Skip to content

Backward fold scale axis to Gemm layer in ONNX Dialect #3131

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
wants to merge 6 commits into
base: main
Choose a base branch
from
Open

Backward fold scale axis to Gemm layer in ONNX Dialect #3131

Show file tree
Hide file tree
Changes from 2 commits
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
0c9ff3c
Backward fold scale axis to gemm layer
Arkar-Hema Apr 18, 2025
f672ddf
Clang format fix
Arkar-Hema Apr 18, 2025
1426876
Backward fold batch to gemm
Arkar-Hema Apr 25, 2025
2c5c861
Merge branch 'main' into backward_fold_scale_to_gemm
AlexandreEichenberger Apr 29, 2025
9619505
Added mean, var and eps contraints
Arkar-Hema May 15, 2025
bfe06b5
Merge branch 'main' into backward_fold_scale_to_gemm
AlexandreEichenberger May 16, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
10 changes: 10 additions & 0 deletions src/Dialect/ONNX/ONNXOps/Canonicalize.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -90,6 +90,15 @@ bool isNotConvProducer(mlir::Value val) {
return true; // If no defining op, assume it's safe
}

bool isTransBFalse(mlir::Attribute attr) {
if (auto intAttr = attr.dyn_cast<mlir::IntegerAttr>()) {
int64_t val =
intAttr.getValue().getSExtValue(); // safe for signless integers
return val == 0; // return true if transB is false (0)
}
return false; // default fallback
}

// Get the index of the axis value in the given permutation array.
IntegerAttr getIndexOfAxisInPerm(
PatternRewriter &rewriter, ArrayAttr permAttr, IntegerAttr axis) {
Expand Down Expand Up @@ -1685,6 +1694,7 @@ void ONNXBatchNormalizationInferenceModeOp::getCanonicalizationPatterns(
results.insert<FuseBatchNormInferenceModeConvPattern>(context);
results.insert<RewriteBatchNormInferenceModeConvPattern1>(context);
results.insert<RewriteBatchNormInferenceModeConvPattern2>(context);
results.insert<BackwardFoldScaleAxisToGemmPattern>(context);
}

/// on the ONNXAddOp.
Expand Down
48 changes: 48 additions & 0 deletions src/Dialect/ONNX/ONNXOps/Canonicalize.td
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -925,6 +925,54 @@ def RewriteBatchNormInferenceModeConvPattern2: Pat<
[(HasRankOf<1> $x)], [], (addBenefit 0)
>;

//===----------------------------------------------------------------------===//
// This is to fold the composition: 'BatchNormalization o Gemm' into 'Gemm'
// by deriving new 'B' and 'C' for 'Gemm' operation after fusing scale and bias.
//
// Given:
// (Gemm) Z = A * B + C
// (BatchNorm) Y = scale * (Z - mean) / sqrt(var + eps) + bias
//
// This transformation corresponds to a recomposition:
// Y = A * (scale * B) + (scale * bias + C)
Copy link
Collaborator

@tungld tungld Apr 24, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@Arkar-Hema could you elaborate how you derived this formula where mean, var and eps are canceled?

Copy link
Contributor Author

@Arkar-Hema Arkar-Hema Apr 25, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Assume mean=0, var=1 and exp=0 (which is usually present in any pre-compiled and normalised models):

Y ≈ scale × Z + bias
Substituting Z = A x B + C
Y ≈ scale × (A × B + C) + bias
Y = A × (scale × B) + (scale × C + bias)

Copy link
Collaborator

@tungld tungld Apr 25, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Assume mean=0, var=1 and exp=0 (which is usually present in any pre-compiled and normalised models):

Then, you have to define this assumption in the constraint part of the rewriting rule. Otherwise, the rewriting rule produces a wrong result.

Anyway, my recommendation is to handle the general case where mean, var, eps are constants (not necessary concrete values, say, of 0, 1 and 0, respectively). New scale and bias values for matmul can be easily computed from mean, var, eps, scale and bias of BatchNorm, and in the inference mode, these values are constants and will be folded automatically by the compiler into a single constant.

