764 lines
24 KiB
C
764 lines
24 KiB
C
/* ----------------------------------------------------------------------
|
|
* Project: CMSIS DSP Library
|
|
* Title: arm_mat_mult_f16.c
|
|
* Description: Floating-point matrix multiplication
|
|
*
|
|
* $Date: 23 April 2021
|
|
* $Revision: V1.9.0
|
|
*
|
|
* Target Processor: Cortex-M and Cortex-A cores
|
|
* -------------------------------------------------------------------- */
|
|
/*
|
|
* Copyright (C) 2010-2021 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 "dsp/matrix_functions_f16.h"
|
|
|
|
#if defined(ARM_FLOAT16_SUPPORTED)
|
|
|
|
|
|
/**
|
|
* @ingroup groupMatrix
|
|
*/
|
|
|
|
|
|
/**
|
|
* @addtogroup MatrixMult
|
|
* @{
|
|
*/
|
|
|
|
/**
|
|
* @brief Floating-point matrix multiplication.
|
|
* @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
|
|
* @return The function returns either
|
|
* <code>ARM_MATH_SIZE_MISMATCH</code> or <code>ARM_MATH_SUCCESS</code> based on the outcome of size checking.
|
|
*/
|
|
|
|
#if defined(ARM_MATH_MVE_FLOAT16) && !defined(ARM_MATH_AUTOVECTORIZE)
|
|
|
|
__STATIC_FORCEINLINE arm_status arm_mat_mult_f16_2x2_mve(
|
|
const arm_matrix_instance_f16 *pSrcA,
|
|
const arm_matrix_instance_f16 *pSrcB,
|
|
arm_matrix_instance_f16 *pDst)
|
|
{
|
|
static const uint16_t offsetA[8] = { 0, 0, 2, 2, 0, 0, 2, 2 };
|
|
/* offsetB allows to read and duplicate 1 row of B */
|
|
static const uint16_t offsetB[8] = { 0, 1, 0, 1, 0, 1, 0, 1 };
|
|
uint16x8_t vecOffsA, vecOffsB;
|
|
f16x8_t vecInA, vecInB, vecDst;
|
|
float16_t *pOut = pDst->pData; /* output data matrix pointer */
|
|
|
|
/*
|
|
* load initial offsets
|
|
*/
|
|
vecOffsA = vldrhq_u16((uint16_t const *) offsetA);
|
|
vecOffsB = vldrhq_u16((uint16_t const *) offsetB);
|
|
/*
|
|
* load {a00 a00 a10 a10 x x x x }
|
|
*/
|
|
vecInA = vldrhq_gather_shifted_offset((float16_t const *) pSrcA->pData, vecOffsA);
|
|
/*
|
|
* load {b00 b01 b00 b01 x x x x }
|
|
*/
|
|
vecInB = vldrhq_gather_shifted_offset((float16_t const *) pSrcB->pData, vecOffsB);
|
|
/*
|
|
* { a00 b00 a00 b01
|
|
* a10 b00 a10 b01
|
|
* x x
|
|
* x x }
|
|
*/
|
|
vecDst = vmulq(vecInA, vecInB);
|
|
/*
|
|
* move to 2nd column of matrix A
|
|
*/
|
|
vecOffsA = vaddq_n_u16(vecOffsA, (uint16_t) 1);
|
|
/*
|
|
* load {a01 a01 a11 a11 x x x x}
|
|
*/
|
|
vecInA = vldrhq_gather_shifted_offset((float16_t const *) pSrcA->pData, vecOffsA);
|
|
/*
|
|
* move to next B row
|
|
*/
|
|
vecOffsB = vaddq_n_u16(vecOffsB, (uint16_t) 2);
|
|
/*
|
|
* load {b10, b11, b10, b11, x x x x }
|
|
*/
|
|
vecInB = vldrhq_gather_shifted_offset((float16_t const *) pSrcB->pData, vecOffsB);
