Skip to content

Fix segfault #19

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
alexkorovko merged 2 commits into from
Jul 13, 2026
Merged

Fix segfault #19

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
d4dd454
Fix segfault
alexkorovko Jul 13, 2026
9c750b3
Fix formatting
alexkorovko Jul 13, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
34 changes: 16 additions & 18 deletions cunls/minimizer/minimizer_state.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -41,10 +41,9 @@ constexpr size_t block_size = 256;
* @param num_states_in_batch Number of state elements in the batch.
* @param num_new_pointers Number of pointers to remap.
*/
__global__ void
set_state_pointers_kernel(float **new_pointers, float *const *old_pointers,
float *old_start_ptr, float *new_start_ptr,
size_t num_states_in_batch, size_t num_new_pointers) {
__global__ void set_state_pointers_kernel(float **new_pointers, float *const *old_pointers,
float *old_start_ptr, float *new_start_ptr,
size_t num_states_in_batch, size_t num_new_pointers) {
int tid = threadIdx.x + blockIdx.x * blockDim.x;
if (tid >= num_new_pointers) {
return;
Expand All @@ -69,16 +68,18 @@ set_state_pointers_kernel(float **new_pointers, float *const *old_pointers,
*/
void MinimizerState::CreateStates(const Problem &problem) {
const auto &state_batches = problem.GetStateBatches();
if (states_.size() < state_batches.size()) {
// Resize to the exact batch count so stale trailing batches from a previous,
// larger problem are dropped when the topology shrinks. Surviving batches keep
// their device buffers; only whole unused batch buffers are freed.
if (states_.size() != state_batches.size()) {
states_.resize(state_batches.size());
}

for (size_t i = 0; i < state_batches.size(); i++) {
const auto &param_batch_ptr = state_batches[i];
auto &state_vec = states_[i];

size_t size =
param_batch_ptr->NumStateBlocks() * param_batch_ptr->AmbientSize();
size_t size = param_batch_ptr->NumStateBlocks() * param_batch_ptr->AmbientSize();

if (state_vec.size() != size) {
state_vec.resize(size);
Expand All @@ -97,7 +98,7 @@ void MinimizerState::CreateStates(const Problem &problem) {
*/
void MinimizerState::CreateStatePointers(const Problem &problem) {
const auto &problem_param_pointers = problem.GetStatePointers();
if (state_pointers_.size() < problem_param_pointers.size()) {
if (state_pointers_.size() != problem_param_pointers.size()) {
state_pointers_.resize(problem_param_pointers.size());
}

Expand All @@ -111,14 +112,13 @@ void MinimizerState::CreateStatePointers(const Problem &problem) {

void MinimizerState::CopyProblemStatePointersFromHost(const Problem &problem) {
const auto &host = problem.GetStatePointers();
if (problem_state_ptrs_device_.size() < host.size()) {
if (problem_state_ptrs_device_.size() != host.size()) {
problem_state_ptrs_device_.resize(host.size());
}
for (size_t i = 0; i < host.size(); ++i) {
problem_state_ptrs_device_[i].resize(host[i].size());
if (!host[i].empty()) {
problem_state_ptrs_device_[i].CopyFromHost(host[i].data(),
host[i].size());
problem_state_ptrs_device_[i].CopyFromHost(host[i].data(), host[i].size());
}
}
}
Expand Down Expand Up @@ -149,8 +149,7 @@ void MinimizerState::Create(cudaStream_t stream, const Problem &problem) {
auto &state_vec = states_[i];

float *ptr = param_batch_ptr->StateBlockDevicePtr(0);
size_t size =
param_batch_ptr->NumStateBlocks() * param_batch_ptr->AmbientSize();
size_t size = param_batch_ptr->NumStateBlocks() * param_batch_ptr->AmbientSize();

thrust::device_ptr<float> src_ptr(ptr);
thrust::device_ptr<float> dst_ptr(state_vec.data());
Expand Down Expand Up @@ -187,8 +186,8 @@ void MinimizerState::Create(cudaStream_t stream, const Problem &problem) {
size_t num_blocks = (param_ptrs.size() + block_size - 1) / block_size;

