EVEngine

640 Commits 1 Branch 0 Tags

3 Commits

Include renames

Author	SHA1	Message	Date
西风逍遥游	34d1fdb63b	refactor(modules): clear remaining back-edges and split the largest TUs (#127 ) * chore(ci): add clang-format gate, CODEOWNERS and agent collaboration rules * refactor(modules): clear remaining back-edges and split the largest TUs Three known dependency back-edges are gone (module_depgraph.py KNOWN_BACK_EDGES is now empty; --check passes with 0 violations, 0 cycles): - window -> graphics: window queries the new IWindowSurfaceHost capability (common/WindowSurfaceHost.h) instead of holding a Graphics pointer; Window::setGraphics was removed and all call sites updated. - window -> image: Window::setIcon was replaced by setIconRGBA(const uint8_t, w, h); the window no longer decodes ImageData. - scene -> graphics: pickScreenAt / collectFrustumIdsAt implementations moved to graphics/ScenePicking.cpp (graphics -> scene is a legal downward edge); scene now depends only on spatial. TU splits (pure moves, verified identical method sets and brace balance): - vulkan/Graphics.cpp (5118 lines) -> Graphics.cpp (lifecycle), GraphicsPipeline, Graphics2D, Graphics3D, GraphicsMesh + shared GraphicsInternal.h. - scene/Scene.cpp -> Scene.cpp (script layer / host lifecycle / serialization) + SceneNodes.cpp + shared SceneInternal.h. Verified with a real build (MSVC 19.51, prebuilt third-party): full build green; window. and Scene.* suites pass, including scene picking/culling and GPU cases. * refactor(graphics): narrow Graphics interfaces and clear public-header leaks Step 4 — interface segregation: - New lean interfaces: IGraphics2D, IGraphics3D, IResourceFactory, IPostFX (pure virtuals, forward-declared types only; graphics/Color.h extracted from Canvas.h so modules that only need the color type stop pulling Canvas.h). - Graphics now implements all four facets; consumers can depend on the narrow interface they use instead of the 129-method god class. - Migrated voxel (VoxelWorld::drawVisible -> IGraphics3D) and model3d (ModelRenderer::buildRenderable(s) -> IResourceFactory&). - Graphics.h slims from 24 sub-header includes to a handful of small type headers + forward declarations; the destructor moved out-of-line so unique_ptr effect members are destroyed where types are complete. Step 5 — header leaks (module_depgraph.py '' marks: 14 -> 7): - map/TileLayer.h, particles/ParticleEmitter.h, avatar/AvatarInstance.h, animation/SpineAnim.h, map/TileSystem.h, model3d/ModelData.h, model3d/ModelRenderer.h, audio/Audio.h + Source.h, image/ImageData.h + Image.h, dialogue/Dialogue.h, procgen Palette/OutputSpec/TextureRecipe/ PbrMaterial -> forward declarations, includes moved to .cpp. - voxel/VoxelWorld.h: terrain sampler Pimpl'd (unique_ptr + out-of-line destructor), removing the procgen header dependency. - gpgpu/vulkan backend headers stop pulling graphics/vulkan/Graphics.h. - Remaining '' marks are deliberate: tiny POD/enum headers (GridConfig, MapObject, BlendMode, Color), an internal backend header (GraphicsInternal.h -> filesystem), and lean interface references. Verified with a full clean build (MSVC) and tests: window/Scene/voxel/ RenderSystem/Quad/shader/TextureSampler/resourceLifetime pass; broad suite only shows the pre-existing PHYSFS-init filesystem-environment failures. feat(resources): revive ResourceManager as unified IAssetReloader-backed cache * fix(resources): leak ResourceManager singleton to avoid exit-time teardown crashes Cached CPU resources may own third-party handles (FreeType faces, Assimp scenes, image decode handlers) whose libraries are torn down at process exit in an unspecified TU order; destroying cached entries from the singleton destructor can segfault at exit (seen on CI Windows in model3d.materialApi.objMtl). Keep the singleton alive until OS reclaim; explicit unload()/clear() still release entries during the run. * refactor(devtools): consume business modules through capability interfaces DevTools / cmdline no longer include business-module headers, so trimming scene / physics / procgen / particles / audio / ui / graphics no longer requires disabling devtools, and devtools can be switched off with -DEVENGINE_MODULE_DEVTOOLS=OFF. New capability interfaces in common/: - ISceneQuery (scene) - host/node inspection + mutation + transform sync - IRenderCapture (graphics) - render status, PNG capture, camera control, visible-entity snapshot, entity-ID mask, G-buffer PNGs - IPhysicsQuery (physics) - 2D world lifecycle + ray casts - IProcgenQuery (procgen) - recipe listing, map/mesh generation - IParticlesQuery (particles) - emitter status + one-shot emit - IAudioQuery (audio) - volume / stop-all - IEditorHost (ui) - declarative editor host (used by cmdline too) Providers register in their module constructors. DevTools rewires McpServer (render/scene/procgen/physics/particles/audio/host tools), SceneInspect (now delegates to ISceneQuery + IRenderCapture) and RenderVision (uses IRenderCapture::capturePngDataUrl); cmdline/Mcp.cpp uses IEditorHost. manifest: devtools DEPS graphics scene physics procgen particles audio ui image event filesystem -> event; cmdline DEPS filesystem ui -> filesystem. Verified: full build green; devtools.mcp.* tests pass; broad suite shows only the pre-existing PHYSFS-init filesystem-environment failures. * style: apply clang-format to changed lines after dev merge * fix(webgpu): include RenderControl/TextureSampler explicitly after Graphics.h slimming * fix(webgpu): provide newFont/print stubs to satisfy IGraphics2D/IResourceFactory vtables * fix(common): make ProcgenQuery.h self-sufficient (cstdint/utility for GCC)	2026-08-20 05:53:33 +08:00
西风逍遥游	cff9bb4d0c	tensor: AITemplate-style compile pipeline, gpgpu Sequence, mini-LLM demo and weight-only quantization Refactor the tensor module around AITemplate-style static compilation: - Optimizer: DCE, constant folding, elementwise chain fusion, matmul/conv bias+activation epilogue fusion, liveness-based static memory planning, topological group execution order. - KernelGen: shape-baked GLSL per fused group; two-pass softmax/layernorm; shared-memory SDPA; naive + 16x16 tiled matmul templates. - GpuBackend: arena buffer allocation per memory plan, matmul autotuning, single command-buffer submission via the new gpgpu::Sequence (uploads, dispatches and downloads recorded once per run, Kompute-style). - shaderc in-process GLSL->SPIR-V on Windows (no glslc.exe dependency), with glslc spawn fallback. - CPU reference kernels shared by eager and the compiled interpreter; no-Vulkan fallback keeps compiled graphs runnable headless. Neural-network coverage: softmax/logsoftmax, layernorm/rmsnorm (with weight/bias), gelu/silu, conv1d/conv2d, pool, embedding, concat/slice, axis reduces, argmax, cast, fused SDPA (+masked), resize2d, batched matmul, seeded random. Weight-only quantization (activations stay fp32): - fp16 (IEEE half), fp8 (e4m3), fp4 (e2m1), int8, int4 with per-group block scales (MXFP-style: scale = maxAbs / format max magnitude). - Quantized weights capture into graph Const nodes; CPU interpreter dequantizes at materialization; GPU matmul/embedding kernels dequantize inline from packed uint buffers (naive matmul variant). - TF::quantizeWeight API + script binding. Mini-LLM validation (TinyStories-gpt2-3M, 8 layers d=64): - test/tensor_llm.cpp builds the full GPT-2 graph from engine ops and matches numpy reference: probs within 3e-7, greedy generation 24/24 identical (~100 ms/token CPU), compiled path exact. - tensor.llm.quantizedDtypes: fp16/int8/fp8/int4 all keep 16/16 greedy top-1 (fp4 16/16 with proper block scaling), weights 27.6 MB -> 3.3-13 MB. - scripts/export_tiny_gpt2.py converts a PyTorch .bin checkpoint (no torch dependency) plus GPT-2 tokenizer into engine assets + numpy reference. Tests: gpgpu 8/8, tensor 31/31 (original suite) + tensor.llm.* + tensor.quant.* Assets in test/assets/mini_llm are gitignored (regenerate via the script).	2026-08-18 17:42:38 +08:00
sunxfancyandCursor	d1034c4c2b	feat: enhance math and tensor modules with new functionalities and documentation updates - Updated the math module to include vector and matrix operations, noise generation, Bezier curves, and random number utilities. - Introduced the tensor module, providing TensorFlow 2-style eager execution and graph building capabilities. - Enhanced documentation to reflect the new features in the math and tensor modules, including usage examples and API references. - Modified CMakeLists.txt to include the new math and tensor modules, ensuring proper build configurations. Co-authored-by: Cursor <cursoragent@cursor.com>	2026-08-07 15:50:02 +08:00

