squash commits into single commit

compiler/plugins/target/AMD-AIE/iree-amd-aie/Transforms/AMDAIECanonicalizeDma.cpp

@@ -13,7 +13,7 @@
 #include "mlir/Pass/Pass.h"
 #include "mlir/Transforms/GreedyPatternRewriteDriver.h"
 
-#define DEBUG_TYPE "iree-amdaie-pack-to-dma"
+#define DEBUG_TYPE "iree-amdaie-canonicalize-dma"
 
 namespace mlir::iree_compiler::AMDAIE {
 

compiler/plugins/target/AMD-AIE/iree-amd-aie/Transforms/AMDAIECanonicalizeDoublyStridedOp.cpp

@@ -5,7 +5,6 @@
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 
 #include "iree-amd-aie/IR/AMDAIEDialect.h"
-#include "iree-amd-aie/IR/AMDAIEOps.h"
 #include "iree-amd-aie/Transforms/AMDAIEDmaUtils.h"
 #include "iree-amd-aie/Transforms/Passes.h"
 #include "mlir/Pass/Pass.h"

compiler/plugins/target/AMD-AIE/iree-amd-aie/Transforms/AMDAIECanonicalizeNpuDmaCpyNd.cpp

@@ -0,0 +1,184 @@
+// Copyright 2024 The IREE Authors
+//
+// Licensed under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+
+#include <memory>
+
+#include "iree-amd-aie/IR/AMDAIEDialect.h"
+#include "iree-amd-aie/IR/AMDAIEOps.h"
+#include "iree-amd-aie/Transforms/Passes.h"
+#include "mlir/Dialect/Utils/StaticValueUtils.h"
+#include "mlir/Pass/Pass.h"
+#include "mlir/Pass/PassManager.h"
+
+#define DEBUG_TYPE "iree-amdaie-canonicalize-npu-dma-cpy-nd"
+
+namespace mlir::iree_compiler::AMDAIE {
+
+class AMDAIECanonicalizeNpuDmaCpyNdPass
+    : public impl::AMDAIECanonicalizeNpuDmaCpyNdBase<
+          AMDAIECanonicalizeNpuDmaCpyNdPass> {
+ public:
+  void getDependentDialects(DialectRegistry &registry) const override {
+    registry.insert<AMDAIEDialect>();
+  }
+
+  void runOnOperation() override {
+    MLIRContext *context = &getContext();
+    ModuleOp moduleOp = getOperation();
+    IRRewriter rewriter(context);
+    Attribute zero = rewriter.getIndexAttr(0);
+    Attribute one = rewriter.getIndexAttr(1);
+
+    WalkResult walkResult = moduleOp->walk([&](NpuDmaCpyNdOp dmaOp) {
+      SmallVector<OpFoldResult> srcOffsets = dmaOp.getSourceMixedOffsets();
+      SmallVector<OpFoldResult> srcSizes = dmaOp.getSourceMixedSizes();
+      SmallVector<OpFoldResult> srcStrides = dmaOp.getSourceMixedStrides();
+
+      SmallVector<OpFoldResult> tgtOffsets = dmaOp.getTargetMixedOffsets();
+      SmallVector<OpFoldResult> tgtSizes = dmaOp.getTargetMixedSizes();
+      SmallVector<OpFoldResult> tgtStrides = dmaOp.getTargetMixedStrides();
+
+      // The first step in canonicalization is padding the offsets/sizes/strides
+      // vectors to be of rank `nbDimensions`. If the rank of any of these
+      // vectors is greater than `nbDimensions`, then this is impossible.
+      bool allValidRanks = srcOffsets.size() <= nbDimensions &&
+                           srcSizes.size() <= nbDimensions &&
+                           srcStrides.size() <= nbDimensions &&
+                           tgtOffsets.size() <= nbDimensions &&
+                           tgtSizes.size() <= nbDimensions &&
+                           tgtStrides.size() <= nbDimensions;
+      if (!allValidRanks) {
+        dmaOp.emitOpError()
+            << " has offsets/sizes/strides attributes that are "
+               "larger than the target canonicalization dimension of "
+            << nbDimensions << ".";
+        return WalkResult::interrupt();
+      }
+
+      // If the source is in L3, then canonicalize the source addressing.
+      // 1) Pad to the correct rank
+      // 2) Move the zero stride (if any) to the outer-most (slowest) dim.
+      if (dmaOp.getSourceMemorySpaceAsUInt() == 0) {
+        if (!dmaOp.hasSourceAddressing()) {
+          dmaOp.emitOpError()
+              << "has source in L3, but does not have source addressing. "
+                 "Source addressing is required to canonicalize here.";
+          return WalkResult::interrupt();
+        }
+        srcOffsets = getPrepended(srcOffsets, zero);
+        srcSizes = getPrepended(srcSizes, one);
+        srcStrides = getPrepended(srcStrides, zero);
+        std::optional<uint32_t> maybeSwapIndex =
+            verifyAndGetZeroStrideIndex(srcSizes, srcStrides, dmaOp);
+        if (!maybeSwapIndex.has_value()) return WalkResult::interrupt();
