Fuse batch normalization into convolution weights

Author: mvpant

stablehlo/testdata/bn_conv_fuse_float32.mlir

@@ -21,6 +21,7 @@
 #include "llvm/ADT/SmallVector.h"
 #include "llvm/ADT/SmallVectorExtras.h"
 #include "llvm/Support/ErrorHandling.h"
+#include "mlir/Dialect/Utils/IndexingUtils.h"
 #include "mlir/IR/Attributes.h"
 #include "mlir/IR/Block.h"
 #include "mlir/IR/Builders.h"
@@ -1467,6 +1468,154 @@ struct ReorderElementwiseAndShapeOp final
   }
 };
 
+// Fuses batch normalization operation with convolution weight:
+// X = conv(input, weight)
+// Y = batch_norm_inference(X, ...)
+// into ->
+// X = conv(input, weight(new))
+// Y = add(X, broadcast_in_dim(Bias(new)))
+//
+struct FuseConvolutionBatchNormalization final
+    : OpRewritePattern<BatchNormInferenceOp> {
+  using OpRewritePattern::OpRewritePattern;
+
+  LogicalResult matchAndRewrite(BatchNormInferenceOp op,
+                                PatternRewriter &rewriter) const override {
+    auto bnOperandType = op.getOperand().getType();
+    auto bnOperandShape = bnOperandType.getShape();
+    auto bnResultType = op.getResult().getType();
+    uint64_t bnFeatureIndex = op.getFeatureIndex();
+
+    auto convOp = op.getOperand().getDefiningOp<ConvolutionOp>();
+    if (!convOp) return failure();
+
+    auto convWeight = convOp.getRhs();
+    auto convWeightType = convWeight.getType();
+    auto convWeightShape = convWeightType.getShape();
+
+    auto dimNumbers = convOp.getDimensionNumbers();
+    if (dimNumbers.getInputBatchDimension() != 0 ||
+        dimNumbers.getInputFeatureDimension() != 1 ||
+        dimNumbers.getOutputBatchDimension() != 0 ||
+        dimNumbers.getOutputFeatureDimension() != 1 ||
+        dimNumbers.getKernelOutputFeatureDimension() != 0 ||
+        dimNumbers.getKernelInputFeatureDimension() != 1)
+      return rewriter.notifyMatchFailure(convOp,
+                                         "Only [b, f, ...]x[o, i, ...]->[b, f, "
+                                         "...] configuration is supported");
+
+    if (convOp.getFeatureGroupCount() > 1 || convOp.getBatchGroupCount() > 1)
+      return rewriter.notifyMatchFailure(
+          convOp, "feature or batch grouping is not supported");
+
+    if (bnOperandShape[bnFeatureIndex] != convWeightShape.front())
+      return failure();
+
+    DenseFPElementsAttr convWeightElems;
+    DenseFPElementsAttr scaleElems;
+    DenseFPElementsAttr offsetElems;
+    DenseFPElementsAttr meanElems;
+    DenseFPElementsAttr varianceElems;
+
+    auto epsilon = op.getEpsilon();
+
+    if (!matchPattern(convWeight, m_Constant(&convWeightElems)))
+      return rewriter.notifyMatchFailure(
+          op, "expected constant convolution weight");
+
+    if (!matchPattern(op.getScale(), m_Constant(&scaleElems)) ||
+        !matchPattern(op.getOffset(), m_Constant(&offsetElems)) ||
+        !matchPattern(op.getMean(), m_Constant(&meanElems)) ||
+        !matchPattern(op.getVariance(), m_Constant(&varianceElems)))
+      return failure();
+
+    const auto &convWeightSemantics =
+        cast<FloatType>(convWeightType.getElementType()).getFloatSemantics();
+
+    // W(new) = W(old) * gamma * rsqrt(variance + epsilon)
+    // B(new) = (B(old) - mean) * rsqrt(variance + epsilon) * gamma + betta
+    // where: gamma - scaling factor
+    //        betta - shifting factor
+    //        rsqrt - reciprocal square root function
+    //        W - weight
+    //        B - bias
+    //
