diff options
| author | Giuseppe Rossini <giuseppe.rossini@arm.com> | 2019-11-06 14:57:49 +0000 |
|---|---|---|
| committer | Pablo Marquez <pablo.tello@arm.com> | 2019-12-04 12:45:12 +0000 |
| commit | f01201abec0a102f6e7a517971f83fef1eaffd50 (patch) | |
| tree | adf844c3c9c8e0e96af9c56de27a094fab515e35 /src/core/NEON/kernels/convolution/depthwise/impl_base.hpp | |
| parent | 6e1791b1bfabc81f08d3117939f6eb5264ed4edf (diff) | |
| download | ComputeLibrary-f01201abec0a102f6e7a517971f83fef1eaffd50.tar.gz | |
COMPMID-2305: NEDepthwiseConvolution 3x3: support for QUANT8_PER_CHANNEL_SYMM
Change-Id: I9a917cff6a089ce6ae16fb4e6066a4194e2e9487
Signed-off-by: Giuseppe Rossini <giuseppe.rossini@arm.com>
Reviewed-on: https://review.mlplatform.org/c/2241
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Pablo Marquez <pablo.tello@arm.com>
Diffstat (limited to 'src/core/NEON/kernels/convolution/depthwise/impl_base.hpp')
| -rw-r--r-- | src/core/NEON/kernels/convolution/depthwise/impl_base.hpp | 3 |
1 files changed, 2 insertions, 1 deletions
diff --git a/src/core/NEON/kernels/convolution/depthwise/impl_base.hpp b/src/core/NEON/kernels/convolution/depthwise/impl_base.hpp index b102a24250..22231cf019 100644 --- a/src/core/NEON/kernels/convolution/depthwise/impl_base.hpp +++ b/src/core/NEON/kernels/convolution/depthwise/impl_base.hpp @@ -292,6 +292,7 @@ MEMBERFN(void)::run( // Parallelise over blocks of channels const auto start_channel = CHANNEL_BLOCK * start; const auto stop_channel = std::min<unsigned int>(_n_channels, CHANNEL_BLOCK * stop); + const auto params_size_per_channel = this->get_packed_params_size()/_n_channels; // Compute top and bottom padding for input and output const int input_pad_top = _padding_top; @@ -325,7 +326,7 @@ MEMBERFN(void)::run( // Get the offset into the packed parameters const auto params_ptr = static_cast<const uint8_t*>(_packed_parameters) + - start_channel*(sizeof(TIn)*KernelRows*KernelColumns + sizeof(TBias)); + start_channel*params_size_per_channel; // Process the row process_tile_row( |