PaddlePaddle
diff --git a/‎paddle/cinn/auto_schedule/search_space/auto_gen_rule/auto_inline_test.cc
Copy file name to clipboard
+3-2Lines changed: 3 additions & 2 deletions b/‎paddle/cinn/auto_schedule/search_space/auto_gen_rule/auto_inline_test.cc
Copy file name to clipboard
+3-2Lines changed: 3 additions & 2 deletions
diff --git a/‎paddle/cinn/auto_schedule/search_space/auto_gen_rule/multi_level_tiling_test.cc
Copy file name to clipboard
+59-62Lines changed: 59 additions & 62 deletions b/‎paddle/cinn/auto_schedule/search_space/auto_gen_rule/multi_level_tiling_test.cc
Copy file name to clipboard
+59-62Lines changed: 59 additions & 62 deletions
diff --git a/‎paddle/cinn/auto_schedule/search_space/auto_gen_rule/test_helper.cc
Copy file name to clipboard
+4-6Lines changed: 4 additions & 6 deletions b/‎paddle/cinn/auto_schedule/search_space/auto_gen_rule/test_helper.cc
Copy file name to clipboard
+4-6Lines changed: 4 additions & 6 deletions
diff --git a/‎paddle/cinn/auto_schedule/task/tune_task.cc
Copy file name to clipboard
+2-1Lines changed: 2 additions & 1 deletion b/‎paddle/cinn/auto_schedule/task/tune_task.cc
Copy file name to clipboard
+2-1Lines changed: 2 additions & 1 deletion
diff --git a/‎paddle/cinn/auto_schedule/tests/performance_comparison_test.cc
Copy file name to clipboard
+3-1Lines changed: 3 additions & 1 deletion b/‎paddle/cinn/auto_schedule/tests/performance_comparison_test.cc
Copy file name to clipboard
+3-1Lines changed: 3 additions & 1 deletion
@@ -71,7 +71,6 @@ TEST(AutoInline, SingleLoopInline) {
                      nullptr,
                      target,
                      true);
-
   VLOG(6) << "Expr after lowering:";
   VLOG(6) << funcs[0]->body;
 
@@ -170,7 +169,9 @@ TEST(AutoInline, AddReluInline) {
 
   EXPECT_EQ(graph->fusion_groups.size(), 1UL);
   std::vector<ir::LoweredFunc> funcs =
-      op_lowerer->LowerWithoutSchedule(graph->fusion_groups[0]);
+      op_lowerer->Lower(graph->fusion_groups[0],
+                        /*apply_op_schedule = */ false,
+                        /*apply_group_schedule=*/false);
 
   VLOG(6) << "Expr before auto inline: " << funcs[0]->body;
 
 
@@ -388,9 +388,9 @@ TEST_F(TestMultiLevelTiling, ReduceSum) {
 
 TEST_F(TestMultiLevelTiling, Pool2d) {
   default_input_names = {"input"};
-  default_output_names = {"var_0"};
-  std::vector<int32_t> input_shape{2, 8, 16, 16};
-  std::vector<int32_t> output_shape{2, 8, 8, 8};
+  default_output_names = {"var_0", "pad_temp_0"};
+  std::vector<std::vector<int32_t>> input_shapes{{2, 8, 16, 16}};
+  std::vector<std::vector<int32_t>> output_shapes{{2, 8, 8, 8}, {2, 8, 18, 18}};
   std::string pooling_type = "max";
   std::vector<int> ksize{3, 3};
   std::vector<int> strides{2, 2};
@@ -402,7 +402,7 @@ TEST_F(TestMultiLevelTiling, Pool2d) {
   bool adaptive = false;
   std::string padding_algorithm = "EXPLICIT";
   frontend::Program pool2d_program = tests::OpBuilder("pool2d").Build(
-      {{"input", input_shape}},
+      {{"input", input_shapes[0]}},
       {{"pool_type", pooling_type},
        {"kernel_size", ksize},
        {"stride_size", strides},
@@ -440,107 +440,104 @@ TEST_F(TestMultiLevelTiling, Pool2d) {
 {
   ScheduleBlock(root)
   {
-    serial for (i, 0, 2)
     {
-      serial for (j, 0, 8)
+      serial for (i, 0, 2)
       {
-        serial for (k, 0, 18)
+        serial for (j, 0, 8)
         {
-          serial for (a, 0, 18)
+          serial for (k, 0, 18)
           {
-            ScheduleBlock(pad_temp_0)
+            serial for (a, 0, 18)
             {
-              i0, i1, i2, i3 = axis.bind(i, j, k, a)
-              pad_temp_0[i, j, k, a] = select(((a < 17) and ((a >= 1) and ((k < 17) and (k >= 1)))), input[i, j, (-1 + k), (-1 + a)], -3.40282347e+38f)
