Remove references to tsmqr/ttmqr/tsmlq/ttmlq/...

compute/pzgelqf.c

@@ -19,7 +19,7 @@
  * @author Mathieu Faverge
  * @author Emmanuel Agullo
  * @author Cedric Castagnede
- * @date 2018-11-09
+ * @date 2018-11-08
  * @precisions normal z -> s d c
  *
  */
@@ -67,7 +67,7 @@ void chameleon_pzgelqf( int genD, CHAM_desc_t *A, CHAM_desc_t *T, CHAM_desc_t *D
     /*
      * zgelqt  = A->nb * (ib+1)
      * zunmlq  = A->nb * ib
-     * ztslqt  = A->nb * (ib+1)
+     * ztplqt  = A->nb * (ib+1)
      * ztpmlqt = A->nb * ib
      */
     ws_worker = A->nb * (ib+1);

compute/pzgeqrf.c

@@ -19,7 +19,7 @@
  * @author Mathieu Faverge
  * @author Emmanuel Agullo
  * @author Cedric Castagnede
- * @date 2018-11-09
+ * @date 2018-11-08
  * @precisions normal z -> s d c
  *
  */
@@ -62,7 +62,7 @@ void chameleon_pzgeqrf( int genD, CHAM_desc_t *A, CHAM_desc_t *T, CHAM_desc_t *D
     /*
      * zgeqrt  = A->nb * (ib+1)
      * zunmqr  = A->nb * ib
-     * ztsqrt  = A->nb * (ib+1)
+     * ztpqrt  = A->nb * (ib+1)
      * ztpmqrt = A->nb * ib
      */
     ws_worker = A->nb * (ib+1);

cudablas/eztrace_module/cudablas_eztrace_module

@@ -273,6 +273,30 @@ int CUDA_ctsmqr(
             void *WORK,  void* LDWORK,
             void *WORKC, void* LDWORKC,
         void* stream);
+int CUDA_cttmlq(
+        void* side, void* trans,
+        void* M1, void* N1,
+        void* M2, void* N2,
+        void* K, void* IB,
+        void *A1, void* LDA1,
+        void *A2, void* LDA2,
+        void *V, void* LDV,
+        void *T, void* LDT,
+              void *WORK,  void* LDWORK,
+              void *WORKC, void* LDWORKC,
+        void* stream);
+int CUDA_cttmqr(
+        void* side, void* trans,
+        void* M1, void* N1,
+        void* M2, void* N2,
+        void* K, void* IB,
+        void *A1, void* LDA1,
+        void *A2, void* LDA2,
+        void *V, void* LDV,
+        void *T, void* LDT,
+            void *WORK,  void* LDWORK,
+            void *WORKC, void* LDWORKC,
+        void* stream);
 int CUDA_ctsqrt(
         void* m, void* n, void* nb,
         void *da1, void* ldda1,
@@ -528,6 +552,30 @@ int CUDA_dtsmqr(
             double *WORK,  void* LDWORK,
             double *WORKC, void* LDWORKC,
         void* stream);
+int CUDA_dttmlq(
+        void* side, void* trans,
+        void* M1, void* N1,
+        void* M2, void* N2,
+        void* K, void* IB,
+        double *A1, void* LDA1,
+        double *A2, void* LDA2,
+        const double *V, void* LDV,
+        const double *T, void* LDT,
+              double *WORK,  void* LDWORK,
+              double *WORKC, void* LDWORKC,
+        void* stream);
+int CUDA_dttmqr(
+        void* side, void* trans,
+        void* M1, void* N1,
+        void* M2, void* N2,
+        void* K, void* IB,
+        double *A1, void* LDA1,
+        double *A2, void* LDA2,
+        const double *V, void* LDV,
+        const double *T, void* LDT,
+            double *WORK,  void* LDWORK,
+            double *WORKC, void* LDWORKC,
+        void* stream);
 int CUDA_dtsqrt(
         void* m, void* n, void* nb,
         double *da1, void* ldda1,
@@ -783,6 +831,30 @@ int CUDA_stsmqr(
             float *WORK,  void* LDWORK,
             float *WORKC, void* LDWORKC,
         void* stream);
+int CUDA_sttmlq(
+        void* side, void* trans,
+        void* M1, void* N1,
+        void* M2, void* N2,
+        void* K, void* IB,
+        float *A1, void* LDA1,
+        float *A2, void* LDA2,
+        const float *V, void* LDV,
+        const float *T, void* LDT,
+              float *WORK,  void* LDWORK,
+              float *WORKC, void* LDWORKC,
+        void* stream);
+int CUDA_sttmqr(
+        void* side, void* trans,
+        void* M1, void* N1,
+        void* M2, void* N2,
+        void* K, void* IB,
+        float *A1, void* LDA1,
+        float *A2, void* LDA2,
+        const float *V, void* LDV,
+        const float *T, void* LDT,
+            float *WORK,  void* LDWORK,
+            float *WORKC, void* LDWORKC,
+        void* stream);
 int CUDA_stsqrt(
         void* m, void* n, void* nb,
         float *da1, void* ldda1,
@@ -1090,6 +1162,30 @@ int CUDA_ztsmqr(
             void *WORK,  void* LDWORK,
             void *WORKC, void* LDWORKC,
         void* stream);
+int CUDA_zttmlq(
+        void* side, void* trans,
+        void* M1, void* N1,
+        void* M2, void* N2,
+        void* K, void* IB,
+        void *A1, void* LDA1,
+        void *A2, void* LDA2,
+        void *V, void* LDV,
+        void *T, void* LDT,
+              void *WORK,  void* LDWORK,
+              void *WORKC, void* LDWORKC,
+        void* stream);
+int CUDA_zttmqr(
+        void* side, void* trans,
+        void* M1, void* N1,
+        void* M2, void* N2,
+        void* K, void* IB,
+        void *A1, void* LDA1,
+        void *A2, void* LDA2,
+        void *V, void* LDV,
+        void *T, void* LDT,
+            void *WORK,  void* LDWORK,
+            void *WORKC, void* LDWORKC,
+        void* stream);
 int CUDA_ztsqrt(
         void* m, void* n, void* nb,
         void *da1, void* ldda1,

