codelets: Add the ipiv_to_perm codelet

include/chameleon/tasks.h

@@ -15,7 +15,8 @@
  * @author Mathieu Faverge
  * @author Cedric Augonnet
  * @author Florent Pruvost
- * @date 2023-07-06
+ * @author Matthieu Kuhn
+ * @date 2023-08-31
  *
  */
 #ifndef _chameleon_tasks_h_
@@ -121,6 +122,10 @@ void INSERT_TASK_hgemm( const RUNTIME_option_t *options,
                                                   const CHAM_desc_t *B, int Bm, int Bn,
                         CHAMELEON_Real16_t beta,  const CHAM_desc_t *C, int Cm, int Cn );
 
+void INSERT_TASK_ipiv_to_perm( const RUNTIME_option_t *options,
+                               int m0, int m, int k,
+                               const CHAM_ipiv_t *ipivdesc, int ipivk );
+
 #include "chameleon/tasks_z.h"
 #include "chameleon/tasks_d.h"
 #include "chameleon/tasks_c.h"

runtime/CMakeLists.txt

@@ -24,7 +24,7 @@
 #  @author Florent Pruvost
 #  @author Philippe Virouleau
 #  @author Matthieu Kuhn
-#  @date 2023-08-22
+#  @date 2023-08-31
 #
 ###
 
@@ -124,6 +124,7 @@ set(CODELETS_ZSRC
 
 set(CODELETS_SRC
   codelets/codelet_map.c
+  codelets/codelet_ipiv_to_perm.c
 )
 
 # Check for the subdirectories

runtime/openmp/codelets/codelet_ipiv_to_perm.c

+/**
+ *
+ * @file openmp/codelet_ipiv_to_perm.c
+ *
+ * @copyright 2012-2023 Bordeaux INP, CNRS (LaBRI UMR 5800), Inria,
+ *                      Univ. Bordeaux. All rights reserved.
+ *
+ ***
+ *
+ * @brief Chameleon OpenMP codelets to convert pivot to permutations
+ *
+ * @version 1.3.0
+ * @author Mathieu Faverge
+ * @author Matthieu Kuhn
+ * @date 2023-08-31
+ *
+ */
+#include "chameleon_openmp.h"
+#include "chameleon/tasks.h"
+#include "coreblas.h"
+
+void INSERT_TASK_ipiv_to_perm( const RUNTIME_option_t *options,
+                               int m0, int m, int k,
+                               const CHAM_ipiv_t *ipivdesc, int ipivk )
+{
+    int *ipiv = NULL; // get pointer from ipivdesc
+    int *perm = NULL; // get pointer from ipivdesc
+    int *invp = NULL; // get pointer from ipivdesc
+
+#pragma omp task firstprivate( m0, m, k ) depend( in:ipiv[0] ) depend( inout:perm[0] ) depend( inout:invp[0] )
+    {
+        CORE_ipiv_to_perm( m0, m, k, ipiv, perm, invp );
+    }
+
+    (void)options;
+    (void)ipivk;
+}

runtime/parsec/codelets/codelet_ipiv_to_perm.c

+/**
+ *
+ * @file parsec/codelet_ipiv_to_perm.c
+ *
+ * @copyright 2023-2023 Bordeaux INP, CNRS (LaBRI UMR 5800), Inria,
+ *                      Univ. Bordeaux. All rights reserved.
+ *
+ ***
+ *
+ * @brief Chameleon Parsec codelets to convert pivot to permutations
+ *
+ * @version 1.3.0
+ * @author Mathieu Faverge
+ * @author Matthieu Kuhn
+ * @date 2023-08-31
+ *
+ */
+#include "chameleon_parsec.h"
+#include "chameleon/tasks.h"
+#include "coreblas.h"
+
+static inline int
+CORE_ipiv_to_perm_parsec( parsec_execution_stream_t *context,
+                          parsec_task_t             *this_task )
+{
+    int m0, m, k;
+    int *ipiv, *perm, *invp;
+
+    parsec_dtd_unpack_args(
+        this_task, &m0, &m, &k, &ipiv, &perm, &invp );
+
+    CORE_ipiv_to_perm( m0, m, k, ipiv, perm, invp );
+}
+
+void INSERT_TASK_ipiv_to_perm( const RUNTIME_option_t *options,
+                               int m0, int m, int k,
+                               const CHAM_ipiv_t *ipivdesc, int ipivk )
+{
+    parsec_taskpool_t* PARSEC_dtd_taskpool = (parsec_taskpool_t *)(options->sequence->schedopt);
+
+    parsec_dtd_taskpool_insert_task(
+        PARSEC_dtd_taskpool, CORE_ipiv_to_perm_parsec, options->priority, "ipiv_to_perm",
+        sizeof(int),         &m0,           VALUE,
+        sizeof(int),         &m,            VALUE,
+        sizeof(int),         &k,            VALUE,
+        PASSED_BY_REF, RUNTIME_ipiv_getaddr( ipivdesc, ipivk ), chameleon_parsec_get_arena_index_ipiv( ipivdesc ) | INPUT,
+        PASSED_BY_REF, RUNTIME_perm_getaddr( ipivdesc, ipivk ), chameleon_parsec_get_arena_index_perm( ipivdesc ) | OUTPUT,
+        PASSED_BY_REF, RUNTIME_invp_getaddr( ipivdesc, ipivk ), chameleon_parsec_get_arena_index_invp( ipivdesc ) | OUTPUT,
+        PARSEC_DTD_ARG_END );
+}

