hashtable optimize, aggresive SIMD via openmp

Makefile

@@ -93,7 +93,7 @@ ifeq ($(AQ_DEBUG), 1)
 	OPTFLAGS = -g3 #-static-libsan -fsanitize=address 
 #	LINKFLAGS = 
 else
-	OPTFLAGS += -Ofast -DNDEBUG -fno-stack-protector 
+	OPTFLAGS += -Ofast -DNDEBUG -fno-stack-protector -fopenmp
 	LINKFLAGS += -flto -s
 endif
 

copyright_stamper.py

@@ -0,0 +1,53 @@
+import os
+
+payload = ('''\
+(C) Bill Sun 2022 - 2023
+All rights reserved. (or some other license stuff)
+''' ).strip().split('\n')
+
+comment_factory = lambda mark, enclosure = '': (f'''\
+{enclosure}{mark}
+{mark} {f'{chr(10)}{mark} '.join(payload)}
+{mark}{enclosure}\n
+''' ).encode()
+
+py_payload = comment_factory('#')
+c_payload = comment_factory('*', '/')
+
+curr = ['.']
+while curr:
+    next = []
+    for dir in curr:
+        items = os.listdir(dir)
+        for file in items:
+            fullpath = f'{dir}{os.sep}{file}'
+            if os.path.isdir(fullpath):
+                next.append(fullpath)
+            else:
+                def write_to_file(payload: str):
+                    with open(fullpath, 'rb+') as f:
+                        content = f.read()
+                        if not content.startswith(payload):
+                            f.seek(0)
+                            f.write(payload + content)
+                            print('processed', fullpath)
+                        else:
+                            print('not processed', fullpath)
+                            
+                if (
+                        file.lower() == 'makefile' or 
+                        file.lower() == 'dockerfile' or 
+                        '.' in file and 
+                        file[file.rfind('.') + 1:].lower() 
+                            in 
+                        ['py', 'sh']
+                    ):
+                    write_to_file(py_payload)
+                elif (
+                        '.' in file and 
+                        file[file.rfind('.') + 1:].lower() 
+                            in 
+                        ['cc', 'c', 'cpp', 'cxx', 'hpp', 'h']
+                    ):
+                    write_to_file(c_payload)
+    curr = next

engine/TODO.md

@@ -8,4 +8,8 @@
 
 ## 2. ColRef supports multiple objects
    - A.a = B.b then in projection A.a B.b will refer to same projection
-   - Colref::ProjEq(ColRef v) => this == v or v in this.proj_eqs
+   - Colref::ProjEq(ColRef v) => this == v or v in this.proj_eqs
+
+## 3. External Optimizing Tools
+   - mold/sold instead of builtin linker will boost linker speed
+   - bolt that have binary optimizer

engine/ast.py

@@ -566,6 +566,7 @@ class scan(ast_node):
         self.parent.context.scans.append(self)
         
     def produce(self, node):
+        self.start += '#pragma openmp simd\n'
         if self.loop_style == scan.LoopStyle.foreach:
             self.colref = node
             self.start += f'for ({self.const}auto& {self.it_var} : {node}) {{\n'
@@ -595,7 +596,7 @@ class scan(ast_node):
                     self.start + 
                     self.front + 
                     b + 
-                    '}'
+                    '\n}'
                 ) for b in self.body])
                 + 
                 self.end
@@ -606,7 +607,7 @@ class scan(ast_node):
                 self.start + 
                 self.front + 
                 '\n'.join(self.body) + 
-                '}' +
+                '\n}' +
                 self.end
             )
         self.context.remove_scan(self, scan_assembly) 
@@ -657,11 +658,13 @@ class groupby_c(ast_node):
         self.context.emitc(f'AQHashTable<{self.group_type}, '
             f'transTypes<{self.group_type}, hasher>> {self.group} {{{self.total_sz}}};')
         self.n_grps = len(self.glist)
-        self.scanner = scan(self, self.total_sz, it_name=scanner_itname)
-        self.scanner.add(f'{self.group}.hashtable_push(forward_as_tuple({g_contents}), {self.scanner.it_var});')
-
+        
+        # self.scanner = scan(self, self.total_sz, it_name=scanner_itname)
+        # self.scanner.add(f'{self.group}.hashtable_push(forward_as_tuple({g_contents}), {self.scanner.it_var});')
+        self.context.emitc(f'{self.group}.hashtable_push_all({g_contents}, {self.total_sz});')
+        
     def consume(self, _):
-        self.scanner.finalize()
+        # self.scanner.finalize()
         self.context.emitc('printf("ht_construct: %lld\\n", (chrono::high_resolution_clock::now() - timer).count()); timer = chrono::high_resolution_clock::now();')
         self.context.emitc(f'auto {self.vecs} = {self.group}.ht_postproc({self.total_sz});')
         self.context.emitc('printf("ht_postproc: %lld\\n", (chrono::high_resolution_clock::now() - timer).count()); timer = chrono::high_resolution_clock::now();')

