[Fix][TIR] LowerCrossThreadReduction with write-back predicate (#14199)

Prior to this PR, the cross-thread reduction lowering pass does not
add a store predicate to the write-back block. This is in consideration
that for a certain write-back buffer position, all values being stored
(by all the threads) in the write-back block are the same. Since all
threads are writing the same value, we were assuming that not having a
write-back block predicate is fine, because the result will not be wrong
in any way.

However, recently we noticed that some GPU backend compiler will capture
this behavior (multiple threads writing a same position) as a race
condition and thus throw compilation error. The compiler does not take
the fact that all values being stored are the same, and insist on
complaining.

This means that we will still need the write-back block predicate to
make things work. And this PR does this change. I have done integration
tests locally to make sure that the generated kernels is right and
produces the right results numerically.

src/tir/transforms/lower_cross_thread_reduction.cc

@@ -407,9 +407,15 @@ Stmt TransformReductionBlock(const BlockRealizeNode* realize,            //
           BufferStore(wb_buffers[i], BufferLoad(ct_buffers[i], {Integer(0)}), wb_indices));
       wb_regions.push_back(BufferRegion(wb_buffers[i], region));
     }
+    PrimExpr wb_predicate = const_true();
+    for (const ForNode* loop : reduction_loops) {
+      if (loop->thread_binding.defined()) {
+        wb_predicate = wb_predicate && (loop->loop_var == IntImm(loop->loop_var->dtype, 0));
+      }
+    }
     stmts.push_back(BlockRealize(
         /*iter_values=*/std::move(bindings),
-        /*predicate=*/const_true(),
+        /*predicate=*/wb_predicate,
         /*block=*/
         Block(/*iter_vars=*/std::move(iter_vars),
               /*reads=*/std::move(ct_buffer_regions),

tests/python/unittest/test_tir_transform_lower_cross_thread_reduction.py

@@ -93,6 +93,7 @@ def lowered_loop_split(a: T.handle, b: T.handle) -> None:
                 )
             with T.block("B_write_back"):
                 vi = T.axis.S(128, i)
+                T.where(ki == 0)
                 T.reads([reduce_temp0[0]])
                 T.writes([B[vi]])
                 B[vi] = reduce_temp0[0]
@@ -136,6 +137,7 @@ def lowered_no_normal_reduction(a: T.handle, b: T.handle) -> None:
                 )
             with T.block("B_write_back"):
                 vi = T.axis.spatial(128, i)
+                T.where(k == 0)
                 T.reads([reduce_temp0[0]])
                 T.writes([B[vi]])
                 B[vi] = reduce_temp0[0]
@@ -183,6 +185,7 @@ def lowered_two_bound_loops(a: T.handle, b: T.handle) -> None:
                     )
                 with T.block("B_write_back"):
                     vi = T.axis.spatial(128, i)
+                    T.where(ko == 0 and ki == 0)
                     T.reads([reduce_temp0[0]])
                     T.writes([B[vi]])
                     B[vi] = reduce_temp0[0]
@@ -264,6 +267,7 @@ def lowered_multiple_blocks_under_reduction_loop(a: T.handle, b: T.handle) -> No
                 )
             with T.block("B_write_back"):
                 vi = T.axis.spatial(16, i)
+                T.where(k0o == 0)
                 T.reads([reduce_temp0[0]])
                 T.writes([B[vi]])
                 B[vi] = reduce_temp0[0]
@@ -326,6 +330,7 @@ def lowered_with_block_predicate(a: T.handle, b: T.handle) -> None:
                 )
             with T.block("B_write_back"):
                 vi = T.axis.spatial(128, i)
+                T.where(ki == 0)
                 T.reads([reduce_temp0[0]])
                 T.writes([B[vi]])
                 B[vi] = reduce_temp0[0]
@@ -428,6 +433,7 @@ def lowered_single_reduction_loop_with_block_predicate(
                     )
                 with T.block("T_softmax_maxelem_write_back"):
                     i0_2 = T.axis.spatial(256, i0 + ax0)
+                    T.where(ax1_1 == 0)
                     T.reads(cross_thread_0[0])
                     T.writes(T_softmax_maxelem_shared[i0_2])
                     T_softmax_maxelem_shared[i0_2] = cross_thread_0[0]
@@ -467,6 +473,7 @@ def lowered_single_reduction_loop_with_block_predicate(
                     )
                 with T.block("T_softmax_expsum_write_back"):
                     i0_4 = T.axis.spatial(256, i0 + ax0)
+                    T.where(ax1_1 == 0)
                     T.reads(cross_thread_1[0])
                     T.writes(T_softmax_expsum_shared[i0_4])
                     T_softmax_expsum_shared[i0_4] = cross_thread_1[0]
@@ -636,6 +643,7 @@ def lowered_reducer_max(a: T.handle, b: T.handle) -> None:
                 )
             with T.block("B_write_back"):
                 vi = T.axis.spatial(128, i)
+                T.where(k == 0)
                 T.reads([reduce_temp0[0]])
                 T.writes([B[vi]])
                 B[vi] = reduce_temp0[0]
@@ -676,6 +684,7 @@ def lowered_zero_rank_buffer(a: T.handle, b: T.handle) -> None:
         with T.block("B_write_back"):
             T.reads([reduce_temp0[0]])
             T.writes([B[()]])
+            T.where(k == 0)
             B[()] = reduce_temp0[0]
 
