Skip to content

[onnx] fix fp16 gelu pattern #503

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

Merged
qingyunqu merged 1 commit into from
Apr 4, 2025
+26 −0
Merged

[onnx] fix fp16 gelu pattern #503

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
9 changes: 9 additions & 0 deletions frontends/onnx-frontend/onnx-frontend/src/Conversion/OFRewriteToCustomCall.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -573,6 +573,15 @@ bool isSplatFPCloseTo(Attribute attr, double value, double eps = 1e-5) {
return false;
if (!elementsAttr.isSplat())
return false;

auto dtype = elementsAttr.getElementType();
auto bitwidth = dtype.getIntOrFloatBitWidth();
if (bitwidth <= 8) {
eps = std::max(fabs(eps), 1e-2);
} else if (bitwidth == 16) {
eps = std::max(fabs(eps), 1e-3);
}

double diff =
elementsAttr.getSplatValue<FloatAttr>().getValueAsDouble() - value;
return fabs(diff) < eps;
Expand Down
17 changes: 17 additions & 0 deletions frontends/onnx-frontend/onnx-frontend/test/of_rewrite_to_custom_call.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -235,6 +235,23 @@ func.func @test_gelu(%37: tensor<1x3x5x5xf32>) -> tensor<1x3x5x5xf32> {

// -----

func.func @test_gelu_fp16(%97 : tensor<16x5638x3072xf16>) -> tensor<16x5638x3072xf16> {
%19 = onnx.Constant dense<5.000000e-01> : tensor<f16>
%20 = onnx.Constant dense<1.000000e+00> : tensor<f16>
%21 = "onnx.Constant"() {value = dense<1.414060e+00> : tensor<f16>} : () -> tensor<f16>
%98 = "onnx.Div"(%97, %21) {onnx_node_name = "Div"} : (tensor<16x5638x3072xf16>, tensor<f16>) -> tensor<16x5638x3072xf16>
%99 = "onnx.Erf"(%98) {onnx_node_name = "Erf"} : (tensor<16x5638x3072xf16>) -> tensor<16x5638x3072xf16>
%100 = "onnx.Add"(%99, %20) {onnx_node_name = "Add"} : (tensor<16x5638x3072xf16>, tensor<f16>) -> tensor<16x5638x3072xf16>
%101 = "onnx.Mul"(%97, %100) {onnx_node_name = "Mul"} : (tensor<16x5638x3072xf16>, tensor<16x5638x3072xf16>) -> tensor<16x5638x3072xf16>
%102 = "onnx.Mul"(%101, %19) {onnx_node_name = "Mul_1"} : (tensor<16x5638x3072xf16>, tensor<f16>) -> tensor<16x5638x3072xf16>
return %102 : tensor<16x5638x3072xf16>
}
// CHECK-LABEL: @test_gelu_fp16
// CHECK-SAME: ([[PARAM_0_:%.+]]: tensor<16x5638x3072xf16>) -> tensor<16x5638x3072xf16> {
// CHECK-NEXT: [[VAR_0_:%.+]] = stablehlo.custom_call @byteir.gelu([[PARAM_0_]]) {byteir_attrs = {approximate = "erf"}} : (tensor<16x5638x3072xf16>) -> tensor<16x5638x3072xf16>

// -----

func.func @test_gelu_without_last_mul(%arg0: tensor<1x3x5x5xf32>, %arg1: tensor<1x3x5x5xf32>) -> tensor<1x3x5x5xf32> {
%38 = "onnx.Constant"() {value = dense<0.000000e+00> : tensor<f32>} : () -> tensor<f32>
%39 = "onnx.Add"(%arg0, %38) : (tensor<1x3x5x5xf32>, tensor<f32>) -> tensor<1x3x5x5xf32>
Expand Down