| Commit message (Collapse) | Author | Age |
| |
|
|
|
|
| |
here. Could also have done a vld1_lane_f32 but doing so here, without the overhead of initializing the unused lane, would have triggered used-of-uninitialized-value errors in tools such as ASan. Note that this code is sub-optimal before or after this change: we should be reading either 2 or 4 float32 values per load-instruction (2 for ARM in-order cores with an affinity for 8-byte loads; 4 for ARM out-of-order cores able to dual-issue 16-byte load instructions with arithmetic instructions). Before or after this patch, we are only loading 4 bytes of useful data here (even if before this patch, we were technically loading 8, only to use only the 4 first).
|
| |
|
| |
|
|
|
|
|
| |
Signed-off-by: Gustavo Lima Chaves <gustavo.lima.chaves@intel.com>
Signed-off-by: Mark D. Ryan <mark.d.ryan@intel.com>
|
|
|
|
| |
It brokes the complex-complex case on SSE.
|
|
|
|
|
|
| |
AVX512).
This commit also removes "half-packet" from data-mappers: it was not used and conceptually broken anyways.
|
| |
|
|
|
|
| |
disable multi-threaded GEMM on non-x86 without c++11.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The major changes are
1. Moving CUDA/PacketMath.h to GPU/PacketMath.h
2. Moving CUDA/MathFunctions.h to GPU/MathFunction.h
3. Moving CUDA/CudaSpecialFunctions.h to GPU/GpuSpecialFunctions.h
The above three changes effectively enable the Eigen "Packet" layer for the HIP platform
4. Merging the "hip_basic" and "cuda_basic" unit tests into one ("gpu_basic")
5. Updating the "EIGEN_DEVICE_FUNC" marking in some places
The change has been tested on the HIP and CUDA platforms.
|
| |
|
| |
|
| |
|
|
|
|
| |
rewriting them as: s*(A.lazyProduct(B)) to save a costly temporary. Measured speedup from 2x to 5x...
|
| |
|
| |
|
| |
|
|
|
|
| |
Found using `codespell` and `grep` from downstream FreeCAD
|
|
|
|
| |
factor: (s*A).triangularView<UpperUnit>()*B
|
|
|
|
| |
some explanations.
|
|
|
|
| |
Work around a compilation error seen with nvcc V8.0.61
|
| |
|
|
|
|
| |
destination of matrix*matrix products.
|
|
|
|
|
| |
* they're used consistently between the declaration and the definition of a function
* we avoid calling host only methods from host device methods.
|
| |
|
|
|
|
| |
columns.
|
| |
|
|
|
|
|
|
|
| |
the common:
A.triangularView() = B*A.sefladjointView()*B.adjoint()
case that used to work in 3.2.
|
|
|
|
| |
(sync is set from and compared to an Index)
|
|
|
|
|
| |
Returned values of omp thread id and numbers are int,
o let's use int instead of Index here.
|
|
|
|
|
|
| |
- div is extremely costly
- this is consistent with the column-major case
- this is consistent with all other BLAS implementations
|
| |
|
| |
|
|
|
|
|
| |
This revised version does not bother about aligned loads/stores,
and rather processes 8 rows at ones for better instruction pipelining.
|
|
|
|
|
|
|
|
|
|
| |
performance of modern CPU.
The previous code has been optimized for Intel core2 for which unaligned loads/stores were prohibitively expensive.
This new version exhibits much higher instruction independence (better pipelining) and explicitly leverage FMA.
According to my benchmark, on Haswell this new kernel is always faster than the previous one, and sometimes even twice as fast.
Even higher performance could be achieved with a better blocking size heuristic and, perhaps, with explicit prefetching.
We should also check triangular product/solve to optimally exploit this new kernel (working on vertical panel of 4 columns is probably not optimal anymore).
|
| |
|
|\ |
|
| | |
|
| | |
|
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| | |
threads when the inner dimension is small.
Timing for square matrices is unchanged, but both CPU and Wall time are significantly improved for skinny matrices. The benchmarks below are for multiplying NxK * KxN matrices with test names of the form BM_OuterishProd/N/K.
