[QNN EP] Update QNN SDK to 2.21 (#20534)

### Description - Updates QNN pipelines to use QNN SDK 2.21 - Downloads QNN SDK from Azure storage to avoid having to rebuild images when a new version is released. ### Motivation and Context Test with the latest QNN SDK.
microsoft · May 2, 2024 · 0dda8b0 · 0dda8b0
1 parent 8707655
commit 0dda8b0
Show file tree

Hide file tree

Showing 14 changed files with 105 additions and 51 deletions.
diff --git a/onnxruntime/core/providers/qnn/builder/opbuilder/resize_op_builder.cc b/onnxruntime/core/providers/qnn/builder/opbuilder/resize_op_builder.cc
@@ -169,10 +169,10 @@ Status ResizeOpBuilder::IsOpSupported(QnnModelWrapper& qnn_model_wrapper,
   //                                                   nearest_mode:
   // coordinate_transformation_mode: | round_prefer_floor  round_prefer_ceil  floor  ceil
   // -----------------------------------------------------------------------------------------
-  //                      half_pixel |     Resize               X              RNN     X
-  //              pytorch_half_pixel |     Resize               X               X      X
-  //                   align_corners |     Resize               X              RNN     X
-  //                      asymmetric |     Resize               X              RNN     X
+  //                      half_pixel |  Resize(QNN < 2.20)        X            RNN     X
+  //              pytorch_half_pixel |  Resize(QNN < 2.20)        X             X      X
+  //                   align_corners |  Resize(QNN < 2.20)  Resize(QNN 2.20)   RNN     X
+  //                      asymmetric |  Resize(QNN < 2.20)        X            RNN     X
 
   if (interp_mode == "nearest") {
     const std::string nearest_mode = GetOnnxAttr(node_helper, onnx_nearest_mode_attr);
@@ -181,13 +181,16 @@ Status ResizeOpBuilder::IsOpSupported(QnnModelWrapper& qnn_model_wrapper,
 
     if (is_npu_backend) {
       // QNN only supports the following nearest_mode values on HTP:
-      // - "round_prefer_floor" via QNN's Resize operator
+      // - QNN 2.19: "round_prefer_floor" via QNN's Resize operator
+      // - QNN 2.20 (API version 2.14): "round_prefer_ceil" via QNN's Resize operator
       // - "floor" via QNN's ResizeNearestNeighbor operator
-      //
-      // QNN validation does not throw an error if unsupported nearest_mode values are used, so we have to
-      // catch them here. Otherwise, accuracy is significantly degraded.
+#if QNN_API_VERSION_MAJOR >= 2 && QNN_API_VERSION_MINOR >= 14
+      ORT_RETURN_IF_NOT(nearest_mode == "round_prefer_ceil" || nearest_mode == "floor",
+                        "QNN EP: Resize on the NPU does not support nearest_mode ", nearest_mode.c_str());
+#else
       ORT_RETURN_IF_NOT(nearest_mode == "round_prefer_floor" || nearest_mode == "floor",
                         "QNN EP: Resize on the NPU does not support nearest_mode ", nearest_mode.c_str());
+#endif
 
       const bool use_resize_nn_op = nearest_mode == "floor";
 
@@ -200,6 +203,13 @@ Status ResizeOpBuilder::IsOpSupported(QnnModelWrapper& qnn_model_wrapper,
       // QNN's ResizeNearestNeighbor requires rank 4 inputs.
       ORT_RETURN_IF(use_resize_nn_op && input_rank != 4,
                     "QNN EP: Resize on the NPU with nearest_mode == 'floor' requires an input with rank 4.");
+
+#if QNN_API_VERSION_MAJOR >= 2 && QNN_API_VERSION_MINOR >= 14
+      // QNN's Resize only supports "round_prefer_ceil" if transformation_mode is "align_corners".
+      ORT_RETURN_IF(!use_resize_nn_op && transformation_mode != "align_corners",
+                    "QNN EP: Resize on the NPU only supports 'round_prefer_ceil' if "
+                    "transformation mode is 'align_corners'");
+#endif
     }
   }
 

diff --git a/onnxruntime/test/providers/qnn/resize_test.cc b/onnxruntime/test/providers/qnn/resize_test.cc
@@ -394,10 +394,21 @@ TEST_F(QnnHTPBackendTests, ResizeU8_2xLinearAsymmetric) {
 
