Topper, Craig via llvm-dev
2020-Aug-19 20:09 UTC
[llvm-dev] Intel AMX programming model discussion.
The width and height can be runtime values that we would just copy into 64 byte configuration block we pass to ldtilecfg. So the code doesn't need to be multiversioned. The user code would also use those values to update pointers in the loops they write using the tiles. If we can't determine that two tiles were defined with the same width and height we need to assume the shape is different and try to avoid ever giving the same tile. Hal, for your suggestion would which physical registers are in which register class be defined dynamically before register allocation? From: Hal Finkel <hfinkel at anl.gov> Sent: Wednesday, August 19, 2020 12:52 PM To: Kaylor, Andrew <andrew.kaylor at intel.com>; Luo, Yuanke <yuanke.luo at intel.com>; Philip Reames <listmail at philipreames.com>; llvm-dev at lists.llvm.org; florian_hahn at apple.com; Topper, Craig <craig.topper at intel.com>; Lu, Hongjiu <hongjiu.lu at intel.com> Subject: Re: [llvm-dev] Intel AMX programming model discussion. On 8/19/20 10:24 AM, Kaylor, Andrew wrote:> When the tile shape is unknown at compile time, how do you plan to do the register allocation of the tiles? My question is: do you do the allocation for this case in the same way as you would if you knew the size was 16x16 (i.e., conservatively assume the largest size)?I think what will happen is that the registers are allocated based on a number of runtime values that are assumed to be different from one another but less than or equal to 16. So, for example, we'll allocate registers for MxN tiles, NxM tiles and MxM tiles without knowing what M and N are. Then at runtime the values of these variables will be used to create the actual tile configuration. The instructions that need to know the shape take these runtime values as operands. So you're going to multiversion the code? In any case, my point is that you probably don't need a custom register allocator. If you just define the tile registers and make sure that the ldtilecfgs implicitly defines them all, then the regular infrastructure likely works. You'll have a bunch of register classes, but that's not necessarily a problem. I recommend trying this, and let us know what you discover, before we go down the road of a new, dedicated allocator just for these registers. -Hal There may be some artifacts coming from the front end that conservatively assume a 16x16 tile, but I think those generally go away in SROA or later specialized passes. Yuanke can confirm or correct my understanding of this. From: Hal Finkel <hfinkel at anl.gov><mailto:hfinkel at anl.gov> Sent: Wednesday, August 19, 2020 5:14 AM To: Luo, Yuanke <yuanke.luo at intel.com><mailto:yuanke.luo 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>; Topper, Craig <craig.topper at intel.com><mailto:craig.topper at intel.com>; Lu, Hongjiu <hongjiu.lu at intel.com><mailto:hongjiu.lu at intel.com> Subject: Re: [llvm-dev] Intel AMX programming model discussion. On 8/19/20 5:34 AM, Luo, Yuanke wrote: There is no problem to have 256 register classes. Just a lot of register classes to me. We don't assume the shape of each physical register be 16x16, it is defined by user. For variable shape, I mean the shape is known in runtime and in compile time the shape is unknown. Take below code as an example, the %row and %col are variable instead of constant. Compiler recognizes llvm.x86.tileloadd64 and deduce the shape of %0 is %row x %col. %0 = tail call <256 x i32> @llvm.x86.tileloadd64(i16 %row, i16 %col, i8* getelementptr inbounds ([1024 x i8], [1024 x i8]* @buf, i64 0, i64 0), i64 32) When the tile shape is unknown at compile time, how do you plan to do the register allocation of the tiles? My question is: do you do the allocation for this case in the same way as you would if you knew the size was 16x16 (i.e., conservatively assume the largest size)? Thanks again, Hal From: Hal Finkel <hfinkel at anl.gov><mailto:hfinkel at anl.gov> Sent: Wednesday, August 19, 2020 4:58 PM To: Luo, Yuanke <yuanke.luo at intel.com><mailto:yuanke.luo 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>; Topper, Craig <craig.topper at intel.com><mailto:craig.topper at intel.com>; Lu, Hongjiu <hongjiu.lu at intel.com><mailto:hongjiu.lu at intel.com> Subject: Re: [llvm-dev] Intel AMX programming model discussion. On 8/19/20 2:21 AM, Luo, Yuanke wrote: Hi Hal, There is 3 aspect to be solved. 1. The HW support max shape 16x16, so there are many register classes from 1x1 to 16x16. We need 256 register classes. 2. We want to support variable shape, so compiler don't know what register class to fit tile shape as it is only known in runtime. 3. The tile configure is to configure physical tile register, so we need to allocate register and then we know the shape of each physical tile register and configure the tile register. I think your suggestion is helpful to reduce the complexity if we only support fixed (constant) tile shape. -Yuanke Thanks, Yuanke. It's not clear to me that having 256 register classes is, in itself, a problem. Is it? What does it mean to support variable-shape tiles in this context? Do you do something other than conservatively assume that they are 16x16 for register-allocation purposes? -Hal From: Hal Finkel <hfinkel at anl.gov><mailto:hfinkel at anl.gov> Sent: Wednesday, August 19, 2020 8:20 AM To: Kaylor, Andrew <andrew.kaylor at intel.com><mailto:andrew.kaylor at intel.com>; Philip Reames <listmail at philipreames.com><mailto:listmail at philipreames.com>; Luo, Yuanke <yuanke.luo at intel.com><mailto:yuanke.luo at intel.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>; Topper, Craig <craig.topper at intel.com><mailto:craig.topper at intel.com>; Lu, Hongjiu <hongjiu.lu at intel.com><mailto:hongjiu.lu at intel.com> Subject: Re: [llvm-dev] Intel AMX programming model discussion. Hi, Andy, I don't quite understand everything that's going on here. Could we model this as: 1. Define a collection of register classes, one for 2x4 tiles, one for 4x2 tiles, etc. each populated with a set of tile registers. Registers can have aliasing relationships (instead of worrying of any kind of subregister/superregister relationships -- these won't be useful anyway). 2. Define the tile-configuration instructions so that they implicitly define all of the registers in all of the classes. Then you would still need to pre-schedule the tile operations as you've described, and collect the configuration information in order to add the ldtilecfgs, but the regular register allocator can handle the allocation itself in the usual way. What do you think? -Hal On 8/18/20 6:58 PM, Kaylor, Andrew via llvm-dev wrote: The AMX registers are complicated. The single configuration register (which is mostly used implicitly, similar to MXCSR for floating point) controls the shape of all the tile registers, and if you change the tile configuration every single tile register is cleared. In practice, if we have to change the the configuration while any of the tile registers are live, performance is going to be terrible. We need to handle this case for correctness, but users of this programming interface will need to have enough awareness of the performance issues and the hardware details to prevent this. We'll also want a diagnostic that lets the user know when this has happened. When the tile configuration is set, the shape of each tile is locked in, so the individual tile registers aren't interchangeable at that point. If a function needs 2x4 tiles, 4x2 tiles, and 4x4 tiles, the configuration needs to be set with this in mind. The shape isn't explicit in every instruction and intrinsic. It must be deduced. And again, we'll need a way to tell the user when efficient allocation can't be done. In practice, I don't expect any function to be using more than three tile shapes. The implication of all this is that I don't think the greedy register allocator is well suited to figure all of this out. We need a special pass to pre-allocate these registers. If the function is written in a way that makes good performance possible, it should be a relatively simple task to allocate everything with minimal spilling. If it isn't possible to get good performance, we don't need to do anything especially clever. We can just do something straightforward that is correct and let the user know that they aren't going to be happy with the results. -Andy From: Philip Reames <listmail at philipreames.com><mailto:listmail at philipreames.com> Sent: Friday, August 14, 2020 8:29 PM To: Luo, Yuanke <yuanke.luo at intel.com><mailto:yuanke.luo at intel.