From 224bee6ed61a1cbd426eff5344978371903ba35b Mon Sep 17 00:00:00 2001
From: Raphael Boucherie <rboucher@zimbra.inria.fr>
Date: Wed, 15 Mar 2017 13:08:22 +0100
Subject: [PATCH] Made libhqr independent, all test work
---
include/{dplasma_qr_param.h => libhqr.h} | 130 ++++---
src/{dplasma_hqr.c => libhqr.c} | 343 +++++++++---------
src/{dplasma_hqr_dbg.c => libhqr_dbg.c} | 94 +++--
...plasma_systolic_qr.c => libhqr_systolic.c} | 92 ++---
testings/testing_pivgen.c | 333 +++++++----------
5 files changed, 477 insertions(+), 515 deletions(-)
rename include/{dplasma_qr_param.h => libhqr.h} (54%)
rename src/{dplasma_hqr.c => libhqr.c} (89%)
rename src/{dplasma_hqr_dbg.c => libhqr_dbg.c} (86%)
rename src/{dplasma_systolic_qr.c => libhqr_systolic.c} (79%)
diff --git a/include/dplasma_qr_param.h b/include/libhqr.h
similarity index 54%
rename from include/dplasma_qr_param.h
rename to include/libhqr.h
index b487899..49b3f42 100644
--- a/include/dplasma_qr_param.h
+++ b/include/libhqr.h
@@ -6,34 +6,70 @@
* @precisions normal z -> z c d s
*
*/
-#ifndef _DPLASMA_QR_PARAM_H_
-#define _DPLASMA_QR_PARAM_H_
+#ifndef _LIBHQR_H_
+#define _LIBHQR_H_
+
+#undef BEGIN_C_DECLS
+#undef END_C_DECLS
+#if defined(c_plusplus) || defined(__cplusplus)
+# define BEGIN_C_DECLS extern "C" {
+# define END_C_DECLS }
+#else
+#define BEGIN_C_DECLS /* empty */
+#define END_C_DECLS /* empty */
+#endif
BEGIN_C_DECLS
+static inline int libhqr_imin(int a, int b){
+ return (a > b) ? b : a;
+}
+
+static inline int libhqr_imax(int a, int b){
+ return (a > b) ? a : b;
+}
+
+static inline int libhqr_iceil(int a, int b){
+ return (a + b - 1) / b;
+}
+
/*
- * DPLASMA_QR_KILLED_BY_TS needs to be set to 0 for all variant of QR
+ * LIBHQR_KILLED_BY_TS needs to be set to 0 for all variant of QR
* factorization to distinguish TT kernels from TS kernels in jdf
*/
-typedef enum dplasma_qr_type_ {
- DPLASMA_QR_KILLED_BY_TS = 0,
- DPLASMA_QR_KILLED_BY_LOCALTREE = 1,
- DPLASMA_QR_KILLED_BY_DOMINO = 2,
- DPLASMA_QR_KILLED_BY_DISTTREE = 3,
-} dplasma_qr_type_e;
-
-typedef enum dplasma_qr_tree_ {
- DPLASMA_FLAT_TREE = 0,
- DPLASMA_GREEDY_TREE = 1,
- DPLASMA_FIBONACCI_TREE = 2,
- DPLASMA_BINARY_TREE = 3,
- DPLASMA_GREEDY1P_TREE = 4,
-} dplasma_qr_tree_e;
-
-struct dplasma_qrtree_s;
-typedef struct dplasma_qrtree_s dplasma_qrtree_t;
-
-struct dplasma_qrtree_s {
+typedef enum libhqr_type_ {
+ LIBHQR_KILLED_BY_TS = 0,
+ LIBHQR_KILLED_BY_LOCALTREE = 1,
+ LIBHQR_KILLED_BY_DOMINO = 2,
+ LIBHQR_KILLED_BY_DISTTREE = 3,
+} libhqr_type_e;
+
+typedef enum libhqr_tree_ {
+ LIBHQR_FLAT_TREE = 0,
+ LIBHQR_GREEDY_TREE = 1,
+ LIBHQR_FIBONACCI_TREE = 2,
+ LIBHQR_BINARY_TREE = 3,
+ LIBHQR_GREEDY1P_TREE = 4,
+} libqr_tree_e;
+
+typedef enum libhqr_typefacto_ {
+ LIBHQR_QR = 0,
+ LIBHQR_LQ = 1,
+} libhqr_typefacto_e;
+
+typedef struct libhqr_tiledesc_s{
+ int mt;
+ int nt;
+ int nodes;
+ int p;
+} libhqr_tiledesc_t;
+
+struct libhqr_tree_s;
+typedef struct libhqr_tree_s libhqr_tree_t;
+
+typedef struct libhqr_context_s libhqr_context_t;
+
+struct libhqr_tree_s {
/**
* getnbgeqrf
* @param[in] arg arguments specific to the reduction tree used
@@ -41,7 +77,7 @@ struct dplasma_qrtree_s {
*
* @return The number of geqrt applied to the panel k
*/
- int (*getnbgeqrf)( const dplasma_qrtree_t *arg, int k );
+ int (*getnbgeqrf)( const libhqr_tree_t *arg, int k );
/**
* getm:
@@ -51,7 +87,7 @@ struct dplasma_qrtree_s {
*
* @return The row index of the i-th geqrt applied on the panel k
*/
- int (*getm)( const dplasma_qrtree_t *arg, int k, int i );
+ int (*getm)( const libhqr_tree_t *arg, int k, int i );
/**
* geti:
@@ -61,7 +97,7 @@ struct dplasma_qrtree_s {
*
* @returns the index in the list of geqrt applied to panel k
*/
- int (*geti)( const dplasma_qrtree_t *qrtree, int k, int m );
+ int (*geti)( const libhqr_tree_t *qrtree, int k, int m );
/**
* gettype:
* @param[in] arg arguments specific to the reduction tree used
@@ -72,7 +108,7 @@ struct dplasma_qrtree_s {
* - 0 if it is a TS kernel
* - >0 otherwise. (TT kernel)
*/
- int (*gettype)( const dplasma_qrtree_t *qrtree, int k, int m );
+ int (*gettype)( const libhqr_tree_t *qrtree, int k, int m );
/**
* currpiv
* @param[in] arg arguments specific to the reduction tree used
@@ -81,7 +117,7 @@ struct dplasma_qrtree_s {
*
* @return The index of the row annihilating the row m at step k
*/
- int (*currpiv)( const dplasma_qrtree_t *qrtree, int k, int m );
+ int (*currpiv)( const libhqr_tree_t *qrtree, int k, int m );
/**
* nextpiv
* @param[in] arg arguments specific to the reduction tree used
@@ -93,7 +129,7 @@ struct dplasma_qrtree_s {
* @return the next line that the row p will kill during step k
* desc->mt if p will never be used again as an annihilator.
*/
- int (*nextpiv)(const dplasma_qrtree_t *qrtree, int k, int p, int m);
+ int (*nextpiv)(const libhqr_tree_t *qrtree, int k, int p, int m);
/**
* prevpiv
* @param[in] arg arguments specific to the reduction tree used
@@ -105,7 +141,7 @@ struct dplasma_qrtree_s {
* @return the previous line killed by the row p during step k
* desc->mt if p has never been used before as an annihilator.
*/
- int (*prevpiv)(const dplasma_qrtree_t *qrtree, int k, int p, int m);
+ int (*prevpiv)(const libhqr_tree_t *qrtree, int k, int p, int m);
/** Descriptor infos associated to the factorization */
int mt;
@@ -117,34 +153,34 @@ struct dplasma_qrtree_s {
void *args;
};
-int dplasma_systolic_init( dplasma_qrtree_t *qrtree,
- PLASMA_enum trans, tiled_matrix_desc_t *A,
+int libhqr_systolic_init( libhqr_tree_t *qrtree,
+ libhqr_typefacto_e trans, libhqr_tiledesc_t *A,
int p, int q );
-void dplasma_systolic_finalize( dplasma_qrtree_t *qrtree );
+void libhqr_systolic_finalize( libhqr_tree_t *qrtree );
-int dplasma_svd_init( dplasma_qrtree_t *qrtree,
- PLASMA_enum trans, tiled_matrix_desc_t *A,
+int libhqr_svd_init( libhqr_tree_t *qrtree,
+ libhqr_typefacto_e trans, libhqr_tiledesc_t *A,
int type_hlvl, int p, int nbcores_per_node, int ratio );
-int dplasma_hqr_init( dplasma_qrtree_t *qrtree,
- PLASMA_enum trans, tiled_matrix_desc_t *A,
+int libhqr_hqr_init( libhqr_tree_t *qrtree,
+ libhqr_typefacto_e trans, libhqr_tiledesc_t *A,
int type_llvl, int type_hlvl,
int a, int p, int domino, int tsrr );
-void dplasma_hqr_finalize( dplasma_qrtree_t *qrtree );
+void libhqr_hqr_finalize( libhqr_tree_t *qrtree );
/*
* Debugging functions
*/
-int dplasma_qrtree_check ( tiled_matrix_desc_t *A, dplasma_qrtree_t *qrtree );
-void dplasma_qrtree_print_dag ( tiled_matrix_desc_t *A, dplasma_qrtree_t *qrtree, char *filename );
-void dplasma_qrtree_print_type ( tiled_matrix_desc_t *A, dplasma_qrtree_t *qrtree );
-void dplasma_qrtree_print_pivot ( tiled_matrix_desc_t *A, dplasma_qrtree_t *qrtree );
-void dplasma_qrtree_print_nbgeqrt( tiled_matrix_desc_t *A, dplasma_qrtree_t *qrtree );
-void dplasma_qrtree_print_perm ( tiled_matrix_desc_t *A, dplasma_qrtree_t *qrtree, int *perm );
-void dplasma_qrtree_print_next_k ( tiled_matrix_desc_t *A, dplasma_qrtree_t *qrtree, int k );
-void dplasma_qrtree_print_prev_k ( tiled_matrix_desc_t *A, dplasma_qrtree_t *qrtree, int k );
-void dplasma_qrtree_print_geqrt_k( tiled_matrix_desc_t *A, dplasma_qrtree_t *qrtree, int k );
+int libhqr_tree_check ( libhqr_tiledesc_t *A, libhqr_tree_t *qrtree );
+void libhqr_tree_print_dag ( libhqr_tiledesc_t *A, libhqr_tree_t *qrtree, char *filename );
+void libhqr_tree_print_type ( libhqr_tiledesc_t *A, libhqr_tree_t *qrtree );
+void libhqr_tree_print_pivot ( libhqr_tiledesc_t *A, libhqr_tree_t *qrtree );
+void libhqr_tree_print_nbgeqrt( libhqr_tiledesc_t *A, libhqr_tree_t *qrtree );
+void libhqr_tree_print_perm ( libhqr_tiledesc_t *A, libhqr_tree_t *qrtree, int *perm );
+void libhqr_tree_print_next_k ( libhqr_tiledesc_t *A, libhqr_tree_t *qrtree, int k );
+void libhqr_tree_print_prev_k ( libhqr_tiledesc_t *A, libhqr_tree_t *qrtree, int k );
+void libhqr_tree_print_geqrt_k( libhqr_tiledesc_t *A, libhqr_tree_t *qrtree, int k );
END_C_DECLS
-#endif /* _DPLASMA_QR_PARAM_H_ */
+#endif /* _LIBHQR_H_ */
diff --git a/src/dplasma_hqr.c b/src/libhqr.c
similarity index 89%
rename from src/dplasma_hqr.c
rename to src/libhqr.c
index f91d4d2..413a4ca 100644
--- a/src/dplasma_hqr.c
+++ b/src/libhqr.c
@@ -70,14 +70,14 @@
* high level tree to reduce communications.
* These lines are defined by (i-k)/p = 0.
