feat/xdg-foreign fdbb95e

frame.go RawHistoryFolder

1package vulkan
2
3import (
4	"fmt"
5	"image"
6	"math"
7	"os"
8	"runtime"
9	"slices"
10	"unsafe"
11
12	"github.com/bnema/neferwl/internal/ports"
13	vk "github.com/bnema/purego-vulkan/vulkan"
14)
15
16// Frames are recorded into a ring of slots, one per frame in flight. The
17// CPU never waits for a frame it just submitted: Render returns a sync
18// file the display waits on (KMS IN_FENCE_FD). A slot's fence is waited
19// only when the slot is reused, and objects a frame read (imports, GPU
20// buffers, client semaphores) are freed only once that frame completed.
21
22// frameSlots is how many frames may be in flight: one per output target.
23const frameSlots = 2
24
25// frameSlot is the command buffer and sync objects of one frame in flight.
26type frameSlot struct {
27	cmd   vk.CommandBuffer
28	fence vk.Fence
29	// done is signalled with the frame and exported as its sync file.
30	done  vk.Semaphore
31	frame uint64 // the frame recorded last, 0 before any
32	busy  bool   // submitted, fence not yet seen signalled
33}
34
35// deferredFree runs once frame completed.
36type deferredFree struct {
37	frame uint64
38	free  func()
39}
40
41func (r *Renderer) createSlots() error {
42	d := r.dd
43	cmds := make([]vk.CommandBuffer, frameSlots)
44	ai := vk.CommandBufferAllocateInfo{SType: vk.StructureTypeCommandBufferAllocateInfo, CommandPool: r.pool, Level: vk.CommandBufferLevelPrimary, CommandBufferCount: frameSlots}
45	if err := checked("vkAllocateCommandBuffers", d.AllocateCommandBuffers(r.device, &ai, &cmds[0])); err != nil {
46		return err
47	}
48	for i := range r.slots {
49		s := &r.slots[i]
50		s.cmd = cmds[i]
51		fi := vk.FenceCreateInfo{SType: vk.StructureTypeFenceCreateInfo}
52		if err := checked("vkCreateFence", d.CreateFence(r.device, &fi, nil, &s.fence)); err != nil {
53			return err
54		}
55		if err := r.newDone(s); err != nil {
56			return err
57		}
58	}
59	return nil
60}
61
62// newDone creates a slot's frame semaphore, exportable as a sync file when
63// the device allows it; a device that refuses falls back to waiting for
64// frames on the CPU (syncFD off).
65func (r *Renderer) newDone(s *frameSlot) error {
66	d := r.dd
67	si := vk.SemaphoreCreateInfo{SType: vk.StructureTypeSemaphoreCreateInfo}
68	export := vk.ExportSemaphoreCreateInfo{SType: vk.StructureTypeExportSemaphoreCreateInfo, HandleTypes: vk.ExternalSemaphoreHandleTypeSyncFDBit}
69	if r.syncFD {
70		si.Next = unsafe.Pointer(&export)
71		if d.CreateSemaphore(r.device, &si, nil, &s.done) == vk.Success {
72			return nil
73		}
74		r.syncFD = false
75		si.Next = nil
76	}
77	return checked("vkCreateSemaphore", d.CreateSemaphore(r.device, &si, nil, &s.done))
78}
79
80func (r *Renderer) destroySlots() {
81	d := r.dd
82	for i := range r.slots {
83		s := &r.slots[i]
84		if s.fence != 0 {
85			d.DestroyFence(r.device, s.fence, nil)
86		}
87		if s.done != 0 {
88			d.DestroySemaphore(r.device, s.done, nil)
89		}
90		if s.cmd != 0 {
91			d.FreeCommandBuffers(r.device, r.pool, 1, &s.cmd)
92		}
93		*s = frameSlot{}
94	}
95}
96
97// waitFrame blocks until frame n completed. Frames not submitted yet and
98// frames known done return at once. Every slot holding a frame up to n
99// is waited and reset, so its fence is unsignalled before reuse.
