Dart Gather Kernels#

void trt_edgellm::kernel::gatherRows(
void *dst,
void const *src,
int32_t const *rowIndices,
int64_t numRows,
int64_t rowBytes,
cudaStream_t stream
)#

Dtype-agnostic row gather: dst[r, :] = src[rowIndices[r], :].

Rows are treated as opaque byte spans of length rowBytes. dst and src must not overlap (gather out-of-place into a scratch buffer, then copy back if compaction in place is desired). Uses 16-byte vectorized copies when rowBytes and both pointers are 16-byte aligned, byte copies otherwise.

Parameters:

  • dst – [out] Destination base pointer (device), at least numRows * rowBytes bytes

  • src – [in] Source base pointer (device)

  • rowIndices – [in] Device INT32 array of source row indices, length numRows

  • numRows – [in] Number of rows to gather

  • rowBytes – [in] Size of one row in bytes

  • stream – [in] CUDA stream for execution

On this page