Add tf.ifft().

[nsthorat]

This is the inverse of tf.fft(). Luckily this is simple as changing -2 * PI to 2 * PI.

Like tf.fft(), only implements 1D on CPU. For GPU, all dimensions are supported.

Towards tensorflow/tfjs#782

For reference:
https://www.mathworks.com/help/matlab/ref/ifft.html
https://www.tensorflow.org/api_docs/python/tf/spectral/ifft

---

This change is 

[dsmilkov]

Reviewed 9 of 9 files at r1.
Reviewable status: 0 of 1 approvals obtained (waiting on @nsthorat and @dsmilkov)

---

src/kernels/backend_cpu.ts, line 2572 at r1 (raw file):

      if (inverse) {
        result = ops.complex(
                     ops.real(result).div(scalar(n)),

place scalar(n) in a temp variable so we don't upload twice to gpu (potentially)

---

src/kernels/webgl/fft_gpu.ts, line 34 at r1 (raw file):

    this.outputShape = inputShape;

    const pi = '3.1415926535897932384626433832795';

use built-in Math.PI

---

src/kernels/webgl/fft_gpu.ts, line 61 at r1 (raw file):

          float imag = getImag(batch, i);

          result += (

why the parenthesis?

[nsthorat]

Reviewable status: complete! 1 of 1 approvals obtained (waiting on @nsthorat)

---

src/kernels/backend_cpu.ts, line 2572 at r1 (raw file):

Previously, dsmilkov (Daniel Smilkov) wrote…

place scalar(n) in a temp variable so we don't upload twice to gpu (potentially)

This is in backend_cpu

---

src/kernels/webgl/fft_gpu.ts, line 34 at r1 (raw file):

Previously, dsmilkov (Daniel Smilkov) wrote…

use built-in Math.PI

Done

---

src/kernels/webgl/fft_gpu.ts, line 61 at r1 (raw file):

Previously, dsmilkov (Daniel Smilkov) wrote…

why the parenthesis?

No good reason, removed.