100func (r *Renderer) waitFrame(n uint64) error {
101	n = min(n, r.submitted)
102	if n <= r.completed {
103		return nil
104	}
105	for i := range r.slots {
106		s := &r.slots[i]
107		if !s.busy || s.frame > n {
108			continue
109		}
110		if err := checked("vkWaitForFences", r.dd.WaitForFences(r.device, 1, &s.fence, 1, math.MaxUint64)); err != nil {
111			return err
112		}
113		if err := checked("vkResetFences", r.dd.ResetFences(r.device, 1, &s.fence)); err != nil {
114			return err
115		}
116		s.busy = false
117	}
118	r.completed = n
119	r.runDeferred()
120	return nil
121}
122
123// retire frees an object once the frame that last used it completed.
124func (r *Renderer) retire(frame uint64, free func()) {
125	if frame <= r.completed {
126		free()
127		return
128	}
129	r.deferred = append(r.deferred, deferredFree{frame: frame, free: free})
130}
131
132func (r *Renderer) runDeferred() {
133	r.deferred = slices.DeleteFunc(r.deferred, func(f deferredFree) bool {
134		if f.frame <= r.completed {
135			f.free()
136			return true
137		}
138		return false
139	})
140}
141
142// idle records that the device finished every submitted frame.
143func (r *Renderer) idle() {
144	for i := range r.slots {
145		s := &r.slots[i]
146		if s.busy {
147			_ = r.dd.ResetFences(r.device, 1, &s.fence)
148			s.busy = false
149		}
150	}
151	// Frames recorded but never submitted have no GPU work either.
152	r.completed = r.frame
153	r.runDeferred()
154}
155
156// Render draws a frame into the current target. done is signalled when
157// the GPU finished it; nil when the device cannot export a sync file,
158// then the frame is finished on return.
159func (r *Renderer) Render(s ports.Scene, contents map[ports.WindowID]ports.SurfaceContent) (done *os.File, err error) {
160	// A frame that fails before its submit keeps its number: objects it
161	// touched are freed once a later frame completes (never early).
162	r.frame++
163	slot := &r.slots[r.frame%frameSlots]
164	// The slot's previous frame must be done before its command buffer
165	// is recorded again.
166	if slot.busy {
167		if err := r.waitFrame(slot.frame); err != nil {
168			return nil, err
169		}
170	}
171	tg := r.target()
172	hdr := r.hdrNits > 0 && len(r.targets) > 0
173	if r.hdrNits > 0 && !hdr {
174		return nil, fmt.Errorf("HDR composition requires an exported target")
175	}
176	if hdr {
177		tg = &r.hdrOwn
178	}
179	dmg := r.frameDamage(tg, s, image.Rect(0, 0, r.width, r.height))
180	ds := r.draws(s, contents, dmg)
181	dmg.finish()
182	r.dropPools()
183	partial := !dmg.all()
184	if partial {
185		// The pass keeps the image: repaint the background under the
186		// region, then everything clipped to it.
187		// Keep the background first without allocating a second draw slice.
188		ds = append(ds, draw{})
189		copy(ds[1:], ds[:len(ds)-1])
190		ds[0] = r.fillDraw(dmg.area, parseColor(s.Background))
191		ds = dmg.clip(ds)
192		r.scratchDraws = ds
193		r.redrawn += dmg.area.Dx() * dmg.area.Dy()
194	} else {
195		r.redrawn += r.width * r.height
196	}
197	d := r.dd
198	cmd := slot.cmd
199	if err := checked("vkResetCommandBuffer", d.ResetCommandBuffer(cmd, 0)); err != nil {
200		return nil, err
201	}
202	begin := vk.CommandBufferBeginInfo{SType: vk.StructureTypeCommandBufferBeginInfo, Flags: vk.CommandBufferUsageOneTimeSubmitBit}
203	if err := checked("vkBeginCommandBuffer", d.BeginCommandBuffer(cmd, &begin)); err != nil {
204		return nil, err
205	}
206	// A full redraw clears the target: its old contents never matter
207	// (UNDEFINED). A partial one keeps them.
