hwpplayer1/suyu
1
0
Fork 0
forked from suyu/suyu
Code Pull requests Activity

vk_stream_buffer: Remove copy code path

Browse source
This commit is contained in:
ReinUsesLisp 2019-02-26 02:09:43 -03:00
parent 33a0597603
commit 730eb1dad7
2 changed files with 18 additions and 53 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

52
src/video_core/renderer_vulkan/vk_stream_buffer.cpp
View file

@ -23,16 +23,15 @@ constexpr u64 WATCHES_RESERVE_CHUNK = 0x1000;
VKStreamBuffer::VKStreamBuffer(const VKDevice& device, VKMemoryManager& memory_manager,
VKScheduler& scheduler, u64 size, vk::BufferUsageFlags usage,
vk::AccessFlags access, vk::PipelineStageFlags pipeline_stage)
: device{device}, scheduler{scheduler},
has_device_exclusive_memory{!memory_manager.IsMemoryUnified()},
buffer_size{size}, access{access}, pipeline_stage{pipeline_stage} {
: device{device}, scheduler{scheduler}, buffer_size{size}, access{access}, pipeline_stage{
pipeline_stage} {
CreateBuffers(memory_manager, usage);
ReserveWatches(WATCHES_INITIAL_RESERVE);
}
VKStreamBuffer::~VKStreamBuffer() = default;
std::tuple<u8*, u64, vk::Buffer, bool> VKStreamBuffer::Reserve(u64 size, bool keep_in_host) {
std::tuple<u8*, u64, bool> VKStreamBuffer::Reserve(u64 size) {
ASSERT(size <= buffer_size);
mapped_size = size;
@ -44,10 +43,7 @@ std::tuple<u8*, u64, vk::Buffer, bool> VKStreamBuffer::Reserve(u64 size, bool ke
offset = 0;
}
use_device = has_device_exclusive_memory && !keep_in_host;
const vk::Buffer buffer = use_device ? *device_buffer : *mappable_buffer;
return {mapped_pointer + offset, offset, buffer, invalidation_mark.has_value()};
return {mapped_pointer + offset, offset, invalidation_mark.has_value()};
}
VKExecutionContext VKStreamBuffer::Send(VKExecutionContext exctx, u64 size) {
@ -61,24 +57,6 @@ VKExecutionContext VKStreamBuffer::Send(VKExecutionContext exctx, u64 size) {
invalidation_mark = std::nullopt;
}
// Only copy to VRAM when requested.
if (use_device) {
const auto& dld = device.GetDispatchLoader();
const u32 graphics_family = device.GetGraphicsFamily();
const auto cmdbuf = exctx.GetCommandBuffer();
// Buffers are mirrored, that's why the copy is done with the same offset on both buffers.
const vk::BufferCopy copy_region(offset, offset, size);
cmdbuf.copyBuffer(*mappable_buffer, *device_buffer, {copy_region}, dld);
// Protect the buffer from GPU usage until the copy has finished.
const vk::BufferMemoryBarrier barrier(vk::AccessFlagBits::eTransferWrite, access,
graphics_family, graphics_family, *device_buffer,
offset, size);
cmdbuf.pipelineBarrier(vk::PipelineStageFlagBits::eTransfer, pipeline_stage, {}, {},
{barrier}, {}, dld);
}
if (used_watches + 1 >= watches.size()) {
// Ensure that there are enough watches.
ReserveWatches(WATCHES_RESERVE_CHUNK);
@ -92,26 +70,14 @@ VKExecutionContext VKStreamBuffer::Send(VKExecutionContext exctx, u64 size) {
}
void VKStreamBuffer::CreateBuffers(VKMemoryManager& memory_manager, vk::BufferUsageFlags usage) {
vk::BufferUsageFlags mappable_usage = usage;
if (has_device_exclusive_memory) {
mappable_usage |= vk::BufferUsageFlagBits::eTransferSrc;
}
const vk::BufferCreateInfo buffer_ci({}, buffer_size, mappable_usage,
vk::SharingMode::eExclusive, 0, nullptr);
const vk::BufferCreateInfo buffer_ci({}, buffer_size, usage, vk::SharingMode::eExclusive, 0,
nullptr);
const auto dev = device.GetLogical();
const auto& dld = device.GetDispatchLoader();
mappable_buffer = dev.createBufferUnique(buffer_ci, nullptr, dld);
mappable_commit = memory_manager.Commit(*mappable_buffer, true);
mapped_pointer = mappable_commit->GetData();
if (has_device_exclusive_memory) {
const vk::BufferCreateInfo buffer_ci({}, buffer_size,
usage | vk::BufferUsageFlagBits::eTransferDst,
vk::SharingMode::eExclusive, 0, nullptr);
device_buffer = dev.createBufferUnique(buffer_ci, nullptr, dld);
device_commit = memory_manager.Commit(*device_buffer, false);
}
buffer = dev.createBufferUnique(buffer_ci, nullptr, dld);
commit = memory_manager.Commit(*buffer, true);
mapped_pointer = commit->GetData();
}
void VKStreamBuffer::ReserveWatches(std::size_t grow_size) {

19
src/video_core/renderer_vulkan/vk_stream_buffer.h
View file

@ -31,15 +31,18 @@ public:
/**
* Reserves a region of memory from the stream buffer.
* @param size Size to reserve.
* @param keep_in_host Mapped buffer will be in host memory, skipping the copy to device local.
* @returns A tuple in the following order: Raw memory pointer (with offset added), buffer
* offset, Vulkan buffer handle, buffer has been invalited.
* offset and a boolean that's true when buffer has been invalidated.
*/
std::tuple<u8*, u64, vk::Buffer, bool> Reserve(u64 size, bool keep_in_host);
std::tuple<u8*, u64, bool> Reserve(u64 size);
/// Ensures that "size" bytes of memory are available to the GPU, potentially recording a copy.
[[nodiscard]] VKExecutionContext Send(VKExecutionContext exctx, u64 size);
vk::Buffer GetBuffer() const {
return *buffer;
}
private:
/// Creates Vulkan buffer handles committing the required the required memory.
void CreateBuffers(VKMemoryManager& memory_manager, vk::BufferUsageFlags usage);
@ -50,19 +53,15 @@ private:
const VKDevice& device; ///< Vulkan device manager.
VKScheduler& scheduler; ///< Command scheduler.
const u64 buffer_size; ///< Total size of the stream buffer.
const bool has_device_exclusive_memory; ///< True if the streaming buffer will use VRAM.
const vk::AccessFlags access; ///< Access usage of this stream buffer.
const vk::PipelineStageFlags pipeline_stage; ///< Pipeline usage of this stream buffer.
UniqueBuffer mappable_buffer; ///< Mapped buffer.
UniqueBuffer device_buffer; ///< Buffer exclusive to the GPU.
VKMemoryCommit mappable_commit; ///< Commit visible from the CPU.
VKMemoryCommit device_commit; ///< Commit stored in VRAM.
u8* mapped_pointer{}; ///< Pointer to the host visible commit
UniqueBuffer buffer; ///< Mapped buffer.
VKMemoryCommit commit; ///< Memory commit.
u8* mapped_pointer{}; ///< Pointer to the host visible commit
u64 offset{}; ///< Buffer iterator.
u64 mapped_size{}; ///< Size reserved for the current copy.
bool use_device{}; ///< True if the current uses VRAM.
std::vector<std::unique_ptr<VKFenceWatch>> watches; ///< Total watches
std::size_t used_watches{}; ///< Count of watches, reset on invalidation.