Using AMD Partition¶
For testing your application on the AMD partition, you need to prepare a job script for that partition or use the interactive job:
salloc -N 1 -c 64 -A PROJECT-ID -p p03-amd --gres=gpu:4 --time=08:00:00
where:
-N 1
means allocating one server,-c 64
means allocating 64 cores,-A
is your project,-p p03-amd
is AMD partition,--gres=gpu:4
means allocating all 4 GPUs of the node,--time=08:00:00
means allocation for 8 hours.
You have also an option to allocate subset of the resources only,
by reducing the -c
and --gres=gpu
to smaller values.
salloc -N 1 -c 48 -A PROJECT-ID -p p03-amd --gres=gpu:3 --time=08:00:00
salloc -N 1 -c 32 -A PROJECT-ID -p p03-amd --gres=gpu:2 --time=08:00:00
salloc -N 1 -c 16 -A PROJECT-ID -p p03-amd --gres=gpu:1 --time=08:00:00
Note
p03-amd01 server has hyperthreading enabled therefore htop shows 128 cores.
p03-amd02 server has hyperthreading disabled therefore htop shows 64 cores.
Using AMD MI100 GPUs¶
The AMD GPUs can be programmed using the ROCm open-source platform.
ROCm and related libraries are installed directly in the system. You can find it here:
/opt/rocm/
The actual version can be found here:
[user@p03-amd02.cs]$ cat /opt/rocm/.info/version
5.5.1-74
Basic HIP Code¶
The first way how to program AMD GPUs is to use HIP.
The basic vector addition code in HIP looks like this.
This a full code and you can copy and paste it into a file.
For this example we use vector_add.hip.cpp
.
#include <cstdio>
#include <hip/hip_runtime.h>
__global__ void add_vectors(float * x, float * y, float alpha, int count)
{
long long idx = blockIdx.x * blockDim.x + threadIdx.x;
if(idx < count)
y[idx] += alpha * x[idx];
}
int main()
{
// number of elements in the vectors
long long count = 10;
// allocation and initialization of data on the host (CPU memory)
float * h_x = new float[count];
float * h_y = new float[count];
for(long long i = 0; i < count; i++)
{
h_x[i] = i;
h_y[i] = 10 * i;
}
// print the input data
printf("X:");
for(long long i = 0; i < count; i++)
printf(" %7.2f", h_x[i]);
printf("\n");
printf("Y:");
for(long long i = 0; i < count; i++)
printf(" %7.2f", h_y[i]);
printf("\n");
// allocation of memory on the GPU device
float * d_x;
float * d_y;
hipMalloc(&d_x, count * sizeof(float));
hipMalloc(&d_y, count * sizeof(float));
// copy the data from host memory to the device
hipMemcpy(d_x, h_x, count * sizeof(float), hipMemcpyHostToDevice);
hipMemcpy(d_y, h_y, count * sizeof(float), hipMemcpyHostToDevice);
int tpb = 256;
int bpg = (count - 1) / tpb + 1;
// launch the kernel on the GPU
add_vectors<<< bpg, tpb >>>(d_x, d_y, 100, count);
// hipLaunchKernelGGL(add_vectors, bpg, tpb, 0, 0, d_x, d_y, 100, count);
// copy the result back to CPU memory
hipMemcpy(h_y, d_y, count * sizeof(float), hipMemcpyDeviceToHost);
// print the results
printf("Y:");
for(long long i = 0; i < count; i++)
printf(" %7.2f", h_y[i]);
printf("\n");
// free the allocated memory
hipFree(d_x);
hipFree(d_y);
delete[] h_x;
delete[] h_y;
return 0;
}
To compile the code we use hipcc
compiler.
For compiler information, use hipcc --version
:
[user@p03-amd02.cs ~]$ hipcc --version
HIP version: 5.5.30202-eaf00c0b
AMD clang version 16.0.0 (https://github.com/RadeonOpenCompute/llvm-project roc-5.5.1 23194 69ef12a7c3cc5b0ccf820bc007bd87e8b3ac3037)
Target: x86_64-unknown-linux-gnu
Thread model: posix
InstalledDir: /opt/rocm-5.5.1/llvm/bin
The code is compiled a follows:
hipcc vector_add.hip.cpp -o vector_add.x
The correct output of the code is:
[user@p03-amd02.cs ~]$ ./vector_add.x
X: 0.00 1.00 2.00 3.00 4.00 5.00 6.00 7.00 8.00 9.00
Y: 0.00 10.00 20.00 30.00 40.00 50.00 60.00 70.00 80.00 90.00
Y: 0.00 110.00 220.00 330.00 440.00 550.00 660.00 770.00 880.00 990.00
More details on HIP programming is in the HIP Programming Guide
HIP and ROCm Libraries¶
The list of official AMD libraries can be found here.
