Hello List!
I am looking at top-down vs. bottom-up list scheduling for simple(r) in-order cores. First, for some context, below is a fairly representative pseudo-code example of the sort of DSP-like codes I am looking at: uint64_t foo(int *pA, int *pB, unsigned N, unsigned C) { uint64_t sum = 0; while (N-- > 0) { A1 = *pA++; A2 = *pA++; B1 = *pB++; B2 = *pB++; ... sum += ((uint64_t) A1 * B1) >> C; sum += ((uint64_t) A2 * B2) >> C; ... } return sum; } These kernels are very tight loops. In this sort of legacy codes it's not uncommon that loops are manually tuned and unrolled with the available registers in mind, so that all values just about fit in registers. In the example above, we read from input streams A and B, and do 2 multiply-accumulate operations. The loop is unrolled 2 times in this example, but 4, 8 or 16 would more realistic, resulting in very high register pressure. But legacy apps written in this way are not the only culprit; we see that with (aggressive) loop unrolling we can end up in exactly the same situation: the loopunroller does exactly what it needs to do, but it results in very high register pressure. Here's the (obvious) problem: all live values in these loops should just about fit in registers, but suboptimal/wrong decisions are made resulting in a lot of spills/reloads; the machine scheduler is making the life of the register allocator very difficult. I am looking at regressions of about 30 - 40% for more than a handful of kernels, and am thus very interested in what I could do about this. The first observation is that it looks like we default to bottom-up scheduling. Starting bottom-up, all these "sum" variables are scheduled first, and after that the loads (this is simplifying things a bit). And thus it looks like we create a lot of very long live-ranges, causing the problems mentioned above. When instead we start top-down, the loads are picked up first, then the MAC, and this repeated for all MACs. The result is a sequence LD, MAC, LD, MAC, etc., which is let's say a sequence more in program-order, also with shorter live-ranges. This does exactly what I want for these sort of kernels, it generates exactly the code we want. While playing with this bottom-up/top-down order, it didn't take long to see that the top-down approach is excellent for these kind of codes, but not for some other benchmarks, and so solving this issue not just a matter of defaulting to a new scheduling policy. I am thinking about prototyping an approach where we start with the bottom-up approach (under a new misched option), but when a register-pressure/live-range trashhold value is reached, we bail and fall back to the top-down approach. This is my first rough idea, but I haven't looked at this problem for very long. My first few data points is suggesting this might be benificial, but I could be missing a lot here. And also I am sure that I am looking at an old/classic problem here and I'm sure others have looked at this problem before. Thus I am wondering if people have experiences/opinions on this. Cheers, Sjoerd. IMPORTANT NOTICE: The contents of this email and any attachments are confidential and may also be privileged. If you are not the intended recipient, please notify the sender immediately and do not disclose the contents to any other person, use it for any purpose, or store or copy the information in any medium. Thank you. _______________________________________________ LLVM Developers mailing list [hidden email] http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev |
On 11/6/2018 6:19 AM, Sjoerd Meijer via llvm-dev wrote:
> Hello List! > > I am looking at top-down vs. bottom-up list scheduling for simple(r) in-order > cores. First, for some context, below is a fairly representative pseudo-code > example of the sort of DSP-like codes I am looking at: > > uint64_t foo(int *pA, int *pB, unsigned N, unsigned C) { > uint64_t sum = 0; > while (N-- > 0) { > A1 = *pA++; > A2 = *pA++; > B1 = *pB++; > B2 = *pB++; > ... > sum += ((uint64_t) A1 * B1) >> C; > sum += ((uint64_t) A2 * B2) >> C; > ... > } > return sum; > } > > These kernels are very tight loops. In this sort of legacy codes it's not > uncommon that loops are manually tuned and unrolled with the available > registers in mind, so that all values just about fit in registers. In the > example above, we read from input streams A and B, and do 2 multiply-accumulate > operations. The loop is unrolled 2 times in this example, but 4, 8 or 16 would > more realistic, resulting in very high register pressure. > > But legacy apps written in this way are not the only culprit; we see that with > (aggressive) loop unrolling we can end up in exactly the same situation: the > loopunroller does exactly what it needs to do, but it results in very high > register pressure. > > Here's the (obvious) problem: all live values in these loops should just about > fit in registers, but suboptimal/wrong decisions are made resulting in a lot of > spills/reloads; the machine scheduler is making the life of the register allocator > very difficult. I am looking at regressions of about 30 - 40% for more > than a handful of kernels, and am thus very interested in what I could do about > this. > > The first observation is that it looks like we default to bottom-up scheduling. > Starting bottom-up, all these "sum" variables are scheduled first, and after > that the loads (this is simplifying things a bit). And thus it looks like we > create a lot of very long live-ranges, causing the problems mentioned above. > When instead we start top-down, the loads are picked up first, then the MAC, and > this repeated for all MACs. The result is a sequence LD, MAC, LD, MAC, etc., > which is let's say a sequence more in program-order, also with shorter > live-ranges. This does exactly what I want for these sort of kernels, it generates > exactly the code we want. Do you know why the existing register pressure heuristics don't work for your testcase with the bottom-up scheduler? -Eli -- Employee of Qualcomm Innovation Center, Inc. Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum, a Linux Foundation Collaborative Project _______________________________________________ LLVM Developers mailing list [hidden email] http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev |
