COMPMID-436, COMPMID-437 - Port NEConvolutionLayer & NEFullyConnectedLayer to support 16 bit fixed point

Change-Id: I69edf2dac242f941bac95c8479d921e7be6abca7
Reviewed-on: http://mpd-gerrit.cambridge.arm.com/79725
Tested-by: Kaizen <jeremy.johnson+kaizengerrit@arm.com>
Reviewed-by: Pablo Tello <pablo.tello@arm.com>

7 files changed, 50 insertions, 20 deletions

diff --git a/src/core/NEON/kernels/NECol2ImKernel.cpp b/src/core/NEON/kernels/NECol2ImKernel.cpp
index e9a73607e6..95a9364082 100644
--- a/src/core/NEON/kernels/NECol2ImKernel.cpp
+++ b/src/core/NEON/kernels/NECol2ImKernel.cpp
@@ -69,7 +69,8 @@ NECol2ImKernel::NECol2ImKernel()
 
 void NECol2ImKernel::configure(const ITensor *input, ITensor *output, std::pair<unsigned int, unsigned int> convolved_dims)
 {
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8, DataType::S8, DataType::QS8, DataType::U16, DataType::S16, DataType::U32, DataType::S32, DataType::F16, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8, DataType::S8, DataType::QS8, DataType::U16, DataType::S16, DataType::QS16, DataType::U32, DataType::S32, DataType::F16,
+                                                  DataType::F32);
     ARM_COMPUTE_ERROR_ON_NULLPTR(output);
 
     TensorShape output_shape = input->info()->tensor_shape();
diff --git a/src/core/NEON/kernels/NEGEMMMatrixAccumulateBiasesKernel.cpp b/src/core/NEON/kernels/NEGEMMMatrixAccumulateBiasesKernel.cpp
index 7a3bae50c0..826a386557 100644
--- a/src/core/NEON/kernels/NEGEMMMatrixAccumulateBiasesKernel.cpp
+++ b/src/core/NEON/kernels/NEGEMMMatrixAccumulateBiasesKernel.cpp
@@ -45,9 +45,9 @@ NEGEMMMatrixAccumulateBiasesKernel::NEGEMMMatrixAccumulateBiasesKernel()
 
 void NEGEMMMatrixAccumulateBiasesKernel::configure(ITensor *accum, const ITensor *biases)
 {
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(accum, 1, DataType::QS8, DataType::F32);
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(biases, 1, DataType::QS8, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(accum, 1, DataType::QS8, DataType::QS16, DataType::F32);
     ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(biases, accum);
+    ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT_POSITION(biases, accum);
     ARM_COMPUTE_ERROR_ON(biases->info()->num_dimensions() != 1);
 
     _biases = biases;
@@ -121,6 +121,21 @@ void NEGEMMMatrixAccumulateBiasesKernel::run(const Window &window)
             in0_out, in1);
             break;
         }
+        case DataType::QS16:
+        {
+            execute_window_loop(window, [&](const Coordinates & id)
+            {
+                qint16x8x2_t       accum  = vld2q_s16(reinterpret_cast<const qint16_t *>(in0_out.ptr()));
+                const qint16x8x2_t biases = vld2q_s16(reinterpret_cast<const qint16_t *>(in1.ptr()));
+
+                accum.val[0] = vqaddq_qs16(accum.val[0], biases.val[0]);
+                accum.val[1] = vqaddq_qs16(accum.val[1], biases.val[1]);
+
+                vst2q_s16(reinterpret_cast<qint16_t *>(in0_out.ptr()), accum);
+            },
+            in0_out, in1);
+            break;
+        }
         default:
             ARM_COMPUTE_ERROR("Data type not supported");
             break;
diff --git a/src/core/NEON/kernels/NEIm2ColKernel.cpp b/src/core/NEON/kernels/NEIm2ColKernel.cpp
index 8c9d12c57c..5bb8b1c22a 100644
--- a/src/core/NEON/kernels/NEIm2ColKernel.cpp
+++ b/src/core/NEON/kernels/NEIm2ColKernel.cpp
@@ -134,10 +134,14 @@ inline void linearize_volume(const uint8_t *const in_ptr,
     // Append 1 if the convolution layer has biases
     if(has_bias)
     {
-        if(std::is_same<T, arm_compute::qint8_t>::value)
+        if(std::is_same<T, qint8_t>::value)
         {
             *out_ptr = scvt_qs8_f32(1.0f, fixed_point_position);
         }
+        else if(std::is_same<T, qint16_t>::value)
+        {
+            *out_ptr = scvt_qs16_f32(1.0f, fixed_point_position);
+        }
         else
         {
             *out_ptr = static_cast<T>(1);
@@ -249,10 +253,14 @@ void NEIm2ColKernel::run_reduced(const Window &window)
         // Add bias
         if(_has_bias)
         {
-            if(std::is_same<T, arm_compute::qint8_t>::value)
+            if(std::is_same<T, qint8_t>::value)
             {
                 *(reinterpret_cast<T *>(out_ptr) + out_width - 1) = scvt_qs8_f32(1.0f, _input->info()->fixed_point_position());
             }
+            else if(std::is_same<T, qint16_t>::value)
+            {
+                *(reinterpret_cast<T *>(out_ptr) + out_width - 1) = scvt_qs16_f32(1.0f, _input->info()->fixed_point_position());
+            }
             else
             {
                 *(reinterpret_cast<T *>(out_ptr) + out_width - 1) = static_cast<T>(1);
@@ -269,8 +277,9 @@ NEIm2ColKernel::NEIm2ColKernel()
 
