LeastSquaresCUDA.cpp

Go to the documentation of this file.

// ----------------------------------------------------------------------------
// -                        CloudViewer: www.cloudViewer.org                  -
// ----------------------------------------------------------------------------
// Copyright (c) 2018-2024 www.cloudViewer.org
// SPDX-License-Identifier: MIT
// ----------------------------------------------------------------------------
 
#include "cloudViewer/core/Blob.h"
#include "cloudViewer/core/CUDAUtils.h"
#include "cloudViewer/core/linalg/BlasWrapper.h"
#include "cloudViewer/core/linalg/LapackWrapper.h"
#include "cloudViewer/core/linalg/LeastSquares.h"
#include "cloudViewer/core/linalg/LinalgUtils.h"
 
namespace cloudViewer {
namespace core {
 
// cusolverDn<t1><t2>gels() is not supported until CUDA 11.0.
// We have to implement for earlier versions via
// Step 1: A = Q*R by geqrf.
// Step 2: B : = Q ^ T* B by ormqr.
// Step 3: solve R* X = B by trsm.
// Ref: https://docs.nvidia.com/cuda/cusolver/index.html#ormqr-example1
void LeastSquaresCUDA(void* A_data,
                      void* B_data,
                      int64_t m,
                      int64_t n,
                      int64_t k,
                      Dtype dtype,
                      const Device& device) {
    cusolverDnHandle_t cusolver_handle =
            CuSolverContext::GetInstance().GetHandle(device);
    cublasHandle_t cublas_handle =
            CuBLASContext::GetInstance().GetHandle(device);
 
    DISPATCH_LINALG_DTYPE_TO_TEMPLATE(dtype, [&]() {
        int len_geqrf, len_ormqr, len;
        Blob dinfo(sizeof(int), device);
 
        OPEN3D_CUSOLVER_CHECK(geqrf_cuda_buffersize<scalar_t>(
                                      cusolver_handle, m, n, m, &len_geqrf),
                              "geqrf_buffersize failed in LeastSquaresCUDA");
        OPEN3D_CUSOLVER_CHECK(ormqr_cuda_buffersize<scalar_t>(
                                      cusolver_handle, CUBLAS_SIDE_LEFT,
                                      CUBLAS_OP_T, m, k, n, m, m, &len_ormqr),
                              "ormqr_buffersize failed in LeastSquaresCUDA");
        len = std::max(len_geqrf, len_ormqr);
 
        Blob workspace(len * sizeof(scalar_t), device);
        Blob tau(n * sizeof(scalar_t), device);
 
        // Step 1: A = QR
        OPEN3D_CUSOLVER_CHECK_WITH_DINFO(
                geqrf_cuda<scalar_t>(
                        cusolver_handle, m, n, static_cast<scalar_t*>(A_data),
                        m, static_cast<scalar_t*>(tau.GetDataPtr()),
                        static_cast<scalar_t*>(workspace.GetDataPtr()), len,
                        static_cast<int*>(dinfo.GetDataPtr())),
                "geqrf failed in LeastSquaresCUDA",
                static_cast<int*>(dinfo.GetDataPtr()), device);
 
        // Step 2: B' = Q^T*B
        OPEN3D_CUSOLVER_CHECK_WITH_DINFO(
                ormqr_cuda<scalar_t>(
                        cusolver_handle, CUBLAS_SIDE_LEFT, CUBLAS_OP_T, m, k, n,
                        static_cast<scalar_t*>(A_data), m,
                        static_cast<scalar_t*>(tau.GetDataPtr()),
                        static_cast<scalar_t*>(B_data), m,
                        static_cast<scalar_t*>(workspace.GetDataPtr()), len,
                        static_cast<int*>(dinfo.GetDataPtr())),
                "ormqr failed in LeastSquaresCUDA",
                static_cast<int*>(dinfo.GetDataPtr()), device);
 
        // Step 3: Solve Rx = B'
        scalar_t alpha = 1.0f;
        OPEN3D_CUBLAS_CHECK(
                trsm_cuda<scalar_t>(cublas_handle, CUBLAS_SIDE_LEFT,
                                    CUBLAS_FILL_MODE_UPPER, CUBLAS_OP_N,
                                    CUBLAS_DIAG_NON_UNIT, n, k, &alpha,
                                    static_cast<scalar_t*>(A_data), m,
                                    static_cast<scalar_t*>(B_data), m),
                "trsm failed in LeastSquaresCUDA");
    });
}
 
}  // namespace core
}  // namespace cloudViewer