fix FLAGS_torch_compatible_kernel to FLAGS_use_accuracy_compatible_kernel (PaddlePaddle#76231)

Author: zhengshengning

paddle/common/flags.cc

@@ -2305,12 +2305,12 @@ PHI_DEFINE_EXPORTED_bool(
 
 /**
  * Torch Compatible related FLAG
- * Name: FLAGS_torch_compatible_kernel
+ * Name: FLAGS_use_accuracy_compatible_kernel
  * Since Version: 3.2.2
  * Value Range: bool, default=false
  * Example:
  * Note: Whether use torch compatible version kernel.
  */
-PHI_DEFINE_EXPORTED_bool(torch_compatible_kernel,
+PHI_DEFINE_EXPORTED_bool(use_accuracy_compatible_kernel,
                          false,
                          "Whether use torch compatible version kernel.");

paddle/fluid/pir/transforms/pd_op_to_kernel_pass.cc

@@ -74,7 +74,7 @@ COMMON_DECLARE_bool(use_onednn);
 
 COMMON_DECLARE_bool(print_ir);
 COMMON_DECLARE_bool(enable_collect_shape);
-COMMON_DECLARE_bool(torch_compatible_kernel);
+COMMON_DECLARE_bool(use_accuracy_compatible_kernel);
 REGISTER_FILE_SYMBOLS(pd_op_to_kernel_pass);
 namespace paddle::dialect {
 
@@ -283,7 +283,7 @@ static bool NeedFallBackFromGPUDNN2GPU(pir::Operation* op,
     }
   } else if ((op->isa<AffineGridOp>() || op->isa<AffineGridGradOp>()) &&
              kernel_key.backend() == phi::Backend::GPUDNN) {
-    if (FLAGS_torch_compatible_kernel) {
+    if (FLAGS_use_accuracy_compatible_kernel) {
       return true;
     }
     bool use_cudnn = true;

paddle/phi/kernels/funcs/pooling.cu

@@ -31,7 +31,7 @@ limitations under the License. */
 #include "paddle/phi/kernels/funcs/random.cuh"
 #include "paddle/phi/kernels/funcs/reduce_function.h"
 
-COMMON_DECLARE_bool(torch_compatible_kernel);
+COMMON_DECLARE_bool(use_accuracy_compatible_kernel);
 
 namespace phi {
 namespace funcs {
@@ -988,7 +988,7 @@ class MaxPool2dGradFunctor<phi::GPUContext, T> {
         output.numel() <= std::numeric_limits<int>::max()) {
       auto pool_divmods = FastDivModForPooling<int>(
           input_channels, output_width, output_height);
-      if (FLAGS_torch_compatible_kernel) {
+      if (FLAGS_use_accuracy_compatible_kernel) {
         int64_t blocks =
             (input_width * input_height + kBlockThreads - 1) / kBlockThreads;
         dim3 grid(blocks, batch_size, input_channels);
@@ -1042,7 +1042,7 @@ class MaxPool2dGradFunctor<phi::GPUContext, T> {
     } else {
       auto pool_divmods = FastDivModForPooling<int64_t>(
           input_channels, output_width, output_height);
-      if (FLAGS_torch_compatible_kernel) {
+      if (FLAGS_use_accuracy_compatible_kernel) {
         int64_t blocks =
             (input_width * input_height + kBlockThreads - 1) / kBlockThreads;
         dim3 grid(blocks, batch_size, input_channels);

python/paddle/nn/functional/vision.py

@@ -100,8 +100,8 @@ def affine_grid(
             False  # ROCM platform do not have MIOPEN kernel for affine_grid
         )
 
-    if paddle.get_flags(["FLAGS_torch_compatible_kernel"]).get(
-        "FLAGS_torch_compatible_kernel", False
+    if paddle.get_flags(["FLAGS_use_accuracy_compatible_kernel"]).get(
+        "FLAGS_use_accuracy_compatible_kernel", False
     ):
         use_cudnn = False
 

test/legacy_test/test_affine_grid_op.py

@@ -253,7 +253,7 @@ def run(place):
 
 class TestAffineGridOpTorchCompatible1(TestAffineGridOp):
     def initTestCase(self):
-        paddle.set_flags({'FLAGS_torch_compatible_kernel': 1})
+        paddle.set_flags({'FLAGS_use_accuracy_compatible_kernel': 1})
         self.theta_shape = (20, 2, 3)
         self.output_shape = np.array([20, 2, 5, 7]).astype("int32")
         self.dynamic_shape = True
@@ -263,7 +263,7 @@ def initTestCase(self):
 
 class TestAffineGridOpTorchCompatible2(TestAffineGridOp):
     def initTestCase(self):
-        paddle.set_flags({'FLAGS_torch_compatible_kernel': 1})
+        paddle.set_flags({'FLAGS_use_accuracy_compatible_kernel': 1})
         self.theta_shape = (20, 3, 4)
         self.output_shape = np.array([20, 1, 2, 5, 7]).astype("int32")
         self.dynamic_shape = True

test/legacy_test/test_pool2d_api.py

@@ -770,7 +770,7 @@ def test_pool2d_static(self):
         paddle.disable_static()
 
     def test_torch_compatible(self):
-        paddle.set_flags({'FLAGS_torch_compatible_kernel': 1})
+        paddle.set_flags({'FLAGS_use_accuracy_compatible_kernel': 1})
         paddle.enable_static()
         for place in self.places:
             self.check_max_static_results(place)