ATen scheduler for the new Matmul/LinearOp IR nodes (#2209)

1. Adds a new scheduler -- `ExprEvalScheduler` that accepts the MatmulOp and LinearOp (next PR) for ATen evaluation. 2. Modify the matmul input generator to test for all cases supported by Thunder. 3. The `eagerMatmul` API is renamed and replaces the existing `matmul` API. `fd.ops.matmul` now creates a `MatmulOp` (except in a few special cases such as scalar dot product, for eg: `[K] x [K]`. Issue #2149, #2092.
NVIDIA · May 14, 2024 · dfba77a · dfba77a
1 parent 03716fc
commit dfba77a
Show file tree

Hide file tree

Showing 15 changed files with 211 additions and 139 deletions.
diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -208,6 +208,7 @@ list(APPEND NVFUSER_SRCS
   ${NVFUSER_SRCS_DIR}/scheduler/transpose.cpp
   ${NVFUSER_SRCS_DIR}/scheduler/utils.cpp
   ${NVFUSER_SRCS_DIR}/scheduler/vectorize_helper.cpp
+  ${NVFUSER_SRCS_DIR}/scheduler/expr_eval_sched.cpp
   ${NVFUSER_SRCS_DIR}/serde/polymorphic_value.cpp
   ${NVFUSER_SRCS_DIR}/serde/utils.cpp
   ${NVFUSER_SRCS_DIR}/swizzle.cpp

diff --git a/csrc/ops/composite.cpp b/csrc/ops/composite.cpp
@@ -54,42 +54,6 @@ TensorView* dropout_backward(TensorView* dy, TensorView* mask, Val* scale) {
   return dx;
 }
 
-// This function will add a castOp to the output of the matrix multiplication
-// The implementation of linear can use this but will skip the cast (set cast
-// flag as false) and add the bias.
-TensorView* matmul(TensorView* a, TensorView* b, bool cast_output_to_input) {
-  NVF_CHECK(
-      a->nDims() == b->nDims(),
-      "The number of dimension of A and B do not match");
-  // TODO: We'll need to suppor nDims == 3 for bmm.
-  NVF_CHECK(
-      a->nDims() == 2,
-      "Only 2-D Tensors are supported, in the future we'll support 3-D as well!");
-
-  std::vector<bool> bcast_dims(a->nDims() + 1, false);
-  // A: [M, K, Bcast]
-  // B: [Bcast, K, N]
-  bcast_dims.at(bcast_dims.size() - 1) = true;
-  auto* tv0b = broadcast(a, bcast_dims);
-  bcast_dims.at(bcast_dims.size() - 1) = false;
-  bcast_dims.at(bcast_dims.size() - 3) = true;
-  auto* tv1b = broadcast(b, bcast_dims);
-
-  NVF_CHECK(
-      a->getDataType().value() == b->getDataType().value(),
-      "data types of inputs to matmul don't match");
-  auto* output = fusedMultiplySum(tv0b, tv1b, {-2});
-  if (cast_output_to_input) {
-    // For matmul, the output dtype should match input.
-    return maybeCastOp(a->getDataType().value(), output);
-  }
-  return output;
-}
-
-TensorView* matmul(TensorView* a, TensorView* b) {
-  return matmul(a, b, true /* cast output to input dtype */);
-}
-
 TensorView* linear(TensorView* a, TensorView* b, TensorView* bias) {
   // TODO: Support 1+ dimensional A.
   NVF_CHECK(
@@ -348,10 +312,7 @@ static TensorView* newForMatmul(TensorView* tv_a, TensorView* tv_b) {
 
 } // namespace
 
-// TODO (Priya): This will be renamed to matmul once we are ready to modify the
-// python API backend. Keeping separate for now, to avoid breaking tests in
-// Thunder.
-TensorView* eagerMatmul(TensorView* tv_a, TensorView* tv_b) {
+TensorView* matmul(TensorView* tv_a, TensorView* tv_b) {
   NVF_CHECK(
       tv_a->nDims() > 0 && tv_b->nDims() > 0,
       "Expected inputs to be atleast 1D, got: ",

diff --git a/csrc/ops/composite.h b/csrc/ops/composite.h
@@ -47,16 +47,6 @@ NVF_API LstmResult lstm(
     TensorView* cell_x,
     TensorView* out_x);
 
