[TIR][PASS] dtype rewrite for indexing variables (#5092)

include/tvm/arith/analyzer.h

@@ -115,6 +115,15 @@ class ConstIntBoundAnalyzer {
   ConstIntBound operator()(const PrimExpr& expr);
 
   /*!
+   * \brief analyze the expr with the intermediate memorized to avoid redundant computation
+   * \param expr The expression of interest.
+   * \param bound The lookup table to store the intermediate results
+   * \return the result of the analysis.
+   */
+  ConstIntBound operator()(const PrimExpr& expr,
+                           std::unordered_map<const PrimExprNode*, ConstIntBound>* bound);
+
+  /*!
    * \brief Update constant int bound information of var.
    *
    * \param var The variable of interest.

include/tvm/tir/ir_pass.h

@@ -358,6 +358,15 @@ Stmt DecorateDeviceScope(Stmt stmt);
 Stmt HoistIfThenElse(Stmt stmt);
 
 /*!
+ * \brief Narrow down PrimExpr datatype in stmt to target_bits.
+ * \note  Run this pass after StorageFlatten.
+ * \param stmt The stmt to do datatype rewrite
+ * \param target_bits the bit of target datatype
+ * \return Transformed stmt.
+ */
+Stmt NarrowDataType(Stmt stmt, int target_bits);
+
+/*!
  * \brief Make an user callable API LoweredFunc.
  *
  *  The main task of this function is to create code to :

include/tvm/tir/transform.h

@@ -87,6 +87,16 @@ TVM_DLL Pass LowerDeviceStorageAccessInfo();
  */
 TVM_DLL Pass LowerWarpMemory();
 
+
+/*!
+ * \brief Narrow down PrimExpr datatype in stmt to target_bits.
+ *
+ * \note Run this pass after StorageFlatten.
+ *
+ * \return The pass.
+ */
+TVM_DLL Pass NarrowDataType();
+
 }  // namespace transform
 }  // namespace tir
 }  // namespace tvm

python/tvm/driver/build_module.py

@@ -159,6 +159,7 @@ def lower(sch,
     # Phase 1
     stmt = ir_pass.RewriteForTensorCore(stmt, sch, binds)
     stmt = ir_pass.StorageFlatten(stmt, binds, 64, cfg.instrument_bound_checkers)
+    stmt = ir_pass.NarrowDataType(stmt, 32)
     stmt = ir_pass.CanonicalSimplify(stmt)
     for f in lower_phase1:
         stmt = f(stmt)

python/tvm/tir/expr.py

@@ -370,7 +370,8 @@ class IterVar(Object, ExprOp):
                 raise TypeError("dom need to be Range")
 
         name = var if var is not None else "iter"
-        var = Var(name, dtype="int32") if not isinstance(var, Var) else var
+        dtype = "int32" if dom is None else dom.extent.dtype
+        var = Var(name, dtype=dtype) if not isinstance(var, Var) else var
         self.__init_handle_by_constructor__(
             _ffi_api.IterVar, dom, var, iter_type, thread_tag)
 

python/tvm/tir/ir_builder.py

@@ -76,7 +76,8 @@ class BufferVar(ObjectGeneric):
     def __getitem__(self, index):
         t = DataType(self._content_type)
         if t.lanes > 1:
-            index = _expr.Ramp(index * t.lanes, 1, t.lanes)
+            base = index * t.lanes
+            index = _expr.Ramp(base, const(1, base.dtype), t.lanes)
         return _expr.Load(self._content_type, self._buffer_var, index)
 
     def __setitem__(self, index, value):
@@ -87,7 +88,8 @@ class BufferVar(ObjectGeneric):
                     value.dtype, self._content_type))
         t = DataType(self._content_type)
         if t.lanes > 1:
-            index = _expr.Ramp(index * t.lanes, 1, t.lanes)
+            base = index * t.lanes
+            index = _expr.Ramp(base, const(1, base.dtype), t.lanes)
         self._builder.emit(_stmt.Store(self._buffer_var, value, index))
 
 

python/tvm/tir/transform/transform.py

@@ -66,3 +66,18 @@ def LowerWarpMemory():
         The result pass
     """
     return _ffi_api.LowerWarpMemory()
+
+
+def NarrowDataType():
+    """Narrow down PrimExpr datatype in stmt to target_bits.
+
+    Returns
+    -------
+    fpass : tvm.ir.transform.Pass
+        The result pass
+
+    Note
+    ----
+    Run this pass after StorageFlatten.
+    """
+    return _ffi_api.NarrowDataType()

src/arith/const_int_bound.cc

@@ -146,9 +146,30 @@ class ConstIntBoundAnalyzer::Impl :
         res = Intersect(res, info.bound);
       }
     }
+    if (bound_) {
+      const PrimExprNode* op = expr.as<PrimExprNode>();
+      auto val = bound_->find(op);
+      if (val != bound_->end()) {
+        CHECK(val->second->min_value == res.min_value &&
+              val->second->max_value == res.max_value)
+          << "Detected bound for " << expr
+          << "conflicts with memorization";
+      }
+      (*bound_)[op] = ConstIntBound(res.min_value, res.max_value);
+    }
     return res;
   }
 
