedVulkan::AllocateMemoryForResource(VkBuffer buf, MemoryScope scope,
MemoryType type)
{
VkDevice d = GetDev();
VkMemoryRequirements mrq = {};
ObjDisp(d)->GetBufferMemoryRequirements(Unwrap(d), Unwrap(buf), &mrq);
return AllocateMemoryForResource(true, mrq, scope, type);
}
MemoryAllocation WrappedVulkan::AllocateMemoryForResource(VkImage im, MemoryScope scope,
MemoryType type)
{
VkDevice d = GetDev();
VkMemoryRequirements mrq = {};
ObjDisp(d)->GetImageMemoryRequirements(Unwrap(d), Unwrap(im), &mrq);
return AllocateMemoryForResource(false, mrq, scope, type);
}
void WrappedVulkan::FreeAllMemory(MemoryScope scope)
{
VkDevice d = GetDev();
std::vector<MemoryAllocation> &allocList = m_MemoryBlocks[(size_t)scope];
for(MemoryAllocation alloc : allocList)
{
ObjDisp(d)->FreeMemory(Unwrap(d), Unwrap(alloc.mem), NULL);
GetResourceManager()->ReleaseWrappedResource(alloc.mem);
}
allocList.clear();
}
bool WrappedVulkan::Serialise_vkCmdWaitEvents(
Serialiser* localSerialiser,
VkCommandBuffer cmdBuffer,
uint32_t eventCount,
const VkEvent* pEvents,
VkPipelineStageFlags srcStageMask,
VkPipelineStageFlags dstStageMask,
uint32_t memoryBarrierCount,
const VkMemoryBarrier* pMemoryBarriers,
uint32_t bufferMemoryBarrierCount,
const VkBufferMemoryBarrier* pBufferMemoryBarriers,
uint32_t imageMemoryBarrierCount,
const VkImageMemoryBarrier* pImageMemoryBarriers)
{
SERIALISE_ELEMENT(ResourceId, cmdid, GetResID(cmdBuffer));
SERIALISE_ELEMENT(VkPipelineStageFlagBits, srcStages, (VkPipelineStageFlagBits)srcStageMask);
SERIALISE_ELEMENT(VkPipelineStageFlagBits, destStages, (VkPipelineStageFlagBits)dstStageMask);
// we don't serialise the original events as we are going to replace this
// with our own
SERIALISE_ELEMENT(uint32_t, memCount, memoryBarrierCount);
SERIALISE_ELEMENT(uint32_t, bufCount, bufferMemoryBarrierCount);
SERIALISE_ELEMENT(uint32_t, imgCount, imageMemoryBarrierCount);
// we keep the original memory barriers
SERIALISE_ELEMENT_ARR(VkMemoryBarrier, memBarriers, pMemoryBarriers, memCount);
SERIALISE_ELEMENT_ARR(VkBufferMemoryBarrier, bufMemBarriers, pBufferMemoryBarriers, bufCount);
SERIALISE_ELEMENT_ARR(VkImageMemoryBarrier, imgMemBarriers, pImageMemoryBarriers, imgCount);
vector<VkImageMemoryBarrier> imgBarriers;
vector<VkBufferMemoryBarrier> bufBarriers;
// it's possible for buffer or image to be NULL if it refers to a resource that is otherwise
// not in the log (barriers do not mark resources referenced). If the resource in question does
// not exist, then it's safe to skip this barrier.
if(m_State < WRITING)
{
for(uint32_t i=0; i < bufCount; i++)
if(bufMemBarriers[i].buffer != VK_NULL_HANDLE)
bufBarriers.push_back(bufMemBarriers[i]);
for(uint32_t i=0; i < imgCount; i++)
{
if(imgMemBarriers[i].image != VK_NULL_HANDLE)
{
imgBarriers.push_back(imgMemBarriers[i]);
ReplacePresentableImageLayout(imgBarriers.back().oldLayout);
ReplacePresentableImageLayout(imgBarriers.back().newLayout);
}
}
}
SAFE_DELETE_ARRAY(bufMemBarriers);
SAFE_DELETE_ARRAY(imgMemBarriers);
// see top of this file for current event/fence handling
if(m_State == EXECUTING)
{
if(ShouldRerecordCmd(cmdid) && InRerecordRange())
{
cmdBuffer = RerecordCmdBuf(cmdid);
VkEventCreateInfo evInfo = {
VK_STRUCTURE_TYPE_EVENT_CREATE_INFO, NULL, 0,
};
VkEvent ev = VK_NULL_HANDLE;
ObjDisp(cmdBuffer)->CreateEvent(Unwrap(GetDev()), &evInfo, NULL, &ev);
// don't wrap this event
ObjDisp(cmdBuffer)->ResetEvent(Unwrap(GetDev()), ev);
ObjDisp(cmdBuffer)->CmdSetEvent(Unwrap(cmdBuffer), ev, VK_PIPELINE_STAGE_ALL_COMMANDS_BIT);
ObjDisp(cmdBuffer)->CmdWaitEvents(Unwrap(cmdBuffer), 1, &ev, (VkPipelineStageFlags)srcStages, (VkPipelineStageFlags)destStages,
memCount, memBarriers,
(uint32_t)bufBarriers.size(), &bufBarriers[0],
(uint32_t)imgBarriers.size(), &imgBarriers[0]);
// register to clean this event up once we're done replaying this section of the log
m_CleanupEvents.push_back(ev);
ResourceId cmd = GetResID(RerecordCmdBuf(cmdid));
