Qualcomm AI Engine Direct - Support QNN 2.28

Differential Revision: D65949627

Pull Request resolved: #6811

Author: shewu-quic

.ci/scripts/build-qnn-sdk.sh

@@ -11,7 +11,7 @@ set -o xtrace
 build_qnn_backend() {
   echo "Start building qnn backend."
   export ANDROID_NDK_ROOT=/opt/ndk
-  export QNN_SDK_ROOT=/tmp/qnn/2.25.0.240728
+  export QNN_SDK_ROOT=/tmp/qnn/2.28.0.241029
   export EXECUTORCH_ROOT="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")/../.." && pwd)"
 
   bash backends/qualcomm/scripts/build.sh --skip_aarch64 --job_number 2 --release

.ci/scripts/setup-qnn-deps.sh

@@ -16,9 +16,9 @@ install_qnn() {
   QNN_INSTALLATION_DIR=/tmp/qnn
   mkdir -p "${QNN_INSTALLATION_DIR}"
 
-  curl -Lo /tmp/v2.25.0.24.07.28.zip "https://softwarecenter.qualcomm.com/api/download/software/qualcomm_neural_processing_sdk/v2.25.0.240728.zip"
+  curl -Lo /tmp/v2.28.0.24.10.29.zip "https://softwarecenter.qualcomm.com/api/download/software/qualcomm_neural_processing_sdk/v2.28.0.241029.zip"
   echo "Finishing downloading qnn sdk."
-  unzip -qo /tmp/v2.25.0.24.07.28.zip -d /tmp
+  unzip -qo /tmp/v2.28.0.24.10.29.zip -d /tmp
   echo "Finishing unzip qnn sdk."
 
 

.ci/scripts/test_llama.sh

@@ -121,7 +121,7 @@ echo "COREML option ${COREML}"
 if [[ "${MODE}" =~ .*qnn.* ]]; then
   QNN=ON
   export EXECUTORCH_ROOT="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")/.." && pwd)"
-  export QNN_SDK_ROOT=/tmp/qnn/2.25.0.240728
+  export QNN_SDK_ROOT=/tmp/qnn/2.28.0.241029
   export LD_LIBRARY_PATH="${QNN_SDK_ROOT}/lib/x86_64-linux-clang"
   export PYTHONPATH=".."
   cp schema/program.fbs exir/_serialize/program.fbs

backends/qualcomm/runtime/backends/QnnBackendCache.cpp

@@ -51,6 +51,11 @@ Error QnnBackendCache::GetQnnGraphInfoFromBinary(
   } else if (binaryinfo->version == QNN_SYSTEM_CONTEXT_BINARY_INFO_VERSION_2) {
     num_graphs = binaryinfo->contextBinaryInfoV2.numGraphs;
     graphs = binaryinfo->contextBinaryInfoV2.graphs;
+#if (QNN_API_VERSION_MAJOR >= 2 && QNN_API_VERSION_MINOR >= 21)
+  } else if (binaryinfo->version == QNN_SYSTEM_CONTEXT_BINARY_INFO_VERSION_3) {
+    num_graphs = binaryinfo->contextBinaryInfoV3.numGraphs;
+    graphs = binaryinfo->contextBinaryInfoV3.graphs;
+#endif
   } else {
     QNN_EXECUTORCH_LOG_WARN(
         "Unknown QNN BinaryInfo version %d.", binaryinfo->version);
@@ -62,6 +67,10 @@ Error QnnBackendCache::GetQnnGraphInfoFromBinary(
       RetrieveGraphInfo<QnnSystemContext_GraphInfoV1_t>(graphs[i].graphInfoV1);
     } else if (graphs->version == QNN_SYSTEM_CONTEXT_GRAPH_INFO_VERSION_2) {
       RetrieveGraphInfo<QnnSystemContext_GraphInfoV2_t>(graphs[i].graphInfoV2);
+#if (QNN_API_VERSION_MAJOR >= 2 && QNN_API_VERSION_MINOR >= 21)
+    } else if (graphs->version == QNN_SYSTEM_CONTEXT_GRAPH_INFO_VERSION_3) {
+      RetrieveGraphInfo<QnnSystemContext_GraphInfoV3_t>(graphs[i].graphInfoV3);
+#endif
     } else {
       QNN_EXECUTORCH_LOG_WARN(
           "Unknown QNN GraphInfo version %d.", binaryinfo->version);

backends/qualcomm/runtime/backends/htpbackend/HtpBackendCache.cpp

@@ -17,34 +17,53 @@ using executorch::runtime::Error;
 Error HtpBackendCache::RetrieveBackendBinaryInfo(
     const QnnSystemContext_BinaryInfo_t* binaryinfo) {
   QnnHtpSystemContext_HwBlobInfo_t* htp_hwblobinfo = nullptr;
+#if (QNN_API_VERSION_MAJOR >= 2 && QNN_API_VERSION_MINOR >= 21)
+  QnnHtpSystemContext_GraphBlobInfo_t* htp_graphblobinfo = nullptr;
+#endif
 
