[PATCH] D72696: [mlir][spirv] mlir-vulkan-runner

[Denis Khalikov via Phabricator via llvm-commits]

denis13 added a comment.

Thanks for looking on this. The VulkanRuntime part is not changed since the last time. This PR covers that was discussed when MLIR was in Tensorflow repo. The instrumentation pass is simple as possible:

1. It creates llvm global with SPIR-V binary shader.
2. It create llvm global with entry point name.
3. It creates 4 runtime calls:  setBinaryShader, setEntryPoint, setNumWorkGroups, runOnVulkan.

Since the MLIR does not define any runtime and this part is only for integration testing, I hope, at this moment, this runtime API is enough. By the way, the runtime API could be improved gradually if needed, the dispatching and command buffer creation could be introduced by instrumented runtime calls as well, the current approach hides inside all Vulkan runtime machinery like command buffer creation, pipeline creation, dispatching and so on. (Vulkan API is huge even to bootstrap computation pipeline).
Thanks!



================
Comment at: mlir/lib/Conversion/GPUToVulkan/ConvertLaunchFuncToVulkanCalls.cpp:27
+#include "mlir/IR/StandardTypes.h"
+#include "mlir/Pass/Pass.h"
+
----------------
mehdi_amini wrote:
> This pass does not seem to include anything that requires a Vulkan environment, it seems like it could be tested (and landed) independently, and not gated in CMake by the runner being enabled (which will help keeping this building in CI).
Yes, it only requires LLVM Dialect and spirv::Serializer, the tests are missing for this pass as well. I actually was thinking to introduce simple working pipeline at the first iteration and if needed I can split the pass which instrument Vulkan runtime calls to the host part and Vulkan runtime to separate PRs. Thanks!


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D72696/new/

https://reviews.llvm.org/D72696