mirror of
https://github.com/doitsujin/dxvk.git
synced 2025-02-21 13:54:18 +01:00
[d3d11] Allocate host-readable images on host-visible memory
Not having to wait for an image->buffer copy to finish on the GPU allows for more efficient synchronization. Significantly improves performance in The Witcher 3.
This commit is contained in:
parent
a4a8e0d6c8
commit
d3e89b20dd
@ -120,17 +120,8 @@ namespace dxvk {
|
||||
cPhysicalSlice.resource()->release();
|
||||
});
|
||||
} else if (MapType != D3D11_MAP_WRITE_NO_OVERWRITE) {
|
||||
// Synchronize with CS thread so that we know whether
|
||||
// the buffer is currently in use by the GPU or not
|
||||
Flush();
|
||||
SynchronizeCsThread();
|
||||
|
||||
while (buffer->isInUse()) {
|
||||
if (MapFlags & D3D11_MAP_FLAG_DO_NOT_WAIT)
|
||||
if (!WaitForResource(buffer->resource(), MapFlags))
|
||||
return DXGI_ERROR_WAS_STILL_DRAWING;
|
||||
|
||||
SynchronizeDevice();
|
||||
}
|
||||
}
|
||||
|
||||
// Use map pointer from previous map operation. This
|
||||
@ -144,11 +135,34 @@ namespace dxvk {
|
||||
pMappedResource->DepthPitch = physicalSlice.length();
|
||||
return S_OK;
|
||||
} else {
|
||||
// Mapping an image is sadly not as simple as mapping a buffer
|
||||
// because applications tend to ignore row and layer strides.
|
||||
// We use a buffer instead and then perform a copy.
|
||||
// Depending on whether the image has been allocated on a
|
||||
// host-visible memory type, we can either use the mapped
|
||||
// memory region directly, or we map a linear buffer.
|
||||
D3D11TextureInfo* textureInfo = GetCommonTextureInfo(pResource);
|
||||
const DxvkImageCreateInfo& imageInfo = textureInfo->image->info();
|
||||
|
||||
textureInfo->mappedSubresource = GetSubresourceFromIndex(
|
||||
VK_IMAGE_ASPECT_COLOR_BIT, imageInfo.mipLevels, Subresource);
|
||||
|
||||
if (textureInfo->image->memFlags() & VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT) {
|
||||
const VkSubresourceLayout subresourceLayout =
|
||||
textureInfo->image->querySubresourceLayout(
|
||||
textureInfo->mappedSubresource);
|
||||
|
||||
if (!WaitForResource(textureInfo->image, MapFlags))
|
||||
return DXGI_ERROR_WAS_STILL_DRAWING;
|
||||
|
||||
pMappedResource->pData = textureInfo->image->mapPtr(subresourceLayout.offset);
|
||||
pMappedResource->RowPitch = subresourceLayout.rowPitch;
|
||||
pMappedResource->DepthPitch = subresourceLayout.rowPitch * imageInfo.extent.height;
|
||||
|
||||
if (imageInfo.type == VK_IMAGE_TYPE_3D)
|
||||
pMappedResource->DepthPitch = subresourceLayout.depthPitch;
|
||||
else if (imageInfo.numLayers > 1)
|
||||
pMappedResource->DepthPitch = subresourceLayout.arrayPitch;
|
||||
|
||||
return S_OK;
|
||||
} else {
|
||||
if (textureInfo->imageBuffer == nullptr) {
|
||||
Logger::err("D3D11: Cannot map a device-local image");
|
||||
return E_INVALIDARG;
|
||||
@ -157,15 +171,10 @@ namespace dxvk {
|
||||
if (pMappedResource == nullptr)
|
||||
return S_FALSE;
|
||||
|
||||
// Query format and subresource in order to compute
|
||||
// Query format info in order to compute
|
||||
// the row pitch and layer pitch properly.
|
||||
const DxvkImageCreateInfo& imageInfo = textureInfo->image->info();
|
||||
const DxvkFormatInfo* formatInfo = imageFormatInfo(imageInfo.format);
|
||||
|
||||
textureInfo->mappedSubresource =
|
||||
GetSubresourceFromIndex(VK_IMAGE_ASPECT_COLOR_BIT,
|
||||
imageInfo.mipLevels, Subresource);
|
||||
|
||||
const VkExtent3D levelExtent = textureInfo->image
|
||||
->mipLevelExtent(textureInfo->mappedSubresource.mipLevel);
|
||||
|
||||
@ -188,6 +197,11 @@ namespace dxvk {
|
||||
cPhysicalSlice.resource()->release();
|
||||
});
|
||||
} else {
|
||||
// We may have to copy the current image contents into the
|
||||
// mapped buffer if the GPU has write access to the image.
