/* ---------------------------------------------------------------------- * 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 * ARM_MATH_SIZE_MISMATCH or ARM_MATH_SUCCESS based on the outcome of size checking. * * @details * Scaling and Overflow Behavior: * * \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 arm_mat_mult_q15() 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 */