[QNN EP] QNN SDK 2.28.2 (#22844)

### Description
- Updates pipelines to use QNN SDK 2.28.2.241116.
- Re-enable LayerNormalization unit tests that failed with accuracy
errors with the previous QNN SDK (2.28.0).
- Update QNN EP to no longer provide a dummy bias for LayerNorm if the
QNN SDK version is >= 2.28.0.


### Motivation and Context
Use the latest QNN SDK. This version improves inference latency for
certain customer models.

Author: adrianlizarraga

onnxruntime/core/providers/qnn/builder/opbuilder/layer_norm_op_builder.cc

@@ -87,10 +87,10 @@ Status LayerNormOpBuilder::ProcessInputs(QnnModelWrapper& qnn_model_wrapper,
     ORT_RETURN_IF_ERROR(ProcessInput(qnn_model_wrapper, inputs[BIAS_IDX], logger, input_names));
   }
 
-#if QNN_API_VERSION_MAJOR == 2 && (QNN_API_VERSION_MINOR >= 17)
+#if QNN_API_VERSION_MAJOR == 2 && QNN_API_VERSION_MINOR >= 17 && QNN_API_VERSION_MINOR <= 20
   if (!has_bias_input && IsNpuBackend(qnn_model_wrapper.GetQnnBackendType())) {
-    // Bias is implicit. QNN SDK 2.24+ (QNN API version 2.17+) has a validation bug for implicit bias inputs,
-    // so provide an explicit bias of all 0 (quantized int32).
+    // Bias is implicit. QNN SDK 2.24 to 2.27 (QNN API version 2.17 to 2.20) has a validation bug for
+    // implicit bias inputs, so provide an explicit bias of all 0 (quantized int32).
     TensorInfo x_input_info = {};
     ORT_RETURN_IF_ERROR(qnn_model_wrapper.GetTensorInfo(inputs[X_IDX], x_input_info));
 

onnxruntime/core/providers/qnn/builder/qnn_backend_manager.cc

@@ -14,6 +14,7 @@
 #include "DSP/QnnDspCommon.h"
 #include "HTP/QnnHtpCommon.h"
 #include "HTP/QnnHtpContext.h"
+#include "Saver/QnnSaver.h"
 #include <gsl/gsl>
 #include "core/framework/endian_utils.h"
 #include "core/common/logging/capture.h"
@@ -1040,7 +1041,14 @@ Status QnnBackendManager::ExtractBackendProfilingInfo() {
   const QnnProfile_EventId_t* profile_events{nullptr};
   uint32_t num_events{0};
   Qnn_ErrorHandle_t result = qnn_interface_.profileGetEvents(profile_backend_handle_, &profile_events, &num_events);
-  ORT_RETURN_IF(QNN_PROFILE_NO_ERROR != result, "Failed to get profile events. Error: ", QnnErrorHandleToString(result));
+  if (!qnn_saver_path_.empty()) {  // Using QNN Saver backend
+    // QNN SDK 2.28.2 returns QNN_SAVER_ERROR_DUMMY_RETVALUE, but previous QNN versions return QNN_PROFILE_NO_ERROR.
+    // We accept both values.
+    ORT_RETURN_IF(QNN_PROFILE_NO_ERROR != result && QNN_SAVER_ERROR_DUMMY_RETVALUE != result,
+                  "Failed to get profile events. Error: ", QnnErrorHandleToString(result));
+  } else {
+    ORT_RETURN_IF(QNN_PROFILE_NO_ERROR != result, "Failed to get profile events. Error: ", QnnErrorHandleToString(result));
+  }
 
   if (num_events > 0) {
     LOGS(*logger_, VERBOSE) << "profile_events: " << profile_events << " num_events: " << num_events;

onnxruntime/test/providers/qnn/gather_op_htp_test.cc

@@ -132,6 +132,7 @@ TEST_F(QnnHTPBackendTests, GatherOp_IndicesDynamicInt32_Axis0) {
 }
 
 // disabled for QNN 2.28.0.241029 failed for accuracy validation
+// Also fails on QNN 2.28.2.
 // qdq@QNN_EP val: 3.6094117164611816 (err: 1.3094117641448975, err/output_range: 22.19342041015625%)
 // qdq@CPU_EP val: 2.2905881404876709 (err: 0.0094118118286132812, err/output_range: 0.15952222049236298%)
 // abs(qdq@QNN_EP - qdq@CPU_EP) / output_range = 22.033897399902344%

onnxruntime/test/providers/qnn/layer_norm_test.cc

@@ -188,15 +188,11 @@ TEST_F(QnnHTPBackendTests, LayerNorm1D_LastAxis_StaticScale_StaticBias_AU8_WU8_B
                                         ExpectedEPNodeAssignment::All);
 }
 