+        uint32_t swapIndex = maybeSwapIndex.value();
+        bubble(srcOffsets, swapIndex);
+        bubble(srcSizes, swapIndex);
+        bubble(srcStrides, swapIndex);
+      }
+
+      if (dmaOp.getTargetMemorySpaceAsUInt() == 0) {
+        if (!dmaOp.hasTargetAddressing()) {
+          dmaOp.emitOpError()
+              << "has target in L3, but does not have target addressing. "
+                 "Target addressing is required to canonicalize here.";
+          return WalkResult::interrupt();
+        }
+        tgtOffsets = getPrepended(tgtOffsets, zero);
+        tgtSizes = getPrepended(tgtSizes, one);
+        tgtStrides = getPrepended(tgtStrides, zero);
+        std::optional<uint32_t> maybeSwapIndex =
+            verifyAndGetZeroStrideIndex(tgtSizes, tgtStrides, dmaOp);
+        if (!maybeSwapIndex.has_value()) return WalkResult::interrupt();
+        uint32_t swapIndex = maybeSwapIndex.value();
+        bubble(tgtOffsets, swapIndex);
+        bubble(tgtSizes, swapIndex);
+        bubble(tgtStrides, swapIndex);
+      }
+
+      rewriter.setInsertionPoint(dmaOp);
+
+      // Replace the npu.dma_cpy_nd with the canonicalized version.
+      dmaOp = rewriter.replaceOpWithNewOp<AMDAIE::NpuDmaCpyNdOp>(
+          dmaOp, dmaOp.getDma(), dmaOp.getTarget(), tgtOffsets, tgtSizes,
+          tgtStrides, dmaOp.getTargetBdId(), dmaOp.getSource(), srcOffsets,
+          srcSizes, srcStrides, dmaOp.getSourceBdId());
+
+      return WalkResult::advance();
+    });
+
+    if (walkResult.wasInterrupted()) return signalPassFailure();
+  }
+
+ private:
+  /// Repeat prepend 'def' to 'tail' to make 'tail' have nbDimensions elements.
+  SmallVector<OpFoldResult> getPrepended(ArrayRef<OpFoldResult> tail,
+                                         Attribute def) {
+    assert(tail.size() <= nbDimensions);
+    SmallVector<OpFoldResult> res(nbDimensions, def);
+    std::copy(tail.begin(), tail.end(),
+              res.begin() + nbDimensions - tail.size());
+    return res;
+  }
+
+  static size_t getLowestIndexMaybeAboveOne(ArrayRef<OpFoldResult> v) {
+    for (size_t i = 0; i < v.size(); i++) {
+      std::optional<int64_t> maybe = getConstantIntValue(v[i]);
+      if (!maybe.has_value() || maybe.value() > 1) return i;
+    }
+    return v.size();
+  }
+
+  static size_t getHighestIndexMaybeZero(ArrayRef<OpFoldResult> v) {
+    for (size_t i = v.size(); i > 0; i--) {
+      std::optional<int64_t> maybe = getConstantIntValue(v[i - 1]);
+      if (!maybe.has_value() || maybe.value() == 0) return i - 1;
+    }
+    return 0;
+  }
+
+  /// Get the highest index where the stride is 0. If this index is greater
+  /// than the lowest index where the size is greater than 1, then fail.
+  std::optional<uint32_t> verifyAndGetZeroStrideIndex(
+      ArrayRef<OpFoldResult> sizes, ArrayRef<OpFoldResult> strides,
+      NpuDmaCpyNdOp dmaOp) {
+    assert(strides.size() == sizes.size() && strides.size() == nbDimensions);
+
+    size_t firstNonUnitDim = getLowestIndexMaybeAboveOne(sizes);
+    size_t lastZeroStrideDim = getHighestIndexMaybeZero(strides);
+
+    if (firstNonUnitDim < lastZeroStrideDim) {
+      // HW limitation.
+      dmaOp.emitOpError("might have stride=0 in dimension ")
+          << lastZeroStrideDim << ", and size>1 in dimension "
+          << firstNonUnitDim << ". As " << firstNonUnitDim << " < "
+          << lastZeroStrideDim
+          << ", this cannot be supported -- the zero stride cannot be moved "
+             "to the outer-most (slowest) dimension, as required by current "
+             "AIE architecture.";
+      return {};
+    }
+    return lastZeroStrideDim;
+  }
+
+  /// Example, for swapIndex = 2.
+  /// Input
+  ///                 [0 1 7 13]
+  /// is mutated to
+  ///                 [7 0 1 13]
+  static void bubble(MutableArrayRef<OpFoldResult> arr, size_t swapIndex) {
+    if (swapIndex > 0) {
+      std::rotate(arr.begin(), arr.begin() + swapIndex,
+                  arr.begin() + swapIndex + 1);
+    }
+  }
+};
+
+std::unique_ptr<Pass> createAMDAIECanonicalizeNpuDmaCpyNdPass() {
+  return std::make_unique<AMDAIECanonicalizeNpuDmaCpyNdPass>();
+}
+
+}  // namespace mlir::iree_compiler::AMDAIE