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

MAINT: Add the _beta variant of gemm

This commit is contained in:
Rohit Goswami 2024-05-27 01:22:57 +00:00 committed by Mateusz Sokół
parent 517eca96c6
commit 891052a3ef
1 changed files with 36 additions and 11 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

47
kernel/meson.build
View File

@ -356,17 +356,6 @@ base_kops = [
}, },
}, },
# Level 3 symbols # Level 3 symbols
# TODO(rg): Handle the if defines set in arch
{ 'base': '?gemm_beta',
'modes': {
's': {'exts': {'': {'dir': 'x86_64', 'kernel': 'gemm_beta.S'}}},
'd': {'exts': {'': {'dir': 'x86_64', 'kernel': 'gemm_beta.S'}}},
'c': {'exts': {'': {'dir': 'x86_64', 'kernel': 'zgemm_beta.S'}}},
'z': {'exts': {'': {'dir': 'x86_64', 'kernel': 'zgemm_beta.S'}}},
# 'q': {'exts': {'': {'dir': 'generic', 'kernel': 'gemm_beta.c'}}},
# 'x': {'exts': {'': {'dir': 'generic', 'kernel': 'zgemm_beta.c'}}},
},
},
{ 'base': '?gemm_kernel', { 'base': '?gemm_kernel',
'modes': { 'modes': {
's': {'exts': {'': {'dir': 'generic', 'kernel': 'gemmkernel_2x2.c'}}}, 's': {'exts': {'': {'dir': 'generic', 'kernel': 'gemmkernel_2x2.c'}}},
@ -508,6 +497,42 @@ base_kops = [
'addl': ['-DNN']}}}, 'addl': ['-DNN']}}},
}, },
}, },
{ 'base': '?gemm',
'modes': {
's': {'exts': {
'_beta': {'dir': 'generic', 'kernel': 'gemm_beta.c'},
# TODO(rg): the _NUM prefixes are arch dependent
'_incopy': {'dir': 'generic', 'kernel': 'gemm_ncopy_8.c'},
'_itcopy': {'dir': 'generic', 'kernel': 'gemm_tcopy_8.c'},
'_oncopy': {'dir': 'generic', 'kernel': 'gemm_ncopy_2.c'},
'_otcopy': {'dir': 'generic', 'kernel': 'gemm_tcopy_4.c'},
# TODO(rg): direct and direct_performant are built only conditionally
'_direct': {'dir': 'x86_64', 'kernel': 'sgemm_direct_skylakex.c'},
'_direct_performant': {'dir': 'x86_64', 'kernel': 'sgemm_direct_performant.c'},
}},
'd': {'exts': {
'_beta': {'dir': 'generic', 'kernel': 'gemm_beta.c'},
'_incopy': {'dir': 'generic', 'kernel': 'gemm_ncopy_8.c'},
'_itcopy': {'dir': 'generic', 'kernel': 'gemm_tcopy_8.c'},
'_oncopy': {'dir': 'generic', 'kernel': 'gemm_ncopy_2.c'},
'_otcopy': {'dir': 'generic', 'kernel': 'gemm_tcopy_4.c'},
}},
'c': {'exts': {
'_beta': {'dir': 'generic', 'kernel': 'zgemm_beta.c'},
'_incopy': {'dir': 'generic', 'kernel': 'zgemm_ncopy_8.c'},
'_itcopy': {'dir': 'generic', 'kernel': 'zgemm_tcopy_8.c'},
'_oncopy': {'dir': 'generic', 'kernel': 'zgemm_ncopy_2.c'},
'_otcopy': {'dir': 'generic', 'kernel': 'zgemm_tcopy_2.c'},
}},
'z': {'exts': {
'_beta': {'dir': 'generic', 'kernel': 'zgemm_beta.c'},
'_incopy': {'dir': 'generic', 'kernel': 'zgemm_ncopy_8.c'},
'_itcopy': {'dir': 'generic', 'kernel': 'zgemm_tcopy_8.c'},
'_oncopy': {'dir': 'generic', 'kernel': 'zgemm_ncopy_2.c'},
'_otcopy': {'dir': 'generic', 'kernel': 'zgemm_tcopy_2.c'},
}},
},
},
{ 'base': '?trmm', { 'base': '?trmm',
'modes': { 'modes': {
's': {'exts': { 's': {'exts': {