Subversion Repositories DashDisplay

/* ----------------------------------------------------------------------

 * Project:      CMSIS DSP Library

 * Title:        arm_mat_mult_fast_q15.c

 * Description:  Q15 matrix multiplication (fast variant)

 *

 * $Date:        27. January 2017

 * $Revision:    V.1.5.1

 *

 * Target Processor: Cortex-M cores

 * -------------------------------------------------------------------- */

/*

 * Copyright (C) 2010-2017 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.

 */

#include "arm_math.h"

/**

 * @ingroup groupMatrix

 */

/**

 * @addtogroup MatrixMult

 * @{

 */

/**

 * @brief Q15 matrix multiplication (fast variant) for Cortex-M3 and Cortex-M4

 * @param[in]       *pSrcA points to the first input matrix structure

 * @param[in]       *pSrcB points to the second input matrix structure

 * @param[out]      *pDst points to output matrix structure

 * @param[in]       *pState points to the array for storing intermediate results

 * @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>Scaling and Overflow Behavior:</b>

 *

 * \par

 * The difference between the function arm_mat_mult_q15() and this fast variant is that

 * the fast variant use a 32-bit rather than a 64-bit accumulator.

 * The result of each 1.15 x 1.15 multiplication is truncated to

 * 2.30 format. These intermediate results are accumulated in a 32-bit register in 2.30

 * format. Finally, the accumulator is saturated and converted to a 1.15 result.

 *

 * \par

 * The fast version has the same overflow behavior as the standard version but provides

 * less precision since it discards the low 16 bits of each multiplication result.

 * In order to avoid overflows completely the input signals must be scaled down.

 * Scale down one of the input matrices by log2(numColsA) bits to

 * avoid overflows, as a total of numColsA additions are computed internally for each

 * output element.

 *

 * \par

 * See <code>arm_mat_mult_q15()</code> for a slower implementation of this function

 * which uses 64-bit accumulation to provide higher precision.

 */

arm_status arm_mat_mult_fast_q15(

  const arm_matrix_instance_q15 * pSrcA,

  const arm_matrix_instance_q15 * pSrcB,

  arm_matrix_instance_q15 * pDst,

  q15_t * pState)

{

  q31_t sum;                                     /* accumulator */

  q15_t *pSrcBT = pState;                        /* input data matrix pointer for transpose */

  q15_t *pInA = pSrcA->pData;                    /* input data matrix pointer A of Q15 type */

  q15_t *pInB = pSrcB->pData;                    /* input data matrix pointer B of Q15 type */

  q15_t *px;                                     /* Temporary output data matrix pointer */

  uint16_t numRowsA = pSrcA->numRows;            /* number of rows of input matrix A    */

  uint16_t numColsB = pSrcB->numCols;            /* number of columns of input matrix B */

  uint16_t numColsA = pSrcA->numCols;            /* number of columns of input matrix A */

  uint16_t numRowsB = pSrcB->numRows;            /* number of rows of input matrix A    */

  uint32_t col, i = 0U, row = numRowsB, colCnt;  /* loop counters */

  arm_status status;                             /* status of matrix multiplication */

#ifndef UNALIGNED_SUPPORT_DISABLE

  q31_t in;                                      /* Temporary variable to hold the input value */

  q31_t inA1, inA2, inB1, inB2;

  q31_t sum2, sum3, sum4;

  q15_t *pInA2, *pInB2, *px2;

  uint32_t j = 0;

#else

  q15_t in;                                      /* Temporary variable to hold the input value */

  q15_t inA1, inA2, inB1, inB2;

#endif /* #ifndef UNALIGNED_SUPPORT_DISABLE */

#ifdef ARM_MATH_MATRIX_CHECK

  /* Check for matrix mismatch condition */

  if ((pSrcA->numCols != pSrcB->numRows) ||

     (pSrcA->numRows != pDst->numRows) || (pSrcB->numCols != pDst->numCols))

  {

    /* Set status as ARM_MATH_SIZE_MISMATCH */

    status = ARM_MATH_SIZE_MISMATCH;

  }

  else

#endif

  {

    /* Matrix transpose */

    do

    {

      /* Apply loop unrolling and exchange the columns with row elements */

      col = numColsB >> 2;

      /* The pointer px is set to starting address of the column being processed */

      px = pSrcBT + i;

      /* 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 (col > 0U)

      {

#ifndef UNALIGNED_SUPPORT_DISABLE

        /* Read two elements from the row */

        in = *__SIMD32(pInB)++;

        /* Unpack and store one element in the destination */

#ifndef ARM_MATH_BIG_ENDIAN

        *px = (q15_t) in;

#else

        *px = (q15_t) ((in & (q31_t) 0xffff0000) >> 16);

#endif /*    #ifndef ARM_MATH_BIG_ENDIAN    */

        /* Update the pointer px to point to the next row of the transposed matrix */

        px += numRowsB;

        /* Unpack and store the second element in the destination */

#ifndef ARM_MATH_BIG_ENDIAN

        *px = (q15_t) ((in & (q31_t) 0xffff0000) >> 16);

