Golang FloatMatrix.Rows方法代码示例

func blockedInverseUpper(A *matrix.FloatMatrix, flags Flags, nb int) (err error) {
	var ATL, ATR, ABL, ABR matrix.FloatMatrix
	var A00, A01, A02, A11, A12, A22 matrix.FloatMatrix

	err = nil
	partition2x2(
		&ATL, &ATR,
		&ABL, &ABR, A, 0, 0, pTOPLEFT)

	for ATL.Rows() < A.Rows() {
		repartition2x2to3x3(&ATL,
			&A00, &A01, &A02,
			nil, &A11, &A12,
			nil, nil, &A22, A, nb, pBOTTOMRIGHT)
		// -------------------------------------------------
		// libflame, variant 1

		// A01 = A00*A01
		MultTrm(&A01, &A00, 1.0, flags)
		// A01 = -A01 / triu(A11)
		SolveTrm(&A01, &A11, -1.0, flags|RIGHT)
		// A11 = inv(A11)
		if flags&UNIT != 0 {
			unblockedInverseUnitUpper(&A11)
		} else {
			unblockedInverseUpper(&A11)
		}
		// -------------------------------------------------
		continue3x3to2x2(
			&ATL, &ATR,
			&ABL, &ABR, &A00, &A11, &A22, A, pBOTTOMRIGHT)
	}
	return
}

// Compute
//      Y = alpha*A*X + beta*Y
//      Y = alpha*A.T*X + beta*Y  ; flags = TRANSA
//
//    A is M*N or N*M generic matrix,
//    X is row or column vector of length N
//    Y is row or column vector of legth M.
//
// MVMult is vector orientation agnostic. It does not matter if Y, X are row or
// column vectors, they are always handled as if they were column vectors.
func MVMult(Y, A, X *matrix.FloatMatrix, alpha, beta float64, flags Flags) error {

	if A.Rows() == 0 || A.Cols() == 0 {
		return nil
	}
	if Y.Rows() != 1 && Y.Cols() != 1 {
		return errors.New("Y not a vector.")
	}
	if X.Rows() != 1 && X.Cols() != 1 {
		return errors.New("X not a vector.")
	}

	Ar := A.FloatArray()
	ldA := A.LeadingIndex()
	Yr := Y.FloatArray()
	incY := 1
	lenY := Y.NumElements()
	if Y.Rows() == 1 {
		// row vector
		incY = Y.LeadingIndex()
	}
	Xr := X.FloatArray()
	incX := 1
	lenX := X.NumElements()
	if X.Rows() == 1 {
		// row vector
		incX = X.LeadingIndex()
	}
	// NOTE: This could diveded to parallel tasks by rows.
	calgo.DMultMV(Yr, Ar, Xr, alpha, beta, calgo.Flags(flags), incY, ldA, incX,
		0, lenX, 0, lenY, vpLen, mB)
	return nil
}

func trmvTest(t *testing.T, A *matrix.FloatMatrix, flags Flags, nb int) bool {
	N := A.Cols()
	//S := 0
	//E := A.Cols()
	X0 := matrix.FloatWithValue(A.Rows(), 1, 2.0)
	X1 := X0.Copy()

	trans := linalg.OptNoTrans
	if flags&TRANS != 0 {
		trans = linalg.OptTrans
	}
	diag := linalg.OptNonUnit
	if flags&UNIT != 0 {
		diag = linalg.OptUnit
	}
	uplo := linalg.OptUpper
	if flags&LOWER != 0 {
		uplo = linalg.OptLower
	}

	blas.TrmvFloat(A, X0, uplo, diag, trans)

	Ar := A.FloatArray()
	Xr := X1.FloatArray()
	if nb == 0 {
		DTrimvUnblkMV(Xr, Ar, flags, 1, A.LeadingIndex(), N)
	}
	result := X0.AllClose(X1)
	t.Logf("   X0 == X1: %v\n", result)
	if !result && A.Rows() < 8 {
		t.Logf("  BLAS TRMV X0:\n%v\n", X0)
		t.Logf("  DTrmv X1:\n%v\n", X1)
	}
	return result
}

