[mlir][spirv][webgpu] Add lowering of IAddCarry to IAdd

[inbelic]

WebGPU does not currently support extended arithmetic, this is an issue when we want to lower from SPIR-V. This commit adds a pattern to transform and emulate spirv.IAddCarry with spirv.IAdd operations

Fixes #65154

[inbelic]

@kuhar @antiagainst I will ping you here for a request to review for a SPIR-V related change. No rush, I understand you are both busy, just before I forget to do so :)

[kuhar]

Thanks for taking a stab at this! I think we can simplify the code, but do let me know if I missed something there.

[kuhar]

Nice, thanks for the changes.

Could you also add a runtime integration test under https://github.com/llvm/llvm-project/tree/main/mlir/test/mlir-vulkan-runner to make sure we are getting the expected results?

[inbelic]

Rebased and force pushed to hopefully remedy the code_formatter error.
I added an integration test, but it is only done by eye. I am not sure how one can test it locally? Is there documentation for the possible ways to test with the mlir-vulkan-runner?

[kuhar]

Rebased and force pushed to hopefully remedy the code_formatter error. I added an integration test, but it is only done by eye. I am not sure how one can test it locally? Is there documentation for the possible ways to test with the mlir-vulkan-runner?

You need to set these two in cmake:

      -DMLIR_INCLUDE_INTEGRATION_TESTS=1 \
      -DMLIR_ENABLE_VULKAN_RUNNER=1 \

[kuhar]

The runner test fails with:

******************** TEST 'MLIR :: mlir-vulkan-runner/iaddcarry_extended.mlir' FAILED ********************
Exit Code: 2

Command Output (stdout):
--
# RUN: at line 4
mlir-vulkan-runner /home/jakub/llvm/llvm-project/mlir/test/mlir-vulkan-runner/iaddcarry_extended.mlir   --shared-libs=/home/jakub/llvm/build/relass/lib/libvulkan-runtime-wrappers.so,/home/jakub/llvm/build/relass/lib/libmlir_runner_utils.so   --entry-point-result=void | /home/jakub/llvm/build/relass/bin/FileCheck /home/jakub/llvm/llvm-project/mlir/test/mlir-vulkan-runner/iaddcarry_extended.mlir
# executed command: mlir-vulkan-runner /home/jakub/llvm/llvm-project/mlir/test/mlir-vulkan-runner/iaddcarry_extended.mlir --shared-libs=/home/jakub/llvm/build/relass/lib/libvulkan-runtime-wrappers.so,/home/jakub/llvm/build/relass/lib/libmlir_runner_utils.so --entry-point-result=void
# .---command stderr------------
# | loc("/home/jakub/llvm/llvm-project/mlir/test/mlir-vulkan-runner/iaddcarry_extended.mlir":25:59): error: expected ','
# | could not parse the input IR
# `-----------------------------
# error: command failed with exit status: 1
# executed command: /home/jakub/llvm/build/relass/bin/FileCheck /home/jakub/llvm/llvm-project/mlir/test/mlir-vulkan-runner/iaddcarry_extended.mlir
# .---command stderr------------
# | FileCheck error: '<stdin>' is empty.
# | FileCheck command line:  /home/jakub/llvm/build/relass/bin/FileCheck /home/jakub/llvm/llvm-project/mlir/test/mlir-vulkan-runner/iaddcarry_extended.mlir
# `-----------------------------
# error: command failed with exit status: 2

[kuhar]

@inbelic I made the test type check and compile but it doesn't give the expected output: kuhar@d20f576

I changed the i8 memrefs to i32 because the type of both iadd_carry results will be the same on the spir-v side anyway, and it doesn't require spir-v capabilities for i8 types. This extends out test coverage to vulkan implementations that don't expose these, e.g., swiftshader.

[inbelic]

@kuhar Awesome thanks. Have got the runner test locally and also get the wrong behaviour when using the negative minimum integers for the input. I believe this is because the Signedness is set for the integer type. While arith.addui_extended and spirv.IAddCarry will interpret the bits as unsigned, spirv.IAddOp does not and that is why it does not provide our desired behaviour to overflow when we add 1 to it. All of this to say, that if you set -2^31-1 to 2^32-1 instead, then we have the expected behaviour.

However, imo it should be that the negative number is interpreted as an unsigned integer and then we unfortunately will need to use a different solution.

[kuhar]

@inbelic Signed and unsigned integer addition is the same operation, that's why there's a single addi instead of addui/addsi. When we embed constants in the IR and print it back the interpretation is signed and we see negative numbers, but it doesn't matter -- it's the same bitpattern as for the unsigned result.

Therefore, addui_extended/IAddCarry both implement signless addition, and the carry bit corresponds to the unsigned overflow bit.

[inbelic]

Resolved my confusion. Updated the test cases with your suggested changes and updated the expected values and inputs.

[kuhar]

LGTM. I ran the integration tests on my machine and everything passed. Thanks!

[kuhar]

@inbelic I went ahead and merged the code