diff --git a/cmake_modules/local_subs.py b/cmake_modules/local_subs.py
index 227afc04908a49920f6917668f9fa5b7b0d7fd85..02af5df38a90e034eebb9438238440267a1a3355 100644
--- a/cmake_modules/local_subs.py
+++ b/cmake_modules/local_subs.py
@@ -43,6 +43,7 @@ _extra_blas = [
('', 'scesca', 'dcesca', 'ccesca', 'zcesca' ),
('', 'sgesum', 'dgesum', 'cgesum', 'zgesum' ),
('', 'sgersum', 'dgersum', 'cgersum', 'zgersum' ),
+ ('', 'sprint', 'dprint', 'cprint', 'zprint' ),
]
_extra_BLAS = [ [ x.upper() for x in row ] for row in _extra_blas ]
diff --git a/compute/CMakeLists.txt b/compute/CMakeLists.txt
index 0106e546cf95b7b0875761e6de3b4209bcf91f4a..bed4df6f140c9449d7c7201bf4249373c14d6146 100644
--- a/compute/CMakeLists.txt
+++ b/compute/CMakeLists.txt
@@ -253,6 +253,7 @@ set(ZSRC
##################
zgenm2.c
pzgenm2.c
+ zprint.c
)
precisions_rules_py(CHAMELEON_SRCS_GENERATED "${ZSRC}"
diff --git a/compute/zprint.c b/compute/zprint.c
new file mode 100644
index 0000000000000000000000000000000000000000..3dbd715e2c9120918817caa944836222ce344292
--- /dev/null
+++ b/compute/zprint.c
@@ -0,0 +1,227 @@
+/**
+ *
+ * @file zprint.c
+ *
+ * @copyright 2012-2023 Bordeaux INP, CNRS (LaBRI UMR 5800), Inria,
+ * Univ. Bordeaux. All rights reserved.
+ *
+ ***
+ *
+ * @brief Chameleon zprint wrappers
+ *
+ * @version 1.2.0
+ * @author Mathieu Faverge
+ * @author Matthieu Kuhn
+ * @date 2022-02-22
+ * @precisions normal z -> s d c
+ *
+ */
+#include "control/common.h"
+#if !defined(CHAMELEON_SIMULATION)
+#include <coreblas/coreblas_z.h>
+#endif
+
+/*
+ * Static variable to know how to handle the data within the kernel
+ * This assumes that only one runtime is enabled at a time.
+ */
+static RUNTIME_id_t zprint_runtime_id = RUNTIME_SCHED_STARPU;
+
+struct zprint_args_s {
+ FILE *file;
+ const char *header;
+};
+
+static inline int
+zprint( const CHAM_desc_t *descA,
+ cham_uplo_t uplo, int m, int n,
+ CHAM_tile_t *tileA, void *op_args )
+{
+ CHAMELEON_Complex64_t *A;
+ struct zprint_args_s *options = (struct zprint_args_s *)op_args;
+
+ int tempmm = m == descA->mt-1 ? descA->m-m*descA->mb : descA->mb;
+ int tempnn = n == descA->nt-1 ? descA->n-n*descA->nb : descA->nb;
+ int minmn = chameleon_min( tempmm, tempnn );
+ int lda, i;
+
+ if ( zprint_runtime_id == RUNTIME_SCHED_PARSEC ) {
+ A = (CHAMELEON_Complex64_t*)tileA;
+ lda = descA->get_blkldd( descA, m );
+ }
+ else {
+ A = CHAM_tile_get_ptr( tileA );
+ lda = tileA->ld;
+ }
+
+#if !defined(CHAMELEON_SIMULATION)
+ CORE_zprint( options->file, options->header, uplo,
+ tempmm, tempnn, m, n, A, lda );
+#endif
+
+ return 0;
+}
+
+/**
+ ********************************************************************************
+ *
+ * @ingroup CHAMELEON_Complex64_t
+ *
+ * @brief Print a matrix tile by tile.
+ *
+ *******************************************************************************
+ *
+ * @param[in] M
+ * The number of rows of the matrix A. M >= 0.
+ *
+ * @param[in] N
+ * The number of columns of the matrix A. N >= 0.
+ *
+ * @param[in] A
+ * The M-by-N matrix A.
+ *
+ * @param[in] LDA
+ * The leading dimension of the array A. LDA >= max(1,M).
