COMPMID-420, COMPMID-414 - Port CLConvolutionLayer and CLFullyConnectedLayer to use 8 bit fixed point

Change-Id: I1cb1b4d7711ad7b569ee691e13a5df1b3430292b
Reviewed-on: http://mpd-gerrit.cambridge.arm.com/79565
Tested-by: Kaizen <jeremy.johnson+kaizengerrit@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>

1 files changed, 36 insertions, 37 deletions

diff --git a/src/runtime/CL/functions/CLConvolutionLayer.cpp b/src/runtime/CL/functions/CLConvolutionLayer.cpp
index b29bf8f136..96d04dc143 100644
--- a/src/runtime/CL/functions/CLConvolutionLayer.cpp
+++ b/src/runtime/CL/functions/CLConvolutionLayer.cpp
@@ -41,7 +41,7 @@ CLConvolutionLayerReshapeWeights::CLConvolutionLayerReshapeWeights()
 
 void CLConvolutionLayerReshapeWeights::configure(const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, bool transpose1xW)
 {
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(weights, 1, DataType::F16, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(weights, 1, DataType::QS8, 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);
@@ -63,8 +63,9 @@ void CLConvolutionLayerReshapeWeights::configure(const ICLTensor *weights, const
         const unsigned int mat_weights_cols = weights->info()->dimension(3);
         const unsigned int mat_weights_rows = weights->info()->dimension(0) * weights->info()->dimension(1) * weights->info()->dimension(2) + (_has_bias ? 1 : 0);
         TensorShape        shape_wr(mat_weights_cols, mat_weights_rows);
-        const DataType     dt = weights->info()->data_type();
-        TensorInfo         info_wr(shape_wr, 1, dt);
+        const DataType     dt                   = weights->info()->data_type();
+        const int          fixed_point_position = weights->info()->fixed_point_position();
+        TensorInfo         info_wr(shape_wr, 1, dt, fixed_point_position);
 
         _weights_reshaped.allocator()->init(info_wr);
         _weights_reshape_kernel.configure(weights, biases, &_weights_reshaped);
@@ -95,23 +96,27 @@ CLConvolutionLayer::CLConvolutionLayer()
 
 void CLConvolutionLayer::configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info, const WeightsInfo &weights_info)
 {
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, 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));
     ARM_COMPUTE_ERROR_ON(weights->info()->num_dimensions() > 4);
 
     if(biases != nullptr)
     {
-        ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(biases, 1, DataType::F16, DataType::F32);
         ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, biases);
+        ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, biases);
         ARM_COMPUTE_ERROR_ON(!weights_info.are_reshaped() && biases->info()->dimension(0) != weights->info()->dimension(3));
         ARM_COMPUTE_ERROR_ON(biases->info()->num_dimensions() > 1);
     }
 
+    const DataType dt                   = input->info()->data_type();
+    const int      fixed_point_position = input->info()->fixed_point_position();
+
     _has_bias             = (biases != nullptr);
     _are_weights_reshaped = weights_info.are_reshaped();
 
-    // Get parameters for conv_info
+    // Get parameters from conv_info
     unsigned int stride_x = 0;
     unsigned int stride_y = 0;
     unsigned int pad_x    = 0;
@@ -123,8 +128,8 @@ void CLConvolutionLayer::configure(const ICLTensor *input, const ICLTensor *weig
     unsigned int conv_w = 0;
     unsigned int conv_h = 0;
 
-    const unsigned int kernel_width  = _are_weights_reshaped ? weights_info.kernel_size().first : weights->info()->dimension(0);
-    const unsigned int kernel_height = _are_weights_reshaped ? weights_info.kernel_size().second : weights->info()->dimension(1);
+    const unsigned int kernel_width  = (_are_weights_reshaped) ? weights_info.kernel_size().first : weights->info()->dimension(0);
+    const unsigned int kernel_height = (_are_weights_reshaped) ? weights_info.kernel_size().second : weights->info()->dimension(1);
     std::tie(conv_w, conv_h) = scaled_dimensions(input->info()->dimension(0), input->info()->dimension(1), kernel_width, kernel_height,
                                                  conv_info);
     ARM_COMPUTE_ERROR_ON_MSG((output->info()->dimension(0) != conv_w) || (output->info()->dimension(1) != conv_h), "Output shape does not match the expected one");
@@ -132,9 +137,10 @@ void CLConvolutionLayer::configure(const ICLTensor *input, const ICLTensor *weig
     // Check if its a "fully connected" convolution
     _is_fully_connected_convolution = ((conv_w == 1) && (conv_h == 1));
 
