232 lines
		
	
	
		
			9.0 KiB
		
	
	
	
		
			CMake
		
	
	
	
			
		
		
	
	
			232 lines
		
	
	
		
			9.0 KiB
		
	
	
	
		
			CMake
		
	
	
	
 | 
						|
include_directories(${PROJECT_SOURCE_DIR})
 | 
						|
include_directories(${PROJECT_BINARY_DIR})
 | 
						|
 | 
						|
 | 
						|
set(BLAS1_SOURCES
 | 
						|
  copy.c
 | 
						|
  nrm2.c
 | 
						|
)
 | 
						|
 | 
						|
set(BLAS1_REAL_ONLY_SOURCES
 | 
						|
  rotm.c rotmg.c # N.B. these do not have complex counterparts
 | 
						|
  rot.c
 | 
						|
  asum.c
 | 
						|
  sum.c
 | 
						|
)
 | 
						|
 | 
						|
# these will have 'z' prepended for the complex version
 | 
						|
set(BLAS1_MANGLED_SOURCES
 | 
						|
  axpy.c swap.c
 | 
						|
  scal.c
 | 
						|
  dot.c
 | 
						|
  rotg.c
 | 
						|
  axpby.c
 | 
						|
)
 | 
						|
 | 
						|
# TODO: USE_NETLIB_GEMV should switch gemv.c to netlib/*gemv.f
 | 
						|
# these all have 'z' sources for complex versions
 | 
						|
set(BLAS2_SOURCES
 | 
						|
  gemv.c ger.c
 | 
						|
  trsv.c trmv.c 
 | 
						|
  syr2.c gbmv.c
 | 
						|
  sbmv.c 
 | 
						|
  spr2.c
 | 
						|
  tbsv.c tbmv.c
 | 
						|
  tpsv.c tpmv.c
 | 
						|
)
 | 
						|
 | 
						|
set(BLAS2_REAL_ONLY_SOURCES
 | 
						|
  symv.c syr.c spmv.c spr.c
 | 
						|
)
 | 
						|
set(BLAS2_COMPLEX_LAPACK_SOURCES
 | 
						|
  symv.c syr.c spmv.c spr.c
 | 
						|
)
 | 
						|
 | 
						|
set(BLAS2_COMPLEX_ONLY_MANGLED_SOURCES
 | 
						|
  hemv.c hbmv.c
 | 
						|
  her.c her2.c
 | 
						|
  hpmv.c hpr.c
 | 
						|
  hpr2.c
 | 
						|
)
 | 
						|
 | 
						|
# these do not have separate 'z' sources
 | 
						|
set(BLAS3_SOURCES
 | 
						|
  gemm.c symm.c
 | 
						|
  trsm.c syrk.c syr2k.c gemmt.c
 | 
						|
)
 | 
						|
 | 
						|
set(BLAS3_MANGLED_SOURCES
 | 
						|
  omatcopy.c imatcopy.c
 | 
						|
  geadd.c
 | 
						|
)
 | 
						|
 | 
						|
# generate the BLAS objs once with and once without cblas
 | 
						|
set (CBLAS_FLAGS "")
 | 
						|
 | 
						|
if (NOT DEFINED NO_FBLAS)
 | 
						|
  list(APPEND CBLAS_FLAGS 0)
 | 
						|
endif ()
 | 
						|
 | 
						|
if (NOT DEFINED NO_CBLAS)
 | 
						|
  list(APPEND CBLAS_FLAGS 1)
 | 
						|
endif ()
 | 
						|
 | 
						|
foreach (CBLAS_FLAG ${CBLAS_FLAGS})
 | 
						|
 | 
						|
  # TODO: don't compile complex sources with cblas for now, the naming schemes are all different and they will have to be handled separately from SINGLE/DOUBLE
 | 
						|
  set(DISABLE_COMPLEX 0)
 | 
						|
  set(MANGLE_COMPLEX 3)
 | 
						|
  if (CBLAS_FLAG EQUAL 1)
 | 
						|
#    set(DISABLE_COMPLEX 1)
 | 
						|
#    set(MANGLE_COMPLEX 1)
 | 
						|
  endif ()
 | 
						|
  GenerateNamedObjects("${BLAS1_SOURCES}" "" "" ${CBLAS_FLAG} "" "" false ${DISABLE_COMPLEX})
 | 
						|
  GenerateNamedObjects("${BLAS1_REAL_ONLY_SOURCES}" "" "" ${CBLAS_FLAG} "" "" false 1)
 | 
						|
  GenerateNamedObjects("${BLAS1_MANGLED_SOURCES}" "" "" ${CBLAS_FLAG} "" "" false ${MANGLE_COMPLEX})