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>; Kaylor, Andrew <andrew.kaylor at intel.com><mailto:andrew.kaylor at intel.com>; Topper, Craig <craig.topper at intel.com><mailto:craig.topper at intel.com>; Lu, Hongjiu <hongjiu.lu at intel.com><mailto:hongjiu.lu at intel.com> Subject: Re: [llvm-dev] Intel AMX programming model discussion. I find your answer unconvincing. I'm not going to debate it as I don't wish to take the time to build the appropriate context, but my initial response is skepticism. Philip On 8/14/20 4:49 PM, Luo, Yuanke wrote: [Yuanke] AMX register is special. It needs to be configured before use and the config instruction is expensive. To avoid unnecessary tile configure, we collect the tile shape information as much as possible and combine them into one ldtilecfg instruction. The ldtilecfg instruction should dominate any AMX instruction that access tile register. On the other side, the ldtilecfg should post-dominated the instruction that define the tile shape. For tile register spill, it should avoid re-config due to the different tile shape, the spilled register should be reloaded to the register that share the same tile shape. Since tile register allocation is special and it may allocate general virtual register to configure tile register, we can add a sperate pass to do it before general register allocation pass. After register allocation, the tile shape information is not needed anymore, so we can transform the pseudo AMX instruction to real AMX instruction by removing the row and column operands. [Philip] This seems complicated. Reading through the documentation, there appears to be a single global tile config for all tile registers at any time. Why not simply model this tile config as a designated special register and the tile instructions as having an implicit use of this register? That would seem to ensure that the register allocator has all the constraints needed. You'd need to teach it how to spill the special registers with the appropriate instructions, but that seems a lot more straight forward? [Yuanke] In that case user need to configure the tile register by themselves. Spilling configure register is very expensive, because it clears all the tile data register to zero. In our proposal, compiler is responsible to deduce the shape for virtual of tile data register, allocate physical registers for them and then configure those physical register. We may build the dependency as you proposed and it can be used for machine IR check to ensure tile data register is configured before use. From: Philip Reames <listmail at philipreames.com><mailto:listmail at philipreames.com> Sent: Saturday, August 15, 2020 1:17 AM To: Luo, Yuanke <yuanke.luo at intel.com><mailto:yuanke.luo at intel.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>; Kaylor, Andrew <andrew.kaylor at intel.com><mailto:andrew.kaylor at intel.com>; Topper, Craig <craig.topper at intel.com><mailto:craig.topper at intel.com>; Lu, Hongjiu <hongjiu.lu at intel.com><mailto:hongjiu.lu at intel.com> Subject: Re: [llvm-dev] Intel AMX programming model discussion. On 8/14/20 6:27 AM, Luo, Yuanke via llvm-dev wrote: Hi, Intel Advanced Matrix Extensions (Intel AMX) is a new programming paradigm consisting of two components: a set of 2-dimensional registers (tiles) representing sub-arrays from a larger 2-dimensional memory image, and accelerators able to operate on tiles. Capability of Intel AMX implementation is enumerated by palettes. Two palettes are supported: palette 0 represents the initialized state and palette 1 consists of 8 tile registers of up to 1 KB size, which is controlled by a tile control register. The instruction manual is posted at https://software.intel.com/content/www/us/en/develop/download/intel-architecture-instruction-set-extensions-programming-reference.html. The AMX abi proposal is posted at https://groups.google.com/g/x86-64-abi/c/NRejFm7pwb4. This email is to discuss the programming model for AMX. Florian has introduced the matrix type and intrinsics in LLVM community. We'd like to adopt some ideas from it. Here is what we propose for the AMX programming model. 1. Data type. We'd like to have fixed vector type for AMX. Since the shape to AMX register can be configurable, the vector size is the maximum size of AMX register. That means the vector size is 1024 bytes. The C code may look like this. typedef int _tile_data __attribute__((__vector_size__(1024), __aligned__(64))); _tile_data tile; And the LLVM IR may look like this. @tile = dso_local local_unnamed_addr global <256 x i32> zeroinitializer, align 64 For llvm IR, it is nice to have a new type x86_amxtile that can be mapped to AMX registers. 2. AMX Intrinsics. The internal intrinsics are 1:1 mapped to AMX instructions. The parameter m, n, k identifies the shape of the tile. The shape can be variable, but it cannot exceed the size that AMX HW can support. Compiler can deduce shape of the tile from the AMX intrinsics. _tile_data _tile_loadd_internal(char m, short n, const void *base, int stride); _tile_data _tile_dpbssd_internal(char m, short n, short k, _tile_data dst, _tile_data src1, _tile_data src2); _tile_data _tile_dpbf16ps_internal(char m, short n, short k, _tile_data dst, _tile_data src1, _tile_data src2); void _tile_stored_internal(char m, short n, void *base, int stride, _tile_data tile); 3. User interfaces. The tile shape and tile data are combined into a struct in C language. The shape of the tile is only allowed to be initialized once. The user interface looks as this. 3 #define __DEFAULT_FN_AMX \ 4 __attribute__((__always_inline__, __nodebug__, __target__("amx-int8"))) 9 typedef struct __tile_str { 10 const char row; 11 const short col; 12 _tile_data tile; 13 }__tile; 14 15 __DEFAULT_FN_AMX 16 void __tile_loadd(__tile *dst, const void *base, long stride) { 17 dst->tile = _tile_loadd_internal(dst->row, dst->col, base, stride); 18 } 19 20 __DEFAULT_FN_AMX 21 void __tile_dpbsud(__tile *dst, __tile src1, __tile src2) { 22 dst->tile = _tile_dpbssd_internal(src1.row, src2.col, src1.col, dst->tile, src1.tile, src2.tile); 23 } 24 25 __DEFAULT_FN_AMX 26 void __tile_stored(void *base, long stride, __tile src) { 27 _tile_stored_internal(src.row, src.col, base, stride, src.tile); 28 } 4. Example code The example shows how to use the user interface in a function. 51 void api(int cond, short row, short col) { 52 __tile a = {row, col}; 53 __tile b = {row, col}; 54 __tile c = {row, col}; 55 56 if(cond) { 57 __tile_loadd(&a, buf, STRIDE); 58 __tile_loadd(&b, buf, STRIDE); 59 __tile_loadd(&c, buf, STRIDE); 60 } else { 61 __tile_loadd(&a, buf2, STRIDE); 62 __tile_loadd(&b, buf2, STRIDE); 63 __tile_loadd(&c, buf2, STRIDE); 64 } 65 __tile_dpbsud(&c, a, b); 66 __tile_stored(buf, STRIDE, c); 67 } 5. LLVM IR The LLVM intrinsics IR take the row and column information as the input parameter, so that compiler can deduce the shape of tile data. The remaining parameters are what AMX instructions require. This is the LLVM IR corresponding to the example code. 12 define dso_local void @api(i32 %cond, i16 signext %row, i16 signext %col) local_unnamed_addr #2 { 13 entry: 14 %tobool = icmp eq i32 %cond, 0 15 %sext = shl i16 %col, 8 16 %conv.i31 = ashr exact i16 %sext, 8 17 br i1 %tobool, label %if.else, label %if.then 18 19 if.then: ; preds = %entry 20 %0 = tail call <256 x i32> @llvm.x86.tileloadd64(i16 %row, i16 %conv.i31, i8* getelementptr inbounds ([1024 x i8], [1024 x i8]* @buf, i64 0, i64 0), i64 32) #3 21 %1 = tail call <256 x i32> @llvm.x86.tileloadd64(i16 %row, i16 %conv.i31, i8* getelementptr inbounds ([1024 x i8], [1024 x i8]* @buf, i64 0, i64 0), i64 32) #3 22 %2 = tail call <256 x i32> @llvm.x86.tileloadd64(i16 %row, i16 %conv.i31, i8* getelementptr inbounds ([1024 x i8], [1024 x i8]* @buf, i64 0, i64 0), i64 32) #3 23 br label %if.end 24 25 if.else: ; preds = %entry 26 %3 = tail call <256 x i32> @llvm.x86.tileloadd64(i16 %row, i16 %conv.i31, i8* getelementptr inbounds ([1024 x i8], [1024 x i8]* @buf2, i64 0, i64 0), i64 32) #3 27 %4 = tail call <256 x i32> @llvm.x86.tileloadd64(i16 %row, i16 %conv.i31, i8* getelementptr inbounds ([1024 x i8], [1024 x i8]* @buf2, i64 0, i64 0), i64 32) #3 28 %5 = tail call <256 x i32> @llvm.x86.tileloadd64(i16 %row, i16 %conv.i31, i8* getelementptr inbounds ([1024 x i8], [1024 x i8]* @buf2, i64 0, i64 0), i64 32) #3 29 br label %if.end 30 31 if.end: ; preds = %if.else, %if.then 32 %a.sroa.1186.0 = phi <256 x i32> [ %3, %if.else ], [ %0, %if.then ] 33 %b.sroa.1068.0 = phi <256 x i32> [ %4, %if.else ], [ %1, %if.then ] 34 %c.sroa.1149.0 = phi <256 x i32> [ %5, %if.else ], [ %2, %if.then ] 35 %6 = tail call <256 x i32> @llvm.x86.tdpbssd(i16 %row, i16 %conv.i31, i16 %conv.i31, <256 x i32> %c.sroa.1149.0, <256 x i32> %a.sroa.1186.0, <256 x i32> %b.sroa.1068.0) #3 36 tail call void @llvm.x86.tilestored64(i16 %row, i16 %conv.i31, i8* getelementptr inbounds ([1024 x i8], [1024 x i8]* @buf, i64 0, i64 0), i64 32, <256 x i32> %6) #3 37 ret void 38 } 6. Shape propagation When in -O0 build, some general load/store for tile vector is generated by front-end. We need to root from AMX intrinsics to propagate the shape information to the virtual tile register. If the an AMX intrinsic use the result of load instruction, the shape is propagated to the load and the load is transformed to tile load intrinsic. If the store instruction uses any result of AMX intrinsic, the shape is propagated to store instruction and the store is transformed to tile store intrinsic 7. Machine IR Since the AMX intrinsics take the row and column as the input parameters, we can create a pseudo instruction corresponding to it. The AMX intrinsics are lowered to the pseudo AMX instruction which has extra row and column operands corresponding to AMX intrinsic. The real AMX instructions don't need the row and column operands. The row and column information should be configured by ldtilecfg before executing any AMX instruction. 