]>
git.gir.st - tmk_keyboard.git/blob - tmk_core/tool/mbed/mbed-sdk/libraries/dsp/cmsis_dsp/MatrixFunctions/arm_mat_mult_fast_q15.c
1 /* ----------------------------------------------------------------------
2 * Copyright (C) 2010-2013 ARM Limited. All rights reserved.
4 * $Date: 17. January 2013
7 * Project: CMSIS DSP Library
8 * Title: arm_mat_mult_fast_q15.c
10 * Description: Q15 matrix multiplication (fast variant)
12 * Target Processor: Cortex-M4/Cortex-M3
14 * Redistribution and use in source and binary forms, with or without
15 * modification, are permitted provided that the following conditions
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
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.
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 * -------------------------------------------------------------------- */
44 * @ingroup groupMatrix
48 * @addtogroup MatrixMult
54 * @brief Q15 matrix multiplication (fast variant) for Cortex-M3 and Cortex-M4
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.
63 * <b>Scaling and Overflow Behavior:</b>
66 * The difference between the function arm_mat_mult_q15() and this fast variant is that
67 * the fast variant use a 32-bit rather than a 64-bit accumulator.
68 * The result of each 1.15 x 1.15 multiplication is truncated to
69 * 2.30 format. These intermediate results are accumulated in a 32-bit register in 2.30
70 * format. Finally, the accumulator is saturated and converted to a 1.15 result.
73 * The fast version has the same overflow behavior as the standard version but provides
74 * less precision since it discards the low 16 bits of each multiplication result.
75 * In order to avoid overflows completely the input signals must be scaled down.
76 * Scale down one of the input matrices by log2(numColsA) bits to
77 * avoid overflows, as a total of numColsA additions are computed internally for each
81 * See <code>arm_mat_mult_q15()</code> for a slower implementation of this function
82 * which uses 64-bit accumulation to provide higher precision.
85 arm_status
arm_mat_mult_fast_q15(
86 const arm_matrix_instance_q15
* pSrcA
,
87 const arm_matrix_instance_q15
* pSrcB
,
88 arm_matrix_instance_q15
* pDst
,
91 q31_t sum
; /* accumulator */
92 q15_t
*pSrcBT
= pState
; /* input data matrix pointer for transpose */
93 q15_t
*pInA
= pSrcA
->pData
; /* input data matrix pointer A of Q15 type */
94 q15_t
*pInB
= pSrcB
->pData
; /* input data matrix pointer B of Q15 type */
95 q15_t
*px
; /* Temporary output data matrix pointer */
96 uint16_t numRowsA
= pSrcA
->numRows
; /* number of rows of input matrix A */
97 uint16_t numColsB
= pSrcB
->numCols
; /* number of columns of input matrix B */
98 uint16_t numColsA
= pSrcA
->numCols
; /* number of columns of input matrix A */
99 uint16_t numRowsB
= pSrcB
->numRows
; /* number of rows of input matrix A */
100 uint16_t col
, i
= 0u, row
= numRowsB
, colCnt
; /* loop counters */
101 arm_status status
; /* status of matrix multiplication */
103 #ifndef UNALIGNED_SUPPORT_DISABLE
105 q31_t in
; /* Temporary variable to hold the input value */
106 q31_t inA1
, inA2
, inB1
, inB2
;
110 q15_t in
; /* Temporary variable to hold the input value */
111 q15_t inA1
, inA2
, inB1
, inB2
;
113 #endif /* #ifndef UNALIGNED_SUPPORT_DISABLE */
115 #ifdef ARM_MATH_MATRIX_CHECK
116 /* Check for matrix mismatch condition */
117 if((pSrcA
->numCols
!= pSrcB
->numRows
) ||
118 (pSrcA
->numRows
!= pDst
->numRows
) || (pSrcB
->numCols
!= pDst
->numCols
))
120 /* Set status as ARM_MATH_SIZE_MISMATCH */
121 status
= ARM_MATH_SIZE_MISMATCH
;
126 /* Matrix transpose */
129 /* Apply loop unrolling and exchange the columns with row elements */
132 /* The pointer px is set to starting address of the column being processed */
135 /* First part of the processing with loop unrolling. Compute 4 outputs at a time.