The libraries are installed in the same directory is ROCm
/opt/rocm/
Following libraries are installed:
drwxr-xr-x 4 root root 44 Jun 7 14:09 hipblas
drwxr-xr-x 3 root root 17 Jun 7 14:09 hipblas-clients
drwxr-xr-x 3 root root 29 Jun 7 14:09 hipcub
drwxr-xr-x 4 root root 44 Jun 7 14:09 hipfft
drwxr-xr-x 3 root root 25 Jun 7 14:09 hipfort
drwxr-xr-x 4 root root 32 Jun 7 14:09 hiprand
drwxr-xr-x 4 root root 44 Jun 7 14:09 hipsolver
drwxr-xr-x 4 root root 44 Jun 7 14:09 hipsparse
and
drwxr-xr-x 4 root root 32 Jun 7 14:09 rocalution
drwxr-xr-x 4 root root 44 Jun 7 14:09 rocblas
drwxr-xr-x 4 root root 44 Jun 7 14:09 rocfft
drwxr-xr-x 4 root root 32 Jun 7 14:09 rocprim
drwxr-xr-x 4 root root 32 Jun 7 14:09 rocrand
drwxr-xr-x 4 root root 44 Jun 7 14:09 rocsolver
drwxr-xr-x 4 root root 44 Jun 7 14:09 rocsparse
drwxr-xr-x 3 root root 29 Jun 7 14:09 rocthrust
Using HipBlas Library¶
The basic code in HIP that uses hipBlas looks like this.
This a full code and you can copy and paste it into a file.
For this example we use hipblas.hip.cpp
.
#include <cstdio>
#include <vector>
#include <cstdlib>
#include <hip/hip_runtime.h>
#include <hipblas/hipblas.h>
int main()
{
srand(9600);
int width = 10;
int height = 7;
int elem_count = width * height;
// initialization of data in CPU memory
float * h_A;
hipHostMalloc(&h_A, elem_count * sizeof(*h_A));
for(int i = 0; i < elem_count; i++)
h_A[i] = (100.0f * rand()) / (float)RAND_MAX;
printf("Matrix A:\n");
for(int r = 0; r < height; r++)
{
for(int c = 0; c < width; c++)
printf("%6.3f ", h_A[r + height * c]);
printf("\n");
}
float * h_x;
hipHostMalloc(&h_x, width * sizeof(*h_x));
for(int i = 0; i < width; i++)
h_x[i] = (100.0f * rand()) / (float)RAND_MAX;
printf("vector x:\n");
for(int i = 0; i < width; i++)
printf("%6.3f ", h_x[i]);
printf("\n");
float * h_y;
hipHostMalloc(&h_y, height * sizeof(*h_y));
for(int i = 0; i < height; i++)
h_x[i] = 100.0f + i;
printf("vector y:\n");
for(int i = 0; i < height; i++)
printf("%6.3f ", h_x[i]);
printf("\n");
// initialization of data in GPU memory
float * d_A;
size_t pitch_A;
hipMallocPitch((void**)&d_A, &pitch_A, height * sizeof(*d_A), width);
hipMemcpy2D(d_A, pitch_A, h_A, height * sizeof(*d_A), height * sizeof(*d_A), width, hipMemcpyHostToDevice);
int lda = pitch_A / sizeof(float);
float * d_x;
hipMalloc(&d_x, width * sizeof(*d_x));
hipMemcpy(d_x, h_x, width * sizeof(*d_x), hipMemcpyHostToDevice);
float * d_y;
hipMalloc(&d_y, height * sizeof(*d_y));
hipMemcpy(d_y, h_y, height * sizeof(*d_y), hipMemcpyHostToDevice);
// basic calculation of the result on the CPU
float alpha=2.0f, beta=10.0f;
for(int i = 0; i < height; i++)
h_y[i] *= beta;
for(int r = 0; r < height; r++)
for(int c = 0; c < width; c++)
h_y[r] += alpha * h_x[c] * h_A[r + height * c];
printf("result y CPU:\n");
for(int i = 0; i < height; i++)
printf("%6.3f ", h_y[i]);
printf("\n");
// calculation of the result on the GPU using the hipBLAS library
hipblasHandle_t blas_handle;
hipblasCreate(&blas_handle);
hipblasSgemv(blas_handle, HIPBLAS_OP_N, height, width, &alpha, d_A, lda, d_x, 1, &beta, d_y, 1);
hipDeviceSynchronize();
hipblasDestroy(blas_handle);
// copy the GPU result to CPU memory and print it
hipMemcpy(h_y, d_y, height * sizeof(*d_y), hipMemcpyDeviceToHost);
printf("result y BLAS:\n");
for(int i = 0; i < height; i++)
printf("%6.3f ", h_y[i]);
printf("\n");
// free all the allocated memory
hipFree(d_A);
hipFree(d_x);
hipFree(d_y);
hipHostFree(h_A);
hipHostFree(h_x);
hipHostFree(h_y);
return 0;
}
The code compilation can be done as follows:
hipcc hipblas.hip.cpp -o hipblas.x -lhipblas
Using HipSolver Library¶
The basic code in HIP that uses hipSolver looks like this.