-// Matmul function which takes in tensors with the shapes
-// A[M,K] B[K,N], but the tensors may have different layouts
-// via strides. All restrictions from the matmul APIs also
-// apply here.
-TensorView* matmul(TensorView* a, TensorView* b);
-// This second matmul function is not exposed via
-// the Python interface, but it does the guts of the work and
-// can be used to create mamtuls without a cast operation following it.
-TensorView* matmul(TensorView* a, TensorView* b, bool cast_output_to_input);
-
 // Linear functions which takes in two tensors of shapes A[M,K] and
 // B[N,K]. Takes in a options bias of shape [N] and performs
 // out = A * B_Transpose + bias. The output dtype matches the dtype
@@ -81,6 +71,9 @@ TensorView* leaky_relu(TensorView* x, Val* negative_slope);
 
 NVF_API TensorView* view_as_real(TensorView* x);
 
-TensorView* eagerMatmul(TensorView* tv_a, TensorView* tv_b);
+// Matmul function which takes in tensors with the shapes
+// A[*, M, K] / A[K] and B[*, K, N] / B[K], but the tensors may have different
+// layouts via strides. This has the same functionality as torch.matmul
+TensorView* matmul(TensorView* tv_a, TensorView* tv_b);
 
 } // namespace nvfuser
diff --git a/csrc/root_domain_map.cpp b/csrc/root_domain_map.cpp
@@ -123,23 +123,55 @@ std::unordered_map<IterDomain*, IterDomain*> PairwiseRootDomainMap::map(
       TensorDomain::noReductions(producer->maybeRFactor());
   const auto& consumer_root = consumer->root();
 
-  // Add key-value iterdomain pair to the map.
-  auto updatePairwiseRootDomainMap =
-      [&root_dims_to_map, producer_to_consumer, &dom_map](
-          IterDomain* map_key_id, IterDomain* map_value_id) {
-        if (!producer_to_consumer) {
-          std::swap(map_key_id, map_value_id);
-        }
-        if (root_dims_to_map.find(map_key_id) != root_dims_to_map.end()) {
-          dom_map.insert(std::make_pair(map_key_id, map_value_id));
-        }
-      };
+  // Check following conditions and add key-value iterdomain pair to domain map:
+  // 1. Do not map broadcast ID to non-broadcast ID unless map_broadcast_ =
+  // true.
+  // 2. Do not map Symbolic ID if the extents are not identical unless
+  // map_symbolic_ = true.
+  auto updatePairwiseRootDomainMap = [&](IterDomain* producer_id,
+                                         IterDomain* consumer_id) {
+    if (!map_broadcast_ &&
+        producer_id->isBroadcast() != consumer_id->isBroadcast()) {
+      return;
+    }
+
+    // Condition: At least one ID is symbolic.
+    //
+    // If map_symbolic_ is true:
+    //   Map these IDs regardless of other considerations.
+    //
+    // If map_symbolic_ is false (default):
+    //   Map these only if their extents are identical. IterType::Symbolic
+    //   reflects that the extent might evaluate to 1 for some inputs, in which
+    //   case it may be valid to use those domains in a broadcast op. If the
+    //   extents are exactly the same between two aligned IterDomains, the
+    //   Symbolic one will be concretized to the same IterType as the other, so
+    //   they should be mapped with one another.
+    if (!map_symbolic_ &&
+        (producer_id->isSymbolic() || consumer_id->isSymbolic()) &&
+        (!producer_id->extent()->sameAs(consumer_id->extent()))) {
+      return;
+    }
+
+    IterDomain* map_key_id = producer_id;
+    IterDomain* map_value_id = consumer_id;
+
+    if (!producer_to_consumer) {
+      std::swap(map_key_id, map_value_id);
+    }
+
+    if (root_dims_to_map.find(map_key_id) != root_dims_to_map.end()) {
+      dom_map.insert(std::make_pair(map_key_id, map_value_id));
+    }
+  };
 
   // For MatmulOp, use the corresponding mapped input iterdomains.
   if (MatmulOp* op = dynamic_cast<MatmulOp*>(consumer_tv_->definition())) {
     // Check if the producer is lhs/rhs input
     MatmulRole input_role =
-        producer->sameAs(op->inA()) ? MatmulRole::INPUT_A : MatmulRole::INPUT_B;
+        producer->sameAs(op->inA()->as<TensorView>()->domain())
+        ? MatmulRole::INPUT_A
+        : MatmulRole::INPUT_B;
     auto out_size = consumer_root.size();
 
