Skip to content

Commit dd7a1d6

Browse files
ChipKerchnerChipKerchner
committed
Merge branch 'develop' of https://github.com/OpenMathLib/OpenBLAS into develop
2 parents b27a118 + aaa5c37 commit dd7a1d6

10 files changed

+848
-4
lines changed

‎Makefile.system‎

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -441,7 +441,7 @@ ifeq (x$(XCVER), x 15)
441441
CCOMMON_OPT += -Wl,-ld_classic
442442
FCOMMON_OPT += -Wl,-ld_classic
443443
endif
444-
ifeq (x$(XCVER), x 16)
444+
ifeq ($(shell [ $(XCVER) -ge 16 ] && echo yes),yes)
445445
override CEXTRALIB := $(filter-out(-lto_library, $(CEXTRALIB)))
446446
endif
447447
endif

‎kernel/power/ddot_microk_power10.c‎

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -119,7 +119,7 @@ static double ddot_kernel_8 (long n, double *x, double *y)
119119
"+b" (y) // 3
120120
:
121121
"m" (*(const double (*)[n]) x),
122-
"m" (*(const double (*)[n]) y),
122+
"m" (*(const double (*)[n]) y)
123123
:
124124
"cr0",
125125
"vs32","vs33","vs34","vs35","vs36","vs37","vs38","vs39",

‎kernel/power/sasum_microk_power10.c‎

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -141,7 +141,7 @@ static float sasum_kernel_32 (long n, float *x)
141141
"=wa" (t2), // 5
142142
"=wa" (t3) // 6
143143
:
144-
"m" (*(const float (*)[n]) x),
144+
"m" (*(const float (*)[n]) x)
145145
:
146146
"cr0",
147147
"vs32","vs33","vs34","vs35","vs36","vs37","vs38","vs39",

‎kernel/power/zaxpy_microk_power10.c‎

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -225,7 +225,7 @@ static void zaxpy_kernel_4 (long n, double *x, double *y,
225225
"=b" (ytmp) // 12
226226
:
227227
"m" (*(const double (*)[n * 2]) x),
228-
"m" (*(const double (*)[2]) mvecp)
228+
"m" (*(const double (*)[2]) mvecp),
229229
230230
"d" (alpha_r), // 15
231231
"d" (alpha_i), // 16

‎kernel/riscv64/KERNEL.RISCV64_ZVL128B‎

Lines changed: 18 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -120,13 +120,21 @@ endif
120120

121121
DGEMMKERNEL = dgemm_kernel_$(DGEMM_UNROLL_M)x$(DGEMM_UNROLL_N)_zvl128b.c
122122
DGEMMONCOPY = ../generic/gemm_ncopy_$(DGEMM_UNROLL_N).c
123+
ifneq ($(filter $(DGEMM_UNROLL_N),4 8 16),)
124+
DGEMMOTCOPY = gemm_tcopy_$(DGEMM_UNROLL_N)_rvv.c
125+
else
123126
DGEMMOTCOPY = ../generic/gemm_tcopy_$(DGEMM_UNROLL_N).c
127+
endif
124128
DGEMMONCOPYOBJ = dgemm_oncopy$(TSUFFIX).$(SUFFIX)
125129
DGEMMOTCOPYOBJ = dgemm_otcopy$(TSUFFIX).$(SUFFIX)
126130

127131
ifneq ($(DGEMM_UNROLL_M), $(DGEMM_UNROLL_N))
128132
DGEMMINCOPY = ../generic/gemm_ncopy_$(DGEMM_UNROLL_M).c
133+
ifneq ($(filter $(DGEMM_UNROLL_M),4 8 16),)
134+
DGEMMITCOPY = gemm_tcopy_$(DGEMM_UNROLL_M)_rvv.c
135+
else
129136
DGEMMITCOPY = ../generic/gemm_tcopy_$(DGEMM_UNROLL_M).c
137+
endif
130138
DGEMMINCOPYOBJ = dgemm_incopy$(TSUFFIX).$(SUFFIX)
131139
DGEMMITCOPYOBJ = dgemm_itcopy$(TSUFFIX).$(SUFFIX)
132140
endif
@@ -258,8 +266,13 @@ endif
258266

