[PASS] copy intrin (#536)

* [PASS] copy intrin

* update comment thanks to derisavi

include/tvm/arithmetic.h

@@ -159,8 +159,8 @@ struct IntSetNode : public Node {
 };
 
 /*!
- * \brief Detect if e can be rewritten as e = sum_{i=0}^n var[i] * coeff[i] + coeff[n]
- *  Where coeff and base are invariant of var.
+ * \brief Detect if e can be rewritten as e = sum_{i=0}^{n-1} var[i] * coeff[i] + coeff[n]
+ *  Where coeff[i] and base are invariant of var[j] for all i and j.
  *
  * \param e The expression to be detected.
  * \param vars List of variables to be used in detection.

include/tvm/buffer.h

@@ -128,7 +128,7 @@ class BufferNode : public Node {
                              Type dtype,
                              Array<Expr> shape,
                              Array<Expr> strides,
-                             Expr byte_offset,
+                             Expr elem_offset,
                              std::string name,
                              std::string scope,
                              int data_alignment,

include/tvm/ir_pass.h

@@ -240,6 +240,24 @@ Stmt InjectPrefetch(Stmt stmt);
 Stmt InjectDoubleBuffer(Stmt stmt, int split_loop);
 
 /*!
+ * \brief Inject copy intrinsics with optional pad.
+ *
+ * \param stmt The statment to be transformed.
+ * \param pragma_key The pragma key for hint of copy.
+ * \param fintrin The function with signature
+ *
+ *   Stmt fintrin(Buffer src,
+ *                Buffer dst,
+ *                Array<Expr> pad_before,
+ *                Array<Expr> pad_after,
+ *                Expr pad_value)
+ * \return Transformed stmt.
+ */
+Stmt InjectCopyIntrin(Stmt stmt,
+                      const std::string& pragma_key,
+                      const runtime::PackedFunc& fintrin);
+
+/*!
  * \brief Rewrite storage allocation pattern.
  *  Moves the allocation to outer most possible scope.
  *  Trying to share space between allocations to make

include/tvm/packed_func_ext.h

@@ -171,8 +171,12 @@ inline TNodeRef TVMRetValue::AsNodeRef() const {
 }
 
 inline void TVMArgsSetter::operator()(size_t i, const NodeRef& other) const {  // NOLINT(*)
+  if (other.defined()) {
     values_[i].v_handle = const_cast<std::shared_ptr<Node>*>(&(other.node_));
     type_codes_[i] = kNodeHandle;
+  } else {
+    type_codes_[i] = kNull;
+  }
 }
 
 // type related stuffs

src/api/api_pass.cc

@@ -92,6 +92,7 @@ REGISTER_PASS3(StorageFlatten);
 REGISTER_PASS4(IRTransform);
 REGISTER_PASS1(VectorizeLoop);
 REGISTER_PASS4(UnrollLoop);
+REGISTER_PASS3(InjectCopyIntrin);
 REGISTER_PASS2(ThreadSync);
 REGISTER_PASS5(MakeAPI);
 REGISTER_PASS2(BindDeviceType);

src/arithmetic/canonical.cc

@@ -307,7 +307,13 @@ class Canonical::Internal : public IRMutator {
     if (!op->is_pure()) {
       stack_.back().has_side_effect = true;
     }
-    return IRMutator::Mutate_(op, e);
+    Expr expr = IRMutator::Mutate_(op, e);
+    op = expr.as<Call>();
+    if (op->is_intrinsic(Call::likely) && is_const(op->args[0])) {
+      return op->args[0];
+    } else {
+      return expr;
+    }
   }
   // For
   Stmt Mutate_(const For* op, const Stmt& s) {
@@ -320,6 +326,13 @@ class Canonical::Internal : public IRMutator {
     --level_counter_;
     return stmt;
   }
+  // IfThenElse
+  Stmt Mutate_(const IfThenElse* op, const Stmt& s) {
+    Stmt stmt  = IRMutator::Mutate_(op, s);
+    op = stmt.as<IfThenElse>();
+    if (is_one(op->condition)) return op->then_case;
+    return stmt;
+  }
   // AttrStmt
   Stmt Mutate_(const AttrStmt* op, const Stmt& s) {
     if (op->attr_key == attr::thread_extent ||

