// single invocation for matops and lapack functions
func runCheck(A *matrix.FloatMatrix, LB int) (bool, time.Duration, time.Duration) {
var flags matops.Flags
N := A.Rows()
ipiv := make([]int, N, N)
ipiv0 := make([]int32, N, N)
flags = matops.LOWER
lopt := linalg.OptLower
if testUpper {
flags = matops.UPPER
lopt = linalg.OptUpper
}
W := matrix.FloatZeros(A.Rows(), LB+2)
fnc := func() {
_, ERRmatops = matops.DecomposeBK(A, W, ipiv, flags, LB)
}
if verbose && N < 10 {
fmt.Fprintf(os.Stderr, "A start:\n%v\n", A)
}
A0 := A.Copy()
mperf.FlushCache()
time0 := mperf.Timeit(fnc)
if verbose && N < 10 {
fmt.Fprintf(os.Stderr, "ipiv:%v\n", ipiv)
fmt.Fprintf(os.Stderr, "A end:\n%v\n", A)
}
fn2 := func() {
ERRlapack = lapack.Sytrf(A0, ipiv0, lopt)
}
if verbose && N < 10 {
fmt.Fprintf(os.Stderr, "A0 start:\n%v\n", A0)
}
mperf.FlushCache()
time2 := mperf.Timeit(fn2)
if verbose && N < 10 {
fmt.Fprintf(os.Stderr, "ipiv0:%v\n", ipiv0)
fmt.Fprintf(os.Stderr, "A0 end:\n%v\n", A0)
}
// now A == A0 && ipiv == ipiv0
ok := A.AllClose(A0)
okip := checkIPIV(ipiv, ipiv0)
if !ok || !okip {
// save result to globals
Rlapack = A0
Rmatops = A
IPIVlapack = ipiv0
IPIVmatops = ipiv
}
return ok && okip, time0, time2
}
// single invocation for matops and lapack functions
func runCheck(A *matrix.FloatMatrix, LB int) (bool, time.Duration, time.Duration) {
M := A.Rows()
N := A.Cols()
nN := N
if M < N {
nN = M
}
ipiv := make([]int, nN, nN)
ipiv0 := make([]int32, nN, nN)
fnc := func() {
_, ERRmatops = matops.DecomposeLU(A, ipiv, LB)
}
if verbose && N < 10 {
fmt.Fprintf(os.Stderr, "A start:\n%v\n", A)
}
A0 := A.Copy()
mperf.FlushCache()
time0 := mperf.Timeit(fnc)
if verbose && N < 10 {
fmt.Fprintf(os.Stderr, "A end:\n%v\n", A)
fmt.Fprintf(os.Stderr, "ipiv:%v\n", ipiv)
}
fn2 := func() {
ERRlapack = lapack.Getrf(A0, ipiv0)
}
if verbose && N < 10 {
fmt.Fprintf(os.Stderr, "A0 start:\n%v\n", A0)
}
mperf.FlushCache()
time2 := mperf.Timeit(fn2)
if verbose && N < 10 {
fmt.Fprintf(os.Stderr, "A0 end:\n%v\n", A0)
fmt.Fprintf(os.Stderr, "ipiv0:%v\n", ipiv0)
}
// now A == A0 && ipiv == ipiv0
ok := A.AllClose(A0)
okip := checkIPIV(ipiv, ipiv0)
_ = okip
if !ok || !okip {
// save result to globals
Rlapack = A0
Rmatops = A
IPIVlapack = ipiv0
IPIVmatops = ipiv
}
return ok && okip, time0, time2
}
// single invocation for matops and lapack functions
func runCheck(A *matrix.FloatMatrix, LB int) (bool, time.Duration, time.Duration) {
var W *matrix.FloatMatrix = nil
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)
}
if verbose && N < 10 {
fmt.Fprintf(os.Stderr, "A start:\n%v\n", A)
}
A0 := A.Copy()
tau0 := tau.Copy()
mperf.FlushCache()
time0 := mperf.Timeit(fnc)
if verbose && N < 10 {
fmt.Fprintf(os.Stderr, "A end:\n%v\n", A)
tau.SetSize(1, N, 1)
fmt.Fprintf(os.Stderr, "tau: %v\n", tau)
}
fn2 := func() {
ERRlapack = lapack.Geqrf(A0, tau0)
}
mperf.FlushCache()
time2 := mperf.Timeit(fn2)
if verbose && N < 10 {
fmt.Fprintf(os.Stderr, "A0 end:\n%v\n", A0)
tau0.SetSize(1, N, 1) // row vector
fmt.Fprintf(os.Stderr, "tau0: %v\n", tau0)
}
// now A == A0 && tau == tau0
ok := A.AllClose(A0)
oktau := tau.AllClose(tau0)
if !ok || !oktau {
// save result to globals
Rlapack = A0
Rmatops = A
TAUlapack = tau0
TAUmatops = tau
}
return ok && oktau, time0, time2
}
func runRefTest(A *matrix.FloatMatrix, ntest, LB int) time.Duration {
var mintime time.Duration
lopt := linalg.OptLower
if testUpper {
lopt = linalg.OptUpper
}
