octo/citra
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

renderer_vulkan: Rewrite stream buffer, again...

Browse Source 
* The previous implemention was fine, but it wasted space. Buckets now are just ticks attached to a particular buffer region, which means we can flush/map arbitrary regions

* A bug in the texture runtime is also fixed which commited to the same buffer twice
This commit is contained in:
GPUCode
2022-10-28 22:25:09 +03:00
parent b4184a3c2b
commit 20496e8ca4
3 changed files with 38 additions and 61 deletions
Download Patch File Download Diff File Expand all files Collapse all files

75
src/video_core/renderer_vulkan/vk_stream_buffer.cpp
View File

@ -120,28 +120,36 @@ StreamBuffer::~StreamBuffer() {
std::tuple<u8*, u32, bool> StreamBuffer::Map(u32 size, u32 alignment) {
ASSERT(size <= total_size && alignment <= total_size);
Bucket& bucket = buckets[bucket_index];
if (alignment > 0) {
bucket.cursor = Common::AlignUp(bucket.cursor, alignment);
buffer_offset = Common::AlignUp(buffer_offset, alignment);
}
// If we reach bucket boundaries move over to the next one
if (bucket.cursor + size > bucket_size) {
bucket.gpu_tick = scheduler.CurrentTick();
MoveNextBucket();
return Map(size, alignment);
bool invalidate = false;
const u32 new_offset = buffer_offset + size;
if (u32 new_index = new_offset / bucket_size; new_index != bucket_index) {
if (new_index >= BUCKET_COUNT) {
if (readback) {
Invalidate();
} else {
Flush();
}
buffer_offset = 0;
flush_offset = 0;
new_index = 0;
invalidate = true;
}
ticks[bucket_index] = scheduler.CurrentTick();
scheduler.Wait(ticks[new_index]);
bucket_index = new_index;
}
const bool invalidate = std::exchange(bucket.invalid, false);
const u32 buffer_offset = bucket_index * bucket_size + bucket.cursor;
u8* mapped = reinterpret_cast<u8*>(staging.mapped.data() + buffer_offset);
return std::make_tuple(mapped, buffer_offset, invalidate);
}
void StreamBuffer::Commit(u32 size) {
buckets[bucket_index].cursor += size;
buffer_offset += size;
}
void StreamBuffer::Flush() {
@ -150,20 +158,18 @@ void StreamBuffer::Flush() {
return;
}
Bucket& bucket = buckets[bucket_index];
const u32 flush_start = bucket_index * bucket_size + bucket.flush_cursor;
const u32 flush_size = bucket.cursor - bucket.flush_cursor;
ASSERT(flush_size <= bucket_size);
ASSERT(flush_start + flush_size <= total_size);
const u32 flush_size = buffer_offset - flush_offset;
ASSERT(flush_size <= total_size);
ASSERT(flush_offset + flush_size <= total_size);
if (flush_size > 0) [[likely]] {
// Ensure all staging writes are visible to the host memory domain
VmaAllocator allocator = instance.GetAllocator();
vmaFlushAllocation(allocator, staging.allocation, flush_start, flush_size);
vmaFlushAllocation(allocator, staging.allocation, flush_offset, flush_size);
if (gpu_buffer) {
scheduler.Record([this, flush_start, flush_size](vk::CommandBuffer, vk::CommandBuffer upload_cmdbuf) {
scheduler.Record([this, flush_offset = flush_offset, flush_size](vk::CommandBuffer, vk::CommandBuffer upload_cmdbuf) {
const vk::BufferCopy copy_region = {
.srcOffset = flush_start, .dstOffset = flush_start, .size = flush_size};
.srcOffset = flush_offset, .dstOffset = flush_offset, .size = flush_size};
upload_cmdbuf.copyBuffer(staging.buffer, gpu_buffer, copy_region);
@ -174,7 +180,7 @@ void StreamBuffer::Flush() {
.srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
.dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
.buffer = gpu_buffer,
.offset = flush_start,
.offset = flush_offset,
.size = flush_size};
upload_cmdbuf.pipelineBarrier(vk::PipelineStageFlagBits::eTransfer, stage_mask,
@ -182,7 +188,7 @@ void StreamBuffer::Flush() {
{});
});
}
bucket.flush_cursor += flush_size;
flush_offset = buffer_offset;
}
}
@ -191,33 +197,16 @@ void StreamBuffer::Invalidate() {
return;
}
Bucket& bucket = buckets[bucket_index];
const u32 flush_start = bucket_index * bucket_size + bucket.flush_cursor;
const u32 flush_size = bucket.cursor - bucket.flush_cursor;
ASSERT(flush_size <= bucket_size);
const u32 flush_size = buffer_offset - flush_offset;
ASSERT(flush_size <= total_size);
ASSERT(flush_offset + flush_size <= total_size);
if (flush_size > 0) [[likely]] {
// Ensure the staging memory can be read by the host
VmaAllocator allocator = instance.GetAllocator();
vmaInvalidateAllocation(allocator, staging.allocation, flush_start, flush_size);
bucket.flush_cursor += flush_size;
vmaInvalidateAllocation(allocator, staging.allocation, flush_offset, flush_size);
flush_offset = buffer_offset;
}
}
void StreamBuffer::MoveNextBucket() {
// Flush and Invalidate are bucket local operations for simplicity so perform them here
if (readback) {
Invalidate();
} else {
Flush();
}
bucket_index = (bucket_index + 1) % BUCKET_COUNT;
Bucket& next_bucket = buckets[bucket_index];
scheduler.Wait(next_bucket.gpu_tick);
next_bucket.cursor = 0;
next_bucket.flush_cursor = 0;
next_bucket.invalid = true;
}
} // namespace Vulkan

