[tosa] Implement torch.linear support. (#535)

Refactor matmul into separate class and derive variants: - matmul - mm, bmm - linear Signed-off-by: Suraj Sudhir <suraj.sudhir@arm.com>
2022-01-25 08:48:58 -08:00 · 2022-01-25 08:48:58 -08:00 · cadea678e5
parent ad4b9e0369
commit cadea678e5
1 changed files with 242 additions and 39 deletions
--- a/lib/Conversion/TorchToTosa/TorchToTosa.cpp
+++ b/lib/Conversion/TorchToTosa/TorchToTosa.cpp
@ -817,40 +817,37 @@ LogicalResult ConvertAtenOp<AtenPowTensorScalarOp>::matchAndRewrite(
  return success();
 }

-// Perform torch matmul, mm and bmm
+// Perform the basic n-dim matmul operation encompassing the handling of
+// broadcasting and dynamic shape propagation.
+// All PyTorch ops that leverage matrix multiplication will derive this and
+// implement their specialized input processing (e.g transpose), and output
+// processing, e.g. GEMM or fully connected bias handling.
 template <typename AtenOpT>
-class ConvertAtenMatMulOp : public OpConversionPattern<AtenOpT> {
+class ConvertAtenMatmulBaseOp : public OpConversionPattern<AtenOpT> {
 public:
  using OpConversionPattern<AtenOpT>::OpConversionPattern;
  using OpAdaptor = typename AtenOpT::Adaptor;
-  LogicalResult
-  matchAndRewrite(AtenOpT op, OpAdaptor adaptor,
-                  ConversionPatternRewriter &rewriter) const override {
-    Value lhs = adaptor.self();
+  // Each variant must implement corresponding parameter parsing options.
+  // Maintain separate input read functions for each variant because it is not
+  // necessarily true with all variants that the first two operands are the lhs
+  // and rhs.
+  virtual LogicalResult readMatMulInputs(AtenOpT op, OpAdaptor adaptor,
+                                         ConversionPatternRewriter &rewriter,
+                                         Value &lhs, Value &rhs) const {
+    return rewriter.notifyMatchFailure(
+        op,
+        "Unimplemented matrix multiplication variant input parsing function");
+  }
+  LogicalResult performMatmul(AtenOpT op, OpAdaptor adaptor,
+                              ConversionPatternRewriter &rewriter, Value &lhs,
+                              Value &rhs, Value &output) const {
+
    auto lhsTy = lhs.getType().cast<RankedTensorType>();
-
-    // Aten matmul, mm and bmm call operand2 by different names.
-    Value rhs = adaptor.getOperands()[1];
    auto rhsTy = rhs.getType().cast<RankedTensorType>();

-    if (!lhsTy || !rhsTy)
-      return op.emitError("Only ranked tensor types supported in TOSA matmul");
-
    auto lhsRank = lhsTy.getRank();
    auto rhsRank = rhsTy.getRank();

-    // Mm takes two 2D tensors
-    if (isa<AtenMmOp>(op)) {
-      assert(lhsRank == 2 && rhsRank == 2 &&
-             "aten.mm called but matrix rank != 2");
-    }
-
-    // Bmm takes two 2D tensors
-    if (isa<AtenBmmOp>(op)) {
-      assert(lhsRank == 3 && rhsRank == 3 &&
-             "aten.bmm called but matrix rank != 2");
-    }
-
    auto lhsShape = lhsTy.getShape();
    auto rhsShape = rhsTy.getShape();

