[llvm-dev] Intel AMX programming model discussion.

Luo, Yuanke via llvm-dev llvm-dev at lists.llvm.org
Fri Sep 4 06:47:10 PDT 2020


Hi Hal,
Generally, your proposal to adapt tile RA to Greedy RA looks good to me. Thank you! I plan to do some prototype for the proposal. Since there is 3 RA in LLVM infrastructure, we need 3 schemes to adapt tile RA to each existing RA. Do you like to finalize the 3 schemes first, or you would like to review the left part of the AMX programming model? We have some limitation to support dynamic shape and I'd like to hear your advice. The dynamic shape requires the ldtilecfg post-dominate the point that define shape, so we encourage user to define their shape in the entry of the function. Take below code as example. Ideally, we hope to insert ldtilecfg at line 57 to config a, b, c, but in this function the c's shape {row, col} is defined in each if/else clause. So that line 57, we don't the shape of c. Do you have any advice for such problem?
52 void kernel(int cond) {
53   _tile a = {row, 8};
54   _tile b = {8, col};
55
56   // copy shape to stack slot
57   // ldtilecfg
58   if(cond) {
59     short row = get_row();
60     short col = get_row();
61     _tile c = {row, col};
62     __tile_loadd(&a, buf, STRIDE);
63     __tile_loadd(&b, buf, STRIDE);
64     __tile_loadd(&c, buf, STRIDE);
65   } else {
66     short row = get_row();
67     short col = get_row();
68     _tile c = {row, col};
69     __tile_loadd(&a, buf2, STRIDE);
70     __tile_loadd(&b, buf2, STRIDE);
71     __tile_loadd(&c, buf2, STRIDE);
72   }
73   __tile_dpbsud(&c, a, b);
74   __tile_stored(buf, STRIDE, c);
75 }

Thanks
Yuanke
From: Hal Finkel <hfinkel at anl.gov>
Sent: Friday, September 4, 2020 5:59 PM
To: Luo, Yuanke <yuanke.luo at intel.com>; Topper, Craig <craig.topper at intel.com>; Kaylor, Andrew <andrew.kaylor at intel.com>; Philip Reames <listmail at philipreames.com>; llvm-dev at lists.llvm.org; florian_hahn at apple.com; Lu, Hongjiu <hongjiu.lu at intel.com>
Subject: Re: [llvm-dev] Intel AMX programming model discussion.



On 9/4/20 3:37 AM, Luo, Yuanke wrote:
Hi Hal,
Thank you for the ideas that help us to improve the design, and sorry for replying late. There is something I am not able to figure out and there some special trait for tile RA.



You're quite welcome.



1.       X86RegisterInfo::getRegAllocationHints can tell RA which physical register is preferred, but it can't force RA to just allocate the hinted register. If the hinted register is not meet, RA would allocate other register.



I addressed this below, but I could have been clearer. Like SystemZRegisterInfo::getRegAllocationHints does sometimes, when hinting the tile registers, the function will return true. This turns the preference into a hard constraint, and the allocator will not allocate any other register. That's my understanding from reading the code.



2.       The shape information should be attached to each virtual register and physical register which is allocated. How to store and get the shape information with limited code change on existing RA?



For each virtual register, getRegAllocationHints could just recompute the shape information. If this isn't a constant-time operation, however, you'll probably want to cache the computed shape requirements in X86MachineFunctionInfo. You can add a map from registers to shape information in that class, and accesses it from getRegAllocationHints. You can store information about the physical registers there too.

Regarding the physical registers, you can grab this information in the pre-rewrite phase. Override addPreRewrite in X86TargetMachine.cpp. You'll need a small pass that records relevant information about the assignments (which, I imagine, is the same small pass that updates the LDTILECFG instructions). For an example of such a pass, see AMDGPU/GCNNSAReassign.cpp



3.       When a tile register is spilled, the shape should also be bound the corresponding spill stack slot, so that it can be assigned the physical tile register with the same shape.