Improvements in Wall time:
Run on [redacted] (12 X 3501 MHz CPUs); 2016-10-05T17:40:02.462497196-07:00
CPU: Intel Haswell with HyperThreading (6 cores) dL1:32KB dL2:256KB dL3:15MB
Benchmark Base (ns) New (ns) Improvement
------------------------------------------------------------------
BM_OuterishProd/64/1 3088 1610 +47.9%
BM_OuterishProd/64/4 3562 2414 +32.2%
BM_OuterishProd/64/32 8861 7815 +11.8%
BM_OuterishProd/128/1 11363 6504 +42.8%
BM_OuterishProd/128/4 11128 9794 +12.0%
BM_OuterishProd/128/64 27691 27396 +1.1%
BM_OuterishProd/256/1 33214 28123 +15.3%
BM_OuterishProd/256/4 34312 36818 -7.3%
BM_OuterishProd/256/128 174866 176398 -0.9%
BM_OuterishProd/512/1 7963684 104224 +98.7%
BM_OuterishProd/512/4 7987913 112867 +98.6%
BM_OuterishProd/512/256 8198378 1306500 +84.1%
BM_OuterishProd/1k/1 7356256 324432 +95.6%
BM_OuterishProd/1k/4 8129616 331621 +95.9%
BM_OuterishProd/1k/512 27265418 7517538 +72.4%
Improvements in CPU time:
Run on [redacted] (12 X 3501 MHz CPUs); 2016-10-05T17:40:02.462497196-07:00
CPU: Intel Haswell with HyperThreading (6 cores) dL1:32KB dL2:256KB dL3:15MB
Benchmark Base (ns) New (ns) Improvement
------------------------------------------------------------------
BM_OuterishProd/64/1 6169 1608 +73.9%
BM_OuterishProd/64/4 7117 2412 +66.1%
BM_OuterishProd/64/32 17702 15616 +11.8%
BM_OuterishProd/128/1 45415 6498 +85.7%
BM_OuterishProd/128/4 44459 9786 +78.0%
BM_OuterishProd/128/64 110657 109489 +1.1%
BM_OuterishProd/256/1 265158 28101 +89.4%
BM_OuterishProd/256/4 274234 183885 +32.9%
BM_OuterishProd/256/128 1397160 1408776 -0.8%
BM_OuterishProd/512/1 78947048 520703 +99.3%
BM_OuterishProd/512/4 86955578 1349742 +98.4%
BM_OuterishProd/512/256 74701613 15584661 +79.1%
BM_OuterishProd/1k/1 78352601 3877911 +95.1%
BM_OuterishProd/1k/4 78521643 3966221 +94.9%
BM_OuterishProd/1k/512 258104736 89480530 +65.3%
|
|\| |
|
| | |
|
| | |
|
| |
| |
| |
| | |
install(DIRECTORY ...) command.
|
| | |
|
| |
| |
| |
| | |
5d51a7f12c69138ed2a43df240bdf27a5313f7ce
|
| |
| |
| |
| |
| |
| |
| |
| |
| | |
e56aabf205a1e8f581dd8a46d7d46ce79c45e158
.
Register blocking sizes are better handled by the cache size heuristics.
The current code introduced very small blocks, for instance for 9x9 matrix,
thus killing performance.
|
| |
| |
| |
| |
| |
| |
| |
| |
| |
| | |
- Replace internal::scalar_product_traits<A,B> by Eigen::ScalarBinaryOpTraits<A,B,OP>
- Remove the "functor_is_product_like" helper (was pretty ugly)
- Currently, OP is not used, but it is available to the user for fine grained tuning
- Currently, only the following operators have been generalized: *,/,+,-,=,*=,/=,+=,-=
- TODO: generalize all other binray operators (comparisons,pow,etc.)
- TODO: handle "scalar op array" operators (currently only * is handled)
- TODO: move the handling of the "void" scalar type to ScalarBinaryOpTraits
|
| | |
|
| |
| |
| |
| |
| |
| |
| |
| | |
conversions.
This fixes "conversion from pointer to same-sized integral type" warnings by ICC.
Ideally, we would use the std::[u]intptr_t types all the time, but since they are C99/C++11 only,
let's be safe.
|