 void NEIm2ColKernel::configure(const ITensor *input, ITensor *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias)
 {
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32, DataType::QS8);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32, DataType::QS8, DataType::QS16);
     ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
+    ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT_POSITION(input, output);
 
     _input          = input;
     _output         = output;
@@ -309,6 +318,9 @@ void NEIm2ColKernel::configure(const ITensor *input, ITensor *output, const Size
             case DataType::QS8:
                 _func = &NEIm2ColKernel::run_reduced<qint8_t>;
                 break;
+            case DataType::QS16:
+                _func = &NEIm2ColKernel::run_reduced<qint16_t>;
+                break;
             default:
                 ARM_COMPUTE_ERROR("Data type not supported");
                 break;
@@ -329,6 +341,9 @@ void NEIm2ColKernel::configure(const ITensor *input, ITensor *output, const Size
             case DataType::QS8:
                 _func = ((pad_x == 0) && (pad_y == 0)) ? &NEIm2ColKernel::run_generic<qint8_t, false> : &NEIm2ColKernel::run_generic<qint8_t, true>;
                 break;
+            case DataType::QS16:
+                _func = ((pad_x == 0) && (pad_y == 0)) ? &NEIm2ColKernel::run_generic<qint16_t, false> : &NEIm2ColKernel::run_generic<qint16_t, true>;
+                break;
             default:
                 ARM_COMPUTE_ERROR("Data type not supported");
                 break;
diff --git a/src/core/NEON/kernels/NETransposeKernel.cpp b/src/core/NEON/kernels/NETransposeKernel.cpp
index a990e9068e..732a0ef4f6 100644
--- a/src/core/NEON/kernels/NETransposeKernel.cpp
+++ b/src/core/NEON/kernels/NETransposeKernel.cpp
@@ -179,7 +179,8 @@ NETransposeKernel::NETransposeKernel()
 
 void NETransposeKernel::configure(const ITensor *input, ITensor *output)
 {
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8, DataType::S8, DataType::QS8, DataType::U16, DataType::S16, DataType::U32, DataType::S32, DataType::F16, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8, DataType::S8, DataType::QS8, DataType::U16, DataType::S16, DataType::QS16, DataType::U32, DataType::S32, DataType::F16,
+                                                  DataType::F32);
     ARM_COMPUTE_ERROR_ON_NULLPTR(output);
 
     TensorShape  output_shape{ input->info()->tensor_shape() };
diff --git a/src/core/NEON/kernels/NEWeightsReshapeKernel.cpp b/src/core/NEON/kernels/NEWeightsReshapeKernel.cpp
index ac688e1381..d685ec7962 100644
--- a/src/core/NEON/kernels/NEWeightsReshapeKernel.cpp
+++ b/src/core/NEON/kernels/NEWeightsReshapeKernel.cpp
@@ -95,7 +95,7 @@ NEWeightsReshapeKernel::NEWeightsReshapeKernel()
 
 void NEWeightsReshapeKernel::configure(const ITensor *input, const ITensor *bias, ITensor *output)
 {
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::F16, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QS16, DataType::F16, DataType::F32);
     ARM_COMPUTE_ERROR_ON_NULLPTR(output);
 
     const int          fixed_point_position = input->info()->fixed_point_position();
@@ -129,26 +129,26 @@ void NEWeightsReshapeKernel::configure(const ITensor *input, const ITensor *bias
     _bias   = bias;
     _output = output;
 
