func TestArrayCopy(test *testing.T) {
runtime.LockOSThread()
// fail test on panic, do not crash
//defer func() {
// if err := recover(); err != nil {
// test.Error(err)
// }
//}()
size := []int{4, 8, 16}
host1, host2 := host.NewArray(3, size), host.NewArray(3, size)
dev1, dev2 := NewArray(3, size), NewArray(3, size)
defer dev1.Free()
defer dev2.Free()
l1 := host1.List
for i := range l1 {
l1[i] = float64(i)
}
dev1.CopyFromHost(host1)
dev2.CopyFromDevice(dev1)
dev2.CopyToHost(host2)
l2 := host2.List
for i := range l1 {
if l2[i] != float64(i) {
if !test.Failed() {
test.Error("expected", i, "got:", l2[i])
}
}
}
}
// Set the multiplier of a MASK or the value of a VALUE
func (q *Quant) SetValue(val []float64) {
//Debug("SetValue", q.name, val)
//~ checkKinds(q, MASK, VALUE)
checkComp(q, len(val))
if q.kind == MASK || q.kind == VALUE {
for i, v := range val {
q.multiplier[i] = v
}
} else if q.kind == FIELD {
q.multiplier = ones(q.nComp)
// use q.Buffer instead?
tempField := host.NewArray(q.nComp, q.array.Size3D())
for c := 0; c < q.nComp; c++ {
for i := 0; i < q.array.Size3D()[X]; i++ {
for j := 0; j < q.array.Size3D()[Y]; j++ {
for k := 0; k < q.array.Size3D()[Z]; k++ {
tempField.Array[c][i][j][k] = float64(val[c])
}
}
}
}
q.SetField(tempField)
// not sure whenever tempBuffer will be destroyed by the GC?
} else {
panic(InputErr(q.name + " is not " + MASK.String() + " or " + FIELD.String() + " or " + VALUE.String() + " but " + q.kind.String()))
}
q.Verify()
q.Invalidate() //!
}
func genWindow(size []int) *host.Array {
window := host.NewArray(1, size)
for i := 0; i < size[0]; i++ {
val0 := 1.0
if size[0] > 16 {
arg0 := float64(i) / float64(size[0]-1)
val0 = gauss(arg0, 0.4)
}
for j := 0; j < size[1]; j++ {
val1 := 1.0
if size[1] > 16 {
arg1 := float64(j) / float64(size[1]-1)
val1 = gauss(arg1, 0.4)
}
for k := 0; k < size[2]; k++ {
val2 := 1.0
if size[2] > 16 {
arg2 := float64(k) / float64(size[2]-1)
val2 = gauss(arg2, 0.4)
}
window.Array[0][i][j][k] = float64(val0 * val1 * val2)
}
}
}
return window
}
// Converts a json vector array to a host.Array.
// Also swaps XYZ - ZYX convention
// TODO: works only for 4D vector arrays
func jsonToHostArray(v interface{}) *host.Array {
defer func() {
err := recover()
if err != nil {
panic(IOErr(fmt.Sprint("Error parsing json array: ", ShortPrint(v), "\ncause: ", err)))
}
}()
err := false
// determine array size as {len(v), len(v[0]), len(v[0][0]), ...}
var size [4]int
v2 := v
for i := range size {
if arr, ok := v2.([]interface{}); ok {
size[i] = len(arr)
if size[i] == 0 {
err = true
break
}
v2 = arr[0]
} else {
err = true
break
}
}
if err {
panic(IOErr(fmt.Sprint("Array with invalid size:", ShortPrint(v))))
}
size3D := size[1:]
arr := host.NewArray(size[0], []int{size3D[X], size3D[Y], size3D[Z]})
a := arr.Array
va := v.([]interface{})
for c := range a {
va_c := va[c].([]interface{})
for i := range a[c] {
va_ci := va_c[i].([]interface{})
for j := range a[c][i] {
va_cij := va_ci[j].([]interface{})
for k := range a[c][i][j] {
a[c][i][j][k] = float64(va_cij[k].(float64)) // convert XYZ-ZYX, works only for 3D
}
}
}
}
runtime.GC() // a LOT of garbage has been made
return convertXYZ(arr)
}
func BenchmarkReduceSumCPU(b *testing.B) {
b.StopTimer()
size := bigsize()
a := host.NewArray(3, size)
b.SetBytes(int64(a.Len()) * SIZEOF_FLOAT)
b.StartTimer()
for i := 0; i < b.N; i++ {
var cpusum float64
for _, num := range a.List {
cpusum += float64(num)
}
}
}
// Should init to zeros
func TestArrayInit(test *testing.T) {
runtime.LockOSThread()
// fail test on panic, do not crash
//defer func() {
// if err := recover(); err != nil {
// test.Error(err)
// }
//}()
size := []int{4, 8, 16}
host1 := host.NewArray(3, size)
dev1 := NewArray(3, size)
defer dev1.Free()
fmt.Printf("%+v\n", dev1)
if dev1.Len() != 3*Prod(size) {
if !test.Failed() {
test.Error("Len(): ", dev1.Len(), "expected: ", 3*Prod(size))
}
}
if dev1.PartLen4D() != 3*Prod(size)/NDevice() {
test.Fail()
}
if dev1.PartLen3D() != Prod(size)/NDevice() {
test.Fail()
}
l1 := host1.List
for i := range l1 {
l1[i] = float64(i)
}
dev1.CopyToHost(host1)
//host1.CopyFromDevice(dev1)
for i := range l1 {
if l1[i] != 0 {
if !test.Failed() {
test.Error(l1[i], "!=0")
}
}
}
}
// Convert mumax's internal ZYX convention to userspace XYZ.
