Integrate improved pooling layer on NEON
Resolves COMPMID-4035
Change-Id: I559f8c4208fba9193dfe5012f03ddaf26c746215
Signed-off-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/4855
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
diff --git a/src/core/NEON/kernels/arm_conv/pooling/pooling_depthfirst_cache_oblivious.hpp b/src/core/NEON/kernels/arm_conv/pooling/pooling_depthfirst_cache_oblivious.hpp
new file mode 100644
index 0000000..4aabd95
--- /dev/null
+++ b/src/core/NEON/kernels/arm_conv/pooling/pooling_depthfirst_cache_oblivious.hpp
@@ -0,0 +1,312 @@
+/*
+ * Copyright (c) 2021 Arm Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#pragma once
+
+#include "pool_common.hpp"
+
+#include <stack>
+#include <vector>
+
+namespace arm_conv {
+namespace pooling {
+
+template <class strategy>
+class PoolingDepthfirstCacheOblivious : public PoolingCommon<typename strategy::operand_type, typename strategy::return_type>
+{
+ using TInput = typename strategy::operand_type;
+ using TOutput = typename strategy::return_type;
+
+ const PoolingArgs m_args; // Copy of arguments
+
+ constexpr static unsigned int input_rows(void)
+ {
+ return (strategy::out_rows() - 1)*strategy::stride_rows() + strategy::pool_rows();
+ }
+
+ constexpr static unsigned int input_cols(void)
+ {
+ return (strategy::out_cols() - 1)*strategy::stride_cols() + strategy::pool_cols();
+ }
+
+ size_t sizeof_input_buffer(void) const
+ {
+ return sizeof(TInput) * m_args.n_channels;
+ }
+
+ size_t sizeof_output_buffer(void) const
+ {
+ return sizeof(TOutput) * m_args.n_channels;
+ }
+
+ public:
+ PoolingDepthfirstCacheOblivious(const PoolingArgs &args) : m_args(args)
+ {
+ }
+
+ PoolingDepthfirstCacheOblivious(PoolingDepthfirstCacheOblivious &) = delete;
+ PoolingDepthfirstCacheOblivious &operator=(PoolingDepthfirstCacheOblivious &) = delete;
+
+ size_t get_working_size(void) const override
+ {
+ // We require an array of pointers for the inputs and outputs, a
+ // channel-length vector in which to dump surplus output, and a
+ // channel-length vector of padding values.
+ return sizeof_input_buffer() + sizeof_output_buffer();
+ }
+
+ void execute(
+ const void *const input,
+ void *const output,
+ void *const working_space
+ ) const override
+ {
+ const size_t ld_input_col = m_args.n_channels;
+ const size_t ld_input_row = ld_input_col * m_args.input_cols;
+ const size_t ld_input_batch = ld_input_row * m_args.input_rows;
+ const size_t ld_output_col = ld_input_col;
+ const size_t ld_output_row = ld_output_col * m_args.output_cols;
+ const size_t ld_output_batch = ld_output_row * m_args.output_rows;
+
+ execute(
+ input, ld_input_col, ld_input_row, ld_input_batch,
+ output, ld_output_col, ld_output_row, ld_output_batch,
+ working_space
+ );
+ }
+
+ void execute(
+ const void *const input,
+ size_t ld_input_col,
+ size_t ld_input_row,
+ size_t ld_input_batch,
+ void *const output,
+ size_t ld_output_col,
+ size_t ld_output_row,
+ size_t ld_output_batch,
+ void *const working_space
+ ) const override
+ {
+ execute(
+ m_args.n_batches, m_args.input_rows, m_args.input_cols,
+ m_args.n_channels,
+ input, ld_input_col, ld_input_row, ld_input_batch,
+ m_args.padding,
+ m_args.output_rows, m_args.output_cols,
+ output, ld_output_col, ld_output_row, ld_output_batch,
+ working_space
+ );
+ }
+
+ void execute(
+ unsigned int batches,
+ unsigned int input_height,
