DelphiMatrixBenchmark - HPC Matrix Multiplication Benchmark Client

DelphiMatrixBenchmark is the desktop client application used in the HPC-Delphi research project to run, validate, and compare matrix multiplication implementations across multiple execution models:

• sequential baseline
• shared-memory parallelism (PPL, OTL, OpenMP)
• vectorized/SIMD kernels (VectorSIMD, AVX2)
• optimized linear algebra backends (ALGLIB, LinAlg/CBLAS, mrMath, MKL)
• distributed-memory MPI variants (MPI configuration)

The application provides a GUI to select algorithms, set benchmark parameters, execute repeated runs, and visualize results in both tabular and chart form.

---

Features

• Unified benchmark runner for heterogeneous implementations under a common interface (IMultiplier).
• Automatic result validation against a deterministic reference computation before accepting timings.
• Configurable workload: matrix dimensions (M, K, N), worker threads (T), repetitions (S).
• Comparative visualization in VCL UI (TStringGrid + TeeChart bar plots).
• Two build configurations with explicit dependency boundaries:
  • Release: local/shared-memory algorithms
  • MPI: adds distributed algorithms and MPI runtime coordination

---

Architecture Overview

flowchart TD
    subgraph MPI_Mode[Only in MPI configuration]
      MPIINIT[DelphiMatrixBenchmark.dpr<br/>MPI_Init · rank split]
    end

    MPIINIT -->|rank 0| UI
    MPIINIT -->|worker ranks| IMPL

    UI[Form.pas<br/>VCL GUI] --> CFG[Benchmark/Config.pas<br/>M,K,N,T,S]
    UI --> FAC[Matrix/Factory.pas<br/>Algorithm registry]
    UI --> RUN[Benchmark/Runner.pas<br/>Execution loop]

    FAC --> IMPL[Matrix/MultImpls.pas<br/>IMultiplier implementations]
    RUN --> IMPL
    RUN --> VAL[Benchmark/Validator.pas<br/>Correctness check]
    RUN --> RES[Benchmark/Result.pas<br/>Total/Avg/Min/Max]

    IMPL --> CORE1[Native Delphi<br/>Base / PPL / OTL / ASM]
    IMPL --> CORE2[HPC-Delphi DLLs<br/>OffC / VectorSIMD / MKL]
    IMPL --> CORE3[Third-party libs<br/>ALGLIB / LinAlg / mrMath / OptiVec]

    RUN --> MPILIB[mpi_delphi + MSMPI<br/>&#40;MPI config only&#41;]
    IMPL --> MPILIB

Loading

---

Repository Structure

DelphiMatrixBenchmark/
│
├── Benchmark/
│   ├── Config.pas          # Benchmark input parameters
│   ├── Result.pas          # Benchmark output model
│   ├── Runner.pas          # Execution orchestration
│   └── Validator.pas       # Numerical correctness validation
│
├── Matrix/
│   ├── Factory.pas         # Name -> IMultiplier factory
│   ├── Multiplier.pas      # IMultiplier interface
│   ├── MultImpls.pas       # Concrete algorithm implementations
│   └── Utils.pas           # Matrix helpers
│
├── Form.pas / Form.dfm     # GUI and interaction logic
├── DelphiMatrixBenchmark.dpr
├── DelphiMatrixBenchmark.dproj
├── CITATION.cff
├── LICENSE
└── README.md

---

Requirements & Development Environment

The project is developed and validated in the following environment:

• Operating System: Windows 11 (64-bit)
• IDE: RAD Studio / Delphi 12.1
• Target platform: Win64
• MPI runtime (MPI config only): Microsoft MPI (MSMPI)

---

Dependencies

This project uses two explicit dependency profiles (Release and MPI). The key idea is:

• Release includes local/shared-memory algorithms and their extra third-party libraries.
• MPI includes distributed algorithms and MPI stack, without requiring Release-only third-party libraries.