func convertXYZ(arr *host.Array) *host.Array {
s := arr.Size3D
n := arr.NComp()
a := arr.Array
transp := host.NewArray(n, []int{s[Z], s[Y], s[X]})
t := transp.Array
for c := 0; c < n; c++ {
for i := 0; i < s[X]; i++ {
for j := 0; j < s[Y]; j++ {
for k := 0; k < s[Z]; k++ {
t[(n-1)-c][k][j][i] = a[c][i][j][k]
}
}
}
}
runtime.GC() // a LOT of garbage has been made
return transp
}
// Returns a new host array of size size2, re-sized from the input array by nearest-neighbor interpolation.
func Resample(in *host.Array, size2 []int) *host.Array {
Assert(len(size2) == 3)
out := host.NewArray(in.NComp(), size2)
out_a := out.Array
in_a := in.Array
size1 := in.Size3D
for c := range out_a {
for i := range out_a[c] {
i1 := (i * size1[X]) / size2[X]
for j := range out_a[0][i] {
j1 := (j * size1[Y]) / size2[Y]
for k := range out_a[0][i][j] {
k1 := (k * size1[Z]) / size2[Z]
out_a[c][i][j][k] = in_a[c][i1][j1][k1]
}
}
}
}
return out
}
func ReadPNG(fname string) *host.Array {
in, err := os.Open(fname)
CheckIO(err)
img, err2 := png.Decode(in)
CheckIO(err2)
width := img.Bounds().Max.X
height := img.Bounds().Max.Y
inside := host.NewArray(1, []int{1, height, width})
for i := 0; i < height; i++ {
for j := 0; j < width; j++ {
r, g, b, _ := img.At(j, height-1-i).RGBA()
if r+g+b < (0xFFFF*3)/2 {
inside.Array[0][0][i][j] = 1
}
}
}
return inside
}
// Extract real or imaginary parts, copy them from src to dst.
// In the meanwhile, check if the other parts are nearly zero
// and scale the kernel to compensate for unnormalized FFTs.
// real_imag = 0: real parts
// real_imag = 1: imag parts
func extract(src *host.Array) *host.Array {
sx := src.Size3D[X]/2 + 1 // antisymmetric
sy := src.Size3D[Y]/2 + 1 // antisymmetric
sz := src.Size3D[Z] / 2 // only real parts should be stored, the value of the imaginary part should stay below the zero threshould
dst := host.NewArray(src.NComp(), []int{sx, sy, sz})
dstArray := dst.Array
srcArray := src.Array
// Normally, the FFT'ed kernel is purely real because of symmetry,
// so we only store the real parts...
maxImg := float64(0.)
maxReal := float64(0.)
for c := range dstArray {
for k := range dstArray[c] {
for j := range dstArray[c][k] {
for i := range dstArray[c][k][j] {
dstArray[c][k][j][i] = srcArray[c][k][j][2*i]
if Abs32(srcArray[c][k][j][2*i+1]) > maxImg {
maxImg = Abs32(srcArray[c][k][j][2*i+1])
}
if Abs32(srcArray[c][k][j][2*i+0]) > maxReal {
maxReal = Abs32(srcArray[c][k][j][2*i+0])
}
}
}
}
}
// ...however, we check that the imaginary parts are nearly zero,
// just to be sure we did not make a mistake during kernel creation.
Debug("FFT Kernel max real part", 0, ":", maxReal)
Debug("FFT Kernel max imag part", 1, ":", maxImg)
Debug("FFT Kernel max imag/real part=", maxImg/maxReal)
if maxImg/maxReal > 1e-12 { // TODO: is this reasonable?
panic(BugF("FFT Kernel max bad/good part=", maxImg/maxReal))
}
return dst
}
// DEBUG: Make a freshly allocated copy on the host.
func (src *Array) LocalCopy() *host.Array {
dst := host.NewArray(src.NComp(), src.Size3D())
src.CopyToHost(dst)
return dst
}