paulhorn/sqlite
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Speed up SQL aggregate functions percentile() and median() by using quickselect to find the required values instead of fully sorting the array of values with quicksort.

Browse Source 
FossilOrigin-Name: 0796d337f6a0ef02a460a24eadda20e33e8fb9ec64b696ae1addee69c1bfddeb
This commit is contained in:
dan
2026-06-15 17:27:05 +00:00
parent f0075cd2e4 b348459070
commit a0dc22438d
3 changed files with 63 additions and 25 deletions
Download Patch File Download Diff File Expand all files Collapse all files
manifest
+8 -7
View File
@@ -1,5 +1,5 @@
C New\sWindows\smakefile\stargets\sfor\scommon\sEXEs\sthat\somit\sthe\s".exe"\ssuffix,\sso\nthat\swhen\smuscle\smemory\skicks\sin\sand\swe\stype\s"make\ssqlite3"\son\swindows,\sit\nstill\sworks.
D 2026-06-15T17:14:12.845
C Speed\sup\sSQL\saggregate\sfunctions\spercentile()\sand\smedian()\sby\susing\squickselect\sto\sfind\sthe\srequired\svalues\sinstead\sof\sfully\ssorting\sthe\sarray\sof\svalues\swith\squicksort.
D 2026-06-15T17:27:05.287
F .fossil-settings/binary-glob 61195414528fb3ea9693577e1980230d78a1f8b0a54c78cf1b9b24d0a409ed6a x
F .fossil-settings/empty-dirs dbb81e8fc0401ac46a1491ab34a7f2c7c0452f2f06b54ebb845d024ca8283ef1
F .fossil-settings/ignore-glob 35175cdfcf539b2318cb04a9901442804be81cd677d8b889fcc9149c21f239ea
@@ -691,7 +691,7 @@ F src/delete.c 59eeca3fb88c29329afc41bb803ee568b120d9dd7470b5f38ab55cc38390b451
F src/expr.c e97dd9f6ada4c448764e225d8963091bf630b3efb2c92e4d0762571cca2a14e5
F src/fault.c 460f3e55994363812d9d60844b2a6de88826e007
F src/fkey.c 931f74cec1dc8038a0217ef340c91ce147dd1bbed08dc40c47ee0ec6edfffb08
F src/func.c 555d5f7686f9eef20fe3574889403f307a6cb16ada5b05cbc6a9288dcf75aaeb
F src/func.c b62f451d670288d4b3fbd8b0a6aa98667de30de7e8c6de9b19112a9296b9ce1a
F src/global.c a19e4b1ca1335f560e9560e590fc13081e21f670643367f99cb9e8f9dc7d615b
F src/hash.c 03c8c0f4be9e8bcb6de65aa26d34a61d48a9430747084a69f9469fbb00ea52ca
F src/hash.h 46b92795a95bfefb210f52f0c316e9d7cdbcdd7e7fcfb0d8be796d3a5767cddf
@@ -2208,8 +2208,9 @@ F tool/warnings-clang.sh bbf6a1e685e534c92ec2bfba5b1745f34fb6f0bc2a362850723a9ee
F tool/warnings.sh a554d13f6e5cf3760f041b87939e3d616ec6961859c3245e8ef701d1eafc2ca2
F tool/win/sqlite.vsix deb315d026cc8400325c5863eef847784a219a2f
F tool/winmain.c 00c8fb88e365c9017db14c73d3c78af62194d9644feaf60e220ab0f411f3604c
P 0b5378678e3d095ef982bb3b5d5ad05dfb15ce4ee12170b8d0a83fa4b75404c0
R 9cbb6763894903672c2691245ef8b1d2
U drh
Z b18666bcbe3bd63a32e5fa1c68a87675
P 1152463a66b47eed27b71e87533e8361b6077dc54ff55b8cd2ccde5cfa8199bd 9b43500f51ae7b43f0e2cc805bb8c93907eddbd8caf21ee5b5a3d2e795962088
R 47d7442e9499fb070afb88cded90736e
T +closed 9b43500f51ae7b43f0e2cc805bb8c93907eddbd8caf21ee5b5a3d2e795962088
U dan
Z 8aad25eee80e8de00de5243ef6acbd82
# Remove this line to create a well-formed Fossil manifest.
manifest.uuid
+1 -1
View File
@@ -1 +1 @@
1152463a66b47eed27b71e87533e8361b6077dc54ff55b8cd2ccde5cfa8199bd
0796d337f6a0ef02a460a24eadda20e33e8fb9ec64b696ae1addee69c1bfddeb
src/func.c
+54 -17
View File
@@ -2967,8 +2967,17 @@ static void percentStep(sqlite3_context *pCtx, int argc, sqlite3_value **argv){
