Details:
- Allow building BLIS with certain framework files (each with the '_amd'
suffix) that have been customized by AMD for Zen-based hardware. These
customized files were derived from portable versions of the same files
(i.e., those without the '_amd' suffix). Whether the portable or AMD-
specific files are compiled is now controlled by a new configure
option, --[en|dis]able-amd-frame-tweaks. This option is disabled by
default in vanilla BLIS, though AMD may choose to enable it by default
in their fork. For now, the added AMD-specific files are:
- bli_gemv_unf_var2_amd.c
- bla_copy_amd.c
- bla_gemv_amd.c
These files reside in 'amd' subdirectories found within the directory
housing their generic counterparts.
- Register optimized real-domain copyv, setv, and swapv kernels in
bli_cntx_init_zen.c.
- Various minor updates to level-1v kernels in 'zen' kernel set.
- Added caxpyf kernel as well as saxpyf and multiple daxpyf kernels to
the 'zen' kernel set
- If the problem passed to ?gemm_() in bla_gemm.c has a unit m or n dim,
call gemv instead and return early.
- Combined variable declarations with their initialization in various
level-2 and level-3 BLAS compatibility files, and also inserted
'const' qualifer in those same declaration statements.
- Moved frame/compat/bla_gemmt.c and .h to frame/compat/extra/ .
- Added copyv and swapv test drivers to 'test' directory.
- Whitespace, comment changes.
Details:
- Fixed a performance regression affecting nearly all level-3 operations
that use the 'haswell' sgemm and dgemm microkernels. This regression
was introduced in 54fa28b, caused by an ill-formed conditional
expression in the assembly code that controls whether cache lines of C
should be prefetched as rows or as columns. Essentially, the two
branches were reversed, causing incomplete prefetching to occur for
both row- and column-stored instances of matrix C. Thanks to Devin
Matthews for his help finding and fixing this bug.
Fixes#613. There are several macros/environment variables which need to be tuned to get good cache block sizes. It would be nice to have a way of getting values automatically.
Details:
- Moved edge-case handling into the gemmtrsm microkernel. This required
changing the microkernel API to take m and n dimension parameters as
well as updating all existing gemmtrsm microkernel function pointer
types, function signatures, and related definitions to take m and n
dimensions. Also updated all existing gemmtrsm kernels in the
'kernels' directory (which for now is limited to haswell and penryn
kernel sets, plus native and 1m-based reference kernels in
'ref_kernels') to take m and n dimensions, and implemented edge-case
handling within those microkernels via a collection of new C
preprocessor macros defined within bli_edge_case_macro_defs.h. Note
that the edge-case handling for gemm-like operations had already
been relocated into the gemm microkernel in 54fa28b.
- Added desriptive comments to GEMM_UKR_SETUP_CT() and related macros in
bli_edge_case_macro_defs.h to allow for easier reading.
- Updated docs/KernelsHowTo.md to reflect above changes. Also cleaned up
the bullet under "Implementation Notes for gemm" that covers alignment
issues. (Thanks to Ivan Korostelev for pointing out the confusing and
outdated language in issue #591.)
- Other minor tweaks to KernelsHowTo.md.
@egaudry and I both saw this issue on Linux with Clang 10.
```
Compiling obj/thunderx2/kernels/armv8a/3/sup/bli_gemmsup_rv_armv8a_asm_d4x8m.o ('thunderx2' CFLAGS for kernels)
kernels/armv8a/3/bli_gemm_armv8a_asm_d6x8.c:171:49: fatal error: invalid symbol redefinition
" \n\t"
^
<inline asm>:90:5: note: instantiated into assembly here
.SLOOPKITER:
^
1 error generated.
```
Signed-off-by: Jeff Hammond <jehammond@nvidia.com>
Add `%=` tag to branch labels, which expands to a unique identifier for each inline assembly block. This prevents duplicate symbol errors on Apple Silicon (#594). Fixes#594. [ci skip] since we can't test Apple Silicon anyways...
Details:
- Moved edge-case handling into the gemm microkernel. This required
changing the microkernel API to take m and n dimension parameters.
