func (p *exchParam) upload() {
// alloc if needed
if p.gpu == nil {
p.gpu = cuda.SymmLUT(cuda.MemAlloc(int64(len(p.lut)) * cu.SIZEOF_FLOAT32))
}
cuda.MemCpyHtoD(unsafe.Pointer(p.gpu), unsafe.Pointer(&p.lut[0]), cu.SIZEOF_FLOAT32*int64(len(p.lut)))
p.gpu_ok = true
}
func (p *exchParam) upload() {
// alloc if needed
if p.gpu == nil {
p.gpu = cuda.SymmLUT(cuda.MemAlloc(int64(len(p.lut)) * cu.SIZEOF_FLOAT32))
}
lut := p.lut // Copy, to work around Go 1.6 cgo pointer limitations.
cuda.MemCpyHtoD(unsafe.Pointer(p.gpu), unsafe.Pointer(&lut[0]), cu.SIZEOF_FLOAT32*int64(len(p.lut)))
p.gpu_ok = true
}
// get an up-to-date version of the lookup-table on GPU
func (p *lut) gpuLUT() cuda.LUTPtrs {
p.source.update()
if !p.gpu_ok {
// upload to GPU
p.assureAlloc()
cuda.Sync() // sync previous kernels, may still be using gpu lut
for c := range p.gpu_buf {
cuda.MemCpyHtoD(p.gpu_buf[c], unsafe.Pointer(&p.cpu_buf[c][0]), cu.SIZEOF_FLOAT32*NREGION)
}
p.gpu_ok = true
cuda.Sync() //sync upload
}
return p.gpu_buf
}