// Copy from device array to device array.
func (dst *Array) CopyFromDevice(src *Array) {
CheckSize(dst.size4D, src.size4D)
d := dst.pointer
s := src.pointer
// copies run concurrently on the individual devices
length := src.partLen4D
cu.MemcpyDtoDAsync(cu.DevicePtr(d), cu.DevicePtr(s), SIZEOF_FLOAT*int64(length), cu.Stream(dst.Stream))
// Synchronize with all copies
dst.Stream.Sync()
}
// Set a single value
func (b *Array) Set(comp, x, y, z int, value float64) {
b.checkBounds(comp, x, y, z)
acomp := b.Comp[comp]
index := acomp.indexOf(x, y, z)
cu.MemcpyHtoD(cu.DevicePtr(offset(uintptr(acomp.pointer), SIZEOF_FLOAT*index)),
cu.HostPtr(unsafe.Pointer(&value)),
1*SIZEOF_FLOAT)
}
// Get a single value
func (b *Array) Get(comp, x, y, z int) float64 {
b.checkBounds(comp, x, y, z)
var value float64
acomp := b.Comp[comp]
index := acomp.indexOf(x, y, z)
cu.MemcpyDtoH(cu.HostPtr(unsafe.Pointer(&value)),
cu.DevicePtr(offset(uintptr(acomp.pointer), SIZEOF_FLOAT*index)),
1*SIZEOF_FLOAT)
return value
}
// Copy from host array to device array.
func (dst *Array) CopyFromHost(src *host.Array) {
CheckSize(dst.size4D, src.Size4D)
partPlaneN := dst.partSize[1] * dst.partSize[2] // floats per YZ plane per GPU
planeN := dst.size3D[1] * dst.size3D[2] // total floats per YZ plane
NPlane := dst.size4D[0] * dst.size3D[0] // total YZ planes (NComp * X size)
partPlaneBytes := SIZEOF_FLOAT * int64(partPlaneN) // bytes per YZ plane per GPU
for i := 0; i < NPlane; i++ {
dstOffset := i * partPlaneN
dstPtr := ArrayOffset(uintptr(dst.pointer), dstOffset)
srcOffset := i * planeN
cu.MemcpyHtoD(cu.DevicePtr(dstPtr), cu.HostPtr(&src.List[srcOffset]), partPlaneBytes)
}
}
// Copy from device array to host array.
func (src *Array) CopyToHost(dst *host.Array) {
CheckSize(dst.Size4D, src.size4D)
partPlaneN := src.partSize[1] * src.partSize[2] // floats per YZ plane per GPU
planeN := src.size3D[1] * src.size3D[2] // total floats per YZ plane
NPlane := src.size4D[0] * src.size3D[0] // total YZ planes (NComp * X size)
partPlaneBytes := SIZEOF_FLOAT * int64(partPlaneN) // bytes per YZ plane per GPU
for i := 0; i < NPlane; i++ {
srcOffset := i * partPlaneN
srcPtr := ArrayOffset(uintptr(src.pointer), srcOffset)
dstOffset := i * planeN
cu.MemcpyDtoH(cu.HostPtr(&dst.List[dstOffset]), cu.DevicePtr(srcPtr), partPlaneBytes)
}
}
// Lets the pointers of an already initialized, but not allocated array (shared)
// point to an allocated array (original) possibly with an offset.
func (shared *Array) PointTo(original *Array, offset int) {
Assert(shared.Len()+offset <= original.Len())
shared.pointer = cu.DevicePtr(ArrayOffset(uintptr(original.pointer), offset))
}
// INTERNAL
// initialize pointers to the component arrays.
// called after the GPU storage has been changed.
func (a *Array) initCompPtrs() {
for c := range a.Comp {
start := c * a.partLen3D
a.Comp[c].pointer = cu.DevicePtr(offset(uintptr(a.pointer), start*SIZEOF_FLOAT))
}
}