Skip to content

[QNN EP] Adjust tolerance for Clip and Transpose tests due to FP16 default in QNN HTP #26499

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
yuslepukhin merged 1 commit into from
Jan 14, 2026
Merged

[QNN EP] Adjust tolerance for Clip and Transpose tests due to FP16 default in QNN HTP #26499

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
28 changes: 11 additions & 17 deletions onnxruntime/test/providers/qnn/clip_op_test.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -22,22 +22,15 @@ static void RunClipTest(const TestInputDef<DataType>& input_def,
ExpectedEPNodeAssignment expected_ep_assignment,
const std::string& backend_name = "cpu",
int opset = 13,
bool enable_fp16_precision = true) {
float fp32_abs_err = 1e-5f) {
ProviderOptions provider_options;
provider_options["backend_type"] = backend_name;

if (backend_name == "htp") {
if (enable_fp16_precision) {
provider_options["enable_htp_fp16_precision"] = "1";
} else {
provider_options["enable_htp_fp16_precision"] = "0";
}
}

RunQnnModelTest(BuildOpTestCase<DataType, DataType>("Clip", {input_def}, min_max_defs, {}),
provider_options,
opset,
expected_ep_assignment);
expected_ep_assignment,
fp32_abs_err);
}

//
Expand Down Expand Up @@ -77,17 +70,19 @@ TEST_F(QnnCPUBackendTests, Clip_5D_f32) {
// HTP tests:
//

// Test Clip with float32 on HTP
// Fails with QNN SDK 2.35.0:
// value pair (-4.54545403, -4.54687548) at index #3 don't match, which is -0.00142145 from -4.54545
TEST_F(QnnHTPBackendTests, DISABLED_Clip_f32) {
// Test Clip with float32 on HTP.
// Since QAIRT 2.35, default float precision on QNN HTP became FP16.
// Converting FP32 -> FP16 -> FP32 may introduce minor accuracy loss.
// For example, a value of -4.54545403 could become -4.54687548 after the conversion.
// The expected difference is approximately 0.00142145, so the tolerance is adjusted to 5e-3f.
TEST_F(QnnHTPBackendTests, Clip_f32) {
RunClipTest<float>(TestInputDef<float>({1, 1, 3, 4}, false, GetFloatDataInRange(-10.0f, 10.0f, 12)),
{TestInputDef<float>({}, true, {-5.0f}),
TestInputDef<float>({}, true, {5.0f})},
ExpectedEPNodeAssignment::All,
"htp",
13,
false);
5e-3f);
}

// Test Clip with int32 on HTP
Expand Down Expand Up @@ -248,8 +243,7 @@ TEST_F(QnnGPUBackendTests, Clip_fp32) {
TestInputDef<float>({}, true, {5.0f})},
ExpectedEPNodeAssignment::All,
"gpu",
13,
false);
13);
}

// Test Clip with int32 on GPU
Expand Down
20 changes: 8 additions & 12 deletions onnxruntime/test/providers/qnn/transpose_htp_test.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -88,21 +88,15 @@ template <typename DataType>
static void RunTransposeNonQDQOnHTP(const TestInputDef<DataType>& input_def,
const std::vector<ONNX_NAMESPACE::AttributeProto>& attrs,
ExpectedEPNodeAssignment expected_ep_assignment,
bool enable_fp16_precision = true) {
float fp32_abs_err = 1e-5f) {
ProviderOptions provider_options;
provider_options["backend_type"] = "htp";

if (enable_fp16_precision) {
provider_options["enable_htp_fp16_precision"] = "1";
} else {
provider_options["enable_htp_fp16_precision"] = "0";
}

RunQnnModelTest(BuildTransposeTestCase<DataType>(input_def, attrs),
provider_options,
13,
expected_ep_assignment,
1e-5f);
fp32_abs_err);
}

// Check that QNN compiles DQ -> Transpose -> Q as a single unit.
Expand All @@ -120,12 +114,14 @@ TEST_F(QnnHTPBackendTests, TransposeInt32OnHTP) {
}

// Check that QNN supports Transpose with float32 data input on HTP
// Fails with QNN SDK 2.35.0:
// value pair (0.183528364, 0.183471695) at index #0 don't match, which is -5.66691e-05 from 0.183528
TEST_F(QnnHTPBackendTests, DISABLED_TransposeFloatOnHTP) {
// Since QAIRT 2.35, default float precision on QNN HTP became FP16.
// Converting FP32 -> FP16 -> FP32 may introduce minor accuracy loss.
// For example, a value of 7.64300251 could become 7.64453173 after the conversion.
// The expected difference is approximately 0.00152922, so the tolerance is adjusted to 5e-3f.
TEST_F(QnnHTPBackendTests, TransposeFloat32OnHTP) {
RunTransposeNonQDQOnHTP<float>(TestInputDef<float>({1, 3, 224, 128}, false, 0, 10.0f),
{utils::MakeAttribute("perm", std::vector<int64_t>{0, 2, 3, 1})},
ExpectedEPNodeAssignment::All, false);
ExpectedEPNodeAssignment::All, 5e-3f);
}

#endif // defined(__aarch64__) || defined(_M_ARM64) || defined(__linux__)
Expand Down
Loading