15
src/video_core/renderer_vulkan/vk_stream_buffer.h
View File

@ -71,17 +71,6 @@ public:
return views[index];
}
private:
/// Moves to the next bucket
void MoveNextBucket();
struct Bucket {
bool invalid = false;
u32 gpu_tick = 0;
u32 cursor = 0;
u32 flush_cursor = 0;
};
private:
const Instance& instance;
Scheduler& scheduler;
@ -90,12 +79,14 @@ private:
VmaAllocation allocation{};
vk::BufferUsageFlagBits usage;
std::array<vk::BufferView, MAX_BUFFER_VIEWS> views{};
std::array<Bucket, BUCKET_COUNT> buckets;
std::size_t view_count = 0;
u32 total_size = 0;
u32 bucket_size = 0;
u32 buffer_offset = 0;
u32 flush_offset = 0;
u32 bucket_index = 0;
bool readback = false;
std::array<u64, BUCKET_COUNT> ticks{};
};
} // namespace Vulkan

9
src/video_core/renderer_vulkan/vk_texture_runtime.cpp
View File

@ -708,12 +708,11 @@ void Surface::Upload(const VideoCore::BufferTextureCopy& upload, const StagingDa
render_cmdbuf.copyBufferToImage(staging.buffer, image, vk::ImageLayout::eTransferDstOptimal,
region_count, copy_regions.data());
});
runtime.upload_buffer.Commit(staging.size);
}
InvalidateAllWatcher();
// Lock this data until the next scheduler switch
runtime.upload_buffer.Commit(staging.size);
}
MICROPROFILE_DEFINE(Vulkan_Download, "VulkanSurface", "Texture Download", MP_RGB(128, 192, 64));
@ -751,10 +750,8 @@ void Surface::Download(const VideoCore::BufferTextureCopy& download, const Stagi
render_cmdbuf.copyImageToBuffer(image, vk::ImageLayout::eTransferSrcOptimal,
staging.buffer, copy_region);
});
runtime.download_buffer.Commit(staging.size);
}
// Lock this data until the next scheduler switch
runtime.download_buffer.Commit(staging.size);
}
u32 Surface::GetInternalBytesPerPixel() const {