rooty/dxvk
1
0
Fork 0
mirror of https://github.com/doitsujin/dxvk.git synced 2025-01-05 19:46:15 +01:00
Code Issues Projects Releases Wiki Activity

[d3d9] Rework UP buffer allocation

Browse Source
This commit is contained in:
Philip Rebohle 2022-07-14 16:15:59 +02:00 committed by Philip Rebohle
parent e884413c49
commit f07a6e160c
2 changed files with 66 additions and 35 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

93
src/d3d9/d3d9_device.cpp
View File

@ -2489,7 +2489,7 @@ namespace dxvk {
const uint32_t dataSize = GetUPDataSize(drawInfo.vertexCount, VertexStreamZeroStride); const uint32_t dataSize = GetUPDataSize(drawInfo.vertexCount, VertexStreamZeroStride);
const uint32_t bufferSize = GetUPBufferSize(drawInfo.vertexCount, VertexStreamZeroStride); const uint32_t bufferSize = GetUPBufferSize(drawInfo.vertexCount, VertexStreamZeroStride);
auto upSlice = AllocTempBuffer<true>(bufferSize); auto upSlice = AllocUPBuffer(bufferSize);
FillUPVertexBuffer(upSlice.mapPtr, pVertexStreamZeroData, dataSize, bufferSize); FillUPVertexBuffer(upSlice.mapPtr, pVertexStreamZeroData, dataSize, bufferSize);
EmitCs([this, EmitCs([this,
@ -2547,7 +2547,7 @@ namespace dxvk {
const uint32_t upSize = vertexBufferSize + indicesSize; const uint32_t upSize = vertexBufferSize + indicesSize;
auto upSlice = AllocTempBuffer<true>(upSize); auto upSlice = AllocUPBuffer(upSize);
uint8_t* data = reinterpret_cast<uint8_t*>(upSlice.mapPtr); uint8_t* data = reinterpret_cast<uint8_t*>(upSlice.mapPtr);
FillUPVertexBuffer(data, pVertexStreamZeroData, vertexDataSize, vertexBufferSize); FillUPVertexBuffer(data, pVertexStreamZeroData, vertexDataSize, vertexBufferSize);
std::memcpy(data + vertexBufferSize, pIndexData, indicesSize); std::memcpy(data + vertexBufferSize, pIndexData, indicesSize);
@ -3967,7 +3967,52 @@ namespace dxvk {
} }
template<bool UpBuffer> D3D9BufferSlice D3D9DeviceEx::AllocUPBuffer(VkDeviceSize size) {
constexpr VkDeviceSize UPBufferSize = 1 << 20;
if (unlikely(m_upBuffer == nullptr)) {
VkMemoryPropertyFlags memoryFlags
= VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
| VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
| VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT;
DxvkBufferCreateInfo info;
info.size = UPBufferSize;
info.usage = VK_BUFFER_USAGE_VERTEX_BUFFER_BIT
| VK_BUFFER_USAGE_INDEX_BUFFER_BIT;
info.access = VK_ACCESS_VERTEX_ATTRIBUTE_READ_BIT
| VK_ACCESS_INDEX_READ_BIT;
info.stages = VK_PIPELINE_STAGE_VERTEX_INPUT_BIT;
m_upBuffer = m_dxvkDevice->createBuffer(info, memoryFlags);
m_upBufferMapPtr = m_upBuffer->mapPtr(0);
}
VkDeviceSize alignedSize = align(size, CACHE_LINE_SIZE);
if (unlikely(m_upBufferOffset + alignedSize > UPBufferSize)) {
auto sliceHandle = m_upBuffer->allocSlice();
m_upBufferOffset = 0;
m_upBufferMapPtr = sliceHandle.mapPtr;
EmitCs([
cBuffer = m_upBuffer,
cSlice = sliceHandle
] (DxvkContext* ctx) {
ctx->invalidateBuffer(cBuffer, cSlice);
});
}
D3D9BufferSlice result;
result.slice = DxvkBufferSlice(m_upBuffer, m_upBufferOffset, size);
result.mapPtr = reinterpret_cast<char*>(m_upBufferMapPtr) + m_upBufferOffset;