+              ScheduleBlock(pad_temp_0)
+              {
+                i0, i1, i2, i3 = axis.bind(i, j, k, a)
+                {
+                  pad_temp_0[i, j, k, a] = select(((a < 17) and ((a >= 1) and ((k < 17) and (k >= 1)))), input[i, j, (-1 + k), (-1 + a)], -3.40282347e+38f)
+                }
+              }
             }
           }
         }
       }
-    }
-  }
-}
-}  // end Expr 0
-Expr 1 {
-{
-  ScheduleBlock(root_0)
-  {
-    {
-      thread_bind[blockIdx.x] for (i_j_k_a_fused, 0, 16)
       {
-        thread_bind[threadIdx.x] for (i_0_j_0_k_0_a_0_fused, 0, 4)
+        thread_bind[blockIdx.x] for (i_j_k_a_fused, 0, 16)
         {
-          serial for (i_1, 0, 1)
+          thread_bind[threadIdx.x] for (i_0_j_0_k_0_a_0_fused, 0, 4)
           {
-            serial for (j_1, 0, 4)
+            serial for (i_1, 0, 1)
             {
-              serial for (k_1, 0, 1)
+              serial for (j_1, 0, 4)
               {
-                serial for (a_1, 0, 4)
+                serial for (k_1, 0, 1)
                 {
-                  ScheduleBlock(var_0__reduce_init)
+                  serial for (a_1, 0, 4)
                   {
-                    i0_0, i1_0, i2_0, i3_0 = axis.bind(((((i_j_k_a_fused / 2) / 2) / 2) + ((i_0_j_0_k_0_a_0_fused / 4) + i_1)), ((4 * (((i_j_k_a_fused / 2) / 2) % 2)) + j_1), ((i_0_j_0_k_0_a_0_fused % 4) + ((4 * ((i_j_k_a_fused / 2) % 2)) + k_1)), ((4 * (i_j_k_a_fused % 2)) + a_1))
+                    ScheduleBlock(var_0__reduce_init)
                     {
-                      var_0__reduce_init[((((i_j_k_a_fused / 2) / 2) / 2) + ((i_0_j_0_k_0_a_0_fused / 4) + i_1)), ((4 * (((i_j_k_a_fused / 2) / 2) % 2)) + j_1), ((4 * ((i_j_k_a_fused / 2) % 2)) + ((i_0_j_0_k_0_a_0_fused % 4) + k_1)), ((4 * (i_j_k_a_fused % 2)) + a_1)] = -3.40282347e+38f
+                      i0_0, i1_0, i2_0, i3_0 = axis.bind(((((i_j_k_a_fused / 2) / 2) / 2) + ((i_0_j_0_k_0_a_0_fused / 4) + i_1)), ((4 * (((i_j_k_a_fused / 2) / 2) % 2)) + j_1), ((i_0_j_0_k_0_a_0_fused % 4) + ((4 * ((i_j_k_a_fused / 2) % 2)) + k_1)), ((4 * (i_j_k_a_fused % 2)) + a_1))
+                      {
+                        var_0__reduce_init[((((i_j_k_a_fused / 2) / 2) / 2) + ((i_0_j_0_k_0_a_0_fused / 4) + i_1)), ((4 * (((i_j_k_a_fused / 2) / 2) % 2)) + j_1), ((4 * ((i_j_k_a_fused / 2) % 2)) + ((i_0_j_0_k_0_a_0_fused % 4) + k_1)), ((4 * (i_j_k_a_fused % 2)) + a_1)] = -3.40282347e+38f
+                      }
                     }
                   }
                 }
               }
             }
-          }
-          {
-            serial for (kernel_idx, 0, 3)
             {
-              serial for (kernel_idx_0, 0, 3)
+              serial for (kernel_idx, 0, 3)
               {
-                serial for (ax0_ax1_ax2_ax3_fused, 0, 28)
+                serial for (kernel_idx_0, 0, 3)
                 {
-                  ScheduleBlock(pad_temp_0_shared_temp_buffer)
+                  serial for (ax0_ax1_ax2_ax3_fused, 0, 28)
                   {
-                    v0, v1, v2, v3 = axis.bind(((((i_j_k_a_fused / 2) / 2) / 2) + ((i_0_j_0_k_0_a_0_fused / 4) + ((ax0_ax1_ax2_ax3_fused / 7) / 4))), (((ax0_ax1_ax2_ax3_fused / 7) % 4) + (4 * (((i_j_k_a_fused / 2) / 2) % 2))), ((8 * ((i_j_k_a_fused / 2) % 2)) + ((2 * (i_0_j_0_k_0_a_0_fused % 4)) + kernel_idx)), ((ax0_ax1_ax2_ax3_fused % 7) + ((8 * (i_j_k_a_fused % 2)) + kernel_idx_0)))
-                    attrs(compute_at_extra_var:ax0,ax1,ax2,ax3, cooperative_process:0)
+                    ScheduleBlock(pad_temp_0_shared_temp_buffer)
                     {
-                      pad_temp_0_shared_temp_buffer[v0, v1, v2, v3] = pad_temp_0[v0, v1, v2, v3]