-    switch(_input->info()->data_type())
+    switch(_input->info()->element_size())
     {
-        case DataType::F32:
+        case 4:
         {
             _func = &weights_reshape<uint32_t>;
             break;
         }
-        case DataType::F16:
+        case 2:
         {
             _func = &weights_reshape<uint16_t>;
             break;
         }
-        case DataType::QS8:
+        case 1:
         {
             _func = &weights_reshape<uint8_t>;
             break;
         }
         default:
         {
-            ARM_COMPUTE_ERROR_ON("Data type not supported");
+            ARM_COMPUTE_ERROR_ON("Element size not supported");
             break;
         }
     }
diff --git a/src/runtime/NEON/functions/NEConvolutionLayer.cpp b/src/runtime/NEON/functions/NEConvolutionLayer.cpp
index dc8652747f..f6481f1918 100644
--- a/src/runtime/NEON/functions/NEConvolutionLayer.cpp
+++ b/src/runtime/NEON/functions/NEConvolutionLayer.cpp
@@ -41,14 +41,13 @@ NEConvolutionLayerReshapeWeights::NEConvolutionLayerReshapeWeights()
 
 void NEConvolutionLayerReshapeWeights::configure(const ITensor *weights, const ITensor *biases, ITensor *output, bool transpose1xW)
 {
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(weights, 1, DataType::QS8, DataType::F16, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(weights, 1, DataType::QS8, DataType::QS16, DataType::F16, DataType::F32);
     ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(weights, output);
     ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(weights, output);
     ARM_COMPUTE_ERROR_ON(weights->info()->num_dimensions() > 4);
 
     if(biases != nullptr)
     {
-        ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(biases, 1, DataType::QS8, DataType::F16, DataType::F32);
         ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(weights, biases);
         ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(weights, biases);
         ARM_COMPUTE_ERROR_ON(biases->info()->dimension(0) != weights->info()->dimension(3));
@@ -96,7 +95,7 @@ NEConvolutionLayer::NEConvolutionLayer()
 
 void NEConvolutionLayer::configure(const ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const PadStrideInfo &conv_info, const WeightsInfo &weights_info)
 {
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::F16, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QS16, DataType::F16, DataType::F32);
     ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights, output);
     ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, weights, output);
     ARM_COMPUTE_ERROR_ON(!weights_info.are_reshaped() && weights->info()->dimension(2) != input->info()->dimension(2));
diff --git a/src/runtime/NEON/functions/NEFullyConnectedLayer.cpp b/src/runtime/NEON/functions/NEFullyConnectedLayer.cpp
index 6e27ed344a..eb84ccaddc 100644
--- a/src/runtime/NEON/functions/NEFullyConnectedLayer.cpp
+++ b/src/runtime/NEON/functions/NEFullyConnectedLayer.cpp
@@ -39,7 +39,7 @@ NEFullyConnectedLayerReshapeWeights::NEFullyConnectedLayerReshapeWeights()
 
 void NEFullyConnectedLayerReshapeWeights::configure(const ITensor *input, ITensor *output, bool transpose_weights, bool is_batched_fc_layer)
 {
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QS16, DataType::F32);
     ARM_COMPUTE_ERROR_ON(output == nullptr);
     ARM_COMPUTE_ERROR_ON(input->info()->num_dimensions() != 2);
     ARM_COMPUTE_ERROR_ON((transpose_weights == false) && (is_batched_fc_layer == false));
@@ -196,10 +196,9 @@ void NEFullyConnectedLayer::configure_fc_fc_nb(const ITensor *input, const ITens
 
 void NEFullyConnectedLayer::configure(const ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, bool transpose_weights, bool are_weights_reshaped)
 {
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::F32);
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(weights, 1, DataType::QS8, DataType::F32);
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::QS8, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QS16, DataType::F32);
     ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights, output);
+    ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT_POSITION(input, weights, output);
     ARM_COMPUTE_ERROR_ON(weights->info()->num_dimensions() != 2);
 
     const DataType dt                   = input->info()->data_type();