// Returns a GPU slice for temporary use. To be returned to the pool with Recycle
func Buffer(nComp int, size [3]int) *data.Slice {
if Synchronous {
Sync()
}
ptrs := make([]unsafe.Pointer, nComp)
// re-use as many buffers as possible form our stack
N := prod(size)
pool := buf_pool[N]
nFromPool := iMin(nComp, len(pool))
for i := 0; i < nFromPool; i++ {
ptrs[i] = pool[len(pool)-i-1]
}
buf_pool[N] = pool[:len(pool)-nFromPool]
// allocate as much new memory as needed
for i := nFromPool; i < nComp; i++ {
if len(buf_check) >= buf_max {
log.Panic("too many buffers in use, possible memory leak")
}
ptrs[i] = MemAlloc(int64(cu.SIZEOF_FLOAT32 * N))
buf_check[ptrs[i]] = struct{}{} // mark this pointer as mine
}
return data.SliceFromPtrs(size, data.GPUMemory, ptrs)
}
func sliceFromList(arr [][]float32, size [3]int) *data.Slice {
ptrs := make([]unsafe.Pointer, len(arr))
for i := range ptrs {
util.Argument(len(arr[i]) == prod(size))
ptrs[i] = unsafe.Pointer(&arr[i][0])
}
return data.SliceFromPtrs(size, data.CPUMemory, ptrs)
}
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)
}
func (c *MFMConvolution) init() {
// init FFT plans
padded := c.kernSize
c.fwPlan = newFFT3DR2C(padded[X], padded[Y], padded[Z])
c.bwPlan = newFFT3DC2R(padded[X], padded[Y], padded[Z])
// init device buffers
nc := fftR2COutputSizeFloats(c.kernSize)
c.fftCBuf = NewSlice(1, nc)
c.fftRBuf = data.SliceFromPtrs(c.kernSize, data.GPUMemory, []unsafe.Pointer{c.fftCBuf.DevPtr(0)})
c.gpuFFTKern[X] = NewSlice(1, nc)
c.gpuFFTKern[Y] = NewSlice(1, nc)
c.gpuFFTKern[Z] = NewSlice(1, nc)
c.initFFTKern3D()
}
func (c *DemagConvolution) init(realKern [3][3]*data.Slice) {
// init device buffers
// 2D re-uses fftBuf[X] as fftBuf[Z], 3D needs all 3 fftBufs.
nc := fftR2COutputSizeFloats(c.realKernSize)
c.fftCBuf[X] = NewSlice(1, nc)
c.fftCBuf[Y] = NewSlice(1, nc)
if c.is2D() {
c.fftCBuf[Z] = c.fftCBuf[X]
} else {
c.fftCBuf[Z] = NewSlice(1, nc)
}
// Real buffer shares storage with Complex buffer
for i := 0; i < 3; i++ {
c.fftRBuf[i] = data.SliceFromPtrs(c.realKernSize, data.GPUMemory, []unsafe.Pointer{c.fftCBuf[i].DevPtr(0)})
}
// init FFT plans
c.fwPlan = newFFT3DR2C(c.realKernSize[X], c.realKernSize[Y], c.realKernSize[Z])
c.bwPlan = newFFT3DC2R(c.realKernSize[X], c.realKernSize[Y], c.realKernSize[Z])
// init FFT kernel
// logic size of FFT(kernel): store real parts only
c.fftKernLogicSize = fftR2COutputSizeFloats(c.realKernSize)
util.Assert(c.fftKernLogicSize[X]%2 == 0)
c.fftKernLogicSize[X] /= 2
// physical size of FFT(kernel): store only non-redundant part exploiting Y, Z mirror symmetry
// X mirror symmetry already exploited: FFT(kernel) is purely real.
physKSize := [3]int{c.fftKernLogicSize[X], c.fftKernLogicSize[Y]/2 + 1, c.fftKernLogicSize[Z]/2 + 1}
output := c.fftCBuf[0]
input := c.fftRBuf[0]
fftKern := data.NewSlice(1, physKSize)
kfull := data.NewSlice(1, output.Size()) // not yet exploiting symmetry
kfulls := kfull.Scalars()
kCSize := physKSize
kCSize[X] *= 2 // size of kernel after removing Y,Z redundant parts, but still complex
kCmplx := data.NewSlice(1, kCSize) // not yet exploiting X symmetry
kc := kCmplx.Scalars()
for i := 0; i < 3; i++ {
for j := i; j < 3; j++ { // upper triangular part
if realKern[i][j] != nil { // ignore 0's
// FW FFT
data.Copy(input, realKern[i][j])
c.fwPlan.ExecAsync(input, output)
data.Copy(kfull, output)
// extract non-redundant part (Y,Z symmetry)
for iz := 0; iz < kCSize[Z]; iz++ {
for iy := 0; iy < kCSize[Y]; iy++ {
for ix := 0; ix < kCSize[X]; ix++ {
kc[iz][iy][ix] = kfulls[iz][iy][ix]
}
}
}
// extract real parts (X symmetry)
scaleRealParts(fftKern, kCmplx, 1/float32(c.fwPlan.InputLen()))
c.kern[i][j] = GPUCopy(fftKern)
}
}
}
}
