aboutsummaryrefslogtreecommitdiff
path: root/src/core/NEON/kernels
diff options
context:
space:
mode:
Diffstat (limited to 'src/core/NEON/kernels')
-rw-r--r--src/core/NEON/kernels/NEDepthwiseConvolutionLayer3x3Kernel.cpp4
-rw-r--r--src/core/NEON/kernels/NEDirectConvolutionLayerOutputStageKernel.cpp2
2 files changed, 4 insertions, 2 deletions
diff --git a/src/core/NEON/kernels/NEDepthwiseConvolutionLayer3x3Kernel.cpp b/src/core/NEON/kernels/NEDepthwiseConvolutionLayer3x3Kernel.cpp
index 92383d9f15..dad4fee837 100644
--- a/src/core/NEON/kernels/NEDepthwiseConvolutionLayer3x3Kernel.cpp
+++ b/src/core/NEON/kernels/NEDepthwiseConvolutionLayer3x3Kernel.cpp
@@ -272,8 +272,8 @@ void NEDepthwiseConvolutionLayer3x3Kernel::configure_generic()
-conv_pad_top,
(num_x_steps - 1) * input_num_elems_processed + num_elems_read_per_iteration,
_input->info()->tensor_shape().y() + conv_pad_bottom);
- AccessWindowStatic weights_access(_weights->info(), 0, 0, _weights->info()->dimension(0), _weights->info()->dimension(1));
- AccessWindowStatic output_access(_output->info(), 0, 0, num_x_steps * _num_elems_written_per_iteration, output_shape.y());
+ AccessWindowStatic weights_access(_weights->info(), 0, 0, _weights->info()->dimension(0), _weights->info()->dimension(1));
+ AccessWindowHorizontal output_access(_output->info(), 0, _num_elems_written_per_iteration);
update_window_and_padding(win, input_access, weights_access, output_access);
output_access.set_valid_region(win, ValidRegion(Coordinates(), _output->info()->tensor_shape()));
diff --git a/src/core/NEON/kernels/NEDirectConvolutionLayerOutputStageKernel.cpp b/src/core/NEON/kernels/NEDirectConvolutionLayerOutputStageKernel.cpp
index 3f33c43b59..08d8f8ce56 100644
--- a/src/core/NEON/kernels/NEDirectConvolutionLayerOutputStageKernel.cpp
+++ b/src/core/NEON/kernels/NEDirectConvolutionLayerOutputStageKernel.cpp
@@ -451,8 +451,10 @@ void NEDirectConvolutionLayerOutputStageKernel::configure(ITensor *input, const
break;
}
case DataType::S32:
+ {
_func = (bias == nullptr) ? &output_stage<int32_t, uint8_t, false, false> : &output_stage<int32_t, uint8_t, false, true>;
break;
+ }
#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
case DataType::F16:
{