GetResourceManager()->RecordBarriers(m_BakedCmdBufferInfo[cmd].imgbarriers, m_ImageLayouts, (uint32_t)imgBarriers.size(), &imgBarriers[0]);
}
}
else if(m_State == READING)
{
cmdBuffer = GetResourceManager()->GetLiveHandle<VkCommandBuffer>(cmdid);
VkEventCreateInfo evInfo = {
VK_STRUCTURE_TYPE_EVENT_CREATE_INFO, NULL, 0,
};
VkEvent ev = VK_NULL_HANDLE;
ObjDisp(cmdBuffer)->CreateEvent(Unwrap(GetDev()), &evInfo, NULL, &ev);
// don't wrap this event
ObjDisp(cmdBuffer)->ResetEvent(Unwrap(GetDev()), ev);
ObjDisp(cmdBuffer)->CmdSetEvent(Unwrap(cmdBuffer), ev, VK_PIPELINE_STAGE_ALL_COMMANDS_BIT);
ObjDisp(cmdBuffer)->CmdWaitEvents(Unwrap(cmdBuffer), 1, &ev, (VkPipelineStageFlags)srcStages, (VkPipelineStageFlags)destStages,
memCount, memBarriers,
(uint32_t)bufBarriers.size(), &bufBarriers[0],
(uint32_t)imgBarriers.size(), &imgBarriers[0]);
// register to clean this event up once we're done replaying this section of the log
m_CleanupEvents.push_back(ev);
ResourceId cmd = GetResID(cmdBuffer);
GetResourceManager()->RecordBarriers(m_BakedCmdBufferInfo[cmd].imgbarriers, m_ImageLayouts, (uint32_t)imgBarriers.size(), &imgBarriers[0]);
}
SAFE_DELETE_ARRAY(memBarriers);
return true;
}
VkResult WrappedVulkan::vkQueueSubmit(VkQueue queue, uint32_t submitCount,
const VkSubmitInfo *pSubmits, VkFence fence)
{
SCOPED_DBG_SINK();
size_t tempmemSize = sizeof(VkSubmitInfo) * submitCount;
// need to count how many semaphore and command buffer arrays to allocate for
for(uint32_t i = 0; i < submitCount; i++)
{
tempmemSize += pSubmits[i].commandBufferCount * sizeof(VkCommandBuffer);
tempmemSize += pSubmits[i].signalSemaphoreCount * sizeof(VkSemaphore);
tempmemSize += pSubmits[i].waitSemaphoreCount * sizeof(VkSemaphore);
}
byte *memory = GetTempMemory(tempmemSize);
VkSubmitInfo *unwrappedSubmits = (VkSubmitInfo *)memory;
VkSemaphore *unwrappedWaitSems = (VkSemaphore *)(unwrappedSubmits + submitCount);
for(uint32_t i = 0; i < submitCount; i++)
{
RDCASSERT(pSubmits[i].sType == VK_STRUCTURE_TYPE_SUBMIT_INFO && pSubmits[i].pNext == NULL);
unwrappedSubmits[i] = pSubmits[i];
unwrappedSubmits[i].pWaitSemaphores =
unwrappedSubmits[i].waitSemaphoreCount ? unwrappedWaitSems : NULL;
for(uint32_t o = 0; o < unwrappedSubmits[i].waitSemaphoreCount; o++)
unwrappedWaitSems[o] = Unwrap(pSubmits[i].pWaitSemaphores[o]);
unwrappedWaitSems += unwrappedSubmits[i].waitSemaphoreCount;
VkCommandBuffer *unwrappedCommandBuffers = (VkCommandBuffer *)unwrappedWaitSems;
unwrappedSubmits[i].pCommandBuffers =
unwrappedSubmits[i].commandBufferCount ? unwrappedCommandBuffers : NULL;
for(uint32_t o = 0; o < unwrappedSubmits[i].commandBufferCount; o++)
unwrappedCommandBuffers[o] = Unwrap(pSubmits[i].pCommandBuffers[o]);
unwrappedCommandBuffers += unwrappedSubmits[i].commandBufferCount;
VkSemaphore *unwrappedSignalSems = (VkSemaphore *)unwrappedCommandBuffers;
unwrappedSubmits[i].pSignalSemaphores =
unwrappedSubmits[i].signalSemaphoreCount ? unwrappedSignalSems : NULL;
for(uint32_t o = 0; o < unwrappedSubmits[i].signalSemaphoreCount; o++)
unwrappedSignalSems[o] = Unwrap(pSubmits[i].pSignalSemaphores[o]);
}
VkResult ret =
ObjDisp(queue)->QueueSubmit(Unwrap(queue), submitCount, unwrappedSubmits, Unwrap(fence));
bool capframe = false;
set<ResourceId> refdIDs;
for(uint32_t s = 0; s < submitCount; s++)
{
for(uint32_t i = 0; i < pSubmits[s].commandBufferCount; i++)
{
ResourceId cmd = GetResID(pSubmits[s].pCommandBuffers[i]);
VkResourceRecord *record = GetRecord(pSubmits[s].pCommandBuffers[i]);
{
SCOPED_LOCK(m_ImageLayoutsLock);
GetResourceManager()->ApplyBarriers(record->bakedCommands->cmdInfo->imgbarriers,
m_ImageLayouts);
}
// need to lock the whole section of code, not just the check on
// m_State, as we also need to make sure we don't check the state,
// start marking dirty resources then while we're doing so the
// state becomes capframe.
// the next sections where we mark resources referenced and add
// the submit chunk to the frame record don't have to be protected.
// Only the decision of whether we're inframe or not, and marking
// dirty.
