@@ -834,18 +834,18 @@ LogicalResult combineLogicalObjectFifos(
834834 // will make an attempt to combine the logical objectFifos as per the
835835 // following algorithm :-
836836 // a. Combine i-th and i+1-th L3->L2 DmaCpyNd ops.
837- // b. Since step a would create a new L2 buffer (with combined shape), we
838- // will
839- // need to update the corresponding two L2->L1 Dma ops by indeed creating
840- // new ones. NOTE: Both of these new L2->L1 Dma ops will be reusing the
841- // same L1 buffers as well.
842- // c. Now pick the unique core ops corresponding to i-th and i+1-th L2->L1
843- // Dma
844- // ops and do the following :-
837+ // b. Form reusable L1 buffer by assigning the cumulative tiles of the
838+ // intended core ops.
839+ // c. Since step a would create a new L2 buffer (with combined shape), we
840+ // will need to update the corresponding two L2->L1 Dma ops by indeed
841+ // creating new ones. NOTE: Both of these new L2-> L1 Dma ops will be
842+ // reusing the same L1 buffers as well.
843+ // d. Now pick the unique core ops corresponding to i-th and i+1-th L2->L1
844+ // Dma ops and do the following :-
845845 // 1. For i-th CoreOp insert an AccessOp from the same L1 buffer towards
846- // the end.
846+ // the end.
847847 // 2. For i+1-th CoreOp insert an AccessOp from the same L1 buffer right
848- // before the corresponding AccessOp within the same CoreOp.
848+ // before the corresponding AccessOp within the same CoreOp.
849849 for (unsigned i = 0 , n = l3ToL2DmaOps.size (); i < n; i += 2 ) {
850850 // Step 1. Combine the picked L3->L2 DmaCpyNd pair.
851851 FailureOr<LogicalObjectFifoFromMemrefOp> maybeNewL2ObjectFifo =
@@ -855,14 +855,56 @@ LogicalResult combineLogicalObjectFifos(
855855 LogicalObjectFifoFromMemrefOp newL2ObjectFifo =
856856 maybeNewL2ObjectFifo.value ();
857857
858- // Step 2. We now have need to create two L2->L1 ops since the size has
858+ // Step 2. Form the reusable L1 buffer by assigning the cumulative tiles of
859+ // the intended core ops.
860+ LogicalObjectFifoFromMemrefOp reuseL1LogicalObjectFifoOp =
861+ l2ToL1DmaOps[i].getTargetObjectFifo ();
862+ SmallVector<Value> tiles;
863+ auto addNewTileFrom = [&](CoreOp coreOp) -> LogicalResult {
864+ OpBuilder::InsertionGuard guard (rewriter);
865+ TileOp tileOp = coreOp.getTileOp ();
866+ std::optional<int64_t > column = getConstantIntValue (tileOp.getCol ());
867+ std::optional<int64_t > row = getConstantIntValue (tileOp.getRow ());
868+ if (!column || !row) {
869+ return coreOp.emitOpError () << " has non-constant tile location"
870+ }
871+ rewriter.setInsertionPoint (reuseL1LogicalObjectFifoOp);
872+ auto colIndex = rewriter.
create <arith::ConstantIndexOp>(
873+ rewriter.getUnknownLoc (), *column);
874+ auto rowIndex = rewriter.
create <arith::ConstantIndexOp>(
875+ rewriter.getUnknownLoc (), *row);
876+ tileOp =
877+ rewriter.create <TileOp>(rewriter.getUnknownLoc (), colIndex, rowIndex);
878+ tiles.push_back (tileOp.getResult ());
879+ return success ();
880+ };
881+ std::optional<CoreOp> maybeFirstCoreOp = fetchUniqueCoreOp (l2ToL1DmaOps[i]);
882+ if (!maybeFirstCoreOp) return failure ();
883+ CoreOp firstCoreOp = maybeFirstCoreOp.value ();
884+ std::optional<CoreOp> maybeSecondCoreOp =
885+ fetchUniqueCoreOp (l2ToL1DmaOps[i + 1 ]);
886+ if (!maybeSecondCoreOp) return failure ();
887+ CoreOp secondCoreOp = maybeSecondCoreOp.value ();
888+ if (failed (addNewTileFrom (firstCoreOp)) ||
889+ failed (addNewTileFrom (secondCoreOp))) {
890+ return failure ();
891+ }
892+ llvm::sort (tiles.begin (), tiles.end (),
893+ AMDAIE::TileOp::tileValueColumnAndRowComparator);
894+ rewriter.setInsertionPoint (reuseL1LogicalObjectFifoOp);
895+ reuseL1LogicalObjectFifoOp =
896+ rewriter.replaceOpWithNewOp <LogicalObjectFifoFromMemrefOp>(
897+ reuseL1LogicalObjectFifoOp,
898+ cast<LogicalObjectFifoType>(
899+ reuseL1LogicalObjectFifoOp.getOutput ().getType ()),
900+ reuseL1LogicalObjectFifoOp.getMemref (), tiles);
901+
902+ // Step 3. We now have need to create two L2->L1 ops since the size has
859903 // changed. But for this we first need to find the new offset for L2 as
860904 // source.
