Summary: 93 instances, 63 unique

Text	Count
// TODO just do a qucikselect	1
// FIXME: convert to int32_t everywhere?	1
// FIXME: we should adjust queryTileSize to deal with this, since	3
# TODO: once deprecated classes are removed, remove the dict and just use .lower() below	1
/// FIXME when MSB of uint16 is set	1
// FIXME: maybe also consider offset in bytes? multiply by sizeof(T)?	1
// FIXME: speed up	1
// FIXME: this is a non-coalesced, unaligned, non-vectorized load	4
// FIXME: what to use for managed memory?	1
// TODO: check as thoroughfully for other index types	1
* TODO: in this file, the read functions that encouter errors may	1
// FIXME: type-specific abs	1
// FIXME: we might ultimately be calling this function with inputs	2
// TODO: parallelize?	2
// TODO shrink global storage if needed	1
// FIXME: compiler doesn't like this expression? compiler bug?	4
// TODO find a better name	3
if (with_id_map) { // FIXME test on q_map instead	1
# FIXME: no rev_swig_ptr equivalent for torch.Tensor, just convert	1
// FIXME: is this a CUDA 9 compiler bug?	2
// FIXME: inherit our same device	1
// FIXME: make sure there are no outstanding memory allocations?	1
// FIXME jhj convert to _n version	1
// FIXME: if we have exact multiples, don't need this	3
: dis > radius; // TODO templatize to remove this test	1
// TODO: Support big endian (currently supporting only little endian)	1
// FIXME: probably impractical for large # of dims?	1
// FIXME: 2 different float16 options?	2
// FIXME: really this can be into pinned memory and a true async	1
// FIXME: tune	1
nbit -= 8; // TODO remove nbit	1
// FIXME: there could be overflow here, but where should we check this?	1
// FIXME: this is a non-coalesced load	4
// FIXME: faiss CPU uses +/-FLT_MAX instead of +/-infinity	1
// TODO check nb of bytes written	1
// FIXME: stride with threads instead of single thread	1
// FIXME: some weird CUDA 11 bug? where cublasSgemmEx on	1
// FIXME: we cannot use temporary memory for new requests because	1
/// FIXME: the output distances must fit in GPU memory	1
// FIXME: some issue with getLaneId() and CUDA 10.1 and P4 GPUs?	4
// TODO should not need stable	1
// FIXME: assumes that nothing is currently running on the sub-indexes, which is	2
// FIXME: why are we doing this?	1
// FIXME: as of CUDA 11, a memory allocation error appears to be	2
# TODO: all result manipulations are in python, should move to C++ if perf	1
// FIXME: this is a non-coalesced, unaligned, 2-vectorized load	3
// FIXME: try always making this centroid id 0 so we can	1
// TODO parallelize	1
# TODO check class name	1
// FIXME: optimize with a dedicated kernel	1
// FIXME: inherit our same device	1
// FIXME: is this a CUDA 9 compiler bug?	2
// FIXME: why does getLaneId() not work when we write out below!?!?!	1
// FIXME jhj: kernel for copy	2
// FIXME: GPUize more of this	3
// FIXME: type-specific abs()	1
// FIXME: this is also slow, since we have to recover the	1
// FIXME: investigate loading separately, so we don't need this	1
// TODO: make tree of partial sums	1
// FIXME: make a SSE version	1
// TODO find a way to provide the nprobes together to do a matmul	1
// FIXME avoid a second pass over the array to sample the threshold	1
// FIXME: parameterize based on algorithm need	1