diff options
author | Pablo Tello <pablo.tello@arm.com> | 2019-02-27 13:32:51 +0000 |
---|---|---|
committer | Giuseppe Rossini <giuseppe.rossini@arm.com> | 2019-03-08 10:41:25 +0000 |
commit | 471043616a869f0e696c8db4e1d0a62b45b4decf (patch) | |
tree | 79f3386c4d6757472835003fbf552566213f3bce /src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.cpp | |
parent | adc5395ad72aceb2c9e7e6beb54d949959d35143 (diff) | |
download | ComputeLibrary-471043616a869f0e696c8db4e1d0a62b45b4decf.tar.gz |
COMPMID-1882: Improve memory coalescence when reshaping the weights for CLDepthwiseConvolution
Change-Id: I97788d9e349f37fcd818d588d668e2d5e22fd568
Signed-off-by: giuros01 <giuseppe.rossini@arm.com>
Reviewed-on: https://review.mlplatform.org/c/818
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Diffstat (limited to 'src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.cpp')
-rw-r--r-- | src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.cpp | 5 |
1 files changed, 4 insertions, 1 deletions
diff --git a/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.cpp b/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.cpp index e8efe6f0a9..770740d180 100644 --- a/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.cpp +++ b/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.cpp @@ -171,7 +171,10 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITen { const bool is_dot8_supported = dot8_supported(CLKernelLibrary::get().get_device()); - kernel_name = is_qasymm ? (std::string("depthwise_convolution_3x3_quantized") + (is_dot8_supported ? "_dot8" : "") + "_nchw") : "depthwise_convolution_3x3"; + kernel_name = is_qasymm ? "dwc_3x3_native_qasymm8" : "depthwise_convolution_3x3"; + kernel_name += (is_qasymm && is_dot8_supported ? "_dot8" : ""); + kernel_name += "_nchw"; + num_elems_written_per_iteration_x = 8 / data_size_from_type(input->data_type()); num_elems_written_per_iteration_y = (is_qasymm && conv_stride_y == 1) ? 2 : 1; num_elems_read_per_iteration_x = 3 + (num_elems_written_per_iteration_x - 1) * conv_stride_x; |