     // For MatmulOp, the input iterdomains at a given index do not necessarily
@@ -150,14 +182,18 @@ std::unordered_map<IterDomain*, IterDomain*> PairwiseRootDomainMap::map(
     // input and output for index=2
     // 2. `B, M, K] x [K, N] -> [B, M, N]`: For  input B, the second iterdomain
     // maps to the third output iterdomain.
-    const std::vector<IterDomain*>& aligned_producer_id =
+    const std::vector<IterDomain*>& aligned_producer_ids =
         ops::mapMatmulOpIterDomains(producer_root, input_role, out_size);
 
     for (auto inx : c10::irange(out_size)) {
-      IterDomain* map_key_id = aligned_producer_id.at(inx);
-      IterDomain* map_value_id = consumer_root.at(inx);
-      updatePairwiseRootDomainMap(map_key_id, map_value_id);
+      IterDomain* producer_id = aligned_producer_ids.at(inx);
+      IterDomain* consumer_id = consumer_root.at(inx);
+      if (producer_id == nullptr) {
+        continue;
+      }
+      updatePairwiseRootDomainMap(producer_id, consumer_id);
     }
+
     return dom_map;
   }
 
@@ -171,8 +207,6 @@ std::unordered_map<IterDomain*, IterDomain*> PairwiseRootDomainMap::map(
     // 2. IDs that may have different extents (e.g., non indexed
     //  domains of torch_gather)
     // 3. Squeeze and unsqueeze
-    // 4. Broadcast and non broadcast
-    // 5. Symbolic ID with different extent from other ID
 
     // Condition 1: when the producer ID is the dim of a select-like op
     if (producer_id == indexed_producer_id) {
@@ -217,38 +251,7 @@ std::unordered_map<IterDomain*, IterDomain*> PairwiseRootDomainMap::map(
       continue;
     }
 
-    // Condition 4
-    if (!map_broadcast_ &&
-        producer_id->isBroadcast() != consumer_id->isBroadcast()) {
-      itc++;
-      itp++;
-      continue;
-    }
-
-    // Condition 5
-    // At least one ID is symbolic.
-    //
-    // If map_symbolic_ is true:
-    //   Map these IDs regardless of other considerations.
-    //
-    // If map_symbolic_ is false (default):
-    //   Map these only if their extents are identical. IterType::Symbolic
-    //   reflects that the extent might evaluate to 1 for some inputs, in which
-    //   case it may be valid to use those domains in a broadcast op. If the
-    //   extents are exactly the same between two aligned IterDomains, the
-    //   Symbolic one will be concretized to the same IterType as the other, so
-    //   they should be mapped with one another.
-    if (!map_symbolic_ &&
-        (producer_id->isSymbolic() || consumer_id->isSymbolic()) &&
-        (!producer_id->extent()->sameAs(consumer_id->extent()))) {
-      itc++;
-      itp++;
-      continue;
-    }
-
-    IterDomain* map_key_id = producer_id;
-    IterDomain* map_value_id = consumer_id;
-    updatePairwiseRootDomainMap(map_key_id, map_value_id);
+    updatePairwiseRootDomainMap(producer_id, consumer_id);
 
     itc++;
     itp++;

diff --git a/csrc/scheduler/all_schedulers.h b/csrc/scheduler/all_schedulers.h
@@ -6,6 +6,7 @@
  */
 // clang-format on
 #pragma once
+#include <scheduler/expr_eval_sched.h>
 #include <scheduler/matmul.h>
 #include <scheduler/no_op.h>
 #include <scheduler/normalization_inner.h>

