Official ARM version: v5.4.0

    Add CMSIS V5.4.0, please refer to index.html available under \docs folder.

    Note: content of \CMSIS\Core\Include has been copied under \Include to keep the same structure
         used in existing projects, and thus avoid projects mass update
    Note: the following components have been removed from ARM original delivery (as not used in ST packages)
        - CMSIS_EW2018.pdf
        - .gitattributes
        - .gitignore
        - \Device
        - \CMSIS
           - \CoreValidation
		   - \DAP
           - \Documentation
           - \DoxyGen
           - \Driver
           - \Pack
           - \RTOS\CMSIS_RTOS_Tutorial.pdf
           - \RTOS\RTX
           - \RTOS\Template
           - \RTOS2\RTX
           - \Utilities
           - All ARM/GCC projects files are deleted from \DSP, \RTOS and \RTOS2

Change-Id: Ia026c3f0f0d016627a4fb5a9032852c33d24b4d3

31 files changed, 7388 insertions, 0 deletions

diff --git a/NN/Source/ActivationFunctions/arm_nn_activations_q15.c b/NN/Source/ActivationFunctions/arm_nn_activations_q15.c
new file mode 100644
index 0000000..9c64e2a
--- /dev/null
+++ b/NN/Source/ActivationFunctions/arm_nn_activations_q15.c
@@ -0,0 +1,101 @@
+/*
+ * Copyright (C) 2010-2018 Arm Limited or its affiliates. All rights reserved.
+ *
+ * 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.
+ */
+
+/* ----------------------------------------------------------------------
+ * Project:      CMSIS NN Library
+ * Title:        arm_nn_activations_q15.c
+ * Description:  Q15 neural network activation function using direct table look-up
+ *
+ * $Date:        17. January 2018
+ * $Revision:    V.1.0.0
+ *
+ * Target Processor:  Cortex-M cores
+ *
+ * -------------------------------------------------------------------- */
+
+#include "arm_math.h"
+#include "arm_common_tables.h"
+#include "arm_nnfunctions.h"
+
+/**
+ *  @ingroup groupNN
+ */
+
+/**
+ * @addtogroup Acti
+ * @{
+ */
+
+  /**
+   * @brief Q15 neural network activation function using direct table look-up
+   * @param[in,out]   data        pointer to input
+   * @param[in]       size        number of elements
+   * @param[in]       int_width   bit-width of the integer part, assume to be smaller than 3
+   * @param[in]       type        type of activation functions
+   * @return none.
+   *
+   * @details
+   * 
+   * This is the direct table look-up approach.
+   *
+   * Assume here the integer part of the fixed-point is <= 3.
+   * More than 3 just not making much sense, makes no difference with
+   * saturation followed by any of these activation functions. 
+   */
+
+void arm_nn_activations_direct_q15(q15_t * data, uint16_t size, uint16_t int_width, arm_nn_activation_type type)
+{
+    uint16_t  i = size;
+    q15_t    *pIn = data;
+    q15_t    *pOut = data;
+    uint16_t  shift_size = 8 + 3 - int_width;
+    uint32_t  bit_mask = 0x7FF >> int_width;
+    uint32_t  full_frac = bit_mask + 1;
+    const q15_t *lookup_table;
+
+    switch (type)
+    {
+    case ARM_SIGMOID:
+        lookup_table = sigmoidTable_q15;
+        break;
+    case ARM_TANH:
+    default:
+        lookup_table = tanhTable_q15;
+        break;
+    }
+
+    while (i)
+    {
+        q15_t     out;
+        q15_t     in = *pIn++;
+        q15_t     frac = (uint32_t) in & bit_mask;
+        q15_t     value = lookup_table[__USAT(in >> shift_size, 8)];
+        q15_t     value2 = lookup_table[__USAT(1 + (in >> shift_size), 8)];
+
+        /* doing the interpolation here for better accuracy */
+        out = ((q31_t) (full_frac - frac) * value + (q31_t) value2 * frac) >> shift_size;
+
+        *pOut++ = out;
+        i--;
+    }
+
+}
+
+/**
+ * @} end of Acti group
+ */
diff --git a/NN/Source/ActivationFunctions/arm_nn_activations_q7.c b/NN/Source/ActivationFunctions/arm_nn_activations_q7.c
new file mode 100644
index 0000000..1ca429f
--- /dev/null
+++ b/NN/Source/ActivationFunctions/arm_nn_activations_q7.c
@@ -0,0 +1,91 @@
+/*
+ * Copyright (C) 2010-2018 Arm Limited or its affiliates. All rights reserved.
+ *
+ * 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.
+ */
+
+/* ----------------------------------------------------------------------
+ * Project:      CMSIS NN Library
+ * Title:        arm_nn_activations_q7.c
+ * Description:  Q7 neural network activation function using direct table look-up
+ *
+ * $Date:        17. January 2018
+ * $Revision:    V.1.0.0
+ *
+ * Target Processor:  Cortex-M cores
+ *
+ * -------------------------------------------------------------------- */
+
+#include "arm_math.h"
+#include "arm_common_tables.h"
+#include "arm_nnfunctions.h"
+
+/**
+ *  @ingroup groupNN
+ */
+
+/**
+ * @addtogroup Acti
+ * @{
+ */
+
+  /**
+   * @brief Q7 neural network activation function using direct table look-up
+   * @param[in,out]   data        pointer to input
+   * @param[in]       size        number of elements
+   * @param[in]       int_width   bit-width of the integer part, assume to be smaller than 3
+   * @param[in]       type        type of activation functions
+   * @return none.
+   *
+   * @details
+   * 
+   * This is the direct table look-up approach.
+   *
+   * Assume here the integer part of the fixed-point is <= 3.
+   * More than 3 just not making much sense, makes no difference with
+   * saturation followed by any of these activation functions. 
+   */
+
+void arm_nn_activations_direct_q7(q7_t * data, uint16_t size, uint16_t int_width, arm_nn_activation_type type)
+{
+    uint16_t  i = size;
+    q7_t     *pIn = data;
+    q7_t     *pOut = data;
+    q7_t      in;
+    q7_t      out;
+    uint16_t  shift_size = 3 - int_width;
+    const q7_t *lookup_table;
+    switch (type)
+    {
+    case ARM_SIGMOID:
+        lookup_table = sigmoidTable_q7;
+        break;
+    case ARM_TANH:
+    default:
+        lookup_table = tanhTable_q7;
+        break;
+    }
+    while (i)
+    {
+        in = *pIn++;
+        out = lookup_table[(uint8_t) (in >> shift_size)];
+        *pOut++ = out;
+        i--;
+    }
+}
+
+/**
+ * @} end of Acti group
+ */
diff --git a/NN/Source/ActivationFunctions/arm_relu_q15.c b/NN/Source/ActivationFunctions/arm_relu_q15.c
new file mode 100644
index 0000000..571d51c
--- /dev/null
+++ b/NN/Source/ActivationFunctions/arm_relu_q15.c
@@ -0,0 +1,106 @@
+/*
+ * Copyright (C) 2010-2018 Arm Limited or its affiliates. All rights reserved.
+ *
+ * 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.
+ */
+
+/* ----------------------------------------------------------------------
+ * Project:      CMSIS NN Library
+ * Title:        arm_relu_q15.c
+ * Description:  Q15 version of ReLU
+ *
+ * $Date:        17. January 2018
+ * $Revision:    V.1.0.0
+ *
+ * Target Processor:  Cortex-M cores
+ *
+ * -------------------------------------------------------------------- */
+
+#include "arm_math.h"
+#include "arm_nnfunctions.h"
+
+/**
+ *  @ingroup groupNN
+ */
+
+/**
+ * @addtogroup Acti
+ * @{
+ */
+
+  /**
+   * @brief Q15 RELU function
+   * @param[in,out]   data        pointer to input
+   * @param[in]       size        number of elements
+   * @return none.
+   * 
+   * @details
+   *
+   * Optimized relu with QSUB instructions.
+   *
+   */
+
+void arm_relu_q15(q15_t * data, uint16_t size)
+{
+
+#if defined (ARM_MATH_DSP)
+    /* Run the following code for Cortex-M4 and Cortex-M7 */
+
+    uint16_t  i = size >> 1;
+    q15_t    *pIn = data;
+    q15_t    *pOut = data;
+    q31_t     in;
+    q31_t     buf;
+    q31_t     mask;
+
+    while (i)
+    {
+        in = *__SIMD32(pIn)++;
+
+        /* extract the first bit */
+        buf = __ROR(in & 0x80008000, 15);
+
+        /* if MSB=1, mask will be 0xFF, 0x0 otherwise */
+        mask = __QSUB16(0x00000000, buf);
+
+        *__SIMD32(pOut)++ = in & (~mask);
+        i--;
+    }
+
+    if (size & 0x1)
+    {
+        if (*pIn < 0)
+        {
+            *pIn = 0;
+        }
+        pIn++;
+    }
+#else
+    /* Run the following code as reference implementation for Cortex-M0 and Cortex-M3 */
+    uint16_t  i;
+
+    for (i = 0; i < size; i++)
+    {
+        if (data[i] < 0)
+            data[i] = 0;
+    }
+
+#endif                          /* ARM_MATH_DSP */
+
+}
+
+/**
+ * @} end of Acti group
+ */
diff --git a/NN/Source/ActivationFunctions/arm_relu_q7.c b/NN/Source/ActivationFunctions/arm_relu_q7.c
new file mode 100644
index 0000000..013325c
--- /dev/null
+++ b/NN/Source/ActivationFunctions/arm_relu_q7.c
@@ -0,0 +1,110 @@
+/*
+ * Copyright (C) 2010-2018 Arm Limited or its affiliates. All rights reserved.
+ *
+ * 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.
+ */
+
+/* ----------------------------------------------------------------------
+ * Project:      CMSIS NN Library
+ * Title:        arm_relu_q7.c
+ * Description:  Q7 version of ReLU
+ *
+ * $Date:        17. January 2018
+ * $Revision:    V.1.0.0
+ *
+ * Target Processor:  Cortex-M cores
+ *
+ * -------------------------------------------------------------------- */
+
+#include "arm_math.h"
+#include "arm_nnfunctions.h"
+
+/**
+ *  @ingroup groupNN
+ */
+
+/**
+ * @addtogroup Acti
+ * @{
+ */
+
+  /**
+   * @brief Q7 RELU function
+   * @param[in,out]   data        pointer to input
+   * @param[in]       size        number of elements
+   * @return none.
+   * 
+   * @details
+   *
+   * Optimized relu with QSUB instructions.
+   *
+   */
+
+void arm_relu_q7(q7_t * data, uint16_t size)
+{
+
+#if defined (ARM_MATH_DSP)
+    /* Run the following code for Cortex-M4 and Cortex-M7 */
+
+    uint16_t  i = size >> 2;
+    q7_t     *pIn = data;
+    q7_t     *pOut = data;
+    q31_t     in;
+    q31_t     buf;
+    q31_t     mask;
+
+    while (i)
+    {
+        in = *__SIMD32(pIn)++;
+
+        /* extract the first bit */
+        buf = __ROR(in & 0x80808080, 7);
+
+        /* if MSB=1, mask will be 0xFF, 0x0 otherwise */
+        mask = __QSUB8(0x00000000, buf);
+
+        *__SIMD32(pOut)++ = in & (~mask);
+        i--;
+    }
+
+    i = size & 0x3;
+    while (i)
+    {
+        if (*pIn < 0)
+        {
+            *pIn = 0;
+        }
+        pIn++;
+        i--;
+    }
+
+#else
+    /* Run the following code as reference implementation for Cortex-M0 and Cortex-M3 */
+
+    uint16_t  i;
+
+    for (i = 0; i < size; i++)
+    {
+        if (data[i] < 0)
+            data[i] = 0;
+    }
+
+#endif                          /* ARM_MATH_DSP */
+
+}
+
+/**
+ * @} end of Acti group
+ */
diff --git a/NN/Source/ConvolutionFunctions/arm_convolve_1x1_HWC_q7_fast_nonsquare.c b/NN/Source/ConvolutionFunctions/arm_convolve_1x1_HWC_q7_fast_nonsquare.c
new file mode 100644
index 0000000..2f4133c
--- /dev/null
+++ b/NN/Source/ConvolutionFunctions/arm_convolve_1x1_HWC_q7_fast_nonsquare.c
@@ -0,0 +1,235 @@
+/*
+ * Copyright (C) 2010-2018 Arm Limited or its affiliates. All rights reserved.
+ *
+ * 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.
+ */
+
+/* ----------------------------------------------------------------------
+ * Project:      CMSIS NN Library
+ * Title:        arm_convolve_1x1_HWC_q7_fast_nonsquare.c
+ * Description:  Fast Q7 version of 1x1 convolution (non-square shape)
+ *
+ * $Date:        17. January 2018
+ * $Revision:    V.1.0.0
+ *
+ * Target Processor:  Cortex-M cores
+ *
+ * -------------------------------------------------------------------- */
+
+#include "arm_math.h"
+#include "arm_nnfunctions.h"
+
+/**
+ *  @ingroup groupNN
+ */
+
+/**
+ * @addtogroup NNConv
+ * @{
+ */
+
+/**
+ * @brief Fast Q7 version of 1x1 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 either
+ * <code>ARM_MATH_SIZE_MISMATCH</code> or <code>ARM_MATH_SUCCESS</code> based on the outcome of size checking.
+ *
+ * This function is optimized for convolution with 1x1 kernel size (i.e., dim_kernel_x=1
+ * and dim_kernel_y=1). It can be used for the second half of MobileNets [1] after depthwise 
+ * separable convolution.
+ *
+ * This function is the version with full list of optimization tricks, but with
+ * some contraints:
+ *   ch_im_in is multiple of 4
+ *   ch_im_out is multiple of 2
+ *
+ * [1] MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications
+ * https://arxiv.org/abs/1704.04861
+ */
+
+arm_status arm_convolve_1x1_HWC_q7_fast_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)
+{
+
+#if defined (ARM_MATH_DSP)
+    /* Run the following code for Cortex-M4 and Cortex-M7 */
+
+    int16_t   i_out_y, i_out_x;
+    int16_t   i_ch_out;
+
+    /* -----------------------
+     *  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;
+
+    if (ch_im_in % 4 != 0 || ch_im_out % 2 != 0 || dim_kernel_x != 1 || dim_kernel_y != 1
+        || padding_x != 0 || padding_y != 0 || stride_x != 1 || stride_y != 1)
+    {
+        /* check if the input dimension meets the constraints */
+        return ARM_MATH_SIZE_MISMATCH;
+    }
+
+    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++)
+        {
+            /* This part implements the im2col function */
+            arm_q7_to_q15_reordered_no_shift((q7_t *) Im_in + (i_out_y * dim_im_in_x + i_out_x) * ch_im_in, pBuffer,
+                                             ch_im_in);
+            pBuffer += ch_im_in;
+
+            if (pBuffer == bufferA + 2 * ch_im_in * dim_kernel_x * dim_kernel_y)
+            {
+                pOut =
+                    arm_nn_mat_mult_kernel_q7_q15_reordered(wt, bufferA, ch_im_out, ch_im_in, bias_shift, out_shift, bias, pOut);
+                /* counter reset */
+                pBuffer = bufferA;
+            }
+        }
+    }
+
+    /* check if there is left-over for compute */
+    if (pBuffer != bufferA)
+    {
+        const q7_t *pA = wt;
+        for (i_ch_out = 0; i_ch_out < ch_im_out; i_ch_out++)
+        {
+            q31_t     sum = ((q31_t)(bias[i_ch_out]) << bias_shift) + NN_ROUND(out_shift);
+            q15_t    *pB = bufferA;
+            /* basically each time it process 4 entries */
+            uint16_t  colCnt = ch_im_in * dim_kernel_x * dim_kernel_y >> 2;
+
+            while (colCnt)
+            {
+
+                q31_t     inA1, inA2;
+                q31_t     inB1, inB2;
+
+                pA = (const q7_t *)read_and_pad_reordered((void *)pA, &inA1, &inA2);
+
+                inB1 = *__SIMD32(pB)++;
+                sum = __SMLAD(inA1, inB1, sum);
+                inB2 = *__SIMD32(pB)++;
+                sum = __SMLAD(inA2, inB2, sum);
+
+                colCnt--;
+            }
+            colCnt = ch_im_in * dim_kernel_y * dim_kernel_x & 0x3;
+            while (colCnt)
+            {
+                q7_t      inA1 = *pA++;
+                q15_t     inB1 = *pB++;
+                sum += inA1 * inB1;
+                colCnt--;
+            }
+            *pOut = (q7_t) __SSAT((sum >> out_shift), 8);
+            pOut++;
+
+        }
+
+    }
+
+#else
+    /* Run the following code as reference implementation for Cortex-M0 and Cortex-M3 */
+		
+    int       i, j, k, l, m, n;
+    int       conv_out;
+    int       in_row, in_col;
+
+    if (ch_im_in % 4 != 0 || ch_im_out % 2 != 0 || dim_kernel_x != 1 || dim_kernel_y != 1
+        || padding_x != 0 || padding_y != 0 || stride_x != 1 || stride_y != 1)
+    {
+        /* check if the input dimension meets the constraints */
+        return ARM_MATH_SIZE_MISMATCH;
+    }
+
+    for (i = 0; i < ch_im_out; i++)
+    {
+        for (j = 0; j < dim_im_out_y; j++)
+        {
+            for (k = 0; k < dim_im_out_x; k++)
+            {
+                conv_out = ((q31_t)(bias[i]) << bias_shift) + NN_ROUND(out_shift);
+                for (m = 0; m < dim_kernel_y; m++)
+                {
+                    for (n = 0; n < dim_kernel_x; n++)
+                    {
+                        // if-for implementation
+                        in_row = stride_y * j + m - padding_y;
+                        in_col = stride_x * k + n - padding_x;
+                        if (in_row >= 0 && in_col >= 0 && in_row < dim_im_in_y && in_col < dim_im_in_x)
+                        {
+                            for (l = 0; l < ch_im_in; l++)
+                            {
+                                conv_out += Im_in[(in_row * dim_im_in_x + in_col) * ch_im_in + l] *
+                                    wt[i * ch_im_in * dim_kernel_y * dim_kernel_x + (m * dim_kernel_y + n) * ch_im_in + l];
+                            }
+                        }
+                    }
+                }
+                Im_out[i + (j * dim_im_out_x + k) * ch_im_out] = (q7_t) __SSAT((conv_out >> out_shift), 8);
+            }
+        }
+    }
+
+#endif                          /* ARM_MATH_DSP */
+
+    /* Return to application */
+    return ARM_MATH_SUCCESS;
+}
+
+/**
+ * @} end of NNConv group
+ */
diff --git a/NN/Source/ConvolutionFunctions/arm_convolve_HWC_q15_basic.c b/NN/Source/ConvolutionFunctions/arm_convolve_HWC_q15_basic.c
new file mode 100644
index 0000000..00b5aa5
--- /dev/null
+++ b/NN/Source/ConvolutionFunctions/arm_convolve_HWC_q15_basic.c
@@ -0,0 +1,207 @@
+/*
+ * Copyright (C) 2010-2018 Arm Limited or its affiliates. All rights reserved.
+ *
+ * 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.
+ */
+
+/* ----------------------------------------------------------------------
+ * Project:      CMSIS NN Library
+ * Title:        arm_convolve_HWC_q15_basic.c
+ * Description:  Q15 version of convolution
+ *
+ * $Date:        17. January 2018
+ * $Revision:    V.1.0.0
+ *
+ * Target Processor:  Cortex-M cores
+ *
+ * -------------------------------------------------------------------- */
+
+#include "arm_math.h"
+#include "arm_nnfunctions.h"
+
+/**
+ *  @ingroup groupNN
+ */
+
+/**
+ * @addtogroup NNConv
+ * @{
+ */
+
+  /**
+   * @brief Basic Q15 convolution function
+   * @param[in]       Im_in       pointer to input tensor
+   * @param[in]       dim_im_in   input tensor dimention
+   * @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  filter kernel size
+   * @param[in]       padding     padding sizes
+   * @param[in]       stride      convolution stride
+   * @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  output tensor dimension
+   * @param[in,out]   bufferA     pointer to buffer space for input
+   * @param[in,out]   bufferB     pointer to buffer space for output
+   * @return     The function returns <code>ARM_MATH_SUCCESS</code> 
+   *
+   * @details
+   *
+   * <b>Buffer size:</b>
+   *
+   * bufferA size: ch_im_in*dim_kernel*dim_kernel
+   *
+   * bufferB size: 0
+   *
+   * This basic version is designed to work for any input tensor and weight
+   * dimension. 
+   */
+
+arm_status
+arm_convolve_HWC_q15_basic(const q15_t * Im_in,
+                           const uint16_t dim_im_in,
+                           const uint16_t ch_im_in,
+                           const q15_t * wt,
+                           const uint16_t ch_im_out,
+                           const uint16_t dim_kernel,
+                           const uint16_t padding,
+                           const uint16_t stride,
+                           const q15_t * bias,
+                           const uint16_t bias_shift,
+                           const uint16_t out_shift,
+                           q15_t * Im_out, 
+                           const uint16_t dim_im_out, 
+                           q15_t * bufferA, 
+                           q7_t * bufferB)
+{
+
+#if defined (ARM_MATH_DSP)
+    /* Run the following code for Cortex-M4 and Cortex-M7 */
+
+    int16_t   i_out_y, i_out_x, i_ker_y, i_ker_x;
+
+    uint16_t  im2col_out_pixel_index = 0;
+    q15_t    *pBuffer = bufferA;
+    q15_t    *pOut = Im_out;
+    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; i_out_y++)
+    {
+        for (i_out_x = 0; i_out_x < dim_im_out; i_out_x++)
+        {
+            for (i_ker_y = i_out_y * stride - padding; i_ker_y < i_out_y * stride - padding + dim_kernel; i_ker_y++)
+            {
+                for (i_ker_x = i_out_x * stride - padding; i_ker_x < i_out_x * stride - padding + dim_kernel; i_ker_x++)
+                {
+                    if (i_ker_y < 0 || i_ker_y >= dim_im_in || i_ker_x < 0 || i_ker_x >= dim_im_in)
+                    {
+                        /* 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
+                    {
+                        /* arm_copy_q15((q15_t *) Im_in + (i_ker_y * dim_im_in + i_ker_x) * ch_im_in, pBuffer, ch_im_in); */
+                        memcpy(pBuffer, (q15_t *) Im_in + (i_ker_y * dim_im_in + i_ker_x) * ch_im_in, sizeof(q15_t)*ch_im_in);
+                    }
+                    pBuffer += ch_im_in;
+                }
+            }
+
+            pA = wt;
+            for (i = 0; i < ch_im_out; i++)
+            {
+                q31_t     sum = ((q31_t)bias[i] << bias_shift) + NN_ROUND(out_shift);
+                q15_t    *pB = im_buffer;
+                uint16_t  colCnt = ch_im_in * dim_kernel * dim_kernel >> 2;
+                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 * dim_kernel & 0x3;
+                while (colCnt)
+                {
+                    q15_t     inA1 = *pA++;
+                    q15_t     inB1 = *pB++;
+                    sum += inA1 * inB1;
+                    colCnt--;
+                }
+                *pOut = (q15_t) __SSAT((sum >> out_shift), 16);
+                pOut++;
+            }
+
+            /* counter reset */
+            pBuffer = im_buffer;
+            im2col_out_pixel_index++;
+        }
+    }
+
+#else
+    /* Run the following code as reference implementation for Cortex-M0 and Cortex-M3 */
+    uint16_t  i, j, k, l, m, n;
+    int       conv_out;
+    signed char in_row, in_col;
+
+    for (i = 0; i < ch_im_out; i++)
+    {
+        for (j = 0; j < dim_im_out; j++)
+        {
+            for (k = 0; k < dim_im_out; k++)
+            {
+                conv_out = ((q31_t)bias[i] << bias_shift) + NN_ROUND(out_shift);
+                for (m = 0; m < dim_kernel; m++)
+                {
+                    for (n = 0; n < dim_kernel; n++)
+                    {
+                        in_row = stride * j + m - padding;
+                        in_col = stride * k + n - padding;
+                        if (in_row >= 0 && in_col >= 0 && in_row < dim_im_in && in_col < dim_im_in)
+                        {
+                            for (l = 0; l < ch_im_in; l++)
+                            {
+                                conv_out +=
+                                    Im_in[(in_row * dim_im_in + in_col) * ch_im_in +
+                                          l] * wt[i * ch_im_in * dim_kernel * dim_kernel + (m * dim_kernel +
+                                                                                            n) * ch_im_in + l];
+                            }
+                        }
+                    }
+                }
+                Im_out[i + (j * dim_im_out + k) * ch_im_out] = (q15_t) __SSAT((conv_out >> out_shift), 16);
+            }
+        }
+    }
+
+#endif                          /* ARM_MATH_DSP */
+
+    /* Return to application */
+    return ARM_MATH_SUCCESS;
+}
+
+/**
+ * @} end of NNConv group
+ */
diff --git a/NN/Source/ConvolutionFunctions/arm_convolve_HWC_q15_fast.c b/NN/Source/ConvolutionFunctions/arm_convolve_HWC_q15_fast.c
new file mode 100644
index 0000000..c9873c1
--- /dev/null
+++ b/NN/Source/ConvolutionFunctions/arm_convolve_HWC_q15_fast.c
@@ -0,0 +1,255 @@
+/*
+ * Copyright (C) 2010-2018 Arm Limited or its affiliates. All rights reserved.
+ *
+ * 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.
+ */
+
+/* ----------------------------------------------------------------------
+ * Project:      CMSIS NN Library
+ * Title:        arm_convolve_HWC_q15_fast.c
+ * Description:  Fast Q15 version of convolution
+ *
+ * $Date:        17. January 2018
+ * $Revision:    V.1.0.0
+ *
+ * Target Processor:  Cortex-M cores
+ *
+ * -------------------------------------------------------------------- */
+
+#include "arm_math.h"
+#include "arm_nnfunctions.h"
+
+/**
+ *  @ingroup groupNN
+ */
+
+/**
+ * @addtogroup NNConv
+ * @{
+ */
+
+  /**
+   * @brief Fast Q15 convolution function
+   * @param[in]       Im_in       pointer to input tensor
+   * @param[in]       dim_im_in   input tensor dimention
+   * @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  filter kernel size
+   * @param[in]       padding     padding sizes
+   * @param[in]       stride      convolution stride
+   * @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  output tensor dimension
+   * @param[in,out]   bufferA     pointer to buffer space for input 
+   * @param[in,out]   bufferB     pointer to buffer space for output
+   * @return     The function returns either
+   * <code>ARM_MATH_SIZE_MISMATCH</code> or <code>ARM_MATH_SUCCESS</code> based on the outcome of size checking.
+   *
+   * @details
+   *
+   * <b>Buffer size:</b>
+   *
+   * bufferA size: 2*ch_im_in*dim_kernel*dim_kernel
+   *
+   * bufferB size: 0
+   *
+   * <b>Input dimension constraints:</b>
+   *
+   * ch_im_in is multiple of 2 
+   *
+   * ch_im_out is multipe of 2
+   *
+   */
+
+arm_status
+arm_convolve_HWC_q15_fast(const q15_t * Im_in,
+                          const uint16_t dim_im_in,
+                          const uint16_t ch_im_in,
+                          const q15_t * wt,
+                          const uint16_t ch_im_out,
+                          const uint16_t dim_kernel,
+                          const uint16_t padding,
+                          const uint16_t stride,
+                          const q15_t * bias,
+                          const uint16_t bias_shift,
+                          const uint16_t out_shift,
+                          q15_t * Im_out, 
+                          const uint16_t dim_im_out, 
+                          q15_t * bufferA, 
+                          q7_t * bufferB)
+{
+
+#if defined (ARM_MATH_DSP)
+    int16_t   i_out_y, i_out_x, i_ker_y, i_ker_x;
+
+    q15_t    *pBuffer = bufferA;
+    q15_t    *im_buffer = bufferA;
+    q15_t    *pOut = Im_out;
+
+    if (ch_im_in % 2 != 0 || ch_im_out % 2 != 0)
+    {
+        /* check if the input dimension meets the constraints */
+        return ARM_MATH_SIZE_MISMATCH;
+    }
+
+    /* Run the following code for Cortex-M4 and Cortex-M7 */
+
+    /* This part implements the im2col function */
+    for (i_out_y = 0; i_out_y < dim_im_out; i_out_y++)
+    {
+        for (i_out_x = 0; i_out_x < dim_im_out; i_out_x++)
+        {
+            for (i_ker_y = i_out_y * stride - padding; i_ker_y < i_out_y * stride - padding + dim_kernel; i_ker_y++)
+            {
+                for (i_ker_x = i_out_x * stride - padding; i_ker_x < i_out_x * stride - padding + dim_kernel; i_ker_x++)
+                {
+                    if (i_ker_y < 0 || i_ker_y >= dim_im_in || i_ker_x < 0 || i_ker_x >= dim_im_in)
+                    {
+                        /* arm_fill_q15(0, pBuffer, ch_im_in); */
+                        memset(pBuffer, 0, sizeof(q15_t)*ch_im_in);
+                    } else
+                    {
+                        /* arm_copy_q15((q15_t *) Im_in + (i_ker_y * dim_im_in + i_ker_x) * ch_im_in, pBuffer, ch_im_in); */
+                        memcpy(pBuffer, (q15_t *) Im_in + (i_ker_y * dim_im_in + i_ker_x) * ch_im_in, sizeof(q15_t)*ch_im_in);
+                    }
+                    pBuffer += ch_im_in;
+                }
+            }
+
+            if (i_out_x & 0x1)
+            {
+                int       i;
+                /* initialize the matrix pointers for A */
+                const q15_t *pA = wt;
+
+                /* set up the second output pointers */
+                q15_t    *pOut2 = pOut + ch_im_out;
+
+                /* this loop over rows in A */
+                for (i = 0; i < ch_im_out; i += 2)
+                {
+                    /* setup pointers for B */
+                    q15_t    *pB = im_buffer;
+                    const q15_t *pB2 = pB + ch_im_in * dim_kernel * dim_kernel;
+
+                    /* aling the second pointer for A */
+                    const q15_t *pA2 = pA + ch_im_in * dim_kernel * dim_kernel;
+
+                    /* init the sum with bias */
+                    q31_t     sum =  ((q31_t)bias[i] << bias_shift) + NN_ROUND(out_shift);
+                    q31_t     sum2 = ((q31_t)bias[i] << bias_shift) + NN_ROUND(out_shift);
+                    q31_t     sum3 = ((q31_t)bias[i + 1] << bias_shift) + NN_ROUND(out_shift);
+                    q31_t     sum4 = ((q31_t)bias[i + 1] << bias_shift) + NN_ROUND(out_shift);
+
+                    uint16_t  colCnt = ch_im_in * dim_kernel * dim_kernel >> 1;
+                    /* accumulate over the vector */
+                    while (colCnt)
+                    {
+                        q31_t     inA1 = *__SIMD32(pA)++;
+                        q31_t     inB1 = *__SIMD32(pB)++;
+                        q31_t     inA2 = *__SIMD32(pA2)++;
+                        q31_t     inB2 = *__SIMD32(pB2)++;
+
+                        sum = __SMLAD(inA1, inB1, sum);
+                        sum2 = __SMLAD(inA1, inB2, sum2);
+                        sum3 = __SMLAD(inA2, inB1, sum3);
+                        sum4 = __SMLAD(inA2, inB2, sum4);
+
+                        colCnt--;
+                    }           /* while over colCnt */
+                    colCnt = ch_im_in * dim_kernel * dim_kernel & 0x1;
+                    while (colCnt)
+                    {
+                        q15_t     inA1 = *pA++;
+                        q15_t     inB1 = *pB++;
+                        q15_t     inA2 = *pA2++;
+                        q15_t     inB2 = *pB2++;
+
+                        sum += inA1 * inB1;
+                        sum2 += inA1 * inB2;
+                        sum3 += inA2 * inB1;
+                        sum4 += inA2 * inB2;
+                        colCnt--;
+                    }           /* while over colCnt */
+                    *pOut++ = (q15_t) __SSAT(sum >> out_shift, 16);
+                    *pOut++ = (q15_t) __SSAT(sum3 >> out_shift, 16);
+                    *pOut2++ = (q15_t) __SSAT(sum2 >> out_shift, 16);
+                    *pOut2++ = (q15_t) __SSAT(sum4 >> out_shift, 16);
+
+                    /* skip the row computed with A2 */
+                    pA += ch_im_in * dim_kernel * dim_kernel;
+                }               /* for over ch_im_out */
+
+                pOut += ch_im_out;
+                /* counter reset */
+                pBuffer = im_buffer;
+            }
+        }
+    }
+
+#else
+    /* Run the following code as reference implementation for Cortex-M0 and Cortex-M3 */
+    uint16_t  i, j, k, l, m, n;
+    int       conv_out;
+    signed char in_row, in_col;
+
+    if (ch_im_in % 2 != 0 || ch_im_out % 2 != 0)
+    {
+        /* check if the input dimension meets the constraints */
+        return ARM_MATH_SIZE_MISMATCH;
+    }
+
+    for (i = 0; i < ch_im_out; i++)
+    {
+        for (j = 0; j < dim_im_out; j++)
+        {
+            for (k = 0; k < dim_im_out; k++)
+            {
+                conv_out = ((q31_t)bias[i] << bias_shift) + NN_ROUND(out_shift);
+                for (m = 0; m < dim_kernel; m++)
+                {
+                    for (n = 0; n < dim_kernel; n++)
+                    {
+                        in_row = stride * j + m - padding;
+                        in_col = stride * k + n - padding;
+                        if (in_row >= 0 && in_col >= 0 && in_row < dim_im_in && in_col < dim_im_in)
+                        {
+                            for (l = 0; l < ch_im_in; l++)
+                            {
+                                conv_out +=
+                                    Im_in[(in_row * dim_im_in + in_col) * ch_im_in +
+                                          l] * wt[i * ch_im_in * dim_kernel * dim_kernel + (m * dim_kernel +
+                                                                                            n) * ch_im_in + l];
+                            }
+                        }
+                    }
+                }
+                Im_out[i + (j * dim_im_out + k) * ch_im_out] = (q15_t) __SSAT((conv_out >> out_shift), 16);
+            }
+        }
+    }
+
+#endif                          /* ARM_MATH_DSP */
+
+    /* Return to application */
+    return ARM_MATH_SUCCESS;
+}
+
+/**
+ * @} end of NNConv group
+ */
diff --git a/NN/Source/ConvolutionFunctions/arm_convolve_HWC_q15_fast_nonsquare.c b/NN/Source/ConvolutionFunctions/arm_convolve_HWC_q15_fast_nonsquare.c
new file mode 100644
index 0000000..0274202
--- /dev/null
+++ b/NN/Source/ConvolutionFunctions/arm_convolve_HWC_q15_fast_nonsquare.c
@@ -0,0 +1,265 @@
+/*
+ * Copyright (C) 2010-2018 Arm Limited or its affiliates. All rights reserved.