@ -1248,11 +1245,8 @@ public:
    // Perform the reshape to output shape. This is always required unless both
    // inputs are rank=3, in which case the tosa.matmul output itself is
    // correctly shaped.
-    bool performOpReshape = !(lhsRank == 3 && rhsRank == 3);
-
-    auto outputTy = OpConversionPattern<AtenOpT>::getTypeConverter()
-                        ->convertType(op.getType())
-                        .template cast<RankedTensorType>();
+    bool performOpReshape =
+        !(lhsRank == 3 && rhsRank == 3 && lhsShape[0] == rhsShape[0]);

    if (performOpReshape) {
      // Since the output shape may be unknown, we construct it
@ -1358,20 +1352,218 @@ public:

        auto transposedOpType =
            RankedTensorType::get(transposedOpShape, outputElemTy);
-        auto transposedOp = rewriter.create<tosa::TransposeOp>(
-            op->getLoc(),
-            OpConversionPattern<AtenOpT>::getTypeConverter()->convertType(
-                transposedOpType),
-            reshapedOp.getResult(), transposedOpShapeConst.getValue());
+        output =
+            rewriter
+                .create<tosa::TransposeOp>(
+                    op->getLoc(),
+                    OpConversionPattern<AtenOpT>::getTypeConverter()
+                        ->convertType(transposedOpType),
+                    reshapedOp.getResult(), transposedOpShapeConst.getValue())
+                .getResult();

-        rewriter.replaceOpWithNewOp<tensor::CastOp>(op, outputTy, transposedOp);
      } else {
-        rewriter.replaceOpWithNewOp<tensor::CastOp>(op, outputTy, reshapedOp);
+        output = reshapedOp.getResult();
      }
    } else {
-      rewriter.replaceOpWithNewOp<tensor::CastOp>(op, outputTy, mmOpResult);
+      output = mmOpResult;
    }