This a full code and you can copy and paste it into a file.
For this example we use hipsolver.hip.cpp
.
#include <cstdio>
#include <vector>
#include <cstdlib>
#include <algorithm>
#include <hipsolver/hipsolver.h>
#include <hipblas/hipblas.h>
int main()
{
srand(63456);
int size = 10;
// allocation and initialization of data on host. this time we use std::vector
int h_A_ld = size;
int h_A_pitch = h_A_ld * sizeof(float);
std::vector<float> h_A(size * h_A_ld);
for(int r = 0; r < size; r++)
for(int c = 0; c < size; c++)
h_A[r * h_A_ld + c] = (10.0 * rand()) / RAND_MAX;
printf("System matrix A:\n");
for(int r = 0; r < size; r++)
{
for(int c = 0; c < size; c++)
printf("%6.3f ", h_A[r * h_A_ld + c]);
printf("\n");
}
std::vector<float> h_b(size);
for(int i = 0; i < size; i++)
h_b[i] = (10.0 * rand()) / RAND_MAX;
printf("RHS vector b:\n");
for(int i = 0; i < size; i++)
printf("%6.3f ", h_b[i]);
printf("\n");
std::vector<float> h_x(size);
// memory allocation on the device and initialization
float * d_A;
size_t d_A_pitch;
hipMallocPitch((void**)&d_A, &d_A_pitch, size, size);
int d_A_ld = d_A_pitch / sizeof(float);
float * d_b;
hipMalloc(&d_b, size * sizeof(float));
float * d_x;
hipMalloc(&d_x, size * sizeof(float));
int * d_piv;
hipMalloc(&d_piv, size * sizeof(int));
int * info;
hipMallocManaged(&info, sizeof(int));
hipMemcpy2D(d_A, d_A_pitch, h_A.data(), h_A_pitch, size * sizeof(float), size, hipMemcpyHostToDevice);
hipMemcpy(d_b, h_b.data(), size * sizeof(float), hipMemcpyHostToDevice);
// solving the system using hipSOLVER
hipsolverHandle_t solverHandle;
hipsolverCreate(&solverHandle);
int wss_trf, wss_trs; // wss = WorkSpace Size
hipsolverSgetrf_bufferSize(solverHandle, size, size, d_A, d_A_ld, &wss_trf);
hipsolverSgetrs_bufferSize(solverHandle, HIPSOLVER_OP_N, size, 1, d_A, d_A_ld, d_piv, d_b, size, &wss_trs);
float * workspace;
int wss = std::max(wss_trf, wss_trs);
hipMalloc(&workspace, wss * sizeof(float));
hipsolverSgetrf(solverHandle, size, size, d_A, d_A_ld, workspace, wss, d_piv, info);
hipsolverSgetrs(solverHandle, HIPSOLVER_OP_N, size, 1, d_A, d_A_ld, d_piv, d_b, size, workspace, wss, info);
hipMemcpy(d_x, d_b, size * sizeof(float), hipMemcpyDeviceToDevice);
hipMemcpy(h_x.data(), d_x, size * sizeof(float), hipMemcpyDeviceToHost);
printf("Solution vector x:\n");
for(int i = 0; i < size; i++)
printf("%6.3f ", h_x[i]);
printf("\n");
hipFree(workspace);
hipsolverDestroy(solverHandle);
// perform matrix-vector multiplication A*x using hipBLAS to check if the solution is correct
hipblasHandle_t blasHandle;
hipblasCreate(&blasHandle);
float alpha = 1;
float beta = 0;
hipMemcpy2D(d_A, d_A_pitch, h_A.data(), h_A_pitch, size * sizeof(float), size, hipMemcpyHostToDevice);
hipblasSgemv(blasHandle, HIPBLAS_OP_N, size, size, &alpha, d_A, d_A_ld, d_x, 1, &beta, d_b, 1);
hipDeviceSynchronize();
hipblasDestroy(blasHandle);
for(int i = 0; i < size; i++)
h_b[i] = 0;
