diff --git a/mlearning/cmsis/Makefile b/mlearning/cmsis/Makefile
index ca02bc43d..dc0881ffe 100644
--- a/mlearning/cmsis/Makefile
+++ b/mlearning/cmsis/Makefile
@@ -24,6 +24,7 @@ cmsis.zip:
$(Q) curl -L https://github.com/ARM-software/CMSIS_5/archive/refs/tags/$(CONFIG_CMSIS_VER).zip -o cmsis.zip
$(Q) unzip -o cmsis.zip
$(Q) mv CMSIS_5-$(CONFIG_CMSIS_VER) CMSIS_5
+ $(Q) patch -p0 < cmsis-nn-support_nnabla.patch
context:: cmsis.zip
diff --git a/mlearning/cmsis/cmsis-nn-support_nnabla.patch b/mlearning/cmsis/cmsis-nn-support_nnabla.patch
new file mode 100644
index 000000000..4de8d6872
--- /dev/null
+++ b/mlearning/cmsis/cmsis-nn-support_nnabla.patch
@@ -0,0 +1,1107 @@
+diff --git a/CMSIS_5/CMSIS/NN/Include/arm_nnfunctions_nnabla.h CMSIS_5/CMSIS/NN/Include/arm_nnfunctions_nnabla.h
+===CHANGE_NOTICE(1/5)===========================================================
+Sony Corporation added this file to 5.4.0
+to add the following function prototypes:
+ - arm_convolve_CHW_f32_basic_nonsquare()
+ - arm_convolve_CHW_q15_basic_nonsquare()
+ - arm_convolve_CHW_q7_basic_nonsquare()
+ - arm_nn_CHW_mat_mult_kernel_q7_q15()
+================================================================================
+--- /dev/null
++++ CMSIS_5/CMSIS/NN/Include/arm_nnfunctions_nnabla.h
+@@ -0,0 +1,217 @@
++/*
++ * Copyright (C) 2010-2018 Arm Limited or its affiliates. All rights reserved.
++ * Copyright 2018 Sony Corporation
++ *
++ * SPDX-License-Identifier: Apache-2.0
++ *
++ * Licensed under the Apache License, Version 2.0 (the License); you may
++ * not use this file except in compliance with the License.
++ * You may obtain a copy of the License at
++ *
++ * www.apache.org/licenses/LICENSE-2.0
++ *
++ * Unless required by applicable law or agreed to in writing, software
++ * distributed under the License is distributed on an AS IS BASIS, WITHOUT
++ * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
++ * See the License for the specific language governing permissions and
++ * limitations under the License.
++ */
++
++/* ----------------------------------------------------------------------
++ * Title: arm_nnfunctions_nnabla.h
++ * Author: Sony Corporation
++ * Description: Sony Corporation added this file to 5.4.0
++ * to add the following function prototypes:
++ * - arm_convolve_CHW_f32_basic_nonsquare()
++ * - arm_convolve_CHW_q15_basic_nonsquare()
++ * - arm_convolve_CHW_q7_basic_nonsquare()
++ * - arm_nn_CHW_mat_mult_kernel_q7_q15()
++ * $Date: 14. September 2018
++ * -------------------------------------------------------------------- */
++
++#ifndef _ARM_NNFUNCTIONS_CHW_H
++#define _ARM_NNFUNCTIONS_CHW_H
++
++#include "arm_nnsupportfunctions.h"
++#include "arm_nn_tables.h"
++
++#define USE_INTRINSIC
++
++//#define ARM_NN_TRUNCATE /* This config the rounding model to floor or round to the nearest int */
++
++#ifdef __cplusplus
++extern "C"
++{
++#endif
++
++ /**
++ * @brief Basic float32 convolution function (non-sqaure shape)
++ * @param[in] Im_in pointer to input tensor
++ * @param[in] dim_im_in_x input tensor dimention x
++ * @param[in] dim_im_in_y input tensor dimention y
++ * @param[in] ch_im_in number of input tensor channels
++ * @param[in] wt pointer to kernel weights
++ * @param[in] ch_im_out number of filters, i.e., output tensor channels
++ * @param[in] dim_kernel_x filter kernel size x
++ * @param[in] dim_kernel_y filter kernel size y
++ * @param[in] padding_x padding size x
++ * @param[in] padding_y padding size y
++ * @param[in] stride_x convolution stride x
++ * @param[in] stride_y convolution stride y
++ * @param[in] bias pointer to bias
++ * @param[in,out] Im_out pointer to output tensor
++ * @param[in] dim_im_out_x output tensor dimension x
++ * @param[in] dim_im_out_y output tensor dimension y
++ * @param[in,out] bufferA pointer to buffer space for input
++ * @param[in,out] bufferB pointer to buffer space for output
++ * @return The function returns ARM_MATH_SUCCESS
++ */
++
++ arm_status
++ arm_convolve_CHW_f32_basic_nonsquare(const float * Im_in,
++ const uint16_t dim_im_in_x,
++ const uint16_t dim_im_in_y,
++ const uint16_t ch_im_in,
++ const float * wt,
++ const uint16_t ch_im_out,
++ const uint16_t dim_kernel_x,
++ const uint16_t dim_kernel_y,
++ const uint16_t padding_x,
++ const uint16_t padding_y,
++ const uint16_t stride_x,
++ const uint16_t stride_y,
++ const float * bias,
++ float * Im_out,
++ const uint16_t dim_im_out_x,
++ const uint16_t dim_im_out_y,
++ float * bufferA,
++ float * bufferB);
++
++
++ /**
++ * @brief Basic Q15 version of CHW convolution (non-sqaure shape)
++ * @param[in] Im_in pointer to input tensor
++ * @param[in] dim_im_in_x input tensor dimention x
++ * @param[in] dim_im_in_y input tensor dimention y
++ * @param[in] ch_im_in number of input tensor channels
++ * @param[in] wt pointer to kernel weights
++ * @param[in] ch_im_out number of filters, i.e., output tensor channels
