Coming from nvmath-python

NVMath.Net follows the structure of nvmath-python: stateful objects with plan and execute steps, option objects, and thin stateless helpers. This page maps the concepts. It describes NVMath.Net as it is today; where nvmath-python has something this library does not, that is stated.

Concept mapping

nvmath-python	NVMath.Net
NumPy / CuPy / PyTorch tensors	DeviceBuffer<T> (explicit GPU memory); shapes are passed separately
obj.plan()	Plan() / PlanAsync()
obj.execute()	Execute() / ExecuteAsync()
obj.free() / context manager	Dispose() / using
options={…} dict or dataclass	MatmulOptions, FftOptions, TensorContractionOptions, DirectSolverOptions
stream= argument	optional CudaStream constructor / method argument
blocking="auto"	Blocking option (default: non-blocking, except where noted)
Row-major arrays (NumPy default)	Column-major buffers; tensors treat the first mode as fastest

Module mapping

nvmath-python	NVMath.Net	Notes
nvmath.fft.fft, ifft, rfft, irfft	Fft.FftAsync, IFftAsync, RFftAsync, IRFftAsync	One-shot helpers
nvmath.fft.FFT	Fft class	N-D and batched plans; FftOptions.Normalization replaces norm scaling. Prolog/epilog callbacks are not available yet
nvmath.linalg.advanced.matmul / Matmul	Matmul	Epilogs (MatmulEpilog), bias, aux outputs, batching, FP8 scale pointers, Autotune, plan cache, conjugate transpose and complex types. Block-scaled MXFP8/NVFP4 is not available yet
nvmath.linalg BLAS-style helpers	Gemm, TriangularSolve, SymmetricMultiply, DiagonalMultiply	Typed per precision (Sgemm, Dgemm, …)
nvmath.sparse.advanced.DirectSolver	SparseDirectSolver<T>	cuDSS; analyse / factorise / refactorise / solve
nvmath.sparse multiplication	SparseLinearAlgebra (SpMVAsync, SpMMAsync, SpGemmAsync, …)	CSR, COO, BSR, CSC
nvmath.tensor.contract / contraction objects	TensorContraction	cuTENSOR 2.x plans
nvmath.tensor element-wise ops	ElementwiseBinary, ElementwiseTrinary	Operator enum CuTensorOperator
nvmath.tensor permutation / reduction	TensorPermutation, TensorReduction
nvmath.bindings.*	NVMathNet.Interop.*Native classes	Raw bindings, public
nvmath.device.* (cuBLASDx, cuFFTDx, …)	—	Device-side APIs need a device compiler; not offered
nvmath.distributed.*	—	Multi-process / NVSHMEM support is not offered. MultiGpuContext covers single-process multi-GPU

Beyond nvmath-python

NVMath.Net also wraps libraries that nvmath-python exposes only through bindings or not at all: dense solvers (DenseSolver: LU, QR, least squares, Cholesky, SVD, eigenvalues, batched and complex variants) and random number generation (CudaRandom).

A side-by-side example

nvmath-python:

import cupy as cp, nvmath
a = cp.random.rand(128, 64, dtype=cp.float32)
b = cp.random.rand(64, 32, dtype=cp.float32)
d = nvmath.linalg.advanced.matmul(a, b)

NVMath.Net:

using NVMathNet;
using NVMathNet.LinAlg;
using NVMathNet.Rand;

const int M = 128, K = 64, N = 32;
using var a = new DeviceBuffer<float>(M * K);
using var b = new DeviceBuffer<float>(K * N);
using var c = new DeviceBuffer<float>(M * N);
using var d = new DeviceBuffer<float>(M * N);

using var rng = new CudaRandom(seed: 1);
rng.FillUniform(a);
rng.FillUniform(b);
c.Clear();

using var matmul = new Matmul(M, N, K, options: new MatmulOptions { Blocking = true });
await matmul.PlanAsync();
await matmul.ExecuteAsync<float>(a, b, c, d);