+ *
+ * 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.
+ */
+
+/* ----------------------------------------------------------------------
+ * Project:      CMSIS NN Library
+ * Title:        arm_convolve_HWC_q15_fast.c
+ * Description:  Fast Q15 version of convolution
+ *
+ * $Date:        24. May 2018
+ * $Revision:    V.1.0.0
+ *
+ * Target Processor:  Cortex-M cores
+ *
+ * -------------------------------------------------------------------- */
+
+#include "arm_math.h"
+#include "arm_nnfunctions.h"
+
+/**
+ *  @ingroup groupNN
+ */
+
+/**
+ * @addtogroup NNConv
+ * @{
+ */
+
+  /**
+   * @brief Fast Q15 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]       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 either
+   * <code>ARM_MATH_SIZE_MISMATCH</code> or <code>ARM_MATH_SUCCESS</code> based on the outcome of size checking.
+   *
+   * @details
+   *
+   * <b>Buffer size:</b>
+   *
+   * bufferA size: 2*ch_im_in*dim_kernel*dim_kernel
+   *
+   * bufferB size: 0
+   *
+   * <b>Input dimension constraints:</b>
+   *
+   * ch_im_in is multiple of 2 
+   *
+   * ch_im_out is multipe of 2
+   *
+   */
+
+arm_status
+arm_convolve_HWC_q15_fast_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)
+{
+
+#if defined (ARM_MATH_DSP)
+    int16_t   i_out_y, i_out_x, i_ker_y, i_ker_x;
+
+    q15_t    *pBuffer = bufferA;
+    q15_t    *im_buffer = bufferA;
+    q15_t    *pOut = Im_out;
+
+    if (ch_im_in % 2 != 0 || ch_im_out % 2 != 0)
+    {
+        /* check if the input dimension meets the constraints */
+        return ARM_MATH_SIZE_MISMATCH;
+    }
+
+    /* Run the following code for Cortex-M4 and Cortex-M7 */
+
+    /* 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++)
+        {
+            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_x)
+                    {
+                        /* arm_fill_q15(0, pBuffer, ch_im_in); */
+                        memset(pBuffer, 0, sizeof(q15_t)*ch_im_in);
+                    } else
+                    {
+                        /* arm_copy_q15((q15_t *) Im_in + (i_ker_y * dim_im_in_x + i_ker_x) * ch_im_in, pBuffer, ch_im_in); */
+                        memcpy(pBuffer, (q15_t *) Im_in + (i_ker_y * dim_im_in_x + i_ker_x) * ch_im_in, sizeof(q15_t)*ch_im_in);
+                    }
+                    pBuffer += ch_im_in;
+                }
+            }
+
+            if (i_out_x & 0x1)
+            {
+                int       i;
+                /* initialize the matrix pointers for A */
+                const q15_t *pA = wt;
+
+                /* set up the second output pointers */
+                q15_t    *pOut2 = pOut + ch_im_out;
+
+                /* this loop over rows in A */
+                for (i = 0; i < ch_im_out; i += 2)
+                {
+                    /* setup pointers for B */
+                    q15_t    *pB = im_buffer;
+                    const q15_t *pB2 = pB + ch_im_in * dim_kernel_y * dim_kernel_x;
+
+                    /* aling the second pointer for A */
+                    const q15_t *pA2 = pA + ch_im_in * dim_kernel_y * dim_kernel_x;
+
+                    /* init the sum with bias */
+                    q31_t     sum =  ((q31_t)bias[i] << bias_shift) + NN_ROUND(out_shift);
+                    q31_t     sum2 = ((q31_t)bias[i] << bias_shift) + NN_ROUND(out_shift);
+                    q31_t     sum3 = ((q31_t)bias[i + 1] << bias_shift) + NN_ROUND(out_shift);
+                    q31_t     sum4 = ((q31_t)bias[i + 1] << bias_shift) + NN_ROUND(out_shift);
+
+                    uint16_t  colCnt = ch_im_in * dim_kernel_y * dim_kernel_x >> 1;
+                    /* accumulate over the vector */
+                    while (colCnt)
+                    {
+                        q31_t     inA1 = *__SIMD32(pA)++;
+                        q31_t     inB1 = *__SIMD32(pB)++;
+                        q31_t     inA2 = *__SIMD32(pA2)++;
+                        q31_t     inB2 = *__SIMD32(pB2)++;
+
+                        sum = __SMLAD(inA1, inB1, sum);
+                        sum2 = __SMLAD(inA1, inB2, sum2);
+                        sum3 = __SMLAD(inA2, inB1, sum3);
+                        sum4 = __SMLAD(inA2, inB2, sum4);
+
+                        colCnt--;
+                    }           /* while over colCnt */
+                    colCnt = ch_im_in * dim_kernel_y * dim_kernel_x & 0x1;
+                    while (colCnt)
+                    {
+                        q15_t     inA1 = *pA++;
+                        q15_t     inB1 = *pB++;
+                        q15_t     inA2 = *pA2++;
+                        q15_t     inB2 = *pB2++;
+
+                        sum += inA1 * inB1;
+                        sum2 += inA1 * inB2;
+                        sum3 += inA2 * inB1;
+                        sum4 += inA2 * inB2;
+                        colCnt--;
+                    }           /* while over colCnt */
+                    *pOut++ = (q15_t) __SSAT(sum >> out_shift, 16);
+                    *pOut++ = (q15_t) __SSAT(sum3 >> out_shift, 16);
+                    *pOut2++ = (q15_t) __SSAT(sum2 >> out_shift, 16);
+                    *pOut2++ = (q15_t) __SSAT(sum4 >> out_shift, 16);
+
+                    /* skip the row computed with A2 */
+                    pA += ch_im_in * dim_kernel_y * dim_kernel_x;
+                }               /* for over ch_im_out */
+
+                pOut += ch_im_out;
+                /* counter reset */
+                pBuffer = im_buffer;
+            }
+        }
+    }
+
+#else
+    /* Run the following code as reference implementation for Cortex-M0 and Cortex-M3 */
+    uint16_t  i, j, k, l, m, n;
+    int       conv_out;
+    signed char in_row, in_col;
+
+    if (ch_im_in % 2 != 0 || ch_im_out % 2 != 0)
+    {
+        /* check if the input dimension meets the constraints */
+        return ARM_MATH_SIZE_MISMATCH;
+    }
+
+    for (i = 0; i < ch_im_out; i++)
+    {
+        for (j = 0; j < dim_im_out_y; j++)
+        {
+            for (k = 0; k < dim_im_out_x; k++)
+            {
+                conv_out = ((q31_t)bias[i] << bias_shift) + NN_ROUND(out_shift);
+                for (m = 0; m < dim_kernel_y; m++)
+                {
+                    for (n = 0; n < dim_kernel_x; n++)
+                    {
+                        in_row = stride_y * j + m - padding_y;
+                        in_col = stride_x * k + n - padding_x;
+                        if (in_row >= 0 && in_col >= 0 && in_row < dim_im_in_y && in_col < dim_im_in_x)
+                        {
+                            for (l = 0; l < ch_im_in; l++)
+                            {
+                                conv_out +=
+                                    Im_in[(in_row * dim_im_in_x + in_col) * ch_im_in +
+                                          l] * wt[i * ch_im_in * dim_kernel_x * dim_kernel_y + (m * dim_kernel_x +
+                                                                                            n) * ch_im_in + l];
+                            }
+                        }
+                    }
+                }
+                Im_out[i + (j * dim_im_out_x + k) * ch_im_out] = (q15_t) __SSAT((conv_out >> out_shift), 16);
+            }
+        }
+    }
+
+#endif                          /* ARM_MATH_DSP */
+
+    /* Return to application */
+    return ARM_MATH_SUCCESS;
+}
+
+/**
+ * @} end of NNConv group
+ */
diff --git a/NN/Source/ConvolutionFunctions/arm_convolve_HWC_q7_RGB.c b/NN/Source/ConvolutionFunctions/arm_convolve_HWC_q7_RGB.c
new file mode 100644
index 0000000..42bfb1f
--- /dev/null
+++ b/NN/Source/ConvolutionFunctions/arm_convolve_HWC_q7_RGB.c
@@ -0,0 +1,279 @@
+/*
+ * Copyright (C) 2010-2018 Arm Limited or its affiliates. All rights reserved.
+ *
+ * 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.
+ */
+
+/* ----------------------------------------------------------------------
+ * Project:      CMSIS NN Library
+ * Title:        arm_convolve_HWC_q7_RGB.c
+ * Description:  Q7 version of convolution for RGB image
+ *
+ * $Date:        17. January 2018
+ * $Revision:    V.1.0.0
+ *
+ * Target Processor:  Cortex-M cores
+ *
+ * -------------------------------------------------------------------- */
+#include "arm_math.h"
+#include "arm_nnfunctions.h"
+
+/**
+ *  @ingroup groupNN
+ */
+
+/**
+ * @addtogroup NNConv
+ * @{
+ */
+
+  /**
+   * @brief Q7 convolution function for RGB image
+   * @param[in]       Im_in       pointer to input tensor
+   * @param[in]       dim_im_in   input tensor dimention
+   * @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  filter kernel size
+   * @param[in]       padding     padding sizes
+   * @param[in]       stride      convolution stride
+   * @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  output tensor dimension
+   * @param[in,out]   bufferA     pointer to buffer space for input
+   * @param[in,out]   bufferB     pointer to buffer space for output
+   * @return     The function returns either
+   * <code>ARM_MATH_SIZE_MISMATCH</code> or <code>ARM_MATH_SUCCESS</code> based on the outcome of size checking.
+   *
+   * @details
+   *
+   * <b>Buffer size:</b>
+   *
+   * bufferA size: 2*ch_im_in*dim_kernel*dim_kernel
+   *
+   * bufferB size: 0
+   *
+   * <b>Input dimension constraints:</b>
+   *
+   * ch_im_in equals 3
+   *
+   * This kernel is written exclusively for convolution with ch_im_in
+   * equals 3. This applies on the first layer of CNNs which has input
+   * image with RGB format.
+   */
+
+arm_status
+arm_convolve_HWC_q7_RGB(const q7_t * Im_in,
+                        const uint16_t dim_im_in,
+                        const uint16_t ch_im_in,
+                        const q7_t * wt,
+                        const uint16_t ch_im_out,
+                        const uint16_t dim_kernel,
+                        const uint16_t padding,
+                        const uint16_t stride,
+                        const q7_t * bias,
+                        const uint16_t bias_shift,
+                        const uint16_t out_shift,
+                        q7_t * Im_out, const uint16_t dim_im_out, q15_t * bufferA, q7_t * bufferB)
+{
+
+#if defined (ARM_MATH_DSP)
+    /* Run the following code for Cortex-M4 and Cortex-M7 */
+    int16_t   i_out_y, i_out_x, i_ker_y, i_ker_x;
+
+    /*
+     *  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;
+
+    // check if number of input channels is 3
+    if (ch_im_in != 3)
+    {
+        return ARM_MATH_SIZE_MISMATCH;
+    }
+    // This part implements the im2col function
+    for (i_out_y = 0; i_out_y < dim_im_out; i_out_y++)
+    {
+        for (i_out_x = 0; i_out_x < dim_im_out; i_out_x++)
+        {
+            for (i_ker_y = i_out_y * stride - padding; i_ker_y < i_out_y * stride - padding + dim_kernel; i_ker_y++)
+            {
+                for (i_ker_x = i_out_x * stride - padding; i_ker_x < i_out_x * stride - padding + dim_kernel; i_ker_x++)
+                {
+                    if (i_ker_y < 0 || i_ker_y >= dim_im_in || i_ker_x < 0 || i_ker_x >= dim_im_in)
+                    {
+                        /* Equivalent to arm_fill_q15(0, pBuffer, ch_im_in) with assumption: ch_im_in = 3 */
+                        *__SIMD32(pBuffer) = 0x0;
+                        *(pBuffer + 2) = 0;
+                        pBuffer += 3;
+                    } else
+                    {
+                        /* 
+                         * Equivalent to:
+                         *  arm_q7_to_q15_no_shift( (q7_t*)Im_in+(i_ker_y*dim_im_in+i_ker_x)*3, pBuffer, 3);
+                         */
+
+                        const q7_t *pPixel = Im_in + (i_ker_y * dim_im_in + i_ker_x) * 3;
+                        q31_t     buf = *__SIMD32(pPixel);
+
+                        union arm_nnword top;
+                        union arm_nnword bottom;
+
+                        top.word = __SXTB16(buf);
+                        bottom.word = __SXTB16(__ROR(buf, 8));
+
+#ifndef ARM_MATH_BIG_ENDIAN
+                        /*
+                         *  little-endian, | omit | 3rd  | 2nd  | 1st  |
+                         *                MSB                         LSB
+                         *   top | 3rd | 1st |; bottom | omit | 2nd |
+                         *
+                         *  version 1, need to swap 2nd and 3rd weight
+                         * *__SIMD32(pBuffer) = top.word;
+                         * *(pBuffer+2) = bottom.half_words[0];
+                         *
+                         *  version 2, no weight shuffling required
+                         */
+                        *pBuffer++ = top.half_words[0];
+                        *__SIMD32(pBuffer) = __PKHBT(bottom.word, top.word, 0);
+#else
+                        /*
+                         *  big-endian,    | 1st  | 2nd  | 3rd  | omit | 
+                         *                MSB                         LSB
+                         *  top | 2nd | omit |; bottom | 1st | 3rd |
+                         * 
+                         *  version 1, need to swap 2nd and 3rd weight
+                         * *__SIMD32(pBuffer) = bottom.word;
+                         * *(pBuffer+2) = top.half_words[1];
+                         * 
+                         *  version 2, no weight shuffling required
+                         */
+                        *pBuffer++ = bottom.half_words[0];
+                        *__SIMD32(pBuffer) = __PKHTB(top.word, bottom.word, 0);
+#endif
+                        pBuffer += 2;
+                    }
+                }
+            }
+
+            if (pBuffer == bufferA + 2 * 3 * dim_kernel * dim_kernel)
+            {
+                pOut =
+                    arm_nn_mat_mult_kernel_q7_q15(wt, bufferA,
+                                                  ch_im_out,
+                                                  3 * dim_kernel * dim_kernel, 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++)
+        {
+            q31_t     sum = ((q31_t)bias[i] << bias_shift) + NN_ROUND(out_shift);
+            q15_t    *pB = bufferA;
+            /* basically each time it process 4 entries */
+            uint16_t  colCnt = 3 * dim_kernel * dim_kernel >> 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 = 3 * dim_kernel * dim_kernel & 0x3;
+            while (colCnt)
+            {
+                q7_t      inA1 = *pA++;
+                q15_t     inB1 = *pB++;
+                sum += inA1 * inB1;
+                colCnt--;
+            }
+            *pOut++ = (q7_t) __SSAT((sum >> out_shift), 8);
+        }
+    }
+#else
+    /* Run the following code as reference implementation for Cortex-M0 and Cortex-M3 */
+
+    uint16_t  i, j, k, l, m, n;
+    int       conv_out;
+    signed char in_row, in_col;
+
+    // check if number of input channels is 3
+    if (ch_im_in != 3)
+    {
+        return ARM_MATH_SIZE_MISMATCH;
+    }
+
+    for (i = 0; i < ch_im_out; i++)
+    {
+        for (j = 0; j < dim_im_out; j++)
+        {
+            for (k = 0; k < dim_im_out; k++)
+            {
+                conv_out = (bias[i] << bias_shift) + NN_ROUND(out_shift);
+                for (m = 0; m < dim_kernel; m++)
+                {
+                    for (n = 0; n < dim_kernel; n++)
+                    {
+                        /* if-for implementation */
+                        in_row = stride * j + m - padding;
+                        in_col = stride * k + n - padding;
+                        if (in_row >= 0 && in_col >= 0 && in_row < dim_im_in && in_col < dim_im_in)
+                        {
+                            for (l = 0; l < ch_im_in; l++)
+                            {
+                                conv_out +=
+                                    Im_in[(in_row * dim_im_in + in_col) * ch_im_in +
+                                          l] * wt[i * ch_im_in * dim_kernel * dim_kernel + (m * dim_kernel +
+                                                                                            n) * ch_im_in + l];
+                            }
+                        }
+                    }
+                }
+                Im_out[i + (j * dim_im_out + k) * ch_im_out] = (q7_t) __SSAT((conv_out >> out_shift), 8);
+            }
+        }
+    }
+
+#endif                          /* ARM_MATH_DSP */
+
+    /* Return to application */
+    return (ARM_MATH_SUCCESS);
+}
+
+/**
+ * @} end of NNConv group
+ */
diff --git a/NN/Source/ConvolutionFunctions/arm_convolve_HWC_q7_basic.c b/NN/Source/ConvolutionFunctions/arm_convolve_HWC_q7_basic.c
new file mode 100644
index 0000000..a926086
--- /dev/null
+++ b/NN/Source/ConvolutionFunctions/arm_convolve_HWC_q7_basic.c
@@ -0,0 +1,230 @@
+/*
+ * Copyright (C) 2010-2018 Arm Limited or its affiliates. All rights reserved.
+ *
+ * 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.
+ */
+
+/* ----------------------------------------------------------------------
+ * Project:      CMSIS NN Library
+ * Title:        arm_convolve_HWC_q7_basic.c
+ * Description:	 Q7 version of convolution
+ *
+ * $Date:        17. January 2018
+ * $Revision:    V.1.0.0
+ *
+ * Target Processor:  Cortex-M cores
+ *
+ * -------------------------------------------------------------------- */
+#include "arm_math.h"
+#include "arm_nnfunctions.h"
+
+/**
+ *  @ingroup groupNN
+ */
+
+/**
+ * @addtogroup NNConv
+ * @{
+ */
+
+  /**
+   * @brief Basic Q7 convolution function
+   * @param[in]       Im_in       pointer to input tensor
+   * @param[in]       dim_im_in   input tensor dimention
+   * @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  filter kernel size
+   * @param[in]       padding     padding sizes
+   * @param[in]       stride      convolution stride
+   * @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  output tensor dimension
+   * @param[in,out]   bufferA     pointer to buffer space for input 
+   * @param[in,out]   bufferB     pointer to buffer space for output
+   * @return     The function returns <code>ARM_MATH_SUCCESS</code> 
+   *
+   * @details
+   *
+   * <b>Buffer size:</b>
+   *
+   * bufferA size: 2*ch_im_in*dim_kernel*dim_kernel
+   *
+   * bufferB size: 0
+   *
+   * This basic version is designed to work for any input tensor and weight
+   * dimension. 
+   */
+
+arm_status
+arm_convolve_HWC_q7_basic(const q7_t * Im_in,
+                          const uint16_t dim_im_in,
+                          const uint16_t ch_im_in,
+                          const q7_t * wt,
+                          const uint16_t ch_im_out,
+                          const uint16_t dim_kernel,
+                          const uint16_t padding,
+                          const uint16_t stride,
+                          const q7_t * bias,
+                          const uint16_t bias_shift,
+                          const uint16_t out_shift,
+                          q7_t * Im_out, 
+                          const uint16_t dim_im_out, 
+                          q15_t * bufferA, 
+                          q7_t * bufferB)
+{
+
+#if defined (ARM_MATH_DSP)
+    /* Run the following code for Cortex-M4 and Cortex-M7 */
+
+    int16_t   i_out_y, i_out_x, i_ker_y, i_ker_x;
+
+    /* 
+     *  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; i_out_y++)
+    {
+        for (i_out_x = 0; i_out_x < dim_im_out; i_out_x++)
+        {
+            for (i_ker_y = i_out_y * stride - padding; i_ker_y < i_out_y * stride - padding + dim_kernel; i_ker_y++)
+            {
+                for (i_ker_x = i_out_x * stride - padding; i_ker_x < i_out_x * stride - padding + dim_kernel; i_ker_x++)
+                {
+                    if (i_ker_y < 0 || i_ker_y >= dim_im_in || i_ker_x < 0 || i_ker_x >= dim_im_in)
+                    {
+                        /* 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
+                    {
+                        /* Copying the pixel data to column */
+                        arm_q7_to_q15_no_shift((q7_t *)
+                                               Im_in + (i_ker_y * dim_im_in + i_ker_x) * ch_im_in, pBuffer, ch_im_in);
+                    }
+                    pBuffer += ch_im_in;
+                }
+            }
+
+            /* Computation is filed for every 2 columns */
+            if (pBuffer == bufferA + 2 * ch_im_in * dim_kernel * dim_kernel)
+            {
+                pOut =
+                    arm_nn_mat_mult_kernel_q7_q15(wt, bufferA,
+                                                  ch_im_out,
+                                                  ch_im_in *
+                                                  dim_kernel * dim_kernel, 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 = ch_im_in * dim_kernel * dim_kernel >> 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 = ch_im_in * dim_kernel * dim_kernel & 0x3;
+            while (colCnt)
+            {
+                q7_t      inA1 = *pA++;
+                q15_t     inB1 = *pB++;
+                sum += inA1 * inB1;
+                colCnt--;
+            }
+            *pOut++ = (q7_t) __SSAT((sum >> out_shift), 8);
+        }
+    }
+#else
+    /* Run the following code as reference implementation for Cortex-M0 and Cortex-M3 */
+
+    uint16_t  i, j, k, l, m, n;
+    int       conv_out;
+    signed char in_row, in_col;
+
+    for (i = 0; i < ch_im_out; i++)
+    {
+        for (j = 0; j < dim_im_out; j++)
+        {
+            for (k = 0; k < dim_im_out; k++)
+            {
+                conv_out = ((q31_t)bias[i] << bias_shift) + NN_ROUND(out_shift);
+                for (m = 0; m < dim_kernel; m++)
+                {
+                    for (n = 0; n < dim_kernel; n++)
+                    {
+                        // if-for implementation
+                        in_row = stride * j + m - padding;
+                        in_col = stride * k + n - padding;
+                        if (in_row >= 0 && in_col >= 0 && in_row < dim_im_in && in_col < dim_im_in)
+                        {
+                            for (l = 0; l < ch_im_in; l++)
+                            {
+                                conv_out +=
+                                    Im_in[(in_row * dim_im_in + in_col) * ch_im_in +
+                                          l] * wt[i * ch_im_in * dim_kernel * dim_kernel + (m * dim_kernel +
+                                                                                            n) * ch_im_in + l];
+                            }
+                        }
+                    }
+                }
+                Im_out[i + (j * dim_im_out + k) * ch_im_out] = (q7_t) __SSAT((conv_out >> out_shift), 8);
+            }
+        }
+    }
+
+#endif                          /* ARM_MATH_DSP */
+
+    /* Return to application */
+    return ARM_MATH_SUCCESS;
+}
+
+/**
+ * @} end of NNConv group
+ */
diff --git a/NN/Source/ConvolutionFunctions/arm_convolve_HWC_q7_basic_nonsquare.c b/NN/Source/ConvolutionFunctions/arm_convolve_HWC_q7_basic_nonsquare.c
new file mode 100644
index 0000000..b426b92
--- /dev/null
+++ b/NN/Source/ConvolutionFunctions/arm_convolve_HWC_q7_basic_nonsquare.c
@@ -0,0 +1,228 @@
+/*
+ * Copyright (C) 2010-2018 Arm Limited or its affiliates. All rights reserved.
+ *
+ * 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.
+ */
+
+/* ----------------------------------------------------------------------
+ * Project:      CMSIS NN Library
+ * Title:        arm_convolve_HWC_q7_basic.c
+ * Description:	 Q7 version of convolution
+ *
+ * $Date:        13. July 2018
+ * $Revision:    V.1.0.0
+ *
+ * Target Processor:  Cortex-M cores
+ *
+ * -------------------------------------------------------------------- */
+#include "arm_math.h"
+#include "arm_nnfunctions.h"
+
+/**
+ *  @ingroup groupNN
+ */
+
+/**
+ * @addtogroup NNConv
+ * @{
+ */
+
+  /**
+   * @brief Basic Q7 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]       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 <code>ARM_MATH_SUCCESS</code>
+   */
+
+arm_status arm_convolve_HWC_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)
+{
+
+#if defined (ARM_MATH_DSP)
+    /* Run the following code for Cortex-M4 and Cortex-M7 */
+
+    int16_t   i_out_y, i_out_x, i_ker_y, i_ker_x;
+
+    /* 
+     *  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++)
+        {
+            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_x)
+                    {
+                        /* 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
+                    {
+                        /* Copying the pixel data to column */
+                        arm_q7_to_q15_no_shift((q7_t *)
+                                               Im_in + (i_ker_y * dim_im_in_x + i_ker_x) * ch_im_in, pBuffer, ch_im_in);
+                    }
+                    pBuffer += ch_im_in;
+                }
+            }
+
+            /* Computation is filed for every 2 columns */
+            if (pBuffer == bufferA + 2 * ch_im_in * dim_kernel_y * dim_kernel_x)
+            {
+                pOut =
+                    arm_nn_mat_mult_kernel_q7_q15(wt, bufferA,
+                                                  ch_im_out,
+                                                  ch_im_in *
+                                                  dim_kernel_y * dim_kernel_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 = ch_im_in * dim_kernel_y * dim_kernel_x >> 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 = ch_im_in * dim_kernel_y * dim_kernel_x & 0x3;
+            while (colCnt)
+            {
+                q7_t      inA1 = *pA++;
+                q15_t     inB1 = *pB++;
+                sum += inA1 * inB1;
+                colCnt--;
+            }
+            *pOut++ = (q7_t) __SSAT((sum >> out_shift), 8);
+        }
+    }
+#else
+    /* Run the following code as reference implementation for Cortex-M0 and Cortex-M3 */
+
+    uint16_t  i, j, k, l, m, n;
+    int       conv_out;
+    signed char in_row, in_col;
+
+    for (i = 0; i < ch_im_out; i++)
+    {
+        for (j = 0; j < dim_im_out_y; j++)
+        {
+            for (k = 0; k < dim_im_out_x; k++)
+            {
+                conv_out = ((q31_t)bias[i] << bias_shift) + NN_ROUND(out_shift);
+                for (m = 0; m < dim_kernel_y; m++)
+                {
+                    for (n = 0; n < dim_kernel_x; n++)
+                    {
+                        // if-for implementation
+                        in_row = stride_y * j + m - padding_y;
+                        in_col = stride_x * k + n - padding_x;
+                        if (in_row >= 0 && in_col >= 0 && in_row < dim_im_in_y && in_col < dim_im_in_x)
+                        {
+                            for (l = 0; l < ch_im_in; l++)
+                            {
+                                conv_out +=
+                                    Im_in[(in_row * dim_im_in_x + in_col) * ch_im_in + l] * 
+                                         wt[i * ch_im_in * dim_kernel_y * dim_kernel_x + 
+                                         (m * dim_kernel_x + n) * ch_im_in + l];
+                            }
+                        }
+                    }
+                }
+                Im_out[i + (j * dim_im_out_x + k) * ch_im_out] = (q7_t) __SSAT((conv_out >> out_shift), 8);
+            }
+        }
+    }
+
+#endif                          /* ARM_MATH_DSP */
+
+    /* Return to application */
+    return ARM_MATH_SUCCESS;
+}
+
+/**
+ * @} end of NNConv group
+ */
diff --git a/NN/Source/ConvolutionFunctions/arm_convolve_HWC_q7_fast.c b/NN/Source/ConvolutionFunctions/arm_convolve_HWC_q7_fast.c
new file mode 100644
index 0000000..7b59d79
--- /dev/null
+++ b/NN/Source/ConvolutionFunctions/arm_convolve_HWC_q7_fast.c
@@ -0,0 +1,408 @@
+/*
+ * Copyright (C) 2010-2018 Arm Limited or its affiliates. All rights reserved.
+ *
+ * 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.
+ */
+
+/* ----------------------------------------------------------------------
+ * Project:      CMSIS NN Library
+ * Title:        arm_convolve_HWC_q7_fast.c
+ * Description:  Fast Q7 version of convolution
+ *
+ * $Date:        17. January 2018
+ * $Revision:    V.1.0.0
+ *
+ * Target Processor:  Cortex-M cores
+ *
+ * -------------------------------------------------------------------- */
+
+#include "arm_math.h"
+#include "arm_nnfunctions.h"
+
+/**
+ *  @ingroup groupNN
+ */
+
+/**
+ * @addtogroup NNConv
+ * @{
+ */
+
+  /**
+   * @brief Fast Q7 convolution function
+   * @param[in]       Im_in       pointer to input tensor
+   * @param[in]       dim_im_in   input tensor dimention
+   * @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  filter kernel size
+   * @param[in]       padding     padding sizes
+   * @param[in]       stride      convolution stride
+   * @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  output tensor dimension
+   * @param[in,out]   bufferA     pointer to buffer space for input 
+   * @param[in,out]   bufferB     pointer to buffer space for output
+   * @return     The function returns either
+   * <code>ARM_MATH_SIZE_MISMATCH</code> or <code>ARM_MATH_SUCCESS</code> based on the outcome of size checking.
