[Mlir-commits] [mlir] [mlir][vector] Add support for linearizing Extract, ExtractStridedSlice, Shuffle VectorOps in VectorLinearize (PR #88204)

[Diego Caballero]

================
@@ -103,6 +108,252 @@ struct LinearizeVectorizable final
     return success();
   }
 
+private:
+  unsigned targetVectorBitWidth;
+};
+
+/// This pattern converts the ExtractStridedSliceOp into a ShuffleOp that works
+/// on a linearized vector.
+/// Following,
+///   vector.extract_strided_slice %source
+///         { offsets = [..], strides = [..], sizes = [..] }
+/// is converted to :
+///   %source_1d = vector.shape_cast %source
+///   %out_1d = vector.shuffle %source_1d, %source_1d [ shuffle_indices_1d ]
+///   %out_nd = vector.shape_cast %out_1d
+/// `shuffle_indices_1d` is computed using the offsets and sizes of the
+/// extraction.
+struct LinearizeVectorExtractStridedSlice final
+    : public mlir::OpConversionPattern<mlir::vector::ExtractStridedSliceOp> {
+  using OpConversionPattern::OpConversionPattern;
+  LinearizeVectorExtractStridedSlice(
+      const TypeConverter &typeConverter, MLIRContext *context,
+      unsigned targetVectBitWidth = std::numeric_limits<unsigned>::max(),
+      PatternBenefit benefit = 1)
+      : OpConversionPattern(typeConverter, context, benefit),
+        targetVectorBitWidth(targetVectBitWidth) {}
+
+  LogicalResult
+  matchAndRewrite(vector::ExtractStridedSliceOp extractOp, OpAdaptor adaptor,
+                  ConversionPatternRewriter &rewriter) const override {
+    auto dstType = getTypeConverter()->convertType(extractOp.getType());
+    assert(!(extractOp.getVector().getType().isScalable() ||
+             dstType.cast<VectorType>().isScalable()) &&
+           "scalable vectors are not supported.");
+    if (!isLessThanTargetBitWidth(extractOp, targetVectorBitWidth))
+      return rewriter.notifyMatchFailure(
+          extractOp, "Can't flatten since targetBitWidth <= OpSize");
+
+    auto offsets = extractOp.getOffsets();
+    auto sizes = extractOp.getSizes();
+    auto strides = extractOp.getStrides();
+    if (!isConstantIntValue(strides[0], 1))
+      return rewriter.notifyMatchFailure(
+          extractOp, "Strided slice with stride != 1 is not supported.");
+    Value srcVector = adaptor.getVector();
+    // If kD offsets are specified for nd source vector (n > k), the granularity
+    // of the extraction is greater than 1. In this case last (n-k) dimensions
+    // form the extraction granularity.
+    // Example :
+    //  vector.extract_strided_slice %src {
+    //      offsets = [0, 0], sizes = [2, 2], strides = [1, 1]} :
+    //      vector<4x8x8xf32> to vector<2x2x8xf32>
+    // Here, extraction granularity is 8.
+    int64_t extractGranularitySize = 1;
+    auto n = extractOp.getSourceVectorType().getRank();
+    int64_t k = (int64_t)offsets.size();
+    if (n > k) {
+      for (unsigned i = 0; i < n - k; i++) {
+        extractGranularitySize *=
+            extractOp.getSourceVectorType().getShape()[i + k];
+      }
+    }
+    // Get total number of extracted slices.
+    int64_t nExtractedSlices = 1;
+    llvm::for_each(sizes, [&](Attribute size) {
+      nExtractedSlices *= size.cast<IntegerAttr>().getInt();
+    });
+    // Compute the strides of the source vector considering first k dimensions.
+    llvm::SmallVector<int64_t, 4> sourceStrides(k, extractGranularitySize);
+    for (int i = k - 2; i >= 0; --i) {
+      sourceStrides[i] = sourceStrides[i + 1] *
+                         extractOp.getSourceVectorType().getShape()[i + 1];
+    }
+    // Final shuffle indices has nExtractedSlices * extractGranularitySize
+    // elements.
+    llvm::SmallVector<int64_t, 4> indices(nExtractedSlices *
+                                          extractGranularitySize);
+    // Compute the strides of the extracted kD vector.
+    llvm::SmallVector<int64_t, 4> extractedStrides(k, 1);