diff --git a/csrc/scheduler/expr_eval_sched.cpp b/csrc/scheduler/expr_eval_sched.cpp
@@ -0,0 +1,33 @@
+// clang-format off
+/*
+ * SPDX-FileCopyrightText: Copyright (c) 2023-present NVIDIA CORPORATION & AFFILIATES.
+ * All rights reserved.
+ * SPDX-License-Identifier: BSD-3-Clause
+ */
+// clang-format on
+
+#include <ir/utils.h>
+#include <scheduler/debug_utils.h>
+#include <scheduler/expr_eval_sched.h>
+#include <scheduler/registry_utils.h>
+
+namespace nvfuser {
+
+// Check if the fusion has a single MatmulOp node
+bool ExprEvalScheduler::canScheduleCompileTime(Fusion* fusion) {
+  auto exprs = fusion->exprs();
+  if (exprs.size() == 1 && exprs.front()->isA<MatmulOp>()) {
+    return true;
+  }
+  scheduler_debug_utils::canScheduleRejectReason(
+      heuristicType(),
+      "Fusion must contain a single expression of type MatmulOp");
+  return false;
+}
+
+void ExprEvalScheduler::schedule(Fusion* fusion) {
+  fusion->aliasOutputToInput(
+      fusion->outputs()[0], /*input=*/nullptr, AllocationType::Evaluate);
+}
+
+} // namespace nvfuser
diff --git a/csrc/scheduler/expr_eval_sched.h b/csrc/scheduler/expr_eval_sched.h
@@ -0,0 +1,49 @@
+// clang-format off
+/*
+ * SPDX-FileCopyrightText: Copyright (c) 2023-present NVIDIA CORPORATION & AFFILIATES.
+ * All rights reserved.
+ * SPDX-License-Identifier: BSD-3-Clause
+ */
+// clang-format on
+#pragma once
+
+#include <scheduler/heuristic.h>
+#include <scheduler/registry.h>
+
+namespace nvfuser {
+
+class Fusion;
+class SchedulerRuntimeInfo;
+class HeuristicSummary;
+
+// ExprEval scheduler represents the case where we allocate outputs directly
+// using EE. No code is generated.
+class ExprEvalScheduler : public SchedulerEntry {
+ public:
+  explicit ExprEvalScheduler(
+      Fusion* fusion,
+      SchedulerRuntimeInfo& runtime_info,
+      HeuristicSummary* data_cache = nullptr)
+      : SchedulerEntry(heuristicType()) {
+    params_ =
+        std::make_shared<HeuristicParams>("", runtime_info.getIndexType());
+  }
+
+  // This scheduler only accepts MatmulOp.
+  static bool canScheduleCompileTime(Fusion* fusion);
+
+  static bool canScheduleRunTime(
+      Fusion* fusion,
+      SchedulerRuntimeInfo& runtime_info,
+      HeuristicSummary* data_cache) {
+    return true;
+  }
+
+  constexpr static ScheduleHeuristic heuristicType() {
+    return ScheduleHeuristic::ExprEval;
+  }
+
+  void schedule(Fusion* fusion) override;
+};
+
+} // namespace nvfuser
diff --git a/csrc/scheduler/heuristic.h b/csrc/scheduler/heuristic.h
@@ -25,11 +25,20 @@ class HeuristicParams : public PolymorphicBase {
     return "Undefined Heuristic Params";
   }
 
-  virtual size_t hash() const = 0;
-
-  virtual bool sameAs(const std::shared_ptr<HeuristicParams>& other) const = 0;
+  virtual size_t hash() const {
+    return 0;
+  };
+
+  virtual bool sameAs(const std::shared_ptr<HeuristicParams>& other) const {
+    if (!other->isStrictlyA<HeuristicParams>()) {
+      return false;
+    }
+    return other->cparams == cparams;
+  }
 
-  virtual std::shared_ptr<HeuristicParams> clone() const = 0;
+  virtual std::shared_ptr<HeuristicParams> clone() const {
+    return std::make_shared<HeuristicParams>();
+  }
 
   HeuristicParams() = default;
   HeuristicParams(std::string tag, PrimDataType index_type)

diff --git a/csrc/scheduler/heuristic_types.cpp b/csrc/scheduler/heuristic_types.cpp
@@ -29,6 +29,8 @@ std::string toString(ScheduleHeuristic sh) {
       return "transpose";
     case ScheduleHeuristic::Matmul:
       return "matmul";
+    case ScheduleHeuristic::ExprEval:
+      return "expr_eval";
     case ScheduleHeuristic::None:
       return "none";
     default:

diff --git a/csrc/scheduler/heuristic_types.h b/csrc/scheduler/heuristic_types.h
@@ -55,11 +55,13 @@ enum class ScheduleHeuristic {
   InnerPersistent,
   InnerOuterPersistent,
   OuterPersistent,
-  Transpose
+  Transpose,
+  ExprEval
 };
 
 //! Define a schedule table to loop over all the heuristics in priority order.
-constexpr std::array<ScheduleHeuristic, 8> all_heuristics_in_priority_order = {
+constexpr std::array<ScheduleHeuristic, 9> all_heuristics_in_priority_order = {
+    ScheduleHeuristic::ExprEval,
     ScheduleHeuristic::NoOp,
     ScheduleHeuristic::Matmul,
     ScheduleHeuristic::Reduction,