gaspersic/triton
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

[BACKEND] Making the warp-level tile "more square" to increase data-reuse for tl.dot. (#442)

Browse Source 
* Increase smem data-reuse for some layouts

* tweak

* Keep the original tiling logic for sm < 80

Co-authored-by: Philippe Tillet <phil@openai.com>
This commit is contained in:
daadaada
2022-01-28 01:59:54 +08:00
committed by GitHub
parent 59d371c6eb
commit e68d6a7776
1 changed files with 29 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files

38
lib/codegen/analysis/layout.cc
View File

@@ -217,16 +217,36 @@ mma_layout::mma_layout(size_t num_warps,
order_ = {0, 1}; order_ = {0, 1};
/* warps per tile */ /* warps per tile */
// try to make things as square as possible to maximize data re-use
wpt_ = {1, 1, 1}; wpt_ = {1, 1, 1};
std::vector<int> wpt_nm1; // try to make warp-level tiles as square as possible to maximize data re-use
do{ if (tgt->as_nvidia()->sm() < 80) {
wpt_nm1 = wpt_; std::vector<int> wpt_nm1;
if(wpt_[0] * wpt_[1] * wpt_[2] < num_warps) do{
wpt_[0] = clamp(wpt_[0]*2, 1, shape_[0] / spw_[0]); wpt_nm1 = wpt_;
if(wpt_[0] * wpt_[1] * wpt_[2] < num_warps) if(wpt_[0] * wpt_[1] * wpt_[2] < num_warps)
wpt_[1] = clamp(wpt_[1]*2, 1, shape_[1] / spw_[1]); wpt_[0] = clamp(wpt_[0]*2, 1, shape_[0] / spw_[0]);
}while(wpt_nm1 != wpt_); if(wpt_[0] * wpt_[1] * wpt_[2] < num_warps)
wpt_[1] = clamp(wpt_[1]*2, 1, shape_[1] / spw_[1]);
}while(wpt_nm1 != wpt_);
} else {
bool changed = false;
do {
changed = false;
if (wpt_[0] * wpt_[1] * wpt_[2] >= num_warps)
break;
if (shape_[0] / spw_[0] / wpt_[0] >= shape_[1] / (spw_[1]*2) / wpt_[1]) {
if (wpt_[0] < shape_[0] / spw_[0]) {
wpt_[0] *= 2;
changed = true;
}
} else {
if (wpt_[1] < shape_[1] / (spw_[1]*2)) {
wpt_[1] *= 2;
changed = true;
}
}
} while (changed);
}
/* shape per block */ /* shape per block */
shape_per_cta_ = {spw_[0]*wpt_[0], spw_[1]*wpt_[1], 1}; shape_per_cta_ = {spw_[0]*wpt_[0], spw_[1]*wpt_[1], 1};