++ * @param[in] dim_kernel_x filter kernel size x
++ * @param[in] dim_kernel_y filter kernel size y
++ * @param[in] padding_x padding sizes x
++ * @param[in] padding_y padding sizes y
++ * @param[in] stride_x convolution stride x
++ * @param[in] stride_y convolution stride y
++ * @param[in] bias pointer to bias
++ * @param[in] bias_shift amount of left-shift for bias
++ * @param[in] out_shift amount of right-shift for output
++ * @param[in,out] Im_out pointer to output tensor
++ * @param[in] dim_im_out_x output tensor dimension x
++ * @param[in] dim_im_out_y output tensor dimension y
++ * @param[in,out] bufferA pointer to buffer space for input
++ * @param[in,out] bufferB pointer to buffer space for output
++ * @return The function returns ARM_MATH_SUCCESS
++ */
++
++ arm_status
++ arm_convolve_CHW_q15_basic_nonsquare(const q15_t * Im_in,
++ const uint16_t dim_im_in_x,
++ const uint16_t dim_im_in_y,
++ const uint16_t ch_im_in,
++ const q15_t * wt,
++ const uint16_t ch_im_out,
++ const uint16_t dim_kernel_x,
++ const uint16_t dim_kernel_y,
++ const uint16_t padding_x,
++ const uint16_t padding_y,
++ const uint16_t stride_x,
++ const uint16_t stride_y,
++ const q15_t * bias,
++ const uint16_t bias_shift,
++ const uint16_t out_shift,
++ q15_t * Im_out,
++ const uint16_t dim_im_out_x,
++ const uint16_t dim_im_out_y,
++ q15_t * bufferA,
++ q7_t * bufferB);
++
++ /**
++ * @brief Basic Q7 version of CHW convolution (non-sqaure shape)
++ * @param[in] Im_in pointer to input tensor
++ * @param[in] dim_im_in_x input tensor dimention x
++ * @param[in] dim_im_in_y input tensor dimention y
++ * @param[in] ch_im_in number of input tensor channels
++ * @param[in] wt pointer to kernel weights
++ * @param[in] ch_im_out number of filters, i.e., output tensor channels
++ * @param[in] dim_kernel_x filter kernel size x
++ * @param[in] dim_kernel_y filter kernel size y
++ * @param[in] padding_x padding size x
++ * @param[in] padding_y padding size y
++ * @param[in] stride_x convolution stride x
++ * @param[in] stride_y convolution stride y
++ * @param[in] bias pointer to bias
++ * @param[in] bias_shift amount of left-shift for bias
++ * @param[in] out_shift amount of right-shift for output
++ * @param[in,out] Im_out pointer to output tensor
++ * @param[in] dim_im_out_x output tensor dimension x
++ * @param[in] dim_im_out_y output tensor dimension y
++ * @param[in,out] bufferA pointer to buffer space for input
++ * @param[in,out] bufferB pointer to buffer space for output
++ * @return The function returns ARM_MATH_SUCCESS
++ */
++
++ arm_status
++ arm_convolve_CHW_q7_basic_nonsquare(const q7_t * Im_in,
++ const uint16_t dim_im_in_x,
++ const uint16_t dim_im_in_y,
++ const uint16_t ch_im_in,
++ const q7_t * wt,
++ const uint16_t ch_im_out,
++ const uint16_t dim_kernel_x,
++ const uint16_t dim_kernel_y,
++ const uint16_t padding_x,
++ const uint16_t padding_y,
++ const uint16_t stride_x,
++ const uint16_t stride_y,
++ const q7_t * bias,
++ const uint16_t bias_shift,
++ const uint16_t out_shift,
++ q7_t * Im_out,
++ const uint16_t dim_im_out_x,
++ const uint16_t dim_im_out_y,
++ q15_t * bufferA,
++ q7_t * bufferB);
++
++ /**
++ * @brief Matrix-multiplication function for convolution with CHW output
++ * @param[in] pA pointer to operand A
++ * @param[in] pInBuffer pointer to operand B, always conssists of 2 vectors
++ * @param[in] ch_im_out numRow of A
++ * @param[in] numCol_A numCol of A
++ * @param[in] out_stride output buffer channel stride
++ * @param[in] bias_shift amount of left-shift for bias
++ * @param[in] out_shift amount of right-shift for output
++ * @param[in] bias the bias
++ * @param[in,out] pOut pointer to output
++ * @return The function returns the incremented output pointer
++ *
++ * @details
++ *
++ * This function does the matrix multiplication with weight matrix
++ * and 2 columns from im2col.
++ */
++
++ q7_t *arm_nn_CHW_mat_mult_kernel_q7_q15(const q7_t * pA,
++ const q15_t * pInBuffer,
++ const uint16_t ch_im_out,
++ const uint16_t numCol_A,
++ const uint16_t out_stride,
++ const uint16_t bias_shift,
++ const uint16_t out_shift,
++ const q7_t * bias,
++ q7_t * pOut);
++
++#ifdef __cplusplus
++}
++#endif
++#endif
+diff --git a/CMSIS_5/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_CHW_f32_basic_nonsquare.c CMSIS_5/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_CHW_f32_basic_nonsquare.c
+===CHANGE_NOTICE(2/5)===========================================================
+Sony Corporation added this file to 5.4.0 for these reasons:
+ - support float version of convolution
+ - support the CHW tensor layout
+================================================================================
+--- /dev/null
++++ CMSIS_5/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_CHW_f32_basic_nonsquare.c
+@@ -0,0 +1,207 @@
++/*
++ * Copyright (C) 2010-2018 Arm Limited or its affiliates. All rights reserved.
