Assignment of ZeroMatrix to DynamicMatrix is extremely slow

[Mikhail Katliar]

Assigment of a ZeroMatrix to DynamicMatrix is >400 times slower than the uniform scalar assignment. A benchmark:

#include <blaze/Math.h>
#include <benchmark/benchmark.h>

template <typename Real, size_t M, size_t N>
static void BM_DynamicMatrixZeroMatrixAssign(::benchmark::State& state)
{
    blaze::DynamicMatrix<Real> A(M, N);

    for (auto _ : state)
        ::benchmark::DoNotOptimize(A = blaze::ZeroMatrix<Real>(M, N));
}

template <typename Real, size_t M, size_t N>
static void BM_DynamicMatrixZeroAssign(::benchmark::State& state)
{
    blaze::DynamicMatrix<Real> A(M, N);

    for (auto _ : state)
        ::benchmark::DoNotOptimize(A = Real {0});
}

BENCHMARK_TEMPLATE(BM_DynamicMatrixZeroMatrixAssign, double, 4, 1);
BENCHMARK_TEMPLATE(BM_DynamicMatrixZeroAssign, double, 4, 1);

Output:

2019-02-19 16:24:13
Running build/bin/tmpc_bench
Run on (12 X 4100 MHz CPU s)
CPU Caches:
  L1 Data 32K (x6)
  L1 Instruction 32K (x6)
  L2 Unified 256K (x6)
  L3 Unified 9216K (x1)
Load Average: 0.27, 0.62, 1.11
-----------------------------------------------------------------------------------------
Benchmark                                               Time             CPU   Iterations
-----------------------------------------------------------------------------------------
BM_DynamicMatrixZeroMatrixAssign<double, 4, 1>       2372 ns         2372 ns       317329
BM_DynamicMatrixZeroAssign<double, 4, 1>             4.99 ns         4.99 ns    118211356

Compiler: gcc-8.2.0, compiler flags: -O2 -g -DNDEBUG

[Mikhail Katliar]

Profiling shows that most of the time is wasted in omp functions:

[Klaus Iglberger]

Hi Mikhail!

Thanks a lot for pointing to this defect: You are correct: Combining DynamicMatrix, ZeroMatrix and OpenMP will result in a significantly more expensive assignment for tiny matrices. We apologize for the inconvenience and will fix the problem as quickly as possible.

Best regards,

Klaus!

[Klaus Iglberger]

Commit b742cdd resolves the significant performance penalty when assigning a ZeroMatrix to a DynamicMatrix while using OpenMP. The fix is immediately available via cloning the Blaze repository and will be officially released in Blaze 3.5.