Dpnp Migration

dpnp migration from NumPy and CuPy

Purpose

Answers one question about an existing program: which of its array calls dpnp implements, and what to do with the rest. Swapping import numpy as np for import dpnp as np moves the calls it covers and raises on the calls it does not, so a port is an inventory problem before it is a performance one.

The method here is to probe the installed release rather than consult a coverage list. A list of supported functions is the single most perishable claim about dpnp: it is accurate for the release someone wrote it against and silently wrong afterwards, in both directions. hasattr is not.

When to Use This Skill

Use this skill when:

• A NumPy or CuPy codebase has to run on Intel hardware and the question is whether it can.
• A call raises NotImplementedError or AttributeError after the import swap.
• The user asks whether dpnp supports a specific function or family.
• CuPy code needs the dpnp spelling of device selection or of a host copy.

Do not use this skill to diagnose a broken install or a missing SYCL runtime (dpnp-troubleshooting), to hand arrays to pandas, scikit-learn, PyTorch, or TensorFlow (dpnp-interop), to choose a device or manage buffers (dpnp-memory), to decide whether the workload belongs on a device at all (dpnp-quickstart), or to migrate a CUDA-based AI repository — model code, kernels, and framework calls are cuda-to-xpu-migration, not this skill.

Quick Start

import dpnp

hasattr(dpnp, "linspace")            # constructor present in this release?
hasattr(dpnp.linalg, "eigh")         # submodule member present?
hasattr(dpnp.fft, "fftn")            # same question, FFT surface

Three lines against the release the user has installed settle more than any table can. Everything below is how to act on the answers.

Implementation Guide

1. Inventory the surface the program actually uses. Grep for the np. call sites and reduce them to a set of names; that set, not the whole NumPy API, is the scope of the port.

2. Probe each name in the installed release. Presence is one question and signature is another:

   import dpnp
   
   wanted = ["sort", "argsort", "einsum", "interp", "unique"]
   missing = [name for name in wanted if not hasattr(dpnp, name)]
   
   import inspect
   inspect.signature(dpnp.sort)      # a present name can still lack a parameter
   

   A name that exists but rejects a keyword the program passes fails at runtime just as hard as an absent one, so read the signature for anything called with optional arguments.

3. Expect three outcomes, and verify each against the installed release rather than this list. The families are stable enough to plan with; the membership is not:

   Outcome	Families that usually land here
   Present	array construction, element-wise arithmetic and ufuncs, reductions, linalg, fft, basic and boolean indexing
   Present with a narrower signature	sorting, some random distributions, anything with a kind= or method= parameter
   Absent by design	string arrays, datetime64/timedelta64, structured and record arrays, polynomials, masked arrays

   The last row is not a gap waiting to be filled. Those families are host data structures rather than numeric kernels, so a device implementation is not pending — plan to keep that code on NumPy.

4. Wrap what is missing, once, at the call site. The fallback converts to the host, runs NumPy there, and comes back:

   import dpnp
   import numpy
   
   def unique_counts(array):
       """dpnp where it implements this, NumPy where it does not."""
       try:
           return dpnp.unique(array, return_counts=True)
       except (NotImplementedError, AttributeError, TypeError):
           values, counts = numpy.unique(dpnp.asnumpy(array), return_counts=True)
           return dpnp.array(values), dpnp.array(counts)
   

   TypeError belongs in that tuple: a narrower signature is how a partially implemented function refuses, and it is the outcome step 2 warns about.

5. Know what the conversions cost. dpnp.array(host_array) copies host to device and dpnp.asnumpy(device_array) copies device to host. dpnp.asarray avoids a copy only when its input already lives in USM memory reachable by the target queue — a NumPy array never does, so treat both directions as copies unless you have measured otherwise.

6. From CuPy, expect the device model to differ more than the array API. CuPy's cupy.cuda.Device(0).use() has no dpnp counterpart: there is no ambient current device to set. Placement is an argument at construction time:

   import dpnp
   
   x = dpnp.zeros(1024, device="gpu")             # explicit at creation
   y = dpnp.zeros(1024, sycl_queue=x.sycl_queue)  # or inherit the queue
   

   cupy.asnumpy maps onto dpnp.asnumpy. For anything else CuPy-specific — .get(), memory pools, RawKernel, cupyx.scipy — probe before promising an equivalent; the pool and kernel APIs in particular have no dpnp analogue to translate into.

7. Record what fell back. A port that ends with four wrapped functions and a note saying which they are is finished. One that ends with a wrapper around every call has hidden its own status, and nothing will tell you later which calls were ever on the device.

Performance

No measured numbers ship with this skill. What to measure once the port runs:

• The fallback rate on the hot path. Each fallback is two transfers plus a host computation, so a wrapped function called per iteration can cost more than the whole device stage saves.
• The end-to-end time against the unported original. A program that runs on the device but falls back inside its inner loop is the failure this skill exists to prevent, and only the whole-program timing shows it.
• Warm-up separately from steady state; first-call compilation is part of a port's measurements too (dpnp-quickstart covers the timing method).

Gotchas & Limitations

• A coverage list is a claim with a shelf life; hasattr is not. Probe the installed release, and say which release an answer was checked against.
• Presence does not imply the same signature. The parameter the program passes is the thing to check, not the name.
• NotImplementedError and AttributeError are different symptoms. The first is a function that exists and declines; the second is a name that is not there at all. A fallback that catches only one of them leaves the other crashing.
• A fallback inside a loop is correct code that loses the port. Wrap the function, not the iteration.
• CuPy's current-device idiom has no translation. Do not offer a context-manager equivalent; pass device= or sycl_queue= instead.
• The host-data families will not arrive. Strings, datetimes, structured arrays, polynomials, and masked arrays are not scheduled work, and telling a user to wait for them is wrong advice.
• Not covered: CUDA kernel sources and RawKernel, cupyx.scipy, framework and model migration, and any judgement about whether the ported workload is large enough to belong on a device.

References

One question here must never be answered from memory: what the installed release implements. The probe in step 2 is cheap, and a remembered coverage table is how this skill would tell a user that a function they need is missing when it is present, or present when it is missing.