digitalmars.D - [GSoC 2026] DCompute Vulkan Backend
Hi everyone, I am making this thread to share my weekly updates for the **DCompute Vulkan Backend** project. Here are the 3 main parts of my project (Mentored by Nicholas Wilson): * **LDC Compiler:** Connect D's data types with Vulkan's memory rules by adding a wrapper kernel. I will also fix pointer sizes to make correct SPIR-V files. * **Testing:** Check the compiled `.spv` files with Vulkan tools to make sure the memory works right. * **DCompute Runtime:** Build the host-side system to start the Vulkan device and run the code. I will post my progress, problems, and questions as replies to this thread every week. This helps keep everything in one place. I am excited to share my work with you all! GitHub: [https://github.com/gnavdev28](https://github.com/gnavdev28)
May 09
Hi everyone, here is my progress update for **Week 1** of the DCompute Vulkan Backend project. * Squashed [Nicholas's PR](https://github.com/ldc-developers/ldc/pull/4958) to thoroughly analyze the required implementation details in `targetVulkan.cpp` and `dcomputecodegenerator.cpp`. * Investigated the `TODO` in `dcomputecodegenerator.cpp` and implemented support for parsing the `vulkan-{version}` target value by adapting the existing OpenCL (OCL) and CUDA logic. * Updated `cl_options.cpp` to include Vulkan in the `--mdcompute-targets` flag's description. * Addressed the hardcoded 64-bit pointer assignment in `targetVulkan.cpp` to fix the layout issue for different pointer sizes. * Refactored relevant files by updating the LLVM version macro checks from `LLVM_LDC_VER >= 2100` to `LLVM_VERSION_MAJOR >= 23`. * **The .spv file output problem:** I used the `--mdcompute-targets` flag to make a `.spv` file, but it was in COFF format, not real SPIR-V. It was very strange. When I added a logging line just to print things out, it made correct SPIR-V code. After that, I built the newest LLVM 23 (on May 24), and it went back to COFF format. Nicholas and I didn't understand why. Finally, I fixed it. I copied the "HACK" part from `DComputeTarget::emit` to `DComputeTarget::writeModule`. Now it works normally. * **Build error:** I got a linker error when building. But it was easy to fix. I just changed the paths in the `CMakeLists.txt` file. * **The TODO part:** There is a `TODO` for `spirv.Layout` in `targetVulkan.cpp`. It is very difficult to understand. So, I haven't touched it yet. I need more time to study it. * **Look into spirv.Layout:** I will spend time studying the `TODO` about `spirv.Layout` in `targetVulkan.cpp` because I didn't finish it this week. * **Write more tests:** I will write some more test files to make sure the compiler always makes the correct `.spv` files. * **Update my PR:** I will wait for Nicholas to finish reviewing my PR. Then, I will update the code and fix any issues he points out. * Commits/PRs: https://github.com/ldc-developers/ldc/pull/5132 * My Personal Blog: Will update soon! Feel free to leave any feedback or suggestions!
May 24
Hi everyone, here is my progress update for **Week 2** of the
DCompute Vulkan Backend project.
* **Updated `target.cpp`:** I added `gABI = abi; gIR = _ir;
gTargetMachine = targetMachine;` to `writeModule` to help the
compiler write the correct file format.
* **Added missing tags:** I added ` CompileFor.deviceOnly` to
`test_vulkan.d` so it can emit the module. I also added the
` kernel()` tag to the test functions so the compiler knows the
entry point.
* **Fixed 3 big crashes in `targetVulkan.cpp`:** This was my
most important work this week. I updated the LLVM intrinsic calls:
* Deleted `i1false` from the `handlefrombinding` call to
fix the bad signature.
* Added `i32zero->getType()` and changed `i32one` to
`i32zero` to fix the `getpointer` invalid index.
* Wrote a `CreateStructGEP` loop to get each `fieldTy` and
load them one by one. This replaced the old `CreateAlignedLoad`
for the big struct and fixed the element type error.
* **Checked `isWritable`:** I used `spirv-dis` to check the
`TargetExtType`. I saw `OpMemberDecorate %_struct_20 0
NonWritable`. I am researching if the number should be `0` or `1`.
* **The COFF file problem:** At first, my compiler made a COFF
file, not a `.spv` file. Sometimes it was right, sometimes wrong.
It was very strange until I updated `writeModule`.
* **Missing Kernel and "not set" spam:** The `.spv` file did
not have `OpEntryPoint` and `_kernel`. Also, when I put a log
inside `addKernelMetadata`, my screen was full of `"->
dcomputetarget: not set"`. I found out I just missed the
` CompileFor.deviceOnly` and ` kernel()` tags.
* **The Compiler Crashes:** Right after the file emitted, the
compiler crashed 3 times in a row! First, it had a bad signature.
Second, it had an invalid index. Third, LLVM did not let me load
a big struct at the same time. I had to debug and fix all of them
in `targetVulkan.cpp`.
* **Update my PR:** I will wait for Nicholas to check my pull
request and I will update my code.
* **Find `isWritable` answer:** I will find the final answer
for the `isWritable` number (0 or 1).
* **Write more tests:** I will write more test files to make
sure everything works perfectly.
* Commits/PRs: https://github.com/ldc-developers/ldc/pull/5132
* My Personal Blog: Updating soon!
Jun 01
Quang Ha <ha.phamquang.dev gmail.com>
Hi everyone, this is my update for Week 3. This week, I did not
write a lot of new code for the compiler. Instead, I focused on
testing to make sure everything works correctly in real life!
* **Added a new test file:** I wrote another test file for the
compiler to generate the `.spv` file.
* **Cloned a test project:** I cloned a repository called
`vulkan_minimal_compute` from GitHub.
* **Tested the Vulkan API:** I changed some code in the
`vulkan_minimal_compute` project. I want to use it to run my
generated `.spv` file. This helps me check if the Vulkan API can
read and use my file correctly.
* **Understanding the real environment:** It takes time to
read and understand the `vulkan_minimal_compute` repo. Setting up
a real Vulkan environment to test the `.spv` file is a bit slow
and requires a lot of checking.
* **Finish the API test:** I will finish modifying the
`vulkan_minimal_compute` project and fully test my `.spv` file.
* **Fix bugs (if any):** If the Vulkan API finds any problems
with my `.spv` file, I will go back to the compiler and fix them.
* **Wait for PR review:** I am still waiting for Nicholas to
review my pull request so I can update my code.
* Commits/PRs: https://github.com/ldc-developers/ldc/pull/5132
* My Personal Blog: https://gnavdev28.github.io (nothing special
yet :D)
Jun 09