runtime/quark/codelets/codelet_ipiv_to_perm.c

+/**
+ *
+ * @file quark/codelet_ipiv_to_perm.c
+ *
+ * @copyright 2023-2023 Bordeaux INP, CNRS (LaBRI UMR 5800), Inria,
+ *                      Univ. Bordeaux. All rights reserved.
+ *
+ ***
+ *
+ * @brief Chameleon Quark codelets to convert pivot to permutations
+ *
+ * @version 1.3.0
+ * @author Mathieu Faverge
+ * @author Matthieu Kuhn
+ * @date 2023-08-31
+ *
+ */
+#include "chameleon_quark.h"
+#include "chameleon/tasks.h"
+#include "coreblas.h"
+
+static inline void
+CORE_ipiv_to_perm_quark( Quark *quark )
+{
+    int m0, m, k;
+    int *ipiv, *perm, *invp;
+
+    quark_unpack_args_6( quark, m0, m, k, ipiv, perm, invp );
+
+    CORE_ipiv_to_perm( m0, m, k, ipiv, perm, invp );
+}
+
+void INSERT_TASK_ipiv_to_perm( const RUNTIME_option_t *options,
+                               int m0, int m, int k,
+                               const CHAM_ipiv_t *ipivdesc, int ipivk )
+{
+    quark_option_t *opt = (quark_option_t*)(options->schedopt);
+
+    QUARK_Insert_Task(
+        opt->quark, CORE_ipiv_to_perm_quark, (Quark_Task_Flags*)opt,
+        sizeof(int),  &m0,  VALUE,
+        sizeof(int),  &m,   VALUE,
+        sizeof(int),  &k,   VALUE,
+        sizeof(int*), RUNTIME_ipiv_getaddr( ipivdesc, ipivk ), INPUT,
+        sizeof(int*), RUNTIME_perm_getaddr( ipivdesc, ipivk ), OUTPUT,
+        sizeof(int*), RUNTIME_invp_getaddr( ipivdesc, ipivk ), OUTPUT,
+        0 );
+}

runtime/starpu/codelets/codelet_ipiv_to_perm.c

+/**
+ *
+ * @file starpu/codelet_ipiv_to_perm.c
+ *
+ * @copyright 2012-2023 Bordeaux INP, CNRS (LaBRI UMR 5800), Inria,
+ *                      Univ. Bordeaux. All rights reserved.
+ *
+ ***
+ *
+ * @brief Chameleon StarPU codelets to convert pivot to permutations
+ *
+ * @version 1.3.0
+ * @author Mathieu Faverge
+ * @author Matthieu Kuhn
+ * @date 2023-08-31
+ *
+ */
+#include "chameleon_starpu.h"
+#include "runtime_codelets.h"
+
+#if !defined(CHAMELEON_SIMULATION)
+static void cl_ipiv_to_perm_cpu_func( void *descr[], void *cl_arg )
+{
+    int m0, m, k;
+    int *ipiv, *perm, *invp;
+
+    starpu_codelet_unpack_args( cl_arg, &m0, &m, &k );
+
+    ipiv = (int*)STARPU_VECTOR_GET_PTR(descr[0]);
+    perm = (int*)STARPU_VECTOR_GET_PTR(descr[1]);
+    invp = (int*)STARPU_VECTOR_GET_PTR(descr[2]);
+
+    CORE_ipiv_to_perm( m0, m, k, ipiv, perm, invp );
+}
+#endif /* !defined(CHAMELEON_SIMULATION) */
+
+/*
+* Codelet definition
+*/
+static struct starpu_codelet cl_ipiv_to_perm = {
+    .where        = STARPU_CPU,
+#if defined(CHAMELEON_SIMULATION)
+    .cpu_funcs[0] = (starpu_cpu_func_t)1,
+#else
+    .cpu_funcs[0] = cl_ipiv_to_perm_cpu_func,
+#endif
+    .nbuffers     = 3,
+    .model        = NULL,
+    .name         = "ipiv_to_perm"
+};
+
+void INSERT_TASK_ipiv_to_perm( const RUNTIME_option_t *options,
+                               int m0, int m, int k,
+                               const CHAM_ipiv_t *ipivdesc, int ipivk )
+{
+    struct starpu_codelet *codelet = &cl_ipiv_to_perm;
+
+    rt_starpu_insert_task(
+        codelet,
+        STARPU_VALUE,             &m0,  sizeof(int),
+        STARPU_VALUE,             &m,   sizeof(int),
+        STARPU_VALUE,             &k,   sizeof(int),
+        STARPU_R,                 RUNTIME_ipiv_getaddr( ipivdesc, ipivk ),
+        STARPU_W,                 RUNTIME_perm_getaddr( ipivdesc, ipivk ),
+        STARPU_W,                 RUNTIME_invp_getaddr( ipivdesc, ipivk ),
+        STARPU_PRIORITY,          options->priority,
+        STARPU_EXECUTE_ON_WORKER, options->workerid,
+        0 );
+}

runtime/starpu/include/runtime_codelets.h

@@ -27,6 +27,8 @@
 #include "runtime_codelet_profile.h"
 
 #if !defined(CHAMELEON_SIMULATION)
+#include "coreblas.h"
+
 #if defined(CHAMELEON_USE_CUDA)
 #include "gpucublas.h"
 #endif