1
0
Fork 0
forked from wry/wry

vulkan: use transfer queue for uploads

This commit is contained in:
Julian Orth 2024-09-17 01:05:43 +02:00
parent 88a528d3bf
commit 3ab5f6981f
6 changed files with 424 additions and 86 deletions

View file

@ -86,8 +86,8 @@ impl VulkanDevice {
drm: &Drm,
) -> Result<VulkanBoAllocator, VulkanError> {
let allocator = self.create_allocator()?;
let pool = self.create_command_pool()?;
let command_buffer = pool.allocate_buffer()?;
let pool = self.create_command_pool(self.graphics_queue_idx)?;
let command_buffer = pool.allocate()?;
let drm = drm.dup_render().map_err(VulkanError::DupDrm)?;
Ok(VulkanBoAllocator {
data: Rc::new(VulkanBoAllocatorData {

View file

@ -1,5 +1,5 @@
use {
crate::gfx_apis::vulkan::{device::VulkanDevice, VulkanError},
crate::gfx_apis::vulkan::{device::VulkanDevice, renderer::CachedCommandBuffers, VulkanError},
ash::vk::{
CommandBuffer, CommandBufferAllocateInfo, CommandBufferLevel, CommandPool,
CommandPoolCreateFlags, CommandPoolCreateInfo,
@ -53,17 +53,24 @@ impl VulkanCommandPool {
}
impl VulkanDevice {
pub fn create_command_pool(self: &Rc<Self>) -> Result<Rc<VulkanCommandPool>, VulkanError> {
pub fn create_command_pool(
self: &Rc<Self>,
queue: u32,
) -> Result<CachedCommandBuffers, VulkanError> {
let info = CommandPoolCreateInfo::default()
.queue_family_index(self.graphics_queue_idx)
.queue_family_index(queue)
.flags(
CommandPoolCreateFlags::TRANSIENT | CommandPoolCreateFlags::RESET_COMMAND_BUFFER,
);
let pool = unsafe { self.device.create_command_pool(&info, None) };
let pool = pool.map_err(VulkanError::AllocateCommandPool)?;
Ok(Rc::new(VulkanCommandPool {
device: self.clone(),
pool,
}))
Ok(CachedCommandBuffers {
pool: Rc::new(VulkanCommandPool {
device: self.clone(),
pool,
}),
buffers: Default::default(),
total_buffers: Default::default(),
})
}
}

View file

@ -63,6 +63,9 @@ pub struct VulkanDevice {
pub(super) memory_types: ArrayVec<MemoryType, MAX_MEMORY_TYPES>,
pub(super) graphics_queue: Queue,
pub(super) graphics_queue_idx: u32,
pub(super) transfer_queue: Option<Queue>,
pub(super) distinct_transfer_queue_family_idx: Option<u32>,
pub(super) transfer_granularity_mask: (u32, u32),
}
impl Drop for VulkanDevice {
@ -185,16 +188,59 @@ impl VulkanInstance {
Err(VulkanError::NoDeviceFound(dev))
}
fn find_graphics_queue(&self, phy_dev: PhysicalDevice) -> Result<u32, VulkanError> {
fn find_queues(
&self,
phy_dev: PhysicalDevice,
) -> Result<(u32, Option<(u32, u32, u32)>), VulkanError> {
let props = unsafe {
self.instance
.get_physical_device_queue_family_properties(phy_dev)
};
props
let gfx_queue = props
.iter()
.position(|p| p.queue_flags.contains(QueueFlags::GRAPHICS))
.map(|v| v as _)
.ok_or(VulkanError::NoGraphicsQueue)
.ok_or(VulkanError::NoGraphicsQueue)?;
let transfer_queue = 'transfer: {
let mut transfer_only = None;
let mut compute_only = None;
let mut separate_gfx = None;
for (idx, props) in props.iter().enumerate() {
if idx == gfx_queue {
continue;
}
let g = &props.min_image_transfer_granularity;
if g.width == 0 || g.height == 0 {
continue;
}
let f = props.queue_flags;
use QueueFlags as F;
if !f.intersects(F::GRAPHICS | F::COMPUTE) && f.intersects(F::TRANSFER) {
transfer_only = Some(idx);
} else if !f.intersects(F::GRAPHICS) && f.intersects(F::COMPUTE) {
compute_only = Some(idx);
} else if f.intersects(F::GRAPHICS) {
separate_gfx = Some(idx);
}
}
if let Some(idx) = transfer_only.or(compute_only).or(separate_gfx) {
break 'transfer Some(idx);
}
if props[gfx_queue].queue_count > 1 {
break 'transfer Some(gfx_queue);
}
None
};
let mut width_mask = 0;
let mut height_mask = 0;
if let Some(idx) = transfer_queue {
let g = &props[idx].min_image_transfer_granularity;
width_mask = g.width.wrapping_sub(1);
height_mask = g.height.wrapping_sub(1);
}
Ok((
gfx_queue as _,
transfer_queue.map(|v| (v as _, width_mask, height_mask)),
))
}
fn supports_semaphore_import(&self, phy_dev: PhysicalDevice) -> bool {
