Skip to content
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

Fuse batch normalization into convolution kernel #2629

Open
wants to merge 4 commits into
base: main
Choose a base branch
from
Open

Fuse batch normalization into convolution kernel #2629

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
3ac0c23
Fuse batch normalization into convolution weights
mvpant Nov 18, 2024
147c226
fix naming, comments
mvpant Dec 23, 2024
b2551e2
add large suffix to numerical test
mvpant Dec 23, 2024
b98530b
add transformation tests
mvpant Dec 25, 2024
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
Prev Previous commit
add transformation tests
  • Loading branch information
@mvpant
mvpant committed Dec 25, 2024
commit b98530bf1364b163bf3b385d5da847350d73d65a
118 changes: 118 additions & 0 deletions stablehlo/tests/transforms/stablehlo_aggressive_simplification.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1970,3 +1970,121 @@ func.func @generic_op(%arg0: tensor<2xf32>, %arg1: tensor<2xf32>) -> tensor<2xf3
%0 = "test_dialect.op"(%arg0, %arg1) : (tensor<2xf32>, tensor<2xf32>) -> (tensor<2xf32>)
return %0 : tensor<2xf32>
}


// -----

/////////
// BatchNormInferenceOp

// CHECK-LABEL: @fuse_conv_bninf
func.func @fuse_conv_bninf() -> (tensor<1x8x5x5xf32>) {
%input = stablehlo.constant dense<33.0> : tensor<1x3x8x8xf32>
%kernel = stablehlo.constant dense<0.1> : tensor<8x3x4x4xf32>
%conv = stablehlo.convolution(%input, %kernel)
dim_numbers = [b, f, 0, 1]x[o, i, 0, 1]->[b, f, 0, 1],
window = {}
{batch_group_count = 1 : i64, feature_group_count = 1 : i64}
: (tensor<1x3x8x8xf32>, tensor<8x3x4x4xf32>) -> tensor<1x8x5x5xf32>

%dummy = stablehlo.constant dense<1.0> : tensor<8xf32>
%out = "stablehlo.batch_norm_inference"(%conv, %dummy, %dummy, %dummy, %dummy)
<{epsilon = 1.0E-6 : f32, feature_index = 1 : i64}>
: (tensor<1x8x5x5xf32>, tensor<8xf32>, tensor<8xf32>, tensor<8xf32>, tensor<8xf32>)
-> tensor<1x8x5x5xf32>

// CHECK-DAG: [[C0:%.+]] = stablehlo.convolution
// CHECK-DAG: [[C1:%.+]] = stablehlo.broadcast_in_dim
// CHECK-NOT: stablehlo.batch_norm_inference
// CHECK: [[C2:%.+]] = stablehlo.add [[C0]], [[C1]]
// CHECK: return [[C2]]
return %out : tensor<1x8x5x5xf32>
}

// CHECK-LABEL: @fuse_conv_bninf_unsupported_group
func.func @fuse_conv_bninf_unsupported_group()
-> (tensor<1x8x5x5xf32>, tensor<1x8x5x5xf32>) {
%input1 = stablehlo.constant dense<33.0> : tensor<2x3x8x8xf32>
%kernel1 = stablehlo.constant dense<0.1> : tensor<8x3x4x4xf32>
%conv1 = stablehlo.convolution(%input1, %kernel1)
dim_numbers = [b, f, 0, 1]x[o, i, 0, 1]->[b, f, 0, 1], window = {}
{batch_group_count = 2 : i64, feature_group_count = 1 : i64}
: (tensor<2x3x8x8xf32>, tensor<8x3x4x4xf32>) -> tensor<1x8x5x5xf32>

%input2 = stablehlo.constant dense<33.0> : tensor<1x6x8x8xf32>
%kernel2 = stablehlo.constant dense<0.1> : tensor<8x3x4x4xf32>
%conv2 = stablehlo.convolution(%input2, %kernel2)
dim_numbers = [b, f, 0, 1]x[o, i, 0, 1]->[b, f, 0, 1], window = {}
{batch_group_count = 1 : i64, feature_group_count = 2 : i64}
: (tensor<1x6x8x8xf32>, tensor<8x3x4x4xf32>) -> tensor<1x8x5x5xf32>

%cst = stablehlo.constant dense<1.0> : tensor<8xf32>
%out1 = "stablehlo.batch_norm_inference"(%conv1, %cst, %cst, %cst, %cst)
<{epsilon = 1.0E-6 : f32, feature_index = 1 : i64}>
: (tensor<1x8x5x5xf32>, tensor<8xf32>, tensor<8xf32>, tensor<8xf32>, tensor<8xf32>)
-> tensor<1x8x5x5xf32>

%out2 = "stablehlo.batch_norm_inference"(%conv2, %cst, %cst, %cst, %cst)
<{epsilon = 1.0E-6 : f32, feature_index = 1 : i64}>
: (tensor<1x8x5x5xf32>, tensor<8xf32>, tensor<8xf32>, tensor<8xf32>, tensor<8xf32>)
-> tensor<1x8x5x5xf32>

// CHECK: [[C0:%.+]] = "stablehlo.batch_norm_inference"
// CHECK: [[C1:%.+]] = "stablehlo.batch_norm_inference"
// CHECK: return [[C0]], [[C1]]
return %out1, %out2 : tensor<1x8x5x5xf32>, tensor<1x8x5x5xf32>
}

