mirror of
https://github.com/baldurk/renderdoc.git
synced 2026-05-06 01:50:38 +00:00
Keep image initial states as buffers in CPU memory, copy from there
* This is potentially slightly less optimal as it means the initial states aren't in GPU memory for a faster copy, but it means we're much less likely to hit OOM due to way more GPU allocs, and it's still pretty fast. * Further optimisation is possible by reducing the number of images that actually either need initial states at all (detect when images are first used to clear via a renderpass), or by detecting images that are frame invariant and we only have initial states for immutable contents and avoid copying them more than once.
This commit is contained in:
baldurk
@@ -1533,8 +1533,6 @@ bool WrappedVulkan::Serialise_InitialState(ResourceId resid, WrappedVkRes *)
|
||||
uint32_t dataSize = 0;
|
||||
m_pSerialiser->Serialise("dataSize", dataSize);
|
||||
|
||||
WrappedVkImage *liveim = (WrappedVkImage *)res;
|
||||
|
||||
VkResult vkr = VK_SUCCESS;
|
||||
|
||||
VkDevice d = GetDev();
|
||||
@@ -1547,16 +1545,17 @@ bool WrappedVulkan::Serialise_InitialState(ResourceId resid, WrappedVkRes *)
|
||||
VK_BUFFER_USAGE_TRANSFER_SRC_BIT | VK_BUFFER_USAGE_TRANSFER_DST_BIT,
|
||||
};
|
||||
|
||||
// short-lived, so not wrapped
|
||||
VkBuffer buf;
|
||||
VkDeviceMemory uploadmem = VK_NULL_HANDLE;
|
||||
|
||||
vkr = ObjDisp(d)->CreateBuffer(Unwrap(d), &bufInfo, NULL, &buf);
|
||||
RDCASSERTEQUAL(vkr, VK_SUCCESS);
|
||||
|
||||
GetResourceManager()->WrapResource(Unwrap(d), buf);
|
||||
|
||||
VkMemoryRequirements mrq = {0};
|
||||
|
||||
ObjDisp(d)->GetBufferMemoryRequirements(Unwrap(d), buf, &mrq);
|
||||
ObjDisp(d)->GetBufferMemoryRequirements(Unwrap(d), Unwrap(buf), &mrq);
|
||||
|
||||
VkMemoryAllocateInfo allocInfo = {
|
||||
VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO, NULL, dataSize,
|
||||
@@ -1568,167 +1567,24 @@ bool WrappedVulkan::Serialise_InitialState(ResourceId resid, WrappedVkRes *)
|
||||
vkr = ObjDisp(d)->AllocateMemory(Unwrap(d), &allocInfo, NULL, &uploadmem);
|
||||
RDCASSERTEQUAL(vkr, VK_SUCCESS);
|
||||
|
||||
vkr = ObjDisp(d)->BindBufferMemory(Unwrap(d), buf, uploadmem, 0);
|
||||
GetResourceManager()->WrapResource(Unwrap(d), uploadmem);
|
||||
|
||||
vkr = ObjDisp(d)->BindBufferMemory(Unwrap(d), Unwrap(buf), Unwrap(uploadmem), 0);
|
||||
RDCASSERTEQUAL(vkr, VK_SUCCESS);
|
||||
|
||||
byte *ptr = NULL;
|
||||
ObjDisp(d)->MapMemory(Unwrap(d), uploadmem, 0, VK_WHOLE_SIZE, 0, (void **)&ptr);
|
||||
ObjDisp(d)->MapMemory(Unwrap(d), Unwrap(uploadmem), 0, VK_WHOLE_SIZE, 0, (void **)&ptr);
|
||||
|
||||
size_t dummy = 0;
|
||||
m_pSerialiser->SerialiseBuffer("data", ptr, dummy);
|
||||
|
||||
ObjDisp(d)->UnmapMemory(Unwrap(d), uploadmem);