++ * Copyright 2018 Sony Corporation
++ *
++ * SPDX-License-Identifier: Apache-2.0
++ *
++ * Licensed under the Apache License, Version 2.0 (the License); you may
++ * not use this file except in compliance with the License.
++ * You may obtain a copy of the License at
++ *
++ * www.apache.org/licenses/LICENSE-2.0
++ *
++ * Unless required by applicable law or agreed to in writing, software
++ * distributed under the License is distributed on an AS IS BASIS, WITHOUT
++ * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
++ * See the License for the specific language governing permissions and
++ * limitations under the License.
++ */
++
++/* ----------------------------------------------------------------------
++ * Title: arm_convolve_CHW_f32_basic_nonsquare.c
++ * Author: Sony Corporation
++ * Description: Sony Corporation added this file to 5.4.0 for these reasons:
++ * - support float version of convolution
++ * - support the CHW tensor layout
++ * $Date: 14. September 2018
++ * -------------------------------------------------------------------- */
++
++#include "arm_math.h"
++#include "arm_nnfunctions_nnabla.h"
++
++/**
++ * @ingroup groupNN
++ */
++
++/**
++ * @addtogroup NNConv
++ * @{
++ */
++
++ /**
++ * @brief Basic float32 version of CHW convolution (non-sqaure shape)
++ * @param[in] Im_in pointer to input tensor
++ * @param[in] dim_im_in_x input tensor dimention x
++ * @param[in] dim_im_in_y input tensor dimention y
++ * @param[in] ch_im_in number of input tensor channels
++ * @param[in] wt pointer to kernel weights
++ * @param[in] ch_im_out number of filters, i.e., output tensor channels
++ * @param[in] dim_kernel_x filter kernel size x
++ * @param[in] dim_kernel_y filter kernel size y
++ * @param[in] padding_x padding sizes x
++ * @param[in] padding_y padding sizes y
++ * @param[in] stride_x convolution stride x
++ * @param[in] stride_y convolution stride y
++ * @param[in] bias pointer to bias
++ * @param[in,out] Im_out pointer to output tensor
++ * @param[in] dim_im_out_x output tensor dimension x
++ * @param[in] dim_im_out_y output tensor dimension y
++ * @param[in,out] bufferA pointer to buffer space for input
++ * @param[in,out] bufferB pointer to buffer space for output
++ * @return The function returns ARM_MATH_SUCCESS
++ *
++ * @details
++ *
++ * Buffer size:
++ *
++ * bufferA size: ch_im_in*dim_kernel_x*dim_kernel_y
++ *
++ * bufferB size: 0
++ *
++ * This basic version is designed to work for any input tensor and weight
++ * dimension.
++ */
++
++arm_status
++arm_convolve_CHW_f32_basic_nonsquare(const float * Im_in,
++ const uint16_t dim_im_in_x,
++ const uint16_t dim_im_in_y,
++ const uint16_t ch_im_in,
++ const float * wt,
++ const uint16_t ch_im_out,
++ const uint16_t dim_kernel_x,
++ const uint16_t dim_kernel_y,
++ const uint16_t padding_x,
++ const uint16_t padding_y,
++ const uint16_t stride_x,
++ const uint16_t stride_y,
++ const float * bias,
++ float * Im_out,
++ const uint16_t dim_im_out_x,
++ const uint16_t dim_im_out_y,
++ float * bufferA,
++ float * bufferB)
++{
++
++ /* Run the following code for Cortex-M4 and Cortex-M7 */
++
++ int16_t i_out_y, i_out_x, i_ker_y, i_ker_x;
++ int16_t i_ker_x_begin, i_ker_y_begin;
++ int16_t i_ker_x_end, i_ker_y_end;
++ int16_t single_in_map_size = dim_im_in_x * dim_im_in_y;
++ int16_t kernel_size_2d = dim_kernel_x * dim_kernel_y;
++
++ uint16_t im2col_out_pixel_index = 0;
++ float *pBuffer = bufferA;
++ float *im_buffer = bufferA;
++ const float *pA;
++ int i;
++
++ /* This part implements the im2col function */
++ for (i_out_y = 0; i_out_y < dim_im_out_y; i_out_y++)
++ {
++ for (i_out_x = 0; i_out_x < dim_im_out_x; i_out_x++)
++ {
++ i_ker_y_begin = i_out_y * stride_y - padding_y;
++ i_ker_y_end = i_ker_y_begin + dim_kernel_y;
++
++ for (i_ker_y = i_ker_y_begin; i_ker_y < i_ker_y_end; i_ker_y++)
++ {
++ i_ker_x_begin = i_out_x * stride_x - padding_x;
++ i_ker_x_end = i_ker_x_begin + dim_kernel_x;
++
++ for (i_ker_x = i_ker_x_begin; i_ker_x < i_ker_x_end; i_ker_x++)
++ {
++ float *pDest = pBuffer + (i_ker_y - i_ker_y_begin) * dim_kernel_x + (i_ker_x - i_ker_x_begin);
++ float *pDestEnd = pDest + ch_im_in * kernel_size_2d;
++
++ if (i_ker_y < 0 || i_ker_y >= dim_im_in_y || i_ker_x < 0 || i_ker_x >= dim_im_in_x)
++ {
++ /* Out of bound zero values */
++ for (; pDest < pDestEnd;)
++ {
++ *pDest = 0;
++ pDest += kernel_size_2d;
++ }
++ } else
++ {