|
|
/*
|
|
* { a00 b00 + a01 b10 a00 b01 + a01 b11
|
|
* a10 b00 + a11 b10 a10 b01 + a11 b11
|
|
* x x
|
|
* x x }
|
|
*/
|
|
vecDst = vfmaq(vecDst, vecInA, vecInB);
|
|
|
|
mve_pred16_t p0 = vctp16q(2*2);
|
|
/*
|
|
* Store the result in the destination buffer
|
|
* (lower half of the vector)
|
|
*/
|
|
vstrhq_p(pOut, vecDst, p0);
|
|
|
|
return (ARM_MATH_SUCCESS);
|
|
}
|
|
|
|
|
|
|
|
|
|
__STATIC_FORCEINLINE arm_status arm_mat_mult_f16_3x3_mve(
|
|
const arm_matrix_instance_f16 *pSrcA,
|
|
const arm_matrix_instance_f16 *pSrcB,
|
|
arm_matrix_instance_f16 *pDst)
|
|
{
|
|
static const uint16_t offsetA[8] = { 0, 0, 0, 3, 3, 3, 6, 6 };
|
|
/* offsetB allows to read and duplicate 1 row of B */
|
|
static const uint16_t offsetB[8] = { 0, 1, 2, 0, 1, 2, 0, 1 };
|
|
uint16x8_t vecOffsA, vecOffsB;
|
|
f16x8_t vecInA, vecInB, vecDst;
|
|
float16_t *pOut = pDst->pData; /* output data matrix pointer */
|
|
|
|
/*
|
|
* load initial offsets
|
|
*/
|
|
vecOffsA = vldrhq_u16((uint16_t const *) offsetA);
|
|
vecOffsB = vldrhq_u16((uint16_t const *) offsetB);
|
|
|
|
/*
|
|
* load {a00 a00 a00 a10 a10 a10 a20 a20}
|
|
*/
|
|
vecInA = vldrhq_gather_shifted_offset((float16_t const *) pSrcA->pData, vecOffsA);
|
|
/*
|
|
* load {b00 b01 b02 b00 b01 b02 b00 b01}
|
|
*/
|
|
vecInB = vldrhq_gather_shifted_offset((float16_t const *) pSrcB->pData, vecOffsB);
|
|
/*
|
|
* { a00 b00 a00 b01 a00 b02
|
|
* a10 b00 a10 b01 a10 b02
|
|
* a20 b00 a20 b01}
|
|
*/
|
|
vecDst = vmulq(vecInA, vecInB);
|
|
|
|
/*
|
|
* move to 2nd column of matrix A
|
|
*/
|
|
vecOffsA = vaddq_n_u16(vecOffsA, (uint16_t) 1);
|
|
/*
|
|
* load {a01 a01 a01 a11 a11 a11 a21 a21}
|
|
*/
|
|
vecInA = vldrhq_gather_shifted_offset((float16_t const *) pSrcA->pData, vecOffsA);
|
|
/*
|
|
* move to next B row
|
|
*/
|
|
vecOffsB = vaddq_n_u16(vecOffsB, (uint16_t) 3);
|
|
/*
|
|
* load {b10, b11, b12, b10, b11, b12, b10, b11}
|
|
*/
|
|
vecInB = vldrhq_gather_shifted_offset((float16_t const *) pSrcB->pData, vecOffsB);
|
|
/*
|
|
* { a00 b00 + a01 b10 a00 b01 + a01 b11 a00 b02 + a01 b12
|
|
* a10 b00 + a11 b10 a10 b01 + a11 b11 a10 b02 + a11 b12
|
|
* a20 b00 + a21 b10 a20 b01 + a21 b11 }
|
|
*/
|
|
vecDst = vfmaq(vecDst, vecInA, vecInB);
|
|
/*
|
|
* move to 3rd column of matrix A
|
|
*/
|
|
vecOffsA = vaddq_n_u16(vecOffsA, (uint16_t) 1);
|
|
/*
|
|
* load {a02 a02 a02 a12 a12 a12 a22 a22}
|
|
*/
|
|
vecInA = vldrhq_gather_shifted_offset((float16_t const *) pSrcA->pData, vecOffsA);
|
|
/*
|
|
* move to next B row
|
|
*/
|
|
vecOffsB = vaddq_n_u16(vecOffsB, (uint16_t) 3);
|
|
/*
|
|
* load {b20, b21, b22, b20, b21, b22, b20, b21}
