uo9xpb74
/
mathnet-numerics
forked from tsai/mathnet-numerics
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity
Browse Source

fixed matrix multiple bug when using more than two threads

la-knuth
Marcus Cuda 16 years ago
parent
4306a730d4
commit
435eaff230
4 changed files with 16 additions and 32 deletions
Whitespace
Show all changes Ignore whitespace when comparing lines Ignore changes in amount of whitespace Ignore changes in whitespace at EOL
Split View
Diff Options
Show Stats Download Patch File Download Diff File
  1. 12
      src/Numerics/Algorithms/LinearAlgebra/ManagedLinearAlgebraProvider.Complex.cs
  2. 12
      src/Numerics/Algorithms/LinearAlgebra/ManagedLinearAlgebraProvider.Complex32.cs
  3. 12
      src/Numerics/Algorithms/LinearAlgebra/ManagedLinearAlgebraProvider.Double.cs
  4. 12
      src/Numerics/Algorithms/LinearAlgebra/ManagedLinearAlgebraProvider.Single.cs

12
src/Numerics/Algorithms/LinearAlgebra/ManagedLinearAlgebraProvider.Complex.cs
View File

@ -438,10 +438,6 @@ namespace MathNet.Numerics.Algorithms.LinearAlgebra
ydata = y;
}
// Start the actual matrix multiplication.
// TODO - For small matrices we should get rid of the parallelism because of startup costs.
// Perhaps the following implementations would be a good one
// http://blog.feradz.com/2009/01/cache-efficient-matrix-multiplication/
MatrixMultiplyWithUpdate(Transpose.DontTranspose, Transpose.DontTranspose, Complex.One, xdata, rowsX, columnsX, ydata, rowsY, columnsY, Complex.Zero, result);
}
@ -813,12 +809,12 @@ namespace MathNet.Numerics.Algorithms.LinearAlgebra
CommonParallel.Invoke(
() => CacheObliviousMatrixMultiply(transposeA, transposeB, alpha, matrixA, shiftArow, shiftAcol, matrixB, shiftBrow, shiftBcol, result, shiftCrow, shiftCcol, m2, n2, k2, constM, constN, constK, false),
() => CacheObliviousMatrixMultiply(transposeA, transposeB, alpha, matrixA, shiftArow, shiftAcol, matrixB, shiftBrow, shiftBcol + n2, result, shiftCrow, shiftCcol + n2, m2, n - n2, k2, constM, constN, constK, false),
() => CacheObliviousMatrixMultiply(transposeA, transposeB, alpha, matrixA, shiftArow, shiftAcol + k2, matrixB, shiftBrow + k2, shiftBcol, result, shiftCrow, shiftCcol, m2, n2, k - k2, constM, constN, constK, false),
() => CacheObliviousMatrixMultiply(transposeA, transposeB, alpha, matrixA, shiftArow, shiftAcol + k2, matrixB, shiftBrow + k2, shiftBcol + n2, result, shiftCrow, shiftCcol + n2, m2, n - n2, k - k2, constM, constN, constK, false));
() => CacheObliviousMatrixMultiply(transposeA, transposeB, alpha, matrixA, shiftArow + m2, shiftAcol, matrixB, shiftBrow, shiftBcol, result, shiftCrow + m2, shiftCcol, m - m2, n2, k2, constM, constN, constK, false),
() => CacheObliviousMatrixMultiply(transposeA, transposeB, alpha, matrixA, shiftArow + m2, shiftAcol, matrixB, shiftBrow, shiftBcol + n2, result, shiftCrow + m2, shiftCcol + n2, m - m2, n - n2, k2, constM, constN, constK, false));
CommonParallel.Invoke(
() => CacheObliviousMatrixMultiply(transposeA, transposeB, alpha, matrixA, shiftArow + m2, shiftAcol, matrixB, shiftBrow, shiftBcol, result, shiftCrow + m2, shiftCcol, m - m2, n2, k2, constM, constN, constK, false),
() => CacheObliviousMatrixMultiply(transposeA, transposeB, alpha, matrixA, shiftArow + m2, shiftAcol, matrixB, shiftBrow, shiftBcol + n2, result, shiftCrow + m2, shiftCcol + n2, m - m2, n - n2, k2, constM, constN, constK, false),
() => CacheObliviousMatrixMultiply(transposeA, transposeB, alpha, matrixA, shiftArow, shiftAcol + k2, matrixB, shiftBrow + k2, shiftBcol, result, shiftCrow, shiftCcol, m2, n2, k - k2, constM, constN, constK, false),
() => CacheObliviousMatrixMultiply(transposeA, transposeB, alpha, matrixA, shiftArow, shiftAcol + k2, matrixB, shiftBrow + k2, shiftBcol + n2, result, shiftCrow, shiftCcol + n2, m2, n - n2, k - k2, constM, constN, constK, false),
() => CacheObliviousMatrixMultiply(transposeA, transposeB, alpha, matrixA, shiftArow + m2, shiftAcol + k2, matrixB, shiftBrow + k2, shiftBcol, result, shiftCrow + m2, shiftCcol, m - m2, n2, k - k2, constM, constN, constK, false),
() => CacheObliviousMatrixMultiply(transposeA, transposeB, alpha, matrixA, shiftArow + m2, shiftAcol + k2, matrixB, shiftBrow + k2, shiftBcol + n2, result, shiftCrow + m2, shiftCcol + n2, m - m2, n - n2, k - k2, constM, constN, constK, false));
}