|
||||
const bool copyExistingData = textureInfo->usage == D3D11_USAGE_STAGING;
|
||||
|
||||
if (copyExistingData) {
|
||||
const VkImageSubresourceLayers subresourceLayers = {
|
||||
textureInfo->mappedSubresource.aspectMask,
|
||||
textureInfo->mappedSubresource.mipLevel,
|
||||
@ -204,10 +218,10 @@ namespace dxvk {
|
||||
cImage, cSubresources, VkOffset3D { 0, 0, 0 },
|
||||
cLevelExtent);
|
||||
});
|
||||
}
|
||||
|
||||
Flush();
|
||||
SynchronizeCsThread();
|
||||
SynchronizeDevice();
|
||||
if (!WaitForResource(textureInfo->imageBuffer->resource(), MapFlags))
|
||||
return DXGI_ERROR_WAS_STILL_DRAWING;
|
||||
|
||||
physicalSlice = textureInfo->imageBuffer->slice();
|
||||
}
|
||||
@ -219,6 +233,7 @@ namespace dxvk {
|
||||
return S_OK;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
|
||||
void STDMETHODCALLTYPE D3D11ImmediateContext::Unmap(
|
||||
@ -233,6 +248,7 @@ namespace dxvk {
|
||||
const D3D11TextureInfo* textureInfo
|
||||
= GetCommonTextureInfo(pResource);
|
||||
|
||||
if (!(textureInfo->image->memFlags() & VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT)) {
|
||||
const VkExtent3D levelExtent = textureInfo->image
|
||||
->mipLevelExtent(textureInfo->mappedSubresource.mipLevel);
|
||||
|
||||
@ -253,6 +269,7 @@ namespace dxvk {
|
||||
});
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
|
||||
void D3D11ImmediateContext::SynchronizeCsThread() {
|
||||
@ -265,14 +282,29 @@ namespace dxvk {
|
||||
|
||||
|
||||
void D3D11ImmediateContext::SynchronizeDevice() {
|
||||
// FIXME waiting until the device finished executing *all*
|
||||
// pending commands is too pessimistic. Instead we should
|
||||
// wait for individual command submissions to complete.
|
||||
// This will require changes in the DxvkDevice class.
|
||||
m_device->waitForIdle();
|
||||
}
|
||||
|
||||
|
||||
bool D3D11ImmediateContext::WaitForResource(
|
||||
const Rc<DxvkResource>& Resource,
|
||||
UINT MapFlags) {
|
||||
// Wait for the any pending D3D11 command to be executed
|
||||
// on the CS thread so that we can determine whether the
|
||||
// resource is currently in use or not.
|
||||
Flush();
|
||||
SynchronizeCsThread();
|
||||
|
||||
if (Resource->isInUse()) {
|
||||
// TODO implement properly in DxvkDevice
|
||||
while (Resource->isInUse())
|
||||
std::this_thread::yield();
|
||||
}
|
||||
|
||||
return true;
|
||||
}
|
||||
|
||||
|
||||
void D3D11ImmediateContext::EmitCsChunk(Rc<DxvkCsChunk>&& chunk) {
|
||||
m_csThread.dispatchChunk(std::move(chunk));
|
||||
}
|
||||
|
@ -50,6 +50,10 @@ namespace dxvk {
|
||||
|
||||
void SynchronizeDevice();
|
||||
|
||||
bool WaitForResource(
|
||||
const Rc<DxvkResource>& Resource,
|
||||
UINT MapFlags);
|
||||
|
||||
void EmitCsChunk(Rc<DxvkCsChunk>&& chunk) final;
|
||||
|
||||
};
|
||||
|
@ -22,7 +22,6 @@ namespace dxvk {
|
||||
return DxgiFormatMode::Any;
|
||||
}
|
||||
|
||||
|
||||
/**
|
||||
* \brief Optimizes image layout based on usage flags
|
||||
*
|
||||
@ -144,8 +143,10 @@ namespace dxvk {
|
||||
if (CPUAccessFlags != 0) {
|
||||
pImageInfo->stages |= VK_PIPELINE_STAGE_HOST_BIT;
|
||||
|
||||
if (CPUAccessFlags & D3D11_CPU_ACCESS_WRITE)
|
||||
if (CPUAccessFlags & D3D11_CPU_ACCESS_WRITE) {
|
||||
pImageInfo->access |= VK_ACCESS_HOST_WRITE_BIT;
|
||||
pImageInfo->tiling = VK_IMAGE_TILING_LINEAR;
|
||||
}
|
||||
|
||||
if (CPUAccessFlags & D3D11_CPU_ACCESS_READ)
|
||||
pImageInfo->access |= VK_ACCESS_HOST_READ_BIT;
|
||||
@ -159,6 +160,29 @@ namespace dxvk {
|
||||
}
|
||||
|
||||
|
||||
/**
|
||||
* \brief Retrieves memory flags for image usage
|
||||
*
|
||||
* If the host requires access to the image, we
|
||||
* should create it on a host-visible memory type.
