Linuxydable/suyu
1
0
Fork 0
forked from suyu/suyu
Code Pull requests Activity
suyu/src/video_core/renderer_vulkan/vk_stream_buffer.cpp

147 lines
5 KiB
C++
Raw Normal View History

vk_stream_buffer: Implement a stream buffer This manages two kinds of streaming buffers: one for unified memory models and one for dedicated GPUs. The first one skips the copy from the staging buffer to the real buffer, since it creates an unified buffer. This implementation waits for all fences to finish their operation before "invalidating". This is suboptimal since it should allocate another buffer or start searching from the beginning. There is room for improvement here. This could also handle AMD's "pinned" memory (a heap with 256 MiB) that seems to be designed for buffer streaming.
2019-02-24 08:22:33 +01:00
// Copyright 2019 yuzu Emulator Project
// Licensed under GPLv2 or any later version
// Refer to the license.txt file included.
#include <algorithm>
#include <optional>
vk_stream_buffer/vk_buffer_cache: Avoid halting and use generic cache The stream buffer before this commit once it was full (no more bytes to write before looping) waiting for all previous operations to finish. This was a temporary solution and had a noticeable performance penalty in performance (from what a profiler showed). To avoid this mark with fences usages of the stream buffer and once it loops wait for them to be signaled. On average this will never wait. Each fence knows where its usage finishes, resulting in a non-paged stream buffer. On the other side, the buffer cache is reimplemented using the generic buffer cache. It makes use of the staging buffer pool and the new stream buffer.
2020-01-06 21:59:20 +01:00
#include <tuple>
vk_stream_buffer: Implement a stream buffer This manages two kinds of streaming buffers: one for unified memory models and one for dedicated GPUs. The first one skips the copy from the staging buffer to the real buffer, since it creates an unified buffer. This implementation waits for all fences to finish their operation before "invalidating". This is suboptimal since it should allocate another buffer or start searching from the beginning. There is room for improvement here. This could also handle AMD's "pinned" memory (a heap with 256 MiB) that seems to be designed for buffer streaming.
2019-02-24 08:22:33 +01:00
#include <vector>
vk_stream_buffer/vk_buffer_cache: Avoid halting and use generic cache The stream buffer before this commit once it was full (no more bytes to write before looping) waiting for all previous operations to finish. This was a temporary solution and had a noticeable performance penalty in performance (from what a profiler showed). To avoid this mark with fences usages of the stream buffer and once it loops wait for them to be signaled. On average this will never wait. Each fence knows where its usage finishes, resulting in a non-paged stream buffer. On the other side, the buffer cache is reimplemented using the generic buffer cache. It makes use of the staging buffer pool and the new stream buffer.
2020-01-06 21:59:20 +01:00
#include "common/alignment.h"
vk_stream_buffer: Implement a stream buffer This manages two kinds of streaming buffers: one for unified memory models and one for dedicated GPUs. The first one skips the copy from the staging buffer to the real buffer, since it creates an unified buffer. This implementation waits for all fences to finish their operation before "invalidating". This is suboptimal since it should allocate another buffer or start searching from the beginning. There is room for improvement here. This could also handle AMD's "pinned" memory (a heap with 256 MiB) that seems to be designed for buffer streaming.
2019-02-24 08:22:33 +01:00
#include "common/assert.h"
#include "video_core/renderer_vulkan/declarations.h"
#include "video_core/renderer_vulkan/vk_device.h"
#include "video_core/renderer_vulkan/vk_resource_manager.h"
#include "video_core/renderer_vulkan/vk_scheduler.h"
#include "video_core/renderer_vulkan/vk_stream_buffer.h"
namespace Vulkan {
vk_stream_buffer/vk_buffer_cache: Avoid halting and use generic cache The stream buffer before this commit once it was full (no more bytes to write before looping) waiting for all previous operations to finish. This was a temporary solution and had a noticeable performance penalty in performance (from what a profiler showed). To avoid this mark with fences usages of the stream buffer and once it loops wait for them to be signaled. On average this will never wait. Each fence knows where its usage finishes, resulting in a non-paged stream buffer. On the other side, the buffer cache is reimplemented using the generic buffer cache. It makes use of the staging buffer pool and the new stream buffer.