+                      v0, v1, v2, v3 = axis.bind(((((i_j_k_a_fused / 2) / 2) / 2) + ((i_0_j_0_k_0_a_0_fused / 4) + ((ax0_ax1_ax2_ax3_fused / 7) / 4))), (((ax0_ax1_ax2_ax3_fused / 7) % 4) + (4 * (((i_j_k_a_fused / 2) / 2) % 2))), ((8 * ((i_j_k_a_fused / 2) % 2)) + ((2 * (i_0_j_0_k_0_a_0_fused % 4)) + kernel_idx)), ((ax0_ax1_ax2_ax3_fused % 7) + ((8 * (i_j_k_a_fused % 2)) + kernel_idx_0)))
+                      attrs(compute_at_extra_var:ax0,ax1,ax2,ax3, cooperative_process:0)
+                      {
+                        pad_temp_0_shared_temp_buffer[v0, v1, v2, v3] = pad_temp_0[v0, v1, v2, v3]
+                      }
                     }
                   }
-                }
-                serial for (i_1, 0, 1)
-                {
-                  serial for (j_1, 0, 4)
+                  serial for (i_1, 0, 1)
                   {
-                    serial for (k_1, 0, 1)
+                    serial for (j_1, 0, 4)
                     {
-                      serial for (a_1, 0, 4)
+                      serial for (k_1, 0, 1)
                       {
-                        ScheduleBlock(var_0_local_temp_buffer)
+                        serial for (a_1, 0, 4)
                         {
-                          i0_1, i1_1, i2_1, i3_1, i4, i5 = axis.bind(((((i_j_k_a_fused / 2) / 2) / 2) + ((i_0_j_0_k_0_a_0_fused / 4) + i_1)), ((4 * (((i_j_k_a_fused / 2) / 2) % 2)) + j_1), ((i_0_j_0_k_0_a_0_fused % 4) + ((4 * ((i_j_k_a_fused / 2) % 2)) + k_1)), ((4 * (i_j_k_a_fused % 2)) + a_1), kernel_idx, kernel_idx_0)
-                          read_buffers(_var_0[i(undefined:undefined), j(undefined:undefined), k(undefined:undefined), a(undefined:undefined)], _pad_temp_0[i(undefined:undefined), j(undefined:undefined)])
-                          write_buffers(_var_0[i(undefined:undefined), j(undefined:undefined), k(undefined:undefined), a(undefined:undefined)])
+                          ScheduleBlock(var_0_local_temp_buffer)
                           {
-                            var_0_local_temp_buffer[((((i_j_k_a_fused / 2) / 2) / 2) + ((i_0_j_0_k_0_a_0_fused / 4) + i_1)), ((4 * (((i_j_k_a_fused / 2) / 2) % 2)) + j_1), ((4 * ((i_j_k_a_fused / 2) % 2)) + ((i_0_j_0_k_0_a_0_fused % 4) + k_1)), ((4 * (i_j_k_a_fused % 2)) + a_1)] = cinn_max(var_0_local_temp_buffer[((((i_j_k_a_fused / 2) / 2) / 2) + ((i_0_j_0_k_0_a_0_fused / 4) + i_1)), ((4 * (((i_j_k_a_fused / 2) / 2) % 2)) + j_1), ((i_0_j_0_k_0_a_0_fused % 4) + ((4 * ((i_j_k_a_fused / 2) % 2)) + k_1)), ((4 * (i_j_k_a_fused % 2)) + a_1)], pad_temp_0_shared_temp_buffer[((((i_j_k_a_fused / 2) / 2) / 2) + ((i_0_j_0_k_0_a_0_fused / 4) + i_1)), ((4 * (((i_j_k_a_fused / 2) / 2) % 2)) + j_1), ((8 * ((i_j_k_a_fused / 2) % 2)) + ((2 * (i_0_j_0_k_0_a_0_fused % 4)) + ((2 * k_1) + kernel_idx))), ((8 * (i_j_k_a_fused % 2)) + ((2 * a_1) + kernel_idx_0))])
+                            i0_1, i1_1, i2_1, i3_1, i4, i5 = axis.bind(((((i_j_k_a_fused / 2) / 2) / 2) + ((i_0_j_0_k_0_a_0_fused / 4) + i_1)), ((4 * (((i_j_k_a_fused / 2) / 2) % 2)) + j_1), ((i_0_j_0_k_0_a_0_fused % 4) + ((4 * ((i_j_k_a_fused / 2) % 2)) + k_1)), ((4 * (i_j_k_a_fused % 2)) + a_1), kernel_idx, kernel_idx_0)
+                            read_buffers(_var_0[i(undefined:undefined), j(undefined:undefined), k(undefined:undefined), a(undefined:undefined)], _pad_temp_0[i(undefined:undefined), j(undefined:undefined)])
+                            write_buffers(_var_0[i(undefined:undefined), j(undefined:undefined), k(undefined:undefined), a(undefined:undefined)])
+                            {