|
||||
|
||||
// create image to copy into from the buffer
|
||||
VkImageCreateInfo imInfo = {
|
||||
VK_STRUCTURE_TYPE_IMAGE_CREATE_INFO, NULL, 0, m_CreationInfo.m_Image[liveim->id].type,
|
||||
m_CreationInfo.m_Image[liveim->id].format, m_CreationInfo.m_Image[liveim->id].extent,
|
||||
(uint32_t)m_CreationInfo.m_Image[liveim->id].mipLevels,
|
||||
(uint32_t)m_CreationInfo.m_Image[liveim->id].arrayLayers,
|
||||
m_CreationInfo.m_Image[liveim->id].samples,
|
||||
VK_IMAGE_TILING_OPTIMAL, // make this optimal since the format/etc is more likely
|
||||
// supported as optimal
|
||||
VK_IMAGE_USAGE_TRANSFER_SRC_BIT | VK_IMAGE_USAGE_TRANSFER_DST_BIT,
|
||||
VK_SHARING_MODE_EXCLUSIVE, 0, NULL, VK_IMAGE_LAYOUT_UNDEFINED,
|
||||
};
|
||||
|
||||
if(IsDepthOrStencilFormat(m_CreationInfo.m_Image[liveim->id].format))
|
||||
{
|
||||
imInfo.usage |= VK_IMAGE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT;
|
||||
}
|
||||
|
||||
VkImage im = VK_NULL_HANDLE;
|
||||
|
||||
vkr = ObjDisp(d)->CreateImage(Unwrap(d), &imInfo, NULL, &im);
|
||||
RDCASSERTEQUAL(vkr, VK_SUCCESS);
|
||||
|
||||
GetResourceManager()->WrapResource(Unwrap(d), im);
|
||||
|
||||
ObjDisp(d)->GetImageMemoryRequirements(Unwrap(d), Unwrap(im), &mrq);
|
||||
|
||||
allocInfo.allocationSize = mrq.size;
|
||||
allocInfo.memoryTypeIndex = GetGPULocalMemoryIndex(mrq.memoryTypeBits);
|
||||
|
||||
VkDeviceMemory mem = VK_NULL_HANDLE;
|
||||
|
||||
vkr = ObjDisp(d)->AllocateMemory(Unwrap(d), &allocInfo, NULL, &mem);
|
||||
RDCASSERTEQUAL(vkr, VK_SUCCESS);
|
||||
|
||||
GetResourceManager()->WrapResource(Unwrap(d), mem);
|
||||
|
||||
vkr = ObjDisp(d)->BindImageMemory(Unwrap(d), Unwrap(im), Unwrap(mem), 0);
|
||||
RDCASSERTEQUAL(vkr, VK_SUCCESS);
|
||||
|
||||
VkCommandBufferBeginInfo beginInfo = {VK_STRUCTURE_TYPE_COMMAND_BUFFER_BEGIN_INFO, NULL,
|
||||
VK_COMMAND_BUFFER_USAGE_ONE_TIME_SUBMIT_BIT};
|
||||
|
||||
// INITSTATEBATCH
|
||||
VkCommandBuffer cmd = GetNextCmd();
|
||||
|
||||
vkr = ObjDisp(d)->BeginCommandBuffer(Unwrap(cmd), &beginInfo);
|
||||
RDCASSERTEQUAL(vkr, VK_SUCCESS);
|
||||
|
||||
VkImageAspectFlags aspectFlags = VK_IMAGE_ASPECT_COLOR_BIT;
|
||||
|
||||
VkFormat fmt = m_CreationInfo.m_Image[liveim->id].format;
|
||||
if(IsStencilOnlyFormat(fmt))
|
||||
aspectFlags = VK_IMAGE_ASPECT_STENCIL_BIT;
|
||||
else if(IsDepthOrStencilFormat(fmt))
|
||||
aspectFlags = VK_IMAGE_ASPECT_DEPTH_BIT;
|
||||
|
||||
VkImageMemoryBarrier srcimBarrier = {
|
||||
VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
|
||||
NULL,
|
||||
0,
|
||||
0,
|
||||
VK_IMAGE_LAYOUT_UNDEFINED,
|
||||
VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL,
|
||||
0,
|
||||
0, // MULTIDEVICE - need to actually pick the right queue family here maybe?
