PyAsc Infrastructure Changes

This patch introduces modifications to the existing PyAsc infrastructure required by the tile-based programming add-on. Currently, there are no changes in public API or user experience. All existing features remain fully functional, as well as unit and system tests.

1. MLIR Changes

1.1 Asc Dialect — IR

• New operations: Broadcast, Reduction (reduce_sum/max/min/prod with ReducePattern), register-based API (~40 ops for C310: binary, unary, vec-scalar, reduce, utility), LayerNorm, Log2, get_buf/rls_buf (BufId sync for C310), data_copy_nd_dma + NdDmaParams, set_loop_mode_para/reset_loop_mode_para, trans_data_to_5hd_tensor, local_tensor_v3, local_tensor.get_phy_addr_v2.
• New types: FixpipeParamsC310, MaskReg, RegTensor, NdDmaParams.
• New attributes: MaskPattern, MemType, ReducePattern, DataCopyMVType.
• New API types: LayerNormTiling, LoopModeParams.
• New interfaces: OpWithSrcInterface (get source tensors), OpWithReusableSrcInterface (marker for safe same-op reuse), CopyToL0OpInterface, MicroAPIOpInterface (+ Binary/Unary/VecScalar variants), ReduceOpInterface, BinaryL0OpInterface, BinaryL2OpInterface.
• Extended interfaces: OpWithDstInterface gained getDstTensors() default; UnaryOpInterface/BinaryOpInterface/VecScalarOpInterface now include OpWithSrcInterface and OpWithReusableSrcInterface; L0/L2 interfaces gained mutable operand accessors and setIsSetMask.
• CopyDirection enum: Replaced with corresponding methods GlobalToLocal/LocalToGlobal/LocalToLocal/GlobalToGlobal.
• Modified operations: SoftMax/SimpleSoftMax changed to VectorOp with src interface, sumTensor/maxTensor made Optional; RmsNorm renamed mnemonic to rms_norm, changed to VectorOp; Fixpipe changed to DataCopyOp with generalized dst type and C310 params support; DataCopyL0/L2 gained src interface; LocalTensorAutoOp gained position attribute; SubIndex ops gained folders (zero-index identity); ReinterpretCastOp fold enhanced to eliminate double-cast chains; MulCast ops moved to BinaryCast templates; VecScalar ops gained Subs and Divs variants; all unary/binary/vec-scalar/math base classes gained default getSrcTensors()/getDstTensors().

1.2 Asc Dialect — Passes

• Renamed: HoistUBAllocation → HoistTensorAllocation (added exclude-in-out option), InsertSync → InsertQueSync.
• Modified: MaterializeTensor (added always-buf option), LegalizeKernelArgs (added set-ffts-addr option, conditional on architecture), DetectKernelType (now classifies as "vector"/"cube"/"mixed" instead of binary flag), InsertSync (extended with V_S/S_V event sync for get_value/set_value), InputOutputTensor (handles input-only tensors used for ubuf_gm copy).
• All existing passes received detailed description fields in TableGen.

1.3 EmitAsc Dialect

• New ops: init_struct (struct initialization with named fields), mask (create bit-by-bit mask pair).
• New type: mask (2-element i64 array for high/low mask).
• variable: Removed Pure trait; added canonicalizer support.
• New utility: InitStructBuilder helper class.

1.4 Code Emitter (Target/AscendC)

• New emitters for all new Asc operations (Broadcast, Reduction, LayerNorm, MicroAPI suite, BufId sync, NdDma, etc.) and arith::NegFOp.
• CodeEmitter: Added methods for C310 types (CO2Layout, FixpipeParamsC310, RegTensor, MaskReg, NdDmaParams).
• Fixpipe emission: Generalized dst type handling; config emitted by name instead of hardcoded constant.
• VecScalar ops: Added Subs and Divs variants to all L0/L1/L2 emitter lists.

1.5 Build System & Tools

• Removed C language requirement from project; added atomic profile update for GCC coverage.
• Added LIT tests for passes and emitters that didn't have them initially.
• ConstantOpBuilder: Widened parameter types to int64_t; added i1() helper.

