arm_mat_mult_q15.c

00001 /* ----------------------------------------------------------------------
00002  * Project:      CMSIS DSP Library
00003  * Title:        arm_mat_mult_q15.c
00004  * Description:  Q15 matrix multiplication
00005  *
00006  * $Date:        27. January 2017
00007  * $Revision:    V.1.5.1
00008  *
00009  * Target Processor: Cortex-M cores
00010  * -------------------------------------------------------------------- */
00011 /*
00012  * Copyright (C) 2010-2017 ARM Limited or its affiliates. All rights reserved.
00013  *
00014  * SPDX-License-Identifier: Apache-2.0
00015  *
00016  * Licensed under the Apache License, Version 2.0 (the License); you may
00017  * not use this file except in compliance with the License.
00018  * You may obtain a copy of the License at
00019  *
00020  * www.apache.org/licenses/LICENSE-2.0
00021  *
00022  * Unless required by applicable law or agreed to in writing, software
00023  * distributed under the License is distributed on an AS IS BASIS, WITHOUT
00024  * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
00025  * See the License for the specific language governing permissions and
00026  * limitations under the License.
00027  */
00028 
00029 #include "arm_math.h"
00030 
00031 /**
00032  * @ingroup groupMatrix
00033  */
00034 
00035 /**
00036  * @addtogroup MatrixMult
00037  * @{
00038  */
00039 
00040 
00041 /**
00042  * @brief Q15 matrix multiplication
00043  * @param[in]       *pSrcA points to the first input matrix structure
00044  * @param[in]       *pSrcB points to the second input matrix structure
00045  * @param[out]      *pDst points to output matrix structure
00046  * @param[in]       *pState points to the array for storing intermediate results (Unused)
00047  * @return          The function returns either
00048  * <code>ARM_MATH_SIZE_MISMATCH</code> or <code>ARM_MATH_SUCCESS</code> based on the outcome of size checking.
00049  *
00050  * @details
00051  * <b>Scaling and Overflow Behavior:</b>
00052  *
00053  * \par
00054  * The function is implemented using a 64-bit internal accumulator. The inputs to the
00055  * multiplications are in 1.15 format and multiplications yield a 2.30 result.
00056  * The 2.30 intermediate
00057  * results are accumulated in a 64-bit accumulator in 34.30 format. This approach
00058  * provides 33 guard bits and there is no risk of overflow. The 34.30 result is then
00059  * truncated to 34.15 format by discarding the low 15 bits and then saturated to
00060  * 1.15 format.
00061  *
00062  * \par
00063  * Refer to <code>arm_mat_mult_fast_q15()</code> for a faster but less precise version of this function for Cortex-M3 and Cortex-M4.
00064  *
00065  */
00066 
00067 arm_status arm_mat_mult_q15(
00068   const arm_matrix_instance_q15 * pSrcA,
00069   const arm_matrix_instance_q15 * pSrcB,
00070   arm_matrix_instance_q15 * pDst,
00071   q15_t * pState)
00072 {
00073   q63_t sum;                                     /* accumulator */
00074 
00075 #if defined (ARM_MATH_DSP)
00076 
00077   /* Run the below code for Cortex-M4 and Cortex-M3 */
00078 
00079   q15_t *pSrcBT = pState;                        /* input data matrix pointer for transpose */
00080   q15_t *pInA = pSrcA->pData;                    /* input data matrix pointer A of Q15 type */
00081   q15_t *pInB = pSrcB->pData;                    /* input data matrix pointer B of Q15 type */
00082   q15_t *px;                                     /* Temporary output data matrix pointer */
00083   uint16_t numRowsA = pSrcA->numRows;            /* number of rows of input matrix A    */
00084   uint16_t numColsB = pSrcB->numCols;            /* number of columns of input matrix B */
00085   uint16_t numColsA = pSrcA->numCols;            /* number of columns of input matrix A */
00086   uint16_t numRowsB = pSrcB->numRows;            /* number of rows of input matrix A    */
00087   uint16_t col, i = 0U, row = numRowsB, colCnt;  /* loop counters */
00088   arm_status status;                             /* status of matrix multiplication */
00089 
00090 #ifndef UNALIGNED_SUPPORT_DISABLE
00091 
00092   q31_t in;                                      /* Temporary variable to hold the input value */
00093   q31_t pSourceA1, pSourceB1, pSourceA2, pSourceB2;
00094 
00095 #else
00096 
00097   q15_t in;                                      /* Temporary variable to hold the input value */
00098   q15_t inA1, inB1, inA2, inB2;
00099 
00100 #endif /* #ifndef UNALIGNED_SUPPORT_DISABLE */
00101 
00102 #ifdef ARM_MATH_MATRIX_CHECK
00103   /* Check for matrix mismatch condition */
00104   if ((pSrcA->numCols != pSrcB->numRows) ||
00105      (pSrcA->numRows != pDst->numRows) || (pSrcB->numCols != pDst->numCols))
00106   {
00107     /* Set status as ARM_MATH_SIZE_MISMATCH */
00108     status = ARM_MATH_SIZE_MISMATCH;
00109   }
00110   else
00111 #endif /*    #ifdef ARM_MATH_MATRIX_CHECK    */
00112   {
00113     /* Matrix transpose */
00114     do
00115     {
00116       /* Apply loop unrolling and exchange the columns with row elements */
00117       col = numColsB >> 2;
00118 
00119       /* The pointer px is set to starting address of the column being processed */
00120       px = pSrcBT + i;
00121 
00122       /* First part of the processing with loop unrolling.  Compute 4 outputs at a time.