*/
-#include "parsec.h"
-#include "dplasma.h"
-#include "dplasma_qr_param.h"
-
+#include "libhqr.h"
+#include <assert.h>
+#include <stdlib.h>
+#include <stdio.h>
#include <math.h>
-#if defined(PARSEC_HAVE_STRING_H)
+//#if defined(LIBHQR_HAVE_STRING_H)
#include <string.h>
-#endif /* defined(PARSEC_HAVE_STRING_H) */
+//#endif /* defined(PARSEC_HAVE_STRING_H) */
#define PRINT_PIVGEN 0
#ifdef PRINT_PIVGEN
@@ -145,14 +145,14 @@ struct hqr_subpiv_s {
/*
* Common functions
*/
-static int hqr_getnbgeqrf( const dplasma_qrtree_t *qrtree, int k );
-static int hqr_getm( const dplasma_qrtree_t *qrtree, int k, int i );
-static int hqr_geti( const dplasma_qrtree_t *qrtree, int k, int m );
-static int hqr_gettype( const dplasma_qrtree_t *qrtree, int k, int m );
+static int hqr_getnbgeqrf( const libhqr_tree_t *qrtree, int k );
+static int hqr_getm( const libhqr_tree_t *qrtree, int k, int i );
+static int hqr_geti( const libhqr_tree_t *qrtree, int k, int m );
+static int hqr_gettype( const libhqr_tree_t *qrtree, int k, int m );
/* Permutation */
-static void hqr_genperm ( dplasma_qrtree_t *qrtree );
-static int hqr_getinvperm( const dplasma_qrtree_t *qrtree, int k, int m );
+static void hqr_genperm ( libhqr_tree_t *qrtree );
+static int hqr_getinvperm( const libhqr_tree_t *qrtree, int k, int m );
/*
* Subtree for low-level
@@ -181,7 +181,7 @@ static void hqr_low_fibonacci_init(hqr_subpiv_t *arg, int minMN);
* Return:
* The number of geqrt to execute in the panel k
*/
-static int hqr_getnbgeqrf( const dplasma_qrtree_t *qrtree, int k ) {
+static int hqr_getnbgeqrf( const libhqr_tree_t *qrtree, int k ) {
hqr_args_t *arg = (hqr_args_t*)(qrtree->args);
int a = qrtree->a;
int p = qrtree->p;
@@ -216,9 +216,9 @@ static int hqr_getnbgeqrf( const dplasma_qrtree_t *qrtree, int k ) {
nb_1 = (nb_12 - nb_11) / a;
/* Add leftover */
- nb_1 += dplasma_imin( p, gmt - nb_12 );
+ nb_1 += libhqr_imin( p, gmt - nb_12 );
- return dplasma_imin( nb_1 + nb_2 + nb_3, gmt - k);
+ return libhqr_imin( nb_1 + nb_2 + nb_3, gmt - k);
}
/*
@@ -227,7 +227,7 @@ static int hqr_getnbgeqrf( const dplasma_qrtree_t *qrtree, int k ) {
* Return:
* The global indice m of the i th geqrt in the panel k
*/
-static int hqr_getm( const dplasma_qrtree_t *qrtree, int k, int i )
+static int hqr_getm( const libhqr_tree_t *qrtree, int k, int i )
{
hqr_args_t *arg = (hqr_args_t*)(qrtree->args);
int a = qrtree->a;
@@ -261,7 +261,7 @@ static int hqr_getm( const dplasma_qrtree_t *qrtree, int k, int i )
* Return:
* The index i of the geqrt in the panel k
*/
-static int hqr_geti( const dplasma_qrtree_t *qrtree, int k, int m )
+static int hqr_geti( const libhqr_tree_t *qrtree, int k, int m )
{
hqr_args_t *arg = (hqr_args_t*)(qrtree->args);
int a = qrtree->a;
@@ -298,7 +298,7 @@ static int hqr_geti( const dplasma_qrtree_t *qrtree, int k, int m )
* 2 - if m is reduced thanks to the bubble tree
* 3 - if m is reduced in distributed
*/
-static int hqr_gettype( const dplasma_qrtree_t *qrtree, int k, int m ) {
+static int hqr_gettype( const libhqr_tree_t *qrtree, int k, int m ) {
hqr_args_t *arg = (hqr_args_t*)(qrtree->args);
int a = qrtree->a;
int p = qrtree->p;
@@ -602,7 +602,7 @@ static void hqr_low_greedy_init(hqr_subpiv_t *arg, int minMN){
int j, k, height, start, end, firstk = 0;
int *nT, *nZ;
- arg->minMN = dplasma_imin( minMN, mt*a );
+ arg->minMN = libhqr_imin( minMN, mt*a );
minMN = arg->minMN;
arg->ipiv = (int*)malloc( mt * minMN * sizeof(int) );
@@ -673,7 +673,7 @@ static void hqr_low_greedy_init(hqr_subpiv_t *arg, int minMN){
nT[0] = mt;
for(k=0; k<lminMN; k++) {
- nT2DO[k] = dplasma_imax( mt - ((k + p - 1 - myrank) / pa), 0 );
+ nT2DO[k] = libhqr_imax( mt - ((k + p - 1 - myrank) / pa), 0 );
if ( nT2DO[k] == 0 ) {
lminMN = k;
break;
@@ -806,7 +806,7 @@ static void hqr_low_greedy1p_init(hqr_subpiv_t *arg, int minMN){
/* This section has not been coded yet, and will perform a classic greedy */
if ( domino )
{
- arg->minMN = dplasma_imin( minMN, mt*a );
+ arg->minMN = libhqr_imin( minMN, mt*a );
minMN = arg->minMN;
arg->ipiv = (int*)malloc( mt * minMN * sizeof(int) );
@@ -818,7 +818,7 @@ static void hqr_low_greedy1p_init(hqr_subpiv_t *arg, int minMN){
*/
for(k=0; k<minMN; k++) {
/* Number of tiles to factorized in this column on this rank */
- nT = dplasma_imax( mt - (k / a), 0 );
+ nT = libhqr_imax( mt - (k / a), 0 );
/* Number of tiles already killed */
nZ = 0;
@@ -852,7 +852,7 @@ static void hqr_low_greedy1p_init(hqr_subpiv_t *arg, int minMN){
*/
for(k=0; k<minMN; k++) {
/* Number of tiles to factorized in this column on this rank */
- nT = dplasma_imax( mt - ((k + p - 1 - myrank) / pa), 0 );
+ nT = libhqr_imax( mt - ((k + p - 1 - myrank) / pa), 0 );
/* Number of tiles already killed */
nZ = 0;
@@ -934,7 +934,7 @@ static int hqr_high_flat_prevpiv(const hqr_subpiv_t *arg, int k, int p, int star
assert( arg->p > 1 );
if ( p == k && arg->ldd > 1 ) {
if ( start == p && p != arg->ldd-1 )
- return dplasma_imin( p + arg->p - 1, arg->ldd - 1 );
+ return libhqr_imin( p + arg->p - 1, arg->ldd - 1 );
else if ( start > p + 1 && (start-k < arg->p))
return start-1;
}
@@ -1000,7 +1000,7 @@ static int hqr_high_binary_prevpiv(const hqr_subpiv_t *arg, int k, int p, int st
if ( (start == p) && ( offset%2 == 0 ) ) {
int i, bit, tmp;
if ( offset == 0 )
- bit = (int)( log( (double)( dplasma_imin(arg->p, arg->ldd - k) ) ) / log( 2. ) );
+ bit = (int)( log( (double)( libhqr_imin(arg->p, arg->ldd - k) ) ) / log( 2. ) );
else {
bit = 0;
while( (offset & (1 << bit )) == 0 )
@@ -1043,7 +1043,7 @@ inline static int hqr_high_fibonacci_prevpiv( const hqr_subpiv_t *qrpiv, int k,
int lp = p - k;
int lstart= start - k;
- int end = dplasma_imin(qrpiv->ldd-k, qrpiv->p);
+ int end = libhqr_imin(qrpiv->ldd-k, qrpiv->p);
for( i=lstart+1; i<end; i++ )
if ( (qrpiv->ipiv)[i] == lp )
return i+k;
@@ -1055,7 +1055,7 @@ inline static int hqr_high_fibonacci_nextpiv( const hqr_subpiv_t *qrpiv, int k,
int i;
myassert( p>=k && (start == qrpiv->ldd || start-k <= qrpiv->p) );
- for( i=dplasma_imin(start-k-1, qrpiv->p-1); i>0; i-- )
+ for( i=libhqr_imin(start-k-1, qrpiv->p-1); i>0; i-- )
if ( (qrpiv->ipiv)[i] == (p-k) )
return i + k;
return (qrpiv->ldd);
@@ -1120,9 +1120,9 @@ static void hqr_high_greedy1p_init(hqr_subpiv_t *arg){
memset( nZ, 0, minMN*sizeof(int));
nT[0] = mt;
- nZ[0] = dplasma_imax( mt - p, 0 );
+ nZ[0] = libhqr_imax( mt - p, 0 );
for(k=1; k<minMN; k++) {
- height = dplasma_imax(mt-k-p, 0);
+ height = libhqr_imax(mt-k-p, 0);
nT[k] = height;
nZ[k] = height;
}
@@ -1173,7 +1173,7 @@ static int hqr_high_greedy_nextpiv(const hqr_subpiv_t *arg, int k, int p, int st
{
int i;
myassert( (start >= k && start < k+arg->p) || start == arg->ldd );
- for( i=dplasma_imin(start-1, k+arg->p-1); i > k; i-- )
+ for( i=libhqr_imin(start-1, k+arg->p-1); i > k; i-- )
if ( (arg->ipiv)[i-k + k* (arg->p)] == p )
return i;
return (arg->ldd);
@@ -1212,9 +1212,9 @@ static void hqr_high_greedy_init(hqr_subpiv_t *arg, int minMN){
memset( nZ, 0, minMN*sizeof(int));
nT[0] = mt;
- nZ[0] = dplasma_imax( mt - p, 0 );
+ nZ[0] = libhqr_imax( mt - p, 0 );
for(k=1; k<minMN; k++) {
- height = dplasma_imax(mt-k-p, 0);
+ height = libhqr_imax(mt-k-p, 0);
nT[k] = height;
nZ[k] = height;
}
@@ -1257,7 +1257,7 @@ static void hqr_high_greedy_init(hqr_subpiv_t *arg, int minMN){
* Generic functions currpiv,prevpiv,nextpiv
*
***************************************************/
-static int hqr_currpiv(const dplasma_qrtree_t *qrtree, int k, int m)
+static int hqr_currpiv(const libhqr_tree_t *qrtree, int k, int m)
{
hqr_args_t *arg = (hqr_args_t*)(qrtree->args);
int tmp, tmpk, perm_m;
@@ -1346,7 +1346,7 @@ static int hqr_currpiv(const dplasma_qrtree_t *qrtree, int k, int m)
* - -1 if start doesn't respect the previous conditions
* - m, the following row killed by p if it exists, A->mt otherwise
*/
-static int hqr_nextpiv(const dplasma_qrtree_t *qrtree, int k, int pivot, int start)
+static int hqr_nextpiv(const libhqr_tree_t *qrtree, int k, int pivot, int start)
{
hqr_args_t *arg = (hqr_args_t*)(qrtree->args);
int tmp, ls, lp, nextp;
@@ -1380,12 +1380,12 @@ static int hqr_nextpiv(const dplasma_qrtree_t *qrtree, int k, int pivot, int sta
{
case -1:
- if ( lp == DPLASMA_QR_KILLED_BY_TS ) {
+ if ( lp == LIBHQR_KILLED_BY_TS ) {
myassert( start == gmt );
return gmt;
}
- case DPLASMA_QR_KILLED_BY_TS:
+ case LIBHQR_KILLED_BY_TS:
/* If the tile is over the diagonal of step k, skip directly to type 2 */
if ( arg->domino && lpivot < k )
@@ -1403,7 +1403,7 @@ static int hqr_nextpiv(const dplasma_qrtree_t *qrtree, int k, int pivot, int sta
start = gmt;
lstart = arg->llvl->ldd * a;
- case DPLASMA_QR_KILLED_BY_LOCALTREE:
+ case LIBHQR_KILLED_BY_LOCALTREE:
/* If the tile is over the diagonal of step k, skip directly to type 2 */
if ( arg->domino && lpivot < k )
@@ -1424,9 +1424,9 @@ static int hqr_nextpiv(const dplasma_qrtree_t *qrtree, int k, int pivot, int sta
start = gmt;
lstart = arg->llvl->ldd * a;
- case DPLASMA_QR_KILLED_BY_DOMINO:
+ case LIBHQR_KILLED_BY_DOMINO:
- if ( lp < DPLASMA_QR_KILLED_BY_DOMINO ) {
+ if ( lp < LIBHQR_KILLED_BY_DOMINO ) {
return gmt;
}
@@ -1442,9 +1442,9 @@ static int hqr_nextpiv(const dplasma_qrtree_t *qrtree, int k, int pivot, int sta
start = gmt;
lstart = arg->llvl->ldd * a;
- case DPLASMA_QR_KILLED_BY_DISTTREE:
+ case LIBHQR_KILLED_BY_DISTTREE:
- if ( lp < DPLASMA_QR_KILLED_BY_DISTTREE ) {
+ if ( lp < LIBHQR_KILLED_BY_DISTTREE ) {
return gmt;
}
@@ -1478,7 +1478,7 @@ static int hqr_nextpiv(const dplasma_qrtree_t *qrtree, int k, int pivot, int sta
* - m, the previous row killed by p if it exists, A->mt otherwise
*/
static int
-hqr_prevpiv(const dplasma_qrtree_t *qrtree, int k, int pivot, int start)
+hqr_prevpiv(const libhqr_tree_t *qrtree, int k, int pivot, int start)
{
hqr_args_t *arg = (hqr_args_t*)(qrtree->args);
int tmp, ls, lp, nextp;
@@ -1505,13 +1505,13 @@ hqr_prevpiv(const dplasma_qrtree_t *qrtree, int k, int pivot, int start)
ls = hqr_gettype( qrtree, k, ostart );
lp = hqr_gettype( qrtree, k, opivot );
- if ( lp == DPLASMA_QR_KILLED_BY_TS )
+ if ( lp == LIBHQR_KILLED_BY_TS )
return gmt;
myassert( lp >= ls );
switch( ls )
{
