Skip to content

MB-59575: Revert memcpy optimizations for flat indexes #23

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
abhinavdangeti merged 4 commits into from
May 23, 2024
Merged

MB-59575: Revert memcpy optimizations for flat indexes #23

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
913cc40
bug fix: moving pointer to skip compiler optimization
Thejas-bhat May 21, 2024
4471234
using volatile keyword
Thejas-bhat May 22, 2024
06f4e48
revert flat codes optimization
Thejas-bhat May 22, 2024
4819bbe
Merge branch 'bleve' into revertFlat
abhinavdangeti May 23, 2024
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
4 changes: 0 additions & 4 deletions faiss/IndexFlat.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -169,10 +169,6 @@ FlatCodesDistanceComputer* IndexFlat::get_FlatCodesDistanceComputer() const {
}

void IndexFlat::reconstruct(idx_t key, float* recons) const {
if (mmaped) {
memcpy(recons, &(codes_ptr[key * code_size]), code_size);
return;
}
memcpy(recons, &(codes[key * code_size]), code_size);
}

Expand Down
6 changes: 0 additions & 6 deletions faiss/IndexFlat.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -54,15 +54,9 @@ struct IndexFlat : IndexFlatCodes {

// get pointer to the floating point data
float* get_xb() {
if (mmaped) {
return (float*)(codes_ptr);
}
return (float*)codes.data();
}
const float* get_xb() const {
if (mmaped) {
return (const float*)(codes_ptr);
}
return (const float*)codes.data();
}

Expand Down
17 changes: 3 additions & 14 deletions faiss/IndexFlatCodes.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -18,20 +18,9 @@ namespace faiss {
IndexFlatCodes::IndexFlatCodes(size_t code_size, idx_t d, MetricType metric)
: Index(d, metric), code_size(code_size) {}

IndexFlatCodes::IndexFlatCodes() :
code_size(0),
mmaped_size(0),
mmaped(false),
codes_ptr(nullptr) {}

IndexFlatCodes::~IndexFlatCodes() {
// setting the pointer to nullptr so that the mmap'd region is zero counted
// from faiss side and safe to be free'd/GC'd etc. on calling application layer
// of faiss.
if (mmaped) {
codes_ptr = nullptr;
}
}
IndexFlatCodes::IndexFlatCodes() : code_size(0) {}

IndexFlatCodes::~IndexFlatCodes() {}

void IndexFlatCodes::add(idx_t n, const float* x) {
FAISS_THROW_IF_NOT(is_trained);
Expand Down
3 changes: 0 additions & 3 deletions faiss/IndexFlatCodes.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -24,9 +24,6 @@ struct IndexFlatCodes : Index {

/// encoded dataset, size ntotal * code_size
std::vector<uint8_t> codes;
bool mmaped; // true if codes_ptr is pointing to a mmaped region
size_t mmaped_size;
uint8_t* codes_ptr;

IndexFlatCodes();

Expand Down
44 changes: 3 additions & 41 deletions faiss/impl/index_read.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -523,40 +523,6 @@ static IndexIVFPQ* read_ivfpq(IOReader* f, uint32_t h, int io_flags) {

int read_old_fmt_hack = 0;

/**
* flat indexes which store the codes directly, can use this API instead to have a
* pointer to the mmaped region to avoid allocation costs. works specifically with
* BufIOReader as of now.
**/
void read_codes_mmaped(IOReader* f, IndexFlat* idxf) {
idxf->mmaped = true;

// read the size of codes data
size_t size;
READANDCHECK(&size, 1);
FAISS_THROW_IF_NOT(size >= 0 && size < (uint64_t{1} << 40));
size *= 4;

// size == ntotal * code_size == ntotal * d * sizeof(float) for IndexFlat
// NOTE: the code_size value can change for indexes with encodings like
// SQ, PQ although the size value will still be equal to ntotal * code_size
// bytes which is accessible via codes_ptr.
FAISS_THROW_IF_NOT(size == idxf->ntotal * idxf->code_size);
idxf->mmaped_size = size;

// BufIOReader is the reader which has a direct pointer to the mmaped
// byte array, so we can directly set the codes_ptr to the mmaped region
BufIOReader* reader = dynamic_cast<BufIOReader*>(f);
FAISS_THROW_IF_NOT_MSG(reader, "reading over mmap'd region is supported only with BufIOReader");
FAISS_THROW_IF_NOT_MSG(reader->buf, "reader buffer is null");

idxf->codes_ptr = const_cast<uint8_t*>(reader->buf);
// seek to the point where the codes section begins
idxf->codes_ptr += reader->rp;
// update read pointer appropriately
reader->rp += size;
}

Index* read_index(IOReader* f, int io_flags) {
Index* idx = nullptr;
uint32_t h;
Expand All @@ -573,13 +539,9 @@ Index* read_index(IOReader* f, int io_flags) {
read_index_header(idxf, f);
idxf->code_size = idxf->d * sizeof(float);

if (io_flags & IO_FLAG_READ_MMAP) {
read_codes_mmaped(f, idxf);
} else {
READXBVECTOR(idxf->codes);
FAISS_THROW_IF_NOT(
idxf->codes.size() == idxf->ntotal * idxf->code_size);
}
READXBVECTOR(idxf->codes);
FAISS_THROW_IF_NOT(
idxf->codes.size() == idxf->ntotal * idxf->code_size);
// leak!
idx = idxf;
} else if (h == fourcc("IxHE") || h == fourcc("IxHe")) {
Expand Down