CUDA CGNR, Part 4: CudaCgnrSolver

* Added CudaCgnrSolver, a new CUDA-accelerated CGNR. * To use CudaCgnrSolver, the user must select CGNR as the linear_solver and CUDA_SPARSE as the sparse_linear_algebra_library. * Updated ConjugateGradientSolver to work with an array of pointers to scratch to support CudaVectors as scratch. * Moved CUDA initialization to run in Solver::Solve as needed. Some performance comparisons on an Ubuntu 20.04 desktop with an Intel i9-9940X CPU @ 3.30GHz, and an nVidia Quadro RTX 6000, all configurations run with 24 threads, and 10 iterations. ================================================= CGNR + CUDA_SPARSE + IDENTITY Preconditioner problem-1778-993923-pre.txt ================================================= Cost: Initial 2.563973e+08 Final 1.724755e+06 Change 2.546725e+08 Minimizer iterations 11 Successful steps 7 Unsuccessful steps 4 Time (in seconds): Preprocessor 4.020158 Residual only evaluation 1.567092 (10) Jacobian & residual evaluation 7.847130 (7) Linear solver 31.688898 (10) Minimizer 46.834987 Postprocessor 0.353974 Total 51.209120 ================================================= SPARSE_SCHUR (CPU) + SUITE_SPARSE + AMD problem-1778-993923-pre.txt ================================================= Cost: Initial 2.563973e+08 Final 1.651617e+06 Change 2.547457e+08 Minimizer iterations 11 Successful steps 11 Unsuccessful steps 0 Time (in seconds): Preprocessor 35.812003 Residual only evaluation 1.658980 (10) Jacobian & residual evaluation 12.218799 (11) Linear solver 76.409992 (10) Minimizer 98.809773 Postprocessor 0.372712 Total 134.994489 ================================================= ITERATIVE_SCHUR (CPU) + JACOBI Preconditioner problem-1778-993923-pre.txt ================================================= Cost: Initial 2.563973e+08 Final 1.684447e+06 Change 2.547128e+08 Minimizer iterations 11 Successful steps 8 Unsuccessful steps 3 Time (in seconds): Preprocessor 15.331614 Residual only evaluation 1.606114 (10) Jacobian & residual evaluation 8.502166 (8) Linear solver 351.910080 (10) Minimizer 368.797327 Postprocessor 0.363536 Total 384.492478 ================================================= CGNR + CUDA_SPARSE + IDENTITY Preconditioner problem-13682-4456117-pre.txt ================================================= Cost: Initial 1.126372e+09 Final 2.269329e+07 Change 1.103678e+09 Minimizer iterations 11 Successful steps 7 Unsuccessful steps 4 Time (in seconds): Preprocessor 19.140087 Residual only evaluation 8.721920 (10) Jacobian & residual evaluation 41.955923 (7) Linear solver 214.121861 (10) Minimizer 296.636890 Postprocessor 1.971827 Total 317.748804 Change-Id: I3a09f31aa6903f661e91f595afd39d427583e856
2026-08-29 08:34:37 +08:00 · 2022-08-14 19:14:04 -05:00
parent 6ab435d774
commit 829089053e
21 changed files with 413 additions and 59 deletions
@@ -327,7 +327,8 @@ bool OptionsAreValidForCgnr(const Solver::Options& options, string* error) {
    return false;
  }

-  if (options.preconditioner_type == SUBSET) {
+  if (options.sparse_linear_algebra_library_type != CUDA_SPARSE &&
+      options.preconditioner_type == SUBSET) {
    if (options.residual_blocks_for_subset_preconditioner.empty()) {
      *error =
          "When using SUBSET preconditioner, "
@@ -341,6 +342,20 @@ bool OptionsAreValidForCgnr(const Solver::Options& options, string* error) {
    }
  }

+  // Check options for CGNR with CUDA_SPARSE.
+  if (options.sparse_linear_algebra_library_type == CUDA_SPARSE) {
+    if (!IsSparseLinearAlgebraLibraryTypeAvailable(CUDA_SPARSE)) {
+      *error = "Can't use CGNR with sparse_linear_algebra_library_type = "
+          "CUDA_SPARSE because support was not enabled when Ceres was built.";
+      return false;
+    }
+    if (options.preconditioner_type != IDENTITY) {
+      StringPrintf("Can't use CGNR with preconditioner_type = %s when "
+                   "sparse_linear_algebra_library_type = CUDA_SPARSE.",
+                    PreconditionerTypeToString(options.preconditioner_type));
+      return false;
+    }
+  }
  return true;
 }

