Extend Transposed Conv. for tiles with N0>1
Partially Resolves: COMPMID-5724
Signed-off-by: Gunes Bayir <gunes.bayir@arm.com>
Change-Id: I0aeddddcdd87c8c79f6dae9a76ffdc2ba0c08e17
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8883
Benchmark: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
diff --git a/src/gpu/cl/kernels/ClTransposedConvolutionKernel.cpp b/src/gpu/cl/kernels/ClTransposedConvolutionKernel.cpp
index 714ca8e..1390451 100644
--- a/src/gpu/cl/kernels/ClTransposedConvolutionKernel.cpp
+++ b/src/gpu/cl/kernels/ClTransposedConvolutionKernel.cpp
@@ -130,8 +130,13 @@
Window win;
output_shape.collapse(2U, 1U); // Collapse width and height into single dimension
+ const unsigned int n0 = adjust_vec_size(16 / output->element_size(), output_channels);
+ const unsigned int m0 = 1;
+ const unsigned int k0 = adjust_vec_size(16 / input->element_size(), input_channels);
+ const unsigned int partial_store_n0 = output_channels % n0;
+
// Create window and update padding
- win = calculate_max_window(output_shape, Steps(1, 1));
+ win = calculate_max_window(output_shape, Steps(n0, m0));
ICLKernel::configure_internal(win);
const std::string kernel_name = "transposed_convolution_nhwc";
@@ -140,11 +145,6 @@
const DataType input_data_type = input->data_type();
const PaddingInfo strides = deconv_info.stride();
- const unsigned int n0 = 1;
- const unsigned int m0 = 1;
- const unsigned int k0 = adjust_vec_size(16 / input->element_size(), input_channels);
- const unsigned int partial_store_n0 = output_channels % n0;
-
if(biases != nullptr)
{
build_options.add_option(std::string("-DHAS_BIAS"));