@ -224,7 +270,15 @@ impl VulkanInstance {
return Err(VulkanError::MissingDeviceExtension(ext));
}
}
let graphics_queue_idx = self.find_graphics_queue(phy_dev)?;
let (graphics_queue_family_idx, transfer_queue_family) = self.find_queues(phy_dev)?;
let mut distinct_transfer_queue_family_idx = None;
let mut transfer_granularity_mask = (0, 0);
if let Some((idx, width_mask, height_mask)) = transfer_queue_family {
if idx != graphics_queue_family_idx {
distinct_transfer_queue_family_idx = Some(idx);
}
transfer_granularity_mask = (width_mask, height_mask);
}
if !self.supports_semaphore_import(phy_dev) {
return Err(VulkanError::SyncobjImport);
}
@ -238,14 +292,24 @@ impl VulkanInstance {
PhysicalDeviceSynchronization2Features::default().synchronization2(true);
let mut dynamic_rendering_features =
PhysicalDeviceDynamicRenderingFeatures::default().dynamic_rendering(true);
let queue_create_info = DeviceQueueCreateInfo::default()
.queue_family_index(graphics_queue_idx)
.queue_priorities(&[1.0]);
let mut queue_create_infos = ArrayVec::<_, 2>::new();
queue_create_infos.push(
DeviceQueueCreateInfo::default()
.queue_family_index(graphics_queue_family_idx)
.queue_priorities(&[1.0]),
);
if let Some((tq, _, _)) = transfer_queue_family {
queue_create_infos.push(
DeviceQueueCreateInfo::default()
.queue_family_index(tq)
.queue_priorities(&[1.0]),
);
}
let device_create_info = DeviceCreateInfo::default()
.push_next(&mut semaphore_features)
.push_next(&mut synchronization2_features)
.push_next(&mut dynamic_rendering_features)
.queue_create_infos(std::slice::from_ref(&queue_create_info))
.queue_create_infos(&queue_create_infos)
.enabled_extension_names(&enabled_extensions);
let device = unsafe {
self.instance
@ -286,7 +350,14 @@ impl VulkanInstance {
.iter()
.copied()
.collect();
let graphics_queue = unsafe { device.get_device_queue(graphics_queue_idx, 0) };
let graphics_queue = unsafe { device.get_device_queue(graphics_queue_family_idx, 0) };
let transfer_queue = transfer_queue_family.map(|(family_idx, _, _)| {
let queue_idx = match family_idx == graphics_queue_family_idx {
true => 1,
false => 0,
};
unsafe { device.get_device_queue(family_idx, queue_idx) }
});
Ok(Rc::new(VulkanDevice {
physical_device: phy_dev,
render_node,
@ -302,7 +373,10 @@ impl VulkanInstance {
formats,
memory_types,
graphics_queue,
graphics_queue_idx,
graphics_queue_idx: graphics_queue_family_idx,
transfer_queue,
distinct_transfer_queue_family_idx,
transfer_granularity_mask,
}))
}
}

View file

@ -58,10 +58,41 @@ pub struct VulkanImage {
pub(super) image: Image,
pub(super) is_undefined: Cell<bool>,
pub(super) contents_are_undefined: Cell<bool>,
pub(super) queue_state: Cell<QueueState>,
pub(super) ty: VulkanImageMemory,
pub(super) bridge: Option<VulkanFramebufferBridge>,
}
#[derive(Copy, Clone, Eq, PartialEq, Debug)]
pub enum QueueState {
Acquired { family: QueueFamily },
Releasing,
Released { to: QueueFamily },
}
#[derive(Copy, Clone, Eq, PartialEq, Debug)]
pub enum QueueFamily {
Gfx,
Transfer,
}
impl QueueState {
pub fn acquire(self, new: QueueFamily) -> QueueTransfer {
match self {
QueueState::Acquired { family } if family == new => QueueTransfer::Unnecessary,
QueueState::Released { to } if to == new => QueueTransfer::Possible,
_ => QueueTransfer::Impossible,
}
}
}
#[derive(Copy, Clone, Eq, PartialEq, Debug)]
pub enum QueueTransfer {
Unnecessary,
Possible,
Impossible,
}
pub enum VulkanImageMemory {
DmaBuf(VulkanDmaBufImage),
Internal(VulkanShmImage),
@ -384,6 +415,9 @@ impl VulkanDmaBufImageTemplate {
format: self.dmabuf.format,
is_undefined: Cell::new(true),
contents_are_undefined: Cell::new(false),
queue_state: Cell::new(QueueState::Acquired {
family: QueueFamily::Gfx,
}),
bridge,
}))
}

View file

@ -13,7 +13,7 @@ use {
descriptor::VulkanDescriptorSetLayout,
device::VulkanDevice,
fence::VulkanFence,
image::{VulkanImage, VulkanImageMemory},
