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

BLD: Add swap and refactor a bit

This commit is contained in:
Rohit Goswami 2024-03-17 13:57:21 +00:00 committed by Mateusz Sokół
parent c76e7c6b95
commit 552f81045d
1 changed files with 252 additions and 26 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

278
kernel/meson.build
View File

@ -13,14 +13,11 @@ endif
# NOTE: The def and undefs are from Makefile.L1 # NOTE: The def and undefs are from Makefile.L1
# Construct all PRECsymbKERNEL from src and dir via files(dir + src) # Construct all PRECsymbKERNEL from src and dir via files(dir + src)
# For the precs array, the following mapping is used for c_args: # For the precs array, the following mapping is used for c_args:
# undef --> -Uwhatever
# def --> -Dwhatever
# addl --> passed AS IS # addl --> passed AS IS
base_kops = [ base_kops = [
# Level 1 BLAS # Level 1 BLAS
{ {
'base': 'rot', 'base': 'rot',
# TODO: A lot of these flags are basically repeating..
'precs': { 'precs': {
's' : { 's' : {
'dir': 'arm', 'dir': 'arm',
@ -35,48 +32,253 @@ base_kops = [
'q' : { 'q' : {
'dir': 'arm', 'dir': 'arm',
'kernel': 'zrot.c', 'kernel': 'zrot.c',
'undef': ['COMPLEX'],
'def': ['DXDOUBLE'],
}, },
# TODO: rot is nonstandard in this instance, taking cs zd xq
# The others (scal, swap) take c z x
'cs' : { 'cs' : {
'dir': 'arm', 'dir': 'arm',
'kernel': 'zrot.c', 'kernel': 'zrot.c',
'undef': ['DOUBLE'],
'def': ['COMPLEX'],
}, },
'zd' : { 'zd' : {
'dir': 'arm', 'dir': 'arm',
'kernel': 'zrot.c', 'kernel': 'zrot.c',
'def': ['COMPLEX', 'DOUBLE'],
}, },
'xq' : { 'xq' : {
'dir': 'arm', 'dir': 'arm',
'kernel': 'zrot.c', 'kernel': 'zrot.c',
'def': ['COMPLEX', 'DXDOUBLE'],
}, },
}, },
}, },
# {'symb': 'swap', {'base': 'swap',
# 'src': {'real': 'swap', 'cmplx': 'zswap', 'ext': '.c'}, 'precs': {
# 'dir': 'arm', 'precs': _std_dk }, 's' : {
# {'symb': 'scal', 'dir': 'arm',
# 'src': {'real': 'scal', 'cmplx': 'zscal', 'ext': '.c'}, 'kernel': 'swap.c',
# 'dir': 'arm', 'precs': _std_dk }, },
# {'symb': 'copy', 'd' : {
# 'src': {'real': 'copy', 'cmplx': 'zcopy', 'ext': '.c'}, 'dir': 'arm',
# 'dir': 'arm', 'precs': _std_dk }, 'kernel': 'swap.c',
# {'symb': 'axpy', },
# 'src': {'real': 'axpy', 'cmplx': 'zaxpy', 'ext': '.c'}, 'q' : {
# 'dir': 'arm', 'precs': _std_dk }, 'dir': 'arm',
# {'symb': 'dot', 'kernel': 'zswap.c',
# 'src': {'real': 'dot', 'cmplx': 'zdot', 'ext': '.c'}, },
# 'dir': 'arm', 'precs': _std_dk }, 'c' : {
'dir': 'arm',
'kernel': 'zswap.c',
},
'z' : {
'dir': 'arm',
'kernel': 'zswap.c',
},
'x' : {
'dir': 'arm',
'kernel': 'zswap.c',
},
},
},
{'base': 'scal',
'precs': {
's' : {
'dir': 'arm',
'kernel': 'scal.c',
},
'd' : {
'dir': 'arm',
'kernel': 'scal.c',
},
'q' : {
'dir': 'arm',
'kernel': 'zscal.c',
},
'c' : {
'dir': 'arm',
'kernel': 'zscal.c',
},
'z' : {
'dir': 'arm',
