[TIR][REFACTOR] RewriteForTensorCore -> te/schedule (apache#5379)

* [TIR][REFACTIR] RewriteForTensorCore -> te/schedule

RewriteForTensor depends on the schedule information, which makes it differ
from a typical pass(which should get all the information from the input TIR).

As a result, we refactor it as a SchedulePostProc step for now.
We should revisit it later as we introduce more support for tensor core patterns in the TIR.

* Fix VTA to fit the new IR Pattern

Author: tqchen

include/tvm/te/schedule_pass.h

@@ -34,6 +34,23 @@
 namespace tvm {
 namespace te {
 
+/*!
+ * \brief To automatically inline the element-wise operations.
+ *
+ * \param sch The schedule to be inlined.
+ */
+void AutoInlineElemWise(Schedule sch);
+
+/*!
+ * \brief To automatically inline operations with injective writes
+ *   (i.e. writes without reduction or sequential loops). Note
+ *   that in this case, guarantees about contiguity, transpose, stride,
+ *   alignemnt and memory footprint in general do not hold.
+ *
+ * \param sch The schedule to be inlined.
+ */
+TVM_DLL void AutoInlineInjective(Schedule sch);
+
 /*!
  * \brief Infer the bound of all iteration variables relates to the schedule.
  *
@@ -55,6 +72,21 @@ Map<IterVar, Range> InferBound(const Schedule& sch);
  */
 Stmt ScheduleOps(Schedule s, Map<IterVar, Range> dom_map, bool debug_keep_trivial_loop);
 
+
+/*!
+ * \brief Try to modify the AST generated by ScheduleOps to support TensorCore.
+ *
+ * \param stmt The stmt to be trasnformed.
+ * \param schedule The original schedule.
+ * \param extern_buffer Map specifies external
+ *    buffer assignment of input and outputs.
+ * \return Transformed stmt.
+ */
+Stmt SchedulePostProcRewriteForTensorCore(
+    Stmt stmt,
+    Schedule schedule,
+    Map<Tensor, Buffer> extern_buffer);
+
 /*!
  * \brief Postprocessing the Stmt generated by ScheduleOps to create
  *        a PrimFunc that can then be used for further TIR optimizations.
@@ -75,23 +107,6 @@ PrimFunc SchedulePostProcToPrimFunc(Array<ObjectRef> arg_list,
                                     Stmt body,
                                     Optional<Map<Tensor, Buffer>> bindings);
 
-/*!
- * \brief To automatically inline the element-wise operations.
- *
- * \param sch The schedule to be inlined.
- */
-void AutoInlineElemWise(Schedule sch);
-
-/*!
- * \brief To automatically inline operations with injective writes
- *   (i.e. writes without reduction or sequential loops). Note
- *   that in this case, guarantees about contiguity, transpose, stride,
- *   alignemnt and memory footprint in general do not hold.
- *
- * \param sch The schedule to be inlined.
- */
-TVM_DLL void AutoInlineInjective(Schedule sch);
-
 }  // namespace te
 }  // namespace tvm
 #endif  // TVM_TE_SCHEDULE_PASS_H_

include/tvm/tir/ir_pass.h

@@ -164,19 +164,6 @@ Stmt Inline(Stmt stmt,
             Array<Var> args,
             PrimExpr body);
 
-/*!
- * \brief Try to modify the AST to support TensorCore
- *
- * \param stmt The stmt to be trasnformed.
- * \param schedule The original schedule.
- * \param extern_buffer Map specifies external
- *    buffer assignment of input and outputs.
- * \return Transformed stmt.
- */
-Stmt RewriteForTensorCore(Stmt stmt,
-                          te::Schedule schedule,
-                          Map<te::Tensor, Buffer> extern_buffer);
-
 /*!
  * \brief Verify if there is any argument bound to compact buffer.
  *

python/tvm/driver/build_module.py

@@ -84,23 +84,43 @@ def get_binds(args, compact=False, binds=None):
     return binds, arg_list
 
 
-def form_body(sch):
+def form_irmodule(sch, args, name, binds):
     """According to the given schedule, form a function.
 
