Improve shader attribute precision on mobile devices

[lucaswoj]

There are two outstanding bugs related to poorly calibrated shader attribute precision on mobile devices (#2055, #2060 and #2020)

@kristfal kindly provided a PR (#2091) with one potential solution to the problem.

We should all take a step back and come up with our ideal blue-sky solution to the problem. Should we

• keep adding highp annotations to the shader source when bugs are reported
• design a "smarter" (than this) shader attribute precision rewriting algorithm
• other?

cc @kristfal @ansis @kkaefer @jfirebaugh

[jfirebaugh]

Previously: #1852, #750.

I think we should use highp universally and unconditionally in vertex shaders. There appears to be little or no downside to doing so.

For fragment shaders, it's more complicated. Do we know of any cases where our fragment shaders produce incorrect results on mobile with mediump?

[ansis]

I think we should use highp universally and unconditionally in vertex shaders. There appears to be little or no downside to doing so.

For fragment shaders, it's more complicated. Do we know of any cases where our fragment shaders produce incorrect results on mobile with mediump?

This sounds right to me. I don't think there are any cases we need high precision in fragment shaders.

[lucaswoj]

Naively switching most shader attributes to highp has led to many rendering errors on desktop devices, even though they claim highp support

        var isHighpSupported = (
            context.getShaderPrecisionFormat &&
            context.getShaderPrecisionFormat(type, context.HIGH_FLOAT).precision > 0
        );

Do we want to continue restricting highp use to mobile devices?

[jfirebaugh]

Naively switching most shader attributes to highp has led to many rendering errors on desktop devices

Can you elaborate on this?

[lucaswoj]

In my prototype branch, #2117, circles and lines fail to render in the browser and in headless testing. I'm not yet sure what the underlying cause is.

[ansis]

done #2096