Initialize HalfConv layout flags

Set MLAS_CONV_PARAMETERS::ChannelsLast in the HalfConv prepare path so
Prepare fully initializes the convolution parameter layout state.

Add a focused MLAS test that verifies Prepare overwrites stale layout flag
values for both NCHW and NHWC HalfConv configurations.

Signed-off-by: Cathal Lawlor <cathal.lawlor@arm.com>

Author: Laan33

onnxruntime/core/mlas/lib/kleidiai/halfconv_kleidiai.cpp

@@ -915,6 +915,7 @@ bool
     Parameters->InputChannels = InputChannels;
     Parameters->FilterCount = FilterCount;
     Parameters->Beta = Beta;
+    Parameters->ChannelsLast = InputOutputChannelsLast;
     Parameters->InputOutputChannelsLast = InputOutputChannelsLast;
 
     size_t input_size = 1;

onnxruntime/test/mlas/unittest/test_conv2d.cpp

@@ -99,6 +99,71 @@ TEST(Conv2d_HalfConv, PrepareReportsWorkingBufferSizeInBytes) {
   EXPECT_EQ(working_buffer_size, parameters.OutputSize * parameters.FilterCount * sizeof(uint16_t));
 }
 
+TEST(Conv2d_HalfConv, PrepareInitializesLayoutFlags) {
+  const int64_t input_shape[] = {5, 5};
+  const int64_t kernel_shape[] = {3, 3};
+  const int64_t dilation_shape[] = {1, 1};
+  const int64_t padding[] = {1, 1, 1, 1};
+  const int64_t stride_shape[] = {1, 1};
+  const int64_t output_shape[] = {5, 5};
+
+  MLAS_ACTIVATION activation;
+  activation.ActivationKind = MlasIdentityActivation;
+
+  MLAS_CONV_PARAMETERS nchw_parameters{};
+  nchw_parameters.ChannelsLast = true;
+  nchw_parameters.InputOutputChannelsLast = true;
+  size_t nchw_working_buffer_size = 0;
+  if (!MlasHalfConvPrepare(&nchw_parameters,
+                           2,
+                           1,
+                           1,
+                           4,
+                           input_shape,
+                           kernel_shape,
+                           dilation_shape,
+                           padding,
+                           stride_shape,
+                           output_shape,
+                           8,
+                           &activation,
+                           &nchw_working_buffer_size,
+                           0.0f,
+                           false,
+                           nullptr,
+                           nullptr)) {
+    GTEST_SKIP() << "HalfConv prepare path unavailable";
+  }
+
+  EXPECT_FALSE(nchw_parameters.ChannelsLast);
+  EXPECT_FALSE(nchw_parameters.InputOutputChannelsLast);
+
+  MLAS_CONV_PARAMETERS nhwc_parameters{};
+  size_t nhwc_working_buffer_size = 0;
+  ASSERT_TRUE(MlasHalfConvPrepare(&nhwc_parameters,
+                                  2,
+                                  1,
+                                  1,
+                                  4,
+                                  input_shape,
+                                  kernel_shape,
+                                  dilation_shape,
+                                  padding,
+                                  stride_shape,
+                                  output_shape,
+                                  8,
+                                  &activation,
+                                  &nhwc_working_buffer_size,
+                                  0.0f,
+                                  true,
+                                  nullptr,
+                                  nullptr));
+
+  EXPECT_TRUE(nhwc_parameters.ChannelsLast);
+  EXPECT_TRUE(nhwc_parameters.InputOutputChannelsLast);
+  EXPECT_EQ(nhwc_working_buffer_size, size_t{0});
+}
+
 static size_t Conv2dRegistLongExecute() {
   size_t count = MlasLongExecuteTests<MlasConv2DTest<false>>::RegisterLongExecute();
   if (GetMlasThreadPool() != nullptr) {