I'm not sure what you mean. If you don't want to just be conservative about the spill size allocation, you do need to know the shape in order to compute the spill-location size. I assume that you can grab that out of X86MachineFunctionInfo from storeRegToStackSlot/loadRegFromStackSlot or eliminateFrameIndex (or copyPhysReg) as needed.



4.       There is no mov/copy instruction for tile register. To copy tile register, we need to store the tile register to memory and load the data from memory to another register. So a lot of code for live interval split in Greedy RA is unnecessary for tile register allocation.



Yes, but this just means that you need to support copying through memory. Setting CopyCost = -1 in X86RegisterInfo.td might help as well.



5.       Compiler can support register spill, but spill should be avoided for performance benefit. We prefer reporting warning on register spill, so that user can realize it and adjust their code to avoid register spill.



If you want to emit a diagnostic, you may be able to do that from storeRegToStackSlot. In any case, please make use of the optimization-remark infrastructure. For an example of how to do this, see RAGreedy::reportNumberOfSplillsReloads in RegAllocGreedy.cpp.



If there is no easy way to take the advantage of current RA infrastructure, there are some pros to have a separate RA for tile register.

1.       We can limit the risk to break RA for general register on each arch. If there are some bugs on tile RA, only application that use AMX is affected.



That's true. But I also worry about that. Any time you need to write non-trivial code that will be used relatively rarely, it's likely to have bugs that take a long time to show up. If you can plug into the generic infrastructure, you benefit from the fact that it's highly-covered, often-used code. Not that you might not run into bugs, of course, especially if you're using it in a new way, but the base logic is likely to already be robust.



2.       We can customize the special trait (config, spilt, spill) of tile register in the sperate RA more freely.



True.

 -Hal



For RegAllocFast, I agree with you. Each region of register is small, and since the performance is not the first priority, we can insert multiply config for each small region.
As you recommend looking at the PBQP solver, I'll take some time to investigate it and go back to you.

Thanks
-Yuanke


From: Hal Finkel <hfinkel at anl.gov><mailto:hfinkel at anl.gov>
Sent: Monday, August 24, 2020 5:03 PM
To: Luo, Yuanke <yuanke.luo at intel.com><mailto:yuanke.luo at intel.com>; Topper, Craig <craig.topper at intel.com><mailto:craig.topper at intel.com>; Kaylor, Andrew <andrew.kaylor at intel.com><mailto:andrew.kaylor at intel.com>; Philip Reames <listmail at philipreames.com><mailto:listmail at philipreames.com>; llvm-dev at lists.llvm.org<mailto:llvm-dev at lists.llvm.org>; florian_hahn at apple.com<mailto:florian_hahn at apple.com>; Lu, Hongjiu <hongjiu.lu at intel.com><mailto:hongjiu.lu at intel.com>
Subject: Re: [llvm-dev] Intel AMX programming model discussion.


Hi, Yuanke,

Thanks for writing this up. Let me back up a bit because the scheme I proposed last week doesn't work without further modification: within a particular "configuration region" (i.e., the code in between the LDTILECFG and the TILERELEASE (or next LDTILECFG)), each tile register can only be used with one shape, and in addition, no register can have its shape changed without zeroing out all of the tile registers. Thus, just using different register classes for the different shapes, as I had suggested, isn't sufficient to model the allocation requirements. That would not prevent the same register from essentially being assigned to differently-shaped virtual registers with non-overlapping live ranges within one configuration region.

Also, as you point out, when multiple non-static tile shapes are in use, if you use one register class for each shape, you would need different register classes for these too. Luckily, I don't think that using the separate register classes actually buys us anything, so please disregard that suggestion of mine. Use only one register class.

Once the configuration regions are identified, you'll know how many tile register shapes are required. If this number is greater than eight, then you'll need to cut the region (requiring all live tiles to be spilled and restored around each re-configuration point). After that, we'll assume that we have eight or fewer distinct shapes.