 // Test 2x QDQ Resize mode: "nearest", coordinate_transformation_mode: "half_pixel", nearest_mode: "round_prefer_floor"
 // Maps to QNN's Resize operator.
-TEST_F(QnnHTPBackendTests, ResizeU8_2xNearestHalfPixelRoundPreferFloor) {
+// UPDATE: "round_prefer_floor" no longer supported in QNN SDK 2.21 (supported in QNN SDK 2.19)
+TEST_F(QnnHTPBackendTests, ResizeU8_2xNearestHalfPixelRoundPreferFloor_Unsupported) {
   std::vector<float> input_data = GetFloatDataInRange(-10.0f, 10.0f, 48);
   RunQDQResizeOpTest<uint8_t>(TestInputDef<float>({1, 3, 4, 4}, false, input_data),
                               {1, 3, 8, 8}, "nearest", "half_pixel", "round_prefer_floor",
+                              ExpectedEPNodeAssignment::None);  // No longer supported as of QNN SDK 2.21
+}
+
+// Test 2x QDQ Resize mode: "nearest", coordinate_transformation_mode: "align_corners", nearest_mode: "round_prefer_ceil"
+// Maps to QNN's Resize operator.
+// UPDATE: "round_prefer_ceil" is supported as of QNN SDK 2.21 if using "align_corners". (Unsupported in QNN SDK 2.19).
+TEST_F(QnnHTPBackendTests, ResizeU8_2xNearestAlignCornersRoundPreferCeil) {
+  std::vector<float> input_data = GetFloatDataInRange(-10.0f, 10.0f, 48);
+  RunQDQResizeOpTest<uint8_t>(TestInputDef<float>({1, 3, 4, 4}, false, input_data),
+                              {1, 3, 8, 8}, "nearest", "align_corners", "round_prefer_ceil",
                               ExpectedEPNodeAssignment::All);
 }
 
@@ -420,11 +431,12 @@ TEST_F(QnnHTPBackendTests, ResizeU8_3xNearestAsymmetricFloor) {
 
 // Test 2x QDQ Resize mode: "nearest", coordinate_transformation_mode: "asymmetric", nearest_mode: "round_prefer_floor"
 // Maps to QNN's Resize operator.
-TEST_F(QnnHTPBackendTests, ResizeU8_2xNearestAsymmetricRoundPreferFloor) {
+// UPDATE: "round_prefer_floor" no longer supported in QNN SDK 2.21 (supported in QNN SDK 2.19)
+TEST_F(QnnHTPBackendTests, ResizeU8_2xNearestAsymmetricRoundPreferFloor_Unsupported) {
   std::vector<float> input_data = GetFloatDataInRange(-10.0f, 10.0f, 8);
   RunQDQResizeOpTest<uint8_t>(TestInputDef<float>({1, 2, 2, 2}, false, input_data),
                               {1, 2, 4, 4}, "nearest", "asymmetric", "round_prefer_floor",
-                              ExpectedEPNodeAssignment::All);
+                              ExpectedEPNodeAssignment::None);  // No longer supported as of QNN SDK 2.21
 }
 
 // Test 3x QDQ Resize mode: "nearest", coordinate_transformation_mode: "asymmetric", nearest_mode: "round_prefer_floor"

diff --git a/onnxruntime/test/providers/qnn/simple_op_htp_test.cc b/onnxruntime/test/providers/qnn/simple_op_htp_test.cc
@@ -616,7 +616,13 @@ TEST_F(QnnHTPBackendTests, UnaryOp_Ceil) {
 }
 
 // Test accuracy of 16-bit QDQ Ceil op.
-TEST_F(QnnHTPBackendTests, UnaryOp_Ceil_U16) {
+// TODO(adrianlizarraga): Fails in QNN SDK 2.21 (linux). On Windows ARM64, fails in QNN SDK 2.19.
+//
+// input: [-12.0, -7.199, -2.399, 2.4, 7.2, 12.0]
+// CPU EP f32 model output: [-12.0, -7.0, -2.0, 3.0, 8.0, 12.0]
+// CPU EP qdq model output: [-12.0, -6.99, -1.99, 3.0, 8.0, 11.99]
+// QNN EP qdq model output: [-11.0 (WRONG), -7.0, -2.0, 2.99, 8.0, 11.99]
+TEST_F(QnnHTPBackendTests, DISABLED_UnaryOp_Ceil_U16) {
   const std::vector<float> input_data = GetFloatDataInRange(-12.0f, 12.0f, 6);
   RunQDQOpTest<uint16_t>("Ceil",
                          {TestInputDef<float>({1, 2, 3}, false, input_data)},

diff --git a/tools/ci_build/github/azure-pipelines/android-arm64-v8a-QNN-crosscompile-ci-pipeline.yml b/tools/ci_build/github/azure-pipelines/android-arm64-v8a-QNN-crosscompile-ci-pipeline.yml
@@ -31,7 +31,7 @@ parameters:
 - name: QnnSdk
   displayName: QNN SDK version
   type: string
-  default: qnn-v2.19.2.240210
+  default: 2.21.0.240401
 
 jobs:
 - job: Build_QNN_EP
@@ -40,10 +40,6 @@ jobs:
   workspace:
     clean: all
 