include/chameleon/tasks.h

@@ -14,7 +14,7 @@
  * @version 1.0.0
  * @author Mathieu Faverge
  * @author Cedric Augonnet
- * @date 2011-06-01
+ * @date 2018-11-08
  *
  */
 #ifndef _chameleon_tasks_h_
@@ -54,16 +54,12 @@ typedef enum chameleon_tasktype_e {
   TASK_ORMQR,
   TASK_POTRF,
   TASK_SSSSM,
+  TASK_TPLQT,
+  TASK_TPMLQT,
+  TASK_TPMQRT,
+  TASK_TPQRT,
   TASK_TRTRI,
-  TASK_TSLQT,
-  TASK_TSMLQ,
-  TASK_TSMQR,
-  TASK_TSQRT,
   TASK_TSTRF,
-  TASK_TTLQT,
-  TASK_TTMLQ,
-  TASK_TTMQR,
-  TASK_TTQRT,
   TASK_UNMLQ,
   TASK_UNMQR,
 
@@ -86,6 +82,15 @@ typedef enum chameleon_tasktype_e {
   TASK_NBKERNELS
 } cham_tasktype_t;
 
+#define TASK_TSLQT TASK_TPLQT
+#define TASK_TSMLQ TASK_TPMLQT
+#define TASK_TSMQR TASK_TPMQRT
+#define TASK_TSQRT TASK_TPQRT
+#define TASK_TTLQT TASK_TPLQT
+#define TASK_TTMLQ TASK_TPMLQT
+#define TASK_TTMQR TASK_TPMQRT
+#define TASK_TTQRT TASK_TPQRT
+
 typedef int (*cham_unary_operator_t)( const CHAM_desc_t *desc,
                                       cham_uplo_t uplo, int m, int n,
                                       void *data, void *op_args );

runtime/quark/codelets/codelet_ztplqt.c

@@ -13,7 +13,7 @@
  *
  * @version 1.0.0
  * @author Mathieu Faverge
- * @date 2016-12-15
+ * @date 2018-11-08
  * @precisions normal z -> s d c
  *
  */
@@ -50,7 +50,7 @@ void INSERT_TASK_ztplqt( const RUNTIME_option_t *options,
                          const CHAM_desc_t *T, int Tm, int Tn, int ldt )
 {
     quark_option_t *opt = (quark_option_t*)(options->schedopt);
-    DAG_CORE_TSLQT;
+    DAG_CORE_TPLQT;
 
