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

[Jakub Kuderski]

================
@@ -1010,6 +1010,53 @@ struct WMMAOpLowering : public ConvertOpToLLVMPattern<WMMAOp> {
   }
 };
 
+struct GatherToLDSOpLowering : public ConvertOpToLLVMPattern<GatherToLDSOp> {
+  GatherToLDSOpLowering(const LLVMTypeConverter &converter, Chipset chipset)
+      : ConvertOpToLLVMPattern<GatherToLDSOp>(converter), chipset(chipset) {}
+
+  Chipset chipset;
+
+  LogicalResult
+  matchAndRewrite(GatherToLDSOp op, GatherToLDSOpAdaptor adaptor,
+                  ConversionPatternRewriter &rewriter) const override {
+    if (chipset < kGfx942)
+      return op.emitOpError("chipset not supported");
+
+    Location loc = op.getLoc();
+
+    auto srcMemRefType = cast<MemRefType>(op.getSrc().getType());
+    auto dstMemRefType = cast<MemRefType>(op.getSrc().getType());
+
+    // 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.
+    size_t loadWidth;
----------------
kuhar wrote:

To avoid future issues with uninitialized variables, I'd assign this `transferType.getIntOrFloatBitWidth() / 8;` and drop the else statement

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