-  variables:
-    - name: QNN_SDK_ROOT
-      value: /data/qnnsdk/${{parameters.QnnSdk}}
-
   steps:
   - task: UsePythonVersion@0
     displayName: Use Python $(pythonVersion)
@@ -56,6 +52,10 @@ jobs:
   - script: sudo chmod go+rw /dev/kvm
     displayName: Update permissions to KVM
 
+  - template: templates/jobs/download_linux_qnn_sdk.yml
+    parameters:
+      QnnSDKVersion: ${{ parameters.QnnSdk }}
+
   - script: |
       export ANDROID_SDK_ROOT=/usr/local/lib/android/sdk
       export ANDROID_HOME=/usr/local/lib/android/sdk
@@ -78,7 +78,7 @@ jobs:
         --android_api=31 \
         --parallel \
         --use_qnn \
-        --qnn_home $(QNN_SDK_ROOT) \
+        --qnn_home $(QnnSDKRootDir) \
         --cmake_generator=Ninja \
         --skip_tests --path_to_protoc_exe $(Build.BinariesDirectory)/installed/bin/protoc
     displayName: Build QNN EP

diff --git a/tools/ci_build/github/azure-pipelines/linux-qnn-ci-pipeline.yml b/tools/ci_build/github/azure-pipelines/linux-qnn-ci-pipeline.yml
@@ -32,7 +32,7 @@ parameters:
 - name: QnnSdk
   displayName: QNN SDK version
   type: string
-  default: qnn-v2.19.2.240210
+  default: 2.21.0.240401
 
 jobs:
   - job: Build_QNN_EP
@@ -41,13 +41,8 @@ jobs:
     workspace:
       clean: all
 
-    variables:
-      - name: QNN_SDK_ROOT
-        value: /data/qnnsdk/${{parameters.QnnSdk}}
-
     steps:
       - script: |
-          ls /data/qnnsdk
           ls -R /data/qnn_test_data
         displayName: Check QNN test data
 
@@ -59,13 +54,17 @@ jobs:
       - script: sudo apt-get update -y && sudo apt-get install -y coreutils ninja-build
         displayName: Install coreutils and ninja
 
+      - template: templates/jobs/download_linux_qnn_sdk.yml
+        parameters:
+          QnnSDKVersion: ${{ parameters.QnnSdk }}
+
       - script: |
           python3 tools/ci_build/build.py \
             --build_dir build \
             --config Release \
             --parallel --use_binskim_compliant_compile_flags \
             --use_qnn \
-            --qnn_home $(QNN_SDK_ROOT) \
+            --qnn_home $(QnnSDKRootDir) \
             --cmake_generator=Ninja \
             --skip_tests
         displayName: Build QNN EP
@@ -75,7 +74,7 @@ jobs:
             --build_dir build \
             --config Release --use_binskim_compliant_compile_flags \
             --test \
-            --qnn_home $(QNN_SDK_ROOT) \
+            --qnn_home $(QnnSDKRootDir) \
             --cmake_generator=Ninja \
             --skip_submodule_sync \
             --ctest_path ""
@@ -86,29 +85,29 @@ jobs:
         inputs:
           script: |
             ./build/Release/onnx_test_runner -e qnn \
-              -v -j 1 -i "backend_path|$(QNN_SDK_ROOT)/lib/x86_64-linux-clang/libQnnCpu.so" \
+              -v -j 1 -i "backend_path|$(QnnSDKRootDir)/lib/x86_64-linux-clang/libQnnCpu.so" \
               cmake/external/onnx/onnx/backend/test/data/node
 
       - task: CmdLine@2
         displayName: Run float32 model tests
         inputs:
           script: |
             ./build/Release/onnx_test_runner -e qnn \
-              -v -j 1 -i "backend_path|$(QNN_SDK_ROOT)/lib/x86_64-linux-clang/libQnnCpu.so" \
+              -v -j 1 -i "backend_path|$(QnnSDKRootDir)/lib/x86_64-linux-clang/libQnnCpu.so" \
               /data/float32_models
 
