Skip to content

Commit 68ebe50

Browse files
nyorainnyorain
committed
Improve vertex viewer
1 parent e91df50 commit 68ebe50

21 files changed

+395
-251
lines changed

‎docs/own/done.md‎

Lines changed: 1 addition & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -1,3 +1,4 @@
1+
- [x] fix vulkan samples vertex input issues (e.g. try gears)
12
- [x] Allow to open where left off?
23
Would require some serialization of frames/commands/resources.
34
Also, would have to completely rework matching to work with those loaded up

‎docs/own/ideas.md‎

Lines changed: 5 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -1,6 +1,11 @@
11
Moved from todo.md. Mostly ideas for experiments that might not even
22
be possible or useful in the end.
33

4+
- Dev window
5+
- *everything* in UI code can output custom stuff into it, just
6+
use tabs/collapsing headers
7+
- Also allows to expose options to be changed on the fly
8+
49
- show custom pipeline stats in the resource viewer using
510
https://registry.khronos.org/vulkan/specs/latest/man/html/VK_KHR_pipeline_executable_properties.html
611

‎docs/own/workstack.md‎

Lines changed: 7 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -1,6 +1,12 @@
1-
- [ ] fix vulkan samples vertex input issues (e.g. try gears)
21
- [ ] fix vulkan samples "Losing open cmd" with trianglevulkan13
32

3+
- [ ] Cleanup "Show All" mess in vertex viewer
4+
- [ ] additionally (orthogonally) have a checkbox that decides
5+
whether output needs to be shown in table, be copied to host
6+
Name it something like "Show Vertex Table (slow)"?
7+
- [ ] optimization: When "Show All" is not active, split up draw calls
8+
Already outlined in vertexCopy.md
9+
410
- [ ] rename vku to vpp?
511
- [ ] Textual "bound descriptor sets" inspector
612
- [ ] Allow to inspect before/after BindDescriptorSets/PushDescriptorSet

‎src/commandHook/hook.cpp‎

Lines changed: 6 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -1079,7 +1079,6 @@ void CommandHook::updateHook(Update&& update) {
10791079

10801080
if (update.newTarget || update.newOps) {
10811081
std::lock_guard lock(dev_->mutex);
1082-
hints_ = {};
10831082

10841083
if(update.newTarget) {
10851084
oldTarget = std::move(target_);
@@ -1102,6 +1101,12 @@ void CommandHook::updateHook(Update&& update) {
11021101
if(update.invalidate) {
11031102
clearCompleted();
11041103
invalidateRecordings();
1104+
1105+
{
1106+
dlg_trace("invalidate Hook");
1107+
std::lock_guard lock(dev_->mutex);
1108+
hints_ = {};
1109+
}
11051110
}
11061111
}
11071112

‎src/commandHook/hook.hpp‎

Lines changed: 1 addition & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -68,9 +68,8 @@ struct CommandHookOps {
6868
// When updating e.g. the id of the ds to be copied, all existing
6969
// recordings have to be invalidated!
7070
bool copyVertexInput {};
71-
u32 vertexInputCmd {}; // for multi draw, specifies the command for which to copy
72-
7371
bool copyXfb {}; // transform feedback
72+
u32 vertexCmd {}; // for multi draw, specifies the command for which to copy
7473

7574
IntrusivePtr<ShaderCaptureHook> shaderCapture {};
7675
Vec3u32 shaderCaptureInput {}; // globalThreadID/vertexID/...

‎src/commandHook/record.cpp‎

Lines changed: 76 additions & 37 deletions
Original file line numberDiff line numberDiff line change
@@ -18,6 +18,7 @@
1818
#include <rp.hpp>
1919
#include <ds.hpp>
2020
#include <vk/format_utils.h>
21+
#include <numeric>
2122

2223
namespace vil {
2324

@@ -429,20 +430,32 @@ void CommandHookRecord::hookRecordAfterDst(Command& dst, RecordInfo& info) {
429430
}
430431

