Raise Normalization for Pytorch (torch-mlir) Layer norm.

This Adds a Raise normalization pattern that matches the tensorrt mlir
that is generated by torch-mlir when lowering the layer norm layer.

Signed-off-by: Matthew Francis-Landau <[email protected]>

add utils function of getSplatConstantElementAttribute

Signed-off-by: Matthew Francis-Landau <[email protected]>

Author: Matthew Francis-Landau

mlir-tensorrt/tensorrt/include/mlir-tensorrt-dialect/TensorRT/IR/TensorRTOps.td

@@ -401,6 +401,7 @@ def TensorRT_CastOp : TensorRT_Op<"cast", [
   let arguments = (ins TensorRT_RankedTensorOf<[I1, UI8, TensorRT_I8, I32, I64, F16, BF16, F32]>:$input);
   let results = (outs TensorRT_RankedTensorOf<[I1, UI8, TensorRT_I8, I32, I64, F16, BF16, F32]>:$result);
   let assemblyFormat = "attr-dict $input `:` type($input) `to` type($result)";
+  let hasFolder = 1;
 
   let extraClassDeclaration = [{
     /// Returns true if created op is valid for TensorRT major version.

mlir-tensorrt/tensorrt/include/mlir-tensorrt-dialect/TensorRT/Utils/Utils.h

@@ -68,6 +68,11 @@ TypedValue<RankedTensorType>
 scatterShapeTensor(RewriterBase &b, Location loc, ArrayRef<int64_t> baseShape,
                    int32_t scatterDim, TypedValue<RankedTensorType> update);
 
+/// Get a splatted constant's attribute by going up a chain of reshape and cast
+/// operations to find the original constant.  The constant can be a different
+/// data type if there is a cast operation in the chain.
+FailureOr<Attribute> getSplatConstantElementAttribute(Value x);
+
 } // namespace tensorrt
 } // namespace mlir
 

mlir-tensorrt/tensorrt/lib/TensorRT/IR/TensorRT.cpp

@@ -2380,6 +2380,17 @@ OpFoldResult IdentityOp::fold(FoldAdaptor adaptor) {
   return foldIdentity(getType(), getInput(), adaptor);
 }
 
+//===----------------------------------------------------------------------===//
+// CastOp
+//===----------------------------------------------------------------------===//
+
+OpFoldResult CastOp::fold(FoldAdaptor adaptor) {
+  if (getInput().getType() == getType()) {
+    return getInput();
+  }
+  return nullptr;
+}
+
 //===----------------------------------------------------------------------===//
 // ReduceOp
 //===----------------------------------------------------------------------===//

mlir-tensorrt/tensorrt/lib/TensorRT/Transforms/RaiseNormalizations.cpp

@@ -23,6 +23,7 @@
 //===----------------------------------------------------------------------===//
 #include "mlir-tensorrt-dialect/TensorRT/IR/TensorRTDialect.h"
 #include "mlir-tensorrt-dialect/TensorRT/Transforms/Passes.h"
+#include "mlir-tensorrt-dialect/TensorRT/Utils/Utils.h"
 #include "mlir/Dialect/PDL/IR/PDL.h"
 #include "mlir/Dialect/PDLInterp/IR/PDLInterp.h"
 #include "mlir/IR/Matchers.h"
@@ -51,6 +52,8 @@ class RaiseNormalizations
     MLIRContext *ctx = &getContext();
     RewritePatternSet patterns(ctx);
     patterns.add<RaiseInstanceNormalization_NCHW>(ctx);
+    patterns.add<RaisePytorchLayerNorm>(ctx);
+    patterns.add<RemoveLayerNormCast>(ctx);
 
     if (failed(applyPatternsGreedily(getOperation(), std::move(patterns)))) {
       emitError(getOperation()->getLoc())

mlir-tensorrt/tensorrt/lib/TensorRT/Transforms/RaiseNormalizations.pdll

@@ -146,6 +146,10 @@ Constraint ReduceSumImpl(val: Value)[{
               (reduceOp.getInput().getType().getRank() - 1)));
 }];
 
