Radeon 上的 OpenCL (aparapi) 简单还原速度慢

OpenCL (aparapi) simple reduction slow on Radeon

我正在尝试对 OpenCL 中的大型双精度数组进行简单归约(在本例中为总和)。我看了网上的教程,发现基本上就是这样解决我的问题:

#pragma OPENCL EXTENSION cl_khr_fp64 : enable

typedef struct This_s{
   __global double *nums;
   int nums__javaArrayLength;
   __local double *buffer;
   __global double *res;
   int passid;
}This;
int get_pass_id(This *this){
   return this->passid;
}
__kernel void run(
   __global double *nums, 
   int nums__javaArrayLength, 
   __local double *buffer, 
   __global double *res, 
   int passid
){
   This thisStruct;
   This* this=&thisStruct;
   this->nums = nums;
   this->nums__javaArrayLength = nums__javaArrayLength;
   this->buffer = buffer;
   this->res = res;
   this->passid = passid;
   {
      int tid = get_local_id(0);
      int i = (get_group_id(0) * get_local_size(0)) + get_local_id(0);
      int gridSize = get_local_size(0) * get_num_groups(0);
      int n = this->nums__javaArrayLength;
      double cur = 0.0;
      for (; i<n; i = i + gridSize){
         cur = cur + this->nums[i];
      }
      this->buffer[tid]  = cur;
      barrier(CLK_LOCAL_MEM_FENCE);
      barrier(CLK_LOCAL_MEM_FENCE);
      if (tid<32){
         this->buffer[tid]  = this->buffer[tid] + this->buffer[(tid + 32)];
      }
      barrier(CLK_LOCAL_MEM_FENCE);
      if (tid<16){
         this->buffer[tid]  = this->buffer[tid] + this->buffer[(tid + 16)];
      }
      barrier(CLK_LOCAL_MEM_FENCE);
      if (tid<8){
         this->buffer[tid]  = this->buffer[tid] + this->buffer[(tid + 8)];
      }
      barrier(CLK_LOCAL_MEM_FENCE);
      if (tid<4){
         this->buffer[tid]  = this->buffer[tid] + this->buffer[(tid + 4)];
      }
      barrier(CLK_LOCAL_MEM_FENCE);
      if (tid<2){
         this->buffer[tid]  = this->buffer[tid] + this->buffer[(tid + 2)];
      }
      barrier(CLK_LOCAL_MEM_FENCE);
      if (tid<1){
         this->buffer[tid]  = this->buffer[tid] + this->buffer[(tid + 1)];
      }
      barrier(CLK_LOCAL_MEM_FENCE);
      if (tid==0){
         this->res[get_group_id(0)]  = this->buffer[0];
      }
      return;
   }
}

如果您想了解奇怪的 this,那是 aparapi 的一个(不幸的是必需的)工件,我用它来将 Java 转换为 OpenCL。

我的内核产生了正确的结果,并且在相当强大的 Nvidia 硬件上,它比 Java 中的顺序求和快大约 10 倍。然而,在 Radeon R9 280 上,它的性能与简单的 Java 代码相当。

我已经使用 CodeXL 分析了内核。它告诉我 MemUnitBusy 仅占 6%。为什么这么低?

原来 OpenCL 没有(直接)错误,但 aparapis 缓冲区管理是。

我在没有 aparapi 的情况下尝试了完全相同的内核,性能很好。我一使用 CL_MEM_USE_HOST_PTR 就变坏了,遗憾的是,这是使用 aparapi 时唯一的选择。似乎 AMD 没有将主机内存复制到具有该选项的设备,即使在运行了几次 "warmup" 之后也是如此。

您可能要考虑转到 aparapi.com. It includes several fixes to bugs and a lot of extra features and performance enhancements over the older library you linked above. It is also in maven central with about a dozen releases. so it is easier to use. The new Github repository is here 上更活跃的项目。