Now the problem is that you need to allocate registers, satisfying all of the usual constraints (non-overlapping live ranges, etc.), but with an additional constraint: once a physical register has been used with some particular tile shape, it cannot be assigned to any other tile shape.

I think that the current infrastructure can support this as follows:

 1. Add an override X86RegisterInfo::getRegAllocationHints. Like SystemZRegisterInfo::getRegAllocationHints does sometimes, when hinting the tile registers, the function will return true (to indicate a hard constraint). As registers are assigned in RegAllocGreedy, getRegAllocationHints is called for each virtual register. For virtual tile registers, look at the passed VirtRegMap, etc. for already-assigned tile virtual registers with different shape requirements as the current virtual register (you'll need to cache the shape requirements in X86MachineFunctionInfo for this to be efficient), and return a hints list consisting of all other non-reserved tile registers.

 2. To support RegAllocFast, which doesn't use getRegAllocationHints, you would need to make the configuration regions small enough that it doesn't matter (and if you're doing this around every tile instruction, this is automatically true).

 3. To support RegAllocPBQP (which is likely a good thing to do, but probably not required), I believe you can support this by adding custom constraints to the solver (kind of like what AArch64PBQPRegAlloc.cpp does).

Once the allocation process is complete, you'll need to go back and update the LDTILECFG data to reflect the chosen shape -> register mapping.

What I don't know, however, is how well the getRegAllocationHints method will work. The benefit is that you don't need to write a custom pre-allocator allocator. On the other hand, it might visit the virtual registers to assign in a suboptimal order because it doesn't really understand the constraint being imposed (generally, we just assign larger live ranges first). On the other hand, it is a greedy algorithm and if you want something systematically closer to optimal, maybe you should be using PBQP anyway. If you do end up needing a custom allocator for these, I recommend looking at the PBQP solver (which, as I recall, is independently reusable).

Hopefully, this is more-helpful advice.

 -Hal
On 8/21/20 9:54 PM, Luo, Yuanke wrote:
It seems I make a mistake on sharing register unit. Can we share register unit for tile register that is within different tile register class (different register class has different tile shape)?  Think about two virtual tile register %2:vtile1x1 and %3:vtile1x2. First %2 is allocated to $tmm0, after that %2 is killed and %t3 is allocated to $tmm0. This is not allowed, because when $tmm0 is allocated to %2, its shape is configured to 1x1. If we reallocated $tmm0 to %3, then we need to re-config $tmm0 to 1x2 which cause $tmm0~$tmm7 be clobbered.

Yuanke

From: Luo, Yuanke
Sent: Friday, August 21, 2020 2:12 PM
To: Hal Finkel <hfinkel at anl.gov><mailto:hfinkel at anl.gov>; Topper, Craig <craig.topper at intel.com><mailto:craig.topper at intel.com>; Kaylor, Andrew <andrew.kaylor at intel.com><mailto:andrew.kaylor at intel.com>; Philip Reames <listmail at philipreames.com><mailto:listmail at philipreames.com>; llvm-dev at lists.llvm.org<mailto:llvm-dev at lists.llvm.org>; florian_hahn at apple.com<mailto:florian_hahn at apple.com>; Lu, Hongjiu <hongjiu.lu at intel.com><mailto:hongjiu.lu at intel.com>
Subject: RE: [llvm-dev] Intel AMX programming model discussion.

Hi Hal,
The proposal is attractive to me, but there is something I still can't figure out. Let's take below MIR as an example. We assume we have 256 register classes (vtile1x1, vtile1x2, ..., tile16x16).