-// QNN 2.27 accuracy issue
-// Inaccuracy detected for output 'output_0', element 0
-// output_range=1.2245157957077026, tolerance=0.40000000596046448%.
-// Expected val (f32@CPU_EP): -0
-// qdq@QNN_EP val: 0.19133351743221283 (err: 0.19133351743221283, err/output_range: 15.625238418579102%)
-// qdq@CPU_EP val: 0 (err: 0, err/output_range: 0%)
-TEST_F(QnnHTPBackendTests, DISABLED_LayerNorm1D_QNN2_24_ImplicitBias_ValidationBug) {
-  // QNN 2.24 LayerNorm fails validation (intermittent) if the bias input is not provided. QNN EP will provide an
-  // explicit bias of all zeros to get around this bug.
+TEST_F(QnnHTPBackendTests, LayerNorm1D_QNN2_24_ImplicitBias_ValidationBug) {
+  // QNN 2.24 to 2.27: LayerNorm fails validation (intermittent) if the bias input is not provided. QNN EP will provide
+  // an explicit bias of all zeros to get around this bug.
+  // QNN 2.28.0: Validation bug is fixed, but get accuracy errors.
+  // QNN 2.28.2: All fixed.
   for (size_t i = 0; i < 15; i++) {  // Run it multiple times since this is an intermittent bug.
     RunLayerNormQDQTest<uint16_t, uint8_t>(TestInputDef<float>({1, 2, 3}, false, GetFloatDataInRange(0.0f, 1.0f, 6)),
                                            TestInputDef<float>({3}, true, GetFloatDataInRange(0.0f, 1.0f, 3)),
@@ -207,14 +203,9 @@ TEST_F(QnnHTPBackendTests, DISABLED_LayerNorm1D_QNN2_24_ImplicitBias_ValidationB
   }
 }
 
-// Test accuracy of 16-bit QDQ LayerNorm with a static scale input.
-// QNN 2.27 accuracy issue
-// Inaccuracy detected for output 'output_0', element 0
-// output_range=1.224743127822876, tolerance=0.40000000596046448%.
-// Expected val (f32@CPU_EP): -0
-// qdq@QNN_EP val: 0.19136904180049896 (err: 0.19136904180049896, err/output_range: 15.625238418579102%)
-// qdq@CPU_EP val: 0 (err: 0, err/output_range: 0%)
-TEST_F(QnnHTPBackendTests, DISABLED_LayerNorm1D_LastAxis_StaticScale_AU16_WU8) {
+TEST_F(QnnHTPBackendTests, LayerNorm1D_LastAxis_StaticScale_AU16_WU8) {
+  // QNN 2.28.0: Get accuracy errors.
+  // QNN 2.28.2: All fixed.
   RunLayerNormQDQTest<uint16_t, uint8_t>(TestInputDef<float>({1, 2, 3}, false, GetFloatDataInRange(0.0f, 10.0f, 6)),
                                          TestInputDef<float>({3}, true, GetFloatDataInRange(0.0f, 1.0f, 3)),  // Static
                                          TestInputDef<float>(),
@@ -225,7 +216,7 @@ TEST_F(QnnHTPBackendTests, DISABLED_LayerNorm1D_LastAxis_StaticScale_AU16_WU8) {
 
 // Test accuracy of 8-bit QDQ LayerNorm with a dynamic scale input.
 //
-// TODO(adrianlizarraga): Fails to finalize with QNN SDK 2.22.
+// TODO(adrianlizarraga): Fails to finalize with QNN SDK 2.22. Still fails on QNN SDK 2.28.2.
 // Verbose logs:
 // Starting stage: Graph Transformations and Optimizations
 // C:\...\QNN\HTP\HTP\src\hexagon\prepare\graph_prepare.cc:203:ERROR:could not create op: q::flat_to_vtcm

onnxruntime/test/providers/qnn/matmul_test.cpp

@@ -273,7 +273,7 @@ TEST_F(QnnHTPBackendTests, MatMulOp_PerChannel_A16_WeightUInt4) {
 }
 
 // Test QDQ per-channel MatMul with int8 act, int4 weights (static)
-// QNN 2.27 regression
+// QNN 2.27 regression. Also fails on QNN 2.28.2.
 // Failed to finalize QNN graph. Error code: 1002
 TEST_F(QnnHTPBackendTests, DISABLED_MatMulOp_PerChannel_AS8_WeightInt4) {
   std::vector<float> input0_data = GetFloatDataInRange(-5.0f, 5.0f, 6);

onnxruntime/test/providers/qnn/simple_op_htp_test.cc

@@ -230,6 +230,7 @@ TEST_F(QnnHTPBackendTests, UnaryOp_Tanh) {
 }
 
 // disabled for QNN 2.28.0.241029 backendValidateOpConfig failed
+// still fails on QNN 2.28.2.
 // QnnDsp <E> [4294967295] has incorrect Value -32768, expected equal to 0.
 // QnnDsp <V> validateNativeOps node_token_6:qti.aisw:Tanh htp op validator failed 3110
 // QnnDsp <V> registered validator failed => 3110

tools/ci_build/github/azure-pipelines/android-arm64-v8a-QNN-crosscompile-ci-pipeline.yml

@@ -32,7 +32,7 @@ parameters:
 - name: QnnSdk
   displayName: QNN SDK version
   type: string
-  default: 2.28.0.241029
+  default: 2.28.2.241116
 
 jobs:
 - job: Build_QNN_EP

tools/ci_build/github/azure-pipelines/c-api-noopenmp-packaging-pipelines.yml

@@ -62,7 +62,7 @@ parameters:
 - name: QnnSdk
   displayName: QNN SDK Version
   type: string
-  default: 2.28.0.241029
+  default: 2.28.2.241116
 
 resources:
   repositories:

tools/ci_build/github/azure-pipelines/linux-qnn-ci-pipeline.yml

@@ -33,7 +33,7 @@ parameters:
 - name: QnnSdk
   displayName: QNN SDK version
   type: string
-  default: 2.28.0.241029
+  default: 2.28.2.241116
 
 jobs:
   - job: Build_QNN_EP

tools/ci_build/github/azure-pipelines/py-packaging-pipeline.yml

@@ -59,7 +59,7 @@ parameters:
 - name: qnn_sdk_version
   type: string
   displayName: 'QNN SDK version. Only for QNN packages.'
-  default: 2.28.0.241029
+  default: 2.28.2.241116
 
 trigger: none