+                              var_0_local_temp_buffer[((((i_j_k_a_fused / 2) / 2) / 2) + ((i_0_j_0_k_0_a_0_fused / 4) + i_1)), ((4 * (((i_j_k_a_fused / 2) / 2) % 2)) + j_1), ((4 * ((i_j_k_a_fused / 2) % 2)) + ((i_0_j_0_k_0_a_0_fused % 4) + k_1)), ((4 * (i_j_k_a_fused % 2)) + a_1)] = cinn_max(var_0_local_temp_buffer[((((i_j_k_a_fused / 2) / 2) / 2) + ((i_0_j_0_k_0_a_0_fused / 4) + i_1)), ((4 * (((i_j_k_a_fused / 2) / 2) % 2)) + j_1), ((i_0_j_0_k_0_a_0_fused % 4) + ((4 * ((i_j_k_a_fused / 2) % 2)) + k_1)), ((4 * (i_j_k_a_fused % 2)) + a_1)], pad_temp_0_shared_temp_buffer[((((i_j_k_a_fused / 2) / 2) / 2) + ((i_0_j_0_k_0_a_0_fused / 4) + i_1)), ((4 * (((i_j_k_a_fused / 2) / 2) % 2)) + j_1), ((8 * ((i_j_k_a_fused / 2) % 2)) + ((2 * (i_0_j_0_k_0_a_0_fused % 4)) + ((2 * k_1) + kernel_idx))), ((8 * (i_j_k_a_fused % 2)) + ((2 * a_1) + kernel_idx_0))])
+                            }
                           }
                         }
                       }
                     }
                   }
                 }
               }
-            }
-            serial for (ax0_0, 0, 1)
-            {
-              serial for (ax1_0, 0, 4)
+              serial for (ax0_0, 0, 1)
               {
-                serial for (ax2_0, 0, 1)
+                serial for (ax1_0, 0, 4)
                 {
-                  serial for (ax3_0, 0, 4)
+                  serial for (ax2_0, 0, 1)
                   {
-                    ScheduleBlock(var_0)
+                    serial for (ax3_0, 0, 4)
                     {
-                      v0, v1, v2, v3 = axis.bind((((((i_j_k_a_fused / 2) / 2) / 2) + (i_0_j_0_k_0_a_0_fused / 4)) + ax0_0), ((4 * (((i_j_k_a_fused / 2) / 2) % 2)) + ax1_0), (((4 * ((i_j_k_a_fused / 2) % 2)) + (i_0_j_0_k_0_a_0_fused % 4)) + ax2_0), ((4 * (i_j_k_a_fused % 2)) + ax3_0))
-                      attrs(reverse_compute_at_extra_var:ax0_0,ax1_0,ax2_0,ax3_0)
+                      ScheduleBlock(var_0)
                       {
-                        var_0[v0, v1, v2, v3] = var_0_local_temp_buffer[v0, v1, v2, v3]
+                        v0, v1, v2, v3 = axis.bind((((((i_j_k_a_fused / 2) / 2) / 2) + (i_0_j_0_k_0_a_0_fused / 4)) + ax0_0), ((4 * (((i_j_k_a_fused / 2) / 2) % 2)) + ax1_0), (((4 * ((i_j_k_a_fused / 2) % 2)) + (i_0_j_0_k_0_a_0_fused % 4)) + ax2_0), ((4 * (i_j_k_a_fused % 2)) + ax3_0))
+                        attrs(reverse_compute_at_extra_var:ax0_0,ax1_0,ax2_0,ax3_0)
+                        {
+                          var_0[v0, v1, v2, v3] = var_0_local_temp_buffer[v0, v1, v2, v3]
+                        }
                       }
                     }
                   }
@@ -553,7 +550,7 @@ Expr 1 {
     }
   }
 }
-}  // end Expr 1
+}  // end Expr 0
 )ROC";
   ASSERT_EQ(ir, expected_ir);
 
@@ -569,8 +566,8 @@ Expr 1 {
           pool2d_program, fixed_rand_seed, /* apply_manual_schedule*/ true))),
       default_input_names,
       default_output_names,
-      {input_shape},
-      {output_shape},
+      input_shapes,
+      output_shapes,
       target_);
 }
 
 
@@ -63,12 +63,10 @@ ir::IRSchedule TestAutoGenRuleBase::MakeIRSchedule(
       absl::flat_hash_map<std::string, hlir::framework::shape_t>>("infershape");
   hlir::framework::OpLowerer op_lowerer(dtype_dict, shape_dict, target_);
 
-  if (apply_manual_schedule) {
-    lowered_funcs_ = op_lowerer.Lower(graph->fusion_groups.front());
-  } else {
-    lowered_funcs_ =
-        op_lowerer.LowerWithoutSchedule(graph->fusion_groups.front());
-  }
+  lowered_funcs_ =
+      op_lowerer.Lower(graph->fusion_groups.front(),
+                       /*apply_op_schedule = */ apply_manual_schedule,
+                       /*apply_group_schedule = */ apply_manual_schedule);
   CHECK(!lowered_funcs_.empty()) << "lowered_funcs_ is empty";
 
   std::vector<Expr> bodys;
 
@@ -39,7 +39,8 @@ void TuneTask::Initialize(
   op_lowerer = lower_handler;
 
   // Set lowered_funcs and analyze output names.
