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

Add copies of GEMMT under its new name GEMMTR

This commit is contained in:
Martin Kroeker 2024-10-30 12:55:14 +01:00 committed by GitHub
parent f66e6d32c2
commit 0cf656fd3e
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
2 changed files with 43 additions and 11 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
interface/CMakeLists.txt
View File

@ -107,6 +107,9 @@ endif ()
# trmm is trsm with a compiler flag set # trmm is trsm with a compiler flag set
GenerateNamedObjects("trsm.c" "TRMM" "trmm" ${CBLAS_FLAG}) GenerateNamedObjects("trsm.c" "TRMM" "trmm" ${CBLAS_FLAG})
# gemmtr is gemmt under the name adopted by the Reference BLAS
GenerateNamedObjects("gemm.c" "" "gemmtr" ${CBLAS_FLAG})
# max and imax are compiled 4 times # max and imax are compiled 4 times
GenerateNamedObjects("max.c" "" "" ${CBLAS_FLAG}) GenerateNamedObjects("max.c" "" "" ${CBLAS_FLAG})
@ -123,6 +126,7 @@ if (BUILD_BFLOAT16)
GenerateNamedObjects("bf16dot.c" "" "sbdot" ${CBLAS_FLAG} "" "" true "BFLOAT16") GenerateNamedObjects("bf16dot.c" "" "sbdot" ${CBLAS_FLAG} "" "" true "BFLOAT16")
GenerateNamedObjects("gemm.c" "" "sbgemm" ${CBLAS_FLAG} "" "" true "BFLOAT16") GenerateNamedObjects("gemm.c" "" "sbgemm" ${CBLAS_FLAG} "" "" true "BFLOAT16")
GenerateNamedObjects("gemmt.c" "" "sbgemmt" ${CBLAS_FLAG} "" "" true "BFLOAT16") GenerateNamedObjects("gemmt.c" "" "sbgemmt" ${CBLAS_FLAG} "" "" true "BFLOAT16")
GenerateNamedObjects("gemmt.c" "" "sbgemmtr" ${CBLAS_FLAG} "" "" true "BFLOAT16")
GenerateNamedObjects("sbgemv.c" "" "sbgemv" ${CBLAS_FLAG} "" "" true "BFLOAT16") GenerateNamedObjects("sbgemv.c" "" "sbgemv" ${CBLAS_FLAG} "" "" true "BFLOAT16")
GenerateNamedObjects("tobf16.c" "SINGLE_PREC" "sbstobf16" ${CBLAS_FLAG} "" "" true "BFLOAT16") GenerateNamedObjects("tobf16.c" "SINGLE_PREC" "sbstobf16" ${CBLAS_FLAG} "" "" true "BFLOAT16")
GenerateNamedObjects("tobf16.c" "DOUBLE_PREC" "sbdtobf16" ${CBLAS_FLAG} "" "" true "BFLOAT16") GenerateNamedObjects("tobf16.c" "DOUBLE_PREC" "sbdtobf16" ${CBLAS_FLAG} "" "" true "BFLOAT16")

50
interface/Makefile
View File

@ -44,12 +44,12 @@ SBLAS3OBJS = \
sgemm.$(SUFFIX) ssymm.$(SUFFIX) strmm.$(SUFFIX) \ sgemm.$(SUFFIX) ssymm.$(SUFFIX) strmm.$(SUFFIX) \