+   *
+   * @details
+   *
+   * <b>Buffer size:</b>
+   *
+   * bufferA size: 2*ch_im_in*dim_kernel*dim_kernel
+   *
+   * bufferB size: 0
+   *
+   * <b>Input dimension constraints:</b>
+   *
+   * ch_im_in is multiple of 4    ( because of the SIMD32 read and swap )
+   *
+   * ch_im_out is multipe of 2    ( bacause 2x2 mat_mult kernel )
+   *
+   * The im2col converts the Q7 tensor input into Q15 column, which is stored in
+   * bufferA. There is reordering happenning during this im2col process with
+   * arm_q7_to_q15_reordered_no_shift. For every four elements, the second and
+   * third elements are swapped. 
+   *
+   * The computation kernel arm_nn_mat_mult_kernel_q7_q15_reordered does the
+   * GEMM computation with the reordered columns.
+   *
+   * To speed-up the determination of the padding condition, we split the
+   * computation into 3x3 parts, i.e., {top, mid, bottom} X {left, mid, right}.
+   * This reduces the total number of boundary condition checks and improves
+   * the data copying performance.
+   */
+
+arm_status
+arm_convolve_HWC_q7_fast(const q7_t * Im_in,
+                         const uint16_t dim_im_in,
+                         const uint16_t ch_im_in,
+                         const q7_t * wt,
+                         const uint16_t ch_im_out,
+                         const uint16_t dim_kernel,
+                         const uint16_t padding,
+                         const uint16_t stride,
+                         const q7_t * bias,
+                         const uint16_t bias_shift,
+                         const uint16_t out_shift,
+                         q7_t * Im_out, 
+                         const uint16_t dim_im_out, 
+                         q15_t * bufferA, 
+                         q7_t * bufferB)
+{
+
+#if defined (ARM_MATH_DSP)
+    /* Run the following code for Cortex-M4 and Cortex-M7 */
+
+    int16_t   i_out_y, i_out_x, i_ker_y, i_ker_x;
+
+    /*
+     *  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;
+
+    if (ch_im_in % 4 != 0 || ch_im_out % 2 != 0)
+    {
+        /* check if the input dimension meets the constraints */
+        return ARM_MATH_SIZE_MISMATCH;
+    }
+
+    /*
+     *  Here we split the entire matrix into three regions depending on the padding situation
+     *    Top: i_out_y from 0 to padding - 1
+     * Middle: i_out_y from padding to dim_im_out-padding-1
+     * Bottom: i_out_y from dim_im_out-padding to dim_im_out-1
+     */
+
+    /* top part */
+    for (i_out_y = 0; i_out_y < padding; i_out_y++)
+    {
+        for (i_out_x = 0; i_out_x < dim_im_out; i_out_x++)
+        {
+            /* This part implements the im2col function */
+            for (i_ker_y = i_out_y * stride - padding; i_ker_y < i_out_y * stride - padding + dim_kernel; i_ker_y++)
+            {
+                for (i_ker_x = i_out_x * stride - padding; i_ker_x < i_out_x * stride - padding + dim_kernel; i_ker_x++)
+                {
+                    if (i_ker_y < 0 || i_ker_y >= dim_im_in || i_ker_x < 0 || i_ker_x >= dim_im_in)
+                    {
+                        /* arm_fill_q15(0, pBuffer, ch_im_in); */
+                        memset(pBuffer, 0, sizeof(q15_t)*ch_im_in);
+                    } else
+                    {
+                        arm_q7_to_q15_reordered_no_shift
+                            ((q7_t *) Im_in + (i_ker_y * dim_im_in + i_ker_x) * ch_im_in, pBuffer, ch_im_in);
+                    }
+                    pBuffer += ch_im_in;
+                }
+            }
+
+            if (pBuffer == bufferA + 2 * ch_im_in * dim_kernel * dim_kernel)
+            {
+                pOut =
+                    arm_nn_mat_mult_kernel_q7_q15_reordered(wt,
+                                                            bufferA,
+                                                            ch_im_out,
+                                                            ch_im_in
+                                                            *
+                                                            dim_kernel * dim_kernel, bias_shift, out_shift, bias, pOut);
+                /* counter reset */
+                pBuffer = bufferA;
+            }
+        }
+    }
+
+    /* middle part, here we also divide the x into left, mid and right */
+    for (; i_out_y < dim_im_out - padding; i_out_y++)
+    {
+
+        /* left part */
+        for (i_out_x = 0; i_out_x < padding; i_out_x++)
+        {
+            /* This part implements the im2col function */
+            for (i_ker_y = i_out_y * stride - padding; i_ker_y < i_out_y * stride - padding + dim_kernel; i_ker_y++)
+            {
+                for (i_ker_x = i_out_x * stride - padding; i_ker_x < i_out_x * stride - padding + dim_kernel; i_ker_x++)
+                {
+                    if (i_ker_x < 0 || i_ker_x >= dim_im_in)
+                    {
+                        /* arm_fill_q15(0, pBuffer, ch_im_in); */
+                        memset(pBuffer, 0, sizeof(q15_t)*ch_im_in);
+                    } else
+                    {
+                        arm_q7_to_q15_reordered_no_shift
+                            ((q7_t *) Im_in + (i_ker_y * dim_im_in + i_ker_x) * ch_im_in, pBuffer, ch_im_in);
+                    }
+                    pBuffer += ch_im_in;
+                }
+            }
+
+            if (pBuffer == bufferA + 2 * ch_im_in * dim_kernel * dim_kernel)
+            {
+                pOut =
+                    arm_nn_mat_mult_kernel_q7_q15_reordered(wt,
+                                                            bufferA,
+                                                            ch_im_out,
+                                                            ch_im_in
+                                                            *
+                                                            dim_kernel * dim_kernel, bias_shift, out_shift, bias, pOut);
+                /* counter reset */
+                pBuffer = bufferA;
+            }
+        }
+
+        /* mid part */
+        for (; i_out_x < dim_im_out - padding; i_out_x++)
+        {
+            /* This part implements the im2col function */
+            for (i_ker_y = i_out_y * stride - padding; i_ker_y < i_out_y * stride - padding + dim_kernel; i_ker_y++)
+            {
+                arm_q7_to_q15_reordered_no_shift((q7_t *) Im_in
+                                                 +
+                                                 (i_ker_y *
+                                                  dim_im_in +
+                                                  i_out_x *
+                                                  stride - padding) * ch_im_in, pBuffer, ch_im_in * dim_kernel);
+                pBuffer += ch_im_in * dim_kernel;
+            }
+
+            if (pBuffer == bufferA + 2 * ch_im_in * dim_kernel * dim_kernel)
+            {
+                pOut =
+                    arm_nn_mat_mult_kernel_q7_q15_reordered(wt,
+                                                            bufferA,
+                                                            ch_im_out,
+                                                            ch_im_in
+                                                            *
+                                                            dim_kernel * dim_kernel, bias_shift, out_shift, bias, pOut);
+                /* counter reset */
+                pBuffer = bufferA;
+            }
+        }
+
+        /* right part */
+        for (; i_out_x < dim_im_out; i_out_x++)
+        {
+            /* This part implements the im2col function */
+            for (i_ker_y = i_out_y * stride - padding; i_ker_y < i_out_y * stride - padding + dim_kernel; i_ker_y++)
+            {
+                for (i_ker_x = i_out_x * stride - padding; i_ker_x < i_out_x * stride - padding + dim_kernel; i_ker_x++)
+                {
+                    if (i_ker_x < 0 || i_ker_x >= dim_im_in)
+                    {
+                        /* arm_fill_q15(0, pBuffer, ch_im_in); */
+                        memset(pBuffer, 0, sizeof(q15_t)*ch_im_in);
+                    } else
+                    {
+                        arm_q7_to_q15_reordered_no_shift
+                            ((q7_t *) Im_in + (i_ker_y * dim_im_in + i_ker_x) * ch_im_in, pBuffer, ch_im_in);
+                    }
+                    pBuffer += ch_im_in;
+                }
+            }
+
+            if (pBuffer == bufferA + 2 * ch_im_in * dim_kernel * dim_kernel)
+            {
+                pOut =
+                    arm_nn_mat_mult_kernel_q7_q15_reordered(wt,
+                                                            bufferA,
+                                                            ch_im_out,
+                                                            ch_im_in
+                                                            *
+                                                            dim_kernel * dim_kernel, bias_shift, out_shift, bias, pOut);
+                /* counter reset */
+                pBuffer = bufferA;
+            }
+        }
+    }
+
+    for (; i_out_y < dim_im_out; i_out_y++)
+    {
+        for (i_out_x = 0; i_out_x < dim_im_out; i_out_x++)
+        {
+            /* This part implements the im2col function */
+            for (i_ker_y = i_out_y * stride - padding; i_ker_y < i_out_y * stride - padding + dim_kernel; i_ker_y++)
+            {
+                for (i_ker_x = i_out_x * stride - padding; i_ker_x < i_out_x * stride - padding + dim_kernel; i_ker_x++)
+                {
+                    if (i_ker_y < 0 || i_ker_y >= dim_im_in || i_ker_x < 0 || i_ker_x >= dim_im_in)
+                    {
+                        /* arm_fill_q15(0, pBuffer, ch_im_in); */
+                        memset(pBuffer, 0, sizeof(q15_t)*ch_im_in);
+                    } else
+                    {
+                        arm_q7_to_q15_reordered_no_shift
+                            ((q7_t *) Im_in + (i_ker_y * dim_im_in + i_ker_x) * ch_im_in, pBuffer, ch_im_in);
+                    }
+                    pBuffer += ch_im_in;
+                }
+            }
+
+            if (pBuffer == bufferA + 2 * ch_im_in * dim_kernel * dim_kernel)
+            {
+                pOut =
+                    arm_nn_mat_mult_kernel_q7_q15_reordered(wt,
+                                                            bufferA,
+                                                            ch_im_out,
+                                                            ch_im_in
+                                                            *
+                                                            dim_kernel * dim_kernel, bias_shift, out_shift, bias, pOut);
+                /* counter reset */
+                pBuffer = bufferA;
+            }
+        }
+    }
+
+    /* check if there is left-over for compute */
+    if (pBuffer != bufferA)
+    {
+        const q7_t *pA = wt;
+        int       i;
+
+        for (i = 0; i < ch_im_out; i++)
+        {
+            q31_t     sum = ((q31_t)bias[i] << bias_shift) + NN_ROUND(out_shift);
+            q15_t    *pB = bufferA;
+            /* each time it process 4 entries */
+            uint16_t  colCnt = ch_im_in * dim_kernel * dim_kernel >> 2;
+
+            while (colCnt)
+            {
+
+                q31_t     inA1, inA2;
+                q31_t     inB1, inB2;
+
+                pA = (q7_t *) read_and_pad_reordered((void *)pA, &inA1, &inA2);
+
+                inB1 = *__SIMD32(pB)++;
+                sum = __SMLAD(inA1, inB1, sum);
+                inB2 = *__SIMD32(pB)++;
+                sum = __SMLAD(inA2, inB2, sum);
+
+                colCnt--;
+            }
+            colCnt = ch_im_in * dim_kernel * dim_kernel & 0x3;
+            while (colCnt)
+            {
+                q7_t      inA1 = *pA++;
+                q15_t     inB1 = *pB++;
+                sum += inA1 * inB1;
+                colCnt--;
+            }
+            *pOut = (q7_t) __SSAT((sum >> out_shift), 8);
+            pOut++;
+
+        }
+
+    }
+#else
+    /* Run the following code as reference implementation for Cortex-M0 and Cortex-M3 */
+
+    uint16_t  i, j, k, l, m, n;
+    int       conv_out;
+    signed char in_row, in_col;
+
+    if (ch_im_in % 4 != 0 || ch_im_out % 2 != 0)
+    {
+        /* check if the input dimension meets the constraints */
+        return ARM_MATH_SIZE_MISMATCH;
+    }
+
+    for (i = 0; i < ch_im_out; i++)
+    {
+        for (j = 0; j < dim_im_out; j++)
+        {
+            for (k = 0; k < dim_im_out; k++)
+            {
+                conv_out = (bias[i] << bias_shift) + NN_ROUND(out_shift);
+                for (m = 0; m < dim_kernel; m++)
+                {
+                    for (n = 0; n < dim_kernel; n++)
+                    {
+                        // if-for implementation
+                        in_row = stride * j + m - padding;
+                        in_col = stride * k + n - padding;
+                        if (in_row >= 0 && in_col >= 0 && in_row < dim_im_in && in_col < dim_im_in)
+                        {
+                            for (l = 0; l < ch_im_in; l++)
+                            {
+                                conv_out +=
+                                    Im_in[(in_row * dim_im_in + in_col) * ch_im_in +
+                                          l] * wt[i * ch_im_in * dim_kernel * dim_kernel + (m * dim_kernel +
+                                                                                            n) * ch_im_in + l];
+                            }
+                        }
+                    }
+                }
+                Im_out[i + (j * dim_im_out + k) * ch_im_out] = (q7_t) __SSAT((conv_out >> out_shift), 8);
+            }
+        }
+    }
+
+#endif                          /* ARM_MATH_DSP */
+
+    /* Return to application */
+    return ARM_MATH_SUCCESS;
+}
+
+/**
+ * @} end of NNConv group
+ */
diff --git a/NN/Source/ConvolutionFunctions/arm_convolve_HWC_q7_fast_nonsquare.c b/NN/Source/ConvolutionFunctions/arm_convolve_HWC_q7_fast_nonsquare.c
new file mode 100644
index 0000000..f2aa4a2
--- /dev/null
+++ b/NN/Source/ConvolutionFunctions/arm_convolve_HWC_q7_fast_nonsquare.c
@@ -0,0 +1,379 @@
+/*
+ * Copyright (C) 2010-2018 Arm Limited or its affiliates. All rights reserved.
+ *
+ * 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.
+ */
+
+/* ----------------------------------------------------------------------
+ * Project:      CMSIS NN Library
+ * Title:        arm_convolve_HWC_q7_fast_nonsquare.c
+ * Description:  Fast Q7 version of convolution (non-sqaure shape)
+ *
+ * $Date:        17. January 2018
+ * $Revision:    V.1.0.0
+ *
+ * Target Processor:  Cortex-M cores
+ *
+ * -------------------------------------------------------------------- */
+
+#include "arm_math.h"
+#include "arm_nnfunctions.h"
+
+/**
+ *  @ingroup groupNN
+ */
+
+/**
+ * @addtogroup NNConv
+ * @{
+ */
+
+/**
+ * @brief Fast Q7 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]       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 either
+ * <code>ARM_MATH_SIZE_MISMATCH</code> or <code>ARM_MATH_SUCCESS</code> based on the outcome of size checking.
+ *
+ * This function is the version with full list of optimization tricks, but with
+ * some contraints:
+ *   ch_im_in is multiple of 4
+ *   ch_im_out is multiple of 2
+ */
+
+arm_status arm_convolve_HWC_q7_fast_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)
+{
+
+#if defined (ARM_MATH_DSP)
+    /* Run the following code for Cortex-M4 and Cortex-M7 */
+
+    int16_t   i_out_y, i_out_x, i_ker_y, i_ker_x;
+
+    /* -----------------------
+     *  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;
+
+    if (ch_im_in % 4 != 0 || ch_im_out % 2 != 0)
+    {
+        /* check if the input dimension meets the constraints */
+        return ARM_MATH_SIZE_MISMATCH;
+    }
+
+    /*
+     *  Here we split the entire matrix into three regions depending on the padding situation
+     *    Top: i_out_y from 0 to padding - 1
+     * Middle: i_out_y from padding to dim_im_out-padding-1
+     * Bottom: i_out_y from dim_im_out-padding to dim_im_out-1
+     */
+
+    /* top part */
+    for (i_out_y = 0; i_out_y < padding_y; i_out_y++)
+    {
+        for (i_out_x = 0; i_out_x < dim_im_out_x; i_out_x++)
+        {
+            /* This part implements the im2col function */
+            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_x)
+                    {
+                        /* arm_fill_q15(0, pBuffer, ch_im_in); */
+                        memset(pBuffer, 0, sizeof(q15_t)*ch_im_in);
+                    } else
+                    {
+                        arm_q7_to_q15_reordered_no_shift((q7_t *) Im_in + (i_ker_y * dim_im_in_x + i_ker_x) * ch_im_in,
+                                                         pBuffer, ch_im_in);
+                    }
+                    pBuffer += ch_im_in;
+                }
+            }
+
+            if (pBuffer == bufferA + 2 * ch_im_in * dim_kernel_x * dim_kernel_y)
+            {
+                pOut =
+                    arm_nn_mat_mult_kernel_q7_q15_reordered(wt, bufferA, ch_im_out, ch_im_in * dim_kernel_x * dim_kernel_y,
+                                                  bias_shift, out_shift, bias, pOut);
+                /* counter reset */
+                pBuffer = bufferA;
+            }
+        }
+    }
+
+    /* middle part, here we also divide the x into left, mid and right */
+    for (; i_out_y < dim_im_out_y - padding_y; i_out_y++)
+    {
+
+        /* left part */
+        for (i_out_x = 0; i_out_x < padding_x; i_out_x++)
+        {
+            /* This part implements the im2col function */
+            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_x < 0 || i_ker_x >= dim_im_in_x)
+                    {
+                        /* arm_fill_q15(0, pBuffer, ch_im_in); */
+                        memset(pBuffer, 0, sizeof(q15_t)*ch_im_in);
+                    } else
+                    {
+                        arm_q7_to_q15_reordered_no_shift((q7_t *) Im_in + (i_ker_y * dim_im_in_x + i_ker_x) * ch_im_in,
+                                                         pBuffer, ch_im_in);
+                    }
+                    pBuffer += ch_im_in;
+                }
+            }
+
+            if (pBuffer == bufferA + 2 * ch_im_in * dim_kernel_x * dim_kernel_y)
+            {
+                pOut =
+                    arm_nn_mat_mult_kernel_q7_q15_reordered(wt, bufferA, ch_im_out, ch_im_in * dim_kernel_x * dim_kernel_y,
+                                                  bias_shift, out_shift, bias, pOut);
+                /* counter reset */
+                pBuffer = bufferA;
+            }
+        }
+
+        /* mid part */
+        for (; i_out_x < dim_im_out_x - padding_x; i_out_x++)
+        {
+            /* This part implements the im2col function */
+            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++)
+            {
+                arm_q7_to_q15_reordered_no_shift((q7_t *) Im_in +
+                                                 (i_ker_y * dim_im_in_x + i_out_x * stride_x - padding_x) * ch_im_in,
+                                                 pBuffer, ch_im_in * dim_kernel_x);
+                pBuffer += ch_im_in * dim_kernel_x;
+            }
+
+            if (pBuffer == bufferA + 2 * ch_im_in * dim_kernel_x * dim_kernel_y)
+            {
+                pOut =
+                    arm_nn_mat_mult_kernel_q7_q15_reordered(wt, bufferA, ch_im_out, ch_im_in * dim_kernel_x * dim_kernel_y,
+                                                  bias_shift, out_shift, bias, pOut);
+                /* counter reset */
+                pBuffer = bufferA;
+            }
+        }
+
+        /* right part */
+        for (; i_out_x < dim_im_out_x; i_out_x++)
+        {
+            /* This part implements the im2col function */
+            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_x < 0 || i_ker_x >= dim_im_in_x)
+                    {
+                        /* arm_fill_q15(0, pBuffer, ch_im_in); */
+                        memset(pBuffer, 0, sizeof(q15_t)*ch_im_in);
+                    } else
+                    {
+                        arm_q7_to_q15_reordered_no_shift((q7_t *) Im_in + (i_ker_y * dim_im_in_x + i_ker_x) * ch_im_in,
+                                                         pBuffer, ch_im_in);
+                    }
+                    pBuffer += ch_im_in;
+                }
+            }
+
+            if (pBuffer == bufferA + 2 * ch_im_in * dim_kernel_x * dim_kernel_y)
+            {
+                pOut =
+                    arm_nn_mat_mult_kernel_q7_q15_reordered(wt, bufferA, ch_im_out, ch_im_in * dim_kernel_x * dim_kernel_y,
+                                                  bias_shift, out_shift, bias, pOut);
+                /* counter reset */
+                pBuffer = bufferA;
+            }
+        }
+    }
+
+    for (; 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++)
+        {
+            /* This part implements the im2col function */
+            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_x)
+                    {
+                        /* arm_fill_q15(0, pBuffer, ch_im_in); */
+                        memset(pBuffer, 0, sizeof(q15_t)*ch_im_in);
+                    } else
+                    {
+                        arm_q7_to_q15_reordered_no_shift((q7_t *) Im_in + (i_ker_y * dim_im_in_x + i_ker_x) * ch_im_in,
+                                                         pBuffer, ch_im_in);
+                    }
+                    pBuffer += ch_im_in;
+                }
+            }
+
+            if (pBuffer == bufferA + 2 * ch_im_in * dim_kernel_x * dim_kernel_y)
+            {
+                pOut =
+                    arm_nn_mat_mult_kernel_q7_q15_reordered(wt, bufferA, ch_im_out, ch_im_in * dim_kernel_x * dim_kernel_y,
+                                                  bias_shift, out_shift, bias, pOut);
+                /* counter reset */
+                pBuffer = bufferA;
+            }
+        }
+    }
+
+    /* check if there is left-over for compute */
+    if (pBuffer != bufferA)
+    {
+        const q7_t *pA = wt;
+        int       i;
+        for (i = 0; i < ch_im_out; i++)
+        {
+            q31_t     sum = ((q31_t)(bias[i]) << bias_shift) + NN_ROUND(out_shift);
+            q15_t    *pB = bufferA;
+            /* basically each time it process 4 entries */
+            uint16_t  colCnt = ch_im_in * dim_kernel_x * dim_kernel_y >> 2;
+
+            while (colCnt)
+            {
+
+                q31_t     inA1, inA2;
+                q31_t     inB1, inB2;
+
+                pA = (const q7_t *)read_and_pad_reordered((void *)pA, &inA1, &inA2);
+
+                inB1 = *__SIMD32(pB)++;
+                sum = __SMLAD(inA1, inB1, sum);
+                inB2 = *__SIMD32(pB)++;
+                sum = __SMLAD(inA2, inB2, sum);
+
+                colCnt--;
+            }
+            colCnt = (ch_im_in * dim_kernel_y * dim_kernel_x) & 0x3;
+            while (colCnt)
+            {
+                q7_t      inA1 = *pA++;
+                q15_t     inB1 = *pB++;
+                sum += inA1 * inB1;
+                colCnt--;
+            }
+            *pOut = (q7_t) __SSAT((sum >> out_shift), 8);
+            pOut++;
+
+        }
+
+    }
+
+#else
+    /* Run the following code as reference implementation for Cortex-M0 and Cortex-M3 */
+    int       i, j, k, l, m, n;
+    int       conv_out;
+    int       in_row, in_col;
+
+    if (ch_im_in % 4 != 0 || ch_im_out % 2 != 0)
+    {
+        /* check if the input dimension meets the constraints */
+        return ARM_MATH_SIZE_MISMATCH;
+    }
+
+    for (i = 0; i < ch_im_out; i++)
+    {
+        for (j = 0; j < dim_im_out_y; j++)
+        {
+            for (k = 0; k < dim_im_out_x; k++)
+            {
+                conv_out = ((q31_t)(bias[i]) << bias_shift) + NN_ROUND(out_shift);
+                for (m = 0; m < dim_kernel_y; m++)
+                {
+                    for (n = 0; n < dim_kernel_x; n++)
+                    {
+                        /* if-for implementation */
+                        in_row = stride_y * j + m - padding_y;
+                        in_col = stride_x * k + n - padding_x;
+                        if (in_row >= 0 && in_col >= 0 && in_row < dim_im_in_y && in_col < dim_im_in_x)
+                        {
+                            for (l = 0; l < ch_im_in; l++)
+                            {
+                                conv_out += Im_in[(in_row * dim_im_in_x + in_col) * ch_im_in + l] *
+                                    wt[i * ch_im_in * dim_kernel_y * dim_kernel_x + (m * dim_kernel_x + n) * ch_im_in + l];      
+                            }
+                        }
+                    }
+                }
+                Im_out[i + (j * dim_im_out_x + k) * ch_im_out] = (q7_t) __SSAT((conv_out >> out_shift), 8);
+            }
+        }
+    }
+
+
+#endif                          /* ARM_MATH_DSP */
+
+    /* Return to application */
+    return ARM_MATH_SUCCESS;
+}
+
+/**
+ * @} end of NNConv group
+ */
diff --git a/NN/Source/ConvolutionFunctions/arm_depthwise_separable_conv_HWC_q7.c b/NN/Source/ConvolutionFunctions/arm_depthwise_separable_conv_HWC_q7.c
new file mode 100644
index 0000000..68ebeb8
--- /dev/null
+++ b/NN/Source/ConvolutionFunctions/arm_depthwise_separable_conv_HWC_q7.c
@@ -0,0 +1,418 @@
+/*
+ * Copyright (C) 2010-2018 Arm Limited or its affiliates. All rights reserved.
+ *
+ * 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.
+ */
+
+/* ----------------------------------------------------------------------
+ * Project:      CMSIS NN Library
+ * Title:        arm_depthwise_separable_conv_HWC_q7.c
+ * Description:  Q7 depthwise separable convolution function
+ *
+ * $Date:        17. January 2018
+ * $Revision:    V.1.0.0
+ *
+ * Target Processor:  Cortex-M cores
+ *
+ * -------------------------------------------------------------------- */
+
+#include "arm_math.h"
+#include "arm_nnfunctions.h"
+
+/**
+ *  @ingroup groupNN
+ */
+
+/**
+ * @addtogroup NNConv
+ * @{
+ */
+
+/**
+ * @brief Q7 depthwise separable convolution function
+ * @param[in]       Im_in       pointer to input tensor
+ * @param[in]       dim_im_in   input tensor dimention
+ * @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  filter kernel size
+ * @param[in]       padding     padding sizes
+ * @param[in]       stride      convolution stride
+ * @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  output tensor dimension
+ * @param[in,out]   bufferA     pointer to buffer space for input
+ * @param[in,out]   bufferB     pointer to buffer space for output
+ * @return     The function returns either
+ * <code>ARM_MATH_SIZE_MISMATCH</code> or <code>ARM_MATH_SUCCESS</code> based on the outcome of size checking.