+  Entry VisitExpr_(const RampNode* op) final {
+    // op = {base + i * stride | 0 <= i < lanes}
+    // Entry(op) = Union(Entry(base + i * stride) | 0 <= i < lanes)
+    // Note that `base + i * stride` is linear w.r.t. `i`
+    // Entry(op) = Union(Entry(base + i * stride) | i = 0, i = lanes-1)
+    Entry a = VisitExpr(op->base);
+    Entry b = VisitExpr(op->base + (op->lanes - 1) * op->stride);
+    return Union(a, b);
+  }
+
   Entry VisitExpr_(const CastNode* op) final {
     Entry a = VisitExpr(op->value);
     Entry b = Everything(op->dtype);
@@ -340,10 +361,13 @@ class ConstIntBoundAnalyzer::Impl :
   }
 
  private:
+  friend class ConstIntBoundAnalyzer;
   // internal variable map
   std::unordered_map<Var, Entry, ObjectHash, ObjectEqual> var_map_;
   // additional bound info
   std::vector<BoundInfo> additional_info_;
+  // look up table for memorization
+  std::unordered_map<const PrimExprNode*, ConstIntBound>* bound_{nullptr};
   // constants: the limit value means umlimited
   // NOTE: kNegInf/kPosInf are used to represent infinity.
   static const constexpr int64_t kNegInf = ConstIntBound::kNegInf;
@@ -536,6 +560,14 @@ ConstIntBound ConstIntBoundAnalyzer::operator()(const PrimExpr& expr) {
   return ConstIntBound(ret.min_value, ret.max_value);
 }
 
+ConstIntBound ConstIntBoundAnalyzer::operator()(const PrimExpr& expr,
+  std::unordered_map<const PrimExprNode*, ConstIntBound>* bound) {
+  impl_->bound_ = bound;
+  Entry ret = impl_->VisitExpr(expr);
+  impl_->bound_ = nullptr;
+  return ConstIntBound(ret.min_value, ret.max_value);
+}
+
 void ConstIntBoundAnalyzer::Update(const Var& var,
                                    const ConstIntBound& info,
                                    bool override) {

src/target/llvm/codegen_cpu.cc

@@ -943,7 +943,7 @@ void CodeGenCPU::VisitStmt_(const ForNode* op) {
         PrimExpr end = MinNode::make((task_id + make_const(t, 1)) * step, op->extent);
         CreateSerialFor(MakeValue(begin),
                         MakeValue(end),
-                        ConstInt32(1),
+                        llvm::ConstantInt::getSigned(GetLLVMType(end), 1),
                         op->loop_var,
                         op->body);
       }

src/target/llvm/codegen_llvm.cc

@@ -1121,7 +1121,8 @@ void CodeGenLLVM::VisitStmt_(const ForNode* op) {
     CHECK(op->for_type == ForType::Serial);
   }
   CreateSerialFor(MakeValue(op->min), MakeValue(op->extent),
-                  ConstInt32(1), op->loop_var, op->body);
+                  llvm::ConstantInt::getSigned(GetLLVMType(op->extent), 1),
+                  op->loop_var, op->body);
 }
 
 

src/tir/ir/buffer.cc

@@ -452,7 +452,7 @@ Buffer BufferNode::make(Var data,
   n->buffer_type = buffer_type;
   if (n->buffer_type == kAutoBroadcast && n->shape.size() > 0 && n->strides.empty()) {
     for (size_t i = 0; i < n->shape.size(); ++i) {
-      n->strides.push_back(Var("stride"));
+      n->strides.push_back(Var("stride", n->shape[i].dtype()));
     }
   }
   return Buffer(n);

src/tir/pass/ffi_api.cc

@@ -156,5 +156,6 @@ REGISTER_PASS(InstrumentBoundCheckers);
 REGISTER_PASS(VerifyCompactBuffer);
 REGISTER_PASS(HoistIfThenElse);
 REGISTER_PASS(InferFragment)
+REGISTER_PASS(NarrowDataType);
 }  // namespace tir
 }  // namespace tvm

src/tir/pass/loop_partition.cc

@@ -587,7 +587,7 @@ inline Stmt LoopPartitioner::MakeFor(const Object *node, PrimExpr extent, Stmt b
     // If the loop extent is 1, do not create the loop anymore
     return Substitute(body, {{Var{for_node->loop_var}, make_const(DataType::Int(32), 0)}});
   } else {
-    return ForNode::make(for_node->loop_var, 0, extent,
+    return ForNode::make(for_node->loop_var, IntImm(for_node->min.dtype(), 0), extent,
                      for_node->for_type, for_node->device_api, body);
   }
 }