strsm.$(SUFFIX) ssyrk.$(SUFFIX) ssyr2k.$(SUFFIX) \ strsm.$(SUFFIX) ssyrk.$(SUFFIX) ssyr2k.$(SUFFIX) \
somatcopy.$(SUFFIX) simatcopy.$(SUFFIX)\ somatcopy.$(SUFFIX) simatcopy.$(SUFFIX)\
sgeadd.$(SUFFIX) sgemmt.$(SUFFIX) sgeadd.$(SUFFIX) sgemmt.$(SUFFIX) sgemmtr.$(SUFFIX)
ifeq ($(BUILD_BFLOAT16),1) ifeq ($(BUILD_BFLOAT16),1)
SBBLAS1OBJS = sbdot.$(SUFFIX) SBBLAS1OBJS = sbdot.$(SUFFIX)
SBBLAS2OBJS = sbgemv.$(SUFFIX) SBBLAS2OBJS = sbgemv.$(SUFFIX)
SBBLAS3OBJS = sbgemm.$(SUFFIX) sbgemmt.$(SUFFIX) SBBLAS3OBJS = sbgemm.$(SUFFIX) sbgemmt.$(SUFFIX) sbgemmtr.$(SUFFIX)
SBEXTOBJS = sbstobf16.$(SUFFIX) sbdtobf16.$(SUFFIX) sbf16tos.$(SUFFIX) dbf16tod.$(SUFFIX) SBEXTOBJS = sbstobf16.$(SUFFIX) sbdtobf16.$(SUFFIX) sbf16tos.$(SUFFIX) dbf16tod.$(SUFFIX)
endif endif
@ -76,7 +76,7 @@ DBLAS3OBJS = \
dgemm.$(SUFFIX) dsymm.$(SUFFIX) dtrmm.$(SUFFIX) \ dgemm.$(SUFFIX) dsymm.$(SUFFIX) dtrmm.$(SUFFIX) \
dtrsm.$(SUFFIX) dsyrk.$(SUFFIX) dsyr2k.$(SUFFIX) \ dtrsm.$(SUFFIX) dsyrk.$(SUFFIX) dsyr2k.$(SUFFIX) \
domatcopy.$(SUFFIX) dimatcopy.$(SUFFIX)\ domatcopy.$(SUFFIX) dimatcopy.$(SUFFIX)\
dgeadd.$(SUFFIX) dgemmt.$(SUFFIX) dgeadd.$(SUFFIX) dgemmt.$(SUFFIX) dgemmtr.$(SUFFIX)
CBLAS1OBJS = \ CBLAS1OBJS = \
caxpy.$(SUFFIX) caxpyc.$(SUFFIX) cswap.$(SUFFIX) \ caxpy.$(SUFFIX) caxpyc.$(SUFFIX) cswap.$(SUFFIX) \
@ -105,7 +105,7 @@ CBLAS3OBJS = \
ctrsm.$(SUFFIX) csyrk.$(SUFFIX) csyr2k.$(SUFFIX) \ ctrsm.$(SUFFIX) csyrk.$(SUFFIX) csyr2k.$(SUFFIX) \
chemm.$(SUFFIX) cherk.$(SUFFIX) cher2k.$(SUFFIX) \ chemm.$(SUFFIX) cherk.$(SUFFIX) cher2k.$(SUFFIX) \
comatcopy.$(SUFFIX) cimatcopy.$(SUFFIX)\ comatcopy.$(SUFFIX) cimatcopy.$(SUFFIX)\
cgeadd.$(SUFFIX) cgemmt.$(SUFFIX) cgeadd.$(SUFFIX) cgemmt.$(SUFFIX) cgemmtr.$(SUFFIX)
ZBLAS1OBJS = \ ZBLAS1OBJS = \
zaxpy.$(SUFFIX) zaxpyc.$(SUFFIX) zswap.$(SUFFIX) \ zaxpy.$(SUFFIX) zaxpyc.$(SUFFIX) zswap.$(SUFFIX) \
@ -134,7 +134,7 @@ ZBLAS3OBJS = \
ztrsm.$(SUFFIX) zsyrk.$(SUFFIX) zsyr2k.$(SUFFIX) \ ztrsm.$(SUFFIX) zsyrk.$(SUFFIX) zsyr2k.$(SUFFIX) \
zhemm.$(SUFFIX) zherk.$(SUFFIX) zher2k.$(SUFFIX) \ zhemm.$(SUFFIX) zherk.$(SUFFIX) zher2k.$(SUFFIX) \
zomatcopy.$(SUFFIX) zimatcopy.$(SUFFIX)\ zomatcopy.$(SUFFIX) zimatcopy.$(SUFFIX)\
