Gabriel Hjort Åkerlund via llvm-dev
2020-Oct-29 04:40 UTC
[llvm-dev] Optimal variant of regbankselect
Hi Dominik, (Sorry for spamming; the first reply only went to you and not the list.) Cool that you want to try it out! There will for sure be some bugs in it, so please let me know if/when you find one and Ill fix it. And if you could make a testcase out of it, that would be superb (as theres currently a complete lack of tests). Although I havent measured it, I expect the compilation time to take about 3x more compared to greedy as it makes three passes over the instructions. However, the most amount of work is done in the first pass, which is comparable to the pass made in greedy. So hopefully its less than 3x, but this should really be measured over a set of functions to get an accurate figure. Also, there will most likely be improvements to be had to decrease compilation time. Cheers, Gabriel From: llvm-dev <llvm-dev-bounces at lists.llvm.org> On Behalf Of Dominik Montada via llvm-dev Sent: den 28 oktober 2020 14:39 To: llvm-dev at lists.llvm.org Subject: Re: [llvm-dev] Optimal variant of regbankselect Hi Gabriel, thank you so much for doing this! I'll try out the patch in our downstream implementation right away. Do you know how big of an impact this has on compile time compared to fast and greedy? Cheers, Dominik Am 28.10.20 um 14:32 schrieb Gabriel Hjort Åkerlund via llvm-dev: Hi all, I have made an attempt of implementing an optimal variant of the register bank selector (regbankselect). The code is available for review at https://reviews.llvm.org/D90304 <https://protect2.fireeye.com/v1/url?k=1bd4b7f0-45642a68-1bd4f76b-861fcb972b fc-40f905c996dc0e0e&q=1&e=2c298e05-1e99-4460-b1a6-8895b7e94086&u=https%3A%2F %2Freviews.llvm.org%2FD90304> , and I would greatly appreciate if anyone interested can provide their comments. I have run a few tests the regbankselect-*.mir testcases for AAarch64 and it seems to work, but more tests are surely needed to increase confidence in the implementation. I also tried using AMDGPU, but that backend does not provide the full list of InstructionMappings for a given MachineInstr, which is needed in order to compute the optimal selection of register banks. Cheers, Gabriel Hjort Åkerlund _______________________________________________ LLVM Developers mailing list llvm-dev at lists.llvm.org <mailto:llvm-dev at lists.llvm.org> https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev <https://protect2.fireeye.com/v1/url?k=7a04028d-24b49f15-7a044216-861fcb972b fc-6323099c52b73708&q=1&e=2c298e05-1e99-4460-b1a6-8895b7e94086&u=https%3A%2F %2Flists.llvm.org%2Fcgi-bin%2Fmailman%2Flistinfo%2Fllvm-dev> -- ---------------------------------------------------------------------- Dominik Montada Email: dominik.montada at hightec-rt.com <mailto:dominik.montada at hightec-rt.com> HighTec EDV-Systeme GmbH Phone: +49 681 92613 19 Europaallee 19 Fax: +49-681-92613-26 D-66113 Saarbrücken WWW: http://www.hightec-rt.com <https://protect2.fireeye.com/v1/url?k=c80bd3cf-96bb4e57-c80b9354-861fcb972b fc-bf8c336a375c6bef&q=1&e=2c298e05-1e99-4460-b1a6-8895b7e94086&u=http%3A%2F% 2Fwww.hightec-rt.com%2F> Managing Director: Vera Strothmann Register Court: Saarbrücken, HRB 10445, VAT ID: DE 138344222 This e-mail may contain confidential and/or privileged information. If you are not the intended recipient please notify the sender immediately and destroy this e-mail. Any unauthorised copying, disclosure or distribution of the material in this e-mail is strictly forbidden. --- -------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20201029/90391071/attachment.html> -------------- next part -------------- A non-text attachment was scrubbed... Name: smime.p7s Type: application/pkcs7-signature Size: 6320 bytes Desc: not available URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20201029/90391071/attachment.bin>
Nicolai Hähnle via llvm-dev
2020-Oct-30 15:49 UTC
[llvm-dev] Optimal variant of regbankselect
