[Driver][SYCL][FPGA] Enable dependency file usage from static archive… #2443
Conversation
mdtoguchi
force-pushed
the
private/mdtoguchi/fpga-deps-from-static-libs
branch
from
1dda16f to
5ed1c55
Compare
…s for FPGA When performing compilations for FPGA, we want to be sure to take advantage of dependency information that could be part of the fat static archives.
mdtoguchi
force-pushed
the
private/mdtoguchi/fpga-deps-from-static-libs
branch
from
5ed1c55 to
0b9e33e
Compare
| /// Check for unbundle and use of deps in static lib | ||
| // RUN: %clang_cl --target=x86_64-pc-windows-msvc -fsycl -fintelfpga %t.lib -### 2>&1 \ | ||
| // RUN: | FileCheck -check-prefix=CHECK_UNBUNDLE %s | ||
| // CHECK_UNBUNDLE: clang-offload-bundler" "-type=aoo" "-targets=sycl-fpga_dep" "-inputs={{.*}}" "-outputs=[[DEPFILES:.+\.txt]]" "-unbundle" |
There was a problem hiding this comment.
Just wondering, why is it that separate unbundler calls are being made using the same archive instead of a single one as a comma? I can't quite trace this back in the code...
There was a problem hiding this comment.
The unbundling of the dep files is separate from the unbundle call after the partial-link. The behavior is similar to what we do for grabbing the dep files for objects.
There was a problem hiding this comment.
Might've formulated this incorrectly - didn't mean this thread to be strictly focused on the current patch :) More of a "why do we have to have separate unbundler calls at all if we're able to get multiple outputs from a single one". But I have a vague idea that the Action construction could be problematic, given the nature of the hack that allows us to have multiple unbundler outputs at all.
There was a problem hiding this comment.
Right. The additional dependency target doesn't fit within the regular unbundler behavior expectations (full toolchains for each target), so we add the dep file extraction as a singular input to the offline compile. There may be a way to accomplish this, but would probably require a bit more investigation/rework.
This change fixes cases when the input function name does not match OpenCL spec but starts with `convert_` prefix, e.g. `convert_float_42`. Original commit: KhronosGroup/SPIRV-LLVM-Translator@848a6d7af70ba28
Check that builtin is valid mostly by matching the regular expression - remove unneeded checks. This is a follow-up to #2443 Original commit: KhronosGroup/SPIRV-LLVM-Translator@04144823462c6c4
[benchmarks] adjust label and legend positions on bar charts
…s for FPGA
When performing compilations for FPGA, we want to be sure to take advantage
of dependency information that could be part of the fat static archives.