- case DPLASMA_QR_KILLED_BY_DISTTREE:
+ case LIBHQR_KILLED_BY_DISTTREE:
if( arg->hlvl != NULL ) {
tmp = arg->hlvl->prevpiv( arg->hlvl, k, pivot, start );
if ( tmp != gmt )
@@ -1521,7 +1521,7 @@ hqr_prevpiv(const dplasma_qrtree_t *qrtree, int k, int pivot, int start)
start = pivot;
lstart = pivot / p;
- case DPLASMA_QR_KILLED_BY_DOMINO:
+ case LIBHQR_KILLED_BY_DOMINO:
/* If the tile is over the diagonal of step k, process it as type 2 */
if ( arg->domino && lpivot < k ) {
@@ -1529,7 +1529,7 @@ hqr_prevpiv(const dplasma_qrtree_t *qrtree, int k, int pivot, int start)
(start+p < gmt ) )
return perm[start+p];
- if ( lp > DPLASMA_QR_KILLED_BY_LOCALTREE )
+ if ( lp > LIBHQR_KILLED_BY_LOCALTREE )
return gmt;
}
@@ -1538,7 +1538,7 @@ hqr_prevpiv(const dplasma_qrtree_t *qrtree, int k, int pivot, int start)
/* If it is the 'local' diagonal block, we go to 1 */
- case DPLASMA_QR_KILLED_BY_LOCALTREE:
+ case LIBHQR_KILLED_BY_LOCALTREE:
/* If the tile is over the diagonal of step k and is of type 2,
it cannot annihilate type 0 or 1 */
if ( arg->domino && lpivot < k )
@@ -1555,7 +1555,7 @@ hqr_prevpiv(const dplasma_qrtree_t *qrtree, int k, int pivot, int start)
start = pivot;
- case DPLASMA_QR_KILLED_BY_TS:
+ case LIBHQR_KILLED_BY_TS:
/* Search for predecessor in TS tree */
/* if ( ( start+p < gmt ) && */
/* ( (((start+p) / p) % a) != 0 ) ) */
@@ -1585,7 +1585,7 @@ hqr_prevpiv(const dplasma_qrtree_t *qrtree, int k, int pivot, int start)
*
***************************************************/
static void
-hqr_genperm( dplasma_qrtree_t *qrtree )
+hqr_genperm( libhqr_tree_t *qrtree )
{
hqr_args_t *arg = (hqr_args_t*)(qrtree->args);
int m = qrtree->mt;
@@ -1593,7 +1593,7 @@ hqr_genperm( dplasma_qrtree_t *qrtree )
int a = qrtree->a;
int p = qrtree->p;
int domino = arg->domino;
- int minMN = dplasma_imin( m, n );
+ int minMN = libhqr_imin( m, n );
int pa = p * a;
int i, j, k;
int nbextra1;
@@ -1618,7 +1618,7 @@ hqr_genperm( dplasma_qrtree_t *qrtree )
end2 = p + ( domino ? k*p : k + nbextra1 );
end2 = (( end2 + pa - 1 ) / pa ) * pa;
- end2 = dplasma_imin( end2, m );
+ end2 = libhqr_imin( end2, m );
/*
* All tiles of type 3, 2 and:
@@ -1656,7 +1656,7 @@ hqr_genperm( dplasma_qrtree_t *qrtree )
}
static int
-hqr_getinvperm( const dplasma_qrtree_t *qrtree, int k, int m )
+hqr_getinvperm( const libhqr_tree_t *qrtree, int k, int m )
{
hqr_args_t *arg = (hqr_args_t*)(qrtree->args);
int gmt = qrtree->mt + 1;
@@ -1664,7 +1664,7 @@ hqr_getinvperm( const dplasma_qrtree_t *qrtree, int k, int m )
int p = qrtree->p;
int pa = p * a;
int start = m / pa * pa;
- int stop = dplasma_imin( start + pa, gmt ) - start;
+ int stop = libhqr_imin( start + pa, gmt ) - start;
int *perm = arg->perm + gmt * k + start;
int i;
@@ -1685,18 +1685,18 @@ hqr_getinvperm( const dplasma_qrtree_t *qrtree, int k, int m )
*
* @ingroup dplasma
*
- * dplasma_hqr_init - Creates the tree structure that will describes the
+ * libhqr_hqr_init - Creates the tree structure that will describes the
* operation performed during QR/LQ factorization with parameterized QR/LQ
* algorithms family.
*
* Trees available parameters are described below. It is recommended to:
* - set p to the same value than the P-by-Q process grid used to distribute
* the data. (P for QR factorization, Q for LQ factorization).
- * - set the low level tree to DPLASMA_GREEDY_TREE.
+ * - set the low level tree to LIBHQR_GREEDY_TREE.
* - set the high level tree to:
- * 1) DPLASMA_FLAT_TREE when the problem is square, because it divides
+ * 1) LIBHQR_FLAT_TREE when the problem is square, because it divides
* by two the volume of communication of any other tree.
- * 2) DPLASMA_FIBONACCI_TREE when the problem is tall and skinny (QR) or
+ * 2) LIBHQR_FIBONACCI_TREE when the problem is tall and skinny (QR) or
* small and fat (LQ), because it reduces the critical path length.
* - Disable the domino effect when problem is square, to keep high efficiency
* kernel proportion high.
@@ -1729,27 +1729,27 @@ hqr_getinvperm( const dplasma_qrtree_t *qrtree, int k, int m )
* Defines the tree used to reduce the main tiles of each local domain
* together. The matrix of those tiles has a lower triangular structure
* with a diagonal by step a.
- * @arg DPLASMA_FLAT_TREE: A Flat tree is used to reduce the local
+ * @arg LIBHQR_FLAT_TREE: A Flat tree is used to reduce the local
* tiles.
- * @arg DPLASMA_GREEDY_TREE: A Greedy tree is used to reduce the local
+ * @arg LIBHQR_GREEDY_TREE: A Greedy tree is used to reduce the local
* tiles.
- * @arg DPLASMA_FIBONACCI_TREE: A Fibonacci tree is used to reduce the
+ * @arg LIBHQR_FIBONACCI_TREE: A Fibonacci tree is used to reduce the
* local tiles.
- * @arg DPLASMA_BINARY_TREE: A Binary tree is used to reduce the local
+ * @arg LIBHQR_BINARY_TREE: A Binary tree is used to reduce the local
* tiles.
- * @arg -1: The default is used (DPLASMA_GREEDY_TREE)
+ * @arg -1: The default is used (LIBHQR_GREEDY_TREE)
*
* @param[in] type_hlvl
* Defines the tree used to reduce the main tiles of each domain. This
* is a band lower diagonal matrix of width p.
- * @arg DPLASMA_FLAT_TREE: A Flat tree is used to reduce the tiles.
- * @arg DPLASMA_GREEDY_TREE: A Greedy tree is used to reduce the tiles.
- * @arg DPLASMA_FIBONACCI_TREE: A Fibonacci tree is used to reduce the
+ * @arg LIBHQR_FLAT_TREE: A Flat tree is used to reduce the tiles.
+ * @arg LIBHQR_GREEDY_TREE: A Greedy tree is used to reduce the tiles.
+ * @arg LIBHQR_FIBONACCI_TREE: A Fibonacci tree is used to reduce the
* tiles.
- * @arg DPLASMA_BINARY_TREE: A Binary tree is used to reduce the tiles.
- * @arg DPLASMA_GREEDY1P_TREE: A Greedy tree is computed for the first
+ * @arg LIBHQR_BINARY_TREE: A Binary tree is used to reduce the tiles.
+ * @arg LIBHQR_GREEDY1P_TREE: A Greedy tree is computed for the first
* column and then duplicated on all others.
- * @arg -1: The default is used (DPLASMA_FIBONACCI_TREE)
+ * @arg -1: The default is used (LIBHQR_FIBONACCI_TREE)
*
* @param[in] a
* Defines the size of the local domains on which a classic flat TS
@@ -1796,8 +1796,8 @@ hqr_getinvperm( const dplasma_qrtree_t *qrtree, int k, int m )
*
*******************************************************************************
*
- * @sa dplasma_hqr_finalize
- * @sa dplasma_systolic_init
+ * @sa libhqr_hqr_finalize
+ * @sa libhqr_systolic_init
* @sa dplasma_zgeqrf_param
* @sa dplasma_cgeqrf_param
* @sa dplasma_dgeqrf_param
@@ -1805,8 +1805,8 @@ hqr_getinvperm( const dplasma_qrtree_t *qrtree, int k, int m )
*
******************************************************************************/
int
-dplasma_hqr_init( dplasma_qrtree_t *qrtree,
- PLASMA_enum trans, tiled_matrix_desc_t *A,
+libhqr_hqr_init( libhqr_tree_t *qrtree,
+ libhqr_typefacto_e trans, libhqr_tiledesc_t *A,
int type_llvl, int type_hlvl,
int a, int p,
int domino, int tsrr )
@@ -1816,30 +1816,29 @@ dplasma_hqr_init( dplasma_qrtree_t *qrtree,
hqr_args_t *arg;
if (qrtree == NULL) {
- dplasma_error("dplasma_hqr_init", "illegal value of qrtree");
- return -1;
+ fprintf(stderr, "libhqr_hqr_init, illegal value of qrtree");
+ return -1;
}
- if ((trans != PlasmaNoTrans) &&
- (trans != PlasmaTrans) &&
- (trans != PlasmaConjTrans)) {
- dplasma_error("dplasma_hqr_init", "illegal value of trans");
- return -2;
+ if ((trans != LIBHQR_QR) &&
+ (trans != LIBHQR_LQ)) {
+ fprintf(stderr, "libhqr_hqr_ini, illegal value of trans");
+ return -2;
}
if (A == NULL) {
- dplasma_error("dplasma_hqr_init", "illegal value of A");
- return -3;
+ fprintf(stderr, "libhqr_hqr_init, illegal value of A");
+ return -3;
}
/* Compute parameters */
- a = (a == -1) ? 4 : dplasma_imax( a, 1 );
- p = dplasma_imax( p, 1 );
+ a = (a == -1) ? 4 : libhqr_imax( a, 1 );
+ p = libhqr_imax( p, 1 );
/* Domino */
if ( domino >= 0 ) {
domino = domino ? 1 : 0;
}
else {
- if (trans == PlasmaNoTrans) {
+ if (trans == LIBHQR_QR) {
ratio = ((double)(A->nt) / (double)(A->mt));
} else {
ratio = ((double)(A->mt) / (double)(A->nt));
@@ -1859,10 +1858,10 @@ dplasma_hqr_init( dplasma_qrtree_t *qrtree,
qrtree->nextpiv = hqr_nextpiv;
qrtree->prevpiv = hqr_prevpiv;
- qrtree->mt = (trans == PlasmaNoTrans) ? A->mt : A->nt;
- qrtree->nt = (trans == PlasmaNoTrans) ? A->nt : A->mt;
+ qrtree->mt = (trans == LIBHQR_QR) ? A->mt : A->nt;
+ qrtree->nt = (trans == LIBHQR_QR) ? A->nt : A->mt;
- a = dplasma_imin( a, qrtree->mt );
+ a = libhqr_imin( a, qrtree->mt );
qrtree->a = a;
qrtree->p = p;
@@ -1876,7 +1875,7 @@ dplasma_hqr_init( dplasma_qrtree_t *qrtree,
arg->llvl = (hqr_subpiv_t*) malloc( sizeof(hqr_subpiv_t) );
arg->hlvl = NULL;
- minMN = dplasma_imin(A->mt, A->nt);
+ minMN = libhqr_imin(A->mt, A->nt);
low_mt = (qrtree->mt + p * a - 1) / ( p * a );
arg->llvl->minMN = minMN;
@@ -1886,19 +1885,19 @@ dplasma_hqr_init( dplasma_qrtree_t *qrtree,
arg->llvl->domino = domino;
switch( type_llvl ) {
- case DPLASMA_FLAT_TREE :
+ case LIBHQR_FLAT_TREE :
hqr_low_flat_init(arg->llvl);
break;
- case DPLASMA_FIBONACCI_TREE :
+ case LIBHQR_FIBONACCI_TREE :
hqr_low_fibonacci_init(arg->llvl, minMN);
break;
- case DPLASMA_BINARY_TREE :
+ case LIBHQR_BINARY_TREE :
hqr_low_binary_init(arg->llvl);
break;
- case DPLASMA_GREEDY1P_TREE :
+ case LIBHQR_GREEDY1P_TREE :
hqr_low_greedy1p_init(arg->llvl, minMN);
break;
- case DPLASMA_GREEDY_TREE :
+ case LIBHQR_GREEDY_TREE :
default:
hqr_low_greedy_init(arg->llvl, minMN);
}
@@ -1913,19 +1912,19 @@ dplasma_hqr_init( dplasma_qrtree_t *qrtree,
arg->hlvl->domino = domino;
switch( type_hlvl ) {
- case DPLASMA_FLAT_TREE :
+ case LIBHQR_FLAT_TREE :
hqr_high_flat_init(arg->hlvl);
break;
- case DPLASMA_GREEDY_TREE :
+ case LIBHQR_GREEDY_TREE :
hqr_high_greedy_init(arg->hlvl, minMN);
break;
- case DPLASMA_GREEDY1P_TREE :
+ case LIBHQR_GREEDY1P_TREE :
hqr_high_greedy1p_init(arg->hlvl);
break;
- case DPLASMA_BINARY_TREE :
+ case LIBHQR_BINARY_TREE :
hqr_high_binary_init(arg->hlvl);
break;
- case DPLASMA_FIBONACCI_TREE :
+ case LIBHQR_FIBONACCI_TREE :
hqr_high_fibonacci_init(arg->hlvl);
break;
default:
@@ -1948,8 +1947,8 @@ dplasma_hqr_init( dplasma_qrtree_t *qrtree,
*
* @ingroup dplasma
*
- * dplasma_hqr_finalize - Cleans the qrtree data structure allocated by call to
- * dplasma_hqr_init().