|
|
*/
|
|
vecInB = vldrhq_gather_shifted_offset((float16_t const *) pSrcB->pData, vecOffsB);
|
|
/*
|
|
* {a00 b00 + a01 b10 + a02 b20 a00 b01 + a01 b11 + a02 b21 a00 b02 + a01 b12 + a02 b22},
|
|
* a10 b00 + a11 b10 + a12 b20 a10 b01 + a11 b11 + a12 b21 a10 b02 + a11 b12 + a12 b22},
|
|
* a20 b00 + a21 b10 + a22 b20 a20 b01 + a21 b11 + a22 b21 }
|
|
*/
|
|
vecDst = vfmaq(vecDst, vecInA, vecInB);
|
|
|
|
/*
|
|
* Store the result in the destination buffer
|
|
*/
|
|
vst1q(pOut, vecDst); pOut += 8;
|
|
|
|
/* last element computed in scalar mode
|
|
* a20 b02 + a21 b12 + a22 b22
|
|
*/
|
|
_Float16 * pA = (_Float16 *)pSrcA->pData;
|
|
_Float16 * pB = (_Float16 *)pSrcB->pData;
|
|
*pOut = pA[2*3] * pB[2] + pA[2*3+1] * pB[3+2] + pA[2*3+2] * pB[2*3+2];
|
|
|
|
return (ARM_MATH_SUCCESS);
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
__STATIC_FORCEINLINE arm_status arm_mat_mult_f16_4x4_mve(
|
|
const arm_matrix_instance_f16 *pSrcA,
|
|
const arm_matrix_instance_f16 *pSrcB,
|
|
arm_matrix_instance_f16 *pDst)
|
|
{
|
|
/* offsetA allows to read and duplicate 2 successive column elements of A */
|
|
static const uint16_t offsetA[8] = { 0, 0, 0, 0, 4, 4, 4, 4 };
|
|
/* offsetB allows to read and duplicate 1 row of B */
|
|
static const uint16_t offsetB[8] = { 0, 1, 2, 3, 0, 1, 2, 3 };
|
|
uint16x8_t vecOffsA, vecOffsB;
|
|
f16x8_t vecInA, vecInB, vecDst0, vecDst1;
|
|
float16_t *pOut = pDst->pData; /* output data matrix pointer */
|
|
|
|
/*
|
|
* load initial offsets
|
|
*/
|
|
vecOffsA = vldrhq_u16((uint16_t const *) offsetA);
|
|
vecOffsB = vldrhq_u16((uint16_t const *) offsetB);
|
|
|
|
/*
|
|
* load {a00 a00 a00 a00 a10 a10 a10 a10}
|
|
*/
|
|
vecInA = vldrhq_gather_shifted_offset((float16_t const *) pSrcA->pData, vecOffsA);
|
|
/*
|
|
* load {b00 b01 b02 b03 b00 b01 b02 b03}
|
|
*/
|
|
vecInB = vldrhq_gather_shifted_offset((float16_t const *) pSrcB->pData, vecOffsB);
|
|
/*
|
|
* { a00 b00 a00 b01 a00 b02 a00 b03
|
|
* a10 b00 a10 b01 a10 b02 a10 b03 }
|
|
*/
|
|
vecDst0 = vmulq(vecInA, vecInB);
|
|
/*
|
|
* jump 2 x A rows (2nd half of matrix)
|
|
*/
|
|
vecOffsA = vaddq_n_u16(vecOffsA, (uint16_t) 8);
|
|
/*
|
|
* load {a20 a20 a20 a20 a30 a30 a30 a30}
|
|
*/
|
|
vecInA = vldrhq_gather_shifted_offset((float16_t const *) pSrcA->pData, vecOffsA);
|
|
/*
|
|
* { a20 b00 a20 b01 a20 b02 a20 b03
|
|
* a30 b00 a30 b01 a30 b02 + a31 b12 }
|
|
*/
|
|
vecDst1 = vmulq(vecInA, vecInB);
|
|
/*
|
|
* rewind back to top half of the A matrix (2nd column)
|
|
*/
|
|
vecOffsA = vsubq(vecOffsA, (uint16_t) 7);
|
|
/*
|
|
* load {a01 a01 a01 a01 a11 a11 a11 a11}