 | 
						|
  GenerateNamedObjects("${BLAS2_SOURCES}" "" "" ${CBLAS_FLAG} "" "" false ${MANGLE_COMPLEX})
 | 
						|
  GenerateNamedObjects("${BLAS2_REAL_ONLY_SOURCES}" "" "" ${CBLAS_FLAG} "" "" false 1)
 | 
						|
  if (NOT DEFINED NO_LAPACK)
 | 
						|
  GenerateNamedObjects("${BLAS2_COMPLEX_LAPACK_SOURCES}" "" "" ${CBLAS_FLAG} "" "" false ${MANGLE_COMPLEX})
 | 
						|
  endif ()
 | 
						|
  GenerateNamedObjects("${BLAS2_COMPLEX_ONLY_MANGLED_SOURCES}" "" "" ${CBLAS_FLAG} "" "" false 4)
 | 
						|
  GenerateNamedObjects("${BLAS3_SOURCES}" "" "" ${CBLAS_FLAG} "" "" false ${DISABLE_COMPLEX})
 | 
						|
  GenerateNamedObjects("${BLAS3_MANGLED_SOURCES}" "" "" ${CBLAS_FLAG} "" "" false ${MANGLE_COMPLEX})
 | 
						|
 | 
						|
  GenerateNamedObjects("xerbla.c" "" "xerbla" ${CBLAS_FLAG} "" "" true)
 | 
						|
  #sdsdot, dsdot
 | 
						|
  if (BUILD_SINGLE OR BUILD_DOUBLE)
 | 
						|
  GenerateNamedObjects("sdsdot.c" "" "sdsdot" ${CBLAS_FLAG} "" "" true "SINGLE")
 | 
						|
endif ()
 | 
						|
if (BUILD_DOUBLE)
 | 
						|
  GenerateNamedObjects("dsdot.c" "" "dsdot" ${CBLAS_FLAG} "" "" true "SINGLE")
 | 
						|
endif ()
 | 
						|
 | 
						|
  # trmm is trsm with a compiler flag set
 | 
						|
  GenerateNamedObjects("trsm.c" "TRMM" "trmm" ${CBLAS_FLAG})
 | 
						|
 | 
						|
  # max and imax are compiled 4 times
 | 
						|
  GenerateNamedObjects("max.c" "" "" ${CBLAS_FLAG})
 | 
						|
  GenerateNamedObjects("max.c" "USE_ABS" "amax" ${CBLAS_FLAG})
 | 
						|
  GenerateNamedObjects("max.c" "USE_ABS;USE_MIN" "amin" ${CBLAS_FLAG})
 | 
						|
  GenerateNamedObjects("max.c" "USE_MIN" "min" ${CBLAS_FLAG})
 | 
						|
 | 
						|
  GenerateNamedObjects("imax.c" "" "i*max" ${CBLAS_FLAG})
 | 
						|
  GenerateNamedObjects("imax.c" "USE_ABS" "i*amax" ${CBLAS_FLAG})
 | 
						|
  GenerateNamedObjects("imax.c" "USE_ABS;USE_MIN" "i*amin" ${CBLAS_FLAG})
 | 
						|
  GenerateNamedObjects("imax.c" "USE_MIN" "i*min" ${CBLAS_FLAG})
 | 
						|
 | 
						|
if (BUILD_BFLOAT16)
 | 
						|
	GenerateNamedObjects("bf16dot.c" "" "sbdot" ${CBLAS_FLAG} "" "" true "BFLOAT16")
 | 
						|
	GenerateNamedObjects("gemm.c" "" "sbgemm" ${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" "DOUBLE_PREC" "sbdtobf16" ${CBLAS_FLAG} "" "" true "BFLOAT16")
 | 
						|
	GenerateNamedObjects("bf16to.c" "SINGLE_PREC" "sbf16tos" ${CBLAS_FLAG} "" "" true "BFLOAT16")
 | 
						|
	GenerateNamedObjects("bf16to.c" "DOUBLE_PREC" "dbf16tod" ${CBLAS_FLAG} "" "" true "BFLOAT16")
 | 
						|
endif ()
 | 
						|
 | 
						|
# complex-specific sources
 | 
						|
foreach (float_type ${FLOAT_TYPES})
 | 
						|
 | 
						|
  if (${float_type} STREQUAL "COMPLEX" OR ${float_type} STREQUAL "ZCOMPLEX")
 | 
						|
    GenerateNamedObjects("zger.c" "" "geru" ${CBLAS_FLAG} "" "" false ${float_type})
 | 
						|
    GenerateNamedObjects("zger.c" "CONJ" "gerc" ${CBLAS_FLAG} "" "" false ${float_type})
 | 
						|