In reply to this post by Sudhindra kulkarni via llvm-dev
Some comments:
- It's all heuristics, so you may be lucky/unlycky in any configuration. I assume you know how to read -debug-only=machine-scheduler output and the GenericScheduler::tryCandidate() code to figure out how most decisions are made. - In my experience bottom-up works better when optimizing for register pressure while top-down tends to work better when optimizing for latencies/stall reduction. - The scheduler has a couple rules dealing with pressure. Hitting any of these limits will make it pick pressure reducing nodes more aggressively (look for tryPressure() in the tryCandidate() code) - Hitting the target limits (number of available registers) - Hitting the maximum of the region before scheduling - Generally attempt to minimmize pressure (this has lower priority though) In the past I looked at some crypto code that also was carefully tuned by a human and that turned out to be very tricky for the heuristics to get right because most values had multiple users. At the time I had a patch that would revert all scheduling decisions if it turned out that the scheduler couldn't find a solution with better or equal register pressure. It wasn't received too enthusiastically and I ended up tweaking some heuristics instead. But we may start that discussion again if you cannot find a know to tune the heuristic for your case. - Matthias > > Hello List! > > I am looking at top-down vs. bottom-up list scheduling for simple(r) in-order > cores. First, for some context, below is a fairly representative pseudo-code > example of the sort of DSP-like codes I am looking at: > > uint64_t foo(int *pA, int *pB, unsigned N, unsigned C) { > uint64_t sum = 0; > while (N-- > 0) { > A1 = *pA++; > A2 = *pA++; > B1 = *pB++; > B2 = *pB++; > ... > sum += ((uint64_t) A1 * B1) >> C; > sum += ((uint64_t) A2 * B2) >> C; > ... > } > return sum; > } > > These kernels are very tight loops. In this sort of legacy codes it's not > uncommon that loops are manually tuned and unrolled with the available > registers in mind, so that all values just about fit in registers. In the > example above, we read from input streams A and B, and do 2 multiply-accumulate > operations. The loop is unrolled 2 times in this example, but 4, 8 or 16 would > more realistic, resulting in very high register pressure. > > But legacy apps written in this way are not the only culprit; we see that with > (aggressive) loop unrolling we can end up in exactly the same situation: the > loopunroller does exactly what it needs to do, but it results in very high > register pressure. > > Here's the (obvious) problem: all live values in these loops should just about > fit in registers, but suboptimal/wrong decisions are made resulting in a lot of > spills/reloads; the machine scheduler is making the life of the register allocator > very difficult. I am looking at regressions of about 30 - 40% for more > than a handful of kernels, and am thus very interested in what I could do about > this. > > The first observation is that it looks like we default to bottom-up scheduling. > Starting bottom-up, all these "sum" variables are scheduled first, and after > that the loads (this is simplifying things a bit). And thus it looks like we > create a lot of very long live-ranges, causing the problems mentioned above. > When instead we start top-down, the loads are picked up first, then the MAC, and > this repeated for all MACs. The result is a sequence LD, MAC, LD, MAC, etc., > which is let's say a sequence more in program-order, also with shorter > live-ranges. This does exactly what I want for these sort of kernels, it generates > exactly the code we want. > > While playing with this bottom-up/top-down order, it didn't take long to see > that the top-down approach is excellent for these kind of codes, but not for > some other benchmarks, and so solving this issue not just a matter of > defaulting to a new scheduling policy. > > I am thinking about prototyping an approach where we start with the bottom-up > approach (under a new misched option), but when a register-pressure/live-range > trashhold value is reached, we bail and fall back to the top-down approach. > This is my first rough idea, but I haven't looked at this problem for very > long. My first few data points is suggesting this might be benificial, but I > could be missing a lot here. And also I am sure that I am looking at an > old/classic problem here and I'm sure others have looked at this problem > before. Thus I am wondering if people have experiences/opinions on this. > > Cheers, > Sjoerd. > IMPORTANT NOTICE: The contents of this email and any attachments are confidential and may also be privileged. If you are not the intended recipient, please notify the sender immediately and do not disclose the contents to any other person, use it for any purpose, or store or copy the information in any medium. Thank you. > _______________________________________________ > LLVM Developers mailing list > [hidden email] > http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev _______________________________________________ LLVM Developers mailing list [hidden email] http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev |
Hi Matthias and Eli,
Thanks for your replies. Yes, I am first going to study this more, to really understand all the different bits and pieces involved here. I indeed need to see if we not accidentally miss something here, or if we are just
unlucky with some heuristics. Your pointers are really useful for that! I will write again when I have more concrete ideas/plans.