8. Register allocation AMX register is special. It needs to be configured before use and the config instruction is expensive. To avoid unnecessary tile configure, we collect the tile shape information as much as possible and combine them into one ldtilecfg instruction. The ldtilecfg instruction should dominate any AMX instruction that access tile register. On the other side, the ldtilecfg should post-dominated the instruction that define the tile shape. For tile register spill, it should avoid re-config due to the different tile shape, the spilled register should be reloaded to the register that share the same tile shape. Since tile register allocation is special and it may allocate general virtual register to configure tile register, we can add a sperate pass to do it before general register allocation pass. After register allocation, the tile shape information is not needed anymore, so we can transform the pseudo AMX instruction to real AMX instruction by removing the row and column operands. This seems complicated. Reading through the documentation, there appears to be a single global tile config for all tile registers at any time. Why not simply model this tile config as a designated special register and the tile instructions as having an implicit use of this register? That would seem to ensure that the register allocator has all the constraints needed. You'd need to teach it how to spill the special registers with the appropriate instructions, but that seems a lot more straight forward? 9. Use recommendation Due to the shape configure issue, we recommend user to define the tile shape at the entry of the function entry and inline function as much as possible. The AMX instructions focus on computation instead of storage, so global variable for tile data is not recommended. Thanks Yuanke _______________________________________________ LLVM Developers mailing list llvm-dev at lists.llvm.org<mailto:llvm-dev at lists.llvm.org> https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev _______________________________________________ LLVM Developers mailing list llvm-dev at lists.llvm.org<mailto:llvm-dev at lists.llvm.org> https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev -- 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 -- 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/20200819/cafea80f/attachment-0001.html>
Hal Finkel via llvm-dev
2020-Aug-20 19:35 UTC
[llvm-dev] Intel AMX programming model discussion.
On 8/19/20 3:09 PM, Topper, Craig wrote:> > The width and height can be runtime values that we would just copy > into 64 byte configuration block we pass to ldtilecfg. So the code > doesn’t need to be multiversioned. The user code would also use those > values to update pointers in the loops they write using the tiles. If > we can’t determine that two tiles were defined with the same width and > height we need to assume the shape is different and try to avoid ever > giving the same tile. > > Hal, for your suggestion would which physical registers are in which > register class be defined dynamically before register allocation? >Here's my thought: First, you have a set of intrinsics that take tile values along with tile configuration parameters (which, presently, seem just to be the sizes). These get lowered into pseudo-instructions that do the same. Thus, you have some register class that represents these arbitrarily-sized tile registers that you'll assign to these pseudo-instruction operands (i.e., they take virtual tile registers right after instruction selection). You might use the 16x16 tile register class for this purpose, but it shouldn't really matter. Second, you run this configuration-placement pass. This pass looks at all of the AMX pseudo-instructions and identifies regions in which the pseudo-instructions use the same configuration parameters (i.e., the same SSA values and/or constants). This pass might reorder the pseudo-instructions when legal in order to form larger regions. Then it places the ldtilecfg at the start of each region (in some common dominating position). ldtilecfg implicitly defines all of the tile registers in every concrete class of tile registers (all 256 of them, or whatever). The pseudo-instructions are replaced by real MI instructions taking a tile register class appropriate for the configuration (which will default to the 16x16 class for cases where the configuration is not a compile-time-known constant). When the configuration is a known constant, the instructions take operands with a register class appropriate for that configuration (e.g., 1x1, 4x4). Third, the rest of the framework runs as usual. Tile registers from the appropriate class are allocated by the register allocator. No live range of any virtual tile register can pass through the ldtilecfg (because it defines them all), but that's okay, none of live ranges will by construction (the configuration-placement pass ensures this). Thanks again, Hal> *From:* Hal Finkel <hfinkel at anl.gov> > *Sent:* Wednesday, August 19, 2020 12:52 PM > *To:* Kaylor, Andrew <andrew.kaylor at intel.com>; Luo, Yuanke > <yuanke.luo at intel.com>; Philip Reames <listmail at philipreames.com>; > llvm-dev at lists.llvm.org; florian_hahn at apple.com; Topper, Craig > <craig.topper at intel.com>; Lu, Hongjiu <hongjiu.lu at intel.com> > *Subject:* Re: [llvm-dev] Intel AMX programming model discussion. > > On 8/19/20 10:24 AM, Kaylor, Andrew wrote: > > > When the tile shape is unknown at compile time, how do you plan > to do the register allocation of the tiles? My question is: do you > do the allocation for this case in the same way as you would if > you knew the size was 16x16 (i.e., conservatively assume the > largest size)? > > I think what will happen is that the registers are allocated based > on a number of runtime values that are assumed to be different > from one another but less than or equal to 16. So, for example, > we’ll allocate registers for MxN tiles, NxM tiles and MxM tiles > without knowing what M and N are. Then at runtime the values of > these variables will be used to create the actual tile > configuration. The instructions that need to know the shape take > these runtime values as operands. > > So you're going to multiversion the code? > > In any case, my point is that you probably don't need a custom > register allocator. If you just define the tile registers and make > sure that the ldtilecfgs implicitly defines them all, then the regular > infrastructure likely works. You'll have a bunch of register classes, > but that's not necessarily a problem. I recommend trying this, and let > us know what you discover, before we go down the road of a new, > dedicated allocator just for these registers. > > -Hal > > There may be some artifacts coming from the front end that > conservatively assume a 16x16 tile, but I think those generally go > away in SROA or later specialized passes. Yuanke can confirm or > correct my understanding of this. > > *From:* Hal Finkel <hfinkel at anl.gov> <mailto:hfinkel at anl.gov> > *Sent:* Wednesday, August 19, 2020 5:14 AM > *To:* Luo, Yuanke <yuanke.luo at intel.com> > <mailto:yuanke.luo 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>; Topper, Craig > <craig.topper at intel.com> <mailto:craig.topper at intel.com>; Lu, > Hongjiu <hongjiu.lu at intel.com> <mailto:hongjiu.lu at intel.com> > *Subject:* Re: [llvm-dev] Intel AMX programming model discussion. > > On 8/19/20 5:34 AM, Luo, Yuanke wrote: > > There is no problem to have 256 register classes. Just a lot > of register classes to me. > > We don’t assume the shape of each physical register be 16x16, > it is defined by user. For variable shape, I mean the shape is > known in runtime and in compile time the shape is unknown. > Take below code as an example, the %row and %col are variable > instead of constant. Compiler recognizes llvm.x86.tileloadd64 > and deduce the shape of %0 is %row x %col. > > %0 = tail call <256 x i32> @llvm.x86.tileloadd64(i16 %row, i16 > %col, i8* getelementptr inbounds ([1024 x i8], [1024 x i8]* > @buf, i64 0, i64 0), i64 32) > > When the tile shape is unknown at compile time, how do you plan to > do the register allocation of the tiles? My question is: do you do > the allocation for this case in the same way as you would if you > knew the size was 16x16 (i.e., conservatively assume the largest > size)? > > Thanks again, > > Hal > > *From:* Hal Finkel <hfinkel at anl.gov> <mailto:hfinkel at anl.gov> > *Sent:* Wednesday, August 19, 2020 4:58 PM > *To:* Luo, Yuanke <yuanke.luo at intel.com> > <mailto:yuanke.luo 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>; Topper, Craig > <craig.topper at intel.com> <mailto:craig.topper at intel.com>; Lu, > Hongjiu <hongjiu.lu at intel.com> <mailto:hongjiu.lu at intel.com> > *Subject:* Re: [llvm-dev] Intel AMX programming model discussion. > > On 8/19/20 2:21 AM, Luo, Yuanke wrote: > > Hi Hal, > > There is 3 aspect to be solved. > > 1.The HW support max shape 16x16, so there are many > register classes from 1x1 to 16x16. We need 256 register > classes. > > 2.We want to support variable shape, so compiler don’t > know what register class to fit tile shape as it is only > known in runtime. > > 3.The tile configure is to configure physical tile > register, so we need to allocate register and then we know > the shape of each physical tile register and configure the > tile register. > > I think your suggestion is helpful to reduce the > complexity if we only support fixed (constant) tile shape. > > -Yuanke > > Thanks, Yuanke. > > It's not clear to me that having 256 register classes is, in > itself, a problem. Is it? > > What does it mean to support variable-shape tiles in this > context? Do you do something other than conservatively assume > that they are 16x16 for register-allocation purposes? > > -Hal > > *From:* Hal Finkel <hfinkel at anl.gov> <mailto:hfinkel at anl.gov> > *Sent:* Wednesday, August 19, 2020 8:20 AM > *To:* Kaylor, Andrew <andrew.kaylor at intel.com> > <mailto:andrew.kaylor at intel.com>; Philip Reames > <listmail at philipreames.com> > <mailto:listmail at philipreames.com>; Luo, Yuanke > <yuanke.luo at intel.com> <mailto:yuanke.luo at intel.