ScopedDefaultUnpackState saved the backend GL unpack state with 6 glGetIntegerv
calls on every construction. glGetIntegerv forces a driver pipeline sync, and
because it ran per dirty texture per frame in the texture upload path, it
dominated the DirectGLES draw path - and stalling the pipeline serialized CPU-GPU
work far beyond its raw CPU cost.
The backend unpack state is set only by MobileGL's own save/restore helpers
(ScopedDefaultUnpackState, TempPixelStoreParameterSync, the R32F copy path), all
of which restore to the resting GL default, so it can be shadow-tracked: read the
previous state from a static shadow (no query), pin the backend to the known
default once up front, and set state with compare-and-set so the paired
glPixelStorei calls also usually no-op.
Device-verified on Adreno 830 (MC 26.3-snapshot3, Espryt, CPU pinned to 1.56/1.96
GHz for a thermally-comparable measurement): rendering correct; fps 105 -> 147
(+40%); render-thread profile: glGetIntegerv ~9% -> below noise, SyncNeccessary-
Textures 25% -> 12%, SyncMipmapsToBackend 23% -> 9%.
Introduce a Mesa pipe_resource-style PipeResource that owns a GL buffer's bytes
and its backend GPU resource, abstracting WHERE the authoritative bytes live:
- Shadow mode (non-persistent buffers): a CPU Vector; the backend keeps its own
GPU copy in sync via BufferBackendOps, exactly as before.
- Persistent mode (coherent GL_MAP_PERSISTENT maps): the backend's host-visible,
COHERENT, persistently-mapped GPU memory is the single source of truth. The app
writes into it directly, every reader resolves against it, and NO per-write
backend transfer happens. The CPU shadow is released.
BufferObject no longer owns a raw shadow Vector; it holds a PipeResource and
exposes one accessor, MappedData(), that all readers go through. Every buffer-data
consumer (UBO payload, PBO texture upload, indirect draws, resident/streamed
uploads, both backends) was migrated from GetDataReadOnly()->data() to
MappedData(), so a persistent buffer's readers see GPU memory - not a stale
shadow. That stale-shadow inconsistency is what corrupted rendering (wrong UBOs ->
misplaced/"lost" vertices) in the first zero-copy attempt (625c8a6, reverted in
896cafc); routing every consumer through one accessor makes it structurally
impossible.
Backends provide the map via BufferBackendOps::AcquirePersistentMap:
- DirectVulkan: a HOST_VISIBLE|HOST_COHERENT (required, not just requested),
persistently mapped resident VkBuffer carrying every usage, seeded from the
shadow, never recreated; AcquireResidentSlice binds it directly.
- DirectGLES: EXT_buffer_storage immutable persistent+coherent glMapBufferRange,
falling back to the shadow when the extension is absent.
Fixes the ~7GB GpuMemory OOM + 100%-CPU/ANR running modern Blaze3D Minecraft on
both Magma and Espryt (per-draw whole-buffer re-upload of the coherent persistent
ring buffer), without the coherency/stale-read hazards of the reverted attempt.
BufferTest: zero-copy stress guard (15,360 draws -> 0 per-draw transfers, and every
reader resolves to GPU memory) + a shadow-fallback test. Host suite: 203/203 pass.
Device verification pending.
glVertexAttribPointer now accepts the GL 3.3 Core packed types
GL_INT_/GL_UNSIGNED_INT_2_10_10_10_REV and the GL_BGRA size, clearing the
two long-standing "// TODO: implement GL_BGRA support" markers. Adds the
format end to end across the frontend, VAO state, and both backends.
- DataType: add Int2101010Rev / Uint2101010Rev with GLToMG / MGToGL /
MGToStr converter cases.
- Validation (ValidateVertexAttribFormat): the full glVertexAttribPointer /
glVertexAttribIPointer error table -- size is 1..4 or GL_BGRA (else
INVALID_VALUE, which takes precedence); a packed type requires size 4 or
GL_BGRA (else INVALID_OPERATION); GL_BGRA requires GL_UNSIGNED_BYTE or a
packed type AND normalized == GL_TRUE (else INVALID_OPERATION); the
integer path rejects packed types (INVALID_ENUM) and GL_BGRA size
(INVALID_VALUE).
