DepthwiseConv Optimizations

PiperOrigin-RevId: 198144118

diff --git a/tensorflow/contrib/lite/kernels/internal/optimized/depthwiseconv_uint8.h b/tensorflow/contrib/lite/kernels/internal/optimized/depthwiseconv_uint8.h
index b85e6c4..3fd00c8 100644 (file)
--- a/tensorflow/contrib/lite/kernels/internal/optimized/depthwiseconv_uint8.h
+++ b/tensorflow/contrib/lite/kernels/internal/optimized/depthwiseconv_uint8.h
@@ -1697,7 +1697,9 @@ inline void DepthwiseConv(const uint8* input_data, const Dims<4>& input_dims,
 #endif
   TFLITE_DCHECK(output_depth == input_depth * depth_multiplier);
 
-#ifdef __aarch64__
+// Enable for arm64 except for the Nvidia Linux 4 Tegra (L4T) running on
+// Jetson TX-2. This compiler does not support the offsetof() macro.
+#if defined(__aarch64__) && !defined(GOOGLE_L4T)
   // Call kernel optimized for depthwise convolutions using 3x3 filters if
   // parameters are supported.
   if (Fast3x3FilterKernelSupported(

diff --git a/tensorflow/contrib/lite/kernels/internal/optimized/depthwiseconv_uint8_3x3_filter.h b/tensorflow/contrib/lite/kernels/internal/optimized/depthwiseconv_uint8_3x3_filter.h
index 51fbd54..8cd7223 100644 (file)
--- a/tensorflow/contrib/lite/kernels/internal/optimized/depthwiseconv_uint8_3x3_filter.h
+++ b/tensorflow/contrib/lite/kernels/internal/optimized/depthwiseconv_uint8_3x3_filter.h
@@ -23,7 +23,9 @@ limitations under the License.
 namespace tflite {
 namespace optimized_ops {
 
-#ifdef __aarch64__
+// Enable for arm64 except for the Nvidia Linux 4 Tegra (L4T) running on
+// Jetson TX-2. This compiler does not support the offsetof() macro.
+#if defined(__aarch64__) && !defined(GOOGLE_L4T)
 
 // clang-format gets confused with this file and ends up formatting lines to
 // be larger than 80 characters. Turn off here and back on at the end of the