OpenXLA-specific changes

Author: Moerafaat

BUILD

@@ -123,6 +123,12 @@ ScratchConfig getScratchConfigForCvt(RankedTensorType srcTy,
 
   std::tie(scratchConfig.inVec, scratchConfig.outVec) =
       getScratchCvtInOutVecLengths(srcTy, dstTy);
+  // We can't write a longer vector than the shape of shared memory.
+  // This shape might be smaller than the tensor shape in case we decided to
+  // do the conversion in multiple iterations.
+  unsigned contiguousShapeDim = scratchConfig.repShape[scratchConfig.order[0]];
+  scratchConfig.inVec = std::min(scratchConfig.inVec, contiguousShapeDim);
+  scratchConfig.outVec = std::min(scratchConfig.outVec, contiguousShapeDim);
 
   // No padding is required if the tensor is 1-D, or if all dimensions except
   // the first accessed dimension have a size of 1.

lib/Analysis/Allocation.cpp

@@ -935,7 +935,7 @@ class ShROpAxisInfoVisitor final : public BinaryOpVisitorImpl<OpTy> {
       // Treat [2^n,2^n+1,...]'s divisibility as 1 instead of 2^n
       lhsDivisibility = 1;
     }
-    return std::max<int64_t>(1, lhsDivisibility / (1 << shift));
+    return std::max<int64_t>(1, lhsDivisibility / (int64_t(1) << shift));
   }
 
   int64_t getConstancy(OpTy op, const AxisInfo &lhs, const AxisInfo &rhs,

lib/Analysis/AxisInfo.cpp

@@ -57,6 +57,12 @@ TritonGPUTypeConverter::TritonGPUTypeConverter(MLIRContext *context,
   addArgumentMaterialization([&](OpBuilder &builder,
                                  RankedTensorType tensorType, ValueRange inputs,
                                  Location loc) -> Value {
+    // Allows partial TTIR to TTGIR conversion by materializing a conversion for
+    // remaining arguments that have been converted to a new type.
+    // We use this to rewrite triton_xla.sparse_dot in a separate pass after
+    // 'convert-triton-to-tritongpu'.
+    return builder.create<triton::gpu::ConvertLayoutOp>(loc, tensorType,
+                                                        inputs);
     llvm_unreachable("Argument rematerialization should not happen in Triton "
                      "-> TritonGPU conversion");
     return {};
@@ -66,6 +72,12 @@ TritonGPUTypeConverter::TritonGPUTypeConverter(MLIRContext *context,
   // convert origValue to newValue
   addSourceMaterialization([&](OpBuilder &builder, RankedTensorType tensorType,
                                ValueRange inputs, Location loc) -> Value {
+    // Allows partial TTIR to TTGIR conversion by materializing a conversion for
+    // remaining uses of values that have been converted to a new type.
+    // We use this to rewrite triton_xla.sparse_dot in a separate pass after
+    // 'convert-triton-to-tritongpu'.
+    return builder.create<triton::gpu::ConvertLayoutOp>(loc, tensorType,
+                                                        inputs);
     llvm_unreachable("Source rematerialization should not happen in Triton -> "
                      "TritonGPU Conversion");
     return {};

lib/Conversion/TritonToTritonGPU/TritonGPUConversion.cpp

@@ -150,6 +150,11 @@ struct CanonicalizeConvertFromAlloc
     auto convert = op.getSrc().getDefiningOp<ConvertLayoutOp>();
     if (!convert)
       return failure();
+    // LocalAllocOp lowering doesn't support going from DotOperandEncoding
+    // to SharedEncoding, so we want to keep this layout conversion.
+    if (mlir::isa<triton::gpu::DotOperandEncodingAttr>(
+            convert.getSrc().getType().getEncoding()))
+      return failure();
     rewriter.replaceOpWithNewOp<triton::gpu::LocalAllocOp>(
         op, op->getResult(0).getType(), convert.getSrc());
     return mlir::success();
@@ -212,13 +217,13 @@ struct CanonicalizeConvertFromConvert
     // heuristic to accommodate fused attention.
     auto srcType = op.getSrc().getType();
     auto dstType = op.getType();
-    if (mlir::isa<DotOperandEncodingAttr>(dstType.getEncoding()) &&
-        mlir::isa<NvidiaMmaEncodingAttr>(srcType.getEncoding()))
+    if (mlir::isa_and_nonnull<DotOperandEncodingAttr>(dstType.getEncoding()) &&
+        mlir::isa_and_nonnull<NvidiaMmaEncodingAttr>(srcType.getEncoding()))
       return failure();
 
