Addressed comments.

Ping for review.

LGTM!

In D159347#4636957, @guraypp wrote:

Adding module=main_kernel is not necessary in --nvvm-attach-target=, that's just to filter to which modules to add the target.

I actually need it to set ptx version. The default version for sm_90 is 7.8, that does not support PTX instructions for TMA. So I set to ptx80+

In D159347#4636557, @guraypp wrote:

@fmorac I use the gpu-module-to-binary pass, you recently introduced, for mlir->llvm->ptx->cubin, eventually link the hosts llvm (has embedded cubin) by clang to generate the executable. Is this the right way to use your Pass?

I used to run gpu mlir integration tests with mlir-cpu-runner, but I guess gpu-module-to-binary is not compatible with it.

Ok, the following patch also does the trick:

To avoid increasing binary size the following patch should be enough to fix the issue:

diff --git a/mlir/tools/mlir-opt/mlir-opt.cpp b/mlir/tools/mlir-opt/mlir-opt.cpp
index b47634476c62..5cd1b40ea6c9 100644
--- a/mlir/tools/mlir-opt/mlir-opt.cpp
+++ b/mlir/tools/mlir-opt/mlir-opt.cpp
@@ -276,7 +276,14 @@ int main(int argc, char **argv) {
   DialectRegistry registry;
   registerAllDialects(registry);
   registerAllExtensions(registry);
-  registerAllToLLVMIRTranslations(registry);
+
+  // TODO: Remove these when a more structured extension registration mechanism
+  // is added.
+  registerBuiltinDialectTranslation(registry);
+  registerGPUDialectTranslation(registry);
+  registerLLVMDialectTranslation(registry);
+  registerNVVMDialectTranslation(registry);
+  registerROCDLDialectTranslation(registry);

Addressed comments. If this patch D158464 gets through, this diff will be repurposed to use the changes in the aforementioned diff.

Addressed comments.

Addressed comments.

In D158277#4600746, @mehdi_amini wrote:

I'm not sure what does the trait provides right now? If you just remove the verifier for ObjectAttr and remove the trait, don't you get basically the same properties?

This patch also solves the need to add registration calls for Target Attributes in Translation registration, see D157703.

Updated the diff to resolve all the conflicts and additions created by pushing all the commits of the new GPU pipeline.

Switched to a single call of std::move

Fix patch application.

Separated gpu-module-to-binary test into 2 tests: nvvm & rocdl.
Also updated lit.cfg.py with:

Addressed the comments.

Moved passes to GPU.

Overall LGTM. My only concern is down-stream users having issues and don't knowing what's happening, because when those registration calls are removed from passes like gpu-to-cubin, they will fail.

Addressed the comments.

Removed private lib.

Rebasing.

Added an assert check for the attribute.
Note: The verifier will also verify that the dyn_cast is valid, as it's a precondition on the elements of the array.

Remove the promised interface, instead use OffloadingTranslationAttrTrait in SelectObjectAttr
and register the interface automatically with the GPUTranslationInterface.

Removing Ods prefix.

Update the gpu.binary attributes to use OffloadingTranslationAttrTrait
instead of OffloadingLLVMTranslationAttrInterface.

Add the OffloadingTranslationAttrTrait trait. This traits decouples gpu.binary op
from LLVM translation, making gpu.binary usable by other translation mechanisms like SPIRV.

Rebasing.

Overall looks good to me, can you add a description & tests for MaxSIOp & MinUIOp?

Ping for review. Solved merge conflicts.

Updated the docs with the comments.

Fixed merge conflicts with D157183.

Changed the description of the CMake variable MLIR_ENABLE_NVPTXCOMPILER. Applied clang-format again.

Adds a couple of syntax tests & fixes a bug not displaying the link option in the attribute.