+ * libhqr_hqr_finalize - Cleans the qrtree data structure allocated by call to
+ * libhqr_hqr_init().
*
*******************************************************************************
*
@@ -1959,11 +1958,11 @@ dplasma_hqr_init( dplasma_qrtree_t *qrtree,
*
*******************************************************************************
*
- * @sa dplasma_hqr_init
+ * @sa libhqr_hqr_init
*
******************************************************************************/
void
-dplasma_hqr_finalize( dplasma_qrtree_t *qrtree )
+libhqr_hqr_finalize( libhqr_tree_t *qrtree )
{
hqr_args_t *arg = (hqr_args_t*)(qrtree->args);
@@ -1991,10 +1990,10 @@ dplasma_hqr_finalize( dplasma_qrtree_t *qrtree )
/*
* Common functions
*/
-static int svd_getnbgeqrf( const dplasma_qrtree_t *qrtree, int k );
-static int svd_getm( const dplasma_qrtree_t *qrtree, int k, int i );
-static int svd_geti( const dplasma_qrtree_t *qrtree, int k, int m );
-static int svd_gettype( const dplasma_qrtree_t *qrtree, int k, int m );
+static int svd_getnbgeqrf( const libhqr_tree_t *qrtree, int k );
+static int svd_getm( const libhqr_tree_t *qrtree, int k, int i );
+static int svd_geti( const libhqr_tree_t *qrtree, int k, int m );
+static int svd_gettype( const libhqr_tree_t *qrtree, int k, int m );
#define svd_getipiv( __qrtree, _k ) ((__qrtree)->llvl->ipiv + ((__qrtree)->llvl->ldd) * (_k) )
#define svd_geta( __qrtree, _k ) ( (svd_getipiv( (__qrtree), (_k) ))[0] )
@@ -2006,7 +2005,7 @@ static int svd_gettype( const dplasma_qrtree_t *qrtree, int k, int m );
* The number of geqrt to execute in the panel k
*/
static int
-svd_getnbgeqrf( const dplasma_qrtree_t *qrtree,
+svd_getnbgeqrf( const libhqr_tree_t *qrtree,
int k )
{
hqr_args_t *arg = (hqr_args_t*)(qrtree->args);
@@ -2033,9 +2032,9 @@ svd_getnbgeqrf( const dplasma_qrtree_t *qrtree,
nb_1 = (nb_12 - nb_11) / a;
/* Add leftover */
- nb_1 += dplasma_imin( p, gmt - nb_12 );
+ nb_1 += libhqr_imin( p, gmt - nb_12 );
- return dplasma_imin( nb_1 + nb_2 + nb_3, gmt - k);
+ return libhqr_imin( nb_1 + nb_2 + nb_3, gmt - k);
}
/*
@@ -2045,7 +2044,7 @@ svd_getnbgeqrf( const dplasma_qrtree_t *qrtree,
* The global indice m of the i th geqrt in the panel k
*/
static int
-svd_getm( const dplasma_qrtree_t *qrtree,
+svd_getm( const libhqr_tree_t *qrtree,
int k, int i )
{
hqr_args_t *arg = (hqr_args_t*)(qrtree->args);
@@ -2075,7 +2074,7 @@ svd_getm( const dplasma_qrtree_t *qrtree,
* The index i of the geqrt in the panel k
*/
static int
-svd_geti( const dplasma_qrtree_t *qrtree,
+svd_geti( const libhqr_tree_t *qrtree,
int k, int m )
{
hqr_args_t *arg = (hqr_args_t*)(qrtree->args);
@@ -2109,7 +2108,7 @@ svd_geti( const dplasma_qrtree_t *qrtree,
* 3 - if m is reduced in distributed
*/
static int
-svd_gettype( const dplasma_qrtree_t *qrtree,
+svd_gettype( const libhqr_tree_t *qrtree,
int k, int m )
{
hqr_args_t *arg = (hqr_args_t*)(qrtree->args);
@@ -2190,7 +2189,7 @@ svd_low_adaptiv_init(hqr_subpiv_t *arg,
int *ipiv;
int mt, a, p, pa, maxmt, myrank;
int j, k, height, start, end, nT, nZ;
- int minMN = dplasma_imin(gmt, gnt);
+ int minMN = libhqr_imin(gmt, gnt);
arg->currpiv = svd_low_adaptiv_currpiv;
arg->nextpiv = svd_low_adaptiv_nextpiv;
@@ -2212,16 +2211,16 @@ svd_low_adaptiv_init(hqr_subpiv_t *arg,
* so,
* a <= mt * (gnt-k) / (ratio * nbcores )
*/
- height = dplasma_iceil( gmt-k, p );
- a = dplasma_imax( height * (gnt-k) / (ratio * nbcores), 1 );
+ height = libhqr_iceil( gmt-k, p );
+ a = libhqr_imax( height * (gnt-k) / (ratio * nbcores), 1 );
/* Now let's make sure all sub-parts are equilibrate */
- j = dplasma_iceil( height, a );
- a = dplasma_iceil( gmt-k, j );
+ j = libhqr_iceil( height, a );
+ a = libhqr_iceil( gmt-k, j );
/* Compute max dimension of the tree */
- mt = dplasma_iceil( gmt, p * a );
- maxmt = dplasma_imax( mt, maxmt );
+ mt = libhqr_iceil( gmt, p * a );
+ maxmt = libhqr_imax( mt, maxmt );
}
arg->ldd = maxmt + 2;
@@ -2243,15 +2242,15 @@ svd_low_adaptiv_init(hqr_subpiv_t *arg,
* so,
* a <= mt * (gnt-k) / (ratio * nbcores )
*/
- height = dplasma_iceil( gmt-k, p );
- a = dplasma_imax( height * (gnt-k) / (ratio * nbcores), 1 );
+ height = libhqr_iceil( gmt-k, p );
+ a = libhqr_imax( height * (gnt-k) / (ratio * nbcores), 1 );
/* Now let's make sure all sub-parts are equilibrate */
- j = dplasma_iceil( height, a );
- a = dplasma_iceil( gmt-k, j );
+ j = libhqr_iceil( height, a );
+ a = libhqr_iceil( gmt-k, j );
pa = p * a;
- mt = dplasma_iceil( gmt, pa );
+ mt = libhqr_iceil( gmt, pa );
ipiv[0] = a;
ipiv[1] = mt;
@@ -2259,7 +2258,7 @@ svd_low_adaptiv_init(hqr_subpiv_t *arg,
assert( mt < arg->ldd-1 );
/* Number of tiles to factorized in this column on this rank */
- nT = dplasma_imax( mt - ((k + p - 1 - myrank) / pa), 0 );
+ nT = libhqr_imax( mt - ((k + p - 1 - myrank) / pa), 0 );
/* Number of tiles already killed */
nZ = 0;
@@ -2321,7 +2320,7 @@ svd_low_adaptiv_init(hqr_subpiv_t *arg,
* Generic functions currpiv,prevpiv,nextpiv
*
***************************************************/
-static int svd_currpiv(const dplasma_qrtree_t *qrtree, int k, int m)
+static int svd_currpiv(const libhqr_tree_t *qrtree, int k, int m)
{
hqr_args_t *arg = (hqr_args_t*)(qrtree->args);
int tmp, tmpk;
@@ -2379,7 +2378,7 @@ static int svd_currpiv(const dplasma_qrtree_t *qrtree, int k, int m)
* - -1 if start doesn't respect the previous conditions
* - m, the following row killed by p if it exists, A->mt otherwise
*/
-static int svd_nextpiv(const dplasma_qrtree_t *qrtree, int k, int pivot, int start)
+static int svd_nextpiv(const libhqr_tree_t *qrtree, int k, int pivot, int start)
{
hqr_args_t *arg = (hqr_args_t*)(qrtree->args);
int tmp, ls, lp, nextp;
@@ -2404,16 +2403,16 @@ static int svd_nextpiv(const dplasma_qrtree_t *qrtree, int k, int pivot, int sta
switch( ls )
{
- case DPLASMA_QR_KILLED_BY_DOMINO:
+ case LIBHQR_KILLED_BY_DOMINO:
assert(0);
case -1:
- if ( lp == DPLASMA_QR_KILLED_BY_TS ) {
+ if ( lp == LIBHQR_KILLED_BY_TS ) {
myassert( start == gmt );
return gmt;
}
- case DPLASMA_QR_KILLED_BY_TS:
+ case LIBHQR_KILLED_BY_TS:
if ( start == gmt )
nextp = pivot + p;
else
@@ -2426,7 +2425,7 @@ static int svd_nextpiv(const dplasma_qrtree_t *qrtree, int k, int pivot, int sta
start = gmt;
lstart = ldd * a;
- case DPLASMA_QR_KILLED_BY_LOCALTREE:
+ case LIBHQR_KILLED_BY_LOCALTREE:
/* Get the next pivot for the low level tree */
tmp = arg->llvl->nextpiv(arg->llvl, k, pivot, lstart / a );
@@ -2441,9 +2440,9 @@ static int svd_nextpiv(const dplasma_qrtree_t *qrtree, int k, int pivot, int sta
start = gmt;
lstart = ldd * a;
- case DPLASMA_QR_KILLED_BY_DISTTREE:
+ case LIBHQR_KILLED_BY_DISTTREE:
- if ( lp < DPLASMA_QR_KILLED_BY_DISTTREE ) {
+ if ( lp < LIBHQR_KILLED_BY_DISTTREE ) {
return gmt;
}
@@ -2477,7 +2476,7 @@ static int svd_nextpiv(const dplasma_qrtree_t *qrtree, int k, int pivot, int sta
* - m, the previous row killed by p if it exists, A->mt otherwise
*/
static int
-svd_prevpiv(const dplasma_qrtree_t *qrtree, int k, int pivot, int start)
+svd_prevpiv(const libhqr_tree_t *qrtree, int k, int pivot, int start)
{
hqr_args_t *arg = (hqr_args_t*)(qrtree->args);
int tmp, ls, lp, nextp;
@@ -2499,15 +2498,15 @@ svd_prevpiv(const dplasma_qrtree_t *qrtree, int k, int pivot, int start)
ls = svd_gettype( qrtree, k, start );
lp = svd_gettype( qrtree, k, pivot );
- if ( lp == DPLASMA_QR_KILLED_BY_TS )
+ if ( lp == LIBHQR_KILLED_BY_TS )
return gmt;
myassert( lp >= ls );
switch( ls )
{
- case DPLASMA_QR_KILLED_BY_DOMINO:
+ case LIBHQR_KILLED_BY_DOMINO:
assert(0);
- case DPLASMA_QR_KILLED_BY_DISTTREE:
+ case LIBHQR_KILLED_BY_DISTTREE:
if( arg->hlvl != NULL ) {
tmp = arg->hlvl->prevpiv( arg->hlvl, k, pivot, start );
if ( tmp != gmt )
@@ -2517,7 +2516,7 @@ svd_prevpiv(const dplasma_qrtree_t *qrtree, int k, int pivot, int start)
start = pivot;
lstart = pivot / p;
- case DPLASMA_QR_KILLED_BY_LOCALTREE:
+ case LIBHQR_KILLED_BY_LOCALTREE:
tmp = arg->llvl->prevpiv(arg->llvl, k, pivot, lstart / a);
if ( (tmp * a * p + rpivot >= gmt)
@@ -2529,7 +2528,7 @@ svd_prevpiv(const dplasma_qrtree_t *qrtree, int k, int pivot, int start)
start = pivot;
- case DPLASMA_QR_KILLED_BY_TS:
+ case LIBHQR_KILLED_BY_TS:
/* Search for predecessor in TS tree */
/* if ( ( start+p < gmt ) && */
/* ( (((start+p) / p) % a) != 0 ) ) */
@@ -2558,17 +2557,17 @@ svd_prevpiv(const dplasma_qrtree_t *qrtree, int k, int pivot, int start)
*
* @ingroup dplasma
*
- * dplasma_svd_init - Create the tree structures that will describes the
+ * libhqr_svd_init - Create the tree structures that will describes the
* operation performed during QR/LQ reduction step of the gebrd_ge2gb operation.
