floraachy
/
OpenBLAS
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Disambiguate whilelt

This commit is contained in:
Chris Sidebottom 2023-07-25 14:50:38 +01:00
parent f971ef55f2
commit 24586bc4ff
21 changed files with 101 additions and 80 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

5
kernel/arm64/cgemm_ncopy_sve_v1.c
View File

@ -1,5 +1,6 @@
/*********************************************************************/ /*********************************************************************/
/* Copyright 2009, 2010 The University of Texas at Austin. */ /* Copyright 2009, 2010 The University of Texas at Austin. */
/* Copyright 2023 The OpenBLAS Project */
/* All rights reserved. */ /* All rights reserved. */
/* */ /* */
/* Redistribution and use in source and binary forms, with or */ /* Redistribution and use in source and binary forms, with or */
@ -52,7 +53,7 @@ int CNAME(BLASLONG m, BLASLONG n, IFLOAT *a, BLASLONG lda, IFLOAT *b){
boffset = b; boffset = b;
j = 0; j = 0;
svbool_t pg = svwhilelt_b32(j, n); svbool_t pg = svwhilelt_b32((uint64_t)j, (uint64_t)n);
uint32_t active = svcntp_b32(svptrue_b32(), pg); uint32_t active = svcntp_b32(svptrue_b32(), pg);
do { do {
@ -69,7 +70,7 @@ int CNAME(BLASLONG m, BLASLONG n, IFLOAT *a, BLASLONG lda, IFLOAT *b){
aoffset += active * lda * 2; aoffset += active * lda * 2;
j += svcntw(); j += svcntw();
pg = svwhilelt_b32(j, n); pg = svwhilelt_b32((uint64_t)j, (uint64_t)n);
active = svcntp_b32(svptrue_b32(), pg); active = svcntp_b32(svptrue_b32(), pg);

5
kernel/arm64/cgemm_tcopy_sve_v1.c
View File

@ -1,5 +1,6 @@
/*********************************************************************/ /*********************************************************************/
/* Copyright 2009, 2010 The University of Texas at Austin. */ /* Copyright 2009, 2010 The University of Texas at Austin. */
/* Copyright 2023 The OpenBLAS Project */
/* All rights reserved. */ /* All rights reserved. */
/* */ /* */
/* Redistribution and use in source and binary forms, with or */ /* Redistribution and use in source and binary forms, with or */
@ -50,7 +51,7 @@ int CNAME(BLASLONG m, BLASLONG n, IFLOAT *a, BLASLONG lda, IFLOAT *b){
boffset = b; boffset = b;
j = 0; j = 0;
svbool_t pg = svwhilelt_b32(j, n); svbool_t pg = svwhilelt_b32((uint64_t)j, (uint64_t)n);
uint32_t active = svcntp_b32(svptrue_b32(), pg); uint32_t active = svcntp_b32(svptrue_b32(), pg);
do { do {
@ -66,7 +67,7 @@ int CNAME(BLASLONG m, BLASLONG n, IFLOAT *a, BLASLONG lda, IFLOAT *b){
aoffset += active * 2; aoffset += active * 2;
j += svcntw(); j += svcntw();
pg = svwhilelt_b32(j, n); pg = svwhilelt_b32((uint64_t)j, (uint64_t)n);
active = svcntp_b32(svptrue_b32(), pg); active = svcntp_b32(svptrue_b32(), pg);
} while (svptest_any(svptrue_b32(), pg)); } while (svptest_any(svptrue_b32(), pg));

9
kernel/arm64/symm_lcopy_sve.c
View File

@ -1,5 +1,6 @@
/*********************************************************************/ /*********************************************************************/
/* Copyright 2009, 2010 The University of Texas at Austin. */ /* Copyright 2009, 2010 The University of Texas at Austin. */
/* Copyright 2023 The OpenBLAS Project */
/* All rights reserved. */ /* All rights reserved. */
/* */ /* */
/* Redistribution and use in source and binary forms, with or */ /* Redistribution and use in source and binary forms, with or */
@ -52,7 +53,7 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
svint64_t one_vec = svdup_s64(1LL); svint64_t one_vec = svdup_s64(1LL);
int64_t j = 0; int64_t j = 0;
svbool_t pg = svwhilelt_b64(j, n); svbool_t pg = svwhilelt_b64((uint64_t)j, (uint64_t)n);
int64_t active = svcntp_b64(svptrue_b64(), pg); int64_t active = svcntp_b64(svptrue_b64(), pg);
svint64_t index_neg = svindex_s64(0LL, -1LL); svint64_t index_neg = svindex_s64(0LL, -1LL);
svint64_t index = svindex_s64(0LL, 1LL); svint64_t index = svindex_s64(0LL, 1LL);
@ -86,7 +87,7 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
posX += sve_size; posX += sve_size;
posX_vec = svdup_s64(posX); posX_vec = svdup_s64(posX);
j += sve_size; j += sve_size;
pg = svwhilelt_b64(j, n); pg = svwhilelt_b64((uint64_t)j, (uint64_t)n);
active = svcntp_b64(svptrue_b64(), pg); active = svcntp_b64(svptrue_b64(), pg);
} while (svptest_any(svptrue_b64(), pg)); } while (svptest_any(svptrue_b64(), pg));
@ -99,7 +100,7 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
int32_t N = n; int32_t N = n;
int32_t j = 0; int32_t j = 0;
svbool_t pg = svwhilelt_b32(j, N); svbool_t pg = svwhilelt_b32((uint32_t)j, (uint32_t)N);
int32_t active = svcntp_b32(svptrue_b32(), pg); int32_t active = svcntp_b32(svptrue_b32(), pg);
svint32_t index_neg = svindex_s32(0, -1); svint32_t index_neg = svindex_s32(0, -1);
svint32_t index = svindex_s32(0, 1); svint32_t index = svindex_s32(0, 1);
@ -133,7 +134,7 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
posX += sve_size; posX += sve_size;
posX_vec = svdup_s32(posX); posX_vec = svdup_s32(posX);
j += sve_size; j += sve_size;
pg = svwhilelt_b32(j, N); pg = svwhilelt_b32((uint32_t)j, (uint32_t)N);
active = svcntp_b32(svptrue_b32(), pg); active = svcntp_b32(svptrue_b32(), pg);
} while (svptest_any(svptrue_b32(), pg)); } while (svptest_any(svptrue_b32(), pg));