#else

        *px = (q15_t) in;

#endif /*    #ifndef ARM_MATH_BIG_ENDIAN    */

        /* Update the pointer px to point to the next row of the transposed matrix */

        px += numRowsB;

        /* Read two elements from the row */

        in = *__SIMD32(pInB)++;

        /* Unpack and store one element in the destination */

#ifndef ARM_MATH_BIG_ENDIAN

        *px = (q15_t) in;

#else

        *px = (q15_t) ((in & (q31_t) 0xffff0000) >> 16);

#endif /*    #ifndef ARM_MATH_BIG_ENDIAN    */

        /* Update the pointer px to point to the next row of the transposed matrix */

        px += numRowsB;

        /* Unpack and store the second element in the destination */

#ifndef ARM_MATH_BIG_ENDIAN

        *px = (q15_t) ((in & (q31_t) 0xffff0000) >> 16);

#else

        *px = (q15_t) in;

#endif /*    #ifndef ARM_MATH_BIG_ENDIAN    */

#else

        /* Read one element from the row */

        in = *pInB++;

        /* Store one element in the destination */

        *px = in;

        /* Update the pointer px to point to the next row of the transposed matrix */

        px += numRowsB;

        /* Read one element from the row */

        in = *pInB++;

        /* Store one element in the destination */

        *px = in;

        /* Update the pointer px to point to the next row of the transposed matrix */

        px += numRowsB;

        /* Read one element from the row */

        in = *pInB++;

        /* Store one element in the destination */

        *px = in;

        /* Update the pointer px to point to the next row of the transposed matrix */

        px += numRowsB;

        /* Read one element from the row */

        in = *pInB++;

        /* Store one element in the destination */

        *px = in;

#endif /* #ifndef UNALIGNED_SUPPORT_DISABLE */

        /* Update the pointer px to point to the next row of the transposed matrix */

        px += numRowsB;

        /* Decrement the column loop counter */

        col--;

      }

      /* If the columns of pSrcB is not a multiple of 4, compute any remaining output samples here.

       ** No loop unrolling is used. */

      col = numColsB % 0x4U;

      while (col > 0U)

      {

        /* Read and store the input element in the destination */

        *px = *pInB++;

        /* Update the pointer px to point to the next row of the transposed matrix */

        px += numRowsB;

        /* Decrement the column loop counter */

        col--;

      }

      i++;

      /* Decrement the row loop counter */

      row--;

    } while (row > 0U);

    /* Reset the variables for the usage in the following multiplication process */

    row = numRowsA;

    i = 0U;

    px = pDst->pData;

#ifndef UNALIGNED_SUPPORT_DISABLE

    /* Process two rows from matrix A at a time and output two rows at a time */

    row = row >> 1;

    px2 = px + numColsB;

#endif

    /* The following loop performs the dot-product of each row in pSrcA with each column in pSrcB */

    /* row loop */

    while (row > 0U)

    {

      /* For every row wise process, the column loop counter is to be initiated */

      col = numColsB;

      /* For every row wise process, the pIn2 pointer is set

       ** to the starting address of the transposed pSrcB data */

      pInB = pSrcBT;

#ifndef UNALIGNED_SUPPORT_DISABLE

      /* Process two (transposed) columns from matrix B at a time */

      col = col >> 1;

      j = 0;

#endif

      /* column loop */

      while (col > 0U)

      {

        /* Set the variable sum, that acts as accumulator, to zero */

        sum = 0;

        /* Initiate the pointer pInA to point to the starting address of the column being processed */

        pInA = pSrcA->pData + i;

#ifndef UNALIGNED_SUPPORT_DISABLE

        sum2 = 0;

        sum3 = 0;

        sum4 = 0;

        pInB  = pSrcBT + j;

        pInA2 = pInA + numColsA;

        pInB2 = pInB + numRowsB;

        /* Read in two elements at once - alows dual MAC instruction */

        colCnt = numColsA >> 1;

#else

        colCnt = numColsA >> 2;

#endif

        /* matrix multiplication */

        while (colCnt > 0U)

        {

          /* c(m,n) = a(1,1)*b(1,1) + a(1,2) * b(2,1) + .... + a(m,p)*b(p,n) */

#ifndef UNALIGNED_SUPPORT_DISABLE

          inA1 = *__SIMD32(pInA)++;

          inB1 = *__SIMD32(pInB)++;

          inA2 = *__SIMD32(pInA2)++;

          inB2 = *__SIMD32(pInB2)++;

          sum  = __SMLAD(inA1, inB1, sum);

          sum2 = __SMLAD(inA1, inB2, sum2);

          sum3 = __SMLAD(inA2, inB1, sum3);

          sum4 = __SMLAD(inA2, inB2, sum4);

#else

          inA1 = *pInA;

          inB1 = *pInB;

          sum += inA1 * inB1;

          inA2 = pInA[1];

          inB2 = pInB[1];

          sum += inA2 * inB2;

          inA1 = pInA[2];

          inB1 = pInB[2];

          sum += inA1 * inB1;

          inA2 = pInA[3];

          inB2 = pInB[3];

          sum += inA2 * inB2;

          pInA += 4;

          pInB += 4;

