Depthwise channel pre-multiplication
Resolves: COMPMID-6337
Change-Id: Ie9097b3f56e8071426c621386a5988bd7f7e8ef2
Signed-off-by: Michael Tyler <michael.tyler@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/9852
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Viet-Hoa Do <viet-hoa.do@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>
diff --git a/src/core/NEON/kernels/assembly/depthwise.hpp b/src/core/NEON/kernels/assembly/depthwise.hpp
index 8eb278c..dbd47cc 100644
--- a/src/core/NEON/kernels/assembly/depthwise.hpp
+++ b/src/core/NEON/kernels/assembly/depthwise.hpp
@@ -27,6 +27,7 @@
#include "arm_gemm.hpp"
#include "arm_gemm_local.hpp"
#include "depthwise_common.hpp"
+#include "premultiply.hpp"
namespace arm_conv
{
@@ -38,8 +39,8 @@
std::string filter = "";
DepthwiseConfig(DepthwiseMethod method)
- : method(method){};
- DepthwiseConfig(){};
+ : method(method) {};
+ DepthwiseConfig() {};
};
struct DepthwiseArgs
@@ -112,17 +113,64 @@
}
};
+template <typename TInput>
+struct Tile
+{
+ TInput *array;
+
+ unsigned int tile_rows = 0;
+ unsigned int tile_cols = 0;
+ unsigned int tile_channels = 0;
+
+ Tile(TInput *array, unsigned int tile_rows, unsigned int tile_cols, unsigned int tile_channels)
+ : array(array), tile_rows(tile_rows), tile_cols(tile_cols), tile_channels(tile_channels)
+ {
+ }
+
+ Tile()
+ : Tile(nullptr, 0, 0, 0)
+ {
+ }
+
+ void load_from(
+ const TInput *input,
+ const unsigned int ld_row, const unsigned int ld_col,
+ const unsigned int n_rows, const unsigned int n_cols,
+ const int input_i, const int input_j,
+ const unsigned int channel_multiplier) const
+ {
+ const auto pad_top = input_i < 0 ? -input_i : 0;
+ const auto pad_left = input_j < 0 ? -input_j : 0;
+
+ const auto padded_rows = std::min(n_rows - input_i, tile_rows) - pad_top;
+ const auto padded_cols = std::min(n_cols - input_j, tile_cols) - pad_left;
+
+ if(padded_rows < tile_rows || padded_cols < tile_cols)
+ {
+ memset(array, 0, tile_rows * tile_cols * tile_channels * sizeof(TInput));
+ }
+
+ do_premultiply<TInput>(
+ (TInput *)input + std::max(input_i, 0) * ld_row + std::max(input_j, 0) * ld_col,
+ ld_row, ld_col,
+ array + pad_top * tile_cols * tile_channels + pad_left * tile_channels,
+ tile_cols * tile_channels, tile_channels,
+ padded_rows, padded_cols, tile_channels / channel_multiplier,
+ channel_multiplier);
+ }
+};
+
template <typename TInput, typename TWeight, typename TOutput>
class DepthwiseCommon : public IDepthwiseCommon
{
- protected:
+protected:
const DepthwiseArgs m_args; // Copy of arguments
std::string m_name{};
- public:
+public:
DepthwiseCommon(const DepthwiseArgs &args)
- : m_args(args){};
- DepthwiseCommon(DepthwiseCommon &) = delete;
+ : m_args(args) {};
+ DepthwiseCommon(DepthwiseCommon &) = delete;
DepthwiseCommon &operator=(DepthwiseCommon &) = delete;
std::string name() const override
@@ -133,7 +181,7 @@
void set_name(std::string name)
{
// Only allow the name to be set once
- if (m_name.empty())
+ if(m_name.empty())
{
m_name = name;
}
@@ -209,47 +257,47 @@
// passed different input/output tensors. Dilation is handled at this
// level; so we set the dilation in the arguments to zero.