9
kernel/arm64/symm_ucopy_sve.c
View File

@ -1,5 +1,6 @@
/*********************************************************************/ /*********************************************************************/
/* Copyright 2009, 2010 The University of Texas at Austin. */ /* Copyright 2009, 2010 The University of Texas at Austin. */
/* Copyright 2023 The OpenBLAS Project */
/* All rights reserved. */ /* All rights reserved. */
/* */ /* */
/* Redistribution and use in source and binary forms, with or */ /* Redistribution and use in source and binary forms, with or */
@ -52,7 +53,7 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
svint64_t one_vec = svdup_s64(1LL); svint64_t one_vec = svdup_s64(1LL);
int64_t j = 0; int64_t j = 0;
svbool_t pg = svwhilelt_b64(j, n); svbool_t pg = svwhilelt_b64((uint64_t)j, (uint64_t)n);
int64_t active = svcntp_b64(svptrue_b64(), pg); int64_t active = svcntp_b64(svptrue_b64(), pg);
svint64_t index_neg = svindex_s64(0LL, -1LL); svint64_t index_neg = svindex_s64(0LL, -1LL);
svint64_t index = svindex_s64(0LL, 1LL); svint64_t index = svindex_s64(0LL, 1LL);
@ -86,7 +87,7 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
posX += sve_size; posX += sve_size;
posX_vec = svdup_s64(posX); posX_vec = svdup_s64(posX);
j += sve_size; j += sve_size;
pg = svwhilelt_b64(j, n); pg = svwhilelt_b64((uint64_t)j, (uint64_t)n);
active = svcntp_b64(svptrue_b64(), pg); active = svcntp_b64(svptrue_b64(), pg);
} while (svptest_any(svptrue_b64(), pg)); } while (svptest_any(svptrue_b64(), pg));
@ -99,7 +100,7 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
int32_t N = n; int32_t N = n;
int32_t j = 0; int32_t j = 0;
svbool_t pg = svwhilelt_b32(j, N); svbool_t pg = svwhilelt_b32((uint32_t)j, (uint32_t)N);
int32_t active = svcntp_b32(svptrue_b32(), pg); int32_t active = svcntp_b32(svptrue_b32(), pg);
svint32_t index_neg = svindex_s32(0, -1); svint32_t index_neg = svindex_s32(0, -1);
svint32_t index = svindex_s32(0, 1); svint32_t index = svindex_s32(0, 1);
@ -133,7 +134,7 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
posX += sve_size; posX += sve_size;
posX_vec = svdup_s32(posX); posX_vec = svdup_s32(posX);
j += sve_size; j += sve_size;
pg = svwhilelt_b32(j, N); pg = svwhilelt_b32((uint32_t)j, (uint32_t)N);
active = svcntp_b32(svptrue_b32(), pg); active = svcntp_b32(svptrue_b32(), pg);
} while (svptest_any(svptrue_b32(), pg)); } while (svptest_any(svptrue_b32(), pg));

9
kernel/arm64/trsm_ltcopy_sve.c
View File

@ -1,5 +1,6 @@
/*********************************************************************/ /*********************************************************************/
/* Copyright 2009, 2010 The University of Texas at Austin. */ /* Copyright 2009, 2010 The University of Texas at Austin. */
/* Copyright 2023 The OpenBLAS Project */
/* All rights reserved. */ /* All rights reserved. */
/* */ /* */
/* Redistribution and use in source and binary forms, with or */ /* Redistribution and use in source and binary forms, with or */
@ -55,12 +56,12 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG offset, FLOAT
jj = offset; jj = offset;
#ifdef DOUBLE #ifdef DOUBLE
int64_t js = 0; int64_t js = 0;
svbool_t pn = svwhilelt_b64(js, n); svbool_t pn = svwhilelt_b64((uint64_t)js, (uint64_t)n);
int n_active = svcntp_b64(svptrue_b64(), pn); int n_active = svcntp_b64(svptrue_b64(), pn);
#else #else
int32_t N = n; int32_t N = n;
int32_t js = 0; int32_t js = 0;
svbool_t pn = svwhilelt_b32(js, N); svbool_t pn = svwhilelt_b32((uint32_t)js, (uint32_t)N);
int n_active = svcntp_b32(svptrue_b32(), pn); int n_active = svcntp_b32(svptrue_b32(), pn);
#endif #endif
do { do {
@ -104,11 +105,11 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG offset, FLOAT
js += n_active; js += n_active;
#ifdef DOUBLE #ifdef DOUBLE
pn = svwhilelt_b64(js, n); pn = svwhilelt_b64((uint64_t)js, (uint64_t)n);
n_active = svcntp_b64(svptrue_b64(), pn); n_active = svcntp_b64(svptrue_b64(), pn);
} while (svptest_any(svptrue_b64(), pn)); } while (svptest_any(svptrue_b64(), pn));
#else #else
pn = svwhilelt_b32(js, N); pn = svwhilelt_b32((uint32_t)js, (uint32_t)N);
n_active = svcntp_b32(svptrue_b32(), pn); n_active = svcntp_b32(svptrue_b32(), pn);
} while (svptest_any(svptrue_b32(), pn)); } while (svptest_any(svptrue_b32(), pn));
#endif #endif

9
kernel/arm64/trsm_uncopy_sve.c
View File

@ -1,5 +1,6 @@
/*********************************************************************/ /*********************************************************************/
/* Copyright 2009, 2010 The University of Texas at Austin. */ /* Copyright 2009, 2010 The University of Texas at Austin. */
/* Copyright 2023 The OpenBLAS Project */
/* All rights reserved. */ /* All rights reserved. */
/* */ /* */
/* Redistribution and use in source and binary forms, with or */ /* Redistribution and use in source and binary forms, with or */
@ -56,13 +57,13 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG offset, FLOAT
#ifdef DOUBLE #ifdef DOUBLE
int64_t js = 0; int64_t js = 0;
svint64_t index = svindex_s64(0LL, lda); svint64_t index = svindex_s64(0LL, lda);
svbool_t pn = svwhilelt_b64(js, n); svbool_t pn = svwhilelt_b64((uint64_t)js, (uint64_t)n);
int n_active = svcntp_b64(svptrue_b64(), pn); int n_active = svcntp_b64(svptrue_b64(), pn);
#else #else
int32_t N = n; int32_t N = n;
int32_t js = 0; int32_t js = 0;
svint32_t index = svindex_s32(0, lda); svint32_t index = svindex_s32(0, lda);
svbool_t pn = svwhilelt_b32(js, N); svbool_t pn = svwhilelt_b32((uint32_t)js, (uint32_t)N);
int n_active = svcntp_b32(svptrue_b32(), pn); int n_active = svcntp_b32(svptrue_b32(), pn);
#endif #endif
do { do {
@ -106,11 +107,11 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG offset, FLOAT
js += n_active; js += n_active;
#ifdef DOUBLE #ifdef DOUBLE
pn = svwhilelt_b64(js, n); pn = svwhilelt_b64((uint64_t)js, (uint64_t)n);
n_active = svcntp_b64(svptrue_b64(), pn); n_active = svcntp_b64(svptrue_b64(), pn);
} while (svptest_any(svptrue_b64(), pn)); } while (svptest_any(svptrue_b64(), pn));
#else #else
pn = svwhilelt_b32(js, N); pn = svwhilelt_b32((uint32_t)js, (uint32_t)N);
n_active = svcntp_b32(svptrue_b32(), pn); n_active = svcntp_b32(svptrue_b32(), pn);
} while (svptest_any(svptrue_b32(), pn)); } while (svptest_any(svptrue_b32(), pn));
#endif #endif