zgeadd.$(SUFFIX) zgemmt.$(SUFFIX) zgeadd.$(SUFFIX) zgemmt.$(SUFFIX) zgemmtr.$(SUFFIX)
ifeq ($(SUPPORT_GEMM3M), 1) ifeq ($(SUPPORT_GEMM3M), 1)
@ -282,12 +282,12 @@ CSBLAS2OBJS = \
CSBLAS3OBJS = \ CSBLAS3OBJS = \
cblas_sgemm.$(SUFFIX) cblas_ssymm.$(SUFFIX) cblas_strmm.$(SUFFIX) cblas_strsm.$(SUFFIX) \ cblas_sgemm.$(SUFFIX) cblas_ssymm.$(SUFFIX) cblas_strmm.$(SUFFIX) cblas_strsm.$(SUFFIX) \
cblas_ssyrk.$(SUFFIX) cblas_ssyr2k.$(SUFFIX) cblas_somatcopy.$(SUFFIX) cblas_simatcopy.$(SUFFIX)\ cblas_ssyrk.$(SUFFIX) cblas_ssyr2k.$(SUFFIX) cblas_somatcopy.$(SUFFIX) cblas_simatcopy.$(SUFFIX)\
cblas_sgeadd.$(SUFFIX) cblas_sgemmt.$(SUFFIX) cblas_sgemm_batch.$(SUFFIX) cblas_sgeadd.$(SUFFIX) cblas_sgemmt.$(SUFFIX) cblas_sgemmtr.$(SUFFIX) cblas_sgemm_batch.$(SUFFIX)
ifeq ($(BUILD_BFLOAT16),1) ifeq ($(BUILD_BFLOAT16),1)
CSBBLAS1OBJS = cblas_sbdot.$(SUFFIX) CSBBLAS1OBJS = cblas_sbdot.$(SUFFIX)
CSBBLAS2OBJS = cblas_sbgemv.$(SUFFIX) CSBBLAS2OBJS = cblas_sbgemv.$(SUFFIX)
CSBBLAS3OBJS = cblas_sbgemm.$(SUFFIX) cblas_sbgemmt.$(SUFFIX) cblas_sbgemm_batch.$(SUFFIX) CSBBLAS3OBJS = cblas_sbgemm.$(SUFFIX) cblas_sbgemmt.$(SUFFIX) cblas_sbgemmtr.$(SUFFIX) cblas_sbgemm_batch.$(SUFFIX)
CSBEXTOBJS = cblas_sbstobf16.$(SUFFIX) cblas_sbdtobf16.$(SUFFIX) cblas_sbf16tos.$(SUFFIX) cblas_dbf16tod.$(SUFFIX) CSBEXTOBJS = cblas_sbstobf16.$(SUFFIX) cblas_sbdtobf16.$(SUFFIX) cblas_sbf16tos.$(SUFFIX) cblas_dbf16tod.$(SUFFIX)
endif endif
@ -308,7 +308,7 @@ CDBLAS2OBJS = \
CDBLAS3OBJS += \ CDBLAS3OBJS += \
cblas_dgemm.$(SUFFIX) cblas_dsymm.$(SUFFIX) cblas_dtrmm.$(SUFFIX) cblas_dtrsm.$(SUFFIX) \ cblas_dgemm.$(SUFFIX) cblas_dsymm.$(SUFFIX) cblas_dtrmm.$(SUFFIX) cblas_dtrsm.$(SUFFIX) \
cblas_dsyrk.$(SUFFIX) cblas_dsyr2k.$(SUFFIX) cblas_domatcopy.$(SUFFIX) cblas_dimatcopy.$(SUFFIX) \ cblas_dsyrk.$(SUFFIX) cblas_dsyr2k.$(SUFFIX) cblas_domatcopy.$(SUFFIX) cblas_dimatcopy.$(SUFFIX) \
cblas_dgeadd.$(SUFFIX) cblas_dgemmt.$(SUFFIX) cblas_dgemm_batch.$(SUFFIX) cblas_dgeadd.$(SUFFIX) cblas_dgemmt.$(SUFFIX) cblas_dgemmtr.$(SUFFIX) cblas_dgemm_batch.$(SUFFIX)
CCBLAS1OBJS = \ CCBLAS1OBJS = \
cblas_icamax.$(SUFFIX) cblas_icamin.$(SUFFIX) cblas_scasum.$(SUFFIX) cblas_caxpy.$(SUFFIX) \ cblas_icamax.$(SUFFIX) cblas_icamin.$(SUFFIX) cblas_scasum.$(SUFFIX) cblas_caxpy.$(SUFFIX) \