m_upBufferOffset += alignedSize;
return result;
}
D3D9BufferSlice D3D9DeviceEx::AllocTempBuffer(VkDeviceSize size) { D3D9BufferSlice D3D9DeviceEx::AllocTempBuffer(VkDeviceSize size) {
constexpr VkDeviceSize DefaultSize = 1 << 20; constexpr VkDeviceSize DefaultSize = 1 << 20;
@ -3975,27 +4020,15 @@ namespace dxvk {
= VK_MEMORY_PROPERTY_HOST_COHERENT_BIT = VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
| VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT; | VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT;
if constexpr (UpBuffer) { D3D9BufferSlice& currentSlice = m_managedUploadBuffer;
memoryFlags |= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT;
}
D3D9BufferSlice& currentSlice = UpBuffer ? m_upBuffer : m_managedUploadBuffer;
if (size <= DefaultSize) { if (size <= DefaultSize) {
if (unlikely(!currentSlice.slice.defined())) { if (unlikely(!currentSlice.slice.defined())) {
DxvkBufferCreateInfo info; DxvkBufferCreateInfo info;
info.size = DefaultSize; info.size = DefaultSize;
if constexpr (UpBuffer) { info.usage = VK_BUFFER_USAGE_TRANSFER_SRC_BIT | VK_BUFFER_USAGE_UNIFORM_TEXEL_BUFFER_BIT;
info.usage = VK_BUFFER_USAGE_VERTEX_BUFFER_BIT info.stages = VK_PIPELINE_STAGE_TRANSFER_BIT | VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT;
| VK_BUFFER_USAGE_INDEX_BUFFER_BIT; info.access = VK_ACCESS_TRANSFER_READ_BIT | VK_ACCESS_SHADER_READ_BIT;
info.access = VK_ACCESS_VERTEX_ATTRIBUTE_READ_BIT
| VK_ACCESS_INDEX_READ_BIT;
info.stages = VK_PIPELINE_STAGE_VERTEX_INPUT_BIT;
} else {
info.usage = VK_BUFFER_USAGE_TRANSFER_SRC_BIT | VK_BUFFER_USAGE_UNIFORM_TEXEL_BUFFER_BIT;
info.stages = VK_PIPELINE_STAGE_TRANSFER_BIT | VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT;
info.access = VK_ACCESS_TRANSFER_READ_BIT | VK_ACCESS_SHADER_READ_BIT;
}
currentSlice.slice = DxvkBufferSlice(m_dxvkDevice->createBuffer(info, memoryFlags)); currentSlice.slice = DxvkBufferSlice(m_dxvkDevice->createBuffer(info, memoryFlags));
currentSlice.mapPtr = currentSlice.slice.mapPtr(0); currentSlice.mapPtr = currentSlice.slice.mapPtr(0);
@ -4024,17 +4057,11 @@ namespace dxvk {
// Create a temporary buffer for very large allocations // Create a temporary buffer for very large allocations
DxvkBufferCreateInfo info; DxvkBufferCreateInfo info;
info.size = size; info.size = size;
if constexpr (UpBuffer) { info.usage = VK_BUFFER_USAGE_VERTEX_BUFFER_BIT
info.usage = VK_BUFFER_USAGE_VERTEX_BUFFER_BIT | VK_BUFFER_USAGE_INDEX_BUFFER_BIT;
| VK_BUFFER_USAGE_INDEX_BUFFER_BIT; info.access = VK_ACCESS_VERTEX_ATTRIBUTE_READ_BIT
info.access = VK_ACCESS_VERTEX_ATTRIBUTE_READ_BIT | VK_ACCESS_INDEX_READ_BIT;
| VK_ACCESS_INDEX_READ_BIT; info.stages = VK_PIPELINE_STAGE_VERTEX_INPUT_BIT;
info.stages = VK_PIPELINE_STAGE_VERTEX_INPUT_BIT;
} else {
info.usage = VK_BUFFER_USAGE_TRANSFER_SRC_BIT | VK_BUFFER_USAGE_UNIFORM_TEXEL_BUFFER_BIT;
info.stages = VK_PIPELINE_STAGE_TRANSFER_BIT | VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT;
info.access = VK_ACCESS_TRANSFER_READ_BIT | VK_ACCESS_SHADER_READ_BIT;