9
kernel/arm64/trsm_utcopy_sve.c
View File

@ -1,5 +1,6 @@
/*********************************************************************/ /*********************************************************************/
/* Copyright 2009, 2010 The University of Texas at Austin. */ /* Copyright 2009, 2010 The University of Texas at Austin. */
/* Copyright 2023 The OpenBLAS Project */
/* All rights reserved. */ /* All rights reserved. */
/* */ /* */
/* Redistribution and use in source and binary forms, with or */ /* Redistribution and use in source and binary forms, with or */
@ -55,12 +56,12 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG offset, FLOAT
jj = offset; jj = offset;
#ifdef DOUBLE #ifdef DOUBLE
int64_t js = 0; int64_t js = 0;
svbool_t pn = svwhilelt_b64(js, n); svbool_t pn = svwhilelt_b64((uint64_t)js, (uint64_t)n);
int n_active = svcntp_b64(svptrue_b64(), pn); int n_active = svcntp_b64(svptrue_b64(), pn);
#else #else
int32_t N = n; int32_t N = n;
int32_t js = 0; int32_t js = 0;
svbool_t pn = svwhilelt_b32(js, N); svbool_t pn = svwhilelt_b32((uint32_t)js, (uint32_t)N);
int n_active = svcntp_b32(svptrue_b32(), pn); int n_active = svcntp_b32(svptrue_b32(), pn);
#endif #endif
do { do {
@ -104,11 +105,11 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG offset, FLOAT
js += n_active; js += n_active;
#ifdef DOUBLE #ifdef DOUBLE
pn = svwhilelt_b64(js, n); pn = svwhilelt_b64((uint64_t)js, (uint64_t)n);
n_active = svcntp_b64(svptrue_b64(), pn); n_active = svcntp_b64(svptrue_b64(), pn);
} while (svptest_any(svptrue_b64(), pn)); } while (svptest_any(svptrue_b64(), pn));
#else #else
pn = svwhilelt_b32(js, N); pn = svwhilelt_b32((uint32_t)js, (uint32_t)N);
n_active = svcntp_b32(svptrue_b32(), pn); n_active = svcntp_b32(svptrue_b32(), pn);
} while (svptest_any(svptrue_b32(), pn)); } while (svptest_any(svptrue_b32(), pn));
#endif #endif

5
kernel/arm64/zgemm_ncopy_sve_v1.c
View File

@ -1,5 +1,6 @@
/*********************************************************************/ /*********************************************************************/
/* Copyright 2009, 2010 The University of Texas at Austin. */ /* Copyright 2009, 2010 The University of Texas at Austin. */
/* Copyright 2023 The OpenBLAS Project */
/* All rights reserved. */ /* All rights reserved. */
/* */ /* */
/* Redistribution and use in source and binary forms, with or */ /* Redistribution and use in source and binary forms, with or */
@ -52,7 +53,7 @@ int CNAME(BLASLONG m, BLASLONG n, IFLOAT *a, BLASLONG lda, IFLOAT *b){
boffset = b; boffset = b;
j = 0; j = 0;
svbool_t pg = svwhilelt_b64(j, n); svbool_t pg = svwhilelt_b64((uint64_t)j, (uint64_t)n);
uint64_t active = svcntp_b64(svptrue_b64(), pg); uint64_t active = svcntp_b64(svptrue_b64(), pg);
do { do {
@ -69,7 +70,7 @@ int CNAME(BLASLONG m, BLASLONG n, IFLOAT *a, BLASLONG lda, IFLOAT *b){
aoffset += active * lda * 2; aoffset += active * lda * 2;
j += svcntd(); j += svcntd();
pg = svwhilelt_b64(j, n); pg = svwhilelt_b64((uint64_t)j, (uint64_t)n);
active = svcntp_b64(svptrue_b64(), pg); active = svcntp_b64(svptrue_b64(), pg);

5
kernel/arm64/zgemm_tcopy_sve_v1.c
View File

@ -1,5 +1,6 @@
/*********************************************************************/ /*********************************************************************/
/* Copyright 2009, 2010 The University of Texas at Austin. */ /* Copyright 2009, 2010 The University of Texas at Austin. */
/* Copyright 2023 The OpenBLAS Project */
/* All rights reserved. */ /* All rights reserved. */
/* */ /* */
/* Redistribution and use in source and binary forms, with or */ /* Redistribution and use in source and binary forms, with or */
@ -50,7 +51,7 @@ int CNAME(BLASLONG m, BLASLONG n, IFLOAT *a, BLASLONG lda, IFLOAT *b){
boffset = b; boffset = b;
j = 0; j = 0;
svbool_t pg = svwhilelt_b64(j, n); svbool_t pg = svwhilelt_b64((uint64_t)j, (uint64_t)n);
uint64_t active = svcntp_b64(svptrue_b64(), pg); uint64_t active = svcntp_b64(svptrue_b64(), pg);
do { do {
@ -66,7 +67,7 @@ int CNAME(BLASLONG m, BLASLONG n, IFLOAT *a, BLASLONG lda, IFLOAT *b){
aoffset += active * 2; aoffset += active * 2;
j += svcntd(); j += svcntd();
pg = svwhilelt_b64(j, n); pg = svwhilelt_b64((uint64_t)j, (uint64_t)n);
active = svcntp_b64(svptrue_b64(), pg); active = svcntp_b64(svptrue_b64(), pg);
} while (svptest_any(svptrue_b64(), pg)); } while (svptest_any(svptrue_b64(), pg));

