arm_compute v18.05
diff --git a/src/runtime/NEON/functions/NELocallyConnectedLayer.cpp b/src/runtime/NEON/functions/NELocallyConnectedLayer.cpp
index 45ddb70..913acf8 100644
--- a/src/runtime/NEON/functions/NELocallyConnectedLayer.cpp
+++ b/src/runtime/NEON/functions/NELocallyConnectedLayer.cpp
@@ -33,39 +33,102 @@
using namespace arm_compute;
+namespace
+{
+void calculate_shapes(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info,
+ TensorShape &shape_wr, TensorShape &shape_im2col, TensorShape &shape_gemm)
+{
+ ARM_COMPUTE_UNUSED(output);
+
+ const unsigned int kernel_width = weights->dimension(0);
+ const unsigned int kernel_height = weights->dimension(1);
+
+ bool has_bias = (biases != nullptr);
+
+ // Get convolved dimensions
+ unsigned int conv_w = 0;
+ unsigned int conv_h = 0;
+ std::tie(conv_w, conv_h) = scaled_dimensions(input->dimension(0), input->dimension(1), kernel_width, kernel_height,
+ conv_info);
+
+ const size_t mat_weights_cols = weights->dimension(3);
+ const size_t mat_weights_rows = weights->dimension(0) * weights->dimension(1) * weights->dimension(2) + ((has_bias) ? 1 : 0);
+ const size_t mat_weights_num = weights->dimension(4);
+
+ shape_wr = TensorShape(mat_weights_cols, mat_weights_rows, mat_weights_num);
+
+ const size_t mat_input_cols = mat_weights_rows;
+ const size_t mat_input_rows = conv_w * conv_h;
+
+ shape_im2col = input->tensor_shape();
+ shape_im2col.set(0, mat_input_cols);
+ shape_im2col.set(1, mat_input_rows);
+ shape_im2col.set(2, 1);
+
+ shape_gemm = shape_im2col;
+ shape_gemm.set(0, mat_weights_cols);
+ shape_gemm.set(1, mat_input_rows);
+}
+} // namespace
+
NELocallyConnectedLayer::NELocallyConnectedLayer(std::shared_ptr<IMemoryManager> memory_manager)
: _memory_group(std::move(memory_manager)), _input_im2col_kernel(), _weights_reshape_kernel(), _mm_kernel(), _output_col2im_kernel(), _input_im2col_reshaped(), _weights_reshaped(), _gemm_output(),
- _is_first_run(false)
+ _is_first_run(false), _original_weights(nullptr)
{
}
+Status NELocallyConnectedLayer::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info)
+{
+ ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, weights, output);
+ ARM_COMPUTE_RETURN_ERROR_ON(weights->dimension(2) != input->dimension(2));
+ ARM_COMPUTE_RETURN_ERROR_ON(!conv_info.padding_is_symmetric());
+
+ bool has_bias = (biases != nullptr);
+
+ if(has_bias)
+ {
+ ARM_COMPUTE_RETURN_ERROR_ON(biases->dimension(0) != weights->dimension(3));
+ ARM_COMPUTE_RETURN_ERROR_ON(biases->num_dimensions() > 2);
+ }
+
+ const unsigned int kernel_width = weights->dimension(0);
+ const unsigned int kernel_height = weights->dimension(1);
+
+ // Get convolved dimensions
+ unsigned int conv_w = 0;
+ unsigned int conv_h = 0;
+ std::tie(conv_w, conv_h) = scaled_dimensions(input->dimension(0), input->dimension(1), kernel_width, kernel_height,
+ conv_info);
+
+ ARM_COMPUTE_RETURN_ERROR_ON_MSG((output->dimension(0) != conv_w) || (output->dimension(1) != conv_h), "Output shape does not match the expected one");
+ ARM_COMPUTE_RETURN_ERROR_ON_MSG(weights->dimension(4) != (conv_w * conv_h), "Weights shape does not match the expected one");
+
+ // Calculate intermediate buffer shapes
+ TensorShape shape_wr;
+ TensorShape shape_im2col;
+ TensorShape shape_gemm;
+ calculate_shapes(input, weights, biases, output, conv_info, shape_wr, shape_im2col, shape_gemm);
+
+ TensorInfo weights_reshaped_info(shape_wr, 1, weights->data_type());
+ TensorInfo input_im2col_reshaped_info(shape_im2col, 1, input->data_type());
+ TensorInfo gemm_output_info(shape_gemm, 1, input->data_type());
+
+ ARM_COMPUTE_RETURN_ON_ERROR(NEIm2ColKernel::validate(input, &input_im2col_reshaped_info, Size2D(kernel_width, kernel_height), conv_info, has_bias, false));