12
src/Numerics/Algorithms/LinearAlgebra/ManagedLinearAlgebraProvider.Complex32.cs
View File

@ -433,10 +433,6 @@ namespace MathNet.Numerics.Algorithms.LinearAlgebra
ydata = y;
}
// Start the actual matrix multiplication.
// TODO - For small matrices we should get rid of the parallelism because of startup costs.
// Perhaps the following implementations would be a good one
// http://blog.feradz.com/2009/01/cache-efficient-matrix-multiplication/
MatrixMultiplyWithUpdate(Transpose.DontTranspose, Transpose.DontTranspose, Complex32.One, xdata, rowsX, columnsX, ydata, rowsY, columnsY, Complex32.Zero, result);
}
@ -808,12 +804,12 @@ namespace MathNet.Numerics.Algorithms.LinearAlgebra
CommonParallel.Invoke(
() => CacheObliviousMatrixMultiply(transposeA, transposeB, alpha, matrixA, shiftArow, shiftAcol, matrixB, shiftBrow, shiftBcol, result, shiftCrow, shiftCcol, m2, n2, k2, constM, constN, constK, false),
() => CacheObliviousMatrixMultiply(transposeA, transposeB, alpha, matrixA, shiftArow, shiftAcol, matrixB, shiftBrow, shiftBcol + n2, result, shiftCrow, shiftCcol + n2, m2, n - n2, k2, constM, constN, constK, false),
() => CacheObliviousMatrixMultiply(transposeA, transposeB, alpha, matrixA, shiftArow, shiftAcol + k2, matrixB, shiftBrow + k2, shiftBcol, result, shiftCrow, shiftCcol, m2, n2, k - k2, constM, constN, constK, false),
() => CacheObliviousMatrixMultiply(transposeA, transposeB, alpha, matrixA, shiftArow, shiftAcol + k2, matrixB, shiftBrow + k2, shiftBcol + n2, result, shiftCrow, shiftCcol + n2, m2, n - n2, k - k2, constM, constN, constK, false));
() => CacheObliviousMatrixMultiply(transposeA, transposeB, alpha, matrixA, shiftArow + m2, shiftAcol, matrixB, shiftBrow, shiftBcol, result, shiftCrow + m2, shiftCcol, m - m2, n2, k2, constM, constN, constK, false),
() => CacheObliviousMatrixMultiply(transposeA, transposeB, alpha, matrixA, shiftArow + m2, shiftAcol, matrixB, shiftBrow, shiftBcol + n2, result, shiftCrow + m2, shiftCcol + n2, m - m2, n - n2, k2, constM, constN, constK, false));
CommonParallel.Invoke(
() => CacheObliviousMatrixMultiply(transposeA, transposeB, alpha, matrixA, shiftArow + m2, shiftAcol, matrixB, shiftBrow, shiftBcol, result, shiftCrow + m2, shiftCcol, m - m2, n2, k2, constM, constN, constK, false),
() => CacheObliviousMatrixMultiply(transposeA, transposeB, alpha, matrixA, shiftArow + m2, shiftAcol, matrixB, shiftBrow, shiftBcol + n2, result, shiftCrow + m2, shiftCcol + n2, m - m2, n - n2, k2, constM, constN, constK, false),
() => CacheObliviousMatrixMultiply(transposeA, transposeB, alpha, matrixA, shiftArow, shiftAcol + k2, matrixB, shiftBrow + k2, shiftBcol, result, shiftCrow, shiftCcol, m2, n2, k - k2, constM, constN, constK, false),
() => CacheObliviousMatrixMultiply(transposeA, transposeB, alpha, matrixA, shiftArow, shiftAcol + k2, matrixB, shiftBrow + k2, shiftBcol + n2, result, shiftCrow, shiftCcol + n2, m2, n - n2, k - k2, constM, constN, constK, false),
() => CacheObliviousMatrixMultiply(transposeA, transposeB, alpha, matrixA, shiftArow + m2, shiftAcol + k2, matrixB, shiftBrow + k2, shiftBcol, result, shiftCrow + m2, shiftCcol, m - m2, n2, k - k2, constM, constN, constK, false),
() => CacheObliviousMatrixMultiply(transposeA, transposeB, alpha, matrixA, shiftArow + m2, shiftAcol + k2, matrixB, shiftBrow + k2, shiftBcol + n2, result, shiftCrow + m2, shiftCcol + n2, m - m2, n - n2, k - k2, constM, constN, constK, false));
}