|
|
*/
|
|
vecInA = vldrhq_gather_shifted_offset((float16_t const *) pSrcA->pData, vecOffsA);
|
|
/*
|
|
* move to next B row
|
|
*/
|
|
vecOffsB = vaddq_n_u16(vecOffsB, (uint16_t) 4);
|
|
/*
|
|
* load {b10, b11, b12, b13, b10, b11, b12, b13}
|
|
*/
|
|
vecInB = vldrhq_gather_shifted_offset((float16_t const *) pSrcB->pData, vecOffsB);
|
|
/*
|
|
* { a00 b00 + a01 b10 a00 b01 + a01 b11 a00 b02 + a01 b12 a00 b03 + a01 b13
|
|
* a10 b00 + a11 b10 a10 b01 + a11 b11 a10 b02 + a11 b12 a10 b03 + a11 b13 }
|
|
*/
|
|
vecDst0 = vfmaq(vecDst0, vecInA, vecInB);
|
|
/*
|
|
* jump 2 x A rows (2nd half of matrix)
|
|
*/
|
|
vecOffsA = vaddq_n_u16(vecOffsA, (uint16_t) 8);
|
|
/*
|
|
* load {a21 a21 a21 a21 a31 a31 a31 a31}
|
|
*/
|
|
vecInA = vldrhq_gather_shifted_offset((float16_t const *) pSrcA->pData, vecOffsA);
|
|
/*
|
|
* {a20 b00 + a21 b10 a20 b01 + a21 b11 a20 b02 + a21 b12 a20 b03 + a21 b13
|
|
* a30 b00 + a31 b10 a30 b01 + a31 b11 a30 b02 + a31 b12 a30 b03 + a31 b13 }
|
|
*/
|
|
vecDst1 = vfmaq(vecDst1, vecInA, vecInB);
|
|
|
|
/*
|
|
* rewind back to top half of the A matrix (3rd column)
|
|
*/
|
|
vecOffsA = vsubq(vecOffsA, (uint16_t) 7);
|
|
/*
|
|
* load {a02 a02 a02 a02 a12 a12 a12 a12}
|
|
*/
|
|
vecInA = vldrhq_gather_shifted_offset((float16_t const *) pSrcA->pData, vecOffsA);
|
|
/*
|
|
* move to next B row
|
|
*/
|
|
vecOffsB = vaddq_n_u16(vecOffsB, (uint16_t) 4);
|
|
/*
|
|
* load {b20, b21, b22, b23, b20, b21, b22, b23}
|
|
*/
|
|
vecInB = vldrhq_gather_shifted_offset((float16_t const *) pSrcB->pData, vecOffsB);
|
|
/*
|
|
* { a00 b00 + a01 b10 + a02 b20 a00 b01 + a01 b11 + a02 b21 a00 b02 + a01 b12 + a02 b22 a00 b03 + a01 b13 + a02 b23
|
|
* a10 b00 + a11 b10 + a12 b20 a10 b01 + a11 b11 + a12 b21 a10 b02 + a11 b12 + a12 b22 a10 b03 + a11 b13 + a12 b23 }
|
|
*/
|
|
vecDst0 = vfmaq(vecDst0, vecInA, vecInB);
|
|
/*
|
|
* jump 2 x A rows
|
|
*/
|
|
vecOffsA = vaddq_n_u16(vecOffsA, (uint16_t) 8);
|
|
|
|
/*
|
|
* load {a22 a22 a22 a22 a32 a32 a32 a32}
|
|
*/
|
|
vecInA = vldrhq_gather_shifted_offset((float16_t const *) pSrcA->pData, vecOffsA);
|
|
/*
|
|
* {a20 b00 + a21 b10 + a22 b20 a20 b01 + a21 b11 + a22 b21 a20 b02 + a21 b12 + a22 b22 a20 b03 + a21 b13 + a22 b23
|
|
* a30 b00 + a31 b10 + a32 b20 a30 b01 + a31 b11 + a32 b21 a30 b02 + a31 b12 + a32 b22 a30 b03 + a31 b13 + a32 b23 }
|
|
*/
|
|
vecDst1 = vfmaq(vecDst1, vecInA, vecInB);
|
|
|
|
/*
|
|
* rewind back to top half of the A matrix (4th column)
|
|
*/
|
|
vecOffsA = vsubq(vecOffsA, (uint16_t) 7);
|
|
/*
|
|
* load {a03 a03 a03 a03 a13 a13 a13 a13}
|
|
*/
|
|