sdk/aquery.h

@@ -1,13 +1,13 @@
 #ifndef _AQUERY_H
 #define _AQUERY_H
 
-enum Log_level {
+enum Log_level : int {
 	LOG_INFO,
 	LOG_ERROR,
 	LOG_SILENT
 };
 
-enum Backend_Type {
+enum Backend_Type : int {
 	BACKEND_AQuery,
 	BACKEND_MonetDB,
 	BACKEND_MariaDB

server/DataSource_conn.h

@@ -14,7 +14,7 @@ struct AQQueryResult {
 
 #ifndef __AQBACKEND_TYPE__
 #define __AQBACKEND_TYPE__ 1
-enum Backend_Type {
+enum Backend_Type : int {
 	BACKEND_AQuery,
 	BACKEND_MonetDB,
 	BACKEND_MariaDB, 

server/hasher.h

@@ -1,3 +1,7 @@
+/*
+* (C) Bill Sun 2022 - 2023
+*/
+
 #pragma once
 
 #include <type_traits>
@@ -5,6 +9,7 @@
 #include <functional>
 #include <string_view>
 #include "types.h"
+#include "vector_type.hpp"
 // #include "robin_hood.h"
 #include "unordered_dense.h"
 
@@ -138,68 +143,145 @@ namespace ankerl::unordered_dense{
 	struct hash<std::tuple<Types...>> : public hasher<Types...>{ };
 }
 