2020-01-06 21:59:20 +01:00
namespace {
vk_stream_buffer: Implement a stream buffer This manages two kinds of streaming buffers: one for unified memory models and one for dedicated GPUs. The first one skips the copy from the staging buffer to the real buffer, since it creates an unified buffer. This implementation waits for all fences to finish their operation before "invalidating". This is suboptimal since it should allocate another buffer or start searching from the beginning. There is room for improvement here. This could also handle AMD's "pinned" memory (a heap with 256 MiB) that seems to be designed for buffer streaming.
2019-02-24 08:22:33 +01:00
constexpr u64 WATCHES_INITIAL_RESERVE = 0x4000;
constexpr u64 WATCHES_RESERVE_CHUNK = 0x1000;
vk_stream_buffer/vk_buffer_cache: Avoid halting and use generic cache The stream buffer before this commit once it was full (no more bytes to write before looping) waiting for all previous operations to finish. This was a temporary solution and had a noticeable performance penalty in performance (from what a profiler showed). To avoid this mark with fences usages of the stream buffer and once it loops wait for them to be signaled. On average this will never wait. Each fence knows where its usage finishes, resulting in a non-paged stream buffer. On the other side, the buffer cache is reimplemented using the generic buffer cache. It makes use of the staging buffer pool and the new stream buffer.
2020-01-06 21:59:20 +01:00
constexpr u64 STREAM_BUFFER_SIZE = 256 * 1024 * 1024;
std::optional<u32> FindMemoryType(const VKDevice& device, u32 filter,
vk::MemoryPropertyFlags wanted) {
const auto properties = device.GetPhysical().getMemoryProperties(device.GetDispatchLoader());
for (u32 i = 0; i < properties.memoryTypeCount; i++) {
if (!(filter & (1 << i))) {
continue;
}
if ((properties.memoryTypes[i].propertyFlags & wanted) == wanted) {
return i;
}
}
return {};
}
} // Anonymous namespace
VKStreamBuffer::VKStreamBuffer(const VKDevice& device, VKScheduler& scheduler,
vk::BufferUsageFlags usage)
: device{device}, scheduler{scheduler} {
CreateBuffers(usage);
ReserveWatches(current_watches, WATCHES_INITIAL_RESERVE);
ReserveWatches(previous_watches, WATCHES_INITIAL_RESERVE);
vk_stream_buffer: Implement a stream buffer This manages two kinds of streaming buffers: one for unified memory models and one for dedicated GPUs. The first one skips the copy from the staging buffer to the real buffer, since it creates an unified buffer. This implementation waits for all fences to finish their operation before "invalidating". This is suboptimal since it should allocate another buffer or start searching from the beginning. There is room for improvement here. This could also handle AMD's "pinned" memory (a heap with 256 MiB) that seems to be designed for buffer streaming.
2019-02-24 08:22:33 +01:00
}
VKStreamBuffer::~VKStreamBuffer() = default;
vk_stream_buffer/vk_buffer_cache: Avoid halting and use generic cache The stream buffer before this commit once it was full (no more bytes to write before looping) waiting for all previous operations to finish. This was a temporary solution and had a noticeable performance penalty in performance (from what a profiler showed). To avoid this mark with fences usages of the stream buffer and once it loops wait for them to be signaled. On average this will never wait. Each fence knows where its usage finishes, resulting in a non-paged stream buffer. On the other side, the buffer cache is reimplemented using the generic buffer cache. It makes use of the staging buffer pool and the new stream buffer.
2020-01-06 21:59:20 +01:00
std::tuple<u8*, u64, bool> VKStreamBuffer::Map(u64 size, u64 alignment) {
ASSERT(size <= STREAM_BUFFER_SIZE);
vk_stream_buffer: Implement a stream buffer This manages two kinds of streaming buffers: one for unified memory models and one for dedicated GPUs. The first one skips the copy from the staging buffer to the real buffer, since it creates an unified buffer. This implementation waits for all fences to finish their operation before "invalidating". This is suboptimal since it should allocate another buffer or start searching from the beginning. There is room for improvement here. This could also handle AMD's "pinned" memory (a heap with 256 MiB) that seems to be designed for buffer streaming.