13
kernel/arm64/zhemm_ltcopy_sve.c
View File

@ -1,5 +1,6 @@
/*********************************************************************/ /*********************************************************************/
/* Copyright 2009, 2010 The University of Texas at Austin. */ /* Copyright 2009, 2010 The University of Texas at Austin. */
/* Copyright 2023 The OpenBLAS Project */
/* All rights reserved. */ /* All rights reserved. */
/* */ /* */
/* Redistribution and use in source and binary forms, with or */ /* Redistribution and use in source and binary forms, with or */
@ -54,7 +55,7 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
svint64_t one_vec = svdup_s64(1LL); svint64_t one_vec = svdup_s64(1LL);
int64_t j = 0; int64_t j = 0;
svbool_t pg = svwhilelt_b64(j, n); svbool_t pg = svwhilelt_b64((uint64_t)j, (uint64_t)n);
int64_t active = svcntp_b64(svptrue_b64(), pg); int64_t active = svcntp_b64(svptrue_b64(), pg);
svint64_t index_neg = svindex_s64(0LL, -1LL); svint64_t index_neg = svindex_s64(0LL, -1LL);
svint64_t index = svindex_s64(0LL, 1LL); svint64_t index = svindex_s64(0LL, 1LL);
@ -79,7 +80,7 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
gat_ind = svadd_m(cmp, gat_ind, lda_vec); gat_ind = svadd_m(cmp, gat_ind, lda_vec);
gat_ind = svadd_m(svnot_z(pg, cmp) , gat_ind, 2); gat_ind = svadd_m(svnot_z(pg, cmp) , gat_ind, 2);
if (offset <= 0) { if (offset <= 0) {
svbool_t off_g = svwhilelt_b64(offset, 0LL); svbool_t off_g = svwhilelt_b64((uint64_t)offset, (uint64_t)0LL);
data_vec_imag = svneg_m(data_vec_imag, off_g, data_vec_imag); data_vec_imag = svneg_m(data_vec_imag, off_g, data_vec_imag);
} }
@ -99,7 +100,7 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
posX += sve_size; posX += sve_size;
posX_vec = svdup_s64(posX); posX_vec = svdup_s64(posX);
j += sve_size; j += sve_size;
pg = svwhilelt_b64(j, n); pg = svwhilelt_b64((uint64_t)j, (uint64_t)n);
active = svcntp_b64(svptrue_b64(), pg); active = svcntp_b64(svptrue_b64(), pg);
} while (svptest_any(svptrue_b64(), pg)); } while (svptest_any(svptrue_b64(), pg));
@ -117,7 +118,7 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
int32_t j = 0; int32_t j = 0;
int32_t N = n; int32_t N = n;
svbool_t pg = svwhilelt_b32(j, N); svbool_t pg = svwhilelt_b32((uint32_t)j, (uint32_t)N);
int32_t active = svcntp_b32(svptrue_b32(), pg); int32_t active = svcntp_b32(svptrue_b32(), pg);
svint32_t index_neg = svindex_s32(0, -1); svint32_t index_neg = svindex_s32(0, -1);
svint32_t index = svindex_s32(0, 1); svint32_t index = svindex_s32(0, 1);
@ -142,7 +143,7 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
gat_ind = svadd_m(cmp, gat_ind, lda_vec); gat_ind = svadd_m(cmp, gat_ind, lda_vec);
gat_ind = svadd_m(svnot_z(pg, cmp) , gat_ind, 2); gat_ind = svadd_m(svnot_z(pg, cmp) , gat_ind, 2);
if (offset <= 0) { if (offset <= 0) {
svbool_t off_g = svwhilelt_b32(offset, 0); svbool_t off_g = svwhilelt_b32((uint32_t)offset, (uint32_t)0);
data_vec_imag = svneg_m(data_vec_imag, off_g, data_vec_imag); data_vec_imag = svneg_m(data_vec_imag, off_g, data_vec_imag);
} }
@ -162,7 +163,7 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
posX += sve_size; posX += sve_size;
posX_vec = svdup_s32(posX); posX_vec = svdup_s32(posX);
j += sve_size; j += sve_size;
pg = svwhilelt_b32(j, N); pg = svwhilelt_b32((uint32_t)j, (uint32_t)N);
active = svcntp_b32(svptrue_b32(), pg); active = svcntp_b32(svptrue_b32(), pg);
} while (svptest_any(svptrue_b32(), pg)); } while (svptest_any(svptrue_b32(), pg));

13
kernel/arm64/zhemm_utcopy_sve.c
View File

@ -1,5 +1,6 @@
/*********************************************************************/ /*********************************************************************/
/* Copyright 2009, 2010 The University of Texas at Austin. */ /* Copyright 2009, 2010 The University of Texas at Austin. */
/* Copyright 2023 The OpenBLAS Project */
/* All rights reserved. */ /* All rights reserved. */
/* */ /* */
/* Redistribution and use in source and binary forms, with or */ /* Redistribution and use in source and binary forms, with or */
@ -54,7 +55,7 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
svint64_t one_vec = svdup_s64(1LL); svint64_t one_vec = svdup_s64(1LL);
int64_t j = 0; int64_t j = 0;
svbool_t pg = svwhilelt_b64(j, n); svbool_t pg = svwhilelt_b64((uint64_t)j, (uint64_t)n);
int64_t active = svcntp_b64(svptrue_b64(), pg); int64_t active = svcntp_b64(svptrue_b64(), pg);
svint64_t index_neg = svindex_s64(0LL, -1LL); svint64_t index_neg = svindex_s64(0LL, -1LL);
svint64_t index = svindex_s64(0LL, 1LL); svint64_t index = svindex_s64(0LL, 1LL);
@ -80,7 +81,7 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
gat_ind = svadd_m(svnot_z(pg, cmp) , gat_ind, lda_vec); gat_ind = svadd_m(svnot_z(pg, cmp) , gat_ind, lda_vec);
data_vec_imag = svneg_z(pg, data_vec_imag); data_vec_imag = svneg_z(pg, data_vec_imag);
if (offset <= 0) { if (offset <= 0) {
svbool_t off_g = svwhilelt_b64(offset, 0LL); svbool_t off_g = svwhilelt_b64((uint64_t)offset, (uint64_t)0LL);
data_vec_imag = svneg_m(data_vec_imag, off_g, data_vec_imag); data_vec_imag = svneg_m(data_vec_imag, off_g, data_vec_imag);
} }
@ -100,7 +101,7 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
posX += sve_size; posX += sve_size;
posX_vec = svdup_s64(posX); posX_vec = svdup_s64(posX);
j += sve_size; j += sve_size;
pg = svwhilelt_b64(j, n); pg = svwhilelt_b64((uint64_t)j, (uint64_t)n);
active = svcntp_b64(svptrue_b64(), pg); active = svcntp_b64(svptrue_b64(), pg);
} while (svptest_any(svptrue_b64(), pg)); } while (svptest_any(svptrue_b64(), pg));
#else #else
@ -116,7 +117,7 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
int32_t j = 0; int32_t j = 0;
int32_t N = n; int32_t N = n;
svbool_t pg = svwhilelt_b32(j, N); svbool_t pg = svwhilelt_b32((uint32_t)j, (uint32_t)N);
int32_t active = svcntp_b32(svptrue_b32(), pg); int32_t active = svcntp_b32(svptrue_b32(), pg);
svint32_t index_neg = svindex_s32(0, -1); svint32_t index_neg = svindex_s32(0, -1);
svint32_t index = svindex_s32(0, 1); svint32_t index = svindex_s32(0, 1);
@ -142,7 +143,7 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
gat_ind = svadd_m(svnot_z(pg, cmp) , gat_ind, lda_vec); gat_ind = svadd_m(svnot_z(pg, cmp) , gat_ind, lda_vec);
data_vec_imag = svneg_z(pg, data_vec_imag); data_vec_imag = svneg_z(pg, data_vec_imag);
if (offset <= 0) { if (offset <= 0) {
svbool_t off_g = svwhilelt_b32(offset, 0); svbool_t off_g = svwhilelt_b32((uint32_t)offset, (uint32_t)0);
data_vec_imag = svneg_m(data_vec_imag, off_g, data_vec_imag); data_vec_imag = svneg_m(data_vec_imag, off_g, data_vec_imag);
} }
@ -162,7 +163,7 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
posX += sve_size; posX += sve_size;
posX_vec = svdup_s32(posX); posX_vec = svdup_s32(posX);
j += sve_size; j += sve_size;
pg = svwhilelt_b32(j, N); pg = svwhilelt_b32((uint32_t)j, (uint32_t)N);
active = svcntp_b32(svptrue_b32(), pg); active = svcntp_b32(svptrue_b32(), pg);
} while (svptest_any(svptrue_b32(), pg)); } while (svptest_any(svptrue_b32(), pg));