{
SCOPED_LOCK(m_CapTransitionLock);
if(m_State == WRITING_CAPFRAME)
{
for(auto it = record->bakedCommands->cmdInfo->dirtied.begin();
it != record->bakedCommands->cmdInfo->dirtied.end(); ++it)
GetResourceManager()->MarkPendingDirty(*it);
capframe = true;
}
else
{
for(auto it = record->bakedCommands->cmdInfo->dirtied.begin();
it != record->bakedCommands->cmdInfo->dirtied.end(); ++it)
GetResourceManager()->MarkDirtyResource(*it);
}
}
if(capframe)
{
// for each bound descriptor set, mark it referenced as well as all resources currently
// bound to it
for(auto it = record->bakedCommands->cmdInfo->boundDescSets.begin();
it != record->bakedCommands->cmdInfo->boundDescSets.end(); ++it)
{
GetResourceManager()->MarkResourceFrameReferenced(GetResID(*it), eFrameRef_Read);
VkResourceRecord *setrecord = GetRecord(*it);
for(auto refit = setrecord->descInfo->bindFrameRefs.begin();
refit != setrecord->descInfo->bindFrameRefs.end(); ++refit)
{
refdIDs.insert(refit->first);
GetResourceManager()->MarkResourceFrameReferenced(refit->first, refit->second.second);
if(refit->second.first & DescriptorSetData::SPARSE_REF_BIT)
{
VkResourceRecord *sparserecord = GetResourceManager()->GetResourceRecord(refit->first);
GetResourceManager()->MarkSparseMapReferenced(sparserecord->sparseInfo);
}
}
}
for(auto it = record->bakedCommands->cmdInfo->sparse.begin();
it != record->bakedCommands->cmdInfo->sparse.end(); ++it)
GetResourceManager()->MarkSparseMapReferenced(*it);
// pull in frame refs from this baked command buffer
record->bakedCommands->AddResourceReferences(GetResourceManager());
record->bakedCommands->AddReferencedIDs(refdIDs);
// ref the parent command buffer by itself, this will pull in the cmd buffer pool
GetResourceManager()->MarkResourceFrameReferenced(record->GetResourceID(), eFrameRef_Read);
for(size_t sub = 0; sub < record->bakedCommands->cmdInfo->subcmds.size(); sub++)
{
record->bakedCommands->cmdInfo->subcmds[sub]->bakedCommands->AddResourceReferences(
GetResourceManager());
record->bakedCommands->cmdInfo->subcmds[sub]->bakedCommands->AddReferencedIDs(refdIDs);
GetResourceManager()->MarkResourceFrameReferenced(
record->bakedCommands->cmdInfo->subcmds[sub]->GetResourceID(), eFrameRef_Read);
record->bakedCommands->cmdInfo->subcmds[sub]->bakedCommands->AddRef();
}
GetResourceManager()->MarkResourceFrameReferenced(GetResID(queue), eFrameRef_Read);
if(fence != VK_NULL_HANDLE)
GetResourceManager()->MarkResourceFrameReferenced(GetResID(fence), eFrameRef_Read);
{
SCOPED_LOCK(m_CmdBufferRecordsLock);
m_CmdBufferRecords.push_back(record->bakedCommands);
for(size_t sub = 0; sub < record->bakedCommands->cmdInfo->subcmds.size(); sub++)
m_CmdBufferRecords.push_back(record->bakedCommands->cmdInfo->subcmds[sub]->bakedCommands);
}
record->bakedCommands->AddRef();
}
record->cmdInfo->dirtied.clear();
}
}
if(capframe)
{
vector<VkResourceRecord *> maps;
{
SCOPED_LOCK(m_CoherentMapsLock);
maps = m_CoherentMaps;
}
for(auto it = maps.begin(); it != maps.end(); ++it)
{
VkResourceRecord *record = *it;
MemMapState &state = *record->memMapState;
// potential persistent map
if(state.mapCoherent && state.mappedPtr && !state.mapFlushed)
{
// only need to flush memory that could affect this submitted batch of work
if(refdIDs.find(record->GetResourceID()) == refdIDs.end())
{
RDCDEBUG("Map of memory %llu not referenced in this queue - not flushing",
record->GetResourceID());
continue;
}
size_t diffStart = 0, diffEnd = 0;
bool found = true;
// enabled as this is necessary for programs with very large coherent mappings
// (> 1GB) as otherwise more than a couple of vkQueueSubmit calls leads to vast
// memory allocation. There might still be bugs lurking in here though
#if 1
// this causes vkFlushMappedMemoryRanges call to allocate and copy to refData
// from serialised buffer. We want to copy *precisely* the serialised data,
// otherwise there is a gap in time between serialising out a snapshot of
// the buffer and whenever we then copy into the ref data, e.g. below.
// during this time, data could be written to the buffer and it won't have
// been caught in the serialised snapshot, and if it doesn't change then
// it *also* won't be caught in any future FindDiffRange() calls.
//
// Likewise once refData is allocated, the call below will also update it
// with the data serialised out for the same reason.
//
// Note: it's still possible that data is being written to by the
// application while it's being serialised out in the snapshot below. That
// is OK, since the application is responsible for ensuring it's not writing
// data that would be needed by the GPU in this submit. As long as the
// refdata we use for future use is identical to what was serialised, we
// shouldn't miss anything
state.needRefData = true;
// if we have a previous set of data, compare.
// otherwise just serialise it all
if(state.refData)
found = FindDiffRange((byte *)state.mappedPtr, state.refData, (size_t)state.mapSize,
diffStart, diffEnd);
else
#endif
diffEnd = (size_t)state.mapSize;
if(found)
{
// MULTIDEVICE should find the device for this queue.
