[Mlir-commits] [mlir] [MLIR][AMDGPU] Add a wrapper for global LDS load intrinsics in AMDGPU (PR #133498)

[Krzysztof Drewniak]

================
@@ -903,6 +903,81 @@ struct WMMAOpLowering : public ConvertOpToLLVMPattern<WMMAOp> {
   }
 };
 
+struct GlobalLoadLDSOpLowering
+    : public ConvertOpToLLVMPattern<GlobalLoadLDSOp> {
+  GlobalLoadLDSOpLowering(const LLVMTypeConverter &converter, Chipset chipset)
+      : ConvertOpToLLVMPattern<GlobalLoadLDSOp>(converter), chipset(chipset) {}
+
+  Chipset chipset;
+
+  LogicalResult
+  matchAndRewrite(GlobalLoadLDSOp op, GlobalLoadLDSOpAdaptor adaptor,
+                  ConversionPatternRewriter &rewriter) const override {
+    Location loc = op.getLoc();
+
+    auto elemType = cast<MemRefType>(op.getDst().getType()).getElementType();
+    size_t elemSizeInBits = elemType.getIntOrFloatBitWidth();
+    if (elemSizeInBits % 8 != 0)
+      return op.emitOpError("element size must be a multiple of 8");
+
+    // TODO: instead of only transfering one element per thread, we could
+    // augment it to transfer multiple elements per thread by issuing multiple
+    // `global_load_lds` instructions.
+    auto loadWidth = elemSizeInBits / 8;
+
+    const Chipset GlobalLoadEnabled{9, 0x4, 0x0};
----------------
krzysz00 wrote:

`kGfx942` above. Especially since gfx940 doesn't exist.

https://github.com/llvm/llvm-project/pull/133498