Input handling rework for SI
Christian König
christian.koenig at amd.com
Wed Mar 6 14:00:51 PST 2013
Am 06.03.2013 21:18, schrieb Tom Stellard:
> On Wed, Mar 06, 2013 at 09:13:38PM +0100, Christian König wrote:
>> [SNIP]
>> Ok, going to do so. Thx for testing.
>>
>> BTW: I've got indirect addressing working with SI, but I'm a bit
>> confused about the common TGSI to LLVM code in mesa for R600 and SI.
>> Did that ever worked for R600? I think I found a serious bug in it,
>> but I'm not 100% sure.
>>
> The TGSI to LLVM indirect addressing code never worked for R600, it's
> not even fully implmented. It just reuses whatever common code is shared
> with llvmpipe.
>
> For indirect addressin on SI, were you able to the
> AMDGPUIndirectAddressingPass?
No, I didn't even tried to go with indirect memory addressing. Instead I
just loaded all temps/inputs/outputs/whatever into into a vector of
appropriate size and then used insert_element/extract_element with an
variable index.
That works great for SI (and actually only needs 117 lines of code on
the LLVM side :D ), but I'm not sure if it's a good idea for R600.
Anyway I need some sleep, not going to commit anything before tomorrow.
Christian.
More information about the llvm-commits
mailing list