tmk_core/tool/mbed/mbed-sdk/libraries/dsp/cmsis_dsp/MatrixFunctions/arm_mat_mult_q15.c

   1 /* ----------------------------------------------------------------------
   2 * Copyright (C) 2010-2013 ARM Limited. All rights reserved.
   3 *
   4 * $Date:        17. January 2013
   5 * $Revision:    V1.4.1
   6 *
   7 * Project:          CMSIS DSP Library
   8 * Title:            arm_mat_mult_q15.c
   9 *
  10 * Description:   Q15 matrix multiplication.
  11 *
  12 * Target Processor: Cortex-M4/Cortex-M3/Cortex-M0
  13 *
  14 * Redistribution and use in source and binary forms, with or without
  15 * modification, are permitted provided that the following conditions
  16 * are met:
  17 *   - Redistributions of source code must retain the above copyright
  18 *     notice, this list of conditions and the following disclaimer.
  19 *   - Redistributions in binary form must reproduce the above copyright
  20 *     notice, this list of conditions and the following disclaimer in
  21 *     the documentation and/or other materials provided with the
  22 *     distribution.
  23 *   - Neither the name of ARM LIMITED nor the names of its contributors
  24 *     may be used to endorse or promote products derived from this
  25 *     software without specific prior written permission.
  26 *
  27 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  28 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  29 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
  30 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
  31 * COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
  32 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
  33 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
  34 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
  35 * CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  36 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN
  37 * ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
  38 * POSSIBILITY OF SUCH DAMAGE.
  39 * -------------------------------------------------------------------- */
  40
  41 #include "arm_math.h"
  42
  43 /**
  44  * @ingroup groupMatrix
  45  */
  46
  47 /**
  48  * @addtogroup MatrixMult
  49  * @{
  50  */
  51
  52
  53 /**
  54  * @brief Q15 matrix multiplication
  55  * @param[in]       *pSrcA points to the first input matrix structure
  56  * @param[in]       *pSrcB points to the second input matrix structure
  57  * @param[out]      *pDst points to output matrix structure
  58  * @param[in]           *pState points to the array for storing intermediate results
  59  * @return              The function returns either
  60  * <code>ARM_MATH_SIZE_MISMATCH</code> or <code>ARM_MATH_SUCCESS</code> based on the outcome of size checking.
  61  *
  62  * @details
  63  * <b>Scaling and Overflow Behavior:</b>
  64  *
  65  * \par
  66  * The function is implemented using a 64-bit internal accumulator. The inputs to the
  67  * multiplications are in 1.15 format and multiplications yield a 2.30 result.
  68  * The 2.30 intermediate
  69  * results are accumulated in a 64-bit accumulator in 34.30 format. This approach
  70  * provides 33 guard bits and there is no risk of overflow. The 34.30 result is then
  71  * truncated to 34.15 format by discarding the low 15 bits and then saturated to
  72  * 1.15 format.
  73  *
  74  * \par
  75  * 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.
  76  *
  77  */
  78
  79 arm_status arm_mat_mult_q15(
  80   const arm_matrix_instance_q15 * pSrcA,
  81   const arm_matrix_instance_q15 * pSrcB,
  82   arm_matrix_instance_q15 * pDst,
  83   q15_t * pState CMSIS_UNUSED)
  84 {
  85   q63_t sum;                                     /* accumulator */
  86
  87 #ifndef ARM_MATH_CM0_FAMILY
  88
  89   /* Run the below code for Cortex-M4 and Cortex-M3 */
  90
  91   q15_t *pSrcBT = pState;                        /* input data matrix pointer for transpose */
  92   q15_t *pInA = pSrcA->pData;                    /* input data matrix pointer A of Q15 type */
  93   q15_t *pInB = pSrcB->pData;                    /* input data matrix pointer B of Q15 type */
  94   q15_t *px;                                     /* Temporary output data matrix pointer */
  95   uint16_t numRowsA = pSrcA->numRows;            /* number of rows of input matrix A    */
  96   uint16_t numColsB = pSrcB->numCols;            /* number of columns of input matrix B */
  97   uint16_t numColsA = pSrcA->numCols;            /* number of columns of input matrix A */
  98   uint16_t numRowsB = pSrcB->numRows;            /* number of rows of input matrix A    */
  99   uint16_t col, i = 0u, row = numRowsB, colCnt;  /* loop counters */