208	b := vk.ImageMemoryBarrier{SType: vk.StructureTypeImageMemoryBarrier, DstAccessMask: vk.AccessColorAttachmentWriteBit, OldLayout: vk.ImageLayoutUndefined, NewLayout: vk.ImageLayoutColorAttachmentOptimal, SrcQueueFamilyIndex: queueFamilyIgnored, DstQueueFamilyIndex: queueFamilyIgnored, Image: tg.image, SubresourceRange: colorRange}
209	if partial {
210		b.OldLayout = tg.layout
211		b.DstAccessMask |= vk.AccessColorAttachmentReadBit
212	}
213	if tg.exported {
214		// Acquire from the display (foreign queue family).
215		b.SrcQueueFamilyIndex, b.DstQueueFamilyIndex = vk.QueueFamilyForeignEXT, r.family
216	}
217	d.CmdPipelineBarrier(cmd, vk.PipelineStageTransferBit|vk.PipelineStageColorAttachmentOutputBit, vk.PipelineStageColorAttachmentOutputBit, 0, 0, nil, 0, nil, 1, &b)
218	// Client buffers come from the foreign queue family (their driver)
219	// and go back to it after the frame, so the next frame acquires them.
220	dmas := r.frameDMAs[:0]
221	if r.frameAcquires == nil {
222		r.frameAcquires = make(map[*imported]*os.File)
223	}
224	clear(r.frameAcquires)
225	acquires := r.frameAcquires
226	for _, dr := range ds {
227		if dr.im != nil && !slices.Contains(dmas, dr.im) {
228			dmas = append(dmas, dr.im)
229			dr.im.last = r.frame
230		}
231		if dr.acquire != nil {
232			acquires[dr.im] = dr.acquire
233		}
234	}
235	r.frameDMAs = dmas
236	r.ownership(cmd, dmas, true)
237	r.recordDraws(cmd, tg.view, parseColor(s.Background), ds, partial)
238	r.ownership(cmd, dmas, false)
239	if hdr {
240		r.recordHDR(cmd, r.targets[r.current])
241	}
242	if !hdr {
243		b.SrcAccessMask, b.OldLayout = vk.AccessColorAttachmentWriteBit, vk.ImageLayoutColorAttachmentOptimal
244		b.SrcQueueFamilyIndex, b.DstQueueFamilyIndex = queueFamilyIgnored, queueFamilyIgnored
245		if tg.exported {
246			// Release to the display: KMS scans it out after the fence.
247			b.NewLayout, b.DstAccessMask = vk.ImageLayoutGeneral, 0
248			b.SrcQueueFamilyIndex, b.DstQueueFamilyIndex = r.family, vk.QueueFamilyForeignEXT
249			d.CmdPipelineBarrier(cmd, vk.PipelineStageColorAttachmentOutputBit, vk.PipelineStageBottomOfPipeBit, 0, 0, nil, 0, nil, 1, &b)
250		} else {
251			b.NewLayout, b.DstAccessMask = vk.ImageLayoutTransferSrcOptimal, vk.AccessTransferReadBit
252			d.CmdPipelineBarrier(cmd, vk.PipelineStageColorAttachmentOutputBit, vk.PipelineStageTransferBit, 0, 0, nil, 0, nil, 1, &b)
253		}
254	}
255	if err := checked("vkEndCommandBuffer", d.EndCommandBuffer(cmd)); err != nil {
256		return nil, err
257	}
258	// Wait for the client's GPU writes to each buffer: its explicit
259	// acquire fence, else the buffer's implicit fences.
