1 /* ----------------------------------------------------------------------
2 * Project: CMSIS DSP Library
3 * Title: arm_mat_cholesky_f16.c
4 * Description: Floating-point Cholesky decomposition
5 *
6 * $Date: 23 April 2021
7 * $Revision: V1.9.0
8 *
9 * Target Processor: Cortex-M and Cortex-A cores
10 * -------------------------------------------------------------------- */
11 /*
12 * Copyright (C) 2010-2021 ARM Limited or its affiliates. All rights reserved.
13 *
14 * SPDX-License-Identifier: Apache-2.0
15 *
16 * Licensed under the Apache License, Version 2.0 (the License); you may
17 * not use this file except in compliance with the License.
18 * You may obtain a copy of the License at
19 *
20 * www.apache.org/licenses/LICENSE-2.0
21 *
22 * Unless required by applicable law or agreed to in writing, software
23 * distributed under the License is distributed on an AS IS BASIS, WITHOUT
24 * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
25 * See the License for the specific language governing permissions and
26 * limitations under the License.
27 */
28
29 #include "dsp/matrix_functions_f16.h"
30
31 #if defined(ARM_FLOAT16_SUPPORTED)
32
33 /**
34 @ingroup groupMatrix
35 */
36
37 /**
38 @addtogroup MatrixChol
39 @{
40 */
41
42 /**
43 * @brief Floating-point Cholesky decomposition of positive-definite matrix.
44 * @param[in] pSrc points to the instance of the input floating-point matrix structure.
45 * @param[out] pDst points to the instance of the output floating-point matrix structure.
46 * @return The function returns ARM_MATH_SIZE_MISMATCH, if the dimensions do not match.
47 * @return execution status
48 - \ref ARM_MATH_SUCCESS : Operation successful
49 - \ref ARM_MATH_SIZE_MISMATCH : Matrix size check failed
50 - \ref ARM_MATH_DECOMPOSITION_FAILURE : Input matrix cannot be decomposed
51 * @par
52 * If the matrix is ill conditioned or only semi-definite, then it is better using the LDL^t decomposition.
53 * The decomposition of A is returning a lower triangular matrix U such that A = U U^t
54 */
55
56 #if defined(ARM_MATH_MVE_FLOAT16) && !defined(ARM_MATH_AUTOVECTORIZE)
57
58 #include "arm_helium_utils.h"
59
arm_mat_cholesky_f16(const arm_matrix_instance_f16 * pSrc,arm_matrix_instance_f16 * pDst)60 arm_status arm_mat_cholesky_f16(
61 const arm_matrix_instance_f16 * pSrc,
62 arm_matrix_instance_f16 * pDst)
63 {
64
65 arm_status status; /* status of matrix inverse */
66
67
68 #ifdef ARM_MATH_MATRIX_CHECK
69
70 /* Check for matrix mismatch condition */
71 if ((pSrc->numRows != pSrc->numCols) ||
72 (pDst->numRows != pDst->numCols) ||
73 (pSrc->numRows != pDst->numRows) )
74 {
75 /* Set status as ARM_MATH_SIZE_MISMATCH */
76 status = ARM_MATH_SIZE_MISMATCH;
77 }
78 else
79
80 #endif /* #ifdef ARM_MATH_MATRIX_CHECK */
81
82 {
83 int i,j,k;
84 int n = pSrc->numRows;
85 _Float16 invSqrtVj;
86 float16_t *pA,*pG;
87 int kCnt;
88
89 mve_pred16_t p0;
90
91 f16x8_t acc, acc0, acc1, acc2, acc3;
92 f16x8_t vecGi;
93 f16x8_t vecGj,vecGj0,vecGj1,vecGj2,vecGj3;
94
95
96 pA = pSrc->pData;
97 pG = pDst->pData;
98
99 for(i=0 ;i < n ; i++)
100 {
101 for(j=i ; j+3 < n ; j+=4)
102 {
103 acc0 = vdupq_n_f16(0.0f16);
104 acc0[0]=pA[(j + 0) * n + i];
105
106 acc1 = vdupq_n_f16(0.0f16);