+    return success();
+  }
+  // The default version just reads two inputs, computes output and returns it.
+  // Other versions may add a bias, apply GEMM-style alpha/beta scaling etc.
+  virtual LogicalResult
+  matchAndRewrite(AtenOpT op, OpAdaptor adaptor,
+                  ConversionPatternRewriter &rewriter) const {
+
+    Value lhs, rhs;
+
+    if (failed(readMatMulInputs(op, adaptor, rewriter, lhs, rhs)))
+      return op.emitError("Failed to read matmul inputs");
+
+    Value output;
+
+    if (failed(performMatmul(op, adaptor, rewriter, lhs, rhs, output)))
+      return op.emitError("Failed to perform matmul operation");
+
+    rewriter.replaceOpWithNewOp<tensor::CastOp>(
+        op,
+        OpConversionPattern<AtenOpT>::getTypeConverter()
+            ->convertType(op.getType())
+            .template cast<RankedTensorType>(),
+        output);
+
+    return success();
+  }
+};
+
+// Legalizes the torch.matmul op for general n-dim matmul.
+template <typename AtenOpT>
+class ConvertAtenMatMulOp : public ConvertAtenMatmulBaseOp<AtenOpT> {
+public:
+  using ConvertAtenMatmulBaseOp<AtenOpT>::ConvertAtenMatmulBaseOp;
+  using OpAdaptor = typename AtenOpT::Adaptor;
+  LogicalResult readMatMulInputs(AtenOpT op, OpAdaptor adaptor,
+                                 ConversionPatternRewriter &rewriter,
+                                 Value &lhs, Value &rhs) const override {
+    lhs = adaptor.self();
+    auto lhsTy = lhs.getType().cast<RankedTensorType>();
+
+    rhs = adaptor.other();
+    auto rhsTy = rhs.getType().cast<RankedTensorType>();
+
+    if (!lhsTy || !rhsTy)
+      return op.emitError("Only ranked tensor types supported in TOSA matmul");
+
+    return success();
+  }
+};
+
+// Implements handling of aten.mm and aten.bmm ops.
+template <typename AtenOpT>
+class ConvertAtenMmOp : public ConvertAtenMatmulBaseOp<AtenOpT> {
+public:
+  using ConvertAtenMatmulBaseOp<AtenOpT>::ConvertAtenMatmulBaseOp;
+  using OpAdaptor = typename AtenOpT::Adaptor;
+  LogicalResult readMatMulInputs(AtenOpT op, OpAdaptor adaptor,
+                                 ConversionPatternRewriter &rewriter,
+                                 Value &lhs, Value &rhs) const override {
+
+    lhs = adaptor.self();
+    auto lhsTy = lhs.getType().cast<RankedTensorType>();
+
+    rhs = adaptor.mat2();
+    auto rhsTy = rhs.getType().cast<RankedTensorType>();
+
+    if (!lhsTy || !rhsTy)
+      return op.emitError("Only ranked tensor types supported in TOSA matmul");
+
+    auto lhsRank = lhsTy.getRank();
+    auto rhsRank = rhsTy.getRank();
+
+    if (isa<AtenMmOp>(op)) {
+      // Mm takes two 2D tensors.
+      if (lhsRank != 2 || rhsRank != 2)
+        return op.emitError("aten.mm called but matrix rank != 2");
+    } else if (isa<AtenBmmOp>(op)) {
+      // Bmm takes two 3D tensors.
+      if (lhsRank != 3 || rhsRank != 3)
+        return op.emitError("aten.bmm called but matrix rank != 3");
+    }
+
+    return success();
+  }
+};
+
+// Implements handling of aten.linear op.
+template <typename AtenOpT>
+class ConvertAtenLinearOp : public ConvertAtenMatmulBaseOp<AtenOpT> {
+public:
+  using ConvertAtenMatmulBaseOp<AtenOpT>::ConvertAtenMatmulBaseOp;
+  using OpAdaptor = typename AtenOpT::Adaptor;
+  LogicalResult readMatMulInputs(AtenOpT op, OpAdaptor adaptor,
+                                 ConversionPatternRewriter &rewriter,
+                                 Value &lhs, Value &rhs) const override {
+
+    lhs = adaptor.input();
+    auto lhsTy = lhs.getType().cast<RankedTensorType>();
+
+    rhs = adaptor.weight();
+    auto rhsTy = rhs.getType().cast<RankedTensorType>();
+
+    if (!lhsTy || !rhsTy)
+      return op.emitError("Only ranked tensor types supported in TOSA matmul");
+
+    auto lhsRank = lhsTy.getRank();
+    auto rhsRank = rhsTy.getRank();
+
+    if (lhsRank != 2 && lhsRank != 3)
+      return op.emitError("aten.Linear called but input rank not 2 or 3");
+    if (rhsRank != 2 && rhsRank != 3)
+      return op.emitError("aten.Linear called but weight rank not 2 or 3");
+
+    // Protection against crash due to unguarded code in TOSA->LinAlg.
+    if (!lhsTy.hasStaticShape() || !rhsTy.hasStaticShape())
+      return op.emitError("aten.Linear needs statically shaped input");
+
+    return success();
