Austin Schuh | 3de38b0 | 2024-06-25 18:25:10 -0700 | [diff] [blame^] | 1 | // Ceres Solver - A fast non-linear least squares minimizer |
| 2 | // Copyright 2023 Google Inc. All rights reserved. |
| 3 | // http://ceres-solver.org/ |
| 4 | // |
| 5 | // Redistribution and use in source and binary forms, with or without |
| 6 | // modification, are permitted provided that the following conditions are met: |
| 7 | // |
| 8 | // * Redistributions of source code must retain the above copyright notice, |
| 9 | // this list of conditions and the following disclaimer. |
| 10 | // * Redistributions in binary form must reproduce the above copyright notice, |
| 11 | // this list of conditions and the following disclaimer in the documentation |
| 12 | // and/or other materials provided with the distribution. |
| 13 | // * Neither the name of Google Inc. nor the names of its contributors may be |
| 14 | // used to endorse or promote products derived from this software without |
| 15 | // specific prior written permission. |
| 16 | // |
| 17 | // THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" |
| 18 | // AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE |
| 19 | // IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE |
| 20 | // ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE |
| 21 | // LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR |
| 22 | // CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF |
| 23 | // SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS |
| 24 | // INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN |
| 25 | // CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) |
| 26 | // ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE |
| 27 | // POSSIBILITY OF SUCH DAMAGE. |
| 28 | // |
| 29 | // Author: joydeepb@cs.utexas.edu (Joydeep Biswas) |
| 30 | // |
| 31 | // A simple CUDA vector class. |
| 32 | |
| 33 | #ifndef CERES_INTERNAL_CUDA_VECTOR_H_ |
| 34 | #define CERES_INTERNAL_CUDA_VECTOR_H_ |
| 35 | |
| 36 | // This include must come before any #ifndef check on Ceres compile options. |
| 37 | // clang-format off |
| 38 | #include "ceres/internal/config.h" |
| 39 | // clang-format on |
| 40 | |
| 41 | #include <math.h> |
| 42 | |
| 43 | #include <memory> |
| 44 | #include <string> |
| 45 | |
| 46 | #include "ceres/context_impl.h" |
| 47 | #include "ceres/internal/export.h" |
| 48 | #include "ceres/types.h" |
| 49 | |
| 50 | #ifndef CERES_NO_CUDA |
| 51 | |
| 52 | #include "ceres/cuda_buffer.h" |
| 53 | #include "ceres/cuda_kernels_vector_ops.h" |
| 54 | #include "ceres/internal/eigen.h" |
| 55 | #include "cublas_v2.h" |
| 56 | #include "cusparse.h" |
| 57 | |
| 58 | namespace ceres::internal { |
| 59 | |
| 60 | // An Nx1 vector, denoted y hosted on the GPU, with CUDA-accelerated operations. |
| 61 | class CERES_NO_EXPORT CudaVector { |
| 62 | public: |
| 63 | // Create a pre-allocated vector of size N and return a pointer to it. The |
| 64 | // caller must ensure that InitCuda() has already been successfully called on |
| 65 | // context before calling this method. |
| 66 | CudaVector(ContextImpl* context, int size); |
| 67 | |
| 68 | CudaVector(CudaVector&& other); |
| 69 | |
| 70 | ~CudaVector(); |
| 71 | |
| 72 | void Resize(int size); |
| 73 | |
| 74 | // Perform a deep copy of the vector. |
| 75 | CudaVector& operator=(const CudaVector&); |
| 76 | |
| 77 | // Return the inner product x' * y. |
| 78 | double Dot(const CudaVector& x) const; |
| 79 | |
| 80 | // Return the L2 norm of the vector (||y||_2). |
| 81 | double Norm() const; |
| 82 | |
| 83 | // Set all elements to zero. |
| 84 | void SetZero(); |
| 85 | |