vecInA = vldrhq_gather_shifted_offset((float16_t const *) pSrcA->pData, vecOffsA);
|
|
/*
|
|
* move to next B row
|
|
*/
|
|
vecOffsB = vaddq_n_u16(vecOffsB, (uint16_t) 4);
|
|
/*
|
|
* load {b30, b31, b32, b33, b30, b31, b32, b33}
|
|
*/
|
|
vecInB = vldrhq_gather_shifted_offset((float16_t const *) pSrcB->pData, vecOffsB);
|
|
/*
|
|
* { a00 b00 +...+ a03 b30, a00 b01 +...+ a03 b31, a00 b02 +...+ a03 b32, a00 b03 +...+ a03 b33
|
|
* a10 b00 +...+ a13 b30, a10 b01 +...+ a13 b31, a10 b02 +...+ a13 b32, a10 b03 +...+ a13 b33 }
|
|
*/
|
|
vecDst0 = vfmaq(vecDst0, vecInA, vecInB);
|
|
/*
|
|
* jump 2 x A rows
|
|
*/
|
|
vecOffsA = vaddq_n_u16(vecOffsA, (uint16_t) 8);
|
|
/*
|
|
* load {a23 a23 a23 a23 a33 a33 a33 a33}
|
|
*/
|
|
vecInA = vldrhq_gather_shifted_offset((float16_t const *) pSrcA->pData, vecOffsA);
|
|
/*
|
|
* {a20 b00 +...+ a23 b30, a20 b01 +...+ a23 b31, a20 b02 +...+ a23 b32, a20 b03 +...+ a23 b33
|
|
* a30 b00 +...+ a33 b30, a30 b01 +...+ a33 b31, a30 b02 +...+ a33 b32, a30 b03 +...+ a33 b33 }
|
|
*/
|
|
vecDst1 = vfmaq(vecDst1, vecInA, vecInB);
|
|
|
|
/*
|
|
* Store the result in the destination buffer
|
|
*/
|
|
vst1q(pOut, vecDst0); pOut += 8;
|
|
vst1q(pOut, vecDst1);
|
|
|
|
return (ARM_MATH_SUCCESS);
|
|
}
|
|
|
|
|
|
arm_status arm_mat_mult_f16(
|
|
const arm_matrix_instance_f16 * pSrcA,
|
|
const arm_matrix_instance_f16 * pSrcB,
|
|
arm_matrix_instance_f16 * pDst)
|
|
{
|
|
float16_t *pInB = pSrcB->pData; /* input data matrix pointer B */
|
|
float16_t *pInA = pSrcA->pData; /* input data matrix pointer A */
|
|
float16_t *pOut = pDst->pData; /* output data matrix pointer */
|
|
int numRowsA = pSrcA->numRows; /* number of rows of input matrix A */
|
|
int numColsB = pSrcB->numCols; /* number of columns of input matrix B */
|
|
int numColsA = pSrcA->numCols; /* number of columns of input matrix A */
|
|
uint32_t blkCnt; /* loop counters */
|
|
int i;
|
|
|
|
|
|
#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 */
|
|
return(ARM_MATH_SIZE_MISMATCH);
|
|
}
|
|
else
|
|
|
|
#endif /* #ifdef ARM_MATH_MATRIX_CHECK */
|
|
{
|
|
/* small squared matrix specialized routines */
|
|
if(numRowsA == numColsB && numColsB == numColsA) {
|
|
if(numRowsA == 2)
|
|
return arm_mat_mult_f16_2x2_mve(pSrcA, pSrcB, pDst);
|
|
else if(numRowsA == 3)
|
|
return arm_mat_mult_f16_3x3_mve(pSrcA, pSrcB, pDst);
|
|
else if(numRowsA == 4)
|
|
return arm_mat_mult_f16_4x4_mve(pSrcA, pSrcB, pDst);
|
|
}
|
|
|
|
/* main loop process 4 rows */
|
|
i = numRowsA / 4;
|
|
while(i > 0)
|
|
{
|
|
float16_t *pInA0, *pInA1, *pInA2, *pInA3;
|
|
float16_t *pInB0;
|
|
float16_t *pOut0, *pOut1, *pOut2, *pOut3;
|