     int shapeB = ( L == 0 ) ? 0 : (QUARK_REGION_L | QUARK_REGION_D);
 

runtime/quark/codelets/codelet_ztpmlqt.c

@@ -13,7 +13,7 @@
  *
  * @version 1.0.0
  * @author Mathieu Faverge
- * @date 2016-12-15
+ * @date 2018-11-08
  * @precisions normal z -> s d c
  *
  */
@@ -57,7 +57,7 @@ void INSERT_TASK_ztpmlqt( const RUNTIME_option_t *options,
                          const CHAM_desc_t *B, int Bm, int Bn, int ldb )
 {
     quark_option_t *opt = (quark_option_t*)(options->schedopt);
-    DAG_CORE_TSMQR;
+    DAG_CORE_TPMLQRT;
 
     int shapeV = ( L == 0 ) ? 0 : (QUARK_REGION_L | QUARK_REGION_D);
 

runtime/quark/codelets/codelet_ztpmqrt.c

@@ -13,7 +13,7 @@
  *
  * @version 1.0.0
  * @author Mathieu Faverge
- * @date 2016-12-15
+ * @date 2018-11-08
  * @precisions normal z -> s d c
  *
  */
@@ -57,7 +57,7 @@ void INSERT_TASK_ztpmqrt( const RUNTIME_option_t *options,
                          const CHAM_desc_t *B, int Bm, int Bn, int ldb )
 {
     quark_option_t *opt = (quark_option_t*)(options->schedopt);
-    DAG_CORE_TSMQR;
+    DAG_CORE_TPMQRT;
 
     int shapeV = ( L == 0 ) ? 0 : (QUARK_REGION_U | QUARK_REGION_D);
 

runtime/quark/codelets/codelet_ztpqrt.c

@@ -13,7 +13,7 @@
  *
  * @version 1.0.0
  * @author Mathieu Faverge
- * @date 2016-12-15
+ * @date 2018-11-08
  * @precisions normal z -> s d c
  *
  */
@@ -50,7 +50,7 @@ void INSERT_TASK_ztpqrt( const RUNTIME_option_t *options,
                          const CHAM_desc_t *T, int Tm, int Tn, int ldt )
 {
     quark_option_t *opt = (quark_option_t*)(options->schedopt);
-    DAG_CORE_TSQRT;
+    DAG_CORE_TPQRT;
 
     int shapeB = ( L == 0 ) ? 0 : (QUARK_REGION_U | QUARK_REGION_D);
 

runtime/quark/include/core_blas_dag.h

@@ -14,7 +14,7 @@
  * @version 1.0.0
  * @author Mathieu Faverge
  * @author Cedric Castagnede
- * @date 2010-11-15
+ * @date 2018-11-08
  *
  */
 #ifndef _core_blas_dag_h_
@@ -71,16 +71,21 @@
 #define DAG_CORE_TRSM       DAG_SET_PROPERTIES( "TRSM"      , "cyan"    )
 #define DAG_CORE_TRSSQ      DAG_SET_PROPERTIES( "TRSSQ"     , "white"   )
 #define DAG_CORE_TRTRI      DAG_SET_PROPERTIES( "TRTRI"     , "white"   )
-#define DAG_CORE_TSLQT      DAG_SET_PROPERTIES( "TSLQT"     , "red"     )
-#define DAG_CORE_TSMLQ      DAG_SET_PROPERTIES( "TSMLQ"     , "yellow"  )
-#define DAG_CORE_TSMQR      DAG_SET_PROPERTIES( "TSMQR"     , "yellow"  )
-#define DAG_CORE_TSQRT      DAG_SET_PROPERTIES( "TSQRT"     , "red"     )
+#define DAG_CORE_TPLQT      DAG_SET_PROPERTIES( "TPLQT"     , "red"     )
+#define DAG_CORE_TPMLQT     DAG_SET_PROPERTIES( "TPMLQT"    , "yellow"  )
+#define DAG_CORE_TPMQRT     DAG_SET_PROPERTIES( "TPMQRT"    , "yellow"  )
+#define DAG_CORE_TPQRT      DAG_SET_PROPERTIES( "TPQRT"     , "red"     )
 #define DAG_CORE_TSTRF      DAG_SET_PROPERTIES( "TSTRF"     , "red"     )
-#define DAG_CORE_TTLQT      DAG_SET_PROPERTIES( "TTLQT"     , "pink"    )
-#define DAG_CORE_TTMLQ      DAG_SET_PROPERTIES( "TTMLQ"     , "magenta" )
-#define DAG_CORE_TTMQR      DAG_SET_PROPERTIES( "TTMQR"     , "magenta" )
-#define DAG_CORE_TTQRT      DAG_SET_PROPERTIES( "TTQRT"     , "pink"    )
 #define DAG_CORE_UNMLQ      DAG_SET_PROPERTIES( "UNMLQ"     , "cyan"    )
 #define DAG_CORE_UNMQR      DAG_SET_PROPERTIES( "UNMQR"     , "cyan"    )
 
