<div style="line-height:1.7;color:#000000;font-size:14px;font-family:Arial"><div>1. clCreateBuffer(CL_MEM_USE_HOST_PTR) 2. clMapxx-->change the data pointed by host ptr--->clUnmapxxx Is that right? I think clUnmapxx must contain cache flush operation although it may not involve extra copy, correct me if am wrong. I am very concerned about the beignet performance on our product. </div> <div></div><div id="divNeteaseMailCard"></div> 在 2014-12-03 10:14:10，"Guo, Yejun" <yejun.guo@intel.com> 写道： <blockquote id="isReplyContent" style="PADDING-LEFT: 1ex; MARGIN: 0px 0px 0px 0.8ex; BORDER-LEFT: #ccc 1px solid"> <style></style> <div class="WordSection1"> Just need to invoke clMapxx function before you access the host ptr from CPU side, and invoke clUnmapxx function after the CPU access. There is no extra copy inside clMap/Unmapxx if all the conditions are satisfied. You can refer to the code in runtime_use_host_ptr_buffer.cpp under beignet/utests.<o:p></o:p> <o:p> </o:p> <o:p> </o:p> From: spring_wind [mailto:<a href="mailto:spring_wind@yeah.net">spring_wind@yeah.net</a>] Sent: Wednesday, December 03, 2014 10:02 AM To: Guo, Yejun Cc: <a href="mailto:beignet@lists.freedesktop.org">beignet@lists.freedesktop.org</a> Subject: Re:RE: [Beignet] CL_MEM_USE_HOST_PTR involve extra copy?<o:p></o:p> <o:p> </o:p> <div> You mean if I use CL_MEM_USE_HOST_PTR and host ptr is page aligned, I change the data pointed by host ptr from CPU side, I don't have to do any flush opertaion or I should still call clMapxx function? <o:p></o:p> 在 2014-12-03 08:37:15，"Guo, Yejun" <<a href="mailto:yejun.guo@intel.com">yejun.guo@intel.com</a>> 写道： <o:p></o:p> <div> Hi,<o:p></o:p> <o:p></o:p> please check the latest code of beignet, there is no copy needed between CPU and GPU if the host_ptr provided by application is page aligned, and the page align limitation is expected to be removed some days later. You can also try CL_MEM_ALLOC_HOST_PTR to avoid the extra copy without align limitation.<o:p></o:p> <o:p></o:p> Btw, this zero-copy is supported starting from linux kernel 3.16 and libdrm 2.4.58.<o:p></o:p> <o:p></o:p> <o:p></o:p> From: Beignet [mailto:<a href="mailto:beignet-bounces@lists.freedesktop.org">beignet-bounces@lists.freedesktop.org</a>] On Behalf Of spring_wind Sent: Tuesday, December 02, 2014 8:38 PM To: <a href="mailto:beignet@lists.freedesktop.org">beignet@lists.freedesktop.org</a> Subject: [Beignet] CL_MEM_USE_HOST_PTR involve extra copy?<o:p></o:p> <o:p></o:p> <div> <div> Hi:<o:p></o:p> </div> <div> Intel opencl optimization guide said using CL_MEM_USE_HOST_PTR can avoid extra copy between CPU and GPU, but I noticed that in beignet's implementation source it was not like that:<o:p></o:p> </div> <div> <div> /* Copy the data if required */<o:p></o:p> </div> <div> if (flags & CL_MEM_COPY_HOST_PTR || flags & CL_MEM_USE_HOST_PTR)<o:p></o:p> </div> <div> cl_buffer_subdata(mem->bo, 0, sz, data);<o:p></o:p> </div> </div> <div> <o:p></o:p> </div> <div> Could someone give me an answer?<o:p></o:p> </div> </div> <o:p></o:p> </div> </div> <o:p> </o:p> </div> </blockquote></div> <a href="#" target="_blank"></a>