// Calculate C = alpha*A*B.T + beta*C, C is M*N, A is M*P and B is N*P
func MMMultTransB(C, A, B *matrix.FloatMatrix, alpha, beta float64) error {
	psize := int64(C.NumElements() * A.Cols())
	Ar := A.FloatArray()
	ldA := A.LeadingIndex()
	Br := B.FloatArray()
	ldB := B.LeadingIndex()
	Cr := C.FloatArray()
	ldC := C.LeadingIndex()
	if nWorker <= 1 || psize <= limitOne {
		calgo.DMult(Cr, Ar, Br, alpha, beta, calgo.TRANSB, ldC, ldA, ldB,
			B.Rows(), 0, C.Cols(), 0, C.Rows(), vpLen, nB, mB)
		return nil
	}

	// here we have more than one worker available
	worker := func(cstart, cend, rstart, rend int, ready chan int) {
		calgo.DMult(Cr, Ar, Br, alpha, beta, calgo.TRANSB, ldC, ldA, ldB, B.Rows(),
			cstart, cend, rstart, rend, vpLen, nB, mB)
		ready <- 1
	}
	colworks, rowworks := divideWork(C.Rows(), C.Cols(), nWorker)
	scheduleWork(colworks, rowworks, C.Cols(), C.Rows(), worker)
	//scheduleWork(colworks, rowworks, worker)
	return nil
}

func runRefTest(A *matrix.FloatMatrix, ntest, LB int) time.Duration {

	var mintime time.Duration

	N := A.Rows()
	ipiv := make([]int32, N, N)
	lopt := linalg.OptLower
	if testUpper {
		lopt = linalg.OptUpper
	}

	fnc := func() {
		ERRlapack = lapack.Sytrf(A, ipiv, lopt)
	}

	A0 := A.Copy()
	for n := 0; n < ntest; n++ {
		if n > 0 {
			// restore original A
			A0.CopyTo(A)
		}
		mperf.FlushCache()
		time0 := mperf.Timeit(fnc)
		if n == 0 || time0 < mintime {
			mintime = time0
		}
	}
	return mintime
}

// Inverse NON-UNIT diagonal tridiagonal matrix
func unblockedInverseUpper(A *matrix.FloatMatrix) (err error) {
	var ATL, ATR, ABL, ABR matrix.FloatMatrix
	var A00, a01, A02, a11, a12t, A22 matrix.FloatMatrix

	err = nil
	partition2x2(
		&ATL, &ATR,
		&ABL, &ABR, A, 0, 0, pTOPLEFT)

	for ATL.Rows() < A.Rows() {
		repartition2x2to3x3(&ATL,
			&A00, &a01, &A02,
			nil, &a11, &a12t,
			nil, nil, &A22, A, 1, pBOTTOMRIGHT)
		// -------------------------------------------------
		aval := a11.Float()

		// a12 = -a12/a11
		InvScale(&a12t, -aval)
		// A02 = A02 + a01*a12
		MVRankUpdate(&A02, &a01, &a12t, 1.0)
		// a01 = a01/a11
		InvScale(&a01, aval)
		// a11 = 1.0/a11
		a11.SetAt(0, 0, 1.0/aval)

		// -------------------------------------------------
		continue3x3to2x2(
			&ATL, &ATR,
			&ABL, &ABR, &A00, &a11, &A22, A, pBOTTOMRIGHT)
	}
	return
}

func runRefTest(A *matrix.FloatMatrix, ntest, LB int) time.Duration {

	var flags matops.Flags
	var mintime time.Duration

	N := A.Rows()
	ipiv := make([]int, N, N)
	flags = matops.LOWER
	if testUpper {
		flags = matops.UPPER
	}

	W := matrix.FloatZeros(A.Rows(), LB+2)
	fnc := func() {
		_, ERRref = matops.DecomposeLDL(A, W, ipiv, flags, 0)
	}

	A0 := A.Copy()
	for n := 0; n < ntest; n++ {
		if n > 0 {
			// restore original A
			A0.CopyTo(A)
		}
		mperf.FlushCache()
		time0 := mperf.Timeit(fnc)
		if n == 0 || time0 < mintime {
			mintime = time0
		}
	}
	return mintime
}

func runTest(A *matrix.FloatMatrix, ntest, LB int) time.Duration {

	var flags matops.Flags
	var mintime time.Duration

	N := A.Rows()
	ipiv := make([]int, N, N)
	flags = matops.LOWER
	if testUpper {
		flags = matops.UPPER
	}

