TileMind

TileMind is a tile size optimization tool designed to automatically determine optimal tile sizes. The method itself is compiler-agnostic and is not bound to any specific compiler infrastructure.

Note on this repository: The test code currently included in this repository is built on top of PolyBench/C, and uses PPCG as the underlying compiler for benchmarking purposes. PPCG is a dependency of the test code only, not of the TileMind method itself.

---

Prerequisites

To run the test code in this repository (based on PolyBench/C), PPCG must be installed and properly configured on your system.

PPCG is required only for the PolyBench-based test code included here. The TileMind method itself is compiler-agnostic and can be adapted to other compiler backends.

---

Hardware Parameters Notice

⚠️ The hardware parameters in the auto-generated config files are profiled specifically for Intel Xeon Platinum 8538C. If you are running on a different platform, you must update the following fields in the config file (<benchmark>_ilp_params_*.py) to match your hardware.

The test platform configuration is as follows:

Parameter	Value
CPU	Intel Xeon Platinum 8538C
Architecture	x86_64
Physical Cores	56
Logical Processors	112 (Hyper-Threading, 2 threads/core)
Base Frequency	~2.7 GHz
L1 Data Cache	2.6 MiB (46 KiB per core × 56)
L2 Cache	112 MiB (2 MiB per core × 56)
L3 Cache	280 MiB (shared)
NUMA Nodes	2 (Node0: CPU 0–27, 56–83 / Node1: CPU 28–55, 84–111)

Cache & thread parameters (per-core values used in config):

# Cache parameters
L1_size = 49152      # L1 data cache size in bytes (~48 KiB per core)
L2_size = 2097152    # L2 cache size in bytes (2 MiB per core)
cache_line = 64      # Cache line size in bytes

# Parallelization information
num_threads = 8      # Number of threads used in current tests
                     # (Physical cores: 56, logical processors: 112, but only 8 used here)

Cache & operation latencies:

# Cache access latencies (nanoseconds)
T_L1 = 3.13   # L1 cache access latency
T_L2 = 7.42   # L2 cache access latency

# Operation latencies (nanoseconds)
T_add = 0.76
T_sub = 0.82
T_mul = 1.55
T_div = 5.19
T_sqrt = 20.00
T_trig = 40.00
T_log_exp = 35.00
T_pow = 45.00

Statement costs in the objective function (the constant coefficients, e.g., in gemm):

# Statement costs (nanoseconds)
T_op1 = 7.42   # Statement 1 cost (block 0)
T_op2 = 16.78  # Statement 2 cost (block 0)

# Block-specific objective functions
block_0_objective = "c0_b0_blocks * c1_b0_blocks * 7.42 + c0_b0_blocks * c1_b0_blocks * c2_b0_blocks * c2_b0_tile * 16.78"

# Auto-generated objective function
objective_function = "(c0_b0_blocks * c1_b0_blocks * 7.42 + c0_b0_blocks * c1_b0_blocks * c2_b0_blocks * c2_b0_tile * 16.78)"

These constants represent profiled per-statement execution costs and must be updated if you retarget to a different platform.

---

Directory Structure

Each benchmark directory (e.g., datamining/correlation/) contains the following key files:

<benchmark>/
├── benchmark_ilp_auto.py            # Tile size solver (sequential)
├── benchmark_ilp_auto_openmp.py     # Tile size solver (OpenMP parallel)
├── benchmark_preprocess.py          # Config file generator (sequential)
├── benchmark_preprocess_openmp.py   # Config file generator (OpenMP parallel)
├── <benchmark>_ilp_params_seq.py    # Config file for sequential (generated)
└── <benchmark>_ilp_params_openmp.py # Config file for OpenMP (generated)

Single config file: If only one <benchmark>_ilp_params.py exists (without _seq / _openmp suffix), it indicates that the loop structure is identical between the sequential and OpenMP versions — the only difference is the number of threads. In this case, the same config file can be used to solve both cases.

---

Workflow

TileMind supports two workflows depending on whether a configuration file already exists.

✅ Case 1: Config File Already Exists

If the config file is already present, directly run the solver:

# Sequential version
python3 benchmark_ilp_auto.py <benchmark>_ilp_params_seq.py

# OpenMP parallel version
python3 benchmark_ilp_auto_openmp.py <benchmark>_ilp_params_openmp.py

# If only a single shared config exists
python3 benchmark_ilp_auto.py <benchmark>_ilp_params.py
python3 benchmark_ilp_auto_openmp.py <benchmark>_ilp_params.py

🔧 Case 2: No Config File (First-time Setup)

If no config file exists, first generate it via the preprocessing script, then run the solver.

Step 1: Generate the config file

# Sequential version
python3 benchmark_preprocess.py <benchmark>.c

# OpenMP parallel version
python3 benchmark_preprocess_openmp.py <benchmark>.c

Step 2: Run the tile size solver

# Sequential version
python3 benchmark_ilp_auto.py <benchmark>_ilp_params_seq.py

# OpenMP parallel version
python3 benchmark_ilp_auto_openmp.py <benchmark>_ilp_params_openmp.py

---

Quick Reference

Scenario	Command
Config exists, sequential	python3 benchmark_ilp_auto.py <benchmark>_ilp_params_seq.py
Config exists, OpenMP	python3 benchmark_ilp_auto_openmp.py <benchmark>_ilp_params_openmp.py
Single shared config, sequential	python3 benchmark_ilp_auto.py <benchmark>_ilp_params.py
Single shared config, OpenMP	python3 benchmark_ilp_auto_openmp.py <benchmark>_ilp_params.py
No config, generate (sequential)	python3 benchmark_preprocess.py <benchmark>.c
No config, generate (OpenMP)	python3 benchmark_preprocess_openmp.py <benchmark>.c

---

Notes

• All benchmarks are based on PolyBench/C 4.2.1.
• PPCG must be configured before running any script.
• The _ilp_params_*.py config files are generated locally and do not need to be committed to the repository.
• Hardware parameters in config files are profiled on Intel Xeon Platinum 8538C. Update them if testing on a different platform.