image::{QueueFamily, QueueState, QueueTransfer, VulkanImage, VulkanImageMemory},
pipeline::{PipelineCreateInfo, VulkanPipeline},
semaphore::VulkanSemaphore,
shaders::{
@ -60,13 +60,12 @@ pub struct VulkanRenderer {
pub(super) formats: Rc<AHashMap<u32, GfxFormat>>,
pub(super) device: Rc<VulkanDevice>,
pub(super) pipelines: CopyHashMap<vk::Format, Rc<VulkanFormatPipelines>>,
pub(super) command_pool: Rc<VulkanCommandPool>,
pub(super) command_buffers: Stack<Rc<VulkanCommandBuffer>>,
pub(super) gfx_command_buffers: CachedCommandBuffers,
pub(super) transfer_command_buffers: Option<CachedCommandBuffers>,
pub(super) wait_semaphores: Stack<Rc<VulkanSemaphore>>,
pub(super) total_buffers: NumCell<usize>,
pub(super) memory: RefCell<Memory>,
pub(super) pending_frames: CopyHashMap<u64, Rc<PendingFrame>>,
pub(super) pending_uploads: CopyHashMap<u64, SpawnedFuture<()>>,
pub(super) pending_submits: CopyHashMap<u64, SpawnedFuture<()>>,
pub(super) allocator: Rc<VulkanAllocator>,
pub(super) last_point: NumCell<u64>,
pub(super) buffer_resv_user: BufferResvUser,
@ -84,6 +83,26 @@ pub struct VulkanRenderer {
pub(super) shm_allocator: Rc<VulkanThreadedAllocator>,
}
pub(super) struct CachedCommandBuffers {
pub(super) pool: Rc<VulkanCommandPool>,
pub(super) buffers: Stack<Rc<VulkanCommandBuffer>>,
pub(super) total_buffers: NumCell<usize>,
}
impl CachedCommandBuffers {
pub(super) fn allocate(&self) -> Result<Rc<VulkanCommandBuffer>, VulkanError> {
zone!("allocate_command_buffer");
let buf = match self.buffers.pop() {
Some(b) => b,
_ => {
self.total_buffers.fetch_add(1);
self.pool.allocate_buffer()?
}
};
Ok(buf)
}
}
pub(super) struct UsedTexture {
tex: Rc<VulkanImage>,
resv: Option<Rc<dyn BufferResv>>,
@ -105,7 +124,8 @@ pub(super) enum TexSourceType {
#[derive(Default)]
pub(super) struct Memory {
sample: Vec<Rc<VulkanImage>>,
dmabuf_sample: Vec<Rc<VulkanImage>>,
queue_transfer: Vec<Rc<VulkanImage>>,
textures: Vec<UsedTexture>,
image_barriers: Vec<ImageMemoryBarrier2<'static>>,
wait_semaphores: Vec<Rc<VulkanSemaphore>>,
@ -143,7 +163,11 @@ impl VulkanDevice {
let tex_frag_shader = self.create_shader(TEX_FRAG)?;
let tex_frag_mult_opaque_shader = self.create_shader(TEX_FRAG_MULT_OPAQUE)?;
let tex_frag_mult_alpha_shader = self.create_shader(TEX_FRAG_MULT_ALPHA)?;
let command_pool = self.create_command_pool()?;
let gfx_command_buffers = self.create_command_pool(self.graphics_queue_idx)?;
let transfer_command_buffers = self
.distinct_transfer_queue_family_idx
.map(|idx| self.create_command_pool(idx))
.transpose()?;
let formats: AHashMap<u32, _> = self
.formats
.iter()
@ -181,13 +205,12 @@ impl VulkanDevice {
formats: Rc::new(formats),
device: self.clone(),
pipelines: Default::default(),
command_pool,
command_buffers: Default::default(),
gfx_command_buffers,
transfer_command_buffers,
wait_semaphores: Default::default(),
total_buffers: Default::default(),
memory: Default::default(),
pending_frames: Default::default(),
pending_uploads: Default::default(),
pending_submits: Default::default(),
allocator,
last_point: Default::default(),
buffer_resv_user: Default::default(),
@ -276,15 +299,21 @@ impl VulkanRenderer {
fn collect_memory(&self, opts: &[GfxApiOpt]) {
zone!("collect_memory");
let mut memory = self.memory.borrow_mut();
memory.sample.clear();
memory.dmabuf_sample.clear();
memory.queue_transfer.clear();
for cmd in opts {
if let GfxApiOpt::CopyTexture(c) = cmd {
let tex = c.tex.clone().into_vk(&self.device.device);
if tex.contents_are_undefined.get() {
continue;
}
match tex.queue_state.get().acquire(QueueFamily::Gfx) {
QueueTransfer::Unnecessary => {}
QueueTransfer::Possible => memory.queue_transfer.push(tex.clone()),
QueueTransfer::Impossible => continue,
}
if let VulkanImageMemory::DmaBuf(_) = &tex.ty {
memory.sample.push(tex.clone())
memory.dmabuf_sample.push(tex.clone())
}
memory.textures.push(UsedTexture {
tex,
@ -340,7 +369,7 @@ impl VulkanRenderer {
});
}
memory.image_barriers.push(fb_image_memory_barrier);