*
* Trees available parameters are described below. It is recommended to:
* - set p to the same value than the P-by-Q process grid used to distribute
* the data. (P for QR factorization, Q for LQ factorization).
- * - set the low level tree to DPLASMA_GREEDY_TREE.
+ * - set the low level tree to LIBHQR_GREEDY_TREE.
* - set the high level tree to:
- * 1) DPLASMA_FLAT_TREE when the problem is square, because it divides
+ * 1) LIBHQR_FLAT_TREE when the problem is square, because it divides
* by two the volume of communication of any other tree.
- * 2) DPLASMA_FIBONACCI_TREE when the problem is tall and skinny (QR) or
+ * 2) LIBHQR_FIBONACCI_TREE when the problem is tall and skinny (QR) or
* small and fat (LQ), because it reduces the critical path length.
* - Disable the domino effect when problem is square, to keep high efficiency
* kernel proportion high.
@@ -2588,9 +2587,8 @@ svd_prevpiv(const dplasma_qrtree_t *qrtree, int k, int pivot, int start)
* On exit, the structure initialized according to the given parameters.
*
* @param[in] trans
- * @arg PlasmaNoTrans: Structure is initialized for the QR steps.
- * @arg PlasmaTrans: Structure is initialized for the LQ steps.
- * @arg PlasmaConjTrans: Structure is initialized for the LQ steps.
+ * @arg LIBHQR_QR: Structure is initialized for the QR steps.
+ * @arg LIBHQR_LQ: Structure is initialized for the LQ steps.
*
* @param[in,out] A
* Descriptor of the distributed matrix A to be factorized, on which
@@ -2602,14 +2600,14 @@ svd_prevpiv(const dplasma_qrtree_t *qrtree, int k, int pivot, int start)
* @param[in] type_hlvl
* Defines the tree used to reduce the main tiles of each domain. This
* is a band lower diagonal matrix of width p.
- * @arg DPLASMA_FLAT_TREE: A Flat tree is used to reduce the tiles.
- * @arg DPLASMA_GREEDY_TREE: A Greedy tree is used to reduce the tiles.
- * @arg DPLASMA_FIBONACCI_TREE: A Fibonacci tree is used to reduce the
+ * @arg LIBHQR_FLAT_TREE: A Flat tree is used to reduce the tiles.
+ * @arg LIBHQR_GREEDY_TREE: A Greedy tree is used to reduce the tiles.
+ * @arg LIBHQR_FIBONACCI_TREE: A Fibonacci tree is used to reduce the
* tiles.
- * @arg DPLASMA_BINARY_TREE: A Binary tree is used to reduce the tiles.
- * @arg DPLASMA_GREEDY1P_TREE: A Greedy tree is computed for the first
+ * @arg LIBHQR_BINARY_TREE: A Binary tree is used to reduce the tiles.
+ * @arg LIBHQR_GREEDY1P_TREE: A Greedy tree is computed for the first
* column and then duplicated on all others.
- * @arg -1: The default is used (DPLASMA_FIBONACCI_TREE)
+ * @arg -1: The default is used (LIBHQR_FIBONACCI_TREE)
*
* @param[in] p
* Defines the number of distributed domains, ie the width of the high
@@ -2617,7 +2615,7 @@ svd_prevpiv(const dplasma_qrtree_t *qrtree, int k, int pivot, int start)
* used. If p == mt, this enforce the high level reduction tree to be
* performed on the full matrix.
* By default, it is recommended to set p to P if trans ==
- * PlasmaNoTrans, and to Q otherwise, where P-by-Q is the process grid
+ * LIBHQR_QR, and to Q otherwise, where P-by-Q is the process grid
* used to distributed the data. (p > 0)
*
* @param[in] nbthread_per_node
@@ -2638,8 +2636,8 @@ svd_prevpiv(const dplasma_qrtree_t *qrtree, int k, int pivot, int start)
*
*******************************************************************************
*
- * @sa dplasma_hqr_finalize
- * @sa dplasma_hqr_init
+ * @sa libhqr_hqr_finalize
+ * @sa libhqr_hqr_init
* @sa dplasma_zgeqrf_param
* @sa dplasma_cgeqrf_param
* @sa dplasma_dgeqrf_param
@@ -2647,30 +2645,29 @@ svd_prevpiv(const dplasma_qrtree_t *qrtree, int k, int pivot, int start)
*
******************************************************************************/
int
-dplasma_svd_init( dplasma_qrtree_t *qrtree,
- PLASMA_enum trans, tiled_matrix_desc_t *A,
+libhqr_svd_init( libhqr_tree_t *qrtree,
+ libhqr_typefacto_e trans, libhqr_tiledesc_t *A,
int type_hlvl, int p, int nbthread_per_node, int ratio )
{
int low_mt, minMN, a = -1;
hqr_args_t *arg;
if (qrtree == NULL) {
- dplasma_error("dplasma_svd_init", "illegal value of qrtree");
+ fprintf(stderr,"libhqr_svd_init, illegal value of qrtree");
return -1;
}
- if ((trans != PlasmaNoTrans) &&
- (trans != PlasmaTrans) &&
- (trans != PlasmaConjTrans)) {
- dplasma_error("dplasma_svd_init", "illegal value of trans");
+ if ((trans != LIBHQR_QR) &&
+ (trans != LIBHQR_LQ)) {
+ fprintf(stderr, "libhqr_svd_init, illegal value of trans");
return -2;
}
if (A == NULL) {
- dplasma_error("dplasma_svd_init", "illegal value of A");
+ fprintf(stderr, "libhqr_svd_init, illegal value of A");
return -3;
}
/* Compute parameters */
- p = dplasma_imax( p, 1 );
+ p = libhqr_imax( p, 1 );
qrtree->getnbgeqrf = svd_getnbgeqrf;
qrtree->getm = svd_getm;
@@ -2680,8 +2677,8 @@ dplasma_svd_init( dplasma_qrtree_t *qrtree,
qrtree->nextpiv = svd_nextpiv;
qrtree->prevpiv = svd_prevpiv;
- qrtree->mt = (trans == PlasmaNoTrans) ? A->mt : A->nt;
- qrtree->nt = (trans == PlasmaNoTrans) ? A->nt : A->mt;
+ qrtree->mt = (trans == LIBHQR_QR) ? A->mt : A->nt;
+ qrtree->nt = (trans == LIBHQR_QR) ? A->nt : A->mt;
qrtree->a = a;
qrtree->p = p;
@@ -2695,7 +2692,7 @@ dplasma_svd_init( dplasma_qrtree_t *qrtree,
arg->llvl = (hqr_subpiv_t*) malloc( sizeof(hqr_subpiv_t) );
arg->hlvl = NULL;
- minMN = dplasma_imin(A->mt, A->nt);
+ minMN = libhqr_imin(A->mt, A->nt);
low_mt = (qrtree->mt + p - 1) / ( p );
arg->llvl->minMN = minMN;
@@ -2705,7 +2702,7 @@ dplasma_svd_init( dplasma_qrtree_t *qrtree,
arg->llvl->domino = 0;
svd_low_adaptiv_init(arg->llvl, qrtree->mt, qrtree->nt,
- nbthread_per_node * (A->super.nodes / p), ratio );
+ nbthread_per_node * (A->nodes / p), ratio );
if ( p > 1 ) {
arg->hlvl = (hqr_subpiv_t*) malloc( sizeof(hqr_subpiv_t) );
@@ -2717,19 +2714,19 @@ dplasma_svd_init( dplasma_qrtree_t *qrtree,
arg->hlvl->domino = 0;
switch( type_hlvl ) {
- case DPLASMA_FLAT_TREE :
+ case LIBHQR_FLAT_TREE :
hqr_high_flat_init(arg->hlvl);
break;
- case DPLASMA_GREEDY_TREE :
+ case LIBHQR_GREEDY_TREE :
hqr_high_greedy_init(arg->hlvl, minMN);
break;
- case DPLASMA_GREEDY1P_TREE :
+ case LIBHQR_GREEDY1P_TREE :
hqr_high_greedy1p_init(arg->hlvl);
break;
- case DPLASMA_BINARY_TREE :
+ case LIBHQR_BINARY_TREE :
hqr_high_binary_init(arg->hlvl);
break;
- case DPLASMA_FIBONACCI_TREE :
+ case LIBHQR_FIBONACCI_TREE :
hqr_high_fibonacci_init(arg->hlvl);
break;
default:
diff --git a/src/dplasma_hqr_dbg.c b/src/libhqr_dbg.c
similarity index 86%
rename from src/dplasma_hqr_dbg.c
rename to src/libhqr_dbg.c
index 33b412f..023f1c0 100644
--- a/src/dplasma_hqr_dbg.c
+++ b/src/libhqr_dbg.c
@@ -70,15 +70,13 @@
* high level tree to reduce communications.
* These lines are defined by (i-k)/p = 0.