107 acc1[0]=pA[(j + 1) * n + i];
108
109 acc2 = vdupq_n_f16(0.0f16);
110 acc2[0]=pA[(j + 2) * n + i];
111
112 acc3 = vdupq_n_f16(0.0f16);
113 acc3[0]=pA[(j + 3) * n + i];
114
115 kCnt = i;
116 for(k=0; k < i ; k+=8)
117 {
118 p0 = vctp16q(kCnt);
119
120 vecGi=vldrhq_z_f16(&pG[i * n + k],p0);
121
122 vecGj0=vldrhq_z_f16(&pG[(j + 0) * n + k],p0);
123 vecGj1=vldrhq_z_f16(&pG[(j + 1) * n + k],p0);
124 vecGj2=vldrhq_z_f16(&pG[(j + 2) * n + k],p0);
125 vecGj3=vldrhq_z_f16(&pG[(j + 3) * n + k],p0);
126
127 acc0 = vfmsq_m(acc0, vecGi, vecGj0, p0);
128 acc1 = vfmsq_m(acc1, vecGi, vecGj1, p0);
129 acc2 = vfmsq_m(acc2, vecGi, vecGj2, p0);
130 acc3 = vfmsq_m(acc3, vecGi, vecGj3, p0);
131
132 kCnt -= 8;
133 }
134 pG[(j + 0) * n + i] = vecAddAcrossF16Mve(acc0);
135 pG[(j + 1) * n + i] = vecAddAcrossF16Mve(acc1);
136 pG[(j + 2) * n + i] = vecAddAcrossF16Mve(acc2);
137 pG[(j + 3) * n + i] = vecAddAcrossF16Mve(acc3);
138 }
139
140 for(; j < n ; j++)
141 {
142
143 kCnt = i;
144 acc = vdupq_n_f16(0.0f16);
145 acc[0] = pA[j * n + i];
146
147 for(k=0; k < i ; k+=8)
148 {
149 p0 = vctp16q(kCnt);
150
151 vecGi=vldrhq_z_f16(&pG[i * n + k],p0);
152 vecGj=vldrhq_z_f16(&pG[j * n + k],p0);
153
154 acc = vfmsq_m(acc, vecGi, vecGj,p0);
155
156 kCnt -= 8;
157 }
158 pG[j * n + i] = vecAddAcrossF16Mve(acc);
159 }
160
161 if (pG[i * n + i] <= 0.0f16)
162 {
163 return(ARM_MATH_DECOMPOSITION_FAILURE);
164 }
165
166 invSqrtVj = (_Float16)1.0f/sqrtf(pG[i * n + i]);
167 for(j=i; j < n ; j++)
168 {
169 pG[j * n + i] = (_Float16)pG[j * n + i] * invSqrtVj ;
170 }
171 }
172
173 status = ARM_MATH_SUCCESS;
174
175 }
176
177
178 /* Return to application */
179 return (status);
180 }
181
182 #else
arm_mat_cholesky_f16(const arm_matrix_instance_f16 * pSrc,arm_matrix_instance_f16 * pDst)183 arm_status arm_mat_cholesky_f16(
184 const arm_matrix_instance_f16 * pSrc,
185 arm_matrix_instance_f16 * pDst)
186 {
187
188 arm_status status; /* status of matrix inverse */
189
190
191 #ifdef ARM_MATH_MATRIX_CHECK
192
193 /* Check for matrix mismatch condition */
194 if ((pSrc->numRows != pSrc->numCols) ||
195 (pDst->numRows != pDst->numCols) ||
196 (pSrc->numRows != pDst->numRows) )
197 {
198 /* Set status as ARM_MATH_SIZE_MISMATCH */
199 status = ARM_MATH_SIZE_MISMATCH;
200 }
201 else
202
203 #endif /* #ifdef ARM_MATH_MATRIX_CHECK */
204
205 {
206 int i,j,k;
207 int n = pSrc->numRows;
208 float16_t invSqrtVj;
209 float16_t *pA,*pG;
210
211 pA = pSrc->pData;
212 pG = pDst->pData;
213
214
215 for(i=0 ; i < n ; i++)
216 {
217 for(j=i ; j < n ; j++)
218 {
219 pG[j * n + i] = pA[j * n + i];
220
221 for(k=0; k < i ; k++)
222 {
223 pG[j * n + i] = pG[j * n + i] - pG[i * n + k] * pG[j * n + k];
224 }
225 }
226
227 if (pG[i * n + i] <= 0.0f)
228 {
229 return(ARM_MATH_DECOMPOSITION_FAILURE);
230 }
231
232 invSqrtVj = 1.0f/sqrtf(pG[i * n + i]);
233 for(j=i ; j < n ; j++)
234 {
235 pG[j * n + i] = pG[j * n + i] * invSqrtVj ;
236 }
237 }
238
239 status = ARM_MATH_SUCCESS;
240
241 }
242
243
244 /* Return to application */
245 return (status);
246 }
247
248 #endif /* defined(ARM_MATH_MVEF) && !defined(ARM_MATH_AUTOVECTORIZE) */
249
250 /**
251 @} end of MatrixChol group
252 */
253 #endif /* #if defined(ARM_FLOAT16_SUPPORTED) */
254
