Skip to content

Migrate CAGRA merge() to caller-concatenated dataset + offsets contract #2507

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

Open
HowardHuang1 wants to merge 2 commits into
base: main
Choose a base branch
from
Open

Migrate CAGRA merge() to caller-concatenated dataset + offsets contract #2507

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
87a0075
Migrate CAGRA merge() to caller-concatenated dataset + offsets contract
HowardHuang1 Aug 26, 2026
50dd606
Merge remote-tracking branch 'upstream/main' into HH-migrate-merge-da…
HowardHuang1 Aug 26, 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
73 changes: 55 additions & 18 deletions c/include/cuvs/neighbors/cagra.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -989,9 +989,39 @@ CUVS_EXPORT cuvsError_t cuvsCagraIndexFromArgs(cuvsResources_t res,
* @{
*/

/**
* @brief Compute per-index write offsets for a merged dataset buffer.
*
* `cuvsCagraMerge`/`cuvsCagraMergeWithParams` require the caller to have already concatenated
* every input index's dataset (in `indices` order, applying `filter` if any) into a single buffer
* and to know each index's starting row within it. For `filter.type == NO_FILTER`, those offsets
* are just the cumulative sizes of `indices` and this function is not needed. For `BITSET`, the
* number of surviving rows per index cannot be derived any other way, so call this first.
*
* @param[in] res cuvsResources_t opaque C handle
* @param[in] indices Array of input cuvsCagraIndex_t handles that will be passed to merge
* @param[in] num_indices Number of input indices
* @param[in] filter Filter that will be passed to merge. Only `NO_FILTER` and `BITSET` supported.
* @param[out] offsets Caller-allocated array of `num_indices + 1` int64_t. Entry `i` is the row at
* which `indices[i]`'s surviving rows must start in the merged buffer; the
* last entry is the total row count of the merged buffer.
* @return cuvsError_t
*/
CUVS_EXPORT cuvsError_t cuvsCagraMergedDatasetOffsets(cuvsResources_t res,
cuvsCagraIndex_t* indices,
size_t num_indices,
cuvsFilter filter,
int64_t* offsets);

/**
* @brief Merge multiple CAGRA indices into a single CAGRA index.
*
* The caller is responsible for concatenating every input index's dataset (applying `filter` if
* any) into a single `merged_dataset` buffer before calling this, and for computing `offsets`
* (see `cuvsCagraMergedDatasetOffsets`). This function only builds/merges the graph and rebinds
* the output index to `merged_dataset` -- it never allocates or copies dataset rows itself. This
* mirrors the `cuvsCagraExtend` contract.
*
* All input indices must have been built with the same data type (`index.dtype`) and
* have the same dimensionality (`index.dims`). The merged index uses the output
* parameters specified in `cuvsCagraIndexParams`. The merge algorithm is selected automatically.
Expand Down Expand Up @@ -1024,12 +1054,14 @@ CUVS_EXPORT cuvsError_t cuvsCagraIndexFromArgs(cuvsResources_t res,
* cuvsCagraIndexParams_t merge_params;
* cuvsError_t params_create_status = cuvsCagraIndexParamsCreate(&merge_params);
*
* cuvsDataset_t merged_dataset;
* cuvsDatasetCreate(&merged_dataset);
* // Build `merged_dataset` as the caller-owned concatenation of index1 || index2 (e.g. via
* // cuvsDatasetMakePadded over a device buffer you populated yourself).
* cuvsDataset_t merged_dataset = ...;
* int64_t offsets[3] = {0, index1_size, index1_size + index2_size};
* cuvsFilter filter = {.type = NO_FILTER, .addr = 0};
*
* cuvsError_t merge_status = cuvsCagraMerge(res, merge_params, (cuvsCagraIndex_t[]){index1,
* index2}, 2, filter, merged_dataset, merged_index);
* index2}, 2, filter, merged_dataset, offsets, merged_index);
*
* // Use merged_index for search operations
*
Expand All @@ -1043,13 +1075,15 @@ CUVS_EXPORT cuvsError_t cuvsCagraIndexFromArgs(cuvsResources_t res,
* @param[in] params cuvsCagraIndexParams_t parameters for the output index
* @param[in] indices Array of input cuvsCagraIndex_t handles to merge
* @param[in] num_indices Number of input indices
* @param[in] filter Filter that can be used to filter out vectors from the merged index
* @param[out] merged_dataset Empty owning dataset handle. Merge first attempts to allocate and
* populate device storage with the same layout as the input indices. For
* an unfiltered merge, if device allocation fails, it falls back to host
* storage and returns a host-backed output index. Keep this dataset alive
* while using \p output_index. A host-backed output index must be updated
* with `cuvsCagraUpdateDataset` before device search.
* @param[in] filter Filter, already applied by the caller while building `merged_dataset`
* @param[in] merged_dataset Caller-owned dataset handle already containing the concatenated (and,
* if `filter` is set, already-filtered) dataset, with the same layout as
* the input indices. Keep this dataset alive while using
* \p output_index. A host-backed dataset must be updated with
* `cuvsCagraUpdateDataset` before device search.
* @param[in] offsets Per-index starting row within `merged_dataset`, as returned by
* `cuvsCagraMergedDatasetOffsets`. Array of `num_indices + 1` int64_t; the last
* entry must equal `merged_dataset`'s row count.
* @param[out] output_index Output handle that will store the merged index.
* Must be initialized using `cuvsCagraIndexCreate` before use.
*/
Expand All @@ -1059,25 +1093,27 @@ CUVS_EXPORT cuvsError_t cuvsCagraMerge(cuvsResources_t res,
size_t num_indices,
cuvsFilter filter,
cuvsDataset_t merged_dataset,
const int64_t* offsets,
cuvsCagraIndex_t output_index);

