Mentions légales du service

Skip to content
Snippets Groups Projects
Commit 401a7513 authored by Mathieu Faverge's avatar Mathieu Faverge
Browse files

Remove COMPLEX/REAL from coreblas

parent 88409dc2
No related branches found
No related tags found
1 merge request!84Fix issue #45 - Remove COMPLEX/REAL
This commit is part of merge request !84. Comments created here will be created in the context of that merge request.
Hide whitespace changes
Inline Side-by-side
Showing
with 33 additions and 77 deletions
coreblas/compute/core_zgessq.c
+ 1
− 4
View file @ 401a7513
...@@ -23,8 +23,6 @@ ...@@ -23,8 +23,6 @@
#include "coreblas/lapacke.h" #include "coreblas/lapacke.h"
#include "coreblas.h" #include "coreblas.h"
#define COMPLEX
#define UPDATE( __nb, __value ) \ #define UPDATE( __nb, __value ) \
if (__value != 0. ){ \ if (__value != 0. ){ \
if ( *scale < __value ) { \ if ( *scale < __value ) { \
...@@ -89,7 +87,6 @@ ...@@ -89,7 +87,6 @@
* \retval -k, the k-th argument had an illegal value * \retval -k, the k-th argument had an illegal value
* *
*/ */
int CORE_zgessq(int M, int N, int CORE_zgessq(int M, int N,
const MORSE_Complex64_t *A, int LDA, const MORSE_Complex64_t *A, int LDA,
double *scale, double *sumsq) double *scale, double *sumsq)
...@@ -104,7 +101,7 @@ int CORE_zgessq(int M, int N, ...@@ -104,7 +101,7 @@ int CORE_zgessq(int M, int N,
tmp = fabs(*ptr); tmp = fabs(*ptr);
UPDATE( 1., tmp ); UPDATE( 1., tmp );
#ifdef COMPLEX #if defined(PRECISION_z) || defined(PRECISION_c)
ptr++; ptr++;
tmp = fabs(*ptr); tmp = fabs(*ptr);
UPDATE( 1., tmp ); UPDATE( 1., tmp );
......
coreblas/compute/core_zhemm.c
+ 0
− 6
View file @ 401a7513
...@@ -25,9 +25,6 @@ ...@@ -25,9 +25,6 @@
*/ */
#include "coreblas.h" #include "coreblas.h"
#undef REAL
#define COMPLEX
#ifdef COMPLEX
/** /**
* *
* @ingroup CORE_MORSE_Complex64_t * @ingroup CORE_MORSE_Complex64_t
...@@ -47,6 +44,3 @@ void CORE_zhemm(MORSE_enum side, MORSE_enum uplo, ...@@ -47,6 +44,3 @@ void CORE_zhemm(MORSE_enum side, MORSE_enum uplo,
B, LDB, B, LDB,
CBLAS_SADDR(beta), C, LDC); CBLAS_SADDR(beta), C, LDC);
} }
#endif
coreblas/compute/core_zher2k.c
+ 0
− 6
View file @ 401a7513
...@@ -25,9 +25,6 @@ ...@@ -25,9 +25,6 @@
*/ */
#include "coreblas.h" #include "coreblas.h"
#undef REAL
#define COMPLEX
#ifdef COMPLEX
/** /**
* *
* @ingroup CORE_MORSE_Complex64_t * @ingroup CORE_MORSE_Complex64_t