for img in &memory.sample {
for img in &memory.dmabuf_sample {
let image_memory_barrier = image_barrier()
.src_queue_family_index(QUEUE_FAMILY_FOREIGN_EXT)
.dst_queue_family_index(self.device.graphics_queue_idx)
@ -351,6 +380,19 @@ impl VulkanRenderer {
.dst_stage_mask(PipelineStageFlags2::FRAGMENT_SHADER);
memory.image_barriers.push(image_memory_barrier);
}
if let Some(family_idx) = self.device.distinct_transfer_queue_family_idx {
for img in &memory.queue_transfer {
let image_memory_barrier = image_barrier()
.src_queue_family_index(family_idx)
.dst_queue_family_index(self.device.graphics_queue_idx)
.image(img.image)
.dst_access_mask(AccessFlags2::SHADER_SAMPLED_READ)
.dst_stage_mask(PipelineStageFlags2::FRAGMENT_SHADER)
.old_layout(ImageLayout::TRANSFER_DST_OPTIMAL)
.new_layout(ImageLayout::SHADER_READ_ONLY_OPTIMAL);
memory.image_barriers.push(image_memory_barrier);
}
}
let dep_info = DependencyInfoKHR::default().image_memory_barriers(&memory.image_barriers);
unsafe {
self.device.device.cmd_pipeline_barrier2(buf, &dep_info);
@ -471,6 +513,11 @@ impl VulkanRenderer {
log::warn!("Ignoring undefined texture");
continue;
}
if tex.queue_state.get().acquire(QueueFamily::Gfx) == QueueTransfer::Impossible
{
log::warn!("Ignoring texture owned by different queue");
continue;
}
let copy_type = match c.alpha.is_some() {
true => TexCopyType::Multiply,
false => TexCopyType::Identity,
@ -616,7 +663,7 @@ impl VulkanRenderer {
.src_stage_mask(PipelineStageFlags2::COLOR_ATTACHMENT_OUTPUT);
}
memory.image_barriers.push(fb_image_memory_barrier);
for img in &memory.sample {
for img in &memory.dmabuf_sample {
let image_memory_barrier = image_barrier()
.src_queue_family_index(self.device.graphics_queue_idx)
.dst_queue_family_index(QUEUE_FAMILY_FOREIGN_EXT)
@ -732,6 +779,8 @@ impl VulkanRenderer {
}
}
};
let attach_async_shm_sync_file = self.device.transfer_queue.is_some()
&& self.device.distinct_transfer_queue_family_idx.is_none();
for texture in &mut memory.textures {
import(
&texture.tex,
@ -739,6 +788,13 @@ impl VulkanRenderer {
texture.resv.take(),
DMA_BUF_SYNC_READ,
);
if attach_async_shm_sync_file {
if let VulkanImageMemory::Internal(shm) = &texture.tex.ty {
if let Some(data) = &shm.async_data {
data.last_sample.set(Some(sync_file.clone()));
}
}
}
}
import(fb, fb_release_sync, None, DMA_BUF_SYNC_WRITE);
}
@ -777,6 +833,12 @@ impl VulkanRenderer {
fn store_layouts(&self, fb: &VulkanImage) {
fb.is_undefined.set(false);
let memory = self.memory.borrow();
for img in &*memory.queue_transfer {
img.queue_state.set(QueueState::Acquired {
family: QueueFamily::Gfx,
});
}
}
fn create_pending_frame(self: &Rc<Self>, buf: Rc<VulkanCommandBuffer>) {
@ -932,7 +994,7 @@ impl VulkanRenderer {
final_barriers =
final_barriers.image_memory_barriers(slice::from_ref(&final_tex_barrier));
}
let buf = self.allocate_command_buffer()?;
let buf = self.gfx_command_buffers.allocate()?;
let mut semaphores = vec![];
let mut semaphore_infos = vec![];
if let VulkanImageMemory::DmaBuf(buf) = &tex.ty {
@ -986,7 +1048,7 @@ impl VulkanRenderer {
.map_err(VulkanError::Submit)?;
}
self.block();
self.command_buffers.push(buf);
self.gfx_command_buffers.buffers.push(buf);
for semaphore in semaphores {
self.wait_semaphores.push(semaphore);
}
@ -1009,7 +1071,8 @@ impl VulkanRenderer {
let sync_file = {
let mut memory = self.memory.borrow_mut();
memory.textures.clear();
memory.sample.clear();
memory.dmabuf_sample.clear();
memory.queue_transfer.clear();
memory.wait_semaphores.clear();
memory.release_fence.take();
memory.release_sync_file.take()
@ -1017,18 +1080,6 @@ impl VulkanRenderer {
res.map(|_| sync_file)
}
pub(super) fn allocate_command_buffer(&self) -> Result<Rc<VulkanCommandBuffer>, VulkanError> {
zone!("allocate_command_buffer");
let buf = match self.command_buffers.pop() {
Some(b) => b,
_ => {
self.total_buffers.fetch_add(1);
self.command_pool.allocate_buffer()?