// CHECK-LABEL: @fuse_conv_bninf_unsupported_configuration
func.func @fuse_conv_bninf_unsupported_configuration()
-> (tensor<1x1x1x1xf32>, tensor<1x1x1x1xf32>, tensor<1x1x1x1xf32>, tensor<1x1x1x1xf32>) {
%input = stablehlo.constant dense<33.0> : tensor<1x1x1x1xf32>
%kernel = stablehlo.constant dense<0.1> : tensor<1x1x1x1xf32>

%conv1 = stablehlo.convolution(%input, %kernel)
dim_numbers = [f, b, 0, 1]x[o, i, 0, 1]->[b, f, 0, 1], window = {}
{batch_group_count = 1 : i64, feature_group_count = 1 : i64}
: (tensor<1x1x1x1xf32>, tensor<1x1x1x1xf32>) -> tensor<1x1x1x1xf32>

%conv2 = stablehlo.convolution(%input, %kernel)
dim_numbers = [0, 1, f, b]x[o, i, 0, 1]->[b, f, 0, 1], window = {}
{batch_group_count = 1 : i64, feature_group_count = 1 : i64}
: (tensor<1x1x1x1xf32>, tensor<1x1x1x1xf32>) -> tensor<1x1x1x1xf32>

%conv3 = stablehlo.convolution(%input, %kernel)
dim_numbers = [b, f, 0, 1]x[i, o, 0, 1]->[b, f, 0, 1], window = {}
{batch_group_count = 1 : i64, feature_group_count = 1 : i64}
: (tensor<1x1x1x1xf32>, tensor<1x1x1x1xf32>) -> tensor<1x1x1x1xf32>

%conv4 = stablehlo.convolution(%input, %kernel)
dim_numbers = [b, f, 0, 1]x[0, 1, o, i]->[b, f, 0, 1], window = {}
{batch_group_count = 1 : i64, feature_group_count = 1 : i64}
: (tensor<1x1x1x1xf32>, tensor<1x1x1x1xf32>) -> tensor<1x1x1x1xf32>

%cst = stablehlo.constant dense<1.0> : tensor<1xf32>

%out1 = "stablehlo.batch_norm_inference"(%conv1, %cst, %cst, %cst, %cst)
<{epsilon = 1.0E-6 : f32, feature_index = 1 : i64}>
: (tensor<1x1x1x1xf32>, tensor<1xf32>, tensor<1xf32>, tensor<1xf32>, tensor<1xf32>)
-> tensor<1x1x1x1xf32>
%out2 = "stablehlo.batch_norm_inference"(%conv2, %cst, %cst, %cst, %cst)
<{epsilon = 1.0E-6 : f32, feature_index = 1 : i64}>
: (tensor<1x1x1x1xf32>, tensor<1xf32>, tensor<1xf32>, tensor<1xf32>, tensor<1xf32>)
-> tensor<1x1x1x1xf32>
%out3 = "stablehlo.batch_norm_inference"(%conv3, %cst, %cst, %cst, %cst)
<{epsilon = 1.0E-6 : f32, feature_index = 1 : i64}>
: (tensor<1x1x1x1xf32>, tensor<1xf32>, tensor<1xf32>, tensor<1xf32>, tensor<1xf32>)
-> tensor<1x1x1x1xf32>
%out4 = "stablehlo.batch_norm_inference"(%conv4, %cst, %cst, %cst, %cst)
<{epsilon = 1.0E-6 : f32, feature_index = 1 : i64}>
: (tensor<1x1x1x1xf32>, tensor<1xf32>, tensor<1xf32>, tensor<1xf32>, tensor<1xf32>)
-> tensor<1x1x1x1xf32>

// CHECK: [[C0:%.+]] = "stablehlo.batch_norm_inference"
// CHECK: [[C1:%.+]] = "stablehlo.batch_norm_inference"
// CHECK: [[C2:%.+]] = "stablehlo.batch_norm_inference"
// CHECK: [[C3:%.+]] = "stablehlo.batch_norm_inference"
// CHECK: return [[C0]], [[C1]], [[C2]], [[C3]]
return %out1, %out2, %out3, %out4 : tensor<1x1x1x1xf32>, tensor<1x1x1x1xf32>,
tensor<1x1x1x1xf32>, tensor<1x1x1x1xf32>
}
6 changes: 3 additions & 3 deletions stablehlo/transforms/StablehloAggressiveSimplification.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1496,10 +1496,10 @@ struct FuseConvolutionBatchNormalization final
auto dimNumbers = convOp.getDimensionNumbers();
if (dimNumbers.getInputBatchDimension() != 0 ||
dimNumbers.getInputFeatureDimension() != 1 ||
dimNumbers.getOutputBatchDimension() != 0 ||
dimNumbers.getOutputFeatureDimension() != 1 ||
dimNumbers.getKernelOutputFeatureDimension() != 0 ||
dimNumbers.getKernelInputFeatureDimension() != 1) {
dimNumbers.getKernelInputFeatureDimension() != 1 ||
dimNumbers.getOutputBatchDimension() != 0 ||
dimNumbers.getOutputFeatureDimension() != 1) {
constexpr StringLiteral msg =
"Only [b, f, ...]x[o, i, ...]->[b, f, ...] configuration is "
"supported";
Expand Down
Loading