2019-02-24 08:22:33 +01:00
mapped_size = size;
vk_stream_buffer/vk_buffer_cache: Avoid halting and use generic cache The stream buffer before this commit once it was full (no more bytes to write before looping) waiting for all previous operations to finish. This was a temporary solution and had a noticeable performance penalty in performance (from what a profiler showed). To avoid this mark with fences usages of the stream buffer and once it loops wait for them to be signaled. On average this will never wait. Each fence knows where its usage finishes, resulting in a non-paged stream buffer. On the other side, the buffer cache is reimplemented using the generic buffer cache. It makes use of the staging buffer pool and the new stream buffer.
2020-01-06 21:59:20 +01:00
if (alignment > 0) {
offset = Common::AlignUp(offset, alignment);
}
WaitPendingOperations(offset);
bool invalidated = false;
if (offset + size > STREAM_BUFFER_SIZE) {
// The buffer would overflow, save the amount of used watches and reset the state.
invalidation_mark = current_watch_cursor;
current_watch_cursor = 0;
vk_stream_buffer: Implement a stream buffer This manages two kinds of streaming buffers: one for unified memory models and one for dedicated GPUs. The first one skips the copy from the staging buffer to the real buffer, since it creates an unified buffer. This implementation waits for all fences to finish their operation before "invalidating". This is suboptimal since it should allocate another buffer or start searching from the beginning. There is room for improvement here. This could also handle AMD's "pinned" memory (a heap with 256 MiB) that seems to be designed for buffer streaming.
2019-02-24 08:22:33 +01:00
offset = 0;
vk_stream_buffer/vk_buffer_cache: Avoid halting and use generic cache The stream buffer before this commit once it was full (no more bytes to write before looping) waiting for all previous operations to finish. This was a temporary solution and had a noticeable performance penalty in performance (from what a profiler showed). To avoid this mark with fences usages of the stream buffer and once it loops wait for them to be signaled. On average this will never wait. Each fence knows where its usage finishes, resulting in a non-paged stream buffer. On the other side, the buffer cache is reimplemented using the generic buffer cache. It makes use of the staging buffer pool and the new stream buffer.
2020-01-06 21:59:20 +01:00
// Swap watches and reset waiting cursors.
std::swap(previous_watches, current_watches);
wait_cursor = 0;
wait_bound = 0;
// Ensure that we don't wait for uncommitted fences.
scheduler.Flush();
invalidated = true;
vk_stream_buffer: Implement a stream buffer This manages two kinds of streaming buffers: one for unified memory models and one for dedicated GPUs. The first one skips the copy from the staging buffer to the real buffer, since it creates an unified buffer. This implementation waits for all fences to finish their operation before "invalidating". This is suboptimal since it should allocate another buffer or start searching from the beginning. There is room for improvement here. This could also handle AMD's "pinned" memory (a heap with 256 MiB) that seems to be designed for buffer streaming.
2019-02-24 08:22:33 +01:00
}
vk_stream_buffer/vk_buffer_cache: Avoid halting and use generic cache The stream buffer before this commit once it was full (no more bytes to write before looping) waiting for all previous operations to finish. This was a temporary solution and had a noticeable performance penalty in performance (from what a profiler showed). To avoid this mark with fences usages of the stream buffer and once it loops wait for them to be signaled. On average this will never wait. Each fence knows where its usage finishes, resulting in a non-paged stream buffer. On the other side, the buffer cache is reimplemented using the generic buffer cache. It makes use of the staging buffer pool and the new stream buffer.