*/
-#include "parsec.h"
-#include "dplasma.h"
-#include "dplasmatypes.h"
-#include "dplasmaaux.h"
-
+#include "libhqr.h"
+#include <stdio.h>
+#include <stdlib.h>
#include <math.h>
-#if defined(PARSEC_HAVE_STRING_H)
+//#if defined(LIBHQR_HAVE_STRING_H)
#include <string.h>
-#endif /* defined(PARSEC_HAVE_STRING_H) */
+//#endif /* defined(LIBHQR_HAVE_STRING_H) */
/* static int dplasma_qrtree_getinon0( const qr_piv_t *arg, */
/* const int k, int i, int mt ); */
@@ -87,9 +85,9 @@
if ( !test ) \
return ret;
-int dplasma_qrtree_check( tiled_matrix_desc_t *A, dplasma_qrtree_t *qrtree)
+int libhqr_tree_check( libhqr_tiledesc_t *A, libhqr_tree_t *qrtree)
{
- int minMN = dplasma_imin(A->mt, A->nt );
+ int minMN = libhqr_imin(A->mt, A->nt );
int i, m, k, nb;
int check;
@@ -100,8 +98,8 @@ int dplasma_qrtree_check( tiled_matrix_desc_t *A, dplasma_qrtree_t *qrtree)
* Check Formula for NB geqrt
*/
{
- /* dplasma_qrtree_print_type( A, qrtree ); */
- /* dplasma_qrtree_print_nbgeqrt( A, qrtree ); */
+ /* libhqr_tree_print_type( A, qrtree ); */
+ /* libhqr_tree_print_nbgeqrt( A, qrtree ); */
check = 1;
for (k=0; k<minMN; k++) {
nb = 0;
@@ -130,7 +128,7 @@ int dplasma_qrtree_check( tiled_matrix_desc_t *A, dplasma_qrtree_t *qrtree)
int prevm = -1;
check = 1;
for (k=0; k<minMN; k++) {
- /* dplasma_qrtree_print_geqrt_k( A, qrtree, k ); */
+ /* libhqr_tree_print_geqrt_k( A, qrtree, k ); */
nb = qrtree->getnbgeqrf( qrtree, k );
prevm = -1;
for (i=0; i < nb; i++) {
@@ -138,7 +136,7 @@ int dplasma_qrtree_check( tiled_matrix_desc_t *A, dplasma_qrtree_t *qrtree)
m = qrtree->getm( qrtree, k, i );
/*
- * getm ahas to be the inverse of geti
+ * getm has to be the inverse of geti
*/
if ( i != qrtree->geti( qrtree, k, m) ) {
check = 0;
@@ -191,8 +189,8 @@ int dplasma_qrtree_check( tiled_matrix_desc_t *A, dplasma_qrtree_t *qrtree)
nb++;
}
if ( nb > 1 ) {
- dplasma_qrtree_print_next_k( A, qrtree, k);
- dplasma_qrtree_print_prev_k( A, qrtree, k);
+ libhqr_tree_print_next_k( A, qrtree, k);
+ libhqr_tree_print_prev_k( A, qrtree, k);
printf(" ----------------------------------------------------\n"
" - a = %d, p = %d, M = %d, N = %d\n"
@@ -203,8 +201,8 @@ int dplasma_qrtree_check( tiled_matrix_desc_t *A, dplasma_qrtree_t *qrtree)
return 3;
}
else if ( nb == 0 ) {
- dplasma_qrtree_print_next_k( A, qrtree, k);
- dplasma_qrtree_print_prev_k( A, qrtree, k);
+ libhqr_tree_print_next_k( A, qrtree, k);
+ libhqr_tree_print_prev_k( A, qrtree, k);
printf(" ----------------------------------------------------\n"
" - a = %d, p = %d, M = %d, N = %d\n"
@@ -234,8 +232,8 @@ int dplasma_qrtree_check( tiled_matrix_desc_t *A, dplasma_qrtree_t *qrtree)
nb++;
}
if ( nb > 1 ) {
- dplasma_qrtree_print_next_k( A, qrtree, k);
- dplasma_qrtree_print_prev_k( A, qrtree, k);
+ libhqr_tree_print_next_k( A, qrtree, k);
+ libhqr_tree_print_prev_k( A, qrtree, k);
printf(" ----------------------------------------------------\n"
" - a = %d, p = %d, M = %d, N = %d\n"
@@ -246,8 +244,8 @@ int dplasma_qrtree_check( tiled_matrix_desc_t *A, dplasma_qrtree_t *qrtree)
return 3;
}
else if ( nb == 0 ) {
- dplasma_qrtree_print_next_k( A, qrtree, k);
- dplasma_qrtree_print_prev_k( A, qrtree, k);
+ libhqr_tree_print_next_k( A, qrtree, k);
+ libhqr_tree_print_prev_k( A, qrtree, k);
printf(" ----------------------------------------------------\n"
" - a = %d, p = %d, M = %d, N = %d\n"
@@ -281,8 +279,8 @@ int dplasma_qrtree_check( tiled_matrix_desc_t *A, dplasma_qrtree_t *qrtree)
prev = qrtree->prevpiv(qrtree, k, m, next);
if ( start != prev ) {
- dplasma_qrtree_print_next_k( A, qrtree, k);
- dplasma_qrtree_print_prev_k( A, qrtree, k);
+ libhqr_tree_print_next_k( A, qrtree, k);
+ libhqr_tree_print_prev_k( A, qrtree, k);
printf(" ----------------------------------------------------\n"
" - a = %d, p = %d, M = %d, N = %d\n"
@@ -303,9 +301,9 @@ int dplasma_qrtree_check( tiled_matrix_desc_t *A, dplasma_qrtree_t *qrtree)
return 0;
}
-void dplasma_qrtree_print_type( tiled_matrix_desc_t *A, dplasma_qrtree_t *qrtree )
+void libhqr_tree_print_type( libhqr_tiledesc_t *A, libhqr_tree_t *qrtree )
{
- int minMN = dplasma_imin(A->mt, A->nt );
+ int minMN = libhqr_imin(A->mt, A->nt );
int m, k;
int lm = 0;
int lmg = 0;
@@ -314,19 +312,19 @@ void dplasma_qrtree_print_type( tiled_matrix_desc_t *A, dplasma_qrtree_t *qrtree
printf("\n------------ Localization = Type of pivot --------------\n");
for(m=0; m<A->mt; m++) {
printf("%3d | ", m);
- for (k=0; k<dplasma_imin(minMN, m+1); k++) {
+ for (k=0; k<libhqr_imin(minMN, m+1); k++) {
printf( "%3d ", qrtree->gettype( qrtree, k, m ) );
}
- for (k=dplasma_imin(minMN, m+1); k<minMN; k++) {
+ for (k=libhqr_imin(minMN, m+1); k<minMN; k++) {
printf( " " );
}
printf(" ");
printf("%2d,%3d | ", rank, lmg);
- for (k=0; k<dplasma_imin(minMN, lmg+1); k++) {
+ for (k=0; k<libhqr_imin(minMN, lmg+1); k++) {
printf( "%3d ", qrtree->gettype( qrtree, k, lmg) );
}
- for (k=dplasma_imin(minMN, lmg+1); k<minMN; k++) {
+ for (k=libhqr_imin(minMN, lmg+1); k<minMN; k++) {
printf( " " );
}
lm++; lmg+=qrtree->p;
@@ -339,9 +337,9 @@ void dplasma_qrtree_print_type( tiled_matrix_desc_t *A, dplasma_qrtree_t *qrtree
}
}
-void dplasma_qrtree_print_pivot( tiled_matrix_desc_t *A, dplasma_qrtree_t *qrtree )
+void libhqr_tree_print_pivot( libhqr_tiledesc_t *A, libhqr_tree_t *qrtree )
{
- int minMN = dplasma_imin(A->mt, A->nt );
+ int minMN = libhqr_imin(A->mt, A->nt );
int m, k;
int lm = 0;
int lmg = 0;
@@ -349,19 +347,19 @@ void dplasma_qrtree_print_pivot( tiled_matrix_desc_t *A, dplasma_qrtree_t *qrtre
printf("\n------------ Current Pivot--------------\n");
for(m=0; m<A->mt; m++) {
printf("%3d | ", m);
- for (k=0; k<dplasma_imin(minMN, m+1); k++) {
+ for (k=0; k<libhqr_imin(minMN, m+1); k++) {
printf( "%3d ", qrtree->currpiv(qrtree, k, m) );
}
- for (k=dplasma_imin(minMN, m+1); k<minMN; k++) {
+ for (k=libhqr_imin(minMN, m+1); k<minMN; k++) {
printf( " " );
}
printf(" ");
printf("%2d,%3d | ", rank, lmg);
- for (k=0; k<dplasma_imin(minMN, lmg+1); k++) {
+ for (k=0; k<libhqr_imin(minMN, lmg+1); k++) {
printf( "%3d ", qrtree->currpiv(qrtree, k, lmg) );
}
- for (k=dplasma_imin(minMN, lmg+1); k<minMN; k++) {
+ for (k=libhqr_imin(minMN, lmg+1); k<minMN; k++) {
printf( " " );
}
lm++; lmg+=qrtree->p;
@@ -374,7 +372,7 @@ void dplasma_qrtree_print_pivot( tiled_matrix_desc_t *A, dplasma_qrtree_t *qrtre
}
}
-void dplasma_qrtree_print_next_k( tiled_matrix_desc_t *A, dplasma_qrtree_t *qrtree, int k )
+void libhqr_tree_print_next_k( libhqr_tiledesc_t *A, libhqr_tree_t *qrtree, int k )
{
int m, s;
printf("\n------------ Next (k = %d)--------------\n", k);
@@ -393,7 +391,7 @@ void dplasma_qrtree_print_next_k( tiled_matrix_desc_t *A, dplasma_qrtree_t *qrtr
}
}
-void dplasma_qrtree_print_prev_k( tiled_matrix_desc_t *A, dplasma_qrtree_t *qrtree, int k )
+void libhqr_tree_print_prev_k( libhqr_tiledesc_t *A, libhqr_tree_t *qrtree, int k )
{
int m, s;
printf("\n------------ Prev (k = %d)--------------\n", k);
@@ -412,9 +410,9 @@ void dplasma_qrtree_print_prev_k( tiled_matrix_desc_t *A, dplasma_qrtree_t *qrtr
}
}
-void dplasma_qrtree_print_perm( tiled_matrix_desc_t *A, dplasma_qrtree_t *qrtree, int *perm )
+void libhqr_tree_print_perm( libhqr_tiledesc_t *A, libhqr_tree_t *qrtree, int *perm )
{
- int minMN = dplasma_imin(A->mt, A->nt );
+ int minMN = libhqr_imin(A->mt, A->nt );
int m, k;
(void)qrtree;
@@ -437,9 +435,9 @@ void dplasma_qrtree_print_perm( tiled_matrix_desc_t *A, dplasma_qrtree_t *qrtree
printf( "\n" );
}
-void dplasma_qrtree_print_nbgeqrt( tiled_matrix_desc_t *A, dplasma_qrtree_t *qrtree )
+void libhqr_tree_print_nbgeqrt( libhqr_tiledesc_t *A, libhqr_tree_t *qrtree )
{
- int minMN = dplasma_imin(A->mt, A->nt );
+ int minMN = libhqr_imin(A->mt, A->nt );
int m, k, nb;
printf("\n------------ Nb GEQRT per k --------------\n");
@@ -465,7 +463,7 @@ void dplasma_qrtree_print_nbgeqrt( tiled_matrix_desc_t *A, dplasma_qrtree_t *qrt
printf( "\n" );
}
-void dplasma_qrtree_print_geqrt_k( tiled_matrix_desc_t *A, dplasma_qrtree_t *qrtree, int k )
+void libhqr_tree_print_geqrt_k( libhqr_tiledesc_t *A, libhqr_tree_t *qrtree, int k )
{
int i, m, nb;
(void)A;
@@ -485,12 +483,12 @@ void dplasma_qrtree_print_geqrt_k( tiled_matrix_desc_t *A, dplasma_qrtree_t *qrt
}
-/* static int dplasma_qrtree_getinon0( const dplasma_qrtree_t *qrtree, */
+/* static int libhqr_tree_getinon0( const libhqr_tree_t *qrtree, */
/* const int k, int i, int mt ) */
/* { */
/* int j; */
/* for(j=k; j<mt; j++) { */
-/* if ( dplasma_qrtree_gettype( qrtree, k, j ) != 0 ) */
+/* if ( libhqr_tree_gettype( qrtree, k, j ) != 0 ) */
/* i--; */
/* if ( i == -1 ) */
/* break; */
@@ -523,11 +521,11 @@ char *color[] = {
};
#define DAG_NBCOLORS 7
-void dplasma_qrtree_print_dag( tiled_matrix_desc_t *A, dplasma_qrtree_t *qrtree, char *filename )
+void libhqr_tree_print_dag( libhqr_tiledesc_t *A, libhqr_tree_t *qrtree, char *filename )
{
int *pos, *next, *done;
int k, m, n, lpos, prev, length;
- int minMN = dplasma_imin( A->mt, A->nt );
+ int minMN = libhqr_imin( A->mt, A->nt );
FILE *f = fopen( filename, "w" );
done = (int*)malloc( A->mt * sizeof(int) );
@@ -557,7 +555,7 @@ void dplasma_qrtree_print_dag( tiled_matrix_desc_t *A, dplasma_qrtree_t *qrtree,
if ( next[n] != A->mt )
continue;
if ( n != A->mt ) {
- lpos = dplasma_imax( pos[m], pos[n] );
+ lpos = libhqr_imax( pos[m], pos[n] );
lpos++;
pos[m] = lpos;
pos[n] = lpos;
@@ -592,7 +590,7 @@ void dplasma_qrtree_print_dag( tiled_matrix_desc_t *A, dplasma_qrtree_t *qrtree,
length = 0;
for(m=0; m < A->mt; m++) {
- length = dplasma_imax(length, pos[m]);
+ length = libhqr_imax(length, pos[m]);
}
length++;
for(k=0; k<length; k++)
diff --git a/src/dplasma_systolic_qr.c b/src/libhqr_systolic.c
similarity index 79%
rename from src/dplasma_systolic_qr.c
rename to src/libhqr_systolic.c
index fa10912..f76dbf8 100644
--- a/src/dplasma_systolic_qr.c
+++ b/src/libhqr_systolic.c
@@ -10,14 +10,15 @@
* used in the Xgeqrf_param.jdf file.
