TempleOSTS_140317.ISO

This commit is contained in:
Terry A. Davis authored and minexew committed 2018-09-08 21:46:02 +02:00
1 parent 2d2a65d572
commit 7ef616a20e
1416 files changed
+223873 -53917

No files matched your search

+164
View File
@@ -0,0 +1,164 @@
#define NUM 1000000
I64 my_mp_cnt=1<<Bsr(mp_cnt);//Power of 2
I32 *a1,*a2;
I32 *b[my_mp_cnt],bn[my_mp_cnt];
I64 mp_not_done_flags;
I64 Compare(I32 *e1,I32 *e2)
{
return *e1-*e2;
}
U0 QSortU32(I32 *base,I64 num)
{
// By customizing, we dramatically improve it!
// Cut and paste from $LK,"QSortI64",A="MN:QSortI64"$().
I64 i;
I32 *less,*greater,pivot;
if (num>1) {
do {
less=base;
greater=base+num;
pivot=base[num/2];
while (less<greater) {
if (*less<=pivot)
less++;
else {
greater--;
SwapU32(less,greater);
}
}
i=less-base;
if (i==num) { //All less or equal to pivot
//Point greater to first less
do greater--;
while (--i && *greater==pivot);
if (i) {
less=base+num/2; //Pivot was not moved, point to it
if (less<greater)
SwapU32(less,greater);
num=i;
} else //All equal
break;
} else if (i<num/2) {
QSortU32(base,i);
num-=i;
base=greater;
} else {
QSortU32(greater,num-i);
num=i;
}
} while (num>1);
}
}
U0 MPSort(I64 dummy=0)
{
nounusedwarn dummy;
QSortU32(b[Gs->num],bn[Gs->num]);
LBtr(&mp_not_done_flags,Gs->num);
}
U0 MPRadixSortDemo(I64 dummy=0)
{
nounusedwarn dummy;
I64 i,j,k1,k2;
F64 t0;
a1=MAlloc(NUM*sizeof(I32));
for (i=0;i<NUM;i++)
a1[i]=RandI32;
a2=MAlloc(NUM*sizeof(I32));
"$$GREEN$$QSort$$FG$$\n";
t0=tS;
MemCpy(a2,a1,sizeof(I32)*NUM);
QSort(a2,NUM,sizeof(I32),&Compare);
"Time:%9.6f\n",tS-t0;
D(a2+NUM/4);
"$$GREEN$$QSortU32$$FG$$\n";
t0=tS;
MemCpy(a2,a1,sizeof(I32)*NUM);
QSortU32(a2,NUM);
"Time:%9.6f\n",tS-t0;
D(a2+NUM/4);
for (i=0;i<my_mp_cnt;i++) {
//$WW,0$We must do full size, just in case.
//There will be uneven split between cores
//depending on the distribution of rand numbers.
b[i]=MAlloc(NUM*sizeof(I32));
bn[i]=0;
}
if (my_mp_cnt<2) throw('MultCore');
"$$GREEN$$MP Radix QSortU32$$FG$$\n";
t0=tS;
k1=32-Bsr(my_mp_cnt);
k2=my_mp_cnt/2;
for (i=0;i<NUM;i++) {
j=a1[i]>>k1+k2; //This is a preliminary radix sort.
b[j][bn[j]++]=a1[i];
}
mp_not_done_flags=1<<my_mp_cnt-1;
for (i=0;i<my_mp_cnt;i++)
