[PATCH vkd3d 00/41] Add support for DXIL shaders

[Hans-Kristian Arntzen]

On 1/29/20 4:31 PM, Hans-Kristian Arntzen wrote:
>
> On 1/29/20 3:09 PM, Rémi Bernon wrote:
>> On 1/29/20 2:54 PM, Andrew Eikum wrote:
>>> On Wed, Jan 29, 2020 at 12:51:25PM +0100, Hans-Kristian Arntzen wrote:
>>>> Hans-Kristian Arntzen (41):
>>>>
>>>>   16 files changed, 6544 insertions(+), 724 deletions(-)
>>>
>>> I can't speak to the content of the series, but this is a huuuuuge
>>> amount of code to dump all at once. I'd suggest sending series of
>>> about 4-5 patches at a time.
>>>
>>> Andrew
>>>
>>
>> Hi,
>>
>> I wanted to have a look and the patches don't apply cleanly either 
>> (it fails at #02, but later as well).
>
> There seems to have been some commits in-between master and the 
> patchset, which caused issue when I tried to apply them to clean 
> master. Doesn't seem like there is much of a mood for getting another 
> 41 commits on the ML, so just uploading a tarball instead if that's 
> okay. Alternatively, a branch is here: 
> https://github.com/HansKristian-Work/vkd3d/tree/dxil-review.
>
>>
>> Also I tried to build dxil-spirv on Debian, and it misses some LLVM 
>> include paths in the CMakeLists.txt. Fixing that is easy, but then it 
>> also requires another include fix to build with llvm-10. Then I tried 
>> several other version and LLVM 8, 9 and 10 seems to build fine, but 
>> it starts to fail compiling with earlier versions of LLVM.
>>
>> In my experience, building a project using LLVM C++ API makes it very 
>> brittle, as they tend to change their API quite often. The projects 
>> fall behind very quickly as they cannot keep up with LLVM development 
>> speed, and it becomes very hard to build.
>>
>> I'm not sure what's the target but I believe that since the LLVM IR 
>> format is advertised as stable since 3.7 -and DXIL format probably 
>> even more- depending on some unstable API to parse it goes a bit 
>> against it.
>>
>> A better solution to this, IMHO, would be to have a custom parser for 
>> DXIL. Although it's a bit more work, the project would be more self 
>> contained and much more portable.
>
> Yes, that's the ideal solution. I am aware of this problem, but I had 
> hoped to defer this. Writing a good BC parser from scratch is likely 
> another several man-months of work on top unless someone have made 
> useful progress on that already.

FWIW, on dxil-spirv master now, LLVM is used as a submodule instead and 
everything is linked statically, so any incompatibility should be gone.

Cheers,
Hans-Kristian

>
> Cheers,
> Hans-Kristian
>
>>
>> Cheers,