+ ARM_COMPUTE_RETURN_ON_ERROR(NEWeightsReshapeKernel::validate(weights, biases, &weights_reshaped_info));
+ ARM_COMPUTE_RETURN_ON_ERROR(NELocallyConnectedMatrixMultiplyKernel::validate(&input_im2col_reshaped_info, &weights_reshaped_info, &gemm_output_info));
+ ARM_COMPUTE_RETURN_ON_ERROR(NECol2ImKernel::validate(&gemm_output_info, output, Size2D(conv_w, conv_h)));
+
+ return Status{};
+}
+
void NELocallyConnectedLayer::configure(const ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const PadStrideInfo &conv_info)
{
- ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32);
- ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(weights, 1, DataType::F32);
- ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::F32);
- ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights, output);
- ARM_COMPUTE_ERROR_ON(weights->info()->dimension(2) != input->info()->dimension(2));
- ARM_COMPUTE_ERROR_ON(!conv_info.padding_is_symmetric());
+ ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, output);
+ ARM_COMPUTE_ERROR_THROW_ON(NELocallyConnectedLayer::validate(input->info(), weights->info(), biases == nullptr ? nullptr : biases->info(), output->info(), conv_info));
- if(biases != nullptr)
- {
- ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(biases, 1, DataType::F32);
- ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, biases);
- ARM_COMPUTE_ERROR_ON(biases->info()->dimension(0) != weights->info()->dimension(3));
- ARM_COMPUTE_ERROR_ON(biases->info()->num_dimensions() > 2);
- }
-
- bool _has_bias = (biases != nullptr);
- _is_first_run = true;
-
- // Get parameters for conv_info
- unsigned int stride_x = 0;
- unsigned int stride_y = 0;
- unsigned int pad_x = 0;
- unsigned int pad_y = 0;
- std::tie(stride_x, stride_y) = conv_info.stride();
- std::tie(pad_x, pad_y) = conv_info.pad();
+ bool _has_bias = (biases != nullptr);
+ _is_first_run = true;
+ _original_weights = weights;
const unsigned int kernel_width = weights->info()->dimension(0);
const unsigned int kernel_height = weights->info()->dimension(1);
@@ -76,32 +139,14 @@
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");
- ARM_COMPUTE_ERROR_ON_MSG(weights->info()->dimension(4) != (conv_w * conv_h), "Weights shape does not match the expected one");
-
- // Create tensor to store the reshaped weights
- const size_t mat_weights_cols = weights->info()->dimension(3);
- const size_t mat_weights_rows = weights->info()->dimension(0) * weights->info()->dimension(1) * weights->info()->dimension(2) + ((_has_bias) ? 1 : 0);
- const size_t mat_weights_num = weights->info()->dimension(4);
-
- const TensorShape shape_wr(mat_weights_cols, mat_weights_rows, mat_weights_num);
+ // Calculate intermediate buffer shapes
+ TensorShape shape_wr;
+ TensorShape shape_im2col;
+ TensorShape shape_gemm;
+ calculate_shapes(input->info(), weights->info(), biases == nullptr ? nullptr : biases->info(), output->info(), conv_info, shape_wr, shape_im2col, shape_gemm);
_weights_reshaped.allocator()->init(TensorInfo(shape_wr, 1, weights->info()->data_type()));
-
- // 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();
- 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()));
-
- // Create locally connected layer 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()));
// Manage intermediate buffers
@@ -125,8 +170,13 @@
// Run weights reshaping (Runs once for every configure)
if(_is_first_run)
{
+ ARM_COMPUTE_ERROR_ON(!_original_weights->is_used());
+
_is_first_run = false;
NEScheduler::get().schedule(&_weights_reshape_kernel, 3);
+
+ // Mark original weights tensor as unused
+ _original_weights->mark_as_unused();
}
_memory_group.acquire();