     // for hopper MMAv3
-    if (mlir::isa<SharedEncodingAttr>(dstType.getEncoding()) &&
-        mlir::isa<NvidiaMmaEncodingAttr>(srcType.getEncoding()) &&
+    if (mlir::isa_and_nonnull<SharedEncodingAttr>(dstType.getEncoding()) &&
+        mlir::isa_and_nonnull<NvidiaMmaEncodingAttr>(srcType.getEncoding()) &&
         llvm::any_of(op.getResult().getUsers(), [](Operation *dot) {
           return dot->hasTrait<OpTrait::DotLike>();
         })) {

lib/Dialect/TritonGPU/IR/Ops.cpp

@@ -21,15 +21,13 @@ namespace mlir {
 namespace triton {
 namespace gpu {
 
-namespace {
-
 // Get the highest version supported for the hardware and the dot.
 static int getMMAVersionSafe(int computeCapability, DotOp op) {
   // List supported mma version in order of preference.
   SmallVector<int> versionsSupported;
   if (computeCapability < 75) {
     versionsSupported = {1};
-  } else if (computeCapability < 90) {
+  } else if (computeCapability < 90 || computeCapability >= 100) {
     versionsSupported = {2};
   } else if (computeCapability < 100) {
     versionsSupported = {3, 2};
@@ -45,8 +43,8 @@ static int getMMAVersionSafe(int computeCapability, DotOp op) {
   return 0;
 }
 
-SmallVector<unsigned> warpsPerTileV2(DotOp dotOp, const ArrayRef<int64_t> shape,
-                                     int numWarps) {
+SmallVector<unsigned>
+warpsPerTileV2(Operation *dotOp, const ArrayRef<int64_t> shape, int numWarps) {
   auto rank = shape.size();
   // Early exit for batched matmul
   if (rank == 3)
@@ -110,10 +108,10 @@ SmallVector<unsigned> warpsPerTileV2(DotOp dotOp, const ArrayRef<int64_t> shape,
 }
 
 SmallVector<unsigned, 2>
-warpsPerTileV3(DotOp dotOp, const ArrayRef<int64_t> shape, int numWarps,
+warpsPerTileV3(Operation *dotOp, const ArrayRef<int64_t> shape, int numWarps,
                const SmallVector<unsigned, 3> &instrShape) {
   SetVector<Operation *> slices;
-  mlir::getForwardSlice(dotOp.getResult(), &slices);
+  mlir::getForwardSlice(dotOp->getResult(0), &slices);
   // Contains a chained dot. We prefer to assign warps to one axis
   // to facilitate use cases like flash attention, allowing reductions within
   // the same warp.
@@ -168,11 +166,26 @@ static Value getSharedMemoryMMAOperand(Value v, mlir::PatternRewriter &rewriter,
   auto newType = MemDescType::get(argType.getShape(), argType.getElementType(),
                                   newLayout, SharedMemorySpace);
   rewriter.setInsertionPointAfterValue(arg);
+
+  // LocalAllocOp lowering doesn't support going from DotOperandEncoding
+  // to SharedEncoding.
+  if (auto dotOpEnc = mlir::dyn_cast<DotOperandEncodingAttr>(
+          argType.getEncoding())) {
+    // Create a layout conversion from DotOperandEncoding to BlockedEncoding
+    // then pass it to the LocalAllocOp.
+    auto newArgType = RankedTensorType::get(
+        argType.getShape(), argType.getElementType(), dotOpEnc.getParent());
+    auto dotOperandToBlockedCvt =
+        rewriter.create<ConvertLayoutOp>(arg.getLoc(), newArgType, arg);
+    return rewriter.create<LocalAllocOp>(arg.getLoc(), newType,
+                                              dotOperandToBlockedCvt);
+  }
+
   return rewriter.create<LocalAllocOp>(arg.getLoc(), newType, arg);
 }
 
 SmallVector<unsigned, 3>
-getWarpsPerTile(DotOp dotOp, const ArrayRef<int64_t> shape, int version,
+getWarpsPerTile(Operation* dotOp, const ArrayRef<int64_t> shape, int version,
                 int numWarps, const SmallVector<unsigned, 3> &instrShape) {
   switch (version) {
   case 2:
@@ -185,18 +198,32 @@ getWarpsPerTile(DotOp dotOp, const ArrayRef<int64_t> shape, int version,
   }
 }
 
