Execution Model

Execution Model

Purpose: Explains the immutable and static nature of the execution pipeline.

VECTORIA employs a strict Compile-Schedule-Execute lifecycle. Once a graph is compiled, its structure and memory footprint are frozen.

1. Immutable Intermediate Representation (IR)

The graph is defined using immutable structures:

• Graph: A collection of Nodes and NodeId outputs.
• Node: A variant of InputNode, ParameterNode, or OpNode.
• OpType: A strictly typed enum (e.g., MatMul, Add, Relu, Transpose, Concat, Attention).

Once constructed, the Graph object is treated as a read-only specification by the Engine.

2. Static Scheduling

During Engine::compile(), the system:

1. Validates the topology (checks for cycles and invalid inputs).
2. Linearizes the graph into a fixed execution schedule_ (std::vector<size_t>).
3. Calculates the exact memory requirement for every tensor.

3. Arena Memory Model

Memory is managed by a linear Arena allocator.

• Allocation: All necessary memory is allocated in a single block during compile().
• Addressing: Each node is assigned a fixed pointer (node_buffers_[i]) into this arena.
• Lifecycle: Memory persists for the lifetime of the Engine and is reset only upon recompilation. This eliminates runtime allocation overhead and fragmentation.

Execution Flow

The execute() method simply iterates through the pre-computed schedule_:

[Start] -> [Fetch Node] -> [Resolve Inputs from Arena] -> [Dispatch Kernel] -> [Write Output to Arena] -> [Next Node]

References

• core/include/vectoria/ir.hpp
• core/src/engine.cpp
• core/src/memory.cpp