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>; > Topper, Craig <craig.topper at intel.com> > <mailto:craig.topper at intel.com>; Lu, Hongjiu > <hongjiu.lu at intel.com> <mailto:hongjiu.lu at intel.com> > *Subject:* Re: [llvm-dev] Intel AMX programming model > discussion. > > Hi, Andy, > > I don't quite understand everything that's going on here. > Could we model this as: > > 1. Define a collection of register classes, one for 2x4 > tiles, one for 4x2 tiles, etc. each populated with a set > of tile registers. Registers can have aliasing > relationships (instead of worrying of any kind of > subregister/superregister relationships -- these won't be > useful anyway). > > 2. Define the tile-configuration instructions so that > they implicitly define all of the registers in all of the > classes. > > Then you would still need to pre-schedule the tile > operations as you've described, and collect the > configuration information in order to add the ldtilecfgs, > but the regular register allocator can handle the > allocation itself in the usual way. What do you think? > > -Hal > > On 8/18/20 6:58 PM, Kaylor, Andrew via llvm-dev wrote: > > The AMX registers are complicated. The single > configuration register (which is mostly used > implicitly, similar to MXCSR for floating point) > controls the shape of all the tile registers, and if > you change the tile configuration every single tile > register is cleared. In practice, if we have to change > the the configuration while any of the tile registers > are live, performance is going to be terrible. We need > to handle this case for correctness, but users of this > programming interface will need to have enough > awareness of the performance issues and the hardware > details to prevent this. We’ll also want a diagnostic > that lets the user know when this has happened. > > When the tile configuration is set, the shape of each > tile is locked in, so the individual tile registers > aren’t interchangeable at that point. If a function > needs 2x4 tiles, 4x2 tiles, and 4x4 tiles, the > configuration needs to be set with this in mind. The > shape isn’t explicit in every instruction and > intrinsic. It must be deduced. And again, we’ll need a > way to tell the user when efficient allocation can’t > be done. In practice, I don’t expect any function to > be using more than three tile shapes. > > The implication of all this is that I don’t think the > greedy register allocator is well suited to figure all > of this out. We need a special pass to pre-allocate > these registers. If the function is written in a way > that makes good performance possible, it should be a > relatively simple task to allocate everything with > minimal spilling. If it isn’t possible to get good > performance, we don’t need to do anything especially > clever. We can just do something straightforward that > is correct and let the user know that they aren’t > going to be happy with the results. > > -Andy > > *From:* Philip Reames <listmail at philipreames.com> > <mailto:listmail at philipreames.com> > *Sent:* Friday, August 14, 2020 8:29 PM > *To:* Luo, Yuanke <yuanke.luo at intel.com> > <mailto:yuanke.luo at intel.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>; Kaylor, Andrew > <andrew.kaylor at intel.com> > <mailto:andrew.kaylor at intel.com>; Topper, Craig > <craig.topper at intel.com> > <mailto:craig.topper at intel.com>; Lu, Hongjiu > <hongjiu.lu at intel.com> <mailto:hongjiu.lu at intel.com> > *Subject:* Re: [llvm-dev] Intel AMX programming model > discussion. > > I find your answer unconvincing. I'm not going to > debate it as I don't wish to take the time to build > the appropriate context, but my initial response is > skepticism. > > Philip > > On 8/14/20 4:49 PM, Luo, Yuanke wrote: > > [Yuanke] AMX register is special. It needs to be > configured before use and the config instruction > is expensive. To avoid unnecessary tile configure, > we collect the tile shape information as much as > possible and combine them into one ldtilecfg > instruction. The ldtilecfg instruction should > dominate any AMX instruction that access tile > register. On the other side, the ldtilecfg should > post-dominated the instruction that define the > tile shape. For tile register spill, it should > avoid re-config due to the different tile shape, > the spilled register should be reloaded to the > register that share the same tile shape. Since > tile register allocation is special and it may > allocate general virtual register to configure > tile register, we can add a sperate pass to do it > before general register allocation pass. After > register allocation, the tile shape information is > not needed anymore, so we can transform the pseudo > AMX instruction to real AMX instruction by > removing the row and column operands. > > [Philip] > > This seems complicated. > > Reading through the documentation, there appears > to be a single global tile config for all tile > registers at any time. > > Why not simply model this tile config as a > designated special register and the tile > instructions as having an implicit use of this > register? That would seem to ensure that the > register allocator has all the constraints needed. > You'd need to teach it how to spill the special > registers with the appropriate instructions, but > that seems a lot more straight forward? > > [Yuanke] In that case user need to configure the > tile register by themselves. Spilling configure > register is very expensive, because it clears all > the tile data register to zero. In our proposal, > compiler is responsible to deduce the shape for > virtual of tile data register, allocate physical > registers for them and then configure those > physical register. We may build the dependency as > you proposed and it can be used for machine IR > check to ensure tile data register is configured > before use. > > *From:* Philip Reames <listmail at philipreames.com> > <mailto:listmail at philipreames.com> > *Sent:* Saturday, August 15, 2020 1:17 AM > *To:* Luo, Yuanke <yuanke.luo at intel.com> > <mailto:yuanke.luo at intel.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>; Kaylor, Andrew > <andrew.kaylor at intel.com> > <mailto:andrew.kaylor at intel.com>; Topper, Craig > <craig.topper at intel.com> > <mailto:craig.topper at intel.com>; Lu, Hongjiu > <hongjiu.lu at intel.com> <mailto:hongjiu.lu at intel.com> > *Subject:* Re: [llvm-dev] Intel AMX programming > model discussion. > > On 8/14/20 6:27 AM, Luo, Yuanke via llvm-dev wrote: > > Hi, > > Intel Advanced Matrix Extensions (Intel AMX) > is a new programming paradigm consisting of > two components: a set of 2-dimensional > registers (tiles) representing sub-arrays from > a larger 2-dimensional memory image, and > accelerators able to operate on tiles. > Capability of Intel AMX implementation is > enumerated by palettes. Two palettes are > supported: palette 0 represents the > initialized state and palette 1 consists of 8 > tile registers of up to 1 KB size, which is > controlled by a tile control register. > > The instruction manual is posted at > https://software.intel.com/content/www/us/en/develop/download/intel-architecture-instruction-set-extensions-programming-reference.html > <https://software.intel.com/content/www/us/en/develop/download/intel-architecture-instruction-set-extensions-programming-reference.html>. > > The AMX abi proposal is posted at > https://groups.google.com/g/x86-64-abi/c/NRejFm7pwb4 > <https://groups.google.com/g/x86-64-abi/c/NRejFm7pwb4>. > > This email is to discuss the programming model > for AMX. Florian has introduced the matrix > type and intrinsics in LLVM community. We’d > like to adopt some ideas from it. > > Here is what we propose for the AMX > programming model. > > 1. Data