[Unity] Support cumsum with pure int32 (#16439)

This PR fixes a bug on attr handling in data type rewriter and enforces i32 buffer in cumsum function definition, which ensures that cumsum can be run on a machine with int32 but not int64.

Author: jinhongyii

include/tvm/tir/data_type_rewriter.h

@@ -104,6 +104,7 @@ class IndexDataTypeRewriter : public DataTypeLegalizer {
   Stmt VisitStmt_(const BlockRealizeNode* op) override;
   Stmt VisitStmt_(const BlockNode* op) override;
   Stmt VisitStmt_(const BufferStoreNode* op) override;
+  Stmt VisitStmt_(const AttrStmtNode* op) override;
   PrimExpr VisitExpr_(const BufferLoadNode* op) override;
   Array<PrimExpr> VisitIndices(Array<PrimExpr> indices);
   Stmt VisitStmt_(const IfThenElseNode* op) override;

python/tvm/topi/cuda/scan.py

@@ -60,8 +60,8 @@ def exclusive_scan_ir(data, output, reduction=None, binop=tvm.tir.generic.add, i
         your operation.
     """
 
-    batch_size = prod(data.shape[:-1])
-    scan_axis_size = data.shape[-1]
+    batch_size = cast(prod(data.shape[:-1]), "int32")
+    scan_axis_size = cast(data.shape[-1], "int32")
 
     ib = tvm.tir.ir_builder.create()
 
@@ -105,7 +105,7 @@ def exclusive_scan_ir(data, output, reduction=None, binop=tvm.tir.generic.add, i
         # Up Sweep of exclusive scan
         lim = ceil_log2(scan_axis_size)
 
-        with ib.for_range(0, cast(lim, "int64"), dtype="int64") as l2_width:
+        with ib.for_range(0, cast(lim, "int32"), dtype="int32") as l2_width:
             width = 2 << l2_width
 
             with ib.new_scope():
@@ -121,9 +121,9 @@ def exclusive_scan_ir(data, output, reduction=None, binop=tvm.tir.generic.add, i
 
                 by = te.thread_axis("blockIdx.y")
                 ib.scope_attr(by, "thread_extent", nthread_by)
-                start = ib.allocate("int64", (1,), name="start", scope="local")
-                middle = ib.allocate("int64", (1,), name="middle", scope="local")
-                end = ib.allocate("int64", (1,), name="end", scope="local")
+                start = ib.allocate("int32", (1,), name="start", scope="local")
+                middle = ib.allocate("int32", (1,), name="middle", scope="local")
+                end = ib.allocate("int32", (1,), name="end", scope="local")
                 start[0] = width * tid
                 with ib.if_scope(start[0] < scan_axis_size):
                     middle[0] = start[0] + tvm.tir.indexdiv(width, 2)
@@ -143,7 +143,7 @@ def exclusive_scan_ir(data, output, reduction=None, binop=tvm.tir.generic.add, i
                     reduction[bx] = output[(bx + 1) * scan_axis_size - 1]
                 output[(bx + 1) * scan_axis_size - 1] = cast(identity_value, out_dtype)
 
-        with ib.for_range(0, cast(lim, "int64"), dtype="int64") as l2_width:
+        with ib.for_range(0, cast(lim, "int32"), dtype="int32") as l2_width:
             width = 2 << (lim - l2_width - 1)
 
             with ib.new_scope():
@@ -159,9 +159,9 @@ def exclusive_scan_ir(data, output, reduction=None, binop=tvm.tir.generic.add, i
 
                 by = te.thread_axis("blockIdx.y")
                 ib.scope_attr(by, "thread_extent", nthread_by)
-                start = ib.allocate("int64", (1,), name="start", scope="local")
-                middle = ib.allocate("int64", (1,), name="middle", scope="local")
-                end = ib.allocate("int64", (1,), name="end", scope="local")
+                start = ib.allocate("int32", (1,), name="start", scope="local")
+                middle = ib.allocate("int32", (1,), name="middle", scope="local")
+                end = ib.allocate("int32", (1,), name="end", scope="local")
                 tmp = ib.allocate(out_dtype, (1,), name="end", scope="local")
                 start[0] = width * tid
                 with ib.if_scope(tvm.tir.all(start[0] < scan_axis_size)):
@@ -206,8 +206,8 @@ def get_reduction_from_exclusive_scan(data, ex_scan_output, binop=tvm.tir.generi
         ex_scan_output = expand_dims(ex_scan_output, axis=0)
 
     def ir(data, data_ex_scan, reduction):
-        batch_size = prod(data.shape[:-1])
-        scan_axis_size = data.shape[-1]
+        batch_size = cast(prod(data.shape[:-1]), "int32")
+        scan_axis_size = cast(data.shape[-1], "int32")
 
         ib = tvm.tir.ir_builder.create()
 

src/tir/ir/data_type_rewriter.cc

@@ -258,6 +258,17 @@ Stmt IndexDataTypeRewriter::VisitStmt_(const AllocateNode* op) {
   }
 }
 
+Stmt IndexDataTypeRewriter::VisitStmt_(const AttrStmtNode* op) {
+  if (op->attr_key == attr::thread_extent || op->attr_key == attr::virtual_thread) {
+    bool is_enabled = is_enabled_;
+    is_enabled_ = true;
+    auto stmt = DataTypeLegalizer::VisitStmt_(op);
+    is_enabled_ = is_enabled;
+    return stmt;
+  }
+  return DataTypeLegalizer::VisitStmt_(op);
+}
+
 Stmt IndexDataTypeRewriter::VisitStmt_(const DeclBufferNode* op) {
   Buffer new_buffer = VisitBuffer(op->buffer);
   DeclBuffer decl_buffer = Downcast<DeclBuffer>(StmtExprMutator::VisitStmt_(op));