LHS Registers Part 2 - Pipelining

cmake/llvm-hash.txt

@@ -1 +1 @@
-df0864e761107b07e38f5503e0cbee0cebb4c5e8
+29b92d07746fac26cd64c914bc9c5c3833974f6d

include/triton/Dialect/TritonGPU/IR/TritonGPUAttrDefs.td

@@ -361,8 +361,8 @@ compared to 1*64 when the hasLeadingOffset is false.
           return get(context, vec, perPhase, maxPhase, order, CTALayout);
         }
 
-        // ---- begin Ampere ----
-        if (mmaEnc.isAmpere()) {
+        // ---- begin Ampere & Hopper ----
+        if (mmaEnc.isAmpere() || mmaEnc.isHopper()) {
           int perPhase = 128 / (shapePerCTA[order[0]] * 4 / dotOpEnc.getKWidth());
           perPhase = std::max<int>(perPhase, 1);
           std::vector<size_t> matShape = {8, 8, 4 * dotOpEnc.getKWidth()};
@@ -397,13 +397,6 @@ compared to 1*64 when the hasLeadingOffset is false.
           llvm_unreachable("invalid operand index");
         }
 
-        // ---- begin version 3 ----
-        if (mmaEnc.isHopper()) {
-          llvm_unreachable("SharedEncodingAttr builder when the MMAEncodingAttr"
-                           " is Hopper has not been implemented yet");
-          return $_get(context, 1, 1, 1, order, CTALayout, true);
-        }
-
         // ---- not implemented ----
         llvm_unreachable("unsupported swizzling for provided MMA version");
     }]>,
@@ -1222,7 +1215,7 @@ For example, the matrix L corresponding to blockTileSize=[32,16] is:
     SmallVector<int> getMMAv1Rep(int opIdx) const;
     SmallVector<int> getMMAv1ShapePerWarp(int opIdx) const;
     int getMMAv1Vec(int opIdx) const;
-    SmallVector<int64_t> getMMAv2Rep(ArrayRef<int64_t> shape,
+    SmallVector<int64_t> getMMAv2OrV3Rep(ArrayRef<int64_t> shape,
                                      int bitwidth, int opIdx) const;
 
     bool supportReduction() const {
@@ -1317,6 +1310,10 @@ The parent field is the layout of d.
 kWidth defines number of consecutive elements stored by one thread along k dimension.
 Some layouts do not use this parameter, either because they have a fixed number of
 elements along the K dim, or they use all elements of the tensor along the K dim.
+
+We require kWidth to be provided for Hopper because the dtype at loading might be
+different from the dtype at WGMMA, due to casting. The kWidth is determined by the
+dtype at WGMMA.
   }];
 
   let parameters = (
@@ -1327,16 +1324,16 @@ elements along the K dim, or they use all elements of the tensor along the K dim
   );
 
   let builders = [
-        // Specially for MMAV1(Volta)
+    // For MMAV2 and V3
     AttrBuilder<(ins "unsigned":$opIdx,
                      "Attribute":$parent,
                      "Type":$eltTy), [{
       NvidiaMmaEncodingAttr parentAttr = mlir::dyn_cast<NvidiaMmaEncodingAttr>(parent);
-      if (!parentAttr || !parentAttr.isAmpere())
+      if (!parentAttr || (!parentAttr.isAmpere() && !parentAttr.isHopper()))
         return $_get(context, opIdx, parent, 0);
       unsigned bitwidth = eltTy.getIntOrFloatBitWidth();
-      unsigned MMAv2kWidth = 32 / bitwidth;
-      return $_get(context, opIdx, parent, MMAv2kWidth);
+      unsigned kWidth = 32 / bitwidth;
+      return $_get(context, opIdx, parent, kWidth);
     }]>
   ];
 

lib/Conversion/TritonGPUToLLVM/ElementwiseOpToLLVM.cpp

@@ -33,14 +33,15 @@ SmallVector<Value> reorderValues(const SmallVector<Value> &values, Type inType,
   // If the parent of the dot operand is in block encoding, we don't need to
   // reorder elements
   auto parentEncoding = dyn_cast<NvidiaMmaEncodingAttr>(ouEncoding.getParent());
-  if (!parentEncoding)
+  if (!parentEncoding || parentEncoding.isHopper())
     return values;
   size_t inBitWidth = inTensorTy.getElementType().getIntOrFloatBitWidth();
   size_t ouBitWidth = ouTensorTy.getElementType().getIntOrFloatBitWidth();
   auto ouEltTy = ouTensorTy.getElementType();
   if (inBitWidth == ouBitWidth)
     return values;
-  if (inBitWidth == 16 && ouBitWidth == 32) {
+  if ((inBitWidth == 16 && ouBitWidth == 32) ||
+      (inBitWidth == 32 && ouBitWidth == 16)) {
     SmallVector<Value> ret;
     for (unsigned i = 0; i < values.size(); i += 8) {
       ret.push_back(values[i]);
@@ -86,8 +87,12 @@ SmallVector<Value> unpackI32(const SmallVector<Value> &inValues, Type srcTy,
   if (!tensorTy)
     return inValues;
   auto encoding = dyn_cast<DotOperandEncodingAttr>(tensorTy.getEncoding());
-  if (!(encoding && isa<NvidiaMmaEncodingAttr>(encoding.getParent())))
+  if (!encoding)
+    return inValues;
+  auto parentEnc = dyn_cast<NvidiaMmaEncodingAttr>(encoding.getParent());
+  if (!parentEnc || parentEnc.isHopper())
     return inValues;
+
   SmallVector<Value> outValues;
   for (auto v : inValues) {
     // cast i32 to appropriate eltType vector and extract elements
@@ -108,8 +113,12 @@ SmallVector<Value> packI32(const SmallVector<Value> &inValues, Type srcTy,
   if (!tensorTy)
     return inValues;
   auto encoding = dyn_cast<DotOperandEncodingAttr>(tensorTy.getEncoding());
-  if (!(encoding && isa<NvidiaMmaEncodingAttr>(encoding.getParent())))
+  if (!encoding)
+    return inValues;
+  auto parentEnc = dyn_cast<NvidiaMmaEncodingAttr>(encoding.getParent());
+  if (!parentEnc || parentEnc.isHopper())
     return inValues;
+
   SmallVector<Value> outValues;
   auto eltType = typeConverter->convertType(tensorTy.getElementType());
   int vecWidth = 32 / eltType.getIntOrFloatBitWidth();