type. > > We’d like to have fixed vector type for AMX. > Since the shape to AMX register can be > configurable, the vector size is the maximum > size of AMX register. That means the vector > size is 1024 bytes. > > The C code may look like this. > > typedef int _tile_data > __attribute__((__vector_size__(1024), > __aligned__(64))); > > _tile_data tile; > > And the LLVM IR may look like this. > > @tile = dso_local local_unnamed_addr global > <256 x i32> zeroinitializer, align 64 > > For llvm IR, it is nice to have a new type > x86_amxtile that can be mapped to AMX registers. > > 2.AMX Intrinsics. > > The internal intrinsics are 1:1 mapped to AMX > instructions. The parameter m, n, k identifies > the shape of the tile. The shape can be > variable, but it cannot exceed the size that > AMX HW can support. Compiler can deduce shape > of the tile from the AMX intrinsics. > > _tile_data _tile_loadd_internal(char m, short > n, const void *base, int stride); > > _tile_data _tile_dpbssd_internal(char m, short > n, short k, _tile_data dst, _tile_data src1, > _tile_data src2); > > _tile_data _tile_dpbf16ps_internal(char m, > short n, short k, _tile_data dst, _tile_data > src1, _tile_data src2); > > void _tile_stored_internal(char m, short n, > void *base, int stride, _tile_data tile); > > 3.User interfaces. > > The tile shape and tile data are combined into > a struct in C language. The shape of the tile > is only allowed to be initialized once. The > user interface looks as this. > > 3 #define __DEFAULT_FN_AMX \ > > 4 __attribute__((__always_inline__, > __nodebug__, __target__("amx-int8"))) > > 9 typedef struct __tile_str { > > 10 const char row; > > 11 const short col; > > 12 _tile_data tile; > > 13 }__tile; > > 14 > > 15 __DEFAULT_FN_AMX > > 16 void __tile_loadd(__tile *dst, const void > *base, long stride) { > > 17 dst->tile > _tile_loadd_internal(dst->row, dst->col, base, > stride); > > 18 } > > 19 > > 20 __DEFAULT_FN_AMX > > 21 void __tile_dpbsud(__tile *dst, __tile > src1, __tile src2) { > > 22 dst->tile > _tile_dpbssd_internal(src1.row, src2.col, > src1.col, dst->tile, src1.tile, src2.tile); > > 23 } > > 24 > > 25 __DEFAULT_FN_AMX > > 26 void __tile_stored(void *base, long stride, > __tile src) { > > 27 _tile_stored_internal(src.row, src.col, > base, stride, src.tile); > > 28 } > > 4.Example code > > The example shows how to use the user > interface in a function. > > 51 void api(int cond, short row, short col) { > > 52 __tile a = {row, col}; > > 53 __tile b = {row, col}; > > 54 __tile c = {row, col}; > > 55 > > 56 if(cond) { > > 57 __tile_loadd(&a, buf, STRIDE); > > 58 __tile_loadd(&b, buf, STRIDE); > > 59 __tile_loadd(&c, buf, STRIDE); > > 60 } else { > > 61 __tile_loadd(&a, buf2, STRIDE); > > 62 __tile_loadd(&b, buf2, STRIDE); > > 63 __tile_loadd(&c, buf2, STRIDE); > > 64 } > > 65 __tile_dpbsud(&c, a, b); > > 66 __tile_stored(buf, STRIDE, c); > > 67 } > > 5.LLVM IR > > The LLVM intrinsics IR take the row and column > information as the input parameter, so that > compiler can deduce the shape of tile data. > The remaining parameters are what AMX > instructions require. This is the LLVM IR > corresponding to the example code. > > 12 define dso_local void @api(i32 %cond, i16 > signext %row, i16 signext %col) > local_unnamed_addr #2 { > > 13 entry: > > 14 %tobool = icmp eq i32 %cond, 0 > > 15 %sext = shl i16 %col, 8 > > 16 %conv.i31 = ashr exact i16 %sext, 8 > > 17 br i1 %tobool, label %if.else, label %if.then > > 18 > > 19 if.then: ; preds = %entry > > 20 %0 = tail call <256 x i32> > @llvm.x86.tileloadd64(i16 %row, i16 %conv.i31, > i8* getelementptr inbounds ([1024 x i8], [1024 > x i8]* @buf, i64 0, i64 0), i64 32) #3 > > 21 %1 = tail call <256 x i32> > @llvm.x86.tileloadd64(i16 %row, i16 %conv.i31, > i8* getelementptr inbounds ([1024 x i8], [1024 > x i8]* @buf, i64 0, i64 0), i64 32) #3 > > 22 %2 = tail call <256 x i32> > @llvm.x86.tileloadd64(i16 %row, i16 %conv.i31, > i8* getelementptr inbounds ([1024 x i8], [1024 > x i8]* @buf, i64 0, i64 0), i64 32) #3 > > 23 br label %if.end > > 24 > > 25 if.else: ; preds = %entry > > 26 %3 = tail call <256 x i32> > @llvm.x86.tileloadd64(i16 %row, i16 %conv.i31, > i8* getelementptr inbounds ([1024 x i8], [1024 > x i8]* @buf2, i64 0, i64 0), i64 32) #3 > > 27 %4 = tail call <256 x i32> > @llvm.x86.tileloadd64(i16 %row, i16 %conv.i31, > i8* getelementptr inbounds ([1024 x i8], [1024 > x i8]* @buf2, i64 0, i64 0), i64 32) #3 > > 28 %5 = tail call <256 x i32> > @llvm.x86.tileloadd64(i16 %row, i16 %conv.i31, > i8* getelementptr inbounds ([1024 x i8], [1024 > x i8]* @buf2, i64 0, i64 0), i64 32) #3 > > 29 br label %if.end > > 30 > > 31 if.end: ; preds = %if.else, %if.then > > 32 %a.sroa.1186.0 = phi <256 x i32> [ %3, > %if.else ], [ %0, %if.then ] > > 33 %b.sroa.1068.0 = phi <256 x i32> [ %4, > %if.else ], [ %1, %if.then ] > > 34 %c.sroa.1149.0 = phi <256 x i32> [ %5, > %if.else ], [ %2, %if.then ] > > 35 %6 = tail call <256 x i32> > @llvm.x86.tdpbssd(i16 %row, i16 %conv.i31, i16 > %conv.i31, <256 x i32> %c.sroa.1149.0, <256 x > i32> %a.sroa.1186.0, <256 x i32> > %b.sroa.1068.0) #3 > > 36 tail call void @llvm.x86.tilestored64(i16 > %row, i16 %conv.i31, i8* getelementptr > inbounds ([1024 x i8], [1024 x i8]* @buf, i64 > 0, i64 0), i64 32, <256 x i32> %6) #3 > > 37 ret void > > 38 } > > 6.Shape propagation > > When in -O0 build, some general load/store for > tile vector is generated by front-end. We need > to root from AMX intrinsics to propagate the > shape information to the virtual tile > register. If the an AMX intrinsic use the > result of load instruction, the shape is > propagated to the load and the load is > transformed to tile load intrinsic. If the > store instruction uses any result of AMX > intrinsic, the shape is propagated to store > instruction and the store is transformed to > tile store intrinsic > > 7.Machine IR > > Since the AMX intrinsics take the row and > column as the input parameters, we can create > a pseudo instruction corresponding to it. The > AMX intrinsics are lowered to the pseudo AMX > instruction which has extra row and column > operands corresponding to AMX intrinsic. The > real AMX instructions don’t need the row and > column operands. The row and column > information should be configured by ldtilecfg > before executing any AMX instruction. > > 8.Register allocation > > AMX register is special. It needs to be > configured before use and the config > instruction is expensive. To avoid unnecessary > tile configure, we collect the tile shape > information as much as possible and combine > them into one ldtilecfg instruction. The > ldtilecfg instruction should dominate any AMX > instruction that access tile register. On the > other side, the ldtilecfg should > post-dominated the instruction that define the > tile shape. For tile register spill, it should > avoid re-config due to the different tile > shape, the spilled register should be reloaded > to the register that share the same tile > shape. Since tile register allocation is > special and it may allocate general virtual > register to configure tile register, we can > add a sperate pass to do it before general > register allocation pass. After register > allocation, the tile shape information is not > needed anymore, so we can transform the pseudo > AMX instruction to real AMX instruction by > removing the row and column operands. > > This seems complicated. > > Reading through the documentation, there appears > to be a single global tile config for all tile > registers at any time. > > Why not simply model this tile config as a > designated special register and the tile > instructions as having an implicit use of this > register? That would seem to ensure that the > register allocator has all the constraints needed. > You'd need to teach it how to spill the special > registers with the appropriate instructions, but > that seems a lot more straight forward? > > 9.Use recommendation > > Due to the shape configure issue, we recommend > user to define the tile shape at the entry of > the function entry and inline function as much > as possible. The AMX instructions focus on > computation instead of storage, so global > variable for tile data is not recommended. > > Thanks > > Yuanke > > > > > > > > > _______________________________________________ > > LLVM Developers mailing list > > llvm-dev at lists.llvm.org <mailto:llvm-dev at lists.llvm.org> > > https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev <https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev> > > > > > > > _______________________________________________ > > LLVM Developers mailing list > > llvm-dev at lists.llvm.org <mailto:llvm-dev at lists.llvm.org> > > https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev <https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev> > > -- > > 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 > > -- > > 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-- 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/20200820/10098fc5/attachment-0001.html>
Topper, Craig via llvm-dev
2020-Aug-20 19:47 UTC
[llvm-dev] Intel AMX programming model discussion.