+template <class Key, class Hash>
+class AQHashTable : public ankerl::unordered_dense::set<Key, Hash> {
+public:
+	uint32_t* reversemap, *mapbase, *ht_base;
+	AQHashTable() = default;
+	explicit AQHashTable(uint32_t sz) 
+		: ankerl::unordered_dense::set<Key, Hash>{} {
+		this->reserve(sz);
+		this->m_values.reserve(sz);
+		reversemap = static_cast<uint32_t *>(malloc(sizeof(uint32_t) * sz * 2));
+		mapbase = reversemap + sz;
+		ht_base =  static_cast<uint32_t *>(calloc(sz, sizeof(uint32_t)));
+	}
+
+	void init(uint32_t sz) {
+		ankerl::unordered_dense::set<Key, Hash>::reserve(sz);
+		reversemap = static_cast<uint32_t *>(malloc(sizeof(uint32_t) * sz * 2));
+		mapbase = reversemap + sz;
+		ht_base =  static_cast<uint32_t *>(calloc(sz, sizeof(uint32_t)));
+	}
+
+	template<typename... Keys_t>
+	inline void hashtable_push_all(Keys_t& ... keys, uint32_t len) {
+		for(uint32_t i = 0; i < len; ++i) 
+			reversemap[i] = ankerl::unordered_dense::set<Key, Hash>::hashtable_push(keys[i]...);
+		for(uint32_t i = 0; i < len; ++i) 
+			++ht_base[reversemap[i]]; 
+	}
+	inline void hashtable_push(Key&& k, uint32_t i){
+		reversemap[i] = ankerl::unordered_dense::set<Key, Hash>::hashtable_push(k);
+		++ht_base[reversemap[i]]; // do this seperately?
+	}
+
+	auto ht_postproc(uint32_t sz) {
+		auto& arr_values = this->values();
+		const auto& len = this->size();
+
+		auto vecs = static_cast<vector_type<uint32_t>*>(malloc(sizeof(vector_type<uint32_t>) * len));
+		vecs[0].init_from(ht_base[0], mapbase);
+		for (uint32_t i = 1; i < len; ++i) {
+			vecs[i].init_from(ht_base[i], mapbase + ht_base[i - 1]);
+			ht_base[i] += ht_base[i - 1];
+		}
+		for (uint32_t i = 0; i < sz; ++i) {
+			auto id = reversemap[i];
+			mapbase[--ht_base[id]] = i;    
+		}
+		return vecs;
+	}
+};
+
+
 template <
-	typename ValueType = bool, 
+	typename ValueType = uint32_t,
 	int PerfectHashingThreshold = 12
 >
 struct PerfectHashTable {
-	// static int m_PerfectHashingThreshold = 12;
-	using key_t = std::conditional_t<PerfectHashingThreshold <= 8, uint8_t, 
-					std::conditional_t<PerfectHashingThreshold <= 16, uint16_t, 
-					std::conditional_t<PerfectHashingThreshold <= 32, uint32_t, 
-					uint64_t
+	using key_t = std::conditional_t<PerfectHashingThreshold <= 8, uint8_t,
+		std::conditional_t<PerfectHashingThreshold <= 16, uint16_t,
+		std::conditional_t<PerfectHashingThreshold <= 32, uint32_t,
+		uint64_t
 		>>>;
-
-	int n_cols, n_rows = 0;
-	// char bits[32];
-	ValueType table[1 << PerfectHashingThreshold];
-	// PerfectHashTable(int n_cols, char* bits) {
-	// 	this->n_cols = n_cols;
-	// 	memcpy(this->bits, bits, 32);
-	// }
-	// template<typename ... Types, template <typename> class VT> 
-	// PerfectHashTable(VT<Types> ... args) {
-		
-	// }
+	constexpr static uint32_t tbl_sz = 1 << PerfectHashingThreshold;
 	template <typename ... Types, template <typename> class VT>
