1 #include "UnaryF32.h"
2 #include "Error.h"
3
4 /* Upper bound of maximum matrix dimension used by Python */
5 #define MAXMATRIXDIM 40
6
7 /*
8
9 Offset in input test pattern for matrix of dimension d * d.
10 Must be coherent with Python script Matrix.py
11
12 */
cholesky_offset(int d)13 static int cholesky_offset(int d)
14 {
15 int offset=14;
16 switch (d)
17 {
18 case 4:
19 offset = 14;
20 break;
21 case 8:
22 offset = 79;
23 break;
24 case 9:
25 offset = 143;
26 break;
27 case 15:
28 offset = 224;
29 break;
30 case 16:
31 offset = 449;
32 break;
33 default:
34 offset = 14;
35 break;
36 }
37
38 return(offset);
39 }
40
41
test_mat_scale_f32()42 void UnaryF32::test_mat_scale_f32()
43 {
44 arm_mat_scale_f32(&this->in1,0.5,&this->out);
45 }
46
test_mat_inverse_f32()47 void UnaryF32::test_mat_inverse_f32()
48 {
49 arm_mat_inverse_f32(&this->in1,&this->out);
50 }
51
test_mat_trans_f32()52 void UnaryF32::test_mat_trans_f32()
53 {
54 arm_mat_trans_f32(&this->in1,&this->out);
55 }
56
test_mat_cmplx_trans_f32()57 void UnaryF32::test_mat_cmplx_trans_f32()
58 {
59 arm_mat_cmplx_trans_f32(&this->in1,&this->out);
60 }
61
test_mat_add_f32()62 void UnaryF32::test_mat_add_f32()
63 {
64 arm_mat_add_f32(&this->in1,&this->in1,&this->out);
65 }
66
test_mat_sub_f32()67 void UnaryF32::test_mat_sub_f32()
68 {
69 arm_mat_sub_f32(&this->in1,&this->in1,&this->out);
70 }
71
test_mat_vec_mult_f32()72 void UnaryF32::test_mat_vec_mult_f32()
73 {
74 arm_mat_vec_mult_f32(&this->in1, vecp, outp);
75 }
76
test_mat_cholesky_dpo_f32()77 void UnaryF32::test_mat_cholesky_dpo_f32()
78 {
79 arm_mat_cholesky_f32(&this->in1,&this->out);
80 }
81
test_solve_upper_triangular_f32()82 void UnaryF32::test_solve_upper_triangular_f32()
83 {
84 arm_mat_solve_upper_triangular_f32(&this->in1,&this->in2,&this->out);
85 }
86
test_solve_lower_triangular_f32()87 void UnaryF32::test_solve_lower_triangular_f32()
88 {
89 arm_mat_solve_lower_triangular_f32(&this->in1,&this->in2,&this->out);
90 }
91
test_ldlt_decomposition_f32()92 void UnaryF32::test_ldlt_decomposition_f32()
93 {
94 arm_mat_ldlt_f32(&this->in1,&this->outll,&this->outd,(uint16_t*)outp);
95 }
96
97
setUp(Testing::testID_t id,std::vector<Testing::param_t> & params,Client::PatternMgr * mgr)98 void UnaryF32::setUp(Testing::testID_t id,std::vector<Testing::param_t>& params,Client::PatternMgr *mgr)
99 {
100
101
102 std::vector<Testing::param_t>::iterator it = params.begin();
103
104 this->nbr = *it++;
105 this->nbc = *it;
106
107
108 switch(id)
109 {
110 case TEST_MAT_VEC_MULT_F32_6:
111 input1.reload(UnaryF32::INPUTA_F32_ID,mgr,this->nbr*this->nbc);
112 vec.reload(UnaryF32::INPUTVEC1_F32_ID,mgr,this->nbc);
113 output.create(this->nbr,UnaryF32::OUT_F32_ID,mgr);
114 vecp=vec.ptr();
115 outp=output.ptr();
116
117 this->in1.numRows = this->nbr;
118 this->in1.numCols = this->nbc;