+Constraint AvgImpl(op: Op) [{
+    return success(cast<tensorrt::ReduceOp>(op).getReduceOperation() == ReduceOperation::kAVG);
+}];
+
 Constraint CheckRank4(val: Value)[{
  RankedTensorType rtt = cast<RankedTensorType>(val.getType());
  return success(rtt.getRank() == 4);
@@ -196,10 +200,10 @@ Constraint ReverseSqrt(val : Value) -> Value{
 }
 
 Constraint FlattenTailDims(val: Value) -> Value {
- CheckRank4(val);
- let reshapeRes = op<tensorrt.reshape>(val);
- FlattenConstraintImpl(reshapeRes);
- return reshapeRes;
+    CheckRank4(val);
+    let reshapeRes = op<tensorrt.reshape>(val);
+    FlattenConstraintImpl(reshapeRes);
+    return reshapeRes;
 }
 
 Constraint ReduceSum(val: Value) -> Value{
@@ -219,6 +223,35 @@ Constraint Mean(input: Value, numHW: Value){
  return Div(ExpandTailDims(ReduceSum(FlattenTailDims(input))), numHW);
 }
 
+Constraint ReduceAvg(input: Value, reduceAxes: Attr) {
+    let avgOp = op<tensorrt.reduce>(input) {keepDimensions = attr<"true">, reduceAxes = reduceAxes};
+    AvgImpl(avgOp);
+    return avgOp;
+}
+
+
+Rewrite GetSplatElementAttr(x: Value) -> Attr [{
+  return *getSplatConstantElementAttribute(x);
+}];
+
+Constraint HasSplatElements(x: Value) [{
+  return LogicalResult(getSplatConstantElementAttribute(x));
+}];
+
+Constraint SameElementType(a: Value, b: Value) [{
+    return success(cast<RankedTensorType>(a.getType()).getElementType() == cast<RankedTensorType>(b.getType()).getElementType());
+}];
+
+Rewrite CreateCast(x: Value, refValue: Value) -> Value [{
+    Type retType = RankedTensorType::Builder(cast<RankedTensorType>(x.getType())).setElementType(cast<RankedTensorType>(refValue.getType()).getElementType());
+    return rewriter.createOrFold<tensorrt::CastOp>(
+        x.getLoc(),
+        retType,
+        x
+    );
+}];
+
+
 Pattern RaiseInstanceNormalization_NCHW {
    let inputType : Type;
    let input : Value<inputType>;
@@ -240,3 +273,57 @@ Pattern RaiseInstanceNormalization_NCHW {
    CheckRank4(addOffset);
   replace addOffset with op<tensorrt.normalization>(input, scale, offset){axis = attr<"array<i64: 2,3>">};
 }
+
+Pattern RaisePytorchLayerNorm {
+    let x: Value;
+    let beta: Value;
+    let gamma: Value;
+    let axis: Attr;
+    let epsilon: Value;
+
+    let mean = ReduceAvg(x, axis);
+    let diffMean = Sub(x, mean);
+
+    let varianceDenominator: Value;
+    let varianceMean = Div(ReduceSum(x), varianceDenominator);  // for some reason Pytorch's lowering computes the mean in 2 different ways....
+    let varianceDiff = Sub(x, varianceMean);
+    let varianceDiffSquared = Mul(varianceDiff, varianceDiff);
+    let varianceNumerator = ReduceSum(varianceDiffSquared);
+    let variance = Div(varianceNumerator, varianceDenominator);
+    let varianceEps = Add(variance, epsilon);
+
+    let inverseSqrt = ReverseSqrt(varianceEps);
+    let normed = Mul(diffMean, inverseSqrt);
+    let prod = Mul(normed, gamma);
+    let root = Add(prod, beta);
+
+    HasSplatElements(epsilon);
+    HasSplatElements(varianceDenominator);
+
+    rewrite root with {
+        let epsilonAttr = GetSplatElementAttr(epsilon);
+        let replacement = op<tensorrt.normalization>(x, gamma, beta) {axis = axis, eps = epsilonAttr};
+        replace root with replacement;
+    };
+}
+
+Pattern RemoveLayerNormCast {
+    let x: Value;
+    let gamma: Value;
+    let beta: Value;
+    let axis: Attr;
+    let epsilonAttr: Attr;
+
+    let castInput = op<tensorrt.cast>(x);
+    let norm = op<tensorrt.normalization>(castInput, gamma, beta) {axis = axis, eps = epsilonAttr};
+    let root = op<tensorrt.cast>(norm);
+
+    SameElementType(x, root);
+
+    rewrite root with {
+        let newGamma = CreateCast(gamma, x);
+        let newBeta = CreateCast(beta, x);
+        let replacement = op<tensorrt.normalization>(x, newGamma, newBeta) {axis = axis, eps = epsilonAttr};
+        replace root with replacement;
+    };
+}