This required updating all existing gemm microkernel function pointer
types, function signatures, and related definitions to take m and n
dimensions. We also updated all existing kernels in the 'kernels'
directory to take m and n dimensions, and implemented edge-case
handling within those microkernels via a collection of new C
preprocessor macros defined within bli_edge_case_macro_defs.h. Also
removed the assembly code that formerly would handle general stride
IO on the microtile, since this can now be handled by the same code
that does edge cases.
- Pass the obj_t.ker_fn (of matrix C) into bli_gemm_cntl_create() and
bli_trsm_cntl_create(), where this function pointer is used in lieu of
the default macrokernel when it is non-NULL, and ignored when it is
NULL.
- Re-implemented macrokernel in bli_gemm_ker_var2.c to be a single
function using byte pointers rather that one function for each
floating-point datatype. Also, obtain the microkernel function pointer
from the .ukr field of the params struct embedded within the obj_t
for matrix C (assuming params is non-NULL and contains a non-NULL
value in the .ukr field). Communicate both the gemm microkernel
pointer to use as well as the params struct to the microkernel via
the auxinfo_t struct.
- Defined gemm_ker_params_t type (for the aforementioned obj_t.params
struct) in bli_gemm_var.h.
- Retired the separate _md macrokernel for mixed datatype computation.
We now use the reimplemented bli_gemm_ker_var2() instead.
- Updated gemmt macrokernels to pass m and n dimensions into microkernel
calls.
- Removed edge-case handling from trmm and trsm macrokernels.
- Moved most of bli_packm_alloc() code into a new helper function,
bli_packm_alloc_ex().
- Fixed a typo bug in bli_gemmtrsm_u_template_noopt_mxn.c.
- Added test/syrk_diagonal and test/tensor_contraction directories with
associated code to test those operations.
Details:
- Added a new 'zen3' subconfiguration targeting support for the AMD Zen3
microarchitecture (#561). Thanks to AMD for this contribution.
- Restructured clang and AOCC support for zen, zen2, and zen3
make_defs.mk files. The clang and AOCC version detection now happens
in configure, not in the subconfigurations' makefile fragments. That
is, we've added logic to configure that detects the version of
clang/AOCC, outputs an appropriate variable to config.mk
(ie: CLANG_OT_*, AOCC_OT_*), and then checks for it within the
makefile fragment (as is currently done for the GCC_OT_* variables).
- Added configure support for a GCC_OT_10_1_0 variable (and associated
substitution anchor) to communicate whether the gcc version is older
than 10.1.0, and use this variable to check for recent enough versions
of gcc to use -march=znver3 in the zen3 subconfig.
- Inlined the contents of config/zen/amd_config.mk into the zen and zen2
make_defs.mk so that the files are self-contained, harmonizing the
format of all three Zen-based subconfigurations' make_defs.mk files.
- Added indenting (with spaces) of GNU make conditionals for easier
reading in zen, zen2, and zen3 make_defs.mk files.
- Adjusted the range of models checked by bli_cpuid_is_zen() (which was
previously 0x00 ~ 0xff and is now 0x00 ~ 0x2f) so that it is
completely disjoint from the models checked by bli_cpuid_is_zen2()
(0x30 ~ 0xff). This is normally necessary because Zen and Zen2
microarchitectures share the same family (23, or 0x17), and so the
model code is the only way to differentiate the two. But in our case,
fixing the model range for zen *wasn't* actually necessary since we
checked for zen2 first, and therefore the wide zen range acted like
the 'else' of an 'if-else' statement. That said, the change helps
improve clarity for the reader by encoding useful knowledge, which
was obtained from https://en.wikichip.org/wiki/amd/cpuid .
- Added zen2.def and zen3.def files to the collection in travis/cpuid.
Note that support for zen, zen2, and zen3 is now present, and while
all the three microarchitectures have identical instruction sets from
the perspective of BLIS microkernels, they each correspond to
different subconfigurations and therefore merit separate testing.
Thanks to Devin Matthews for his guidance in hacking these files as
slight modifications of zen.def.
- Enabled testing of zen2 and zen3 via the SDE in travis/do_sde.sh.
Now, zen, zen2, and zen3 are tested through the SDE via Travis CI
builds.
- Updated travis/do_sde.sh to grab the SDE tarball from a new ci-utils
repository on GitHub rather than on Intel's website. This change was
made in an attempt to circumvent recent troubles with Travis CI not
being able to download the SDE directly from Intel's website via curl.
