[Refactor] Extract per-task adstack codegen state and helpers from TaskCodeGenLLVM into TaskCodeGenLLVMAdStack by duburcqa · Pull Request #612 · Genesis-Embodied-AI/quadrants

[AutoDiff] Stage 1.1: recognize MaxOverRange specs reducible by a parallel max-reducer dispatch (option D)

duburcqa

committed

[AutoDiff] Stage 1.2: SPIR-V max-reducer shader (option D); body bytecode interpreter for kConst/kBoundVariable/kExternalTensorRead/kAdd/kSub/kMul/kMax

duburcqa

committed

[AutoDiff] Stage 1.3: LLVM runtime function for the option-D max reducer (single-thread serial walk over body bytecode)

duburcqa

committed

[AutoDiff] Stage 1.4a: AdStackCache max-reducer cache methods + body bytecode encoder (encode_max_reducer_body_bytecode)

duburcqa

committed

[AutoDiff] Stage 1.6: substitute_precomputed_max_over_range helper (replaces captured MaxOverRange nodes with Const after reducer dispatch)

duburcqa

committed

[AutoDiff] Stage 1.4b: GfxRuntime::dispatch_max_reducers + adstack_max_reducer_launch.cpp (cmdlist + buffer binding for the option-D max reducer)

duburcqa

committed

[AutoDiff] Stage 1.4+1.6: launch_kernel wires dispatch_max_reducers and substitution into the SPIR-V sizer eval paths

duburcqa

committed

[AutoDiff] Hard-require PSB+Int64 at the adstack reverse-mode entry; drop redundant per-helper cap gates

duburcqa

committed

[AutoDiff] Stage 1.5 + comment cleanup: LLVM dispatch_max_reducers_for_tasks; strip plan-specific noise from comments

duburcqa

committed

[AutoDiff] Adstack max-reducer: dispatch fixes, Metal u32 atomic, cap-hit tripwires, regression tests

duburcqa

committed

[AutoDiff] Adstack: short-circuit MaxOverRange walk on cap-hit (avoid GPU TDR on out-of-grammar shapes)

duburcqa

committed

[AutoDiff] Adstack: drop LLVM device sizer overflow-flag write to avoid clobbering task_id under contention

duburcqa

committed

[AutoDiff] Adstack: scope cap-hit tripwire test to backends with explicit device-sizer signals

duburcqa

committed

[AutoDiff] Adstack: drop arch restriction on cap-hit tripwire test

duburcqa

committed

[Docs] Document the per-task sizer iteration cap and its parallel-eval bypass

duburcqa

committed

[AutoDiff] Adstack max-reducer: capture nested MaxOverRange chains across distinct bound variables (multi-axis)

duburcqa

committed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[Refactor] Extract per-task adstack codegen state and helpers from TaskCodeGenLLVM into TaskCodeGenLLVMAdStack#612

[Refactor] Extract per-task adstack codegen state and helpers from TaskCodeGenLLVM into TaskCodeGenLLVMAdStack#612
duburcqa wants to merge 20 commits into
mainfrom
duburcqa/codegen_llvm_adstack_factor_out

Commits on May 6, 2026

[AutoDiff] Stage 1.1: recognize MaxOverRange specs reducible by a parallel max-reducer dispatch (option D)

[AutoDiff] Stage 1.2: SPIR-V max-reducer shader (option D); body bytecode interpreter for kConst/kBoundVariable/kExternalTensorRead/kAdd/kSub/kMul/kMax

[AutoDiff] Stage 1.3: LLVM runtime function for the option-D max reducer (single-thread serial walk over body bytecode)

[AutoDiff] Stage 1.4a: AdStackCache max-reducer cache methods + body bytecode encoder (encode_max_reducer_body_bytecode)

[AutoDiff] Stage 1.6: substitute_precomputed_max_over_range helper (replaces captured MaxOverRange nodes with Const after reducer dispatch)

[AutoDiff] Stage 1.4b: GfxRuntime::dispatch_max_reducers + adstack_max_reducer_launch.cpp (cmdlist + buffer binding for the option-D max reducer)

[AutoDiff] Stage 1.4+1.6: launch_kernel wires dispatch_max_reducers and substitution into the SPIR-V sizer eval paths

[AutoDiff] Hard-require PSB+Int64 at the adstack reverse-mode entry; drop redundant per-helper cap gates

[AutoDiff] Stage 1.5 + comment cleanup: LLVM dispatch_max_reducers_for_tasks; strip plan-specific noise from comments

[AutoDiff] Adstack max-reducer: dispatch fixes, Metal u32 atomic, cap-hit tripwires, regression tests

[AutoDiff] Adstack: short-circuit MaxOverRange walk on cap-hit (avoid GPU TDR on out-of-grammar shapes)

[AutoDiff] Adstack: drop LLVM device sizer overflow-flag write to avoid clobbering task_id under contention

[AutoDiff] Adstack: scope cap-hit tripwire test to backends with explicit device-sizer signals

[AutoDiff] Adstack: drop arch restriction on cap-hit tripwire test

[Docs] Document the per-task sizer iteration cap and its parallel-eval bypass

[AutoDiff] Adstack max-reducer: capture nested MaxOverRange chains across distinct bound variables (multi-axis)

Commits on May 7, 2026

[AutoDiff] Adstack max-reducer: round-based dispatch substitutes captured deps before host-eval

[AutoDiff] Adstack max-reducer: support bound-var-indexed FieldLoad in body grammar

[AutoDiff] LLVM adstack lazy-claim: split into stage-grouped subdir (bound_eval / metadata_publish / heap_grow)

[CI debug] fprintf trace dispatch_max_reducers + cuda allocate_memory(_runtime)