fnc := func() {
ERRlapack = lapack.Potrf(A, 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
}
func runTest(A *matrix.FloatMatrix, ntest, LB int) time.Duration {
var flags matops.Flags
var mintime time.Duration
flags = matops.LOWER
if testUpper {
flags = matops.UPPER
}
fnc := func() {
_, ERRmatops = matops.DecomposeCHOL(A, 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
}
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
}
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
}
func runRefTest(A *matrix.FloatMatrix, ntest, LB int) time.Duration {
var mintime time.Duration
N := A.Cols()
tau := matrix.FloatZeros(N, 1)
fnc := func() {
ERRlapack = lapack.Geqrf(A, tau)
}
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
}
}
return mintime
}
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 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
}
// single invocation for matops and lapack functions
func runCheck(A *matrix.FloatMatrix, LB int) (bool, time.Duration, time.Duration) {
var flags matops.Flags
N := A.Rows()
flags = matops.LOWER
lopt := linalg.OptLower
if testUpper {
flags = matops.UPPER
lopt = linalg.OptUpper
}
fnc := func() {
_, ERRmatops = matops.DecomposeCHOL(A, flags, LB)
}
if verbose && N < 10 {
fmt.Fprintf(os.Stderr, "A start:\n%v\n", A)
}
A0 := A.Copy()
mperf.FlushCache()
time0 := mperf.Timeit(fnc)
if verbose && N < 10 {
fmt.Fprintf(os.Stderr, "A end:\n%v\n", A)
}
fn2 := func() {
ERRlapack = lapack.Potrf(A0, lopt)
}
if verbose && N < 10 {
fmt.Fprintf(os.Stderr, "A0 start:\n%v\n", A0)
}
mperf.FlushCache()
time2 := mperf.Timeit(fn2)
if verbose && N < 10 {
fmt.Fprintf(os.Stderr, "A0 end:\n%v\n", A0)
}
// now A == A0 && ipiv == ipiv0
ok := A.AllClose(A0)
if !ok {
// save result to globals
Rlapack = A0
Rmatops = A
}
return ok, time0, time2
}
func main() {
flag.Parse()
runtime.GOMAXPROCS(nWorker)
matops.NumWorkers(nWorker)
rand.Seed(time.Now().UnixNano())
testFunc, ok := tests[testName]
if !ok {
fmt.Printf("Error: test %s does not exists.\nKnown tests:\n", testName)
for tname := range tests {
fmt.Printf("\t%s\n", tname)
}
return
}
var checkFunc mperf.MatrixCheckFunc
if transpose[0] == 'A' {
checkFunc = CheckTransA
} else {
checkFunc = CheckNoTrans
}
if singleTest {
fnc, A, X, Y0 := testFunc(M, N, P)
mperf.FlushCache()
tm := mperf.Timeit(fnc)
if check {
reftime, ok := mperf.CheckWithFunc(A, X, Y0, checkFunc)
if verbose {
fmt.Fprintf(os.Stderr, "%s: %v\n", testName, tm)
fmt.Fprintf(os.Stderr, "Reference: [%v] %v (%.2f) \n",
ok, reftime, tm.Seconds()/reftime.Seconds())
}
}
//sec, _ := mperf.SingleTest(testName, testFunc, M, N, P, check, verbose)
if asGflops {
gflops := 2.0 * float64(int64(M)*int64(N)*int64(P)) / tm.Seconds() * 1e-9
fmt.Printf("%.4f Gflops\n", gflops)
} else if asEps {
eps := float64(int64(M)*int64(N)) / tm.Seconds()
fmt.Printf("%.4f Eps\n", eps)
} else {
fmt.Printf("%vs\n", tm.Seconds())
}
return
}
if len(sizeList) > 0 {
sizes = parseSizeList(sizeList)
}
times := mperf.MultipleSizeTests(testFunc, sizes, testCount, verbose)
if asGflops || asEps {
if verbose {
fmt.Printf("calculating Gflops ...\n")
}
for sz := range times {
n := int64(sz)
if asGflops {
times[sz] = 2.0 * float64(n*n) / times[sz] * 1e-9
} else {
times[sz] = float64(n) / times[sz]
}
}
}
// print out as python dictionary
fmt.Printf("{")
i := 0
for sz := range times {
if i > 0 {
fmt.Printf(", ")
}
fmt.Printf("%d: %v", sz, times[sz])
i++
}
fmt.Printf("}\n")
}