+    const SmallVector<double> multipliers = llvm::map_to_vector(
+        llvm::zip_equal(varianceElems, scaleElems),
+        [&epsilon](const std::tuple<APFloat, APFloat> &pack) -> double {
+          const auto &[variance, scale] = pack;
+          auto varEps = (variance + epsilon).convertToDouble();
+          auto rsqrt = 1.0 / std::sqrt(varEps);
+          return rsqrt * scale.convertToDouble();
+        });
+
+    SmallVector<APFloat> newWeight;
+    newWeight.reserve(convWeightType.getNumElements());
+
+    const size_t outFeatureTileSize =
+        computeProduct(convWeightShape.drop_front());
+    auto it = convWeightElems.begin();
+    for (const auto &multiplier : multipliers) {
+      for (size_t i = 0; i < outFeatureTileSize; ++i) {
+        double v = (*it).convertToDouble() * multiplier;
+        APFloat result(v);
+        bool losesInfo;
+        if (APFloat::opStatus::opInvalidOp ==
+            result.convert(convWeightSemantics, APFloat::rmNearestTiesToEven,
+                           &losesInfo))
+          return failure();
+        newWeight.push_back(result);
+        ++it;
+      }
+    }
+
+    SmallVector<APFloat> biasValues;
+    biasValues.reserve(multipliers.size());
+
+    for (const auto &[off, multiplier, mean] :
+         llvm::zip_equal(offsetElems, multipliers, meanElems)) {
+      // stablehlo convolution operation doesn't have a builtin bias
+      double convBias = 0;
+      double v = (convBias - mean.convertToDouble()) * multiplier +
+                 off.convertToDouble();
+      APFloat result(v);
+
+      bool losesInfo;
+      if (APFloat::opStatus::opInvalidOp ==
+          result.convert(convWeightSemantics, APFloat::rmNearestTiesToEven,
+                         &losesInfo))
+        return failure();
+
+      biasValues.push_back(result);
+    }
+
+    rewriter.setInsertionPoint(op);
+    auto newConvWeight = rewriter.create<ConstantOp>(
+        convWeight.getLoc(), convWeightType,
+        DenseFPElementsAttr::get(convWeightType, newWeight));
+
+    // Keep old convolution as it might have other users
+    auto newConvOp = rewriter.create<ConvolutionOp>(
+        convOp.getLoc(), convOp->getResultTypes(),
+        ValueRange{convOp.getLhs(), newConvWeight}, convOp->getAttrs());
+
+    SmallVector<int64_t> biasShape{static_cast<int64_t>(biasValues.size())};
+    auto biasType =
+        convWeightType.cloneWith(biasShape, convWeightType.getElementType());
+    auto bias = rewriter.create<ConstantOp>(
+        op.getLoc(), biasType, DenseFPElementsAttr::get(biasType, biasValues));
+
+    auto indices =
+        rewriter.getDenseI64ArrayAttr({static_cast<int64_t>(bnFeatureIndex)});
+    auto bcast = rewriter.create<BroadcastInDimOp>(op.getLoc(), bnResultType,
+                                                   bias, indices);
+    auto add = rewriter.create<AddOp>(op.getLoc(), newConvOp, bcast);
+
+    rewriter.replaceOp(op, add);
+    return success();
+  }
+};
+
 struct StablehloAggressiveSimplificationPass final
     : impl::StablehloAggressiveSimplificationPassBase<
           StablehloAggressiveSimplificationPass> {
@@ -1513,6 +1662,8 @@ void populateStablehloCanonicalizationPatterns(MLIRContext *context,
   patterns
       ->add<GetDimensionSizeOpCanon, DynamicBroadcastInDimOpNotActuallyDynamic,
             DynamicReshapeOpIsStatic, DynamicIotaIsStatic>(context);
+
+  patterns->add<FuseConvolutionBatchNormalization>(context);
 }
 
 }  // namespace stablehlo