Chip Kerchner
|
36bd3eeddf
|
Vectorize BF16 GEMV (VSX & MMA). Use GEMM_GEMV_FORWARD_BF16 (for Power).
|
1 year ago |
Chip Kerchner
|
89702e1f4a
|
Fix zero element GEMV test.
|
1 year ago |
Chip Kerchner
|
77f85c7c00
|
GEMV tests don't like zero elements.
|
1 year ago |
Chip Kerchner
|
868aa857bc
|
Change malloc zero to return one byte and update the SBGEMM test to again use sizes of zero.
|
1 year ago |
Chip Kerchner
|
b1802f4dc8
|
Fix unit test to start at 1 instead of 0 - since malloc zero bytes fails on some systems.
|
1 year ago |
Chip Kerchner
|
c23897f585
|
Add GEMV testing to SBGEMx vs SGEMx testing.
|
1 year ago |
Chip Kerchner
|
31226740d6
|
Cleanup of SBGEMM unit test.
|
1 year ago |
Chip Kerchner
|
b1737698db
|
Fix DEFAULTS in SBGEMM for POWER10. Also comparisons for SBGEMM unit test can be exactly due to epilison differences.
|
1 year ago |
Chip Kerchner
|
f708944fea
|
Add all 4 variations of the SBGEMM to compare_sgemm_sbgemm
|
1 year ago |
Martin Kroeker
|
e9f480111e
|
fix sbgemm bfloat16 conversion errors introduced in PR 4488
|
2 years ago |
Martin Kroeker
|
fb99fc2e6e
|
fix type conversion warnings
|
2 years ago |
Chip-Kerchner
|
d46eba06a7
|
Pack structure only on AIX.
|
2 years ago |
Chip-Kerchner
|
e98e3c4783
|
Fix float32_bits union so that it always the sizeof float.
|
2 years ago |
Chip-Kerchner
|
97a61d0577
|
Fix bfloat16_bits union so that it always the sizeof unsigned short.
|
2 years ago |
Martin Kroeker
|
3d338b57de
|
remove spurious loops
|
3 years ago |
Martin Kroeker
|
d9dc015cfc
|
Use blasint for INTERFACE64 compatibility
|
3 years ago |
Rajalakshmi Srinivasaraghavan
|
1d97405c02
|
POWER: Enable bfloat16 kernels by default
This patch enables bfloat16 kernels by default for POWER processors.
Tested on Linux POWER8, POWER9, POWER10 and AIX POWER10 systems.
|
3 years ago |
Martin Kroeker
|
924fd806d0
|
Rename "HALF" and "sh" to "BFLOAT16" and "sb"
|
5 years ago |
Martin Kroeker
|
4db09c6cec
|
Rename compare_sgemm_shgemm.c to compare_sgemm_sbgemm.c
|
5 years ago |