-    // Create tensor to store the reshaped weights
-    size_t mat_weights_cols = weights->info()->dimension(3);
-    size_t mat_weights_rows = weights->info()->dimension(0) * weights->info()->dimension(1) * weights->info()->dimension(2) + ((_has_bias) ? 1 : 0);
+    unsigned int mat_weights_cols = weights->info()->dimension(3);
+    unsigned int mat_weights_rows = weights->info()->dimension(0) * weights->info()->dimension(1) * weights->info()->dimension(2) + (_has_bias ? 1 : 0);
+
+    // Reshape weights if needed
     if(_are_weights_reshaped)
     {
         mat_weights_cols                         = output->info()->dimension(2);
@@ -147,49 +153,48 @@ void CLConvolutionLayer::configure(const ICLTensor *input, const ICLTensor *weig
         {
             // Create tensor to store the reshaped weights
             TensorShape shape_wr(mat_weights_cols, mat_weights_rows);
-            TensorInfo  info_wr(shape_wr, 1, weights->info()->data_type());
+            TensorInfo  info_wr(shape_wr, 1, dt, fixed_point_position);
             _weights_reshaped.allocator()->init(info_wr);
-            _reshape_weights.configure(weights, biases, &_weights_reshaped, false);
-            weights = &_weights_reshaped;
+            _reshape_weights.configure(weights, biases, &_weights_reshaped, false /* 1xW transpose */);
         }
         else
         {
             // Create tensor to store transposed weights
-            TensorShape shape_wt(mat_weights_rows * 4, static_cast<size_t>(std::ceil(mat_weights_cols / 4.f)));
-            TensorInfo  info_wt(shape_wt, 1, weights->info()->data_type());
-            _weights_transposed.allocator()->init(info_wt);
-            _reshape_weights.configure(weights, biases, &_weights_transposed, true);
-            weights = &_weights_transposed;
+            const float transpose_width = 16.0f / input->info()->element_size();
+            TensorShape shape_wt(mat_weights_rows * static_cast<unsigned int>(transpose_width), static_cast<unsigned int>(std::ceil(mat_weights_cols / transpose_width)));
+            TensorInfo  info_wt(shape_wt, 1, dt, fixed_point_position);
+            _weights_reshaped.allocator()->init(info_wt);
+            _reshape_weights.configure(weights, biases, &_weights_reshaped, true /* 1xW transpose */);
         }
+        weights = &_weights_reshaped;
     }
+
     // Create tensor to store im2col reshaped inputs
-    const size_t mat_input_cols = mat_weights_rows;
-    const size_t mat_input_rows = conv_w * conv_h;
-    TensorShape  shape_im2col   = input->info()->tensor_shape();
+    const unsigned int mat_input_cols = mat_weights_rows;
+    const unsigned int mat_input_rows = conv_w * conv_h;
+    TensorShape        shape_im2col   = input->info()->tensor_shape();
     shape_im2col.set(0, mat_input_cols);
     shape_im2col.set(1, mat_input_rows);
     shape_im2col.set(2, 1);
-    _input_im2col_reshaped.allocator()->init(TensorInfo(shape_im2col, 1, input->info()->data_type()));
+    _input_im2col_reshaped.allocator()->init(TensorInfo(shape_im2col, 1, dt, fixed_point_position));
 
     // Create tensor (interleave) to prepare input tensor for GEMM
     if(!_is_fully_connected_convolution)
     {
         TensorShape shape_interleaved = shape_im2col;
         shape_interleaved.set(0, shape_interleaved.x() * 4);
-        shape_interleaved.set(1, std::ceil(static_cast<float>(shape_interleaved.y()) / 4.f));
-        _input_interleaved_reshaped.allocator()->init(TensorInfo(shape_interleaved, 1, input->info()->data_type()));
+        shape_interleaved.set(1, std::ceil(shape_interleaved.y() / 4.f));
+        _input_interleaved_reshaped.allocator()->init(TensorInfo(shape_interleaved, 1, dt, fixed_point_position));
     }
 
     // Create GEMM output tensor
     TensorShape shape_gemm = _input_im2col_reshaped.info()->tensor_shape();
     shape_gemm.set(0, mat_weights_cols);
     shape_gemm.set(1, mat_input_rows);
-    _gemm_output.allocator()->init(TensorInfo(shape_gemm, 1, input->info()->data_type()));
+    _gemm_output.allocator()->init(TensorInfo(shape_gemm, 1, dt, fixed_point_position));
 
     // Configure kernels
     _input_im2col_kernel.configure(input, &_input_im2col_reshaped, Size2D(kernel_width, kernel_height), conv_info, _has_bias);
-    _output_col2im_kernel.configure(&_gemm_output, output, std::make_pair(conv_w, conv_h));
-
     if(_is_fully_connected_convolution)
     {
         _mm_kernel.configure(&_input_im2col_reshaped, weights, &_gemm_output, 1.0f);
@@ -199,19 +204,13 @@ void CLConvolutionLayer::configure(const ICLTensor *input, const ICLTensor *weig
         _input_interleave_kernel.configure(&_input_im2col_reshaped, &_input_interleaved_reshaped);
         _mm_kernel.configure(&_input_interleaved_reshaped, weights, &_gemm_output, 1.0f);
     }
+    _output_col2im_kernel.configure(&_gemm_output, output, std::make_pair(conv_w, conv_h));
 
+    // Allocate intermediate tensor
     if(!_are_weights_reshaped)
     {
-        if(!_is_fully_connected_convolution)
-        {
-            _weights_transposed.allocator()->allocate();
-        }
-        else
-        {
-            _weights_reshaped.allocator()->allocate();
-        }
+        _weights_reshaped.allocator()->allocate();
     }
-
     _input_im2col_reshaped.allocator()->allocate();
     if(!_is_fully_connected_convolution)
     {