9
kernel/arm64/zsymm_lcopy_sve.c
View File

@ -1,5 +1,6 @@
/*********************************************************************/ /*********************************************************************/
/* Copyright 2009, 2010 The University of Texas at Austin. */ /* Copyright 2009, 2010 The University of Texas at Austin. */
/* Copyright 2023 The OpenBLAS Project */
/* All rights reserved. */ /* All rights reserved. */
/* */ /* */
/* Redistribution and use in source and binary forms, with or */ /* Redistribution and use in source and binary forms, with or */
@ -53,7 +54,7 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
svint64_t one_vec = svdup_s64(1LL); svint64_t one_vec = svdup_s64(1LL);
int64_t j = 0; int64_t j = 0;
svbool_t pg = svwhilelt_b64(j, n); svbool_t pg = svwhilelt_b64((uint64_t)j, (uint64_t)n);
int64_t active = svcntp_b64(svptrue_b64(), pg); int64_t active = svcntp_b64(svptrue_b64(), pg);
svint64_t index_neg = svindex_s64(0LL, -1LL); svint64_t index_neg = svindex_s64(0LL, -1LL);
svint64_t index = svindex_s64(0LL, 1LL); svint64_t index = svindex_s64(0LL, 1LL);
@ -90,7 +91,7 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
posX += sve_size; posX += sve_size;
posX_vec = svdup_s64(posX); posX_vec = svdup_s64(posX);
j += sve_size; j += sve_size;
pg = svwhilelt_b64(j, n); pg = svwhilelt_b64((uint64_t)j, (uint64_t)n);
active = svcntp_b64(svptrue_b64(), pg); active = svcntp_b64(svptrue_b64(), pg);
} while (svptest_any(svptrue_b64(), pg)); } while (svptest_any(svptrue_b64(), pg));
@ -103,7 +104,7 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
int32_t N = n; int32_t N = n;
int32_t j = 0; int32_t j = 0;
svbool_t pg = svwhilelt_b32(j, N); svbool_t pg = svwhilelt_b32((uint32_t)j, (uint32_t)N);
int32_t active = svcntp_b32(svptrue_b32(), pg); int32_t active = svcntp_b32(svptrue_b32(), pg);
svint32_t index_neg = svindex_s32(0, -1); svint32_t index_neg = svindex_s32(0, -1);
svint32_t index = svindex_s32(0, 1); svint32_t index = svindex_s32(0, 1);
@ -140,7 +141,7 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
posX += sve_size; posX += sve_size;
posX_vec = svdup_s32(posX); posX_vec = svdup_s32(posX);
j += sve_size; j += sve_size;
pg = svwhilelt_b32(j, N); pg = svwhilelt_b32((uint32_t)j, (uint32_t)N);
active = svcntp_b32(svptrue_b32(), pg); active = svcntp_b32(svptrue_b32(), pg);
} while (svptest_any(svptrue_b32(), pg)); } while (svptest_any(svptrue_b32(), pg));

9
kernel/arm64/zsymm_ucopy_sve.c
View File

@ -1,5 +1,6 @@
/*********************************************************************/ /*********************************************************************/
/* Copyright 2009, 2010 The University of Texas at Austin. */ /* Copyright 2009, 2010 The University of Texas at Austin. */
/* Copyright 2023 The OpenBLAS Project */
/* All rights reserved. */ /* All rights reserved. */
/* */ /* */
/* Redistribution and use in source and binary forms, with or */ /* Redistribution and use in source and binary forms, with or */
@ -53,7 +54,7 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
svint64_t one_vec = svdup_s64(1LL); svint64_t one_vec = svdup_s64(1LL);
int64_t j = 0; int64_t j = 0;
svbool_t pg = svwhilelt_b64(j, n); svbool_t pg = svwhilelt_b64((uint64_t)j, (uint64_t)n);
int64_t active = svcntp_b64(svptrue_b64(), pg); int64_t active = svcntp_b64(svptrue_b64(), pg);
svint64_t index_neg = svindex_s64(0LL, -1LL); svint64_t index_neg = svindex_s64(0LL, -1LL);
svint64_t index = svindex_s64(0LL, 1LL); svint64_t index = svindex_s64(0LL, 1LL);
@ -90,7 +91,7 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
posX += sve_size; posX += sve_size;
posX_vec = svdup_s64(posX); posX_vec = svdup_s64(posX);
j += sve_size; j += sve_size;
pg = svwhilelt_b64(j, n); pg = svwhilelt_b64((uint64_t)j, (uint64_t)n);
active = svcntp_b64(svptrue_b64(), pg); active = svcntp_b64(svptrue_b64(), pg);
} while (svptest_any(svptrue_b64(), pg)); } while (svptest_any(svptrue_b64(), pg));
@ -103,7 +104,7 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
int32_t N = n; int32_t N = n;
int32_t j = 0; int32_t j = 0;
svbool_t pg = svwhilelt_b32(j, N); svbool_t pg = svwhilelt_b32((uint32_t)j, (uint32_t)N);
int32_t active = svcntp_b32(svptrue_b32(), pg); int32_t active = svcntp_b32(svptrue_b32(), pg);
svint32_t index_neg = svindex_s32(0, -1); svint32_t index_neg = svindex_s32(0, -1);
svint32_t index = svindex_s32(0, 1); svint32_t index = svindex_s32(0, 1);
@ -140,7 +141,7 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
posX += sve_size; posX += sve_size;
posX_vec = svdup_s32(posX); posX_vec = svdup_s32(posX);
j += sve_size; j += sve_size;
pg = svwhilelt_b32(j, N); pg = svwhilelt_b32((uint32_t)j, (uint32_t)N);
active = svcntp_b32(svptrue_b32(), pg); active = svcntp_b32(svptrue_b32(), pg);
} while (svptest_any(svptrue_b32(), pg)); } while (svptest_any(svptrue_b32(), pg));