119 this->in1.pData = input1.ptr();
120 break;
121 case TEST_MAT_TRANS_F32_3:
122 input1.reload(UnaryF32::INPUTA_F32_ID,mgr,this->nbr*this->nbc);
123 output.create(this->nbr*this->nbc,UnaryF32::OUT_F32_ID,mgr);
124
125 this->out.numRows = this->nbc;
126 this->out.numCols = this->nbr;
127 this->out.pData = output.ptr();
128
129 this->in1.numRows = this->nbr;
130 this->in1.numCols = this->nbc;
131 this->in1.pData = input1.ptr();
132 break;
133 case TEST_MAT_CMPLX_TRANS_F32_7:
134 input1.reload(UnaryF32::INPUTAC_F32_ID,mgr,2*this->nbr*this->nbc);
135 output.create(2*this->nbr*this->nbc,UnaryF32::OUT_F32_ID,mgr);
136
137 this->out.numRows = this->nbc;
138 this->out.numCols = this->nbr;
139 this->out.pData = output.ptr();
140
141 this->in1.numRows = this->nbr;
142 this->in1.numCols = this->nbc;
143 this->in1.pData = input1.ptr();
144 break;
145
146 case TEST_MAT_CHOLESKY_DPO_F32_8:
147 {
148 int offset=14;
149 float32_t *p;
150 float32_t *aPtr;
151 input1.reload(UnaryF32::INPUTSCHOLESKY1_DPO_F32_ID,mgr);
152 output.create(this->nbc * this->nbr,UnaryF32::OUT_F32_ID,mgr);
153
154 a.create(this->nbr*this->nbc,UnaryF32::TMPA_F32_ID,mgr);
155
156 /* Offsets must be coherent with the sizes used in python script
157 Matrix.py for pattern generation */
158 offset=cholesky_offset(this->nbr);
159
160 p = input1.ptr();
161 aPtr = a.ptr();
162
163
164
165 memcpy(aPtr,p + offset,sizeof(float32_t)*this->nbr*this->nbr);
166
167 this->out.numRows = this->nbr;
168 this->out.numCols = this->nbc;
169 this->out.pData = output.ptr();
170
171 this->in1.numRows = this->nbr;
172 this->in1.numCols = this->nbc;
173 this->in1.pData = aPtr;
174
175
176
177 }
178 break;
179
180 case TEST_SOLVE_UPPER_TRIANGULAR_F32_9:
181 {
182 int offset=14;
183 float32_t *p;
184 float32_t *aPtr;
185 float32_t *bPtr;
186
187 input1.reload(UnaryF32::INPUT_UT_DPO_F32_ID,mgr);
188 input2.reload(UnaryF32::INPUT_RNDA_DPO_F32_ID,mgr);
189 output.create(this->nbc * this->nbr,UnaryF32::OUT_F32_ID,mgr);
190
191 a.create(this->nbr*this->nbc,UnaryF32::TMPA_F32_ID,mgr);
192 b.create(this->nbr*this->nbc,UnaryF32::TMPB_F32_ID,mgr);
193
194 /* Offsets must be coherent with the sizes used in python script
195 Matrix.py for pattern generation */
196 offset=cholesky_offset(this->nbr);
197
198 p = input1.ptr();
199 aPtr = a.ptr();
200 memcpy(aPtr,&p[offset],sizeof(float32_t)*this->nbr*this->nbr);
201
202 p = input2.ptr();
203 bPtr = b.ptr();
204 memcpy(bPtr,&p[offset],sizeof(float32_t)*this->nbr*this->nbr);
205
206 this->out.numRows = this->nbr;
207 this->out.numCols = this->nbc;
208 this->out.pData = output.ptr();
209
210 this->in1.numRows = this->nbr;
211 this->in1.numCols = this->nbc;
212 this->in1.pData = aPtr;
213
214 this->in2.numRows = this->nbr;
215 this->in2.numCols = this->nbc;
216 this->in2.pData = bPtr;
217 }