...@@ -46,6 +43,3 @@ void CORE_zher2k(MORSE_enum uplo, MORSE_enum trans, ...@@ -46,6 +43,3 @@ void CORE_zher2k(MORSE_enum uplo, MORSE_enum trans,
CBLAS_SADDR(alpha), A, LDA, B, LDB, CBLAS_SADDR(alpha), A, LDA, B, LDB,
beta, C, LDC); beta, C, LDC);
} }
#endif
coreblas/compute/core_zherk.c
+ 0
− 6
View file @ 401a7513
...@@ -25,9 +25,6 @@ ...@@ -25,9 +25,6 @@
*/ */
#include "coreblas.h" #include "coreblas.h"
#undef REAL
#define COMPLEX
#ifdef COMPLEX
/** /**
* *
* @ingroup CORE_MORSE_Complex64_t * @ingroup CORE_MORSE_Complex64_t
...@@ -45,6 +42,3 @@ void CORE_zherk(MORSE_enum uplo, MORSE_enum trans, ...@@ -45,6 +42,3 @@ void CORE_zherk(MORSE_enum uplo, MORSE_enum trans,
alpha, A, LDA, alpha, A, LDA,
beta, C, LDC); beta, C, LDC);
} }
#endif
coreblas/compute/core_zhessq.c
+ 4
− 6
View file @ 401a7513
...@@ -23,8 +23,6 @@ ...@@ -23,8 +23,6 @@
#include "coreblas/lapacke.h" #include "coreblas/lapacke.h"
#include "coreblas.h" #include "coreblas.h"
#define COMPLEX
#define UPDATE( __nb, __value ) \ #define UPDATE( __nb, __value ) \
if (__value != 0. ){ \ if (__value != 0. ){ \
if ( *scale < __value ) { \ if ( *scale < __value ) { \
...@@ -113,7 +111,7 @@ int CORE_zhessq(MORSE_enum uplo, int N, ...@@ -113,7 +111,7 @@ int CORE_zhessq(MORSE_enum uplo, int N,
tmp = fabs(*ptr); tmp = fabs(*ptr);
UPDATE( 2., tmp ); UPDATE( 2., tmp );
#ifdef COMPLEX #if defined(PRECISION_z) || defined(PRECISION_c)
ptr++; ptr++;
tmp = fabs(*ptr); tmp = fabs(*ptr);
UPDATE( 2., tmp ); UPDATE( 2., tmp );
...@@ -124,7 +122,7 @@ int CORE_zhessq(MORSE_enum uplo, int N, ...@@ -124,7 +122,7 @@ int CORE_zhessq(MORSE_enum uplo, int N,
tmp = fabs(*ptr); tmp = fabs(*ptr);
UPDATE( 1., tmp ); UPDATE( 1., tmp );
#ifdef COMPLEX #if defined(PRECISION_z) || defined(PRECISION_c)
ptr++; ptr++;
#endif #endif
} }
...@@ -138,7 +136,7 @@ int CORE_zhessq(MORSE_enum uplo, int N, ...@@ -138,7 +136,7 @@ int CORE_zhessq(MORSE_enum uplo, int N,
UPDATE( 1., tmp ); UPDATE( 1., tmp );
ptr++; ptr++;
#ifdef COMPLEX #if defined(PRECISION_z) || defined(PRECISION_c)
ptr++; ptr++;
#endif #endif
...@@ -147,7 +145,7 @@ int CORE_zhessq(MORSE_enum uplo, int N, ...@@ -147,7 +145,7 @@ int CORE_zhessq(MORSE_enum uplo, int N,
tmp = fabs(*ptr); tmp = fabs(*ptr);
UPDATE( 2., tmp ); UPDATE( 2., tmp );
#ifdef COMPLEX #if defined(PRECISION_z) || defined(PRECISION_c)
ptr++; ptr++;
tmp = fabs(*ptr); tmp = fabs(*ptr);
UPDATE( 2., tmp ); UPDATE( 2., tmp );
......