++ const float *pSrc = Im_in + i_ker_y * dim_im_in_x + i_ker_x;
++ for (; pDest < pDestEnd;)
++ {
++ *pDest = *pSrc;
++ pSrc += single_in_map_size;
++ pDest += kernel_size_2d;
++ }
++ }
++ }
++ }
++
++ pA = wt;
++ float *pOut = Im_out++;
++ int16_t map_size_out = dim_im_out_x * dim_im_out_y;
++ for (i = 0; i < ch_im_out; i++)
++ {
++ float sum = 0;
++ float *pB = im_buffer;
++ uint16_t colCnt = (ch_im_in * dim_kernel_x * dim_kernel_y) >> 2;
++
++ if (bias)
++ {
++ sum = bias[i];
++ }
++
++ while (colCnt)
++ {
++ float inA1 = *pA++;
++ float inB1 = *pB++;
++ float inA2 = *pA++;
++ float inB2 = *pB++;
++
++ sum += inA1 * inB1;
++ sum += inA2 * inB2;
++
++ inA1 = *pA++;
++ inB1 = *pB++;
++ inA2 = *pA++;
++ inB2 = *pB++;
++
++ sum += inA1 * inB1;
++ sum += inA2 * inB2;
++
++ colCnt--;
++ }
++ colCnt = (ch_im_in * dim_kernel_x * dim_kernel_y) & 0x3;
++ while (colCnt)
++ {
++ float inA1 = *pA++;
++ float inB1 = *pB++;
++ sum += inA1 * inB1;
++ colCnt--;
++ }
++ *pOut = sum;
++ pOut += map_size_out;
++ }
++
++ /* counter reset */
++ pBuffer = im_buffer;
++ im2col_out_pixel_index++;
++ }
++ }
++
++ /* Return to application */
++ return ARM_MATH_SUCCESS;
++}
++
++/**
++ * @} end of NNConv group
++ */
+diff --git a/CMSIS_5/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_CHW_q15_basic_nonsquare.c CMSIS_5/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_CHW_q15_basic_nonsquare.c
+===CHANGE_NOTICE(3/5)===========================================================
+Sony Corporation added this file to 5.4.0 to support the CHW tensor layout
+================================================================================
+--- /dev/null
++++ CMSIS_5/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_CHW_q15_basic_nonsquare.c
+@@ -0,0 +1,231 @@
++/*
++ * Copyright (C) 2010-2018 Arm Limited or its affiliates. All rights reserved.
++ * Copyright 2018 Sony Corporation
++ *
++ * SPDX-License-Identifier: Apache-2.0
++ *
++ * Licensed under the Apache License, Version 2.0 (the License); you may
++ * not use this file except in compliance with the License.
++ * You may obtain a copy of the License at
++ *
++ * www.apache.org/licenses/LICENSE-2.0
++ *
++ * Unless required by applicable law or agreed to in writing, software
++ * distributed under the License is distributed on an AS IS BASIS, WITHOUT
++ * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
++ * See the License for the specific language governing permissions and
++ * limitations under the License.
++ */
++
++/* ----------------------------------------------------------------------
++ * Title: arm_convolve_CHW_q15_basic_nonsquare.c
++ * Author: Sony Corporation
++ * Description: Sony Corporation added this file to 5.4.0
++ * to support the CHW tensor layout
++ * $Date: 14. September 2018
++ * -------------------------------------------------------------------- */
++
++#include "arm_math.h"
++#include "arm_nnfunctions_nnabla.h"
++
++/**
++ * @ingroup groupNN
++ */
++
++/**
++ * @addtogroup NNConv
++ * @{
++ */
++
++ /**
++ * @brief Basic Q15 version of CHW convolution (non-sqaure shape)
++ * @param[in] Im_in pointer to input tensor
++ * @param[in] dim_im_in_x input tensor dimention x
++ * @param[in] dim_im_in_y input tensor dimention y
++ * @param[in] ch_im_in number of input tensor channels
++ * @param[in] wt pointer to kernel weights
++ * @param[in] ch_im_out number of filters, i.e., output tensor channels
++ * @param[in] dim_kernel_x filter kernel size x
++ * @param[in] dim_kernel_y filter kernel size y
++ * @param[in] padding_x padding sizes x
++ * @param[in] padding_y padding sizes y
++ * @param[in] stride_x convolution stride x
++ * @param[in] stride_y convolution stride y
++ * @param[in] bias pointer to bias
++ * @param[in] bias_shift amount of left-shift for bias
++ * @param[in] out_shift amount of right-shift for output
++ * @param[in,out] Im_out pointer to output tensor
++ * @param[in] dim_im_out_x output tensor dimension x
++ * @param[in] dim_im_out_y output tensor dimension y
++ * @param[in,out] bufferA pointer to buffer space for input
++ * @param[in,out] bufferB pointer to buffer space for output
++ * @return The function returns ARM_MATH_SUCCESS
++ *
++ * @details
++ *
++ * Buffer size:
++ *
++ * bufferA size: ch_im_in*dim_kernel_x*dim_kernel_y
++ *
++ * bufferB size: 0
++ *
++ * This basic version is designed to work for any input tensor and weight
++ * dimension.