// MULTIDEVICE only want to flush maps associated with this queue
VkDevice dev = GetDev();
{
RDCLOG("Persistent map flush forced for %llu (%llu -> %llu)", record->GetResourceID(),
(uint64_t)diffStart, (uint64_t)diffEnd);
VkMappedMemoryRange range = {VK_STRUCTURE_TYPE_MAPPED_MEMORY_RANGE, NULL,
(VkDeviceMemory)(uint64_t)record->Resource,
state.mapOffset + diffStart, diffEnd - diffStart};
vkFlushMappedMemoryRanges(dev, 1, &range);
state.mapFlushed = false;
}
GetResourceManager()->MarkPendingDirty(record->GetResourceID());
}
else
{
RDCDEBUG("Persistent map flush not needed for %llu", record->GetResourceID());
}
}
}
{
CACHE_THREAD_SERIALISER();
for(uint32_t s = 0; s < submitCount; s++)
{
SCOPED_SERIALISE_CONTEXT(QUEUE_SUBMIT);
Serialise_vkQueueSubmit(localSerialiser, queue, 1, &pSubmits[s], fence);
m_FrameCaptureRecord->AddChunk(scope.Get());
for(uint32_t sem = 0; sem < pSubmits[s].waitSemaphoreCount; sem++)
GetResourceManager()->MarkResourceFrameReferenced(
GetResID(pSubmits[s].pWaitSemaphores[sem]), eFrameRef_Read);
for(uint32_t sem = 0; sem < pSubmits[s].signalSemaphoreCount; sem++)
GetResourceManager()->MarkResourceFrameReferenced(
GetResID(pSubmits[s].pSignalSemaphores[sem]), eFrameRef_Read);
}
}
}
return ret;
}
VkResult WrappedVulkan::vkQueuePresentKHR(
VkQueue queue,
const VkPresentInfoKHR* pPresentInfo)
{
if(m_State == WRITING_IDLE)
{
RenderDoc::Inst().Tick();
GetResourceManager()->FlushPendingDirty();
}
m_FrameCounter++; // first present becomes frame #1, this function is at the end of the frame
if(pPresentInfo->swapchainCount > 1 && (m_FrameCounter % 100) == 0)
{
RDCWARN("Presenting multiple swapchains at once - only first will be processed");
}
vector<VkSwapchainKHR> unwrappedSwaps;
vector<VkSemaphore> unwrappedSems;
VkPresentInfoKHR unwrappedInfo = *pPresentInfo;
for(uint32_t i=0; i < unwrappedInfo.swapchainCount; i++)
unwrappedSwaps.push_back(Unwrap(unwrappedInfo.pSwapchains[i]));
for(uint32_t i=0; i < unwrappedInfo.waitSemaphoreCount; i++)
unwrappedSems.push_back(Unwrap(unwrappedInfo.pWaitSemaphores[i]));
unwrappedInfo.pSwapchains = unwrappedInfo.swapchainCount ? &unwrappedSwaps[0] : NULL;
unwrappedInfo.pWaitSemaphores = unwrappedInfo.waitSemaphoreCount ? &unwrappedSems[0] : NULL;
// Don't support any extensions for present info
RDCASSERT(pPresentInfo->pNext == NULL);
VkResourceRecord *swaprecord = GetRecord(pPresentInfo->pSwapchains[0]);
RDCASSERT(swaprecord->swapInfo);
SwapchainInfo &swapInfo = *swaprecord->swapInfo;
bool activeWindow = RenderDoc::Inst().IsActiveWindow(LayerDisp(m_Instance), swapInfo.wndHandle);
// need to record which image was last flipped so we can get the correct backbuffer
// for a thumbnail in EndFrameCapture
swapInfo.lastPresent = pPresentInfo->pImageIndices[0];
m_LastSwap = swaprecord->GetResourceID();
VkImage backbuffer = swapInfo.images[pPresentInfo->pImageIndices[0]].im;
if(m_State == WRITING_IDLE)
{
m_FrameTimes.push_back(m_FrameTimer.GetMilliseconds());
m_TotalTime += m_FrameTimes.back();
m_FrameTimer.Restart();
// update every second
if(m_TotalTime > 1000.0)
{
m_MinFrametime = 10000.0;
m_MaxFrametime = 0.0;
m_AvgFrametime = 0.0;
m_TotalTime = 0.0;
for(size_t i=0; i < m_FrameTimes.size(); i++)
{
m_AvgFrametime += m_FrameTimes[i];
if(m_FrameTimes[i] < m_MinFrametime)
m_MinFrametime = m_FrameTimes[i];
if(m_FrameTimes[i] > m_MaxFrametime)
m_MaxFrametime = m_FrameTimes[i];
}
m_AvgFrametime /= double(m_FrameTimes.size());
m_FrameTimes.clear();
}
uint32_t overlay = RenderDoc::Inst().GetOverlayBits();
if(overlay & eRENDERDOC_Overlay_Enabled)
{
VkRenderPass rp = swapInfo.rp;
VkImage im = swapInfo.images[pPresentInfo->pImageIndices[0]].im;
VkFramebuffer fb = swapInfo.images[pPresentInfo->pImageIndices[0]].fb;
VkLayerDispatchTable *vt = ObjDisp(GetDev());
TextPrintState textstate = { GetNextCmd(), rp, fb, swapInfo.extent.width, swapInfo.extent.height };
VkCommandBufferBeginInfo beginInfo = { VK_STRUCTURE_TYPE_COMMAND_BUFFER_BEGIN_INFO, NULL, VK_COMMAND_BUFFER_USAGE_ONE_TIME_SUBMIT_BIT };
VkResult vkr = vt->BeginCommandBuffer(Unwrap(textstate.cmd), &beginInfo);