+ unsigned int input_width,
+ unsigned int channels,
+ const void *const _input,
+ size_t ld_input_col,
+ size_t ld_input_row,
+ size_t ld_input_batch,
+ const PaddingValues &padding,
+ unsigned int output_height,
+ unsigned int output_width,
+ void *const _output,
+ size_t ld_output_col,
+ size_t ld_output_row,
+ size_t ld_output_batch,
+ void *const _working_space
+ ) const override
+ {
+ strategy strat(m_args.cpu_info);
+#ifdef CYCLE_PROFILING
+ arm_gemm::profiler prof;
+#endif // CYCLE_PROFILING
+
+ // Cast input and output pointers into the right types
+ const TInput *const inptr = static_cast<const TInput *>(_input);
+ TOutput *const outptr = static_cast<TOutput *>(_output);
+
+ // Allocate portions of the working space
+ uint8_t *const working_space = static_cast<uint8_t *>(_working_space);
+ TOutput *const output_buffer = reinterpret_cast<TOutput *>(working_space);
+ TInput *const input_buffer = reinterpret_cast<TInput *>(working_space + sizeof_output_buffer());
+
+ // Fill the input buffer
+ const TInput pad_value = (m_args.pool_type == PoolingType::AVERAGE)
+ ? static_cast<TInput>(0)
+ : (std::numeric_limits<TInput>::has_infinity
+ ? -std::numeric_limits<TInput>::infinity()
+ : std::numeric_limits<TInput>::lowest());
+ for (unsigned int i = 0; i < channels; i++)
+ {
+ input_buffer[i] = pad_value;
+ }
+
+ // Keep subdividing the output plane across the longest dimension until we
+ // reach the size of the tile. Queue items for later processing. Note - we
+ // can determine the largest size of the queue a priori from the input
+ // tensor size, this would allow us to allocate memory within the working
+ // space and improve performance.
+ struct WorkItem
+ {
+ unsigned int output_i, output_j;
+ unsigned int output_height, output_width;
+
+ WorkItem(unsigned int i, unsigned int j, unsigned int height, unsigned int width)
+ : output_i(i), output_j(j), output_height(height), output_width(width) {}
+ };
+
+ auto execute = [&] (const WorkItem &item) {
+ // Create an array for the output pointers
+ TOutput * _outptr_array[strategy::out_rows() * strategy::out_cols()];
+ TOutput **const outptr_array = _outptr_array;
+
+ // Construct the output pointer array
+ {
+ const auto output_pad_right = strategy::out_rows() - item.output_width;
+ auto outptr_element = outptr_array;
+ auto outptr_row = outptr + item.output_i * ld_output_row + item.output_j * ld_output_col;
+
+ // Fill the array with pointers to the output buffer
+ for (unsigned int i = 0; i < strategy::out_rows() * strategy::out_cols(); i++)
+ {
+ outptr_array[i] = output_buffer;
+ }
+
+ // Fill in the valid portion of the array
+ for (unsigned int i = 0; i < item.output_height; i++)
+ {
+ auto outptr_col = outptr_row;
+ for (unsigned int j = 0; j < item.output_width; j++)
+ {
+ *(outptr_element++) = outptr_col;
+ outptr_col += ld_output_col;
+ }
+ outptr_element += output_pad_right;
+ outptr_row += ld_output_row;
+ }
+ }
+
+ const int start_i = item.output_i * strategy::stride_rows() - padding.top;
+ const int end_i = start_i + input_rows();
+ const unsigned int pad_top = std::max(0, 0 - start_i);
+ const unsigned int pad_bottom = std::max(0, end_i - static_cast<int>(input_height));
+
+ const int start_j = item.output_j * strategy::stride_cols() - padding.left;
+ const int end_j = start_j + input_cols();
+ const unsigned int pad_left = std::max(0, 0 - start_j);