I think I'm still missing something here. The configuration is per tile. The multiply instructions take a MxK tile and multiply it by a KxN tile and accumulate into an MxN tile. So the configuration needs to know how many of each size of tile it needs to avoid a spill. Wouldn't the register allocator then need to know which physical tiles have been configured to which sizes so that it only chooses those tiles for an operand that needs that size? ~Craig From: Hal Finkel <hfinkel at anl.gov> Sent: Thursday, August 20, 2020 12:35 PM To: Topper, Craig <craig.topper at intel.com>; Kaylor, Andrew <andrew.kaylor at intel.com>; Luo, Yuanke <yuanke.luo 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 8/19/20 3:09 PM, Topper, Craig wrote: The width and height can be runtime values that we would just copy into 64 byte configuration block we pass to ldtilecfg. So the code doesn't need to be multiversioned. The user code would also use those values to update pointers in the loops they write using the tiles. If we can't determine that two tiles were defined with the same width and height we need to assume the shape is different and try to avoid ever giving the same tile. Hal, for your suggestion would which physical registers are in which register class be defined dynamically before register allocation? Here's my thought: First, you have a set of intrinsics that take tile values along with tile configuration parameters (which, presently, seem just to be the sizes). These get lowered into pseudo-instructions that do the same. Thus, you have some register class that represents these arbitrarily-sized tile registers that you'll assign to these pseudo-instruction operands (i.e., they take virtual tile registers right after instruction selection). You might use the 16x16 tile register class for this purpose, but it shouldn't really matter. Second, you run this configuration-placement pass. This pass looks at all of the AMX pseudo-instructions and identifies regions in which the pseudo-instructions use the same configuration parameters (i.e., the same SSA values and/or constants). This pass might reorder the pseudo-instructions when legal in order to form larger regions. Then it places the ldtilecfg at the start of each region (in some common dominating position). ldtilecfg implicitly defines all of the tile registers in every concrete class of tile registers (all 256 of them, or whatever). The pseudo-instructions are replaced by real MI instructions taking a tile register class appropriate for the configuration (which will default to the 16x16 class for cases where the configuration is not a compile-time-known constant). When the configuration is a known constant, the instructions take operands with a register class appropriate for that configuration (e.g., 1x1, 4x4). Third, the rest of the framework runs as usual. Tile registers from the appropriate class are allocated by the register allocator. No live range of any virtual tile register can pass through the ldtilecfg (because it defines them all), but that's okay, none of live ranges will by construction (the configuration-placement pass ensures this). Thanks again, Hal From: Hal Finkel <hfinkel at anl.gov><mailto:hfinkel at anl.gov> Sent: Wednesday, August 19, 2020 12:52 PM To: Kaylor, Andrew <andrew.kaylor at intel.com><mailto:andrew.kaylor at intel.com>; Luo, Yuanke <yuanke.luo at intel.com><mailto:yuanke.luo 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>; Topper, Craig <craig.topper at intel.com><mailto:craig.topper at intel.com>; Lu, Hongjiu <hongjiu.lu at intel.com><mailto:hongjiu.lu at intel.com> Subject: Re: [llvm-dev] Intel AMX programming model discussion. On 8/19/20 10:24 AM, Kaylor, Andrew wrote:> When the tile shape is unknown at compile time, how do you plan to do the register allocation of the tiles? My question is: do you do the allocation for this case in the same way as you would if you knew the size was 16x16 (i.e., conservatively assume the largest size)?I think what will happen is that the registers are allocated based on a number of runtime values that are assumed to be different from one another but less than or equal to 16. So, for example, we'll allocate registers for MxN tiles, NxM tiles and MxM tiles without knowing what M and N are. Then at runtime the values of these variables will be used to create the actual tile configuration. The instructions that need to know the shape take these runtime values as operands. So you're going to multiversion the code? In any case, my point is that you probably don't need a custom register allocator. If you just define the tile registers and make sure that the ldtilecfgs implicitly defines them all, then the regular infrastructure likely works. You'll have a bunch of register classes, but that's not necessarily a problem. I recommend trying this, and let us know what you discover, before we go down the road of a new, dedicated allocator just for these registers. -Hal There may be some artifacts coming from the front end that conservatively assume a 16x16 tile, but I think those generally go away in SROA or later specialized passes. Yuanke can confirm or correct my understanding of this. From: Hal Finkel <hfinkel at anl.gov><mailto:hfinkel at anl.gov> Sent: Wednesday, August 19, 2020 5:14 AM To: Luo, Yuanke <yuanke.luo at intel.com><mailto:yuanke.luo 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>; Topper, Craig <craig.topper at intel.com><mailto:craig.topper at intel.com>; Lu, Hongjiu <hongjiu.lu at intel.com><mailto:hongjiu.lu at intel.com> Subject: Re: [llvm-dev] Intel AMX programming model discussion. On 8/19/20 5:34 AM, Luo, Yuanke wrote: There is no problem to have 256 register classes. Just a lot of register classes to me. We don't assume the shape of each physical register be 16x16, it is defined by user. For variable shape, I mean the shape is known in runtime and in compile time the shape is unknown. Take below code as an example, the %row and %col are variable instead of constant. Compiler recognizes llvm.x86.tileloadd64 and deduce the shape of %0 is %row x %col. %0 = tail call <256 x i32> @llvm.x86.tileloadd64(i16 %row, i16 %col, i8* getelementptr inbounds ([1024 x i8], [1024 x i8]* @buf, i64 0, i64 0), i64 32) When the tile shape is unknown at compile time, how do you plan to do the register allocation of the tiles? My question is: do you do the allocation for this case in the same way as you would if you knew the size was 16x16 (i.e., conservatively assume the largest size)? Thanks again, Hal From: Hal Finkel <hfinkel at anl.gov><mailto:hfinkel at anl.gov> Sent: Wednesday, August 19, 2020 4:58 PM To: Luo, Yuanke <yuanke.luo at intel.com><mailto:yuanke.luo 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>; Topper, Craig <craig.topper at intel.com><mailto:craig.topper at intel.com>; Lu, Hongjiu <hongjiu.lu at intel.com><mailto:hongjiu.lu at intel.com> Subject: Re: [llvm-dev] Intel AMX programming model discussion. On 8/19/20 2:21 AM, Luo, Yuanke wrote: Hi Hal, There is 3 aspect to be solved. 1. The HW support max shape 16x16, so there are many register classes from 1x1 to 16x16. We need 256 register classes. 2. We want to support variable shape, so compiler don't know what register class to fit tile shape as it is only known in runtime. 3. The tile configure is to configure physical tile register, so we need to allocate register and then we know the shape of each physical tile register and configure the tile register. I think your suggestion is helpful to reduce the complexity if we only support fixed (constant) tile shape. -Yuanke Thanks, Yuanke. It's not clear to me that having 256 register classes is, in itself, a problem. Is it? What does it mean to support variable-shape tiles in this context? Do you do something other than conservatively assume that they are 16x16 for register-allocation purposes? -Hal From: Hal Finkel <hfinkel at anl.gov><mailto:hfinkel at anl.gov> Sent: Wednesday, August 19, 2020 8:20 AM To: Kaylor, Andrew <andrew.kaylor at intel.com><mailto:andrew.kaylor at intel.com>; Philip Reames <listmail at philipreames.com><mailto:listmail at philipreames.com>; Luo, Yuanke <yuanke.luo at intel.com><mailto:yuanke.luo at intel.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>; Topper, Craig <craig.topper at intel.com><mailto:craig.topper at intel.com>; Lu, Hongjiu <hongjiu.lu at intel.com><mailto:hongjiu.lu at intel.com> Subject: Re: [llvm-dev] Intel AMX programming model discussion. Hi, Andy, I don't quite understand everything that's going on here. Could we model this as: 1. Define a collection of register classes, one for 2x4 tiles, one for 4x2 tiles, etc. each populated with a set of tile registers. Registers can have aliasing relationships (instead of worrying of any kind of subregister/superregister relationships -- these won't be useful anyway). 