-  this->lowered_funcs = op_lowerer->LowerWithoutSchedule(subgraph);
+  this->lowered_funcs = op_lowerer->Lower(
+      subgraph, /*apply_op_schedule = */ false, /*apply_group_schedule=*/false);
   this->output_names = GetOutputNamesFromLoweredFunc(this->lowered_funcs);
   this->serialized_key = SerializeToString(shape_dict, dtype_dict);
 }
 
@@ -157,7 +157,9 @@ class PerformanceTester : public ::testing::Test {
 
     for (auto group : graph->fusion_groups) {
       compile_options.lowered_funcs.push_back(
-          op_lowerer->LowerWithoutSchedule(group));
+          op_lowerer->Lower(group,
+                            /*apply_op_schedule = */ false,
+                            /*apply_group_schedule=*/false));
     }
 
     VLOG(3) << "===========================No Schedule LoweredFunc "
Original file line number	Diff line number	Diff line change
`@@ -388,9 +388,9 @@ TEST_F(TestMultiLevelTiling, ReduceSum) {`
`388`	`388`
`389`	`389`	`TEST_F(TestMultiLevelTiling, Pool2d) {`
`390`	`390`	`default_input_names = {"input"};`
`391`		`- default_output_names = {"var_0"};`
`392`		`- std::vector<int32_t> input_shape{2, 8, 16, 16};`
`393`		`- std::vector<int32_t> output_shape{2, 8, 8, 8};`
	`391`	`+ default_output_names = {"var_0", "pad_temp_0"};`
	`392`	`+ std::vector<std::vector<int32_t>> input_shapes{{2, 8, 16, 16}};`
	`393`	`+ std::vector<std::vector<int32_t>> output_shapes{{2, 8, 8, 8}, {2, 8, 18, 18}};`
`394`	`394`	`std::string pooling_type = "max";`
`395`	`395`	`std::vector<int> ksize{3, 3};`
`396`	`396`	`std::vector<int> strides{2, 2};`
`@@ -402,7 +402,7 @@ TEST_F(TestMultiLevelTiling, Pool2d) {`
`402`	`402`	`bool adaptive = false;`
`403`	`403`	`std::string padding_algorithm = "EXPLICIT";`
`404`	`404`	`frontend::Program pool2d_program = tests::OpBuilder("pool2d").Build(`
`405`		`- {{"input", input_shape}},`
	`405`	`+ {{"input", input_shapes[0]}},`
`406`	`406`	`{{"pool_type", pooling_type},`
`407`	`407`	`{"kernel_size", ksize},`
`408`	`408`	`{"stride_size", strides},`
`@@ -440,107 +440,104 @@ TEST_F(TestMultiLevelTiling, Pool2d) {`
`440`	`440`	`{`
`441`	`441`	`ScheduleBlock(root)`
`442`	`442`	`{`
`443`		`- serial for (i, 0, 2)`
`444`	`443`	`{`
`445`		`- serial for (j, 0, 8)`
	`444`	`+ serial for (i, 0, 2)`
`446`	`445`	`{`
`447`		`- serial for (k, 0, 18)`
	`446`	`+ serial for (j, 0, 8)`
`448`	`447`	`{`
`449`		`- serial for (a, 0, 18)`
	`448`	`+ serial for (k, 0, 18)`
`450`	`449`	`{`
`451`		`- ScheduleBlock(pad_temp_0)`
	`450`	`+ serial for (a, 0, 18)`
`452`	`451`	`{`
`453`		`- i0, i1, i2, i3 = axis.bind(i, j, k, a)`
`454`		`- pad_temp_0[i, j, k, a] = select(((a < 17) and ((a >= 1) and ((k < 17) and (k >= 1)))), input[i, j, (-1 + k), (-1 + a)], -3.40282347e+38f)`
	`452`	`+ ScheduleBlock(pad_temp_0)`
	`453`	`+ {`
	`454`	`+ i0, i1, i2, i3 = axis.bind(i, j, k, a)`
	`455`	`+ {`
	`456`	`+ pad_temp_0[i, j, k, a] = select(((a < 17) and ((a >= 1) and ((k < 17) and (k >= 1)))), input[i, j, (-1 + k), (-1 + a)], -3.40282347e+38f)`
	`457`	`+ }`
	`458`	`+ }`
`455`	`459`	`}`
`456`	`460`	`}`
`457`	`461`	`}`
`458`	`462`	`}`
`459`		`- }`
`460`		`- }`
`461`		`-}`
`462`		`-} // end Expr 0`
`463`		`-Expr 1 {`
`464`		`-{`
`465`		`- ScheduleBlock(root_0)`
`466`		`- {`
`467`		`- {`
`468`		`- thread_bind[blockIdx.x] for (i_j_k_a_fused, 0, 16)`
`469`	`463`	`{`
`470`		`- thread_bind[threadIdx.x] for (i_0_j_0_k_0_a_0_fused, 0, 4)`
