Introduce cpu quant8 fully connected kernel (#4918)

Introduce cpu quantized int8 fully connected kernel from tflite
Use kernel in neurun cpu backend

Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>

1 files changed, 52 insertions, 0 deletions

diff --git a/libs/cker/include/cker/operation/FullyConnected.h b/libs/cker/include/cker/operation/FullyConnected.h
index b253b21fd..60218e34a 100644
--- a/libs/cker/include/cker/operation/FullyConnected.h
+++ b/libs/cker/include/cker/operation/FullyConnected.h
@@ -82,6 +82,58 @@ inline void FullyConnected(const FullyConnectedParams &params, const Shape &inpu
   }
 }
 
+inline void FullyConnected(const FullyConnectedParams &params, const Shape &input_shape,
+                           const uint8_t *input_data, const Shape &filter_shape,
+                           const uint8_t *filter_data, const Shape &bias_shape,
+                           const int32_t *bias_data, const Shape &output_shape,
+                           uint8_t *output_data)
+{
+  const int32_t input_offset = params.input_offset;
+  const int32_t filter_offset = params.weights_offset;
+  const int32_t output_offset = params.output_offset;
+  const int32_t output_multiplier = params.output_multiplier;
+  const int output_shift = params.output_shift;
+  const int32_t output_activation_min = params.quantized_activation_min;
+  const int32_t output_activation_max = params.quantized_activation_max;
+  TFLITE_DCHECK_GE(filter_shape.DimensionsCount(), 2);
+  TFLITE_DCHECK_GE(output_shape.DimensionsCount(), 1);
+
+  TFLITE_DCHECK_LE(output_activation_min, output_activation_max);
+  // TODO(benoitjacob): This really should be:
+  //     const int batches = ArraySize(output_dims, 1);
+  // but the current --variable_batch hack consists in overwriting the 3rd
+  // dimension with the runtime batch size, as we don't keep track for each
+  // array of which dimension is the batch dimension in it.
+  const int output_dim_count = output_shape.DimensionsCount();
+  const int filter_dim_count = filter_shape.DimensionsCount();
+  const int batches = FlatSizeSkipDim(output_shape, output_dim_count - 1);
+  const int output_depth =
+      MatchingDim(filter_shape, filter_dim_count - 2, output_shape, output_dim_count - 1);
+  const int accum_depth = filter_shape.Dims(filter_dim_count - 1);
+  for (int b = 0; b < batches; ++b)
+  {
+    for (int out_c = 0; out_c < output_depth; ++out_c)
+    {
+      int32_t acc = 0;
+      for (int d = 0; d < accum_depth; ++d)
+      {
+        int32_t input_val = input_data[b * accum_depth + d];
+        int32_t filter_val = filter_data[out_c * accum_depth + d];
+        acc += (filter_val + filter_offset) * (input_val + input_offset);
+      }
+      if (bias_data)
+      {
+        acc += bias_data[out_c];
+      }
+      acc = MultiplyByQuantizedMultiplier(acc, output_multiplier, output_shift);
+      acc += output_offset;
+      acc = std::max(acc, output_activation_min);
+      acc = std::min(acc, output_activation_max);
+      output_data[out_c + output_depth * b] = static_cast<uint8>(acc);
+    }
+  }
+}
+
 } // namespace cker
 } // namespace nnfw