[SYCL] Expanded Ahead Of Time Compilation support #262

mdtoguchi · 2019-06-27T15:55:46Z

Adds support for AOT compilation in a single pass. Support is triggered by
using a sub architecture value in the triple. The following 3 are supported:

spir64_fpga, spir64_x86_64, spir64_gen

When one of these are used, the respective backend compiler (aoc, ioc or cloc)
is invoked to perform the AOT compilation. That device binary is then wrapped
and included in the final binary.

Additional options can also be added to the SYCL device compilation, the
Backend compile or the device link. These are handled by -Xsycl-compiler,
-Xsycl-target or -Xsycl-linker respectively

Signed-off-by: Michael D Toguchi michael.d.toguchi@intel.com

llvm/lib/Support/Triple.cpp

asavonic · 2019-06-27T21:01:10Z

clang/lib/Driver/ToolChains/SYCL.cpp

+  CmdArgs.push_back("-cmd=link");
+  CmdArgs.push_back("-device=cpu");
+  for (const auto &II : Inputs) {
+    CmdArgs.push_back(Args.MakeArgString(Twine("-binary=") + II.getFilename()));


What are the Inputs here? LLVM IR Bitcode with spir64 triple?

What are the Inputs here? LLVM IR Bitcode with spir64 triple?

Yes, that is the expected input for ioc. Depending on expectations of the tool, this can either be LLVM-IR or SPIR-V.

This may work for now, but LLVM IR is not a stable interface and older/newer CPU runtime may not consume LLVM IR generated by SYCL clang.

Be optimistic... :-)

mdtoguchi · 2019-07-02T18:50:26Z

Changes refactored to remove CPU support for now. CPU support will be added later.

asavonic · 2019-07-09T13:56:17Z

Can you include an example of compilation commands for 1-step offline compilation for FPGA and GPU?

mdtoguchi · 2019-07-09T15:31:41Z

Can you include an example of compilation commands for 1-step offline compilation for FPGA and GPU?

FPGA: clang++ -fsycl -fsycl-targets=spir64_fpga-unknown-linux-sycldevice
testfile.cpp -Xsycl-target "-DFOO"
GEN: clang++ -fsycl -fsycl-targets=spir64_gen-unknown-linux-sycldevice
testfile.cpp -Xsycl-target "-device skl"

I also added these command lines to the commit message.

agozillon · 2019-07-09T17:51:45Z

In general I would be quite interested at some point in the future in seeing some shiny new documentation on the commands added to the driver and how to use all these new commands, maybe that's a silly request though which I can understand.

But I think it does help to elaborate a little on it outside of the table gen files/commit logs. I know there is some coverage of this inside of: https://github.com/intel/llvm/blob/sycl/sycl/doc/SYCL_compiler_and_runtime_design.md but I'm not too sure how up-to-date it is (and perhaps there is some documentation elsewhere I'm unaware of).

mdtoguchi · 2019-07-09T17:54:42Z

Definitely not a silly request. We are working on adding documentation which covers all of the new options being provided. From: agozillon <notifications@github.com> Sent: Tuesday, July 09, 2019 10:52 AM To: intel/llvm <llvm@noreply.github.com> Cc: Toguchi, Michael D <michael.d.toguchi@intel.com>; Author <author@noreply.github.com> Subject: Re: [intel/llvm] [SYCL] Expanded Ahead Of Time Compilation support (#262) In general I would be quite interested at some point in the future in seeing some shiny new documentation on the commands added to the driver and how to use all these new commands, maybe that's a silly request though which I can understand. But I think it does help to elaborate a little on it outside of the table gen files/commit logs. I know there is some coverage of this inside of: https://github.com/intel/llvm/blob/sycl/sycl/doc/SYCL_compiler_and_runtime_design.md but I'm not too sure how up-to-date it is (and perhaps there is some documentation elsewhere I'm unaware of). — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub<#262?email_source=notifications&email_token=ALNNPVCZJ7QORZY55OYO4LTP6TF3JA5CNFSM4H35PKCKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGODZRATSQ#issuecomment-509741514>, or mute the thread<https://github.com/notifications/unsubscribe-auth/ALNNPVBBF64YPTSWQ7D5NYLP6TF3JANCNFSM4H35PKCA>.

mdtoguchi · 2019-07-10T17:17:31Z

Any other comments? Would be great to get this in so I can (more easily) work on the AOT CPU enabling work. :)

bader

@mdtoguchi, looks good.
I have a couple of suggestions.
Let me know how do you feel about applying them.

clang/lib/Driver/Driver.cpp

clang/include/clang/Basic/DiagnosticDriverKinds.td

clang/lib/Driver/Driver.cpp

clang/lib/Driver/ToolChains/SYCL.cpp

mdtoguchi · 2019-07-10T18:18:16Z

@mdtoguchi, looks good.
I have a couple of suggestions.
Let me know how do you feel about applying them.

I can work on your suggestions and also modify the option names added to match what was decided for the -Xsycl*.

bader · 2019-07-10T18:21:43Z

@mdtoguchi, looks good.
I have a couple of suggestions.
Let me know how do you feel about applying them.

I can work on your suggestions and also modify the option names added to match what was decided for the -Xsycl*.

Great. I'll wait for updated version then.
Thanks for working on this and sorry for the late review.