coreblas/compute/core_zplghe.c
+ 5
− 8
View file @ 401a7513
...@@ -25,9 +25,6 @@ ...@@ -25,9 +25,6 @@
*/ */
#include "coreblas.h" #include "coreblas.h"
#define COMPLEX
#undef REAL
/* /*
Rnd64seed is a global variable but it doesn't spoil thread safety. All matrix Rnd64seed is a global variable but it doesn't spoil thread safety. All matrix
generating threads only read Rnd64seed. It is safe to set Rnd64seed before generating threads only read Rnd64seed. It is safe to set Rnd64seed before
...@@ -41,7 +38,7 @@ ...@@ -41,7 +38,7 @@
#define RndF_Mul 5.4210108624275222e-20f #define RndF_Mul 5.4210108624275222e-20f
#define RndD_Mul 5.4210108624275222e-20 #define RndD_Mul 5.4210108624275222e-20
#ifdef COMPLEX #if defined(PRECISION_z) || defined(PRECISION_c)
#define NBELEM 2 #define NBELEM 2
#else #else
#define NBELEM 1 #define NBELEM 1
...@@ -87,7 +84,7 @@ void CORE_zplghe( double bump, int m, int n, MORSE_Complex64_t *A, int lda, ...@@ -87,7 +84,7 @@ void CORE_zplghe( double bump, int m, int n, MORSE_Complex64_t *A, int lda,
for (i = j; i < m; i++) { for (i = j; i < m; i++) {
*tmp = 0.5f - ran * RndF_Mul; *tmp = 0.5f - ran * RndF_Mul;
ran = Rnd64_A * ran + Rnd64_C; ran = Rnd64_A * ran + Rnd64_C;
#ifdef COMPLEX #if defined(PRECISION_z) || defined(PRECISION_c)
*tmp += I*(0.5f - ran * RndF_Mul); *tmp += I*(0.5f - ran * RndF_Mul);
ran = Rnd64_A * ran + Rnd64_C; ran = Rnd64_A * ran + Rnd64_C;
#endif #endif
...@@ -98,7 +95,7 @@ void CORE_zplghe( double bump, int m, int n, MORSE_Complex64_t *A, int lda, ...@@ -98,7 +95,7 @@ void CORE_zplghe( double bump, int m, int n, MORSE_Complex64_t *A, int lda,
} }
for (j = 0; j < n; j++) { for (j = 0; j < n; j++) {
#ifdef COMPLEX #if defined(PRECISION_z) || defined(PRECISION_c)
A[j+j*lda] += bump - I*cimag( A[j+j*lda] ); A[j+j*lda] += bump - I*cimag( A[j+j*lda] );
#else #else
A[j+j*lda] += bump; A[j+j*lda] += bump;
...@@ -119,7 +116,7 @@ void CORE_zplghe( double bump, int m, int n, MORSE_Complex64_t *A, int lda, ...@@ -119,7 +116,7 @@ void CORE_zplghe( double bump, int m, int n, MORSE_Complex64_t *A, int lda,
for (i = 0; i < m; i++) { for (i = 0; i < m; i++) {
*tmp = 0.5f - ran * RndF_Mul; *tmp = 0.5f - ran * RndF_Mul;
ran = Rnd64_A * ran + Rnd64_C; ran = Rnd64_A * ran + Rnd64_C;
#ifdef COMPLEX #if defined(PRECISION_z) || defined(PRECISION_c)
*tmp += I*(0.5f - ran * RndF_Mul); *tmp += I*(0.5f - ran * RndF_Mul);
ran = Rnd64_A * ran + Rnd64_C; ran = Rnd64_A * ran + Rnd64_C;
#endif #endif
...@@ -142,7 +139,7 @@ void CORE_zplghe( double bump, int m, int n, MORSE_Complex64_t *A, int lda, ...@@ -142,7 +139,7 @@ void CORE_zplghe( double bump, int m, int n, MORSE_Complex64_t *A, int lda,
for (j = 0; j < n; j++) { for (j = 0; j < n; j++) {
A[j*lda+i] = 0.5f - ran * RndF_Mul; A[j*lda+i] = 0.5f - ran * RndF_Mul;
ran = Rnd64_A * ran + Rnd64_C; ran = Rnd64_A * ran + Rnd64_C;
#ifdef COMPLEX #if defined(PRECISION_z) || defined(PRECISION_c)
A[j*lda+i] -= I*(0.5f - ran * RndF_Mul); A[j*lda+i] -= I*(0.5f - ran * RndF_Mul);
ran = Rnd64_A * ran + Rnd64_C; ran = Rnd64_A * ran + Rnd64_C;
#endif #endif
......
