-
Notifications
You must be signed in to change notification settings - Fork 64
Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
Add sparse index support to knowhere with linscan and wand algorithm;…
… added related unit test Signed-off-by: Buqian Zheng <zhengbuqian@gmail.com>
- Loading branch information
1 parent
5230b6a
commit 950a375
Showing
12 changed files
with
1,377 additions
and
2 deletions.
There are no files selected for viewing
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,225 @@ | ||
| // Copyright (C) 2019-2023 Zilliz. All rights reserved. | ||
| // | ||
| // Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance | ||
| // with the License. You may obtain a copy of the License at | ||
| // | ||
| // http://www.apache.org/licenses/LICENSE-2.0 | ||
| // | ||
| // Unless required by applicable law or agreed to in writing, software valributed under the License | ||
| // is valributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express | ||
| // or implied. See the License for the specific language governing permissions and limitations under the License. | ||
|
|
||
| #pragma once | ||
|
|
||
| #include <algorithm> | ||
| #include <cstddef> | ||
| #include <cstdint> | ||
| #include <cstring> | ||
| #include <type_traits> | ||
| #include <vector> | ||
|
|
||
| #include "knowhere/operands.h" | ||
|
|
||
| namespace knowhere::sparse { | ||
|
|
||
| // integer type in SparseRow | ||
| using table_t = uint32_t; | ||
| // type used to represent the id of a vector in the index interface. | ||
| // this is same as other index types. | ||
| using label_t = int64_t; | ||
|
|
||
| template <typename T> | ||
| struct IdVal { | ||
| table_t id; | ||
| T val; | ||
|
|
||
| IdVal() = default; | ||
| IdVal(table_t id, T val) : id(id), val(val) { | ||
| } | ||
|
|
||
| inline friend bool | ||
| operator<(const IdVal& lhs, const IdVal& rhs) { | ||
| return lhs.val < rhs.val || (lhs.val == rhs.val && lhs.id < rhs.id); | ||
| } | ||
| inline friend bool | ||
| operator>(const IdVal& lhs, const IdVal& rhs) { | ||
| return !(lhs < rhs); | ||
| } | ||
|
|
||
| inline friend bool | ||
| operator==(const IdVal& lhs, const IdVal& rhs) { | ||
| return lhs.id == rhs.id && lhs.val == rhs.val; | ||
| } | ||
| }; | ||
|
|
||
| template <typename T> | ||
| class SparseRow { | ||
| static_assert(std::is_same_v<T, fp32>, "SparseRow supports float only"); | ||
| public: | ||
| SparseRow() = default; | ||
|
|
||
| SparseRow(const SparseRow& other) { | ||
| copy_from(other); | ||
| } | ||
|
|
||
| SparseRow& | ||
| operator=(const SparseRow& other) { | ||
| copy_from(other); | ||
| return *this; | ||
| } | ||
|
|
||
| ~SparseRow() { | ||
| if (own_data && data != nullptr) { | ||
| delete[] data; | ||
| } | ||
| } | ||
|
|
||
| size_t | ||
| size() const { | ||
| return count; | ||
| } | ||
|
|
||
| size_t | ||
| byte_size() const { | ||
| return count * (sizeof(table_t) + sizeof(T)) + sizeof(*this); | ||
| } | ||
|
|
||
| void* | ||
| data_ptr() const { | ||
| return data; | ||
| } | ||
|
|
||
| // dim of a sparse vector is the max index + 1, or 0 for an empty vector. | ||
| int64_t | ||
| dim() const { | ||
| if (count == 0) | ||
| return 0; | ||
| return (*this)[count - 1].id + 1; | ||
| } | ||
|
|
||
| IdVal<T> | ||
| operator[](size_t i) const { | ||
| table_t index = *(table_t*)(data + i * (sizeof(table_t) + sizeof(T))); | ||