 100   arm_status status;                             /* status of matrix multiplication */
 101
 102 #ifndef UNALIGNED_SUPPORT_DISABLE
 103
 104   q31_t in;                                      /* Temporary variable to hold the input value */
 105   q31_t pSourceA1, pSourceB1, pSourceA2, pSourceB2;
 106
 107 #else
 108
 109   q15_t in;                                      /* Temporary variable to hold the input value */
 110   q15_t inA1, inB1, inA2, inB2;
 111
 112 #endif  /*      #ifndef UNALIGNED_SUPPORT_DISABLE       */
 113
 114 #ifdef ARM_MATH_MATRIX_CHECK
 115   /* Check for matrix mismatch condition */
 116   if((pSrcA->numCols != pSrcB->numRows) ||
 117      (pSrcA->numRows != pDst->numRows) || (pSrcB->numCols != pDst->numCols))
 118   {
 119     /* Set status as ARM_MATH_SIZE_MISMATCH */
 120     status = ARM_MATH_SIZE_MISMATCH;
 121   }
 122   else
 123 #endif /*    #ifdef ARM_MATH_MATRIX_CHECK    */
 124   {
 125     /* Matrix transpose */
 126     do
 127     {
 128       /* Apply loop unrolling and exchange the columns with row elements */
 129       col = numColsB >> 2;
 130
 131       /* The pointer px is set to starting address of the column being processed */
 132       px = pSrcBT + i;
 133
 134       /* First part of the processing with loop unrolling.  Compute 4 outputs at a time.
 135        ** a second loop below computes the remaining 1 to 3 samples. */
 136       while(col > 0u)
 137       {
 138 #ifndef UNALIGNED_SUPPORT_DISABLE
 139
 140         /* Read two elements from the row */
 141         in = *__SIMD32(pInB)++;
 142
 143         /* Unpack and store one element in the destination */
 144 #ifndef ARM_MATH_BIG_ENDIAN
 145
 146         *px = (q15_t) in;
 147
 148 #else
 149
 150         *px = (q15_t) ((in & (q31_t) 0xffff0000) >> 16);
 151
 152 #endif /*    #ifndef ARM_MATH_BIG_ENDIAN    */
 153
 154         /* Update the pointer px to point to the next row of the transposed matrix */
 155         px += numRowsB;
 156
 157         /* Unpack and store the second element in the destination */
 158 #ifndef ARM_MATH_BIG_ENDIAN
 159
 160         *px = (q15_t) ((in & (q31_t) 0xffff0000) >> 16);
 161
 162 #else
 163
 164         *px = (q15_t) in;
 165
 166 #endif /*    #ifndef ARM_MATH_BIG_ENDIAN    */
 167
 168         /* Update the pointer px to point to the next row of the transposed matrix */
 169         px += numRowsB;
 170
 171         /* Read two elements from the row */
 172         in = *__SIMD32(pInB)++;
 173
 174         /* Unpack and store one element in the destination */
 175 #ifndef ARM_MATH_BIG_ENDIAN
 176
 177         *px = (q15_t) in;
 178
 179 #else
 180
 181         *px = (q15_t) ((in & (q31_t) 0xffff0000) >> 16);
 182
 183 #endif /*    #ifndef ARM_MATH_BIG_ENDIAN    */
 184
 185         /* Update the pointer px to point to the next row of the transposed matrix */
 186         px += numRowsB;
 187
 188         /* Unpack and store the second element in the destination */
 189
 190 #ifndef ARM_MATH_BIG_ENDIAN
 191
 192         *px = (q15_t) ((in & (q31_t) 0xffff0000) >> 16);
 193
 194 #else
 195
 196         *px = (q15_t) in;
 197
 198 #endif /*    #ifndef ARM_MATH_BIG_ENDIAN    */
 199
 200         /* Update the pointer px to point to the next row of the transposed matrix */
 201         px += numRowsB;
 202
 203 #else
 204
 205         /* Read one element from the row */
 206         in = *pInB++;
 207
 208         /* Store one element in the destination */
 209         *px = in;
 210
 211         /* Update the pointer px to point to the next row of the transposed matrix */
 212         px += numRowsB;
 213
 214         /* Read one element from the row */
 215         in = *pInB++;
 216
 217         /* Store one element in the destination */
 218         *px = in;
 219
 220         /* Update the pointer px to point to the next row of the transposed matrix */
 221         px += numRowsB;
 222
 223         /* Read one element from the row */
 224         in = *pInB++;
 225
 226         /* Store one element in the destination */
 227         *px = in;
 228
 229         /* Update the pointer px to point to the next row of the transposed matrix */
 230         px += numRowsB;
 231
 232         /* Read one element from the row */
 233         in = *pInB++;
 234
 235         /* Store one element in the destination */
 236         *px = in;
 237
 238         /* Update the pointer px to point to the next row of the transposed matrix */
 239         px += numRowsB;
 240
 241 #endif  /*      #ifndef UNALIGNED_SUPPORT_DISABLE       */
 242
 243        /* Decrement the column loop counter */
 244         col--;
 245       }
 246
 247       /* If the columns of pSrcB is not a multiple of 4, compute any remaining output samples here.