src/pass/inject_copy_intrin.cc

+/*!
+ *  Copyright (c) 2017 by Contributors
+ * \brief Replace certain copy with copy intrinsics.
+ * \file copy_intrin_rewrite.cc
+ */
+#include <tvm/ir.h>
+#include <tvm/packed_func_ext.h>
+#include <tvm/ir_mutator.h>
+#include <tvm/ir_pass.h>
+
+namespace tvm {
+namespace ir {
+
+using runtime::PackedFunc;
+
+class CopyIntrinInjector : public IRMutator {
+ public:
+  CopyIntrinInjector(const std::string& pragma_key,
+                     const PackedFunc& flower_copy_fromto)
+      : pragma_key_(pragma_key),
+        flower_copy_fromto_(flower_copy_fromto) {
+  }
+
+  Stmt Mutate_(const AttrStmt* op, const Stmt& s) final {
+    if (op->attr_key == attr::storage_scope) {
+      const Variable* buf = op->node.as<Variable>();
+      storage_scope_[buf] = op->value.as<StringImm>()->value;
+    } else if (op->attr_key == ir::attr::pragma_scope) {
+      const std::string& pname = op->value.as<StringImm>()->value;
+      if (pname == pragma_key_) {
+        Stmt ret;
+        CHECK(MatchCopyPattern(op->body, &ret))
+            << "Cannot match copy pattern of " << op->body;
+        return ret;
+      }
+    }
+    return IRMutator::Mutate_(op, s);
+  }
+
+ private:
+  bool MatchCopyPattern(Stmt stmt, Stmt *out) {
+    Stmt body = stmt;
+
+    // strip the loops
+    std::vector<const For*> loops;
+    while (const For* op = body.as<For>()) {
+      if (!is_zero(op->min)) return false;
+      loops.push_back(op);
+      body = op->body;
+    }
+    const Store* store = body.as<Store>();
+    if (store == nullptr) return false;
+    const Select* select = store->value.as<Select>();
+    const Load* load = store->value.as<Load>();
+
+    // for now only support true condition matching
+    if (select != nullptr) {
+      load = select->true_value.as<Load>();
+    }
+    if (load == nullptr) return false;
+    if (load->type.lanes() != 1) return false;
+    Array<Var> loop_vars;
+    for (const For* op : loops) {
+      loop_vars.push_back(Var(op->loop_var.node_));
+    }
+    Array<Expr> store_strides =
+        arith::DetectLinearEquation(store->index, loop_vars);
+    Array<Expr> load_strides =
+        arith::DetectLinearEquation(load->index, loop_vars);
+    if (load_strides.size()  == 0 || store_strides.size() == 0) return false;
+    Array<Expr> dst_shape;
+    for (const For* op : loops) {
+      dst_shape.push_back(op->extent);
+    }
+    Array<Expr> src_shape = dst_shape;
+    Array<Expr> pad_before, pad_after;
+    Expr pad_value;
+    Expr src_elem_offset = load_strides[loop_vars.size()];
+    if (select != nullptr) {
+      Array<Expr> clip_bound =
+          arith::DetectClipBound(select->condition, loop_vars);
+      pad_value = select->false_value;
+      if (clip_bound.size() == 0) return false;
+      CHECK_EQ(src_shape.size(), loop_vars.size());
+      CHECK_EQ(clip_bound.size(), loop_vars.size() * 2);
+      for (size_t i = 0; i < src_shape.size(); ++i) {
+        Expr min_value = clip_bound[2 * i];
+        Expr max_value = clip_bound[2 * i + 1];
+        Type t = loop_vars[i].type();
+        Expr svalue = src_shape[i];
+        if (min_value.defined()) {
+          Expr pbefore = Simplify(Max::make(min_value, make_zero(t)));
+          src_elem_offset = src_elem_offset + pbefore * load_strides[i];
+          svalue = svalue - pbefore;
+          pad_before.push_back(pbefore);
+        } else {
+          pad_before.push_back(make_zero(t));
+        }
+        if (max_value.defined()) {
+          Expr pafter = Simplify(Max::make(loops[i]->extent - max_value - make_const(t, 1),
+                                           make_zero(t)));
+          svalue = svalue - pafter;
+          pad_after.push_back(pafter);
+        } else {
+          pad_after.push_back(make_zero(t));
+        }
+        src_shape.Set(i, Simplify(svalue));
+      }
+      src_elem_offset = Simplify(src_elem_offset);
+    }
+    CHECK_EQ(load_strides.size(), store_strides.size());
+    CHECK_EQ(load_strides.size(), loop_vars.size() + 1);
+    Array<Expr> src_strides(load_strides.begin(), load_strides.begin() + loop_vars.size());
+    Array<Expr> dst_strides(store_strides.begin(), store_strides.begin() + loop_vars.size());
+    Buffer dst = BufferNode::make(
+        Var(store->buffer_var.node_),
+        load->type,
+        dst_shape,
+        dst_strides,
+        store_strides[loop_vars.size()],
+        store->buffer_var->name_hint,
+        GetStorageScope(store->buffer_var.get()),
+        0, 0);
+    Buffer src = BufferNode::make(
+        Var(load->buffer_var.node_),
+        load->type,
+        src_shape,
+        src_strides,
+        src_elem_offset,
+        load->buffer_var->name_hint,
+        GetStorageScope(load->buffer_var.get()),
+        0, 0);
+    *out = flower_copy_fromto_(src, dst, pad_before, pad_after, pad_value);
+    CHECK(out->defined()) << "flower function did not return correct stmt";
+    return true;
+  }
+  // Get storage scope
+  std::string GetStorageScope(const Variable* var) const {
+    auto it = storage_scope_.find(var);
+    if (it != storage_scope_.end()) {
+      return it->second;
+    } else {
+      return "";
+    }
+  }
+  // pragma key
+  const std::string& pragma_key_;
+  // function to lower copy intrinsics.
+  const PackedFunc& flower_copy_fromto_;
+  // Storage scope
+  std::unordered_map<const Variable*, std::string> storage_scope_;
+};
+
+Stmt InjectCopyIntrin(Stmt stmt,
+                      const std::string& pragma_key,
+                      const PackedFunc& flower_copy_fromto) {
+  return CopyIntrinInjector(pragma_key, flower_copy_fromto)
+      .Mutate(stmt);
+}
+
+}  // namespace ir
+}  // namespace tvm