       - task: CmdLine@2
         displayName: Run QDQ model tests
         inputs:
           script: |
             ./build/Release/onnx_test_runner -e qnn \
-              -v -j 1 -i "backend_path|$(QNN_SDK_ROOT)/lib/x86_64-linux-clang/libQnnHtp.so" \
+              -v -j 1 -i "backend_path|$(QnnSDKRootDir)/lib/x86_64-linux-clang/libQnnHtp.so" \
               /data/qdq_models
 
       - task: CmdLine@2
         displayName: Run QDQ model tests with context cache enabled
         inputs:
           script: |
             ./build/Release/onnx_test_runner -e qnn \
-              -v -f -j 1 -i "backend_path|$(QNN_SDK_ROOT)/lib/x86_64-linux-clang/libQnnHtp.so" \
+              -v -f -j 1 -i "backend_path|$(QnnSDKRootDir)/lib/x86_64-linux-clang/libQnnHtp.so" \
               /data/qdq_models/mobilenetv2-1.0_add_transpose_quant
diff --git a/tools/ci_build/github/azure-pipelines/py-packaging-pipeline.yml b/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: 2.19.2.240210
+  default: 2.21.0.240401
 
 trigger: none
 

diff --git a/tools/ci_build/github/azure-pipelines/qnn-ep-nuget-packaging-pipeline.yml b/tools/ci_build/github/azure-pipelines/qnn-ep-nuget-packaging-pipeline.yml
@@ -2,7 +2,7 @@ parameters:
 - name: QnnSdk
   displayName: QNN SDK Version
   type: string
-  default: 2.19.2.240210
+  default: 2.21.0.240401
 
 - name: build_config
   displayName: Build Configuration

diff --git a/tools/ci_build/github/azure-pipelines/templates/jobs/download_linux_qnn_sdk.yml b/tools/ci_build/github/azure-pipelines/templates/jobs/download_linux_qnn_sdk.yml
@@ -0,0 +1,25 @@
+parameters:
+  - name: QnnSDKVersion
+    type: string
+    default: '2.21.0.240401'
+
+steps:
+  - script: |
+      azcopy cp --recursive https://lotusscus.blob.core.windows.net/models/qnnsdk/qnn-v${{ parameters.QnnSDKVersion }} $(Agent.TempDirectory)
+    displayName: 'Download QNN SDK v${{ parameters.QnnSDKVersion }}'
+
+  - bash: |
+      echo "##vso[task.setvariable variable=QnnSDKRootDir]$(Agent.TempDirectory)/qnn-v${{ parameters.QnnSDKVersion }}"
+    displayName: Set QnnSDKRootDir
+
+  - script: |
+      echo $(QnnSDKRootDir)
+    displayName: 'Print QnnSDKRootDir after downloading QNN SDK'
+
+  - script: |
+      azcopy cp --recursive 'https://lotusscus.blob.core.windows.net/models/qnnsdk/Qualcomm AI Hub Proprietary License.pdf' $(QnnSDKRootDir)
+    displayName: 'Download Qualcomm AI Hub license'
+
+  - script: |
+      ls -al $(QnnSDKRootDir)
+    displayName: 'Print contents of QNN SDK'
diff --git a/tools/ci_build/github/azure-pipelines/templates/jobs/download_win_qnn_sdk.yml b/tools/ci_build/github/azure-pipelines/templates/jobs/download_win_qnn_sdk.yml
@@ -1,7 +1,7 @@
 parameters:
   - name: QnnSDKVersion
     type: string
-    default: '2.19.2.240210'
+    default: '2.21.0.240401'
 
 steps:
   - powershell: |

diff --git a/tools/ci_build/github/azure-pipelines/templates/py-packaging-stage.yml b/tools/ci_build/github/azure-pipelines/templates/py-packaging-stage.yml
@@ -60,7 +60,7 @@ parameters:
 - name: qnn_sdk_version
   type: string
   displayName: 'QNN SDK version. Only for QNN packages.'
-  default: 2.19.2.240210
+  default: 2.21.0.240401
 
 stages:
 - ${{ if eq(parameters.enable_windows_cpu, true) }}:

diff --git a/tools/ci_build/github/azure-pipelines/templates/py-win-arm64-qnn.yml b/tools/ci_build/github/azure-pipelines/templates/py-win-arm64-qnn.yml
@@ -7,7 +7,7 @@ parameters:
 - name: QNN_SDK
   displayName: QNN SDK Version
   type: string
-  default: 2.19.2.240210
+  default: 2.21.0.240401
 
 - name: PYTHON_VERSION
   type: string

diff --git a/tools/ci_build/github/azure-pipelines/templates/py-win-x64-qnn.yml b/tools/ci_build/github/azure-pipelines/templates/py-win-x64-qnn.yml
@@ -7,7 +7,7 @@ parameters:
 - name: QNN_SDK
   displayName: QNN SDK Version
   type: string
-  default: 2.19.2.240210
+  default: 2.21.0.240401
 