+ *
+ * @details
+ *
+ * <b>Buffer size:</b>
+ *
+ * bufferA size: 2*ch_im_in*dim_kernel*dim_kernel
+ *
+ * bufferB size: 0
+ *
+ * <b>Input dimension constraints:</b>
+ *
+ * ch_im_in equals ch_im_out
+ *
+ * Implementation:
+ * There are 3 nested loop here:
+ * Inner loop: calculate each output value with MAC instruction over an accumulator
+ * Mid   loop: loop over different output channel
+ * Outer loop: loop over different output (x, y)
+ */
+
+arm_status arm_depthwise_separable_conv_HWC_q7(const q7_t * Im_in,
+                                               const uint16_t dim_im_in,
+                                               const uint16_t ch_im_in,
+                                               const q7_t * wt,
+                                               const uint16_t ch_im_out,
+                                               const uint16_t dim_kernel,
+                                               const uint16_t padding,
+                                               const uint16_t stride,
+                                               const q7_t * bias,
+                                               const uint16_t bias_shift,
+                                               const uint16_t out_shift,
+                                               q7_t * Im_out, 
+                                               const uint16_t dim_im_out, 
+                                               q15_t * bufferA, 
+                                               q7_t * bufferB)
+{
+
+#if defined (ARM_MATH_DSP)
+    /* Run the following code for Cortex-M4 and Cortex-M7 */
+
+    int16_t   i_out_y, i_out_x;
+    int16_t   i_ker_y, i_ker_x;
+    q7_t     *colBuffer = (q7_t *) bufferA;
+    q7_t     *pBuffer = colBuffer;
+    const q7_t *pBias = bias;
+    q7_t     *pOut = Im_out;
+    uint16_t  rowCnt;
+    uint16_t  row_shift;
+
+    /* do some checking here, basically ch_im_in == ch_im_out */
+    if (ch_im_in != ch_im_out)
+    {
+        return ARM_MATH_SIZE_MISMATCH;
+    }
+
+    for (i_out_y = 0; i_out_y < dim_im_out; i_out_y++)
+    {
+        for (i_out_x = 0; i_out_x < dim_im_out; i_out_x++)
+        {
+            /* we first do im2col here */
+            for (i_ker_y = i_out_y * stride - padding; i_ker_y < i_out_y * stride - padding + dim_kernel; i_ker_y++)
+            {
+                for (i_ker_x = i_out_x * stride - padding; i_ker_x < i_out_x * stride - padding + dim_kernel; i_ker_x++)
+                {
+                    if (i_ker_y < 0 || i_ker_y >= dim_im_in || i_ker_x < 0 || i_ker_x >= dim_im_in)
+                    {
+                        /* arm_fill_q7(0, pBuffer, ch_im_in); */
+                        memset(pBuffer, 0, ch_im_in);
+                    } else
+                    {
+                        /* arm_copy_q7((q7_t *) Im_in + (i_ker_y * dim_im_in + i_ker_x) * ch_im_in, pBuffer, ch_im_in); */
+                        memcpy(pBuffer, (q7_t *) Im_in + (i_ker_y * dim_im_in + i_ker_x) * ch_im_in, ch_im_in);
+                    }
+                    pBuffer += ch_im_in;
+                }
+            }
+
+            /* we will do the computation here for each channel */
+            rowCnt = ch_im_out >> 2;
+            row_shift = 0;
+            pBias = bias;
+
+            while (rowCnt)
+            {
+                q31_t     sum =  ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift);
+                q31_t     sum2 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift);
+                q31_t     sum3 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift);
+                q31_t     sum4 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift);
+
+                uint16_t  colCnt = (dim_kernel * dim_kernel) >> 1;
+                q7_t     *pB = colBuffer + row_shift;
+                const q7_t *pA = wt + row_shift;
+                row_shift += 4;
+
+#ifdef USE_INTRINSIC
+
+#ifndef ARM_MATH_BIG_ENDIAN
+
+                while (colCnt)
+                {
+                    q31_t     inA1, inA2, inB1, inB2, opA, opB;
+
+                    inB1 = *__SIMD32(pB);
+                    pB += ch_im_in;
+                    opB = *__SIMD32(pB);
+                    pB += ch_im_in;
+                    inB2 = __PKHTB(opB, inB1, 16);
+                    inB1 = __PKHBT(inB1, opB, 16);
+                    inA1 = *__SIMD32(pA);
+                    pA += ch_im_in;
+                    opB = *__SIMD32(pA);
+                    pA += ch_im_in;
+                    inA2 = __PKHTB(opB, inA1, 16);
+                    inA1 = __PKHBT(inA1, opB, 16);
+                    opA = __SXTB16(inA1);
+                    opB = __SXTB16(inB1);
+                    sum = __SMLAD(opA, opB, sum);
+                    opA = __SXTB16(__ROR(inA1, 8));
+                    opB = __SXTB16(__ROR(inB1, 8));
+                    sum2 = __SMLAD(opA, opB, sum2);
+                    opA = __SXTB16(inA2);
+                    opB = __SXTB16(inB2);
+                    sum3 = __SMLAD(opA, opB, sum3);
+                    opA = __SXTB16(__ROR(inA2, 8));
+                    opB = __SXTB16(__ROR(inB2, 8));
+                    sum4 = __SMLAD(opA, opB, sum4);
+                    colCnt--;
+                }
+#else
+
+                while (colCnt)
+                {
+                    q31_t     inA1, inA2, inB1, inB2, opA, opB;
+
+                    inB1 = *__SIMD32(pB);
+                    pB += ch_im_in;
+                    opB = *__SIMD32(pB);
+                    pB += ch_im_in;
+                    inB2 = __PKHBT(opB, inB1, 16);
+                    inB1 = __PKHTB(inB1, opB, 16);
+                    inA1 = *__SIMD32(pA);
+                    pA += ch_im_in;
+                    opB = *__SIMD32(pA);
+                    pA += ch_im_in;
+                    inA2 = __PKHBT(opB, inA1, 16);
+                    inA1 = __PKHTB(inA1, opB, 16);
+                    opA = __SXTB16(inA1);
+                    opB = __SXTB16(inB1);
+                    sum2 = __SMLAD(opA, opB, sum2);
+                    opA = __SXTB16(__ROR(inA1, 8));
+                    opB = __SXTB16(__ROR(inB1, 8));
+                    sum = __SMLAD(opA, opB, sum);
+                    opA = __SXTB16(inA2);
+                    opB = __SXTB16(inB2);
+                    sum4 = __SMLAD(opA, opB, sum4);
+                    opA = __SXTB16(__ROR(inA2, 8));
+                    opB = __SXTB16(__ROR(inB2, 8));
+                    sum3 = __SMLAD(opA, opB, sum3);
+                    colCnt--;
+                }
+
+#endif                          /* ARM_MATH_BIG_ENDIAN */
+
+#else
+
+#ifndef ARM_MATH_BIG_ENDIAN
+                /*
+                 *   r0    r1    r2    r3    r4   r5
+                 *  inA1, inA2, inB1, inB2, opA, opB
+                 */
+
+                asm volatile ("COL_LOOP_%=:\n"
+                              "ldr.w r2, [%[pB], #0]\n"
+                              "add.w %[pB], %[pB], %[ch_im_in]\n"
+                              "ldr.w r5, [%[pB], #0]\n"
+                              "add.w %[pB], %[pB], %[ch_im_in]\n"
+                              "pkhtb r3, r5, r2, ASR #16\n"
+                              "pkhbt r2, r2, r5, LSL #16\n"
+                              "ldr.w r0, [%[pA], #0]\n"
+                              "add.w %[pA], %[pA], %[ch_im_in]\n"
+                              "ldr.w r5, [%[pA], #0]\n"
+                              "add.w %[pA], %[pA], %[ch_im_in]\n"
+                              "pkhtb r1, r5, r0, ASR #16\n"
+                              "pkhbt r0, r0, r5, LSL #16\n"
+                              "sxtb16 r4, r0\n"
+                              "sxtb16 r5, r2\n"
+                              "smlad %[sum], r4, r5, %[sum]\n"
+                              "mov.w r4, r0, ror #8\n"
+                              "mov.w r5, r2, ror #8\n"
+                              "sxtb16 r4, r4\n"
+                              "sxtb16 r5, r5\n"
+                              "smlad %[sum2], r4, r5, %[sum2]\n"
+                              "sxtb16 r4, r1\n"
+                              "sxtb16 r5, r3\n"
+                              "smlad %[sum3], r4, r5, %[sum3]\n"
+                              "mov.w r4, r1, ror #8\n"
+                              "mov.w r5, r3, ror #8\n"
+                              "sxtb16 r4, r4\n"
+                              "sxtb16 r5, r5\n"
+                              "smlad %[sum4], r4, r5, %[sum4]\n"
+                              "subs %[colCnt], #1\n"
+                              "bne COL_LOOP_%=\n":[sum]
+                              "+r"(sum),[sum2] "+r"(sum2),
+                              [sum3] "+r"(sum3),
+                              [sum4] "+r"(sum4),[pB] "+r"(pB),
+                              [pA] "+r"(pA):[colCnt]
+                              "r"(colCnt),[ch_im_in] "r"(ch_im_in):"r0", "r1", "r2", "r3", "r4", "r5");
+#else
+                /*
+                 *  r0    r1    r2    r3    r4   r5
+                 * inA1, inA2, inB1, inB2, opA, opB
+                 */
+                asm volatile ("COL_LOOP_%=:\n"
+                              "ldr.w r2, [%[pB], #0]\n"
+                              "add.w %[pB], %[pB], %[ch_im_in]\n"
+                              "ldr.w r5, [%[pB], #0]\n"
+                              "add.w %[pB], %[pB], %[ch_im_in]\n"
+                              "pkhbt r3, r5, r2, LSL #16\n"
+                              "pkhtb r2, r2, r5, ASR #16\n"
+                              "ldr.w r0, [%[pA], #0]\n"
+                              "add.w %[pA], %[pA], %[ch_im_in]\n"
+                              "ldr.w r5, [%[pA], #0]\n"
+                              "add.w %[pA], %[pA], %[ch_im_in]\n"
+                              "pkhbt r1, r5, r0, LSL #16\n"
+                              "pkhtb r0, r0, r5, ASR #16\n"
+                              "sxtb16 r4, r0\n"
+                              "sxtb16 r5, r2\n"
+                              "smlad %[sum2], r4, r5, %[sum2]\n"
+                              "mov.w r4, r0, ror #8\n"
+                              "mov.w r5, r2, ror #8\n"
+                              "sxtb16 r4, r4\n"
+                              "sxtb16 r5, r5\n"
+                              "smlad %[sum], r4, r5, %[sum]\n"
+                              "sxtb16 r4, r1\n"
+                              "sxtb16 r5, r3\n"
+                              "smlad %[sum4], r4, r5, %[sum4]\n"
+                              "mov.w r4, r1, ror #8\n"
+                              "mov.w r5, r3, ror #8\n"
+                              "sxtb16 r4, r4\n"
+                              "sxtb16 r5, r5\n"
+                              "smlad %[sum3], r4, r5, %[sum3]\n"
+                              "subs %[colCnt], #1\n"
+                              "bne COL_LOOP_%=\n":[sum]
+                              "+r"(sum),[sum2] "+r"(sum2),
+                              [sum3] "+r"(sum3),
+                              [sum4] "+r"(sum4),[pB] "+r"(pB),
+                              [pA] "+r"(pA):[colCnt]
+                              "r"(colCnt),[ch_im_in] "r"(ch_im_in):"r0", "r1", "r2", "r3", "r4", "r5");
+
+#endif                          /* ARM_MATH_BIG_ENDIAN */
+
+#endif                          /* USE_INTRINSIC */
+
+                colCnt = (dim_kernel * dim_kernel) & 0x1;
+                while (colCnt)
+                {
+                    union arm_nnword inA, inB;
+                    inA.word = *__SIMD32(pA);
+                    pA += ch_im_in;
+                    inB.word = *__SIMD32(pB);
+                    pB += ch_im_in;
+                    sum += inA.bytes[0] * inB.bytes[0];
+                    sum2 += inA.bytes[1] * inB.bytes[1];
+                    sum3 += inA.bytes[2] * inB.bytes[2];
+                    sum4 += inA.bytes[3] * inB.bytes[3];
+                    colCnt--;
+                }
+
+                *pOut++ = (q7_t) __SSAT((sum >> out_shift), 8);
+                *pOut++ = (q7_t) __SSAT((sum2 >> out_shift), 8);
+                *pOut++ = (q7_t) __SSAT((sum3 >> out_shift), 8);
+                *pOut++ = (q7_t) __SSAT((sum4 >> out_shift), 8);
+
+                rowCnt--;
+            }
+
+            rowCnt = ch_im_out & 0x3;
+            while (rowCnt)
+            {
+                q7_t     *pB = colBuffer + row_shift;
+                const q7_t *pA = wt + row_shift;
+                q31_t     sum = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift);
+                uint16_t  colCnt = (dim_kernel * dim_kernel);
+
+                row_shift += 1;
+
+                while (colCnt)
+                {
+                    q7_t      A1 = *pA;
+                    q7_t      B1 = *pB;
+                    pA += ch_im_in;
+                    pB += ch_im_in;
+                    sum += A1 * B1;
+
+                    colCnt--;
+                }
+                *pOut++ = (q7_t) __SSAT((sum >> out_shift), 8);
+                rowCnt--;
+            }
+
+            /* clear counter and pointers */
+            pBuffer = colBuffer;
+        }
+    }
+
+#else
+    /* Run the following code as reference implementation for Cortex-M0 and Cortex-M3 */
+    int       i_out_y, i_out_x, i_ch_out, i_ker_x, i_ker_y;
+    int       conv_out;
+
+    /* do some checking here, basically ch_im_in == ch_im_out */
+    if (ch_im_in != ch_im_out)
+    {
+        return ARM_MATH_SIZE_MISMATCH;
+    }
+
+    for (i_out_y = 0; i_out_y < dim_im_out; i_out_y++)
+    {
+        for (i_out_x = 0; i_out_x < dim_im_out; i_out_x++)
+        {
+            for (i_ch_out = 0; i_ch_out < ch_im_out; i_ch_out++)
+            {
+                // for each output
+                conv_out = ((q31_t)(bias[i_ch_out]) << bias_shift) + NN_ROUND(out_shift);
+                for (i_ker_y = 0; i_ker_y < dim_kernel; i_ker_y++)
+                {
+                    for (i_ker_x = 0; i_ker_x < dim_kernel; i_ker_x++)
+                    {
+                        int       in_row = stride * i_out_y + i_ker_y - padding;
+                        int       in_col = stride * i_out_x + i_ker_x - padding;
+                        if (in_row >= 0 && in_col >= 0 && in_row < dim_im_in && in_col < dim_im_in)
+                        {
+                            conv_out +=
+                                Im_in[(in_row *
+                                       dim_im_in +
+                                       in_col) *
+                                      ch_im_in +
+                                      i_ch_out] * wt[(i_ker_y * dim_kernel + i_ker_x) * ch_im_out + i_ch_out];
+                        }
+                    }
+                }
+                Im_out[(i_out_y * dim_im_out +
+                        i_out_x) * ch_im_out + i_ch_out] = (q7_t) __SSAT((conv_out >> out_shift), 8);
+            }
+        }
+    }
+
+#endif                          /* ARM_MATH_DSP */
+
+    /* Return to application */
+    return ARM_MATH_SUCCESS;
+
+}
+
+/**
+ * @} end of NNConv group
+ */
diff --git a/NN/Source/ConvolutionFunctions/arm_depthwise_separable_conv_HWC_q7_nonsquare.c b/NN/Source/ConvolutionFunctions/arm_depthwise_separable_conv_HWC_q7_nonsquare.c
new file mode 100644
index 0000000..397f233
--- /dev/null
+++ b/NN/Source/ConvolutionFunctions/arm_depthwise_separable_conv_HWC_q7_nonsquare.c
@@ -0,0 +1,411 @@
+/*
+ * Copyright (C) 2010-2018 Arm Limited or its affiliates. All rights reserved.
+ *
+ * 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.
+ */
+
+/* ----------------------------------------------------------------------
+ * Project:      CMSIS NN Library
+ * Title:        arm_depthwise_separable_conv_HWC_q7_nonsquare.c
+ * Description:  Q7 depthwise separable convolution function (non-square shape)
+ *
+ * $Date:        17. January 2018
+ * $Revision:    V.1.0.0
+ *
+ * Target Processor:  Cortex-M cores
+ *
+ * -------------------------------------------------------------------- */
+
+#include "arm_math.h"
+#include "arm_nnfunctions.h"
+
+/**
+ *  @ingroup groupNN
+ */
+
+/**
+ * @addtogroup NNConv
+ * @{
+ */
+
+/**
+ * @brief Q7 depthwise separable convolution function (non-square 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 either
+ * <code>ARM_MATH_SIZE_MISMATCH</code> or <code>ARM_MATH_SUCCESS</code> based on the outcome of size checking.
+ *
+ * This function is the version with full list of optimization tricks, but with
+ * some contraints:
+ *   ch_im_in is multiple of 2
+ *   ch_im_out is multiple of 2
+ */
+
+arm_status arm_depthwise_separable_conv_HWC_q7_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)
+{
+
+#if defined (ARM_MATH_DSP)
+    /* Run the following code for Cortex-M4 and Cortex-M7 */
+
+/*
+ * Implementation:
+ * There are 3 nested loop here:
+ * Inner loop: calculate each output value with MAC instruction over an accumulator
+ * Mid   loop: loop over different output channel
+ * Outer loop: loop over different output (x, y)
+ *
+ */
+
+    int16_t   i_out_y, i_out_x;
+    int16_t   i_ker_y, i_ker_x;
+    q7_t     *colBuffer = (q7_t *) bufferA;
+    q7_t     *pBuffer = colBuffer;
+    const q7_t *pBias = bias;
+    q7_t     *pOut = Im_out;
+    uint16_t  rowCnt;
+    uint16_t  row_shift;
+
+    /* do some checking here, basically ch_im_in == ch_im_out */
+    if (ch_im_in != ch_im_out)
+    {
+        return ARM_MATH_SIZE_MISMATCH;
+    }
+
+    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++)
+        {
+            /* we first do im2col here */
+            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_x)
+                    {
+                        /* arm_fill_q7(0, pBuffer, ch_im_in); */
+                        memset(pBuffer, 0, ch_im_in);
+                    } else
+                    {
+                        /* arm_copy_q7((q7_t *) Im_in + (i_ker_y * dim_im_in_x + i_ker_x) * ch_im_in, pBuffer, ch_im_in); */
+                        memcpy(pBuffer, (q7_t *) Im_in + (i_ker_y * dim_im_in_x + i_ker_x) * ch_im_in, ch_im_in);
+                    }
+                    pBuffer += ch_im_in;
+                }
+            }
+
+            /* we will do the computation here for each channel */
+            rowCnt = ch_im_out >> 2;
+            row_shift = 0;
+            pBias = bias;
+
+            while (rowCnt)
+            {
+                q31_t     sum =  ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift);
+                q31_t     sum2 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift);
+                q31_t     sum3 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift);
+                q31_t     sum4 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift);
+
+                uint16_t  colCnt = (dim_kernel_x * dim_kernel_y) >> 1;
+                q7_t     *pB = colBuffer + row_shift;
+                const q7_t *pA = wt + row_shift;
+                row_shift += 4;
+
+#ifdef USE_INTRINSIC
+
+#ifndef ARM_MATH_BIG_ENDIAN
+
+                while (colCnt)
+                {
+                    q31_t     inA1, inA2, inB1, inB2, opA, opB;
+
+                    inB1 = *__SIMD32(pB);
+                    pB += ch_im_in;
+                    opB = *__SIMD32(pB);
+                    pB += ch_im_in;
+                    inB2 = __PKHTB(opB, inB1, 16);
+                    inB1 = __PKHBT(inB1, opB, 16);
+                    inA1 = *__SIMD32(pA);
+                    pA += ch_im_in;
+                    opB = *__SIMD32(pA);
+                    pA += ch_im_in;
+                    inA2 = __PKHTB(opB, inA1, 16);
+                    inA1 = __PKHBT(inA1, opB, 16);
+                    opA = __SXTB16(inA1);
+                    opB = __SXTB16(inB1);
+                    sum = __SMLAD(opA, opB, sum);
+                    opA = __SXTB16(__ROR(inA1, 8));
+                    opB = __SXTB16(__ROR(inB1, 8));
+                    sum2 = __SMLAD(opA, opB, sum2);
+                    opA = __SXTB16(inA2);
+                    opB = __SXTB16(inB2);
+                    sum3 = __SMLAD(opA, opB, sum3);
+                    opA = __SXTB16(__ROR(inA2, 8));
+                    opB = __SXTB16(__ROR(inB2, 8));
+                    sum4 = __SMLAD(opA, opB, sum4);
+                    colCnt--;
+                }
+#else
+
+                while (colCnt)
+                {
+                    q31_t     inA1, inA2, inB1, inB2, opA, opB;
+
+                    inB1 = *__SIMD32(pB);
+                    pB += ch_im_in;
+                    opB = *__SIMD32(pB);
+                    pB += ch_im_in;
+                    inB2 = __PKHBT(opB, inB1, 16);
+                    inB1 = __PKHTB(inB1, opB, 16);
+                    inA1 = *__SIMD32(pA);
+                    pA += ch_im_in;
+                    opB = *__SIMD32(pA);
+                    pA += ch_im_in;
+                    inA2 = __PKHBT(opB, inA1, 16);
+                    inA1 = __PKHTB(inA1, opB, 16);
+                    opA = __SXTB16(inA1);
+                    opB = __SXTB16(inB1);
+                    sum2 = __SMLAD(opA, opB, sum2);
+                    opA = __SXTB16(__ROR(inA1, 8));
+                    opB = __SXTB16(__ROR(inB1, 8));
+                    sum = __SMLAD(opA, opB, sum);
+                    opA = __SXTB16(inA2);
+                    opB = __SXTB16(inB2);
+                    sum4 = __SMLAD(opA, opB, sum4);
+                    opA = __SXTB16(__ROR(inA2, 8));
+                    opB = __SXTB16(__ROR(inB2, 8));
+                    sum3 = __SMLAD(opA, opB, sum3);
+                    colCnt--;
+                }
+
+#endif                          /* ARM_MATH_BIG_ENDIAN */
+
+#else
+
+#ifndef ARM_MATH_BIG_ENDIAN
+                //  r0    r1    r2    r3    r4   r5
+                // inA1, inA2, inB1, inB2, opA, opB
+                asm volatile ("COL_LOOP:\n"
+                              "ldr.w r2, [%[pB], #0]\n"
+                              "add.w %[pB], %[pB], %[ch_im_in]\n"
+                              "ldr.w r5, [%[pB], #0]\n"
+                              "add.w %[pB], %[pB], %[ch_im_in]\n"
+                              "pkhtb r3, r5, r2, ASR #16\n"
+                              "pkhbt r2, r2, r5, LSL #16\n"
+                              "ldr.w r0, [%[pA], #0]\n"
+                              "add.w %[pA], %[pA], %[ch_im_in]\n"
+                              "ldr.w r5, [%[pA], #0]\n"
+                              "add.w %[pA], %[pA], %[ch_im_in]\n"
+                              "pkhtb r1, r5, r0, ASR #16\n"
+                              "pkhbt r0, r0, r5, LSL #16\n"
+                              "sxtb16 r4, r0\n"
+                              "sxtb16 r5, r2\n"
+                              "smlad %[sum], r4, r5, %[sum]\n"
+                              "mov.w r4, r0, ror #8\n"
+                              "mov.w r5, r2, ror #8\n"
+                              "sxtb16 r4, r4\n"
+                              "sxtb16 r5, r5\n"
+                              "smlad %[sum2], r4, r5, %[sum2]\n"
+                              "sxtb16 r4, r1\n"
+                              "sxtb16 r5, r3\n"
+                              "smlad %[sum3], r4, r5, %[sum3]\n"
+                              "mov.w r4, r1, ror #8\n"
+                              "mov.w r5, r3, ror #8\n"
+                              "sxtb16 r4, r4\n"
+                              "sxtb16 r5, r5\n"
+                              "smlad %[sum4], r4, r5, %[sum4]\n"
+                              "subs %[colCnt], #1\n"
+                              "bne COL_LOOP\n":[sum] "+r"(sum),[sum2] "+r"(sum2),[sum3] "+r"(sum3),
+                              [sum4] "+r"(sum4),[pB] "+r"(pB),[pA] "+r"(pA):[colCnt] "r"(colCnt),
+                              [ch_im_in] "r"(ch_im_in):"r0", "r1", "r2", "r3", "r4", "r5");
+#else
+                //  r0    r1    r2    r3    r4   r5
+                // inA1, inA2, inB1, inB2, opA, opB
+                asm volatile ("COL_LOOP:\n"
+                              "ldr.w r2, [%[pB], #0]\n"
+                              "add.w %[pB], %[pB], %[ch_im_in]\n"
+                              "ldr.w r5, [%[pB], #0]\n"
+                              "add.w %[pB], %[pB], %[ch_im_in]\n"
+                              "pkhbt r3, r5, r2, LSL #16\n"
+                              "pkhtb r2, r2, r5, ASR #16\n"
+                              "ldr.w r0, [%[pA], #0]\n"
+                              "add.w %[pA], %[pA], %[ch_im_in]\n"
+                              "ldr.w r5, [%[pA], #0]\n"
+                              "add.w %[pA], %[pA], %[ch_im_in]\n"
+                              "pkhbt r1, r5, r0, LSL #16\n"
+                              "pkhtb r0, r0, r5, ASR #16\n"
+                              "sxtb16 r4, r0\n"
+                              "sxtb16 r5, r2\n"
+                              "smlad %[sum2], r4, r5, %[sum2]\n"
+                              "mov.w r4, r0, ror #8\n"
+                              "mov.w r5, r2, ror #8\n"
+                              "sxtb16 r4, r4\n"
+                              "sxtb16 r5, r5\n"
+                              "smlad %[sum], r4, r5, %[sum]\n"
+                              "sxtb16 r4, r1\n"
+                              "sxtb16 r5, r3\n"
+                              "smlad %[sum4], r4, r5, %[sum4]\n"
+                              "mov.w r4, r1, ror #8\n"
+                              "mov.w r5, r3, ror #8\n"
+                              "sxtb16 r4, r4\n"
+                              "sxtb16 r5, r5\n"
+                              "smlad %[sum3], r4, r5, %[sum3]\n"
+                              "subs %[colCnt], #1\n"
+                              "bne COL_LOOP\n":[sum] "+r"(sum),[sum2] "+r"(sum2),[sum3] "+r"(sum3),
+                              [sum4] "+r"(sum4),[pB] "+r"(pB),[pA] "+r"(pA):[colCnt] "r"(colCnt),
+                              [ch_im_in] "r"(ch_im_in):"r0", "r1", "r2", "r3", "r4", "r5");
+#endif                          /*ARM_MATH_BIG_ENDIAN */
+
+#endif                          /* USE_INTRINSIC */
+
+                colCnt = (dim_kernel_x * dim_kernel_y) & 0x1;
+                while (colCnt)
+                {
+                    union arm_nnword inA, inB;
+                    inA.word = *__SIMD32(pA);
+                    pA += ch_im_in;
+                    inB.word = *__SIMD32(pB);
+                    pB += ch_im_in;
+                    sum += inA.bytes[0] * inB.bytes[0];
+                    sum2 += inA.bytes[1] * inB.bytes[1];
+                    sum3 += inA.bytes[2] * inB.bytes[2];
+                    sum4 += inA.bytes[3] * inB.bytes[3];
+                    colCnt--;
+                }
+
+                *pOut++ = (q7_t) __SSAT((sum >> out_shift), 8);
+                *pOut++ = (q7_t) __SSAT((sum2 >> out_shift), 8);
+                *pOut++ = (q7_t) __SSAT((sum3 >> out_shift), 8);
+                *pOut++ = (q7_t) __SSAT((sum4 >> out_shift), 8);
+
+                rowCnt--;
+            }
+
+            rowCnt = ch_im_out & 0x3;
+            while (rowCnt)
+            {
+                q7_t     *pB = colBuffer + row_shift;
+                const q7_t *pA = wt + row_shift;
+                q31_t     sum = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift);
+                uint16_t  colCnt = (dim_kernel_x * dim_kernel_y);
+
+                row_shift += 1;
+
+                while (colCnt)
+                {
+                    q7_t      A1 = *pA;
+                    q7_t      B1 = *pB;
+                    pA += ch_im_in;
+                    pB += ch_im_in;
+                    sum += A1 * B1;
+
+                    colCnt--;
+                }
+                *pOut++ = (q7_t) __SSAT((sum >> out_shift), 8);
+                rowCnt--;
+            }
+
+            // clear counter and pointers
+            pBuffer = colBuffer;
+        }
+    }
+
+#else
+    /* Run the following code as reference implementation for Cortex-M0 and Cortex-M3 */
+    int       i_out_y, i_out_x, i_ch_out;
+    int       i_ker_y, i_ker_x; 
+
+    /* do some checking here, basically ch_im_in == ch_im_out */
+    if (ch_im_in != ch_im_out)
+    {
+        return ARM_MATH_SIZE_MISMATCH;
+    }
+
+    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++)
+        {
+            for (i_ch_out = 0; i_ch_out < ch_im_out; i_ch_out++)
+            {
+                // for each output 
+                int       conv_out = ((q31_t)(bias[i_ch_out]) << bias_shift) + NN_ROUND(out_shift);
+                for (i_ker_y = 0; i_ker_y < dim_kernel_y; i_ker_y++)
+                {
+                    for (i_ker_x = 0; i_ker_x < dim_kernel_x; i_ker_x++)
+                    {
+                        int       in_row = stride_y * i_out_y + i_ker_y - padding_y;
+                        int       in_col = stride_x * i_out_x + i_ker_x - padding_x;
+                        if (in_row >= 0 && in_col >= 0 && in_row < dim_im_in_y && in_col < dim_im_in_x)
+                        {
+                            conv_out += Im_in[(in_row * dim_im_in_x + in_col) * ch_im_in + i_ch_out] *                        
+                                wt[(i_ker_y * dim_kernel_x + i_ker_x) * ch_im_out + i_ch_out];
+                        }
+                    }
+                }
+                Im_out[(i_out_y * dim_im_out_x + i_out_x) * ch_im_out + i_ch_out] =
+                    (q7_t) __SSAT((conv_out >> out_shift), 8);
+            }
+        }
+    }
+
+#endif                          /* ARM_MATH_DSP */
+
+
+    /* Return to application */
+    return ARM_MATH_SUCCESS;
+
+}
+
+/**
+ * @} end of NNConv group
+ */
diff --git a/NN/Source/ConvolutionFunctions/arm_nn_mat_mult_kernel_q7_q15.c b/NN/Source/ConvolutionFunctions/arm_nn_mat_mult_kernel_q7_q15.c
new file mode 100644
index 0000000..a4adc5d
--- /dev/null
+++ b/NN/Source/ConvolutionFunctions/arm_nn_mat_mult_kernel_q7_q15.c
@@ -0,0 +1,187 @@
+/*
+ * Copyright (C) 2010-2018 Arm Limited or its affiliates. All rights reserved.
+ *
+ * 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.
+ */
+
+/* ----------------------------------------------------------------------
+ * Project:      CMSIS NN Library
+ * Title:        arm_nn_mat_mult_kernel_q7_q15.c
+ * Description:  Matrix-multiplication function for convolution
+ *
+ * $Date:        17. January 2018
+ * $Revision:    V.1.0.0
+ *
+ * Target Processor:  Cortex-M cores
+ * -------------------------------------------------------------------- */
+
+#include "arm_math.h"
+#include "arm_nnfunctions.h"
+
+  /**
+   * @brief Matrix-multiplication function for convolution
+   * @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]       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_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 bias_shift,
+                                        const uint16_t out_shift, 
+                                        const q7_t * bias, 
+                                        q7_t * pOut)
+{
+#if defined (ARM_MATH_DSP)
+    /* set up the second output pointers */
+    q7_t     *pOut2 = pOut + ch_im_out;
+    const q7_t *pBias = bias;
+
+    uint16_t  rowCnt = ch_im_out >> 1;
+    /* this loop over rows in A */
+    while (rowCnt)
+    {
+        /* 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;
+
+        /* init the sum with bias */
+        q31_t     sum =  ((q31_t)(*pBias) << bias_shift) + NN_ROUND(out_shift);
+        q31_t     sum2 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift);
+        q31_t     sum3 = ((q31_t)(*pBias) << bias_shift) + NN_ROUND(out_shift);
+        q31_t     sum4 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift);
+
+        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 = (q7_t *) read_and_pad((void *)pA, &inA11, &inA12);
+            pA2 = (q7_t *) read_and_pad((void *)pA2, &inA21, &inA22);
+
+            sum = __SMLAD(inA11, inB1, sum);
+            sum2 = __SMLAD(inA11, inB2, sum2);
+            sum3 = __SMLAD(inA21, inB1, sum3);
+            sum4 = __SMLAD(inA21, inB2, sum4);
+
+            inB1 = *__SIMD32(pB)++;
+            inB2 = *__SIMD32(pB2)++;
+
+            sum = __SMLAD(inA12, inB1, sum);
+            sum2 = __SMLAD(inA12, inB2, sum2);
+            sum3 = __SMLAD(inA22, inB1, sum3);
+            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;
+            sum2 += inA1 * inB2;
+            sum3 += 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;
+        rowCnt--;
+    }                           /* for over ch_im_out */
+
+    /* compute left-over row if any */
+    if (ch_im_out & 0x1)
+    {
+        /* 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     sum2 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift);
+
+        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);
+            sum2 = __SMLAD(inA11, inB2, sum2);
+
+            inB1 = *__SIMD32(pB)++;
+            inB2 = *__SIMD32(pB2)++;
+            sum = __SMLAD(inA12, inB1, sum);
+            sum2 = __SMLAD(inA12, inB2, sum2);
+
+            colCnt--;
+        }
+        colCnt = numCol_A & 0x3;
+        while (colCnt)
+        {
+            q7_t      inA1 = *pA++;
+            q15_t     inB1 = *pB++;
+            q15_t     inB2 = *pB2++;
+
+            sum += inA1 * inB1;
+            sum2 += inA1 * inB2;
+            colCnt--;
+        }
+
+        *pOut++ = (q7_t) __SSAT((sum >> out_shift), 8);
+        *pOut2++ = (q7_t) __SSAT((sum2 >> out_shift), 8);
+    }
+
+    pOut += ch_im_out;
+
+    /* return the new output pointer with offset */
+    return pOut;
+#else
+    /* To be completed */
+    return NULL;
+#endif                          /* ARM_MATH_DSP */
+
+}
diff --git a/NN/Source/ConvolutionFunctions/arm_nn_mat_mult_kernel_q7_q15_reordered.c b/NN/Source/ConvolutionFunctions/arm_nn_mat_mult_kernel_q7_q15_reordered.c
new file mode 100644
index 0000000..deef7c6
--- /dev/null
+++ b/NN/Source/ConvolutionFunctions/arm_nn_mat_mult_kernel_q7_q15_reordered.c
@@ -0,0 +1,138 @@
+/*
+ * Copyright (C) 2010-2018 Arm Limited or its affiliates. All rights reserved.
+ *
+ * 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.
+ */
+
+/* ----------------------------------------------------------------------
+ * Project:      CMSIS NN Library
+ * Title:        arm_nn_mat_mult_kernel_q7_q15_reordered.c
+ * Description:  Matrix-multiplication function for convolution with reordered columns
+ *
+ * $Date:        17. January 2018
+ * $Revision:    V.1.0.0
+ *
+ * Target Processor:  Cortex-M cores
+ * -------------------------------------------------------------------- */
+
+#include "arm_nnfunctions.h"
+#include "arm_math.h"
+
+  /**
+   * @brief Matrix-multiplication function for convolution with reordered columns
+   * @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]       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 assumes that data in pInBuffer are reordered
+   */
+
+q7_t     *arm_nn_mat_mult_kernel_q7_q15_reordered(const q7_t * pA,
+                                                  const q15_t * pInBuffer,
+                                                  const uint16_t ch_im_out,
+                                                  const uint16_t numCol_A,
+                                                  const uint16_t bias_shift,
+                                                  const uint16_t out_shift, 
+                                                  const q7_t * bias, 
+                                                  q7_t * pOut)
+{
+
+#if defined (ARM_MATH_DSP)
+    /* set up the second output pointers */
+    q7_t     *pOut2 = pOut + ch_im_out;
+    int       i;
+
+    /* this loop over rows in A */
+    for (i = 0; i < ch_im_out; i += 2)
+    {
+        /* 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;
+
+        /* init the sum with bias */
+        q31_t     sum =  ((q31_t)(bias[i]) << bias_shift) + NN_ROUND(out_shift);
+        q31_t     sum2 = ((q31_t)(bias[i]) << bias_shift) + NN_ROUND(out_shift);
+        q31_t     sum3 = ((q31_t)(bias[i + 1]) << bias_shift) + NN_ROUND(out_shift);
+        q31_t     sum4 = ((q31_t)(bias[i + 1]) << bias_shift) + NN_ROUND(out_shift);
+
+        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 = (q7_t *) read_and_pad_reordered((void *)pA, &inA11, &inA12);
+            pA2 = (q7_t *) read_and_pad_reordered((void *)pA2, &inA21, &inA22);
+
+            sum = __SMLAD(inA11, inB1, sum);
+            sum2 = __SMLAD(inA11, inB2, sum2);
+            sum3 = __SMLAD(inA21, inB1, sum3);
+            sum4 = __SMLAD(inA21, inB2, sum4);
+
+            inB1 = *__SIMD32(pB)++;
+            inB2 = *__SIMD32(pB2)++;
+
+            sum = __SMLAD(inA12, inB1, sum);
+            sum2 = __SMLAD(inA12, inB2, sum2);
+            sum3 = __SMLAD(inA22, inB1, sum3);
+            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;
+            sum2 += inA1 * inB2;
+            sum3 += 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 */
+
+    pOut += ch_im_out;
+
+    /* return the new output pointer with offset */
+    return pOut;
+#else
+    /* To be completed */
+    return NULL;
+#endif                          /* ARM_MATH_DSP */
+}
diff --git a/NN/Source/FullyConnectedFunctions/arm_fully_connected_mat_q7_vec_q15.c b/NN/Source/FullyConnectedFunctions/arm_fully_connected_mat_q7_vec_q15.c
new file mode 100644
index 0000000..2746967
--- /dev/null
+++ b/NN/Source/FullyConnectedFunctions/arm_fully_connected_mat_q7_vec_q15.c
@@ -0,0 +1,199 @@
+/*
+ * Copyright (C) 2010-2018 Arm Limited or its affiliates. All rights reserved.
