用OpenCL实现HEVC中ME模块的测试数据分析

HappinessCat 2013-06-06

使用opencl来实现编码算法中运动搜索模块!

下面测试数据时在GTX570上的测试结果:

LCU为32x32, 100帧720P, CPU上纯C算法使用搜索时间是67s, GPU上是0.915s

LCU为16x16, CPU 是76.8s, GPU上是1.6s

LCU为8x8, CPU 是82.5s, GPU上是4.2s

 

用OpenCL实现HEVC中ME模块的测试数据分析

 

同样的程序, CPU改为SSE实现, GPU做一个小的改动, 使用缩减算法! 结果如下:

 

用OpenCL实现HEVC中ME模块的测试数据分析

从上面数据可以看出, sse 比C语言快5倍左右, 新的GPU kenel快了20%左右, 其中LCU为8x8的快了好几倍!

 

综合看来

OPENCL实现 比C语言实现接近100倍的级别, 比SSE快了接近20倍左右!

另外提一句, 如果OPENCL不适用__local 内存的话, 会慢一半!

下面贴出部分代码供参考:

 #define SearchRange 16
#define Edge_SIZE_T 48
//32x32 version of kernel
__kernel void opencl_me_32x32(const __global short* p_ref, __global short* p_cur, __global int* outputBuf, __local int* local_refBuf, __local int* local_curBuf, __local int* mv_cost)
{
 int searchrange = SearchRange;
 int edeg = Edge_SIZE_T;
 int width = get_global_size(0);
 int height = get_global_size(1);
 int block_w = get_local_size(0);
 int block_h = get_local_size(1);
 int local_x = get_local_id(0);
 int local_y = get_local_id(1);
 int lcu_x = get_group_id(0);
 int lcu_y = get_group_id(1);
 int stride = width + 2 * edeg;
 int lcu_adr_offset = edeg * stride + edeg;
 int local_refBuf_stride = block_w + 2 * searchrange;
   
 //LCU blcok adr
 lcu_adr_offset += lcu_y * stride * block_h + lcu_x * block_w;
 int ref_lcu_adr_offset = lcu_adr_offset - searchrange - searchrange * stride;
 //thread adr
 int global_thread_adr_offset = local_y * stride + local_x;
 int thread_adr_offset  = local_y * local_refBuf_stride + local_x;

 local_curBuf[local_y * block_w + local_x]          = p_cur[lcu_adr_offset + global_thread_adr_offset];

 local_refBuf[thread_adr_offset]            = p_ref[ref_lcu_adr_offset + global_thread_adr_offset ];
 local_refBuf[thread_adr_offset + block_w]         = p_ref[ref_lcu_adr_offset + global_thread_adr_offset + block_w];
 local_refBuf[thread_adr_offset + local_refBuf_stride * block_h]    = p_ref[ref_lcu_adr_offset + global_thread_adr_offset + stride * block_h];
 local_refBuf[thread_adr_offset + local_refBuf_stride * block_h + block_w] = p_ref[ref_lcu_adr_offset + global_thread_adr_offset + stride * block_h + block_w];

 barrier(CLK_LOCAL_MEM_FENCE);
  {
  int i;
  int uiSum = 0;
  for( int i = 0; i < block_h; i++ )
  {

计算sad

}
{
  比较最小SAD 保存bestcost
}
 if((local_y ==0) && (local_x == 0))
 {
  int best_sad  = mv_cost[local_y*2*SearchRange*3 + local_x*3 + 2];
  int best_mvx  = mv_cost[local_y*2*SearchRange*3 + local_x*3 + 0];
  int best_mvy  = mv_cost[local_y*2*SearchRange*3 + local_x*3 + 1];
  outputBuf[(lcu_y * get_num_groups(0) + lcu_x)*3 + 0] = best_mvx;
  outputBuf[(lcu_y * get_num_groups(0) + lcu_x)*3 + 1] = best_mvy;
  outputBuf[(lcu_y * get_num_groups(0) + lcu_x)*3 + 2] = best_sad;
  //printf("\nxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx = %d, y = %d, sad = %d",best_mvx, best_mvy, best_sad);
 }