Made tf.abs underflow-safe on CPU & GPU and added test for it.

[DirkToewe]

Description

Fixes part of #895. The CPU backend now uses Math.hypot and the WebGL backend uses length albeit with some workaround to ensure underflow-safety (which is otherwise not given, at least on Intel GPUs).

---

For repository owners only:

Please remember to apply all applicable tags to your pull request.
Tags: FEATURE, BREAKING, BUG, PERF, DEV, DOC, SECURITY

For more info see: /~https://github.com/tensorflow/tfjs/blob/master/DEVELOPMENT.md

---

This change is 

[nsthorat]

Thank you for the fix, really nice work! I sent some minor styling comments just so you get familiar with our style for the future :)

Reviewable status: 0 of 1 approvals obtained (waiting on @DirkToewe)

---

src/kernels/webgl/complex_abs_gpu.ts, line 29 at r1 (raw file):

    this.userCode = `
      void main() {
        float re = abs( getRealAtOutCoords() );

could you remove the extra white space? after and before the parens (here and below)

---

src/kernels/webgl/complex_abs_gpu.ts, line 31 at r1 (raw file):

        float re = abs( getRealAtOutCoords() );
        float im = abs( getImagAtOutCoords() );
        float mx = max(re,im);

there should be an extra space between re, im

---

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

        // sadly the length function in glsl is not underflow-safe
        // (at least not on Integl GPUs). So the safe solution is

Intel

---

src/kernels/webgl/complex_abs_gpu.ts, line 37 at r1 (raw file):

        // to ensure underflow-safety in all cases.
        setOutput(
          mx * length(  vec2(1, min(re,im)/mx)  )

remove the extra spaces inside of length()

---

src/ops/unary_ops_test.ts, line 173 at r1 (raw file):

  it('is underflow-safe for complex64', () => {
    const a = tf.complex(
       /*re=*/[ 1e-30, 0     ],

no need for these comments :)

[DirkToewe]

Done, hope I didn't forget anything.

[nsthorat]

LGTM -- I will merge after one small change which will fix the travis build

Reviewable status: 0 of 1 approvals obtained (waiting on @DirkToewe)

---

src/ops/unary_ops_test.ts, line 173 at r2 (raw file):

  it('is underflow-safe for complex64', () => {
    const a = tf.complex(
       [1e-30, 0    ],

One last comment, instead of 1e-30 can you use tf.ENV.get('EPSILON') which is the smallest representable value larger than zero (works on all devices, this is why travis is failing for iOS which is float16 which has a larger epsilon).

[DirkToewe]

eps² will not underflow, but I tried to improvise...

There was a 0/0 bug that caused the NaN and which should now be fixed.

[nsthorat]

Hm, I know what the issue is (we're uploading this as a uniform) so I'll merge this change and then fix it :)

[nsthorat]

Really nice work!

[DirkToewe]

Thank You!
What issue are You referring to? The 0/0 was in complex_abs_gpu and I hope I fixed it. My other statement about eps² generally not rounding down to zero is not a bug. What would be necessary is another constant like ENV.get('MIN_VALUE') to determine a value that is close to underflow.

Take float64 for example: Number.EPSILON=2.220446049250313e-16 is actually a ginormously huge number compared to Number.MIN_VALUE=5e-324. It is only small compared to one in that 1+eps = 1.