+ *
+ * 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.
+ */
+
+/* ----------------------------------------------------------------------
+ * Project:      CMSIS NN Library
+ * Title:        arm_fully_connected_mat_q7_vec_q15.c
+ * Description:  Mixed Q15-Q7 fully-connected layer function
+ *
+ * $Date:        17. January 2018
+ * $Revision:    V.1.0.0
+ *
+ * Target Processor:  Cortex-M cores
+ *
+ * -------------------------------------------------------------------- */
+
+#include "arm_math.h"
+#include "arm_nnfunctions.h"
+
+/**
+ *  @ingroup groupNN
+ */
+
+/**
+ * @addtogroup FC
+ * @{
+ */
+
+  /**
+   * @brief Mixed Q15-Q7 fully-connected layer function
+   * @param[in]       pV          pointer to input vector
+   * @param[in]       pM          pointer to matrix weights
+   * @param[in]       dim_vec     length of the vector
+   * @param[in]       num_of_rows number of rows in weight matrix
+   * @param[in]       bias_shift  amount of left-shift for bias
+   * @param[in]       out_shift   amount of right-shift for output
+   * @param[in]       bias        pointer to bias
+   * @param[in,out]   pOut        pointer to output vector
+   * @param[in,out]   vec_buffer  pointer to buffer space for input
+   * @return     The function returns <code>ARM_MATH_SUCCESS</code>
+   *
+   * @details
+   *
+   * <b>Buffer size:</b>
+   *
+   * vec_buffer size: 0
+   *
+   *  Q7_Q15 version of the fully connected layer
+   *
+   *  Weights are in q7_t and Activations are in q15_t
+   *
+   */
+
+arm_status
+arm_fully_connected_mat_q7_vec_q15(const q15_t * pV,
+                                   const q7_t * pM,
+                                   const uint16_t dim_vec,
+                                   const uint16_t num_of_rows,
+                                   const uint16_t bias_shift,
+                                   const uint16_t out_shift, 
+                                   const q7_t * bias, 
+                                   q15_t * pOut, 
+                                   q15_t * vec_buffer)
+{
+
+#if defined (ARM_MATH_DSP)
+    /* Run the following code for Cortex-M4 and Cortex-M7 */
+
+    const q7_t *pB = pM;
+    const q7_t *pB2;
+    q15_t    *pO = pOut;
+    const q7_t *pBias = bias;
+    const q15_t *pA = pV;
+
+    uint16_t  rowCnt = num_of_rows >> 1;
+
+    while (rowCnt)
+    {
+        q31_t     sum =  ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift);
+        q31_t     sum2 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift);
+        uint16_t  colCnt = dim_vec >> 2;
+
+        pA = pV;
+        pB2 = pB + dim_vec;
+
+        while (colCnt)
+        {
+            q31_t     inV, inM11, inM12, inM21, inM22;
+            pB = (q7_t *) read_and_pad((void *)pB, &inM11, &inM12);
+            pB2 = (q7_t *) read_and_pad((void *)pB2, &inM21, &inM22);
+
+            inV = *__SIMD32(pA)++;
+
+            sum = __SMLAD(inV, inM11, sum);
+            sum2 = __SMLAD(inV, inM21, sum2);
+
+            inV = *__SIMD32(pA)++;
+
+            sum = __SMLAD(inV, inM12, sum);
+            sum2 = __SMLAD(inV, inM22, sum2);
+
+            colCnt--;
+        }
+        colCnt = dim_vec & 0x3;
+        while (colCnt)
+        {
+            q15_t     inV = *pA++;
+            q7_t      inM = *pB++;
+            q7_t      inM2 = *pB2++;
+
+            sum += inV * inM;
+            sum2 += inV * inM2;
+            colCnt--;
+        }                       /* while over colCnt */
+        *pO++ = (q15_t) (__SSAT((sum >> out_shift), 16));
+        *pO++ = (q15_t) (__SSAT((sum2 >> out_shift), 16));
+
+        /*adjust the pointers and counters */
+        pB += dim_vec;
+        rowCnt--;
+    }
+
+    /* left-over part of the rows */
+    rowCnt = num_of_rows & 0x1;
+
+    while (rowCnt)
+    {
+        q31_t     sum = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift);
+        uint16_t  colCnt = dim_vec >> 2;
+
+        pA = pV;
+
+        while (colCnt)
+        {
+            q31_t     inV1, inV2, inM11, inM12;
+
+            pB = (q7_t *) read_and_pad((void *)pB, &inM11, &inM12);
+
+            inV1 = *__SIMD32(pA)++;
+            sum = __SMLAD(inV1, inM11, sum);
+
+            inV2 = *__SIMD32(pA)++;
+            sum = __SMLAD(inV2, inM12, sum);
+
+            colCnt--;
+        }
+
+        /* left-over of the vector */
+        colCnt = dim_vec & 0x3;
+        while (colCnt)
+        {
+            q15_t     inV = *pA++;
+            q7_t      inM = *pB++;
+            sum += inV * inM;
+            colCnt--;
+        }
+
+        *pO++ = (q15_t) (__SSAT((sum >> out_shift), 16));
+
+        rowCnt--;
+    }
+
+#else
+    int       i, j;
+    /* Run the following code as reference implementation for Cortex-M0 and Cortex-M3 */
+    for (i = 0; i < num_of_rows; i++)
+    {
+        int       ip_out = ((q31_t)(bias[i]) << bias_shift) + NN_ROUND(out_shift);
+        for (j = 0; j < dim_vec; j++)
+        {
+            ip_out += pV[j] * pM[i * dim_vec + j];
+        }
+        pOut[i] = (q15_t) __SSAT((ip_out >> out_shift), 16);
+    }
+
+#endif                          /* ARM_MATH_DSP */
+
+    /* Return to ARM_MATH_SUCCESS */
+    return (ARM_MATH_SUCCESS);
+
+}
+
+/**
+ * @} end of FC group
+ */
diff --git a/NN/Source/FullyConnectedFunctions/arm_fully_connected_mat_q7_vec_q15_opt.c b/NN/Source/FullyConnectedFunctions/arm_fully_connected_mat_q7_vec_q15_opt.c
new file mode 100644
index 0000000..7be156f
--- /dev/null
+++ b/NN/Source/FullyConnectedFunctions/arm_fully_connected_mat_q7_vec_q15_opt.c
@@ -0,0 +1,403 @@
+/*
+ * Copyright (C) 2010-2018 Arm Limited or its affiliates. All rights reserved.
+ *
+ * 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.
+ */
+
+/* ----------------------------------------------------------------------
+ * Project:      CMSIS NN Library
+ * Title:        arm_fully_connected_mat_q7_vec_q15_opt.c
+ * Description:  Mixed Q15-Q7 opt fully-connected layer function
+ *
+ * $Date:        17. January 2018
+ * $Revision:    V.1.0.0
+ *
+ * Target Processor:  Cortex-M cores
+ *
+ * -------------------------------------------------------------------- */
+
+#include "arm_math.h"
+#include "arm_nnfunctions.h"
+
+/**
+ *  @ingroup groupNN
+ */
+
+/**
+ * @addtogroup FC
+ * @{
+ */
+
+  /**
+   * @brief Mixed Q15-Q7 opt fully-connected layer function
+   * @param[in]       pV          pointer to input vector
+   * @param[in]       pM          pointer to matrix weights
+   * @param[in]       dim_vec     length of the vector
+   * @param[in]       num_of_rows number of rows in weight matrix
+   * @param[in]       bias_shift  amount of left-shift for bias
+   * @param[in]       out_shift   amount of right-shift for output
+   * @param[in]       bias        pointer to bias
+   * @param[in,out]   pOut        pointer to output vector
+   * @param[in,out]   vec_buffer  pointer to buffer space for input
+   * @return     The function returns <code>ARM_MATH_SUCCESS</code>
+   *
+   * @details
+   *
+   * <b>Buffer size:</b>
+   *
+   * vec_buffer size: 0
+   *
+   *  Q7_Q15 version of the fully connected layer
+   *
+   *  Weights are in q7_t and Activations are in q15_t
+   *
+   *  Limitation: x4 version requires weight reordering to work
+   *
+   *  Here we use only one pointer to read 4 rows in the weight
+   *  matrix. So if the original q7_t matrix looks like this:
+   *
+   *  | a11 | a12 | a13 | a14 | a15 | a16 | a17 |
+   *
+   *  | a21 | a22 | a23 | a24 | a25 | a26 | a27 |
+   *
+   *  | a31 | a32 | a33 | a34 | a35 | a36 | a37 |
+   *
+   *  | a41 | a42 | a43 | a44 | a45 | a46 | a47 |
+   *
+   *  | a51 | a52 | a53 | a54 | a55 | a56 | a57 |
+   *
+   *  | a61 | a62 | a63 | a64 | a65 | a66 | a67 |
+   *
+   *  We operates on multiple-of-4 rows, so the first four rows becomes
+   *
+   *  | a11 | a21 | a12 | a22 | a31 | a41 | a32 | a42 |
+   *
+   *  | a13 | a23 | a14 | a24 | a33 | a43 | a34 | a44 |
+   *
+   *  | a15 | a25 | a16 | a26 | a35 | a45 | a36 | a46 |
+   *
+   *  The column left over will be in-order.
+   *  which is:
+   *  | a17 | a27 | a37 | a47 |
+   *
+   *  For the left-over rows, we do 1x1 computation, so the data remains
+   *  as its original order. 
+   *
+   *  So the stored weight matrix looks like this:
+   *
+   *  | a11 | a21 | a12 | a22 | a31 | a41 |
+   *
+   *  | a32 | a42 | a13 | a23 | a14 | a24 |
+   *
+   *  | a33 | a43 | a34 | a44 | a15 | a25 |
+   *
+   *  | a16 | a26 | a35 | a45 | a36 | a46 |
+   *
+   *  | a17 | a27 | a37 | a47 | a51 | a52 |
+   *
+   *  | a53 | a54 | a55 | a56 | a57 | a61 |
+   *
+   *  | a62 | a63 | a64 | a65 | a66 | a67 |
+   *
+   */
+
+arm_status
+arm_fully_connected_mat_q7_vec_q15_opt(const q15_t * pV,
+                                       const q7_t * pM,
+                                       const uint16_t dim_vec,
+                                       const uint16_t num_of_rows,
+                                       const uint16_t bias_shift,
+                                       const uint16_t out_shift, const q7_t * bias, q15_t * pOut, q15_t * vec_buffer)
+{
+
+#if defined (ARM_MATH_DSP)
+    /* Run the following code for Cortex-M4 and Cortex-M7 */
+
+    const q7_t *pB = pM;
+    q15_t    *pO = pOut;
+    const q7_t *pBias = bias;
+    const q15_t *pA = pV;
+
+    uint16_t  rowCnt = num_of_rows >> 2;
+
+    while (rowCnt)
+    {
+        q31_t     sum =  ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift);
+        q31_t     sum2 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift);
+        q31_t     sum3 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift);
+        q31_t     sum4 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift);
+
+        uint16_t  colCnt = dim_vec >> 1;
+
+        pA = pV;
+
+#ifdef USE_INTRINSIC
+
+#ifndef ARM_MATH_BIG_ENDIAN
+
+        while (colCnt)
+        {
+            q31_t     inM11, inM12, inM13, inM14;
+            q31_t     inV;
+
+            inV = *__SIMD32(pA)++;
+            inM11 = *__SIMD32(pB)++;
+            inM12 = __SXTB16(__ROR(inM11, 8));
+            inM11 = __SXTB16(inM11);
+            sum = __SMLAD(inM11, inV, sum);
+            sum2 = __SMLAD(inM12, inV, sum2);
+            inM13 = *__SIMD32(pB)++;
+            inM14 = __SXTB16(__ROR(inM13, 8));
+            inM13 = __SXTB16(inM13);
+            sum3 = __SMLAD(inM13, inV, sum3);
+            sum4 = __SMLAD(inM14, inV, sum4);
+            colCnt--;
+        }
+
+#else
+
+        while (colCnt)
+        {
+            q31_t     inM11, inM12, inM13, inM14;
+            q31_t     inV;
+
+            inV = *__SIMD32(pA)++;
+            inM11 = *__SIMD32(pB)++;
+            inM12 = __SXTB16(__ROR(inM11, 8));
+            inM11 = __SXTB16(inM11);
+            sum = __SMLAD(inM12, inV, sum);
+            sum2 = __SMLAD(inM11, inV, sum2);
+            inM13 = *__SIMD32(pB)++;
+            inM14 = __SXTB16(__ROR(inM13, 8));
+            inM13 = __SXTB16(inM13);
+            sum3 = __SMLAD(inM14, inV, sum3);
+            sum4 = __SMLAD(inM13, inV, sum4);
+            colCnt--;
+        }
+
+#endif                          /* ARM_MATH_BIG_ENDIAN */
+
+#else
+
+        /*
+         * register needed:
+         * loop counter: colCnt
+         * accumulators: sum, sum2, sum3, sum4
+         * pointers: pB, pA
+         * weight data: inM11, inM12, inM13, inM14
+         * activation data: inV
+         */
+
+#ifndef ARM_MATH_BIG_ENDIAN
+        asm volatile ("COL_LOOP_%=:\n"
+                      "ldr.w r4, [%[pA]], #4\n"
+                      "ldr.w r1, [%[pB]], #8\n"
+                      "mov.w r0, r1, ror #8\n"
+                      "sxtb16 r0, r0\n"
+                      "sxtb16 r1, r1\n"
+                      "smlad %[sum], r4, r1, %[sum]\n"
+                      "smlad %[sum2], r4, r0, %[sum2]\n"
+                      "ldr.w r3, [%[pB], #-4]\n"
+                      "mov.w r2, r3, ror #8\n"
+                      "sxtb16 r2, r2\n"
+                      "sxtb16 r3, r3\n"
+                      "smlad %[sum3], r4, r3, %[sum3]\n"
+                      "smlad %[sum4], r4, r2, %[sum4]\n"
+                      "subs %[colCnt], #1\n"
+                      "bne COL_LOOP_%=\n":[sum] "+r"(sum),
+                      [sum2] "+r"(sum2),[sum3] "+r"(sum3),
+                      [sum4] "+r"(sum4),[pB] "+r"(pB),[pA] "+r"(pA):[colCnt] "r"(colCnt):"r0", "r1", "r2", "r3", "r4");
+#else
+        asm volatile ("COL_LOOP_%=:\n"
+                      "ldr.w r4, [%[pA]], #4\n"
+                      "ldr.w r1, [%[pB]], #8\n"
+                      "mov.w r0, r1, ror #8\n"
+                      "sxtb16 r0, r0\n"
+                      "sxtb16 r1, r1\n"
+                      "smlad %[sum], r4, r0, %[sum]\n"
+                      "smlad %[sum2], r4, r1, %[sum2]\n"
+                      "ldr.w r3, [%[pB], #-4]\n"
+                      "mov.w r2, r3, ror #8\n"
+                      "sxtb16 r2, r2\n"
+                      "sxtb16 r3, r3\n"
+                      "smlad %[sum3], r4, r2, %[sum3]\n"
+                      "smlad %[sum4], r4, r3, %[sum4]\n"
+                      "subs %[colCnt], #1\n"
+                      "bne COL_LOOP_%=\n":[sum] "+r"(sum),
+                      [sum2] "+r"(sum2),[sum3] "+r"(sum3),
+                      [sum4] "+r"(sum4),[pB] "+r"(pB),[pA] "+r"(pA):[colCnt] "r"(colCnt):"r0", "r1", "r2", "r3", "r4");
+#endif                          /* ARM_MATH_BIG_ENDIAN */
+
+#endif                          /* USE_INTRINSIC */
+
+        colCnt = dim_vec & 0x1;
+        while (colCnt)
+        {
+            q15_t     inV = *pA++;
+            q7_t      inM = *pB++;
+            q7_t      inM2 = *pB++;
+            q7_t      inM3 = *pB++;
+            q7_t      inM4 = *pB++;
+
+            sum += inV * inM;
+            sum2 += inV * inM2;
+            sum3 += inV * inM3;
+            sum4 += inV * inM4;
+            colCnt--;
+        }                       /* while over colCnt */
+        *pO++ = (q15_t) (__SSAT((sum >> out_shift), 16));
+        *pO++ = (q15_t) (__SSAT((sum2 >> out_shift), 16));
+        *pO++ = (q15_t) (__SSAT((sum3 >> out_shift), 16));
+        *pO++ = (q15_t) (__SSAT((sum4 >> out_shift), 16));
+
+        /* adjust the pointers and counters */
+        rowCnt--;
+    }
+
+    /* left-over part of the rows */
+    rowCnt = num_of_rows & 0x3;
+
+    while (rowCnt)
+    {
+        q31_t     sum = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift);
+
+        uint16_t  colCnt = dim_vec >> 2;
+
+        pA = pV;
+
+        while (colCnt)
+        {
+            q31_t     inV1, inV2, inM11, inM12;
+
+            pB = (q7_t *) read_and_pad((void *)pB, &inM11, &inM12);
+
+            inV1 = *__SIMD32(pA)++;
+            sum = __SMLAD(inV1, inM11, sum);
+
+            inV2 = *__SIMD32(pA)++;
+            sum = __SMLAD(inV2, inM12, sum);
+
+            colCnt--;
+        }
+
+        /* left-over of the vector */
+        colCnt = dim_vec & 0x3;
+        while (colCnt)
+        {
+            q15_t     inV = *pA++;
+            q7_t      inM = *pB++;
+            sum += inV * inM;
+            colCnt--;
+        }
+
+        *pO++ = (q15_t) (__SSAT((sum >> out_shift), 16));
+
+        rowCnt--;
+    }
+
+#else
+    /* Run the following code as reference implementation for Cortex-M0 and Cortex-M3 */
+    uint16_t  rowCnt = num_of_rows >> 2;
+    const q7_t *pB = pM;
+    const q15_t *pA;
+    q15_t    *pO = pOut;
+    const q7_t *pBias = bias;
+
+    while (rowCnt)
+    {
+        q31_t     sum =  ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift);
+        q31_t     sum2 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift);
+        q31_t     sum3 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); 
+        q31_t     sum4 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); 
+        uint16_t  colCnt = dim_vec >> 1;
+
+        pA = pV;
+
+        while (colCnt)
+        {
+            q15_t     inA1 = *pA++;
+            q15_t     inA2 = *pA++;
+
+            q7_t      inB1 = *pB++;
+            q7_t      inB3 = *pB++;
+            q7_t      inB2 = *pB++;
+            q7_t      inB4 = *pB++;
+
+            sum += inA1 * inB1 + inA2 * inB2;
+            sum2 += inA1 * inB3 + inA2 * inB4;
+
+            inB1 = *pB++;
+            inB3 = *pB++;
+            inB2 = *pB++;
+            inB4 = *pB++;
+
+            sum3 += inA1 * inB1 + inA2 * inB2;
+            sum4 += inA1 * inB3 + inA2 * inB4;
+
+            colCnt--;
+        }
+
+        colCnt = dim_vec & 0x1;
+        while (colCnt)
+        {
+            q15_t     inA = *pA++;
+            q7_t      inB = *pB++;
+            sum += inA * inB;
+            inB = *pB++;
+            sum2 += inA * inB;
+            inB = *pB++;
+            sum3 += inA * inB;
+            inB = *pB++;
+            sum4 += inA * inB;
+
+            colCnt--;
+        }
+        *pO++ = (q15_t) __SSAT((sum >> out_shift), 16);
+        *pO++ = (q15_t) __SSAT((sum2 >> out_shift), 16);
+        *pO++ = (q15_t) __SSAT((sum3 >> out_shift), 16);
+        *pO++ = (q15_t) __SSAT((sum4 >> out_shift), 16);
+
+        rowCnt--;
+    }
+
+    rowCnt = num_of_rows & 0x3;
+
+    while (rowCnt)
+    {
+        int       ip_out = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift);
+        int       j;
+
+        pA = pV;
+        for (j = 0; j < dim_vec; j++)
+        {
+            q15_t     inA = *pA++;
+            q7_t      inB = *pB++;
+            ip_out += inA * inB;
+        }
+        *pO++ = (q15_t) __SSAT((ip_out >> out_shift), 16);
+
+        rowCnt--;
+    }
+
+#endif                          /* ARM_MATH_DSP */
+
+    /* Return to ARM_MATH_SUCCESS */
+    return (ARM_MATH_SUCCESS);
+
+}
+
+/**
+ * @} end of FC group
+ */
diff --git a/NN/Source/FullyConnectedFunctions/arm_fully_connected_q15.c b/NN/Source/FullyConnectedFunctions/arm_fully_connected_q15.c
new file mode 100644
index 0000000..c3e7cf2
--- /dev/null
+++ b/NN/Source/FullyConnectedFunctions/arm_fully_connected_q15.c
@@ -0,0 +1,193 @@
+/*
+ * Copyright (C) 2010-2018 Arm Limited or its affiliates. All rights reserved.
+ *
+ * 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.
+ */
+
+/* ----------------------------------------------------------------------
+ * Project:      CMSIS NN Library
+ * Title:        arm_fully_connected_q15.c
+ * Description:  Q15 basic fully-connected layer function
+ *
+ * $Date:        17. January 2018
+ * $Revision:    V.1.0.0
+ *
+ * Target Processor:  Cortex-M cores
+ *
+ * -------------------------------------------------------------------- */
+
+#include "arm_math.h"
+#include "arm_nnfunctions.h"
+
+/**
+ *  @ingroup groupNN
+ */
+
+/**
+ * @addtogroup FC
+ * @{
+ */
+
+  /**
+   * @brief Q15 opt fully-connected layer function
+   * @param[in]       pV          pointer to input vector
+   * @param[in]       pM          pointer to matrix weights
+   * @param[in]       dim_vec     length of the vector
+   * @param[in]       num_of_rows number of rows in weight matrix
+   * @param[in]       bias_shift  amount of left-shift for bias
+   * @param[in]       out_shift   amount of right-shift for output
+   * @param[in]       bias        pointer to bias
+   * @param[in,out]   pOut        pointer to output vector
+   * @param[in,out]   vec_buffer  pointer to buffer space for input
+   * @return     The function returns <code>ARM_MATH_SUCCESS</code>
+   *
+   *
+   * @details
+   *
+   * <b>Buffer size:</b>
+   *
+   * vec_buffer size: 0
+   *
+   */
+
+arm_status
+arm_fully_connected_q15(const q15_t * pV,
+                        const q15_t * pM,
+                        const uint16_t dim_vec,
+                        const uint16_t num_of_rows,
+                        const uint16_t bias_shift,
+                        const uint16_t out_shift, 
+                        const q15_t * bias, 
+                        q15_t * pOut,
+                        q15_t * vec_buffer)
+{
+
+#if defined (ARM_MATH_DSP)
+    /* Run the following code for Cortex-M4 and Cortex-M7 */
+
+    const q15_t *pB = pM;
+    const q15_t *pB2 = pB + dim_vec;
+    q15_t    *pO = pOut;
+    const q15_t    *pA;
+    const q15_t    *pBias = bias;
+    uint16_t rowCnt = num_of_rows >> 1;
+
+    /* this loop loops over different output */
+    while (rowCnt) {
+        q31_t     sum =  ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift);
+        q31_t     sum2 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift);
+
+        uint16_t  colCnt = dim_vec >> 2;
+
+        pA = pV;
+        pB2 = pB + dim_vec;
+
+        while (colCnt)
+        {
+            q31_t     inV1, inM1, inM2;
+            inV1 = *__SIMD32(pA)++;
+            inM1 = *__SIMD32(pB)++;
+            sum = __SMLAD(inV1, inM1, sum);
+            inM2 = *__SIMD32(pB2)++;
+            sum2 = __SMLAD(inV1, inM2, sum2);
+
+            inV1 = *__SIMD32(pA)++;
+            inM1 = *__SIMD32(pB)++;
+            sum = __SMLAD(inV1, inM1, sum);
+            inM2 = *__SIMD32(pB2)++;
+            sum2 = __SMLAD(inV1, inM2, sum2);
+
+            colCnt--;
+        }
+        colCnt = dim_vec & 0x3;
+        while (colCnt)
+        {
+            q15_t     inV = *pA++;
+            q15_t     inM = *pB++;
+            q15_t     inM2 = *pB2++;
+
+            sum += inV * inM;
+            sum2 += inV * inM2;
+            colCnt--;
+        }                       /* while over colCnt */
+        *pO++ =  (q15_t) (__SSAT((sum >> out_shift), 16));
+        *pO++ = (q15_t) (__SSAT((sum2>> out_shift), 16));
+		
+        /* adjust the pointers and counters */
+        pB = pB + dim_vec;
+        rowCnt --;
+    }
+
+    rowCnt = num_of_rows & 0x1;
+
+    while (rowCnt) {
+        q31_t     sum = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift);
+
+        uint16_t  colCnt = dim_vec >> 2;
+
+        pA = pV;
+      
+        while (colCnt) {
+            q31_t     inV1, inM1;
+            inV1 = *__SIMD32(pA)++;
+            inM1 = *__SIMD32(pB)++;
+            sum = __SMLAD(inV1, inM1, sum);
+            
+            inV1 = *__SIMD32(pA)++;
+            inM1 = *__SIMD32(pB)++;
+            sum = __SMLAD(inV1, inM1, sum);
+				
+            colCnt--;
+	}
+			
+	/* left-over of the vector */
+	colCnt = dim_vec & 0x3;
+	while(colCnt) {
+            q15_t     inV = *pA++;
+            q15_t     inM = *pB++;
+
+            sum += inV * inM;
+
+            colCnt--;
+	}
+
+        *pO++ =  (q15_t) (__SSAT((sum >> out_shift), 16));
+			
+        rowCnt --;
+    }
+
+#else
+    int       i, j;
+    /* Run the following code as reference implementation for Cortex-M0 and Cortex-M3 */
+    for (i = 0; i < num_of_rows; i++)
+    {
+        int       ip_out = ((q31_t)(bias[i]) << bias_shift) + NN_ROUND(out_shift);
+        for (j = 0; j < dim_vec; j++)
+        {
+            ip_out += pV[j] * pM[i * dim_vec + j];
+        }
+        pOut[i] = (q15_t) __SSAT((ip_out >> out_shift), 16);
+    }
+
+#endif                          /* ARM_MATH_DSP */
+
+    /* Return to application */
+    return (ARM_MATH_SUCCESS);
+
+}
+
+/**
+ * @} end of FC group
+ */
diff --git a/NN/Source/FullyConnectedFunctions/arm_fully_connected_q15_opt.c b/NN/Source/FullyConnectedFunctions/arm_fully_connected_q15_opt.c
new file mode 100644
index 0000000..f7a3915
--- /dev/null
+++ b/NN/Source/FullyConnectedFunctions/arm_fully_connected_q15_opt.c
@@ -0,0 +1,332 @@
+/*
+ * Copyright (C) 2010-2018 Arm Limited or its affiliates. All rights reserved.
+ *
+ * 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.
+ */
+
+/* ----------------------------------------------------------------------
+ * Project:      CMSIS NN Library
+ * Title:        arm_fully_connected_q15_opt.c
+ * Description:  Q15 opt fully-connected layer function
+ *
+ * $Date:        17. January 2018
+ * $Revision:    V.1.0.0
+ *
+ * Target Processor:  Cortex-M cores
+ *
+ * -------------------------------------------------------------------- */
+
+#include "arm_math.h"
+#include "arm_nnfunctions.h"
+
+/**
+ *  @ingroup groupNN
+ */
+
+/**
+ * @addtogroup FC
+ * @{
+ */
+
+  /**
+   * @brief Q15 opt fully-connected layer function
+   * @param[in]       pV          pointer to input vector
+   * @param[in]       pM          pointer to matrix weights
+   * @param[in]       dim_vec     length of the vector
+   * @param[in]       num_of_rows number of rows in weight matrix
+   * @param[in]       bias_shift  amount of left-shift for bias
+   * @param[in]       out_shift   amount of right-shift for output
+   * @param[in]       bias        pointer to bias
+   * @param[in,out]   pOut        pointer to output vector
+   * @param[in,out]   vec_buffer  pointer to buffer space for input
+   * @return     The function returns <code>ARM_MATH_SUCCESS</code>
+   *
+   *
+   * @details
+   *
+   * <b>Buffer size:</b>
+   *
+   * vec_buffer size: 0
+   *
+   *  Here we use only one pointer to read 4 rows in the weight
+   *  matrix. So if the original matrix looks like this:
+   *
+   *  | a11 | a12 | a13 |
+   *
+   *  | a21 | a22 | a23 |
+   *
+   *  | a31 | a32 | a33 |
+   *
+   *  | a41 | a42 | a43 |
+   *
+   *  | a51 | a52 | a53 |
+   *
+   *  | a61 | a62 | a63 |
+   *
+   *  We operates on multiple-of-4 rows, so the first four rows becomes
+   *
+   *  | a11 | a12 | a21 | a22 | a31 | a32 | a41 | a42 |
+   *
+   *  | a13 | a23 | a33 | a43 |
+   *
+   *  Remaining rows are kept the same original order.