- VAO: store GL_BGRA as size 4 plus a new IsBgra flag (reset on the
binding-format path).
- DirectVulkan: map the packed/BGRA formats to
VK_FORMAT_A2B10G10R10_* (normal) and VK_FORMAT_A2R10G10B10_* /
VK_FORMAT_B8G8R8A8_UNORM (BGRA reversed), fold IsBgra into the pipeline
hash, and size packed/BGRA elements as one 4-byte word via
GetAttributeByteSize. (Vulkan *_SNORM decodes with the GL 4.2 symmetric
rule, a documented deviation from the 3.3 signed formula.)
- DirectGLES: round-trip the packed enum through the loader, pass GL_BGRA
as the driver size argument, and size client uploads with the packed
4-byte word.
Tests: 4 VertexArrayTest cases covering packed/BGRA storage and the full
float/integer error table; the packed-size hard-fail is mutation-verified.
VertexArrayTest 42/42, SanityTest 30/30, library builds clean.
MG_Config::FeaturesTable snapshots every MOBILEGL_* toggle once in
ConfigLoader::Init with a single truthy rule (non-empty, not '0', not
'false' case-insensitively), replacing 13 scattered std::getenv sites
that used four different parsing conventions. Renderer-derived bits
(IsAngleRenderer/IsAngleLlvmpipeRenderer/AvoidSamplerMipmapMinFilter)
move into GLESCapabilities, set once in FillInGLESCapabilities, so hot
paths (glMemoryBarrier ANGLE flush, sampler min-filter sync) stop doing
per-call string scans. MOBILEGL_PRESENT_DUMP_CALL/_CURRENT_CALL stay
live getenv (the retrace harness mutates them at runtime) and
MOBILEGL_LOG_FILE_PATH stays in Log.cpp (log init precedes config
init); both are documented in Config.h. Known semantic unification:
MOBILEGL_DISABLE_SUBGROUP previously required exactly 'true' and
MOBILEGL_PRESENT_STATS exactly '1'; both now follow the shared rule
(CI's 0/1 values parse identically). Also bumps CoreVersion to 26.07.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
ES keeps gl_InstanceID zero-based and ignores the indirect command's
'reserved, must be zero' word, but ANGLE-on-Vulkan forwards the command
verbatim to vkCmdDraw*Indirect and compiles gl_InstanceID to SPIR-V
InstanceIndex, which includes firstInstance. Shaders computing
gl_BaseInstance + gl_InstanceID (Flywheel indirect) then add the base
twice, scrambling instance-to-mesh association.
Probe the actual driver semantics at capability-fill time with a tiny
indirect draw (an ES indirect draw needs a non-default VAO) and, on
leaking drivers, rewrite vertex shaders that use the native indirect
SSBO machinery so gl_InstanceID subtracts the command's baseInstance
word during native indirect draws.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- BackendProgramObjectImpl::CacheResourceLocations resolves every
glGetUniformBlockIndex / glGetUniformLocation string query once per
link and establishes the block binding points there. Per draw,
BindCurrentProgramWithResources now uses the cached indices, re-issues
glUniform1i only when a sampler's unit actually changed (program state
persists), uploads the global UBO only when its content version moved,
and skips redundant glUseProgram binds (guard reset on program-name
reuse, MakeCurrent, and every explicit glUseProgram(0)). The caches are
invalidated through ProgramObject's link version, which also makes a
relinked program finally re-sync its backend program.
- Track a texture-unit high-water mark (fed by glBindTexture /
glBindTextureUnit / glBindSampler / glBindImageTexture) so the two
per-draw unit scans (MAX_TEXTURE_IMAGE_UNITS is 192) and the
texture-deletion unbind loop only walk units that were ever touched.
- Forward the app's eglSwapInterval to the native EGL surface through a
new BackendObject::SetEGLSwapInterval hook (applied immediately when
the surface exists, otherwise deferred to surface creation /
MakeCurrent). "VSync off" finally reaches the hardware - DirectGLES
was hard-locked to the display refresh before.
