func MemCpy(dst, src unsafe.Pointer, bytes int64) {
Sync()
timer.Start("memcpy")
cu.MemcpyAsync(cu.DevicePtr(uintptr(dst)), cu.DevicePtr(uintptr(src)), bytes, stream0)
Sync()
timer.Stop("memcpy")
}
// Frees the GPU memory and disables the slice.
func (b *Bytes) Free() {
if b.Ptr != nil {
cu.MemFree(cu.DevicePtr(uintptr(b.Ptr)))
}
b.Ptr = nil
b.Len = 0
}
func MemCpyHtoD(dst, src unsafe.Pointer, bytes int64) {
Sync() // sync previous kernels
timer.Start("memcpyHtoD")
cu.MemcpyHtoD(cu.DevicePtr(uintptr(dst)), src, bytes)
Sync() // sync copy
timer.Stop("memcpyHtoD")
}
func MemCpyDtoH(dst, src unsafe.Pointer, bytes int64) {
Sync() // sync previous kernels
timer.Start("memcpyDtoH")
cu.MemcpyDtoH(dst, cu.DevicePtr(uintptr(src)), bytes)
Sync() // sync copy
timer.Stop("memcpyDtoH")
}
// return a 1-float CUDA reduction buffer from a pool
// initialized to initVal
func reduceBuf(initVal float32) unsafe.Pointer {
if reduceBuffers == nil {
initReduceBuf()
}
buf := <-reduceBuffers
cu.MemsetD32Async(cu.DevicePtr(uintptr(buf)), math.Float32bits(initVal), 1, stream0)
return buf
}
// Execute the FFT plan, asynchronous.
// src and dst are 3D arrays stored 1D arrays.
func (p *fft3DC2RPlan) ExecAsync(src, dst *data.Slice) {
if Synchronous {
Sync()
timer.Start("fft")
}
oksrclen := p.InputLenFloats()
if src.Len() != oksrclen {
panic(fmt.Errorf("fft size mismatch: expecting src len %v, got %v", oksrclen, src.Len()))
}
okdstlen := p.OutputLenFloats()
if dst.Len() != okdstlen {
panic(fmt.Errorf("fft size mismatch: expecting dst len %v, got %v", okdstlen, dst.Len()))
}
p.handle.ExecC2R(cu.DevicePtr(uintptr(src.DevPtr(0))), cu.DevicePtr(uintptr(dst.DevPtr(0))))
if Synchronous {
Sync()
timer.Stop("fft")
}
}
// Execute the FFT plan, asynchronous.
// src and dst are 3D arrays stored 1D arrays.
func (p *fft3DR2CPlan) ExecAsync(src, dst *data.Slice) {
if Synchronous {
Sync()
timer.Start("fft")
}
util.Argument(src.NComp() == 1 && dst.NComp() == 1)
oksrclen := p.InputLen()
if src.Len() != oksrclen {
log.Panicf("fft size mismatch: expecting src len %v, got %v", oksrclen, src.Len())
}
okdstlen := p.OutputLen()
if dst.Len() != okdstlen {
log.Panicf("fft size mismatch: expecting dst len %v, got %v", okdstlen, dst.Len())
}
p.handle.ExecR2C(cu.DevicePtr(uintptr(src.DevPtr(0))), cu.DevicePtr(uintptr(dst.DevPtr(0))))
if Synchronous {
Sync()
timer.Stop("fft")
}
}
// Frees all buffers. Called after mesh resize.
func FreeBuffers() {
Sync()
for _, size := range buf_pool {
for i := range size {
cu.DevicePtr(uintptr(size[i])).Free()
size[i] = nil
}
}
buf_pool = make(map[int][]unsafe.Pointer)
buf_check = make(map[unsafe.Pointer]struct{})
}
func newSlice(nComp int, size [3]int, alloc func(int64) unsafe.Pointer, memType int8) *data.Slice {
data.EnableGPU(memFree, cu.MemFreeHost, MemCpy, MemCpyDtoH, MemCpyHtoD)
length := prod(size)
bytes := int64(length) * cu.SIZEOF_FLOAT32
ptrs := make([]unsafe.Pointer, nComp)
for c := range ptrs {
ptrs[c] = unsafe.Pointer(alloc(bytes))
cu.MemsetD32(cu.DevicePtr(uintptr(ptrs[c])), 0, int64(length))
}
return data.SliceFromPtrs(size, memType, ptrs)
}
// Memset sets the Slice's components to the specified values.
// To be carefully used on unified slice (need sync)
func Memset(s *data.Slice, val ...float32) {
if Synchronous { // debug
Sync()
timer.Start("memset")
}
util.Argument(len(val) == s.NComp())
for c, v := range val {
cu.MemsetD32Async(cu.DevicePtr(uintptr(s.DevPtr(c))), math.Float32bits(v), int64(s.Len()), stream0)
}
if Synchronous { //debug
Sync()
timer.Stop("memset")
}
}
// zero 1-component slice
func zero1_async(dst *data.Slice) {
cu.MemsetD32Async(cu.DevicePtr(uintptr(dst.DevPtr(0))), 0, int64(dst.Len()), stream0)
}
// Construct new byte slice with given length,
// initialised to zeros.
func NewBytes(Len int) *Bytes {
ptr := cu.MemAlloc(int64(Len))
cu.MemsetD8(cu.DevicePtr(ptr), 0, int64(Len))
return &Bytes{unsafe.Pointer(uintptr(ptr)), Len}
}
func memFree(ptr unsafe.Pointer) { cu.MemFree(cu.DevicePtr(uintptr(ptr))) }