1) Dependency matrix (by type and configuration)

Type	Component	Release	MPI	Why
Toolkit	RAD Studio / Delphi 12.1 (Win64)	✅	✅	Build system / IDE
Library (HPC-Delphi)	offc_delphi	✅	✅	Core + MPI hybrid (MPI+OffC-OMP+AVX2)
Library (HPC-Delphi)	vector_simd_delphi	✅	✅	Core + MPI hybrid (MPI+PPL+VectorSIMD)
Library (HPC-Delphi)	mkl_delphi	✅	✅	Core + MPI hybrid (MPI+OffC-MKL)
Library (HPC-Delphi)	mpi_delphi	❌	✅	Delphi bindings for MPI primitives
Library (third-party)	ALGLIB	✅	❌	Release-only implementations
Library (third-party)	Winsoft Linear Algebra (LinAlg)	✅	❌	Release-only implementations
Library (third-party)	mrMath / FastMath units	✅	❌	Release-only implementations
Library (third-party)	OptiVec	✅	❌	Release-only implementations
Library (third-party)	OmniThreadLibrary (OTL)	✅	❌	Release-only implementations
Library (UI)	TeeChart VCL	✅	✅	Result visualization
Toolkit	Intel oneAPI Base Toolkit	✅	✅	MKL/OpenMP environment
Runtime	MSMPI Runtime	❌	✅	Required to execute MPI binaries

2) Dependency catalog

HPC-Delphi libraries

• offc_delphi
• vector_simd_delphi
• mkl_delphi
• mpi_delphi (MPI only)

Third-party libraries

• ALGLIB
• Winsoft Linear Algebra
• mrMath
• OptiVec
• FastMath
• OmniThreadLibrary
• Steema TeeChart VCL

Toolkits / SDKs

• RAD Studio / Delphi 12.1 (Win64)
• Intel oneAPI Base Toolkit
• Microsoft MPI SDK (MPI only)

Runtimes

• Microsoft MPI Runtime (MPI only)

Dependency setup in RAD Studio

Configure search paths in: Project -> Options -> Building -> Delphi Compiler -> Search path

Common paths (Release + MPI)

path\to\offc_delphi\interface
path\to\vector_simd_delphi\interface
path\to\mkl_delphi\interface

Release-only paths

path\to\alglib-delphi\wrapper
path\to\linalg\Delphi12-Win64
path\to\mrMath
path\to\OptiVec_for_Delphi\win64\Lib8
path\to\FastMath\FastMath

MPI-only path

path\to\mpi_delphi\interface

Note: these profiles are intentionally separated so MPI builds do not require Release-only third-party libraries.

---

Build configurations

The project defines two Win64 configurations with exclusive algorithm sets.

Release

• Compiler define: RELEASE
• Available implementations:
  • Base, OffC-Base
  • PPL, OTL, OffC-OMP
  • VectorSIMD, OffC-AVX2, ASM
  • PPL+VectorSIMD, OffC-OMP+AVX2
  • ALGLIB, LINALG, mrmath, OffC-MKL
• Requires Release-only third-party dependencies listed above.

MPI

• Compiler define: MPI
• Available implementations:
  • MPI+Base
  • MPI+PPL+VectorSIMD
  • MPI+OffC-OMP+AVX2
  • MPI+OffC-MKL
• Requires common dependencies listed above and mpi_delphi + MSMPI.
• Runtime behavior (DelphiMatrixBenchmark.dpr):
  • rank 0 hosts GUI and orchestrates runs
  • worker ranks execute broadcasted MPI algorithms

---

DLL deployment

To run the executable successfully, all required DLLs must be discoverable by Windows. Choose one of the following methods:

Method A: Add DLL directories to system PATH (Recommended for Development)

During active development, this avoids copying DLLs after every rebuild.

