Updated

This patch is related to https://github.com/tensorflow/mlir/issues/196, and the pass is based on https://github.com/google/clspv/blob/master/lib/RewriteInsertsPass.cpp

LGTM! Thanks for simplifying this!

Addressed review comments.

LGTM! Thanks!

Thanks! It fixes a build problem!

@antiagainst thanks for review, can you please take a look, to make sure I've updated the patch in right way. Thanks.

Addresses comments.

Rebase on master.

Remove new line.

Addresses comments.

Was fixed at https://reviews.llvm.org/D78934

Rebase on master.

Addresses review comments.

@antiagainst thanks for review, updated!

@mravishankar thanks for review!

@antiagainst thanks for review!I'll fix it.

LGTM! Looks more pretty than it was, thanks!

This patch is related to an issue https://bugs.llvm.org/show_bug.cgi?id=45392

LGTM!

@antiagainst thanks a lot for review! Sure, I would like to help!

LGTM! Thanks!
I'm currently have access only to "mobile" gpu "GeForce GTX 1050 Mobile":
Compute shader execution time: 15.360us
Command buffer submit time: 7us
Wait idle time: 155us

@mehdi_amini thanks, recently obtained it.

Address review comments.

Address review comments.

Fix LINT issues.

This patch eliminates VulkanLaunchOpOperandAdaptor and enables emitCWrappers while lowering to LLVM dialect. By doing this we follow the official guideline for calling convention https://mlir.llvm.org/docs/ConversionToLLVMDialect/#calling-convention.
emitCWrappers enables the following steps:

@rriddle thanks for review, will address it in follow up pr.

In D75192#1915612, @jpienaar wrote:

This seems to have broken build on gcc 5:
error: converting to 'std::tuple<mlir::Value, mlir::Value>' from initializer list would use explicit constructor 'constexpr std::tuple<_T1, _T2>::tuple(const _T1&, const _T2&) [with _T1 = mlir::Value; _T2 = mlir::Value]'

operands[numConfigOps + index * loweredMemRefNumOps1D + 3]};
                                                          ^

LGTM! Thanks!

Address comments, sorry for delay.

@antiagainst thanks a lot for detailed review! I'll update soon according to your suggestions, thanks!

Rebase on master.

Disable clang-diagnostic for include vulkan header file.

To eliminate static construction and desctruction of VulkanRuntimeManager, we need to instrument runtime calls with first argument as a generic pointer to VulkanRuntimeManager, but the setResource function based on std dialect, which fills the given memref with the given value, could not accept generic pointers, because std dialect does not have generic pointer types. So, to solve it, I added new pass which serializes spirv::Module to binary form appends it as a string attribute to vulkan launch call op. Erases all gpu.module and spirv::Module, so we can lower it to LLVM directly, then we instrument host part with runtime calls as before, but adding initVulkan, deinitVulkan calls and for each lowered memref we create call to bindResource according the current GPUToSPIRV convention (Each resource binds to 0 descriptor set and descriptor binding argument order lined up in argument order.

LGTM!

Since no users anymore for this patch, closing it :)

Thanks, looking forward for the feedback from Lei.

Rebased on master to trigger a rebuild.

Addresses review comments.

Addresses review comments.

@rriddle thanks, I'll address your comments D72696

Addressed review comments.

Rebased on master.

Fix typo in comment.

Thanks a lot for review! I addressed comments related to pass in separate revision https://reviews.llvm.org/D74549 and will update mlir-vulkan-runner, VulkanRuntime as well.

Added simple test to check GpuLaunchFuncToVulkanCalssPass.
Rebased on D74012.
Rebased on 5a1778 which changes MemRef calling convention.