2020-01-06 21:59:20 +01:00
const auto dev = device.GetLogical();
const auto& dld = device.GetDispatchLoader();
const auto pointer = reinterpret_cast<u8*>(dev.mapMemory(*memory, offset, size, {}, dld));
return {pointer, offset, invalidated};
vk_stream_buffer: Implement a stream buffer This manages two kinds of streaming buffers: one for unified memory models and one for dedicated GPUs. The first one skips the copy from the staging buffer to the real buffer, since it creates an unified buffer. This implementation waits for all fences to finish their operation before "invalidating". This is suboptimal since it should allocate another buffer or start searching from the beginning. There is room for improvement here. This could also handle AMD's "pinned" memory (a heap with 256 MiB) that seems to be designed for buffer streaming.
2019-02-24 08:22:33 +01:00
}
vk_stream_buffer/vk_buffer_cache: Avoid halting and use generic cache The stream buffer before this commit once it was full (no more bytes to write before looping) waiting for all previous operations to finish. This was a temporary solution and had a noticeable performance penalty in performance (from what a profiler showed). To avoid this mark with fences usages of the stream buffer and once it loops wait for them to be signaled. On average this will never wait. Each fence knows where its usage finishes, resulting in a non-paged stream buffer. On the other side, the buffer cache is reimplemented using the generic buffer cache. It makes use of the staging buffer pool and the new stream buffer.
2020-01-06 21:59:20 +01:00
void VKStreamBuffer::Unmap(u64 size) {
vk_stream_buffer: Implement a stream buffer This manages two kinds of streaming buffers: one for unified memory models and one for dedicated GPUs. The first one skips the copy from the staging buffer to the real buffer, since it creates an unified buffer. This implementation waits for all fences to finish their operation before "invalidating". This is suboptimal since it should allocate another buffer or start searching from the beginning. There is room for improvement here. This could also handle AMD's "pinned" memory (a heap with 256 MiB) that seems to be designed for buffer streaming.
2019-02-24 08:22:33 +01:00
ASSERT_MSG(size <= mapped_size, "Reserved size is too small");
vk_stream_buffer/vk_buffer_cache: Avoid halting and use generic cache The stream buffer before this commit once it was full (no more bytes to write before looping) waiting for all previous operations to finish. This was a temporary solution and had a noticeable performance penalty in performance (from what a profiler showed). To avoid this mark with fences usages of the stream buffer and once it loops wait for them to be signaled. On average this will never wait. Each fence knows where its usage finishes, resulting in a non-paged stream buffer. On the other side, the buffer cache is reimplemented using the generic buffer cache. It makes use of the staging buffer pool and the new stream buffer.
2020-01-06 21:59:20 +01:00
const auto dev = device.GetLogical();
dev.unmapMemory(*memory, device.GetDispatchLoader());
offset += size;
vk_stream_buffer: Implement a stream buffer This manages two kinds of streaming buffers: one for unified memory models and one for dedicated GPUs. The first one skips the copy from the staging buffer to the real buffer, since it creates an unified buffer. This implementation waits for all fences to finish their operation before "invalidating". This is suboptimal since it should allocate another buffer or start searching from the beginning. There is room for improvement here. This could also handle AMD's "pinned" memory (a heap with 256 MiB) that seems to be designed for buffer streaming.
2019-02-24 08:22:33 +01:00
vk_stream_buffer/vk_buffer_cache: Avoid halting and use generic cache The stream buffer before this commit once it was full (no more bytes to write before looping) waiting for all previous operations to finish. This was a temporary solution and had a noticeable performance penalty in performance (from what a profiler showed). To avoid this mark with fences usages of the stream buffer and once it loops wait for them to be signaled. On average this will never wait. Each fence knows where its usage finishes, resulting in a non-paged stream buffer. On the other side, the buffer cache is reimplemented using the generic buffer cache. It makes use of the staging buffer pool and the new stream buffer.
2020-01-06 21:59:20 +01:00
if (current_watch_cursor + 1 >= current_watches.size()) {
vk_stream_buffer: Implement a stream buffer This manages two kinds of streaming buffers: one for unified memory models and one for dedicated GPUs. The first one skips the copy from the staging buffer to the real buffer, since it creates an unified buffer. This implementation waits for all fences to finish their operation before "invalidating". This is suboptimal since it should allocate another buffer or start searching from the beginning. There is room for improvement here. This could also handle AMD's "pinned" memory (a heap with 256 MiB) that seems to be designed for buffer streaming.