/**
* @brief Merge multiple CAGRA indices with explicit merge parameters.
*
* See `cuvsCagraMerge` for the full `merged_dataset`/`offsets` contract.
*
* @param[in] res cuvsResources_t opaque C handle
* @param[in] params cuvsCagraIndexParams_t parameters for the output index
* @param[in] merge_params cuvsCagraMergeParams_t parameters controlling the merge algorithm, or
* NULL to use AUTO defaults
* @param[in] indices Array of input cuvsCagraIndex_t handles to merge
* @param[in] num_indices Number of input indices
* @param[in] filter Filter that can be used to filter out vectors from the merged index
* @param[out] merged_dataset Empty owning dataset handle. Merge first attempts to allocate and
* populate device storage with the same layout as the input indices. For
* an unfiltered merge, AUTO and REBUILD can fall back to host storage if
* device allocation fails; explicit FASTENER reports the allocation
* failure instead. Keep this dataset alive while using `output_index`.
* A host-backed output index must be updated with
* `cuvsCagraUpdateDataset` before device search.
* @param[in] filter Filter, already applied by the caller while building `merged_dataset`
* @param[in] merged_dataset Caller-owned dataset handle already containing the concatenated (and,
* if `filter` is set, already-filtered) dataset. Keep this dataset alive
* while using `output_index`. A host-backed dataset must be updated with
* `cuvsCagraUpdateDataset` before device search.
* @param[in] offsets Per-index starting row within `merged_dataset`, as returned by
* `cuvsCagraMergedDatasetOffsets`. Array of `num_indices + 1` int64_t.
* @param[out] output_index Output handle initialized with `cuvsCagraIndexCreate`
*/
CUVS_EXPORT cuvsError_t cuvsCagraMergeWithParams(cuvsResources_t res,
Expand All @@ -1087,6 +1123,7 @@ CUVS_EXPORT cuvsError_t cuvsCagraMergeWithParams(cuvsResources_t res,
size_t num_indices,
cuvsFilter filter,
cuvsDataset_t merged_dataset,
const int64_t* offsets,
cuvsCagraIndex_t output_index);

/**
Expand Down
Loading
Loading