Hi Gabriel, it's cool that you try to do this, but I think you need some more proof before calling this optimal :) Put briefly, dynamic programming approaches for this kind of mapping/matching can be optimal if the underlying dependency structure is a tree. If it's a more general DAG, optimality goes out of the window. In your dynamic programming approach, my understanding is that you remember the cost of every possible "realization"/"mapping" of every node individually. So, you end up computing N*M pieces of information, where N is the number of nodes and M is the number of options to choose from at each node. Unfortunately, the final selection of choices only works in a tree, because you basically have to choose each node's "realization"/"mapping" based on a single successor. As soon as you're in a general DAG, where a node has to satisfy *multiple* successors, that no longer works. You could theoretically extend the dynamic programming approach, but only by computing information about the cost of *correlated* choices for multiple nodes simultaneously. But then, your performance can go out of the window because you need to compute something like N*M^K pieces of information, where K is a bound on the number of nodes whose choices you need to consider simultaneously. K cannot be bounded in general (other than by some trivial function of the number of nodes in the graph), which means you end up with an exponential worst case if you want to solve this optimally. I would actually expect optimal register bank selection on a general DAG to be NP-complete, but I haven't thought about it too deeply. Cheers, Nicolai On Thu, Oct 29, 2020 at 5:41 AM Gabriel Hjort Åkerlund via llvm-dev <llvm-dev at lists.llvm.org> wrote:> > Hi Dominik, > > > > (Sorry for spamming; the first reply only went to you and not the list.) > > > > Cool that you want to try it out! There will for sure be some bugs in it, so please let me know if/when you find one and I’ll fix it. And if you could make a testcase out of it, that would be superb (as there’s currently a complete lack of tests). > > > > Although I haven’t measured it, I expect the compilation time to take about 3x more compared to greedy as it makes three passes over the instructions. However, the most amount of work is done in the first pass, which is comparable to the pass made in greedy. So hopefully it’s less than 3x, but this should really be measured over a set of functions to get an accurate figure. Also, there will most likely be improvements to be had to decrease compilation time. > > > > Cheers, > > Gabriel > > > > From: llvm-dev <llvm-dev-bounces at lists.llvm.org> On Behalf Of Dominik Montada via llvm-dev > Sent: den 28 oktober 2020 14:39 > To: llvm-dev at lists.llvm.org > Subject: Re: [llvm-dev] Optimal variant of regbankselect > > > > Hi Gabriel, > > thank you so much for doing this! I'll try out the patch in our downstream implementation right away. Do you know how big of an impact this has on compile time compared to fast and greedy? > > Cheers, > > Dominik > > Am 28.10.20 um 14:32 schrieb Gabriel Hjort Åkerlund via llvm-dev: > > Hi all, > > > > I have made an attempt of implementing an optimal variant of the register bank selector (regbankselect). The code is available for review at https://reviews.llvm.org/D90304, and I would greatly appreciate if anyone interested can provide their comments. I have run a few tests the regbankselect-*.mir testcases for AAarch64 and it seems to work, but more tests are surely needed to increase confidence in the implementation. I also tried using AMDGPU, but that backend does not provide the full list of InstructionMappings for a given MachineInstr, which is needed in order to compute the optimal selection of register banks. > > > > Cheers, > > Gabriel Hjort Åkerlund > > > > _______________________________________________ > > LLVM Developers mailing list > > llvm-dev at lists.llvm.org > > https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev > > -- > > ---------------------------------------------------------------------- > > Dominik Montada Email: dominik.montada at hightec-rt.com > > HighTec EDV-Systeme GmbH Phone: +49 681 92613 19 > > Europaallee 19 Fax: +49-681-92613-26 > > D-66113 Saarbrücken WWW: http://www.hightec-rt.com > > > > Managing Director: Vera Strothmann > > Register Court: Saarbrücken, HRB 10445, VAT ID: DE 138344222 > > > > This e-mail may contain confidential and/or privileged information. If > > you are not the intended recipient please notify the sender immediately > > and destroy this e-mail. Any unauthorised copying, disclosure or > > distribution of the material in this e-mail is strictly forbidden. > > --- > > _______________________________________________ > LLVM Developers mailing list > llvm-dev at lists.llvm.org > https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev-- Lerne, wie die Welt wirklich ist, aber vergiss niemals, wie sie sein sollte.