+   *
+   *  So the stored weight matrix looks like this:
+   *
+   *
+   *  | a11 | a12 | a21 | a22 | a31 | a32 | a41 | a42 |
+   *
+   *  | a13 | a23 | a33 | a43 | a51 | a52 | a53 | a61 |
+   *
+   *  | a62 | a63 |
+   */
+
+arm_status
+arm_fully_connected_q15_opt(const q15_t * pV,
+                            const q15_t * pM,
+                            const uint16_t dim_vec,
+                            const uint16_t num_of_rows,
+                            const uint16_t bias_shift,
+                            const uint16_t out_shift, 
+                            const q15_t * bias, 
+                            q15_t * pOut, 
+                            q15_t * vec_buffer)
+{
+
+#if defined (ARM_MATH_DSP)
+    /* Run the following code for Cortex-M4 and Cortex-M7 */
+
+    const q15_t *pB = pM;
+    q15_t    *pO = pOut;
+    const q15_t *pBias = bias;
+    const q15_t *pA = pV;
+
+    uint16_t  rowCnt = num_of_rows >> 2;
+
+    while (rowCnt)
+    {
+        q31_t     sum =  ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift);
+        q31_t     sum2 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); 
+        q31_t     sum3 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); 
+        q31_t     sum4 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); 
+
+        uint16_t  colCnt = dim_vec >> 1;
+
+        pA = pV;
+
+#ifdef USE_INTRINSIC
+
+        while (colCnt)
+        {
+            q31_t     inM11, inM12, inM13, inM14;
+            q31_t     inV;
+
+            inV = *__SIMD32(pA)++;
+            inM11 = *__SIMD32(pB)++;
+            sum = __SMLAD(inV, inM11, sum);
+            inM12 = *__SIMD32(pB)++;
+            sum2 = __SMLAD(inV, inM12, sum2);
+            inM13 = *__SIMD32(pB)++;
+            sum3 = __SMLAD(inV, inM13, sum3);
+            inM14 = *__SIMD32(pB)++;
+            sum4 = __SMLAD(inV, inM14, sum4);
+            colCnt--;
+        }
+
+#else
+
+        /*
+         * register needed:
+         * loop counter: colCnt
+         * accumulators: sum, sum2, sum3, sum4
+         * pointers: pB, pA
+         * weight data: inM11, inM12, inM13, inM14
+         * activation data: inV
+         */
+
+        asm volatile ("COL_LOOP_%=:\n"
+                      "ldr.w r4, [%[pA]], #4\n"
+                      "ldr.w r0, [%[pB]], #16\n"
+                      "smlad %[sum], r4, r0, %[sum]\n"
+                      "ldr.w r1, [%[pB] , #-12]\n"
+                      "smlad %[sum2], r4, r1, %[sum2]\n"
+                      "ldr.w r2, [%[pB] , #-8]\n"
+                      "smlad %[sum3], r4, r2, %[sum3]\n"
+                      "ldr.w r3, [%[pB] , #-4]\n"
+                      "smlad %[sum4], r4, r3, %[sum4]\n"
+                      "subs %[colCnt], #1\n"
+                      "bne COL_LOOP_%=\n":[sum] "+r"(sum),
+                      [sum2] "+r"(sum2),[sum3] "+r"(sum3),
+                      [sum4] "+r"(sum4),[pB] "+r"(pB),[pA] "+r"(pA):[colCnt] "r"(colCnt):"r0", "r1", "r2", "r3", "r4");
+
+#endif                          /* USE_INTRINSIC */
+
+        colCnt = dim_vec & 0x1;
+        while (colCnt)
+        {
+
+            q15_t     inV = *pA++;
+            q15_t     inM = *pB++;
+            q15_t     inM2 = *pB++;
+            q15_t     inM3 = *pB++;
+            q15_t     inM4 = *pB++;
+
+            sum += inV * inM;
+            sum2 += inV * inM2;
+            sum3 += inV * inM3;
+            sum4 += inV * inM4;
+            colCnt--;
+        }                       /* while over colCnt */
+        *pO++ = (q15_t) (__SSAT((sum >> out_shift), 16));
+        *pO++ = (q15_t) (__SSAT((sum2 >> out_shift), 16));
+        *pO++ = (q15_t) (__SSAT((sum3 >> out_shift), 16));
+        *pO++ = (q15_t) (__SSAT((sum4 >> out_shift), 16));
+
+        /* adjust the pointers and counters */
+        rowCnt--;
+    }
+
+    /* left-over part of the rows */
+    rowCnt = num_of_rows & 0x3;
+
+    while (rowCnt)
+    {
+        q31_t     sum = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift);
+
+        uint16_t  colCnt = dim_vec >> 2;
+
+        pA = pV;
+
+        while (colCnt)
+        {
+            q31_t     inV1, inV2, inM1, inM2;
+
+            inM1 = *__SIMD32(pB)++;
+            inV1 = *__SIMD32(pA)++;
+            sum = __SMLAD(inV1, inM1, sum);
+
+            inM2 = *__SIMD32(pB)++;
+            inV2 = *__SIMD32(pA)++;
+            sum = __SMLAD(inV2, inM2, sum);
+
+            colCnt--;
+        }
+
+        /* left-over of the vector */
+        colCnt = dim_vec & 0x3;
+        while (colCnt)
+        {
+            q15_t     inV = *pA++;
+            q15_t     inM = *pB++;
+            sum += inV * inM;
+            colCnt--;
+        }
+
+        *pO++ = (q15_t) (__SSAT((sum >> out_shift), 16));
+
+        rowCnt--;
+    }
+
+#else
+    /* Run the following code as reference implementation for Cortex-M0 and Cortex-M3 */
+    uint16_t  rowCnt = num_of_rows >> 2;
+    const q15_t *pB = pM;
+    const q15_t *pA;
+    q15_t    *pO = pOut;
+    const q15_t *pBias = bias;
+
+    while (rowCnt)
+    {
+        q31_t     sum =  ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift);
+        q31_t     sum2 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift);
+        q31_t     sum3 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift);
+        q31_t     sum4 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift);
+
+        uint16_t  colCnt = dim_vec >> 1;
+
+        pA = pV;
+        while (colCnt)
+        {
+            q15_t     inA1 = *pA++;
+            q15_t     inA2 = *pA++;
+
+            q15_t     inB1 = *pB++;
+            q15_t     inB2 = *pB++;
+            sum += inA1 * inB1 + inA2 * inB2;
+
+            inB1 = *pB++;
+            inB2 = *pB++;
+            sum2 += inA1 * inB1 + inA2 * inB2;
+
+            inB1 = *pB++;
+            inB2 = *pB++;
+            sum3 += inA1 * inB1 + inA2 * inB2;
+
+            inB1 = *pB++;
+            inB2 = *pB++;
+            sum4 += inA1 * inB1 + inA2 * inB2;
+
+            colCnt--;
+        }
+        colCnt = dim_vec & 0x1;
+        while (colCnt)
+        {
+            q15_t     inA = *pA++;
+            q15_t     inB = *pB++;
+            sum += inA * inB;
+            inB = *pB++;
+            sum2 += inA * inB;
+            inB = *pB++;
+            sum3 += inA * inB;
+            inB = *pB++;
+            sum4 += inA * inB;
+            colCnt--;
+        }
+        *pO++ = (q15_t) __SSAT((sum >> out_shift), 16);
+        *pO++ = (q15_t) __SSAT((sum2 >> out_shift), 16);
+        *pO++ = (q15_t) __SSAT((sum3 >> out_shift), 16);
+        *pO++ = (q15_t) __SSAT((sum4 >> out_shift), 16);
+
+        rowCnt--;
+    }
+    rowCnt = num_of_rows & 0x3;
+
+    while (rowCnt)
+    {
+        int       ip_out = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift);
+        int       j;
+
+        pA = pV;
+        for (j = 0; j < dim_vec; j++)
+        {
+            q15_t     inA = *pA++;
+            q15_t     inB = *pB++;
+            ip_out += inA * inB;
+        }
+        *pO++ = (q15_t) __SSAT((ip_out >> out_shift), 16);
+
+        rowCnt--;
+    }
+
+#endif                          /* ARM_MATH_DSP */
+
+    /* Return to ARM_MATH_SUCCESS */
+    return (ARM_MATH_SUCCESS);
+
+}
+
+/**
+ * @} end of FC group
+ */
diff --git a/NN/Source/FullyConnectedFunctions/arm_fully_connected_q7.c b/NN/Source/FullyConnectedFunctions/arm_fully_connected_q7.c
new file mode 100644
index 0000000..d8efc04
--- /dev/null
+++ b/NN/Source/FullyConnectedFunctions/arm_fully_connected_q7.c
@@ -0,0 +1,198 @@
+/*
+ * Copyright (C) 2010-2018 Arm Limited or its affiliates. All rights reserved.
+ *
+ * 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.
+ */
+
+/* ----------------------------------------------------------------------
+ * Project:      CMSIS NN Library
+ * Title:        arm_fully_connected_q7.c
+ * Description:  Q7 basic fully-connected layer function
+ *
+ * $Date:        17. January 2018
+ * $Revision:    V.1.0.0
+ *
+ * Target Processor:  Cortex-M cores
+ *
+ * -------------------------------------------------------------------- */
+
+#include "arm_math.h"
+#include "arm_nnfunctions.h"
+
+/**
+ *  @ingroup groupNN
+ */
+
+/**
+ * @addtogroup FC
+ * @{
+ */
+
+  /**
+   * @brief Q7 basic fully-connected layer function
+   * @param[in]       pV          pointer to input vector
+   * @param[in]       pM          pointer to matrix weights
+   * @param[in]       dim_vec     length of the vector
+   * @param[in]       num_of_rows number of rows in weight matrix
+   * @param[in]       bias_shift  amount of left-shift for bias
+   * @param[in]       out_shift   amount of right-shift for output
+   * @param[in]       bias        pointer to bias
+   * @param[in,out]   pOut        pointer to output vector
+   * @param[in,out]   vec_buffer  pointer to buffer space for input
+   * @return     The function returns <code>ARM_MATH_SUCCESS</code>
+   *
+   * @details
+   *
+   * <b>Buffer size:</b>
+   *
+   * vec_buffer size: dim_vec
+   *
+   * This basic function is designed to work with regular weight
+   * matrix without interleaving.
+   *
+   */
+
+arm_status
+arm_fully_connected_q7(const q7_t * pV,
+                       const q7_t * pM,
+                       const uint16_t dim_vec,
+                       const uint16_t num_of_rows,
+                       const uint16_t bias_shift,
+                       const uint16_t out_shift, const q7_t * bias, q7_t * pOut, q15_t * vec_buffer)
+{
+
+#if defined (ARM_MATH_DSP)
+    /* Run the following code for Cortex-M4 and Cortex-M7 */
+
+    const q7_t *pB = pM;
+    const q7_t *pB2;
+    q7_t     *pO = pOut;
+    const q7_t *pBias = bias;
+    q15_t    *pA;
+    uint16_t  rowCnt = num_of_rows >> 1;
+
+    /* expand the vector into the buffer */
+    arm_q7_to_q15_reordered_no_shift(pV, vec_buffer, dim_vec);
+
+    while (rowCnt)
+    {
+        q31_t     sum =  ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift);
+        q31_t     sum2 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift);
+        uint16_t  colCnt = dim_vec >> 2;
+
+        pA = vec_buffer;
+        pB2 = pB + dim_vec;
+
+        while (colCnt)
+        {
+            q31_t     inV, inM11, inM12, inM21, inM22;
+            pB = (q7_t *) read_and_pad_reordered((void *)pB, &inM11, &inM12);
+            pB2 = (q7_t *) read_and_pad_reordered((void *)pB2, &inM21, &inM22);
+
+            inV = *__SIMD32(pA)++;
+
+            sum = __SMLAD(inV, inM11, sum);
+            sum2 = __SMLAD(inV, inM21, sum2);
+
+            inV = *__SIMD32(pA)++;
+
+            sum = __SMLAD(inV, inM12, sum);
+            sum2 = __SMLAD(inV, inM22, sum2);
+
+            colCnt--;
+        }
+        colCnt = dim_vec & 0x3;
+        while (colCnt)
+        {
+            q7_t      inV = *pA++;
+            q15_t     inM = *pB++;
+            q15_t     inM2 = *pB2++;
+
+            sum += inV * inM;
+            sum2 += inV * inM2;
+            colCnt--;
+        }                       /* while over colCnt */
+        *pO++ = (q7_t) (__SSAT((sum >> out_shift), 8));
+        *pO++ = (q7_t) (__SSAT((sum2 >> out_shift), 8));
+
+        /* adjust the pointers and counters */
+        pB += dim_vec;
+        rowCnt--;
+    }
+
+    /* left-over part of the rows */
+    rowCnt = num_of_rows & 0x1;
+
+    while (rowCnt)
+    {
+        uint16_t  colCnt = dim_vec >> 2;
+        q31_t     sum = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift);
+
+        pA = vec_buffer;
+
+        while (colCnt)
+        {
+            q31_t     inV1, inV2, inM11, inM12;
+
+            pB = (q7_t *) read_and_pad_reordered((void *)pB, &inM11, &inM12);
+
+            inV1 = *__SIMD32(pA)++;
+            sum = __SMLAD(inV1, inM11, sum);
+
+            inV2 = *__SIMD32(pA)++;
+            sum = __SMLAD(inV2, inM12, sum);
+
+            colCnt--;
+        }
+
+        /* left-over of the vector */
+        colCnt = dim_vec & 0x3;
+        while (colCnt)
+        {
+            q7_t      inV = *pA++;
+            q15_t     inM = *pB++;
+            sum += inV * inM;
+            colCnt--;
+        }
+
+        *pO++ = (q7_t) (__SSAT((sum >> out_shift), 8));
+
+        rowCnt--;
+    }
+
+#else
+    int       i, j;
+
+    /* Run the following code as reference implementation for Cortex-M0 and Cortex-M3 */
+    for (i = 0; i < num_of_rows; i++)
+    {
+        int       ip_out = ((q31_t)(bias[i]) << bias_shift) + NN_ROUND(out_shift);
+        for (j = 0; j < dim_vec; j++)
+        {
+            ip_out += pV[j] * pM[i * dim_vec + j];
+        }
+        pOut[i] = (q7_t) __SSAT((ip_out >> out_shift), 8);
+    }
+
+#endif                          /* ARM_MATH_DSP */
+
+    /* Return to ARM_MATH_SUCCESS */
+    return (ARM_MATH_SUCCESS);
+
+}
+
+/**
+ * @} end of FC group
+ */
diff --git a/NN/Source/FullyConnectedFunctions/arm_fully_connected_q7_opt.c b/NN/Source/FullyConnectedFunctions/arm_fully_connected_q7_opt.c
new file mode 100644
index 0000000..e3d0874
--- /dev/null
+++ b/NN/Source/FullyConnectedFunctions/arm_fully_connected_q7_opt.c
@@ -0,0 +1,484 @@
+/*
+ * Copyright (C) 2010-2018 Arm Limited or its affiliates. All rights reserved.
+ *
+ * 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.
+ */
+
+/* ----------------------------------------------------------------------
+ * Project:      CMSIS NN Library
+ * Title:        arm_fully_connected_q7_opt.c
+ * Description:  Q7 basic fully-connected layer function
+ *
+ * $Date:        17. January 2018
+ * $Revision:    V.1.0.0
+ *
+ * Target Processor:  Cortex-M cores
+ *
+ * -------------------------------------------------------------------- */
+
+#include "arm_math.h"
+#include "arm_nnfunctions.h"
+
+/**
+ *  @ingroup groupNN
+ */
+
+/**
+ * @addtogroup FC
+ * @{
+ */
+
+  /**
+   * @brief Q7 opt fully-connected layer function
+   * @param[in]       pV          pointer to input vector
+   * @param[in]       pM          pointer to matrix weights
+   * @param[in]       dim_vec     length of the vector
+   * @param[in]       num_of_rows number of rows in weight matrix
+   * @param[in]       bias_shift  amount of left-shift for bias
+   * @param[in]       out_shift   amount of right-shift for output
+   * @param[in]       bias        pointer to bias
+   * @param[in,out]   pOut        pointer to output vector
+   * @param[in,out]   vec_buffer  pointer to buffer space for input
+   * @return     The function returns <code>ARM_MATH_SUCCESS</code>
+   *
+   * @details
+   *
+   * <b>Buffer size:</b>
+   *
+   * vec_buffer size: dim_vec
+   *
+   * This opt function is designed to work with interleaved weight
+   * matrix. The vector input is assumed in q7_t format, we call
+   *  arm_q7_to_q15_no_shift_shuffle function to expand into
+   *  q15_t format with certain weight re-ordering, refer to the function
+   *  comments for more details.
+   *  Here we use only one pointer to read 4 rows in the weight
+   *  matrix. So if the original q7_t matrix looks like this:
+   *
+   *  | a11 | a12 | a13 | a14 | a15 | a16 | a17 |
+   *
+   *  | a21 | a22 | a23 | a24 | a25 | a26 | a27 |
+   *
+   *  | a31 | a32 | a33 | a34 | a35 | a36 | a37 |
+   *
+   *  | a41 | a42 | a43 | a44 | a45 | a46 | a47 |
+   *
+   *  | a51 | a52 | a53 | a54 | a55 | a56 | a57 |
+   *
+   *  | a61 | a62 | a63 | a64 | a65 | a66 | a67 |
+   *
+   *
+   *  We operates on multiple-of-4 rows, so the first four rows becomes
+   *
+   *  | a11 | a21 | a13 | a23 | a31 | a41 | a33 | a43 |
+   *
+   *  | a12 | a22 | a14 | a24 | a32 | a42 | a34 | a44 |
+   *
+   *  | a15 | a25 | a35 | a45 | a16 | a26 | a36 | a46 |
+   *
+   *  So within the kernel, we first read the re-ordered vector in as:
+   *
+   *  | b1  | b3  | and | b2  | b4  |
+   *
+   *  the four q31_t weights will look like
+   *
+   *  | a11 | a13 |, | a21 | a23 |, | a31 | a33 |, | a41 | a43 |
+   *
+   *  | a12 | a14 |, | a22 | a24 |, | a32 | a34 |, | a42 | a44 |
+   *
+   *  The column left over will be in-order.
+   *  which is:
+   *
+   *  | a17 | a27 | a37 | a47 |
+   *
+   *  For the left-over rows, we do 1x1 computation, so the data remains
+   *  as its original order. 
+   *
+   *  So the stored weight matrix looks like this:
+   *
+   *  | a11 | a21 | a13 | a23 | a31 | a41 |
+   *
+   *  | a33 | a43 | a12 | a22 | a14 | a24 |
+   *
+   *  | a32 | a42 | a34 | a44 | a15 | a25 |
+   *
+   *  | a35 | a45 | a16 | a26 | a36 | a46 |
+   *
+   *  | a17 | a27 | a37 | a47 | a51 | a52 |
+   *
+   *  | a53 | a54 | a55 | a56 | a57 | a61 |
+   *
+   *  | a62 | a63 | a64 | a65 | a66 | a67 |
+   *
+   *
+   */
+
+arm_status
+arm_fully_connected_q7_opt(const q7_t * pV,
+                           const q7_t * pM,
+                           const uint16_t dim_vec,
+                           const uint16_t num_of_rows,
+                           const uint16_t bias_shift,
+                           const uint16_t out_shift, 
+                           const q7_t * bias, 
+                           q7_t * pOut, 
+                           q15_t * vec_buffer)
+{
+
+#if defined (ARM_MATH_DSP)
+    /* Run the following code for Cortex-M4 and Cortex-M7 */
+
+    const q7_t *pB = pM;
+    q7_t     *pO = pOut;
+    const q7_t *pBias = bias;
+    q15_t    *pA;
+    uint16_t  rowCnt = num_of_rows >> 2;
+
+    arm_q7_to_q15_reordered_no_shift(pV, vec_buffer, dim_vec);
+
+    while (rowCnt)
+    {
+
+        q31_t     sum =  ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift);
+        q31_t     sum2 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift);
+        q31_t     sum3 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift);
+        q31_t     sum4 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift);
+
+        uint16_t  colCnt = dim_vec >> 2;
+
+        pA = vec_buffer;
+
+#ifdef USE_INTRINSIC
+
+#ifndef ARM_MATH_BIG_ENDIAN
+        while (colCnt)
+        {
+            q31_t     inM11, inM12, inM13, inM14;
+            q31_t     inV;
+
+            inV = *__SIMD32(pA)++;
+            inM11 = *__SIMD32(pB)++;
+            inM12 = __SXTB16(__ROR(inM11, 8));
+            inM11 = __SXTB16(inM11);
+            sum = __SMLAD(inM11, inV, sum);
+            sum2 = __SMLAD(inM12, inV, sum2);
+            inM13 = *__SIMD32(pB)++;
+            inM14 = __SXTB16(__ROR(inM13, 8));
+            inM13 = __SXTB16(inM13);
+            sum3 = __SMLAD(inM13, inV, sum3);
+            sum4 = __SMLAD(inM14, inV, sum4);
+
+            inV = *__SIMD32(pA)++;
+            inM11 = *__SIMD32(pB)++;
+            inM12 = __SXTB16(__ROR(inM11, 8));
+            inM11 = __SXTB16(inM11);
+            sum = __SMLAD(inM11, inV, sum);
+            sum2 = __SMLAD(inM12, inV, sum2);
+            inM13 = *__SIMD32(pB)++;
+            inM14 = __SXTB16(__ROR(inM13, 8));
+            inM13 = __SXTB16(inM13);
+            sum3 = __SMLAD(inM13, inV, sum3);
+            sum4 = __SMLAD(inM14, inV, sum4);
+            colCnt--;
+        }
+#else
+        while (colCnt)
+        {
+            q31_t     inM11, inM12, inM13, inM14;
+            q31_t     inV;
+
+            inV = *__SIMD32(pA)++;
+            inM11 = *__SIMD32(pB)++;
+            inM12 = __SXTB16(__ROR(inM11, 8));
+            inM11 = __SXTB16(inM11);
+            sum = __SMLAD(inM12, inV, sum);
+            sum2 = __SMLAD(inM11, inV, sum2);
+            inM13 = *__SIMD32(pB)++;
+            inM14 = __SXTB16(__ROR(inM13, 8));
+            inM13 = __SXTB16(inM13);
+            sum3 = __SMLAD(inM14, inV, sum3);
+            sum4 = __SMLAD(inM13, inV, sum4);
+
+            inV = *__SIMD32(pA)++;
+            inM11 = *__SIMD32(pB)++;
+            inM12 = __SXTB16(__ROR(inM11, 8));
+            inM11 = __SXTB16(inM11);
+            sum = __SMLAD(inM12, inV, sum);
+            sum2 = __SMLAD(inM11, inV, sum2);
+            inM13 = *__SIMD32(pB)++;
+            inM14 = __SXTB16(__ROR(inM13, 8));
+            inM13 = __SXTB16(inM13);
+            sum3 = __SMLAD(inM14, inV, sum3);
+            sum4 = __SMLAD(inM13, inV, sum4);
+            colCnt--;
+        }
+#endif                          /* ARM_MATH_BIG_ENDIAN */
+
+#else
+
+        /*
+         * register needed:
+         * loop counter: colCnt
+         * accumulators: sum, sum2, sum3, sum4
+         * pointers: pB, pA
+         * weight data: inM11, inM12, inM13, inM14
+         * activation data: inV
+         */
+
+#ifndef ARM_MATH_BIG_ENDIAN
+        asm volatile ("COL_LOOP_%=:\n"
+                      "ldr.w r4, [%[pA]], #8\n"
+                      "ldr.w r1, [%[pB]], #16\n"
+                      "mov.w r0, r1, ror #8\n"
+                      "sxtb16 r0, r0\n"
+                      "sxtb16 r1, r1\n"
+                      "smlad %[sum], r4, r1, %[sum]\n"
+                      "smlad %[sum2], r4, r0, %[sum2]\n"
+                      "ldr.w r3, [%[pB], #-12]\n"
+                      "mov.w r2, r3, ror #8\n"
+                      "sxtb16 r2, r2\n"
+                      "sxtb16 r3, r3\n"
+                      "smlad %[sum3], r4, r3, %[sum3]\n"
+                      "smlad %[sum4], r4, r2, %[sum4]\n"
+                      "ldr.w r4, [%[pA], #-4]\n"
+                      "ldr.w r1, [%[pB], #-8]\n"
+                      "mov.w r0, r1, ror #8\n"
+                      "sxtb16 r0, r0\n"
+                      "sxtb16 r1, r1\n"
+                      "smlad %[sum], r4, r1, %[sum]\n"
+                      "smlad %[sum2], r4, r0, %[sum2]\n"
+                      "ldr.w r3, [%[pB], #-4]\n"
+                      "mov.w r2, r3, ror #8\n"
+                      "sxtb16 r2, r2\n"
+                      "sxtb16 r3, r3\n"
+                      "smlad %[sum3], r4, r3, %[sum3]\n"
+                      "smlad %[sum4], r4, r2, %[sum4]\n"
+                      "subs %[colCnt], #1\n"
+                      "bne COL_LOOP_%=\n":[sum] "+r"(sum),
+                      [sum2] "+r"(sum2),[sum3] "+r"(sum3),
+                      [sum4] "+r"(sum4),[pB] "+r"(pB),[pA] "+r"(pA):[colCnt] "r"(colCnt):"r0", "r1", "r2", "r3", "r4");
+#else
+        asm volatile ("COL_LOOP_%=:\n"
+                      "ldr.w r4, [%[pA]], #8\n"
+                      "ldr.w r1, [%[pB]], #16\n"
+                      "mov.w r0, r1, ror #8\n"
+                      "sxtb16 r0, r0\n"
+                      "sxtb16 r1, r1\n"
+                      "smlad %[sum], r4, r0, %[sum]\n"
+                      "smlad %[sum2], r4, r1, %[sum2]\n"
+                      "ldr.w r3, [%[pB], #-12]\n"
+                      "mov.w r2, r3, ror #8\n"
+                      "sxtb16 r2, r2\n"
+                      "sxtb16 r3, r3\n"
+                      "smlad %[sum3], r4, r2, %[sum3]\n"
+                      "smlad %[sum4], r4, r3, %[sum4]\n"
+                      "ldr.w r4, [%[pA], #-4]\n"
+                      "ldr.w r1, [%[pB], #-8]\n"
+                      "mov.w r0, r1, ror #8\n"
+                      "sxtb16 r0, r0\n"
+                      "sxtb16 r1, r1\n"
+                      "smlad %[sum], r4, r0, %[sum]\n"
+                      "smlad %[sum2], r4, r1, %[sum2]\n"
+                      "ldr.w r3, [%[pB], #-4]\n"
+                      "mov.w r2, r3, ror #8\n"
+                      "sxtb16 r2, r2\n"
+                      "sxtb16 r3, r3\n"
+                      "smlad %[sum3], r4, r2, %[sum3]\n"
+                      "smlad %[sum4], r4, r3, %[sum4]\n"
+                      "subs %[colCnt], #1\n"
+                      "bne COL_LOOP_%=\n":[sum] "+r"(sum),
+                      [sum2] "+r"(sum2),[sum3] "+r"(sum3),
+                      [sum4] "+r"(sum4),[pB] "+r"(pB),[pA] "+r"(pA):[colCnt] "r"(colCnt):"r0", "r1", "r2", "r3", "r4");
+#endif                          /* ARM_MATH_BIG_ENDIAN */
+
+#endif                          /* USE_INTRINSIC */
+
+        colCnt = dim_vec & 0x3;
+        while (colCnt)
+        {
+            q15_t     inV = *pA++;
+            q7_t      inM = *pB++;
+            q7_t      inM2 = *pB++;
+            q7_t      inM3 = *pB++;
+            q7_t      inM4 = *pB++;
+
+            sum += inV * inM;
+            sum2 += inV * inM2;
+            sum3 += inV * inM3;
+            sum4 += inV * inM4;
+            colCnt--;
+        }                       /* while over colCnt */
+        *pO++ = (q7_t) (__SSAT((sum >> out_shift), 8));
+        *pO++ = (q7_t) (__SSAT((sum2 >> out_shift), 8));
+        *pO++ = (q7_t) (__SSAT((sum3 >> out_shift), 8));
+        *pO++ = (q7_t) (__SSAT((sum4 >> out_shift), 8));
+
+        /* adjust the pointers and counters */
+        rowCnt--;
+    }
+
+    /* left-over part of the rows */
+    rowCnt = num_of_rows & 0x3;
+
+    while (rowCnt)
+    {
+        q31_t     sum = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift);
+        uint16_t  colCnt = dim_vec >> 2;
+
+        pA = vec_buffer;
+
+        while (colCnt)
+        {
+            q31_t     inV1, inV2, inM11, inM12;
+
+            pB = (q7_t *) read_and_pad_reordered((void *)pB, &inM11, &inM12);
+
+            inV1 = *__SIMD32(pA)++;
+            sum = __SMLAD(inV1, inM11, sum);
+
+            inV2 = *__SIMD32(pA)++;
+            sum = __SMLAD(inV2, inM12, sum);
+
+            colCnt--;
+        }
+
+        /* left-over of the vector */
+        colCnt = dim_vec & 0x3;
+        while (colCnt)
+        {
+            q15_t     inV = *pA++;
+            q7_t      inM = *pB++;
+            sum += inV * inM;
+            colCnt--;
+        }
+
+        *pO++ = (q7_t) (__SSAT((sum >> out_shift), 8));
+
+        rowCnt--;
+    }
+
+#else
+    /* Run the following code as reference implementation for Cortex-M0 and Cortex-M3 */
+    uint16_t  rowCnt = num_of_rows >> 2;
+    const q7_t *pB = pM;
+    const q7_t *pA;
+    q7_t     *pO = pOut;
+    const q7_t *pBias = bias;
+
+    while (rowCnt)
+    {
+        q31_t     sum =  ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift);
+        q31_t     sum2 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift);
+        q31_t     sum3 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift);
+        q31_t     sum4 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift);
+
+        uint16_t  colCnt = dim_vec >> 2;
+
+        pA = pV;
+
+        while (colCnt)
+        {
+            q7_t      inA1 = *pA++;
+            q7_t      inA3 = *pA++;
+            q7_t      inA2 = *pA++;
+            q7_t      inA4 = *pA++;
+
+            q7_t      inB1 = *pB++;
+            q7_t      inB3 = *pB++;
+            q7_t      inB2 = *pB++;
+            q7_t      inB4 = *pB++;
+
+            sum += inA1 * inB1 + inA2 * inB2;
+            sum2 += inA1 * inB3 + inA2 * inB4;
+
+            inB1 = *pB++;
+            inB3 = *pB++;
+            inB2 = *pB++;
+            inB4 = *pB++;
+
+            sum3 += inA1 * inB1 + inA2 * inB2;
+            sum4 += inA1 * inB3 + inA2 * inB4;
+
+            inB1 = *pB++;
+            inB3 = *pB++;
+            inB2 = *pB++;
+            inB4 = *pB++;
+
+            sum += inA3 * inB1 + inA4 * inB2;
+            sum2 += inA3 * inB3 + inA4 * inB4;
+
+            inB1 = *pB++;
+            inB3 = *pB++;
+            inB2 = *pB++;
+            inB4 = *pB++;
+
+            sum3 += inA3 * inB1 + inA4 * inB2;
+            sum4 += inA3 * inB3 + inA4 * inB4;
+
+            colCnt--;
+        }
+        colCnt = dim_vec & 0x3;
+        while (colCnt)
+        {
+            q7_t      inA = *pA++;
+            q7_t      inB = *pB++;
+            sum += inA * inB;
+            inB = *pB++;
+            sum2 += inA * inB;
+            inB = *pB++;
+            sum3 += inA * inB;
+            inB = *pB++;
+            sum4 += inA * inB;
+
+            colCnt--;
+        }
+        *pO++ = (q7_t) __SSAT((sum >> out_shift), 8);
+        *pO++ = (q7_t) __SSAT((sum2 >> out_shift), 8);
+        *pO++ = (q7_t) __SSAT((sum3 >> out_shift), 8);
+        *pO++ = (q7_t) __SSAT((sum4 >> out_shift), 8);
+
+        rowCnt--;
+    }
+
+    rowCnt = num_of_rows & 0x3;
+
+    while (rowCnt)
+    {
+        int       ip_out = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift);
+
+        int       j;
+
+        pA = pV;
+        for (j = 0; j < dim_vec; j++)
+        {
+            q7_t      inA = *pA++;
+            q7_t      inB = *pB++;
+            ip_out += inA * inB;
+        }
+        *pO++ = (q7_t) __SSAT((ip_out >> out_shift), 8);
+
+        rowCnt--;
+    }
+
+#endif                          /* ARM_MATH_DSP */
+
+    /* Return to ARM_MATH_SUCCESS */
+    return (ARM_MATH_SUCCESS);
+
+}
+
+/**
+ * @} end of FC group
+ */
diff --git a/NN/Source/NNSupportFunctions/arm_nn_mult_q15.c b/NN/Source/NNSupportFunctions/arm_nn_mult_q15.c
new file mode 100644
index 0000000..5a60459
--- /dev/null
+++ b/NN/Source/NNSupportFunctions/arm_nn_mult_q15.c
@@ -0,0 +1,147 @@
+/*
+ * Copyright (C) 2010-2018 Arm Limited or its affiliates. All rights reserved.
+ *
+ * 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.
+ */
+
+/* ----------------------------------------------------------------------
+ * Project:      CMSIS NN Library
+ * Title:        arm_nn_mult_q15.c
+ * Description:  Q15 vector multiplication with variable output shifts
+ *
+ * $Date:        13. July 2018
+ * $Revision:    V.1.0.0
+ *
+ * Target Processor:  Cortex-M cores
+ *
+ * -------------------------------------------------------------------- */
+
+#include "arm_nnfunctions.h"
+
+/**    
+ * @ingroup groupSupport    
+ */
+
+/**
+ * @addtogroup NNBasicMath
+ * @{
+ */
+
+
+/**
+ * @brief           Q7 vector multiplication with variable output shifts
+ * @param[in]       *pSrcA        pointer to the first input vector
+ * @param[in]       *pSrcB        pointer to the second input vector
+ * @param[out]      *pDst         pointer to the output vector
+ * @param[in]       out_shift     amount of right-shift for output
+ * @param[in]       blockSize     number of samples in each vector
+ * @return none.
