Add support for bare pointer calling convention in gpu-to-llvm

[bondhugula]

Summary

Add support for the bare pointer calling convention in the gpu-to-llvm
pass. This wasn't being exposed and is needed when GPU-compiled MLIR is
to be called with this convention.

[mehdi_amini]

Can you add a test for this?

mlir/lib/Conversion/GPUCommon/GPUToLLVMConversion.cpp

720

options.useBarePtrCallConv = hostBarePtrCallConv; ?

[bondhugula]

Address review comment.

[bondhugula]

In D152477#4407366, @mehdi_amini wrote:

Can you add a test for this?

Do we really need a test for this? bare pointer lowering for memrefs has its own tests. This pass is just a user, just like gpu-to-nvvm and gpu-to-rocdl, and it's expected to work. I can add a test though if something new is exercised or we want to anyway to test to ensure the functionality is there.

mlir/lib/Conversion/GPUCommon/GPUToLLVMConversion.cpp

720

Thanks, done.

[mehdi_amini]

I wasn’t sure if the lowering is expected to after the signature of a gpu.func for example?

[krzysz00]

Seems reasonable to me, surprised that this didn't break before (did something break?), approved.

[bondhugula]

In D152477#4410690, @mehdi_amini wrote:

I wasn’t sure if the lowering is expected to after the signature of a gpu.func for example?

The gpu-to-llvm pass is applied to lower host-side IR: typically, builtin.module(gpu-to-llvm). gpu.func aren't even processed by it. The other revision I have is the one that adds that option for gpu-to-nvvm which is the one that lowers the signature for gpu.func. We need both of these changes.
https://reviews.llvm.org/D152480

[bondhugula]

In D152477#4414275, @krzysz00 wrote:

Seems reasonable to me, surprised that this didn't break before (did something break?), approved.

gpu-to-nvvm never provided that option - the other revision I have adds it to gpu-to-nvvm (https://reviews.llvm.org/D152480). If we used the bare ptr lowering here and not in gpu-to-nvvm or vice versa, things would break. That said, I can update an integration test case there to use the bare ptr lowering that will exercise both of these changes.

mlir/include/mlir/Conversion/Passes.td

let description = [{
     calls.
 
     This pass does not generate code to call GPU runtime APIs directly but
     instead uses a small wrapper library that exports a stable and conveniently
     typed ABI on top of GPU runtimes such as CUDA or ROCm (HIP).
   }];
 
   let options = [
+    Option<"hostBarePtrCallConv", "use-bare-pointers-for-host", "bool",
+           /*default=*/"false",
+           "Use bare pointers to pass memref arguments to host functions. "
+           "All memrefs must have static shape.">,
     Option<"kernelBarePtrCallConv", "use-bare-pointers-for-kernels", "bool",
            /*default=*/"false",
              "Use bare pointers to pass memref arguments to kernels. "
              "The kernel must use the same setting for this option."
            >,
     Option<"gpuBinaryAnnotation", "gpu-binary-annotation", "std::string",
                /*default=*/"gpu::getDefaultGpuBinaryAnnotation()",
                "Annotation attribute string for GPU binary"

mlir/lib/Conversion/GPUCommon/GPUToLLVMConversion.cpp

matchAndRewrite(gpu::SDDMMOp op, OpAdaptor adaptor,
                   ConversionPatternRewriter &rewriter) const override;
 };
 
 } // namespace
 
 void GpuToLLVMConversionPass::runOnOperation() {
   LowerToLLVMOptions options(&getContext());
   options.useOpaquePointers = useOpaquePointers;
+  options.useBarePtrCallConv = hostBarePtrCallConv;
 
   LLVMTypeConverter converter(&getContext(), options);
   RewritePatternSet patterns(&getContext());
   LLVMConversionTarget target(getContext());
 
   target.addIllegalDialect<gpu::GPUDialect>();
 
   mlir::arith::populateArithToLLVMConversionPatterns(converter, patterns);
   mlir::cf::populateControlFlowToLLVMConversionPatterns(converter, patterns);