hipMemcpy(h_b.data(), d_b, size * sizeof(float), hipMemcpyDeviceToHost);
printf("Check multiplication vector Ax:\n");
for(int i = 0; i < size; i++)
printf("%6.3f ", h_b[i]);
printf("\n");
// free all the allocated memory
hipFree(info);
hipFree(d_piv);
hipFree(d_x);
hipFree(d_b);
hipFree(d_A);
return 0;
}
The code compilation can be done as follows:
hipcc hipsolver.hip.cpp -o hipsolver.x -lhipblas -lhipsolver
Using OpenMP Offload to Program AMD GPUs¶
The ROCm™ installation includes an LLVM-based implementation that fully supports the OpenMP 4.5 standard and a subset of the OpenMP 5.0 standard. Fortran, C/C++ compilers, and corresponding runtime libraries are included.
The OpenMP toolchain is automatically installed as part of the standard ROCm installation
and is available under /opt/rocm/llvm
. The sub-directories are:
bin
: Compilers (flang and clang) and other binaries.examples
: The usage section below shows how to compile and run these programs.include
: Header files.lib
: Libraries including those required for target offload.lib-debug
: Debug versions of the above libraries.
More information can be found in the AMD OpenMP Support Guide.
Compilation of OpenMP Code¶
Basic example that uses OpenMP offload is here.
Again, code is complete and can be copied and pasted into a file.
Here we use vadd.cpp
.
#include <cstdio>
#include <cstdlib>
int main(int argc, char ** argv)
{
long long count = 1 << 20;
if(argc > 1)
count = atoll(argv[1]);
long long print_count = 16;
if(argc > 2)
print_count = atoll(argv[2]);
long long * a = new long long[count];
long long * b = new long long[count];
long long * c = new long long[count];
#pragma omp parallel for
for(long long i = 0; i < count; i++)
{
a[i] = i;
b[i] = 10 * i;
}
printf("A: ");
for(long long i = 0; i < print_count; i++)
printf("%3lld ", a[i]);
printf("\n");
printf("B: ");
for(long long i = 0; i < print_count; i++)
printf("%3lld ", b[i]);
printf("\n");
#pragma omp target map(to: a[0:count],b[0:count]) map(from: c[0:count])
#pragma omp teams distribute parallel for
for(long long i = 0; i < count; i++)
{
c[i] = a[i] + b[i];
}
printf("C: ");
for(long long i = 0; i < print_count; i++)
printf("%3lld ", c[i]);
printf("\n");
delete[] a;
delete[] b;
delete[] c;
return 0;
}
This code can be compiled like this:
/opt/rocm/llvm/bin/clang++ -O3 -target x86_64-pc-linux-gnu -fopenmp -fopenmp-targets=amdgcn-amd-amdhsa -Xopenmp-target=amdgcn-amd-amdhsa -march=gfx908 vadd.cpp -o vadd.x
These options are required for target offload from an OpenMP program:
-target x86_64-pc-linux-gnu
-fopenmp
-fopenmp-targets=amdgcn-amd-amdhsa
-Xopenmp-target=amdgcn-amd-amdhsa
This flag specifies the GPU architecture of targeted GPU.
You need to chage this when moving for instance to LUMI with MI250X GPU.
The MI100 GPUs presented in CS have code gfx908
:
-march=gfx908
Note: You also have to include the O0
, O2
, O3
or O3
flag.
Without this flag the execution of the compiled code fails.