OpenCL (aparapi) simple reduction slow on Radeon

I am trying to code a simple reduction (in this case a sum) over a large double array in OpenCL. I have looked at online tutorials and found that this is essentially the way to solve my problem:

#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;
   }
}

If you are wondering about the strange this, that is an (unfortunately necessary) artifact of aparapi, which I use to translate Java to OpenCL.

My kernel produces the correct results and, on reasonably beefy Nvidia Hardware, it is about 10x faster than a sequential sum in Java. On a Radeon R9 280 however it is comparable in performance to the simple Java code.

I have profiled the kernel with CodeXL. It tells me that MemUnitBusy is at just 6%. Why is it so low?

Turns out OpenCL is not (directly) at fault, but aparapis buffer management is.

I tried out the exact same kernel without aparapi, and the performance is good. It turns bad as soon as I use CL_MEM_USE_HOST_PTR, which is sadly the only option when using aparapi. It seems AMD is not copying host memory to the device with that option, even after several "warmup" runs.