The driver-side cost of the per-draw string lookups was about half of a
30% Adreno driver hotspot; libMobileGL's share of the vanilla render
thread fell from 22% to 9% (simpleperf, Adreno 830).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
139de763 started preserving layout(binding) on SSBO/image declarations in
transpiled ESSL (ES cannot rebind either through the API). That is correct
for SSBOs and for images whose GL source carries an explicit binding
(Flywheel), but wrong for image uniforms without one: glslang auto-assigns
a binding during transpile, while the app addresses the unit through
desktop-GL semantics - the link-time default (0) or glUniform1i, which ES
forbids on image uniforms. Iris/Photon picks image units with glUniform1i,
so its compute passes (auto exposure / colored light) read and wrote the
transpiler-invented units instead: the photon-v1.3b retrace came out dark
and orange-tinted (ssim 0.65 vs golden).
Rewrite every image uniform declaration's binding qualifier to the
frontend-tracked unit (layout binding reflected at link, overridden by any
later glUniform1i) when transpiling for the backend. Flywheel's explicit
bindings rewrite to the same value; Iris packs get the unit the app
actually bound with glBindImageTexture.
Verified on llvmpipe DirectGLES: photon-v1.3b retrace 0.652 -> 0.9988,
photon-v1.1 control stays at 0.9991, all 147 unit tests pass.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Create 6 / Flywheel 1.0.6 now renders correctly with both flywheel:instancing
and flywheel:indirect on DirectGLES and DirectVulkan (verified in-game on
Adreno 830: waterwheels and cogwheels solid, animated, correct pairing, no
crashes across all four combinations).
- MG_State/MG_Impl: sync explicitly-ranged SSBO bindings of FLUSH_EXPLICIT
persistent maps to the backend before compute dispatches. Flywheel writes
its scatter-copy descriptors into the staging ring's persistent map and
never flushes that span (UB per spec, works on drivers whose maps alias
GPU-visible memory); our maps alias the CPU shadow, so the descriptors
never reached the GPU: the scatter compute copied nothing (GLES: empty
draw commands) or stale garbage (Vulkan: wild indirect commands ending in
VK_ERROR_DEVICE_LOST).
- MG_Impl/MG_Backend: real glFenceSync objects backed by backend fences
(GLES: native ES syncs guarded by context generation and owner thread;
Vulkan: buffer-manager frame serials), replacing always-signaled stubs
that let Flywheel reclaim staging memory the GPU still reads.
- MG_Backend/DirectGLES: compute dispatches now run the same per-program
resource sync as draws (uniform-block bindings and sampler units must be
re-established through the API because layout(binding) is stripped from
transpiled ESSL) and rebind texture units afterwards; the cull shader
used to read a stale _FlwFrameUniforms binding and the depth-pyramid
downsample sampled a stale unit-0 texture, zeroing the Hi-Z pyramid and
occlusion-culling all Flywheel geometry. Image uniforms are excluded from
glUniform1i (ES bakes their unit via layout(binding)); image-unit sync is
clamped to the device limit; eliminated/SSBO-classified uniform blocks
are skipped.
- MG_Backend/DirectGLES: gl_BaseInstance in native indirect draws reads the
GPU-written command buffer through an injected mg_IndirectParams SSBO
view addressed per draw instead of the zero CPU shadow; layout(binding)
is preserved for SSBO/image declarations (ES has no API rebinding for
them); the ES context ownership claim moved to a global atomic owner
thread with an EGL ground-truth check, and deferred buffer op state is
mutex-guarded, so ops cannot silently no-op after context migration.