12
src/Numerics/Algorithms/LinearAlgebra/ManagedLinearAlgebraProvider.Double.cs
View File

@ -432,10 +432,6 @@ namespace MathNet.Numerics.Algorithms.LinearAlgebra
ydata = y;
}
// Start the actual matrix multiplication.
// TODO - For small matrices we should get rid of the parallelism because of startup costs.
// Perhaps the following implementations would be a good one
// http://blog.feradz.com/2009/01/cache-efficient-matrix-multiplication/
MatrixMultiplyWithUpdate(Transpose.DontTranspose, Transpose.DontTranspose, 1.0, xdata, rowsX, columnsX, ydata, rowsY, columnsY, 0.0, result);
}
@ -698,12 +694,12 @@ namespace MathNet.Numerics.Algorithms.LinearAlgebra
CommonParallel.Invoke(
() => CacheObliviousMatrixMultiply(transposeA, transposeB, alpha, matrixA, shiftArow, shiftAcol, matrixB, shiftBrow, shiftBcol, result, shiftCrow, shiftCcol, m2, n2, k2, constM, constN, constK, false),
() => CacheObliviousMatrixMultiply(transposeA, transposeB, alpha, matrixA, shiftArow, shiftAcol, matrixB, shiftBrow, shiftBcol + n2, result, shiftCrow, shiftCcol + n2, m2, n - n2, k2, constM, constN, constK, false),
() => CacheObliviousMatrixMultiply(transposeA, transposeB, alpha, matrixA, shiftArow, shiftAcol + k2, matrixB, shiftBrow + k2, shiftBcol, result, shiftCrow, shiftCcol, m2, n2, k - k2, constM, constN, constK, false),
() => CacheObliviousMatrixMultiply(transposeA, transposeB, alpha, matrixA, shiftArow, shiftAcol + k2, matrixB, shiftBrow + k2, shiftBcol + n2, result, shiftCrow, shiftCcol + n2, m2, n - n2, k - k2, constM, constN, constK, false));
() => CacheObliviousMatrixMultiply(transposeA, transposeB, alpha, matrixA, shiftArow + m2, shiftAcol, matrixB, shiftBrow, shiftBcol, result, shiftCrow + m2, shiftCcol, m - m2, n2, k2, constM, constN, constK, false),
() => CacheObliviousMatrixMultiply(transposeA, transposeB, alpha, matrixA, shiftArow + m2, shiftAcol, matrixB, shiftBrow, shiftBcol + n2, result, shiftCrow + m2, shiftCcol + n2, m - m2, n - n2, k2, constM, constN, constK, false));
CommonParallel.Invoke(
() => CacheObliviousMatrixMultiply(transposeA, transposeB, alpha, matrixA, shiftArow + m2, shiftAcol, matrixB, shiftBrow, shiftBcol, result, shiftCrow + m2, shiftCcol, m - m2, n2, k2, constM, constN, constK, false),
() => CacheObliviousMatrixMultiply(transposeA, transposeB, alpha, matrixA, shiftArow + m2, shiftAcol, matrixB, shiftBrow, shiftBcol + n2, result, shiftCrow + m2, shiftCcol + n2, m - m2, n - n2, k2, constM, constN, constK, false),
() => CacheObliviousMatrixMultiply(transposeA, transposeB, alpha, matrixA, shiftArow, shiftAcol + k2, matrixB, shiftBrow + k2, shiftBcol, result, shiftCrow, shiftCcol, m2, n2, k - k2, constM, constN, constK, false),
() => CacheObliviousMatrixMultiply(transposeA, transposeB, alpha, matrixA, shiftArow, shiftAcol + k2, matrixB, shiftBrow + k2, shiftBcol + n2, result, shiftCrow, shiftCcol + n2, m2, n - n2, k - k2, constM, constN, constK, false),
() => CacheObliviousMatrixMultiply(transposeA, transposeB, alpha, matrixA, shiftArow + m2, shiftAcol + k2, matrixB, shiftBrow + k2, shiftBcol, result, shiftCrow + m2, shiftCcol, m - m2, n2, k - k2, constM, constN, constK, false),
() => CacheObliviousMatrixMultiply(transposeA, transposeB, alpha, matrixA, shiftArow + m2, shiftAcol + k2, matrixB, shiftBrow + k2, shiftBcol + n2, result, shiftCrow + m2, shiftCcol + n2, m - m2, n - n2, k - k2, constM, constN, constK, false));
}