@@ -652,6 +667,18 @@ std::string SchurStructureToString(const int row_block_size,
  return internal::StringPrintf("%s,%s,%s", row.c_str(), e.c_str(), f.c_str());
 }

+bool IsCudaRequired(const Solver::Options& options) {
+  if (options.linear_solver_type == DENSE_NORMAL_CHOLESKY ||
+      options.linear_solver_type == DENSE_SCHUR ||
+      options.linear_solver_type == DENSE_QR) {
+    return (options.dense_linear_algebra_library_type == CUDA);
+  }
+  if (options.linear_solver_type == CGNR) {
+    return (options.sparse_linear_algebra_library_type == CUDA_SPARSE);
+  }
+  return false;
+}
+
 }  // namespace

 bool Solver::Options::IsValid(string* error) const {
@@ -697,6 +724,13 @@ void Solver::Solve(const Solver::Options& options,
  Program* program = problem_impl->mutable_program();
  PreSolveSummarize(options, problem_impl, summary);

+  if (IsCudaRequired(options)) {
+    if (!problem_impl->context()->InitCUDA(&summary->message)) {
+      LOG(ERROR) << "Terminating: " << summary->message;
+      return;
+    }
+  }
+
  // If gradient_checking is enabled, wrap all cost functions in a
  // gradient checker and install a callback that terminates if any gradient
  // error is detected.
@@ -866,22 +900,40 @@ string Solver::Summary::FullReport() const {
      }
    }

-    if (linear_solver_type_used == SPARSE_NORMAL_CHOLESKY ||
+    const bool used_sparse_linear_algebra_library =
+        linear_solver_type_used == SPARSE_NORMAL_CHOLESKY ||
        linear_solver_type_used == SPARSE_SCHUR ||
-        (linear_solver_type_used == CGNR &&
-         preconditioner_type_used == SUBSET) ||
+        linear_solver_type_used == CGNR ||
        (linear_solver_type_used == ITERATIVE_SCHUR &&
-         (preconditioner_type_used == CLUSTER_JACOBI ||
-          preconditioner_type_used == CLUSTER_TRIDIAGONAL))) {
+            (preconditioner_type_used == CLUSTER_JACOBI ||
+            preconditioner_type_used == CLUSTER_TRIDIAGONAL));
+
+    const bool linear_solver_ordering_required =
+        linear_solver_type_used == SPARSE_SCHUR ||
+        (linear_solver_type_used == ITERATIVE_SCHUR &&
+            (preconditioner_type_used == CLUSTER_JACOBI ||
+            preconditioner_type_used == CLUSTER_TRIDIAGONAL)) ||
+        (linear_solver_type_used == CGNR && preconditioner_type_used == SUBSET);
+
+    if (used_sparse_linear_algebra_library) {
      const char* mixed_precision_suffix =
          (mixed_precision_solves_used ? "(Mixed Precision)" : "");
-      StringAppendF(
-          &report,
-          "\nSparse linear algebra library %15s + %s %s\n",
-          SparseLinearAlgebraLibraryTypeToString(
-              sparse_linear_algebra_library_type),
-          LinearSolverOrderingTypeToString(linear_solver_ordering_type),
-          mixed_precision_suffix);
+      if (linear_solver_ordering_required) {
+        StringAppendF(
+            &report,
+            "\nSparse linear algebra library %15s + %s %s\n",
+            SparseLinearAlgebraLibraryTypeToString(
+                sparse_linear_algebra_library_type),
+            LinearSolverOrderingTypeToString(linear_solver_ordering_type),
+            mixed_precision_suffix);
+      } else {
+        StringAppendF(
+            &report,
+            "\nSparse linear algebra library %15s %s\n",
+            SparseLinearAlgebraLibraryTypeToString(
+                sparse_linear_algebra_library_type),
+            mixed_precision_suffix);
+      }
    }

    StringAppendF(&report, "\n");