coreblas/compute/core_zplgsy.c
+ 4
− 7
View file @ 401a7513
...@@ -25,9 +25,6 @@ ...@@ -25,9 +25,6 @@
*/ */
#include "coreblas.h" #include "coreblas.h"
#define COMPLEX
#undef REAL
/* /*
Rnd64seed is a global variable but it doesn't spoil thread safety. All matrix Rnd64seed is a global variable but it doesn't spoil thread safety. All matrix
generating threads only read Rnd64seed. It is safe to set Rnd64seed before generating threads only read Rnd64seed. It is safe to set Rnd64seed before
...@@ -41,7 +38,7 @@ ...@@ -41,7 +38,7 @@
#define RndF_Mul 5.4210108624275222e-20f #define RndF_Mul 5.4210108624275222e-20f
#define RndD_Mul 5.4210108624275222e-20 #define RndD_Mul 5.4210108624275222e-20
#ifdef COMPLEX #if defined(PRECISION_z) || defined(PRECISION_c)
#define NBELEM 2 #define NBELEM 2
#else #else
#define NBELEM 1 #define NBELEM 1
...@@ -88,7 +85,7 @@ void CORE_zplgsy( MORSE_Complex64_t bump, int m, int n, MORSE_Complex64_t *A, in ...@@ -88,7 +85,7 @@ void CORE_zplgsy( MORSE_Complex64_t bump, int m, int n, MORSE_Complex64_t *A, in
for (i = j; i < m; i++) { for (i = j; i < m; i++) {
*tmp = 0.5f - ran * RndF_Mul; *tmp = 0.5f - ran * RndF_Mul;
ran = Rnd64_A * ran + Rnd64_C; ran = Rnd64_A * ran + Rnd64_C;
#ifdef COMPLEX #if defined(PRECISION_z) || defined(PRECISION_c)
*tmp += I*(0.5f - ran * RndF_Mul); *tmp += I*(0.5f - ran * RndF_Mul);
ran = Rnd64_A * ran + Rnd64_C; ran = Rnd64_A * ran + Rnd64_C;
#endif #endif
...@@ -116,7 +113,7 @@ void CORE_zplgsy( MORSE_Complex64_t bump, int m, int n, MORSE_Complex64_t *A, in ...@@ -116,7 +113,7 @@ void CORE_zplgsy( MORSE_Complex64_t bump, int m, int n, MORSE_Complex64_t *A, in
for (i = 0; i < m; i++) { for (i = 0; i < m; i++) {
*tmp = 0.5f - ran * RndF_Mul; *tmp = 0.5f - ran * RndF_Mul;
ran = Rnd64_A * ran + Rnd64_C; ran = Rnd64_A * ran + Rnd64_C;
#ifdef COMPLEX #if defined(PRECISION_z) || defined(PRECISION_c)
*tmp += I*(0.5f - ran * RndF_Mul); *tmp += I*(0.5f - ran * RndF_Mul);
ran = Rnd64_A * ran + Rnd64_C; ran = Rnd64_A * ran + Rnd64_C;
#endif #endif
...@@ -139,7 +136,7 @@ void CORE_zplgsy( MORSE_Complex64_t bump, int m, int n, MORSE_Complex64_t *A, in ...@@ -139,7 +136,7 @@ void CORE_zplgsy( MORSE_Complex64_t bump, int m, int n, MORSE_Complex64_t *A, in
for (j = 0; j < n; j++) { for (j = 0; j < n; j++) {
A[j*lda+i] = 0.5f - ran * RndF_Mul; A[j*lda+i] = 0.5f - ran * RndF_Mul;
ran = Rnd64_A * ran + Rnd64_C; ran = Rnd64_A * ran + Rnd64_C;
#ifdef COMPLEX #if defined(PRECISION_z) || defined(PRECISION_c)
A[j*lda+i] += I*(0.5f - ran * RndF_Mul); A[j*lda+i] += I*(0.5f - ran * RndF_Mul);
ran = Rnd64_A * ran + Rnd64_C; ran = Rnd64_A * ran + Rnd64_C;
#endif #endif
......