259267
ifeq ($(BUILD_HFLOAT16), 1)
260268
SHGEMMKERNEL = shgemm_kernel_$(SHGEMM_UNROLL_M)x$(SHGEMM_UNROLL_N)_zvl128b.c
269+
ifneq ($(filter $(SHGEMM_UNROLL_N),8 16),)
270+
SHGEMMONCOPY = gemm_ncopy_$(SHGEMM_UNROLL_N)fp_rvv.c
271+
SHGEMMOTCOPY = gemm_tcopy_$(SHGEMM_UNROLL_N)fp_rvv.c
272+
else
261273
SHGEMMONCOPY = ../generic/gemm_ncopy_$(SHGEMM_UNROLL_N).c
262274
SHGEMMOTCOPY = ../generic/gemm_tcopy_$(SHGEMM_UNROLL_N).c
275+
endif
263276
SHGEMMONCOPYOBJ = shgemm_oncopy$(TSUFFIX).$(SUFFIX)
264277
SHGEMMOTCOPYOBJ = shgemm_otcopy$(TSUFFIX).$(SUFFIX)
265278
ifndef SHGEMM_BETA
@@ -269,8 +282,13 @@ endif
269282

270283
ifeq ($(BUILD_BFLOAT16), 1)
271284
SBGEMMKERNEL = sbgemm_kernel_$(SBGEMM_UNROLL_M)x$(SBGEMM_UNROLL_N)_zvl128b.c
285+
ifneq ($(filter $(SBGEMM_UNROLL_N),8 16),)
286+
SBGEMMONCOPY = gemm_ncopy_$(SBGEMM_UNROLL_N)fp_rvv.c
287+
SBGEMMOTCOPY = gemm_tcopy_$(SBGEMM_UNROLL_N)fp_rvv.c
288+
else
272289
SBGEMMONCOPY = ../generic/gemm_ncopy_$(SBGEMM_UNROLL_N).c
273290
SBGEMMOTCOPY = ../generic/gemm_tcopy_$(SBGEMM_UNROLL_N).c
291+
endif
274292
SBGEMMONCOPYOBJ = sbgemm_oncopy$(TSUFFIX).$(SUFFIX)
275293
SBGEMMOTCOPYOBJ = sbgemm_otcopy$(TSUFFIX).$(SUFFIX)
276294
ifndef SBGEMM_BETA

‎kernel/riscv64/KERNEL.RISCV64_ZVL256B‎

Lines changed: 28 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -119,12 +119,20 @@ endif
119119