set_state_pointers_kernel<<<num_blocks, block_size, 0, stream>>>(
new_pointers, old_pointers, state_batch_ptr, new_param_ptr,
num_states_in_batch, param_ptrs.size());
new_pointers, old_pointers, state_batch_ptr, new_param_ptr, num_states_in_batch,
param_ptrs.size());
THROW_ON_CUDA_ERROR(cudaGetLastError());
}
}
Expand All @@ -204,8 +203,7 @@ void MinimizerState::Create(cudaStream_t stream, const Problem &problem) {
* @param stream CUDA stream for GPU operations.
* @param from Source state vectors to copy from.
*/
void MinimizerState::Copy(cudaStream_t stream,
const std::vector<dvector<float>> &from) {
void MinimizerState::Copy(cudaStream_t stream, const std::vector<dvector<float>> &from) {
if (states_.size() != from.size()) {
states_.resize(from.size());
}
Expand Down Expand Up @@ -251,4 +249,4 @@ void Copy(cudaStream_t stream, const MinimizerState &state, Problem &problem) {
thrust::copy(stream_policy, src_ptr, src_ptr + dvec.size(), dst_ptr);
}
}
} // namespace cunls
} // namespace cunls
131 changes: 88 additions & 43 deletions tests/gauss_newton_test.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -50,7 +50,7 @@ namespace cunls {
*/
template <class TestParam>
class GaussNewtonMinimizerTest : public ::testing::Test {
public:
public:
static constexpr int kDim = TestParam::vector_size;
using StatesType = VectorStateBatch<kDim>;
using VectorType = Vector<kDim>;
Expand All @@ -76,16 +76,14 @@ class GaussNewtonMinimizerTest : public ::testing::Test {
state_values_[i].fill(x);
}

minimizer_options_.sparse_linear_solver_type =
SparseLinearSolverType::cuDSS;
minimizer_options_.sparse_linear_solver_type = SparseLinearSolverType::cuDSS;

cuDSSLinearSolverOptions cudss_solver_options = {
.mode = static_cast<cuDSSLinearSolverMode>(TestParam::solver_id),
.nthreads = 1,
.threading_lib_path = "",
};
minimizer_options_.sparse_linear_solver_config = {.cudss_solver_options =
cudss_solver_options};
minimizer_options_.sparse_linear_solver_config = {.cudss_solver_options = cudss_solver_options};
}

/**
Expand All @@ -97,23 +95,19 @@ class GaussNewtonMinimizerTest : public ::testing::Test {
*
* @param states Optimized state batch to verify.
*/
void CheckConvergence(const StatesType &states,
const std::vector<int> &const_state_ids = {}) {
void CheckConvergence(const StatesType &states, const std::vector<int> &const_state_ids = {}) {
size_t num_blocks = states.NumStateBlocks();
auto ptr =
reinterpret_cast<const VectorType *>(states.StateBlockDevicePtr(0));
auto ptr = reinterpret_cast<const VectorType *>(states.StateBlockDevicePtr(0));

std::vector<VectorType> host_states(num_blocks);
THROW_ON_CUDA_ERROR(cudaMemcpy(host_states.data(), ptr,
num_blocks * sizeof(VectorType),
THROW_ON_CUDA_ERROR(cudaMemcpy(host_states.data(), ptr, num_blocks * sizeof(VectorType),
cudaMemcpyDeviceToHost));

ASSERT_EQ(host_states.size(), observations_.size());
for (size_t i = 0; i < num_blocks; i++) {
const auto &obs = observations_[i];
const auto &state_vals = host_states[i];
auto const_state_it =
std::find(const_state_ids.begin(), const_state_ids.end(), i);
auto const_state_it = std::find(const_state_ids.begin(), const_state_ids.end(), i);
if (const_state_it != const_state_ids.end()) {
// State is constant, verify it wasn't changed during optimization
float x = static_cast<float>(i);
Expand All @@ -128,31 +122,30 @@ class GaussNewtonMinimizerTest : public ::testing::Test {
}
}

const size_t num_vectors_ = 10000; ///< Number of state blocks in test.
const size_t num_vectors_ = 10000; ///< Number of state blocks in test.

std::vector<VectorType> observations_; ///< Target values for optimization.
std::vector<VectorType> state_values_; ///< Initial state values.
std::vector<VectorType> observations_; ///< Target values for optimization.
std::vector<VectorType> state_values_; ///< Initial state values.

MinimizerOptions minimizer_options_{.disable_safety_checks = false};

profiler::Domain profiler_domain_{
"GaussNewtonMinimizerTest"}; ///< Profiling domain.
profiler::Domain profiler_domain_{"GaussNewtonMinimizerTest"}; ///< Profiling domain.
};

/**
* @brief Helper struct for parameterized test dimensions.
*
* @tparam Value Vector dimension (1, 2, 3, or 4).
*/
template <int VectorSize, int SolverId> struct TestParam {
template <int VectorSize, int SolverId>
struct TestParam {
static constexpr int vector_size = VectorSize;
static constexpr int solver_id = SolverId;
};