tests/python/unittest/test_pass_inject_copy_intrin.py

+import tvm
+
+def test_copy2d():
+    m = tvm.var('m')
+    l = tvm.var('l')
+    A = tvm.placeholder((m, l), name='A')
+    B = tvm.compute((m, l), lambda i, j: A[i, j], name='B')
+    s = tvm.create_schedule(B.op)
+    s[B].pragma(B.op.axis[0], "memcpy")
+    bounds = tvm.schedule.InferBound(s)
+    stmt = tvm.schedule.ScheduleOps(s, bounds)
+    Ab = tvm.decl_buffer(A.shape, A.dtype, name='A')
+    Bb = tvm.decl_buffer(B.shape, B.dtype, name='B')
+    stmt = tvm.ir_pass.StorageFlatten(stmt, {A: Ab, B: Bb}, 64)
+    def cb(src, dst, pad_before, pad_after, pad_value):
+        assert dst.strides[0] == l
+        assert dst.strides[1].value == 1
+        assert src.strides[0] == l
+        assert tuple(src.shape) == (m, l)
+        return tvm.make.Evaluate(0)
+    stmt = tvm.ir_pass.InjectCopyIntrin(stmt, "memcpy", cb)
+
+def test_copy_pad():
+    m = tvm.var('m')
+    l = tvm.var('l')
+    A = tvm.placeholder((m, l), name='A')
+    B = tvm.compute((m + 2, l), lambda i, j:
+                    tvm.select(tvm.all(i >= 1, i < m + 1),
+                                       A[i - 1, j], 1.0), name='B')
+    s = tvm.create_schedule(B.op)
+    s[B].pragma(B.op.axis[0], "memcpy")
+    bounds = tvm.schedule.InferBound(s)
+    stmt = tvm.schedule.ScheduleOps(s, bounds)
+    Ab = tvm.decl_buffer(A.shape, A.dtype, name='A')
+    Bb = tvm.decl_buffer(B.shape, B.dtype, name='B')
+    stmt = tvm.ir_pass.StorageFlatten(stmt, {A: Ab, B: Bb}, 64)
+    def cb(src, dst, pad_before, pad_after, pad_value):
+        assert tvm.ir_pass.Simplify(src.elem_offset).value == 0
+        assert pad_before[0].value == 1
+        assert pad_before[1].value == 0
+        assert pad_after[0].value == 1
+        assert pad_after[1].value == 0
+        assert pad_value.value == 1.0
+        return tvm.make.Evaluate(0)
+    stmt = tvm.ir_pass.InjectCopyIntrin(stmt, "memcpy", cb)
+
+def assert_expr_equal(a, b):
+    assert tvm.ir_pass.Simplify(a - b).value == 0
+
+def test_copy_pad_split():
+    m = 4 * 3
+    A = tvm.placeholder((m, ), name="A")
+    Apad = tvm.compute((m + 2,), lambda i:
+                       tvm.select(tvm.all(i >= 1, i <= m),
+                                  A[i - 1], 0.0), "Apad")
+    B = tvm.compute((m,), lambda i: Apad[i] + Apad[i + 1] + Apad[i + 2])
+    s = tvm.create_schedule(B.op)
+    xo, xi = s[B].split(B.op.axis[0], factor=4)
+    s[Apad].compute_at(s[B], xo)
+    s[Apad].pragma(s[Apad].op.axis[0], "memcpy")
+    bounds = tvm.schedule.InferBound(s)
+    stmt = tvm.schedule.ScheduleOps(s, bounds)
+    Ab = tvm.decl_buffer(A.shape, A.dtype, name='A')
+    Bb = tvm.decl_buffer(B.shape, B.dtype, name='B')
+    stmt = tvm.ir_pass.StorageFlatten(stmt, {A: Ab, B: Bb}, 64)
+    stmt = tvm.ir_pass.CanonicalSimplify(stmt)
+    def cb(src, dst, pad_before, pad_after, pad_value):
+        assert(dst.elem_offset.value == 0)
+        assert_expr_equal(src.elem_offset, tvm.max(xo * 4, 1) - 1)
+        rpad_before = tvm.max(1 - xo * 4, 0)
+        rpad_after = tvm.max(xo * 4 - 7, 0)
+        assert_expr_equal(pad_before[0], rpad_before)
+        assert_expr_equal(pad_after[0], rpad_after)
+        assert_expr_equal(src.shape[0], 6 - rpad_before - rpad_after)
+        return tvm.make.Evaluate(0)
+    stmt = tvm.ir_pass.InjectCopyIntrin(stmt, "memcpy", cb)
+
+
+if __name__ == "__main__":
+    test_copy2d()
+    test_copy_pad()
+    test_copy_pad_split()