85bfd3abb0
cpu: rnn: implement merged iter calculation using matmul
2025-10-17 05:32:45 -07:00
24379b3bf2
cpu: rnn: use merged layer execution for all sizes with matmul
2025-10-17 05:32:45 -07:00
24c39c357b
cpu: rnn: fix int overflows
2025-10-17 05:32:45 -07:00
b115e85606
cpu: rnn: use matmul wrapper for fwd prop
2025-10-17 05:32:45 -07:00
cda1a3ed3e
cpu: rnn: precreate matmul ctx
2025-10-17 05:32:45 -07:00
6397e09a99
cpu: rnn: use matmul for bwd prop
2025-10-17 05:32:45 -07:00
1c100e7eaa
cpu: rnn: fix memory descriptors of matmul arguments
...
Arguments src and weights of Matmul primitive had wrong memory
descriptors.
This was not detected e.g. by tests as these runtime descriptors
were not used (correct memory descriptors were used during Matmul
primitive descriptor creation).
Local variables were renamed because parameters a_, b_, c_ can
correspond to arbitrary matrices, and binding them to RNN's src
and weights was misleading.
2025-10-17 05:32:44 -07:00
9f5a86a440
cpu: rnn: implement lstm projection using matmul
2025-10-17 05:32:44 -07:00
e13d8d10f8
cpu: rnn: implement merged gemm layer step using matmul
...
One layer case is supported.
2025-10-17 05:32:44 -07:00
53be46a45a
cpu: rnn: use merge_gemm_layer with matmul
2025-10-17 05:32:44 -07:00
53bf0a482b
cpu: rnn: use matmul for FWD_D for avx512 and fp32
2025-10-17 05:32:44 -07:00
d2b59266e1
github: workflows: bump KyleMayes/install-llvm-action from 2.0.7 to 2.0.8 ( #4121 )
...
Signed-off-by: dependabot[bot] <support@github.com >
Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>
2025-10-17 11:50:39 +01:00
5365a85d64
github: workflows: bump github/codeql-action from 3.30.6 to 4.30.8
...
Bumps [github/codeql-action](https://github.com/github/codeql-action ) from 3.30.6 to 4.30.8.
- [Release notes](https://github.com/github/codeql-action/releases )
- [Changelog](https://github.com/github/codeql-action/blob/main/CHANGELOG.md )
- [Commits](64d10c1313...f443b600d9support@github.com >
2025-10-16 15:40:30 -07:00
7b351ae955
src: gpu: intel: gemm: add strided batch support to group sums
2025-10-16 14:30:24 -07:00
dc1d8c3d55
doc: update supported attributes section for matmul
2025-10-16 13:38:24 -07:00
52077ed87c
api, doc: fix documentation warn wrt parameter name
2025-10-16 13:38:24 -07:00
c16cf34b28
doc: fix a few misspellings
2025-10-16 13:38:24 -07:00
cff4bb8aaa
doc: add section for quantization-related APIs and refactor
2025-10-16 13:38:24 -07:00
954f7f76f2
xe: jit: fix shr usage
2025-10-16 12:47:19 -07:00
3a756b982b
cpu: aarch64: skip f16 to f32 upcast for clip and clip_v2 eltwise ( #4101 )
2025-10-16 15:06:10 +01:00
e2aa79e849
ci: aarch64: make comparison script friendlier for local testing ( #4157 )
2025-10-16 13:12:46 +01:00
34caf5740b
cpu: aarch64: jit_reorder: fix dispatch issue for direct copy ( #4150 )
2025-10-16 08:57:31 +01:00
bf6617dd1f
tests: benchdnn: utils: improve --buffer-prefix messaging
2025-10-15 18:09:06 -07:00
b5773a0826
xe: sdpa: workaround for xe3 regressions for PTL
2025-10-15 17:55:25 -04:00
34c81a1ca8
xe: gemm: convert HHS kernel to FHS to fix perf regression
2025-10-15 13:53:19 -07:00
7b3f2af13c
xe: jit: gemm: fix stream-k slab count calculation
2025-10-15 10:05:31 -07:00
a4874e48ea
tests: sdpa: fix fallthrough
2025-10-15 09:08:56 -07:00
21adf4d977
tests: sdpa: fix typo
2025-10-15 09:08:56 -07:00
ff30885a40
xe: sdpa: add check for host scalar fetch
2025-10-15 09:08:56 -07:00
f8ef3008b1
xe: rnn: avoid potential null pointer dereference
2025-10-15 09:08:56 -07:00
201e204682
xe: conv: make conversion explicit
2025-10-15 09:08:56 -07:00
3f66dc5049
third_party: ngen: avoid unnecessary copy
2025-10-15 09:08:56 -07:00
2b06b5c1b5
xe: jit: avoid unnecessary copies
2025-10-15 09:08:56 -07:00
4060edd33c
xe: gemm: assert host scalar scale is nonzero
2025-10-15 09:08:56 -07:00
f1302e4804
xe: ir: remove dead code
2025-10-15 09:08:56 -07:00
8df1253973
xe: lrn: remove dead code
2025-10-15 09:08:56 -07:00
27e186ccd9
gtest: check returned ptr on expected md init failure
2025-10-15 10:42:00 +02:00
de84f7f345
benchdnn: properly lock before checking allocations size
2025-10-15 10:42:00 +02:00
34663d344a
benchdnn: fix reference check in init_memory
2025-10-15 10:42:00 +02:00
67cd11b889
cpu: x64: matmul fix wei_k_blk query
2025-10-14 23:10:32 -04:00
5ab84bf3df
xe: ukernel: Add new ukernel entries for MoE support
2025-10-14 20:05:31 -04:00
5a2b8ccedb
scripts: update converter with quantization mode
2025-10-15 02:01:20 +02:00
eed2287b10
doc: document quantization mode and clean up quantization page
2025-10-15 02:01:20 +02:00
5331f9e068
gtest: cover mx input and output scaling in iface tests
2025-10-15 02:01:20 +02:00
76fd406dd4
benchdnn: matmul: add support for mx quantization testing
2025-10-15 02:01:20 +02:00
a3abba35b3
cpu: add mx quantization mode support to matmul primitive
2025-10-15 02:01:20 +02:00
1934858751
common: introduce quantization mode
2025-10-15 02:01:20 +02:00
98ae9938af
cpu: enable e8m0->f32 reorder
2025-10-15 02:01:20 +02:00
8f70dafafd
third_party: ngen: auto-SWSB: fix early A@ dependency clearing bug
2025-10-14 15:38:41 -07:00
d5e144c943
tests: benchdnn: enable mode p on windows
2025-10-14 14:28:23 -07:00
