COMPMID-1277 - Optimizing CLIm2ColKernel for NHWC.
This patch includes:
- Im2Col optimizations for NHWC using a new data layout
- Refactoring of CLIm2ColKernel adding validation method and auto-init
- Removed im2col_reduced from CLIm2ColKernel and created a new kernel CLFlattenLayerKernel
Change-Id: I1620640b6796baa268324b33ae92cdd8de53e27c
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/141241
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Giorgio Arena <giorgio.arena@arm.com>
diff --git a/src/runtime/NEON/functions/NEFullyConnectedLayer.cpp b/src/runtime/NEON/functions/NEFullyConnectedLayer.cpp
index 25b8adc..c2f0283 100644
--- a/src/runtime/NEON/functions/NEFullyConnectedLayer.cpp
+++ b/src/runtime/NEON/functions/NEFullyConnectedLayer.cpp
@@ -113,7 +113,7 @@
// If the fully connected layer is called after a convolution layer, the input tensor must be linearized
// Initialize output tensor for im2col
- TensorShape shape_im2col = compute_im2col_fc_shape(input->info());
+ TensorShape shape_im2col = compute_flatten_shape(input->info());
_im2col_output.allocator()->init(input->info()->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(shape_im2col));
// Configure im2col kernel
@@ -249,7 +249,7 @@
bool is_fc_after_conv = true;
bool is_quantized = is_data_type_quantized_asymmetric(input->data_type());
- const ITensorInfo &im2col_input = TensorInfo(input->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(compute_im2col_fc_shape(input)));
+ const ITensorInfo &im2col_input = TensorInfo(input->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(compute_flatten_shape(input)));
const ITensorInfo &reshaped_weights = TensorInfo(weights->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(compute_transposed_shape(*weights)));
const ITensorInfo &converted_weights = weights_reshaped ? TensorInfo(weights->clone()->set_is_resizable(true).reset_padding()) : TensorInfo(*reshaped_weights.clone());
const ITensorInfo &gemmlowp_output = TensorInfo(output->clone()->set_is_resizable(true).reset_padding().set_data_type(DataType::S32));
@@ -420,4 +420,4 @@
_is_prepared = true;
}
-}
+}
\ No newline at end of file
diff --git a/src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp b/src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp
index c0a5d0a..df4a040 100644
--- a/src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp
+++ b/src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp
@@ -223,7 +223,7 @@
{
// Calculate im2col shape
// For NEON the batch size is on the fourth dimension
- // TODO (giaiod01): Use auto-init COMPMID-1277
+ // TODO (giaiod01): Auto-initialize the output shape of im2col COMPMID-1482
TensorShape shape_im2col = input->info()->tensor_shape();
shape_im2col.set(0, mat_weights_rows);
shape_im2col.set(1, conv_w * conv_h);
@@ -232,7 +232,7 @@
_im2col_output.allocator()->init(input->info()->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(shape_im2col));
_memory_group.manage(&_im2col_output);
- // Configure and tune im2col
+ // Configure
_im2col_kernel.configure(input, &_im2col_output, Size2D(kernel_width, kernel_height), conv_info, _append_bias, false, false, dilation);
// Update GEMM input