User:Cantonios/3.4

New Major Features in Core

• New support for bfloat16

The 16-bit Brain floating point format[1] is now available as Eigen::bfloat16. The constructor must be called explicitly, but it can otherwise be used as any other scalar type. To convert back-and-forth between uint16_t to extract the bit representation, use Eigen::numext::bit_cast.

  bfloat16 s(0.25);                                 // explicit construction
  uint16_t s_bits = numext::bit_cast<uint16_t>(s);  // bit representation
 
  using MatrixBf16 = Matrix<bfloat16, Dynamic, Dynamic>;
  MatrixBf16 X = s * MatrixBf16::Random(3, 3);

New backends

• AMD ROCm HIP:
  • Unified with CUDA to create a generic GPU backend for NVIDIA/AMD.

Improvements/Cleanups to Core modules

• Dense matrix decompositions and solvers
  • SVD implementations now have an info() method for checking convergence.

  MatrixXf m = MatrixXf::Random(3,2);
  JacobiSVD<MatrixXf> svd(m, ComputeThinU | ComputeThinV);
  if (svd.info() == ComputationInfo::Success) {
    // SVD computation was successful.
    VectorXf x = svd.solve(b);
  }

• • Decompositions now fail quickly for detected invalid inputs.
  • Fixed aliasing issues with in-place small matrix inversions.
  • Fixed several edge-cases with empty or zero inputs.
• Sparse matrix support, decompositions and solvers
  • Enable assignment and addition with diagonal matrices.

  SparseMatrix<float> A(10, 10);
  VectorXf x = VectorXf::Random(10);
  A = x.asDiagonal();
  A += x.asDiagonal();

• • Added new IRDS iterative linear solver.

  A.makeCompressed();   // Recommendation is to compress input before calling sparse solvers.
  IDRS<SparseMatrix<float>, DiagonalPreconditioner<float> > idrs(A);
  if (idrs.info() == ComputationInfo::Success) {
    VectorXf x = idrs.solve(b);
  }

• • Support added for SuiteSparse KLU routines.

  A.makeCompressed();   // Recommendation is to compress input before calling sparse solvers.
  KLU<SparseMatrix<T> > klu(A);
  if (klu.info() == ComputationInfo::Success) {
    VectorXf x = klu.solve(b);
  }

• • SparseCholesky now works with row-major matrices.
  • Various bug fixes and performance improvements.

• Improved support for half
  • Native support for ARM __fp16, CUDA/HIP __half, Clang F16C.
  • Better vectorization support across backends.
• Improved bool support
  • Partial vectorization support for boolean operations.
  • Significantly improved performance (x25) for logical operations with Matrix or Tensor of bool.
• Improved support for custom types
  • More custom types work out-of-the-box (see #2201[2]).
• Improved Geometry Module
  • Transform::computeRotationScaling() and Transform::computeScalingRotation() are now more continuous across degeneracies (see !349[3]).
  • New minimal vectorization support.

Backend-specific improvements

• SSE/AVX/AVX512
  • Enable AVX512 instructions by default if available.
  • New std::complex, half, bfloat16 vectorization support.
  • Better accuracy for several vectorized math functions including exp, log, pow, sqrt.
  • Many missing packet functions added.
• GPU (CUDA and HIP)
  • Several optimized math functions added, better support for std::complex.
  • Option to disable CUDA entirely by defining EIGEN_NO_CUDA.
  • Many more functions can now be used in device code (e.g. comparisons, matrix inversion).
• SYCL
  • Redesigned SYCL implementation for use with the Tensor[4] module, which can be enabled by defining EIGEN_USE_SYCL.
  • New generic memory model used by TensorDeviceSycl.
  • Better integration with OpenCL devices.
  • Added many math function specializations.