func (c *DemagConvolution) exec3D(outp, inp, vol *data.Slice, Bsat float64) {
padded := c.kernSize
// FW FFT
for i := 0; i < 3; i++ {
zero1(c.fftRBuf[i], c.stream)
in := inp.Comp(i)
copyPadMul(c.fftRBuf[i], in, padded, c.size, vol, Bsat, c.stream)
c.fwPlan.ExecAsync(c.fftRBuf[i], c.fftCBuf[i])
}
// kern mul
N0, N1, N2 := c.fftKernSize[0], c.fftKernSize[1], c.fftKernSize[2] // TODO: rm these
kernMulRSymm3D(c.fftCBuf,
c.gpuFFTKern[0][0], c.gpuFFTKern[1][1], c.gpuFFTKern[2][2],
c.gpuFFTKern[1][2], c.gpuFFTKern[0][2], c.gpuFFTKern[0][1],
N0, N1, N2, c.stream)
// BW FFT
for i := 0; i < 3; i++ {
c.bwPlan.ExecAsync(c.fftCBuf[i], c.fftRBuf[i])
out := outp.Comp(i)
copyPad(out, c.fftRBuf[i], c.size, padded, c.stream)
}
c.stream.Synchronize()
}
func (c *DemagConvolution) exec2D(outp, inp, vol *data.Slice, Bsat float64) {
// Convolution is separated into
// a 1D convolution for x and a 2D convolution for yz.
// So only 2 FFT buffers are needed at the same time.
// FFT x
zero1(c.fftRBuf[0], c.stream)
in := inp.Comp(0)
padded := c.kernSize
copyPadMul(c.fftRBuf[0], in, padded, c.size, vol, Bsat, c.stream)
c.fwPlan.ExecAsync(c.fftRBuf[0], c.fftCBuf[0])
// kern mul X
N1, N2 := c.fftKernSize[1], c.fftKernSize[2] // TODO: rm these
kernMulRSymm2Dx(c.fftCBuf[0], c.gpuFFTKern[0][0], N1, N2, c.stream)
// bw FFT x
c.bwPlan.ExecAsync(c.fftCBuf[0], c.fftRBuf[0])
out := outp.Comp(0)
copyPad(out, c.fftRBuf[0], c.size, padded, c.stream)
// FW FFT yz
for i := 1; i < 3; i++ {
zero1(c.fftRBuf[i], c.stream)
in := inp.Comp(i)
copyPadMul(c.fftRBuf[i], in, padded, c.size, vol, Bsat, c.stream)
c.fwPlan.ExecAsync(c.fftRBuf[i], c.fftCBuf[i])
}
// kern mul yz
kernMulRSymm2Dyz(c.fftCBuf[1], c.fftCBuf[2],
c.gpuFFTKern[1][1], c.gpuFFTKern[2][2], c.gpuFFTKern[1][2],
N1, N2, c.stream)
// BW FFT yz
for i := 1; i < 3; i++ {
c.bwPlan.ExecAsync(c.fftCBuf[i], c.fftRBuf[i])
out := outp.Comp(i)
copyPad(out, c.fftRBuf[i], c.size, padded, c.stream)
}
c.stream.Synchronize()
}