Qualcomm AI Engine Direct - Support QNN 2.28

Note that if merged, QNN version less than 2.28 will not support it.

Author: shewu-quic

backends/qualcomm/runtime/backends/QnnBackendCache.cpp

@@ -51,6 +51,9 @@ Error QnnBackendCache::GetQnnGraphInfoFromBinary() {
   } else if (binaryinfo->version == QNN_SYSTEM_CONTEXT_BINARY_INFO_VERSION_2) {
     num_graphs = binaryinfo->contextBinaryInfoV2.numGraphs;
     graph = binaryinfo->contextBinaryInfoV2.graphs;
+  } else if (binaryinfo->version == QNN_SYSTEM_CONTEXT_BINARY_INFO_VERSION_3) {
+    num_graphs = binaryinfo->contextBinaryInfoV3.numGraphs;
+    graph = binaryinfo->contextBinaryInfoV3.graphs;
   } else {
     QNN_EXECUTORCH_LOG_WARN(
         "Unknown QNN BinaryInfo version %d.", binaryinfo->version);
@@ -65,28 +68,45 @@ Error QnnBackendCache::GetQnnGraphInfoFromBinary() {
     return Error::Internal;
   }
 
-  // only have version_1 now
-  if (graph[0].version != QNN_SYSTEM_CONTEXT_GRAPH_INFO_VERSION_1) {
+  if (graph[0].version == QNN_SYSTEM_CONTEXT_GRAPH_INFO_VERSION_1) {
+    // get graph name from metadata
+    graph_name_ = graph->graphInfoV1.graphName;
+
+    // get graph inputs from metadata
+    uint32_t numGraphInputs = graph->graphInfoV1.numGraphInputs;
+    input_tensor_structs_.reserve(numGraphInputs);
+    for (std::uint32_t i = 0; i < numGraphInputs; ++i) {
+      input_tensor_structs_.emplace_back(graph->graphInfoV1.graphInputs[i]);
+    }
+
+    // get graph outputs from metadata
+    uint32_t numGraphOutputs = graph->graphInfoV1.numGraphOutputs;
+    output_tensor_structs_.reserve(numGraphOutputs);
+    for (std::uint32_t i = 0; i < numGraphOutputs; ++i) {
+      output_tensor_structs_.emplace_back(graph->graphInfoV1.graphOutputs[i]);
+    }
+  } else if (graph[0].version == QNN_SYSTEM_CONTEXT_GRAPH_INFO_VERSION_3) {
+    // get graph name from metadata
+    graph_name_ = graph->graphInfoV3.graphName;
+
+    // get graph inputs from metadata
+    uint32_t numGraphInputs = graph->graphInfoV3.numGraphInputs;
+    input_tensor_structs_.reserve(numGraphInputs);
+    for (std::uint32_t i = 0; i < numGraphInputs; ++i) {
+      input_tensor_structs_.emplace_back(graph->graphInfoV3.graphInputs[i]);
+    }
+
+    // get graph outputs from metadata
+    uint32_t numGraphOutputs = graph->graphInfoV3.numGraphOutputs;
+    output_tensor_structs_.reserve(numGraphOutputs);
+    for (std::uint32_t i = 0; i < numGraphOutputs; ++i) {
+      output_tensor_structs_.emplace_back(graph->graphInfoV3.graphOutputs[i]);
+    }
+  } else {
     QNN_EXECUTORCH_LOG_WARN(
         "Unknown QNN GraphInfo version %d.", graph[0].version);
     return Error::Internal;
   }
-  // get graph name from metadata
-  graph_name_ = graph->graphInfoV1.graphName;
-
-  // get graph inputs from metadata
-  uint32_t numGraphInputs = graph->graphInfoV1.numGraphInputs;
-  input_tensor_structs_.reserve(numGraphInputs);
-  for (std::uint32_t i = 0; i < numGraphInputs; ++i) {
-    input_tensor_structs_.emplace_back(graph->graphInfoV1.graphInputs[i]);
-  }
-
-  // get graph outputs from metadata
-  uint32_t numGraphOutputs = graph->graphInfoV1.numGraphOutputs;
-  output_tensor_structs_.reserve(numGraphOutputs);
-  for (std::uint32_t i = 0; i < numGraphOutputs; ++i) {
-    output_tensor_structs_.emplace_back(graph->graphInfoV1.graphOutputs[i]);
-  }
 
   return Error::Ok;
 }

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

@@ -17,33 +17,61 @@ using executorch::runtime::Error;
 Error HtpBackendCache::RetrieveBackendBinaryInfo(
     const QnnSystemContext_BinaryInfo_t* binaryinfo) {
   QnnHtpSystemContext_HwBlobInfo_t* htp_hwblobinfo = nullptr;
+  QnnHtpSystemContext_GraphBlobInfo_t* htp_graphblobinfo = nullptr;
 
   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);
+  } else if (binaryinfo->version == QNN_SYSTEM_CONTEXT_BINARY_INFO_VERSION_3) {
+    if (binaryinfo->contextBinaryInfoV3.numGraphs > 1) {
+      QNN_EXECUTORCH_LOG_WARN(
+          "The context binary contains %lu graphs. But now "
+          "assume that one context binary contains one graph.",
+          binaryinfo->contextBinaryInfoV3.numGraphs);
+      return Error::Internal;
+    }
+    htp_graphblobinfo = static_cast<QnnHtpSystemContext_GraphBlobInfo_t*>(
+        binaryinfo->contextBinaryInfoV3.graphs->graphInfoV3.graphBlobInfo);
   } else {
     QNN_EXECUTORCH_LOG_WARN(
         "Unknown QNN BinaryInfo version %d.", binaryinfo->version);
     return Error::Internal;
   }
 
-  if (htp_hwblobinfo == nullptr) {
+  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;
+    }
+  } else {
     QNN_EXECUTORCH_LOG_WARN(
         "Htp hardware blob information is not found in binary information.");
     return Error::Ok;
   }
 
-  if (htp_hwblobinfo->version ==
-      QNN_SYSTEM_CONTEXT_HTP_HW_INFO_BLOB_VERSION_V1) {
-    spill_fill_buf_ =
-        (*htp_hwblobinfo).contextBinaryHwInfoBlobV1_t.spillFillBufferSize;
+  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;
+    }
   } else {
     QNN_EXECUTORCH_LOG_WARN(
-        "Unknown QNN Htp hw blob info version %d.", htp_hwblobinfo->version);
-    return Error::Internal;
+        "Htp graph blob information is not found in binary information.");
+    return Error::Ok;
   }
 
   return Error::Ok;