coreblas/compute/core_zplrnt.c
+ 2
− 5
View file @ 401a7513
...@@ -25,9 +25,6 @@ ...@@ -25,9 +25,6 @@
*/ */
#include "coreblas.h" #include "coreblas.h"
#define COMPLEX
#undef REAL
/* /*
Rnd64seed is a global variable but it doesn't spoil thread safety. All matrix Rnd64seed is a global variable but it doesn't spoil thread safety. All matrix
generating threads only read Rnd64seed. It is safe to set Rnd64seed before generating threads only read Rnd64seed. It is safe to set Rnd64seed before
...@@ -41,7 +38,7 @@ ...@@ -41,7 +38,7 @@
#define RndF_Mul 5.4210108624275222e-20f #define RndF_Mul 5.4210108624275222e-20f
#define RndD_Mul 5.4210108624275222e-20 #define RndD_Mul 5.4210108624275222e-20
#ifdef COMPLEX #if defined(PRECISION_z) || defined(PRECISION_c)
#define NBELEM 2 #define NBELEM 2
#else #else
#define NBELEM 1 #define NBELEM 1
...@@ -82,7 +79,7 @@ void CORE_zplrnt( int m, int n, MORSE_Complex64_t *A, int lda, ...@@ -82,7 +79,7 @@ void CORE_zplrnt( int m, int n, MORSE_Complex64_t *A, int lda,
for (i = 0; i < m; ++i) { for (i = 0; i < m; ++i) {
*tmp = 0.5f - ran * RndF_Mul; *tmp = 0.5f - ran * RndF_Mul;
ran = Rnd64_A * ran + Rnd64_C; ran = Rnd64_A * ran + Rnd64_C;
#ifdef COMPLEX #if defined(PRECISION_z) || defined(PRECISION_c)
*tmp += I*(0.5f - ran * RndF_Mul); *tmp += I*(0.5f - ran * RndF_Mul);
ran = Rnd64_A * ran + Rnd64_C; ran = Rnd64_A * ran + Rnd64_C;
#endif #endif
......
coreblas/compute/core_zsyssq.c
+ 4
− 6
View file @ 401a7513
...@@ -23,8 +23,6 @@ ...@@ -23,8 +23,6 @@
#include "coreblas/lapacke.h" #include "coreblas/lapacke.h"
#include "coreblas.h" #include "coreblas.h"
#define COMPLEX
#define UPDATE( __nb, __value ) \ #define UPDATE( __nb, __value ) \
if (__value != 0. ){ \ if (__value != 0. ){ \
if ( *scale < __value ) { \ if ( *scale < __value ) { \
...@@ -113,7 +111,7 @@ int CORE_zsyssq(MORSE_enum uplo, int N, ...@@ -113,7 +111,7 @@ int CORE_zsyssq(MORSE_enum uplo, int N,
tmp = fabs(*ptr); tmp = fabs(*ptr);
UPDATE( 2., tmp ); UPDATE( 2., tmp );
#ifdef COMPLEX #if defined(PRECISION_z) || defined(PRECISION_c)
ptr++; ptr++;
tmp = fabs(*ptr); tmp = fabs(*ptr);
UPDATE( 2., tmp ); UPDATE( 2., tmp );
...@@ -124,7 +122,7 @@ int CORE_zsyssq(MORSE_enum uplo, int N, ...@@ -124,7 +122,7 @@ int CORE_zsyssq(MORSE_enum uplo, int N,
tmp = fabs(*ptr); tmp = fabs(*ptr);
UPDATE( 1., tmp ); UPDATE( 1., tmp );
#ifdef COMPLEX #if defined(PRECISION_z) || defined(PRECISION_c)
ptr++; ptr++;
tmp = fabs(*ptr); tmp = fabs(*ptr);
UPDATE( 1., tmp ); UPDATE( 1., tmp );
...@@ -140,7 +138,7 @@ int CORE_zsyssq(MORSE_enum uplo, int N, ...@@ -140,7 +138,7 @@ int CORE_zsyssq(MORSE_enum uplo, int N,
UPDATE( 1., tmp ); UPDATE( 1., tmp );
ptr++; ptr++;
#ifdef COMPLEX #if defined(PRECISION_z) || defined(PRECISION_c)
tmp = fabs(*ptr); tmp = fabs(*ptr);
UPDATE( 1., tmp ); UPDATE( 1., tmp );
ptr++; ptr++;
...@@ -151,7 +149,7 @@ int CORE_zsyssq(MORSE_enum uplo, int N, ...@@ -151,7 +149,7 @@ int CORE_zsyssq(MORSE_enum uplo, int N,
tmp = fabs(*ptr); tmp = fabs(*ptr);
UPDATE( 2., tmp ); UPDATE( 2., tmp );
#ifdef COMPLEX #if defined(PRECISION_z) || defined(PRECISION_c)
ptr++; ptr++;
tmp = fabs(*ptr); tmp = fabs(*ptr);
UPDATE( 2., tmp ); UPDATE( 2., tmp );
......