|
||||
Unwrap(im),
|
||||
{aspectFlags, 0, 1, 0, 1}};
|
||||
|
||||
if(aspectFlags == VK_IMAGE_ASPECT_DEPTH_BIT && !IsDepthOnlyFormat(fmt))
|
||||
srcimBarrier.subresourceRange.aspectMask |= VK_IMAGE_ASPECT_STENCIL_BIT;
|
||||
|
||||
VkDeviceSize bufOffset = 0;
|
||||
|
||||
// must ensure offset remains valid. Must be multiple of block size, or 4, depending on format
|
||||
VkDeviceSize bufAlignment = 4;
|
||||
if(IsBlockFormat(imInfo.format))
|
||||
bufAlignment = (VkDeviceSize)GetByteSize(1, 1, 1, imInfo.format, 0);
|
||||
|
||||
// copy each slice/mip individually
|
||||
for(uint32_t a = 0; a < imInfo.arrayLayers; a++)
|
||||
{
|
||||
VkExtent3D extent = imInfo.extent;
|
||||
|
||||
for(uint32_t m = 0; m < imInfo.mipLevels; m++)
|
||||
{
|
||||
VkBufferImageCopy region = {
|
||||
0,
|
||||
0,
|
||||
0,
|
||||
{aspectFlags, m, a, 1},
|
||||
{
|
||||
0, 0, 0,
|
||||
},
|
||||
extent,
|
||||
};
|
||||
|
||||
bufOffset = AlignUp(bufOffset, bufAlignment);
|
||||
|
||||
region.bufferOffset = bufOffset;
|
||||
|
||||
bufOffset += GetByteSize(imInfo.extent.width, imInfo.extent.height, imInfo.extent.depth,
|
||||
imInfo.format, m);
|
||||
|
||||
srcimBarrier.subresourceRange.baseArrayLayer = a;
|
||||
srcimBarrier.subresourceRange.baseMipLevel = m;
|
||||
|
||||
// first we update layout from undefined to destination optimal, for the copy from the
|
||||
// buffer
|
||||
srcimBarrier.oldLayout = VK_IMAGE_LAYOUT_UNDEFINED;
|
||||
srcimBarrier.newLayout = VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL;
|
||||
srcimBarrier.dstAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT;
|
||||
|
||||
DoPipelineBarrier(cmd, 1, &srcimBarrier);
|
||||
|
||||
ObjDisp(d)->CmdCopyBufferToImage(Unwrap(cmd), buf, Unwrap(im),
|
||||
VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL, 1, ®ion);
|
||||
|
||||
// then update layout into source optimal, for all subsequent copies from this immutable
|
||||
// initial
|
||||
// state image, to the live image.
|
||||
srcimBarrier.oldLayout = srcimBarrier.newLayout;
|
||||
srcimBarrier.newLayout = VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL;
|
||||
srcimBarrier.srcAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT;
|
||||
srcimBarrier.dstAccessMask = VK_ACCESS_TRANSFER_READ_BIT;
|
||||
|
||||
DoPipelineBarrier(cmd, 1, &srcimBarrier);
|
||||
|
||||
extent.width = RDCMAX(extent.width >> 1, 1U);
|
||||
extent.height = RDCMAX(extent.height >> 1, 1U);
|
||||
extent.depth = RDCMAX(extent.depth >> 1, 1U);
|
||||
}
|
||||
}
|
||||
|
||||
vkr = ObjDisp(d)->EndCommandBuffer(Unwrap(cmd));
|
||||
RDCASSERTEQUAL(vkr, VK_SUCCESS);
|
||||
|
||||
// INITSTATEBATCH
|
||||
SubmitCmds();
|
||||
FlushQ();
|
||||
|
||||
// destroy the temporary buffer for uploading - we just keep the image
|
||||
ObjDisp(d)->DestroyBuffer(Unwrap(d), buf, NULL);
|
||||
ObjDisp(d)->FreeMemory(Unwrap(d), uploadmem, NULL);
|
||||
ObjDisp(d)->UnmapMemory(Unwrap(d), Unwrap(uploadmem));
|
||||
|
||||
// remember to free this memory on shutdown
|
||||
m_CleanupMems.push_back(mem);
|
||||
m_CleanupMems.push_back(uploadmem);
|
||||
|
||||
GetResourceManager()->SetInitialContents(
|
||||
id, VulkanResourceManager::InitialContentData(GetWrapped(im), 0, NULL));