2. Python Changes

2.1 Codegen (asc/codegen/)

• FunctionVisitor:
  • Added CustomBuiltins support for injecting custom builtins into the name scope.
  • Added MatMult (@ operator) support.
  • Boolean operators (and/or/not) and IfExp: compile-time constant folding before falling back to IR.
  • visit_Call: Handles ast.Starred arguments (*args unpacking).
  • visit_For: Generalized to accept any BaseRange subclass; calls handle_op() on the range object.
  • compute_inout: Materializes initial handles outside the block; accepts init_handles for if/else merge.
  • visit_If: Else-branch inherits init handles from then-branch.
• NameScope: builtins → default_builtins; constructor accepts custom_builtins; inherit() propagates them.
• Function: Added compute_globals() for closure variable resolution; improved AST line number tracking for decorated functions; ConstExpr detection checks runtime isinstance in addition to annotations.
• Specialization: PointerArgType.to_ir() switched to get_unranked_memref_type.

2.2 Runtime & Compiler (asc/runtime/)

• Platform support: Added C310 architecture — CompilationArch enum (C220/C310), platform_to_arch() mapping, 12 new platforms (Ascend950PR_95xx); set_platform() accepts string values; CompilePlatform enum removed in favor of CompilationArch.
• Compiler: Compiler sets compilation_arch, soc_version on the module before passes; added -fcce-vf-vl Bisheng flag for C310; bisheng_options type fixed to Tuple[str, ...]; opt_level default changed to int = 3; compilation pipeline extended with new pass stages.
• JIT: Refactored compile/cache/launch into separated phases using CompilePrereqs dataclass; kernel handles cached and reused across calls (LaunchedKernel); __getitem__ returns self; compute_globals() used for closures; jit() injects CustomBuiltins(range=asc_range) by default; 0-dim torch tensors treated as scalars; LaunchOptions.core_num changed to Optional[int] = None (auto-detect). Cache hash algorithm abstracted via HashDriver protocol (default: Blake2b instead of SHA-256).
• Launcher: expand_kernel_args handles torch 0-dim tensors as scalars; core_num auto-detected if None; DRY_RUN env var renamed to PYASC_DRY_RUN; kernel binary handles optionally cached and unregistered after use.

2.3 Language Core (asc/language/core/)

• bfloat16 dtype: Added throughout — DataType regex/kind mapping, KnownTypes, exports, builder bindings.
• IRValue/PlainValue/materialize_ir_value: __bool__ raises TypeError on IR values; __not__ replaced by logical_not(); added max()/min() methods; type inference refactored (bool→int1, int→int_, float→float_).
• BaseRange/range: BaseRange extracted as base class with handle_op() hook; range inherits from it.

2.4 Pybind11 Bindings (python/src/)

• IR.cpp: Added ir.attr submodule, bindings for more builtin types (bfloat16, math dialect), clone_shaped_type, TypedAttr, get_splat_attr, get_dict_of_int_attr, ub_block_size; registered math dialect in context.
• OpBuilder.cpp: New PyOpBuilder(Operation*) constructor; new bindings for bfloat16 type/constants/attrs, bool/i8/i16/f16/f32/f64 attrs, i32/bool array attrs; new arith ops (MaxSI, MinSI, MaximumF, MinimumF, ShLI, ShRSI, generic ConstantOp); full math dialect op bindings.
• Passes.cpp: New pass bindings for all new Asc passes (with parameterized options for hoist_tensor_allocation, legalize_kernel_args, materialize_tensor, reuse_ub_allocation); add_hoist_ub_allocation removed (replaced by add_hoist_tensor_allocation).

2.5 Build System (setup.py)

• LLVM auto-download removed; LLVM must be provided via LLVM_INSTALL_PREFIX.
• Compiler/linker configuration generalized: PYASC_SETUP_COMPILER, PYASC_SETUP_LINKER, PYASC_SETUP_CMAKE_APPEND, PYASC_SETUP_JOBS env vars; PYASC_SETUP_CLANG_LLD now sets defaults instead of hardcoding.