1.       After instruction selection, the pseudo AMX instruction is generated. The name of pseudo instructions have 'P' prefix. Now all the AMX pseudo instruction take vtile as register class. Let's assume %13 is constant 3, %10 is constant 4 and %14 is variable.
  %1:vtile = PTILELOADDV %13:gr16, %10:gr16, %17:gr64, 1, %18:gr64_nosp, 0, $noreg
  %2:vtile = PTILELOADDV %10:gr16, %14:gr16, %17:gr64, 1, %18:gr64_nosp, 0, $noreg
  %3:vtile = PTILELOADDV %13:gr16, %14:gr16, %17:gr64, 1, %18:gr64_nosp, 0, $noreg
%21:vtile = PTDPBSSDV %13:gr16, %10:gr16, %14:gr16, %3:vtile(tied-def 0), %1:vtile, %2:vtile

2.       The configuration-placement pass looks at all of the AMX pseudo-instructions and identifies regions in which the pseudo-instructions use the same configuration parameters. It first replaces the register class for all tile registers whose shape is known in compile-time. Since the shape of %1 is constant, so it replaces %1:vtile with %1:vtile3x4 which change the register class and morph pseudo instruction into AMX real instruction. The shape of %2 and %3 is unknown in compile-time, so it arbitrarily picks up a tile register class which is not assigned before and assign the register class to %2 and %3. After register class allocation, the code is transformed as this. The register class for %2:vtile1x1 and %3:vtile1x2 is allocated.
   PLDTILECFG
  %1:vtile3x4  = TILELOADDV %17:gr64, 1, %18:gr64_nosp, 0, $noreg
  %2:vtile1x1 = TILELOADDV %17:gr64, 1, %18:gr64_nosp, 0, $noreg
  %3:vtile1x2 = TILELOADDV %17:gr64, 1, %18:gr64_nosp, 0, $noreg
%21:vtile1x2 = TDPBSSDV %9:vtile1x2(tied-def 0), %1:vtile3x4, %2:vtile1x1
Something I am not figured out.

1.       I not sure if we can have AMX instruction's inputs and outputs fit multiple register classes (vtile1x1, ..., vtile16x16), otherwise we need 256 pseudo instructions.

2.       Whether 256 register class is enough to be allocated. There may be more 256 unknow shape tile registers.

3.       In this pass we also find the proper pointer (common dominator) to insert ldtilecfg, but at this time the register is allocated, we don't know the shape of each physical tile register. So we just insert a pseudo tile config instruction.

3.       All tile register class share the same register unit. We do register allocation by the framework, and the code is transformed as this.
  $tmm0  = TILELOADDV %17:gr64, 1, %18:gr64_nosp, 0, $noreg
  $tmm1 = TILELOADDV %17:gr64, 1, %18:gr64_nosp, 0, $noreg
  $tmm2 = TILELOADDV %17:gr64, 1, %18:gr64_nosp, 0, $noreg
$tmm2 = TDPBSSDV $tmm2(tied-def 0), $tmm0, $tmm1

4.       Run config pass to collect the shape of each physical tile register and config them. The code can be generated as below. Here is the problem, how can we know the shape of the physical tile register?
   MOV row, col info to %stack.0 for each physical tile register   ??????
  LDTILECFG %stack.0, 1, $noreg, 0, $noreg, implicit-def $tmm0, implicit-def $tmm1, implicit-def $tmm2, implicit-def $tmm3, implicit-def $tmm4, implicit-def $tmm5, implicit-def $tmm6, implicit-def $tmm7
  $tmm0  = TILELOADDV %17:gr64, 1, %18:gr64_nosp, 0, $noreg
  $tmm1 = TILELOADDV %17:gr64, 1, %18:gr64_nosp, 0, $noreg
  $tmm2 = TILELOADDV %17:gr64, 1, %18:gr64_nosp, 0, $noreg
$tmm2 = TDPBSSDV $tmm2(tied-def 0), $tmm0, $tmm1

Thanks
Yuanke

...

--

Hal Finkel

Lead, Compiler Technology and Programming Languages

Leadership Computing Facility

Argonne National Laboratory

--

Hal Finkel

Lead, Compiler Technology and Programming Languages

Leadership Computing Facility

Argonne National Laboratory
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20200904/22b4c4b8/attachment.html>


More information about the llvm-dev mailing list