12
src/Numerics/Algorithms/LinearAlgebra/ManagedLinearAlgebraProvider.Single.cs
View File

@ -428,10 +428,6 @@ namespace MathNet.Numerics.Algorithms.LinearAlgebra
ydata = y;
}
// Start the actual matrix multiplication.
// TODO - For small matrices we should get rid of the parallelism because of startup costs.
// Perhaps the following implementations would be a good one
// http://blog.feradz.com/2009/01/cache-efficient-matrix-multiplication/
MatrixMultiplyWithUpdate(Transpose.DontTranspose, Transpose.DontTranspose, 1.0f, xdata, rowsX, columnsX, ydata, rowsY, columnsY, 0.0f, result);
}
@ -694,12 +690,12 @@ namespace MathNet.Numerics.Algorithms.LinearAlgebra
CommonParallel.Invoke(
() => CacheObliviousMatrixMultiply(transposeA, transposeB, alpha, matrixA, shiftArow, shiftAcol, matrixB, shiftBrow, shiftBcol, result, shiftCrow, shiftCcol, m2, n2, k2, constM, constN, constK, false),
() => CacheObliviousMatrixMultiply(transposeA, transposeB, alpha, matrixA, shiftArow, shiftAcol, matrixB, shiftBrow, shiftBcol + n2, result, shiftCrow, shiftCcol + n2, m2, n - n2, k2, constM, constN, constK, false),
() => CacheObliviousMatrixMultiply(transposeA, transposeB, alpha, matrixA, shiftArow, shiftAcol + k2, matrixB, shiftBrow + k2, shiftBcol, result, shiftCrow, shiftCcol, m2, n2, k - k2, constM, constN, constK, false),
() => CacheObliviousMatrixMultiply(transposeA, transposeB, alpha, matrixA, shiftArow, shiftAcol + k2, matrixB, shiftBrow + k2, shiftBcol + n2, result, shiftCrow, shiftCcol + n2, m2, n - n2, k - k2, constM, constN, constK, false));
() => CacheObliviousMatrixMultiply(transposeA, transposeB, alpha, matrixA, shiftArow + m2, shiftAcol, matrixB, shiftBrow, shiftBcol, result, shiftCrow + m2, shiftCcol, m - m2, n2, k2, constM, constN, constK, false),
() => CacheObliviousMatrixMultiply(transposeA, transposeB, alpha, matrixA, shiftArow + m2, shiftAcol, matrixB, shiftBrow, shiftBcol + n2, result, shiftCrow + m2, shiftCcol + n2, m - m2, n - n2, k2, constM, constN, constK, false));
CommonParallel.Invoke(
() => CacheObliviousMatrixMultiply(transposeA, transposeB, alpha, matrixA, shiftArow + m2, shiftAcol, matrixB, shiftBrow, shiftBcol, result, shiftCrow + m2, shiftCcol, m - m2, n2, k2, constM, constN, constK, false),
() => CacheObliviousMatrixMultiply(transposeA, transposeB, alpha, matrixA, shiftArow + m2, shiftAcol, matrixB, shiftBrow, shiftBcol + n2, result, shiftCrow + m2, shiftCcol + n2, m - m2, n - n2, k2, constM, constN, constK, false),
() => CacheObliviousMatrixMultiply(transposeA, transposeB, alpha, matrixA, shiftArow, shiftAcol + k2, matrixB, shiftBrow + k2, shiftBcol, result, shiftCrow, shiftCcol, m2, n2, k - k2, constM, constN, constK, false),
() => CacheObliviousMatrixMultiply(transposeA, transposeB, alpha, matrixA, shiftArow, shiftAcol + k2, matrixB, shiftBrow + k2, shiftBcol + n2, result, shiftCrow, shiftCcol + n2, m2, n - n2, k - k2, constM, constN, constK, false),
() => CacheObliviousMatrixMultiply(transposeA, transposeB, alpha, matrixA, shiftArow + m2, shiftAcol + k2, matrixB, shiftBrow + k2, shiftBcol, result, shiftCrow + m2, shiftCcol, m - m2, n2, k - k2, constM, constN, constK, false),
() => CacheObliviousMatrixMultiply(transposeA, transposeB, alpha, matrixA, shiftArow + m2, shiftAcol + k2, matrixB, shiftBrow + k2, shiftBcol + n2, result, shiftCrow + m2, shiftCcol + n2, m - m2, n - n2, k - k2, constM, constN, constK, false));
}

Write Preview
Loading…
Cancel
Save