00123        ** a second loop below computes the remaining 1 to 3 samples. */
00124       while (col > 0U)
00125       {
00126 #ifndef UNALIGNED_SUPPORT_DISABLE
00127 
00128         /* Read two elements from the row */
00129         in = *__SIMD32(pInB)++;
00130 
00131         /* Unpack and store one element in the destination */
00132 #ifndef ARM_MATH_BIG_ENDIAN
00133 
00134         *px = (q15_t) in;
00135 
00136 #else
00137 
00138         *px = (q15_t) ((in & (q31_t) 0xffff0000) >> 16);
00139 
00140 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
00141 
00142         /* Update the pointer px to point to the next row of the transposed matrix */
00143         px += numRowsB;
00144 
00145         /* Unpack and store the second element in the destination */
00146 #ifndef ARM_MATH_BIG_ENDIAN
00147 
00148         *px = (q15_t) ((in & (q31_t) 0xffff0000) >> 16);
00149 
00150 #else
00151 
00152         *px = (q15_t) in;
00153 
00154 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
00155 
00156         /* Update the pointer px to point to the next row of the transposed matrix */
00157         px += numRowsB;
00158 
00159         /* Read two elements from the row */
00160         in = *__SIMD32(pInB)++;
00161 
00162         /* Unpack and store one element in the destination */
00163 #ifndef ARM_MATH_BIG_ENDIAN
00164 
00165         *px = (q15_t) in;
00166 
00167 #else
00168 
00169         *px = (q15_t) ((in & (q31_t) 0xffff0000) >> 16);
00170 
00171 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
00172 
00173         /* Update the pointer px to point to the next row of the transposed matrix */
00174         px += numRowsB;
00175 
00176         /* Unpack and store the second element in the destination */
00177 
00178 #ifndef ARM_MATH_BIG_ENDIAN
00179 
00180         *px = (q15_t) ((in & (q31_t) 0xffff0000) >> 16);
00181 
00182 #else
00183 
00184         *px = (q15_t) in;
00185 
00186 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
00187 
00188         /* Update the pointer px to point to the next row of the transposed matrix */
00189         px += numRowsB;
00190 
00191 #else
00192 
00193         /* Read one element from the row */
00194         in = *pInB++;
00195 
00196         /* Store one element in the destination */
00197         *px = in;
00198 
00199         /* Update the pointer px to point to the next row of the transposed matrix */
00200         px += numRowsB;
00201 
00202         /* Read one element from the row */
00203         in = *pInB++;
00204 
00205         /* Store one element in the destination */
00206         *px = in;
00207 
00208         /* Update the pointer px to point to the next row of the transposed matrix */
00209         px += numRowsB;
00210 
00211         /* Read one element from the row */
00212         in = *pInB++;
00213 
00214         /* Store one element in the destination */
00215         *px = in;
00216 
00217         /* Update the pointer px to point to the next row of the transposed matrix */
00218         px += numRowsB;
00219 
00220         /* Read one element from the row */
00221         in = *pInB++;
00222 
00223         /* Store one element in the destination */
00224         *px = in;
00225 
00226         /* Update the pointer px to point to the next row of the transposed matrix */
00227         px += numRowsB;
00228 
00229 #endif /* #ifndef UNALIGNED_SUPPORT_DISABLE */
00230 
00231        /* Decrement the column loop counter */
00232         col--;
00233       }
00234 
00235       /* If the columns of pSrcB is not a multiple of 4, compute any remaining output samples here.
