This is a placeholder issue for dealing with OpenCL tuning for the SD820 platform. I plan to add/edit details as warranted, time/effort/interest permitting. In general, more complete information and/or reproducers are available upon request. However, the goal is to get something up sooner rather than later for these issues.
Currently, I'm just learning about the SD820 platform and its tools. I suspect there may be various 'level 0' things I'm doing wrong or am generally unaware of. That said, here's a summary of my current issues:
1) lack of insight into hardware and toolchain makes tuning a matter of error prone and time consuming guesswork
2) general lack of documentation: hardware, tools, software
3) bug: some 512x512 sgemm variants fail. compiler bug?
4) profiler(s) UI buggy and very slow/awkward to use (both AP and SDP).
(1) is pretty general. my wishlist includes a disassembler, a tuning guide, a hardware guide, examples of high-perf low-level code, etc ... all the things you get for Maxwell/Pascal currently (more or less) -- and maybe now for AMD too with ROCmumble.
(2) again pretty general. wishlist includes ISA documentation, compiler source/documentation, hardware descriptions, profiler documentation (i.e. what the kernel metrics actually mean). it looks like the current SDK doesn't even include any OpenCL docs or examples anymore -- that's pretty worrisome!
for (3), i probably need to minimize the example and wrap it up in it's own issue before i'd expect anyone would want to look at it. but if anyone is eager/willing, let me know, and i'll bump up the priority. for now, see doc/sgemm-notes.txt#L54
(4) command line, non-broken, automated profiler would be nice. or failing that, support for the existing one(s) (community or otherwise). the forums are littered with unanswered posts ...
This is a placeholder issue for dealing with OpenCL tuning for the SD820 platform. I plan to add/edit details as warranted, time/effort/interest permitting. In general, more complete information and/or reproducers are available upon request. However, the goal is to get something up sooner rather than later for these issues.
Currently, I'm just learning about the SD820 platform and its tools. I suspect there may be various 'level 0' things I'm doing wrong or am generally unaware of. That said, here's a summary of my current issues:
1) lack of insight into hardware and toolchain makes tuning a matter of error prone and time consuming guesswork 2) general lack of documentation: hardware, tools, software 3) bug: some 512x512 sgemm variants fail. compiler bug? 4) profiler(s) UI buggy and very slow/awkward to use (both AP and SDP).
(1) is pretty general. my wishlist includes a disassembler, a tuning guide, a hardware guide, examples of high-perf low-level code, etc ... all the things you get for Maxwell/Pascal currently (more or less) -- and maybe now for AMD too with ROCmumble. (2) again pretty general. wishlist includes ISA documentation, compiler source/documentation, hardware descriptions, profiler documentation (i.e. what the kernel metrics actually mean). it looks like the current SDK doesn't even include any OpenCL docs or examples anymore -- that's pretty worrisome! for (3), i probably need to minimize the example and wrap it up in it's own issue before i'd expect anyone would want to look at it. but if anyone is eager/willing, let me know, and i'll bump up the priority. for now, see doc/sgemm-notes.txt#L54 (4) command line, non-broken, automated profiler would be nice. or failing that, support for the existing one(s) (community or otherwise). the forums are littered with unanswered posts ...