+ const unsigned int pad_right = std::max(0, end_j - static_cast<int>(input_width));
+
+ // Create an array for the input pointers
+ const TInput * _inptr_array[input_rows() * input_cols()];
+ const TInput **const inptr_array = _inptr_array;
+ {
+ const unsigned int row_padding = pad_top + pad_bottom;
+ const unsigned int valid_rows = input_rows() - row_padding;
+
+ const unsigned int col_padding = pad_left + pad_right;
+ const unsigned int valid_cols = input_cols() - col_padding;
+
+ // Fill the array with pointers to the input buffer
+ for (unsigned int i = 0; i < input_rows() * input_cols(); i++)
+ {
+ inptr_array[i] = input_buffer;
+ }
+
+ // Compute valid initial pointer
+ auto inptr_row = inptr + std::max(start_i, 0) * ld_input_row + std::max(start_j, 0) * ld_input_col;
+
+ // Fill in the valid portion of the input array
+ auto inptr_element = inptr_array + pad_top * input_cols() + pad_left;
+ for (unsigned int i = 0; i < valid_rows; i++)
+ {
+ auto inptr_col = inptr_row;
+ for (unsigned int j = 0; j < valid_cols; j++)
+ {
+ *(inptr_element++) = inptr_col;
+ inptr_col += ld_input_col;
+ }
+
+ inptr_row += ld_input_row;
+ inptr_element += col_padding; // Skip the padding elements
+ }
+ }
+
+ // Call the kernel
+#ifdef CYCLE_PROFILING
+ // TODO Work number
+ auto p = prof.ScopedProfiler(PROFILE_KERNEL, (unsigned long)(item.output_height * item.output_width * strategy::pool_rows() * strategy::pool_cols()));
+#endif // CYCLE_PROFILING
+ strat.kernel(channels, inptr_array, outptr_array,
+ pad_left, pad_top, pad_right, pad_bottom);
+ };
+
+ // Add the initial work item to the stack of work.
+ std::stack<WorkItem, std::vector<WorkItem>> stack;
+ stack.push(WorkItem(0, 0, output_height, output_width));
+ while (!stack.empty())
+ {
+ // Pop an item from the stack, bisect the largest dimension and either
+ // execute the resulting tiles or add them to the stack if they are too
+ // large.
+ const WorkItem item(stack.top());
+ stack.pop();
+
+ if (item.output_height <= strategy::out_rows() &&
+ item.output_width <= strategy::out_cols())
+ {
+ execute(item);
+ }
+ else
+ {
+ // Split the largest dimension, such that we get an exact number of
+ // tiles in the first partition.
+ if (item.output_height >= item.output_width)
+ {
+ const unsigned int height_in_tiles = (item.output_height + strategy::out_rows() - 1) / strategy::out_rows();
+ const unsigned int tiles_first = height_in_tiles - height_in_tiles / 2;
+
+ const unsigned int height_first = tiles_first * strategy::out_rows();
+ const unsigned int height_second = item.output_height - height_first;
+
+ stack.push(WorkItem(item.output_i + height_first, item.output_j, height_second, item.output_width));
+ stack.push(WorkItem(item.output_i, item.output_j, height_first, item.output_width));
+ }
+ else
+ {
+ const unsigned int width_in_tiles = item.output_width / strategy::out_cols();
+ const unsigned int tiles_first = width_in_tiles - width_in_tiles / 2;
+
+ const unsigned int width_first = tiles_first * strategy::out_cols();
+ const unsigned int width_second = item.output_width - width_first;
+
+ stack.push(WorkItem(item.output_i, item.output_j + width_first, item.output_height, width_second));
+ stack.push(WorkItem(item.output_i, item.output_j, item.output_height, width_first));
+ }
+ }
+ }
+ }
+};
+
+} // namespace pooling
+} // namespace arm_conv