-	// std::enable_if_t<std::is_same_v<ValueType, bool>, void>
-	void
+	static vector_type<uint32_t>*
 	construct(VT<Types>&... args) { // construct a hash set
-		((this->n_cols = args.size), ...);
+		AQTmr();
+		int n_cols, n_rows = 0;
+
+		((n_cols = args.size), ...);
 		static_assert(
 			(sizeof...(Types) < PerfectHashingThreshold) &&
-			//(sizeof(Types) + ...) < PerfectHashingThreshold && 
 			(std::is_integral_v<Types> && ...),
 			"Types must be integral and less than 12 wide in total."
 			);
-		// this should be an attrib of VT.
-		key_t* // this better be automatically determined by Threshould 
-			hash_values = static_cast<key_t*>(
-					calloc(this->n_cols, sizeof(key_t))
-				);
-		//new short[this->n_cols] {0}; // use calloc/delete
+		key_t* 
+		hash_values = static_cast<key_t*>(
+				calloc(n_cols, sizeof(key_t))
+		);
 		auto get_hash = [&hash_values](auto& arg, int idx) {
-			uint32_t i = 0;
-			if(idx > 0)
-				for (auto& a : arg) {
+			
+			if (idx > 0) {
+#pragma omp simd
+				for (uint32_t i = 0; i < arg.size; ++i) {
 					hash_values[i] =
 						(hash_values[i] << arg.stats.bits) +
-						(a - arg.stats.minima);
-					++i;
+						(arg.container[i] - arg.stats.minima);
 				}
-			else 
-				for (auto& a : arg) {
-					hash_values[i] = a - arg.stats.minima;
-					++i;
+			}
+			else {
+#pragma omp simd
+				for (uint32_t i = 0; i < arg.size; ++i) {
+						hash_values[i] = arg.container[i] - arg.stats.minima;
+					}
 				}
-		};
+			};
 		int idx = 0;
 		(get_hash(args, idx++), ...);
-		for (uint32_t i = 0; i < this->n_cols; ++i) {
-			this->table[hash_values[i]] = true;
-			// problem: random memory access
+		uint32_t cnt[tbl_sz];
+		uint32_t n_grps = 0;
+		memset(cnt, 0, tbl_sz * sizeof(tbl_sz));
+#pragma omp simd
+		for (uint32_t i = 0; i < n_cols; ++i) {
+			++cnt[hash_values[i]];
+		}
+		ValueType grp_ids[tbl_sz];
+#pragma omp simd
+		for (ValueType i = 0; i < tbl_sz; ++i) {
+			if (cnt[i] != 0) {
+				cnt[n_grps] = cnt[i];
+				grp_ids[i] = n_grps++;
+			}
+		}
+		uint32_t* idxs = static_cast<uint32_t*>(
+			malloc(n_cols * sizeof(uint32_t))
+		);
+		uint32_t** idxs_ptr = static_cast<uint32_t**>(
+			malloc(n_grps * sizeof(uint32_t*))
+		);
+		idxs_ptr[0] = idxs;
+#ifdef _MSCVER
+#pragma omp simd
+#endif
+		for (int i = 1; i < n_grps; ++i) {
+			idxs_ptr[i] = idxs_ptr[i - 1] + cnt[i - 1];
+		}
+#pragma omp simd 
+		for (int i = 0; i < n_cols; ++i) {
+			*(idxs_ptr[grp_ids[hash_values[i]]]++) = i;
+		}
+		vector_type<uint32_t>* idxs_vec = static_cast<vector_type<uint32_t>*>(
+			malloc(n_grps * sizeof(vector_type<uint32_t>))
+		);
+#pragma omp simd
+		for (int i = 0; i < n_grps; ++i) {
+			idxs_vec[i].container = idxs_ptr[i];
+			idxs_vec[i].size = cnt[i];
 		}
-		// delete[] hash_values;
-		free(hash_values); // dispatch to gc
+		free(hash_values);
+		return idxs_vec;
 	}
 };