DepthwiseArgs args(this->m_args);
- args.n_batches = batches;
- args.input_rows = input_height;
- args.input_cols = input_width;
+ args.n_batches = batches;
+ args.input_rows = input_height;
+ args.input_cols = input_width;
args.input_channels = channels;
- args.output_rows = output_height;
- args.output_cols = output_width;
- args.padding = padding;
+ args.output_rows = output_height;
+ args.output_cols = output_width;
+ args.padding = padding;
args.dilation_rows = args.dilation_cols = 1;
- auto ld_input_col_d = ld_input_col * m_args.dilation_cols;
- auto ld_input_row_d = ld_input_row * m_args.dilation_rows;
+ auto ld_input_col_d = ld_input_col * m_args.dilation_cols;
+ auto ld_input_row_d = ld_input_row * m_args.dilation_rows;
auto ld_output_col_d = ld_output_col * m_args.dilation_cols;
auto ld_output_row_d = ld_output_row * m_args.dilation_rows;
- for (size_t drow = 0; drow < m_args.dilation_rows; drow++)
+ for(size_t drow = 0; drow < m_args.dilation_rows; drow++)
{
size_t start_i;
std::tie(args.output_rows, args.input_rows, start_i,
args.padding.top, args.padding.bottom) =
- get_reduced_view_for_dilation(
- output_height, input_height, drow, m_args.dilation_rows,
- m_args.kernel_rows, m_args.stride_rows, padding.top);
+ get_reduced_view_for_dilation(
+ output_height, input_height, drow, m_args.dilation_rows,
+ m_args.kernel_rows, m_args.stride_rows, padding.top);
- auto input_row = static_cast<const TInput *>(input) + start_i * ld_input_row;
+ auto input_row = static_cast<const TInput *>(input) + start_i * ld_input_row;
auto output_row = static_cast<TOutput *>(output) + drow * ld_output_row;
- if (args.output_rows)
+ if(args.output_rows)
{
- for (size_t dcol = 0; dcol < m_args.dilation_cols; dcol++)
+ for(size_t dcol = 0; dcol < m_args.dilation_cols; dcol++)
{
size_t start_j;
std::tie(args.output_cols, args.input_cols, start_j,
args.padding.left, args.padding.right) =
- get_reduced_view_for_dilation(
- output_width, input_width, dcol, m_args.dilation_cols,
- m_args.kernel_cols, m_args.stride_cols, padding.left);
+ get_reduced_view_for_dilation(
+ output_width, input_width, dcol, m_args.dilation_cols,
+ m_args.kernel_cols, m_args.stride_cols, padding.left);
- const TInput *input_col = input_row + start_j * ld_input_col;
- TOutput *output_col = output_row + dcol * ld_output_col;
+ const TInput *input_col = input_row + start_j * ld_input_col;
+ TOutput *output_col = output_row + dcol * ld_output_col;
- if (args.output_cols)
+ if(args.output_cols)
{
this->execute_internal(
args, input_col, ld_input_col_d, ld_input_row_d, ld_input_batch, parameters,
@@ -261,7 +309,7 @@
}
}
- protected:
+protected:
virtual void execute_internal(
const DepthwiseArgs &instance_args,
const void *input,
@@ -276,6 +324,11 @@
void *working_space,
unsigned int thread_id,
unsigned int n_threads) const = 0;
+
+ virtual bool uses_premultiply() const
+ {
+ return true;
+ }
};
template <typename TInput, typename TWeight = TInput, typename TOutput = TInput>
diff --git a/src/core/NEON/kernels/assembly/depthwise_common.hpp b/src/core/NEON/kernels/assembly/depthwise_common.hpp
index fea6326..a5db793 100644
--- a/src/core/NEON/kernels/assembly/depthwise_common.hpp
+++ b/src/core/NEON/kernels/assembly/depthwise_common.hpp
@@ -85,7 +85,7 @@
size_t ld_weight_row = 0) = 0;
// Determine the amount of working space required
- virtual size_t get_working_size(unsigned int n_threads, unsigned int n_input_channels) const = 0;
+ virtual size_t get_working_size(unsigned int n_threads) const = 0;
// Execute the convolution over the specified area of memory.
virtual void execute(
diff --git a/src/core/NEON/kernels/assembly/pool_common.hpp b/src/core/NEON/kernels/assembly/pool_common.hpp
index 599e18a..f1f70cf 100644
--- a/src/core/NEON/kernels/assembly/pool_common.hpp
+++ b/src/core/NEON/kernels/assembly/pool_common.hpp
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2021-2022 Arm Limited.