+// Move anonymous namespace down, so getWarpsPerTile is visible to the sparsity
+// extension.
+namespace {
+
 class BlockedToMMA : public mlir::OpRewritePattern<DotOp> {
   int computeCapability;
   mutable llvm::DenseMap<Operation *, unsigned> dotOpInstNs;
 
   static bool bwdFilter(Operation *op) {
+    // Dot operand layout assignment to Predicates are not currently supported
+    // during lowering from TritonGPU to LLVM in Triton for MMA cases. This
+    // condition limits visibility of the original bit-width so that predicate
+    // are not considered, hence, kwidth can never be = 32.
+    if (isa<arith::UIToFPOp>(op)) {
+      Type srcType = getElementTypeOrSelf(op->getOperand(0));
+      if (srcType.isInteger(1))
+        return false;
+    }
     return op->getNumOperands() == 1 &&
            (isa<FpToFpOp, BitcastOp, ConvertLayoutOp>(op) ||
             isPureUnaryInlineAsm(op) ||
             op->getDialect()->getTypeID() ==
                 mlir::TypeID::get<arith::ArithDialect>());
   }
 
+public:
   // Finds the first different bitwidth in the chain of shape-preserving
   // unary ops that x depends on.
   // There are two primary scenarios:
@@ -720,6 +747,15 @@ class TritonGPUAccelerateMatmulPass
   }
 };
 
+// Expose helper functions from BlockedToMMA to be reused for sparse matmul.
+int computeOrigBitWidth(Value x) {
+  return BlockedToMMA::computeOrigBitWidth(x);
+}
+Value getSharedMemMMAOperand(Value v, mlir::PatternRewriter &rewriter,
+                                int opIdx, bool allowTranspose) {
+  return getSharedMemoryMMAOperand(v, rewriter, opIdx, allowTranspose);
+}
+
 } // namespace gpu
 } // namespace triton
 } // namespace mlir

lib/Dialect/TritonGPU/Transforms/AccelerateMatmul.cpp

@@ -284,6 +284,7 @@ class HoistLayoutConversion : public OpRewritePattern<ConvertLayoutOp> {
     if (!foundInitializer)
       return failure();
 
+    rewriter.setInsertionPointAfter(src);
     SmallVector<ConvertLayoutOp> newOperands;
     for (auto operand : src->getOperands()) {
       // We checked earlier that all operands are ranked tensors.

lib/Dialect/TritonGPU/Transforms/OptimizeDotOperands.cpp

@@ -131,6 +131,7 @@ static int createAsyncCopy(scf::ForOp forOp, tt::LoadOp loadOp, Value alloc,
 
   Value zero = builder.createWithStage<arith::ConstantIntOp>(
       forOp.getLoc(), stage, clusterId, 0, 32);
+
   // Replace the load with insert/extract slice.
   builder.setInsertionPoint(loadOp);
   Location loc = loadOp.getLoc();
@@ -491,7 +492,8 @@ assignMemoryLayouts(scf::ForOp &forOp,
     });
 
     bool isTMALoad = isa<tt::ExperimentalDescriptorLoadOp>(op);
-    loadsToPipeline.insert(&op);
+    // TODO: b/381421713 - Uncomment this once pipelining is fixed.
+    // loadsToPipeline.insert(&op);
     LoadInfo loadInfo;
     for (auto use : users) {
       if (use->hasTrait<OpTrait::DotLike>()) {
@@ -527,6 +529,11 @@ assignMemoryLayouts(scf::ForOp &forOp,
               getBlockedEncoding(loadOp, axisInfoAnalysis);
       }
     }
+
+    // TODO: b/381421713 - Remove this once pipelining is fixed.
+    if (!loadInfo.sharedEncoding) continue;
+    loadsToPipeline.insert(&op);
+
     loadToInfo[&op] = loadInfo;
   }
   // Make sure all loads in loadsToPipeline are in loadToInfo.

lib/Dialect/TritonGPU/Transforms/Pipeliner/MatmulLoopPipeline.cpp

@@ -116,7 +116,7 @@ Value Prefetcher::generatePrefetch(Value v, unsigned opIdx, bool isPrologue,
   // opIdx: 0 => a, 1 => b
   auto type = cast<triton::gpu::MemDescType>(v.getType());
   SmallVector<int64_t> shape{type.getShape().begin(), type.getShape().end()};
-  SmallVector<int64_t> offset{0, 0};
+  SmallVector<int64_t> offset(shape.size(), 0);
   Type elementType = type.getElementType();
 
   // k => (prefetchWidth, k - prefetchWidth)
@@ -141,8 +141,14 @@ Value Prefetcher::generatePrefetch(Value v, unsigned opIdx, bool isPrologue,
           type.getMutableMemory(), type.getAllocShape()),
       v, offsetsVal);
 