|
f16x8_t vecMac0, vecMac1, vecMac2, vecMac3;
|
|
f16x8_t vecInB;
|
|
|
|
/* pointers to 4 consecutive output rows */
|
|
pOut0 = pOut;
|
|
pOut1 = pOut0 + numColsB;
|
|
pOut2 = pOut1 + numColsB;
|
|
pOut3 = pOut2 + numColsB;
|
|
pInB0 = pInB;
|
|
|
|
int k = numColsB >> 3;
|
|
while(k > 0)
|
|
{
|
|
/* pointers to 4 consecutive Matrix A rows */
|
|
pInA0 = pInA;
|
|
pInA1 = pInA0 + numColsA;
|
|
pInA2 = pInA1 + numColsA;
|
|
pInA3 = pInA2 + numColsA;
|
|
|
|
vecMac0 = vdupq_n_f16(0.0f16);
|
|
vecMac1 = vdupq_n_f16(0.0f16);
|
|
vecMac2 = vdupq_n_f16(0.0f16);
|
|
vecMac3 = vdupq_n_f16(0.0f16);
|
|
|
|
blkCnt = numColsA;
|
|
|
|
while (blkCnt > 0U)
|
|
{
|
|
/*
|
|
* load {bi,4n+0, bi,4n+1, bi,4n+2, bi,4n+3..., bi,4n+7}
|
|
*/
|
|
vecInB = *(f16x8_t *)pInB0; /* vldrhq_f16(pInB0, 0); */
|
|
|
|
vecMac0 = vfmaq(vecMac0, vecInB, *pInA0++);
|
|
vecMac1 = vfmaq(vecMac1, vecInB, *pInA1++);
|
|
vecMac2 = vfmaq(vecMac2, vecInB, *pInA2++);
|
|
vecMac3 = vfmaq(vecMac3, vecInB, *pInA3++);
|
|
|
|
pInB0 = pInB0 + numColsB;
|
|
/*
|
|
* Decrement the blockSize loop counter
|
|
*/
|
|
blkCnt--;
|
|
}
|
|
|
|
/* Store the results (4 x 8 block) in the destination buffer */
|
|
vst1q(pOut0, vecMac0); pOut0 += 8;
|
|
vst1q(pOut1, vecMac1); pOut1 += 8;
|
|
vst1q(pOut2, vecMac2); pOut2 += 8;
|
|
vst1q(pOut3, vecMac3); pOut3 += 8;
|
|
/*
|
|
* rewind
|
|
*/
|
|
pInB0 -= (numColsB * numColsA) - 8;
|
|
k--;
|
|
}
|
|
|
|
int colBLeft = numColsB & 7;
|
|
if (colBLeft)
|
|
{
|
|
pInA0 = pInA;
|
|
pInA1 = pInA0 + numColsA;
|
|
pInA2 = pInA1 + numColsA;
|
|
pInA3 = pInA2 + numColsA;
|
|
mve_pred16_t p0 = vctp16q(colBLeft);
|
|
|
|
vecMac0 = vdupq_n_f16(0.0f16);
|
|
vecMac1 = vdupq_n_f16(0.0f16);
|
|
vecMac2 = vdupq_n_f16(0.0f16);
|
|
vecMac3 = vdupq_n_f16(0.0f16);
|
|
|
|
blkCnt = numColsA;
|
|
|
|
while (blkCnt > 0U)
|
|
{
|
|
/*
|
|
* load {bi,4n+0, bi,4n+1, bi,4n+2, ..bi,4n+colBLeft-1, 0, ..}
|
|
*/
|
|
vecInB = vldrhq_z_f16(pInB0, p0);
|
|
|
|
vecMac0 = vfmaq(vecMac0, vecInB, *pInA0++);
|
|
vecMac1 = vfmaq(vecMac1, vecInB, *pInA1++);
|
|
vecMac2 = vfmaq(vecMac2, vecInB, *pInA2++);
|
|
vecMac3 = vfmaq(vecMac3, vecInB, *pInA3++);
|
|
|
|
pInB0 = pInB0 + numColsB;
|
|
/*
|
|
* Decrement the blockSize loop counter
|
|
*/
|
|
blkCnt--;
|
|
}
|
|
|
|
/* Store the results (4 x colBLeft block) in the destination buffer */
|
|
vstrhq_p_f16(pOut0, vecMac0, p0);
|
|
vstrhq_p_f16(pOut1, vecMac1, p0);
|
|
vstrhq_p_f16(pOut2, vecMac2, p0);
|
|
vstrhq_p_f16(pOut3, vecMac3, p0);
|
|
}
|
|
|
|
pInA += 4 * numColsA;
|
|
pOut += 4 * numColsB;
|
|
i--;
|
|
}
|
|
|
|
/*
|
|
* non multiple of 4 rows for Matrix A