#endif /* #ifndef UNALIGNED_SUPPORT_DISABLE */

          /* Decrement the loop counter */

          colCnt--;

        }

        /* process odd column samples */

#ifndef UNALIGNED_SUPPORT_DISABLE

        if (numColsA & 1U) {

          inA1 = *pInA++;

          inB1 = *pInB++;

          inA2 = *pInA2++;

          inB2 = *pInB2++;

          sum  += inA1 * inB1;

          sum2 += inA1 * inB2;

          sum3 += inA2 * inB1;

          sum4 += inA2 * inB2;

        }

#else

        colCnt = numColsA % 0x4U;

        while (colCnt > 0U)

        {

          /* c(m,n) = a(1,1)*b(1,1) + a(1,2) * b(2,1) + .... + a(m,p)*b(p,n) */

          sum += (q31_t) (*pInA++) * (*pInB++);

          colCnt--;

        }

#endif

        /* Saturate and store the result in the destination buffer */

        *px++  = (q15_t) (sum >> 15);

#ifndef UNALIGNED_SUPPORT_DISABLE

        *px++  = (q15_t) (sum2 >> 15);

        *px2++ = (q15_t) (sum3 >> 15);

        *px2++ = (q15_t) (sum4 >> 15);

        j += numRowsB * 2;

#endif

        /* Decrement the column loop counter */

        col--;

      }

      i = i + numColsA;

#ifndef UNALIGNED_SUPPORT_DISABLE

      i = i + numColsA;

      px = px2 + (numColsB & 1U);

      px2 = px + numColsB;

#endif

      /* Decrement the row loop counter */

      row--;

    }

    /* Compute any remaining odd row/column below */

#ifndef UNALIGNED_SUPPORT_DISABLE

    /* Compute remaining output column */

    if (numColsB & 1U) {

      /* Avoid redundant computation of last element */

      row = numRowsA & (~0x1);

      /* Point to remaining unfilled column in output matrix */

      px = pDst->pData+numColsB-1;

      pInA = pSrcA->pData;

      /* row loop */

      while (row > 0)

      {

        /* point to last column in matrix B */

        pInB  = pSrcBT + numRowsB*(numColsB-1);

        /* Set the variable sum, that acts as accumulator, to zero */

        sum  = 0;

        /* Compute 4 columns at once */

        colCnt = numColsA >> 2;

        /* matrix multiplication */

        while (colCnt > 0U)

        {

          inA1 = *__SIMD32(pInA)++;

          inA2 = *__SIMD32(pInA)++;

          inB1 = *__SIMD32(pInB)++;

          inB2 = *__SIMD32(pInB)++;

          sum  = __SMLAD(inA1, inB1, sum);

          sum  = __SMLAD(inA2, inB2, sum);

          /* Decrement the loop counter */

          colCnt--;

        }

        colCnt = numColsA & 3U;

        while (colCnt > 0U) {

          sum += (q31_t) (*pInA++) * (*pInB++);

          colCnt--;

        }

        /* Store the result in the destination buffer */

        *px  = (q15_t) (sum  >> 15);

        px += numColsB;

        /* Decrement the row loop counter */

        row--;

      }

    }

    /* Compute remaining output row */

    if (numRowsA & 1U) {

      /* point to last row in output matrix */

      px = pDst->pData+(numColsB)*(numRowsA-1);

      pInB  = pSrcBT;

      col = numColsB;

      i = 0U;

      /* col loop */

      while (col > 0)

      {

        /* point to last row in matrix A */

        pInA = pSrcA->pData + (numRowsA-1)*numColsA;

        /* Set the variable sum, that acts as accumulator, to zero */

        sum  = 0;

        /* Compute 4 columns at once */

        colCnt = numColsA >> 2;

        /* matrix multiplication */

        while (colCnt > 0U)

        {

          inA1 = *__SIMD32(pInA)++;

          inA2 = *__SIMD32(pInA)++;

          inB1 = *__SIMD32(pInB)++;

          inB2 = *__SIMD32(pInB)++;

          sum  = __SMLAD(inA1, inB1, sum);

          sum  = __SMLAD(inA2, inB2, sum);

          /* Decrement the loop counter */

          colCnt--;

        }

        colCnt = numColsA & 3U;

        while (colCnt > 0U) {

          sum += (q31_t) (*pInA++) * (*pInB++);

          colCnt--;

        }

        /* Store the result in the destination buffer */

        *px++  = (q15_t) (sum  >> 15);

        /* Decrement the col loop counter */

        col--;

      }

    }

#endif /* #ifndef UNALIGNED_SUPPORT_DISABLE */

    /* set status as ARM_MATH_SUCCESS */

    status = ARM_MATH_SUCCESS;

  }

  /* Return to application */

  return (status);

}

/**

 * @} end of MatrixMult group

 */