@ -333,7 +333,7 @@ CCBLAS3OBJS = \
cblas_csyrk.$(SUFFIX) cblas_csyr2k.$(SUFFIX) \ cblas_csyrk.$(SUFFIX) cblas_csyr2k.$(SUFFIX) \
cblas_chemm.$(SUFFIX) cblas_cherk.$(SUFFIX) cblas_cher2k.$(SUFFIX) \ cblas_chemm.$(SUFFIX) cblas_cherk.$(SUFFIX) cblas_cher2k.$(SUFFIX) \
cblas_comatcopy.$(SUFFIX) cblas_cimatcopy.$(SUFFIX)\ cblas_comatcopy.$(SUFFIX) cblas_cimatcopy.$(SUFFIX)\
cblas_cgeadd.$(SUFFIX) cblas_cgemmt.$(SUFFIX) cblas_cgemm_batch.$(SUFFIX) cblas_cgeadd.$(SUFFIX) cblas_cgemmt.$(SUFFIX) cblas_cgemmtr.$(SUFFIX) cblas_cgemm_batch.$(SUFFIX)
CXERBLAOBJ = \ CXERBLAOBJ = \
cblas_xerbla.$(SUFFIX) cblas_xerbla.$(SUFFIX)
@ -364,7 +364,7 @@ CZBLAS3OBJS = \
cblas_zsyrk.$(SUFFIX) cblas_zsyr2k.$(SUFFIX) \ cblas_zsyrk.$(SUFFIX) cblas_zsyr2k.$(SUFFIX) \
cblas_zhemm.$(SUFFIX) cblas_zherk.$(SUFFIX) cblas_zher2k.$(SUFFIX)\ cblas_zhemm.$(SUFFIX) cblas_zherk.$(SUFFIX) cblas_zher2k.$(SUFFIX)\
cblas_zomatcopy.$(SUFFIX) cblas_zimatcopy.$(SUFFIX) \ cblas_zomatcopy.$(SUFFIX) cblas_zimatcopy.$(SUFFIX) \
cblas_zgeadd.$(SUFFIX) cblas_zgemmt.$(SUFFIX) cblas_zgemm_batch.$(SUFFIX) cblas_zgeadd.$(SUFFIX) cblas_zgemmt.$(SUFFIX) cblas_zgemmtr.$(SUFFIX) cblas_zgemm_batch.$(SUFFIX)
ifeq ($(SUPPORT_GEMM3M), 1) ifeq ($(SUPPORT_GEMM3M), 1)
@ -1305,6 +1305,8 @@ sbgemm.$(SUFFIX) sbgemm.$(PSUFFIX) : gemm.c ../param.h
$(CC) -c $(CFLAGS) $< -o $(@F) $(CC) -c $(CFLAGS) $< -o $(@F)
sbgemmt.$(SUFFIX) sbgemmt.$(PSUFFIX) : sbgemmt.c ../param.h sbgemmt.$(SUFFIX) sbgemmt.$(PSUFFIX) : sbgemmt.c ../param.h
$(CC) -c $(CFLAGS) $< -o $(@F) $(CC) -c $(CFLAGS) $< -o $(@F)
sbgemmtr.$(SUFFIX) sbgemmtr.$(PSUFFIX) : sbgemmt.c ../param.h
$(CC) -c $(CFLAGS) $< -o $(@F)
endif endif
sgemm.$(SUFFIX) sgemm.$(PSUFFIX) : gemm.c ../param.h sgemm.$(SUFFIX) sgemm.$(PSUFFIX) : gemm.c ../param.h
@ -1340,7 +1342,19 @@ cgemmt.$(SUFFIX) cgemmt.$(PSUFFIX) : gemmt.c ../param.h
zgemmt.$(SUFFIX) zgemmt.$(PSUFFIX) : gemmt.c ../param.h zgemmt.$(SUFFIX) zgemmt.$(PSUFFIX) : gemmt.c ../param.h
$(CC) -c $(CFLAGS) $< -o $(@F) $(CC) -c $(CFLAGS) $< -o $(@F)
xgemmt.$(SUFFIX) xgemmt.$(PSUFFIX) : gemmt.c ../param.h sgemmtr.$(SUFFIX) sgemmtr.$(PSUFFIX) : gemmt.c ../param.h
$(CC) -c $(CFLAGS) $< -o $(@F)