Thanks to Devin Matthews for suggesting the idea.
- Updated travis/do_sde.sh to grab the latest version (8.69.1) of the
Intel SDE from the flame/ci-utils repository.
- Updated .travis.yml to use gcc 9. The file was previously using gcc 8,
which did not support -march=znver2.
- Created amd64_legacy umbrella family in config_registry for targeting
older (bulldozer, piledriver, steamroller, and excavator)
microarchitectures and moved those same subconfigs out of the amd64
umbrella family. However, x86_64 retains amd64_legacy as a constituent
member.
- Fixed a bug in configure related to the building of the so-called
config list. When processing the contents of config_registry,
configure creates a series of structures and lists that allow for
various mappings related to configuration families, subconfigs, and
kernel sets. Two of those lists are built via substitution of
umbrella families with their subconfig members, and one of those
lists was improperly performing the substitution in a way that would
erroneously match on partial umbrella family names. That code was
changed to match the code that was already doing the substitution
properly, via substitute_words(). Also added comments noting the
importance of using substitute_words() in both instances.
- Comment updates.
Details:
- Renamed herk macrokernels and supporting files and functions to gemmt,
which is possible since at the macrokernel level they are identical.
Then recast herk/her2k/syrk/syr2k in terms of gemmt within the expert
level-3 oapi (bli_l3_oapi_ex.c) while also redefining them as literal
functions rather than cpp macros that instantiate multiple functions.
Thanks to Devin Matthews for his efforts on this issue (#531).
- Check that the maximum stack buffer size is sufficiently large
relative to the register blocksizes for each datatype, and do so when
the context is initialized rather than when an operation is called.
Note that with this change, users who pass in their own contexts into
the expert interfaces currently will *not* have any checks performed.
Thanks to Devin Matthews for suggesting this change.
Details:
- Fixed a bug that broke the use of 1m for dcomplex when the single-
precision real and double-precision real ukernels had opposing I/O
preferences (row-preferential sgemm ukernel + column-preferential
dgemm ukernel, or vice versa). The fix involved adjusting the API
to bli_cntx_set_ind_blkszs() so that the induced method context init
function (e.g., bli_cntx_init_<subconfig>_ind()) could call that
function for only one datatype at a time. This allowed the blocksize
scaling (which varies depending on whether we're doing 1m_r or 1m_c)
to happen on a per-datatype basis. This fixes issue #557. Thanks to
Devin Matthews and RuQing Xu for helping discover and report this bug.
- The aforementioned 1m fix required moving the 1m_r/1m_c logic from
bli_cntx_ref.c into a new function, bli_l3_set_schemas(), which is
called from each level-3 _front() function. The pack_t schemas in the
cntx_t were also removed entirely, along with the associated accessor
functions. This in turn required updating the trsm1m-related virtual
ukernels to read the pack schema for B from the auxinfo_t struct
rather than the context. This also required slight tweaks to
bli_gemm_md.c.
- Repositioned the logic for transposing the operation to accommodate
the microkernel IO preference. This mostly only affects gemm. Thanks
to Devin Matthews for his help with this.
- Updated dpackm pack ukernels in the 'armsve' kernel set to avoid
querying pack_t schemas from the context.
- Removed the num_t dt argument from the ind_cntx_init_ft type defined
in bli_gks.c. The context initialization functions for induced methods
were previously passed a dt argument, but I can no longer figure out
*why* they were passed this value. To reduce confusion, I've removed
the dt argument (including also from the function defintion +
prototype).
- Commented out setting of cntx_t schemas in bli_cntx_ind_stage.c. This
breaks high-leve implementations of 3m and 4m, but this is okay since
those implementations will be removed very soon.
- Removed some older blocks of preprocessor-disabled code.
- Comment update to test_libblis.c.
- `ref2` call in `bli_gemmsup_rv_armv8a_asm_d6x8m.c` is commented out.
- `bli_gemmsup_rv_armv8a_asm_d4x8m.c` contains a tail `ref2` call but
it's not called by any upper routine.
Ref cannot handle panel strides (packed cases) thus cannot be called
from the beginning of `gemmsup` (i.e. cannot be dispatch target of
gemmsup to other sizes.)