}
D3D9BufferSlice result; D3D9BufferSlice result;
result.slice = DxvkBufferSlice(m_dxvkDevice->createBuffer(info, memoryFlags)); result.slice = DxvkBufferSlice(m_dxvkDevice->createBuffer(info, memoryFlags));
@ -4502,7 +4529,7 @@ namespace dxvk {
DxvkBufferSlice copySrcSlice; DxvkBufferSlice copySrcSlice;
if (pSrcTexture->DoesStagingBufferUploads(SrcSubresource)) { if (pSrcTexture->DoesStagingBufferUploads(SrcSubresource)) {
VkDeviceSize dirtySize = extentBlockCount.width * extentBlockCount.height * extentBlockCount.depth * formatInfo->elementSize; VkDeviceSize dirtySize = extentBlockCount.width * extentBlockCount.height * extentBlockCount.depth * formatInfo->elementSize;
D3D9BufferSlice slice = AllocTempBuffer<false>(dirtySize); D3D9BufferSlice slice = AllocTempBuffer(dirtySize);
copySrcSlice = slice.slice; copySrcSlice = slice.slice;
void* srcData = reinterpret_cast<uint8_t*>(srcSlice.mapPtr) + copySrcOffset; void* srcData = reinterpret_cast<uint8_t*>(srcSlice.mapPtr) + copySrcOffset;
util::packImageData( util::packImageData(
@ -4552,7 +4579,7 @@ namespace dxvk {
} }
// the converter can not handle the 4 aligned pitch so we always repack into a staging buffer // the converter can not handle the 4 aligned pitch so we always repack into a staging buffer
D3D9BufferSlice slice = AllocTempBuffer<false>(srcSlice.length); D3D9BufferSlice slice = AllocTempBuffer(srcSlice.length);
VkDeviceSize pitch = align(srcTexLevelExtentBlockCount.width * formatInfo->elementSize, 4); VkDeviceSize pitch = align(srcTexLevelExtentBlockCount.width * formatInfo->elementSize, 4);
util::packImageData( util::packImageData(
@ -4712,7 +4739,7 @@ namespace dxvk {
DxvkBufferSlice copySrcSlice; DxvkBufferSlice copySrcSlice;
if (pResource->DoesStagingBufferUploads()) { if (pResource->DoesStagingBufferUploads()) {
D3D9BufferSlice slice = AllocTempBuffer<false>(range.max - range.min); D3D9BufferSlice slice = AllocTempBuffer(range.max - range.min);
copySrcSlice = slice.slice; copySrcSlice = slice.slice;
void* srcData = reinterpret_cast<uint8_t*>(srcSlice.mapPtr) + range.min; void* srcData = reinterpret_cast<uint8_t*>(srcSlice.mapPtr) + range.min;
memcpy(slice.mapPtr, srcData, range.max - range.min); memcpy(slice.mapPtr, srcData, range.max - range.min);

8
src/d3d9/d3d9_device.h
View File

@ -969,7 +969,8 @@ namespace dxvk {
void DetermineConstantLayouts(bool canSWVP); void DetermineConstantLayouts(bool canSWVP);
template<bool UpBuffer> D3D9BufferSlice AllocUPBuffer(VkDeviceSize size);
D3D9BufferSlice AllocTempBuffer(VkDeviceSize size); D3D9BufferSlice AllocTempBuffer(VkDeviceSize size);
bool ShouldRecord(); bool ShouldRecord();
@ -1163,7 +1164,10 @@ namespace dxvk {
D3D9ConstantBuffer m_psFixedFunction; D3D9ConstantBuffer m_psFixedFunction;
D3D9ConstantBuffer m_psShared; D3D9ConstantBuffer m_psShared;
D3D9BufferSlice m_upBuffer; Rc<DxvkBuffer> m_upBuffer;
VkDeviceSize m_upBufferOffset = 0ull;
void* m_upBufferMapPtr = nullptr;
D3D9BufferSlice m_managedUploadBuffer; D3D9BufferSlice m_managedUploadBuffer;
D3D9Cursor m_cursor; D3D9Cursor m_cursor;