RDCASSERTEQUAL(vkr, VK_SUCCESS);
VkImageMemoryBarrier bbBarrier = {
VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER, NULL,
0, 0, VK_IMAGE_LAYOUT_PRESENT_SRC_KHR, VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL,
VK_QUEUE_FAMILY_IGNORED, VK_QUEUE_FAMILY_IGNORED,
Unwrap(im),
{ VK_IMAGE_ASPECT_COLOR_BIT, 0, 1, 0, 1 }
};
bbBarrier.srcAccessMask = VK_ACCESS_ALL_READ_BITS;
bbBarrier.dstAccessMask = VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT;
DoPipelineBarrier(textstate.cmd, 1, &bbBarrier);
GetDebugManager()->BeginText(textstate);
if(activeWindow)
{
vector<RENDERDOC_InputButton> keys = RenderDoc::Inst().GetCaptureKeys();
string overlayText = "Vulkan. ";
for(size_t i=0; i < keys.size(); i++)
{
if(i > 0)
overlayText += ", ";
overlayText += ToStr::Get(keys[i]);
}
if(!keys.empty())
overlayText += " to capture.";
if(overlay & eRENDERDOC_Overlay_FrameNumber)
{
overlayText += StringFormat::Fmt(" Frame: %d.", m_FrameCounter);
}
if(overlay & eRENDERDOC_Overlay_FrameRate)
{
overlayText += StringFormat::Fmt(" %.2lf ms (%.2lf .. %.2lf) (%.0lf FPS)",
m_AvgFrametime, m_MinFrametime, m_MaxFrametime, 1000.0f/m_AvgFrametime);
}
float y=0.0f;
if(!overlayText.empty())
{
GetDebugManager()->RenderText(textstate, 0.0f, y, overlayText.c_str());
y += 1.0f;
}
if(overlay & eRENDERDOC_Overlay_CaptureList)
{
GetDebugManager()->RenderText(textstate, 0.0f, y, "%d Captures saved.\n", (uint32_t)m_FrameRecord.size());
y += 1.0f;
uint64_t now = Timing::GetUnixTimestamp();
for(size_t i=0; i < m_FrameRecord.size(); i++)
{
if(now - m_FrameRecord[i].frameInfo.captureTime < 20)
{
GetDebugManager()->RenderText(textstate, 0.0f, y, "Captured frame %d.\n", m_FrameRecord[i].frameInfo.frameNumber);
y += 1.0f;
}
}
}
#if !defined(RELEASE)
GetDebugManager()->RenderText(textstate, 0.0f, y, "%llu chunks - %.2f MB", Chunk::NumLiveChunks(), float(Chunk::TotalMem())/1024.0f/1024.0f);
y += 1.0f;
#endif
}
else
{
vector<RENDERDOC_InputButton> keys = RenderDoc::Inst().GetFocusKeys();
string str = "Vulkan. Inactive swapchain.";
for(size_t i=0; i < keys.size(); i++)
{
if(i == 0)
str += " ";
else
str += ", ";
str += ToStr::Get(keys[i]);
}
if(!keys.empty())
str += " to cycle between swapchains";
GetDebugManager()->RenderText(textstate, 0.0f, 0.0f, str.c_str());
}
GetDebugManager()->EndText(textstate);
std::swap(bbBarrier.oldLayout, bbBarrier.newLayout);
bbBarrier.srcAccessMask = VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT;
bbBarrier.dstAccessMask = VK_ACCESS_ALL_READ_BITS;
DoPipelineBarrier(textstate.cmd, 1, &bbBarrier);
ObjDisp(textstate.cmd)->EndCommandBuffer(Unwrap(textstate.cmd));
SubmitCmds();
FlushQ();
}
}
VkResult vkr = ObjDisp(queue)->QueuePresentKHR(Unwrap(queue), &unwrappedInfo);
if(!activeWindow)
return vkr;
RenderDoc::Inst().SetCurrentDriver(RDC_Vulkan);
// kill any current capture that isn't application defined
if(m_State == WRITING_CAPFRAME && !m_AppControlledCapture)
RenderDoc::Inst().EndFrameCapture(LayerDisp(m_Instance), swapInfo.wndHandle);
if(RenderDoc::Inst().ShouldTriggerCapture(m_FrameCounter) && m_State == WRITING_IDLE)
{
RenderDoc::Inst().StartFrameCapture(LayerDisp(m_Instance), swapInfo.wndHandle);
m_AppControlledCapture = false;
}
return vkr;
}
bool WrappedVulkan::ReleaseResource(WrappedVkRes *res)
{
if(res == NULL) return true;
// MULTIDEVICE need to get the actual device that created this object
VkDevice dev = GetDev();
const VkLayerDispatchTable *vt = ObjDisp(dev);
WrappedVkNonDispRes *nondisp = (WrappedVkNonDispRes *)res;
WrappedVkDispRes *disp = (WrappedVkDispRes *)res;
uint64_t handle = (uint64_t)nondisp;
switch(IdentifyTypeByPtr(res))
{
case eResSurface:
case eResSwapchain:
if(m_State >= WRITING)
RDCERR("Swapchain/swapchain object is leaking");
else
RDCERR("Should be no swapchain/surface objects created on replay");
break;
case eResUnknown:
RDCERR("Unknown resource type!");
break;
case eResCommandBuffer:
// special case here, on replay we don't have the tracking
// to remove these with the parent object so do it here.