| 86 | // Copy from Eigen vector. |
| 87 | void CopyFromCpu(const Vector& x); |
| 88 | |
| 89 | // Copy from CPU memory array. |
| 90 | void CopyFromCpu(const double* x); |
| 91 | |
| 92 | // Copy to Eigen vector. |
| 93 | void CopyTo(Vector* x) const; |
| 94 | |
| 95 | // Copy to CPU memory array. It is the caller's responsibility to ensure |
| 96 | // that the array is large enough. |
| 97 | void CopyTo(double* x) const; |
| 98 | |
| 99 | // y = a * x + b * y. |
| 100 | void Axpby(double a, const CudaVector& x, double b); |
| 101 | |
| 102 | // y = diag(d)' * diag(d) * x + y. |
| 103 | void DtDxpy(const CudaVector& D, const CudaVector& x); |
| 104 | |
| 105 | // y = s * y. |
| 106 | void Scale(double s); |
| 107 | |
| 108 | int num_rows() const { return num_rows_; } |
| 109 | int num_cols() const { return 1; } |
| 110 | |
| 111 | const double* data() const { return data_.data(); } |
| 112 | double* mutable_data() { return data_.data(); } |
| 113 | |
| 114 | const cusparseDnVecDescr_t& descr() const { return descr_; } |
| 115 | |
| 116 | private: |
| 117 | CudaVector(const CudaVector&) = delete; |
| 118 | void DestroyDescriptor(); |
| 119 | |
| 120 | int num_rows_ = 0; |
| 121 | ContextImpl* context_ = nullptr; |
| 122 | CudaBuffer<double> data_; |
| 123 | // CuSparse object that describes this dense vector. |
| 124 | cusparseDnVecDescr_t descr_ = nullptr; |
| 125 | }; |
| 126 | |
| 127 | // Blas1 operations on Cuda vectors. These functions are needed as an |
| 128 | // abstraction layer so that we can use different versions of a vector style |
| 129 | // object in the conjugate gradients linear solver. |
| 130 | // Context and num_threads arguments are not used by CUDA implementation, |
| 131 | // context embedded into CudaVector is used instead. |
| 132 | inline double Norm(const CudaVector& x, |
| 133 | ContextImpl* context = nullptr, |
| 134 | int num_threads = 1) { |
| 135 | (void)context; |
| 136 | (void)num_threads; |
| 137 | return x.Norm(); |
| 138 | } |
| 139 | inline void SetZero(CudaVector& x, |
| 140 | ContextImpl* context = nullptr, |
| 141 | int num_threads = 1) { |
| 142 | (void)context; |
| 143 | (void)num_threads; |
| 144 | x.SetZero(); |
| 145 | } |
| 146 | inline void Axpby(double a, |
| 147 | const CudaVector& x, |
| 148 | double b, |
| 149 | const CudaVector& y, |
| 150 | CudaVector& z, |
| 151 | ContextImpl* context = nullptr, |
| 152 | int num_threads = 1) { |
| 153 | (void)context; |
| 154 | (void)num_threads; |
| 155 | if (&x == &y && &y == &z) { |
| 156 | // z = (a + b) * z; |
| 157 | z.Scale(a + b); |
| 158 | } else if (&x == &z) { |
| 159 | // x is aliased to z. |
| 160 | // z = x |
| 161 | // = b * y + a * x; |
| 162 | z.Axpby(b, y, a); |
| 163 | } else if (&y == &z) { |
| 164 | // y is aliased to z. |
| 165 | // z = y = a * x + b * y; |
| 166 | z.Axpby(a, x, b); |
| 167 | } else { |
| 168 | // General case: all inputs and outputs are distinct. |
| 169 | z = y; |
| 170 | z.Axpby(a, x, b); |
| 171 | } |
| 172 | } |
| 173 | inline double Dot(const CudaVector& x, |
| 174 | const CudaVector& y, |
| 175 | ContextImpl* context = nullptr, |
| 176 | int num_threads = 1) { |
| 177 | (void)context; |
| 178 | (void)num_threads; |
| 179 | return x.Dot(y); |
| 180 | } |
| 181 | inline void Copy(const CudaVector& from, |
| 182 | CudaVector& to, |
| 183 | ContextImpl* context = nullptr, |
| 184 | int num_threads = 1) { |
| 185 | (void)context; |
| 186 | (void)num_threads; |
| 187 | to = from; |
| 188 | } |
| 189 | |
| 190 | } // namespace ceres::internal |
| 191 | |
| 192 | #endif // CERES_NO_CUDA |
| 193 | #endif // CERES_INTERNAL_CUDA_SPARSE_LINEAR_OPERATOR_H_ |