861905 // TODO: For now I'm hardcoding the offsets but later it'd just depend
862906 // on combining/non-combining dimensions.
863907 // Offset = 0,0
864- LogicalObjectFifoFromMemrefOp reuseL1LogicalObjectFifoOp =
865- l2ToL1DmaOps[i].getTargetObjectFifo ();
866908 SmallVector<OpFoldResult> newL2AsSourceOffsets =
867909 l2ToL1DmaOps[i].getSourceMixedOffsets ();
868910 DmaCpyNdOp newFirstL2ToL1DmaOp = createL2ToL1ForReuse (
@@ -872,31 +914,24 @@ LogicalResult combineLogicalObjectFifos(
872914 // the first L2->L1 Dma.
873915 newL2AsSourceOffsets = l2ToL1DmaOps[i + 1 ].getSourceMixedOffsets ();
874916 newL2AsSourceOffsets[1 ] = rewriter.getIndexAttr (1 );
875- DmaCpyNdOp newSecondL2ToL1DmaOp = createL2ToL1ForReuse (
876- rewriter, l2ToL1DmaOps[i + 1 ], reuseL1LogicalObjectFifoOp,
877- newL2ObjectFifo, newL2AsSourceOffsets);
917+ createL2ToL1ForReuse (rewriter, l2ToL1DmaOps[i + 1 ],
918+ reuseL1LogicalObjectFifoOp, newL2ObjectFifo ,
919+ newL2AsSourceOffsets);
878920
879- // Step 3. PICK the CoreOps associated with the 1:1 L2->L1.
921+ // Step 4. Pick the CoreOps associated with the 1:1 L2->L1.
880922 // For the first Core op we'll insert Read at the end. It doesn't matter
881923 // for now so we're gonna insert it right before amdaie.end op.
882- std::optional<CoreOp> maybeFirstCoreOp =
883- fetchUniqueCoreOp (newFirstL2ToL1DmaOp);
884- if (!maybeFirstCoreOp) return failure ();
885- CoreOp firstCoreOp = maybeFirstCoreOp.value ();
886- firstCoreOp.walk ([&](AMDAIE::EndOp endOp) {
887- OpBuilder::InsertionGuard guard (rewriter);
888- // Hardcoding to `AMDAIE::MemoryAccess::Read`.
889- rewriter.setInsertionPoint (endOp);
890- rewriter.
create <AMDAIE::LogicalObjectFifoAccessOp>(
891- rewriter.getUnknownLoc (), reuseL1LogicalObjectFifoOp.getOutput (),
892- AMDAIE::MemoryAccess::Read);
924+ firstCoreOp.walk ([&](AMDAIE::LogicalObjectFifoAccessOp accessOp) {
925+ if (accessOp.getInput () == newFirstL2ToL1DmaOp.getTargetObjectFifo ()) {
926+ OpBuilder::InsertionGuard guard (rewriter);
927+ rewriter.setInsertionPointAfter (accessOp);
928+ rewriter.
create <AMDAIE::LogicalObjectFifoAccessOp>(
929+ rewriter.getUnknownLoc (), reuseL1LogicalObjectFifoOp.getOutput (),
930+ accessOp.getAccessType ());
931+ }
893932 });
894933 // For the second Core op we'll insert `Read` right before the first read
895934 // from the corresponding L1 logicalobjectFifo.
896- std::optional<CoreOp> maybeSecondCoreOp =
897- fetchUniqueCoreOp (newSecondL2ToL1DmaOp);
898- if (!maybeSecondCoreOp) return failure ();
899- CoreOp secondCoreOp = maybeSecondCoreOp.value ();
900935 secondCoreOp.walk ([&](AMDAIE::LogicalObjectFifoAccessOp accessOp) {
901936 if (accessOp.getInput () == l2ToL1DmaOps[i + 1 ].getTargetObjectFifo ()) {
902937 OpBuilder::InsertionGuard guard (rewriter);
0 commit comments