Cheers, Sjoerd. From: [hidden email] <[hidden email]> on behalf of Matthias Braun <[hidden email]>
Sent: 06 November 2018 20:55 To: Sjoerd Meijer Cc: [hidden email] Subject: Re: [llvm-dev] top-down vs. bottom-up list scheduling Some comments:
- It's all heuristics, so you may be lucky/unlycky in any configuration. I assume you know how to read -debug-only=machine-scheduler output and the GenericScheduler::tryCandidate() code to figure out how most decisions are made. - In my experience bottom-up works better when optimizing for register pressure while top-down tends to work better when optimizing for latencies/stall reduction. - The scheduler has a couple rules dealing with pressure. Hitting any of these limits will make it pick pressure reducing nodes more aggressively (look for tryPressure() in the tryCandidate() code) - Hitting the target limits (number of available registers) - Hitting the maximum of the region before scheduling - Generally attempt to minimmize pressure (this has lower priority though) In the past I looked at some crypto code that also was carefully tuned by a human and that turned out to be very tricky for the heuristics to get right because most values had multiple users. At the time I had a patch that would revert all scheduling decisions if it turned out that the scheduler couldn't find a solution with better or equal register pressure. It wasn't received too enthusiastically and I ended up tweaking some heuristics instead. But we may start that discussion again if you cannot find a know to tune the heuristic for your case. - Matthias > > Hello List! > > I am looking at top-down vs. bottom-up list scheduling for simple(r) in-order > cores. First, for some context, below is a fairly representative pseudo-code > example of the sort of DSP-like codes I am looking at: > > uint64_t foo(int *pA, int *pB, unsigned N, unsigned C) { > uint64_t sum = 0; > while (N-- > 0) { > A1 = *pA++; > A2 = *pA++; > B1 = *pB++; > B2 = *pB++; > ... > sum += ((uint64_t) A1 * B1) >> C; > sum += ((uint64_t) A2 * B2) >> C; > ... > } > return sum; > } > > These kernels are very tight loops. In this sort of legacy codes it's not > uncommon that loops are manually tuned and unrolled with the available > registers in mind, so that all values just about fit in registers. In the > example above, we read from input streams A and B, and do 2 multiply-accumulate > operations. The loop is unrolled 2 times in this example, but 4, 8 or 16 would > more realistic, resulting in very high register pressure. > > But legacy apps written in this way are not the only culprit; we see that with > (aggressive) loop unrolling we can end up in exactly the same situation: the > loopunroller does exactly what it needs to do, but it results in very high > register pressure. > > Here's the (obvious) problem: all live values in these loops should just about > fit in registers, but suboptimal/wrong decisions are made resulting in a lot of > spills/reloads; the machine scheduler is making the life of the register allocator > very difficult. I am looking at regressions of about 30 - 40% for more > than a handful of kernels, and am thus very interested in what I could do about > this. > > The first observation is that it looks like we default to bottom-up scheduling. > Starting bottom-up, all these "sum" variables are scheduled first, and after > that the loads (this is simplifying things a bit). And thus it looks like we > create a lot of very long live-ranges, causing the problems mentioned above. > When instead we start top-down, the loads are picked up first, then the MAC, and > this repeated for all MACs. The result is a sequence LD, MAC, LD, MAC, etc., > which is let's say a sequence more in program-order, also with shorter > live-ranges. This does exactly what I want for these sort of kernels, it generates > exactly the code we want. > > While playing with this bottom-up/top-down order, it didn't take long to see > that the top-down approach is excellent for these kind of codes, but not for > some other benchmarks, and so solving this issue not just a matter of > defaulting to a new scheduling policy. > > I am thinking about prototyping an approach where we start with the bottom-up > approach (under a new misched option), but when a register-pressure/live-range > trashhold value is reached, we bail and fall back to the top-down approach. > This is my first rough idea, but I haven't looked at this problem for very > long. My first few data points is