// This ensures we clean up after ourselves with a well-
// behaved application.
if(m_State < WRITING)
GetResourceManager()->ReleaseWrappedResource((VkCommandBuffer)res);
break;
case eResDescriptorSet:
if(m_State < WRITING)
GetResourceManager()->ReleaseWrappedResource(VkDescriptorSet(handle));
break;
case eResPhysicalDevice:
if(m_State < WRITING)
GetResourceManager()->ReleaseWrappedResource((VkPhysicalDevice)disp);
break;
case eResQueue:
if(m_State < WRITING)
GetResourceManager()->ReleaseWrappedResource((VkQueue)disp);
break;
case eResDevice:
// these are explicitly released elsewhere, do not need to destroy
// any API objects.
// On replay though we do need to tidy up book-keeping for these.
if(m_State < WRITING)
{
GetResourceManager()->ReleaseCurrentResource(disp->id);
GetResourceManager()->RemoveWrapper(ToTypedHandle(disp->real.As<VkDevice>()));
}
break;
case eResInstance:
if(m_State < WRITING)
{
GetResourceManager()->ReleaseCurrentResource(disp->id);
GetResourceManager()->RemoveWrapper(ToTypedHandle(disp->real.As<VkInstance>()));
}
break;
case eResDeviceMemory:
{
VkDeviceMemory real = nondisp->real.As<VkDeviceMemory>();
GetResourceManager()->ReleaseWrappedResource(VkDeviceMemory(handle));
vt->FreeMemory(Unwrap(dev), real, NULL);
break;
}
case eResBuffer:
{
VkBuffer real = nondisp->real.As<VkBuffer>();
GetResourceManager()->ReleaseWrappedResource(VkBuffer(handle));
vt->DestroyBuffer(Unwrap(dev), real, NULL);
break;
}
case eResBufferView:
{
VkBufferView real = nondisp->real.As<VkBufferView>();
GetResourceManager()->ReleaseWrappedResource(VkBufferView(handle));
vt->DestroyBufferView(Unwrap(dev), real, NULL);
break;
}
case eResImage:
{
VkImage real = nondisp->real.As<VkImage>();
GetResourceManager()->ReleaseWrappedResource(VkImage(handle));
vt->DestroyImage(Unwrap(dev), real, NULL);
break;
}
case eResImageView:
{
VkImageView real = nondisp->real.As<VkImageView>();
GetResourceManager()->ReleaseWrappedResource(VkImageView(handle));
vt->DestroyImageView(Unwrap(dev), real, NULL);
break;
}
case eResFramebuffer:
{
VkFramebuffer real = nondisp->real.As<VkFramebuffer>();
GetResourceManager()->ReleaseWrappedResource(VkFramebuffer(handle));
vt->DestroyFramebuffer(Unwrap(dev), real, NULL);
break;
}
case eResRenderPass:
{
VkRenderPass real = nondisp->real.As<VkRenderPass>();
GetResourceManager()->ReleaseWrappedResource(VkRenderPass(handle));
vt->DestroyRenderPass(Unwrap(dev), real, NULL);
break;
}
case eResShaderModule:
{
VkShaderModule real = nondisp->real.As<VkShaderModule>();
GetResourceManager()->ReleaseWrappedResource(VkShaderModule(handle));
vt->DestroyShaderModule(Unwrap(dev), real, NULL);
break;
}
case eResPipelineCache:
{
VkPipelineCache real = nondisp->real.As<VkPipelineCache>();
GetResourceManager()->ReleaseWrappedResource(VkPipelineCache(handle));
vt->DestroyPipelineCache(Unwrap(dev), real, NULL);
break;
}
case eResPipelineLayout:
{
VkPipelineLayout real = nondisp->real.As<VkPipelineLayout>();
GetResourceManager()->ReleaseWrappedResource(VkPipelineLayout(handle));
vt->DestroyPipelineLayout(Unwrap(dev), real, NULL);
break;
}
case eResPipeline:
{
VkPipeline real = nondisp->real.As<VkPipeline>();
GetResourceManager()->ReleaseWrappedResource(VkPipeline(handle));
vt->DestroyPipeline(Unwrap(dev), real, NULL);
break;
}
case eResSampler:
{
VkSampler real = nondisp->real.As<VkSampler>();
GetResourceManager()->ReleaseWrappedResource(VkSampler(handle));
vt->DestroySampler(Unwrap(dev), real, NULL);
break;
}
case eResDescriptorPool:
{
VkDescriptorPool real = nondisp->real.As<VkDescriptorPool>();
GetResourceManager()->ReleaseWrappedResource(VkDescriptorPool(handle));
vt->DestroyDescriptorPool(Unwrap(dev), real, NULL);
break;
}
case eResDescriptorSetLayout:
{
VkDescriptorSetLayout real = nondisp->real.As<VkDescriptorSetLayout>();
GetResourceManager()->ReleaseWrappedResource(VkDescriptorSetLayout(handle));
vt->DestroyDescriptorSetLayout(Unwrap(dev), real, NULL);
break;
}
case eResCommandPool:
{
VkCommandPool real = nondisp->real.As<VkCommandPool>();
GetResourceManager()->ReleaseWrappedResource(VkCommandPool(handle));