|
||||
id, VulkanResourceManager::InitialContentData(GetWrapped(buf), 0, NULL));
|
||||
}
|
||||
else if(type == eResDeviceMemory)
|
||||
{
|
||||
@@ -2068,7 +1924,7 @@ void WrappedVulkan::Apply_InitialState(WrappedVkRes *live,
|
||||
return;
|
||||
}
|
||||
|
||||
WrappedVkImage *im = (WrappedVkImage *)initial.resource;
|
||||
WrappedVkBuffer *buf = (WrappedVkBuffer *)initial.resource;
|
||||
|
||||
VkCommandBuffer cmd = GetNextCmd();
|
||||
|
||||
@@ -2100,56 +1956,81 @@ void WrappedVulkan::Apply_InitialState(WrappedVkRes *live,
|
||||
if(aspectFlags == VK_IMAGE_ASPECT_DEPTH_BIT && !IsDepthOnlyFormat(fmt))
|
||||
dstimBarrier.subresourceRange.aspectMask |= VK_IMAGE_ASPECT_STENCIL_BIT;
|
||||
|
||||
// loop over every mip
|
||||
for(int m = 0; m < m_CreationInfo.m_Image[id].mipLevels; m++)
|
||||
VkDeviceSize bufOffset = 0;
|
||||
|
||||
// must ensure offset remains valid. Must be multiple of block size, or 4, depending on format
|
||||
VkDeviceSize bufAlignment = 4;
|
||||
if(IsBlockFormat(fmt))
|
||||
bufAlignment = (VkDeviceSize)GetByteSize(1, 1, 1, fmt, 0);
|
||||
|
||||
// copy each slice/mip individually
|
||||
for(int a = 0; a < m_CreationInfo.m_Image[id].arrayLayers; a++)
|
||||
{
|
||||
VkImageCopy region = {
|
||||
{aspectFlags, (uint32_t)m, 0, (uint32_t)m_CreationInfo.m_Image[id].arrayLayers},
|
||||
{0, 0, 0},
|
||||
{aspectFlags, (uint32_t)m, 0, (uint32_t)m_CreationInfo.m_Image[id].arrayLayers},
|
||||
{0, 0, 0},
|
||||
extent,
|
||||
};
|
||||
extent = m_CreationInfo.m_Image[id].extent;
|
||||
|
||||
dstimBarrier.subresourceRange.baseMipLevel = m;
|
||||
|
||||
// first update the live image layout into destination optimal (the initial state
|
||||
// image is always and permanently in source optimal already).
|
||||
dstimBarrier.newLayout = VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL;
|
||||
dstimBarrier.dstAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT;
|
||||
|
||||
for(size_t si = 0; si < m_ImageLayouts[id].subresourceStates.size(); si++)
|
||||
for(int m = 0; m < m_CreationInfo.m_Image[id].mipLevels; m++)
|
||||
{
|
||||
dstimBarrier.subresourceRange = m_ImageLayouts[id].subresourceStates[si].subresourceRange;
|
||||
dstimBarrier.oldLayout = m_ImageLayouts[id].subresourceStates[si].newLayout;
|
||||
dstimBarrier.srcAccessMask =
|
||||
VK_ACCESS_ALL_WRITE_BITS | MakeAccessMask(dstimBarrier.oldLayout);
|
||||
DoPipelineBarrier(cmd, 1, &dstimBarrier);
|
||||
VkBufferImageCopy region = {
|
||||
0,
|
||||
0,
|
||||
0,
|
||||
{aspectFlags, m, a, 1},
|
||||
{
|
||||
0, 0, 0,
|
||||
},
|
||||
extent,
|
||||
};
|
||||
|
||||
bufOffset = AlignUp(bufOffset, bufAlignment);
|
||||
|
||||
region.bufferOffset = bufOffset;
|
||||
|
||||
// pass 0 for mip since we've already pre-downscaled extent
|
||||
bufOffset += GetByteSize(extent.width, extent.height, extent.depth, fmt, 0);
|
||||
|
||||
dstimBarrier.subresourceRange.baseArrayLayer = a;
|
||||
dstimBarrier.subresourceRange.baseMipLevel = m;
|
||||
|
||||
dstimBarrier.subresourceRange.baseMipLevel = m;
|
||||
|
||||
// first update the live image layout into destination optimal (the initial state
|
||||
// image is always and permanently in source optimal already).