+#define DAG_CORE_TSLQT      DAG_CORE_TPLQT
+#define DAG_CORE_TSMLQ      DAG_CORE_TPMLQT
+#define DAG_CORE_TSMQR      DAG_CORE_TPMQRT
+#define DAG_CORE_TSQRT      DAG_CORE_TPQRT
+#define DAG_CORE_TTLQT      DAG_CORE_TPLQT
+#define DAG_CORE_TTMLQ      DAG_CORE_TPMLQT
+#define DAG_CORE_TTMQR      DAG_CORE_TPMQRT
+#define DAG_CORE_TTQRT      DAG_CORE_TPQRT
+
 #endif /* _core_blas_dag_h_ */

runtime/starpu/codelets/codelet_zcallback.c

@@ -15,7 +15,7 @@
  *  @author Mathieu Faverge
  *  @author Cedric Augonnet
  *  @author Florent Pruvost
- *  @date 2015-09-16
+ *  @date 2018-11-08
  *  @precisions normal z -> c d s
  *
  */
@@ -69,16 +69,8 @@ CHAMELEON_CL_CB(ztrasm,        starpu_matrix_get_nx(task->handles[0]), starpu_ma
 CHAMELEON_CL_CB(ztrmm,         starpu_matrix_get_nx(task->handles[1]), starpu_matrix_get_ny(task->handles[1]), 0,                                               M*M*N)
 CHAMELEON_CL_CB(ztrsm,         starpu_matrix_get_nx(task->handles[1]), starpu_matrix_get_ny(task->handles[1]), 0,                                               M*M*N)
 CHAMELEON_CL_CB(ztrtri,        starpu_matrix_get_nx(task->handles[0]), starpu_matrix_get_nx(task->handles[0]), starpu_matrix_get_nx(task->handles[0]), (1./3.)*M *M*M)
-CHAMELEON_CL_CB(ztslqt,        starpu_matrix_get_nx(task->handles[0]), starpu_matrix_get_nx(task->handles[0]), starpu_matrix_get_nx(task->handles[0]),     2. *M* M*M)
-CHAMELEON_CL_CB(ztsmlq,        starpu_matrix_get_nx(task->handles[0]), starpu_matrix_get_nx(task->handles[0]), starpu_matrix_get_nx(task->handles[0]), (4.0*M+starpu_matrix_get_nx(task->handles[3]))*M*M)
-CHAMELEON_CL_CB(ztsmqr,        starpu_matrix_get_nx(task->handles[0]), starpu_matrix_get_nx(task->handles[0]), starpu_matrix_get_nx(task->handles[0]), (4.0*M+starpu_matrix_get_nx(task->handles[3]))*M*M)
 CHAMELEON_CL_CB(ztsmlq_hetra1, starpu_matrix_get_nx(task->handles[0]), starpu_matrix_get_nx(task->handles[0]), starpu_matrix_get_nx(task->handles[0]), (4.0*M+starpu_matrix_get_nx(task->handles[3]))*M*M)
 CHAMELEON_CL_CB(ztsmqr_hetra1, starpu_matrix_get_nx(task->handles[0]), starpu_matrix_get_nx(task->handles[0]), starpu_matrix_get_nx(task->handles[0]), (4.0*M+starpu_matrix_get_nx(task->handles[3]))*M*M)
-CHAMELEON_CL_CB(ztsqrt,        starpu_matrix_get_nx(task->handles[0]), starpu_matrix_get_nx(task->handles[0]), starpu_matrix_get_nx(task->handles[0]),     2. *M* M*M)
 CHAMELEON_CL_CB(ztstrf,        starpu_matrix_get_nx(task->handles[0]), starpu_matrix_get_nx(task->handles[0]), starpu_matrix_get_nx(task->handles[0]),         M* M*M)
-CHAMELEON_CL_CB(zttlqt,        starpu_matrix_get_nx(task->handles[0]), starpu_matrix_get_nx(task->handles[0]), starpu_matrix_get_nx(task->handles[0]),     1. *M* M*M)
-CHAMELEON_CL_CB(zttmlq,        starpu_matrix_get_nx(task->handles[0]), starpu_matrix_get_nx(task->handles[0]), starpu_matrix_get_nx(task->handles[0]), (2.0*M+starpu_matrix_get_nx(task->handles[3]))*M*M)
-CHAMELEON_CL_CB(zttmqr,        starpu_matrix_get_nx(task->handles[0]), starpu_matrix_get_nx(task->handles[0]), starpu_matrix_get_nx(task->handles[0]), (2.0*M+starpu_matrix_get_nx(task->handles[3]))*M*M)
-CHAMELEON_CL_CB(zttqrt,        starpu_matrix_get_nx(task->handles[0]), starpu_matrix_get_nx(task->handles[0]), starpu_matrix_get_nx(task->handles[0]),     1. *M* M*M)
 CHAMELEON_CL_CB(zunmlq,        starpu_matrix_get_nx(task->handles[0]), starpu_matrix_get_nx(task->handles[0]), starpu_matrix_get_nx(task->handles[0]),     2. *M* M*M)
 CHAMELEON_CL_CB(zunmqr,        starpu_matrix_get_nx(task->handles[0]), starpu_matrix_get_nx(task->handles[0]), starpu_matrix_get_nx(task->handles[0]),     2. *M* M*M)