	W := matrix.FloatZeros(A.Rows(), LB+2)
	fnc := func() {
		_, ERRmatops = matops.DecomposeBK(A, W, ipiv, flags, LB)
	}

	A0 := A.Copy()
	for n := 0; n < ntest; n++ {
		if n > 0 {
			// restore original A
			A0.CopyTo(A)
		}
		mperf.FlushCache()
		time0 := mperf.Timeit(fnc)
		if n == 0 || time0 < mintime {
			mintime = time0
		}
		if verbose {
			fmt.Printf("%.4f ms\n", time0.Seconds()*1000.0)
		}
	}
	return mintime
}

// Inverse UNIT diagonal tridiagonal matrix
func unblockedInverseUnitLower(A *matrix.FloatMatrix) (err error) {
	var ATL, ATR, ABL, ABR matrix.FloatMatrix
	var A00, a10t, a11, A20, a21, A22 matrix.FloatMatrix

	err = nil
	partition2x2(
		&ATL, &ATR,
		&ABL, &ABR, A, 0, 0, pTOPLEFT)

	for ATL.Rows() < A.Rows() {
		repartition2x2to3x3(&ATL,
			&A00, nil, nil,
			&a10t, &a11, nil,
			&A20, &a21, &A22, A, 1, pBOTTOMRIGHT)
		// -------------------------------------------------

		// a21 = -a21
		Scale(&a21, -1.0)
		// A20 = A20 + a21*a10.t
		MVRankUpdate(&A20, &a21, &a10t, 1.0)

		// -------------------------------------------------
		continue3x3to2x2(
			&ATL, &ATR,
			&ABL, &ABR, &A00, &a11, &A22, A, pBOTTOMRIGHT)
	}
	return
}

func runTest(A *matrix.FloatMatrix, ntest, LB int) time.Duration {

	var mintime time.Duration

	M := A.Rows()
	N := A.Cols()
	nN := N
	if M < N {
		nN = M
	}
	ipiv := make([]int, nN, nN)

	fnc := func() {
		_, ERRmatops = matops.DecomposeLU(A, ipiv, LB)
	}

	A0 := A.Copy()
	for n := 0; n < ntest; n++ {
		if n > 0 {
			// restore original A
			A0.CopyTo(A)
		}
		mperf.FlushCache()
		time0 := mperf.Timeit(fnc)
		if n == 0 || time0 < mintime {
			mintime = time0
		}
		if verbose {
			fmt.Printf("%.4f ms\n", time0.Seconds()*1000.0)
		}
	}
	return mintime
}

/*
 *  ( A11  a12 )   ( U11 u12 )( D1  0  )( U11.t 0 )
 *  ( a21  a22 )   (  0   1  )(  0  d2 )( u12.t 1 )
 *
 *   a22  =   d2
 *   a01  =   u12*d2                       => u12 = a12/d2
 *   A11  =   u12*d2*u12.t + U11*D1*U11.t  => U11 = A11 - u12*d2*u12.t
 */
func unblkUpperLDLnoPiv(A *matrix.FloatMatrix) (err error) {
	var ATL, ATR, ABL, ABR matrix.FloatMatrix
	var A00, a01, A02, a11, a12, A22 matrix.FloatMatrix

	err = nil
	partition2x2(
		&ATL, &ATR,
		&ABL, &ABR, A, 0, 0, pBOTTOMRIGHT)

	for ATL.Rows() > 0 {
		repartition2x2to3x3(&ATL,
			&A00, &a01, &A02,
			nil, &a11, &a12,
			nil, nil, &A22, A, 1, pTOPLEFT)

		// --------------------------------------------------------

		// A00 = A00 - u01*d11*u01.T = A00 - a01*a01.T/a11; triangular update
		err = MVUpdateTrm(&A00, &a01, &a01, -1.0/a11.Float(), UPPER)

		// u01 = a01/a11
		InvScale(&a01, a11.Float())
		// ---------------------------------------------------------

		continue3x3to2x2(
			&ATL, &ATR,
			&ABL, &ABR, &A00, &a11, &A22, A, pTOPLEFT)
	}
	return
}

func runTest(A *matrix.FloatMatrix, ntest, LB int) time.Duration {
	var W *matrix.FloatMatrix = nil
	var mintime time.Duration

