align computation with triton

intel · Jul 11, 2024 · 9ed0b84 · 9ed0b84
1 parent edc6cc0
commit 9ed0b84
Showing 1 changed file with 32 additions and 42 deletions.
diff --git a/lib/gc/Transforms/FlashAttentionConversion.cpp b/lib/gc/Transforms/FlashAttentionConversion.cpp
@@ -331,6 +331,12 @@ struct MHAToFlashAttention
                                 ValueRange{curSumSlice, rescaledPrevSumSlice},
                                 ValueRange{reducedShapeOut})
                             .getResult(0);
+    Value newSumSliceRecip =
+        rewriter
+            .create<linalg::ReciprocalOp>(loc, reducedShapeOut.getType(),
+                                          ValueRange{newSumSlice},
+                                          ValueRange{reducedShapeOut})
+            .getResult(0);
     SmallVector<int64_t> VShape{cfg.RowBlockSize, headDim};
     Value VShapeOut = rewriter.create<tensor::EmptyOp>(loc, VShape, dtype);
     Value matmulVOutFilled =
@@ -341,38 +347,40 @@ struct MHAToFlashAttention
                                       ValueRange{PSlice, collapsedVSlice},
                                       ValueRange{matmulVOutFilled})
             .getResult(0);
-    Value expMaxDiffBroadcasted =
+    Value newSumSliceRecipBroadcasted =
         rewriter
-            .create<linalg::BroadcastOp>(loc, expMaxDiff, VShapeOut,
+            .create<linalg::BroadcastOp>(loc, newSumSliceRecip, VShapeOut,
                                          SmallVector<int64_t>{1})
             .getResults()[0];
-    Value expMaxDiffBroadcastedEps =
+    Value rescaledPrevSumSliceBroadcasted =
         rewriter
-            .create<linalg::GenericOp>(
-                loc, VShapeOut.getType(), ValueRange{expMaxDiffBroadcasted},
-                ValueRange{VShapeOut}, indexingMaps,
-                SmallVector<utils::IteratorType>(2,
-                                                 utils::IteratorType::parallel),
-                [&](OpBuilder &nestedBuilder, Location nestedLoc,
-                    ValueRange args) {
-                  Value eps = nestedBuilder.create<arith::ConstantOp>(
-                      loc, nestedBuilder.getFloatAttr(dtype, 1e-9));
-                  Value added =
-                      nestedBuilder.create<arith::AddFOp>(loc, args[0], eps);
-                  nestedBuilder.create<linalg::YieldOp>(nestedLoc, added);
-                })
+            .create<linalg::BroadcastOp>(loc, rescaledPrevSumSlice, VShapeOut,
+                                         SmallVector<int64_t>{1})
+            .getResults()[0];
+    Value rescaledMatmulV =
+        rewriter
+            .create<linalg::MulOp>(
+                loc, matmulVOutFilled.getType(),
+                ValueRange{matmulV, newSumSliceRecipBroadcasted},
+                ValueRange{matmulVOutFilled})
+            .getResult(0);
+    Value sumSliceQuotient =
+        rewriter
+            .create<linalg::MulOp>(loc, matmulVOutFilled.getType(),
+                                   ValueRange{rescaledPrevSumSliceBroadcasted,
+                                              newSumSliceRecipBroadcasted},
+                                   ValueRange{matmulVOutFilled})
             .getResult(0);
     Value rescaledOSlice =
         rewriter
-            .create<linalg::DivOp>(
-                loc, VShapeOut.getType(),
-                ValueRange{prevOSlice, expMaxDiffBroadcastedEps},
-                ValueRange{VShapeOut})
+            .create<linalg::MulOp>(loc, matmulVOutFilled.getType(),
+                                   ValueRange{prevOSlice, sumSliceQuotient},
+                                   ValueRange{matmulVOutFilled})
             .getResult(0);
     Value newOSlice =
         rewriter
             .create<linalg::AddOp>(loc, VShapeOut.getType(),
-                                   ValueRange{rescaledOSlice, matmulV},
+                                   ValueRange{rescaledOSlice, rescaledMatmulV},
                                    ValueRange{VShapeOut})
             .getResult(0);
     // yield all the results of the innermost loop.
@@ -381,25 +389,7 @@ struct MHAToFlashAttention
     // yield rowBlockLoop results
     rewriter.setInsertionPointToEnd(rowBlockLoop.getBody());
     auto innermostLoopResults = columnBlockLoop->getResults();
-    Value OSliceFinal = innermostLoopResults[0],
-          sumSliceFinal = innermostLoopResults[2];
-    Value sliceShapeOut =
-        rewriter.create<tensor::EmptyOp>(loc, reducedShape, dtype);
-    Value broadcastedSliceShapeOut =
-        rewriter.create<tensor::EmptyOp>(loc, VShape, dtype);
-    Value sumSliceFinalBroadcasted =
-        rewriter
-            .create<linalg::BroadcastOp>(loc, sumSliceFinal,
-                                         broadcastedSliceShapeOut,
-                                         SmallVector<int64_t>{1})
-            .getResults()[0];
-    Value rescaledOSliceFinal =
-        rewriter
-            .create<linalg::DivOp>(
-                loc, broadcastedSliceShapeOut.getType(),
-                ValueRange{OSliceFinal, sumSliceFinalBroadcasted},
-                ValueRange{broadcastedSliceShapeOut})
-            .getResult(0);
+    Value OSliceFinal = innermostLoopResults[0];
     SmallVector<OpFoldResult> outputOffsets;
     outputOffsets.push_back(getAsOpFoldResult(ivs[0]));
     outputOffsets.push_back(getAsOpFoldResult(ivs[1]));
@@ -409,8 +399,8 @@ struct MHAToFlashAttention
     outputSizes[2] = rewriter.getIndexAttr(cfg.RowBlockSize);
     outputSizes[3] = rewriter.getIndexAttr(headDim);
     Value insertedRescaledOSlice = rewriter.create<tensor::InsertSliceOp>(
-        loc, rescaledOSliceFinal, rowBlockLoop.getRegionIterArgs()[0],
-        outputOffsets, outputSizes, strides);
+        loc, OSliceFinal, rowBlockLoop.getRegionIterArgs()[0], outputOffsets,
+        outputSizes, strides);
     rewriter.create<scf::YieldOp>(loc, ValueRange{insertedRescaledOSlice});
     // Add the scf.yield operations for all the outer loops.
     for (auto [outerLoop, innerLoop] :