Author

SHA1

Message

Date

西风逍遥游

34d1fdb63b

refactor(modules): clear remaining back-edges and split the largest TUs (#127 )

* chore(ci): add clang-format gate, CODEOWNERS and agent collaboration rules

* refactor(modules): clear remaining back-edges and split the largest TUs

Three known dependency back-edges are gone (module_depgraph.py KNOWN_BACK_EDGES
is now empty; --check passes with 0 violations, 0 cycles):

- window -> graphics: window queries the new IWindowSurfaceHost capability
  (common/WindowSurfaceHost.h) instead of holding a Graphics pointer;
  Window::setGraphics was removed and all call sites updated.
- window -> image: Window::setIcon was replaced by setIconRGBA(const uint8_t*,
  w, h); the window no longer decodes ImageData.
- scene -> graphics: pickScreenAt / collectFrustumIdsAt implementations moved
  to graphics/ScenePicking.cpp (graphics -> scene is a legal downward edge);
  scene now depends only on spatial.

TU splits (pure moves, verified identical method sets and brace balance):

- vulkan/Graphics.cpp (5118 lines) -> Graphics.cpp (lifecycle), GraphicsPipeline,
  Graphics2D, Graphics3D, GraphicsMesh + shared GraphicsInternal.h.
- scene/Scene.cpp -> Scene.cpp (script layer / host lifecycle / serialization)
  + SceneNodes.cpp + shared SceneInternal.h.

Verified with a real build (MSVC 19.51, prebuilt third-party): full build green;
window.* and Scene.* suites pass, including scene picking/culling and GPU cases.

* refactor(graphics): narrow Graphics interfaces and clear public-header leaks

Step 4 — interface segregation:
- New lean interfaces: IGraphics2D, IGraphics3D, IResourceFactory, IPostFX
  (pure virtuals, forward-declared types only; graphics/Color.h extracted from
  Canvas.h so modules that only need the color type stop pulling Canvas.h).
- Graphics now implements all four facets; consumers can depend on the narrow
  interface they use instead of the 129-method god class.
- Migrated voxel (VoxelWorld::drawVisible -> IGraphics3D*) and model3d
  (ModelRenderer::buildRenderable(s) -> IResourceFactory&).
- Graphics.h slims from 24 sub-header includes to a handful of small type
  headers + forward declarations; the destructor moved out-of-line so
  unique_ptr effect members are destroyed where types are complete.

Step 5 — header leaks (module_depgraph.py '*' marks: 14 -> 7):
- map/TileLayer.h, particles/ParticleEmitter.h, avatar/AvatarInstance.h,
  animation/SpineAnim.h, map/TileSystem.h, model3d/ModelData.h,
  model3d/ModelRenderer.h, audio/Audio.h + Source.h, image/ImageData.h +
  Image.h, dialogue/Dialogue.h, procgen Palette/OutputSpec/TextureRecipe/
  PbrMaterial -> forward declarations, includes moved to .cpp.
- voxel/VoxelWorld.h: terrain sampler Pimpl'd (unique_ptr + out-of-line
  destructor), removing the procgen header dependency.
- gpgpu/vulkan backend headers stop pulling graphics/vulkan/Graphics.h.
- Remaining '*' marks are deliberate: tiny POD/enum headers (GridConfig,
  MapObject, BlendMode, Color), an internal backend header
  (GraphicsInternal.h -> filesystem), and lean interface references.

Verified with a full clean build (MSVC) and tests: window/Scene/voxel/
RenderSystem/Quad/shader/TextureSampler/resourceLifetime pass; broad suite
only shows the pre-existing PHYSFS-init filesystem-environment failures.

* feat(resources): revive ResourceManager as unified IAssetReloader-backed cache

* fix(resources): leak ResourceManager singleton to avoid exit-time teardown crashes

Cached CPU resources may own third-party handles (FreeType faces, Assimp
scenes, image decode handlers) whose libraries are torn down at process
exit in an unspecified TU order; destroying cached entries from the
singleton destructor can segfault at exit (seen on CI Windows in
model3d.materialApi.objMtl). Keep the singleton alive until OS reclaim;
explicit unload()/clear() still release entries during the run.

* refactor(devtools): consume business modules through capability interfaces

DevTools / cmdline no longer include business-module headers, so trimming
scene / physics / procgen / particles / audio / ui / graphics no longer
requires disabling devtools, and devtools can be switched off with
-DEVENGINE_MODULE_DEVTOOLS=OFF.