mlir-tensorrt/tensorrt/lib/TensorRT/Utils/Utils.cpp

@@ -22,6 +22,7 @@
 #include "mlir-tensorrt-dialect/TensorRT/Utils/Utils.h"
 
 #include "mlir-tensorrt-dialect/TensorRT/IR/TensorRTDialect.h"
+#include "mlir/IR/Matchers.h"
 #include "mlir/Interfaces/FunctionInterfaces.h"
 
 using namespace mlir;
@@ -158,3 +159,34 @@ tensorrt::scatterShapeTensor(RewriterBase &b, Location loc,
 
   return b.create<tensorrt::ConcatenationOp>(loc, parts, 0);
 }
+
+FailureOr<Attribute> tensorrt::getSplatConstantElementAttribute(Value x) {
+  while (true) {
+    if (auto expandRank = x.getDefiningOp<tensorrt::ExpandRankOp>())
+      x = expandRank.getInput();
+    else if (auto collapseRank = x.getDefiningOp<tensorrt::CollapseRankOp>())
+      x = collapseRank.getInput();
+    else if (auto reshape = x.getDefiningOp<tensorrt::ReshapeOp>())
+      x = reshape.getInput();
+    else if (auto broadcast = x.getDefiningOp<tensorrt::BroadcastOp>())
+      x = broadcast.getInput();
+    else if (auto cast = x.getDefiningOp<tensorrt::CastOp>())
+      x = cast.getInput();
+    else if (auto identity = x.getDefiningOp<tensorrt::IdentityOp>())
+      x = identity.getInput();
+    else if (auto slice = x.getDefiningOp<tensorrt::SliceOp>())
+      x = slice.getInput();
+    else if (auto constant = x.getDefiningOp<tensorrt::ConstantOp>()) {
+      SplatElementsAttr els{};
+      if (!matchPattern(x, m_Constant(&els)))
+        return failure();
+      Attribute value = els.getSplatValue<Attribute>();
+      if (!isa<FloatAttr, IntegerAttr>(value))
+        return failure();
+      return value;
+    } else {
+      return failure();
+    }
+  }
+  return failure();
+}

mlir-tensorrt/tensorrt/test/Dialect/TensorRT/raise-normalizations.mlir

@@ -53,3 +53,38 @@ func.func @raise_inst_norm_nchw(%arg0: tensor<1x3x1x1xf32>, %arg1: tensor<1x3x1x
 
