Subversion Repositories AFRtranscoder

/*

 * 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

 */