136 ** a second loop below computes the remaining 1 to 3 samples. */
139 #ifndef UNALIGNED_SUPPORT_DISABLE
140 /* Read two elements from the row */
141 in
= *__SIMD32(pInB
)++;
143 /* Unpack and store one element in the destination */
144 #ifndef ARM_MATH_BIG_ENDIAN
150 *px
= (q15_t
) ((in
& (q31_t
) 0xffff0000) >> 16);
152 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
154 /* Update the pointer px to point to the next row of the transposed matrix */
157 /* Unpack and store the second element in the destination */
158 #ifndef ARM_MATH_BIG_ENDIAN
160 *px
= (q15_t
) ((in
& (q31_t
) 0xffff0000) >> 16);
166 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
168 /* Update the pointer px to point to the next row of the transposed matrix */
171 /* Read two elements from the row */
172 in
= *__SIMD32(pInB
)++;
174 /* Unpack and store one element in the destination */
175 #ifndef ARM_MATH_BIG_ENDIAN
181 *px
= (q15_t
) ((in
& (q31_t
) 0xffff0000) >> 16);
183 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
185 /* Update the pointer px to point to the next row of the transposed matrix */
188 /* Unpack and store the second element in the destination */
190 #ifndef ARM_MATH_BIG_ENDIAN
192 *px
= (q15_t
) ((in
& (q31_t
) 0xffff0000) >> 16);
198 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
202 /* Read one element from the row */
205 /* Store one element in the destination */
208 /* Update the pointer px to point to the next row of the transposed matrix */
211 /* Read one element from the row */
214 /* Store one element in the destination */
217 /* Update the pointer px to point to the next row of the transposed matrix */
220 /* Read one element from the row */
223 /* Store one element in the destination */
226 /* Update the pointer px to point to the next row of the transposed matrix */
229 /* Read one element from the row */
232 /* Store one element in the destination */
235 #endif /* #ifndef UNALIGNED_SUPPORT_DISABLE */
237 /* Update the pointer px to point to the next row of the transposed matrix */
240 /* Decrement the column loop counter */
244 /* If the columns of pSrcB is not a multiple of 4, compute any remaining output samples here.
245 ** No loop unrolling is used. */
246 col
= numColsB
% 0x4u
;
250 /* Read and store the input element in the destination */
253 /* Update the pointer px to point to the next row of the transposed matrix */
256 /* Decrement the column loop counter */
262 /* Decrement the row loop counter */
267 /* Reset the variables for the usage in the following multiplication process */
272 /* The following loop performs the dot-product of each row in pSrcA with each column in pSrcB */
276 /* For every row wise process, the column loop counter is to be initiated */
279 /* For every row wise process, the pIn2 pointer is set
280 ** to the starting address of the transposed pSrcB data */
286 /* Set the variable sum, that acts as accumulator, to zero */
289 /* Apply loop unrolling and compute 2 MACs simultaneously. */
290 colCnt
= numColsA
>> 2;
292 /* Initiate the pointer pIn1 to point to the starting address of the column being processed */
293 pInA
= pSrcA
->pData
+ i
;
295 /* matrix multiplication */
298 /* c(m,n) = a(1,1)*b(1,1) + a(1,2) * b(2,1) + .... + a(m,p)*b(p,n) */
299 #ifndef UNALIGNED_SUPPORT_DISABLE
301 inA1
= *__SIMD32(pInA
)++;
302 inB1
= *__SIMD32(pInB
)++;
303 inA2
= *__SIMD32(pInA
)++;
304 inB2
= *__SIMD32(pInB
)++;
306 sum
= __SMLAD(inA1
, inB1
, sum
);
307 sum
= __SMLAD(inA2
, inB2
, sum
);
326 #endif /* #ifndef UNALIGNED_SUPPORT_DISABLE */
328 /* Decrement the loop counter */
332 /* process odd column samples */
333 colCnt
= numColsA
% 0x4u
;
337 /* c(m,n) = a(1,1)*b(1,1) + a(1,2) * b(2,1) + .... + a(m,p)*b(p,n) */
338 sum
+= (q31_t
) (*pInA
++) * (*pInB
++);
343 /* Saturate and store the result in the destination buffer */
344 *px
= (q15_t
) (sum
>> 15);
347 /* Decrement the column loop counter */
354 /* Decrement the row loop counter */
359 /* set status as ARM_MATH_SUCCESS */
360 status
= ARM_MATH_SUCCESS
;
363 /* Return to application */
368 * @} end of MatrixMult group