*
*/
-#include "parsec.h"
-#include "dplasma.h"
-#include "dplasma_qr_param.h"
+#include "libhqr.h"
+#include <assert.h>
+#include <stdio.h>
+#include <stdlib.h>
#include <math.h>
-#if defined(PARSEC_HAVE_STRING_H)
+//#if defined(LIBHQR_HAVE_STRING_H)
#include <string.h>
-#endif /* defined(PARSEC_HAVE_STRING_H) */
+//#endif /* defined(LIBHQR_HAVE_STRING_H) */
#define PRINT_PIVGEN 0
#ifdef PRINT_PIVGEN
@@ -28,19 +29,19 @@
#define nbextra1_formula ( (k % pa) > (pa - p) ) ? (-k)%pa + pa : 0
-static int systolic_getnbgeqrf( const dplasma_qrtree_t *qrtree, int k )
+static int systolic_getnbgeqrf( const libhqr_tree_t *qrtree, int k )
{
int pq = qrtree->p * qrtree->a;
- return dplasma_imin( pq, qrtree->mt - k);
+ return libhqr_imin( pq, qrtree->mt - k);
}
-static int systolic_getm( const dplasma_qrtree_t *qrtree, int k, int i )
+static int systolic_getm( const libhqr_tree_t *qrtree, int k, int i )
{
(void)qrtree;
return k+i;
}
-static int systolic_geti( const dplasma_qrtree_t *qrtree, int k, int m )
+static int systolic_geti( const libhqr_tree_t *qrtree, int k, int m )
{
(void)qrtree;
return m-k;
@@ -55,7 +56,7 @@ static int systolic_geti( const dplasma_qrtree_t *qrtree, int k, int m )
* 1 - if m is reduced thanks to the 2nd coordinate flat tree
* 3 - if m is reduced thanks to the 1st coordinate flat tree
*/
-static int systolic_gettype( const dplasma_qrtree_t *qrtree, int k, int m ) {
+static int systolic_gettype( const libhqr_tree_t *qrtree, int k, int m ) {
int p = qrtree->p;
int q = qrtree->a;
int pq = p * q;
@@ -78,7 +79,7 @@ static int systolic_gettype( const dplasma_qrtree_t *qrtree, int k, int m ) {
* Generic functions currpiv,prevpiv,nextpiv
*
***************************************************/
-static int systolic_currpiv(const dplasma_qrtree_t *qrtree, int k, int m)
+static int systolic_currpiv(const libhqr_tree_t *qrtree, int k, int m)
{
int p = qrtree->p;
int q = qrtree->a;
@@ -119,7 +120,7 @@ static int systolic_currpiv(const dplasma_qrtree_t *qrtree, int k, int m)
* - -1 if start doesn't respect the previous conditions
* - m, the following row killed by p if it exists, A->mt otherwise
*/
-static int systolic_nextpiv(const dplasma_qrtree_t *qrtree, int k, int pivot, int start)
+static int systolic_nextpiv(const libhqr_tree_t *qrtree, int k, int pivot, int start)
{
int ls, lp, nextp;
int q = qrtree->a;
@@ -138,12 +139,12 @@ static int systolic_nextpiv(const dplasma_qrtree_t *qrtree, int k, int pivot, in
{
case -1:
- if ( lp == DPLASMA_QR_KILLED_BY_TS ) {
+ if ( lp == LIBHQR_KILLED_BY_TS ) {
myassert( start == mt );
return mt;
}
- case DPLASMA_QR_KILLED_BY_TS:
+ case LIBHQR_KILLED_BY_TS:
if ( start == mt )
nextp = pivot + pq;
@@ -155,9 +156,9 @@ static int systolic_nextpiv(const dplasma_qrtree_t *qrtree, int k, int pivot, in
start = mt;
- case DPLASMA_QR_KILLED_BY_LOCALTREE:
+ case LIBHQR_KILLED_BY_LOCALTREE:
- if (lp < DPLASMA_QR_KILLED_BY_DISTTREE)
+ if (lp < LIBHQR_KILLED_BY_DISTTREE)
return mt;
if ( start == mt )
@@ -172,7 +173,7 @@ static int systolic_nextpiv(const dplasma_qrtree_t *qrtree, int k, int pivot, in
start = mt;
- case DPLASMA_QR_KILLED_BY_DISTTREE:
+ case LIBHQR_KILLED_BY_DISTTREE:
if (pivot > k)
return mt;
@@ -208,7 +209,7 @@ static int systolic_nextpiv(const dplasma_qrtree_t *qrtree, int k, int pivot, in
* - -1 if start doesn't respect the previous conditions
* - m, the previous row killed by p if it exists, A->mt otherwise
*/
-static int systolic_prevpiv(const dplasma_qrtree_t *qrtree, int k, int pivot, int start)
+static int systolic_prevpiv(const libhqr_tree_t *qrtree, int k, int pivot, int start)
{
int ls, lp, nextp;
int rpivot;
@@ -226,13 +227,13 @@ static int systolic_prevpiv(const dplasma_qrtree_t *qrtree, int k, int pivot, in
ls = systolic_gettype( qrtree, k, start );
lp = systolic_gettype( qrtree, k, pivot );
- if ( lp == DPLASMA_QR_KILLED_BY_TS )
+ if ( lp == LIBHQR_KILLED_BY_TS )
return mt;
myassert( lp >= ls );
switch( ls )
{
- case DPLASMA_QR_KILLED_BY_DISTTREE:
+ case LIBHQR_KILLED_BY_DISTTREE:
if ( pivot == k ) {
if ( start == pivot ) {
@@ -250,9 +251,9 @@ static int systolic_prevpiv(const dplasma_qrtree_t *qrtree, int k, int pivot, in
}
start = pivot;
- case DPLASMA_QR_KILLED_BY_LOCALTREE:
+ case LIBHQR_KILLED_BY_LOCALTREE:
- if ( lp > DPLASMA_QR_KILLED_BY_LOCALTREE ) {
+ if ( lp > LIBHQR_KILLED_BY_LOCALTREE ) {
if ( start == pivot ) {
nextp = start + (q-1) * p;
@@ -269,9 +270,9 @@ static int systolic_prevpiv(const dplasma_qrtree_t *qrtree, int k, int pivot, in
}
start = pivot;
- case DPLASMA_QR_KILLED_BY_TS:
+ case LIBHQR_KILLED_BY_TS:
/* Search for predecessor in TS tree */
- if ( lp > DPLASMA_QR_KILLED_BY_TS ) {
+ if ( lp > LIBHQR_KILLED_BY_TS ) {
if ( start == pivot ) {
nextp = mt - (mt - rpivot - 1)%pq - 1;
@@ -293,9 +294,9 @@ static int systolic_prevpiv(const dplasma_qrtree_t *qrtree, int k, int pivot, in
/**
*******************************************************************************
*
- * @ingroup dplasma
+ * @ingroup libhqr
*
- * dplasma_systolic_init - Creates the tree structure that will describes the
+ * libhqr_systolic_init - Creates the tree structure that will describes the
* operation performed during QR/LQ factorization with parameterized QR/LQ
* algorithms family. The trees created here correspond to systolic arrays with
* 1, 2, or 3 dimensions of flat trees.
@@ -338,8 +339,8 @@ static int systolic_prevpiv(const dplasma_qrtree_t *qrtree, int k, int pivot, in
*
*******************************************************************************
*
- * @sa dplasma_systolic_finalize
- * @sa dplasma_hqr_init
+ * @sa libhqr_systolic_finalize
+ * @sa libhqr_hqr_init
* @sa dplasma_zgeqrf_param
* @sa dplasma_cgeqrf_param
* @sa dplasma_dgeqrf_param
@@ -347,30 +348,29 @@ static int systolic_prevpiv(const dplasma_qrtree_t *qrtree, int k, int pivot, in
*
******************************************************************************/
int
-dplasma_systolic_init( dplasma_qrtree_t *qrtree,
- PLASMA_enum trans, tiled_matrix_desc_t *A,
+libhqr_systolic_init( libhqr_tree_t *qrtree,
+ libhqr_typefacto_e trans, libhqr_tiledesc_t *A,
int p, int q )
{
if (qrtree == NULL) {
- dplasma_error("dplasma_systolic_init", "illegal value of qrtree");
+ fprintf(stderr, "libhqr_systolic_init, illegal value of qrtree");
return -1;
}
- if ((trans != PlasmaNoTrans) &&
- (trans != PlasmaTrans) &&
- (trans != PlasmaConjTrans)) {
- dplasma_error("dplasma_systolic_init", "illegal value of trans");
+ if ((trans != LIBHQR_QR) &&
+ (trans != LIBHQR_LQ)) {
+ fprintf(stderr, "libhqr_systolic_init, illegal value of trans");
return -2;
}
if (A == NULL) {
- dplasma_error("dplasma_systolic_init", "illegal value of A");
+ fprintf(stderr, "libhqr_systolic_init, illegal value of A");
return -3;
}
if ( p < 0 ) {
- dplasma_error("dplasma_systolic_init", "illegal value of p");
+ fprintf(stderr, "libhqr_systolic_init, illegal value of p");
return -4;
}
if ( q < -1 ) {
- dplasma_error("dplasma_systolic_init", "illegal value of q");
+ fprintf(stderr, "libhqr_systolic_init, illegal value of q");
return -5;
}
@@ -382,11 +382,11 @@ dplasma_systolic_init( dplasma_qrtree_t *qrtree,
qrtree->nextpiv = systolic_nextpiv;
qrtree->prevpiv = systolic_prevpiv;
- qrtree->mt = (trans == PlasmaNoTrans) ? A->mt : A->nt;
- qrtree->nt = (trans == PlasmaNoTrans) ? A->nt : A->mt;
+ qrtree->mt = (trans == LIBHQR_QR) ? A->mt : A->nt;
+ qrtree->nt = (trans == LIBHQR_QR) ? A->nt : A->mt;
- qrtree->a = dplasma_imax( q, 1 );
- qrtree->p = dplasma_imax( p, 1 );
+ qrtree->a = libhqr_imax( q, 1 );
+ qrtree->p = libhqr_imax( p, 1 );
qrtree->args = NULL;
return 0;
@@ -397,8 +397,8 @@ dplasma_systolic_init( dplasma_qrtree_t *qrtree,
*
* @ingroup dplasma
*
- * dplasma_systolic_finalize - Cleans the qrtree data structure allocated by
- * call to dplasma_systolic_init().
+ * libhqr_systolic_finalize - Cleans the qrtree data structure allocated by
+ * call to libhqr_systolic_init().