|
|
* process single row
|
|
*/
|
|
if (numRowsA & 3)
|
|
{
|
|
i = numRowsA & 3;
|
|
do
|
|
{
|
|
float16_t *pInA0;
|
|
float16_t *pInB0;
|
|
float16_t *pOut0;
|
|
f16x8_t vecInB;
|
|
f16x8_t vecMac0;
|
|
|
|
pOut0 = pOut;
|
|
pInB0 = pInB;
|
|
|
|
int k = numColsB >> 3;
|
|
while(k > 0)
|
|
{
|
|
pInA0 = pInA;
|
|
|
|
vecMac0 = vdupq_n_f16(0.0f16);
|
|
blkCnt = numColsA;
|
|
|
|
while (blkCnt > 0U)
|
|
{
|
|
/*
|
|
* load {bi,4n+0, bi,4n+1, bi,4n+2, bi,4n+3, ...bi,4n+7}
|
|
*/
|
|
vecInB = *(f16x8_t *)pInB0; /* vldrhq_f16(pInB0, 0); */
|
|
|
|
vecMac0 = vfmaq(vecMac0, vecInB, *pInA0++);
|
|
|
|
pInB0 = pInB0 + numColsB;
|
|
/*
|
|
* Decrement the blockSize loop counter
|
|
*/
|
|
blkCnt--;
|
|
}
|
|
/* Store the results (1 x 8 block) in the destination buffer */
|
|
vst1q(pOut0, vecMac0); pOut0 += 8;
|
|
/*
|
|
* rewind
|
|
*/
|
|
pInB0 -= (numColsB * numColsA) - 8;
|
|
k--;
|
|
}
|
|
|
|
int colBLeft = numColsB & 7;
|
|
if (colBLeft)
|
|
{
|
|
pInA0 = pInA;
|
|
mve_pred16_t p0 = vctp16q(colBLeft);
|
|
|
|
vecMac0 = vdupq_n_f16(0.0f16);
|
|
blkCnt = numColsA;
|
|
|
|
while (blkCnt > 0U)
|
|
{
|
|
/*
|
|
* load {bi,4n+0, bi,4n+1, bi,4n+2, ..., bi,4n+colBLeft, 0, ...}
|
|
*/
|
|
vecInB = vldrhq_z_f16(pInB0, p0);
|
|
|
|
vecMac0 = vfmaq(vecMac0, vecInB, *pInA0++);
|
|
|
|
pInB0 = pInB0 + numColsB;
|
|
/*
|
|
* Decrement the blockSize loop counter
|
|
*/
|
|
blkCnt--;
|
|
}
|
|
/* Store the results (1 x colBLeft block) in the destination buffer */
|
|
vstrhq_p_f16(pOut0, vecMac0, p0);
|
|
}
|
|
|
|
pInA += 1 * numColsA;
|
|
pOut += 1 * numColsB;
|
|
}
|
|
while (--i);
|
|
}
|
|
/*
|
|
* Return to application
|
|
*/
|
|
return (ARM_MATH_SUCCESS);
|
|
}
|
|
}
|
|
#else
|
|
|
|
|
|
arm_status arm_mat_mult_f16(
|
|
const arm_matrix_instance_f16 * pSrcA,
|
|
const arm_matrix_instance_f16 * pSrcB,
|
|
arm_matrix_instance_f16 * pDst)
|
|
{
|
|
float16_t *pIn1 = pSrcA->pData; /* Input data matrix pointer A */
|
|
float16_t *pIn2 = pSrcB->pData; /* Input data matrix pointer B */
|
|
float16_t *pInA = pSrcA->pData; /* Input data matrix pointer A */
|
|
float16_t *pInB = pSrcB->pData; /* Input data matrix pointer B */
|
|
float16_t *pOut = pDst->pData; /* Output data matrix pointer */
|
|
float16_t *px; /* Temporary output data matrix pointer */
|
|
_Float16 sum; /* Accumulator */
|
|
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 */
|
|
uint32_t col, i = 0U, row = numRowsA, colCnt; /* Loop counters */
|
|
arm_status status; /* Status of matrix multiplication */
|
|
|
|
#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 /* #ifdef ARM_MATH_MATRIX_CHECK */
