[Arm-netbook] mali gpu reverse engineering lkcl may ignore | libre GPU discussion

[Luke Kenneth Casson Leighton]

On Sat, Jun 24, 2017 at 9:55 AM, Bluey <bluey at smallfootprint.info> wrote:

> Please excuse my ignorance in such matters but would it be possible to
> use a RISC-V or FPGA chip as an interim eGPU until such time that a more
> specialised chip can be developed and released?

 on its own (as-is), no.  however with certain very very specific and
in some cases specialised SIMD instructions a reasonable approximation
can be had.  these operations are:

 * SIMD "and" for a bit-wise zero check (as large as possible)
 * inverse-squared function (for 1/x^2) - a very common operation in 3D
 * SIMD 12-14 bit accurate divide operation.

this latter turns out to be "good enough" for the majority of 3D
operations, where the accuracy on screens which only *have* 1920
pixels (11 bits being sufficient) division calculations beyond 12-14
bits is completely and utterly redundant... *under certain
circumstances*.

the point being that a divide operation which only requires 12-14 bits
of accuracy may complete in half the time, thus dramatically saving on
CPU cycles.

> I appreciate that RISC-V/FPGA chips are not likely to be well-suited to the task
> of GPU processing but perhaps they would be better than no GPU at all.

 not "and be power-efficient at the same time"

> Once a specialised libre GPU has been developed, the RISC-V / FPGA chips
> could be repurposed as a CPU for other projects/computers/laptops/etc. and,
> hence, ensuring that they don't go to waste.

 this was the reasoning behind ICubeCorp's "UPU" - Unified Processing
Unit - which unfortunately they kept proprietary.  i tried to help
them to understand the need to release the full boot initialisation
source code and to comply with the GPL but they did not follow up.

l.