suggesting this might be benificial, but I > could be missing a lot here. And also I am sure that I am looking at an > old/classic problem here and I'm sure others have looked at this problem > before. Thus I am wondering if people have experiences/opinions on this. > > Cheers, > Sjoerd. > IMPORTANT NOTICE: The contents of this email and any attachments are confidential and may also be privileged. If you are not the intended recipient, please notify the sender immediately and do not disclose the contents to any other person, use it for any purpose, or store or copy the information in any medium. Thank you. > _______________________________________________ > LLVM Developers mailing list > [hidden email] > http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev _______________________________________________ LLVM Developers mailing list [hidden email] http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev |
In reply to this post by Sudhindra kulkarni via llvm-dev
> On Nov 6, 2018, at 6:19 AM, Sjoerd Meijer via llvm-dev <[hidden email]> wrote: > > Hello List! > > I am looking at top-down vs. bottom-up list scheduling for simple(r) in-order > cores. First, for some context, below is a fairly representative pseudo-code > example of the sort of DSP-like codes I am looking at: > > uint64_t foo(int *pA, int *pB, unsigned N, unsigned C) { > uint64_t sum = 0; > while (N-- > 0) { > A1 = *pA++; > A2 = *pA++; > B1 = *pB++; > B2 = *pB++; > ... > sum += ((uint64_t) A1 * B1) >> C; > sum += ((uint64_t) A2 * B2) >> C; > ... > } > return sum; > } Bottom-up scheduling would work for you in this case if the heuristics knew to minimize pressure at every point in the loop. Ideally, that could be detected from the shape of the DAG before list scheduling begins. The closest thing we have to doing that in the current source is the “subtree” scheduling. See `computeDFSResult`. …or maybe you find a simpler way to control the heurstics! I’m not entirely sure why the top-down heuristics are working better for you in this case. What you’re proposing is similar in spirit to LLVM's bidirectional scheduling. The problem is that a list scheduler can dig itself into a hole in either direction so bidirectional doesn’t really save you unless you throw away the previous schedule and start over. LLVM’s normal no-backtracking approach is really designed to avoid compile-time issues. -Andy _______________________________________________ LLVM Developers mailing list [hidden email] http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev |
In reply to this post by Sudhindra kulkarni via llvm-dev
Hi Matthias, > In my experience bottom-up works better when optimizing for register pressure while top-down tends to work better when optimizing for latencies/stall reduction. That's a very interesting insight, do you mind sharing the intuition behind it? Thanks, Alex From: llvm-dev [mailto:[hidden email]] On Behalf Of Matthias Braun via llvm-dev _______________________________________________ LLVM Developers mailing list [hidden email] http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev |
In the broadest generalization of the scheduling problem, if your DAG is shaped like a tree, then the number of available nodes at the beginning and end of scheduling will be affected by the order that you schedule. If you schedule the “pointy” end of the tree first (aka the root), the there aren’t many choices at first, and the leaves gradually become available at the end of scheduling. If you schedule the leaves first, then the heuristic must choose from many available nodes at once. When scheduling for resource utilization, it’s hard to correctly choose among many nodes that use the same resources. Scheduling the root first means less chance to get things wrong early on. In fact, if the DAG is strictly a tree, then bottom-up scheduling with Sethi-Ullman numbers easily minimizes registers. Of course, DAG’s aren’t actually tress, so there’s no clear right or wrong direction to schedule, but bottom-up tends to be pretty effective at minimizing registers if that’s the main constraint. When scheduling for latency, you just need the critical path from each node, which is computed before scheduling begins and doesn’t change. So having more nodes available at first doesn’t hurt. -Andy
_______________________________________________ LLVM Developers mailing list [hidden email] http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev |
Free forum by Nabble | Edit this page |