9
kernel/arm64/ztrmm_lncopy_sve_v1.c
View File

@ -1,5 +1,6 @@
/*********************************************************************/ /*********************************************************************/
/* Copyright 2009, 2010 The University of Texas at Austin. */ /* Copyright 2009, 2010 The University of Texas at Austin. */
/* Copyright 2023 The OpenBLAS Project */
/* All rights reserved. */ /* All rights reserved. */
/* */ /* */
/* Redistribution and use in source and binary forms, with or */ /* Redistribution and use in source and binary forms, with or */
@ -54,11 +55,11 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
FLOAT *ao; FLOAT *ao;
#ifdef DOUBLE #ifdef DOUBLE
svint64_t index = svindex_s64(0LL, lda); svint64_t index = svindex_s64(0LL, lda);
svbool_t pn = svwhilelt_b64(js, n); svbool_t pn = svwhilelt_b64((uint64_t)js, (uint64_t)n);
int n_active = svcntp_b64(svptrue_b64(), pn); int n_active = svcntp_b64(svptrue_b64(), pn);
#else #else
svint32_t index = svindex_s32(0, lda); svint32_t index = svindex_s32(0, lda);
svbool_t pn = svwhilelt_b32(js, n); svbool_t pn = svwhilelt_b32((uint64_t)js, (uint64_t)n);
int n_active = svcntp_b32(svptrue_b32(), pn); int n_active = svcntp_b32(svptrue_b32(), pn);
#endif #endif
do do
@ -132,11 +133,11 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
posY += n_active; posY += n_active;
js += n_active; js += n_active;
#ifdef DOUBLE #ifdef DOUBLE
pn = svwhilelt_b64(js, n); pn = svwhilelt_b64((uint64_t)js, (uint64_t)n);
n_active = svcntp_b64(svptrue_b64(), pn); n_active = svcntp_b64(svptrue_b64(), pn);
} while (svptest_any(svptrue_b64(), pn)); } while (svptest_any(svptrue_b64(), pn));
#else #else
pn = svwhilelt_b32(js, n); pn = svwhilelt_b32((uint64_t)js, (uint64_t)n);
n_active = svcntp_b32(svptrue_b32(), pn); n_active = svcntp_b32(svptrue_b32(), pn);
} while (svptest_any(svptrue_b32(), pn)); } while (svptest_any(svptrue_b32(), pn));
#endif #endif

9
kernel/arm64/ztrmm_ltcopy_sve_v1.c
View File

@ -1,5 +1,6 @@
/*********************************************************************/ /*********************************************************************/
/* Copyright 2009, 2010 The University of Texas at Austin. */ /* Copyright 2009, 2010 The University of Texas at Austin. */
/* Copyright 2023 The OpenBLAS Project */
/* All rights reserved. */ /* All rights reserved. */
/* */ /* */
/* Redistribution and use in source and binary forms, with or */ /* Redistribution and use in source and binary forms, with or */
@ -53,10 +54,10 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
FLOAT *ao; FLOAT *ao;
js = 0; js = 0;
#ifdef DOUBLE #ifdef DOUBLE
svbool_t pn = svwhilelt_b64(js, n); svbool_t pn = svwhilelt_b64((uint64_t)js, (uint64_t)n);
int n_active = svcntp_b64(svptrue_b64(), pn); int n_active = svcntp_b64(svptrue_b64(), pn);
#else #else
svbool_t pn = svwhilelt_b32(js, n); svbool_t pn = svwhilelt_b32((uint64_t)js, (uint64_t)n);
int n_active = svcntp_b32(svptrue_b32(), pn); int n_active = svcntp_b32(svptrue_b32(), pn);
#endif #endif
do do
@ -129,11 +130,11 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
posY += n_active; posY += n_active;
js += n_active; js += n_active;
#ifdef DOUBLE #ifdef DOUBLE
pn = svwhilelt_b64(js, n); pn = svwhilelt_b64((uint64_t)js, (uint64_t)n);
n_active = svcntp_b64(svptrue_b64(), pn); n_active = svcntp_b64(svptrue_b64(), pn);
} while (svptest_any(svptrue_b64(), pn)); } while (svptest_any(svptrue_b64(), pn));
#else #else
pn = svwhilelt_b32(js, n); pn = svwhilelt_b32((uint64_t)js, (uint64_t)n);
n_active = svcntp_b32(svptrue_b32(), pn); n_active = svcntp_b32(svptrue_b32(), pn);
} while (svptest_any(svptrue_b32(), pn)); } while (svptest_any(svptrue_b32(), pn));
#endif #endif