runtime/starpu/control/runtime_zlocality.c

@@ -15,7 +15,7 @@
  * @author Cedric Augonnet
  * @author Mathieu Faverge
  * @author Cedric Castagnede
- * @date 2011-06-01
+ * @date 2018-11-08
  * @precisions normal z -> s d c
  *
  */
@@ -58,23 +58,15 @@ void RUNTIME_zlocality_allrestrict( uint32_t where )
 
     /* QR */
     cl_zgeqrt_restrict_where( where );
-    cl_ztsqrt_restrict_where( where );
+    cl_ztpqrt_restrict_where( where );
     cl_zunmqr_restrict_where( where );
-    cl_ztsmqr_restrict_where( where );
-
-    /* QR-RH */
-/*     cl_zttqrt_restrict_where( where ); */
-/*     cl_zttmqr_restrict_where( where ); */
+    cl_ztpmqrt_restrict_where( where );
 
     /* LQ */
    cl_zgelqt_restrict_where( where );
-   cl_ztslqt_restrict_where( where );
+   cl_ztplqt_restrict_where( where );
    cl_zunmlq_restrict_where( where );
-   cl_ztsmlq_restrict_where( where );
-
-    /* LQ-RH */
-/*     cl_zttlqt_restrict_where( where ); */
-/*     cl_zttmlq_restrict_where( where ); */
+   cl_ztpmlqt_restrict_where( where );
 
 }
 
@@ -112,23 +104,15 @@ void RUNTIME_zlocality_onerestrict( cham_tasktype_t kernel, uint32_t where )
 
     /* QR */
     case TASK_GEQRT:  cl_zgeqrt_restrict_where( where ); break;
+    case TASK_TPQRT:  cl_ztpqrt_restrict_where( where ); break;
     case TASK_UNMQR:  cl_zunmqr_restrict_where( where ); break;
-    case TASK_TSMQR:  cl_ztsmqr_restrict_where( where ); break;
-    case TASK_TSQRT:  cl_ztsqrt_restrict_where( where ); break;
-
-    /* QR-RH */
-/*     case TASK_TTMQR:  cl_zttmqr_restrict_where( where ); break; */
-/*     case TASK_TTQRT:  cl_zttqrt_restrict_where( where ); break; */
+    case TASK_TPMQRT:  cl_ztpmqrt_restrict_where( where ); break;
 