    GenerateNamedObjects("zdot.c" "CONJ" "dotc" ${CBLAS_FLAG} "" "" false ${float_type})
 | 
						|
    GenerateNamedObjects("zdot.c" "" "dotu" ${CBLAS_FLAG} "" "" false ${float_type})
 | 
						|
 | 
						|
    GenerateNamedObjects("symm.c" "HEMM" "hemm" ${CBLAS_FLAG} "" "" false ${float_type})
 | 
						|
    GenerateNamedObjects("syrk.c" "HEMM" "herk" ${CBLAS_FLAG} "" "" false ${float_type})
 | 
						|
    GenerateNamedObjects("syr2k.c" "HEMM" "her2k" ${CBLAS_FLAG} "" "" false ${float_type})
 | 
						|
 | 
						|
    if (USE_GEMM3M)
 | 
						|
      GenerateNamedObjects("gemm.c" "GEMM3M" "gemm3m" ${CBLAS_FLAG} "" "" false ${float_type})
 | 
						|
    endif()
 | 
						|
  endif ()
 | 
						|
  if (${float_type} STREQUAL "COMPLEX")
 | 
						|
    GenerateNamedObjects("zscal.c" "SSCAL" "sscal" ${CBLAS_FLAG} "" "" false "COMPLEX")
 | 
						|
    GenerateNamedObjects("nrm2.c" "" "scnrm2" ${CBLAS_FLAG} "" "" true "COMPLEX")
 | 
						|
    GenerateNamedObjects("zrot.c" "" "csrot" ${CBLAS_FLAG} "" "" true "COMPLEX")
 | 
						|
    GenerateNamedObjects("max.c" "USE_ABS;USE_MIN" "scamin" ${CBLAS_FLAG} "" "" true "COMPLEX")
 | 
						|
    GenerateNamedObjects("max.c" "USE_ABS" "scamax" ${CBLAS_FLAG} "" "" true "COMPLEX")
 | 
						|
    GenerateNamedObjects("asum.c" "" "scasum" ${CBLAS_FLAG} "" "" true "COMPLEX")
 | 
						|
    GenerateNamedObjects("sum.c" "" "scsum" ${CBLAS_FLAG} "" "" true "COMPLEX")
 | 
						|
  endif ()
 | 
						|
  if (${float_type} STREQUAL "ZCOMPLEX")
 | 
						|
    GenerateNamedObjects("zscal.c" "SSCAL" "dscal" ${CBLAS_FLAG} "" "" false "ZCOMPLEX")
 | 
						|
    GenerateNamedObjects("nrm2.c" "" "dznrm2" ${CBLAS_FLAG} "" "" true "ZCOMPLEX")
 | 
						|
    GenerateNamedObjects("zrot.c" "" "zdrot" ${CBLAS_FLAG} "" "" true "ZCOMPLEX")
 | 
						|
    GenerateNamedObjects("max.c" "USE_ABS;USE_MIN" "dzamin" ${CBLAS_FLAG} "" "" true "ZCOMPLEX")
 | 
						|
    GenerateNamedObjects("max.c" "USE_ABS" "dzamax" ${CBLAS_FLAG} "" "" true "ZCOMPLEX")
 | 
						|
    GenerateNamedObjects("asum.c" "" "dzasum" ${CBLAS_FLAG} "" "" true "ZCOMPLEX")
 | 
						|
    GenerateNamedObjects("sum.c" "" "dzsum" ${CBLAS_FLAG} "" "" true "ZCOMPLEX")
 | 
						|
  endif ()
 | 
						|
endforeach ()
 | 
						|
 | 
						|
endforeach ()
 | 
						|
 | 
						|
#Special functions for CBLAS
 | 
						|
if (NOT DEFINED NO_CBLAS)
 | 
						|
  foreach (float_type ${FLOAT_TYPES})
 | 
						|
  if (${float_type} STREQUAL "COMPLEX" OR ${float_type} STREQUAL "ZCOMPLEX")
 | 
						|
    #cblas_dotc_sub cblas_dotu_sub
 | 
						|
    GenerateNamedObjects("zdot.c" "FORCE_USE_STACK" "dotu_sub" 1 "" "" false ${float_type})
 | 
						|
    GenerateNamedObjects("zdot.c" "FORCE_USE_STACK;CONJ" "dotc_sub" 1 "" "" false ${float_type})
 | 
						|
  endif()
 | 
						|
  endforeach ()
 | 
						|
endif()
 | 
						|
 | 
						|
if (NOT DEFINED NO_LAPACK)
 | 
						|
  set(LAPACK_SOURCES
 | 
						|
    lapack/gesv.c