431432
u32 CommandHookRecord::vertexCountHint(const DrawCmdBase& bcmd, const RecordInfo& info) const {
433+
// NOTE: for xfb hints, we always use vertexCountHint
434+
432435
if(auto* dcmd = commandCast<const DrawCmd*>(&bcmd); dcmd) {
433436
return dcmd->vertexCount;
434437
} else if(auto* dcmd = commandCast<const DrawIndexedCmd*>(&bcmd); dcmd) {
435438
return dcmd->indexCount;
436439
} else if(auto* dcmd = commandCast<const DrawIndirectCmd*>(&bcmd); dcmd) {
437-
return dcmd->isIndexed() ? info.hints.indexCountHint : info.hints.vertexCountHint;
440+
return info.hints.vertexCountHint;
438441
} else if(auto* dcmd = commandCast<const DrawIndirectCountCmd*>(&bcmd); dcmd) {
439-
return dcmd->isIndexed() ? info.hints.indexCountHint : info.hints.vertexCountHint;
442+
return info.hints.vertexCountHint;
440443
} else if(auto* dcmd = commandCast<const DrawMultiCmd*>(&bcmd); dcmd) {
441-
auto& drawInfo = dcmd->vertexInfos[info.ops.vertexInputCmd];
442-
return drawInfo.vertexCount;
444+
auto cmds = dcmd->vertexInfos;
445+
if(info.ops.vertexCmd != u32(-1)) {
446+
auto count = std::min(info.ops.vertexCmd + 1, u32(cmds.size()));
447+
cmds = cmds.first(count);
448+
}
449+
return std::accumulate(cmds.begin(), cmds.end(), 0u,
450+
[](auto& a, auto& b) { return a + b.vertexCount; });
443451
} else if(auto* dcmd = commandCast<const DrawMultiIndexedCmd*>(&bcmd); dcmd) {
444-
auto& drawInfo = dcmd->indexInfos[info.ops.vertexInputCmd];
445-
return drawInfo.indexCount;
452+
auto cmds = dcmd->indexInfos;
453+
if(info.ops.vertexCmd != u32(-1)) {
454+
auto count = std::min(info.ops.vertexCmd + 1, u32(cmds.size()));
455+
cmds = cmds.first(count);
456+
}
457+
return std::accumulate(cmds.begin(), cmds.end(), 0u,
458+
[](auto& a, auto& b) { return a + b.indexCount; });
446459
} else {
447460
dlg_fatal("Unsupported draw command");
448461
return 0u;
@@ -473,12 +486,17 @@ void CommandHookRecord::hookRecordDst(Command& cmd, RecordInfo& info) {
473486
vertexCount = 1 * 1024 * 1024;
474487
}
475488

489+
// TODO: for large captures, allow only capturing position
490+
auto memType = info.ops.vertexCmd == u32(-1) ?
491+
OwnBuffer::Type::deviceLocal : OwnBuffer::Type::hostVisible;
492+
476493
const auto xfbSize = drawCmd->state->pipe->xfbPatch->stride * vertexCount;
477494
const auto usage =
478495
VK_BUFFER_USAGE_TRANSFER_DST_BIT |
479496
VK_BUFFER_USAGE_TRANSFORM_FEEDBACK_BUFFER_BIT_EXT |
480497
VK_BUFFER_USAGE_VERTEX_BUFFER_BIT;
481-
state->transformFeedback.ensure(dev, xfbSize, usage);
498+
state->transformFeedback.ensure(dev, xfbSize, usage,
499+
{}, {}, memType);
482500

483501
auto offset = VkDeviceSize(0u);
484502
dev.dispatch.CmdBindTransformFeedbackBuffersEXT(cb, 0u, 1u,
@@ -1221,11 +1239,7 @@ void CommandHookRecord::copyVertexInput(Command& bcmd, RecordInfo& info) {
12211239
// later on so we have to care about queue families
12221240
auto queueFams = combineQueueFamilies({{record->queueFamily, dev.gfxQueue->family}});
12231241

1224-
// See vertexCopy.md for more information on this beast
1225-
// TODO:
1226-
// - we currently assume that index buffer offsets are aligned
1227-
// to allow binding them as storage buffers. Can't assume that!
1228-
// - missing barriers!
1242+
// See vertexCopy.md for more (possibly outdated) information on this beast
12291243

12301244
using Metadata = VertexCopyMetadata;
12311245

@@ -1251,7 +1265,10 @@ void CommandHookRecord::copyVertexInput(Command& bcmd, RecordInfo& info) {
12511265