++ */
++
++arm_status
++arm_convolve_CHW_q15_basic_nonsquare(const q15_t * Im_in,
++ const uint16_t dim_im_in_x,
++ const uint16_t dim_im_in_y,
++ const uint16_t ch_im_in,
++ const q15_t * wt,
++ const uint16_t ch_im_out,
++ const uint16_t dim_kernel_x,
++ const uint16_t dim_kernel_y,
++ const uint16_t padding_x,
++ const uint16_t padding_y,
++ const uint16_t stride_x,
++ const uint16_t stride_y,
++ const q15_t * bias,
++ const uint16_t bias_shift,
++ const uint16_t out_shift,
++ q15_t * Im_out,
++ const uint16_t dim_im_out_x,
++ const uint16_t dim_im_out_y,
++ q15_t * bufferA,
++ q7_t * bufferB)
++{
++
++ /* Run the following code for Cortex-M4 and Cortex-M7 */
++
++ int16_t i_out_y, i_out_x, i_ker_y, i_ker_x;
++ int16_t i_ker_x_begin, i_ker_y_begin;
++ int16_t i_ker_x_end, i_ker_y_end;
++ int16_t single_in_map_size = dim_im_in_x * dim_im_in_y;
++ int16_t kernel_size_2d = dim_kernel_x * dim_kernel_y;
++
++ uint16_t im2col_out_pixel_index = 0;
++ q15_t *pBuffer = bufferA;
++ q15_t *im_buffer = bufferA;
++ const q15_t *pA;
++ int i;
++
++ /* This part implements the im2col function */
++ for (i_out_y = 0; i_out_y < dim_im_out_y; i_out_y++)
++ {
++ for (i_out_x = 0; i_out_x < dim_im_out_x; i_out_x++)
++ {
++#define USE_CHW_IN_COL
++#ifdef USE_CHW_IN_COL
++
++ i_ker_y_begin = i_out_y * stride_y - padding_y;
++ i_ker_y_end = i_ker_y_begin + dim_kernel_y;
++
++ for (i_ker_y = i_ker_y_begin; i_ker_y < i_ker_y_end; i_ker_y++)
++ {
++ i_ker_x_begin = i_out_x * stride_x - padding_x;
++ i_ker_x_end = i_ker_x_begin + dim_kernel_x;
++
++ for (i_ker_x = i_ker_x_begin; i_ker_x < i_ker_x_end; i_ker_x++)
++ {
++ q15_t *pDest = pBuffer + (i_ker_y - i_ker_y_begin) * dim_kernel_x + (i_ker_x - i_ker_x_begin);
++ q15_t *pDestEnd = pDest + ch_im_in * kernel_size_2d;
++
++ if (i_ker_y < 0 || i_ker_y >= dim_im_in_y || i_ker_x < 0 || i_ker_x >= dim_im_in_x)
++ {
++ /* Out of bound zero values */
++ for (; pDest < pDestEnd;)
++ {
++ *pDest = 0;
++ pDest += kernel_size_2d;
++ }
++ } else
++ {
++ const q15_t *pSrc = Im_in + i_ker_y * dim_im_in_x + i_ker_x;
++ for (; pDest < pDestEnd;)
++ {
++ *pDest = *pSrc;
++ pSrc += single_in_map_size;
++ pDest += kernel_size_2d;
++ }
++ }
++ }
++ }
++#else
++ // HWC in columns
++ for (i_ker_y = i_out_y * stride_y - padding_y; i_ker_y < i_out_y * stride_y - padding_y + dim_kernel_y; i_ker_y++)
++ {
++ for (i_ker_x = i_out_x * stride_x - padding_x; i_ker_x < i_out_x * stride_x - padding_x + dim_kernel_x; i_ker_x++)
++ {
++ if (i_ker_y < 0 || i_ker_y >= dim_im_in_y || i_ker_x < 0 || i_ker_x >= dim_im_in_y)
++ {
++ /* Filling 0 for out-of-bound paddings */
++ /* arm_fill_q15(0, pBuffer, ch_im_in); */
++ memset(pBuffer, 0, sizeof(q15_t)*ch_im_in);
++ } else
++ {
++ /* load CHW patch to HWC column */
++ const q15_t *pSrc = Im_in + i_ker_y * dim_im_in_x + i_ker_x;
++ for (int16_t ch_idx = 0; ch_idx < ch_im_in; ch_idx++)
++ {
++ pBuffer[ch_idx++] = *pSrc;
++ pSrc += single_in_map_size;
++ }
++ }
++
++ pBuffer += ch_im_in;
++ }
++ }
++#endif
++
++ pA = wt;
++ q15_t *pOut = Im_out++;
++ int16_t map_size_out = dim_im_out_x * dim_im_out_y;
++ for (i = 0; i < ch_im_out; i++)
++ {
++ q31_t sum = 0;
++ q15_t *pB = im_buffer;
++ uint16_t colCnt = (ch_im_in * dim_kernel_x * dim_kernel_y) >> 2;
++
++ if (bias)
++ {
++ sum = ((q31_t)bias[i] << bias_shift) + NN_ROUND(out_shift);
++ }
++
++ while (colCnt)
++ {
++ q31_t inA1 = *__SIMD32(pA)++;
++ q31_t inB1 = *__SIMD32(pB)++;
++ q31_t inA2 = *__SIMD32(pA)++;
++ q31_t inB2 = *__SIMD32(pB)++;
++
++ sum = __SMLAD(inA1, inB1, sum);
++ sum = __SMLAD(inA2, inB2, sum);
++
++ colCnt--;
++ }
++ colCnt = (ch_im_in * dim_kernel_x * dim_kernel_y) & 0x3;
++ while (colCnt)
++ {
++ q15_t inA1 = *pA++;
++ q15_t inB1 = *pB++;
++ sum += inA1 * inB1;
++ colCnt--;
++ }
++ *pOut = (q15_t) __SSAT((sum >> out_shift), 16);
++ pOut += map_size_out;
++ }
++
++ /* counter reset */
++ pBuffer = im_buffer;
++ im2col_out_pixel_index++;
++ }
++ }
++
++ /* Return to application */
++ return ARM_MATH_SUCCESS;
++}
++
++/**
++ * @} end of NNConv group
++ */
+diff --git a/CMSIS_5/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_CHW_q7_basic_nonsquare.c CMSIS_5/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_CHW_q7_basic_nonsquare.c
+===CHANGE_NOTICE(4/5)===========================================================
+Sony Corporation added this file to 5.4.0 to support the CHW tensor layout
+================================================================================
+--- /dev/null
++++ CMSIS_5/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_CHW_q7_basic_nonsquare.c
+@@ -0,0 +1,214 @@
++/*
++ * Copyright (C) 2010-2018 Arm Limited or its affiliates. All rights reserved.