     Parameters
     ----------
     sch : tvm.te.schedule.Schedule
-    The given scheduler to form the raw body
+        The given scheduler to form the raw body
+
+    args : list of Buffer or Tensor or Var
+        The argument lists to the function.
+
+    name : str
+        The name of result function.
+
+    binds : dict of :any:`Tensor` to :any:`Buffer`, optional
+        The binds information
 
     Returns
     -------
     The body formed according to the given schedule
     """
     # normalize schedule first
+    cfg = BuildConfig.current()
     sch = sch.normalize()
     bounds = schedule.InferBound(sch)
     stmt = schedule.ScheduleOps(sch, bounds)
-    return stmt
+
+    compact = ir_pass.VerifyCompactBuffer(stmt)
+    binds, arg_list = get_binds(args, compact, binds)
+
+    stmt = schedule.SchedulePostProcRewriteForTensorCore(stmt, sch, binds)
+    func = schedule.SchedulePostProcToPrimFunc(arg_list, stmt, binds)
+
+    func = func.with_attr("global_symbol", name)
+    if cfg.restricted_func:
+        func = func.with_attr("tir.noalias", True)
+    return tvm.IRModule({name: func})
 
 
 def _wrap_as_prim_func_pass(flist, name):
@@ -166,24 +186,13 @@ def lower(sch,
 
     # Phase 0
     if isinstance(sch, schedule.Schedule):
-        stmt = form_body(sch)
-
-    for f in lower_phase0:
-        stmt = f(stmt)
-
-    compact = ir_pass.VerifyCompactBuffer(stmt)
-    binds, arg_list = get_binds(args, compact, binds)
-    stmt = ir_pass.RewriteForTensorCore(stmt, sch, binds)
-
-    # Start the new style pass manager.
-    func = schedule.SchedulePostProcToPrimFunc(arg_list, stmt, binds)
-    func = func.with_attr("global_symbol", name)
-    if cfg.restricted_func:
-        func = func.with_attr("tir.noalias", True)
-    mod = tvm.IRModule({name: func})
+        mod = form_irmodule(sch, args, name, binds)
+    else:
+        mod = sch
 
     # Phase 1
     pass_list = [
+        _wrap_as_prim_func_pass(lower_phase0, "Custom-Phase0"),
         tvm.tir.transform.InjectPrefetch(),
         tvm.tir.transform.StorageFlatten(64, cfg.instrument_bound_checkers),
         tvm.tir.transform.NarrowDataType(32),

python/tvm/te/hybrid/__init__.py

@@ -30,7 +30,7 @@
 # 2. Support multi-level HalideIR
 import inspect
 import tvm._ffi
-from tvm.driver.build_module import form_body
+import tvm.te.schedule
 from tvm._ffi.base import decorate
 
 from .module import HybridModule
@@ -87,8 +87,10 @@ def build(sch, inputs, outputs, name="hybrid_func"):
         The built results is wrapped in a HybridModule.
         The usage of HybridModule is roughly the same as normal TVM-built modules.
     """
+    sch = sch.normalize()
+    bounds = tvm.te.schedule.InferBound(sch)
+    stmt = tvm.te.schedule.ScheduleOps(sch, bounds)
 
-    stmt = form_body(sch)
     src = _Dump(stmt, inputs, outputs, name)
 
     return HybridModule(src, name)

src/tir/pass/tensor_core.cc renamed to src/te/schedule/schedule_postproc_rewrite_for_tensor_core.cc