12521266
auto stage0 = VK_PIPELINE_STAGE_BOTTOM_OF_PIPE_BIT;
12531267
dev.dispatch.CmdWriteTimestamp(cb, stage0, this->queryPool, timeStampID);
1254-
dlg_trace(" vertexCopy timeStamp {}: {}", timeStampID, name);
1268+
1269+
if(dev.printVertexCaptureTimings) {
1270+
dlg_trace(" vertexCopy timeStamp {}: {}", timeStampID, name);
1271+
}
12551272

12561273
ownTimingNames.push_back(name);
12571274

@@ -1395,7 +1412,6 @@ void CommandHookRecord::copyVertexInput(Command& bcmd, RecordInfo& info) {
13951412

13961413
u32 vertexHint = info.hints.vertexCountHint;
13971414
if(vertexHint == 0u) {
1398-
vertexHint = fallbackVertexCountHint;
13991415
vertexHint = std::max(vertexHint, numIndices); // for indexed drawing
14001416
}
14011417

@@ -1428,13 +1444,15 @@ void CommandHookRecord::copyVertexInput(Command& bcmd, RecordInfo& info) {
14281444
VK_BUFFER_USAGE_STORAGE_BUFFER_BIT);
14291445

14301446
// write copyType and other meta information
1431-
auto wbuf = state->indexBufCopy.writeData();
14321447
Metadata md {};
1433-
md.copyTypeOrIndexOffset = copyTypeVertices;
1448+
md.firstIndex = copyTypeVertices;
14341449
md.firstVertex = params.firstVertex;
14351450
md.firstInstance = params.firstInstance;
1436-
nytl::write(wbuf, md);
1437-
state->indexBufCopy.flushMap();
1451+
dev.dispatch.CmdUpdateBuffer(cb, state->indexBufCopy.buf, 0u,
1452+
sizeof(md), &md);
1453+
vku::cmdBarrier(dev, cb, state->indexBufCopy.asSpan(),
1454+
vku::SyncScope::transferWrite(),
1455+
vku::SyncScope::computeReadWrite());
14381456

14391457
copyVertexBufs(false, params.vertexCount, params.instanceCount, VK_INDEX_TYPE_NONE_KHR);
14401458
};
@@ -1448,6 +1466,8 @@ void CommandHookRecord::copyVertexInput(Command& bcmd, RecordInfo& info) {
14481466
hook.processIndices32_ :
14491467
hook.processIndices16_;
14501468

1469+
dlg_assert(cmd->state->indices.offset % indSize == 0u);
1470+
14511471
auto& dst = state->indexBufCopy;
14521472
const auto usage =
14531473
VK_BUFFER_USAGE_STORAGE_BUFFER_BIT |
@@ -1457,9 +1477,9 @@ void CommandHookRecord::copyVertexInput(Command& bcmd, RecordInfo& info) {
14571477
usage, queueFams);
14581478

14591479
// init meta information
1460-
auto wbuf = dst.writeData();
14611480
Metadata md {};
1462-
md.copyTypeOrIndexOffset = params.firstIndex; // indexOffset
1481+
md.firstIndex = params.firstIndex; // indexOffset
1482+
md.indexBufOffset = cmd->state->indices.offset / indSize;
14631483
md.minIndex = params.indexCount > 0 ? 0xFFFFFFFFu : 0u;
14641484
md.maxIndex = 0u;
14651485
md.indexCount = params.indexCount;
@@ -1470,23 +1490,27 @@ void CommandHookRecord::copyVertexInput(Command& bcmd, RecordInfo& info) {
14701490
md.dispatchPerVertexZ = 1u;
14711491
md.firstInstance = params.firstInstance;
14721492
md.firstVertex = params.vertexOffset;
1473-
nytl::write(wbuf, md);
1474-
dst.flushMap();
1493+
dev.dispatch.CmdUpdateBuffer(cb, dst.buf, 0u,
1494+
sizeof(md), &md);
1495+
vku::cmdBarrier(dev, cb, state->indexBufCopy.asSpan(),
1496+
vku::SyncScope::transferWrite(),
1497+
vku::SyncScope::computeReadWrite());
14751498