   if (binaryinfo->version == QNN_SYSTEM_CONTEXT_BINARY_INFO_VERSION_1) {
     htp_hwblobinfo = static_cast<QnnHtpSystemContext_HwBlobInfo_t*>(
         binaryinfo->contextBinaryInfoV1.hwInfoBlob);
   } else if (binaryinfo->version == QNN_SYSTEM_CONTEXT_BINARY_INFO_VERSION_2) {
     htp_hwblobinfo = static_cast<QnnHtpSystemContext_HwBlobInfo_t*>(
         binaryinfo->contextBinaryInfoV2.hwInfoBlob);
+#if (QNN_API_VERSION_MAJOR >= 2 && QNN_API_VERSION_MINOR >= 21)
+  } else if (binaryinfo->version == QNN_SYSTEM_CONTEXT_BINARY_INFO_VERSION_3) {
+    htp_graphblobinfo = static_cast<QnnHtpSystemContext_GraphBlobInfo_t*>(
+        binaryinfo->contextBinaryInfoV3.graphs->graphInfoV3.graphBlobInfo);
+#endif
   } else {
     QNN_EXECUTORCH_LOG_WARN(
         "Unknown QNN BinaryInfo version %d.", binaryinfo->version);
     return Error::Internal;
   }
 
-  if (htp_hwblobinfo == nullptr) {
-    QNN_EXECUTORCH_LOG_WARN(
-        "Htp hardware blob information is not found in binary information.");
-    return Error::Ok;
+  if (htp_hwblobinfo) {
+    if (htp_hwblobinfo->version ==
+        QNN_SYSTEM_CONTEXT_HTP_HW_INFO_BLOB_VERSION_V1) {
+      spill_fill_buf_ =
+          (*htp_hwblobinfo).contextBinaryHwInfoBlobV1_t.spillFillBufferSize;
+    } else {
+      QNN_EXECUTORCH_LOG_WARN(
+          "Unknown QNN Htp hw blob info version %d.", htp_hwblobinfo->version);
+      return Error::Internal;
+    }
   }
 
-  if (htp_hwblobinfo->version ==
-      QNN_SYSTEM_CONTEXT_HTP_HW_INFO_BLOB_VERSION_V1) {
-    spill_fill_buf_ =
-        (*htp_hwblobinfo).contextBinaryHwInfoBlobV1_t.spillFillBufferSize;
-  } else {
-    QNN_EXECUTORCH_LOG_WARN(
-        "Unknown QNN Htp hw blob info version %d.", htp_hwblobinfo->version);
-    return Error::Internal;
+#if (QNN_API_VERSION_MAJOR >= 2 && QNN_API_VERSION_MINOR >= 21)
+  if (htp_graphblobinfo) {
+    if (htp_graphblobinfo->version ==
+        QNN_SYSTEM_CONTEXT_HTP_GRAPH_INFO_BLOB_VERSION_V1) {
+      spill_fill_buf_ =
+          (*htp_graphblobinfo).contextBinaryGraphBlobInfoV1.spillFillBufferSize;
+    } else {
+      QNN_EXECUTORCH_LOG_WARN(
+          "Unknown QNN Htp graph blob info version %d.",
+          htp_graphblobinfo->version);
+      return Error::Internal;
+    }
   }
+#endif
 
   return Error::Ok;
 }

docs/source/build-run-qualcomm-ai-engine-direct-backend.md

@@ -59,7 +59,7 @@ This example is verified with SM8550 and SM8450.
    - Click the "Get Software" button to download a version of QNN SDK.
    - However, at the moment of updating this tutorial, the above website doesn't provide QNN SDK newer than 2.22.6.
    - The below is public links to download various QNN versions. Hope they can be publicly discoverable soon.
-   - [QNN 2.26.0](https://softwarecenter.qualcomm.com/api/download/software/qualcomm_neural_processing_sdk/v2.26.0.240828.zip)
+   - [QNN 2.28.0](https://softwarecenter.qualcomm.com/api/download/software/qualcomm_neural_processing_sdk/v2.28.0.241029.zip)
 
 The directory with installed Qualcomm AI Engine Direct SDK looks like:
 ```

docs/source/llm/build-run-llama3-qualcomm-ai-engine-direct-backend.md

@@ -9,7 +9,7 @@ This tutorial demonstrates how to export Llama 3 8B Instruct for Qualcomm AI Eng
 - Follow [the README for executorch llama](https://github.com/pytorch/executorch/tree/main/examples/models/llama) to know how to run a llama model on mobile via ExecuTorch.
 - A Qualcomm device with 16GB RAM
   - We are continuing to optimize our memory usage to ensure compatibility with lower memory devices.
-- The version of [Qualcomm AI Engine Direct SDK](https://developer.qualcomm.com/software/qualcomm-ai-engine-direct-sdk) is 2.26.0 or above.
+- The version of [Qualcomm AI Engine Direct SDK](https://developer.qualcomm.com/software/qualcomm-ai-engine-direct-sdk) is 2.28.0 or above.
 
 ## Instructions
 

shim/xplat/executorch/backends/qualcomm/qnn_version.bzl

@@ -1,2 +1,2 @@
 def get_qnn_library_verision():
-    return "2.26"
+    return "2.28"