2. Define the tile-configuration instructions so that they implicitly define all of the registers in all of the classes. Then you would still need to pre-schedule the tile operations as you've described, and collect the configuration information in order to add the ldtilecfgs, but the regular register allocator can handle the allocation itself in the usual way. What do you think? -Hal On 8/18/20 6:58 PM, Kaylor, Andrew via llvm-dev wrote: The AMX registers are complicated. The single configuration register (which is mostly used implicitly, similar to MXCSR for floating point) controls the shape of all the tile registers, and if you change the tile configuration every single tile register is cleared. In practice, if we have to change the the configuration while any of the tile registers are live, performance is going to be terrible. We need to handle this case for correctness, but users of this programming interface will need to have enough awareness of the performance issues and the hardware details to prevent this. We'll also want a diagnostic that lets the user know when this has happened. When the tile configuration is set, the shape of each tile is locked in, so the individual tile registers aren't interchangeable at that point. If a function needs 2x4 tiles, 4x2 tiles, and 4x4 tiles, the configuration needs to be set with this in mind. The shape isn't explicit in every instruction and intrinsic. It must be deduced. And again, we'll need a way to tell the user when efficient allocation can't be done. In practice, I don't expect any function to be using more than three tile shapes. The implication of all this is that I don't think the greedy register allocator is well suited to figure all of this out. We need a special pass to pre-allocate these registers. If the function is written in a way that makes good performance possible, it should be a relatively simple task to allocate everything with minimal spilling. If it isn't possible to get good performance, we don't need to do anything especially clever. We can just do something straightforward that is correct and let the user know that they aren't going to be happy with the results. -Andy From: Philip Reames <listmail at philipreames.com><mailto:listmail at philipreames.com> Sent: Friday, August 14, 2020 8:29 PM To: Luo, Yuanke <yuanke.luo at intel.com><mailto:yuanke.luo at intel.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>; Kaylor, Andrew <andrew.kaylor at intel.com><mailto:andrew.kaylor at intel.com>; Topper, Craig <craig.topper at intel.com><mailto:craig.topper at intel.com>; Lu, Hongjiu <hongjiu.lu at intel.com><mailto:hongjiu.lu at intel.com> Subject: Re: [llvm-dev] Intel AMX programming model discussion. I find your answer unconvincing. I'm not going to debate it as I don't wish to take the time to build the appropriate context, but my initial response is skepticism. Philip On 8/14/20 4:49 PM, Luo, Yuanke wrote: [Yuanke] AMX register is special. It needs to be configured before use and the config instruction is expensive. To avoid unnecessary tile configure, we collect the tile shape information as much as possible and combine them into one ldtilecfg instruction. The ldtilecfg instruction should dominate any AMX instruction that access tile register. On the other side, the ldtilecfg should post-dominated the instruction that define the tile shape. For tile register spill, it should avoid re-config due to the different tile shape, the spilled register should be reloaded to the register that share the same tile shape. Since tile register allocation is special and it may allocate general virtual register to configure tile register, we can add a sperate pass to do it before general register allocation pass. After register allocation, the tile shape information is not needed anymore, so we can transform the pseudo AMX instruction to real AMX instruction by removing the row and column operands. [Philip] This seems complicated. Reading through the documentation, there appears to be a single global tile config for all tile registers at any time. Why not simply model this tile config as a designated special register and the tile instructions as having an implicit use of this register? That would seem to ensure that the register allocator has all the constraints needed. You'd need to teach it how to spill the special registers with the appropriate instructions, but that seems a lot more straight forward? [Yuanke] In that case user need to configure the tile register by themselves. Spilling configure register is very expensive, because it clears all the tile data register to zero. In our proposal, compiler is responsible to deduce the shape for virtual of tile data register, allocate physical registers for them and then configure those physical register. We may build the dependency as you proposed and it can be used for machine IR check to ensure tile data register is configured before use. From: Philip Reames <listmail at philipreames.com><mailto:listmail at philipreames.com> Sent: Saturday, August 15, 2020 1:17 AM To: Luo, Yuanke <yuanke.luo at intel.com><mailto:yuanke.luo at intel.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>; Kaylor, Andrew <andrew.kaylor at intel.com><mailto:andrew.kaylor at intel.com>; Topper, Craig <craig.topper at intel.com><mailto:craig.topper at intel.com>; Lu, Hongjiu <hongjiu.lu at intel.com><mailto:hongjiu.lu at intel.com> Subject: Re: [llvm-dev] Intel AMX programming model discussion. On 8/14/20 6:27 AM, Luo, Yuanke via llvm-dev wrote: Hi, Intel Advanced Matrix Extensions (Intel AMX) is a new programming paradigm consisting of two components: a set of 2-dimensional registers (tiles) representing sub-arrays from a larger 2-dimensional memory image, and accelerators able to operate on tiles. Capability of Intel AMX implementation is enumerated by palettes. Two palettes are supported: palette 0 represents the initialized state and palette 1 consists of 8 tile registers of up to 1 KB size, which is controlled by a tile control register. The instruction manual is posted at https://software.intel.com/content/www/us/en/develop/download/intel-architecture-instruction-set-extensions-programming-reference.html. The AMX abi proposal is posted at https://groups.google.com/g/x86-64-abi/c/NRejFm7pwb4. This email is to discuss the programming model for AMX. Florian has introduced the matrix type and intrinsics in LLVM community. We'd like to adopt some ideas from it. Here is what we propose for the AMX programming model. 1. Data type. We'd like to have fixed vector type for AMX. Since the shape to AMX register can be configurable, the vector size is the maximum size of AMX register. That means the vector size is 1024 bytes. The C code may look like this. typedef int _tile_data __attribute__((__vector_size__(1024), __aligned__(64))); _tile_data tile; And the LLVM IR may look like this. @tile = dso_local local_unnamed_addr global <256 x i32> zeroinitializer, align 64 For llvm IR, it is nice to have a new type x86_amxtile that can be mapped to AMX registers. 2. AMX Intrinsics. The internal intrinsics are 1:1 mapped to AMX instructions. The parameter m, n, k identifies the shape of the tile. The shape can be variable, but it cannot exceed the size that AMX HW can support. Compiler can deduce shape of the tile from the AMX intrinsics. _tile_data _tile_loadd_internal(char m, short n, const void *base, int stride); _tile_data _tile_dpbssd_internal(char m, short n, short k, _tile_data dst, _tile_data src1, _tile_data src2); _tile_data _tile_dpbf16ps_internal(char m, short n, short k, _tile_data dst, _tile_data src1, _tile_data src2); void _tile_stored_internal(char m, short n, void *base, int stride, _tile_data tile); 3. User interfaces. The tile shape and tile data are combined into a struct in C language. The shape of the tile is only allowed to be initialized once. The user interface looks as this. 3 #define __DEFAULT_FN_AMX \ 4 __attribute__((__always_inline__, __nodebug__, __target__("amx-int8"))) 9 typedef struct __tile_str { 10 const char row; 11 const short col; 12 _tile_data tile; 13 }__tile; 14 15 __DEFAULT_FN_AMX 16 void __tile_loadd(__tile *dst, const void *base, long stride) { 17 dst->tile = _tile_loadd_internal(dst->row, dst->col, base, stride); 18 } 19 20 __DEFAULT_FN_AMX 21 void __tile_dpbsud(__tile *dst, __tile src1, __tile src2) { 22 dst->tile = _tile_dpbssd_internal(src1.row, src2.col, src1.col, dst->tile, src1.tile, src2.tile); 23 } 24 25 __DEFAULT_FN_AMX 26 void __tile_stored(void *base, long stride, __tile src) { 27 _tile_stored_internal(src.row, src.col, base, stride, src.tile); 28 } 4. Example code The example shows how to use the user interface in a function. 