Check CUDA error and raise exception if needed.

Check if annotation is const[...]

Check if annotation is template[...]

Check if annotation is Tuple[...] or tuple[...]

check all scalar arguments or constexpr arguments. no check pointer dtypes argument signature should be: [*]dtype[:annotation] *

Split list into chunks for parallel processing

Collect local variables defined in the target function

Collect all variable names in the calling context

The pure computation logic for one micro-batch.

Create AST node for ll.Tensor[elem_dtype]

Recursively dump a node with custom attributes.

Evaluate an AST expression using the function's global context

Extract assembly template, constraints, and operands from an asm() call. Args: asm_call: AST node representing asm() call

Extract member variables and template parameters from __init__ method.

Extract struct arguments from constructor call. This is a generic implementation that extracts all arguments (positional and

Find the first asm() call in an AST node. Supports: - asm(...) - ll.asm(...) - ll.asm_volatile(...)

Find the return variable name from function body.

Format an attribute value for display.

Generate a codegen function for the simple_builtin. The codegen function will be called during code generation with C++ code strings

Generate the complete C++ struct definition.

Generate eval_return_type function from function signature. Args: func: Function with type annotations func_ast: Functio

Generate code for instance method call.

Generate code for static/class method call.

Generate a unique, valid C++ struct name from StructType fields.

Generate a unique, valid C++ struct name from StructType fields.

Return (major, minor) CUDA compute capability, or None if no CUDA GPU.

Get default value string for C++ initialization.

Convert tensor dtype to CUDA tensor map dtype.

Query current SM clock frequency via nvidia-smi.

Try to detect if a Call node is calling an intrin/builtin function and return its info. Args: node: The Call AST node ct

Try to detect if a Call node is calling an intrin/builtin function.

Calculate optimal number of 1D blocks per group.

Uses `rocm-smi --showtoponuma --json` and returns {"card0": 0, ...}

Returns total NVLink bandwidth in GB/s for specified GPU

Returns the maximum PCIe link speed in GB/s for specified GPU

Convert swizzle mode to CUDA enum.

Handle `ll.align_memory` calls, extract align_bytes and scope, then update memory analysis.

Handle builtin function calls.

Handle function calls where func is a Constant.

Handle `ll.empty` calls.

Handle single name assignment (e.g., x = 5).

Handle special struct method calls.

Handle struct stub constructor assignment.

Handle struct stub constructor calls.

Handle struct stub method calls.

Handle tuple unpacking for struct stub method calls.

Handle subscript assignment (e.g., latency[0] = value).

Handle tuple assignment (e.g., (x, y) = (5, 3.14)).

Handle tuple unpacking from function call: a, b = xxx.func() -> a = xxx.func(..., b). First target receives return value, others are

Handle `ll.zeros` like ll.empty (same shape/dtype/scope, alloc+zero).