grace.md

# Using NVIDIA Grace Partition

For testing your application on the IBM Power partition,
you need to prepare a job script for that partition or use the interactive job:

```console
salloc -N 1 -c 144 -A PROJECT-ID -p p11-grace --time=08:00:00
```

where:

- `-N 1` means allocation single node,
- `-c 144` means allocation 144 cores,
- `-p p11-grace` is NVIDIA Grace partition,
- `--time=08:00:00` means allocation for 8 hours.

## Available Toolchains

The platform offers three toolchains:
- Standard GCC (as a module `ml GCC`)
- [NVHPC](https://developer.nvidia.com/hpc-sdk) (as a module `ml NVHPC`)
- [Clang for NVIDIA Grace](https://developer.nvidia.com/grace/clang) (installed in `/opt/nvidia/clang`)

!!! note
    The NVHPC toolchain showed strong results with minimal amount of tunning necessary in our initial evaluation.

### GCC Toolchain

The GCC compiler seems to struggle with vectorization of short (constant length) loops, which tend to get completely unrolled/eliminated instead of being vectorized. For example simple nested loop such as

```cpp
for(int i = 0; i < 1000000; ++i) {
    // Iterations dependent in "i"
    // ...
    for(int j = 0; j < 8; ++j) {
        // but independent in "j"
        // ...
    }
}
```
may emit scalar code for the inner loop leading to no vectorization being used at all.

### Clang (for Grace) Toolchain

The Clang/LLVM tends to behave similarly, but can be guided to properly vectorize the inner loop with either flags `-O3 -ffast-math -march=native -fno-unroll-loops -mllvm -force-vector-width=8` or pragmas such as `#pragma clang loop vectorize_width(8)` and `#pragma clang loop unroll(disable)`.

```cpp
for(int i = 0; i < 1000000; ++i) {
    // Iterations dependent in "i"
    // ...
    #pragma clang loop unroll(disable) vectorize_width(8)
    for(int j = 0; j < 8; ++j) {
        // but independent in "j"
        // ...
    }
}
```

!!! note
    Our basic experiments show that fixed width vectorization (NEON) tends to perform better in the case of short (register-length) loops than SVE. In cases (like above), where specified `vectorize_width` is larger than avaliable vector unit width, Clang will emit multiple NEON instructions (eg. 4 instructions will be emitted to process 8 64-bit operations in 128-bit units of Grace).

### NVHPC Toolchain

The NVHPC toolchain handled aforementioned case without any additional tunning. Simple `-O3 -march=native -fast` should be therefore sufficient.

## Basic Math Libraries

The basic libraries (BLAS and LAPACK) are included in NVHPC toolchain and can be used simply as `-lblas` and `-llapack` for BLAS and LAPACK respectively (`lp64` and `ilp64` versions are also included).

!!! note
    The Grace platform doesn't include CUDA-capable GPU, therefore `nvcc` will fail with an error. This means that `nvc`, `nvc++` and `nvfortran` should be used instead.

### NVIDIA Performance Libraries

The [NVPL](https://developer.nvidia.com/nvpl) package includes more extensive set of libraries in both sequential and multi-threaded versions:
- BLACS: `-lnvpl_blacs_{lp64,ilp64}_{mpich,openmpi3,openmpi4,openmpi5}`
- BLAS: `-lnvpl_blas_{lp64,ilp64}_{seq,gomp}`
- FFTW: `-lnvpl_fftw`
- LAPACK: `-lnvpl_lapack_{lp64,ilp64}_{seq,gomp}`
- ScaLAPACK: `-lnvpl_scalapack_{lp64,ilp64}`
- RAND: `-lnvpl_rand` or `-lnvpl_rand_mt`
- SPARSE: `-lnvpl_sparse`

This package should be compatible with all avaliable toolchains and includes CMake module files for easy integration into CMake-based projects. For further documentation see also [NVPL](https://docs.nvidia.com/nvpl).

## Basic Communication Libraries

The OpenMPI 4 implementation is included with NVHPC toolchain and is exposed as a module (`ml OpenMPI`). The following example

```cpp
#include <mpi.h>
#include <sched.h>
#include <omp.h>

int main(int argc, char **argv)
{
        int rank;
        MPI_Init(&argc, &argv);
        MPI_Comm_rank(MPI_COMM_WORLD, &rank);
        #pragma omp parallel
        {
                printf("Hello on rank %d, thread %d on CPU %d\n", rank, omp_get_thread_num(), sched_getcpu());
        }
        MPI_Finalize();
}
```

can be compiled and run as follows

```console
ml OpenMPI
mpic++ -fast -fopenmp hello.cpp -o hello
OMP_PROC_BIND=close OMP_NUM_THREADS=4 mpirun -np 4 --map-by slot:pe=36 ./hello
```
In this configuration we run 4 ranks bound to one quarter of cores each with 4 OpenMP threads.

## Simple BLAS Application

The `hello world` example application (written in `C++` and `Fortran`) uses simple stationary probability vector estimation to illustrate use of GEMM (BLAS 3 routine).

