[Mlir-commits] [mlir] [mlir][gpu] Introduce `gpu.dynamic.shared.memory` Op (PR #71516)

[Nicolas Vasilache]

================
@@ -433,6 +433,74 @@ def GPU_GPUFuncOp : GPU_Op<"func", [
   let hasVerifier = 1;
 }
 
+def GPU_DynamicSharedMemoryOp : GPU_Op<"dynamic.shared.memory", 
+                              [MemoryEffects<[MemWrite]>] > {
+  let summary = "Get the memref for dynamic shared memory";
+  
+  let description = [{
+    This operation returns shared memory, also referred to as workgroup memory,
+    using given offsets.
+
+    It is possible to use both constants and SSA values as offsets. 
+
+    If this operation is used within a `gpu.launch`, the verifier will make an
+    attempt to verify that the offsets fall within bounds by utilizing the 
+    `dynamic_shared_memory_size` argument of `gpu.launch` when the values are
+    compile-time constants. Otherwise, the verifier does not perform offset 
+    checks.
+    
+    Examples: 
+    ```mlir
+    // Constant value, offset = 32 * 64 * sizeof(f32) * 1
+    %0 = gpu.dynamic.shared.memory [1] : memref<32x64xf32, #gpu.address_space<workgroup>>
+
+    // Multi-dimensional constant values, offset = (32 * 64 * 1 + 8) * sizeof(f32)
+    %0 = gpu.dynamic.shared.memory [1, 0, 8] : memref<32x64xf32, #gpu.address_space<workgroup>>
----------------
nicolasvasilache wrote:

I find these semantics extremely confusing, see my comment in  https://discourse.llvm.org/t/rfc-simplifying-dynamic-shared-memory-access-in-gpu/

Let's get a converged design in the RFC first.

https://github.com/llvm/llvm-project/pull/71516