Add OpenACC GEMM implementation (src/01_matmul)

[Copilot]

Adds a new src/01_matmul/ directory with an OpenACC port of the BLAS GEMM benchmark (C = α·A·B + β·C), supporting float and double precision.

Key files

• main.cpp — OpenACC kernel using #pragma acc parallel loop collapse(2) with inner reduction(+:s), wrapped in an #pragma acc data region; includes a host reference implementation for correctness validation and a timed benchmark loop
• utils.h — Shared helpers: rand_matrix, performance (GFLOP/s reporting), print_2x2_matrix_values (DEBUG only)
• Makefile — Auto-detects compiler: NVHPC (nvc++ -acc=gpu -gpu=$(GPU_ARCH)) or GCC (-fopenacc)

Kernel structure

#pragma acc parallel loop collapse(2) \
    present(A[0:M*K], B[0:K*N], C[0:M*N])
for (int i = 0; i < M; i++) {
  for (int j = 0; j < N; j++) {
    fp s = fp(0);
    #pragma acc loop reduction(+:s)
    for (int k = 0; k < K; k++)
      s += A[i * K + k] * B[k * N + j];
    C[i * N + j] = alpha * s + beta * C[i * N + j];
  }
}

Build

# NVHPC
make CC=nvc++ GPU_ARCH=cc80

# GCC
make CC=g++

./main 1024 1024 1024 100

Original prompt

01_matmul内にOpenACCのmatmulコードを作成してください。

[Chronological Review: The conversation began with user requests related to file operations and compiling code. The user then requested to port a BLAS GEMM implementation to Kokkos, followed by commands to compile and run the code. The user faced compilation errors and sought assistance with running the code on a GPU. The conversation evolved to include requests for performance comparisons across different implementations and pushing changes to a repository. The user also inquired about saving conversation context and creating new notes, leading to a request for OpenACC code in a specific file.]

[Intent Mapping:

1. "mv 02_kokkos" - User requested to move a directory.
2. "test" - User initiated a test command.
3. "Please port blas-gemm in HeCBench to Kokkos implementation" - User aimed to convert a specific implementation.
4. "Can you compile and run blas-gemm-kokkos/main.cpp?" - User requested compilation and execution of a specific file.
5. "Continue: 'Continue to iterate?'" - User sought to continue a previous discussion.
6. "How to compile main.cpp? I got this error..." - User needed help with compilation errors.
7. "main.cpp を GPU 上で実行するためのコマンドを教えて" - User requested commands for GPU execution.
8. "以下のエラーが出ます＞" - User reported an error encountered during execution.
9. "OpenMPではなく、CUDAで動かしたい場合はどうすれば良いですか？" - User inquired about running code with CUDA instead of OpenMP.
10. "Please run blas-gemm in cuda, hip, kokkos, omp, sycl versions..." - User requested performance comparisons across different implementations.
11. "Can you push the changes to the repository?" - User asked to push changes to a remote repository.
12. "Can you push the changes in HeCBench (only blas-gemm-kokkos) to the remote repo?" - User specified pushing changes to a particular repository.
13. "この window を閉じたらこれまでの会話は消えてしまいますか？" - User inquired about the persistence of conversation history.
14. "以下に保存してください。" - User requested to save conversation context to a specific file.
15. "間違って消してしまいました。再度メモを作成してください。" - User asked to recreate a note after deletion.
16. "テスト" - User initiated another test command.
17. "Cancel: 'Delegate to cloud agent'" - User canceled a previous command.
18. "01_matmul内にOpenACCのmatmulコードを作成してください。" - User requested to create OpenACC code in a specific file.]

[Technical Inventory:

• Technologies: Kokkos, CUDA, OpenMP, OpenACC, HeCBench.
• Tools: Makefile for compilation, Git for version control.
• Concepts: Performance benchmarking, GPU execution, code porting.]

[Code Archaeology:

• Files discussed: main.cpp, 01_matmul.
• Functions: BLAS GEMM implementation, Kokkos implementation, OpenACC code creation.
• Changes: User requested modifications and troubleshooting for compilation and execution.]

[Progress Assessment:

• Completed: User has compiled and run some code, identified errors, and requested specific implementations.
• Pending: User requests for performance comparisons, pushing changes, and creating OpenACC code.]

[Context Validation: All user requests and technical details are captured for continuation.]

[Recent Commands Analysis:

• Last Agent Commands: User requested to create OpenACC code in 01_matmul.
• Tool Results Summary: No specific tool results were returned as the focus was on user requests.
• Pre-Summary State: The agent was actively addressing user requests related to code implementation and troubleshooting.
• Operation Context: User's recent commands were aimed at enhancing code functionality and performance.]

1. Conversation Overview: - Primary Objectives: User requested to move directories, compile code, port implementations, troubleshoot errors, and create performance comparisons. - Session Context: The conversation transitioned from file operations to code compilation and performance benchmarking, with a focus on Kokkos and GPU execution. - User Intent Evolution: User's needs evolved from basic file management to complex code implementation and performance analysis.

2. Technical Foundation:

• Kokkos: A C++ library for parallel programming.
• CUDA: A parallel computing platform and application programming interface model created by NVIDIA.
• OpenMP: An API that supports multi-platform shared memory multiprocessing programming.
• OpenACC: A directive-based programming model for parallel computing.

3. Codebase Status:

• main.cpp:
  • Purpose: Contains the BLAS GEMM implementation.
  • Current State: User faced compilation errors related to CUDA flags.
  • Key Code Segments: Not specified, but related to matrix multiplication.
  • Dependencies: Kokkos and CUDA libraries.
• 01_matmul:
  • Purpose: User requested to create OpenACC code.
  • Current State: Pending creation of OpenACC code.

4. Problem Resolution:

• Issues Encountered: Compilation errors with unrecognized command-line options.
• Solutions Implemented: User sought guidance on how to compile and run code on a GPU.
• Debugging Context: Ongoing tro...

Created from VS Code.