14761499
auto& ds = allocDs(indexPipe);
14771500
{
14781501
vku::DescriptorUpdate dsu(ds);
1479-
// TODO: respect storage buf alignment
1480-
dsu(vku::BufferSpan{cmd->state->indices.buffer->handle,
1481-
{cmd->state->indices.offset, VK_WHOLE_SIZE}});
1502+
dsu(vku::BufferSpan{cmd->state->indices.buffer->handle, {0u, VK_WHOLE_SIZE}});
14821503
dsu(vku::BufferSpan{dst.buf, {0u, VK_WHOLE_SIZE}});
14831504
}
14841505

1506+
constexpr auto indicesPerInvoc = 8u; // see vertexCopy.glsl
1507+
const auto groupsX = ceilDivide(params.indexCount, 64u * indicesPerInvoc);
1508+
14851509
dev.dispatch.CmdBindPipeline(cb, VK_PIPELINE_BIND_POINT_COMPUTE,
14861510
indexPipe.pipe());
14871511
dev.dispatch.CmdBindDescriptorSets(cb, VK_PIPELINE_BIND_POINT_COMPUTE,
14881512
indexPipe.pipeLayout().vkHandle(), 0u, 1u, &ds.vkHandle(), 0u, nullptr);
1489-
dev.dispatch.CmdDispatch(cb, ceilDivide(params.indexCount, 64u), 1u, 1u);
1513+
dev.dispatch.CmdDispatch(cb, groupsX, 1u, 1u);
14901514
cmdBarrierCompute(dev, cb, state->indexBufCopy);
14911515

14921516
cmdTimestamp("after_processIndices");
@@ -1527,7 +1551,13 @@ void CommandHookRecord::copyVertexInput(Command& bcmd, RecordInfo& info) {
15271551
dev.dispatch.CmdBindDescriptorSets(cb, VK_PIPELINE_BIND_POINT_COMPUTE,
15281552
cmdPipe.pipeLayout().vkHandle(), 0u, 1u, &ds.vkHandle(), 0u, nullptr);
15291553
dlg_assert(cmdStride % 4u == 0u);
1530-
u32 pcr[] = {cmdBufOffsetShader, countBufOffset, info.ops.vertexInputCmd, cmdStride / 4u};
1554+
1555+
u32 pcr[] = {
1556+
cmdBufOffsetShader,
1557+
countBufOffset,
1558+
info.ops.vertexCmd,
1559+
cmdStride / 4u
1560+
};
15311561
dev.dispatch.CmdPushConstants(cb, cmdPipe.pipeLayout().vkHandle(),
15321562
VK_SHADER_STAGE_COMPUTE_BIT, 0u, sizeof(pcr), pcr);
15331563
dev.dispatch.CmdDispatch(cb, 1u, 1u, 1u);
@@ -1552,7 +1582,8 @@ void CommandHookRecord::copyVertexInput(Command& bcmd, RecordInfo& info) {
15521582
numIndices = fallbackIndexCountHint;
15531583
}
15541584

1555-
state->indexBufCopy.ensure(dev, sizeof(Metadata) + numIndices * indexSize(indexType),
1585+
const auto indSize = indexSize(indexType);
1586+
state->indexBufCopy.ensure(dev, sizeof(Metadata) + numIndices * indSize,
15561587
VK_BUFFER_USAGE_STORAGE_BUFFER_BIT |
15571588
VK_BUFFER_USAGE_INDEX_BUFFER_BIT |
15581589
VK_BUFFER_USAGE_INDIRECT_BUFFER_BIT);
@@ -1579,17 +1610,26 @@ void CommandHookRecord::copyVertexInput(Command& bcmd, RecordInfo& info) {
15791610
dsu(countBufShader);
15801611
}
15811612

1582-
dlg_trace("countBufOffset: {}", countBufOffset);
1583-
dlg_trace("drawID: {}", info.ops.vertexInputCmd);
1613+
// dlg_trace("countBufOffset: {}", countBufOffset);
1614+
// dlg_trace("drawID: {}", info.ops.vertexCmd);
15841615