	`464`	`+ thread_bind[blockIdx.x] for (i_j_k_a_fused, 0, 16)`
`471`	`465`	`{`
`472`		`- serial for (i_1, 0, 1)`
	`466`	`+ thread_bind[threadIdx.x] for (i_0_j_0_k_0_a_0_fused, 0, 4)`
`473`	`467`	`{`
`474`		`- serial for (j_1, 0, 4)`
	`468`	`+ serial for (i_1, 0, 1)`
`475`	`469`	`{`
`476`		`- serial for (k_1, 0, 1)`
	`470`	`+ serial for (j_1, 0, 4)`
`477`	`471`	`{`
`478`		`- serial for (a_1, 0, 4)`
	`472`	`+ serial for (k_1, 0, 1)`
`479`	`473`	`{`
`480`		`- ScheduleBlock(var_0__reduce_init)`
	`474`	`+ serial for (a_1, 0, 4)`
`481`	`475`	`{`
`482`		`- i0_0, i1_0, i2_0, i3_0 = axis.bind(((((i_j_k_a_fused / 2) / 2) / 2) + ((i_0_j_0_k_0_a_0_fused / 4) + i_1)), ((4 * (((i_j_k_a_fused / 2) / 2) % 2)) + j_1), ((i_0_j_0_k_0_a_0_fused % 4) + ((4 * ((i_j_k_a_fused / 2) % 2)) + k_1)), ((4 * (i_j_k_a_fused % 2)) + a_1))`
	`476`	`+ ScheduleBlock(var_0__reduce_init)`
`483`	`477`	`{`
`484`		`- var_0__reduce_init[((((i_j_k_a_fused / 2) / 2) / 2) + ((i_0_j_0_k_0_a_0_fused / 4) + i_1)), ((4 * (((i_j_k_a_fused / 2) / 2) % 2)) + j_1), ((4 * ((i_j_k_a_fused / 2) % 2)) + ((i_0_j_0_k_0_a_0_fused % 4) + k_1)), ((4 * (i_j_k_a_fused % 2)) + a_1)] = -3.40282347e+38f`
	`478`	`+ i0_0, i1_0, i2_0, i3_0 = axis.bind(((((i_j_k_a_fused / 2) / 2) / 2) + ((i_0_j_0_k_0_a_0_fused / 4) + i_1)), ((4 * (((i_j_k_a_fused / 2) / 2) % 2)) + j_1), ((i_0_j_0_k_0_a_0_fused % 4) + ((4 * ((i_j_k_a_fused / 2) % 2)) + k_1)), ((4 * (i_j_k_a_fused % 2)) + a_1))`
	`479`	`+ {`
	`480`	`+ var_0__reduce_init[((((i_j_k_a_fused / 2) / 2) / 2) + ((i_0_j_0_k_0_a_0_fused / 4) + i_1)), ((4 * (((i_j_k_a_fused / 2) / 2) % 2)) + j_1), ((4 * ((i_j_k_a_fused / 2) % 2)) + ((i_0_j_0_k_0_a_0_fused % 4) + k_1)), ((4 * (i_j_k_a_fused % 2)) + a_1)] = -3.40282347e+38f`
	`481`	`+ }`
`485`	`482`	`}`
`486`	`483`	`}`
`487`	`484`	`}`
`488`	`485`	`}`
`489`	`486`	`}`
`490`		`- }`
`491`		`- {`
`492`		`- serial for (kernel_idx, 0, 3)`
`493`	`487`	`{`
`494`		`- serial for (kernel_idx_0, 0, 3)`
	`488`	`+ serial for (kernel_idx, 0, 3)`
`495`	`489`	`{`
`496`		`- serial for (ax0_ax1_ax2_ax3_fused, 0, 28)`
	`490`	`+ serial for (kernel_idx_0, 0, 3)`
`497`	`491`	`{`
`498`		`- ScheduleBlock(pad_temp_0_shared_temp_buffer)`
	`492`	`+ serial for (ax0_ax1_ax2_ax3_fused, 0, 28)`
`499`	`493`	`{`
`500`		`- v0, v1, v2, v3 = axis.bind(((((i_j_k_a_fused / 2) / 2) / 2) + ((i_0_j_0_k_0_a_0_fused / 4) + ((ax0_ax1_ax2_ax3_fused / 7) / 4))), (((ax0_ax1_ax2_ax3_fused / 7) % 4) + (4 * (((i_j_k_a_fused / 2) / 2) % 2))), ((8 * ((i_j_k_a_fused / 2) % 2)) + ((2 * (i_0_j_0_k_0_a_0_fused % 4)) + kernel_idx)), ((ax0_ax1_ax2_ax3_fused % 7) + ((8 * (i_j_k_a_fused % 2)) + kernel_idx_0)))`
`501`		`- attrs(compute_at_extra_var:ax0,ax1,ax2,ax3, cooperative_process:0)`
	`494`	`+ ScheduleBlock(pad_temp_0_shared_temp_buffer)`
`502`	`495`	`{`
`503`		`- pad_temp_0_shared_temp_buffer[v0, v1, v2, v3] = pad_temp_0[v0, v1, v2, v3]`
	`496`	`+ v0, v1, v2, v3 = axis.bind(((((i_j_k_a_fused / 2) / 2) / 2) + ((i_0_j_0_k_0_a_0_fused / 4) + ((ax0_ax1_ax2_ax3_fused / 7) / 4))), (((ax0_ax1_ax2_ax3_fused / 7) % 4) + (4 * (((i_j_k_a_fused / 2) / 2) % 2))), ((8 * ((i_j_k_a_fused / 2) % 2)) + ((2 * (i_0_j_0_k_0_a_0_fused % 4)) + kernel_idx)), ((ax0_ax1_ax2_ax3_fused % 7) + ((8 * (i_j_k_a_fused % 2)) + kernel_idx_0)))`
	`497`	`+ attrs(compute_at_extra_var:ax0,ax1,ax2,ax3, cooperative_process:0)`
	`498`	`+ {`