+ * Copyright (c) 2021-2023 Arm Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -66,7 +66,6 @@
// Determine the amount of working space required.
virtual size_t get_working_size(unsigned int num_threads) const = 0;
- virtual size_t get_working_size(unsigned int num_threads, unsigned int n_channels) const = 0;
// Execute pooling over the specified area of memory.
virtual void execute(
diff --git a/src/core/NEON/kernels/assembly/pooling.hpp b/src/core/NEON/kernels/assembly/pooling.hpp
index 1b47853..e8db35c 100644
--- a/src/core/NEON/kernels/assembly/pooling.hpp
+++ b/src/core/NEON/kernels/assembly/pooling.hpp
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2021-2022 Arm Limited.
+ * Copyright (c) 2021-2023 Arm Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -122,11 +122,7 @@
PoolingCommon(PoolingCommon &) = delete;
PoolingCommon &operator=(PoolingCommon &) = delete;
- size_t get_working_size(unsigned int, unsigned int) const override = 0;
- size_t get_working_size(unsigned int n_threads) const override
- {
- return this->get_working_size(n_threads, m_args.n_channels);
- }
+ size_t get_working_size(unsigned int) const override = 0;
// Execute pooling over the specified area of memory.
void execute(
diff --git a/src/core/NEON/kernels/assembly/premultiply.hpp b/src/core/NEON/kernels/assembly/premultiply.hpp
new file mode 100644
index 0000000..16f26de
--- /dev/null
+++ b/src/core/NEON/kernels/assembly/premultiply.hpp
@@ -0,0 +1,81 @@
+/*
+ * Copyright (c) 2023 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.
+ */
+
+void do_premultiply_float_6(const float *in_ptr,
+ const unsigned int ld_row,
+ const unsigned int ld_col,
+ float *out_ptr,
+ const unsigned int out_ld_row,
+ const unsigned int out_ld_col,
+ const unsigned int tile_rows,
+ const unsigned int tile_cols,
+ const unsigned input_channels);
+
+template <typename T>
+void do_premultiply(const T *in_ptr,
+ const unsigned int ld_row,
+ const unsigned int ld_col,
+ T *out_ptr,
+ const unsigned int out_ld_row,
+ const unsigned int out_ld_col,
+ const unsigned int tile_rows,
+ const unsigned int tile_cols,
+ const unsigned input_channels,
+ const unsigned int channel_multiplier)
+{
+ if(sizeof(T) == 4 && channel_multiplier == 6)
+ {
+ do_premultiply_float_6(
+ (const float *)in_ptr, ld_row, ld_col,
+ (float *)out_ptr, out_ld_row, out_ld_col,
+ tile_rows, tile_cols,
+ input_channels);
+ }
+ else
+ {
+ for(unsigned int i = 0; i < tile_rows; i++)
+ {
+ const T *ip2 = in_ptr + i * ld_row;
+ T *op2 = out_ptr + i * out_ld_row;
+ for(unsigned int j = 0; j < tile_cols; j++)
+ {
+ const T *ip = ip2;
+ T *op = op2;
+ for(unsigned int c = 0; c < input_channels; c++)
+ {
+ T val = *ip;
+ ip++;
+
+ for(unsigned int r = 0; r < channel_multiplier; r++)
+ {
+ op[r] = val;
+ }
+ op += channel_multiplier;
+ }
+ ip2 += ld_col;
+ op2 += out_ld_col;
+ }
+ }
+ }
+}