'kernel': 'zscal.c',
},
'x' : {
'dir': 'arm',
'kernel': 'zscal.c',
},
},
},
{'base': 'copy',
'precs': {
's' : {
'dir': 'arm',
'kernel': 'copy.c',
},
'd' : {
'dir': 'arm',
'kernel': 'copy.c',
},
'q' : {
'dir': 'arm',
'kernel': 'zcopy.c',
},
'c' : {
'dir': 'arm',
'kernel': 'zcopy.c',
},
'z' : {
'dir': 'arm',
'kernel': 'zcopy.c',
},
'x' : {
'dir': 'arm',
'kernel': 'zcopy.c',
},
},
},
{'base': 'axpy',
'precs': {
's' : {
'dir': 'arm',
'kernel': 'axpy.c',
},
'd' : {
'dir': 'arm',
'kernel': 'axpy.c',
},
'q' : {
'dir': 'arm',
'kernel': 'zaxpy.c',
},
'c' : {
'dir': 'arm',
'kernel': 'zaxpy.c',
},
'z' : {
'dir': 'arm',
'kernel': 'zaxpy.c',
},
'x' : {
'dir': 'arm',
'kernel': 'zaxpy.c',
},
},
},
{'base': 'dot',
'precs': {
's' : {
'dir': 'arm',
'kernel': 'dot.c',
},
'd' : {
'dir': 'arm',
'kernel': 'dot.c',
},
'q' : {
'dir': 'arm',
'kernel': 'zdot.c',
},
'c' : {
'dir': 'arm',
'kernel': 'zdot.c',
},
'z' : {
'dir': 'arm',
'kernel': 'zdot.c',
},
'x' : {
'dir': 'arm',
'kernel': 'zdot.c',
},
},
},
# xDOTU xDOTC xxDOT aren't present
{'base': 'nrm2',
'precs': {
's' : {
'dir': 'arm',
'kernel': 'nrm2.c',
},
'd' : {
'dir': 'arm',
'kernel': 'nrm2.c',
},
'q' : {
'dir': 'arm',
'kernel': 'znrm2.c',
},
'c' : {
'dir': 'arm',
'kernel': 'znrm2.c',
},
'z' : {
'dir': 'arm',
'kernel': 'znrm2.c',
},
'x' : {
'dir': 'arm',
'kernel': 'znrm2.c',
},
},
},
{'base': 'asum',
'precs': {
's' : {
'dir': 'arm',
'kernel': 'asum.c',
},
'd' : {
'dir': 'arm',
'kernel': 'asum.c',
},
'q' : {
'dir': 'arm',
'kernel': 'zasum.c',
},
'c' : {
'dir': 'arm',
'kernel': 'zasum.c',
},
'z' : {
'dir': 'arm',
'kernel': 'zasum.c',
},
'x' : {
'dir': 'arm',
'kernel': 'zasum.c',
},
},
},
{'base': 'amax',
'precs': {
's' : {
'dir': 'arm',
'kernel': 'amax.c',
},
'd' : {
'dir': 'arm',
'kernel': 'amax.c',
},
'q' : {
'dir': 'arm',
'kernel': 'zamax.c',
},
'c' : {
'dir': 'arm',
'kernel': 'zamax.c',
},
'z' : {
'dir': 'arm',
'kernel': 'zamax.c',
},
'x' : {
'dir': 'arm',
'kernel': 'zamax.c',
},
},
},
] ]
kernel_confs = [] kernel_confs = []
# For the precs array, the following mapping is used for c_args:
# undef --> -Uwhatever
# def --> -Dwhatever
foreach _kop : base_kops foreach _kop : base_kops
foreach pkey, pval : _kop['precs'] foreach pkey, pval : _kop['precs']
kcfg = { kcfg = {
@ -97,6 +299,30 @@ foreach _kop : base_kops
'def': ['DOUBLE'], 'def': ['DOUBLE'],
} }
endif endif
if 'q' == pkey
kcfg += {
'undef': ['COMPLEX'],
'def': ['DXDOUBLE'],
}
endif
# TODO: rot is nonstandard in this instance, taking cs zd xq
# The others (scal, swap) take c z x, so we need to use in
if 'c' in pkey
kcfg += {
'undef': ['DOUBLE'],
'def': ['COMPLEX'],
}
endif
if 'z' in pkey
kcfg += {
'def': ['COMPLEX', 'DOUBLE'],
}
endif
if 'x' in pkey
kcfg += {
'def': ['COMPLEX', 'DXDOUBLE'],
}
endif
message(kcfg) message(kcfg)
endforeach endforeach
endforeach endforeach