	N := A.Cols()
	tau := matrix.FloatZeros(N, 1)
	if LB > 0 {
		W = matrix.FloatZeros(A.Rows(), LB)
	}
	fnc := func() {
		_, ERRmatops = matops.DecomposeQR(A, tau, W, LB)
	}

	A0 := A.Copy()
	for n := 0; n < ntest; n++ {
		if n > 0 {
			// restore original A
			A0.CopyTo(A)
			tau.Scale(0.0)
		}
		mperf.FlushCache()
		time0 := mperf.Timeit(fnc)
		if n == 0 || time0 < mintime {
			mintime = time0
		}
		if verbose {
			fmt.Printf("%.4f ms\n", time0.Seconds()*1000.0)
		}
	}
	return mintime
}

func runRefTest(A *matrix.FloatMatrix, ntest, LB int) time.Duration {

	var mintime time.Duration

	M := A.Rows()
	N := A.Cols()
	nN := N
	if M < N {
		nN = M
	}
	ipiv := make([]int32, nN, nN)

	fnc := func() {
		ERRlapack = lapack.Getrf(A, ipiv)
	}

	A0 := A.Copy()
	for n := 0; n < ntest; n++ {
		if n > 0 {
			// restore original A
			A0.CopyTo(A)
		}
		mperf.FlushCache()
		time0 := mperf.Timeit(fnc)
		if n == 0 || time0 < mintime {
			mintime = time0
		}
	}
	return mintime
}

// unblocked LU decomposition w/o pivots, FLAME LU nopivots variant 5
func unblockedLUnoPiv(A *matrix.FloatMatrix) (err error) {
	var ATL, ATR, ABL, ABR matrix.FloatMatrix
	var A00, a01, A02, a10, a11, a12, A20, a21, A22 matrix.FloatMatrix

	err = nil
	partition2x2(
		&ATL, &ATR,
		&ABL, &ABR, A, 0, 0, pTOPLEFT)

	for ATL.Rows() < A.Rows() {
		repartition2x2to3x3(&ATL,
			&A00, &a01, &A02,
			&a10, &a11, &a12,
			&A20, &a21, &A22, A, 1, pBOTTOMRIGHT)

		// a21 = a21/a11
		//a21.Scale(1.0/a11.Float())
		InvScale(&a21, a11.Float())
		// A22 = A22 - a21*a12
		err = MVRankUpdate(&A22, &a21, &a12, -1.0)

		continue3x3to2x2(
			&ATL, &ATR,
			&ABL, &ABR, &A00, &a11, &A22, A, pBOTTOMRIGHT)
	}
	return
}

// blocked LU decomposition w/o pivots, FLAME LU nopivots variant 5
func blockedLUnoPiv(A *matrix.FloatMatrix, nb int) (err error) {
	var ATL, ATR, ABL, ABR matrix.FloatMatrix
	var A00, A01, A02, A10, A11, A12, A20, A21, A22 matrix.FloatMatrix

	err = nil
	partition2x2(
		&ATL, &ATR,
		&ABL, &ABR, A, 0, 0, pTOPLEFT)

	for ATL.Rows() < A.Rows() {
		repartition2x2to3x3(&ATL,
			&A00, &A01, &A02,
			&A10, &A11, &A12,
			&A20, &A21, &A22, A, nb, pBOTTOMRIGHT)

		// A00 = LU(A00)
		unblockedLUnoPiv(&A11)
		// A12 = trilu(A00)*A12.-1  (TRSM)
		SolveTrm(&A12, &A11, 1.0, LEFT|LOWER|UNIT)
		// A21 = A21.-1*triu(A00) (TRSM)
		SolveTrm(&A21, &A11, 1.0, RIGHT|UPPER)
		// A22 = A22 - A21*A12
		Mult(&A22, &A21, &A12, -1.0, 1.0, NOTRANS)

		continue3x3to2x2(
			&ATL, &ATR,
			&ABL, &ABR, &A00, &A11, &A22, A, pBOTTOMRIGHT)
	}
	return
}