@@ -18,9 +18,12 @@
  */
 
 /*!
- * \file tensor_core.cc
+ * \file schedule_postproc_rewrite_for_tensor_core.cc
+ *
+ * \brief Rewrite the Stmt generated by ScheduleOps
+ *        to accomondate tensorcore.
  */
-// IR Passes for TensorCore CodeGen
+#include <tvm/runtime/registry.h>
 #include <tvm/tir/expr.h>
 #include <tvm/tir/stmt.h>
 #include <tvm/te/operation.h>
@@ -32,12 +35,11 @@
 #include <tvm/target/target.h>
 #include <tvm/runtime/device_api.h>
 #include <unordered_map>
-#include "ir_util.h"
 #include "../../arith/compute_expr.h"
 #include "../../runtime/thread_storage_scope.h"
 
 namespace tvm {
-namespace tir {
+namespace te {
 
 using namespace te;
 using runtime::StorageRank;
@@ -86,10 +88,10 @@ class MMAMatcher: public StmtVisitor {
   }
 
   void VisitStmt_(const AttrStmtNode* op) final {
-    if (op->attr_key == attr::pragma_tensor_core) {
+    if (op->attr_key == tir::attr::pragma_tensor_core) {
       tensor_core_on_ = true;
       StmtVisitor::VisitStmt_(op);
-    } else if (op->attr_key == attr::realize_scope) {
+    } else if (op->attr_key == tir::attr::realize_scope) {
       storage_scope_[op->node.get()] = op->value.as<StringImmNode>()->value;
       this->VisitStmt(op->body);
     } else {
@@ -414,18 +416,18 @@ class BufferAnalyser : public StmtExprVisitor {
   }
 
   void VisitStmt_(const AttrStmtNode* op) final {
-    if (op->attr_key == attr::thread_extent) {
+    if (op->attr_key == tir::attr::thread_extent) {
       if (const IntImmNode* value = op->value.as<IntImmNode>()) {
         thread_extent_.insert(
             std::make_pair(
                 op->node.as<IterVarNode>()->var->name_hint,
                 value->value));
       }
       StmtExprVisitor::VisitStmt_(op);
-    } else if (op->attr_key == attr::realize_scope) {
+    } else if (op->attr_key == tir::attr::realize_scope) {
       storage_scope_[op->node.get()] = op->value.as<StringImmNode>()->value;
       this->VisitStmt(op->body);
-    } else if (op->attr_key == attr::buffer_dim_align) {
+    } else if (op->attr_key == tir::attr::buffer_dim_align) {
       te::Tensor tensor = Downcast<te::Tensor>(op->node);
       const CallNode* tuple = op->value.as<CallNode>();
       CHECK(tuple && tuple->is_intrinsic(intrinsic::tvm_tuple));
@@ -850,7 +852,7 @@ class TensorCoreIRMutator : public StmtExprMutator {
 
   Stmt VisitStmt_(const AttrStmtNode* op) final {
     Stmt stmt = StmtExprMutator::VisitStmt_(op);
-    if (op->attr_key == attr::realize_scope) {
+    if (op->attr_key == tir::attr::realize_scope) {
       auto node = op->node.as<te::OperationNode>();
       if (node != nullptr) {
         if (!frag_reg_.count(node->name)) {
@@ -1186,9 +1188,10 @@ class TensorCoreIRMutator : public StmtExprMutator {
   int warp_threads_y_{-1};
 };
 
-Stmt RewriteForTensorCore(Stmt stmt,
-                          Schedule schedule,
-                          Map<Tensor, Buffer> extern_buffer) {
+Stmt SchedulePostProcRewriteForTensorCore(
+    Stmt stmt,
+    Schedule schedule,
+    Map<Tensor, Buffer> extern_buffer) {
   // Check if current lower target is CUDA
   auto target = tvm::Target::Current(true);
   if (target.defined() && target->target_name != "cuda") {
@@ -1223,5 +1226,13 @@ Stmt RewriteForTensorCore(Stmt stmt,
   return TensorCoreIRMutator(schedule_analyser, buffer_analyser)(std::move(stmt));
 }
 