|
|
|
|
{
|
|
/* The following loop performs the dot-product of each row in pSrcA with each column in pSrcB */
|
|
/* row loop */
|
|
do
|
|
{
|
|
/* Output pointer is set to starting address of row being processed */
|
|
px = pOut + i;
|
|
|
|
/* For every row wise process, column loop counter is to be initiated */
|
|
col = numColsB;
|
|
|
|
/* For every row wise process, pIn2 pointer is set to starting address of pSrcB data */
|
|
pIn2 = pSrcB->pData;
|
|
|
|
/* column loop */
|
|
do
|
|
{
|
|
/* Set the variable sum, that acts as accumulator, to zero */
|
|
sum = 0.0f16;
|
|
|
|
/* Initialize pointer pIn1 to point to starting address of column being processed */
|
|
pIn1 = pInA;
|
|
|
|
#if defined (ARM_MATH_LOOPUNROLL)
|
|
|
|
/* Loop unrolling: Compute 4 MACs at a time. */
|
|
colCnt = numColsA >> 2U;
|
|
|
|
/* 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) */
|
|
|
|
/* Perform the multiply-accumulates */
|
|
sum += (_Float16)*pIn1++ * (_Float16)*pIn2;
|
|
pIn2 += numColsB;
|
|
|
|
sum += (_Float16)*pIn1++ * (_Float16)*pIn2;
|
|
pIn2 += numColsB;
|
|
|
|
sum += (_Float16)*pIn1++ * (_Float16)*pIn2;
|
|
pIn2 += numColsB;
|
|
|
|
sum += (_Float16)*pIn1++ * (_Float16)*pIn2;
|
|
pIn2 += numColsB;
|
|
|
|
/* Decrement loop counter */
|
|
colCnt--;
|
|
}
|
|
|
|
/* Loop unrolling: Compute remaining MACs */
|
|
colCnt = numColsA % 0x4U;
|
|
|
|
#else
|
|
|
|
/* Initialize cntCnt with number of columns */
|
|
colCnt = numColsA;
|
|
|
|
#endif /* #if defined (ARM_MATH_LOOPUNROLL) */
|
|
|
|
while (colCnt > 0U)
|
|
{
|
|
/* c(m,n) = a(1,1) * b(1,1) + a(1,2) * b(2,1) + .... + a(m,p) * b(p,n) */
|
|
|
|
/* Perform the multiply-accumulates */
|
|
sum += (_Float16)*pIn1++ * (_Float16)*pIn2;
|
|
pIn2 += numColsB;
|
|
|
|
/* Decrement loop counter */
|
|
colCnt--;
|
|
}
|
|
|
|
/* Store result in destination buffer */
|
|
*px++ = sum;
|
|
|
|
/* Decrement column loop counter */
|
|
col--;
|
|
|
|
/* Update pointer pIn2 to point to starting address of next column */
|
|
pIn2 = pInB + (numColsB - col);
|
|
|
|
} while (col > 0U);
|
|
|
|
/* Update pointer pInA to point to starting address of next row */
|
|
i = i + numColsB;
|
|
pInA = pInA + numColsA;
|
|
|
|
/* Decrement row loop counter */
|
|
row--;
|
|
|
|
} while (row > 0U);
|
|
|
|
/* Set status as ARM_MATH_SUCCESS */
|
|
status = ARM_MATH_SUCCESS;
|
|
}
|
|
|
|
/* Return to application */
|
|
return (status);
|
|
}
|
|
|
|
#endif /* defined(ARM_MATH_MVEF) && !defined(ARM_MATH_AUTOVECTORIZE) */
|
|
|
|
/**
|
|
* @} end of MatrixMult group
|
|
*/
|
|
|
|
#endif /* #if defined(ARM_FLOAT16_SUPPORTED) */
|
|
|