Modify the implementation of BlockYReduce to fit more scenes (#39170)

paddle/pten/kernels/primitive/compute_primitives.h

@@ -118,7 +118,7 @@ __device__ __forceinline__ T BlockXReduce(T val, ReduceOp reducer) {
  */
 template <typename T, typename ReduceOp>
 __device__ __forceinline__ T BlockYReduce(T val, ReduceOp reducer) {
-  __shared__ T shared_memory[details::kReduceMaxThread];
+  __shared__ T shared_memory[1024];
   shared_memory[SharedMemoryIndex(0)] = val;
   for (int stride = blockDim.y / 2; stride > 0; stride >>= 1) {
     __syncthreads();
@@ -128,7 +128,8 @@ __device__ __forceinline__ T BlockYReduce(T val, ReduceOp reducer) {
     }
     shared_memory[SharedMemoryIndex(0)] = val;
   }
-  return val;
+  __syncthreads();
+  return shared_memory[threadIdx.x];
 }
 
 }  // namespace details