 // CHECK-LABEL: @neg_raise_nhwc
 //   CHECK-NOT: tensorrt.normalization
+
+// -----
+
+// CHECK: @raise_layer_norm_pytorch(%[[arg0:.+]]: tensor<16x1024x1024xf32>)
+// CHECK: %[[ret:.+]] = tensorrt.normalization [[attr:.+]](%[[arg0]] : tensor<16x1024x1024xf32>, %[[gamma:.+]] : tensor<1x1x1024xf32>, %[[beta:.+]] : tensor<1x1x1024xf32>)
+// CHECK: return %[[ret]]
+func.func @raise_layer_norm_pytorch(%arg0: tensor<16x1024x1024xf32>) -> tensor<16x1024x1024xf32> {
+    %cst_i64 = tensorrt.constant dense<1024> : tensor<i64>
+    %cst_f32 = tensorrt.constant dense<9.99999974E-6> : tensor<1x1x1xf32>
+
+    %cst_bf16_1 = tensorrt.constant dense_resource<__elided__> : tensor<1024xbf16>  // beta (added)
+    %cst_bf16_2 = tensorrt.constant dense_resource<__elided__> : tensor<1024xbf16>   // gamma (multiplied)
+
+    %6 = tensorrt.reduce <kSUM> %arg0 {keepDimensions = true, reduceAxes = array<i64: 2>} : tensor<16x1024x1024xf32> -> tensor<16x1024x1xf32>
+    %7 = tensorrt.cast %cst_i64 : tensor<i64> to tensor<f32>
+    %8 = tensorrt.expand_rank %7 : tensor<f32> to tensor<1x1x1xf32>
+    %9 = tensorrt.element_wise <kDIV>(%6, %8 : tensor<16x1024x1xf32>, tensor<1x1x1xf32>) -> tensor<16x1024x1xf32>
+    %10 = tensorrt.element_wise <kSUB>(%arg0, %9 : tensor<16x1024x1024xf32>, tensor<16x1024x1xf32>) -> tensor<16x1024x1024xf32>
+    %11 = tensorrt.element_wise <kPROD>(%10, %10 : tensor<16x1024x1024xf32>, tensor<16x1024x1024xf32>) -> tensor<16x1024x1024xf32>
+    %12 = tensorrt.reduce <kSUM> %11 {keepDimensions = true, reduceAxes = array<i64: 2>} : tensor<16x1024x1024xf32> -> tensor<16x1024x1xf32>
+    %13 = tensorrt.element_wise <kDIV>(%12, %8 : tensor<16x1024x1xf32>, tensor<1x1x1xf32>) -> tensor<16x1024x1xf32>  // Var[x]
+    %15 = tensorrt.reduce <kAVG> %arg0 {keepDimensions = true, reduceAxes = array<i64: 2>} : tensor<16x1024x1024xf32> -> tensor<16x1024x1xf32>  // E[x]
+    %16 = tensorrt.element_wise <kSUM>(%13, %cst_f32 : tensor<16x1024x1xf32>, tensor<1x1x1xf32>) -> tensor<16x1024x1xf32> // Var[x] + epsilon
+    %17 = tensorrt.unary {unaryOperation = #tensorrt.unary_operation<kRECIP>} %16 : tensor<16x1024x1xf32>
+    %18 = tensorrt.unary {unaryOperation = #tensorrt.unary_operation<kSQRT>} %17 : tensor<16x1024x1xf32>    // compute 1/sqrt(...)
+    %19 = tensorrt.element_wise <kSUB>(%arg0, %15 : tensor<16x1024x1024xf32>, tensor<16x1024x1xf32>) -> tensor<16x1024x1024xf32>
+    %20 = tensorrt.element_wise <kPROD>(%19, %18 : tensor<16x1024x1024xf32>, tensor<16x1024x1xf32>) -> tensor<16x1024x1024xf32>  // multiply for division
+    %21 = tensorrt.cast %cst_bf16_2 : tensor<1024xbf16> to tensor<1024xf32>
+    %22 = tensorrt.expand_rank %21 : tensor<1024xf32> to tensor<1x1x1024xf32>
+    %23 = tensorrt.element_wise <kPROD>(%20, %22 : tensor<16x1024x1024xf32>, tensor<1x1x1024xf32>) -> tensor<16x1024x1024xf32> // multiply gamma
+    %24 = tensorrt.cast %cst_bf16_1 : tensor<1024xbf16> to tensor<1024xf32>
+    %25 = tensorrt.expand_rank %24 : tensor<1024xf32> to tensor<1x1x1024xf32>
+    %26 = tensorrt.element_wise <kSUM>(%23, %25 : tensor<16x1024x1024xf32>, tensor<1x1x1024xf32>) -> tensor<16x1024x1024xf32> // add beta
+    return %26 : tensor<16x1024x1024xf32>
+}