	`499`	`+ pad_temp_0_shared_temp_buffer[v0, v1, v2, v3] = pad_temp_0[v0, v1, v2, v3]`
	`500`	`+ }`
`504`	`501`	`}`
`505`	`502`	`}`
`506`		`- }`
`507`		`- serial for (i_1, 0, 1)`
`508`		`- {`
`509`		`- serial for (j_1, 0, 4)`
	`503`	`+ serial for (i_1, 0, 1)`
`510`	`504`	`{`
`511`		`- serial for (k_1, 0, 1)`
	`505`	`+ serial for (j_1, 0, 4)`
`512`	`506`	`{`
`513`		`- serial for (a_1, 0, 4)`
	`507`	`+ serial for (k_1, 0, 1)`
`514`	`508`	`{`
`515`		`- ScheduleBlock(var_0_local_temp_buffer)`
	`509`	`+ serial for (a_1, 0, 4)`
`516`	`510`	`{`
`517`		`- i0_1, i1_1, i2_1, i3_1, i4, i5 = axis.bind(((((i_j_k_a_fused / 2) / 2) / 2) + ((i_0_j_0_k_0_a_0_fused / 4) + i_1)), ((4 * (((i_j_k_a_fused / 2) / 2) % 2)) + j_1), ((i_0_j_0_k_0_a_0_fused % 4) + ((4 * ((i_j_k_a_fused / 2) % 2)) + k_1)), ((4 * (i_j_k_a_fused % 2)) + a_1), kernel_idx, kernel_idx_0)`
`518`		`- read_buffers(_var_0[i(undefined:undefined), j(undefined:undefined), k(undefined:undefined), a(undefined:undefined)], _pad_temp_0[i(undefined:undefined), j(undefined:undefined)])`
`519`		`- write_buffers(_var_0[i(undefined:undefined), j(undefined:undefined), k(undefined:undefined), a(undefined:undefined)])`
	`511`	`+ ScheduleBlock(var_0_local_temp_buffer)`
`520`	`512`	`{`
`521`		- var_0_local_temp_buffer[((((i_j_k_a_fused / 2) / 2) / 2) + ((i_0_j_0_k_0_a_0_fused / 4) + i_1)), ((4 * (((i_j_k_a_fused / 2) / 2) % 2)) + j_1), ((4 * ((i_j_k_a_fused / 2) % 2)) + ((i_0_j_0_k_0_a_0_fused % 4) + k_1)), ((4 * (i_j_k_a_fused % 2)) + a_1)] = cinn_max(var_0_local_temp_buffer[((((i_j_k_a_fused / 2) / 2) / 2) + ((i_0_j_0_k_0_a_0_fused / 4) + i_1)), ((4 * (((i_j_k_a_fused / 2) / 2) % 2)) + j_1), ((i_0_j_0_k_0_a_0_fused % 4) + ((4 * ((i_j_k_a_fused / 2) % 2)) + k_1)), ((4 * (i_j_k_a_fused % 2)) + a_1)], pad_temp_0_shared_temp_buffer[((((i_j_k_a_fused / 2) / 2) / 2) + ((i_0_j_0_k_0_a_0_fused / 4) + i_1)), ((4 * (((i_j_k_a_fused / 2) / 2) % 2)) + j_1), ((8 * ((i_j_k_a_fused / 2) % 2)) + ((2 * (i_0_j_0_k_0_a_0_fused % 4)) + ((2 * k_1) + kernel_idx))), ((8 * (i_j_k_a_fused % 2)) + ((2 * a_1) + kernel_idx_0))])
	`513`	`+ i0_1, i1_1, i2_1, i3_1, i4, i5 = axis.bind(((((i_j_k_a_fused / 2) / 2) / 2) + ((i_0_j_0_k_0_a_0_fused / 4) + i_1)), ((4 * (((i_j_k_a_fused / 2) / 2) % 2)) + j_1), ((i_0_j_0_k_0_a_0_fused % 4) + ((4 * ((i_j_k_a_fused / 2) % 2)) + k_1)), ((4 * (i_j_k_a_fused % 2)) + a_1), kernel_idx, kernel_idx_0)`
	`514`	`+ read_buffers(_var_0[i(undefined:undefined), j(undefined:undefined), k(undefined:undefined), a(undefined:undefined)], _pad_temp_0[i(undefined:undefined), j(undefined:undefined)])`
	`515`	`+ write_buffers(_var_0[i(undefined:undefined), j(undefined:undefined), k(undefined:undefined), a(undefined:undefined)])`
	`516`	`+ {`
	`517`	+ var_0_local_temp_buffer[((((i_j_k_a_fused / 2) / 2) / 2) + ((i_0_j_0_k_0_a_0_fused / 4) + i_1)), ((4 * (((i_j_k_a_fused / 2) / 2) % 2)) + j_1), ((4 * ((i_j_k_a_fused / 2) % 2)) + ((i_0_j_0_k_0_a_0_fused % 4) + k_1)), ((4 * (i_j_k_a_fused % 2)) + a_1)] = cinn_max(var_0_local_temp_buffer[((((i_j_k_a_fused / 2) / 2) / 2) + ((i_0_j_0_k_0_a_0_fused / 4) + i_1)), ((4 * (((i_j_k_a_fused / 2) / 2) % 2)) + j_1), ((i_0_j_0_k_0_a_0_fused % 4) + ((4 * ((i_j_k_a_fused / 2) % 2)) + k_1)), ((4 * (i_j_k_a_fused % 2)) + a_1)], pad_temp_0_shared_temp_buffer[((((i_j_k_a_fused / 2) / 2) / 2) + ((i_0_j_0_k_0_a_0_fused / 4) + i_1)), ((4 * (((i_j_k_a_fused / 2) / 2) % 2)) + j_1), ((8 * ((i_j_k_a_fused / 2) % 2)) + ((2 * (i_0_j_0_k_0_a_0_fused % 4)) + ((2 * k_1) + kernel_idx))), ((8 * (i_j_k_a_fused % 2)) + ((2 * a_1) + kernel_idx_0))])