src/tir/pass/unroll_loop.cc

@@ -160,7 +160,9 @@ class LoopUnroller : public StmtExprMutator {
     PrimExpr extent = tir::Simplify(op->extent);
     const IntImmNode  *v1 = extent.as<IntImmNode>();
     int value = -1;
-    if (v1 != nullptr) {
+    // integers that do not fit in int32_t are treated as symbolic,
+    // as it's impossible to unroll such large loops
+    if (v1 != nullptr && v1->value <= std::numeric_limits<int>::max()) {
       value = static_cast<int>(v1->value);
     }
     return value;

tests/python/unittest/test_tir_transform_narrow_datatype.py

+# Licensed to the Apache Software Foundation (ASF) under one
+# or more contributor license agreements.  See the NOTICE file
+# distributed with this work for additional information
+# regarding copyright ownership.  The ASF licenses this file
+# to you under the Apache License, Version 2.0 (the
+# "License"); you may not use this file except in compliance
+# with the License.  You may obtain a copy of the License at
+#
+#   http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an
+# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+# KIND, either express or implied.  See the License for the
+# specific language governing permissions and limitations
+# under the License.
+import tvm
+from tvm import te
+from tvm.tir import const
+
+
+def lower_stmt(params, stmt, target_bits):
+    func = tvm.tir.PrimFunc(params, stmt).with_attr(
+        "target_bits", target_bits)
+    func = tvm.tir.transform.NarrowDataType()(tvm.IRModule.from_expr(func))["main"]
+    stmt = func.body
+    return stmt
+
+
+def lower_sch(sch, args, target_bits):
+    binds = {}
+    arg_list = []
+    for x in args:
+        if isinstance(x, te.tensor.Tensor):
+            buf = tvm.tir.decl_buffer(x.shape, dtype=x.dtype, name=x.name)
+            assert x not in binds
+            binds[x] = buf
+            arg_list.append(buf)
+        else:
+            raise ValueError("args must be Tensor, Buffer or Var")
+    bounds = te.schedule.InferBound(sch)
+    stmt = te.schedule.ScheduleOps(sch, bounds)
+    stmt = tvm.tir.ir_pass.StorageFlatten(stmt, binds, 64, False)
+    return lower_stmt(arg_list, stmt, target_bits)
+
+
+def test_basic():
+    def check(m, n, target_bits, target_dtype):
+        ib = tvm.tir.ir_builder.create()
+        Ab = tvm.tir.decl_buffer((m, n), name='A')
+        A = ib.buffer_ptr(Ab)
+        Bb = tvm.tir.decl_buffer((m, n), name='B')
+        B = ib.buffer_ptr(Bb)
+        with ib.for_range(0, m, name='i') as i:
+            with ib.for_range(0, n, name='j') as j:
+                B[i * n + j] = A[i * n + j] + 1
+        stmt = ib.get()
+        stmt = lower_stmt([Ab, Bb], stmt, target_bits)
+        assert stmt.loop_var.dtype == target_dtype
+        assert stmt.body.loop_var.dtype == target_dtype
+
+    # const shape
+    # i32 -> i32
+    check(2, 2, 32, "int32")
+    check(2**16, 2**16, 32, "int32")  # i32 + i32 is not promoted to i64 even if overflow
+    # i64 -> i32
+    check(const(2, dtype='int64'), const(2, dtype='int64'), 32, "int32")
+    check(const(2**16, dtype='int64'), const(2**16, dtype='int64'), 32, "int64")
+    # i32 -> i16
+    check(2, 2, 16, "int16")
+    check(2**10, 2**10, 16, "int32")
+
+    # symbolic shape
+    check(te.size_var(name='m', dtype='int32'), te.size_var(name='n', dtype='int32'), 32, "int32")
+    check(te.size_var(name='m', dtype='int64'), te.size_var(name='n', dtype='int64'), 32, "int64")
+
+
+def test_thread_axis():
+    def check(m, n, target_bits, target_dtype):
+        ib = tvm.tir.ir_builder.create()
+        Ab = tvm.tir.decl_buffer((m, n), name='A')
+        A = ib.buffer_ptr(Ab)
+        Bb = tvm.tir.decl_buffer((m, n), name='B')
+        B = ib.buffer_ptr(Bb)
+        bx = te.thread_axis("blockIdx.x")
+        tx = te.thread_axis("threadIdx.x")
+        ib.scope_attr(bx, "thread_extent", m)
+        ib.scope_attr(tx, "thread_extent", n)
+        B[bx * n + tx] = A[bx * n + tx] + 1
