[Mlir-commits] [mlir] [mlir][gpu][NVPTX] Enable NVIDIA GPU JIT compilation path (PR #66220)

[Aart Bik]

================
@@ -144,6 +144,22 @@ struct SparseCompilerOptions
                                            desc("GPU target architecture")};
   PassOptions::Option<std::string> gpuFeatures{*this, "gpu-features",
                                                desc("GPU target features")};
+  /// For NVIDIA GPUs there are 3 compilation format options:
+  /// 1. `isa`: the compiler generates PTX and the driver JITs the PTX.
+  /// 2. `bin`: generates a CUBIN object for `chip=gpuChip`.
+  /// 3. `fatbin`: generates a fat binary with a CUBIN object for `gpuChip` and
+  /// also embeds the PTX in the fat binary.
+  /// Notes:
+  /// Option 1 adds a significant runtime performance hit, however, tests are
+  /// more likely to pass with this option.
+  /// Option 2 is better for execution time as there is no JIT; however, the
+  /// program will fail if there's an arch mismatch between `gpuChip` and the
----------------
aartbik wrote:

nit: can you please spell out "architecture" (unless this is NVidia convention to write it that way)

https://github.com/llvm/llvm-project/pull/66220