+  }
+  // Override the default rewriter to perform RHS transpose and bias addition as
+  // well.
+  LogicalResult
+  matchAndRewrite(AtenOpT op, OpAdaptor adaptor,
+                  ConversionPatternRewriter &rewriter) const override {
+
+    Value lhs, rhs;
+
+    if (failed(readMatMulInputs(op, adaptor, rewriter, lhs, rhs)))
+      return op.emitError("Failed to read matmul inputs");
+
+    // The aten.Linear op has a bias tensor that is added to the matmul output.
+    auto bias = adaptor.bias();
+    auto biasTy = bias.getType();
+
+    // TOSA does not mandate that elementwise op tensors need to be ranked.
+    if (!biasTy.template isa<Torch::NoneType>() &&
+        !biasTy.template isa<TensorType>())
+      return op.emitError("Only tensor types supported in GEMM to "
+                          "TOSA for bias tensor");
+
+    // RHS must have its last two dims transposed prior to matrix
+    // multiplication.
+    auto rhsTy = rhs.getType().cast<RankedTensorType>();
+    auto rhsRank = rhsTy.getRank();
+    auto rhsShape = rhsTy.getShape();
+    auto rhsElemTy = rhsTy.getElementType();
+
+    // Create a non-const shape array to transpose dims.
+    SmallVector<int64_t> transposedRhsShape;
+    for (auto &shape : rhsShape)
+      transposedRhsShape.push_back(shape);
+    SmallVector<int32_t> transposedRhsDims;
+    for (int32_t i = 0; i < rhsRank; i++)
+      transposedRhsDims.push_back(i);
+
+    // Swap the last two dims.
+    std::swap(transposedRhsShape[rhsRank - 1], transposedRhsShape[rhsRank - 2]);
+    std::swap(transposedRhsDims[rhsRank - 1], transposedRhsDims[rhsRank - 2]);
+
+    llvm::Optional<Value> transposedRhsShapeConst =
+        tosa::getConstTensor<int32_t>(
+            rewriter, op,
+            /*vec=*/transposedRhsDims,
+            /*shape=*/{static_cast<int32_t>(transposedRhsDims.size())});
+
+    auto transposedRhsType =
+        RankedTensorType::get(transposedRhsShape, rhsElemTy);
+    rhs = rewriter.create<tosa::TransposeOp>(
+        op->getLoc(),
+        OpConversionPattern<AtenOpT>::getTypeConverter()->convertType(
+            transposedRhsType),
+        rhs, transposedRhsShapeConst.getValue());
+
+    Value matmulOutput;
+    if (failed(
+            this->performMatmul(op, adaptor, rewriter, lhs, rhs, matmulOutput)))
+      return op.emitError("Failed to perform matmul operation");
+
+    Value matmulPlusBias = matmulOutput;
+    if (!biasTy.template isa<Torch::NoneType>()) {
+      // Bias addition broadcasts to the matmul output shape.
+      matmulPlusBias =
+          rewriter
+              .create<tosa::AddOp>(op->getLoc(), matmulOutput.getType(),
+                                   matmulOutput, bias)
+              .getResult();
+    }
+
+    rewriter.replaceOpWithNewOp<tensor::CastOp>(
+        op,
+        OpConversionPattern<AtenOpT>::getTypeConverter()
+            ->convertType(op.getType())
+            .template cast<RankedTensorType>(),
+        matmulPlusBias);
+
    return success();
  }
 };
@ -1544,10 +1736,21 @@ public:
  target.addIllegalOp<AtenOp>();                                               \
  patterns.add<ConvertAtenMatMulOp<AtenOp>>(typeConverter, context);
    INSERT_MATMUL_ATENOP_PATTERN(AtenMatmulOp);
-    INSERT_MATMUL_ATENOP_PATTERN(AtenMmOp);
-    INSERT_MATMUL_ATENOP_PATTERN(AtenBmmOp);
 #undef INSERT_MATMUL_ATEMOP_PATTERN

+#define INSERT_MM_ATENOP_PATTERN(AtenOp)                                       \
+  target.addIllegalOp<AtenOp>();                                               \
+  patterns.add<ConvertAtenMmOp<AtenOp>>(typeConverter, context);
+    INSERT_MM_ATENOP_PATTERN(AtenMmOp);
+    INSERT_MM_ATENOP_PATTERN(AtenBmmOp);
+#undef INSERT_MM_ATEMOP_PATTERN
+
+#define INSERT_LINEAR_ATENOP_PATTERN(AtenOp)                                   \
+  target.addIllegalOp<AtenOp>();                                               \
+  patterns.add<ConvertAtenLinearOp<AtenOp>>(typeConverter, context);
+    INSERT_LINEAR_ATENOP_PATTERN(AtenLinearOp);
+#undef INSERT_LINEAR_ATEMOP_PATTERN
+
 #define INSERT_ATENOP_PATTERN(AtenOp)                                          \
  target.addIllegalOp<AtenOp>();                                               \
  patterns.add<ConvertAtenOp<AtenOp>>(typeConverter, context);