server/libaquery.h

@@ -57,7 +57,7 @@ char* intToString(T val, char* buf){
 }
 
 
-enum Log_level {
+enum Log_level : int {
 	LOG_INFO,
 	LOG_ERROR,
 	LOG_SILENT
@@ -65,7 +65,7 @@ enum Log_level {
 
 #ifndef __AQBACKEND_TYPE__
 #define __AQBACKEND_TYPE__ 1
-enum Backend_Type {
+enum Backend_Type : int {
 	BACKEND_AQuery,
 	BACKEND_MonetDB,
 	BACKEND_MariaDB, 

server/monetdb_ext.c

@@ -1,3 +1,7 @@
+/*
+* (C) Bill Sun 2022 - 2023
+*/
+
 #include "pch_msc.hpp"
 // Non-standard Extensions for MonetDBe, may break concurrency control!
 

server/server.cpp

@@ -80,7 +80,7 @@ extern "C" void __DLLEXPORT__ receive_args(int argc, char**argv){
     n_recvd = argv;
 }
 
-enum BinaryInfo_t {
+enum BinaryInfo_t : int { // For ABI consistency between compiler
 	MSVC, MSYS, GCC, CLANG, AppleClang
 };
 

server/table.h

@@ -35,7 +35,7 @@ struct ColRef_cstorage {
 	void* container;
 	unsigned int size, capacity;
 	const char* name;
-	int ty; // what if enum is not int?
+	int ty; 
 };
 
 template <template <class...> class VT, class T, 

server/types.h

@@ -32,7 +32,7 @@ constexpr static bool is_vector_type = is_vector_impl<T>::value;
 template <class T>
 constexpr size_t aq_szof = sizeof(T);
 template <>
-inline constexpr size_t aq_szof<void> = 0;
+constexpr size_t aq_szof<void> = 0;
 template <class T1, class T2>
 struct aqis_same_impl {
 	constexpr static bool value = 
@@ -68,7 +68,7 @@ template <class T1, class T2>
 constexpr bool aqis_same<T1, T2> = aqis_same_impl<T1, T2>::value;
 
 namespace types {
-	enum Type_t {
+	enum Type_t : int {
 		__AQUERY_TYPES__
 	};
 	static constexpr const char* printf_str[] = { "%d", "%f", "%s", "%lf", "%Lf", "%ld", "%s", "%hi", "%s", "%s", "%hhd",

server/vector_type.hpp

@@ -1,5 +1,5 @@
 /*
-* Bill Sun 2022
+* (C) Bill Sun 2022 - 2023
 */
 
 
@@ -13,7 +13,11 @@
 #include <initializer_list>
 #include <unordered_set>
 #include <iostream>
-#include "hasher.h"
+
+template <typename _Ty>
+class vector_type;
+
+// #include "hasher.h"
 #include "types.h"
 #include "gc.h"
 #pragma pack(push, 1)
@@ -484,56 +488,6 @@ public:
 };
 #pragma pack(pop)
 
-template <class Key, class Hash>
-class AQHashTable : public ankerl::unordered_dense::set<Key, Hash> {
-public:
-	uint32_t* reversemap, *mapbase, *ht_base;
-	AQHashTable() = default;
-	explicit AQHashTable(uint32_t sz) 
-		: ankerl::unordered_dense::set<Key, Hash>{} {
-		this->reserve(sz);
-		this->m_values.reserve(sz);
-		reversemap = static_cast<uint32_t *>(malloc(sizeof(uint32_t) * sz * 2));
-		mapbase = reversemap + sz;
-		ht_base =  static_cast<uint32_t *>(calloc(sz, sizeof(uint32_t)));
-	}
-
-	void init(uint32_t sz) {
-		ankerl::unordered_dense::set<Key, Hash>::reserve(sz);
-		reversemap = static_cast<uint32_t *>(malloc(sizeof(uint32_t) * sz * 2));
-		mapbase = reversemap + sz;
-		ht_base =  static_cast<uint32_t *>(calloc(sz, sizeof(uint32_t)));
-	}
-
-	template<typename... Keys_t>
-	inline void hashtable_push_all(Keys_t& ... keys, uint32_t len) {
-		for(uint32_t i = 0; i < len; ++i) 
-			reversemap[i] = ankerl::unordered_dense::set<Key, Hash>::hashtable_push(keys[i]...);
-		for(uint32_t i = 0; i < len; ++i) 
-			++ht_base[reversemap[i]]; 
-	}
-	inline void hashtable_push(Key&& k, uint32_t i){
-		reversemap[i] = ankerl::unordered_dense::set<Key, Hash>::hashtable_push(k);
-		++ht_base[reversemap[i]]; // do this seperately?
-	}
-
-	auto ht_postproc(uint32_t sz) {
-		auto& arr_values = this->values();
-		const auto& len = this->size();
-
-		auto vecs = static_cast<vector_type<uint32_t>*>(malloc(sizeof(vector_type<uint32_t>) * len));
-		vecs[0].init_from(ht_base[0], mapbase);
-		for (uint32_t i = 1; i < len; ++i) {
-			vecs[i].init_from(ht_base[i], mapbase + ht_base[i - 1]);
-			ht_base[i] += ht_base[i - 1];
-		}
-		for (uint32_t i = 0; i < sz; ++i) {
-			auto id = reversemap[i];
-			mapbase[--ht_base[id]] = i;    
-		}
-		return vecs;
-	}
-};
 
 template<>
 vector_type<std::string_view>::vector_type(const uint32_t size, void* data);