}
};
Ok(buf)
}
fn allocate_semaphore(&self) -> Result<Rc<VulkanSemaphore>, VulkanError> {
zone!("allocate_semaphore");
let semaphore = match self.wait_semaphores.pop() {
@ -1047,7 +1098,7 @@ impl VulkanRenderer {
clear: Option<&Color>,
) -> Result<(), VulkanError> {
self.check_defunct()?;
let buf = self.allocate_command_buffer()?;
let buf = self.gfx_command_buffers.allocate()?;
self.collect_memory(opts);
self.begin_command_buffer(buf.buffer)?;
self.initial_barriers(buf.buffer, fb);
@ -1078,7 +1129,7 @@ impl VulkanRenderer {
pub fn on_drop(&self) {
self.defunct.set(true);
let mut pending_frames = self.pending_frames.lock();
let mut pending_uploads = self.pending_uploads.lock();
let mut pending_uploads = self.pending_submits.lock();
if pending_frames.is_not_empty() || pending_uploads.is_not_empty() {
log::warn!("Context dropped with pending frames.");
self.block();
@ -1164,7 +1215,7 @@ async fn await_release(
frame.renderer.block();
}
if let Some(buf) = frame.cmd.take() {
frame.renderer.command_buffers.push(buf);
frame.renderer.gfx_command_buffers.buffers.push(buf);
}
for wait_semaphore in frame.wait_semaphores.take() {
frame.renderer.wait_semaphores.push(wait_semaphore);

View file

@ -14,7 +14,7 @@ use {
allocator::VulkanAllocation,
command::VulkanCommandBuffer,
fence::VulkanFence,
image::{VulkanImage, VulkanImageMemory},
image::{QueueFamily, QueueState, QueueTransfer, VulkanImage, VulkanImageMemory},
renderer::{image_barrier, VulkanRenderer},
staging::VulkanStagingBuffer,
VulkanError,
@ -22,9 +22,10 @@ use {
rect::{Rect, Region},
utils::{clonecell::CloneCell, errorfmt::ErrorFmt, on_drop::OnDrop},
},
arrayvec::ArrayVec,
ash::vk::{
AccessFlags2, BufferImageCopy2, BufferMemoryBarrier2, CommandBufferBeginInfo,
CommandBufferSubmitInfo, CommandBufferUsageFlags, CopyBufferToImageInfo2,
CommandBufferSubmitInfo, CommandBufferUsageFlags, CopyBufferToImageInfo2, DependencyInfo,
DependencyInfoKHR, DeviceSize, Extent3D, ImageAspectFlags, ImageCreateInfo, ImageLayout,
ImageSubresourceLayers, ImageSubresourceRange, ImageTiling, ImageType, ImageUsageFlags,
ImageViewCreateInfo, ImageViewType, Offset3D, PipelineStageFlags2, SampleCountFlags,
@ -58,6 +59,8 @@ pub struct VulkanShmImageAsyncData {
pub(super) callback_id: Cell<u64>,
pub(super) regions: RefCell<Vec<BufferImageCopy2<'static>>>,
pub(super) cpu: Rc<CpuWorker>,
pub(super) last_sample: Cell<Option<SyncFile>>,
pub(super) data_copied: Cell<bool>,
}
impl VulkanShmImage {
@ -160,7 +163,7 @@ impl VulkanShmImage {
}
})?;
let Some((cmd, fence, sync_file, point)) =
self.submit_buffer_to_image_copy(img, &staging, cpy)?
self.submit_buffer_to_image_copy(img, &staging, cpy, false)?
