devtools/eigen
1
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2026-04-10 11:34:33 +08:00
Code Issues Packages Projects Releases Wiki Activity
13,188 Commits 15 Branches 93 Tags
gpu-cg-interop
Commit Graph

4 Commits

Author SHA1 Message Date
Rasmus Munk Larsen
014f12f11a GPU: Add BLAS-1 ops, DeviceScalar, device-resident SpMV, and CG interop (5/5) 
Add the operator interface needed for GPU iterative solvers:

- BLAS Level-1 on DeviceMatrix: dot(), norm(), squaredNorm(), setZero(),
  noalias(), operator+=/-=/\*= dispatching to cuBLAS axpy/scal/dot/nrm2.
- DeviceScalar<Scalar>: device-resident scalar returned by reductions.
  Defers host sync until value is read (implicit conversion). Device-side
  division via NPP for real types.
- GpuContext: stream-borrowing constructor, setThreadLocal(), cublasLtHandle(),
  cusparseHandle().
- GEMM upgraded from cublasGemmEx to cublasLtMatmul with heuristic algorithm
  selection and plan caching.
- GpuSparseContext: GpuContext& constructor for same-stream execution,
  deviceView() returning DeviceSparseView with operator* for device-resident
  SpMV (d_y = d_A * d_x).
- geam expressions: d_C = d_A + alpha * d_B via cublasXgeam.
- GpuSVD::matrixV() convenience wrapper.

These additions make DeviceMatrix usable as a VectorType in Eigen algorithm
templates. Conjugate gradient is the motivating example and is tested against
CPU ConjugateGradient for correctness.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
 2026-04-09 20:19:59 -07:00
Rasmus Munk Larsen
43a95b62bb GPU: Add sparse solvers, FFT, and SpMV (cuDSS, cuFFT, cuSPARSE) 
Add GPU sparse direct solvers (Cholesky, LDL^T, LU) via cuDSS, 1D/2D FFT
via cuFFT with plan caching, and sparse matrix-vector/matrix multiply
(SpMV/SpMM) via cuSPARSE.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
 2026-04-09 19:11:49 -07:00
Rasmus Munk Larsen
8593c7f5a1 GPU: Add dense cuSOLVER solvers (QR, SVD, EigenSolver) 
Add QR (geqrf + ormqr + trsm), SVD (gesvd), and self-adjoint eigenvalue
decomposition (syevd) via cuSOLVER. All support host and DeviceMatrix input.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
 2026-04-09 19:11:34 -07:00
Rasmus Munk Larsen
58c44ef36d GPU: Add library dispatch module (DeviceMatrix, cuBLAS, cuSOLVER) 
Add Eigen/GPU module: A standalone GPU library dispatch layer where
DeviceMatrix<Scalar> operations map 1:1 to cuBLAS/cuSOLVER calls.
CPU and GPU solvers coexist in the same binary with compatible syntax.

Core infrastructure:
- DeviceMatrix<Scalar>: RAII dense column-major GPU memory wrapper with
  async host transfer (fromHost/toHost) and CUDA event-based cross-stream
  synchronization.
- GpuContext: Unified execution context owning a CUDA stream + cuBLAS
  handle + cuSOLVER handle. Thread-local default with explicit override
  via setThreadLocal(). Stream-borrowing constructor for integration.
- DeviceBuffer: Typed RAII device allocation with move semantics.

cuBLAS dispatch (expression syntax):
- GEMM: d_C = d_A.adjoint() * d_B (cublasXgemm)
- TRSM: d_X = d_A.triangularView<Lower>().solve(d_B) (cublasXtrsm)
- SYMM/HEMM: d_C = d_A.selfadjointView<Lower>() * d_B (cublasXsymm)
- SYRK/HERK: d_C = d_A * d_A.adjoint() (cublasXsyrk)

cuSOLVER dispatch:
- GpuLLT: Cached Cholesky factorization (cusolverDnXpotrf + Xpotrs)
- GpuLU: Cached LU factorization (cusolverDnXgetrf + Xgetrs)
- Solver chaining: auto x = d_A.llt().solve(d_B)
- Solver expressions with .device(ctx) for explicit stream control.

CI: Bump CUDA container to Ubuntu 22.04 (CMake 3.22), GCC 10->11,
Clang 12->14. Bump cmake_minimum_required to 3.17 for FindCUDAToolkit.

Tests: gpu_cublas.cpp, gpu_cusolver_llt.cpp, gpu_cusolver_lu.cpp,
gpu_device_matrix.cpp, gpu_library_example.cu
Benchmarks: bench_gpu_solvers.cpp, bench_gpu_chaining.cpp,
bench_gpu_batching.cpp
 2026-04-09 19:05:25 -07:00