/** @brief Test types: 1D, 2D, 3D, and 4D vectors. */
typedef ::testing::Types<TestParam<1, 0>, TestParam<2, 0>, TestParam<3, 0>,
TestParam<4, 0>, TestParam<1, 1>, TestParam<2, 1>,
TestParam<3, 1>, TestParam<4, 1>>
typedef ::testing::Types<TestParam<1, 0>, TestParam<2, 0>, TestParam<3, 0>, TestParam<4, 0>,
TestParam<1, 1>, TestParam<2, 1>, TestParam<3, 1>, TestParam<4, 1>>
TestParams;
TYPED_TEST_CASE(GaussNewtonMinimizerTest, TestParams);

Expand Down Expand Up @@ -193,11 +186,9 @@ TYPED_TEST(GaussNewtonMinimizerTest, SimpleGN) {
* states.
*/
TYPED_TEST(GaussNewtonMinimizerTest, GNWithConstantStates) {
auto test_range =
this->profiler_domain_.CreateDomainRange("GNWithConstantStates");
auto test_range = this->profiler_domain_.CreateDomainRange("GNWithConstantStates");
std::vector<int> const_state_ids = {0, 9, 99, 999};
typename TestFixture::StateData state_data(this->state_values_,
const_state_ids);
typename TestFixture::StateData state_data(this->state_values_, const_state_ids);
auto &vector_states = state_data.get();
auto device_pointers = test_utils::CollectStatePointers(vector_states);
typename TestFixture::FactorData factor_data(this->observations_);
Expand Down Expand Up @@ -253,11 +244,9 @@ TYPED_TEST(GaussNewtonMinimizerTest, SimpleLM) {
* during optimization while still optimizing other states.
*/
TYPED_TEST(GaussNewtonMinimizerTest, LMWithConstantStates) {
auto test_range =
this->profiler_domain_.CreateDomainRange("LMWithConstantStates");
auto test_range = this->profiler_domain_.CreateDomainRange("LMWithConstantStates");
std::vector<int> const_state_ids = {0, 9, 99, 999};
typename TestFixture::StateData state_data(this->state_values_,
const_state_ids);
typename TestFixture::StateData state_data(this->state_values_, const_state_ids);
auto &vector_states = state_data.get();
auto device_pointers = test_utils::CollectStatePointers(vector_states);
typename TestFixture::FactorData factor_data(this->observations_);
Expand All @@ -282,8 +271,7 @@ TYPED_TEST(GaussNewtonMinimizerTest, LMWithConstantStates) {
* @brief Levenberg-Marquardt with Hessian-diagonal column scaling.
*/
TYPED_TEST(GaussNewtonMinimizerTest, LMColumnScalingHessianDiagonal) {
auto test_range = this->profiler_domain_.CreateDomainRange(
"LMColumnScalingHessianDiagonal");
auto test_range = this->profiler_domain_.CreateDomainRange("LMColumnScalingHessianDiagonal");
typename TestFixture::StateData state_data(this->state_values_);
auto &vector_states = state_data.get();
auto device_pointers = test_utils::CollectStatePointers(vector_states);
Expand All @@ -309,8 +297,7 @@ TYPED_TEST(GaussNewtonMinimizerTest, LMColumnScalingHessianDiagonal) {
* @brief Levenberg-Marquardt with Jacobian column-norm scaling.
*/
TYPED_TEST(GaussNewtonMinimizerTest, LMColumnScalingJacobianColumnNorm) {
auto test_range = this->profiler_domain_.CreateDomainRange(
"LMColumnScalingJacobianColumnNorm");
auto test_range = this->profiler_domain_.CreateDomainRange("LMColumnScalingJacobianColumnNorm");
typename TestFixture::StateData state_data(this->state_values_);
auto &vector_states = state_data.get();
auto device_pointers = test_utils::CollectStatePointers(vector_states);
Expand All @@ -336,8 +323,7 @@ TYPED_TEST(GaussNewtonMinimizerTest, LMColumnScalingJacobianColumnNorm) {
* @brief Gauss-Newton with column scaling (shared MinimizerOptions path).
*/
TYPED_TEST(GaussNewtonMinimizerTest, GNColumnScalingHessianDiagonal) {
auto test_range = this->profiler_domain_.CreateDomainRange(
"GNColumnScalingHessianDiagonal");
auto test_range = this->profiler_domain_.CreateDomainRange("GNColumnScalingHessianDiagonal");
typename TestFixture::StateData state_data(this->state_values_);
auto &vector_states = state_data.get();
auto device_pointers = test_utils::CollectStatePointers(vector_states);
Expand Down Expand Up @@ -389,8 +375,7 @@ TEST(MinimizeBufferReuse, GaussNewtonTwiceIdenticalSummaries) {
.nthreads = 1,
.threading_lib_path = "",
};
opts.sparse_linear_solver_config = {.cudss_solver_options =
cudss_solver_options};
opts.sparse_linear_solver_config = {.cudss_solver_options = cudss_solver_options};
opts.disable_safety_checks = false;

