swung0x48/MobileGL
1
0
Fork 0
mirror of https://github.com/MobileGL-Dev/MobileGL synced 2026-09-12 14:18:31 +09:00
Code Issues Packages Projects Releases Wiki Activity

[Perf] (MG_Backend): key DirectVulkan's pipeline memo on state values, not a version that never repeats

Browse Source 
Two per-draw churn costs, one cause each.

A blend toggle switched pipelines through a memo keyed on a monotonic
pipeline-state version - which never repeats, so flipping GL_BLEND off and back
on produced a "new" key both times, forced the full SetupDraw and rebuilt the
whole pipeline payload for a pipeline the cache already held. The memo now keys
on a value hash of the pipeline-relevant fixed-function state, recomputed only
when the state version moved, and the consecutive-draw fast path re-resolves
just the pipeline through it when nothing but render state changed. Blaze3D
brackets every batch with exactly this toggle; mc_state_toggle drops 36%
(6629 -> 4230 ns/op, 4.8x native to 3.7x).

The sampler-churn cost had the same shape as the Espryt side fixed separately:
glBindSampler bumps the frontend texture-bind generation even when it re-binds
the sampler the unit already holds, so the per-draw fast path died every draw.
The fast path now proves each binding's descriptor inputs unchanged - texture
and sampler lifetime ids, parameter and content sums, the sampling-resolution
generation, image epochs and exact layouts - and reuses the binding's cached
VkDescriptorImageInfo instead of re-running the resolve chain.
mc_sampler_churn drops 30% (1597 -> 1125), and the proof machinery pays for
itself on the uniform-range case too (-17%).

mc_tex_param stays where it is on this backend deliberately: profiling shows its
remaining cost is frontend validation with zero backend work, unreachable from
Renderer/.