vt->DestroyCommandPool(Unwrap(dev), real, NULL);
break;
}
case eResFence:
{
VkFence real = nondisp->real.As<VkFence>();
GetResourceManager()->ReleaseWrappedResource(VkFence(handle));
vt->DestroyFence(Unwrap(dev), real, NULL);
break;
}
case eResEvent:
{
VkEvent real = nondisp->real.As<VkEvent>();
GetResourceManager()->ReleaseWrappedResource(VkEvent(handle));
vt->DestroyEvent(Unwrap(dev), real, NULL);
break;
}
case eResQueryPool:
{
VkQueryPool real = nondisp->real.As<VkQueryPool>();
GetResourceManager()->ReleaseWrappedResource(VkQueryPool(handle));
vt->DestroyQueryPool(Unwrap(dev), real, NULL);
break;
}
case eResSemaphore:
{
VkSemaphore real = nondisp->real.As<VkSemaphore>();
GetResourceManager()->ReleaseWrappedResource(VkSemaphore(handle));
vt->DestroySemaphore(Unwrap(dev), real, NULL);
break;
}
}
return true;
}
MemoryAllocation WrappedVulkan::AllocateMemoryForResource(bool buffer, VkMemoryRequirements mrq,
MemoryScope scope, MemoryType type)
{
MemoryAllocation ret;
ret.scope = scope;
ret.type = type;
ret.buffer = buffer;
ret.size = AlignUp(mrq.size, mrq.alignment);
RDCDEBUG("Allocating 0x%llx with alignment 0x%llx in 0x%x for a %s (%s in %s)", ret.size,
mrq.alignment, mrq.memoryTypeBits, buffer ? "buffer" : "image", ToStr(type).c_str(),
ToStr(scope).c_str());
std::vector<MemoryAllocation> &blockList = m_MemoryBlocks[(size_t)scope];
// first try to find a match
int i = 0;
for(MemoryAllocation &block : blockList)
{
RDCDEBUG(
"Considering block %d: memory type %u and type %s. Total size 0x%llx, current offset "
"0x%llx, last alloc was %s",
i, block.memoryTypeIndex, ToStr(block.type).c_str(), block.size, block.offs,
block.buffer ? "buffer" : "image");
i++;
// skip this block if it's not the memory type we want
if(ret.type != block.type || (mrq.memoryTypeBits & (1 << block.memoryTypeIndex)) == 0)
{
RDCDEBUG("block type %d or memory type %d is incompatible", block.type, block.memoryTypeIndex);
continue;
}
// offs is where we can put our next sub-allocation
VkDeviceSize offs = block.offs;
// if we are on a buffer/image, account for any alignment we might have to do
if(ret.buffer != block.buffer)
offs = AlignUp(offs, m_PhysicalDeviceData.props.limits.bufferImageGranularity);
// align as required by the resource
offs = AlignUp(offs, mrq.alignment);
if(offs > block.size)
{
RDCDEBUG("Next offset 0x%llx would be off the end of the memory (size 0x%llx).", offs,
block.size);
continue;
}
VkDeviceSize avail = block.size - offs;
RDCDEBUG("At next offset 0x%llx, there's 0x%llx bytes available for 0x%llx bytes requested",
offs, avail, ret.size);
// if the allocation will fit, we've found our candidate.
if(ret.size <= avail)
{
// update the block offset and buffer/image bit
block.offs = offs + ret.size;
block.buffer = ret.buffer;
// update our return value
ret.offs = offs;
ret.mem = block.mem;
RDCDEBUG("Allocating using this block: 0x%llx -> 0x%llx", ret.offs, block.offs);
// stop searching
break;
}
}
if(ret.mem == VK_NULL_HANDLE)
{
RDCDEBUG("No available block found - allocating new block");
VkDeviceSize &allocSize = m_MemoryBlockSize[(size_t)scope];
// we start allocating 32M, then increment each time we need a new block.
switch(allocSize)
{
case 0: allocSize = 32; break;
case 32: allocSize = 64; break;
case 64: allocSize = 128; break;
case 128:
case 256: allocSize = 256; break;
default:
RDCDEBUG("Unexpected previous allocation size 0x%llx bytes, allocating 256MB", allocSize);
allocSize = 256;
break;
}
uint32_t memoryTypeIndex = 0;
switch(ret.type)
{
case MemoryType::Upload: memoryTypeIndex = GetUploadMemoryIndex(mrq.memoryTypeBits); break;
case MemoryType::GPULocal:
memoryTypeIndex = GetGPULocalMemoryIndex(mrq.memoryTypeBits);
break;
case MemoryType::Readback:
memoryTypeIndex = GetReadbackMemoryIndex(mrq.memoryTypeBits);
break;
}
VkMemoryAllocateInfo info = {
VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO, NULL, allocSize * 1024 * 1024, memoryTypeIndex,
};
if(ret.size > info.allocationSize)
{
// if we get an over-sized allocation, first try to immediately jump to the largest block
// size.
