Module tensor_cores 

Advanced GPU Tensor Core utilization for spatial algorithms

This module provides cutting-edge implementations that leverage modern GPU tensor cores (NVIDIA’s Tensor Cores, AMD’s Matrix Cores, Intel’s XMX units) for maximum performance in spatial computing. It includes mixed-precision operations, automatic layout optimization, and hardware-specific kernel selection for optimal throughput.

Features

• Tensor Core acceleration for matrix operations in spatial algorithms
• Mixed-precision computing (FP16, BF16, INT8, INT4) for maximum throughput
• Automatic tensor layout optimization for memory coalescing
• Hierarchical tiling strategies for large datasets
• Multi-GPU tensor parallelism for distributed spatial computation
• Dynamic precision selection based on numerical stability requirements
• Fused kernel operations to minimize memory bandwidth
• Async execution pipelines for maximum GPU utilization

Supported Hardware

• NVIDIA: V100, A100, H100, RTX 30/40 series (Tensor Cores)
• AMD: MI250X, MI300 series (Matrix Cores)
• Intel: Ponte Vecchio, Arc GPUs (XMX units)
• Automatic fallback to standard compute units when tensor cores unavailable

Examples

use scirs2_spatial::tensor_cores::{TensorCoreDistanceMatrix, TensorCoreClustering, PrecisionMode};
use scirs2_core::ndarray::array;

// Tensor core distance matrix computation
let points = array![[0.0, 0.0], [1.0, 0.0], [0.0, 1.0], [1.0, 1.0]];

let mut tensor_matrix = TensorCoreDistanceMatrix::new()?
    .with_precision_mode(PrecisionMode::Mixed16)
    .with_tensor_layout_optimization(true)
    .with_hierarchical_tiling(true);

let distances = tensor_matrix.compute_parallel(&points.view()).await?;
println!("Tensor core distance matrix: {:?}", distances);

// Tensor core k-means clustering
let mut tensor_kmeans = TensorCoreClustering::new(2)?
    .with_tensor_cores(true)
    .with_mixed_precision(true)
    .with_dynamic_precision_scaling(true);

let (centroids, assignments) = tensor_kmeans.fit(&points.view()).await?;
println!("Tensor core centroids: {:?}", centroids);

Structs

AdvancedTensorCoreDistanceMatrix

Tensor core distance matrix computer with advanced stability monitoring

DynamicPrecisionConfig

Dynamic precision scaling configuration

ErrorRecoverySystem

Advanced error recovery system

NumericalStabilityMonitor

Real-time numerical stability monitor

PerformanceAccuracyAnalyzer

Performance-accuracy trade-off analyzer

RecoveryAttempt

Recovery attempt record

StabilityMetrics

Numerical stability metrics

TensorCoreCapabilities

Tensor core capabilities

TensorCoreClustering

Tensor core clustering algorithm

TensorCoreDistanceMatrix

Tensor core distance matrix computer

TradeOffParams

Trade-off optimization parameters

Enums

GpuArchitecture

GPU architecture types

NumericalErrorType

Numerical error types

OptimizationObjective

Optimization objectives

PrecisionMode

Precision modes for tensor core operations

RecoveryAction

Recovery action types

ScalingStrategy

Dynamic precision scaling strategy

StabilityLevel

Numerical stability level

TensorCoreType

Tensor core types

TensorLayout

Tensor layout optimization strategies

Functions

detect_tensor_core_capabilities

Detect tensor core capabilities of available GPU hardware