All nine cases measured on both backends, interleaved A/B, no case worse than
noise. Unit tests 421/421.
This commit is contained in:
BZLZHH
2026-08-06 11:21:26 -04:00
parent 9be5d95440
commit 4b3fd11462
4 changed files with 378 additions and 43 deletions
Download Patch File Download Diff File Expand all files Collapse all files
MobileGL/MG_Backend/DirectVulkan/Renderer/VulkanRenderer.h
+46 -3
View File
@@ -531,7 +531,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
// Per-pipeline provoking-vertex mode. capturesXfbFromGeometryStage must be a LINK-TIME
// property of the program, never the dynamic "is transform feedback active" flag: the
// 8-entry m_pipelineMemo and the SetupDrawSnapshot fast path key on programObj.hash and
// GetRenderStateParametersVersion(), neither of which moves when glBeginTransformFeedback is
// the pipeline-state value hash, neither of which moves when glBeginTransformFeedback is
// called, so a dynamic input here would hand back a stale VkPipeline.
VkProvokingVertexModeEXT SelectProvokingVertexMode(VkPrimitiveTopology topology,
Bool capturesXfbFromGeometryStage) const;
@@ -623,7 +623,12 @@ namespace MobileGL::MG_Backend::DirectVulkan {
Uint64 programHash = 0;
Uint64 vertexInputHash = 0;
Uint64 renderPassHash = 0;
Uint renderStateVersion = 0;
// VALUE hash of the pipeline-relevant fixed-function state (see
// ComputePipelineStateHash), not the monotonic pipeline-state version:
// the version never repeats, so a per-draw GL_BLEND toggle would miss
// all entries forever even though the state alternates between two
// values the memo already holds.
Uint64 pipelineStateHash = 0;
ProgramFactory::CompileOptionFlags transformFlags = {};
VkPipeline pipeline = VK_NULL_HANDLE;
};
@@ -631,11 +636,26 @@ namespace MobileGL::MG_Backend::DirectVulkan {
PipelineMemoEntry m_pipelineMemo[kPipelineMemoSize];
Uint32 m_pipelineMemoCount = 0;
Uint32 m_pipelineMemoNext = 0;
// Hash of every fixed-function GL state the pipeline payload reads that the
// memo key's other fields (mode / program / vertex input / render pass /
// transform flags) do not already pin down. Equal hash under an equal rest
// of key => byte-identical PipelineCreatePayload. Cached per pipeline-state
// version: the version is monotonic and bumps on every pipeline-state
// change, so an unchanged (version, colorAttachmentCount) proves the state
// bytes are unchanged and the hash can be reused without re-reading them.
Uint64 ComputePipelineStateHash(Uint32 colorAttachmentCount) const;
Uint m_pipelineStateHashVersion = 0;
Uint32 m_pipelineStateHashColorCount = 0;
Uint64 m_pipelineStateHash = 0;
Bool m_pipelineStateHashValid = false;
// Drops every memoized pipeline handle. Required at command-buffer
// boundaries and whenever any pipeline may have been destroyed.
// boundaries and whenever any pipeline may have been destroyed. Also drops
// the cached pipeline-state hash: the same boundaries can retire the GL
// context whose monotonic version the cache is keyed on.
void InvalidatePipelineMemo() {
m_pipelineMemoCount = 0;
m_pipelineMemoNext = 0;
m_pipelineStateHashValid = false;
}
UnorderedMap<ProgramFactory::HashType, VkPipeline> m_computePipelines;
UniquePtr<ProgramFactory> m_programFactory;
@@ -707,6 +727,16 @@ namespace MobileGL::MG_Backend::DirectVulkan {
Uint64 renderbufferImageEpoch = 0;
Uint64 sampledContentSum = 0;
Uint64 sampledParamsSum = 0;
// Guards the sampler-descriptor reuse hint: bumped by any sampler-object
// parameter or texture shape change (see GetSamplingResolutionGeneration),
// none of which the sums above cover.
Uint64 samplingResolutionGeneration = 0;
// Render-pass flavor input (DepthTest || StencilTest at snapshot time).
// A pipeline-state change that leaves this equal cannot change which
// render pass GetOrCreateRenderPass would pick, so the fast path may
// re-resolve just the pipeline against the active pass; a change that
// flips it must fall back to the full path's pass selection.
Bool drawUsesDepthStencil = false;
IntVec2 renderPassExtent = {0, 0};
VkPipeline pipeline = VK_NULL_HANDLE;
};
@@ -715,11 +745,24 @@ namespace MobileGL::MG_Backend::DirectVulkan {
// Per-draw scratch buffers (clear keeps capacity) — these paths run for every
// draw call and must not allocate.
Vector<MG_State::GLState::ITextureObject*> m_sampledTexturesScratch;
// Per-binding (texture, effective sampler) lifetime-id records from the same
// CollectSampledTextures walk that filled m_sampledTexturesScratch. The fast
// path shadow-compares against them (SampledBindingsUnchanged) when the
// texture bind generation moved, so a redundant glBindSampler/glBindTexture
// storm that resolves to the same bindings keeps the fast path.
Vector<UniformManager::SampledBindingRecord> m_sampledBindingRecordsScratch;
// Parallel to m_sampledTexturesScratch, refilled by every SetupDraw's
// first sampled-texture loop: the resolved backend resources, so the
// post-transition loop can skip re-resolving textures whose layout is
// already sampleable.
Vector<VkTextureManager::TextureResource*> m_sampledResourcesScratch;
// Layout VALUE of each sampled resource when the snapshot (and so the cached
// sampler descriptors) was built, parallel to m_sampledResourcesScratch. The
// fast path's validity check only proves the layout is still sampleable; the
// descriptor-reuse hint additionally needs it to be the SAME sampleable
// layout (a mid-frame compute dispatch can move a sampled texture from
// READ_ONLY_OPTIMAL to GENERAL, both valid, different descriptor).
Vector<VkImageLayout> m_sampledLayoutSnapshots;
Vector<MG_State::GLState::ITextureObject*> m_storageImageTexturesScratch;
Vector<VkBuffer> m_vertexBuffersScratch;
Vector<VkDeviceSize> m_vertexOffsetsScratch;