++ * Copyright 2018 Sony Corporation
++ *
++ * SPDX-License-Identifier: Apache-2.0
++ *
++ * Licensed under the Apache License, Version 2.0 (the License); you may
++ * not use this file except in compliance with the License.
++ * You may obtain a copy of the License at
++ *
++ * www.apache.org/licenses/LICENSE-2.0
++ *
++ * Unless required by applicable law or agreed to in writing, software
++ * distributed under the License is distributed on an AS IS BASIS, WITHOUT
++ * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
++ * See the License for the specific language governing permissions and
++ * limitations under the License.
++ */
++
++/* ----------------------------------------------------------------------
++ * Title: arm_convolve_CHW_q7_basic_nonsquare.c
++ * Author: Sony Corporation
++ * Description: Sony Corporation added this file to 5.4.0
++ * to support the CHW tensor layout
++ * $Date: 14. September 2018
++ * -------------------------------------------------------------------- */
++
++#include "arm_math.h"
++#include "arm_nnfunctions_nnabla.h"
++
++/**
++ * @ingroup groupNN
++ */
++
++/**
++ * @addtogroup NNConv
++ * @{
++ */
++
++ /**
++ * @brief Basic Q7 version of CHW convolution (non-sqaure shape)
++ * @param[in] Im_in pointer to input tensor
++ * @param[in] dim_im_in_x input tensor dimention x
++ * @param[in] dim_im_in_y input tensor dimention y
++ * @param[in] ch_im_in number of input tensor channels
++ * @param[in] wt pointer to kernel weights
++ * @param[in] ch_im_out number of filters, i.e., output tensor channels
++ * @param[in] dim_kernel_x filter kernel size x
++ * @param[in] dim_kernel_y filter kernel size y
++ * @param[in] padding_x padding size x
++ * @param[in] padding_y padding size y
++ * @param[in] stride_x convolution stride x
++ * @param[in] stride_y convolution stride y
++ * @param[in] bias pointer to bias
++ * @param[in] bias_shift amount of left-shift for bias
++ * @param[in] out_shift amount of right-shift for output
++ * @param[in,out] Im_out pointer to output tensor
++ * @param[in] dim_im_out_x output tensor dimension x
++ * @param[in] dim_im_out_y output tensor dimension y
++ * @param[in,out] bufferA pointer to buffer space for input
++ * @param[in,out] bufferB pointer to buffer space for output
++ * @return The function returns ARM_MATH_SUCCESS
++ */
++
++arm_status arm_convolve_CHW_q7_basic_nonsquare(const q7_t * Im_in,
++ const uint16_t dim_im_in_x,
++ const uint16_t dim_im_in_y,
++ const uint16_t ch_im_in,
++ const q7_t * wt,
++ const uint16_t ch_im_out,
++ const uint16_t dim_kernel_x,
++ const uint16_t dim_kernel_y,
++ const uint16_t padding_x,
++ const uint16_t padding_y,
++ const uint16_t stride_x,
++ const uint16_t stride_y,
++ const q7_t * bias,
++ const uint16_t bias_shift,
++ const uint16_t out_shift,
++ q7_t * Im_out,
++ const uint16_t dim_im_out_x,
++ const uint16_t dim_im_out_y,
++ q15_t * bufferA,
++ q7_t * bufferB)
++{
++
++ /* Run the following code for Cortex-M4 and Cortex-M7 */
++
++ int16_t i_out_y, i_out_x, i_ker_y, i_ker_x;
++ int16_t i_ker_x_begin, i_ker_y_begin;
++ int16_t i_ker_x_end, i_ker_y_end;
++ int16_t single_in_map_size = dim_im_in_x * dim_im_in_y;
++ int16_t kernel_size_2d = dim_kernel_x * dim_kernel_y;
++ int16_t kernel_size_3d = ch_im_in * kernel_size_2d;
++
++ /*
++ * Here we use bufferA as q15_t internally as computation are done with q15_t level
++ * im2col are done to output in q15_t format from q7_t input
++ */
++ q15_t *pBuffer = bufferA;
++ q7_t *pOut = Im_out;
++
++ /* This part implements the im2col function */
++ for (i_out_y = 0; i_out_y < dim_im_out_y; i_out_y++)
++ {
++ for (i_out_x = 0; i_out_x < dim_im_out_x; i_out_x++)
++ {
++ i_ker_y_begin = i_out_y * stride_y - padding_y;
++ i_ker_y_end = i_out_y * stride_y - padding_y + dim_kernel_y;
++
++ for (i_ker_y = i_ker_y_begin; i_ker_y < i_ker_y_end; i_ker_y++)
++ {
++
++ i_ker_x_begin = i_out_x * stride_x - padding_x;
++ i_ker_x_end = i_out_x * stride_x - padding_x + dim_kernel_x;
++
++ for (i_ker_x = i_ker_x_begin; i_ker_x < i_ker_x_end; i_ker_x++)
++ {
++
++ q15_t *pDest = pBuffer + (i_ker_y - i_ker_y_begin) * dim_kernel_x + (i_ker_x - i_ker_x_begin);
++ q15_t *pDestEnd = pDest + ch_im_in * kernel_size_2d;
++
++ if (i_ker_y < 0 || i_ker_y >= dim_im_in_y || i_ker_x < 0 || i_ker_x >= dim_im_in_x)