260	waits := r.frameWaits[:0]
261	for _, im := range dmas {
262		if f := acquires[im]; f != nil {
263			if sem := r.importFence(f); sem != 0 {
264				waits = append(waits, sem)
265			}
266		} else {
267			waits = r.appendReadFences(waits, im)
268		}
269	}
270	r.frameWaits = waits
271	frame := r.frame
272	for _, sem := range waits {
273		sem := sem
274		r.retire(frame, func() { d.DestroySemaphore(r.device, sem, nil) })
275	}
276	submit := vk.SubmitInfo{SType: vk.StructureTypeSubmitInfo, CommandBufferCount: 1, CommandBuffers: &cmd}
277	if cap(r.frameStages) < len(waits) {
278		r.frameStages = make([]vk.PipelineStageFlags, len(waits))
279	}
280	stages := r.frameStages[:len(waits)]
281	if len(waits) > 0 {
282		for i := range stages {
283			stages[i] = vk.PipelineStageFragmentShaderBit
284		}
285		submit.WaitSemaphoreCount, submit.WaitSemaphores, submit.WaitDstStageMask = uint32(len(waits)), &waits[0], &stages[0]
286	}
287	if r.syncFD {
288		submit.SignalSemaphoreCount, submit.SignalSemaphores = 1, &slot.done
289	}
290	if err := checked("vkQueueSubmit", d.QueueSubmit(r.queue, 1, &submit, slot.fence)); err != nil {
291		return nil, err
292	}
293	runtime.KeepAlive(stages)
294	runtime.KeepAlive(waits)
295	slot.frame, slot.busy, r.submitted = frame, true, frame
296	if hdr {
297		// recordHDR left the composed linear image in TRANSFER_SRC for capture.
298		tg.layout = vk.ImageLayoutTransferSrcOptimal
299		r.targets[r.current].layout = vk.ImageLayoutGeneral
300	} else {
301		tg.layout = b.NewLayout
302	}
303	oldWindows := tg.windows
304	tg.hold(s, dmg)
305	r.damageDrawn = oldWindows
306	r.last = tg
307	r.dropUnused()
308	r.dropShm()
309	if !r.syncFD {
310		return nil, r.waitFrame(frame)
311	}
312	var fd int32
313	get := vk.SemaphoreGetFdInfoKHR{SType: vk.StructureTypeSemaphoreGetFDInfoKHR, Semaphore: slot.done, HandleType: vk.ExternalSemaphoreHandleTypeSyncFDBit}
314	if d.GetSemaphoreFdKHR(r.device, &get, &fd) != vk.Success {
315		// The frame is drawn: finish it on the CPU. The semaphore stays
316		// signalled: replace it, so the slot's next submit signals a
317		// fresh one.
318		if err := r.waitFrame(frame); err != nil {
319			return nil, err
320		}
321		d.DestroySemaphore(r.device, slot.done, nil)
322		slot.done = 0
323		return nil, r.newDone(slot)
324	}
325	if fd < 0 {
326		// Already signalled: the driver may return -1.
327		return nil, r.waitFrame(frame)
328	}
329	f := os.NewFile(uintptr(fd), "neferwl-frame")
330	if f == nil {
331		return nil, fmt.Errorf("invalid frame sync file %d", fd)
332	}
333	return f, nil
334}
335
336// ownership moves client images between their driver (foreign queue
337// family, GENERAL: their contents survive) and the fragment shader.
338func (r *Renderer) ownership(cmd vk.CommandBuffer, dmas []*imported, acquire bool) {
339	for _, im := range dmas {
340		b := vk.ImageMemoryBarrier{SType: vk.StructureTypeImageMemoryBarrier, Image: im.image, SubresourceRange: colorRange}
341		if im.yuv {
342			b.SubresourceRange.AspectMask = aspectPlane0 | aspectPlane1
343		}
344		if acquire {
345			b.OldLayout, b.NewLayout = vk.ImageLayoutGeneral, vk.ImageLayoutShaderReadOnlyOptimal
346			b.SrcQueueFamilyIndex, b.DstQueueFamilyIndex = vk.QueueFamilyForeignEXT, r.family
347			b.DstAccessMask = vk.AccessShaderReadBit
348			// The frame waits on the client's fence at the fragment
349			// stage: the transfer must come after that wait.