coreblas/compute/core_ztrssq.c
+ 2
− 4
View file @ 401a7513
...@@ -23,8 +23,6 @@ ...@@ -23,8 +23,6 @@
#include "coreblas/lapacke.h" #include "coreblas/lapacke.h"
#include "coreblas.h" #include "coreblas.h"
#define COMPLEX
#define UPDATE( __nb, __value ) \ #define UPDATE( __nb, __value ) \
if (__value != 0. ){ \ if (__value != 0. ){ \
if ( *scale < __value ) { \ if ( *scale < __value ) { \
...@@ -115,7 +113,7 @@ int CORE_ztrssq(MORSE_enum uplo, MORSE_enum diag, int M, int N, ...@@ -115,7 +113,7 @@ int CORE_ztrssq(MORSE_enum uplo, MORSE_enum diag, int M, int N,
tmp = fabs(*ptr); tmp = fabs(*ptr);
UPDATE( 1., tmp ); UPDATE( 1., tmp );
#ifdef COMPLEX #if defined(PRECISION_z) || defined(PRECISION_c)
ptr++; ptr++;
tmp = fabs(*ptr); tmp = fabs(*ptr);
UPDATE( 1., tmp ); UPDATE( 1., tmp );
...@@ -133,7 +131,7 @@ int CORE_ztrssq(MORSE_enum uplo, MORSE_enum diag, int M, int N, ...@@ -133,7 +131,7 @@ int CORE_ztrssq(MORSE_enum uplo, MORSE_enum diag, int M, int N,
tmp = fabs(*ptr); tmp = fabs(*ptr);
UPDATE( 1., tmp ); UPDATE( 1., tmp );
#ifdef COMPLEX #if defined(PRECISION_z) || defined(PRECISION_c)
ptr++; ptr++;
tmp = fabs(*ptr); tmp = fabs(*ptr);
UPDATE( 1., tmp ); UPDATE( 1., tmp );
......
coreblas/compute/core_ztslqt.c
+ 2
− 4
View file @ 401a7513
...@@ -25,8 +25,6 @@ ...@@ -25,8 +25,6 @@
*/ */
#include "coreblas/lapacke.h" #include "coreblas/lapacke.h"
#include "coreblas.h" #include "coreblas.h"
#undef REAL
#define COMPLEX
/** /**
* *
...@@ -142,7 +140,7 @@ int CORE_ztslqt(int M, int N, int IB, ...@@ -142,7 +140,7 @@ int CORE_ztslqt(int M, int N, int IB,
/* /*
* Generate elementary reflector H( II*IB+I ) to annihilate A( II*IB+I, II*IB+I:N ). * Generate elementary reflector H( II*IB+I ) to annihilate A( II*IB+I, II*IB+I:N ).
*/ */
#ifdef COMPLEX #if defined(PRECISION_z) || defined(PRECISION_c)
LAPACKE_zlacgv_work(N, &A2[ii+i], LDA2); LAPACKE_zlacgv_work(N, &A2[ii+i], LDA2);
LAPACKE_zlacgv_work(1, &A1[LDA1*(ii+i)+ii+i], LDA1); LAPACKE_zlacgv_work(1, &A1[LDA1*(ii+i)+ii+i], LDA1);
#endif #endif
...@@ -184,7 +182,7 @@ int CORE_ztslqt(int M, int N, int IB, ...@@ -184,7 +182,7 @@ int CORE_ztslqt(int M, int N, int IB,
CBLAS_SADDR(alpha), &A2[ii], LDA2, CBLAS_SADDR(alpha), &A2[ii], LDA2,
&A2[ii+i], LDA2, &A2[ii+i], LDA2,
CBLAS_SADDR(zzero), &T[LDT*(ii+i)], 1); CBLAS_SADDR(zzero), &T[LDT*(ii+i)], 1);
#ifdef COMPLEX #if defined(PRECISION_z) || defined(PRECISION_c)
LAPACKE_zlacgv_work(N, &A2[ii+i], LDA2 ); LAPACKE_zlacgv_work(N, &A2[ii+i], LDA2 );
LAPACKE_zlacgv_work(1, &A1[LDA1*(ii+i)+ii+i], LDA1 ); LAPACKE_zlacgv_work(1, &A1[LDA1*(ii+i)+ii+i], LDA1 );
#endif #endif
......