CudaStream stream;
Expand All @@ -399,15 +384,13 @@ TEST(MinimizeBufferReuse, GaussNewtonTwiceIdenticalSummaries) {
float *state_base = vector_states.StateBlockDevicePtr(0);
const size_t num_floats = n * 1;
std::vector<float> initial_host(num_floats);
THROW_ON_CUDA_ERROR(cudaMemcpy(initial_host.data(), state_base,
num_floats * sizeof(float),
THROW_ON_CUDA_ERROR(cudaMemcpy(initial_host.data(), state_base, num_floats * sizeof(float),
cudaMemcpyDeviceToHost));

MinimizerSummary s1 = minimizer.Minimize(stream.GetStream(), problem);
THROW_ON_CUDA_ERROR(cudaStreamSynchronize(stream.GetStream()));

THROW_ON_CUDA_ERROR(cudaMemcpy(state_base, initial_host.data(),
num_floats * sizeof(float),
THROW_ON_CUDA_ERROR(cudaMemcpy(state_base, initial_host.data(), num_floats * sizeof(float),
cudaMemcpyHostToDevice));

MinimizerSummary s2 = minimizer.Minimize(stream.GetStream(), problem);
Expand All @@ -418,4 +401,66 @@ TEST(MinimizeBufferReuse, GaussNewtonTwiceIdenticalSummaries) {
EXPECT_NEAR(s1.initial_cost, s2.initial_cost, 1e-4f);
}

} // namespace cunls
/**
* @brief Reusing a minimizer after the number of state batches decreases must
* not retain stale state vectors from the previous problem.
*
* Regression test for a two-batch to one-batch topology transition. Before the
* fix, MinimizerState::CreateStates only grew its outer vector. UpdateStates
* consequently passed two state pointers to StateBatchOps configured for one
* batch, causing an assertion failure in debug builds and an out-of-bounds
* access in release builds.
*/
TEST(MinimizeBufferReuse, GaussNewtonStateBatchCountDecreases) {
constexpr size_t n = 32;
const auto observations = test_utils::MakeZeroVectors<1>(n);
const auto initial_states = test_utils::MakeConstantVectors<1>(n, 1.0f);

MinimizerOptions opts;
opts.sparse_linear_solver_type = SparseLinearSolverType::cuDSS;
cuDSSLinearSolverOptions cudss_solver_options = {
.mode = cuDSSLinearSolverMode::SlowInitFastSolve,
.nthreads = 1,
.threading_lib_path = "",
};
opts.sparse_linear_solver_config = {.cudss_solver_options = cudss_solver_options};
opts.disable_safety_checks = false;

CudaStream stream;
GaussNewtonMinimizer minimizer(opts);

{
test_utils::VectorStateData<1> first_state_data(initial_states);
test_utils::VectorStateData<1> second_state_data(initial_states);
test_utils::PriorFactorData<1> first_factor_data(observations);
test_utils::PriorFactorData<1> second_factor_data(observations);

Problem two_batch_problem;
two_batch_problem.AddStateBatch(first_state_data.ptr());
two_batch_problem.AddStateBatch(second_state_data.ptr());
two_batch_problem.AddFactorBatch(&first_factor_data.get(),
test_utils::CollectStatePointers(first_state_data.get()));
two_batch_problem.AddFactorBatch(&second_factor_data.get(),
test_utils::CollectStatePointers(second_state_data.get()));
ASSERT_TRUE(two_batch_problem.CheckConsistency());

minimizer.Minimize(stream.GetStream(), two_batch_problem);
THROW_ON_CUDA_ERROR(cudaStreamSynchronize(stream.GetStream()));
}

{
test_utils::VectorStateData<1> state_data(initial_states);
test_utils::PriorFactorData<1> factor_data(observations);

Problem one_batch_problem;
one_batch_problem.AddStateBatch(state_data.ptr());
one_batch_problem.AddFactorBatch(&factor_data.get(),
test_utils::CollectStatePointers(state_data.get()));
ASSERT_TRUE(one_batch_problem.CheckConsistency());

EXPECT_NO_THROW(minimizer.Minimize(stream.GetStream(), one_batch_problem));
THROW_ON_CUDA_ERROR(cudaStreamSynchronize(stream.GetStream()));
}
}

} // namespace cunls