Skip to content

Commit 9f7a9a3

Browse files
authored
Merge pull request #2252 from thrasibule/trtrs
Optimized ?trtrs
2 parents 3f1077c + 5d6525c commit 9f7a9a3

File tree

12 files changed

+1536
-18
lines changed

12 files changed

+1536
-18
lines changed

common_lapack.h

Lines changed: 146 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -293,4 +293,150 @@ blasint zlarf_R(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLO
293293
blasint xlarf_L(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
294294
blasint xlarf_R(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
295295

296+
blasint strtrs_UNU_single(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
297+
blasint strtrs_UNN_single(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
298+
blasint strtrs_UTU_single(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
299+
blasint strtrs_UTN_single(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
300+
blasint strtrs_LNU_single(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
301+
blasint strtrs_LNN_single(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
302+
blasint strtrs_LTU_single(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
303+
blasint strtrs_LTN_single(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
304+
blasint dtrtrs_UNU_single(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
305+
blasint dtrtrs_UNN_single(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
306+
blasint dtrtrs_UTU_single(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
307+
blasint dtrtrs_UTN_single(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
308+
blasint dtrtrs_LNU_single(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
309+
blasint dtrtrs_LNN_single(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
310+
blasint dtrtrs_LTU_single(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
311+
blasint dtrtrs_LTN_single(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
312+
blasint qtrtrs_UNU_single(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
313+
blasint qtrtrs_UNN_single(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
314+
blasint qtrtrs_UTU_single(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
315+
blasint qtrtrs_UTN_single(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
316+
blasint qtrtrs_LNU_single(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
317+
blasint qtrtrs_LNN_single(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
318+
blasint qtrtrs_LTU_single(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
319+
blasint qtrtrs_LTN_single(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
320+
blasint ctrtrs_UNU_single(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
321+
blasint ctrtrs_UNN_single(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
322+
blasint ctrtrs_UTU_single(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
323+
blasint ctrtrs_UTN_single(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
324+
blasint ctrtrs_URU_single(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
325+
blasint ctrtrs_URN_single(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
326+
blasint ctrtrs_UCU_single(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
327+
blasint ctrtrs_UCN_single(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
328+
blasint ctrtrs_LNU_single(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
329+
blasint ctrtrs_LNN_single(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
330+
blasint ctrtrs_LTU_single(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
331+
blasint ctrtrs_LTN_single(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
332+
blasint ctrtrs_LRU_single(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
333+
blasint ctrtrs_LRN_single(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
334+
blasint ctrtrs_LCU_single(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
335+
blasint ctrtrs_LCN_single(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
336+
blasint ztrtrs_UNU_single(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
337+
blasint ztrtrs_UNN_single(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
338+
blasint ztrtrs_UTU_single(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
339+
blasint ztrtrs_UTN_single(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
340+
blasint ztrtrs_URU_single(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
341+
blasint ztrtrs_URN_single(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
342+
blasint ztrtrs_UCU_single(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
343+
blasint ztrtrs_UCN_single(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
344+
blasint ztrtrs_LNU_single(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
345+
blasint ztrtrs_LNN_single(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
346+
blasint ztrtrs_LTU_single(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
347+
blasint ztrtrs_LTN_single(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
348+
blasint ztrtrs_LRU_single(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
349+
blasint ztrtrs_LRN_single(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
350+
blasint ztrtrs_LCU_single(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
351+
blasint ztrtrs_LCN_single(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
352+
blasint xtrtrs_UNU_single(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
353+
blasint xtrtrs_UNN_single(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
354+
blasint xtrtrs_UTU_single(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
355+
blasint xtrtrs_UTN_single(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
356+
blasint xtrtrs_URU_single(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
357+
blasint xtrtrs_URN_single(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
358+
blasint xtrtrs_UCU_single(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
359+
blasint xtrtrs_UCN_single(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
360+
blasint xtrtrs_LNU_single(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
361+
blasint xtrtrs_LNN_single(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
362+
blasint xtrtrs_LTU_single(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
363+
blasint xtrtrs_LTN_single(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
364+
blasint xtrtrs_LRU_single(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
365+
blasint xtrtrs_LRN_single(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
366+
blasint xtrtrs_LCU_single(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
367+
blasint xtrtrs_LCN_single(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
368+
369+
blasint strtrs_UNU_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
370+
blasint strtrs_UNN_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
371+
blasint strtrs_UTU_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
372+
blasint strtrs_UTN_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
373+
blasint strtrs_LNU_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
374+
blasint strtrs_LNN_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
375+
blasint strtrs_LTU_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
376+
blasint strtrs_LTN_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
377+
blasint dtrtrs_UNU_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
378+
blasint dtrtrs_UNN_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
379+
blasint dtrtrs_UTU_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
380+
blasint dtrtrs_UTN_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
381+
blasint dtrtrs_LNU_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
382+
blasint dtrtrs_LNN_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
383+
blasint dtrtrs_LTU_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
384+
blasint dtrtrs_LTN_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
385+
blasint qtrtrs_UNU_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
386+
blasint qtrtrs_UNN_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
387+
blasint qtrtrs_UTU_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
388+
blasint qtrtrs_UTN_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
389+
blasint qtrtrs_LNU_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
390+
blasint qtrtrs_LNN_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
391+
blasint qtrtrs_LTU_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
392+
blasint qtrtrs_LTN_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
393+
blasint ctrtrs_UNU_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
394+
blasint ctrtrs_UNN_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
395+
blasint ctrtrs_UTU_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
396+
blasint ctrtrs_UTN_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
397+
blasint ctrtrs_URU_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
398+
blasint ctrtrs_URN_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
399+
blasint ctrtrs_UCU_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
400+
blasint ctrtrs_UCN_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
401+
blasint ctrtrs_LNU_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
402+
blasint ctrtrs_LNN_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
403+
blasint ctrtrs_LTU_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
404+
blasint ctrtrs_LTN_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
405+
blasint ctrtrs_LRU_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
406+
blasint ctrtrs_LRN_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
407+
blasint ctrtrs_LCU_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
408+
blasint ctrtrs_LCN_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
409+
blasint ztrtrs_UNU_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
410+
blasint ztrtrs_UNN_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
411+
blasint ztrtrs_UTU_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
412+
blasint ztrtrs_UTN_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
413+
blasint ztrtrs_URU_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
414+
blasint ztrtrs_URN_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
415+
blasint ztrtrs_UCU_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
416+
blasint ztrtrs_UCN_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
417+
blasint ztrtrs_LNU_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
418+
blasint ztrtrs_LNN_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
419+
blasint ztrtrs_LTU_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
420+
blasint ztrtrs_LTN_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
421+
blasint ztrtrs_LRU_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
422+
blasint ztrtrs_LRN_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
423+
blasint ztrtrs_LCU_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
424+
blasint ztrtrs_LCN_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
425+
blasint xtrtrs_UNU_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
426+
blasint xtrtrs_UNN_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
427+
blasint xtrtrs_UTU_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
428+
blasint xtrtrs_UTN_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
429+
blasint xtrtrs_URU_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
430+
blasint xtrtrs_URN_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
431+
blasint xtrtrs_UCU_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
432+
blasint xtrtrs_UCN_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
433+
blasint xtrtrs_LNU_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
434+
blasint xtrtrs_LNN_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
435+
blasint xtrtrs_LTU_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
436+
blasint xtrtrs_LTN_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
437+
blasint xtrtrs_LRU_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
438+
blasint xtrtrs_LRN_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
439+
blasint xtrtrs_LCU_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
440+
blasint xtrtrs_LCN_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
441+
296442
#endif

0 commit comments

Comments
 (0)