else {
return Ok(());
};
@ -168,7 +171,7 @@ impl VulkanShmImage {
"await upload",
await_upload(point, img.clone(), cmd, sync_file, fence, staging),
);
img.renderer.pending_uploads.set(point, future);
img.renderer.pending_submits.set(point, future);
Ok(())
}
@ -177,6 +180,7 @@ impl VulkanShmImage {
img: &Rc<VulkanImage>,
staging: &VulkanStagingBuffer,
regions: &[BufferImageCopy2],
use_transfer_queue: bool,
) -> Result<Option<(Rc<VulkanCommandBuffer>, Rc<VulkanFence>, SyncFile, u64)>, VulkanError>
{
let memory_barrier = |sam, ssm, dam, dsm| {
@ -189,18 +193,29 @@ impl VulkanShmImage {
.dst_access_mask(dam)
.dst_stage_mask(dsm)
};
let initial_image_barrier = image_barrier()
let mut transfer_queue_family_idx = img.renderer.device.graphics_queue_idx;
if use_transfer_queue {
if let Some(idx) = img.renderer.device.distinct_transfer_queue_family_idx {
transfer_queue_family_idx = idx;
}
}
let mut initial_image_barrier = image_barrier()
.image(img.image)
.src_access_mask(AccessFlags2::SHADER_SAMPLED_READ)
.src_stage_mask(PipelineStageFlags2::FRAGMENT_SHADER)
.src_queue_family_index(img.renderer.device.graphics_queue_idx)
.dst_queue_family_index(transfer_queue_family_idx)
.dst_access_mask(AccessFlags2::TRANSFER_WRITE)
.dst_stage_mask(PipelineStageFlags2::TRANSFER)
.old_layout(if img.is_undefined.get() {
ImageLayout::UNDEFINED
} else {
ImageLayout::SHADER_READ_ONLY_OPTIMAL
})
.new_layout(ImageLayout::TRANSFER_DST_OPTIMAL)
.dst_access_mask(AccessFlags2::TRANSFER_WRITE)
.dst_stage_mask(PipelineStageFlags2::TRANSFER);
.new_layout(ImageLayout::TRANSFER_DST_OPTIMAL);
if transfer_queue_family_idx == img.renderer.device.graphics_queue_idx {
initial_image_barrier = initial_image_barrier
.src_access_mask(AccessFlags2::SHADER_SAMPLED_READ)
.src_stage_mask(PipelineStageFlags2::FRAGMENT_SHADER)
}
let initial_buffer_barrier = memory_barrier(
AccessFlags2::HOST_WRITE,
PipelineStageFlags2::HOST,
@ -210,14 +225,19 @@ impl VulkanShmImage {
let initial_dep_info = DependencyInfoKHR::default()
.buffer_memory_barriers(slice::from_ref(&initial_buffer_barrier))
.image_memory_barriers(slice::from_ref(&initial_image_barrier));
let final_image_barrier = image_barrier()
let mut final_image_barrier = image_barrier()
.image(img.image)
.src_queue_family_index(transfer_queue_family_idx)
.dst_queue_family_index(img.renderer.device.graphics_queue_idx)
.src_access_mask(AccessFlags2::TRANSFER_WRITE)
.src_stage_mask(PipelineStageFlags2::TRANSFER)
.old_layout(ImageLayout::TRANSFER_DST_OPTIMAL)
.new_layout(ImageLayout::SHADER_READ_ONLY_OPTIMAL)
.dst_access_mask(AccessFlags2::SHADER_SAMPLED_READ)
.dst_stage_mask(PipelineStageFlags2::FRAGMENT_SHADER);
.new_layout(ImageLayout::SHADER_READ_ONLY_OPTIMAL);
if transfer_queue_family_idx == img.renderer.device.graphics_queue_idx {
final_image_barrier = final_image_barrier
.dst_access_mask(AccessFlags2::SHADER_SAMPLED_READ)
.dst_stage_mask(PipelineStageFlags2::FRAGMENT_SHADER);
}
let final_buffer_barrier = memory_barrier(
AccessFlags2::TRANSFER_READ,
PipelineStageFlags2::TRANSFER,
@ -232,7 +252,10 @@ impl VulkanShmImage {
.dst_image(img.image)
.dst_image_layout(ImageLayout::TRANSFER_DST_OPTIMAL)
.regions(regions);
let cmd = img.renderer.allocate_command_buffer()?;
let cmd = match &img.renderer.transfer_command_buffers {
Some(b) if use_transfer_queue => b.allocate()?,
_ => img.renderer.gfx_command_buffers.allocate()?,
};
let dev = &img.renderer.device.device;
let command_buffer_info = CommandBufferSubmitInfo::default().command_buffer(cmd.buffer);
let submit_info =
@ -249,7 +272,10 @@ impl VulkanShmImage {
dev.end_command_buffer(cmd.buffer)
.map_err(VulkanError::EndCommandBuffer)?;
dev.queue_submit2(
img.renderer.device.graphics_queue,
match img.renderer.device.transfer_queue {
Some(q) if use_transfer_queue => q,
_ => img.renderer.device.graphics_queue,
},
slice::from_ref(&submit_info),
release_fence.fence,
)
@ -286,8 +312,8 @@ async fn await_upload(
);