+    // Compute extractedStrides.
+    for (int i = k - 2; i >= 0; --i) {
+      extractedStrides[i] =
+          extractedStrides[i + 1] * sizes[i + 1].cast<IntegerAttr>().getInt();
+    }
+    // Iterate over all extracted slices from 0 to nExtractedSlices - 1
+    // and compute the multi-dimensional index and the corresponding linearized
+    // index within the source vector.
+    for (int64_t i = 0; i < nExtractedSlices; ++i) {
+      int64_t index = i;
+      // Compute the corresponding multi-dimensional index.
+      llvm::SmallVector<int64_t, 4> multiDimIndex(k, 0);
+      for (int64_t j = 0; j < k; ++j) {
+        multiDimIndex[j] = (index / extractedStrides[j]);
+        index -= multiDimIndex[j] * extractedStrides[j];
+      }
+      // Compute the corresponding linearized index in the source vector
+      // i.e. shift the multiDimIndex by the offsets.
+      int64_t linearizedIndex = 0;
+      for (int64_t j = 0; j < k; ++j) {
+        linearizedIndex +=
+            (offsets[j].cast<IntegerAttr>().getInt() + multiDimIndex[j]) *
+            sourceStrides[j];
+      }
+      // Fill the indices array form linearizedIndex to linearizedIndex +
+      // extractGranularitySize.
+      for (int64_t j = 0; j < extractGranularitySize; ++j) {
+        indices[i * extractGranularitySize + j] = linearizedIndex + j;
+      }
+    }
+    // Perform a shuffle to extract the kD vector.
+    rewriter.replaceOpWithNewOp<vector::ShuffleOp>(
+        extractOp, dstType, srcVector, srcVector,
+        rewriter.getI64ArrayAttr(indices));
+    return success();
+  }
+
+private:
+  unsigned targetVectorBitWidth;
+};
+
+/// This pattern converts the ShuffleOp that works on nD (n > 1)
+/// vectors to a ShuffleOp that works on linearized vectors.
+/// Following,
+///   vector.shuffle %v1, %v2 [ shuffle_indices ]
+/// is converted to :
+///   %v1_1d = vector.shape_cast %v1
+///   %v2_1d = vector.shape_cast %v2
+///   %out_1d = vector.shuffle %v1_1d, %v2_1d [ shuffle_indices_1d ]
+///   %out_nd = vector.shape_cast %out_1d
+// `shuffle_indices_1d` is computed using the sizes and `shuffle_indices`
+/// of the original shuffle operation.
+struct LinearizeVectorShuffle final
+    : public OpConversionPattern<vector::ShuffleOp> {
+  using OpConversionPattern::OpConversionPattern;
+  LinearizeVectorShuffle(
+      const TypeConverter &typeConverter, MLIRContext *context,
+      unsigned targetVectBitWidth = std::numeric_limits<unsigned>::max(),
+      PatternBenefit benefit = 1)
+      : OpConversionPattern(typeConverter, context, benefit),
+        targetVectorBitWidth(targetVectBitWidth) {}
+
+  LogicalResult
+  matchAndRewrite(vector::ShuffleOp shuffleOp, OpAdaptor adaptor,
+                  ConversionPatternRewriter &rewriter) const override {
+    auto dstType = getTypeConverter()->convertType(shuffleOp.getType());
+    assert(!(shuffleOp.getV1VectorType().isScalable() ||
+             shuffleOp.getV2VectorType().isScalable() ||
+             dstType.cast<VectorType>().isScalable()) &&
+           "scalable vectors are not supported.");
+    if (!isLessThanTargetBitWidth(shuffleOp, targetVectorBitWidth))
+      return rewriter.notifyMatchFailure(
+          shuffleOp, "Can't flatten since targetBitWidth <= OpSize");
+
+    auto vec1 = adaptor.getV1();
+    auto vec2 = adaptor.getV2();
+    int shuffleSliceLen = 1;
+    int rank = shuffleOp.getV1().getType().getRank();
+
+    // If rank > 1, we need to do the shuffle in the granularity of slices
+    // instead of scalars. Size of the slice is equal to the rank-1 innermost
+    // dims. Mask of the shuffle op specifies which slice to take from the
+    // outermost dim.
+    if (rank > 1) {
+      auto shape = shuffleOp.getV1().getType().getShape();
+      for (unsigned i = 1; i < shape.size(); i++) {
----------------
dcaballe wrote:

pre-increment

https://github.com/llvm/llvm-project/pull/88204