Update CMSIS-NN vector_sum_s8 calls to new API (#2685)

BUG=Add rhs_offset argument

Author: AdrianLundell

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

@@ -147,8 +147,11 @@ TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) {
         data->kernel_sums = static_cast<int32_t*>(
             context->AllocatePersistentBuffer(context, buf_size));
 
+        int32_t input_offset = -data->reference_op_data.input_zero_point;
+        int32_t filter_offset = -data->reference_op_data.filter_zero_point;
         arm_vector_sum_s8(data->kernel_sums, filter_dims.n, data->output_depth,
-                          filter_data, 1, nullptr);
+                          filter_data, input_offset, filter_offset,
+                          tflite::GetTensorData<int32_t>(bias));
 
         // Do not request a scratch buffer since using persistent memory
         buf_size = 0;
@@ -321,7 +324,8 @@ TfLiteStatus EvalQuantizedInt8(TfLiteContext* context, TfLiteNode* node,
       // 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);
+          tflite::micro::GetTensorData<int8_t>(filter), fc_params.input_offset,
+          fc_params.filter_offset, bias_data);
     }
 
     TF_LITE_ENSURE_EQ(

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

@@ -193,7 +193,9 @@ TfLiteStatus CmsisNnPrepareSvdf(TfLiteContext* context, TfLiteNode* node) {
           context->AllocatePersistentBuffer(context, buf_size));
 
       arm_vector_sum_s8(data->kernel_sums, input_size, num_filters,
-                        GetTensorData<int8_t>(weights_feature), 1, nullptr);
+                        GetTensorData<int8_t>(weights_feature),
+                        -data->input_zero_point,
+                        -data->activation_state_zero_point, nullptr);
     }
 
   } else {

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

@@ -53,7 +53,7 @@ LSTMBuffers<int16_t> CMSIS_NN_CreateLSTMBuffers(TfLiteContext* context,
 void CMSIS_NN_VectorSum(int32_t* kernel_sum, const int32_t size1,
                         const int32_t size2, const int8_t* weights,
                         const int32_t offset, const int32_t* biases) {
-  arm_vector_sum_s8(kernel_sum, size1, size2, weights, offset, biases);
+  arm_vector_sum_s8(kernel_sum, size1, size2, weights, offset, 0, biases);
 }
 
 void CMSIS_NN_VectorSum(int64_t* kernel_sum, const int32_t size1,

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

@@ -38,9 +38,9 @@ source ${TENSORFLOW_ROOT}tensorflow/lite/micro/tools/make/bash_helpers.sh
 DOWNLOADS_DIR=${1}
 DOWNLOADED_CMSIS_NN_PATH=${DOWNLOADS_DIR}/cmsis_nn
 
-ZIP_PREFIX_NN="95f293df19c9a38806868fe12a64a4f9b457f9c1"
+ZIP_PREFIX_NN="f2cb41ca1450a4eb4307b2779dd5aae9028285a5"
 CMSIS_NN_URL="http://github.com/ARM-software/CMSIS-NN/archive/${ZIP_PREFIX_NN}.zip"
-CMSIS_NN_MD5="5e0c4cd60a5f074c4d26d1be236caefd"
+CMSIS_NN_MD5="4d0e623432d6f8d3b201cbcd89218adf"
 
 should_download=$(check_should_download ${DOWNLOADS_DIR})