120120
DGEMMKERNEL = dgemm_kernel_$(DGEMM_UNROLL_M)x$(DGEMM_UNROLL_N)_zvl256b.c
121121
DGEMMONCOPY = ../generic/gemm_ncopy_$(DGEMM_UNROLL_N).c
122+
ifneq ($(filter $(DGEMM_UNROLL_N),4 8 16),)
123+
DGEMMOTCOPY = gemm_tcopy_$(DGEMM_UNROLL_N)_rvv.c
124+
else
122125
DGEMMOTCOPY = ../generic/gemm_tcopy_$(DGEMM_UNROLL_N).c
126+
endif
123127
DGEMMONCOPYOBJ = dgemm_oncopy$(TSUFFIX).$(SUFFIX)
124128
DGEMMOTCOPYOBJ = dgemm_otcopy$(TSUFFIX).$(SUFFIX)
125129
ifneq ($(DGEMM_UNROLL_M), $(DGEMM_UNROLL_N))
126130
DGEMMINCOPY = ../generic/gemm_ncopy_$(DGEMM_UNROLL_M).c
131+
ifneq ($(filter $(DGEMM_UNROLL_M),4 8 16),)
132+
DGEMMITCOPY = gemm_tcopy_$(DGEMM_UNROLL_M)_rvv.c
133+
else
127134
DGEMMITCOPY = ../generic/gemm_tcopy_$(DGEMM_UNROLL_M).c
135+
endif
128136
DGEMMINCOPYOBJ = dgemm_incopy$(TSUFFIX).$(SUFFIX)
129137
DGEMMITCOPYOBJ = dgemm_itcopy$(TSUFFIX).$(SUFFIX)
130138
endif
@@ -226,13 +234,23 @@ SOMATCOPY_CT = omatcopy_ct_rvv.c
226234
ifeq ($(BUILD_HFLOAT16), 1)
227235
SHGEMMKERNEL = shgemm_kernel_$(SHGEMM_UNROLL_M)x$(SHGEMM_UNROLL_N)_zvl256b.c
228236
ifneq ($(SHGEMM_UNROLL_M), $(SHGEMM_UNROLL_N))
237+
ifneq ($(filter $(SHGEMM_UNROLL_M),8 16),)
238+
SHGEMMINCOPY = gemm_ncopy_$(SHGEMM_UNROLL_M)fp_rvv.c
239+
SHGEMMITCOPY = gemm_tcopy_$(SHGEMM_UNROLL_M)fp_rvv.c
240+
else
229241
SHGEMMINCOPY = ../generic/gemm_ncopy_$(SHGEMM_UNROLL_M).c
230242
SHGEMMITCOPY = ../generic/gemm_tcopy_$(SHGEMM_UNROLL_M).c
243+
endif
231244
SHGEMMINCOPYOBJ = shgemm_incopy$(TSUFFIX).$(SUFFIX)
232245
SHGEMMITCOPYOBJ = shgemm_itcopy$(TSUFFIX).$(SUFFIX)
233246
endif
247+
ifneq ($(filter $(SHGEMM_UNROLL_N),8 16),)
248+
SHGEMMONCOPY = gemm_ncopy_$(SHGEMM_UNROLL_N)fp_rvv.c
249+
SHGEMMOTCOPY = gemm_tcopy_$(SHGEMM_UNROLL_N)fp_rvv.c
250+
else
234251
SHGEMMONCOPY = ../generic/gemm_ncopy_$(SHGEMM_UNROLL_N).c
235252
SHGEMMOTCOPY = ../generic/gemm_tcopy_$(SHGEMM_UNROLL_N).c
253+
endif
236254
SHGEMMONCOPYOBJ = shgemm_oncopy$(TSUFFIX).$(SUFFIX)
237255
SHGEMMOTCOPYOBJ = shgemm_otcopy$(TSUFFIX).$(SUFFIX)
238256
ifndef SHGEMM_BETA
@@ -243,13 +261,23 @@ endif
243261
ifeq ($(BUILD_BFLOAT16), 1)
244262
SBGEMMKERNEL = sbgemm_kernel_$(SBGEMM_UNROLL_M)x$(SBGEMM_UNROLL_N)_zvl256b.c
245263
ifneq ($(SBGEMM_UNROLL_M), $(SBGEMM_UNROLL_N))
264+
ifneq ($(filter $(SBGEMM_UNROLL_M),8 16),)
265+
SBGEMMINCOPY = gemm_ncopy_$(SBGEMM_UNROLL_M)fp_rvv.c
266+
SBGEMMITCOPY = gemm_tcopy_$(SBGEMM_UNROLL_M)fp_rvv.c
267+
else
246268
SBGEMMINCOPY = ../generic/gemm_ncopy_$(SBGEMM_UNROLL_M).c
247269
SBGEMMITCOPY = ../generic/gemm_tcopy_$(SBGEMM_UNROLL_M).c
270+
endif
248271
SBGEMMINCOPYOBJ = sbgemm_incopy$(TSUFFIX).$(SUFFIX)
249272
SBGEMMITCOPYOBJ = sbgemm_itcopy$(TSUFFIX).$(SUFFIX)
250273
endif
274+
ifneq ($(filter $(SBGEMM_UNROLL_N),8 16),)
275+
SBGEMMONCOPY = gemm_ncopy_$(SBGEMM_UNROLL_N)fp_rvv.c
276+
SBGEMMOTCOPY = gemm_tcopy_$(SBGEMM_UNROLL_N)fp_rvv.c
277+
else
251278
SBGEMMONCOPY = ../generic/gemm_ncopy_$(SBGEMM_UNROLL_N).c
252279
SBGEMMOTCOPY = ../generic/gemm_tcopy_$(SBGEMM_UNROLL_N).c
280+
endif
253281
SBGEMMONCOPYOBJ = sbgemm_oncopy$(TSUFFIX).$(SUFFIX)
254282
SBGEMMOTCOPYOBJ = sbgemm_otcopy$(TSUFFIX).$(SUFFIX)
255283
ifndef SBGEMM_BETA

0 commit comments

Comments
 (0)