     /* LQ */
    case TASK_GELQT:  cl_zgelqt_restrict_where( where ); break;
+   case TASK_TPLQT:  cl_ztplqt_restrict_where( where ); break;
    case TASK_UNMLQ:  cl_zunmlq_restrict_where( where ); break;
-   case TASK_TSMLQ:  cl_ztsmlq_restrict_where( where ); break;
-   case TASK_TSLQT:  cl_ztslqt_restrict_where( where ); break;
-
-    /* LQ-RH */
-/*     case TASK_TTMLQ:  cl_zttmlq_restrict_where( where ); break; */
-/*     case TASK_TTLQT:  cl_zttlqt_restrict_where( where ); break; */
+   case TASK_TPMLQT:  cl_ztpmlqt_restrict_where( where ); break;
 
     default:
       return;
@@ -167,23 +151,15 @@ void RUNTIME_zlocality_allrestore( )
 
     /* QR */
     cl_zgeqrt_restore_where();
-    cl_ztsqrt_restore_where();
+    cl_ztpqrt_restore_where();
     cl_zunmqr_restore_where();
-    cl_ztsmqr_restore_where();
-
-    /* QR-RH */
-/*     cl_zttqrt_restore_where(); */
-/*     cl_zttmqr_restore_where(); */
+    cl_ztpmqrt_restore_where();
 
     /* LQ */
    cl_zgelqt_restore_where();
-   cl_ztslqt_restore_where();
+   cl_ztplqt_restore_where();
    cl_zunmlq_restore_where();
-   cl_ztsmlq_restore_where();
-
-    /* LQ-RH */
-/*     cl_zttlqt_restore_where(); */
-/*     cl_zttmlq_restore_where(); */
+   cl_ztpmlqt_restore_where();
 
 }
 
@@ -221,23 +197,15 @@ void RUNTIME_zlocality_onerestore( cham_tasktype_t kernel )
 
     /* QR */
     case TASK_GEQRT:  cl_zgeqrt_restore_where(); break;
+    case TASK_TPQRT:  cl_ztpqrt_restore_where(); break;
     case TASK_UNMQR:  cl_zunmqr_restore_where(); break;
-    case TASK_TSMQR:  cl_ztsmqr_restore_where(); break;
-    case TASK_TSQRT:  cl_ztsqrt_restore_where(); break;
-
-    /* QR-RH */
-/*     case TASK_TTMQR:  cl_zttmqr_restore_where(); break; */
-/*     case TASK_TTQRT:  cl_zttqrt_restore_where(); break; */
+    case TASK_TPMQR:  cl_ztpmqrt_restore_where(); break;
 
     /* LQ */
    case TASK_GELQT:  cl_zgelqt_restore_where(); break;
+   case TASK_TPLQT:  cl_ztplqt_restore_where(); break;
    case TASK_UNMLQ:  cl_zunmlq_restore_where(); break;
-   case TASK_TSMLQ:  cl_ztsmlq_restore_where(); break;
-   case TASK_TSLQT:  cl_ztslqt_restore_where(); break;
-
-    /* LQ-RH */
-/*     case TASK_TTMLQ:  cl_zttmlq_restore_where(); break; */
-/*     case TASK_TTLQT:  cl_zttlqt_restore_where(); break; */
+   case TASK_TPMLQR: cl_ztpmlqp_restore_where(); break;
 
     default:
       return;

runtime/starpu/control/runtime_zprofiling.c

@@ -15,7 +15,7 @@
  * @author Cedric Augonnet
  * @author Mathieu Faverge
  * @author Cedric Castagnede
- * @date 2011-06-01
+ * @date 2018-11-08
  * @precisions normal z -> s d c
  *
  */
@@ -43,21 +43,18 @@ void RUNTIME_zdisplay_allprofile()
     profiling_display_zgelqt_info();
     profiling_display_zgeqrt_info();
     profiling_display_zgessm_info();
-    profiling_display_zgetrf_info();
     profiling_display_zgetrf_incpiv_info();