allocSize = 256;
info.allocationSize = allocSize * 1024 * 1024;
// if it's still over-sized, just allocate precisely enough and give it a dedicated allocation
if(ret.size > info.allocationSize)
{
RDCDEBUG("Over-sized allocation for 0x%llx bytes", ret.size);
info.allocationSize = ret.size;
}
}
RDCDEBUG("Creating new allocation of 0x%llx bytes", info.allocationSize);
MemoryAllocation chunk;
chunk.buffer = ret.buffer;
chunk.memoryTypeIndex = memoryTypeIndex;
chunk.scope = scope;
chunk.type = type;
chunk.size = info.allocationSize;
// the offset starts immediately after this allocation
chunk.offs = ret.size;
VkDevice d = GetDev();
// do the actual allocation
VkResult vkr = ObjDisp(d)->AllocateMemory(Unwrap(d), &info, NULL, &chunk.mem);
RDCASSERTEQUAL(vkr, VK_SUCCESS);
GetResourceManager()->WrapResource(Unwrap(d), chunk.mem);
// push the new chunk
blockList.push_back(chunk);
// return the first bytes in the new chunk
ret.offs = 0;
ret.mem = chunk.mem;
}
return ret;
}
VkResult WrappedVulkan::vkQueuePresentKHR(VkQueue queue, const VkPresentInfoKHR *pPresentInfo)
{
if(m_State == WRITING_IDLE)
{
RenderDoc::Inst().Tick();
GetResourceManager()->FlushPendingDirty();
}
m_FrameCounter++; // first present becomes frame #1, this function is at the end of the frame
if(pPresentInfo->swapchainCount > 1 && (m_FrameCounter % 100) == 0)
{
RDCWARN("Presenting multiple swapchains at once - only first will be processed");
}
vector<VkSwapchainKHR> unwrappedSwaps;
vector<VkSemaphore> unwrappedSems;
VkPresentInfoKHR unwrappedInfo = *pPresentInfo;
for(uint32_t i = 0; i < unwrappedInfo.swapchainCount; i++)
unwrappedSwaps.push_back(Unwrap(unwrappedInfo.pSwapchains[i]));
for(uint32_t i = 0; i < unwrappedInfo.waitSemaphoreCount; i++)
unwrappedSems.push_back(Unwrap(unwrappedInfo.pWaitSemaphores[i]));
unwrappedInfo.pSwapchains = unwrappedInfo.swapchainCount ? &unwrappedSwaps[0] : NULL;
unwrappedInfo.pWaitSemaphores = unwrappedInfo.waitSemaphoreCount ? &unwrappedSems[0] : NULL;
// Don't support any extensions for present info
RDCASSERT(pPresentInfo->pNext == NULL);
VkResourceRecord *swaprecord = GetRecord(pPresentInfo->pSwapchains[0]);
RDCASSERT(swaprecord->swapInfo);
SwapchainInfo &swapInfo = *swaprecord->swapInfo;
bool activeWindow = RenderDoc::Inst().IsActiveWindow(LayerDisp(m_Instance), swapInfo.wndHandle);
// need to record which image was last flipped so we can get the correct backbuffer
// for a thumbnail in EndFrameCapture
swapInfo.lastPresent = pPresentInfo->pImageIndices[0];
m_LastSwap = swaprecord->GetResourceID();
if(m_State == WRITING_IDLE)
{
uint32_t overlay = RenderDoc::Inst().GetOverlayBits();
if(overlay & eRENDERDOC_Overlay_Enabled)
{
VkRenderPass rp = swapInfo.rp;
VkImage im = swapInfo.images[pPresentInfo->pImageIndices[0]].im;
VkFramebuffer fb = swapInfo.images[pPresentInfo->pImageIndices[0]].fb;
VkLayerDispatchTable *vt = ObjDisp(GetDev());
TextPrintState textstate = {
GetNextCmd(), rp, fb, swapInfo.extent.width, swapInfo.extent.height, swapInfo.format};
VkCommandBufferBeginInfo beginInfo = {VK_STRUCTURE_TYPE_COMMAND_BUFFER_BEGIN_INFO, NULL,
VK_COMMAND_BUFFER_USAGE_ONE_TIME_SUBMIT_BIT};
VkResult vkr = vt->BeginCommandBuffer(Unwrap(textstate.cmd), &beginInfo);
RDCASSERTEQUAL(vkr, VK_SUCCESS);
VkImageMemoryBarrier bbBarrier = {VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
NULL,
0,
0,
VK_IMAGE_LAYOUT_PRESENT_SRC_KHR,
VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL,
VK_QUEUE_FAMILY_IGNORED,
VK_QUEUE_FAMILY_IGNORED,
Unwrap(im),
{VK_IMAGE_ASPECT_COLOR_BIT, 0, 1, 0, 1}};
bbBarrier.srcAccessMask = VK_ACCESS_ALL_READ_BITS;
bbBarrier.dstAccessMask = VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT;
DoPipelineBarrier(textstate.cmd, 1, &bbBarrier);
GetDebugManager()->BeginText(textstate);
int flags = activeWindow ? RenderDoc::eOverlay_ActiveWindow : 0;
string overlayText = RenderDoc::Inst().GetOverlayText(RDC_Vulkan, m_FrameCounter, flags);
if(!overlayText.empty())
GetDebugManager()->RenderText(textstate, 0.0f, 0.0f, overlayText.c_str());
GetDebugManager()->EndText(textstate);
std::swap(bbBarrier.oldLayout, bbBarrier.newLayout);
bbBarrier.srcAccessMask = VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT;
bbBarrier.dstAccessMask = VK_ACCESS_ALL_READ_BITS;
DoPipelineBarrier(textstate.cmd, 1, &bbBarrier);
ObjDisp(textstate.cmd)->EndCommandBuffer(Unwrap(textstate.cmd));
SubmitCmds();
FlushQ();
}
}
VkResult vkr = ObjDisp(queue)->QueuePresentKHR(Unwrap(queue), &unwrappedInfo);
if(!activeWindow)
return vkr;
RenderDoc::Inst().SetCurrentDriver(RDC_Vulkan);
// kill any current capture that isn't application defined
if(m_State == WRITING_CAPFRAME && !m_AppControlledCapture)
RenderDoc::Inst().EndFrameCapture(LayerDisp(m_Instance), swapInfo.wndHandle);
if(RenderDoc::Inst().ShouldTriggerCapture(m_FrameCounter) && m_State == WRITING_IDLE)
{
RenderDoc::Inst().StartFrameCapture(LayerDisp(m_Instance), swapInfo.wndHandle);
m_AppControlledCapture = false;
}
return vkr;
}
