修改注释

S1/41/reduce_sum_algorithm.maca

@@ -26,13 +26,12 @@ constexpr double REDUCE_ERROR_TOLERANCE = 0.005; // 0.5%
 
 #if !USE_DEFAULT_REF_IMPL
 
-constexpr int BLOCK_SIZE = 256;
+constexpr int BLOCK_SIZE = 512;
 constexpr int WARP_SIZE = 32;
 
 // 1. Warp 级归约：使用寄存器洗牌指令，无需 Shared Mem，速度极快
 template <typename T>
 __device__ __forceinline__ T warpReduceSum(T val) {
-    // 假设 warpSize 为 32
     #pragma unroll
     for (int offset = WARP_SIZE / 2; offset > 0; offset /= 2) {
         val += __shfl_down_sync(0xffffffff, val, offset);
@@ -79,7 +78,6 @@ __global__ void reduceKernel(const T* __restrict__ d_in, T* __restrict__ d_out,
     T sum = 0;
 
     // Grid-Stride Loop: 处理数据量大于线程总数的情况
-    // 这种模式能保证良好的内存合并访问
     int thread_id = blockIdx.x * blockDim.x + threadIdx.x;
     int stride = blockDim.x * gridDim.x;