lib/Dialect/TritonGPU/IR/Dialect.cpp

@@ -1022,13 +1022,17 @@ LogicalResult DotOperandEncodingAttr::verify(
     return emitError() << "triton_gpu.dot_op parent paramenter cannot be null";
   }
   if (auto parentAttr = mlir::dyn_cast<NvidiaMmaEncodingAttr>(parent)) {
-    if (kWidth != 0 && !parentAttr.isAmpere())
+    if (kWidth != 0 && !(parentAttr.isAmpere() || parentAttr.isHopper()))
       return emitError() << "triton_gpu.dot_op kWidth parameter can only be "
-                            "non-zero for Ampere MMA parent";
-    if (kWidth == 0 && parentAttr.isAmpere())
+                            "non-zero for Ampere or Hopper MMA parent";
+    if (kWidth == 0 && (parentAttr.isAmpere() || parentAttr.isHopper()))
       return emitError()
              << "triton_gpu.dot_op kWidth parameter is mandatory for "
-                "Ampere MMA parent";
+                "Ampere or Hopper MMA parent";
+    if (opIdx != 0 && parentAttr.isHopper())
+      return emitError()
+             << "triton_gpu.dot_op opIdx parameter must be 0 for "
+                "Hopper MMA parent";
     return success();
   }
 
@@ -1957,17 +1961,17 @@ SmallVector<int> NvidiaMmaEncodingAttr::getMMAv1ShapePerWarp(int opIdx) const {
 int NvidiaMmaEncodingAttr::getMMAv1Vec(int opIdx) const {
   return 2 * getMMAv1Rep(opIdx)[opIdx];
 }
-SmallVector<int64_t> NvidiaMmaEncodingAttr::getMMAv2Rep(ArrayRef<int64_t> shape,
+SmallVector<int64_t> NvidiaMmaEncodingAttr::getMMAv2OrV3Rep(ArrayRef<int64_t> shape,
                                                         int bitwidth,
                                                         int opIdx) const {
+  assert(isAmpere() || isHopper());
   auto rank = shape.size();
   auto warpsPerCTA = getWarpsPerCTA();
   SmallVector<int> shapePerWarp = {1, 16, 8, 4 * 64 / bitwidth};
   int numRepBatch =
       rank == 3
           ? std::max<int64_t>(1, shape[0] / (shapePerWarp[0] * warpsPerCTA[0]))
           : 1;
-  assert(isAmpere());
 
   if (opIdx == 0)
     return {numRepBatch,
@@ -1982,18 +1986,25 @@ SmallVector<int64_t> NvidiaMmaEncodingAttr::getMMAv2Rep(ArrayRef<int64_t> shape,
                                                     warpsPerCTA[rank - 1]))};
   }
 }
+
 unsigned NvidiaMmaEncodingAttr::getTotalElemsPerThreadForOperands(
     ArrayRef<int64_t> shape, Type eltTy, int kWidth, int opIdx) const {
   auto shapePerCTA = getShapePerCTA(*this, shape);
   int warpsPerCTAM = getWarpsPerCTA()[0];
   int warpsPerCTAN = getWarpsPerCTA()[1];
   // H100
   if (isHopper()) {
-    return getTotalElemsPerThread(shape, eltTy);
+    assert(opIdx == 0);
+    auto instrMNK = getInstrShape();
+    int repM = ceil<unsigned>(shapePerCTA[0], instrMNK[0] * warpsPerCTAM);
+    int repK = ceil<unsigned>(shapePerCTA[1], instrMNK[2]);
+    // For each WGMMA instr, a 2x2 matrix fragment is loaded. Each thread holds
+    // kWidth elements for each quadrant. WGMMA is repeated repM * repK times.
+    return 4 * kWidth * repM * repK;
   }
   // A100
   if (isAmpere()) {
-    auto rep = getMMAv2Rep(shapePerCTA, eltTy.getIntOrFloatBitWidth(), opIdx);
+    auto rep = getMMAv2OrV3Rep(shapePerCTA, eltTy.getIntOrFloatBitWidth(), opIdx);
     if (opIdx == 0)
       return 4 * rep[0] * rep[1] * rep[2];
     if (opIdx == 1)
@@ -2720,6 +2731,11 @@ 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();

lib/Dialect/TritonGPU/Transforms/AccelerateMatmul.cpp

@@ -153,6 +153,21 @@ 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);
 }
 
@@ -162,6 +177,15 @@ class BlockedToMMA : public mlir::OpRewritePattern<DotOp> {
   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) ||