//
// Therefore, we rewrite:
// onnx.BatchNormalizationInferenceMode(
// onnx.Gemm(A, B, C, alpha, beta, transA, transB),
// scale, bias, mean, var
// ) {epsilon = ..., momentum = ...}
//
// as:
// onnx.Gemm(
// A,
// onnx.Mul(B, scale),
// onnx.Add(onnx.Mul(bias, scale), C),
// alpha, beta, transA, transB)
//
// This transformation is only valid when transB = 0
// to maintain the correct computation shape alignment.
//
//===----------------------------------------------------------------------===//

def isTransBFalse : Constraint<CPred<
"onnx_mlir::isTransBFalse($0)">, "TransB is 1 not 0"
>;

def BackwardFoldScaleAxisToGemmPattern : Pat<
(ONNXBatchNormalizationInferenceModeOp:$res
(ONNXGemmOp $A, $B, $C, $alpha, $beta, $transA, $transB),
$scale, $bias, $_mean, $_var, $_epsilon, $_momentum),
(ONNXGemmOp
$A,
(ONNXMulOp $B, $scale),
(ONNXAddOp
(ONNXMulOp $bias, $scale),
$C),
(GemmAlpha), (GemmBeta), (GemmTransA), (GemmTransB)),
[(isTransBFalse $transB)],
[], (addBenefit 1)
>;

//===----------------------------------------------------------------------===//
// Canonicalization for ONNXShapeOp
//===----------------------------------------------------------------------===//
Expand Down
26 changes: 26 additions & 0 deletions test/mlir/onnx/onnx_canonicalization.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -784,6 +784,32 @@ func.func @test_rewrite_batchnormtestmode_1d_f16(%arg0 : tensor<64xf16>, %scale

// -----

func.func @test_backward_fold_scale_axis(%arg0: tensor<1x256xf32>) -> tensor<1x128xf32> {
%0 = onnx.Constant dense<0.00999999977> : tensor<256x128xf32>
%1 = onnx.Constant dense<0.00999999977> : tensor<128xf32>
%2 = onnx.Constant dense<0.00999999977> : tensor<128xf32>
%3 = onnx.Constant dense<0.00999999977> : tensor<128xf32>
%4 = onnx.Constant dense<0.00999999977> : tensor<128xf32>
%5 = onnx.Constant dense<0.00999999977> : tensor<128xf32>
%6 = "onnx.Gemm"(%arg0, %0, %1) {alpha = 1.000000e+00 : f32, beta = 1.000000e+00 : f32, onnx_node_name = "onnx.Gemm_0", transA = 0 : si64, transB = 0 : si64} : (tensor<1x256xf32>, tensor<256x128xf32>, tensor<128xf32>) -> tensor<1x128xf32>
%7 = "onnx.BatchNormalizationInferenceMode"(%6, %2, %3, %4, %5) {epsilon = 9.99999974E-6 : f32, momentum = 0.899999976 : f32, onnx_node_name = "onnx.BatchNormalizationInferenceMode_1"} : (tensor<1x128xf32>, tensor<128xf32>, tensor<128xf32>, tensor<128xf32>, tensor<128xf32>) -> tensor<1x128xf32>
%8 = "onnx.Relu"(%7) {onnx_node_name = "onnx.Relu_2"} : (tensor<1x128xf32>) -> tensor<1x128xf32>
return %8 : tensor<1x128xf32>
// CHECK-LABEL: func @test_backward_fold_scale_axis
// CHECK-SAME: ([[PARAM_0_:%.+]]: tensor<1x256xf32>) -> tensor<1x128xf32> {
// CHECK: [[VAR_0_:%.+]] = onnx.Constant dense<{{.*}}> : tensor<256x128xf32>
// CHECK: [[VAR_1_:%.+]] = onnx.Constant dense<{{.*}}> : tensor<128xf32>
// CHECK: [[MUL_0_:%.+]] = "onnx.Mul"([[VAR_0_]], [[VAR_1_]])
// CHECK: [[MUL_1_:%.+]] = "onnx.Mul"([[VAR_1_]], [[VAR_1_]])
// CHECK: [[ADD_0_:%.+]] = "onnx.Add"([[MUL_1_]], [[VAR_1_]])
// CHECK: [[VAR_2_:%.+]] = "onnx.Gemm"([[PARAM_0_]], [[MUL_0_]], [[ADD_0_]])
// CHECK-SAME: : (tensor<1x256xf32>, tensor<256x128xf32>, tensor<128xf32>) -> tensor<1x128xf32>
// CHECK: [[VAR_3_:%.+]] = "onnx.Relu"([[VAR_2_]]) {onnx_node_name = "onnx.Relu_2"} : (tensor<1x128xf32>) -> tensor<1x128xf32>
// CHECK-NEXT: return [[VAR_3_]] : tensor<1x128xf32>
}

// -----

func.func @test_normalize_add(%arg0 : tensor<2xf32>) -> tensor<2xf32> {
%cst = "onnx.NoValue"() {value} : () -> none
%0 = onnx.Constant dense<[0.0, 1.0]> : tensor<2xf32>
Expand Down
Loading