Spawn(&MPSort,NULL,NULL,i);
while (mp_not_done_flags)
Yield;
j=0;
for (i=0;i<my_mp_cnt;i++) {
MemCpy(&a2[j],b[i],bn[i]*sizeof(I32));
j+=bn[i];
}
"Time:%9.6f\n",tS-t0;
D(a2+NUM/4);
Free(a1);
Free(a2);
for (i=0;i<my_mp_cnt;i++)
Free(b[i]);
}
MPRadixSortDemo;
/*$HL,0$ Results on 8 Cores 2.660GHz Core i7:
$FG,2$QSort$FG$
Time: 1.535474
00000000 EA 1D F6 BF 16 2E F6 BF 0B 3C F6 BF CD 44 F6 BF ê.ö¿..ö¿.<ö¿ÍDö¿
00000010 B1 45 F6 BF 52 4A F6 BF 96 71 F6 BF 9F 78 F6 BF ±Eö¿RJö¿–qö¿Ÿxö¿
00000020 52 87 F6 BF 23 9F F6 BF EF AD F6 BF 2D B4 F6 BF R‡ö¿#Ÿö¿ï­ö¿-´ö¿
00000030 4B EE F6 BF 60 FA F6 BF 31 08 F7 BF 1A 0E F7 BF Kîö¿`úö¿1.÷¿..÷¿
00000040 5C 39 F7 BF 26 3B F7 BF C7 3D F7 BF CF 3D F7 BF \9÷¿&;÷¿Ç=÷¿Ï=÷¿
00000050 7D 51 F7 BF 2F 79 F7 BF 7B 7A F7 BF CF 8C F7 BF }Q÷¿/y÷¿{z÷¿ÏŒ÷¿
00000060 09 95 F7 BF 04 A3 F7 BF D2 A8 F7 BF 61 EA F7 BF .•÷¿.£÷¿Ò¨÷¿aê÷¿
00000070 2D ED F7 BF F2 03 F8 BF 4C 09 F8 BF BB 09 F8 BF -í÷¿ò.ø¿L.ø¿».ø¿
$FG,2$QSortU32$FG$
Time: 0.132599
00000000 EA 1D F6 BF 16 2E F6 BF 0B 3C F6 BF CD 44 F6 BF ê.ö¿..ö¿.<ö¿ÍDö¿
00000010 B1 45 F6 BF 52 4A F6 BF 96 71 F6 BF 9F 78 F6 BF ±Eö¿RJö¿–qö¿Ÿxö¿
00000020 52 87 F6 BF 23 9F F6 BF EF AD F6 BF 2D B4 F6 BF R‡ö¿#Ÿö¿ï­ö¿-´ö¿
00000030 4B EE F6 BF 60 FA F6 BF 31 08 F7 BF 1A 0E F7 BF Kîö¿`úö¿1.÷¿..÷¿
00000040 5C 39 F7 BF 26 3B F7 BF C7 3D F7 BF CF 3D F7 BF \9÷¿&;÷¿Ç=÷¿Ï=÷¿
00000050 7D 51 F7 BF 2F 79 F7 BF 7B 7A F7 BF CF 8C F7 BF }Q÷¿/y÷¿{z÷¿ÏŒ÷¿
00000060 09 95 F7 BF 04 A3 F7 BF D2 A8 F7 BF 61 EA F7 BF .•÷¿.£÷¿Ò¨÷¿aê÷¿
00000070 2D ED F7 BF F2 03 F8 BF 4C 09 F8 BF BB 09 F8 BF -í÷¿ò.ø¿L.ø¿».ø¿
$FG,2$MP Radix QSortU32$FG$
Time: 0.024471
00000000 EA 1D F6 BF 16 2E F6 BF 0B 3C F6 BF CD 44 F6 BF ê.ö¿..ö¿.<ö¿ÍDö¿
00000010 B1 45 F6 BF 52 4A F6 BF 96 71 F6 BF 9F 78 F6 BF ±Eö¿RJö¿–qö¿Ÿxö¿
00000020 52 87 F6 BF 23 9F F6 BF EF AD F6 BF 2D B4 F6 BF R‡ö¿#Ÿö¿ï­ö¿-´ö¿
00000030 4B EE F6 BF 60 FA F6 BF 31 08 F7 BF 1A 0E F7 BF Kîö¿`úö¿1.÷¿..÷¿
00000040 5C 39 F7 BF 26 3B F7 BF C7 3D F7 BF CF 3D F7 BF \9÷¿&;÷¿Ç=÷¿Ï=÷¿
00000050 7D 51 F7 BF 2F 79 F7 BF 7B 7A F7 BF CF 8C F7 BF }Q÷¿/y÷¿{z÷¿ÏŒ÷¿
00000060 09 95 F7 BF 04 A3 F7 BF D2 A8 F7 BF 61 EA F7 BF .•÷¿.£÷¿Ò¨÷¿aê÷¿
00000070 2D ED F7 BF F2 03 F8 BF 4C 09 F8 BF BB 09 F8 BF -í÷¿ò.ø¿L.ø¿».ø¿
$HL,1$*/