coreblas/compute/core_ztsqrt.c
+ 3
− 5
View file @ 401a7513
...@@ -25,8 +25,6 @@ ...@@ -25,8 +25,6 @@
*/ */
#include "coreblas/lapacke.h" #include "coreblas/lapacke.h"
#include "coreblas.h" #include "coreblas.h"
#undef REAL
#define COMPLEX
/** /**
* *
...@@ -144,7 +142,7 @@ int CORE_ztsqrt(int M, int N, int IB, ...@@ -144,7 +142,7 @@ int CORE_ztsqrt(int M, int N, int IB,
sb-i-1, sb-i-1,
&A1[LDA1*(ii+i+1)+(ii+i)], LDA1, &A1[LDA1*(ii+i+1)+(ii+i)], LDA1,
WORK, 1); WORK, 1);
#ifdef COMPLEX #if defined(PRECISION_z) || defined(PRECISION_c)
LAPACKE_zlacgv_work(sb-i-1, WORK, 1); LAPACKE_zlacgv_work(sb-i-1, WORK, 1);
#endif #endif
cblas_zgemv( cblas_zgemv(
...@@ -153,14 +151,14 @@ int CORE_ztsqrt(int M, int N, int IB, ...@@ -153,14 +151,14 @@ int CORE_ztsqrt(int M, int N, int IB,
CBLAS_SADDR(zone), &A2[LDA2*(ii+i+1)], LDA2, CBLAS_SADDR(zone), &A2[LDA2*(ii+i+1)], LDA2,
&A2[LDA2*(ii+i)], 1, &A2[LDA2*(ii+i)], 1,
CBLAS_SADDR(zone), WORK, 1); CBLAS_SADDR(zone), WORK, 1);
#ifdef COMPLEX #if defined(PRECISION_z) || defined(PRECISION_c)
LAPACKE_zlacgv_work(sb-i-1, WORK, 1 ); LAPACKE_zlacgv_work(sb-i-1, WORK, 1 );
#endif #endif
cblas_zaxpy( cblas_zaxpy(
sb-i-1, CBLAS_SADDR(alpha), sb-i-1, CBLAS_SADDR(alpha),
WORK, 1, WORK, 1,
&A1[LDA1*(ii+i+1)+ii+i], LDA1); &A1[LDA1*(ii+i+1)+ii+i], LDA1);
#ifdef COMPLEX #if defined(PRECISION_z) || defined(PRECISION_c)
LAPACKE_zlacgv_work(sb-i-1, WORK, 1 ); LAPACKE_zlacgv_work(sb-i-1, WORK, 1 );
#endif #endif
cblas_zgerc( cblas_zgerc(
......
coreblas/compute/core_zttlqt.c
+ 3
− 5
View file @ 401a7513
...@@ -25,8 +25,6 @@ ...@@ -25,8 +25,6 @@
*/ */
#include "coreblas/lapacke.h" #include "coreblas/lapacke.h"
#include "coreblas.h" #include "coreblas.h"
#undef REAL
#define COMPLEX
/** /**
* *
...@@ -111,7 +109,7 @@ int CORE_zttlqt(int M, int N, int IB, ...@@ -111,7 +109,7 @@ int CORE_zttlqt(int M, int N, int IB,
{ {
static MORSE_Complex64_t zone = 1.0; static MORSE_Complex64_t zone = 1.0;
static MORSE_Complex64_t zzero = 0.0; static MORSE_Complex64_t zzero = 0.0;
#ifdef COMPLEX #if defined(PRECISION_z) || defined(PRECISION_c)
static int ione = 1; static int ione = 1;
#endif #endif
...@@ -154,7 +152,7 @@ int CORE_zttlqt(int M, int N, int IB, ...@@ -154,7 +152,7 @@ int CORE_zttlqt(int M, int N, int IB,
/* /*
* Generate elementary reflector H( II*IB+I ) to annihilate A( II*IB+I, II*IB+I:M ). * Generate elementary reflector H( II*IB+I ) to annihilate A( II*IB+I, II*IB+I:M ).
