https://community.intel.com/t5/OpenCL-for-CPU/Performance-of-quot-intel-sub-group-block-readN-writeN-quot-vs/m-p/1099493#M5105 <P>The driver can help a lot with optimizing memory transfers and can often get to similar results. &nbsp; However, if you're already going to the trouble of using the subgroup reads and writes this can help guarantee you're getting optimal memory buffer bandwidth which could give some advantages over vload/vstore, using vector data types for memory I/O, etc. &nbsp;</P> <UL> <LI>Subgroup read/write is closer to what the driver would try to optimize for in any case. &nbsp;If you've already arranged your data I/O to work this way this should be an optimal data access approach (for linear buffers) which will touch a minimal # of cache lines to maximize cache efficienc</LI> <LI>This also optimizes the calculations needed to compute addresses and the # of addresses that need to be passed to the driver. With &nbsp;subgroups only the address of the first item in the block and a length is sent, vs. an address for every work item in the subgroup</LI> </UL> Fri, 02 Dec 2016 16:36:37 GMT Jeffrey_M_Intel1 2016-12-02T16:36:37Z https://community.intel.com/t5/OpenCL-for-CPU/Performance-of-quot-intel-sub-group-block-readN-writeN-quot-vs/m-p/1099492#M5104 <P>Does&nbsp;subgroup extension API "intel_sub_group_block_readN/writeN" have better performance than "vloadN/vstoreN"? I did some testing, but don't see much difference between them.&nbsp; Can you elaborate the read/write&nbsp; performance expectation between them?</P> Tue, 29 Nov 2016 08:21:40 GMT https://community.intel.com/t5/OpenCL-for-CPU/Performance-of-quot-intel-sub-group-block-readN-writeN-quot-vs/m-p/1099492#M5104 Shengquan_Y_Intel 2016-11-29T08:21:40Z https://community.intel.com/t5/OpenCL-for-CPU/Performance-of-quot-intel-sub-group-block-readN-writeN-quot-vs/m-p/1099493#M5105 <P>The driver can help a lot with optimizing memory transfers and can often get to similar results. &nbsp; However, if you're already going to the trouble of using the subgroup reads and writes this can help guarantee you're getting optimal memory buffer bandwidth which could give some advantages over vload/vstore, using vector data types for memory I/O, etc. &nbsp;</P> <UL> <LI>Subgroup read/write is closer to what the driver would try to optimize for in any case. &nbsp;If you've already arranged your data I/O to work this way this should be an optimal data access approach (for linear buffers) which will touch a minimal # of cache lines to maximize cache efficienc</LI> <LI>This also optimizes the calculations needed to compute addresses and the # of addresses that need to be passed to the driver. With &nbsp;subgroups only the address of the first item in the block and a length is sent, vs. an address for every work item in the subgroup</LI> </UL> Fri, 02 Dec 2016 16:36:37 GMT https://community.intel.com/t5/OpenCL-for-CPU/Performance-of-quot-intel-sub-group-block-readN-writeN-quot-vs/m-p/1099493#M5105 Jeffrey_M_Intel1 2016-12-02T16:36:37Z