New capability interfaces in common/:
- ISceneQuery (scene) - host/node inspection + mutation + transform sync
- IRenderCapture (graphics) - render status, PNG capture, camera control,
  visible-entity snapshot, entity-ID mask, G-buffer PNGs
- IPhysicsQuery (physics) - 2D world lifecycle + ray casts
- IProcgenQuery (procgen) - recipe listing, map/mesh generation
- IParticlesQuery (particles) - emitter status + one-shot emit
- IAudioQuery (audio) - volume / stop-all
- IEditorHost (ui) - declarative editor host (used by cmdline too)

Providers register in their module constructors. DevTools rewires McpServer
(render/scene/procgen/physics/particles/audio/host tools), SceneInspect
(now delegates to ISceneQuery + IRenderCapture) and RenderVision (uses
IRenderCapture::capturePngDataUrl); cmdline/Mcp.cpp uses IEditorHost.

manifest: devtools DEPS graphics scene physics procgen particles audio ui
image event filesystem -> event; cmdline DEPS filesystem ui -> filesystem.

Verified: full build green; devtools.mcp.* tests pass; broad suite shows
only the pre-existing PHYSFS-init filesystem-environment failures.

* style: apply clang-format to changed lines after dev merge

* fix(webgpu): include RenderControl/TextureSampler explicitly after Graphics.h slimming

* fix(webgpu): provide newFont/print stubs to satisfy IGraphics2D/IResourceFactory vtables

* fix(common): make ProcgenQuery.h self-sufficient (cstdint/utility for GCC)

2026-08-20 05:53:33 +08:00

西风逍遥游

cff9bb4d0c

tensor: AITemplate-style compile pipeline, gpgpu Sequence, mini-LLM demo and weight-only quantization

Refactor the tensor module around AITemplate-style static compilation:

- Optimizer: DCE, constant folding, elementwise chain fusion, matmul/conv
  bias+activation epilogue fusion, liveness-based static memory planning,
  topological group execution order.
- KernelGen: shape-baked GLSL per fused group; two-pass softmax/layernorm;
  shared-memory SDPA; naive + 16x16 tiled matmul templates.
- GpuBackend: arena buffer allocation per memory plan, matmul autotuning,
  single command-buffer submission via the new gpgpu::Sequence (uploads,
  dispatches and downloads recorded once per run, Kompute-style).
- shaderc in-process GLSL->SPIR-V on Windows (no glslc.exe dependency),
  with glslc spawn fallback.
- CPU reference kernels shared by eager and the compiled interpreter;
  no-Vulkan fallback keeps compiled graphs runnable headless.

Neural-network coverage: softmax/logsoftmax, layernorm/rmsnorm (with
weight/bias), gelu/silu, conv1d/conv2d, pool, embedding, concat/slice,
axis reduces, argmax, cast, fused SDPA (+masked), resize2d, batched
matmul, seeded random.

Weight-only quantization (activations stay fp32):
- fp16 (IEEE half), fp8 (e4m3), fp4 (e2m1), int8, int4 with per-group
  block scales (MXFP-style: scale = maxAbs / format max magnitude).
- Quantized weights capture into graph Const nodes; CPU interpreter
  dequantizes at materialization; GPU matmul/embedding kernels dequantize
  inline from packed uint buffers (naive matmul variant).
- TF::quantizeWeight API + script binding.

Mini-LLM validation (TinyStories-gpt2-3M, 8 layers d=64):
- test/tensor_llm.cpp builds the full GPT-2 graph from engine ops and
  matches numpy reference: probs within 3e-7, greedy generation 24/24
  identical (~100 ms/token CPU), compiled path exact.
- tensor.llm.quantizedDtypes: fp16/int8/fp8/int4 all keep 16/16 greedy
  top-1 (fp4 16/16 with proper block scaling), weights 27.6 MB -> 3.3-13 MB.
- scripts/export_tiny_gpt2.py converts a PyTorch .bin checkpoint (no torch
  dependency) plus GPT-2 tokenizer into engine assets + numpy reference.

Tests: gpgpu 8/8, tensor 31/31 (original suite) + tensor.llm.* + tensor.quant.*
Assets in test/assets/mini_llm are gitignored (regenerate via the script).

2026-08-18 17:42:38 +08:00

sunxfancyandCursor

d1034c4c2b

feat: enhance math and tensor modules with new functionalities and documentation updates

- Updated the math module to include vector and matrix operations, noise generation, Bezier curves, and random number utilities.
- Introduced the tensor module, providing TensorFlow 2-style eager execution and graph building capabilities.
- Enhanced documentation to reflect the new features in the math and tensor modules, including usage examples and API references.
- Modified CMakeLists.txt to include the new math and tensor modules, ensuring proper build configurations.

Co-authored-by: Cursor <cursoragent@cursor.com>

2026-08-07 15:50:02 +08:00