9
kernel/arm64/ztrmm_uncopy_sve_v1.c
View File

@ -1,5 +1,6 @@
/*********************************************************************/ /*********************************************************************/
/* Copyright 2009, 2010 The University of Texas at Austin. */ /* Copyright 2009, 2010 The University of Texas at Austin. */
/* Copyright 2023 The OpenBLAS Project */
/* All rights reserved. */ /* All rights reserved. */
/* */ /* */
/* Redistribution and use in source and binary forms, with or */ /* Redistribution and use in source and binary forms, with or */
@ -54,11 +55,11 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
FLOAT *ao; FLOAT *ao;
#ifdef DOUBLE #ifdef DOUBLE
svint64_t index = svindex_s64(0LL, lda); svint64_t index = svindex_s64(0LL, lda);
svbool_t pn = svwhilelt_b64(js, n); svbool_t pn = svwhilelt_b64((uint64_t)js, (uint64_t)n);
int n_active = svcntp_b64(svptrue_b64(), pn); int n_active = svcntp_b64(svptrue_b64(), pn);
#else #else
svint32_t index = svindex_s32(0, lda); svint32_t index = svindex_s32(0, lda);
svbool_t pn = svwhilelt_b32(js, n); svbool_t pn = svwhilelt_b32((uint64_t)js, (uint64_t)n);
int n_active = svcntp_b32(svptrue_b32(), pn); int n_active = svcntp_b32(svptrue_b32(), pn);
#endif #endif
do do
@ -132,11 +133,11 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
posY += n_active; posY += n_active;
js += n_active; js += n_active;
#ifdef DOUBLE #ifdef DOUBLE
pn = svwhilelt_b64(js, n); pn = svwhilelt_b64((uint64_t)js, (uint64_t)n);
n_active = svcntp_b64(svptrue_b64(), pn); n_active = svcntp_b64(svptrue_b64(), pn);
} while (svptest_any(svptrue_b64(), pn)); } while (svptest_any(svptrue_b64(), pn));
#else #else
pn = svwhilelt_b32(js, n); pn = svwhilelt_b32((uint64_t)js, (uint64_t)n);
n_active = svcntp_b32(svptrue_b32(), pn); n_active = svcntp_b32(svptrue_b32(), pn);
} while (svptest_any(svptrue_b32(), pn)); } while (svptest_any(svptrue_b32(), pn));
#endif #endif

9
kernel/arm64/ztrmm_utcopy_sve_v1.c
View File

@ -1,5 +1,6 @@
/*********************************************************************/ /*********************************************************************/
/* Copyright 2009, 2010 The University of Texas at Austin. */ /* Copyright 2009, 2010 The University of Texas at Austin. */
/* Copyright 2023 The OpenBLAS Project */
/* All rights reserved. */ /* All rights reserved. */
/* */ /* */
/* Redistribution and use in source and binary forms, with or */ /* Redistribution and use in source and binary forms, with or */
@ -53,10 +54,10 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
FLOAT *ao; FLOAT *ao;
js = 0; js = 0;
#ifdef DOUBLE #ifdef DOUBLE
svbool_t pn = svwhilelt_b64(js, n); svbool_t pn = svwhilelt_b64((uint64_t)js, (uint64_t)n);
int n_active = svcntp_b64(svptrue_b64(), pn); int n_active = svcntp_b64(svptrue_b64(), pn);
#else #else
svbool_t pn = svwhilelt_b32(js, n); svbool_t pn = svwhilelt_b32((uint64_t)js, (uint64_t)n);
int n_active = svcntp_b32(svptrue_b32(), pn); int n_active = svcntp_b32(svptrue_b32(), pn);
#endif #endif
do do
@ -128,11 +129,11 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
posY += n_active; posY += n_active;
js += n_active; js += n_active;
#ifdef DOUBLE #ifdef DOUBLE
pn = svwhilelt_b64(js, n); pn = svwhilelt_b64((uint64_t)js, (uint64_t)n);
n_active = svcntp_b64(svptrue_b64(), pn); n_active = svcntp_b64(svptrue_b64(), pn);
} while (svptest_any(svptrue_b64(), pn)); } while (svptest_any(svptrue_b64(), pn));
#else #else
pn = svwhilelt_b32(js, n); pn = svwhilelt_b32((uint64_t)js, (uint64_t)n);
n_active = svcntp_b32(svptrue_b32(), pn); n_active = svcntp_b32(svptrue_b32(), pn);
} while (svptest_any(svptrue_b32(), pn)); } while (svptest_any(svptrue_b32(), pn));
#endif #endif

9
kernel/arm64/ztrsm_lncopy_sve.c
View File

@ -1,5 +1,6 @@
/*********************************************************************/ /*********************************************************************/
/* Copyright 2009, 2010 The University of Texas at Austin. */ /* Copyright 2009, 2010 The University of Texas at Austin. */
/* Copyright 2023 The OpenBLAS Project */
/* All rights reserved. */ /* All rights reserved. */
/* */ /* */
/* Redistribution and use in source and binary forms, with or */ /* Redistribution and use in source and binary forms, with or */
@ -52,13 +53,13 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG offset, FLOAT
#ifdef DOUBLE #ifdef DOUBLE
int64_t js = 0; int64_t js = 0;
svint64_t index = svindex_s64(0LL, lda); svint64_t index = svindex_s64(0LL, lda);
svbool_t pn = svwhilelt_b64(js, n); svbool_t pn = svwhilelt_b64((uint64_t)js, (uint64_t)n);
int n_active = svcntp_b64(svptrue_b64(), pn); int n_active = svcntp_b64(svptrue_b64(), pn);
#else #else
int32_t N = n; int32_t N = n;
int32_t js = 0; int32_t js = 0;
svint32_t index = svindex_s32(0, lda); svint32_t index = svindex_s32(0, lda);
svbool_t pn = svwhilelt_b32(js, N); svbool_t pn = svwhilelt_b32((uint32_t)js, (uint32_t)N);
int n_active = svcntp_b32(svptrue_b32(), pn); int n_active = svcntp_b32(svptrue_b32(), pn);
#endif #endif
do { do {
@ -106,11 +107,11 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG offset, FLOAT
js += n_active; js += n_active;
#ifdef DOUBLE #ifdef DOUBLE
pn = svwhilelt_b64(js, n); pn = svwhilelt_b64((uint64_t)js, (uint64_t)n);
n_active = svcntp_b64(svptrue_b64(), pn); n_active = svcntp_b64(svptrue_b64(), pn);
} while (svptest_any(svptrue_b64(), pn)); } while (svptest_any(svptrue_b64(), pn));
#else #else
pn = svwhilelt_b32(js, N); pn = svwhilelt_b32((uint32_t)js, (uint32_t)N);
n_active = svcntp_b32(svptrue_b32(), pn); n_active = svcntp_b32(svptrue_b32(), pn);
} while (svptest_any(svptrue_b32(), pn)); } while (svptest_any(svptrue_b32(), pn));
#endif #endif