2019-02-24 08:22:33 +01:00
// Ensure that there are enough watches.
vk_stream_buffer/vk_buffer_cache: Avoid halting and use generic cache The stream buffer before this commit once it was full (no more bytes to write before looping) waiting for all previous operations to finish. This was a temporary solution and had a noticeable performance penalty in performance (from what a profiler showed). To avoid this mark with fences usages of the stream buffer and once it loops wait for them to be signaled. On average this will never wait. Each fence knows where its usage finishes, resulting in a non-paged stream buffer. On the other side, the buffer cache is reimplemented using the generic buffer cache. It makes use of the staging buffer pool and the new stream buffer.
2020-01-06 21:59:20 +01:00
ReserveWatches(current_watches, WATCHES_RESERVE_CHUNK);
vk_stream_buffer: Implement a stream buffer This manages two kinds of streaming buffers: one for unified memory models and one for dedicated GPUs. The first one skips the copy from the staging buffer to the real buffer, since it creates an unified buffer. This implementation waits for all fences to finish their operation before "invalidating". This is suboptimal since it should allocate another buffer or start searching from the beginning. There is room for improvement here. This could also handle AMD's "pinned" memory (a heap with 256 MiB) that seems to be designed for buffer streaming.
2019-02-24 08:22:33 +01:00
}
vk_stream_buffer/vk_buffer_cache: Avoid halting and use generic cache The stream buffer before this commit once it was full (no more bytes to write before looping) waiting for all previous operations to finish. This was a temporary solution and had a noticeable performance penalty in performance (from what a profiler showed). To avoid this mark with fences usages of the stream buffer and once it loops wait for them to be signaled. On average this will never wait. Each fence knows where its usage finishes, resulting in a non-paged stream buffer. On the other side, the buffer cache is reimplemented using the generic buffer cache. It makes use of the staging buffer pool and the new stream buffer.
2020-01-06 21:59:20 +01:00
auto& watch = current_watches[current_watch_cursor++];
watch.upper_bound = offset;
watch.fence.Watch(scheduler.GetFence());
vk_stream_buffer: Implement a stream buffer This manages two kinds of streaming buffers: one for unified memory models and one for dedicated GPUs. The first one skips the copy from the staging buffer to the real buffer, since it creates an unified buffer. This implementation waits for all fences to finish their operation before "invalidating". This is suboptimal since it should allocate another buffer or start searching from the beginning. There is room for improvement here. This could also handle AMD's "pinned" memory (a heap with 256 MiB) that seems to be designed for buffer streaming.
2019-02-24 08:22:33 +01:00
}
vk_stream_buffer/vk_buffer_cache: Avoid halting and use generic cache The stream buffer before this commit once it was full (no more bytes to write before looping) waiting for all previous operations to finish. This was a temporary solution and had a noticeable performance penalty in performance (from what a profiler showed). To avoid this mark with fences usages of the stream buffer and once it loops wait for them to be signaled. On average this will never wait. Each fence knows where its usage finishes, resulting in a non-paged stream buffer. On the other side, the buffer cache is reimplemented using the generic buffer cache. It makes use of the staging buffer pool and the new stream buffer.
2020-01-06 21:59:20 +01:00
void VKStreamBuffer::CreateBuffers(vk::BufferUsageFlags usage) {
const vk::BufferCreateInfo buffer_ci({}, STREAM_BUFFER_SIZE, usage, vk::SharingMode::eExclusive,
0, nullptr);
vk_stream_buffer: Implement a stream buffer This manages two kinds of streaming buffers: one for unified memory models and one for dedicated GPUs. The first one skips the copy from the staging buffer to the real buffer, since it creates an unified buffer. This implementation waits for all fences to finish their operation before "invalidating". This is suboptimal since it should allocate another buffer or start searching from the beginning. There is room for improvement here. This could also handle AMD's "pinned" memory (a heap with 256 MiB) that seems to be designed for buffer streaming.