00236        ** No loop unrolling is used. */
00237       col = numColsB % 0x4U;
00238 
00239       while (col > 0U)
00240       {
00241         /* Read and store the input element in the destination */
00242         *px = *pInB++;
00243 
00244         /* Update the pointer px to point to the next row of the transposed matrix */
00245         px += numRowsB;
00246 
00247         /* Decrement the column loop counter */
00248         col--;
00249       }
00250 
00251       i++;
00252 
00253       /* Decrement the row loop counter */
00254       row--;
00255 
00256     } while (row > 0U);
00257 
00258     /* Reset the variables for the usage in the following multiplication process */
00259     row = numRowsA;
00260     i = 0U;
00261     px = pDst->pData;
00262 
00263     /* The following loop performs the dot-product of each row in pSrcA with each column in pSrcB */
00264     /* row loop */
00265     do
00266     {
00267       /* For every row wise process, the column loop counter is to be initiated */
00268       col = numColsB;
00269 
00270       /* For every row wise process, the pIn2 pointer is set
00271        ** to the starting address of the transposed pSrcB data */
00272       pInB = pSrcBT;
00273 
00274       /* column loop */
00275       do
00276       {
00277         /* Set the variable sum, that acts as accumulator, to zero */
00278         sum = 0;
00279 
00280         /* Apply loop unrolling and compute 2 MACs simultaneously. */
00281         colCnt = numColsA >> 2;
00282 
00283         /* Initiate the pointer pIn1 to point to the starting address of the column being processed */
00284         pInA = pSrcA->pData + i;
00285 
00286 
00287         /* matrix multiplication */
00288         while (colCnt > 0U)
00289         {
00290           /* c(m,n) = a(1,1)*b(1,1) + a(1,2) * b(2,1) + .... + a(m,p)*b(p,n) */
00291 #ifndef UNALIGNED_SUPPORT_DISABLE
00292 
00293           /* read real and imag values from pSrcA and pSrcB buffer */
00294           pSourceA1 = *__SIMD32(pInA)++;
00295           pSourceB1 = *__SIMD32(pInB)++;
00296 
00297           pSourceA2 = *__SIMD32(pInA)++;
00298           pSourceB2 = *__SIMD32(pInB)++;
00299 
00300           /* Multiply and Accumlates */
00301           sum = __SMLALD(pSourceA1, pSourceB1, sum);
00302           sum = __SMLALD(pSourceA2, pSourceB2, sum);
00303 
00304 #else
00305           /* read real and imag values from pSrcA and pSrcB buffer */
00306           inA1 = *pInA++;
00307           inB1 = *pInB++;
00308           inA2 = *pInA++;
00309           /* Multiply and Accumlates */
00310           sum += inA1 * inB1;
00311           inB2 = *pInB++;
00312 
00313           inA1 = *pInA++;
00314           inB1 = *pInB++;
00315           /* Multiply and Accumlates */
00316           sum += inA2 * inB2;
00317           inA2 = *pInA++;
00318           inB2 = *pInB++;
00319 
00320           /* Multiply and Accumlates */
00321           sum += inA1 * inB1;
00322           sum += inA2 * inB2;
00323 
00324 #endif /* #ifndef UNALIGNED_SUPPORT_DISABLE */
00325 
00326           /* Decrement the loop counter */
00327           colCnt--;
00328         }
00329 
00330         /* process remaining column samples */
00331         colCnt = numColsA & 3U;
00332 
00333         while (colCnt > 0U)
00334         {
00335           /* c(m,n) = a(1,1)*b(1,1) + a(1,2) * b(2,1) + .... + a(m,p)*b(p,n) */
00336           sum += *pInA++ * *pInB++;
00337 
00338           /* Decrement the loop counter */
00339           colCnt--;
00340         }
00341 
00342         /* Saturate and store the result in the destination buffer */
00343         *px = (q15_t) (__SSAT((sum >> 15), 16));
00344         px++;
00345 
00346         /* Decrement the column loop counter */
00347         col--;
00348 
00349       } while (col > 0U);
00350 
00351       i = i + numColsA;
00352 
00353       /* Decrement the row loop counter */
00354       row--;
00355 
00356     } while (row > 0U);
00357 
00358 #else
00359 
