make bilinear interpolate stable. (#48644)

* make bilinear interpolate stable.

* fix code

paddle/phi/kernels/gpu/interpolate_grad_kernel.cu

@@ -25,6 +25,8 @@
 #include "paddle/phi/kernels/funcs/math_function.h"
 #include "paddle/phi/kernels/primitive/datamover_primitives.h"
 
+DECLARE_bool(cudnn_deterministic);
+
 namespace phi {
 
 template <typename T>
@@ -1034,6 +1036,12 @@ static void Interpolate2DCUDABwd(
 #endif
 
     if (optimize_flag & is_nchw) {
+      if (FLAGS_cudnn_deterministic) {
+        VLOG(2)
+            << "Run grad kernel of bilinear interpolate 2d with single thread.";
+        config.block_per_grid = 1;
+        config.thread_per_block = 1;
+      }
       KeBilinearInterpBwShareMemory<T><<<config.block_per_grid,
                                          config.thread_per_block,
                                          0,
@@ -1052,11 +1060,17 @@ static void Interpolate2DCUDABwd(
     } else if (!optimize_flag & is_nchw) {
       const int num_kernels = n * c * out_h * out_w;
       const int num_threads = std::min(dev_ctx.GetMaxThreadsPerBlock(), 1024);
+      int block_per_grid = backends::gpu::DivUp(num_kernels, num_threads);
+      int thread_per_block = num_threads;
+      if (FLAGS_cudnn_deterministic) {
+        VLOG(2)
+            << "Run grad kernel of bilinear interpolate 2d with single thread.";
+        block_per_grid = 1;
+        thread_per_block = 1;
+      }
       KeBilinearInterpNCHWBw<T>
-          <<<backends::gpu::DivUp(num_kernels, num_threads),
-             num_threads,
-             0,
-             dev_ctx.stream()>>>(input_grad_data,
+          <<<block_per_grid, thread_per_block, 0, dev_ctx.stream()>>>(
+              input_grad_data,
               in_h,
               in_w,
               out_h,