Chip Kerchner
|
03f646b7e3
|
New VSX version of BF16 GEMV (Power) - up to 6.7X faster
|
2023-04-21 17:06:59 +00:00 |
|
Charles Schlosser
|
29c8e3c754
|
fix pow for uint32_t, disable pmul<Packet4ul>
|
2023-04-21 05:47:56 +00:00 |
|
Juraj Oršulić
|
7f06bcae2c
|
Geometry/EulerAngles: make sure that returned solution has canonical ranges
|
2023-04-19 19:12:24 +00:00 |
|
Rasmus Munk Larsen
|
a347dbbab2
|
Delete last few occurences of HasHalfPacket.
|
2023-04-19 10:36:59 -07:00 |
|
Charles Schlosser
|
2b954be663
|
fix typo in sse packetmath
|
2023-04-18 18:17:41 +00:00 |
|
Rasmus Munk Larsen
|
25685c90ad
|
Fix incorrect packet type for unsigned int version of pfirst() in MSVC workaround in PacketMath.h.
|
2023-04-18 17:46:23 +00:00 |
|
Chip Kerchner
|
3f3ce214e6
|
New BF16 pcast functions and move type casting to TypeCasting.h
|
2023-04-18 02:38:38 +00:00 |
|
Pedro Gonnet
|
17b5b4de58
|
Add Packet4ui, Packet8ui, and Packet4ul to the SSE/AVX PacketMath.h headers
|
2023-04-17 23:33:59 +00:00 |
|
Chip Kerchner
|
1148f0a9ec
|
Add dynamic dispatch to BF16 GEMM (Power) and new VSX version
|
2023-04-14 22:20:42 +00:00 |
|
Rasmus Munk Larsen
|
554fe02ae3
|
Enable new AVX512 GEMM kernel by default.
|
2023-04-12 13:39:06 -07:00 |
|
b-shi
|
15fbddaf9b
|
ASAN fixes for AVX512 GEMM/TRSM
|
2023-04-04 15:54:24 -07:00 |
|
Rasmus Munk Larsen
|
df1049ddf4
|
Small packet math cleanup.
|
2023-04-04 16:14:32 +00:00 |
|
Antoine Hoarau
|
9b48d10215
|
Guard all malloc, realloc and free() fonctions with check_that_malloc_is_allowed()
|
2023-04-04 04:24:22 +00:00 |
|
Rasmus Munk Larsen
|
c730290fa0
|
Use the correct truncating intrinsic for double->int casting.
|
2023-04-03 13:56:41 -07:00 |
|
Charles Schlosser
|
bfbc66e078
|
refactor indexedviewmethods, enable non-const ref access with symbolic indices
|
2023-03-29 01:35:26 +00:00 |
|
Rasmus Munk Larsen
|
1a5dfd7c0f
|
Fix incorrect casting in AVX512DQ path.
|
2023-03-27 09:28:06 -07:00 |
|
Charles Schlosser
|
a08649994f
|
Optimize generic_rsqrt_newton_step
|
2023-03-24 22:42:57 +00:00 |
|
Rasmus Munk Larsen
|
b8b8a26145
|
Add more missing vectorized casts for int on x86, and remove redundant unit tests
|
2023-03-24 16:02:00 +00:00 |
|
Rasmus Munk Larsen
|
d57a79e512
|
Optimize float->bool cast for AVX2, based on Charles Schlosser's comments.
|
2023-03-21 20:59:25 -07:00 |
|
Rasmus Munk Larsen
|
a5ae832773
|
Fix reversal of arguments to _mm256_set_m128() in pcast<Packet4d, Packet8f>.
|
2023-03-22 03:21:44 +00:00 |
|
Rasmus Munk Larsen
|
09945f2cc1
|
Optimize casting for x86_64.
|
2023-03-21 18:24:16 +00:00 |
|
Colin Broderick
|
8f9b8e3630
|
Replaced all instances of internal::(U)IntPtr with std::(u)intptr_t. Remove ICC workaround.
|
2023-03-21 16:50:23 +00:00 |
|
Antonio Sánchez
|
2c8011c2dd
|
Fix arm builds.
|
2023-03-20 16:59:38 +00:00 |
|
Jonas Schulze
|
81cb6a51d0
|
Fix some typos
|
2023-03-16 23:11:43 +00:00 |
|
Rasmus Munk Larsen
|
0488b708b4
|
Vectorize tensor.isnan() by using typed predicates.