350			r.dd.CmdPipelineBarrier(cmd, vk.PipelineStageFragmentShaderBit, vk.PipelineStageFragmentShaderBit, 0, 0, nil, 0, nil, 1, &b)
351		} else {
352			b.OldLayout, b.NewLayout = vk.ImageLayoutShaderReadOnlyOptimal, vk.ImageLayoutGeneral
353			b.SrcQueueFamilyIndex, b.DstQueueFamilyIndex = r.family, vk.QueueFamilyForeignEXT
354			b.SrcAccessMask = vk.AccessShaderReadBit
355			r.dd.CmdPipelineBarrier(cmd, vk.PipelineStageFragmentShaderBit, vk.PipelineStageBottomOfPipeBit, 0, 0, nil, 0, nil, 1, &b)
356		}
357	}
358}
359
360// oneShot records and runs commands outside the frame ring and waits for
361// them (setup and screenshots only, never per frame).
362func (r *Renderer) oneShot(record func(vk.CommandBuffer)) error {
363	d := r.dd
364	var cmd vk.CommandBuffer
365	ai := vk.CommandBufferAllocateInfo{SType: vk.StructureTypeCommandBufferAllocateInfo, CommandPool: r.pool, Level: vk.CommandBufferLevelPrimary, CommandBufferCount: 1}
366	if err := checked("vkAllocateCommandBuffers", d.AllocateCommandBuffers(r.device, &ai, &cmd)); err != nil {
367		return err
368	}
369	defer d.FreeCommandBuffers(r.device, r.pool, 1, &cmd)
370	var fence vk.Fence
371	fi := vk.FenceCreateInfo{SType: vk.StructureTypeFenceCreateInfo}
372	if err := checked("vkCreateFence", d.CreateFence(r.device, &fi, nil, &fence)); err != nil {
373		return err
374	}
375	defer d.DestroyFence(r.device, fence, nil)
376	begin := vk.CommandBufferBeginInfo{SType: vk.StructureTypeCommandBufferBeginInfo, Flags: vk.CommandBufferUsageOneTimeSubmitBit}
377	if err := checked("vkBeginCommandBuffer", d.BeginCommandBuffer(cmd, &begin)); err != nil {
378		return err
379	}
380	record(cmd)
381	if err := checked("vkEndCommandBuffer", d.EndCommandBuffer(cmd)); err != nil {
382		return err
383	}
384	submit := vk.SubmitInfo{SType: vk.StructureTypeSubmitInfo, CommandBufferCount: 1, CommandBuffers: &cmd}
385	if err := checked("vkQueueSubmit", d.QueueSubmit(r.queue, 1, &submit, fence)); err != nil {
386		return err
387	}
388	return checked("vkWaitForFences", d.WaitForFences(r.device, 1, &fence, 1, math.MaxUint64))
389}
390
391// Clear draws a background-only frame and waits for it.
392func (r *Renderer) Clear(rgb [3]uint8) error {
393	done, err := r.Render(ports.Scene{Background: fmt.Sprintf("#%02x%02x%02x", rgb[0], rgb[1], rgb[2])}, nil)
394	if done != nil {
395		done.Close()
396	}
397	if err != nil {
398		return err
399	}
400	return r.waitFrame(r.frame)
401}

Keyboard shortcuts

Single-key shortcuts are on. ctrl k and esc work whatever this says.

ctrl k :
go to a section, branch, tag, file or repository
?
this list
j k
move the row cursor down / up
enter
open the row under the cursor; on Files, a directory opens or closes in place
↑ ↓ ← → home end
in the file tree (focus on a row): previous / next row, close / open a directory or go to its parent / first entry, first / last row. They never open a file
l h
on Files: open a directory or a file / close a directory or go up to its parent, then to the parent page. Elsewhere: into the side panel (on a merge request, its files) / back to the content; there j k move between its controls
g g G
first / last row
/
filter the list; on Files, find a file anywhere in the repository (esc clears, enter opens the best match)
b
switch branch or tag
g s g f g c g b g m g i g r g a
tabs: summary, files, commits, branches & tags, merge requests, issues, releases, actions (when available)
g h
home
w
wide mode: use the whole width of the screen, or not (remembered)
y
copy a permalink pinned to the commit
d u x
diff: split / unified / collapse or expand all
e E
diff: the whole file in view / every file, with the lines the hunks hide
n
merge request: next note in the diff
esc
close a dialog, leave a field

Branches 23

Tags 10

↑↓ moveenter openesc close