++ {
++ /* Filling 0 for out-of-bound paddings */
++ for (; pDest < pDestEnd;)
++ {
++ *pDest = 0;
++ pDest += kernel_size_2d;
++ }
++ } else
++ {
++ /* Copying the pixel data to column */
++ const q7_t *pSrc = Im_in + i_ker_y * dim_im_in_x + i_ker_x;
++ for (; pDest < pDestEnd;)
++ {
++ *pDest = *pSrc;
++ pSrc += single_in_map_size;
++ pDest += kernel_size_2d;
++ }
++ }
++ }
++ }
++
++ pBuffer += kernel_size_3d;
++
++ /* Computation is filed for every 2 columns */
++ if (pBuffer == bufferA + 2 * kernel_size_3d)
++ {
++ pOut =
++ arm_nn_CHW_mat_mult_kernel_q7_q15(wt, bufferA,
++ ch_im_out,
++ ch_im_in *
++ dim_kernel_y * dim_kernel_x,
++ dim_im_out_y * dim_im_out_x,
++ bias_shift, out_shift, bias, pOut);
++
++ /* counter reset */
++ pBuffer = bufferA;
++ }
++ }
++ }
++
++ /* left-over because odd number of output pixels */
++ if (pBuffer != bufferA)
++ {
++ const q7_t *pA = wt;
++ int i;
++
++ for (i = 0; i < ch_im_out; i++)
++ {
++ /* Load the accumulator with bias first */
++ q31_t sum = ((q31_t)bias[i] << bias_shift) + NN_ROUND(out_shift);
++
++ /* Point to the beging of the im2col buffer */
++ q15_t *pB = bufferA;
++
++ /* Each time it process 4 entries */
++ uint16_t colCnt = kernel_size_3d >> 2;
++
++ while (colCnt)
++ {
++ q31_t inA1, inA2;
++ q31_t inB1, inB2;
++
++ pA = (q7_t *) read_and_pad((void *)pA, &inA1, &inA2);
++
++ inB1 = *__SIMD32(pB)++;
++ sum = __SMLAD(inA1, inB1, sum);
++ inB2 = *__SIMD32(pB)++;
++ sum = __SMLAD(inA2, inB2, sum);
++
++ colCnt--;
++ }
++ colCnt = kernel_size_3d & 0x3;
++ while (colCnt)
++ {
++ q7_t inA1 = *pA++;
++ q15_t inB1 = *pB++;
++ sum += inA1 * inB1;
++ colCnt--;
++ }
++ *pOut = (q7_t) __SSAT((sum >> out_shift), 8);
++ pOut += dim_im_out_y * dim_im_out_x;
++ }
++ }
++
++ /* Return to application */
++ return ARM_MATH_SUCCESS;
++}
++
++/**
++ * @} end of NNConv group
++ */
+diff --git a/CMSIS_5/CMSIS/NN/Source/ConvolutionFunctions/arm_nn_CHW_mat_mult_kernel_q7_q15.c CMSIS_5/CMSIS/NN/Source/ConvolutionFunctions/arm_nn_CHW_mat_mult_kernel_q7_q15.c
+===CHANGE_NOTICE(5/5)===========================================================
+Sony Corporation added this file to 5.4.0 to support the CHW tensor layout
+================================================================================
+--- /dev/null
++++ CMSIS_5/CMSIS/NN/Source/ConvolutionFunctions/arm_nn_CHW_mat_mult_kernel_q7_q15.c
+@@ -0,0 +1,196 @@
++/*
++ * Copyright (C) 2010-2018 Arm Limited or its affiliates. All rights reserved.
++ * Copyright 2018 Sony Corporation
++ *
++ * SPDX-License-Identifier: Apache-2.0
++ *
++ * Licensed under the Apache License, Version 2.0 (the License); you may
++ * not use this file except in compliance with the License.
++ * You may obtain a copy of the License at
++ *
++ * www.apache.org/licenses/LICENSE-2.0
++ *
++ * Unless required by applicable law or agreed to in writing, software
++ * distributed under the License is distributed on an AS IS BASIS, WITHOUT
++ * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
++ * See the License for the specific language governing permissions and
++ * limitations under the License.
++ */
++
++/* ----------------------------------------------------------------------
++ * Title: arm_nn_CHW_mat_mult_kernel_q7_q15.c
++ * Author: Sony Corporation
++ * Description: Sony Corporation added this file to 5.4.0
++ * to support the CHW tensor layout in convolution
++ * $Date: 14. September 2018
++ * -------------------------------------------------------------------- */
++
++#include "arm_math.h"
++#include "arm_nnfunctions_nnabla.h"
++
++ /**
++ * @brief Matrix-multiplication function for convolution with CHW output
++ * @param[in] pA pointer to operand A
++ * @param[in] pInBuffer pointer to operand B, always conssists of 2 vectors
++ * @param[in] ch_im_out numRow of A
++ * @param[in] numCol_A numCol of A
++ * @param[in] out_stride output buffer channel stride
++ * @param[in] bias_shift amount of left-shift for bias
++ * @param[in] out_shift amount of right-shift for output
++ * @param[in] bias the bias
++ * @param[in,out] pOut pointer to output
++ * @return The function returns the incremented output pointer
++ *
++ * @details
++ *
++ * This function does the matrix multiplication with weight matrix
++ * and 2 columns from im2col.