|
||||
dstimBarrier.newLayout = VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL;
|
||||
dstimBarrier.dstAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT;
|
||||
|
||||
for(size_t si = 0; si < m_ImageLayouts[id].subresourceStates.size(); si++)
|
||||
{
|
||||
dstimBarrier.subresourceRange = m_ImageLayouts[id].subresourceStates[si].subresourceRange;
|
||||
dstimBarrier.oldLayout = m_ImageLayouts[id].subresourceStates[si].newLayout;
|
||||
dstimBarrier.srcAccessMask =
|
||||
VK_ACCESS_ALL_WRITE_BITS | MakeAccessMask(dstimBarrier.oldLayout);
|
||||
DoPipelineBarrier(cmd, 1, &dstimBarrier);
|
||||
}
|
||||
|
||||
ObjDisp(cmd)->CmdCopyBufferToImage(Unwrap(cmd), buf->real.As<VkBuffer>(),
|
||||
ToHandle<VkImage>(live),
|
||||
VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL, 1, ®ion);
|
||||
|
||||
// update the live image layout back
|
||||
dstimBarrier.oldLayout = dstimBarrier.newLayout;
|
||||
|
||||
// make sure the apply completes before any further work
|
||||
dstimBarrier.srcAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT;
|
||||
dstimBarrier.dstAccessMask = VK_ACCESS_ALL_READ_BITS;
|
||||
|
||||
for(size_t si = 0; si < m_ImageLayouts[id].subresourceStates.size(); si++)
|
||||
{
|
||||
dstimBarrier.subresourceRange = m_ImageLayouts[id].subresourceStates[si].subresourceRange;
|
||||
dstimBarrier.newLayout = m_ImageLayouts[id].subresourceStates[si].newLayout;
|
||||
dstimBarrier.dstAccessMask |= MakeAccessMask(dstimBarrier.newLayout);
|
||||
DoPipelineBarrier(cmd, 1, &dstimBarrier);
|
||||
}
|
||||
|
||||
// update the extent for the next mip
|
||||
extent.width = RDCMAX(extent.width >> 1, 1U);
|
||||
extent.height = RDCMAX(extent.height >> 1, 1U);
|
||||
extent.depth = RDCMAX(extent.depth >> 1, 1U);
|
||||
}
|
||||
|
||||
ObjDisp(cmd)->CmdCopyImage(Unwrap(cmd), im->real.As<VkImage>(),
|
||||
VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL, ToHandle<VkImage>(live),
|
||||
VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL, 1, ®ion);
|
||||
|
||||
// update the live image layout back
|
||||
dstimBarrier.oldLayout = dstimBarrier.newLayout;
|
||||
|
||||
// make sure the apply completes before any further work
|
||||
dstimBarrier.srcAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT;
|
||||
dstimBarrier.dstAccessMask = VK_ACCESS_ALL_READ_BITS;
|
||||
|
||||
for(size_t si = 0; si < m_ImageLayouts[id].subresourceStates.size(); si++)
|
||||
{
|
||||
dstimBarrier.subresourceRange = m_ImageLayouts[id].subresourceStates[si].subresourceRange;
|
||||
dstimBarrier.newLayout = m_ImageLayouts[id].subresourceStates[si].newLayout;
|
||||
dstimBarrier.dstAccessMask |= MakeAccessMask(dstimBarrier.newLayout);
|
||||
DoPipelineBarrier(cmd, 1, &dstimBarrier);
|
||||
}
|
||||
|
||||
// update the extent for the next mip
|
||||
extent.width = RDCMAX(extent.width >> 1, 1U);
|
||||
extent.height = RDCMAX(extent.height >> 1, 1U);
|
||||
extent.depth = RDCMAX(extent.depth >> 1, 1U);
|
||||
}
|
||||
|
||||
vkr = ObjDisp(cmd)->EndCommandBuffer(Unwrap(cmd));
|
||||
|
||||
Reference in New Issue
Block a user