-}  // namespace tir
+TVM_REGISTER_GLOBAL("schedule.SchedulePostProcRewriteForTensorCore")
+.set_body_typed([](Stmt stmt,
+                   Schedule schedule,
+                   Map<te::Tensor, Buffer> extern_buffer) {
+  return SchedulePostProcRewriteForTensorCore(
+      stmt, schedule, extern_buffer);
+});
+
+}  // namespace te
 }  // namespace tvm

src/tir/pass/ffi_api.cc

@@ -75,14 +75,6 @@ TVM_REGISTER_GLOBAL("ir_pass.Substitute")
     }
   });
 
-TVM_REGISTER_GLOBAL("ir_pass.RewriteForTensorCore")
-.set_body_typed
-  ([](const Stmt& stmt,
-      const te::Schedule& schedule,
-      const Map<te::Tensor, Buffer>& extern_buffer) {
-      return RewriteForTensorCore(stmt, schedule, extern_buffer);
-  });
-
 TVM_REGISTER_GLOBAL("ir_pass.ExprUseVar")
 .set_body([](TVMArgs args, TVMRetValue *ret) {
     *ret = ExprUseVar(args[0].operator PrimExpr(), args[1].operator Var());

vta/python/vta/ir_pass.py

@@ -638,7 +638,7 @@ def inject_conv2d_transpose_skip(stmt_in):
     selects = []
 
     def _find_basics(op):
-        if isinstance(op, tvm.tir.Call):
+        if isinstance(op, tvm.tir.BufferLoad):
             calls.append(op)
         elif isinstance(op, tvm.tir.Select):
             selects.append(op)
@@ -664,18 +664,18 @@ def _do_fold(op):
                 body = op.body.body
                 while isinstance(body, tvm.tir.IfThenElse):
                     body = body.then_case
-                args = body.args
-                res_tensor = body.func.output(0)
+                args = body.indices
+                res_buffer = body.buffer
                 tpl = (args[0], 1, args[1], 1, args[2], 1, args[3], 1, 0, 1, 0, env.BLOCK_OUT)
                 inner = tvm.tir.AttrStmt(
-                    [dout, res_tensor], 'buffer_bind_scope',
+                    [dout, res_buffer], 'buffer_bind_scope',
                     tvm.tir.call_intrin('handle', 'tvm_tuple', *tpl), inner)
                 return inner
             else:
                 conv_call, data_call, kernel_call = calls[-3:]
-                pad_data_tensor = data_call.func.output(0)
-                kernel_tensor = kernel_call.func.output(0)
-                res_tensor = conv_call.func.output(0)
+                pad_data_tensor = data_call.buffer
+                kernel_tensor = kernel_call.buffer
+                res_tensor = conv_call.buffer
 
                 if selects:
                     condition = selects[0].condition
@@ -696,19 +696,19 @@ def _do_fold(op):
                                                  0, 0, 0))
                 inner = irb.get()
 
-                args = conv_call.args
+                args = conv_call.indices
                 tpl = (args[0], 1, args[1], 1, args[2], 1, args[3],
                        1, 0, 1, 0, env.BLOCK_OUT)
                 inner = tvm.tir.AttrStmt(
                     [dout, res_tensor], 'buffer_bind_scope',
                     tvm.tir.call_intrin('handle', 'tvm_tuple', *tpl), inner)
-                args = kernel_call.args
+                args = kernel_call.indices
                 tpl = (args[0], 1, args[1], 1, args[2], 1, args[3],
                        1, 0, env.BLOCK_OUT, 0, env.BLOCK_IN)
                 inner = tvm.tir.AttrStmt(
                     [dwgt, kernel_tensor], 'buffer_bind_scope',
                     tvm.tir.call_intrin('handle', 'tvm_tuple', *tpl), inner)
-                args = data_call.args
+                args = data_call.indices
                 tpl = (args[0], 1, args[1], 1, args[2], 1, args[3],
                        1, 0, 1, 0, env.BLOCK_IN)
                 inner = tvm.tir.AttrStmt(