Stationary probability vector estimation in `C++`:

```cpp
#include <iostream>
#include <vector>
#include <chrono>
#include "cblas.h"

const size_t ITERATIONS  = 32;
const size_t MATRIX_SIZE = 1024;

int main(int argc, char *argv[])
{
    const size_t matrixElements = MATRIX_SIZE*MATRIX_SIZE;

    std::vector<float> a(matrixElements, 1.0f / float(MATRIX_SIZE));

    for(size_t i = 0; i < MATRIX_SIZE; ++i)
        a[i] = 0.5f / (float(MATRIX_SIZE) - 1.0f);
    a[0] = 0.5f;

    std::vector<float> w1(matrixElements, 0.0f);
    std::vector<float> w2(matrixElements, 0.0f);

    std::copy(a.begin(), a.end(), w1.begin());

    std::vector<float> *t1, *t2;
    t1 = &w1;
    t2 = &w2;

    auto c1 = std::chrono::steady_clock::now();

    for(size_t i = 0; i < ITERATIONS; ++i)
    {
        std::fill(t2->begin(), t2->end(), 0.0f);

        cblas_sgemm(CblasRowMajor, CblasNoTrans, CblasNoTrans, MATRIX_SIZE, MATRIX_SIZE, MATRIX_SIZE,
                    1.0f, t1->data(), MATRIX_SIZE,
                    a.data(), MATRIX_SIZE,
                    1.0f, t2->data(), MATRIX_SIZE);

        std::swap(t1, t2);
    }

    auto c2 = std::chrono::steady_clock::now();

    for(size_t i = 0; i < MATRIX_SIZE; ++i)
    {
        std::cout << (*t1)[i*MATRIX_SIZE + i] << " ";
    }

    std::cout << std::endl;

    std::cout << "Elapsed Time: " << std::chrono::duration<double>(c2 - c1).count() << std::endl;

    return 0;
}
```

Stationary probability vector estimation in `Fortran`:

```fortran
program main
    implicit none

    integer :: matrix_size, iterations
    integer :: i
    real, allocatable, target :: a(:,:), w1(:,:), w2(:,:)
    real, dimension(:,:), contiguous, pointer :: t1, t2, tmp
    real, pointer :: out_data(:), out_diag(:)
    integer :: cr, cm, c1, c2

    iterations  = 32
    matrix_size = 1024

    call system_clock(count_rate=cr)
    call system_clock(count_max=cm)

    allocate(a(matrix_size, matrix_size))
    allocate(w1(matrix_size, matrix_size))
    allocate(w2(matrix_size, matrix_size))

    a(:,:) = 1.0 / real(matrix_size)
    a(:,1) = 0.5 / real(matrix_size - 1)
    a(1,1) = 0.5

    w1 = a
    w2(:,:) = 0.0

    t1 => w1
    t2 => w2

    call system_clock(c1)

    do i = 0, iterations
        t2(:,:) = 0.0

        call sgemm('N', 'N', matrix_size, matrix_size, matrix_size, 1.0, t1, matrix_size, a, matrix_size, 1.0, t2, matrix_size)

        tmp => t1
        t1  => t2
        t2  => tmp
    end do

    call system_clock(c2)

    out_data(1:size(t1)) => t1
    out_diag => out_data(1::matrix_size+1)

    print *, out_diag
    print *, "Elapsed Time: ", (c2 - c1) / real(cr)

    deallocate(a)
    deallocate(w1)
    deallocate(w2)
end program main
```

### Using NVHPC Toolchain

The C++ version of the example can be compiled with NVHPC and ran as folows

```console
ml NVHPC
nvc++ -O3 -march=native -fast -I$NVHPC/Linux_aarch64/$EBVERSIONNVHPC/compilers/include/lp64 -lblas main.cpp -o main
OMP_NUM_THREADS=144 OMP_PROC_BIND=spread ./main
```

The Fortran version is just as simple:

```console
ml NVHPC
nvfortran -O3 -march=native -fast -lblas main.f90 -o main.x
OMP_NUM_THREADS=144 OMP_PROC_BIND=spread ./main
```

!!! note
    It may be advantageous to use NVPL libraries instead NVHPC ones. For example DGEMM BLAS 3 routine from NVPL is almost 30% faster than NVHPC one.

### Using Clang (for Grace) Toolchain

Similarly Clang for Grace toolchain with NVPL BLAS can be used to compile C++ version of the example.

```console
ml NVHPC
/opt/nvidia/clang/17.23.11/bin/clang++ -O3 -march=native -ffast-math -I$NVHPC/Linux_aarch64/$EBVERSIONNVHPC/compilers/include/lp64 -lnvpl_blas_lp64_gomp main.cpp -o main
```

!!! note
    NVHPC module is used just for the `cblas.h` include in this case. This can be avoided by changing the code to use `nvpl_blas.h` instead.

## Additional Resources

- [https://www.nvidia.com/en-us/data-center/grace-cpu-superchip/][1]
- [https://developer.nvidia.com/hpc-sdk][2]
- [https://developer.nvidia.com/grace/clang][3]
- [https://docs.nvidia.com/nvpl][4]

[1]: https://www.nvidia.com/en-us/data-center/grace-cpu-superchip/
[2]: https://developer.nvidia.com/hpc-sdk
[3]: https://developer.nvidia.com/grace/clang
[4]: https://docs.nvidia.com/nvpl