prenex/magyarsort
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

trying more ILP in spsort - not much success and will be reverted

Browse Source
This commit is contained in:
Richard Thier 2022-08-16 19:24:27 +02:00
parent f8793f71ab
commit c7e794b7ea
1 changed files with 32 additions and 6 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

38
space_partitioning_sort/spsort.h
View File

@ -133,11 +133,19 @@ inline uint32_t internal_mid(uint32_t low, uint32_t high) {
inline void spsort(uint32_t *t, int n, int m = 32); inline void spsort(uint32_t *t, int n, int m = 32);
/** Helper function that puts elements higher then mid to the top of the array and lower to the bottom. Returns number of bottoms */ /** Helper function that puts elements higher then mid to the top of the array and lower to the bottom. Returns number of bottoms */
inline int internal_array_separate(uint32_t *t, int n, uint32_t mid) { inline int internal_array_separate(uint32_t *t, int n, uint32_t mid, int bulk_xchg = 32) {
if(n > 0) { if(n > 0) {
// Two heads that also read & write (both) // Two heads that also read & write (both)
int left = 0; int left = 0;
int right = n - 1; int right = n - 1;
// These are needed for more ILP so that we can do the xchg operations in bulk
// and without data dependencies just do it in an unrolled loop from time to time!
std::vector<int> xchg_left(0);
std::vector<int> xchg_right(0);
xchg_left.reserve(bulk_xchg);
xchg_right.reserve(bulk_xchg);
while(left < right) { while(left < right) {
// Step over already good positioned values from left // Step over already good positioned values from left
while((left < right) && (t[left] < mid)) { while((left < right) && (t[left] < mid)) {
@ -152,12 +160,30 @@ inline int internal_array_separate(uint32_t *t, int n, uint32_t mid) {
// Extra check needed for edge-case! // Extra check needed for edge-case!
if(left < right) { if(left < right) {
// Both in wrong location - xchg them! // Both in wrong location - xchg them!
auto tmp = t[right]; // instead of doing it right here, collect them up! (*)
t[right] = t[left]; xchg_left.push_back(left);
t[left] = tmp; xchg_right.push_back(right);
++left;
--right;
} }
// See if we can do some bulk-exchange now (*)
// This loop the compiler should more easily unroll
// and the CPU should be able to schedule ILP-wise!
if(xchg_left.size() <= bulk_xchg) {
for(int i = 0; i < xchg_left.size(); ++i) {
auto tmp = t[xchg_left[i]];
t[xchg_left[i]] = t[xchg_right[i]];
t[xchg_right[i]] = tmp;
}
xchg_left.resize(0);
xchg_right.resize(0);
}
}
// Finish the remaining bulk exchanges (*)
for(int i = 0; i < xchg_left.size(); ++i) {
auto tmp = t[xchg_left[i]];
t[xchg_left[i]] = t[xchg_right[i]];
t[xchg_right[i]] = tmp;
} }
// Edge-case increment if single elem happens in middle in the end // Edge-case increment if single elem happens in middle in the end