+ *
+ *******************************************************************************
+ *
+ * @retval the two-norm estimate.
+ *
+ *******************************************************************************
+ *
+ * @sa CHAMELEON_zlatms_Tile
+ * @sa CHAMELEON_zlatms_Tile_Async
+ * @sa CHAMELEON_clange
+ * @sa CHAMELEON_dlange
+ * @sa CHAMELEON_slange
+ *
+ */
+int CHAMELEON_zprint( FILE *file, const char *header,
+ cham_uplo_t uplo, int M, int N,
+ CHAMELEON_Complex64_t *A, int LDA )
+{
+ int NB;
+ int status;
+ CHAM_context_t *chamctxt;
+ RUNTIME_sequence_t *sequence = NULL;
+ RUNTIME_request_t request = RUNTIME_REQUEST_INITIALIZER;
+ CHAM_desc_t descAl, descAt;
+ struct zprint_args_s options = {
+ .file = file,
+ .header = header,
+ };
+
+ chamctxt = chameleon_context_self();
+ if (chamctxt == NULL) {
+ chameleon_fatal_error("CHAMELEON_zlatms", "CHAMELEON not initialized");
+ return CHAMELEON_ERR_NOT_INITIALIZED;
+ }
+ /* Check input arguments */
+ if (M < 0) {
+ chameleon_error("CHAMELEON_zlatms", "illegal value of M");
+ return -1;
+ }
+ if (N < 0) {
+ chameleon_error("CHAMELEON_zlatms", "illegal value of N");
+ return -2;
+ }
+ if (LDA < chameleon_max(1, M)) {
+ chameleon_error("CHAMELEON_zlatms", "illegal value of LDA");
+ return -4;
+ }
+
+ /* Quick return */
+ if (chameleon_min(N, M) == 0) {
+ return (double)0.0;
+ }
+
+ /* Tune NB depending on M, N & NRHS; Set NBNB */
+ status = chameleon_tune(CHAMELEON_FUNC_ZGEMM, M, N, 0);
+ if (status != CHAMELEON_SUCCESS) {
+ chameleon_error("CHAMELEON_zlatms", "chameleon_tune() failed");
+ return status;
+ }
+
+ /* Set NT */
+ NB = CHAMELEON_NB;
+
+ chameleon_sequence_create( chamctxt, &sequence );
+
+ /* Submit the matrix conversion */
+ chameleon_zlap2tile( chamctxt, &descAl, &descAt, ChamDescInput, uplo,
+ A, NB, NB, LDA, N, M, N, sequence, &request );
+
+ /* Call the tile interface */
+ zprint_runtime_id = chamctxt->scheduler;
+ chameleon_pmap( uplo, &descAt, zprint, &options, sequence, &request );
+
+ /* Submit the matrix conversion back */
+ chameleon_ztile2lap( chamctxt, &descAl, &descAt,
+ ChamDescInput, uplo, sequence, &request );
+
+ chameleon_sequence_wait( chamctxt, sequence );
+
+ /* Cleanup the temporary data */
+ chameleon_ztile2lap_cleanup( chamctxt, &descAl, &descAt );
+
+ chameleon_sequence_destroy( chamctxt, sequence );
+ return status;
+}
+
+/**
+ ********************************************************************************
+ *
+ * @ingroup CHAMELEON_Complex64_t_Tile
+ *
+ * @brief Tile equivalent of CHAMELEON_zprint().
+ *
+ * Operates on matrices stored by tiles.
+ * All matrices are passed through descriptors.
+ * All dimensions are taken from the descriptors.
+ *
+ *******************************************************************************
+ *
+ * @param[in] A
+ * On entry, the input matrix A.