img.renderer.block();
}
img.renderer.command_buffers.push(buf);
img.renderer.pending_uploads.remove(&id);
img.renderer.gfx_command_buffers.buffers.push(buf);
img.renderer.pending_submits.remove(&id);
}
impl VulkanShmImageAsyncData {
@ -334,8 +360,9 @@ impl VulkanShmImage {
return Err(VulkanError::InvalidBufferSize);
}
data.busy.set(true);
data.data_copied.set(false);
if img.contents_are_undefined.get() {
damage = Region::new2(Rect::new_sized(0, 0, img.width as _, img.height as _).unwrap())
damage = Region::new2(Rect::new_sized(0, 0, img.width as _, img.height as _).unwrap());
}
let copies = &mut *data.regions.borrow_mut();
@ -361,13 +388,18 @@ impl VulkanShmImage {
.buffer_row_length(img.stride / self.shm_info.bpp);
copies.push(copy);
};
let (width_mask, height_mask) = img.renderer.device.transfer_granularity_mask;
let width_mask = width_mask as i32;
let height_mask = height_mask as i32;
for damage in damage.rects() {
let Some(damage) = Rect::new(
damage.x1().max(0),
damage.y1().max(0),
damage.x2().min(img.width as i32),
damage.y2().min(img.height as i32),
) else {
if damage.x2() < 0 || damage.y2() < 0 {
continue;
}
let x1 = damage.x1().max(0) & !width_mask;
let y1 = damage.y1().max(0) & !height_mask;
let x2 = ((damage.x2() + width_mask) & !width_mask).min(img.width as i32);
let y2 = ((damage.y2() + height_mask) & !height_mask).min(img.height as i32);
let Some(damage) = Rect::new(x1, y1, x2, y2) else {
continue;
};
if damage.is_empty() {
@ -381,6 +413,8 @@ impl VulkanShmImage {
);
}
self.async_release_from_gfx_queue(img, data)?;
if let Some(staging) = data.staging.get() {
return self.async_upload_initiate_copy(img, data, &staging, copies, client_mem);
}
@ -404,6 +438,95 @@ impl VulkanShmImage {
)
}
fn async_release_from_gfx_queue(
&self,
img: &Rc<VulkanImage>,
data: &VulkanShmImageAsyncData,
) -> Result<(), VulkanError> {
img.renderer.check_defunct()?;
let Some(transfer_queue_idx) = img.renderer.device.distinct_transfer_queue_family_idx
else {
let Some(sync_file) = data.last_sample.take() else {
img.queue_state.set(QueueState::Released {
to: QueueFamily::Transfer,
});
return Ok(());
};
let id = img.renderer.allocate_point();
let pending = img.renderer.eng.spawn(
"await_transfer_to_transfer",
await_gfx_queue_release(id, img.clone(), None, None, sync_file),
);
img.renderer.pending_submits.set(id, pending);
img.queue_state.set(QueueState::Releasing);
return Ok(());
};
let mut barriers = ArrayVec::<_, 2>::new();
match img.queue_state.get() {
QueueState::Acquired { family } => {
assert_eq!(family, QueueFamily::Gfx);
}
QueueState::Releasing => {
unreachable!();
}
QueueState::Released { to } => {
assert_eq!(to, QueueFamily::Gfx);
let barrier = image_barrier()
.image(img.image)
.src_queue_family_index(transfer_queue_idx)
.dst_queue_family_index(img.renderer.device.graphics_queue_idx)
.dst_stage_mask(PipelineStageFlags2::ALL_COMMANDS)
.old_layout(ImageLayout::TRANSFER_DST_OPTIMAL)
.new_layout(ImageLayout::SHADER_READ_ONLY_OPTIMAL);
barriers.push(barrier);
}
}
let barrier = image_barrier()
.image(img.image)
.src_queue_family_index(img.renderer.device.graphics_queue_idx)
.dst_queue_family_index(transfer_queue_idx)
.src_access_mask(AccessFlags2::SHADER_SAMPLED_READ)
.src_stage_mask(PipelineStageFlags2::ALL_COMMANDS)
.old_layout(if img.is_undefined.get() {
ImageLayout::UNDEFINED
} else {
ImageLayout::SHADER_READ_ONLY_OPTIMAL
})
.new_layout(ImageLayout::TRANSFER_DST_OPTIMAL);
barriers.push(barrier);
let dep_info = DependencyInfo::default().image_memory_barriers(&barriers);
let release_fence = img.renderer.device.create_fence()?;
let dev = &img.renderer.device.device;
let begin_info =
CommandBufferBeginInfo::default().flags(CommandBufferUsageFlags::ONE_TIME_SUBMIT);
let cmd = img.renderer.gfx_command_buffers.allocate()?;
let command_buffer_info = CommandBufferSubmitInfo::default().command_buffer(cmd.buffer);
let submit_info =