|
||||
* \param [in] Usage Image usage flags
|
||||
* \returns Image memory properties
|
||||
*/
|
||||
static VkMemoryPropertyFlags GetImageMemoryFlags(UINT CPUAccessFlags) {
|
||||
if (CPUAccessFlags & D3D11_CPU_ACCESS_READ) {
|
||||
return VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
||||
| VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
|
||||
| VK_MEMORY_PROPERTY_HOST_CACHED_BIT;
|
||||
} else {
|
||||
// If only write access is required, we will emulate
|
||||
// image mapping through a buffer. Some games ignore
|
||||
// the row pitch when mapping images, which leads to
|
||||
// incorrect rendering.
|
||||
return VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT;
|
||||
}
|
||||
}
|
||||
|
||||
|
||||
D3D11Texture1D::D3D11Texture1D(
|
||||
D3D11Device* pDevice,
|
||||
const D3D11_TEXTURE1D_DESC* pDesc)
|
||||
@ -201,7 +225,7 @@ namespace dxvk {
|
||||
// Create the image and, if necessary, the image buffer
|
||||
m_texInfo.formatMode = formatMode;
|
||||
m_texInfo.image = pDevice->GetDXVKDevice()->createImage(
|
||||
info, VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT);
|
||||
info, GetImageMemoryFlags(m_desc.CPUAccessFlags));
|
||||
m_texInfo.imageBuffer = m_desc.CPUAccessFlags != 0
|
||||
? CreateImageBuffer(pDevice->GetDXVKDevice(), info.format, info.extent)
|
||||
: nullptr;
|
||||
@ -301,7 +325,7 @@ namespace dxvk {
|
||||
// Create the image and, if necessary, the image buffer
|
||||
m_texInfo.formatMode = formatMode;
|
||||
m_texInfo.image = pDevice->GetDXVKDevice()->createImage(
|
||||
info, VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT);
|
||||
info, GetImageMemoryFlags(m_desc.CPUAccessFlags));
|
||||
m_texInfo.imageBuffer = m_desc.CPUAccessFlags != 0
|
||||
? CreateImageBuffer(pDevice->GetDXVKDevice(), info.format, info.extent)
|
||||
: nullptr;
|
||||
@ -354,7 +378,7 @@ namespace dxvk {
|
||||
|
||||
|
||||
///////////////////////////////////////////
|
||||
// D 3 D 1 1 T E X T U R E 2 D
|
||||
// D 3 D 1 1 T E X T U R E 3 D
|
||||
D3D11Texture3D::D3D11Texture3D(
|
||||
D3D11Device* pDevice,
|
||||
const D3D11_TEXTURE3D_DESC* pDesc)
|
||||
@ -398,7 +422,7 @@ namespace dxvk {
|
||||
// Create the image and, if necessary, the image buffer
|
||||
m_texInfo.formatMode = formatMode;
|
||||
m_texInfo.image = pDevice->GetDXVKDevice()->createImage(
|
||||
info, VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT);
|
||||
info, GetImageMemoryFlags(m_desc.CPUAccessFlags));
|
||||
m_texInfo.imageBuffer = m_desc.CPUAccessFlags != 0
|
||||
? CreateImageBuffer(pDevice->GetDXVKDevice(), info.format, info.extent)
|
||||
: nullptr;
|
||||
|
@ -28,7 +28,6 @@ namespace dxvk {
|
||||
};
|
||||
|
||||
|
||||
|
||||
///////////////////////////////////////////
|
||||
// D 3 D 1 1 T E X T U R E 1 D
|
||||
class D3D11Texture1D : public D3D11DeviceChild<ID3D11Texture1D> {
|
||||
|
Loading…
x
Reference in New Issue
Block a user