| T value = *(T*)(data + i * (sizeof(table_t) + sizeof(T)) + sizeof(table_t)); | ||
| return {index, value}; | ||
| } | ||
|
|
||
| void | ||
| set_at(size_t i, table_t index, T value) { | ||
| *(table_t*)(data + i * (sizeof(table_t) + sizeof(T))) = index; | ||
| *(T*)(data + i * (sizeof(table_t) + sizeof(T)) + sizeof(table_t)) = value; | ||
| } | ||
|
|
||
| float | ||
| dot(const SparseRow<T>& other) const { | ||
| float product_sum = 0.0f; | ||
| size_t i = 0, j = 0; | ||
| while (i < count && j < other.count) { | ||
| auto [index1, value1] = (*this)[i]; | ||
| auto [index2, value2] = other[j]; | ||
|
|
||
| if (index1 < index2) { | ||
| ++i; | ||
| } else if (index1 > index2) { | ||
| ++j; | ||
| } else { | ||
| product_sum += value1 * value2; | ||
| ++i; | ||
| ++j; | ||
| } | ||
| } | ||
| return product_sum; | ||
| } | ||
|
|
||
| void | ||
| reserve(size_t count) { | ||
| if (own_data && data != nullptr) { | ||
| delete[] data; | ||
| } | ||
| data = new uint8_t[count * (sizeof(table_t) + sizeof(T))]; | ||
| own_data = true; | ||
| this->count = count; | ||
| } | ||
|
|
||
| // performs deep copy of data | ||
| void | ||
| copy_from(const SparseRow<T>& other) { | ||
| reserve(other.count); | ||
| memcpy(data, other.data, count * (sizeof(table_t) + sizeof(T))); | ||
| } | ||
|
|
||
| private: | ||
| // data must be sorted by column id. use raw pointer for easy mmap and zero | ||
| // copy. | ||
| uint8_t* data = nullptr; | ||
| size_t count = 0; | ||
| bool own_data = false; | ||
| }; | ||
|
|
||
| // When pushing new elements into a MaxMinHeap, only `capacity` elements with the | ||
| // largest val are kept. pop()/top() returns the smallest element out of them. | ||
| template <typename T> | ||
| class MaxMinHeap { | ||
| public: | ||
| explicit MaxMinHeap(int capacity) : capacity_(capacity), pool_(capacity) { | ||
| } | ||
| void | ||
| push(table_t id, T val) { | ||
| if (size_ < capacity_) { | ||
| pool_[size_] = {id, val}; | ||
| size_ += 1; | ||
| std::push_heap(pool_.begin(), pool_.begin() + size_, std::greater<IdVal<T>>()); | ||
| } else if (val > pool_[0].val) { | ||
| sift_down(id, val); | ||
| } | ||
| } | ||
| table_t | ||
| pop() { | ||
| std::pop_heap(pool_.begin(), pool_.begin() + size_, std::greater<IdVal<T>>()); | ||
| size_ -= 1; | ||
| return pool_[size_].id; | ||
| } | ||
| [[nodiscard]] size_t | ||
| size() const { | ||
| return size_; | ||
| } | ||
| [[nodiscard]] bool | ||
| empty() const { | ||
| return size() == 0; | ||
| } | ||
| IdVal<T> | ||
| top() const { | ||
| return pool_[0]; | ||
| } | ||
| [[nodiscard]] bool | ||
| full() const { | ||
| return size_ == capacity_; | ||
| } | ||
|
|
||
| private: | ||
| void | ||
| sift_down(table_t id, T val) { | ||
| size_t i = 0; | ||
| for (; 2 * i + 1 < size_;) { | ||
| size_t j = i; | ||
| size_t l = 2 * i + 1, r = 2 * i + 2; | ||
| if (pool_[l].val < val) { | ||
| j = l; | ||
| } | ||
| if (r < size_ && pool_[r].val < std::min(pool_[l].val, val)) { | ||
| j = r; | ||
| } | ||
| if (i == j) { | ||
| break; | ||
| } | ||
| pool_[i] = pool_[j]; | ||
| i = j; | ||
| } | ||
| pool_[i] = {id, val}; | ||
| } | ||
|
|
||
| size_t size_ = 0, capacity_; | ||
| std::vector<IdVal<T>> pool_; | ||
| }; // class MaxMinHeap | ||
|
|
||
| } // namespace knowhere::sparse |
Oops, something went wrong.