Add 3D (NCW) channels-last dim_order support to Neutron runtime

backends/nxp/runtime/NeutronBackend.cpp

@@ -95,6 +95,21 @@ void wait(uint32_t channel) {}
 static NeutronConfig neutronMemCopyConfig = {copy, wait};
 #endif
 
+// Extended dim_order check that handles 3D tensors (NCW → NWC).
+// The core is_channels_last_dim_order() only supports 4D/5D.
+bool is_channels_last_dim_order_extended(
+    const uint8_t* dim_order,
+    size_t dims) {
+  if (is_channels_last_dim_order(dim_order, dims)) {
+    return true;
+  }
+  // 3D: dim_order [0, 2, 1] means channel dim (1) is innermost → NWC layout.
+  if (dims == 3) {
+    return dim_order[0] == 0 && dim_order[1] == 2 && dim_order[2] == 1;
+  }
+  return false;
+}
+
 // Applied on outputs.
 template <typename T>
 void transposeToChannelFirst(
@@ -149,12 +164,21 @@ void transposeInput(
   if (length < 3) {
     return;
   }
-  size_t N = 1;
-  size_t C = sizes[length - 3];
-  size_t H = sizes[length - 2];
-  size_t W = sizes[length - 1];
-  for (size_t i = 0; i < length - 3; i++) {
-    N *= sizes[i];
+  size_t N, C, H, W;
+  if (length == 3) {
+    // 3D [N, C, W] → treat as [N, C, 1, W]
+    N = sizes[0];
+    C = sizes[1];
+    H = 1;
+    W = sizes[2];
+  } else {
+    N = 1;
+    C = sizes[length - 3];
+    H = sizes[length - 2];
+    W = sizes[length - 1];
+    for (size_t i = 0; i < length - 3; i++) {
+      N *= sizes[i];
+    }
   }
   switch (element_size) {
     case 1:
@@ -204,12 +228,21 @@ void transposeOutput(
   if (length < 3) {
     return;
   }
-  size_t N = 1;
-  size_t C = sizes[length - 3];
-  size_t H = sizes[length - 2];
-  size_t W = sizes[length - 1];
-  for (size_t i = 0; i < length - 3; i++) {
-    N *= sizes[i];
+  size_t N, C, H, W;
+  if (length == 3) {
+    // 3D [N, C, W] → treat as [N, C, 1, W]
+    N = sizes[0];
+    C = sizes[1];
+    H = 1;
+    W = sizes[2];
+  } else {
+    N = 1;
+    C = sizes[length - 3];
+    H = sizes[length - 2];
+    W = sizes[length - 1];
+    for (size_t i = 0; i < length - 3; i++) {
+      N *= sizes[i];
+    }
   }
   switch (element_size) {
     case 1:
@@ -252,7 +285,8 @@ bool multipleChannelsPresent(const ArrayRef<exec_aten::SizesType>& sizes) {
   if (length < 3) {
     return true;
   }
-  exec_aten::SizesType C = sizes[length - 3];
+  // For 3D [N, C, W], channel dim is sizes[1]. For 4D+, it's sizes[length-3].
+  exec_aten::SizesType C = (length == 3) ? sizes[1] : sizes[length - 3];
   return C != 1;
 }
 
@@ -431,7 +465,7 @@ class NeutronBackend final : public PyTorchBackendInterface {
           return Error::InvalidProgram;
         }
 
-        if (is_channels_last_dim_order(dim_order, arg.dim())) {
+        if (is_channels_last_dim_order_extended(dim_order, arg.dim())) {
           // The tensor is already permuted.
           ET_LOG(Debug, "Using channels last dim order for input %d.\n", i);
           cfg->dcfg.inputs[i] = arg.const_data_ptr();
@@ -481,7 +515,7 @@ class NeutronBackend final : public PyTorchBackendInterface {
           multipleChannelsPresent(arg.sizes())) {
         // The output will have to be transposed.
 
-        if (is_channels_last_dim_order(dim_order, arg.dim())) {
+        if (is_channels_last_dim_order_extended(dim_order, arg.dim())) {
           // The tensor will already be correctly permuted. No transposition
           //  needed.
           cfg->dcfg.outputs[i] = arg.mutable_data_ptr();
@@ -539,7 +573,7 @@ class NeutronBackend final : public PyTorchBackendInterface {
         }
 
         auto dim_order = arg.dim_order().data();
-        if (is_channels_last_dim_order(dim_order, arg.dim())) {
+        if (is_channels_last_dim_order_extended(dim_order, arg.dim())) {
           // The rest of the model expects the `channels_last` dim order, which
           //  the data already matches.
           ET_LOG(Debug, "Using channels last dim order for output %d.\n", i);
@@ -587,4 +621,4 @@ static auto registered = register_backend(backend_id);
 } // namespace
 } // namespace neutron
 } // namespace executor
-} // namespace torch
+} // namespace torch