*
*******************************************************************************
*
@@ -408,11 +408,11 @@ dplasma_systolic_init( dplasma_qrtree_t *qrtree,
*
*******************************************************************************
*
- * @sa dplasma_systolic_init
+ * @sa libhqr_systolic_init
*
******************************************************************************/
void
-dplasma_systolic_finalize( dplasma_qrtree_t *qrtree )
+libhqr_systolic_finalize( libhqr_tree_t *qrtree )
{
if ( qrtree->args != NULL) {
free( qrtree->args );
diff --git a/testings/testing_pivgen.c b/testings/testing_pivgen.c
index b86b441..135f61a 100644
--- a/testings/testing_pivgen.c
+++ b/testings/testing_pivgen.c
@@ -7,213 +7,144 @@
*
*/
-#include "common.h"
-#include "data_dist/matrix/two_dim_rectangle_cyclic.h"
+#include "libhqr.h"
+#include <stdlib.h>
+#include <stdio.h>
+#include <string.h>
+#include <math.h>
int main(int argc, char ** argv)
{
- parsec_context_t* parsec;
- dplasma_qrtree_t qrtree;
+ libhqr_tree_t qrtree;
+ libhqr_tiledesc_t matrix;
+ int alltreel[] = { 0, 1, 2, 3, 4 };
+ int alltreeh[] = { 0, 1, 2, 3, 4 };
+ int allP[] = { 3, 5, 7, 8 };
+ int allA[] = { 1, 2, 4, 7 };
+ int allM[] = { 1, 3, 4, 10, 17, 25, 128 };
+ int allN[] = { 1, 2, 5, 13, 26, 58 };
+ int nbtreel = 4;
+ int nbtreeh = 5;
+ int nbP = 4;
+ int nbA = 4;
+ int nbM = 7;
+ int nbN = 6;
int rc, ret = 0;
- int iparam[IPARAM_SIZEOF];
- char *dot_filename;
-
- /* Set defaults for non argv iparams */
- iparam_default_facto(iparam);
- iparam_default_ibnbmb(iparam, 1, 1, 1);
- iparam[IPARAM_LDA] = -'m';
- iparam[IPARAM_LDB] = -'m';
-
- /* Initialize PaRSEC */
- parsec = setup_parsec(argc, argv, iparam);
- PASTE_CODE_IPARAM_LOCALS(iparam);
-
- if (check) {
- tiled_matrix_desc_t *B;
- int alltreel[] = { 0, 1, 2, 3, 4 };
- int alltreeh[] = { 0, 1, 2, 3, 4 };
- int allP[] = { 3, 5, 7, 8 };
- int allA[] = { 1, 2, 4, 7 };
- int allM[] = { 1, 3, 4, 10, 17, 25, 128 };
- int allN[] = { 1, 2, 5, 13, 26, 58 };
- int nbtreel = 4;
- int nbtreeh = 5;
- int nbP = 4;
- int nbA = 4;
- int nbM = 7;
- int nbN = 6;
- int l, h, p, a, m, n, d, r;
- int done, todo;
- todo = 0;
- done = 0;
-
- /* HQR */
- todo += nbtreel * nbM * nbN * (2 * nbA - 1) * (1 + 2 * nbtreeh * nbP);
- /* systolic */
- todo += nbM * nbN * nbA * nbP;
-
- LDA = max(allM[ nbM-1 ], LDA);
- /* initializing matrix structure */
- PASTE_CODE_ALLOCATE_MATRIX(ddescA, 1,
- two_dim_block_cyclic, (&ddescA, matrix_ComplexDouble, matrix_Tile,
- nodes, rank, MB, NB, LDA, allN[ nbN-1 ], 0, 0,
- allM[ nbM-1 ], allN[ nbN-1 ], SMB, SNB, P));
-
- /*
- *
- * Tests for HQR code
- *
- */
- for( l=0; l<nbtreel; l++) {
- /* No High Level */
- h = 0; d = 0; p = -1;
- for( a=0; a<nbA; a++) {
- for( m=0; m<nbM; m++) {
- for( n=0; n<nbN; n++) {
- for( r=0; r<2; r++) {
- if (r==1 && a==1)
- continue;
-
- B = tiled_matrix_submatrix((tiled_matrix_desc_t*)&ddescA, 0, 0, allM[m], allN[n] );
- dplasma_hqr_init( &qrtree, PlasmaNoTrans, B, alltreel[l], 0, allA[a], -1, 0, r );
-
- rc = dplasma_qrtree_check( B, &qrtree );
- if (rc != 0) {
- fprintf(stderr, "-M %d -N %d --treel=%d --qr_a=%d --tsrr=%d FAILED(%d)\n",
- allM[m], allN[n], alltreel[l], allA[a], r, rc);
- ret |= 1;
- }
- dplasma_hqr_finalize( &qrtree );
- free(B);
-
- done++;
- printf("\r%6d / %6d", done, todo);
- }
- }
- }
- }
- /* With High level */
- for( d=0; d<2; d++) { /* Domino */
- if (d == 1 && alltreel[l] == DPLASMA_GREEDY1P_TREE)
- continue;
- for( h=0; h<nbtreeh; h++) {
- for( p=0; p<nbP; p++) {
- for( a=0; a<nbA; a++) {
- for( m=0; m<nbM; m++) {
- for( n=0; n<nbN; n++) {
- for( r=0; r<2; r++) {
- if (r==1 && a==1)
- continue;
-
- B = tiled_matrix_submatrix((tiled_matrix_desc_t*)&ddescA, 0, 0, allM[m], allN[n] );
- dplasma_hqr_init( &qrtree, PlasmaNoTrans, B, alltreel[l], alltreeh[h], allA[a], allP[p], d, r);
-
- rc = dplasma_qrtree_check( B, &qrtree );
- if (rc != 0) {
- fprintf(stderr, "-M %d -N %d --treel=%d --qr_a=%d --tsrr=%d --qr_p=%d --treeh=%d --domino=%d FAILED(%d)\n",
- allM[m], allN[n], alltreel[l], allA[a], r, allP[p], alltreeh[h], d, rc);
- ret |= 1;
- }
-
- dplasma_hqr_finalize( &qrtree );
- free(B);
-
- done++;
- printf("\r%6d / %6d", done, todo);
- }
- }
- }
- }
- }
- }
- }
- }
-
- /*
- *
- * Tests for systolic code
- *
- */
- /* With High level */
- for( p=0; p<nbP; p++) {
- for( a=0; a<nbA; a++) {
- for( m=0; m<nbM; m++) {
- for( n=0; n<nbN; n++) {
- B = tiled_matrix_submatrix((tiled_matrix_desc_t*)&ddescA, 0, 0, allM[m], allN[n] );
- dplasma_systolic_init( &qrtree, PlasmaNoTrans, B, allA[a], allP[p]);
-
- rc = dplasma_qrtree_check( B, &qrtree );
- if (rc != 0) {
- fprintf(stderr, "systolic: -M %d -N %d --qr_a=%d --qr_p=%d FAILED(%d)\n",
- allM[m], allN[n], allA[a], allP[p], rc);
- ret |= 1;
- }
-
- dplasma_systolic_finalize( &qrtree );
- free(B);
-
- done++;
- printf("\r%6d / %6d", done, todo);
- }
- }
- }
- }
-
- parsec_data_free(ddescA.mat);
- tiled_matrix_desc_destroy( (tiled_matrix_desc_t*)&ddescA);
-
- } else {
-
- LDA = max(M, LDA);
- /* initializing matrix structure */
- PASTE_CODE_ALLOCATE_MATRIX(ddescA, 1,
- two_dim_block_cyclic, (&ddescA, matrix_ComplexDouble, matrix_Tile,
- nodes, rank, MB, NB, LDA, N, 0, 0,
- M, N, SMB, SNB, P));
-
-#if defined(SYSTOLIC)
- dplasma_systolic_init( &qrtree,
- PlasmaNoTrans, (tiled_matrix_desc_t *)&ddescA,
- iparam[IPARAM_QR_HLVL_SZE],
- iparam[IPARAM_QR_TS_SZE] );
-#else
- dplasma_hqr_init( &qrtree,
- PlasmaNoTrans, (tiled_matrix_desc_t*)&ddescA,
- iparam[IPARAM_LOWLVL_TREE], iparam[IPARAM_HIGHLVL_TREE],
- iparam[IPARAM_QR_TS_SZE], iparam[IPARAM_QR_HLVL_SZE],
- iparam[IPARAM_QR_DOMINO], iparam[IPARAM_QR_TSRR] );
-#endif
-
- asprintf(&dot_filename, "tree-%dx%d-a%d-p%d-l%d-h%d-d%d.dot",
- M, N,
- iparam[IPARAM_QR_TS_SZE],
- iparam[IPARAM_QR_HLVL_SZE],
- iparam[IPARAM_LOWLVL_TREE],
- iparam[IPARAM_HIGHLVL_TREE],
- iparam[IPARAM_QR_DOMINO]);
-
- /*dplasma_qrtree_print_dag( (tiled_matrix_desc_t*)&ddescA, &qrtree, dot_filename );*/
- ret = dplasma_qrtree_check( (tiled_matrix_desc_t*)&ddescA, &qrtree );
-
- /* dplasma_qrtree_print_pivot( (tiled_matrix_desc_t*)&ddescA, &qrtree); */
- /* dplasma_qrtree_print_next_k( (tiled_matrix_desc_t*)&ddescA, &qrtree, 0); */
- /* dplasma_qrtree_print_prev_k( (tiled_matrix_desc_t*)&ddescA, &qrtree, 0); */
- /* dplasma_qrtree_print_nbgeqrt( (tiled_matrix_desc_t*)&ddescA, &qrtree ); */
- /* dplasma_qrtree_print_type ( (tiled_matrix_desc_t*)&ddescA, &qrtree ); */
-
-#if defined(SYSTOLIC)
- dplasma_systolic_finalize( &qrtree );
-#else
- dplasma_hqr_finalize( &qrtree );
-#endif
-
- free(dot_filename);
-
- parsec_data_free(ddescA.mat);
- tiled_matrix_desc_destroy( (tiled_matrix_desc_t*)&ddescA);
+ int nbnodes = 1;
+ int P = 1;
+ int l, h, p, a, m, n, d, r;
+ int done, todo;
+ todo = 0;
+ done = 0;
+
+ /* HQR */
+ todo += nbtreel * nbM * nbN * (2 * nbA - 1) * (1 + 2 * nbtreeh * nbP);
+ /* systolic */
+ todo += nbM * nbN * nbA * nbP;
+
+ /*
+ *
+ * Tests for HQR code
+ *
+ */
+ matrix.nodes = nbnodes;
+ matrix.p = P;
+ for( l=0; l<nbtreel; l++) {
+ /* No High Level */
+ h = 0; d = 0; p = -1;
+ for( a=0; a<nbA; a++) {
+ for( m=0; m<nbM; m++) {
+ matrix.mt = allM[m];
+ for( n=0; n<nbN; n++) {
+ matrix.nt = allN[n];
+ for( r=0; r<2; r++) {
+ if (r==1 && a==1)
+ continue;
+
+ libhqr_hqr_init( &qrtree, LIBHQR_QR, &matrix, alltreel[l], 0, allA[a], -1, 0, r );
+
+ rc = libhqr_tree_check( &matrix, &qrtree );
+ if (rc != 0) {
+ fprintf(stderr, "-M %d -N %d --treel=%d --qr_a=%d --tsrr=%d FAILED(%d)\n",
+ allM[m], allN[n], alltreel[l], allA[a], r, rc);
+ ret |= 1;
+ }
+ libhqr_hqr_finalize( &qrtree );
+
+ done++;
+ printf("\r%6d / %6d", done, todo);
+ }
+ }
+ }
+ }
+ /* With High level */
+ matrix.nodes = nbnodes;
+ for( d=0; d<2; d++) { /* Domino */
+ if (d == 1 && alltreel[l] == LIBHQR_GREEDY1P_TREE)
+ continue;
+ for( h=0; h<nbtreeh; h++) {
+ for( p=0; p<nbP; p++) {
+ matrix.p = allP[p];
+ for( a=0; a<nbA; a++) {
+ for( m=0; m<nbM; m++) {
+ matrix.mt = allM[m];
+ for( n=0; n<nbN; n++) {
+ matrix.nt = allN[n];
+ for( r=0; r<2; r++) {
+ if (r==1 && a==1)
+ continue;
+ libhqr_hqr_init( &qrtree, LIBHQR_QR, &matrix, alltreel[l], alltreeh[h], allA[a], allP[p], d, r);
+
+ rc = libhqr_tree_check( &matrix, &qrtree );
+ if (rc != 0) {
+ fprintf(stderr, "-M %d -N %d --treel=%d --qr_a=%d --tsrr=%d --qr_p=%d --treeh=%d --domino=%d FAILED(%d)\n",
+ allM[m], allN[n], alltreel[l], allA[a], r, allP[p], alltreeh[h], d, rc);
+ ret |= 1;
+ }
+
+ libhqr_hqr_finalize( &qrtree );
+
+ done++;
+ printf("\r%6d / %6d", done, todo);
+ }
+ }
+ }
+ }
+ }
+ }
+ }
}
- cleanup_parsec(parsec, iparam);
+ /*
+ *
+ * Tests for systolic code
+ *
+ */
+ /* With High level */
+ for( p=0; p<nbP; p++) {
+ matrix.p = allP[p];
+ for( a=0; a<nbA; a++) {
+ for( m=0; m<nbM; m++) {
+ matrix.mt = allM[m];
+ for( n=0; n<nbN; n++) {
+ matrix.nt = allN[n];
+ libhqr_systolic_init( &qrtree, LIBHQR_QR, &matrix, allA[a], allP[p]);
+
+ rc = libhqr_tree_check( &matrix, &qrtree );
+ if (rc != 0) {
+ fprintf(stderr, "systolic: -M %d -N %d --qr_a=%d --qr_p=%d FAILED(%d)\n",
+ allM[m], allN[n], allA[a], allP[p], rc);
+ ret |= 1;
+ }
+
+ libhqr_systolic_finalize( &qrtree );
+
+ done++;
+ printf("\r%6d / %6d", done, todo);
+ }
+ }
+ }
+ }
if ( ret == 0 )
return EXIT_SUCCESS;
--
GitLab