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

[BACKEND] Porting the legacy heuristic rule in assigning shared layout for A/B of MMAv1 (#948)

Browse Source
This commit is contained in:
goostavz
2022-12-06 03:30:23 +08:00
committed by GitHub
parent 99c7e0e008
commit e057c65cf0
2 changed files with 8 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
include/triton/Dialect/TritonGPU/IR/TritonGPUAttrDefs.td
View File

@@ -92,12 +92,16 @@ A_{3, 2} A_{3, 3} A_{3, 0} A_{3, 1} ... [phase 1] /
unsigned inner = (opIdx == 0) ? 0 : 1;
// ---- begin version 1 ----
// TODO: handle rep (see
// https://github.com/openai/triton/blob/master/lib/codegen/analysis/layout.cc#L209)
if (version == 1) {
bool is_row = order[0] != 0;
bool is_vec4 = opIdx == 0 ? is_row && (shape[order[0]] <= 16) :
!is_row && (shape[order[0]] <= 16);
int pack_size = opIdx == 0 ? ((is_row || is_vec4) ? 1 : 2) :
((is_row && !is_vec4) ? 2 : 1);
int rep = 2 * pack_size;
int maxPhase = (order[inner] == 1 ? 8 : 4) / perPhase;
return $_get(context, 1, perPhase, maxPhase, order);
}
return $_get(context, 2 * rep, perPhase, maxPhase, order);
}
// ---- begin version 2 ----
if (version == 2) {

2
lib/Conversion/TritonGPUToLLVM/TritonGPUToLLVM.cpp
View File

@@ -3035,7 +3035,6 @@ void ConvertLayoutOpConversion::processReplica(
currVal = zext(llvmElemTy, currVal);
else if (isPtr)
currVal = ptrtoint(llvmElemTy, currVal);
valVec = insert_element(vecTy, valVec, currVal, idx_val(v));
}
store(valVec, ptr);
@@ -3143,7 +3142,6 @@ LogicalResult ConvertLayoutOpConversion::lowerBlockedToShared(
triton::gpu::ConvertLayoutOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const {
auto loc = op.getLoc();
Value src = op.src();
Value dst = op.result();
auto srcTy = src.getType().cast<RankedTensorType>();