[KleidiAI] Always attempt activation packing

ghstack-source-id: 52488f3
ghstack-comment-id: 3169012590
Pull Request resolved: #13232

Author: mcr229

backends/xnnpack/runtime/XNNCompiler.cpp

@@ -609,9 +609,6 @@ bool isQP8(const fb_xnnpack::XNNGraph* graph, const NodePtr node) {
   auto cvt_output_id = graph_node->output_id();
 
   auto check_dtype = [graph](uint32_t id, DataType dtype) -> bool {
-    assert(
-        dtype == DataType::xnn_datatype_qdint8 ||
-        dtype == DataType::xnn_datatype_qbint4);
     for (auto value : *graph->xvalues()) {
       if (value->xvalue_union_type() !=
           fb_xnnpack::XValueUnion::XNNQuantizedTensorValue) {
@@ -631,16 +628,21 @@ bool isQP8(const fb_xnnpack::XNNGraph* graph, const NodePtr node) {
     return false;
   }
 
+  std::vector<DataType> supported_filter_dtypes{
+      DataType::xnn_datatype_qcint8,
+      DataType::xnn_datatype_qcint4,
+      DataType::xnn_datatype_qbint4};
   // Find if the convert output is going to the right linear node.
   // Assuming if we can find one valid linear node, then we can use QP8
   // for all the linear nodes consuming this convert output.
   for (auto node : *graph->xnodes()) {
     if (node->xnode_union_type() == fb_xnnpack::XNodeUnion::XNNFullyConnected) {
       auto linear_node = node->xnode_union_as_XNNFullyConnected();
       if (linear_node->input1_id() == cvt_output_id) {
-        if (check_dtype(
-                linear_node->filter_id(), DataType::xnn_datatype_qbint4)) {
-          return true;
+        for (auto supported_filter : supported_filter_dtypes) {
+          if (check_dtype(linear_node->filter_id(), supported_filter)) {
+            return true;
+          }
         }
       }
     }
@@ -659,21 +661,20 @@ Error defineConvertNode(
     const fb_xnnpack::XNNGraph* flatbuffer_graph) noexcept {
   MAYBE_UNUSED(flatbuffer_graph);
   auto graph_node = node->xnode_union_as_XNNConvert();
-
-  int32_t flags = graph_node->flags();
 #ifdef ENABLE_XNNPACK_KLEIDI
-// This is not currently exposed at include/xnnpack.h yet once it is
-// we can remove this runtime logic and do this ahead-of-time
-#define XNN_FLAG_MAYBE_PACK_FOR_QB4W_GEMM 0x00000100;
-  if (isQP8(flatbuffer_graph, node)) {
-    flags |= XNN_FLAG_MAYBE_PACK_FOR_QB4W_GEMM;
-    ET_LOG(
-        Debug,
-        "Setting XNN_FLAG_MAYBE_PACK_FOR_QB4W_GEMM flag for convert node %i",
-        node->debug_handle());
-  }
+  // This is not currently exposed at include/xnnpack.h yet once it is
+  // we can remove this runtime logic and do this ahead-of-time
+  #define XNN_FLAG_MAYBE_PACK_FOR_QB4W_GEMM 0x00000100;
+    if (isQP8(flatbuffer_graph, node)) {
+      flags |= XNN_FLAG_MAYBE_PACK_FOR_QB4W_GEMM;
+      ET_LOG(
+          Debug,
+          "Setting XNN_FLAG_MAYBE_PACK_FOR_QB4W_GEMM flag for convert node %i",
+          node->debug_handle());
+    }
 #endif
 
+  int32_t flags = graph_node->flags();
   xnn_status status = xnn_define_convert(
       subgraph_ptr,
       remapped_ids.at(graph_node->input_id()),