*/ */
#ifdef COMPLEX #if defined(PRECISION_z) || defined(PRECISION_c)
LAPACKE_zlacgv_work(ni, &A2[j], LDA2); LAPACKE_zlacgv_work(ni, &A2[j], LDA2);
LAPACKE_zlacgv_work(ione, &A1[LDA1*j+j], LDA1); LAPACKE_zlacgv_work(ione, &A1[LDA1*j+j], LDA1);
#endif #endif
...@@ -216,7 +214,7 @@ int CORE_zttlqt(int M, int N, int IB, ...@@ -216,7 +214,7 @@ int CORE_zttlqt(int M, int N, int IB,
} }
#ifdef COMPLEX #if defined(PRECISION_z) || defined(PRECISION_c)
LAPACKE_zlacgv_work(ni, &A2[j], LDA2 ); LAPACKE_zlacgv_work(ni, &A2[j], LDA2 );
LAPACKE_zlacgv_work(ione, &A1[LDA1*j+j], LDA1 ); LAPACKE_zlacgv_work(ione, &A1[LDA1*j+j], LDA1 );
#endif #endif
......
coreblas/compute/core_zttqrt.c
+ 3
− 5
View file @ 401a7513
...@@ -25,8 +25,6 @@ ...@@ -25,8 +25,6 @@
*/ */
#include "coreblas/lapacke.h" #include "coreblas/lapacke.h"
#include "coreblas.h" #include "coreblas.h"
#undef REAL
#define COMPLEX
/** /**
* *
...@@ -166,7 +164,7 @@ int CORE_zttqrt(int M, int N, int IB, ...@@ -166,7 +164,7 @@ int CORE_zttqrt(int M, int N, int IB,
&A1[LDA1*(j+1)+j], LDA1, &A1[LDA1*(j+1)+j], LDA1,
WORK, 1); WORK, 1);
#ifdef COMPLEX #if defined(PRECISION_z) || defined(PRECISION_c)
LAPACKE_zlacgv_work(ni, WORK, 1); LAPACKE_zlacgv_work(ni, WORK, 1);
#endif #endif
cblas_zgemv( cblas_zgemv(
...@@ -175,7 +173,7 @@ int CORE_zttqrt(int M, int N, int IB, ...@@ -175,7 +173,7 @@ int CORE_zttqrt(int M, int N, int IB,
CBLAS_SADDR(zone), &A2[LDA2*(j+1)], LDA2, CBLAS_SADDR(zone), &A2[LDA2*(j+1)], LDA2,
&A2[LDA2*j], 1, &A2[LDA2*j], 1,
CBLAS_SADDR(zone), WORK, 1); CBLAS_SADDR(zone), WORK, 1);
#ifdef COMPLEX #if defined(PRECISION_z) || defined(PRECISION_c)
LAPACKE_zlacgv_work(ni, WORK, 1); LAPACKE_zlacgv_work(ni, WORK, 1);
#endif #endif
alpha = -conj(TAU[j]); alpha = -conj(TAU[j]);
...@@ -183,7 +181,7 @@ int CORE_zttqrt(int M, int N, int IB, ...@@ -183,7 +181,7 @@ int CORE_zttqrt(int M, int N, int IB,
ni, CBLAS_SADDR(alpha), ni, CBLAS_SADDR(alpha),
WORK, 1, WORK, 1,
&A1[LDA1*(j+1)+j], LDA1); &A1[LDA1*(j+1)+j], LDA1);
#ifdef COMPLEX #if defined(PRECISION_z) || defined(PRECISION_c)
LAPACKE_zlacgv_work(ni, WORK, 1); LAPACKE_zlacgv_work(ni, WORK, 1);
#endif #endif
cblas_zgerc( cblas_zgerc(
......
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment