Intel Posts FP16 GLSL Patches For Mesa

[phoronix]

Phoronix: Intel Posts FP16 GLSL Patches For Mesa

Topi Pohjolainen of Intel has been working on adding proper FP16/half-precision support to Mesa's GLSL code...

http://www.phoronix.com/scan.php?page=news_item&px=Intel-FP16-GLSL-Patches

[microcode]

I wonder if their next big arch update will include actual FP16 (and int16) compute (or does it already exist?).

[Kayden]

The shader ISA has supported W types (int16) since the original 965. HF (16-bit float) is newer, but has existed for quite some time as well.

[mzs.112000]

So, because it is in GLSL, that means that everything that supports OpenGL 3.3 will automatically support FP16(even if it is slower due to lack of native hardware) functions? Or am I missing something?
I wonder what ever happened to that project that was building a GLSL-based FP64 library for cards without native FP64 support?

[coder]

I wonder if their next big arch update will include actual FP16 (and int16) compute (or does it already exist?).

Intel iGPUs added native fp16 in Gen8 (coincident with Broadwell). As with Nvidia's GP100 and AMD's Vega, its throughput is 2x that of fp32.
In contrast, the way Nvidia's tensor engine works, in Volta, is rather different and might not be easily utilized for ordinary shader arithmetic.


This is worth a look:

https://software.intel.com/sites/def...ics%20Gen8.pdf

Here's a nice comparison of their various generations and models:

https://en.wikipedia.org/wiki/List_o...hth_generation