From 435eaff2303714bc1dce36fb28391679d27a656b Mon Sep 17 00:00:00 2001 From: Marcus Cuda Date: Fri, 1 Apr 2011 18:28:53 +0800 Subject: [PATCH] fixed matrix multiple bug when using more than two threads --- .../ManagedLinearAlgebraProvider.Complex.cs | 12 ++++-------- .../ManagedLinearAlgebraProvider.Complex32.cs | 12 ++++-------- .../ManagedLinearAlgebraProvider.Double.cs | 12 ++++-------- .../ManagedLinearAlgebraProvider.Single.cs | 12 ++++-------- 4 files changed, 16 insertions(+), 32 deletions(-) diff --git a/src/Numerics/Algorithms/LinearAlgebra/ManagedLinearAlgebraProvider.Complex.cs b/src/Numerics/Algorithms/LinearAlgebra/ManagedLinearAlgebraProvider.Complex.cs index 870e0b64..14f65dab 100644 --- a/src/Numerics/Algorithms/LinearAlgebra/ManagedLinearAlgebraProvider.Complex.cs +++ b/src/Numerics/Algorithms/LinearAlgebra/ManagedLinearAlgebraProvider.Complex.cs @@ -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)); } diff --git a/src/Numerics/Algorithms/LinearAlgebra/ManagedLinearAlgebraProvider.Complex32.cs b/src/Numerics/Algorithms/LinearAlgebra/ManagedLinearAlgebraProvider.Complex32.cs index cdca7299..005f9c82 100644 --- a/src/Numerics/Algorithms/LinearAlgebra/ManagedLinearAlgebraProvider.Complex32.cs +++ b/src/Numerics/Algorithms/LinearAlgebra/ManagedLinearAlgebraProvider.Complex32.cs @@ -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)); } diff --git a/src/Numerics/Algorithms/LinearAlgebra/ManagedLinearAlgebraProvider.Double.cs b/src/Numerics/Algorithms/LinearAlgebra/ManagedLinearAlgebraProvider.Double.cs index 0d459542..f2f3b822 100644 --- a/src/Numerics/Algorithms/LinearAlgebra/ManagedLinearAlgebraProvider.Double.cs +++ b/src/Numerics/Algorithms/LinearAlgebra/ManagedLinearAlgebraProvider.Double.cs @@ -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)); } diff --git a/src/Numerics/Algorithms/LinearAlgebra/ManagedLinearAlgebraProvider.Single.cs b/src/Numerics/Algorithms/LinearAlgebra/ManagedLinearAlgebraProvider.Single.cs index 108a2bb4..820c6de2 100644 --- a/src/Numerics/Algorithms/LinearAlgebra/ManagedLinearAlgebraProvider.Single.cs +++ b/src/Numerics/Algorithms/LinearAlgebra/ManagedLinearAlgebraProvider.Single.cs @@ -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)); }