	`518`	`+ }`
`522`	`519`	`}`
`523`	`520`	`}`
`524`	`521`	`}`
`525`	`522`	`}`
`526`	`523`	`}`
`527`	`524`	`}`
`528`	`525`	`}`
`529`		`- }`
`530`		`- serial for (ax0_0, 0, 1)`
`531`		`- {`
`532`		`- serial for (ax1_0, 0, 4)`
	`526`	`+ serial for (ax0_0, 0, 1)`
`533`	`527`	`{`
`534`		`- serial for (ax2_0, 0, 1)`
	`528`	`+ serial for (ax1_0, 0, 4)`
`535`	`529`	`{`
`536`		`- serial for (ax3_0, 0, 4)`
	`530`	`+ serial for (ax2_0, 0, 1)`
`537`	`531`	`{`
`538`		`- ScheduleBlock(var_0)`
	`532`	`+ serial for (ax3_0, 0, 4)`
`539`	`533`	`{`
`540`		`- v0, v1, v2, v3 = axis.bind((((((i_j_k_a_fused / 2) / 2) / 2) + (i_0_j_0_k_0_a_0_fused / 4)) + ax0_0), ((4 * (((i_j_k_a_fused / 2) / 2) % 2)) + ax1_0), (((4 * ((i_j_k_a_fused / 2) % 2)) + (i_0_j_0_k_0_a_0_fused % 4)) + ax2_0), ((4 * (i_j_k_a_fused % 2)) + ax3_0))`
`541`		`- attrs(reverse_compute_at_extra_var:ax0_0,ax1_0,ax2_0,ax3_0)`
	`534`	`+ ScheduleBlock(var_0)`
`542`	`535`	`{`
`543`		`- var_0[v0, v1, v2, v3] = var_0_local_temp_buffer[v0, v1, v2, v3]`
	`536`	`+ v0, v1, v2, v3 = axis.bind((((((i_j_k_a_fused / 2) / 2) / 2) + (i_0_j_0_k_0_a_0_fused / 4)) + ax0_0), ((4 * (((i_j_k_a_fused / 2) / 2) % 2)) + ax1_0), (((4 * ((i_j_k_a_fused / 2) % 2)) + (i_0_j_0_k_0_a_0_fused % 4)) + ax2_0), ((4 * (i_j_k_a_fused % 2)) + ax3_0))`
	`537`	`+ attrs(reverse_compute_at_extra_var:ax0_0,ax1_0,ax2_0,ax3_0)`
	`538`	`+ {`
	`539`	`+ var_0[v0, v1, v2, v3] = var_0_local_temp_buffer[v0, v1, v2, v3]`
	`540`	`+ }`
`544`	`541`	`}`
`545`	`542`	`}`
`546`	`543`	`}`
`@@ -553,7 +550,7 @@ Expr 1 {`
`553`	`550`	`}`
`554`	`551`	`}`
`555`	`552`	`}`
`556`		`-} // end Expr 1`
	`553`	`+} // end Expr 0`
`557`	`554`	`)ROC";`
`558`	`555`	`ASSERT_EQ(ir, expected_ir);`
`559`	`556`
`@@ -569,8 +566,8 @@ Expr 1 {`
`569`	`566`	`pool2d_program, fixed_rand_seed, /* apply_manual_schedule*/ true))),`
`570`	`567`	`default_input_names,`
`571`	`568`	`default_output_names,`
`572`		`- {input_shape},`
`573`		`- {output_shape},`
	`569`	`+ input_shapes,`
	`570`	`+ output_shapes,`
`574`	`571`	`target_);`
`575`	`572`	`}`
`576`	`573`
Original file line number	Diff line number	Diff line change
`@@ -39,7 +39,8 @@ void TuneTask::Initialize(`
`39`	`39`	`op_lowerer = lower_handler;`
`40`	`40`
`41`	`41`	`// Set lowered_funcs and analyze output names.`
`42`		`- this->lowered_funcs = op_lowerer->LowerWithoutSchedule(subgraph);`
	`42`	`+ this->lowered_funcs = op_lowerer->Lower(`
	`43`	`+ subgraph, /apply_op_schedule = / false, /apply_group_schedule=/false);`
`43`	`44`	`this->output_names = GetOutputNamesFromLoweredFunc(this->lowered_funcs);`
`44`	`45`	`this->serialized_key = SerializeToString(shape_dict, dtype_dict);`
`45`	`46`	`}`
Original file line number	Diff line number	Diff line change
`@@ -157,7 +157,9 @@ class PerformanceTester : public ::testing::Test {`
`157`	`157`
`158`	`158`	`for (auto group : graph->fusion_groups) {`
`159`	`159`	`compile_options.lowered_funcs.push_back(`
`160`		`- op_lowerer->LowerWithoutSchedule(group));`
	`160`	`+ op_lowerer->Lower(group,`
	`161`	`+ /apply_op_schedule = / false,`
	`162`	`+ /apply_group_schedule=/false));`
`161`	`163`	`}`
`162`	`164`
`163`	`165`	`VLOG(3) << "===========================No Schedule LoweredFunc "`