Fix gru as small frame_size has error. (#20922)

seems shuffle_sync cannot handle small size

test=develop

Signed-off-by: zhaoyuchen <zhaoyuchen01@baidu.com>

paddle/fluid/operators/math/detail/gru_gpu_kernel.h

@@ -105,7 +105,7 @@ __global__ void KeGruForwardFinalOutput(OpFinalOutput op_final_output,
  * threads(tile_size, 1)
  * grid(frame_blocks, 1)
  */
-template <class T, int Tiled_size>
+template <class T>
 __global__ void KeFastCollectiveGruGate(T *gate_value, T *prev_output_value,
                                         T *gate_weight, T *reset_output,
                                         int frame_size,
@@ -113,7 +113,9 @@ __global__ void KeFastCollectiveGruGate(T *gate_value, T *prev_output_value,
   T xt_0 = 0.0f;
   T a0 = 0.0f;
   T c0 = 0.0f;
-  T b0[Tiled_size];
+
+  int Tiled_size = blockDim.x;
+  T b0[16];
 
   int COL = blockIdx.x * blockDim.x + threadIdx.x;
   int Tiled_mask = ((1 << Tiled_size) - 1);
@@ -163,7 +165,7 @@ __global__ void KeFastCollectiveGruGate(T *gate_value, T *prev_output_value,
  * threads(tile_size, 1)
  * grid(frame_blocks, 1)
  */
-template <class T, int Tiled_size>
+template <class T>
 __global__ void KeFastCollectiveGruOut(T *gate_weight, T *prev_out_value,
                                        T *output_value, T *gate_value,
                                        T *reset_value, int frame_size,
@@ -172,9 +174,10 @@ __global__ void KeFastCollectiveGruOut(T *gate_weight, T *prev_out_value,
   int COL = blockIdx.x * blockDim.x + threadIdx.x;
 
   T a0 = 0.0f;
-  T b0[Tiled_size];
+  T b0[16];
   T c0 = 0.0f;
 
+  int Tiled_size = blockDim.x;
   int Tiled_mask = ((1 << Tiled_size) - 1);
   //- Tiled  matrix multiply with register shift
   if (prev_out_value) {

paddle/fluid/operators/math/gru_compute.cu

@@ -31,19 +31,25 @@ struct GRUUnitFunctor<platform::CUDADeviceContext, T> {
     dim3 grid;
     if (batch_size == 1) {
       if (context.GetComputeCapability() >= 70) {
-        constexpr int tiled_size = 16;
+        auto ComputeTiledSize = [](int frame_size) {
+          if (frame_size >= 16)
+            return 16;
+          else if (frame_size < 16)
+            return 8;
+        };
+
+        auto tiled_size = ComputeTiledSize(frame_size);
         int frame_blocks = (frame_size * 2 + tiled_size - 1) / tiled_size;
         threads = dim3(tiled_size, 1);
         grid = dim3(frame_blocks, 1);
-        detail::KeFastCollectiveGruGate<
-            T, tiled_size><<<grid, threads, 0, stream>>>(
+
+        detail::KeFastCollectiveGruGate<T><<<grid, threads, 0, stream>>>(
             value.gate_value, value.prev_out_value, value.gate_weight,
             value.reset_output_value, frame_size, active_gate);
 
         frame_blocks = (frame_size + tiled_size - 1) / tiled_size;
         grid = dim3(frame_blocks, 1);
-        detail::KeFastCollectiveGruOut<
-            T, tiled_size><<<grid, threads, 0, stream>>>(
+        detail::KeFastCollectiveGruOut<T><<<grid, threads, 0, stream>>>(
             value.state_weight, value.prev_out_value, value.output_value,
             value.gate_value, value.reset_output_value, frame_size, active_node,
             origin_mode);