|
2023-03-16 04:04:22 +00:00 |
|
Rasmus Munk Larsen
|
f02856c640
|
Use EIGEN_NOT_A_MACRO macro (oh the irony!) to avoid build issue in TensorFlow.
|
2023-03-15 11:42:57 -07:00 |
|
Rasmus Munk Larsen
|
690ae9502f
|
Use C++11 standard features for detecting presence of Inf and NaN
|
2023-03-15 16:52:44 +00:00 |
|
Chip Kerchner
|
d71ac6a755
|
Fix recent PowerPC warnings and clang warning
|
2023-03-15 16:50:46 +00:00 |
|
Chip Kerchner
|
23e1541863
|
Put deadcode checks back in from previous change.
|
2023-03-14 00:57:16 +00:00 |
|
Chip Kerchner
|
6c58f0fe1f
|
Revert changes that made BF16 GEMM to cause bad register spillage for LLVM (Power)
|
2023-03-13 23:36:06 +00:00 |
|
Rasmus Munk Larsen
|
79de101d23
|
Handle PropagateFast the same way as PropagateNaN in minmax visitor to
|
2023-03-13 20:47:11 +00:00 |
|
Chip Kerchner
|
9d72412385
|
Add MMA to BF16 GEMV - 5.0-6.3X faster (for Power)
|
2023-03-13 19:37:13 +00:00 |
|
Rasmus Munk Larsen
|
2067b54b13
|
Fix bug in minmax_coeff_visitor for matrix of all NaNs.
|
2023-03-13 18:25:22 +00:00 |
|
Rasmus Munk Larsen
|
ee0ff0ab3a
|
Fix typo in MathFunctions.h
|
2023-03-13 15:50:40 +00:00 |
|
Rasmus Munk Larsen
|
21c49e8f8e
|
Delete mystery character from Eigen/src/Core/arch/NEON/MathFunctions.h
|
2023-03-10 23:27:24 +00:00 |
|
Rasmus Munk Larsen
|
6bb9609bcb
|
Make new Select implementation backwards compatible.
|
2023-03-10 23:07:47 +00:00 |
|
Antonio Sánchez
|
394aabb0a3
|
Fix failing MSVC tests due to compiler bugs.
|
2023-03-10 22:36:57 +00:00 |
|
Rasmus Munk Larsen
|
d6235d76db
|
Clean up generic packetmath specializations for various backends with the help of a macro.
|
2023-03-10 22:02:23 +00:00 |
|
Rasmus Munk Larsen
|
2419632cf5
|
Revert change to allFinite(), since the new version does not work for complex numbers.
|
2023-03-09 21:50:43 +00:00 |
|
Charles Schlosser
|
7bf2968fed
|
Specify Permutation Index for PartialPivLU and FullPivLU
|
2023-03-07 20:28:05 +00:00 |
|
Charles Schlosser
|
1ce8b25825
|
Vectorize any() / all()
|
2023-03-06 23:54:02 +00:00 |
|
Chip Kerchner
|
2b513ca2a0
|
Added partial linear access for LHS & Output - 30% faster for bfloat16 GEMM MMA (Power)
|
2023-03-02 19:22:43 +00:00 |
|
Charles Schlosser
|
0b396c3167
|
Scalarize comps
|
2023-03-02 17:06:23 +00:00 |
|
Antonio Sánchez
|
62d5cfe835
|
Fix ODR issues with Intel's AVX512 TRSM kernels.
|
2023-02-27 07:54:52 +00:00 |
|
Charles Schlosser
|
826627f653
|
vectorize comparisons and select by enabling typed comparisons
|
2023-02-25 20:52:11 +00:00 |
|
Rasmus Munk Larsen
|
2e9b945baf
|
Fix bug that disabled vectorization for coeffMin/coeffMax.
|
2023-02-25 20:03:54 +00:00 |
|
Antonio Sánchez
|
bc5cdc7a67
|
Guard use of long double on GPU device.
|
2023-02-24 21:49:59 +00:00 |
|
Chip Kerchner
|
e4598fedbe
|
Fix compiler versions for certain instructions on Power.
|
2023-02-23 23:24:41 +00:00 |
|
Rasmus Munk Larsen
|
1c0a6cf228
|
Get rid of EIGEN_HAS_AVX512_MATH workaround.
|
2023-02-23 23:16:41 +00:00 |
|
Rasmus Munk Larsen
|
6bcd941ee3
|
Use pmsub in twoprod. This speeds up pow() on Skylake by ~1%.
|
2023-02-21 20:09:29 +00:00 |
|