15851616
dsu.apply();
15861617

15871618
dev.dispatch.CmdBindPipeline(cb, VK_PIPELINE_BIND_POINT_COMPUTE,
15881619
cmdPipe.pipe());
15891620
dev.dispatch.CmdBindDescriptorSets(cb, VK_PIPELINE_BIND_POINT_COMPUTE,
15901621
cmdPipe.pipeLayout().vkHandle(), 0u, 1u, &ds.vkHandle(), 0u, nullptr);
1622+
15911623
dlg_assert(cmdStride % 4u == 0u);
1592-
u32 pcr[] = {cmdBufOffsetShader, countBufOffset, info.ops.vertexInputCmd, cmdStride / 4u};
1624+
dlg_assert(cmd->state->indices.offset % indSize == 0u);
1625+
1626+
u32 pcr[] = {
1627+
cmdBufOffsetShader,
1628+
countBufOffset,
1629+
info.ops.vertexCmd,
1630+
cmdStride / 4u,
1631+
u32(cmd->state->indices.offset / indSize),
1632+
};
15931633
dev.dispatch.CmdPushConstants(cb, cmdPipe.pipeLayout().vkHandle(),
15941634
VK_SHADER_STAGE_COMPUTE_BIT, 0u, sizeof(pcr), pcr);
15951635
dev.dispatch.CmdDispatch(cb, 1u, 1u, 1u);
@@ -1608,8 +1648,7 @@ void CommandHookRecord::copyVertexInput(Command& bcmd, RecordInfo& info) {
16081648
vku::DescriptorUpdate dsu(ds);
16091649

16101650
auto& inds = cmd->state->indices;
1611-
// TODO: respect storage buffer alignment
1612-
dsu(vku::BufferSpan{inds.buffer->handle, {inds.offset, VK_WHOLE_SIZE}});
1651+
dsu(vku::BufferSpan{inds.buffer->handle, {0u, VK_WHOLE_SIZE}});
16131652
dsu(state->indexBufCopy.asSpan());
16141653
dsu.apply();
16151654

@@ -1665,10 +1704,10 @@ void CommandHookRecord::copyVertexInput(Command& bcmd, RecordInfo& info) {
16651704
} else if(auto* dcmd = commandCast<DrawMultiCmd*>(&bcmd); dcmd) {
16661705
cmd = dcmd;
16671706

1668-
dlg_assertm_or(info.ops.vertexInputCmd < dcmd->vertexInfos.size(), return,
1707+
dlg_assertm_or(info.ops.vertexCmd < dcmd->vertexInfos.size(), return,
16691708
"Command to copy out-of-range");
16701709

1671-
auto& drawInfo = dcmd->vertexInfos[info.ops.vertexInputCmd];
1710+
auto& drawInfo = dcmd->vertexInfos[info.ops.vertexCmd];
16721711

16731712
VkDrawIndirectCommand params;
16741713
params.vertexCount = drawInfo.vertexCount;
@@ -1680,10 +1719,10 @@ void CommandHookRecord::copyVertexInput(Command& bcmd, RecordInfo& info) {
16801719
} else if(auto* dcmd = commandCast<DrawMultiIndexedCmd*>(&bcmd); dcmd) {
16811720
cmd = dcmd;
16821721

1683-
dlg_assertm_or(info.ops.vertexInputCmd < dcmd->indexInfos.size(), return,
1722+
dlg_assertm_or(info.ops.vertexCmd < dcmd->indexInfos.size(), return,
16841723
"Command to copy out-of-range");
16851724

1686-
auto& drawInfo = dcmd->indexInfos[info.ops.vertexInputCmd];
1725+
auto& drawInfo = dcmd->indexInfos[info.ops.vertexCmd];
16871726

16881727
VkDrawIndexedIndirectCommand params;
16891728
params.indexCount = drawInfo.indexCount;

‎src/commandHook/record.hpp‎

Lines changed: 5 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -132,10 +132,14 @@ struct CommandHookRecord {
132132
u32 dispatchPerInstanceZ;
133133
u32 firstInstance;
134134

135+
u32 firstIndex;
135136
u32 indexCount;
136137
u32 minIndex;
137138
u32 maxIndex;
138-
u32 copyTypeOrIndexOffset;
139+
140+
u32 pad_[2];
141+
u32 indexBufOffset; // divided by indexSize
142+
u32 copyType;
139143
};
140144

141145
static constexpr auto copyTypeVertices = 1u;

0 commit comments

Comments
 (0)