[Mlir-commits] [mlir] [mlir][amdgpu] Lower amdgpu.make_dma_base (PR #169817)

[Erick Ochoa Lopez]

================
@@ -2264,6 +2264,76 @@ struct AMDGPUPermlaneLowering : public ConvertOpToLLVMPattern<PermlaneSwapOp> {
   }
 };
 
+struct AMDGPUMakeDmaBaseLowering
+    : public ConvertOpToLLVMPattern<MakeDmaBaseOp> {
+  using ConvertOpToLLVMPattern::ConvertOpToLLVMPattern;
+
+  AMDGPUMakeDmaBaseLowering(const LLVMTypeConverter &converter, Chipset chipset)
+      : ConvertOpToLLVMPattern<MakeDmaBaseOp>(converter), chipset(chipset) {}
+  Chipset chipset;
+
+  LogicalResult
+  matchAndRewrite(MakeDmaBaseOp op, OpAdaptor adaptor,
+                  ConversionPatternRewriter &rewriter) const override {
+    if (chipset < kGfx1250)
+      return op->emitOpError("make_dma_base is only supported on gfx1250");
+
+    Location loc = op.getLoc();
+
+    ValueRange srcIndices = adaptor.getSrcIndices();
+    Value src = adaptor.getSrc();
+    auto srcMemRefType = cast<MemRefType>(op.getSrc().getType());
+
+    Value srcPtr =
+        getStridedElementPtr(rewriter, loc, srcMemRefType, src, srcIndices);
+
+    ValueRange dstIndices = adaptor.getDstIndices();
+    Value dst = adaptor.getDst();
+    auto dstMemRefType = cast<MemRefType>(op.getDst().getType());
+
+    Value dstPtr =
+        getStridedElementPtr(rewriter, loc, dstMemRefType, dst, dstIndices);
+
+    bool storeFrom = hasWorkgroupMemorySpace(srcMemRefType.getMemorySpace());
+    Value ldsAddr = storeFrom ? srcPtr : dstPtr;
+    Value globalAddr = storeFrom ? dstPtr : srcPtr;
+
+    Type i32 = rewriter.getI32Type();
+    Type i64 = rewriter.getI64Type();
+
+    Value castForLdsAddr =
+        LLVM::PtrToIntOp::create(rewriter, loc, i32, ldsAddr);
+    Value castForGlobalAddr =
+        LLVM::PtrToIntOp::create(rewriter, loc, i64, globalAddr);
+
+    Value mask = createI64Constant(rewriter, loc, 0x1FFFFFFFFFFFFFF);
+    Value first57BitsOfGlobalAddr =
+        LLVM::AndOp::create(rewriter, loc, castForGlobalAddr, mask);
+    Value shift = LLVM::LShrOp::create(rewriter, loc, first57BitsOfGlobalAddr,
+                                       createI64Constant(rewriter, loc, 32));
+
+    Value lowHalf =
+        LLVM::TruncOp::create(rewriter, loc, i32, first57BitsOfGlobalAddr);
+    Value highHalf = LLVM::TruncOp::create(rewriter, loc, i32, shift);
+
+    Value c0 = createI32Constant(rewriter, loc, 0);
+    Value c1 = createI32Constant(rewriter, loc, 1);
+    Value c2 = createI32Constant(rewriter, loc, 2);
+    Value c3 = createI32Constant(rewriter, loc, 3);
+
+    Type v4i32 = this->typeConverter->convertType(VectorType::get(4, i32));
+    Value result = LLVM::UndefOp::create(rewriter, loc, v4i32);
----------------
amd-eochoalo wrote:

1. https://github.com/llvm/llvm-project/pull/169817/commits/8a0202d7f1121da3824008af13e2b41954565b55
2. https://github.com/llvm/llvm-project/pull/169817/commits/af068e0812ff63eafa8c6876a1e929efae3f729f

Thanks for reviewing this!

https://github.com/llvm/llvm-project/pull/169817