00360   /* Run the below code for Cortex-M0 */
00361 
00362   q15_t *pIn1 = pSrcA->pData;                    /* input data matrix pointer A */
00363   q15_t *pIn2 = pSrcB->pData;                    /* input data matrix pointer B */
00364   q15_t *pInA = pSrcA->pData;                    /* input data matrix pointer A of Q15 type */
00365   q15_t *pInB = pSrcB->pData;                    /* input data matrix pointer B of Q15 type */
00366   q15_t *pOut = pDst->pData;                     /* output data matrix pointer */
00367   q15_t *px;                                     /* Temporary output data matrix pointer */
00368   uint16_t numColsB = pSrcB->numCols;            /* number of columns of input matrix B */
00369   uint16_t numColsA = pSrcA->numCols;            /* number of columns of input matrix A */
00370   uint16_t numRowsA = pSrcA->numRows;            /* number of rows of input matrix A    */
00371   uint16_t col, i = 0U, row = numRowsA, colCnt;  /* loop counters */
00372   arm_status status;                             /* status of matrix multiplication */
00373 
00374 #ifdef ARM_MATH_MATRIX_CHECK
00375 
00376   /* Check for matrix mismatch condition */
00377   if ((pSrcA->numCols != pSrcB->numRows) ||
00378      (pSrcA->numRows != pDst->numRows) || (pSrcB->numCols != pDst->numCols))
00379   {
00380     /* Set status as ARM_MATH_SIZE_MISMATCH */
00381     status = ARM_MATH_SIZE_MISMATCH;
00382   }
00383   else
00384 #endif /* #ifdef ARM_MATH_MATRIX_CHECK */
00385 
00386   {
00387     /* The following loop performs the dot-product of each row in pSrcA with each column in pSrcB */
00388     /* row loop */
00389     do
00390     {
00391       /* Output pointer is set to starting address of the row being processed */
00392       px = pOut + i;
00393 
00394       /* For every row wise process, the column loop counter is to be initiated */
00395       col = numColsB;
00396 
00397       /* For every row wise process, the pIn2 pointer is set
00398        ** to the starting address of the pSrcB data */
00399       pIn2 = pSrcB->pData;
00400 
00401       /* column loop */
00402       do
00403       {
00404         /* Set the variable sum, that acts as accumulator, to zero */
00405         sum = 0;
00406 
00407         /* Initiate the pointer pIn1 to point to the starting address of pSrcA */
00408         pIn1 = pInA;
00409 
00410         /* Matrix A columns number of MAC operations are to be performed */
00411         colCnt = numColsA;
00412 
00413         /* matrix multiplication */
00414         while (colCnt > 0U)
00415         {
00416           /* c(m,n) = a(1,1)*b(1,1) + a(1,2) * b(2,1) + .... + a(m,p)*b(p,n) */
00417           /* Perform the multiply-accumulates */
00418           sum += (q31_t) * pIn1++ * *pIn2;
00419           pIn2 += numColsB;
00420 
00421           /* Decrement the loop counter */
00422           colCnt--;
00423         }
00424 
00425         /* Convert the result from 34.30 to 1.15 format and store the saturated value in destination buffer */
00426         /* Saturate and store the result in the destination buffer */
00427         *px++ = (q15_t) __SSAT((sum >> 15), 16);
00428 
00429         /* Decrement the column loop counter */
00430         col--;
00431 
00432         /* Update the pointer pIn2 to point to the  starting address of the next column */
00433         pIn2 = pInB + (numColsB - col);
00434 
00435       } while (col > 0U);
00436 
00437       /* Update the pointer pSrcA to point to the  starting address of the next row */
00438       i = i + numColsB;
00439       pInA = pInA + numColsA;
00440 
00441       /* Decrement the row loop counter */
00442       row--;
00443 
00444     } while (row > 0U);
00445 
00446 #endif /* #if defined (ARM_MATH_DSP) */
00447     /* set status as ARM_MATH_SUCCESS */
00448     status = ARM_MATH_SUCCESS;
00449   }
00450 
00451   /* Return to application */
00452   return (status);
00453 }
00454 
00455 /**
00456  * @} end of MatrixMult group
00457  */
00458