+ *
+ * <b>Scaling and Overflow Behavior:</b>
+ * \par
+ * The function uses saturating arithmetic.
+ * Results outside of the allowable Q15 range [0x8000 0x7FFF] will be saturated.
+ */
+
+void arm_nn_mult_q15(
+  q15_t * pSrcA,
+  q15_t * pSrcB,
+  q15_t * pDst,
+  const uint16_t out_shift,
+  uint32_t blockSize)
+{
+  uint32_t blkCnt;                               /* loop counters */
+
+#if defined (ARM_MATH_DSP)
+
+/* Run the below code for Cortex-M4 and Cortex-M3 */
+  q31_t inA1, inA2, inB1, inB2;                  /* temporary input variables */
+  q15_t out1, out2, out3, out4;                  /* temporary output variables */
+  q31_t mul1, mul2, mul3, mul4;                  /* temporary variables */
+
+  /* loop Unrolling */
+  blkCnt = blockSize >> 2U;
+
+  /* First part of the processing with loop unrolling.  Compute 4 outputs at a time.
+   ** a second loop below computes the remaining 1 to 3 samples. */
+  while (blkCnt > 0U)
+  {
+    /* read two samples at a time from sourceA */
+    inA1 = *__SIMD32(pSrcA)++;
+    /* read two samples at a time from sourceB */
+    inB1 = *__SIMD32(pSrcB)++;
+    /* read two samples at a time from sourceA */
+    inA2 = *__SIMD32(pSrcA)++;
+    /* read two samples at a time from sourceB */
+    inB2 = *__SIMD32(pSrcB)++;
+
+    /* multiply mul = sourceA * sourceB */
+    mul1 = (q31_t) ((q15_t) (inA1 >> 16) * (q15_t) (inB1 >> 16));
+    mul2 = (q31_t) ((q15_t) inA1 * (q15_t) inB1);
+    mul3 = (q31_t) ((q15_t) (inA2 >> 16) * (q15_t) (inB2 >> 16));
+    mul4 = (q31_t) ((q15_t) inA2 * (q15_t) inB2);
+
+    /* saturate result to 16 bit */
+    out1 = (q15_t) __SSAT((mul1 + NN_ROUND(out_shift)) >> out_shift, 16);
+    out2 = (q15_t) __SSAT((mul2 + NN_ROUND(out_shift)) >> out_shift, 16);
+    out3 = (q15_t) __SSAT((mul3 + NN_ROUND(out_shift)) >> out_shift, 16);
+    out4 = (q15_t) __SSAT((mul4 + NN_ROUND(out_shift)) >> out_shift, 16);
+
+    /* store the result */
+#ifndef ARM_MATH_BIG_ENDIAN
+
+    *__SIMD32(pDst)++ = __PKHBT(out2, out1, 16);
+    *__SIMD32(pDst)++ = __PKHBT(out4, out3, 16);
+
+#else
+
+    *__SIMD32(pDst)++ = __PKHBT(out2, out1, 16);
+    *__SIMD32(pDst)++ = __PKHBT(out4, out3, 16);
+
+#endif /* #ifndef ARM_MATH_BIG_ENDIAN */
+
+    /* Decrement the blockSize loop counter */
+    blkCnt--;
+  }
+
+  /* If the blockSize is not a multiple of 4, compute any remaining output samples here.
+   ** No loop unrolling is used. */
+  blkCnt = blockSize % 0x4U;
+
+#else
+
+  /* Run the below code for Cortex-M0 */
+
+  /* Initialize blkCnt with number of samples */
+  blkCnt = blockSize;
+
+#endif /* #if defined (ARM_MATH_DSP) */
+
+
+  while (blkCnt > 0U)
+  {
+    /* C = A * B */
+    /* Multiply the inputs and store the result in the destination buffer */
+    *pDst++ = (q15_t) __SSAT((((q31_t) (*pSrcA++) * (*pSrcB++) + NN_ROUND(out_shift)) >> out_shift), 16);
+
+    /* Decrement the blockSize loop counter */
+    blkCnt--;
+  }
+}
+
+/**
+ * @} end of NNBasicMath group
+ */
+
diff --git a/NN/Source/NNSupportFunctions/arm_nn_mult_q7.c b/NN/Source/NNSupportFunctions/arm_nn_mult_q7.c
new file mode 100644
index 0000000..3735c04
--- /dev/null
+++ b/NN/Source/NNSupportFunctions/arm_nn_mult_q7.c
@@ -0,0 +1,119 @@
+/*
+ * Copyright (C) 2010-2018 Arm Limited or its affiliates. All rights reserved.
+ *
+ * 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.
+ */
+
+/* ----------------------------------------------------------------------
+ * Project:      CMSIS NN Library
+ * Title:        arm_nn_mult_q7.c
+ * Description:  Q7 vector multiplication with variable output shifts
+ *
+ * $Date:        13. July 2018
+ * $Revision:    V.1.0.0
+ *
+ * Target Processor:  Cortex-M cores
+ *
+ * -------------------------------------------------------------------- */
+
+#include "arm_nnfunctions.h"
+
+/**    
+ * @ingroup groupSupport    
+ */
+
+/**
+ * @addtogroup NNBasicMath
+ * @{
+ */
+
+/**
+ * @brief           Q7 vector multiplication with variable output shifts
+ * @param[in]       *pSrcA        pointer to the first input vector
+ * @param[in]       *pSrcB        pointer to the second input vector
+ * @param[out]      *pDst         pointer to the output vector
+ * @param[in]       out_shift     amount of right-shift for output
+ * @param[in]       blockSize     number of samples in each vector
+ * @return none.
+ *
+ * <b>Scaling and Overflow Behavior:</b>
+ * \par
+ * The function uses saturating arithmetic.
+ * Results outside of the allowable Q7 range [0x80 0x7F] will be saturated.
+ */
+
+void arm_nn_mult_q7(
+  q7_t * pSrcA,
+  q7_t * pSrcB,
+  q7_t * pDst,
+  const uint16_t out_shift,
+  uint32_t blockSize)
+{
+  uint32_t blkCnt;                               /* loop counters */
+
+#if defined (ARM_MATH_DSP)
+
+/* Run the below code for Cortex-M4 and Cortex-M3 */
+  q7_t out1, out2, out3, out4;                   /* Temporary variables to store the product */
+
+  /* loop Unrolling */
+  blkCnt = blockSize >> 2U;
+
+  /* First part of the processing with loop unrolling.  Compute 4 outputs at a time.
+   ** a second loop below computes the remaining 1 to 3 samples. */
+  while (blkCnt > 0U)
+  {
+    /* C = A * B */
+    /* Multiply the inputs and store the results in temporary variables */
+    out1 = (q7_t) __SSAT((((q15_t) (*pSrcA++) * (*pSrcB++) + NN_ROUND(out_shift)) >> out_shift), 8);
+    out2 = (q7_t) __SSAT((((q15_t) (*pSrcA++) * (*pSrcB++) + NN_ROUND(out_shift)) >> out_shift), 8);
+    out3 = (q7_t) __SSAT((((q15_t) (*pSrcA++) * (*pSrcB++) + NN_ROUND(out_shift)) >> out_shift), 8);
+    out4 = (q7_t) __SSAT((((q15_t) (*pSrcA++) * (*pSrcB++) + NN_ROUND(out_shift)) >> out_shift), 8);
+
+    /* Store the results of 4 inputs in the destination buffer in single cycle by packing */
+    *__SIMD32(pDst)++ = __PACKq7(out1, out2, out3, out4);
+
+    /* Decrement the blockSize loop counter */
+    blkCnt--;
+  }
+
+  /* If the blockSize is not a multiple of 4, compute any remaining output samples here.
+   ** No loop unrolling is used. */
+  blkCnt = blockSize % 0x4U;
+
+#else
+
+  /* Run the below code for Cortex-M0 */
+
+  /* Initialize blkCnt with number of samples */
+  blkCnt = blockSize;
+
+#endif /* #if defined (ARM_MATH_DSP) */
+
+
+  while (blkCnt > 0U)
+  {
+    /* C = A * B */
+    /* Multiply the inputs and store the result in the destination buffer */
+    *pDst++ = (q7_t) __SSAT((((q15_t) (*pSrcA++) * (*pSrcB++) + NN_ROUND(out_shift)) >> out_shift), 8);
+
+    /* Decrement the blockSize loop counter */
+    blkCnt--;
+  }
+}
+
+/**
+ * @} end of NNBasicMath group
+ */
diff --git a/NN/Source/NNSupportFunctions/arm_nntables.c b/NN/Source/NNSupportFunctions/arm_nntables.c
new file mode 100644
index 0000000..c28f1a6
--- /dev/null
+++ b/NN/Source/NNSupportFunctions/arm_nntables.c
@@ -0,0 +1,297 @@
+/*
+ * Copyright (C) 2010-2018 Arm Limited or its affiliates. All rights reserved.
+ *
+ * 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.
+ */
+
+/* ----------------------------------------------------------------------
+ * Project:      CMSIS NN Library
+ * Title:        arm_nntables.c
+ * Description:  Converts the elements of the Q7 vector to Q15 vector without left-shift
+ *
+ * $Date:        17. January 2018
+ * $Revision:    V.1.0.0
+ *
+ * Target Processor:  Cortex-M cores
+ *
+ * -------------------------------------------------------------------- */
+
+#include "arm_nnsupportfunctions.h"
+
+/**
+ * @brief tables for various activation functions
+ *
+ * This file include the declaration of common tables.
+ * Most of them are used for activation functions 
+ *
+ * Assumption:
+ * Unified table: input is 3.x format, i.e, range of [-8, 8)
+ * sigmoid(8) = 0.9996646498695336
+ * tanh(8) = 0.9999997749296758
+ * The accuracy here should be good enough
+ *
+ * 2-stage HL table: 
+ *
+ * The entire input range is divided into two parts:
+ *
+ * Low range table: 0x000x xxxx or 0x111x xxxx 
+ * table entry will be the binary number excluding the first
+ * two digits, i.e., 0x0x xxxx or 0x1x xxxx
+ * 
+ *
+ *
+ * High range table 0x0010 0000 -- 0x0111 1111
+ *                  0x1000 0000 -- 0x1101 1111
+ * 
+ * For positive numbers, table entry will be
+ * 0x0010 0000 -- 0x0111 1111 minus 0x0010 0000
+ * i.e., 0x0000 0000 - 0x0101 11111
+ *
+ * same thing for the negative numbers, table entry will be
+ * 0x1000 0000 -- 0x1101 1111 minux 0x0010 0000
+ * i.e., 0x0110 0000 - 0x1011 1111
+ */
+
+const q7_t sigmoidTable_q7[256] = {
+    0x40, 0x42, 0x44, 0x46, 0x48, 0x4a, 0x4c, 0x4e,
+    0x50, 0x52, 0x53, 0x55, 0x57, 0x59, 0x5a, 0x5c,
+    0x5e, 0x5f, 0x61, 0x62, 0x63, 0x65, 0x66, 0x67,
+    0x69, 0x6a, 0x6b, 0x6c, 0x6d, 0x6e, 0x6f, 0x70,
+    0x71, 0x72, 0x72, 0x73, 0x74, 0x74, 0x75, 0x76,
+    0x76, 0x77, 0x77, 0x78, 0x78, 0x79, 0x79, 0x7a,
+    0x7a, 0x7a, 0x7b, 0x7b, 0x7b, 0x7c, 0x7c, 0x7c,
+    0x7c, 0x7c, 0x7d, 0x7d, 0x7d, 0x7d, 0x7d, 0x7e,
+    0x7e, 0x7e, 0x7e, 0x7e, 0x7e, 0x7e, 0x7e, 0x7f,
+    0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+    0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+    0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+    0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+    0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+    0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+    0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+    0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+    0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+    0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+    0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+    0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+    0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01,
+    0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01,
+    0x01, 0x01, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02,
+    0x02, 0x02, 0x03, 0x03, 0x03, 0x03, 0x03, 0x04,
+    0x04, 0x04, 0x04, 0x04, 0x05, 0x05, 0x05, 0x06,
+    0x06, 0x06, 0x07, 0x07, 0x08, 0x08, 0x09, 0x09,
+    0x0a, 0x0a, 0x0b, 0x0c, 0x0c, 0x0d, 0x0e, 0x0e,
+    0x0f, 0x10, 0x11, 0x12, 0x13, 0x14, 0x15, 0x16,
+    0x17, 0x19, 0x1a, 0x1b, 0x1d, 0x1e, 0x1f, 0x21,
+    0x22, 0x24, 0x26, 0x27, 0x29, 0x2b, 0x2d, 0x2e,
+    0x30, 0x32, 0x34, 0x36, 0x38, 0x3a, 0x3c, 0x3e,
+};
+
+const q15_t sigmoidTable_q15[256] = {
+    0x4000, 0x4200, 0x43ff, 0x45fc, 0x47f5, 0x49eb, 0x4bdc, 0x4dc8,
+    0x4fad, 0x518a, 0x5360, 0x552c, 0x56ef, 0x58a8, 0x5a57, 0x5bfb,
+    0x5d93, 0x5f20, 0x60a1, 0x6216, 0x637f, 0x64db, 0x662b, 0x676f,
+    0x68a6, 0x69d2, 0x6af1, 0x6c05, 0x6d0d, 0x6e09, 0x6efb, 0x6fe2,
+    0x70be, 0x7190, 0x7258, 0x7316, 0x73cc, 0x7478, 0x751b, 0x75b7,
+    0x764a, 0x76d6, 0x775b, 0x77d8, 0x784f, 0x78c0, 0x792a, 0x798f,
+    0x79ee, 0x7a48, 0x7a9d, 0x7aed, 0x7b39, 0x7b80, 0x7bc4, 0x7c03,
+    0x7c3f, 0x7c78, 0x7cad, 0x7ce0, 0x7d0f, 0x7d3c, 0x7d66, 0x7d8d,
+    0x7db3, 0x7dd6, 0x7df7, 0x7e16, 0x7e33, 0x7e4f, 0x7e69, 0x7e81,
+    0x7e98, 0x7eae, 0x7ec2, 0x7ed5, 0x7ee7, 0x7ef8, 0x7f08, 0x7f17,
+    0x7f25, 0x7f32, 0x7f3e, 0x7f4a, 0x7f55, 0x7f5f, 0x7f69, 0x7f72,
+    0x7f7b, 0x7f83, 0x7f8a, 0x7f91, 0x7f98, 0x7f9e, 0x7fa4, 0x7faa,
+    0x7faf, 0x7fb4, 0x7fb8, 0x7fbd, 0x7fc1, 0x7fc5, 0x7fc8, 0x7fcc,
+    0x7fcf, 0x7fd2, 0x7fd5, 0x7fd7, 0x7fda, 0x7fdc, 0x7fde, 0x7fe0,
+    0x7fe2, 0x7fe4, 0x7fe6, 0x7fe7, 0x7fe9, 0x7fea, 0x7feb, 0x7fed,
+    0x7fee, 0x7fef, 0x7ff0, 0x7ff1, 0x7ff2, 0x7ff3, 0x7ff4, 0x7ff4,
+    0x000b, 0x000c, 0x000c, 0x000d, 0x000e, 0x000f, 0x0010, 0x0011,
+    0x0012, 0x0013, 0x0015, 0x0016, 0x0017, 0x0019, 0x001a, 0x001c,
+    0x001e, 0x0020, 0x0022, 0x0024, 0x0026, 0x0029, 0x002b, 0x002e,
+    0x0031, 0x0034, 0x0038, 0x003b, 0x003f, 0x0043, 0x0048, 0x004c,
+    0x0051, 0x0056, 0x005c, 0x0062, 0x0068, 0x006f, 0x0076, 0x007d,
+    0x0085, 0x008e, 0x0097, 0x00a1, 0x00ab, 0x00b6, 0x00c2, 0x00ce,
+    0x00db, 0x00e9, 0x00f8, 0x0108, 0x0119, 0x012b, 0x013e, 0x0152,
+    0x0168, 0x017f, 0x0197, 0x01b1, 0x01cd, 0x01ea, 0x0209, 0x022a,
+    0x024d, 0x0273, 0x029a, 0x02c4, 0x02f1, 0x0320, 0x0353, 0x0388,
+    0x03c1, 0x03fd, 0x043c, 0x0480, 0x04c7, 0x0513, 0x0563, 0x05b8,
+    0x0612, 0x0671, 0x06d6, 0x0740, 0x07b1, 0x0828, 0x08a5, 0x092a,
+    0x09b6, 0x0a49, 0x0ae5, 0x0b88, 0x0c34, 0x0cea, 0x0da8, 0x0e70,
+    0x0f42, 0x101e, 0x1105, 0x11f7, 0x12f3, 0x13fb, 0x150f, 0x162e,
+    0x175a, 0x1891, 0x19d5, 0x1b25, 0x1c81, 0x1dea, 0x1f5f, 0x20e0,
+    0x226d, 0x2405, 0x25a9, 0x2758, 0x2911, 0x2ad4, 0x2ca0, 0x2e76,
+    0x3053, 0x3238, 0x3424, 0x3615, 0x380b, 0x3a04, 0x3c01, 0x3e00,
+};
+
+const q15_t sigmoidLTable_q15[128] = {
+    0x4000, 0x4100, 0x4200, 0x42ff, 0x43ff, 0x44fd, 0x45fc, 0x46f9,
+    0x47f5, 0x48f1, 0x49eb, 0x4ae5, 0x4bdc, 0x4cd3, 0x4dc8, 0x4ebb,
+    0x4fad, 0x509c, 0x518a, 0x5276, 0x5360, 0x5447, 0x552c, 0x560f,
+    0x56ef, 0x57cd, 0x58a8, 0x5981, 0x5a57, 0x5b2a, 0x5bfb, 0x5cc9,
+    0x5d93, 0x5e5b, 0x5f20, 0x5fe2, 0x60a1, 0x615d, 0x6216, 0x62cc,
+    0x637f, 0x642e, 0x64db, 0x6584, 0x662b, 0x66ce, 0x676f, 0x680c,
+    0x68a6, 0x693d, 0x69d2, 0x6a63, 0x6af1, 0x6b7c, 0x6c05, 0x6c8a,
+    0x6d0d, 0x6d8d, 0x6e09, 0x6e84, 0x6efb, 0x6f70, 0x6fe2, 0x7051,
+    0x0f42, 0x0faf, 0x101e, 0x1090, 0x1105, 0x117c, 0x11f7, 0x1273,
+    0x12f3, 0x1376, 0x13fb, 0x1484, 0x150f, 0x159d, 0x162e, 0x16c3,
+    0x175a, 0x17f4, 0x1891, 0x1932, 0x19d5, 0x1a7c, 0x1b25, 0x1bd2,
+    0x1c81, 0x1d34, 0x1dea, 0x1ea3, 0x1f5f, 0x201e, 0x20e0, 0x21a5,
+    0x226d, 0x2337, 0x2405, 0x24d6, 0x25a9, 0x267f, 0x2758, 0x2833,
+    0x2911, 0x29f1, 0x2ad4, 0x2bb9, 0x2ca0, 0x2d8a, 0x2e76, 0x2f64,
+    0x3053, 0x3145, 0x3238, 0x332d, 0x3424, 0x351b, 0x3615, 0x370f,
+    0x380b, 0x3907, 0x3a04, 0x3b03, 0x3c01, 0x3d01, 0x3e00, 0x3f00,
+};
+
+const q15_t sigmoidHTable_q15[192] = {
+    0x70be, 0x7190, 0x7258, 0x7316, 0x73cc, 0x7478, 0x751b, 0x75b7,
+    0x764a, 0x76d6, 0x775b, 0x77d8, 0x784f, 0x78c0, 0x792a, 0x798f,
+    0x79ee, 0x7a48, 0x7a9d, 0x7aed, 0x7b39, 0x7b80, 0x7bc4, 0x7c03,
+    0x7c3f, 0x7c78, 0x7cad, 0x7ce0, 0x7d0f, 0x7d3c, 0x7d66, 0x7d8d,
+    0x7db3, 0x7dd6, 0x7df7, 0x7e16, 0x7e33, 0x7e4f, 0x7e69, 0x7e81,
+    0x7e98, 0x7eae, 0x7ec2, 0x7ed5, 0x7ee7, 0x7ef8, 0x7f08, 0x7f17,
+    0x7f25, 0x7f32, 0x7f3e, 0x7f4a, 0x7f55, 0x7f5f, 0x7f69, 0x7f72,
+    0x7f7b, 0x7f83, 0x7f8a, 0x7f91, 0x7f98, 0x7f9e, 0x7fa4, 0x7faa,
+    0x7faf, 0x7fb4, 0x7fb8, 0x7fbd, 0x7fc1, 0x7fc5, 0x7fc8, 0x7fcc,
+    0x7fcf, 0x7fd2, 0x7fd5, 0x7fd7, 0x7fda, 0x7fdc, 0x7fde, 0x7fe0,
+    0x7fe2, 0x7fe4, 0x7fe6, 0x7fe7, 0x7fe9, 0x7fea, 0x7feb, 0x7fed,
+    0x7fee, 0x7fef, 0x7ff0, 0x7ff1, 0x7ff2, 0x7ff3, 0x7ff4, 0x7ff4,
+    0x000b, 0x000c, 0x000c, 0x000d, 0x000e, 0x000f, 0x0010, 0x0011,
+    0x0012, 0x0013, 0x0015, 0x0016, 0x0017, 0x0019, 0x001a, 0x001c,
+    0x001e, 0x0020, 0x0022, 0x0024, 0x0026, 0x0029, 0x002b, 0x002e,
+    0x0031, 0x0034, 0x0038, 0x003b, 0x003f, 0x0043, 0x0048, 0x004c,
+    0x0051, 0x0056, 0x005c, 0x0062, 0x0068, 0x006f, 0x0076, 0x007d,
+    0x0085, 0x008e, 0x0097, 0x00a1, 0x00ab, 0x00b6, 0x00c2, 0x00ce,
+    0x00db, 0x00e9, 0x00f8, 0x0108, 0x0119, 0x012b, 0x013e, 0x0152,
+    0x0168, 0x017f, 0x0197, 0x01b1, 0x01cd, 0x01ea, 0x0209, 0x022a,
+    0x024d, 0x0273, 0x029a, 0x02c4, 0x02f1, 0x0320, 0x0353, 0x0388,
+    0x03c1, 0x03fd, 0x043c, 0x0480, 0x04c7, 0x0513, 0x0563, 0x05b8,
+    0x0612, 0x0671, 0x06d6, 0x0740, 0x07b1, 0x0828, 0x08a5, 0x092a,
+    0x09b6, 0x0a49, 0x0ae5, 0x0b88, 0x0c34, 0x0cea, 0x0da8, 0x0e70,
+};
+
+const q7_t tanhTable_q7[256] = {
+    0x00, 0x08, 0x10, 0x18, 0x1f, 0x27, 0x2e, 0x35,
+    0x3b, 0x41, 0x47, 0x4c, 0x51, 0x56, 0x5a, 0x5e,
+    0x61, 0x65, 0x68, 0x6a, 0x6d, 0x6f, 0x71, 0x72,
+    0x74, 0x75, 0x76, 0x78, 0x78, 0x79, 0x7a, 0x7b,
+    0x7b, 0x7c, 0x7c, 0x7d, 0x7d, 0x7e, 0x7e, 0x7e,
+    0x7e, 0x7e, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+    0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+    0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+    0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+    0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+    0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+    0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+    0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+    0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+    0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+    0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+    0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+    0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+    0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+    0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+    0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+    0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+    0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+    0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+    0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+    0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x81,
+    0x81, 0x81, 0x81, 0x81, 0x81, 0x81, 0x81, 0x82,
+    0x82, 0x82, 0x82, 0x82, 0x83, 0x83, 0x84, 0x84,
+    0x85, 0x85, 0x86, 0x87, 0x88, 0x88, 0x8a, 0x8b,
+    0x8c, 0x8e, 0x8f, 0x91, 0x93, 0x96, 0x98, 0x9b,
+    0x9f, 0xa2, 0xa6, 0xaa, 0xaf, 0xb4, 0xb9, 0xbf,
+    0xc5, 0xcb, 0xd2, 0xd9, 0xe1, 0xe8, 0xf0, 0xf8,
+};
+
+const q15_t tanhTable_q15[256] = {
+    0x0000, 0x07fd, 0x0feb, 0x17b9, 0x1f59, 0x26bf, 0x2ddf, 0x34ae,
+    0x3b27, 0x4142, 0x46fd, 0x4c56, 0x514d, 0x55e2, 0x5a1a, 0x5df6,
+    0x617c, 0x64b0, 0x6797, 0x6a37, 0x6c95, 0x6eb5, 0x709e, 0x7254,
+    0x73dc, 0x753a, 0x7672, 0x7788, 0x787f, 0x795b, 0x7a1e, 0x7acb,
+    0x7b65, 0x7bee, 0x7c66, 0x7cd1, 0x7d30, 0x7d84, 0x7dce, 0x7e0f,
+    0x7e49, 0x7e7d, 0x7eaa, 0x7ed2, 0x7ef5, 0x7f14, 0x7f30, 0x7f48,
+    0x7f5e, 0x7f71, 0x7f82, 0x7f91, 0x7f9e, 0x7fa9, 0x7fb3, 0x7fbc,
+    0x7fc4, 0x7fcb, 0x7fd1, 0x7fd7, 0x7fdc, 0x7fe0, 0x7fe4, 0x7fe7,
+    0x7fea, 0x7fed, 0x7fef, 0x7ff1, 0x7ff3, 0x7ff4, 0x7ff6, 0x7ff7,
+    0x7ff8, 0x7ff9, 0x7ffa, 0x7ffa, 0x7ffb, 0x7ffc, 0x7ffc, 0x7ffd,
+    0x7ffd, 0x7ffd, 0x7ffe, 0x7ffe, 0x7ffe, 0x7ffe, 0x7fff, 0x7fff,
+    0x7fff, 0x7fff, 0x7fff, 0x7fff, 0x7fff, 0x7fff, 0x7fff, 0x7fff,
+    0x7fff, 0x7fff, 0x7fff, 0x7fff, 0x7fff, 0x7fff, 0x7fff, 0x7fff,
+    0x7fff, 0x7fff, 0x7fff, 0x7fff, 0x7fff, 0x7fff, 0x7fff, 0x7fff,
+    0x7fff, 0x7fff, 0x7fff, 0x7fff, 0x7fff, 0x7fff, 0x7fff, 0x7fff,
+    0x7fff, 0x7fff, 0x7fff, 0x7fff, 0x7fff, 0x7fff, 0x7fff, 0x7fff,
+    0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000,
+    0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000,
+    0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000,
+    0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000,
+    0x8000, 0x8000, 0x8001, 0x8001, 0x8001, 0x8001, 0x8001, 0x8001,
+    0x8001, 0x8001, 0x8001, 0x8002, 0x8002, 0x8002, 0x8002, 0x8003,
+    0x8003, 0x8003, 0x8004, 0x8004, 0x8005, 0x8006, 0x8006, 0x8007,
+    0x8008, 0x8009, 0x800a, 0x800c, 0x800d, 0x800f, 0x8011, 0x8013,
+    0x8016, 0x8019, 0x801c, 0x8020, 0x8024, 0x8029, 0x802f, 0x8035,
+    0x803c, 0x8044, 0x804d, 0x8057, 0x8062, 0x806f, 0x807e, 0x808f,
+    0x80a2, 0x80b8, 0x80d0, 0x80ec, 0x810b, 0x812e, 0x8156, 0x8183,
+    0x81b7, 0x81f1, 0x8232, 0x827c, 0x82d0, 0x832f, 0x839a, 0x8412,
+    0x849b, 0x8535, 0x85e2, 0x86a5, 0x8781, 0x8878, 0x898e, 0x8ac6,
+    0x8c24, 0x8dac, 0x8f62, 0x914b, 0x936b, 0x95c9, 0x9869, 0x9b50,
+    0x9e84, 0xa20a, 0xa5e6, 0xaa1e, 0xaeb3, 0xb3aa, 0xb903, 0xbebe,
+    0xc4d9, 0xcb52, 0xd221, 0xd941, 0xe0a7, 0xe847, 0xf015, 0xf803,
+};
+
+const q15_t tanhLTable_q15[128] = {
+    0x0000, 0x0400, 0x07fd, 0x0bf7, 0x0feb, 0x13d7, 0x17b9, 0x1b90,
+    0x1f59, 0x2314, 0x26bf, 0x2a58, 0x2ddf, 0x3151, 0x34ae, 0x37f6,
+    0x3b27, 0x3e40, 0x4142, 0x442c, 0x46fd, 0x49b6, 0x4c56, 0x4edd,
+    0x514d, 0x53a3, 0x55e2, 0x580a, 0x5a1a, 0x5c13, 0x5df6, 0x5fc4,
+    0x617c, 0x6320, 0x64b0, 0x662d, 0x6797, 0x68f0, 0x6a37, 0x6b6e,
+    0x6c95, 0x6dac, 0x6eb5, 0x6fb0, 0x709e, 0x717f, 0x7254, 0x731e,
+    0x73dc, 0x7490, 0x753a, 0x75da, 0x7672, 0x7701, 0x7788, 0x7807,
+    0x787f, 0x78f0, 0x795b, 0x79bf, 0x7a1e, 0x7a77, 0x7acb, 0x7b1b,
+    0x849b, 0x84e5, 0x8535, 0x8589, 0x85e2, 0x8641, 0x86a5, 0x8710,
+    0x8781, 0x87f9, 0x8878, 0x88ff, 0x898e, 0x8a26, 0x8ac6, 0x8b70,
+    0x8c24, 0x8ce2, 0x8dac, 0x8e81, 0x8f62, 0x9050, 0x914b, 0x9254,
+    0x936b, 0x9492, 0x95c9, 0x9710, 0x9869, 0x99d3, 0x9b50, 0x9ce0,
+    0x9e84, 0xa03c, 0xa20a, 0xa3ed, 0xa5e6, 0xa7f6, 0xaa1e, 0xac5d,
+    0xaeb3, 0xb123, 0xb3aa, 0xb64a, 0xb903, 0xbbd4, 0xbebe, 0xc1c0,
+    0xc4d9, 0xc80a, 0xcb52, 0xceaf, 0xd221, 0xd5a8, 0xd941, 0xdcec,
+    0xe0a7, 0xe470, 0xe847, 0xec29, 0xf015, 0xf409, 0xf803, 0xfc00,
+};
+
+const q15_t tanhHTable_q15[192] = {
+    0x7b65, 0x7bee, 0x7c66, 0x7cd1, 0x7d30, 0x7d84, 0x7dce, 0x7e0f,
+    0x7e49, 0x7e7d, 0x7eaa, 0x7ed2, 0x7ef5, 0x7f14, 0x7f30, 0x7f48,
+    0x7f5e, 0x7f71, 0x7f82, 0x7f91, 0x7f9e, 0x7fa9, 0x7fb3, 0x7fbc,
+    0x7fc4, 0x7fcb, 0x7fd1, 0x7fd7, 0x7fdc, 0x7fe0, 0x7fe4, 0x7fe7,
+    0x7fea, 0x7fed, 0x7fef, 0x7ff1, 0x7ff3, 0x7ff4, 0x7ff6, 0x7ff7,
+    0x7ff8, 0x7ff9, 0x7ffa, 0x7ffa, 0x7ffb, 0x7ffc, 0x7ffc, 0x7ffd,
+    0x7ffd, 0x7ffd, 0x7ffe, 0x7ffe, 0x7ffe, 0x7ffe, 0x7fff, 0x7fff,
+    0x7fff, 0x7fff, 0x7fff, 0x7fff, 0x7fff, 0x7fff, 0x7fff, 0x7fff,
+    0x7fff, 0x7fff, 0x7fff, 0x7fff, 0x7fff, 0x7fff, 0x7fff, 0x7fff,
+    0x7fff, 0x7fff, 0x7fff, 0x7fff, 0x7fff, 0x7fff, 0x7fff, 0x7fff,
+    0x7fff, 0x7fff, 0x7fff, 0x7fff, 0x7fff, 0x7fff, 0x7fff, 0x7fff,
+    0x7fff, 0x7fff, 0x7fff, 0x7fff, 0x7fff, 0x7fff, 0x7fff, 0x7fff,
+    0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000,
+    0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000,
+    0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000,
+    0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000,
+    0x8000, 0x8000, 0x8001, 0x8001, 0x8001, 0x8001, 0x8001, 0x8001,
+    0x8001, 0x8001, 0x8001, 0x8002, 0x8002, 0x8002, 0x8002, 0x8003,
+    0x8003, 0x8003, 0x8004, 0x8004, 0x8005, 0x8006, 0x8006, 0x8007,
+    0x8008, 0x8009, 0x800a, 0x800c, 0x800d, 0x800f, 0x8011, 0x8013,
+    0x8016, 0x8019, 0x801c, 0x8020, 0x8024, 0x8029, 0x802f, 0x8035,
+    0x803c, 0x8044, 0x804d, 0x8057, 0x8062, 0x806f, 0x807e, 0x808f,
+    0x80a2, 0x80b8, 0x80d0, 0x80ec, 0x810b, 0x812e, 0x8156, 0x8183,
+    0x81b7, 0x81f1, 0x8232, 0x827c, 0x82d0, 0x832f, 0x839a, 0x8412,
+};
diff --git a/NN/Source/NNSupportFunctions/arm_q7_to_q15_no_shift.c b/NN/Source/NNSupportFunctions/arm_q7_to_q15_no_shift.c
new file mode 100644
index 0000000..264e760
--- /dev/null
+++ b/NN/Source/NNSupportFunctions/arm_q7_to_q15_no_shift.c
@@ -0,0 +1,134 @@
+/*
+ * Copyright (C) 2010-2018 Arm Limited or its affiliates. All rights reserved.
