[BACKEND] Commenting out reduce fast path shared mem optimization for… (#1761)

… MMA layout that ended up being a cause for failure
This commit is contained in:
Zahi Moudallal
2023-06-08 14:39:38 -07:00
committed by GitHub
parent 9101736da1
commit fdf2bd5801

View File

@@ -60,9 +60,9 @@ SmallVector<SmallVector<unsigned>> ReduceOpHelper::getScratchConfigsFast() {
auto argLayout = getSrcLayout();
auto argLayoutMma = argLayout.dyn_cast<triton::gpu::MmaEncodingAttr>();
if (argLayoutMma && argLayoutMma.getVersionMajor() == 2 &&
triton::gpu::getWarpsPerCTA(argLayout)[axis] == 1)
return {{1, 1}, {1, 1}};
// if (argLayoutMma && argLayoutMma.getVersionMajor() == 2 &&
// triton::gpu::getWarpsPerCTA(argLayout)[axis] == 1)
// return {{1, 1}, {1, 1}};
/// shared memory block0
smemShapes[0] = convertType<unsigned>(getSrcShape());