Adds support for AOT compilation in a single pass. Support is triggered by using a sub architecture value in the triple. The following 2 are supported: spir64_fpga, spir64_gen When one of these are used, the respective backend compiler (aoc, ocloc) is invoked to perform the AOT compilation. That device binary is then wrapped and included in the final binary. Additional options can also be added to the SYCL device compilation, the Backend compile or the device link. These are handled by -Xsycl-target-frontend, -Xsycl-target-backend or -Xsycl-target-linker respectively Example command lines: FPGA: clang++ -fsycl -fsycl-targets=spir64_fpga-unknown-linux-sycldevice testfile.cpp -Xsycl-target-backend "-DFOO" GEN: clang++ -fsycl -fsycl-targets=spir64_gen-unknown-linux-sycldevice testfile.cpp -Xsycl-target-backend "-device skl" Signed-off-by: Michael D Toguchi <michael.d.toguchi@intel.com>

mdtoguchi · 2019-07-11T14:04:00Z

I didn't even realize that was still in there. That was from an early implementation! triples are definitely kept around and can be used for compilation. I'll clean it up with the AOT for CPU work. From: Alexey Bader <notifications@github.com> Sent: Thursday, July 11, 2019 1:25 AM To: intel/llvm <llvm@noreply.github.com> Cc: Toguchi, Michael D <michael.d.toguchi@intel.com>; Mention <mention@noreply.github.com> Subject: Re: [intel/llvm] [SYCL] Expanded Ahead Of Time Compilation support (#262) @bader commented on this pull request.

________________________________ In clang/lib/Driver/Driver.cpp<#262 (comment)>:

@@ -3263,6 +3282,43 @@ class OffloadingActionBuilder final {

} } } + // Gather information about the SYCL Ahead of Time targets. The targets + // are determined on the SubArch values passed along in the triple. + // The SubArch information for SYCL offload is not used during the + // compilation and is only used to determine additional compilation steps + // needed in the driver toolchain. + Arg *SYCLTargets = + C.getInputArgs().getLastArg(options::OPT_fsycl_targets_EQ); + if (SYCLTargets) { + llvm::StringMap<const char *> FoundNormalizedTriples; + for (const char *Val : SYCLTargets->getValues()) { Okay. Let's ignore this comment. — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub<#262?email_source=notifications&email_token=ALNNPVCHDDNNIUR3H7KVMNLP63U5FA5CNFSM4H35PKCKYY3PNVWWK3TUL52HS4DFWFIHK3DMKJSXC5LFON2FEZLWNFSXPKTDN5WW2ZLOORPWSZGOB6DYNWQ#discussion_r302420164>, or mute the thread<https://github.com/notifications/unsubscribe-auth/ALNNPVFCSKGIH4RX62TG7PDP63U5FANCNFSM4H35PKCA>.

mdtoguchi requested review from kbobrovs and vzakhari June 27, 2019 15:55

kbobrovs reviewed Jun 27, 2019

View reviewed changes

llvm/lib/Support/Triple.cpp Outdated Show resolved Hide resolved

asavonic reviewed Jun 27, 2019

View reviewed changes

mdtoguchi force-pushed the private/mdtoguchi/aot-enhancements branch from 96d3590 to 2f8c780 Compare June 29, 2019 01:03

agozillon mentioned this pull request Jul 1, 2019

[SYCL] Driver/ToolChain Extension #53

Closed

mdtoguchi force-pushed the private/mdtoguchi/aot-enhancements branch from 2f8c780 to 552d378 Compare July 2, 2019 18:48

bader requested a review from asavonic July 9, 2019 13:44

mdtoguchi force-pushed the private/mdtoguchi/aot-enhancements branch from 552d378 to deeec85 Compare July 9, 2019 15:29

bader reviewed Jul 10, 2019

View reviewed changes

mdtoguchi force-pushed the private/mdtoguchi/aot-enhancements branch from deeec85 to 788e384 Compare July 11, 2019 01:11

bader approved these changes Jul 11, 2019

View reviewed changes

bader merged commit c451ac3 into intel:sycl Jul 11, 2019

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[SYCL] Expanded Ahead Of Time Compilation support #262

[SYCL] Expanded Ahead Of Time Compilation support #262

mdtoguchi commented Jun 27, 2019

asavonic Jun 27, 2019

mdtoguchi Jun 27, 2019

asavonic Jun 28, 2019

keryell Jun 28, 2019

mdtoguchi commented Jul 2, 2019

asavonic commented Jul 9, 2019

mdtoguchi commented Jul 9, 2019

agozillon commented Jul 9, 2019

mdtoguchi commented Jul 9, 2019 via email

mdtoguchi commented Jul 10, 2019

bader left a comment

mdtoguchi commented Jul 10, 2019

bader commented Jul 10, 2019

mdtoguchi commented Jul 11, 2019 via email

[SYCL] Expanded Ahead Of Time Compilation support #262

[SYCL] Expanded Ahead Of Time Compilation support #262

Conversation

mdtoguchi commented Jun 27, 2019

asavonic Jun 27, 2019

Choose a reason for hiding this comment

mdtoguchi Jun 27, 2019

Choose a reason for hiding this comment

asavonic Jun 28, 2019

Choose a reason for hiding this comment

keryell Jun 28, 2019

Choose a reason for hiding this comment

mdtoguchi commented Jul 2, 2019

asavonic commented Jul 9, 2019

mdtoguchi commented Jul 9, 2019

agozillon commented Jul 9, 2019

mdtoguchi commented Jul 9, 2019 via email

mdtoguchi commented Jul 10, 2019

bader left a comment

Choose a reason for hiding this comment

mdtoguchi commented Jul 10, 2019

bader commented Jul 10, 2019

mdtoguchi commented Jul 11, 2019 via email