218 break;
219
220 case TEST_SOLVE_LOWER_TRIANGULAR_F32_10:
221 {
222 int offset=14;
223 float32_t *p;
224 float32_t *aPtr;
225 float32_t *bPtr;
226
227 input1.reload(UnaryF32::INPUT_LT_DPO_F32_ID,mgr);
228 input2.reload(UnaryF32::INPUT_RNDA_DPO_F32_ID,mgr);
229 output.create(this->nbc * this->nbr,UnaryF32::OUT_F32_ID,mgr);
230
231 a.create(this->nbr*this->nbc,UnaryF32::TMPA_F32_ID,mgr);
232 b.create(this->nbr*this->nbc,UnaryF32::TMPB_F32_ID,mgr);
233
234 /* Offsets must be coherent with the sizes used in python script
235 Matrix.py for pattern generation */
236 offset=cholesky_offset(this->nbr);
237
238 p = input1.ptr();
239 aPtr = a.ptr();
240 memcpy(aPtr,&p[offset],sizeof(float32_t)*this->nbr*this->nbr);
241
242 p = input2.ptr();
243 bPtr = b.ptr();
244 memcpy(bPtr,&p[offset],sizeof(float32_t)*this->nbr*this->nbr);
245
246 this->out.numRows = this->nbr;
247 this->out.numCols = this->nbc;
248 this->out.pData = output.ptr();
249
250 this->in1.numRows = this->nbr;
251 this->in1.numCols = this->nbc;
252 this->in1.pData = aPtr;
253
254 this->in2.numRows = this->nbr;
255 this->in2.numCols = this->nbc;
256 this->in2.pData = bPtr;
257 }
258 break;
259
260 case TEST_LDLT_DECOMPOSITION_F32_11:
261 {
262 float32_t *p, *aPtr;
263
264 int offset=14;
265 input1.reload(UnaryF32::INPUTSCHOLESKY1_DPO_F32_ID,mgr);
266
267
268 outputll.create(this->nbr*this->nbr,UnaryF32::LL_F32_ID,mgr);
269 outputd.create(this->nbr*this->nbr,UnaryF32::D_F32_ID,mgr);
270 outputp.create(this->nbr,UnaryF32::PERM_S16_ID,mgr);
271
272
273 a.create(MAXMATRIXDIM*MAXMATRIXDIM,UnaryF32::TMPA_F32_ID,mgr);
274
275 /* Offsets must be coherent with the sizes used in python script
276 Matrix.py for pattern generation */
277 offset=cholesky_offset(this->nbr);
278
279 p = input1.ptr();
280 aPtr = a.ptr();
281 memcpy(aPtr,&p[offset],sizeof(float32_t)*this->nbr*this->nbr);
282
283 this->in1.numRows = this->nbr;
284 this->in1.numCols = this->nbc;
285 this->in1.pData = aPtr;
286
287 this->outll.numRows = this->nbr;
288 this->outll.numCols = this->nbc;
289 this->outll.pData = outputll.ptr();
290
291 this->outd.numRows = this->nbr;
292 this->outd.numCols = this->nbc;
293 this->outd.pData = outputd.ptr();
294
295
296 outpp = outputp.ptr();
297 }
298 break;
299
300 default:
301 input1.reload(UnaryF32::INPUTA_F32_ID,mgr,this->nbr*this->nbc);
302 output.create(this->nbr*this->nbc,UnaryF32::OUT_F32_ID,mgr);
303
304 this->out.numRows = this->nbr;
305 this->out.numCols = this->nbc;
306 this->out.pData = output.ptr();
307
308 this->in1.numRows = this->nbr;
309 this->in1.numCols = this->nbc;
310 this->in1.pData = input1.ptr();
311 break;
312 }
313
314
315
316
317
318
319
320 }
321
tearDown(Testing::testID_t id,Client::PatternMgr * mgr)322 void UnaryF32::tearDown(Testing::testID_t id,Client::PatternMgr *mgr)
323 {
324 }
325