 248        ** No loop unrolling is used. */
 249       col = numColsB % 0x4u;
 250
 251       while(col > 0u)
 252       {
 253         /* Read and store the input element in the destination */
 254         *px = *pInB++;
 255
 256         /* Update the pointer px to point to the next row of the transposed matrix */
 257         px += numRowsB;
 258
 259         /* Decrement the column loop counter */
 260         col--;
 261       }
 262
 263       i++;
 264
 265       /* Decrement the row loop counter */
 266       row--;
 267
 268     } while(row > 0u);
 269
 270     /* Reset the variables for the usage in the following multiplication process */
 271     row = numRowsA;
 272     i = 0u;
 273     px = pDst->pData;
 274
 275     /* The following loop performs the dot-product of each row in pSrcA with each column in pSrcB */
 276     /* row loop */
 277     do
 278     {
 279       /* For every row wise process, the column loop counter is to be initiated */
 280       col = numColsB;
 281
 282       /* For every row wise process, the pIn2 pointer is set
 283        ** to the starting address of the transposed pSrcB data */
 284       pInB = pSrcBT;
 285
 286       /* column loop */
 287       do
 288       {
 289         /* Set the variable sum, that acts as accumulator, to zero */
 290         sum = 0;
 291
 292         /* Apply loop unrolling and compute 2 MACs simultaneously. */
 293         colCnt = numColsA >> 2;
 294
 295         /* Initiate the pointer pIn1 to point to the starting address of the column being processed */
 296         pInA = pSrcA->pData + i;
 297
 298
 299         /* matrix multiplication */
 300         while(colCnt > 0u)
 301         {
 302           /* c(m,n) = a(1,1)*b(1,1) + a(1,2) * b(2,1) + .... + a(m,p)*b(p,n) */
 303 #ifndef UNALIGNED_SUPPORT_DISABLE
 304
 305           /* read real and imag values from pSrcA and pSrcB buffer */
 306           pSourceA1 = *__SIMD32(pInA)++;
 307           pSourceB1 = *__SIMD32(pInB)++;
 308
 309           pSourceA2 = *__SIMD32(pInA)++;
 310           pSourceB2 = *__SIMD32(pInB)++;
 311
 312           /* Multiply and Accumlates */
 313           sum = __SMLALD(pSourceA1, pSourceB1, sum);
 314           sum = __SMLALD(pSourceA2, pSourceB2, sum);
 315
 316 #else
 317           /* read real and imag values from pSrcA and pSrcB buffer */
 318           inA1 = *pInA++;
 319           inB1 = *pInB++;
 320           inA2 = *pInA++;
 321           /* Multiply and Accumlates */
 322           sum += inA1 * inB1;
 323           inB2 = *pInB++;
 324
 325           inA1 = *pInA++;
 326           inB1 = *pInB++;
 327           /* Multiply and Accumlates */
 328           sum += inA2 * inB2;
 329           inA2 = *pInA++;
 330           inB2 = *pInB++;
 331
 332           /* Multiply and Accumlates */
 333           sum += inA1 * inB1;
 334           sum += inA2 * inB2;
 335
 336 #endif  /*      #ifndef UNALIGNED_SUPPORT_DISABLE       */
 337
 338           /* Decrement the loop counter */
 339           colCnt--;
 340         }
 341
 342         /* process remaining column samples */
 343         colCnt = numColsA & 3u;
 344
 345         while(colCnt > 0u)
 346         {
 347           /* c(m,n) = a(1,1)*b(1,1) + a(1,2) * b(2,1) + .... + a(m,p)*b(p,n) */
 348           sum += *pInA++ * *pInB++;
 349
 350           /* Decrement the loop counter */
 351           colCnt--;
 352         }
 353
 354         /* Saturate and store the result in the destination buffer */
 355         *px = (q15_t) (__SSAT((sum >> 15), 16));
 356         px++;
 357
 358         /* Decrement the column loop counter */
 359         col--;
 360
 361       } while(col > 0u);
 362
 363       i = i + numColsA;
 364
 365       /* Decrement the row loop counter */
 366       row--;
 367
 368     } while(row > 0u);
 369
 370 #else
 371