+  // We need to assign kwidth to zero in the case where the parent layout is
+  // Blocked, otherwise the verifier emits a failure. The parent layout is
+  // Blocked only when Tensor Cores are disabled.
+  int kwidth = dyn_cast<triton::gpu::BlockedEncodingAttr>(dotEncoding)
+                   ? 0
+                   : prefetchWidth / 8;
   auto dotOperandEnc = triton::gpu::DotOperandEncodingAttr::get(
-      builder.getContext(), opIdx, dotEncoding, prefetchWidth / 8);
+      builder.getContext(), opIdx, dotEncoding, kwidth);
   Value prefetchSlice = builder.create<triton::gpu::LocalLoadOp>(
       v.getLoc(), RankedTensorType::get(shape, elementType, dotOperandEnc),
       newSmem);
@@ -191,6 +197,22 @@ LogicalResult Prefetcher::initialize() {
         break;
       if (!op->getResult(0).hasOneUse())
         break;
+      // Similar to issues faced in HoistLayoutConversion pattern in
+      // OptimizeDotOperands.cpp, we can't propagate through type casts from
+      // predicates as they aren't supported in Triton when encoded with dot_op
+      // layout.
+      if (isa<arith::UIToFPOp>(op)) {
+        Type srcType = getElementTypeOrSelf(op->getOperand(0));
+        if (srcType.isInteger(1))
+          break;
+      }
+      // Propagation through ExpandDims is currently not supported. This blindly
+      // replaces the encoding with dot encoding & but ExpandDims requires a
+      // SliceEncoding. This could be rewritten to support it somehow, but I
+      // don't think it's trivial & it's currently crashing.
+      if (isa<ExpandDimsOp>(op)) {
+        break;
+      }
       rets.push_back(op->getOperand(0));
       if (auto cvt = dyn_cast<triton::gpu::LocalLoadOp>(op)) {
         foundConvertFromShared = true;

lib/Dialect/TritonGPU/Transforms/Prefetch.cpp

@@ -994,18 +994,26 @@ getSharedEncIfAllUsersAreDotEnc(Value val, bool &incompatible) {
     } else {
       if (!isa<ttg::LocalLoadOp, ttg::ConvertLayoutOp>(user))
         return std::nullopt;
-      auto dotOpEnc = dyn_cast<ttg::DotOperandEncodingAttr>(
-          cast<triton::gpu::TensorOrMemDesc>(user->getResult(0).getType())
-              .getEncoding());
-      if (!dotOpEnc)
+      auto enc =
+          cast<triton::gpu::TensorOrMemDesc>(user->getResult(0).getType()).getEncoding();
+      if (isa<ttg::DotOperandEncodingAttr>(enc)) {
+        auto srcTy = cast<triton::gpu::TensorOrMemDesc>(val.getType());
+        auto CTALayout = ttg::getCTALayout(srcTy.getEncoding());
+        auto order = ttg::getOrder(srcTy.getEncoding());
+        unsigned bitWidth = srcTy.getElementType().getIntOrFloatBitWidth();
+        tempAttr = ttg::SharedEncodingAttr::get(
+            val.getContext(), cast<ttg::DotOperandEncodingAttr>(enc),
+            srcTy.getShape(), order, CTALayout, bitWidth, /*needTrans=*/false);
+      } else if (enc.getAbstractAttribute().getName().str() ==
+                 "triton.gpu.sparse_dot_meta_encoding") {
+        auto srcTy = cast<triton::gpu::TensorOrMemDesc>(val.getType());
+        tempAttr = ttg::SharedEncodingAttr::get(
+            val.getContext(), /*vec=*/1, /*perPhase=*/1, /*maxPhase=*/1,
+            ttg::getOrder(srcTy.getEncoding()),
+            ttg::getCTALayout(srcTy.getEncoding()));
+      } else {
         return std::nullopt;
-      auto srcTy = cast<triton::gpu::TensorOrMemDesc>(val.getType());
-      auto CTALayout = ttg::getCTALayout(srcTy.getEncoding());
-      auto order = ttg::getOrder(srcTy.getEncoding());
-      unsigned bitWidth = srcTy.getElementType().getIntOrFloatBitWidth();
-      tempAttr = ttg::SharedEncodingAttr::get(
-          val.getContext(), dotOpEnc, srcTy.getShape(), order, CTALayout,
-          bitWidth, /*needTrans=*/false);
+      }
     }
     // Check that the shared encodings needed by the users are compatible.
     if (attr != nullptr && attr != tempAttr) {

lib/Dialect/TritonGPU/Transforms/Utility.cpp

@@ -44,7 +44,8 @@ struct FenceInsertionPass
       return;
     ModuleOp mod = getOperation();
     mod.walk([&](Operation *op) {
-      if (!isa<ttng::WarpGroupDotOp>(op))
+      if (!isa<ttng::WarpGroupDotOp>(op) &&
+          op->getName().getStringRef() != "triton_xla.sparse_dot")
         return WalkResult::advance();
       OpBuilder builder(op);
       auto a = op->getOperand(0);