9
kernel/arm64/ztrsm_ltcopy_sve.c
View File

@ -1,5 +1,6 @@
/*********************************************************************/ /*********************************************************************/
/* Copyright 2009, 2010 The University of Texas at Austin. */ /* Copyright 2009, 2010 The University of Texas at Austin. */
/* Copyright 2023 The OpenBLAS Project */
/* All rights reserved. */ /* All rights reserved. */
/* */ /* */
/* Redistribution and use in source and binary forms, with or */ /* Redistribution and use in source and binary forms, with or */
@ -51,12 +52,12 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG offset, FLOAT
jj = offset; jj = offset;
#ifdef DOUBLE #ifdef DOUBLE
int64_t js = 0; int64_t js = 0;
svbool_t pn = svwhilelt_b64(js, n); svbool_t pn = svwhilelt_b64((uint64_t)js, (uint64_t)n);
int n_active = svcntp_b64(svptrue_b64(), pn); int n_active = svcntp_b64(svptrue_b64(), pn);
#else #else
int32_t N = n; int32_t N = n;
int32_t js = 0; int32_t js = 0;
svbool_t pn = svwhilelt_b32(js, N); svbool_t pn = svwhilelt_b32((uint32_t)js, (uint32_t)N);
int n_active = svcntp_b32(svptrue_b32(), pn); int n_active = svcntp_b32(svptrue_b32(), pn);
#endif #endif
do { do {
@ -102,11 +103,11 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG offset, FLOAT
js += n_active; js += n_active;
#ifdef DOUBLE #ifdef DOUBLE
pn = svwhilelt_b64(js, n); pn = svwhilelt_b64((uint64_t)js, (uint64_t)n);
n_active = svcntp_b64(svptrue_b64(), pn); n_active = svcntp_b64(svptrue_b64(), pn);
} while (svptest_any(svptrue_b64(), pn)); } while (svptest_any(svptrue_b64(), pn));
#else #else
pn = svwhilelt_b32(js, N); pn = svwhilelt_b32((uint32_t)js, (uint32_t)N);
n_active = svcntp_b32(svptrue_b32(), pn); n_active = svcntp_b32(svptrue_b32(), pn);
} while (svptest_any(svptrue_b32(), pn)); } while (svptest_any(svptrue_b32(), pn));
#endif #endif

9
kernel/arm64/ztrsm_uncopy_sve.c
View File

@ -1,5 +1,6 @@
/*********************************************************************/ /*********************************************************************/
/* Copyright 2009, 2010 The University of Texas at Austin. */ /* Copyright 2009, 2010 The University of Texas at Austin. */
/* Copyright 2023 The OpenBLAS Project */
/* All rights reserved. */ /* All rights reserved. */
/* */ /* */
/* Redistribution and use in source and binary forms, with or */ /* Redistribution and use in source and binary forms, with or */
@ -52,13 +53,13 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG offset, FLOAT
#ifdef DOUBLE #ifdef DOUBLE
int64_t js = 0; int64_t js = 0;
svint64_t index = svindex_s64(0LL, lda); svint64_t index = svindex_s64(0LL, lda);
svbool_t pn = svwhilelt_b64(js, n); svbool_t pn = svwhilelt_b64((uint64_t)js, (uint64_t)n);
int n_active = svcntp_b64(svptrue_b64(), pn); int n_active = svcntp_b64(svptrue_b64(), pn);
#else #else
int32_t N = n; int32_t N = n;
int32_t js = 0; int32_t js = 0;
svint32_t index = svindex_s32(0, lda); svint32_t index = svindex_s32(0, lda);
svbool_t pn = svwhilelt_b32(js, N); svbool_t pn = svwhilelt_b32((uint32_t)js, (uint32_t)N);
int n_active = svcntp_b32(svptrue_b32(), pn); int n_active = svcntp_b32(svptrue_b32(), pn);
#endif #endif
do { do {
@ -106,11 +107,11 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG offset, FLOAT
js += n_active; js += n_active;
#ifdef DOUBLE #ifdef DOUBLE
pn = svwhilelt_b64(js, n); pn = svwhilelt_b64((uint64_t)js, (uint64_t)n);
n_active = svcntp_b64(svptrue_b64(), pn); n_active = svcntp_b64(svptrue_b64(), pn);
} while (svptest_any(svptrue_b64(), pn)); } while (svptest_any(svptrue_b64(), pn));
#else #else
pn = svwhilelt_b32(js, N); pn = svwhilelt_b32((uint32_t)js, (uint32_t)N);
n_active = svcntp_b32(svptrue_b32(), pn); n_active = svcntp_b32(svptrue_b32(), pn);
} while (svptest_any(svptrue_b32(), pn)); } while (svptest_any(svptrue_b32(), pn));
#endif #endif

9
kernel/arm64/ztrsm_utcopy_sve.c
View File

@ -1,5 +1,6 @@
/*********************************************************************/ /*********************************************************************/
/* Copyright 2009, 2010 The University of Texas at Austin. */ /* Copyright 2009, 2010 The University of Texas at Austin. */
/* Copyright 2023 The OpenBLAS Project */
/* All rights reserved. */ /* All rights reserved. */
/* */ /* */
/* Redistribution and use in source and binary forms, with or */ /* Redistribution and use in source and binary forms, with or */
@ -51,12 +52,12 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG offset, FLOAT
jj = offset; jj = offset;
#ifdef DOUBLE #ifdef DOUBLE
int64_t js = 0; int64_t js = 0;
svbool_t pn = svwhilelt_b64(js, n); svbool_t pn = svwhilelt_b64((uint64_t)js, (uint64_t)n);
int n_active = svcntp_b64(svptrue_b64(), pn); int n_active = svcntp_b64(svptrue_b64(), pn);
#else #else
int32_t N = n; int32_t N = n;
int32_t js = 0; int32_t js = 0;
svbool_t pn = svwhilelt_b32(js, N); svbool_t pn = svwhilelt_b32((uint32_t)js, (uint32_t)N);
int n_active = svcntp_b32(svptrue_b32(), pn); int n_active = svcntp_b32(svptrue_b32(), pn);
#endif #endif
do { do {
@ -102,11 +103,11 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG offset, FLOAT
js += n_active; js += n_active;
#ifdef DOUBLE #ifdef DOUBLE
pn = svwhilelt_b64(js, n); pn = svwhilelt_b64((uint64_t)js, (uint64_t)n);
n_active = svcntp_b64(svptrue_b64(), pn); n_active = svcntp_b64(svptrue_b64(), pn);
} while (svptest_any(svptrue_b64(), pn)); } while (svptest_any(svptrue_b64(), pn));
#else #else
pn = svwhilelt_b32(js, N); pn = svwhilelt_b32((uint32_t)js, (uint32_t)N);
n_active = svcntp_b32(svptrue_b32(), pn); n_active = svcntp_b32(svptrue_b32(), pn);
} while (svptest_any(svptrue_b32(), pn)); } while (svptest_any(svptrue_b32(), pn));
#endif #endif