Add fast shuffled fully-connected path also for the case where the

RHS has 4 columns (so far was only for the GEMV case where RHS has
1 column).

Also pre-shuffle / pre-xor the input activations, not just the
weights. We need a buffer for that, so the shuffled FullyConnected
operator gets a second output acting as its workspace, similar
to what we have been doing for Conv operators needed a im2col
workspace buffer.

PiperOrigin-RevId: 193958461

1 files changed, 1 insertions, 14 deletions

diff --git a/tensorflow/contrib/lite/toco/tooling_util.cc b/tensorflow/contrib/lite/toco/tooling_util.cc
index cf2cbeedc7..5a341294db 100644
--- a/tensorflow/contrib/lite/toco/tooling_util.cc
+++ b/tensorflow/contrib/lite/toco/tooling_util.cc
@@ -1405,20 +1405,7 @@ void ResolveModelFlags(const ModelFlags& model_flags, Model* model) {
     }
     input_minmax.min = (qmin - mean_value) / std_value;
     input_minmax.max = (qmax - mean_value) / std_value;
-    if (input_array.minmax) {
-      if (input_array_proto.has_mean_value() ||
-          input_array_proto.has_std_value()) {
-        const double width = input_minmax.max - input_minmax.min;
-        const double kMinMaxAllowedDiff = 1e-6 * width;
-        CHECK(std::abs(input_minmax.min - input_array.minmax->min) <
-                  kMinMaxAllowedDiff &&
-              std::abs(input_minmax.max - input_array.minmax->max) <
-                  kMinMaxAllowedDiff)
-            << input_minmax.min << ", " << input_minmax.max
-            << " != " << input_array.minmax->min << ", "
-            << input_array.minmax->max;
-      }
-    } else {
+    if (!input_array.minmax) {
       input_array.GetOrCreateMinMax() = input_minmax;
     }
   }