dgemmtr.$(SUFFIX) dgemmtr.$(PSUFFIX) : gemmt.c ../param.h
$(CC) -c $(CFLAGS) $< -o $(@F)
qgemmtr.$(SUFFIX) qgemmtr.$(PSUFFIX) : gemmt.c ../param.h
$(CC) -c $(CFLAGS) $< -o $(@F)
cgemmtr.$(SUFFIX) cgemmtr.$(PSUFFIX) : gemmt.c ../param.h
$(CC) -c $(CFLAGS) $< -o $(@F)
zgemmtr.$(SUFFIX) zgemmtr.$(PSUFFIX) : gemmt.c ../param.h
$(CC) -c $(CFLAGS) $< -o $(@F) $(CC) -c $(CFLAGS) $< -o $(@F)
ssymm.$(SUFFIX) ssymm.$(PSUFFIX) : symm.c ssymm.$(SUFFIX) ssymm.$(PSUFFIX) : symm.c
@ -1966,9 +1980,14 @@ cblas_zgemm.$(SUFFIX) cblas_zgemm.$(PSUFFIX) : gemm.c ../param.h
cblas_sgemmt.$(SUFFIX) cblas_sgemmt.$(PSUFFIX) : gemmt.c ../param.h cblas_sgemmt.$(SUFFIX) cblas_sgemmt.$(PSUFFIX) : gemmt.c ../param.h
$(CC) -DCBLAS -c $(CFLAGS) $< -o $(@F) $(CC) -DCBLAS -c $(CFLAGS) $< -o $(@F)
cblas_sgemmtr.$(SUFFIX) cblas_sgemmtr.$(PSUFFIX) : gemmt.c ../param.h
$(CC) -DCBLAS -c $(CFLAGS) $< -o $(@F)
ifeq ($(BUILD_BFLOAT16),1) ifeq ($(BUILD_BFLOAT16),1)
cblas_sbgemmt.$(SUFFIX) cblas_sbgemmt.$(PSUFFIX) : sbgemmt.c ../param.h cblas_sbgemmt.$(SUFFIX) cblas_sbgemmt.$(PSUFFIX) : sbgemmt.c ../param.h
$(CC) -DCBLAS -c $(CFLAGS) $< -o $(@F) $(CC) -DCBLAS -c $(CFLAGS) $< -o $(@F)
cblas_sbgemmtr.$(SUFFIX) cblas_sbgemmtr.$(PSUFFIX) : sbgemmt.c ../param.h
$(CC) -DCBLAS -c $(CFLAGS) $< -o $(@F)
endif endif
cblas_dgemmt.$(SUFFIX) cblas_dgemmt.$(PSUFFIX) : gemmt.c ../param.h cblas_dgemmt.$(SUFFIX) cblas_dgemmt.$(PSUFFIX) : gemmt.c ../param.h
@ -1980,6 +1999,15 @@ cblas_cgemmt.$(SUFFIX) cblas_cgemmt.$(PSUFFIX) : gemmt.c ../param.h
cblas_zgemmt.$(SUFFIX) cblas_zgemmt.$(PSUFFIX) : gemmt.c ../param.h cblas_zgemmt.$(SUFFIX) cblas_zgemmt.$(PSUFFIX) : gemmt.c ../param.h
$(CC) -DCBLAS -c $(CFLAGS) $< -o $(@F) $(CC) -DCBLAS -c $(CFLAGS) $< -o $(@F)
cblas_dgemmtr.$(SUFFIX) cblas_dgemmtr.$(PSUFFIX) : gemmt.c ../param.h
$(CC) -DCBLAS -c $(CFLAGS) $< -o $(@F)
cblas_cgemmtr.$(SUFFIX) cblas_cgemmtr.$(PSUFFIX) : gemmt.c ../param.h
$(CC) -DCBLAS -c $(CFLAGS) $< -o $(@F)
cblas_zgemmtr.$(SUFFIX) cblas_zgemmtr.$(PSUFFIX) : gemmt.c ../param.h
$(CC) -DCBLAS -c $(CFLAGS) $< -o $(@F)
cblas_ssymm.$(SUFFIX) cblas_ssymm.$(PSUFFIX) : symm.c cblas_ssymm.$(SUFFIX) cblas_ssymm.$(PSUFFIX) : symm.c
$(CC) -DCBLAS -c $(CFLAGS) $< -o $(@F) $(CC) -DCBLAS -c $(CFLAGS) $< -o $(@F)