+        stmt = ib.get()
+        stmt = lower_stmt([Ab, Bb], stmt, target_bits)
+        assert stmt.node.var.dtype == target_dtype
+        assert stmt.body.node.var.dtype == target_dtype
+
+    # i32 -> i32
+    check(2, 32,
+          target_bits=32, target_dtype='int32')
+    check(2**30, 32,  # i32 + i32 is not promoted to i64 even in the case of overflow
+          target_bits=32, target_dtype='int32')
+    # i64 -> i32
+    check(const(2, dtype='int64'),
+          const(32, dtype='int64'),
+          target_bits=32, target_dtype='int32')
+    check(const(2**30, dtype='int64'),
+          const(32, dtype='int64'),
+          target_bits=32, target_dtype='int64')
+    # i32 -> i16
+    check(2, 32,
+          target_bits=16, target_dtype='int16')
+    check(2**14, 32,
+          target_bits=16, target_dtype='int32')
+
+
+def test_multilanes():
+    def check(m, lanes, target_bits, target_dtype):
+        ib = tvm.tir.ir_builder.create()
+        Ab = tvm.tir.decl_buffer((m,), dtype='float32x{}'.format(lanes), name='A')
+        A = ib.buffer_ptr(Ab)
+        Bb = tvm.tir.decl_buffer((m,), dtype='float32x{}'.format(lanes), name='B')
+        B = ib.buffer_ptr(Bb)
+        with ib.for_range(0, m, name='i', dtype=m.dtype) as i:
+            B[i] = A[i] + 1
+        stmt = ib.get()
+        stmt = lower_stmt([Ab, Bb], stmt, target_bits)
+        assert stmt.loop_var.dtype == target_dtype
+
+    # i32 -> i32
+    check(const(2 ** 10, dtype='int32'), 2,
+          target_bits=32, target_dtype='int32')
+    check(const(2 ** 32, dtype='int32'), 2,
+          target_bits=32, target_dtype='int32')
+    # i64 -> i32
+    check(const(2 ** 10, dtype='int64'), 2,
+          target_bits=32, target_dtype='int32')
+    check(const(2 ** 32, dtype='int64'), 2,
+          target_bits=32, target_dtype='int64')
+    # i32 -> i16
+    check(const(2 ** 10, dtype='int32'), 2,
+          target_bits=16, target_dtype='int16')
+    check(const(2 ** 16, dtype='int32'), 2,
+          target_bits=16, target_dtype='int32')
+
+
+def test_reduce():
+    def check(m, target_bits, target_dtype):
+        A = te.placeholder((m,), name='A', dtype='float32')
+        k = te.reduce_axis((0, m), "k")
+        B = te.compute((), lambda *idx: te.sum(A[k], axis=k), name='B')
+        s = te.create_schedule(B.op)
+        stmt = lower_sch(s, [A, B], target_bits)
+        assert stmt.body[1].loop_var.dtype == target_dtype
+
+    # i32 -> i32
+    check(const(64, dtype='int32'), 32, 'int32')
+    # i64 -> i32
+    check(const(64, dtype='int64'), 32, 'int32')
+    # i32 -> i16
+    check(const(64, dtype='int32'), 16, 'int16')
+    check(const(2**16, dtype='int32'), 16, 'int32')
+    # symbolic
+    check(te.var('n', dtype='int32'), 32, 'int32')
+    check(te.var('n', dtype='int64'), 32, 'int64')
+
+
+def test_slice():
+    def check(m, n, target_bits, target_dtype):
+        # The index may overflow in B, while not in A
+        ib = tvm.tir.ir_builder.create()
+        Ab = tvm.tir.decl_buffer((m, n), name='A')
+        A = ib.buffer_ptr(Ab)
+        Bb = tvm.tir.decl_buffer((m, n * 2), name='B')
+        B = ib.buffer_ptr(Bb)
+        with ib.for_range(0, m, name='i') as i:
+            with ib.for_range(0, n, name='j') as j:
+                A[i * n + j] = B[i * 2 * n + 2 * j] + 1
+        stmt = ib.get()
+        stmt = lower_stmt([Ab, Bb], stmt, target_bits)
+        assert stmt.loop_var.dtype == target_dtype
+        assert stmt.body.loop_var.dtype == target_dtype
+
+    # The maximum index is (2**15 * 2**15 - 1) * 2 <= 2**31 - 1
+    check(const(2**15, 'int64'), const(2**15, 'int64'),
+          target_bits=32, target_dtype='int32')
+    # The maximum index is (2**15 * 2**15 - 1 + 2**15) * 2 > 2**31 - 1
+    check(const(2**15, 'int64'), const((2**15 + 1), 'int64'),
+          target_bits=32, target_dtype='int64')
+
+
+if __name__ == "__main__":
+    test_basic()
+    test_thread_axis()
+    test_multilanes()
+    test_reduce()
+    test_slice()