cmsis-nn: update fully connected int8 (#2469)

- Adds non zero filter offset support.
- Adds support for batch matmul like behavior where weights are like input, i.e. not initialized before eval.

BUG=non zero filter offset not supported for CMSIS-NN

tensorflow/lite/micro/kernels/cmsis_nn/fully_connected.cc

diff --git a/tensorflow/lite/micro/kernels/cmsis_nn/fully_connected.cc b/tensorflow/lite/micro/kernels/cmsis_nn/fully_connected.cc
index 0c4f8aa..dc7b78c 100644
--- a/tensorflow/lite/micro/kernels/cmsis_nn/fully_connected.cc
+++ b/tensorflow/lite/micro/kernels/cmsis_nn/fully_connected.cc
@@ -104,9 +104,7 @@
     TF_LITE_ENSURE_EQ(context, input->params.zero_point, 0);
     TF_LITE_ENSURE_EQ(context, output->params.zero_point, 0);
     buf_size = arm_fully_connected_s16_get_buffer_size(&filter_dims);
-  } else if (input->type == kTfLiteInt8 &&
-             data->reference_op_data.filter_zero_point == 0 &&
-             filter->type != kTfLiteInt4) {
+  } else if (input->type == kTfLiteInt8 && filter->type != kTfLiteInt4) {
     const RuntimeShape input_shape = GetTensorShape(input);
 
     TFLITE_DCHECK_GE(output_dim_count, 2);
@@ -130,11 +128,13 @@
     } else {
       buf_size = arm_fully_connected_s8_get_buffer_size(&filter_dims);
 
-      if (buf_size > 0) {
+      int8_t* filter_data = GetTensorData<int8_t>(filter);
+      data->kernel_sums = nullptr;
+
+      if (buf_size > 0 && filter_data != nullptr) {
         data->kernel_sums = static_cast<int32_t*>(
             context->AllocatePersistentBuffer(context, buf_size));
 
-        int8_t* filter_data = GetTensorData<int8_t>(filter);
         arm_vector_sum_s8(data->kernel_sums, filter_dims.n, data->output_depth,
                           filter_data, 1, nullptr);
 
@@ -298,12 +298,20 @@
   } else {
     cmsis_nn_fc_params fc_params;
     fc_params.input_offset = -data.reference_op_data.input_zero_point;
+    fc_params.filter_offset = -data.reference_op_data.filter_zero_point;
     fc_params.output_offset = data.reference_op_data.output_zero_point;
-    fc_params.filter_offset = 0;
     fc_params.activation.min = data.reference_op_data.output_activation_min;
     fc_params.activation.max = data.reference_op_data.output_activation_max;
 
-    ctx.buf = data.kernel_sums;
+    if (data.kernel_sums != nullptr) {
+      ctx.buf = data.kernel_sums;
+    } else if (ctx.buf != nullptr) {
+      // If behaving like batch matmul we calculate kernel sums in eval.
+      arm_vector_sum_s8(
+          static_cast<int32_t*>(ctx.buf), filter_dims.n, data.output_depth,
+          tflite::micro::GetTensorData<int8_t>(filter), 1, nullptr);
+    }
+
     TF_LITE_ENSURE_EQ(
         context,
         arm_fully_connected_s8(
@@ -393,22 +401,8 @@
           return EvalQuantizedInt4(context, node, data, input, filter, bias,
                                    output);
         case kTfLiteInt8:
-          if (data.reference_op_data.filter_zero_point == 0) {
-            return EvalQuantizedInt8(context, node, data, input, filter, bias,
-                                     output);
-          } else {
-            tflite::reference_integer_ops::FullyConnected(
-                FullyConnectedParamsQuantized(data.reference_op_data),
-                tflite::micro::GetTensorShape(input),
-                tflite::micro::GetTensorData<int8_t>(input),
-                tflite::micro::GetTensorShape(filter),
-                tflite::micro::GetTensorData<int8_t>(filter),
-                tflite::micro::GetTensorShape(bias),
-                tflite::micro::GetOptionalTensorData<int32_t>(bias),
-                tflite::micro::GetTensorShape(output),
-                tflite::micro::GetTensorData<int8_t>(output));
-            return kTfLiteOk;
-          }
+          return EvalQuantizedInt8(context, node, data, input, filter, bias,
+                                   output);
         default:
           MicroPrintf("Filter Type %s (%d) not supported.",
                       TfLiteTypeGetName(filter->type), filter->type);

tensorflow/lite/micro/tools/make/ext_libs/cmsis_nn_download.sh

diff --git a/tensorflow/lite/micro/tools/make/ext_libs/cmsis_nn_download.sh b/tensorflow/lite/micro/tools/make/ext_libs/cmsis_nn_download.sh
index cc79116..601c4d4 100755
--- a/tensorflow/lite/micro/tools/make/ext_libs/cmsis_nn_download.sh
+++ b/tensorflow/lite/micro/tools/make/ext_libs/cmsis_nn_download.sh
@@ -47,9 +47,9 @@
   echo >&2 "${DOWNLOADED_CMSIS_NN_PATH} already exists, skipping the download."
 else
 
-  ZIP_PREFIX_NN="2a999a2fd887c98042353accac77479f00b5f99d"
+  ZIP_PREFIX_NN="72e1ebf623ab1660a3e14e4e36fdcddce46f1991"
   CMSIS_NN_URL="http://github.com/ARM-software/CMSIS-NN/archive/${ZIP_PREFIX_NN}.zip"
-  CMSIS_NN_MD5="c6cfe1f8e0f6518c92f7e42ed7b7afd4"
+  CMSIS_NN_MD5="23a623f4eca6c8f11ee5366c2cf61a44"
 
   # wget is much faster than git clone of the entire repo. So we wget a specific
   # version and can then apply a patch, as needed.