PolymerPIM

PolymerPIM is a C++ and Julia library for programming UPMEM DPUs with vector expressions. Operations are automatically fused and JIT-compiled into DPU kernels.

Build

source /usr/upmem_env.sh
make install BACKEND=hw PIPELINE=1 JIT=1
make test PIPELINE=1 JIT=1

make installs the repository-local dependencies on first use. Use BACKEND=simulator when DPU hardware is unavailable.

C++ API

Include <polymerpim.h> and link against libpolymerpim:

#include <polymerpim.h>

using namespace polymerpim;

init(64);
{
  std::vector<int32_t> host_a{1, 2, 3, 4};
  std::vector<int32_t> host_b{10, 10, 10, 10};
  int32_t centroid = 3;

  DPUVector<int32_t> a(host_a);
  DPUVector<int32_t> b(host_b);
  auto result = sqr(a - centroid) + b;
  auto total = sum(result);             // lazy reduction

  std::vector<int32_t> values = result.to_cpu();
  auto scalar = total.get();
}
shutdown();

Expressions remain lazy until a read, fence(vector), or sync(). Small indexed reductions can stay local to each DPU until read:

DPULocalVector<int32_t> bins(256);
bins[index] += 1;
std::vector<int32_t> histogram = bins.to_cpu();

The public header is installed under include/polymerpim; implementation headers under host/detail are private.

Julia API

PolymerPIM.jl provides the same lazy execution model through Julia arrays, broadcasts, reductions, and local accumulators:

using PolymerPIM

a = DPUVector(Int32[1, 2, 3, 4])
b = DPUVector(fill(Int32(10), 4))

result = abs2.(a .- 3) .+ b
total = sum(result)

Array(result)
total[]

See the Julia API guide for installation, supported operations, synchronization, JIT inspection, and tests.

Benchmarks

The self-contained benchmark suite and runner are documented in benchmarks/README.md.