Skip to content
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

improve the preformance for hgraph bulid & add #216

Merged
merged 1 commit into from
Dec 17, 2024
Merged

improve the preformance for hgraph bulid & add #216

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
42 changes: 31 additions & 11 deletions src/io/memory_block_io.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -37,13 +37,15 @@ class MemoryBlockIO : public BasicIO<MemoryBlockIO> {
public:
explicit MemoryBlockIO(Allocator* allocator, uint64_t block_size = DEFAULT_BLOCK_SIZE)
: block_size_(block_size), allocator_(allocator), blocks_(0, allocator) {
this->update_by_block_size();
}

MemoryBlockIO(const JsonType& io_param, const IndexCommonParam& common_param)
: allocator_(common_param.allocator_), blocks_(0, common_param.allocator_) {
: MemoryBlockIO(common_param.allocator_) {
if (io_param.contains(BLOCK_IO_BLOCK_SIZE_KEY)) {
this->block_size_ =
io_param[BLOCK_IO_BLOCK_SIZE_KEY]; // TODO(LHT): trans str to uint64_t
this->update_by_block_size();
}
}

Expand Down Expand Up @@ -83,22 +85,34 @@ class MemoryBlockIO : public BasicIO<MemoryBlockIO> {
private:
[[nodiscard]] inline bool
check_valid_offset(uint64_t size) const {
return size <= blocks_.size() * block_size_;
return size <= (blocks_.size() << block_bit_);
}

inline void
update_by_block_size() {
block_bit_ = 0;
while (block_size_ > 0) {
block_size_ >>= 1;
block_bit_ += 1;
}
block_bit_ -= 1;
jiaweizone marked this conversation as resolved.
Show resolved Hide resolved
in_block_mask_ = (1ULL << block_bit_) - 1;
block_size_ = in_block_mask_ + 1;
}

inline void
check_and_realloc(uint64_t size);

[[nodiscard]] inline const uint8_t*
get_data_ptr(uint64_t offset) const {
auto block_no = offset / block_size_;
auto block_off = offset % block_size_;
auto block_no = offset >> block_bit_;
auto block_off = offset & in_block_mask_;
return blocks_[block_no] + block_off;
}

[[nodiscard]] inline bool
check_in_one_block(uint64_t off1, uint64_t off2) const {
return (off1 / block_size_) == (off2 / block_size_);
return (off1 ^ off2) < block_size_;
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

block_size_ here is equal to 0

}

private:
Expand All @@ -108,15 +122,21 @@ class MemoryBlockIO : public BasicIO<MemoryBlockIO> {

Allocator* const allocator_{nullptr};

static const uint64_t DEFAULT_BLOCK_SIZE = 128 * 1024 * 1024; // 128MB
static constexpr uint64_t DEFAULT_BLOCK_SIZE = 128 * 1024 * 1024; // 128MB

static constexpr uint64_t DEFAULT_BLOCK_BIT = 27;

uint64_t block_bit_{DEFAULT_BLOCK_BIT};

uint64_t in_block_mask_ = (1 << DEFAULT_BLOCK_BIT) - 1;
};

void
MemoryBlockIO::WriteImpl(const uint8_t* data, uint64_t size, uint64_t offset) {
check_and_realloc(size + offset);
uint64_t cur_size = 0;
auto start_no = offset / block_size_;
auto start_off = offset % block_size_;
auto start_no = offset >> block_bit_;
auto start_off = offset & in_block_mask_;
auto max_size = block_size_ - start_off;
while (cur_size < size) {
uint8_t* cur_write = blocks_[start_no] + start_off;
Expand All @@ -134,8 +154,8 @@ MemoryBlockIO::ReadImpl(uint64_t size, uint64_t offset, uint8_t* data) const {
bool ret = check_valid_offset(size + offset);
if (ret) {
uint64_t cur_size = 0;
auto start_no = offset / block_size_;
auto start_off = offset % block_size_;
auto start_no = offset >> block_bit_;
auto start_off = offset & in_block_mask_;
auto max_size = block_size_ - start_off;
while (cur_size < size) {
const uint8_t* cur_read = blocks_[start_no] + start_off;
Expand Down Expand Up @@ -189,7 +209,7 @@ MemoryBlockIO::check_and_realloc(uint64_t size) {
if (check_valid_offset(size)) {
return;
}
const uint64_t new_block_count = (size + this->block_size_ - 1) / block_size_;
const uint64_t new_block_count = (size + this->block_size_ - 1) >> block_bit_;
auto cur_block_size = this->blocks_.size();
while (cur_block_size < new_block_count) {
this->blocks_.emplace_back((uint8_t*)(allocator_->Allocate(block_size_)));
Expand Down
20 changes: 8 additions & 12 deletions src/simd/avx512.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -299,9 +299,10 @@ SQ8ComputeL2Sqr(const float* query,
for (; i + 15 < dim; i += 16) {
// Load data into registers
__m128i code_values = _mm_loadu_si128(reinterpret_cast<const __m128i*>(codes + i));
__m512 diff_values = _mm512_loadu_ps(diff + i);

__m512i codes_512 = _mm512_cvtepu8_epi32(code_values);
__m512 code_floats = _mm512_div_ps(_mm512_cvtepi32_ps(codes_512), _mm512_set1_ps(255.0f));
__m512 diff_values = _mm512_loadu_ps(diff + i);
__m512 lowerBound_values = _mm512_loadu_ps(lowerBound + i);
__m512 query_values = _mm512_loadu_ps(query + i);

Expand Down Expand Up @@ -373,19 +374,14 @@ SQ8ComputeCodesL2Sqr(const uint8_t* codes1,
for (; i + 15 < dim; i += 16) {
__m128i code1_values = _mm_loadu_si128(reinterpret_cast<const __m128i*>(codes1 + i));
__m128i code2_values = _mm_loadu_si128(reinterpret_cast<const __m128i*>(codes2 + i));
__m512i codes1_512 = _mm512_cvtepu8_epi32(code1_values);
__m512i codes2_512 = _mm512_cvtepu8_epi32(code2_values);
__m512 code1_floats = _mm512_div_ps(_mm512_cvtepi32_ps(codes1_512), _mm512_set1_ps(255.0f));
__m512 code2_floats = _mm512_div_ps(_mm512_cvtepi32_ps(codes2_512), _mm512_set1_ps(255.0f));
__m512 diff_values = _mm512_loadu_ps(diff + i);
__m512 lowerBound_values = _mm512_loadu_ps(lowerBound + i);

// Perform calculations
__m512 scaled_codes1 = _mm512_fmadd_ps(code1_floats, diff_values, lowerBound_values);
__m512 scaled_codes2 = _mm512_fmadd_ps(code2_floats, diff_values, lowerBound_values);
__m512 val = _mm512_sub_ps(scaled_codes1, scaled_codes2);
val = _mm512_mul_ps(val, val);
sum = _mm512_add_ps(sum, val);
__m512i codes1_512 = _mm512_cvtepu8_epi32(code1_values);
jiaweizone marked this conversation as resolved.
Show resolved Hide resolved
__m512i codes2_512 = _mm512_cvtepu8_epi32(code2_values);
__m512 sub = _mm512_cvtepi32_ps(_mm512_sub_epi32(codes1_512, codes2_512));
__m512 scaled = _mm512_mul_ps(sub, _mm512_set1_ps(1.0 / 255.0f));
__m512 val = _mm512_mul_ps(scaled, diff_values);
sum = _mm512_fmadd_ps(val, val, sum);
}

// Horizontal addition
Expand Down
Loading