Expand description
TPT GPU kernel parameter optimizer.
Three-phase search strategy:
- Grid search — exhaustive sweep of the full parameter space
- Hill-climbing — greedy local improvement from the best grid result
- AI-guided search — LLM suggests next candidates based on history
Re-exports§
pub use real_evaluator::generate_milestone_json;pub use real_evaluator::generate_milestone_report;pub use real_evaluator::optimize_all_gemm_problems;pub use real_evaluator::optimize_gemm_problem;pub use real_evaluator::standard_gemm_problems;pub use real_evaluator::GemmOptResult;pub use real_evaluator::GemmProblemConfig;pub use real_evaluator::RealGemmEvaluator;pub use attention_eval::generate_attention_milestone_json;pub use attention_eval::generate_attention_milestone_report;pub use attention_eval::optimize_all_attention_problems;pub use attention_eval::optimize_attention_problem;pub use attention_eval::standard_attention_problems;pub use attention_eval::AttentionOptResult;pub use attention_eval::AttentionProblemConfig;pub use attention_eval::RealAttentionEvaluator;pub use fused_eval::fused_gemm_problems;pub use fused_eval::generate_beat_cublas_json;pub use fused_eval::generate_beat_cublas_report;pub use fused_eval::optimize_fused_problem;pub use fused_eval::run_beat_cublas_campaign;pub use fused_eval::FusedGemmEvaluator;pub use fused_eval::FusedGemmOptResult;pub use fused_eval::FusedGemmProblem;pub use conv2d_eval::generate_conv2d_milestone_json;pub use conv2d_eval::generate_conv2d_milestone_report;pub use conv2d_eval::optimize_all_conv2d_problems;pub use conv2d_eval::optimize_conv2d_problem;pub use conv2d_eval::standard_conv2d_problems;pub use conv2d_eval::Conv2dOptResult;pub use conv2d_eval::Conv2dProblemConfig;pub use conv2d_eval::RealConv2dEvaluator;pub use normalization_eval::generate_normalization_milestone_json;pub use normalization_eval::generate_normalization_milestone_report;pub use normalization_eval::optimize_all_normalization_problems;pub use normalization_eval::optimize_normalization_problem;pub use normalization_eval::standard_normalization_problems;pub use normalization_eval::NormalizationOptResult;pub use normalization_eval::NormalizationProblemConfig;pub use normalization_eval::RealNormalizationEvaluator;pub use vector_add_eval::generate_vector_add_milestone_json;pub use vector_add_eval::generate_vector_add_milestone_report;pub use vector_add_eval::optimize_all_vector_add_problems;pub use vector_add_eval::optimize_vector_add_problem;pub use vector_add_eval::standard_vector_add_problems;pub use vector_add_eval::RealVectorAddEvaluator;pub use vector_add_eval::VectorAddOptResult;pub use vector_add_eval::VectorAddProblemConfig;pub use kernel_category::KernelCategory;
Modules§
- attention_
eval - Attention evaluator for the optimizer loop.
- conv2d_
eval - Conv2D evaluator for the optimizer loop.
- fused_
eval - Fused GEMM evaluator — models GEMM+bias+activation fusion vs cuBLAS pipeline.
- kernel_
category - Kernel category classification and kernel-specific parameter spaces.
- normalization_
eval - Normalization evaluator for the optimizer loop.
- real_
evaluator - Real-hardware evaluators for the optimizer loop.
- vector_
add_ eval - VectorAdd evaluator for the optimizer loop.
Structs§
- OptResult
- Optimization result for a single parameter configuration.
- Param
Dim - A single tuning dimension: name + allowed discrete values.
- Param
Space - The full discrete parameter space.
- Simulated
Evaluator - Synthetic evaluator — no hardware required.
- Tuning
Params - A concrete point in the tuning parameter space.
Traits§
Functions§
- ai_
guided_ search - LLM-guided search: ask the AI for the next candidate, evaluate, repeat.
- grid_
search - Evaluate every point in the space; return sorted results (best first).
- hill_
climb - Greedy hill-climb from
start; stops when no neighbor improves the score.