++ */
++
++q7_t *arm_nn_CHW_mat_mult_kernel_q7_q15(const q7_t * pA,
++ const q15_t * pInBuffer,
++ const uint16_t ch_im_out,
++ const uint16_t numCol_A,
++ const uint16_t out_stride,
++ const uint16_t bias_shift,
++ const uint16_t out_shift,
++ const q7_t * bias,
++ q7_t * pOut)
++{
++ /* set up the second output pointers */
++ q7_t *pOut_base = pOut;
++ q7_t *pOut2;
++ const q7_t *pBias = bias;
++ int16_t i_row;
++
++ uint16_t rowCnt = ch_im_out >> 1;
++ /* this loop over rows in A */
++ for (i_row = 0; i_row < rowCnt; ++i_row)
++ {
++ /* setup output pointers */
++ pOut = pOut_base + 2 * i_row * out_stride;
++ pOut2 = pOut + out_stride;
++
++ /* setup pointers for B */
++ const q15_t *pB = pInBuffer;
++ const q15_t *pB2 = pB + numCol_A;
++
++ /* align the second pointer for A */
++ const q7_t *pA2 = pA + numCol_A;
++
++ /* sum & sum3 belong to same outmap, sum2 & sum4 belong to another outmap
++ *
++ * sum sum3
++ * sum2 sum4
++ *
++ */
++ /* init the sum with bias */
++ q31_t sum = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift);
++ q31_t sum3 = sum;
++ q31_t sum2 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift);
++ q31_t sum4 = sum2;
++
++ uint16_t colCnt = numCol_A >> 2;
++ /* accumulate over the vector */
++ while (colCnt)
++ {
++ q31_t inA11, inA12, inA21, inA22;
++ q31_t inB1 = *__SIMD32(pB)++;
++ q31_t inB2 = *__SIMD32(pB2)++;
++
++ /* pA is in CHW -> inA11 & inA12 belong to same out-map weight */
++ pA = (q7_t *) read_and_pad((void *)pA, &inA11, &inA12);
++ pA2 = (q7_t *) read_and_pad((void *)pA2, &inA21, &inA22);
++
++ /* inB1 belongs to the first columns, inB2 is the second column */
++ sum = __SMLAD(inA11, inB1, sum);
++ sum3 = __SMLAD(inA11, inB2, sum3);
++ sum2 = __SMLAD(inA21, inB1, sum2);
++ sum4 = __SMLAD(inA21, inB2, sum4);
++
++ inB1 = *__SIMD32(pB)++;
++ inB2 = *__SIMD32(pB2)++;
++
++ sum = __SMLAD(inA12, inB1, sum);
++ sum3 = __SMLAD(inA12, inB2, sum3);
++ sum2 = __SMLAD(inA22, inB1, sum2);
++ sum4 = __SMLAD(inA22, inB2, sum4);
++
++ colCnt--;
++ } /* while over colCnt */
++ colCnt = numCol_A & 0x3;
++ while (colCnt)
++ {
++ q7_t inA1 = *pA++;
++ q15_t inB1 = *pB++;
++ q7_t inA2 = *pA2++;
++ q15_t inB2 = *pB2++;
++
++ sum += inA1 * inB1;
++ sum3 += inA1 * inB2;
++ sum2 += inA2 * inB1;
++ sum4 += inA2 * inB2;
++ colCnt--;
++ } /* while over colCnt */
++ *pOut++ = (q7_t) __SSAT((sum >> out_shift), 8);
++ *pOut = (q7_t) __SSAT((sum3 >> out_shift), 8);
++ *pOut2++ = (q7_t) __SSAT((sum2 >> out_shift), 8);
++ *pOut2 = (q7_t) __SSAT((sum4 >> out_shift), 8);
++
++ /* skip the row computed with A2 */
++ pA += numCol_A;
++ } /* for over ch_im_out */
++
++ /* compute left-over row if any */
++ if (ch_im_out & 0x1)
++ {
++ /* setup output pointers */
++ pOut = pOut_base + (ch_im_out - 1) * out_stride;
++
++ /* setup pointers for B */
++ const q15_t *pB = pInBuffer;
++ const q15_t *pB2 = pB + numCol_A;
++
++ /* load the bias */
++ q31_t sum = ((q31_t)(*pBias) << bias_shift) + NN_ROUND(out_shift);
++ q31_t sum3 = sum;
++
++ uint16_t colCnt = numCol_A >> 2;
++ while (colCnt)
++ {
++ q31_t inA11, inA12;
++ q31_t inB1 = *__SIMD32(pB)++;
++ q31_t inB2 = *__SIMD32(pB2)++;
++
++ pA = (q7_t *) read_and_pad((void *)pA, &inA11, &inA12);
++
++ sum = __SMLAD(inA11, inB1, sum);
++ sum3 = __SMLAD(inA11, inB2, sum3);
++
++ inB1 = *__SIMD32(pB)++;
++ inB2 = *__SIMD32(pB2)++;
++
++ sum = __SMLAD(inA12, inB1, sum);
++ sum3 = __SMLAD(inA12, inB2, sum3);
++
++ colCnt--;
++ }
++ colCnt = numCol_A & 0x3;
++ while (colCnt)
++ {
++ q7_t inA1 = *pA++;
++ q15_t inB1 = *pB++;
++ q15_t inB2 = *pB2++;
++
++ sum += inA1 * inB1;
++ sum3 += inA1 * inB2;
++ colCnt--;
++ }
++
++ *pOut++ = (q7_t) __SSAT((sum >> out_shift), 8);
++ *pOut = (q7_t) __SSAT((sum3 >> out_shift), 8);
++ }
++
++ /* return the new output pointer with offset */
++ return pOut_base + 2;
++}