Add these directories to your Windows Path environment variable:

• path\to\offc_delphi\build
• path\to\vector_simd_delphi\build
• path\to\mkl_delphi\build
• path\to\alglib-delphi\wrapper (Release only)
• path\to\mpi_delphi\build (MPI only)

Method B: Copy DLLs into the executable folder (Recommended for Distribution)

Copy required DLLs into the selected build output directory:

• Win64\Release\
• Win64\MPI\

At minimum:

• offc_delphi.dll
• vector_simd_delphi.dll
• mkl_delphi.dll
• alglib405_64hpc.dll (Release only)
• mpi_delphi.dll (MPI only)
• MSMPI runtime available in system PATH (e.g., msmpi.dll) (MPI only)

---

RAD Studio build guide (step-by-step)

A) Build Release

1. Open DelphiMatrixBenchmark.dproj in RAD Studio.
2. Set Target Platform to Win64.
3. Select Build Configuration = Release.
4. Verify common + Release-only search paths listed above.
5. Ensure required runtime DLLs are discoverable using one of these options:
   • add their directories to system Path (development), or
   • copy them into Win64\Release\ (distribution).
6. Confirm at least: offc_delphi.dll, vector_simd_delphi.dll, mkl_delphi.dll, alglib405_64hpc.dll.
7. Build project: Project -> Build DelphiMatrixBenchmark.

B) Build MPI

1. Select Build Configuration = MPI.
2. Confirm mpi_delphi\interface is present in search path.
3. Ensure MSMPI runtime is installed.
4. Ensure required DLLs are discoverable using one of these options:
   • add their directories to system Path (including MSMPI runtime), or
   • copy all required DLLs to Win64\MPI\.
5. Confirm at least:
   • mpi_delphi.dll
   • MSMPI runtime available in PATH (e.g., msmpi.dll)
   • offc_delphi.dll for MPI+OffC-OMP+AVX2
   • vector_simd_delphi.dll for MPI+PPL+VectorSIMD
   • mkl_delphi.dll for MPI+OffC-MKL
6. Build project.

C) Run MPI executable

Run using mpiexec (example with 4 ranks):

mpiexec -n 4 Win64\MPI\DelphiMatrixBenchmark.exe

---

Running benchmarks

1. Launch the executable (Release or MPI build output).
2. Configure benchmark parameters:
   • M, K, N (matrix dimensions)
   • T (thread count for relevant algorithms)
   • S (number of repetitions)
3. Select one or more algorithm implementations.
4. Execute benchmark.
5. Review:
   • tabular metrics: Total, Avg, Min, Max
   • comparative bar chart

The runner validates numerical correctness before accepting timing results.

---

Known Constraints

Dependency availability and licensing

Some third-party dependencies used in this project may not be permanently available to all users:

• some were used under trial/evaluation licenses for the paper,
• some are open source,
• some are free-to-use but not open source.

Because of this, certain compute/linear-algebra implementations may be unavailable in other environments.

Building without unavailable third-party compute libraries

If one or more compute/linear-algebra dependencies are missing, the project can still be compiled by disabling those implementations.

1) Remove algorithm entries from the registry

Edit Matrix/Factory.pas:

• remove the corresponding name(s) from TFactory.GetAvailable
• remove the corresponding else if Name = '...' then Result := ... branch(es) from TFactory.CreateByName

2) Remove/comment the corresponding implementations

Edit Matrix/MultImpls.pas:

• remove unavailable unit(s) from the uses clause
• remove or comment class declarations and implementations tied to those unit(s)

Dependency -> code mapping

• ALGLIB

  • Unit: XALGLIB
  • Class: TALGLIB
  • Factory name: 'ALGLIB'

• Winsoft Linear Algebra

  • Unit: LinAlg
  • Class: TLinearAlgebra
  • Factory name: 'LINALG'

• OptiVec

  • Units: vecLib, VDstd, VDmath
  • Class: TOptiVecVectorLib
  • Factory name: 'OptiVec-VectorLib'

• mrMath / FastMath-dependent paths

  • Units: FMAMatrixMultOperationsx64, FMAMatrixMultTransposedOperationsx64
  • Classes: Tmrmath, TASM
  • Factory names: 'mrmath', 'ASM'

• OmniThreadLibrary (OTL)

  • Unit: OTLParallel
  • Class: TOTL
  • Factory name: 'OTL'

---

Academic citation

If you use this software in your research, please cite it using the metadata provided in CITATION.cff.

---

License

This project is licensed under the MIT License. See LICENSE for details.