SubmitInfo2::default().command_buffer_infos(slice::from_ref(&command_buffer_info));
unsafe {
dev.begin_command_buffer(cmd.buffer, &begin_info)
.map_err(VulkanError::BeginCommandBuffer)?;
dev.cmd_pipeline_barrier2(cmd.buffer, &dep_info);
dev.end_command_buffer(cmd.buffer)
.map_err(VulkanError::EndCommandBuffer)?;
dev.queue_submit2(
img.renderer.device.graphics_queue,
slice::from_ref(&submit_info),
release_fence.fence,
)
.map_err(VulkanError::Submit)?;
}
let sync_file = release_fence.export_sync_file()?;
let id = img.renderer.allocate_point();
let pending = img.renderer.eng.spawn(
"await_transfer_to_transfer",
await_gfx_queue_release(id, img.clone(), Some(cmd), Some(release_fence), sync_file),
);
img.renderer.pending_submits.set(id, pending);
img.queue_state.set(QueueState::Releasing);
Ok(())
}
fn async_upload_after_allocation(
&self,
img: &Rc<VulkanImage>,
@ -501,25 +624,28 @@ impl VulkanShmImage {
&self,
img: &Rc<VulkanImage>,
data: &VulkanShmImageAsyncData,
res: Result<(), ReadWriteJobError>,
) -> Result<(), VulkanError> {
if let Err(e) = res {
return Err(VulkanError::AsyncCopyToStaging(e));
if !data.data_copied.get() {
return Ok(());
}
if img.queue_state.get().acquire(QueueFamily::Transfer) == QueueTransfer::Impossible {
return Ok(());
}
img.renderer.check_defunct()?;
let regions = &*data.regions.borrow();
let staging = data.staging.get().unwrap();
staging.upload(|_, _| ())?;
let Some((cmd, fence, sync_file, point)) =
self.submit_buffer_to_image_copy(img, &staging, regions)?
self.submit_buffer_to_image_copy(img, &staging, regions, true)?
else {
return Ok(());
};
img.queue_state.set(QueueState::Releasing);
let future = img.renderer.eng.spawn(
"await async upload",
await_async_upload(point, img.clone(), cmd, fence, sync_file),
);
img.renderer.pending_uploads.set(point, future);
img.renderer.pending_submits.set(point, future);
Ok(())
}
}
@ -577,7 +703,42 @@ fn complete_async_upload(
};
let data = shm.async_data.as_ref().unwrap();
store(data);
if let Err(e) = shm.async_upload_copy_buffer_to_image(img, data, res) {
if let Err(e) = res {
data.complete(Err(VulkanError::AsyncCopyToStaging(e)));
}
data.data_copied.set(true);
if let Err(e) = shm.async_upload_copy_buffer_to_image(img, data) {
data.complete(Err(e));
}
}
async fn await_gfx_queue_release(
id: u64,
img: Rc<VulkanImage>,
buf: Option<Rc<VulkanCommandBuffer>>,
_fence: Option<Rc<VulkanFence>>,
sync_file: SyncFile,
) {
let res = img.renderer.ring.readable(&sync_file.0).await;
if let Err(e) = res {
log::error!(
"Could not wait for sync file to become readable: {}",
ErrorFmt(e)
);
img.renderer.block();
}
if let Some(buf) = buf {
img.renderer.gfx_command_buffers.buffers.push(buf);
}
img.renderer.pending_submits.remove(&id);
img.queue_state.set(QueueState::Released {
to: QueueFamily::Transfer,
});
let VulkanImageMemory::Internal(shm) = &img.ty else {
unreachable!();
};
let data = shm.async_data.as_ref().unwrap();
if let Err(e) = shm.async_upload_copy_buffer_to_image(&img, data) {
data.complete(Err(e));
}
}
@ -597,8 +758,14 @@ async fn await_async_upload(
);
img.renderer.block();
}
img.renderer.command_buffers.push(buf);
img.renderer.pending_uploads.remove(&id);
match &img.renderer.transfer_command_buffers {
Some(b) => b.buffers.push(buf),
None => img.renderer.gfx_command_buffers.buffers.push(buf),
}
img.queue_state.set(QueueState::Released {
to: QueueFamily::Gfx,
});
img.renderer.pending_submits.remove(&id);
let VulkanImageMemory::Internal(shm) = &img.ty else {
unreachable!();
};
@ -701,6 +868,8 @@ impl VulkanRenderer {
callback_id: Cell::new(0),
regions: Default::default(),
cpu: cpu.clone(),
last_sample: Default::default(),
data_copied: Default::default(),
});
}
let shm = VulkanShmImage {
@ -722,6 +891,9 @@ impl VulkanRenderer {
image,
is_undefined: Cell::new(true),
contents_are_undefined: Cell::new(true),
queue_state: Cell::new(QueueState::Acquired {
family: QueueFamily::Gfx,
}),
ty: VulkanImageMemory::Internal(shm),
bridge: None,
});