 | 
						|
  )
 | 
						|
 | 
						|
  # prepend z for complex versions
 | 
						|
  set(LAPACK_MANGLED_SOURCES
 | 
						|
    lapack/getrf.c lapack/getrs.c lapack/potrf.c lapack/getf2.c
 | 
						|
    lapack/potf2.c lapack/laswp.c lapack/lauu2.c
 | 
						|
    lapack/lauum.c lapack/trti2.c lapack/trtri.c
 | 
						|
  )
 | 
						|
 | 
						|
  GenerateNamedObjects("${LAPACK_SOURCES}")
 | 
						|
  if (NOT RELAPACK_REPLACE)
 | 
						|
  GenerateNamedObjects("${LAPACK_MANGLED_SOURCES}" "" "" 0 "" "" 0 3)
 | 
						|
  else ()
 | 
						|
  GenerateNamedObjects("lapack/getrs.c" "" "" 0 "" "" 0 3)
 | 
						|
  GenerateNamedObjects("lapack/getf2.c" "" "" 0 "" "" 0 3)
 | 
						|
  GenerateNamedObjects("lapack/potf2.c" "" "" 0 "" "" 0 3)
 | 
						|
  GenerateNamedObjects("lapack/laswp.c" "" "" 0 "" "" 0 3)
 | 
						|
  GenerateNamedObjects("lapack/lauu2.c" "" "" 0 "" "" 0 3)
 | 
						|
  GenerateNamedObjects("lapack/trti2.c" "" "" 0 "" "" 0 3)
 | 
						|
  endif()
 | 
						|
endif ()
 | 
						|
 | 
						|
if ( BUILD_COMPLEX AND NOT  BUILD_SINGLE)
 | 
						|
	GenerateNamedObjects("scal.c" "" "scal" 0 "" "" false "SINGLE")
 | 
						|
	GenerateNamedObjects("copy.c" "" "copy" 0 "" "" false "SINGLE")
 | 
						|
	GenerateNamedObjects("dot.c" "" "dot" 0 "" "" false "SINGLE")
 | 
						|
	GenerateNamedObjects("rot.c" "" "rot" 0 "" "" false "SINGLE")
 | 
						|
	GenerateNamedObjects("nrm2.c" "" "nrm2" 0 "" "" false "SINGLE")
 | 
						|
	GenerateNamedObjects("gemv.c" "" "gemv" 0 "" "" false "SINGLE")
 | 
						|
	GenerateNamedObjects("gemm.c" "" "gemm" 0 "" "" false "SINGLE")
 | 
						|
	GenerateNamedObjects("asum.c" "" "asum" 0 "" "" false "SINGLE")
 | 
						|
	GenerateNamedObjects("swap.c" "" "swap" 0 "" "" false "SINGLE")
 | 
						|
	GenerateNamedObjects("axpy.c" "" "axpy" 0 "" "" false "SINGLE")
 | 
						|
	GenerateNamedObjects("imax.c" "USE_ABS" "i*amax" 0 "" "" false "SINGLE")
 | 
						|
endif ()
 | 
						|
if ( BUILD_COMPLEX16 AND NOT  BUILD_DOUBLE)
 | 
						|
	GenerateNamedObjects("scal.c" "" "scal" 0 "" "" false "DOUBLE")
 | 
						|
	GenerateNamedObjects("copy.c" "" "copy" 0 "" "" false "DOUBLE")
 | 
						|
	GenerateNamedObjects("dot.c" "" "dot" 0 "" "" false "DOUBLE")
 | 
						|
	GenerateNamedObjects("rot.c" "" "rot" 0 "" "" false "DOUBLE")
 | 
						|
	GenerateNamedObjects("nrm2.c" "" "nrm2" 0 "" "" false "DOUBLE")
 | 
						|
	GenerateNamedObjects("gemv.c" "" "gemv" 0 "" "" false "DOUBLE")
 | 
						|
	GenerateNamedObjects("gemm.c" "" "gemm" 0 "" "" false "DOUBLE")
 | 
						|
	GenerateNamedObjects("asum.c" "" "asum" 0 "" "" false "DOUBLE")
 | 
						|
	GenerateNamedObjects("swap.c" "" "swap" 0 "" "" false "DOUBLE")
 | 
						|
	GenerateNamedObjects("axpy.c" "" "axpy" 0 "" "" false "DOUBLE")
 | 
						|
	GenerateNamedObjects("imax.c" "USE_ABS" "i*amax" 0 "" "" false "DOUBLE")
 | 
						|
endif ()
 | 
						|
 | 
						|
add_library(interface OBJECT ${OPENBLAS_SRC})
 |