 
@@ -865,6 +874,7 @@ def lowered_softmax(var_A: T.handle, var_T_softmax_norm: T.handle) -> None:
                 )
             with T.block("T_softmax_maxelem_write_back"):
                 i0_2 = T.axis.spatial(256, i0)
+                T.where(ax0_1 == 0)
                 T.reads([reduce_temp0[0]])
                 T.writes([T_softmax_maxelem_shared[i0_2]])
                 T_softmax_maxelem_shared[i0_2] = reduce_temp0[0]
@@ -907,6 +917,7 @@ def lowered_softmax(var_A: T.handle, var_T_softmax_norm: T.handle) -> None:
                 )
             with T.block("T_softmax_expsum_write_back"):
                 i0_4 = T.axis.spatial(256, i0)
+                T.where(ax0_1 == 0)
                 T.reads([reduce_temp1[0]])
                 T.writes([T_softmax_expsum_shared[i0_4]])
                 T_softmax_expsum_shared[i0_4] = reduce_temp1[0]
@@ -1018,6 +1029,7 @@ def lowered_argmax_split(
                 )
             with T.block("argmax_write_back"):
                 i = T.axis.spatial(128, i0)
+                T.where(i1_1 == 0)
                 T.reads(cross_thread_argmax_v0[0], cross_thread_argmax_v1[0])
                 T.writes(argmax_v0[i], argmax_v1[i])
                 argmax_v0[i] = cross_thread_argmax_v0[0]
@@ -1109,6 +1121,7 @@ def lowered_argmin_split_init_update_reordered(
                 )
             with T.block("argmin_write_back"):
                 i = T.axis.spatial(128, i0)
+                T.where(i1_1 == 0)
                 T.reads(cross_thread_argmin_v0[0], cross_thread_argmin_v1[0])
                 T.writes(argmin_v0[i], argmin_v1[i])
                 argmin_v0[i] = cross_thread_argmin_v0[0]
@@ -1227,6 +1240,7 @@ def lowered_layer_norm_tuple_sum(
                 )
             with T.block("data_red_temp_write_back"):
                 ax0 = T.axis.spatial(128, i0_fused)
+                T.where(i1_1 == 0)
                 T.reads(cross_thread_data_red_temp_v0[0], cross_thread_data_red_temp_v1[0])
                 T.writes(data_red_temp_v0[ax0], data_red_temp_v1[ax0])
                 data_red_temp_v0[ax0] = cross_thread_data_red_temp_v0[0]