Gabriel Hjort Åkerlund via llvm-dev
2020-Nov-02 14:03 UTC
[llvm-dev] Optimal variant of regbankselect
Hi Nicolai, Thanks for your response! Fair enough, I will rename it to "global" as that is a property that we can all agree on. And if it happens to compute the optimal selection, then that's just a bonus. =) Cheers, Gabriel -----Original Message----- From: Nicolai Hähnle <nhaehnle at gmail.com> Sent: den 30 oktober 2020 16:49 To: Gabriel Hjort Åkerlund <gabriel.hjort.akerlund at ericsson.com> Cc: llvm-dev at lists.llvm.org Subject: Re: [llvm-dev] Optimal variant of regbankselect Hi Gabriel, it's cool that you try to do this, but I think you need some more proof before calling this optimal :) Put briefly, dynamic programming approaches for this kind of mapping/matching can be optimal if the underlying dependency structure is a tree. If it's a more general DAG, optimality goes out of the window. In your dynamic programming approach, my understanding is that you remember the cost of every possible "realization"/"mapping" of every node individually. So, you end up computing N*M pieces of information, where N is the number of nodes and M is the number of options to choose from at each node. Unfortunately, the final selection of choices only works in a tree, because you basically have to choose each node's "realization"/"mapping" based on a single successor. As soon as you're in a general DAG, where a node has to satisfy *multiple* successors, that no longer works. You could theoretically extend the dynamic programming approach, but only by computing information about the cost of *correlated* choices for multiple nodes simultaneously. But then, your performance can go out of the window because you need to compute something like N*M^K pieces of information, where K is a bound on the number of nodes whose choices you need to consider simultaneously. K cannot be bounded in general (other than by some trivial function of the number of nodes in the graph), which means you end up with an exponential worst case if you want to solve this optimally. I would actually expect optimal register bank selection on a general DAG to be NP-complete, but I haven't thought about it too deeply. Cheers, Nicolai On Thu, Oct 29, 2020 at 5:41 AM Gabriel Hjort Åkerlund via llvm-dev <llvm-dev at lists.llvm.org> wrote:> > Hi Dominik, > > > > (Sorry for spamming; the first reply only went to you and not the > list.) > > > > Cool that you want to try it out! There will for sure be some bugs in it, so please let me know if/when you find one and I’ll fix it. And if you could make a testcase out of it, that would be superb (as there’s currently a complete lack of tests). > > > > Although I haven’t measured it, I expect the compilation time to take about 3x more compared to greedy as it makes three passes over the instructions. However, the most amount of work is done in the first pass, which is comparable to the pass made in greedy. So hopefully it’s less than 3x, but this should really be measured over a set of functions to get an accurate figure. Also, there will most likely be improvements to be had to decrease compilation time. > > > > Cheers, > > Gabriel > > > > From: llvm-dev <llvm-dev-bounces at lists.llvm.org> On Behalf Of Dominik > Montada via llvm-dev > Sent: den 28 oktober 2020 14:39 > To: llvm-dev at lists.llvm.org > Subject: Re: [llvm-dev] Optimal variant of regbankselect > > > > Hi Gabriel, > > thank you so much for doing this! I'll try out the patch in our downstream implementation right away. Do you know how big of an impact this has on compile time compared to fast and greedy? > > Cheers, > > Dominik > > Am 28.10.20 um 14:32 schrieb Gabriel Hjort Åkerlund via llvm-dev: > > Hi all, > > > > I have made an attempt of implementing an optimal variant of the register bank selector (regbankselect). The code is available for review at https://reviews.llvm.org/D90304, and I would greatly appreciate if anyone interested can provide their comments. I have run a few tests the regbankselect-*.mir testcases for AAarch64 and it seems to work, but more tests are surely needed to increase confidence in the implementation. I also tried using AMDGPU, but that backend does not provide the full list of InstructionMappings for a given MachineInstr, which is needed in order to compute the optimal selection of register banks. > > > > Cheers, > > Gabriel Hjort Åkerlund > > > > _______________________________________________ > > LLVM Developers mailing list > > llvm-dev at lists.llvm.org > > https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev > > -- > > ---------------------------------------------------------------------- > > Dominik Montada Email: dominik.montada at hightec-rt.com > > HighTec EDV-Systeme GmbH Phone: +49 681 92613 19 > > Europaallee 19 Fax: +49-681-92613-26 > > D-66113 Saarbrücken WWW: https://protect2.fireeye.com/v1/url?k=1a1aa8a6-44ab73c6-1a1ae83d-86e2237f51fb-3c96ae1cad03cc84&q=1&e=083f110d-1869-45bf-91fb-ac31bacb8334&u=http%3A%2F%2Fwww.hightec-rt.com%2F > > > > Managing Director: Vera Strothmann > > Register Court: Saarbrücken, HRB 10445, VAT ID: DE 138344222 > > > > This e-mail may contain confidential and/or privileged information. If > > you are not the intended recipient please notify the sender > immediately > > and destroy this e-mail. Any unauthorised copying, disclosure or > > distribution of the material in this e-mail is strictly forbidden. > > --- > > _______________________________________________ > LLVM Developers mailing list > llvm-dev at lists.llvm.org > https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev-- Lerne, wie die Welt wirklich ist, aber vergiss niemals, wie sie sein sollte. -------------- next part -------------- A non-text attachment was scrubbed... Name: smime.p7s Type: application/pkcs7-signature Size: 6320 bytes Desc: not available URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20201102/0b851839/attachment.bin>