2019-02-24 08:22:33 +01:00
const auto dev = device.GetLogical();
const auto& dld = device.GetDispatchLoader();
vk_stream_buffer: Remove copy code path
2019-02-26 06:09:43 +01:00
buffer = dev.createBufferUnique(buffer_ci, nullptr, dld);
vk_stream_buffer/vk_buffer_cache: Avoid halting and use generic cache The stream buffer before this commit once it was full (no more bytes to write before looping) waiting for all previous operations to finish. This was a temporary solution and had a noticeable performance penalty in performance (from what a profiler showed). To avoid this mark with fences usages of the stream buffer and once it loops wait for them to be signaled. On average this will never wait. Each fence knows where its usage finishes, resulting in a non-paged stream buffer. On the other side, the buffer cache is reimplemented using the generic buffer cache. It makes use of the staging buffer pool and the new stream buffer.
2020-01-06 21:59:20 +01:00
const auto requirements = dev.getBufferMemoryRequirements(*buffer, dld);
// Prefer device local host visible allocations (this should hit AMD's pinned memory).
auto type = FindMemoryType(device, requirements.memoryTypeBits,
vk::MemoryPropertyFlagBits::eHostVisible |
vk::MemoryPropertyFlagBits::eHostCoherent |
vk::MemoryPropertyFlagBits::eDeviceLocal);
if (!type) {
// Otherwise search for a host visible allocation.
type = FindMemoryType(device, requirements.memoryTypeBits,
vk::MemoryPropertyFlagBits::eHostVisible |
vk::MemoryPropertyFlagBits::eHostCoherent);
ASSERT_MSG(type, "No host visible and coherent memory type found");
}
const vk::MemoryAllocateInfo alloc_ci(requirements.size, *type);
memory = dev.allocateMemoryUnique(alloc_ci, nullptr, dld);
dev.bindBufferMemory(*buffer, *memory, 0, dld);
vk_stream_buffer: Implement a stream buffer This manages two kinds of streaming buffers: one for unified memory models and one for dedicated GPUs. The first one skips the copy from the staging buffer to the real buffer, since it creates an unified buffer. This implementation waits for all fences to finish their operation before "invalidating". This is suboptimal since it should allocate another buffer or start searching from the beginning. There is room for improvement here. This could also handle AMD's "pinned" memory (a heap with 256 MiB) that seems to be designed for buffer streaming.
2019-02-24 08:22:33 +01:00
}
vk_stream_buffer/vk_buffer_cache: Avoid halting and use generic cache The stream buffer before this commit once it was full (no more bytes to write before looping) waiting for all previous operations to finish. This was a temporary solution and had a noticeable performance penalty in performance (from what a profiler showed). To avoid this mark with fences usages of the stream buffer and once it loops wait for them to be signaled. On average this will never wait. Each fence knows where its usage finishes, resulting in a non-paged stream buffer. On the other side, the buffer cache is reimplemented using the generic buffer cache. It makes use of the staging buffer pool and the new stream buffer.
2020-01-06 21:59:20 +01:00
void VKStreamBuffer::ReserveWatches(std::vector<Watch>& watches, std::size_t grow_size) {
watches.resize(watches.size() + grow_size);
}
void VKStreamBuffer::WaitPendingOperations(u64 requested_upper_bound) {
if (!invalidation_mark) {
return;
}
while (requested_upper_bound < wait_bound && wait_cursor < *invalidation_mark) {
auto& watch = previous_watches[wait_cursor];
wait_bound = watch.upper_bound;
watch.fence.Wait();
++wait_cursor;
}
vk_stream_buffer: Implement a stream buffer This manages two kinds of streaming buffers: one for unified memory models and one for dedicated GPUs. The first one skips the copy from the staging buffer to the real buffer, since it creates an unified buffer. This implementation waits for all fences to finish their operation before "invalidating". This is suboptimal since it should allocate another buffer or start searching from the beginning. There is room for improvement here. This could also handle AMD's "pinned" memory (a heap with 256 MiB) that seems to be designed for buffer streaming.
2019-02-24 08:22:33 +01:00
}
} // namespace Vulkan
Copy permalink