** (1) To avoid a dependency on qsort()
** (2) To avoid the function call to the comparison routine for each
** comparison.
**
** If parameter iReq is non-negative, then the caller will only access
** elements a[iReq] and a[iReq+1] (if it exists) of the sorted array and
** so it is not necessary to position any other elements. Or if iReq is
** negative, then the final array must be fully sorted.
*/
static void percentSort(double *a, unsigned int n){
static void percentSort(
double *a, /* Array to sort */
unsigned int n, /* Number of elements in array a[] */
int iReq /* Element caller cares about (or -ve) */
){
int iLt; /* Entries before a[iLt] are less than rPivot */
int iGt; /* Entries at or after a[iGt] are greater than rPivot */
int i; /* Loop counter */
@@ -3005,17 +3014,41 @@ static void percentSort(double *a, unsigned int n){
}
}while( i<iGt );
/* Recurse on the smaller partition only. The smaller partition
** will hold n/2 or fewer entries, which assures that the stack
** depth will not exceed O(log(n)), even for pathological cases.
** Loop without recursion for the larger partition. */
if( iLt>(int)(n/2) ){
if( n-iGt>=2 ) percentSort(a+iGt, n-iGt);
n = iLt;
assert( a[iLt]==rPivot );
assert( iGt>iLt );
if( iReq>=0 ){
/* In this case, the only elements that the caller requires sorted into
** the correct positions are elements a[iReq] and a[iReq+1]. At this
** point we know that element a[iLt] is in the correct position and
** all elements smaller than a[iLt] are in the left-hand partition.
** So if (iReq<iLt), then it is only necessary to sort the left
** partition.
**
** If (iReq>=iLt), then elements iReq and iReq+1 are either in the
** right partition or the equal partition (elements for which
** iLt<=iElem<iGt). Therefore it is always sufficient to sort only
** the right partition in this case. */
if( iReq<iLt ){
n = iLt;
}else{
a += iGt;
n -= iGt;
iReq = MAX(0, iReq-iGt);
}
}else{
if( iLt>=2 ) percentSort(a, iLt);
a += iGt;
n -= iGt;
/* Recurse on the smaller partition only. The smaller partition
** will hold n/2 or fewer entries, which assures that the stack
** depth will not exceed O(log(n)), even for pathological cases.
** Loop without recursion for the larger partition. */
if( iLt>(int)(n/2) ){
if( n-iGt>=2 ) percentSort(a+iGt, n-iGt, -1);
n = iLt;
}else{
if( iLt>=2 ) percentSort(a, iLt, -1);
a += iGt;
n -= iGt;
}
}
}while( n>=2 );
}
@@ -3052,7 +3085,7 @@ static void percentInverse(sqlite3_context *pCtx,int argc,sqlite3_value **argv){
}
if( p->bSorted==0 ){
assert( p->nUsed>1 );
percentSort(p->a, p->nUsed);
percentSort(p->a, p->nUsed, -1);
p->bSorted = 1;
}
p->bKeepSorted = 1;
@@ -3081,13 +3114,17 @@ static void percentCompute(sqlite3_context *pCtx, int bIsFinal){
if( p==0 ) return;
if( p->a==0 ) return;
if( p->nUsed ){
if( p->bSorted==0 ){
assert( p->nUsed>1 );
percentSort(p->a, p->nUsed);
p->bSorted = 1;
}
ix = p->rPct*(p->nUsed-1);
i1 = (unsigned)ix;
if( p->bSorted==0 ){
/* In cases where bIsFinal is non-zero, setting Percentile.bSorted
** after the percentSort() call here is not technically correct, as
** the array is not fully sorted. But in this case the object will be
** freed below anyway, so it doesn't matter. */
assert( p->nUsed>1 );
percentSort(p->a, p->nUsed, (bIsFinal ? (int)i1 : -1));
p->bSorted = 1;
}
if( settings & 1 ){
vx = p->a[i1];
}else{