Struct SpecializedGpuKernels 

pub struct SpecializedGpuKernels { /* private fields */ }

Enhanced GPU kernel manager for specialized gates

Implementations

impl SpecializedGpuKernels

pub fn new(config: OptimizationConfig) -> QuantRS2Result<Self>

Create a new specialized GPU kernel manager

pub fn apply_holonomic_gate( &self, state: &mut [Complex64], holonomy_matrix: &[Complex64], target_qubits: &[QubitId], ) -> QuantRS2Result<()>

Apply a holonomic (unitary) gate to the targeted qubits.

Dispatches to a GPU kernel when a specialized GPU context is available; otherwise computes the result on the CPU. The CPU path is a real matrix-vector application over the targeted-qubit subspace (it is not a no-op). Because the specialized GPU gate kernels are still DEFERRED, the GPU contexts are currently None, so this resolves to the CPU path — which performs the genuine computation.

pub fn apply_post_quantum_hash_gate( &self, _state: &mut [Complex64], _hash_circuit: &[Complex64], compression_type: PostQuantumCompressionType, ) -> QuantRS2Result<()>

Apply a post-quantum cryptographic hash gate.

These compression schemes (quantum sponge / Merkle tree / Grover) are implemented as GPU device kernels that are not yet authored, and they have no defined CPU reference here. Rather than silently returning success without touching state (the previous fabricated behavior), this returns an honest error. DEFERRED until the real kernels exist.

pub fn apply_quantum_ml_attention( &self, _state: &mut [Complex64], _query_params: &[Complex64], _key_params: &[Complex64], _value_params: &[Complex64], _num_heads: usize, ) -> QuantRS2Result<()>

Apply a quantum ML attention mechanism.

The specialized attention kernels (CUDA/WebGPU) are not yet authored and there is no defined CPU reference for the previously no-op fallback. Returns an honest error rather than silently returning success without transforming state. DEFERRED until the real kernels exist.

pub fn apply_fused_gate_sequence( &self, state: &mut [Complex64], gates: &[Box<dyn GateOp>], ) -> QuantRS2Result<()>

Apply a sequence of gates to the state vector.

Each gate is applied with a real matrix-vector update on the CPU via [apply_single_gate_optimized]. Gate fusion (merging adjacent gates into a single combined matrix for fewer passes) is DEFERRED: [analyze_gate_fusion_opportunities] currently finds no chains, so this reduces to honest per-gate application. The result is numerically correct; only the fused-pass optimization is missing.

pub fn get_performance_report(&self) -> PerformanceReport

Get performance report