diff options
Diffstat (limited to 'src/core/NEON')
-rw-r--r-- | src/core/NEON/kernels/arm_conv/pooling/pooling_depthfirst_generic.hpp | 17 | ||||
-rw-r--r-- | src/core/NEON/kernels/arm_conv/pooling/pooling_depthfirst_generic_quantized.hpp | 17 |
2 files changed, 32 insertions, 2 deletions
diff --git a/src/core/NEON/kernels/arm_conv/pooling/pooling_depthfirst_generic.hpp b/src/core/NEON/kernels/arm_conv/pooling/pooling_depthfirst_generic.hpp index fa06a0078b..5979862ed8 100644 --- a/src/core/NEON/kernels/arm_conv/pooling/pooling_depthfirst_generic.hpp +++ b/src/core/NEON/kernels/arm_conv/pooling/pooling_depthfirst_generic.hpp @@ -191,6 +191,13 @@ class PoolingDepthfirstGeneric : public PoolingCommon<typename strategy::operand const auto pad_bottom = static_cast<unsigned int>(std::max<int>(end_in_i - height, 0)); const auto valid_rows = input_rows() - pad_top - pad_bottom; + // Compute the number of pooling window rows which are contained in + // either the valid region of the input tensor, or the padding. + const auto padded_bottom = std::min<unsigned int>( + start_in_i + m_args.pool_window.rows, height + padding.bottom + ); + const auto n_total_rows = padded_bottom - start_in_i; + auto outptr_col = outptr_row; auto inptr_row = inptr_batch + (start_in_i + pad_top) * ld_input_row; @@ -205,6 +212,13 @@ class PoolingDepthfirstGeneric : public PoolingCommon<typename strategy::operand const auto pad_right = static_cast<unsigned int>(std::max<int>(0, end_in_j - width)); const auto valid_cols = input_cols() - pad_left - pad_right; + // Compute the number of pooling window columns which are contained + // in either the valid region of the input tensor, or the padding. + const auto padded_right = std::min<unsigned int>( + start_in_j + m_args.pool_window.cols, width + padding.right + ); + const auto n_total_cols = padded_right - start_in_j; + // Construct the input pointer array - fill in all valid points // contiguously. const TInput **ptrs = inptr_array; @@ -222,7 +236,8 @@ class PoolingDepthfirstGeneric : public PoolingCommon<typename strategy::operand // Compute the number of valid cells const auto valid_cells = valid_rows * valid_cols; - const auto window_cells = m_args.exclude_padding ? valid_cells : input_rows() * input_cols(); + const auto cells_in_range = n_total_rows * n_total_cols; + const auto window_cells = m_args.exclude_padding ? valid_cells : cells_in_range; // Get the output pointer for this call TOutput *outptr = outptr_col; diff --git a/src/core/NEON/kernels/arm_conv/pooling/pooling_depthfirst_generic_quantized.hpp b/src/core/NEON/kernels/arm_conv/pooling/pooling_depthfirst_generic_quantized.hpp index 1f2891f814..f3cb9a1d1f 100644 --- a/src/core/NEON/kernels/arm_conv/pooling/pooling_depthfirst_generic_quantized.hpp +++ b/src/core/NEON/kernels/arm_conv/pooling/pooling_depthfirst_generic_quantized.hpp @@ -191,6 +191,13 @@ class PoolingDepthfirstGenericQuantized : public PoolingCommon<typename strategy const auto pad_top = static_cast<unsigned int>(-std::min(start_in_i, 0)); const auto pad_bottom = static_cast<unsigned int>(-std::min(static_cast<int>(height) - end_in_i, 0)); + // Compute the number of pooling window rows which are contained in + // either the valid region of the input tensor, or the padding. + const auto padded_bottom = std::min<unsigned int>( + start_in_i + m_args.pool_window.rows, height + padding.bottom + ); + const auto n_total_rows = padded_bottom - start_in_i; + for (int out_j = 0, start_in_j = -padding.left; out_j < static_cast<int>(output_width); out_j++, start_in_j += m_args.pool_stride.cols) @@ -201,6 +208,13 @@ class PoolingDepthfirstGenericQuantized : public PoolingCommon<typename strategy const auto pad_left = static_cast<unsigned int>(-std::min(start_in_j, 0)); const auto pad_right = static_cast<unsigned int>(-std::min(static_cast<int>(width) - end_in_j, 0)); + // Compute the number of pooling window columns which are contained + // in either the valid region of the input tensor, or the padding. + const auto padded_right = std::min<unsigned int>( + start_in_j + m_args.pool_window.cols, width + padding.right + ); + const auto n_total_cols = padded_right - start_in_j; + // Construct the input pointer array - fill in all valid points // contiguously. const TInput **ptrs = inptr_array; @@ -221,7 +235,8 @@ class PoolingDepthfirstGenericQuantized : public PoolingCommon<typename strategy const auto valid_rows = input_rows() - pad_top - pad_bottom; const auto valid_cols = input_cols() - pad_left - pad_right; const auto valid_cells = valid_rows * valid_cols; - const auto window_cells = m_args.exclude_padding ? valid_cells : input_rows() * input_cols(); + const auto cells_in_range = n_total_rows * n_total_cols; + const auto window_cells = m_args.exclude_padding ? valid_cells : cells_in_range; // Get the output pointer for this call TOutput *outptr = outptr_batch + out_i * ld_output_row + out_j * ld_output_col; |