 - name: ENV_SETUP_SCRIPT
   type: string

diff --git a/tools/ci_build/github/azure-pipelines/win-qnn-arm64-ci-pipeline.yml b/tools/ci_build/github/azure-pipelines/win-qnn-arm64-ci-pipeline.yml
@@ -32,7 +32,7 @@ parameters:
 - name: QnnSdk
   displayName: QNN SDK version
   type: string
-  default: qnn-v2.19.2.240210_win
+  default: 2.21.0.240401
 
 jobs:
 - job: 'build'
@@ -42,7 +42,6 @@ jobs:
     buildArch: arm64
     BuildConfig: 'RelWithDebInfo'
     ALLOW_RELEASED_ONNX_OPSET_ONLY: '1'
-    QNN_SDK_ROOT: 'C:\data\qnnsdk\${{parameters.QnnSdk}}'
   timeoutInMinutes: 240
   workspace:
     clean: all
@@ -57,10 +56,6 @@ jobs:
       DIR $(Agent.ToolsDirectory)\Python\3.11.0\arm64
     displayName: Copy python 3.11.0 version to agent tools directory
 
-  - script: |
-      DIR C:\data\qnnsdk
-    displayName: Check available QNN SDKs
-
   - task: UsePythonVersion@0
     inputs:
       versionSpec: '3.x'
@@ -71,11 +66,15 @@ jobs:
     inputs:
       versionSpec: 6.4.x
 
+  - template: templates/jobs/download_win_qnn_sdk.yml
+    parameters:
+      QnnSDKVersion: ${{ parameters.QnnSdk }}
+
   - task: PythonScript@0
     displayName: 'Build'
     inputs:
       scriptPath: '$(Build.SourcesDirectory)\tools\ci_build\build.py'
-      arguments: '--config $(BuildConfig) --build_dir $(Build.BinariesDirectory) --skip_tests --cmake_generator "Visual Studio 17 2022" --use_qnn --qnn_home $(QNN_SDK_ROOT) --parallel'
+      arguments: '--config $(BuildConfig) --build_dir $(Build.BinariesDirectory) --skip_tests --cmake_generator "Visual Studio 17 2022" --use_qnn --qnn_home $(QnnSDKRootDir) --parallel'
       workingDirectory: '$(Build.BinariesDirectory)'
 
   - powershell: |
@@ -84,17 +83,17 @@ jobs:
     displayName: 'Run unit tests'
 
   - script: |
-     .\$(BuildConfig)\onnx_test_runner -j 1 -v -e qnn -i "backend_path|$(QNN_SDK_ROOT)\lib\aarch64-windows-msvc\QnnCpu.dll" $(Build.SourcesDirectory)\cmake\external\onnx\onnx\backend\test\data\node
+     .\$(BuildConfig)\onnx_test_runner -j 1 -v -e qnn -i "backend_path|$(QnnSDKRootDir)\lib\aarch64-windows-msvc\QnnCpu.dll" $(Build.SourcesDirectory)\cmake\external\onnx\onnx\backend\test\data\node
     workingDirectory: '$(Build.BinariesDirectory)\$(BuildConfig)'
     displayName: 'Run ONNX Tests'
 
   - script: |
-     .\$(BuildConfig)\onnx_test_runner -j 1 -v -e qnn -i "backend_path|$(QNN_SDK_ROOT)\lib\aarch64-windows-msvc\QnnCpu.dll" C:\data\float32_models
+     .\$(BuildConfig)\onnx_test_runner -j 1 -v -e qnn -i "backend_path|$(QnnSDKRootDir)\lib\aarch64-windows-msvc\QnnCpu.dll" C:\data\float32_models
     workingDirectory: '$(Build.BinariesDirectory)\$(BuildConfig)'
     displayName: 'Run float32 model tests'
 
   - script: |
-     .\$(BuildConfig)\onnx_test_runner -j 1 -v -e qnn -i "backend_path|$(QNN_SDK_ROOT)\lib\aarch64-windows-msvc\QnnHtp.dll" C:\data\qdq_models
+     .\$(BuildConfig)\onnx_test_runner -j 1 -v -e qnn -i "backend_path|$(QnnSDKRootDir)\lib\aarch64-windows-msvc\QnnHtp.dll" C:\data\qdq_models
     workingDirectory: '$(Build.BinariesDirectory)\$(BuildConfig)'
     displayName: 'Run QDQ model tests'
     enabled: false