- MG_Backend/DirectVulkan: new RebaseInstanceIndexPass rewrites vertex
InstanceIndex loads to (InstanceIndex - BaseInstance). glslang's relaxed
Vulkan mode aliases gl_InstanceID to InstanceIndex, which includes
firstInstance, but GL's gl_InstanceID is zero-based - draws with nonzero
baseInstance paired meshes with wrong instance data (cogwheel drawn as a
waterwheel, another wheel collapsed invisible). Gated on the
shaderDrawParameters device feature. Sampled-read barriers additionally
cover the compute stage (the Hi-Z downsample samples the depth
attachment from compute), and short uniform-buffer ranges keep the
existing zero-padding.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- Advertise ARB_gpu_shader5 / ARB_multi_bind / ARB_shading_language_420pack /
ARB_vertex_attrib_binding / ARB_shader_image_size so LWJGL reports
SUPPORTS_INDIRECT.
- New LowerDrawParametersPass demotes DrawIndex/BaseInstance/BaseVertex
builtins to Private globals (mg_DrawID/mg_BaseInstance/mg_BaseVertex) for
the ESSL transpile; SPIRV-Cross otherwise throws for ES profiles. The
program manager promotes the emitted globals to uniforms and feeds them
per (sub-)draw.
- Indirect draws now execute natively on the GPU (glDrawElementsIndirect /
glDrawArraysIndirect per command) when an indirect buffer is bound, so
compute-written command fields (Flywheel culling updates instanceCount)
are honored; detects GL_EXT_base_instance and falls back to the CPU loop
when the command's baseInstance cannot be consumed natively.
- Sync SSBO binding points for graphics draws, not just compute (Flywheel
vertex shaders read instance data from SSBOs).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Track context generation + synced change serial per resource; re-register
ops on MakeCurrent. Fixes frozen buffer contents after the trace replayer's
probe context teardown.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Replace the application-specific PackPhotonSharedVec3Memory GLSL regex
patch with a general DecomposeWorkgroupVec3Pass SPIR-V optimization pass.
The new pass decomposes vec3/ivec3/uvec3/bvec3 Workgroup (shared) memory
variables into scalar arrays (e.g. shared vec3 arr[N][M] -> shared float
arr[N][M][3]), rewriting whole-vector loads/stores into per-component
scalar loads/stores. Component-level accesses (e.g. arr[i].x) are
unchanged since a trailing component index into a float[3] yields the
same scalar pointer as it did for a vec3.
Unlike the regex hack, the pass is application-agnostic: it does not
match on variable names, array dimensions, or shader pack identity, and
runs at the SPIR-V level before SPIRV-Cross decompilation.
Registered in SanitizeAndOptimizeBinary after AggressiveDCE so dead
workgroup accesses are already eliminated. Asserts on unsupported
OpAtomic*/OpCopyMemory targeting vec3 workgroup pointers.
Adds ProgramUtilTest.DecomposeWorkgroupVec3InSpirvPass covering array
declaration, +=, whole load/store, component access, and row-copy loop.
- add a DirectGLES ANGLE fallback control for 8-bit SNORM texture formats
- normalize SNORM8 textures to float storage so ANGLE can render Complementary intermediate framebuffers
- reuse the normalized upload conversion path for SNORM8 and existing norm16 float fallbacks
Implemented:
- Advertise Voxy-required DirectGLES extensions without raising the reported OpenGL version.
- Add DirectGLES multi draw indirect count emulation and preserve GL draw indirect baseInstance semantics on GLES.
- Add DirectGLES DSA framebuffer clear/blit paths used by Minecraft and Voxy presentation.
Fixed:
- Rewrite gl_BaseInstance in DirectGLES vertex shaders and provide a backend uniform for indirect draw emulation.
- Materialize framebuffer attachment textures during DirectGLES FBO sync so named framebuffer operations do not desync backend attachment state.
- Avoid redundant texture buffer rebinding and handle texture buffers without bound storage during backend sync.
Tests:
- Add MG_Test coverage for DirectGLES Voxy extension advertising, baseInstance shader rewriting, and DSA named framebuffer clear/blit backend wiring.
- deal with legacy GLSL syntax (attribute/varying/gl_FragColor/texture2D/etc.)
- implement glGet GL_SHADER_SOURCE_LENGTH, and make sure returns
original shader source
- expose proper extensions (GL_ARB_depth_texture)
- support env var MOBILEGL_LOG_FILE_PATH
- unit tests to test against those changes
swung0x48