 372   /* Run the below code for Cortex-M0 */
 373
 374   q15_t *pIn1 = pSrcA->pData;                    /* input data matrix pointer A */
 375   q15_t *pIn2 = pSrcB->pData;                    /* input data matrix pointer B */
 376   q15_t *pInA = pSrcA->pData;                    /* input data matrix pointer A of Q15 type */
 377   q15_t *pInB = pSrcB->pData;                    /* input data matrix pointer B of Q15 type */
 378   q15_t *pOut = pDst->pData;                     /* output data matrix pointer */
 379   q15_t *px;                                     /* Temporary output data matrix pointer */
 380   uint16_t numColsB = pSrcB->numCols;            /* number of columns of input matrix B */
 381   uint16_t numColsA = pSrcA->numCols;            /* number of columns of input matrix A */
 382   uint16_t numRowsA = pSrcA->numRows;            /* number of rows of input matrix A    */
 383   uint16_t col, i = 0u, row = numRowsA, colCnt;  /* loop counters */
 384   arm_status status;                             /* status of matrix multiplication */
 385
 386 #ifdef ARM_MATH_MATRIX_CHECK
 387
 388   /* Check for matrix mismatch condition */
 389   if((pSrcA->numCols != pSrcB->numRows) ||
 390      (pSrcA->numRows != pDst->numRows) || (pSrcB->numCols != pDst->numCols))
 391   {
 392     /* Set status as ARM_MATH_SIZE_MISMATCH */
 393     status = ARM_MATH_SIZE_MISMATCH;
 394   }
 395   else
 396 #endif /*    #ifdef ARM_MATH_MATRIX_CHECK    */
 397
 398   {
 399     /* The following loop performs the dot-product of each row in pSrcA with each column in pSrcB */
 400     /* row loop */
 401     do
 402     {
 403       /* Output pointer is set to starting address of the row being processed */
 404       px = pOut + i;
 405
 406       /* For every row wise process, the column loop counter is to be initiated */
 407       col = numColsB;
 408
 409       /* For every row wise process, the pIn2 pointer is set
 410        ** to the starting address of the pSrcB data */
 411       pIn2 = pSrcB->pData;
 412
 413       /* column loop */
 414       do
 415       {
 416         /* Set the variable sum, that acts as accumulator, to zero */
 417         sum = 0;
 418
 419         /* Initiate the pointer pIn1 to point to the starting address of pSrcA */
 420         pIn1 = pInA;
 421
 422         /* Matrix A columns number of MAC operations are to be performed */
 423         colCnt = numColsA;
 424
 425         /* matrix multiplication */
 426         while(colCnt > 0u)
 427         {
 428           /* c(m,n) = a(1,1)*b(1,1) + a(1,2) * b(2,1) + .... + a(m,p)*b(p,n) */
 429           /* Perform the multiply-accumulates */
 430           sum += (q31_t) * pIn1++ * *pIn2;
 431           pIn2 += numColsB;
 432
 433           /* Decrement the loop counter */
 434           colCnt--;
 435         }
 436
 437         /* Convert the result from 34.30 to 1.15 format and store the saturated value in destination buffer */
 438         /* Saturate and store the result in the destination buffer */
 439         *px++ = (q15_t) __SSAT((sum >> 15), 16);
 440
 441         /* Decrement the column loop counter */
 442         col--;
 443
 444         /* Update the pointer pIn2 to point to the  starting address of the next column */
 445         pIn2 = pInB + (numColsB - col);
 446
 447       } while(col > 0u);
 448
 449       /* Update the pointer pSrcA to point to the  starting address of the next row */
 450       i = i + numColsB;
 451       pInA = pInA + numColsA;
 452
 453       /* Decrement the row loop counter */
 454       row--;
 455
 456     } while(row > 0u);
 457
 458 #endif /* #ifndef ARM_MATH_CM0_FAMILY */
 459     /* set status as ARM_MATH_SUCCESS */
 460     status = ARM_MATH_SUCCESS;
 461   }
 462
 463   /* Return to application */
 464   return (status);
 465 }
 466
 467 /**
 468  * @} end of MatrixMult group
 469  */