+ *
+ *******************************************************************************
+ *
+ * @retval CHAMELEON_SUCCESS successful exit
+ *
+ *******************************************************************************
+ *
+ * @sa CHAMELEON_zlatms
+ * @sa CHAMELEON_zlatms_Tile_Async
+ * @sa CHAMELEON_clange_Tile
+ * @sa CHAMELEON_dlange_Tile
+ * @sa CHAMELEON_slange_Tile
+ *
+ */
+int CHAMELEON_zprint_Tile( FILE *file, const char *header,
+ cham_uplo_t uplo, CHAM_desc_t *A )
+{
+ CHAM_context_t *chamctxt;
+ RUNTIME_sequence_t *sequence = NULL;
+ RUNTIME_request_t request = RUNTIME_REQUEST_INITIALIZER;
+ struct zprint_args_s options = {
+ .file = file,
+ .header = header,
+ };
+ int status;
+
+ chamctxt = chameleon_context_self();
+ if (chamctxt == NULL) {
+ chameleon_fatal_error("CHAMELEON_zlatms_Tile", "CHAMELEON not initialized");
+ return CHAMELEON_ERR_NOT_INITIALIZED;
+ }
+ chameleon_sequence_create( chamctxt, &sequence );
+
+ zprint_runtime_id = chamctxt->scheduler;
+ chameleon_pmap( uplo, A, zprint, &options, sequence, &request );
+ CHAMELEON_Desc_Flush( A, sequence );
+
+ chameleon_sequence_wait( chamctxt, sequence );
+ status = sequence->status;
+ chameleon_sequence_destroy( chamctxt, sequence );
+ return status;
+}
diff --git a/coreblas/compute/CMakeLists.txt b/coreblas/compute/CMakeLists.txt
index d880039fe2d83e112d882414c24bec1a1e24ec69..40baddc598df9da3ce04b58a54b7fe645435b564 100644
--- a/coreblas/compute/CMakeLists.txt
+++ b/coreblas/compute/CMakeLists.txt
@@ -107,6 +107,7 @@ set(ZSRC
core_zttqrt.c
core_zunmlq.c
core_zunmqr.c
+ core_zprint.c
)
if( CHAMELEON_USE_HMAT )
list( APPEND ZSRC
diff --git a/coreblas/compute/core_zprint.c b/coreblas/compute/core_zprint.c
new file mode 100644
index 0000000000000000000000000000000000000000..0e4fdc887f0c47ca85d4900e1ff52341e9f8bc83
--- /dev/null
+++ b/coreblas/compute/core_zprint.c
@@ -0,0 +1,46 @@
+/**
+ *
+ * @file core_zprint.c
+ *
+ * @copyright 2012-2023 Bordeaux INP, CNRS (LaBRI UMR 5800), Inria,
+ * Univ. Bordeaux. All rights reserved.
+ *
+ ***
+ *
+ * @brief Chameleon core_zprint CPU kernel
+ *
+ * @version 1.0.0
+ * @author Mathieu Faverge
+ * @date 2020-03-03
+ * @precisions normal z -> c d s
+ *
+ */
+#include "coreblas/lapacke.h"
+#include "coreblas.h"
+
+void
+CORE_zprint( FILE *file, const char *header,
+ cham_uplo_t uplo, int M, int N, int Am, int An,
+ const CHAMELEON_Complex64_t *A, int lda )
+{
+ FILE *output = (file == NULL) ? stdout : file;
+ int i, j;
+
+ fflush( output );
+ fprintf( output, "--- %10s (%2d, %2d) / %p, %d:\n", header, Am, An, A, lda );
+ for(i=0; i<M; i++) {
+ fprintf( output, " " );
+ for(j=0; j<N; j++) {
+#if defined(PRECISION_z) || defined(PRECISION_c)
+ fprintf( output, " (% e, % e)",
+ creal( A[j*lda + i] ),
+ cimag( A[j*lda + i] ));
+#else
+ fprintf( output, " % e", A[j*lda + i] );
+#endif
+ }
+ fprintf( output, "\n" );
+ }
+ fprintf( output, "-------------------------\n" );
+ fflush( output );
+}
diff --git a/coreblas/compute/core_ztile.c b/coreblas/compute/core_ztile.c
index 986d587f80bf16abe4a1e6479aeeb88e09a4f686..6de9b2c63cd546ddb8216fa0f5c0bdfd522fdb1d 100644
--- a/coreblas/compute/core_ztile.c
+++ b/coreblas/compute/core_ztile.c
@@ -1072,3 +1072,13 @@ TCORE_zgram( cham_uplo_t uplo,
return CORE_zgram(
uplo, M, N, Mt, Nt, CHAM_tile_get_ptr( Di ), Di->ld, CHAM_tile_get_ptr( Dj ), Dj->ld, CHAM_tile_get_ptr( D ), CHAM_tile_get_ptr( A ), A->ld );