+ *
+ * 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.
+ */
+
+/* ----------------------------------------------------------------------
+ * Project:      CMSIS NN Library
+ * Title:        arm_q7_to_q15_no_shift.c
+ * Description:  Converts the elements of the Q7 vector to Q15 vector without left-shift
+ *
+ * $Date:        17. January 2018
+ * $Revision:    V.1.0.0
+ *
+ * Target Processor:  Cortex-M cores
+ *
+ * -------------------------------------------------------------------- */
+
+#include "arm_nnsupportfunctions.h"
+
+/**    
+ * @ingroup groupSupport    
+ */
+
+/**    
+ * @addtogroup nndata_convert    
+ * @{    
+ */
+
+/**    
+ * @brief Converts the elements of the Q7 vector to Q15 vector without left-shift 
+ * @param[in]       *pSrc points to the Q7 input vector    
+ * @param[out]      *pDst points to the Q15 output vector   
+ * @param[in]       blockSize length of the input vector    
+ * @return none.    
+ *    
+ * \par Description:    
+ *    
+ * The equation used for the conversion process is:    
+ *   
+ * <pre>    
+ * 	pDst[n] = (q15_t) pSrc[n];   0 <= n < blockSize.    
+ * </pre>    
+ *   
+ */
+
+void arm_q7_to_q15_no_shift(const q7_t * pSrc, q15_t * pDst, uint32_t blockSize)
+{
+    const q7_t *pIn = pSrc;     /* Src pointer */
+    uint32_t  blkCnt;           /* loop counter */
+
+#ifndef ARM_MATH_CM0_FAMILY
+    q31_t     in;
+    q31_t     in1, in2;
+    q31_t     out1, out2;
+
+    /* Run the below code for Cortex-M4 and Cortex-M3 */
+
+    /*loop Unrolling */
+    blkCnt = blockSize >> 2u;
+
+    /* First part of the processing with loop unrolling.  Compute 4 outputs at a time.    
+     ** a second loop below computes the remaining 1 to 3 samples. */
+    while (blkCnt > 0u)
+    {
+        /* C = (q15_t) A << 8 */
+        /* convert from q7 to q15 and then store the results in the destination buffer */
+        in = *__SIMD32(pIn)++;
+
+        /* rotatate in by 8 and extend two q7_t values to q15_t values */
+        in1 = __SXTB16(__ROR(in, 8));
+
+        /* extend remainig two q7_t values to q15_t values */
+        in2 = __SXTB16(in);
+
+#ifndef ARM_MATH_BIG_ENDIAN
+
+        out2 = __PKHTB(in1, in2, 16);
+        out1 = __PKHBT(in2, in1, 16);
+
+#else
+
+        out1 = __PKHTB(in1, in2, 16);
+        out2 = __PKHBT(in2, in1, 16);
+
+#endif
+
+        *__SIMD32(pDst)++ = out1;
+        *__SIMD32(pDst)++ = out2;
+
+        /* Decrement the loop counter */
+        blkCnt--;
+    }
+
+    /* If the blockSize is not a multiple of 4, compute any remaining output samples here.    
+     ** No loop unrolling is used. */
+    blkCnt = blockSize % 0x4u;
+
+#else
+
+    /* Run the below code for Cortex-M0 */
+
+    /* Loop over blockSize number of values */
+    blkCnt = blockSize;
+
+#endif                          /* #ifndef ARM_MATH_CM0_FAMILY */
+
+    while (blkCnt > 0u)
+    {
+        /* C = (q15_t) A << 8 */
+        /* convert from q7 to q15 and then store the results in the destination buffer */
+        *pDst++ = (q15_t) * pIn++;
+
+        /* Decrement the loop counter */
+        blkCnt--;
+    }
+
+}
+
+/**    
+ * @} end of nndata_convert group   
+ */
diff --git a/NN/Source/NNSupportFunctions/arm_q7_to_q15_reordered_no_shift.c b/NN/Source/NNSupportFunctions/arm_q7_to_q15_reordered_no_shift.c
new file mode 100644
index 0000000..7d29aa4
--- /dev/null
+++ b/NN/Source/NNSupportFunctions/arm_q7_to_q15_reordered_no_shift.c
@@ -0,0 +1,145 @@
+/*
+ * Copyright (C) 2010-2018 Arm Limited or its affiliates. All rights reserved.
+ *
+ * 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.
+ */
+
+/* ----------------------------------------------------------------------
+ * Project:      CMSIS NN Library
+ * Title:        arm_q7_to_q15_reordered_no_shift.c
+ * Description:  Converts the elements of the Q7 vector to reordered Q15 vector without left-shift
+ *
+ * $Date:        17. January 2018
+ * $Revision:    V.1.0.0
+ *
+ * Target Processor:  Cortex-M cores
+ *
+ * -------------------------------------------------------------------- */
+
+#include "arm_nnsupportfunctions.h"
+
+/**    
+ * @ingroup groupSupport    
+ */
+
+/**    
+ * @addtogroup nndata_convert    
+ * @{    
+ */
+
+/**    
+ * @brief Converts the elements of the Q7 vector to reordered Q15 vector without left-shift
+ * @param[in]       *pSrc points to the Q7 input vector    
+ * @param[out]      *pDst points to the Q15 output vector   
+ * @param[in]       blockSize length of the input vector    
+ * @return none.    
+ *    
+ * @details
+ *
+ * This function does the q7 to q15 expansion with re-ordering 
+ *
+ * <pre>
+ *                          |   A1   |   A2   |   A3   |   A4   |
+ *
+ *                           0      7 8     15 16    23 24    31
+ * </pre>
+ *
+ * is converted into:
+ *
+ * <pre>
+ *  |       A1       |       A3       |   and  |       A2       |       A4       |
+ *
+ *   0             15 16            31          0             15 16            31
+ * </pre>
+ *
+ *
+ * This looks strange but is natural considering how sign-extension is done at
+ * assembly level. 
+ *
+ * The expansion of other other oprand will follow the same rule so that the end 
+ * results are the same.
+ *
+ * The tail (i.e., last (N % 4) elements) will still be in original order.
+ *   
+ */
+
+void arm_q7_to_q15_reordered_no_shift(const q7_t * pSrc, q15_t * pDst, uint32_t blockSize)
+{
+    const q7_t *pIn = pSrc;     /* Src pointer */
+    uint32_t  blkCnt;           /* loop counter */
+
+#ifndef ARM_MATH_CM0_FAMILY
+    q31_t     in;
+    q31_t     in1, in2;
+
+    /* Run the below code for Cortex-M4 and Cortex-M3 */
+
+    /*loop Unrolling */
+    blkCnt = blockSize >> 2u;
+
+    /* First part of the processing with loop unrolling.  Compute 4 outputs at a time.    
+     ** a second loop below computes the remaining 1 to 3 samples. */
+    while (blkCnt > 0u)
+    {
+        /* C = (q15_t) A << 8 */
+        /* convert from q7 to q15 and then store the results in the destination buffer */
+        in = *__SIMD32(pIn)++;
+
+        /* rotatate in by 8 and extend two q7_t values to q15_t values */
+        in1 = __SXTB16(__ROR(in, 8));
+
+        /* extend remainig two q7_t values to q15_t values */
+        in2 = __SXTB16(in);
+
+#ifndef ARM_MATH_BIG_ENDIAN
+        *__SIMD32(pDst)++ = in2;
+        *__SIMD32(pDst)++ = in1;
+#else
+        *__SIMD32(pDst)++ = in1;
+        *__SIMD32(pDst)++ = in2;
+#endif
+
+        /* Decrement the loop counter */
+        blkCnt--;
+    }
+
+    /* If the blockSize is not a multiple of 4, compute any remaining output samples here.    
+     ** No loop unrolling is used. */
+    blkCnt = blockSize % 0x4u;
+
+#else
+
+    /* Run the below code for Cortex-M0 */
+
+    /* Loop over blockSize number of values */
+    blkCnt = blockSize;
+
+#endif                          /* #ifndef ARM_MATH_CM0_FAMILY */
+
+    while (blkCnt > 0u)
+    {
+        /* C = (q15_t) A << 8 */
+        /* convert from q7 to q15 and then store the results in the destination buffer */
+        *pDst++ = (q15_t) * pIn++;
+
+        /* Decrement the loop counter */
+        blkCnt--;
+    }
+
+}
+
+/**    
+ * @} end of q7_to_x group    
+ */
diff --git a/NN/Source/PoolingFunctions/arm_pool_q7_HWC.c b/NN/Source/PoolingFunctions/arm_pool_q7_HWC.c
new file mode 100644
index 0000000..b451f5e
--- /dev/null
+++ b/NN/Source/PoolingFunctions/arm_pool_q7_HWC.c
@@ -0,0 +1,448 @@
+/*
+ * Copyright (C) 2010-2018 Arm Limited or its affiliates. All rights reserved.
+ *
+ * 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.
+ */
+
+/* ----------------------------------------------------------------------
+ * Project:      CMSIS NN Library
+ * Title:        arm_pool_q7_HWC.c
+ * Description:  Pooling function implementations
+ *
+ * $Date:        17. January 2018
+ * $Revision:    V.1.0.0
+ *
+ * Target Processor:  Cortex-M cores
+ *
+ * -------------------------------------------------------------------- */
+
+#include "arm_math.h"
+#include "arm_nnfunctions.h"
+
+#if defined (ARM_MATH_DSP)
+
+/**
+ * @brief A few utility functions used by pooling functions
+ *
+ * 
+ */
+
+static void buffer_scale_back_q15_to_q7(q15_t * buffer, q7_t * target, uint16_t length, uint16_t scale)
+{
+    int       i;
+
+    for (i = 0; i < length; i++)
+    {
+        target[i] = (q7_t) (buffer[i] / scale);
+    }
+}
+
+static void compare_and_replace_if_larger_q7(q7_t * base,   // base data
+                                             q7_t * target, // compare target
+                                             const uint16_t length  // data length
+    )
+{
+    q7_t     *pIn = base;
+    q7_t     *pCom = target;
+    union arm_nnword in;
+    union arm_nnword com;
+    uint16_t  cnt = length >> 2;
+
+    while (cnt > 0u)
+    {
+        in.word = *__SIMD32(pIn);
+        com.word = *__SIMD32(pCom)++;
+
+        // if version
+        if (com.bytes[0] > in.bytes[0])
+            in.bytes[0] = com.bytes[0];
+        if (com.bytes[1] > in.bytes[1])
+            in.bytes[1] = com.bytes[1];
+        if (com.bytes[2] > in.bytes[2])
+            in.bytes[2] = com.bytes[2];
+        if (com.bytes[3] > in.bytes[3])
+            in.bytes[3] = com.bytes[3];
+
+        *__SIMD32(pIn)++ = in.word;
+
+        cnt--;
+    }
+}
+
+static void accumulate_q7_to_q15(q15_t * base, q7_t * target, const uint16_t length)
+{
+    q15_t    *pCnt = base;
+    q7_t     *pV = target;
+    q31_t     v1, v2, vo1, vo2;
+    uint16_t  cnt = length >> 2;
+    q31_t     in;
+
+    while (cnt > 0u)
+    {
+        q31_t     value = *__SIMD32(pV)++;
+        v1 = __SXTB16(__ROR(value, 8));
+        v2 = __SXTB16(value);
+#ifndef ARM_MATH_BIG_ENDIAN
+
+        vo2 = __PKHTB(v1, v2, 16);
+        vo1 = __PKHBT(v2, v1, 16);
+
+#else
+
+        vo1 = __PKHTB(v1, v2, 16);
+        vo2 = __PKHBT(v2, v1, 16);
+
+#endif
+
+        in = *__SIMD32(pCnt);
+        *__SIMD32(pCnt)++ = __QADD16(vo1, in);
+
+        in = *__SIMD32(pCnt);
+        *__SIMD32(pCnt)++ = __QADD16(vo2, in);
+
+        cnt--;
+    }
+    cnt = length & 0x3;
+    while (cnt > 0u)
+    {
+        *pCnt++ += *pV++;
+        cnt--;
+    }
+}
+
+#endif                          // ARM_MATH_DSP
+
+/**
+ *  @ingroup groupNN
+ */
+
+/**
+ * @addtogroup Pooling
+ * @{
+ */
+
+  /**
+   * @brief Q7 max pooling function
+   * @param[in, out]  Im_in       pointer to input tensor
+   * @param[in]       dim_im_in   input tensor dimention
+   * @param[in]       ch_im_in    number of input tensor channels
+   * @param[in]       dim_kernel  filter kernel size
+   * @param[in]       padding     padding sizes
+   * @param[in]       stride      convolution stride
+   * @param[in]       dim_im_out  output tensor dimension
+   * @param[in,out]   bufferA     pointer to buffer space for input
+   * @param[in,out]   Im_out      pointer to output tensor
+   * @return none.
+   *
+   * @details
+   *
+   * <b>Buffer size:</b>
+   *
+   * bufferA size:  0
+   *
+   * The pooling function is implemented as split x-pooling then
+   * y-pooling.
+   *
+   * This pooling function is input-destructive. Input data is undefined
+   * after calling this function.
+   *
+   */
+
+void
+arm_maxpool_q7_HWC(q7_t * Im_in,
+                   const uint16_t dim_im_in,
+                   const uint16_t ch_im_in,
+                   const uint16_t dim_kernel,
+                   const uint16_t padding,
+                   const uint16_t stride, const uint16_t dim_im_out, q7_t * bufferA, q7_t * Im_out)
+{
+
+#if defined (ARM_MATH_DSP)
+    /* Run the following code for Cortex-M4 and Cortex-M7 */
+
+    int16_t   i_x, i_y;
+
+    /* first does the pooling along x axis */
+    for (i_y = 0; i_y < dim_im_in; i_y++)
+    {
+
+        for (i_x = 0; i_x < dim_im_out; i_x++)
+        {
+            /* for each output pixel */
+            q7_t     *target = Im_in + (i_y * dim_im_in + i_x) * ch_im_in;
+            q7_t     *win_start;
+            q7_t     *win_stop;
+            if (i_x * stride - padding < 0)
+            {
+                win_start = target;
+            } else
+            {
+                win_start = Im_in + (i_y * dim_im_in + i_x * stride - padding) * ch_im_in;
+            }
+
+            if (i_x * stride - padding + dim_kernel >= dim_im_in)
+            {
+                win_stop = Im_in + (i_y * dim_im_in + dim_im_in) * ch_im_in;
+            } else
+            {
+                win_stop = Im_in + (i_y * dim_im_in + i_x * stride - padding + dim_kernel) * ch_im_in;
+            }
+
+            /* first step is to copy over initial data */
+            /* arm_copy_q7(win_start, target, ch_im_in); */
+            memmove(target, win_start, ch_im_in);
+
+            /* start the max operation from the second part */
+            win_start += ch_im_in;
+            for (; win_start < win_stop; win_start += ch_im_in)
+            {
+                compare_and_replace_if_larger_q7(target, win_start, ch_im_in);
+            }
+        }
+    }
+
+    /* then does the pooling along y axis */
+    for (i_y = 0; i_y < dim_im_out; i_y++)
+    {
+
+        /* for each output row */
+        q7_t     *target = Im_out + i_y * dim_im_out * ch_im_in;
+        q7_t     *row_start;
+        q7_t     *row_end;
+        /* setting the starting row */
+        if (i_y * stride - padding < 0)
+        {
+            row_start = Im_in;
+        } else
+        {
+            row_start = Im_in + (i_y * stride - padding) * dim_im_in * ch_im_in;
+        }
+        /* setting the stopping row */
+        if (i_y * stride - padding + dim_kernel >= dim_im_in)
+        {
+            row_end = Im_in + dim_im_in * dim_im_in * ch_im_in;
+        } else
+        {
+            row_end = Im_in + (i_y * stride - padding + dim_kernel) * dim_im_in * ch_im_in;
+        }
+
+        /* copy over the first row */
+        /* arm_copy_q7(row_start, target, dim_im_out * ch_im_in); */
+        memmove(target, row_start, dim_im_out * ch_im_in);
+
+        /* move over to next row */
+        row_start += ch_im_in * dim_im_in;
+
+        for (; row_start < row_end; row_start += dim_im_in * ch_im_in)
+        {
+            compare_and_replace_if_larger_q7(target, row_start, dim_im_out * ch_im_in);
+        }
+    }
+
+#else
+    /* Run the following code as reference implementation for Cortex-M0 and Cortex-M3 */
+
+    int16_t   i_ch_in, i_x, i_y;
+    int16_t   k_x, k_y;
+
+    for (i_ch_in = 0; i_ch_in < ch_im_in; i_ch_in++)
+    {
+        for (i_y = 0; i_y < dim_im_out; i_y++)
+        {
+            for (i_x = 0; i_x < dim_im_out; i_x++)
+            {
+                int       max = -129;
+                for (k_y = i_y * stride - padding; k_y < i_y * stride - padding + dim_kernel; k_y++)
+                {
+                    for (k_x = i_x * stride - padding; k_x < i_x * stride - padding + dim_kernel; k_x++)
+                    {
+                        if (k_y >= 0 && k_x >= 0 && k_y < dim_im_in && k_x < dim_im_in)
+                        {
+                            if (Im_in[i_ch_in + ch_im_in * (k_x + k_y * dim_im_in)] > max)
+                            {
+                                max = Im_in[i_ch_in + ch_im_in * (k_x + k_y * dim_im_in)];
+                            }
+                        }
+                    }
+                }
+                Im_out[i_ch_in + ch_im_in * (i_x + i_y * dim_im_out)] = max;
+            }
+        }
+    }
+
+#endif                          /* ARM_MATH_DSP */
+
+}
+
+  /**
+   * @brief Q7 average pooling function
+   * @param[in,out]   Im_in       pointer to input tensor
+   * @param[in]       dim_im_in   input tensor dimention
+   * @param[in]       ch_im_in    number of input tensor channels
+   * @param[in]       dim_kernel  filter kernel size
+   * @param[in]       padding     padding sizes
+   * @param[in]       stride      convolution stride
+   * @param[in]       dim_im_out  output tensor dimension
+   * @param[in,out]   bufferA     pointer to buffer space for input
+   * @param[in,out]   Im_out      pointer to output tensor
+   * @return none.
+   *
+   * @details
+   *
+   * <b>Buffer size:</b>
+   *
+   * bufferA size:  2*dim_im_out*ch_im_in
+   *
+   * The pooling function is implemented as split x-pooling then
+   * y-pooling.
+   *
+   * This pooling function is input-destructive. Input data is undefined
+   * after calling this function.
+   *
+   */
+
+void
+arm_avepool_q7_HWC(q7_t * Im_in,
+                   const uint16_t dim_im_in,
+                   const uint16_t ch_im_in,
+                   const uint16_t dim_kernel,
+                   const uint16_t padding,
+                   const uint16_t stride, const uint16_t dim_im_out, q7_t * bufferA, q7_t * Im_out)
+{
+
+#if defined (ARM_MATH_DSP)
+    /* Run the following code for Cortex-M4 and Cortex-M7 */
+
+    q15_t    *buffer = (q15_t *) bufferA;
+    int16_t   i_x, i_y;
+    int16_t   count = 0;
+
+    /* first does the pooling along x axis */
+    for (i_y = 0; i_y < dim_im_in; i_y++)
+    {
+
+        for (i_x = 0; i_x < dim_im_out; i_x++)
+        {
+            /* for each output pixel */
+            q7_t     *target = Im_in + (i_y * dim_im_in + i_x) * ch_im_in;
+            q7_t     *win_start;
+            q7_t     *win_stop;
+            if (i_x * stride - padding < 0)
+            {
+                win_start = target;
+            } else
+            {
+                win_start = Im_in + (i_y * dim_im_in + i_x * stride - padding) * ch_im_in;
+            }
+
+            if (i_x * stride - padding + dim_kernel >= dim_im_in)
+            {
+                win_stop = Im_in + (i_y * dim_im_in + dim_im_in) * ch_im_in;
+            } else
+            {
+                win_stop = Im_in + (i_y * dim_im_in + i_x * stride - padding + dim_kernel) * ch_im_in;
+            }
+
+            /* first step is to copy over initial data */
+            arm_q7_to_q15_no_shift(win_start, buffer, ch_im_in);
+            count = 1;
+
+            /* start the max operation from the second part */
+            win_start += ch_im_in;
+            for (; win_start < win_stop; win_start += ch_im_in)
+            {
+                accumulate_q7_to_q15(buffer, win_start, ch_im_in);
+                count++;
+            }
+            buffer_scale_back_q15_to_q7(buffer, target, ch_im_in, count);
+        }
+    }
+
+    /* then does the pooling along y axis */
+    for (i_y = 0; i_y < dim_im_out; i_y++)
+    {
+        /* for each output row */
+        q7_t     *target = Im_out + i_y * dim_im_out * ch_im_in;
+        q7_t     *row_start;
+        q7_t     *row_end;
+        /* setting the starting row */
+        if (i_y * stride - padding < 0)
+        {
+            row_start = Im_in;
+        } else
+        {
+            row_start = Im_in + (i_y * stride - padding) * dim_im_in * ch_im_in;
+        }
+        /* setting the stopping row */
+        if (i_y * stride - padding + dim_kernel >= dim_im_in)
+        {
+            row_end = Im_in + dim_im_in * dim_im_in * ch_im_in;
+        } else
+        {
+            row_end = Im_in + (i_y * stride - padding + dim_kernel) * dim_im_in * ch_im_in;
+        }
+
+        /* copy over the first row */
+        arm_q7_to_q15_no_shift(row_start, buffer, dim_im_out * ch_im_in);
+        count = 1;
+
+        /* move over to next row */
+        row_start += ch_im_in * dim_im_in;
+
+        for (; row_start < row_end; row_start += dim_im_in * ch_im_in)
+        {
+            accumulate_q7_to_q15(buffer, row_start, dim_im_out * ch_im_in);
+            count++;
+        }
+        buffer_scale_back_q15_to_q7(buffer, target, dim_im_out * ch_im_in, count);
+    }
+
+#else
+    /* Run the following code as reference implementation for Cortex-M0 and Cortex-M3 */
+
+    int16_t   i_ch_in, i_x, i_y;
+    int16_t   k_x, k_y;
+
+    for (i_ch_in = 0; i_ch_in < ch_im_in; i_ch_in++)
+    {
+        for (i_y = 0; i_y < dim_im_out; i_y++)
+        {
+            for (i_x = 0; i_x < dim_im_out; i_x++)
+            {
+                int       sum = 0;
+                int       count = 0;
+                for (k_y = i_y * stride - padding; k_y < i_y * stride - padding + dim_kernel; k_y++)
+                {
+                    for (k_x = i_x * stride - padding; k_x < i_x * stride - padding + dim_kernel; k_x++)
+                    {
+                        if (k_y >= 0 && k_x >= 0 && k_y < dim_im_in && k_x < dim_im_in)
+                        {
+                            sum += Im_in[i_ch_in + ch_im_in * (k_x + k_y * dim_im_in)];
+                            count++;
+                        }
+                    }
+                }
+                Im_out[i_ch_in + ch_im_in * (i_x + i_y * dim_im_out)] = sum / count;
+            }
+        }
+    }
+
+#endif                          /* ARM_MATH_DSP */
+
+}
+
+/**
+ * @} end of Pooling group
+ */
diff --git a/NN/Source/SoftmaxFunctions/arm_softmax_q15.c b/NN/Source/SoftmaxFunctions/arm_softmax_q15.c
new file mode 100644
index 0000000..abc2737
--- /dev/null
+++ b/NN/Source/SoftmaxFunctions/arm_softmax_q15.c
@@ -0,0 +1,120 @@
+/*
+ * Copyright (C) 2010-2018 Arm Limited or its affiliates. All rights reserved.
+ *
+ * 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.
+ */
+
+/* ----------------------------------------------------------------------
+ * Project:      CMSIS NN Library
+ * Title:        arm_softmax_q15.c
+ * Description:  Q15 softmax function
+ *
+ * $Date:        20. February 2018
+ * $Revision:    V.1.0.0
+ *
+ * Target Processor:  Cortex-M cores
+ *
+ * -------------------------------------------------------------------- */
+
+#include "arm_math.h"
+#include "arm_nnfunctions.h"
+
+/**
+ *  @ingroup groupNN
+ */
+
+/**
+ * @addtogroup Softmax
+ * @{
+ */
+
+  /**
+   * @brief Q15 softmax function
+   * @param[in]       vec_in      pointer to input vector
+   * @param[in]       dim_vec     input vector dimention
+   * @param[out]      p_out       pointer to output vector
+   * @return none.
+   *
+   * @details
+   *
+   *  Here, instead of typical e based softmax, we use
+   *  2-based softmax, i.e.,:
+   *
+   *  y_i = 2^(x_i) / sum(2^x_j)
+   *
+   *  The relative output will be different here.
+   *  But mathematically, the gradient will be the same
+   *  with a log(2) scaling factor.
+   *
+   */
+
+void arm_softmax_q15(const q15_t * vec_in, const uint16_t dim_vec, q15_t * p_out)
+{
+    q31_t     sum;
+    int16_t   i;
+    uint8_t   shift;
+    q31_t     base;
+    base = -1 * 0x100000;
+    for (i = 0; i < dim_vec; i++)
+    {
+        if (vec_in[i] > base)
+        {
+            base = vec_in[i];
+        }
+    }
+
+    /* we ignore really small values  
+     * anyway, they will be 0 after shrinking
+     * to q15_t
+     */
+    base = base - 16;
+
+    sum = 0;
+
+    for (i = 0; i < dim_vec; i++)
+    {
+        if (vec_in[i] > base)
+        {
+            shift = (uint8_t)__USAT(vec_in[i] - base, 5);
+            sum += 0x1 << shift;
+        }
+    }
+
+    /* This is effectively (0x1 << 32) / sum */
+    int64_t div_base = 0x100000000LL;
+    int output_base = (int32_t)(div_base / sum);
+
+    /* Final confidence will be output_base >> ( 17 - (vec_in[i] - base) )
+     * so 32768 (0x1<<15) -> 100% confidence when sum = 0x1 << 16, output_base = 0x1 << 16
+     * and vec_in[i]-base = 16
+     */
+    for (i = 0; i < dim_vec; i++)
+    {
+        if (vec_in[i] > base) 
+        {
+            /* Here minimum value of 17+base-vec[i] will be 1 */
+            shift = (uint8_t)__USAT(17+base-vec_in[i], 5);
+            p_out[i] = (q15_t) __SSAT((output_base >> shift), 16);
+        } else
+        {
+            p_out[i] = 0;
+        }
+    }
+
+}
+
+/**
+ * @} end of Softmax group
+ */
diff --git a/NN/Source/SoftmaxFunctions/arm_softmax_q7.c b/NN/Source/SoftmaxFunctions/arm_softmax_q7.c
new file mode 100644
index 0000000..a4e2548
--- /dev/null
+++ b/NN/Source/SoftmaxFunctions/arm_softmax_q7.c
@@ -0,0 +1,121 @@
+/*
+ * Copyright (C) 2010-2018 Arm Limited or its affiliates. All rights reserved.
+ *
+ * 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.
+ */
+
+/* ----------------------------------------------------------------------
+ * Project:      CMSIS NN Library
+ * Title:        arm_softmax_q7.c
+ * Description:  Q7 softmax function
+ *
+ * $Date:        20. February 2018
+ * $Revision:    V.1.0.0
+ *
+ * Target Processor:  Cortex-M cores
+ *
+ * -------------------------------------------------------------------- */
+
+#include "arm_math.h"
+#include "arm_nnfunctions.h"
+
+/**
+ *  @ingroup groupNN
+ */
+
+/**
+ * @addtogroup Softmax
+ * @{
+ */
+
+  /**
+   * @brief Q7 softmax function
+   * @param[in]       vec_in      pointer to input vector
+   * @param[in]       dim_vec     input vector dimention
+   * @param[out]      p_out       pointer to output vector
+   * @return none.
+   *
+   * @details
+   *
+   *  Here, instead of typical natural logarithm e based softmax, we use
+   *  2-based softmax here, i.e.,:
+   * 
+   *  y_i = 2^(x_i) / sum(2^x_j)
+   *
+   *  The relative output will be different here.
+   *  But mathematically, the gradient will be the same
+   *  with a log(2) scaling factor.
+   *
+   */
+
+void arm_softmax_q7(const q7_t * vec_in, const uint16_t dim_vec, q7_t * p_out)
+{
+    q31_t     sum;
+    int16_t   i;
+    uint8_t   shift;
+    q15_t     base;
+    base = -257;
+
+    /* We first search for the maximum */
+    for (i = 0; i < dim_vec; i++)
+    {
+        if (vec_in[i] > base)
+        {
+            base = vec_in[i];
+        }
+    }
+
+    /* 
+     * So the base is set to max-8, meaning 
+     * that we ignore really small values. 
+     * anyway, they will be 0 after shrinking to q7_t.
+     */
+    base = base - 8;
+
+    sum = 0;
+
+    for (i = 0; i < dim_vec; i++)
+    {
+        if (vec_in[i] > base) 
+        {
+            shift = (uint8_t)__USAT(vec_in[i] - base, 5);
+            sum += 0x1 << shift;
+        }
+    }
+
+    /* This is effectively (0x1 << 20) / sum */
+    int output_base = 0x100000 / sum;
+
+    /* 
+     * Final confidence will be output_base >> ( 13 - (vec_in[i] - base) )
+     * so 128 (0x1<<7) -> 100% confidence when sum = 0x1 << 8, output_base = 0x1 << 12 
+     * and vec_in[i]-base = 8
+     */
+    for (i = 0; i < dim_vec; i++) 
+    {
+        if (vec_in[i] > base) 
+        {
+            /* Here minimum value of 13+base-vec_in[i] will be 5 */
+            shift = (uint8_t)__USAT(13+base-vec_in[i], 5);
+            p_out[i] = (q7_t) __SSAT((output_base >> shift), 8);
+        } else {
+            p_out[i] = 0;
+        }
+    }
+}
+
+/**
+ * @} end of Softmax group
+ */