}
+
+void
+TCORE_zprint( FILE *file, const char *header,
+ cham_uplo_t uplo, int M, int N,
+ int Am, int An, const CHAM_tile_t *A )
+{
+ coreblas_kernel_trace( A );
+ assert( A->format & CHAMELEON_TILE_FULLRANK );
+ CORE_zprint( file, header, uplo, M, N, Am, An, CHAM_tile_get_ptr( A ), A->ld );
+}
diff --git a/coreblas/include/coreblas/coreblas_z.h b/coreblas/include/coreblas/coreblas_z.h
index b0bcf3bdbbf7dfc8e465aa3f77a29451c66fe03a..050a7968769a5c37a138faee18ef0a577ee6be19 100644
--- a/coreblas/include/coreblas/coreblas_z.h
+++ b/coreblas/include/coreblas/coreblas_z.h
@@ -393,4 +393,8 @@ int CORE_zgram( cham_uplo_t uplo, int M, int N, int Mt, int Nt,
const double *D,
double *A, int LDA );
+void CORE_zprint( FILE *file, const char *header,
+ cham_uplo_t uplo, int m, int n, int Am, int An,
+ const CHAMELEON_Complex64_t *A, int lda );
+
#endif /* _coreblas_z_h_ */
diff --git a/coreblas/include/coreblas/coreblas_ztile.h b/coreblas/include/coreblas/coreblas_ztile.h
index aa66fdd751657288475070c6dec606ff08fd714a..74cd413168d44653f27c408eee7dea77eca603ad 100644
--- a/coreblas/include/coreblas/coreblas_ztile.h
+++ b/coreblas/include/coreblas/coreblas_ztile.h
@@ -86,5 +86,6 @@ int TCORE_zunmqr( cham_side_t side, cham_trans_t trans, int M, int N, int K, in
int TCORE_zgesum( cham_store_t storev, int M, int N, const CHAM_tile_t *A, CHAM_tile_t *sum );
int TCORE_zcesca( int center, int scale, cham_store_t axis, int M, int N, int Mt, int Nt, const CHAM_tile_t *Gi, const CHAM_tile_t *Gj, const CHAM_tile_t *G, const CHAM_tile_t *Di, const CHAM_tile_t *Dj, CHAM_tile_t *A );
int TCORE_zgram( cham_uplo_t uplo, int M, int N, int Mt, int Nt, const CHAM_tile_t *Di, const CHAM_tile_t *Dj, const CHAM_tile_t *D, CHAM_tile_t *A );
+void TCORE_zprint( FILE *file, const char *header, cham_uplo_t uplo, int M, int N, int Am, int An, const CHAM_tile_t *A );
#endif /* _coreblas_ztile_h_ */
diff --git a/include/chameleon.h b/include/chameleon.h
index 4f75ba8306c7da129df957100edf313cb121987c..c0a9ea9a9f6896be35a78ffab0b2e264c2bf6772 100644
--- a/include/chameleon.h
+++ b/include/chameleon.h
@@ -27,6 +27,7 @@
* CHAMELEON types and constants
*/
#include "chameleon/config.h"
+#include <stdio.h>
#include "chameleon/constants.h"
#include "chameleon/types.h"
#include "chameleon/struct.h"
@@ -37,7 +38,6 @@
*/
#include "chameleon/runtime.h"
-
/* ****************************************************************************
* CHAMELEON Simulation mode
*/
diff --git a/include/chameleon/chameleon_z.h b/include/chameleon/chameleon_z.h
index c36e6289dcfc7dd19122cd671f25a66717c586f2..e587d015c4239a90fb10c228d9bb41faaca7dda1 100644
--- a/include/chameleon/chameleon_z.h
+++ b/include/chameleon/chameleon_z.h
@@ -382,6 +382,9 @@ int CHAMELEON_zgram( cham_uplo_t uplo, int N, CHAMELEON_Complex64_t *A, int LDA
int CHAMELEON_zgram_Tile( cham_uplo_t uplo, CHAM_desc_t *A );
int CHAMELEON_zgram_Tile_Async( cham_uplo_t uplo, CHAM_desc_t *A, void *user_ws, RUNTIME_sequence_t *sequence, RUNTIME_request_t *request );
+int CHAMELEON_zprint( FILE *file, const char *header, cham_uplo_t uplo, int M, int N, CHAMELEON_Complex64_t *A, int LDA );
+int CHAMELEON_zprint_Tile( FILE *file, const char *header, cham_uplo_t uplo, CHAM_desc_t *descA );
+
END_C_DECLS
#endif /* _chameleon_z_h_ */