51 void api(int cond, short row, short col) { 52 __tile a = {row, col}; 53 __tile b = {row, col}; 54 __tile c = {row, col}; 55 56 if(cond) { 57 __tile_loadd(&a, buf, STRIDE); 58 __tile_loadd(&b, buf, STRIDE); 59 __tile_loadd(&c, buf, STRIDE); 60 } else { 61 __tile_loadd(&a, buf2, STRIDE); 62 __tile_loadd(&b, buf2, STRIDE); 63 __tile_loadd(&c, buf2, STRIDE); 64 } 65 __tile_dpbsud(&c, a, b); 66 __tile_stored(buf, STRIDE, c); 67 } 5. LLVM IR The LLVM intrinsics IR take the row and column information as the input parameter, so that compiler can deduce the shape of tile data. The remaining parameters are what AMX instructions require. This is the LLVM IR corresponding to the example code. 12 define dso_local void @api(i32 %cond, i16 signext %row, i16 signext %col) local_unnamed_addr #2 { 13 entry: 14 %tobool = icmp eq i32 %cond, 0 15 %sext = shl i16 %col, 8 16 %conv.i31 = ashr exact i16 %sext, 8 17 br i1 %tobool, label %if.else, label %if.then 18 19 if.then: ; preds = %entry 20 %0 = tail call <256 x i32> @llvm.x86.tileloadd64(i16 %row, i16 %conv.i31, i8* getelementptr inbounds ([1024 x i8], [1024 x i8]* @buf, i64 0, i64 0), i64 32) #3 21 %1 = tail call <256 x i32> @llvm.x86.tileloadd64(i16 %row, i16 %conv.i31, i8* getelementptr inbounds ([1024 x i8], [1024 x i8]* @buf, i64 0, i64 0), i64 32) #3 22 %2 = tail call <256 x i32> @llvm.x86.tileloadd64(i16 %row, i16 %conv.i31, i8* getelementptr inbounds ([1024 x i8], [1024 x i8]* @buf, i64 0, i64 0), i64 32) #3 23 br label %if.end 24 25 if.else: ; preds = %entry 26 %3 = tail call <256 x i32> @llvm.x86.tileloadd64(i16 %row, i16 %conv.i31, i8* getelementptr inbounds ([1024 x i8], [1024 x i8]* @buf2, i64 0, i64 0), i64 32) #3 27 %4 = tail call <256 x i32> @llvm.x86.tileloadd64(i16 %row, i16 %conv.i31, i8* getelementptr inbounds ([1024 x i8], [1024 x i8]* @buf2, i64 0, i64 0), i64 32) #3 28 %5 = tail call <256 x i32> @llvm.x86.tileloadd64(i16 %row, i16 %conv.i31, i8* getelementptr inbounds ([1024 x i8], [1024 x i8]* @buf2, i64 0, i64 0), i64 32) #3 29 br label %if.end 30 31 if.end: ; preds = %if.else, %if.then 32 %a.sroa.1186.0 = phi <256 x i32> [ %3, %if.else ], [ %0, %if.then ] 33 %b.sroa.1068.0 = phi <256 x i32> [ %4, %if.else ], [ %1, %if.then ] 34 %c.sroa.1149.0 = phi <256 x i32> [ %5, %if.else ], [ %2, %if.then ] 35 %6 = tail call <256 x i32> @llvm.x86.tdpbssd(i16 %row, i16 %conv.i31, i16 %conv.i31, <256 x i32> %c.sroa.1149.0, <256 x i32> %a.sroa.1186.0, <256 x i32> %b.sroa.1068.0) #3 36 tail call void @llvm.x86.tilestored64(i16 %row, i16 %conv.i31, i8* getelementptr inbounds ([1024 x i8], [1024 x i8]* @buf, i64 0, i64 0), i64 32, <256 x i32> %6) #3 37 ret void 38 } 6. Shape propagation When in -O0 build, some general load/store for tile vector is generated by front-end. We need to root from AMX intrinsics to propagate the shape information to the virtual tile register. If the an AMX intrinsic use the result of load instruction, the shape is propagated to the load and the load is transformed to tile load intrinsic. If the store instruction uses any result of AMX intrinsic, the shape is propagated to store instruction and the store is transformed to tile store intrinsic 7. Machine IR Since the AMX intrinsics take the row and column as the input parameters, we can create a pseudo instruction corresponding to it. The AMX intrinsics are lowered to the pseudo AMX instruction which has extra row and column operands corresponding to AMX intrinsic. The real AMX instructions don't need the row and column operands. The row and column information should be configured by ldtilecfg before executing any AMX instruction. 8. Register allocation AMX register is special. It needs to be configured before use and the config instruction is expensive. To avoid unnecessary tile configure, we collect the tile shape information as much as possible and combine them into one ldtilecfg instruction. The ldtilecfg instruction should dominate any AMX instruction that access tile register. On the other side, the ldtilecfg should post-dominated the instruction that define the tile shape. For tile register spill, it should avoid re-config due to the different tile shape, the spilled register should be reloaded to the register that share the same tile shape. Since tile register allocation is special and it may allocate general virtual register to configure tile register, we can add a sperate pass to do it before general register allocation pass. After register allocation, the tile shape information is not needed anymore, so we can transform the pseudo AMX instruction to real AMX instruction by removing the row and column operands. This seems complicated. Reading through the documentation, there appears to be a single global tile config for all tile registers at any time. Why not simply model this tile config as a designated special register and the tile instructions as having an implicit use of this register? That would seem to ensure that the register allocator has all the constraints needed. You'd need to teach it how to spill the special registers with the appropriate instructions, but that seems a lot more straight forward? 9. Use recommendation Due to the shape configure issue, we recommend user to define the tile shape at the entry of the function entry and inline function as much as possible. The AMX instructions focus on computation instead of storage, so global variable for tile data is not recommended. Thanks Yuanke _______________________________________________ LLVM Developers mailing list llvm-dev at lists.llvm.org<mailto:llvm-dev at lists.llvm.org> https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev _______________________________________________ LLVM Developers mailing list llvm-dev at lists.llvm.org<mailto:llvm-dev at lists.llvm.org> https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev -- 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 -- 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 -- 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/20200820/46696943/attachment.html>
Luo, Yuanke via llvm-dev
2020-Aug-21 06:12 UTC
[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. a. 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. b. Whether 256 register class is enough to be allocated. There may be more 256 unknow shape tile registers. c. 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 From: Hal Finkel <hfinkel at anl.gov> Sent: Friday, August 21, 2020 3:35 AM To: Topper, Craig <craig.topper at intel.com>; Kaylor, Andrew <andrew.kaylor at intel.com>; Luo, Yuanke <yuanke.luo 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 8/19/20 3:09 PM, Topper, Craig wrote: The width and height can be runtime values that we would just copy into 64 byte configuration block we pass to ldtilecfg. So the code doesn't need to be multiversioned. The user code would also use those values to update pointers in the loops they write using the tiles. If we can't determine that two tiles were defined with the same width and height we need to assume the shape is different and try to avoid ever giving the same tile. Hal, for your suggestion would which physical registers are in which register class be defined dynamically before register allocation? Here's my thought: First, you have a set of intrinsics that take tile values along with tile configuration parameters (which, presently, seem just to be the sizes). These get lowered into pseudo-instructions that do the same. Thus, you have some register class that represents these arbitrarily-sized tile registers that you'll assign to these pseudo-instruction operands (i.e., they take virtual tile registers right after instruction selection). You might use the 16x16 tile register class for this purpose, but it shouldn't really matter. Second, you run this configuration-placement pass. This pass looks at all of the AMX pseudo-instructions and identifies regions in which the pseudo-instructions use the same configuration parameters (i.e., the same SSA values and/or constants). This pass might reorder the pseudo-instructions when legal in order to form larger regions. Then it places the ldtilecfg at the start of each region (in some common dominating position). ldtilecfg implicitly defines all of the tile registers in every concrete class of tile registers (all 256 of them, or whatever). The pseudo-instructions are replaced by real MI instructions taking a tile register class appropriate for the configuration (which will default to the 16x16 class for cases where the configuration is not a compile-time-known constant). When the configuration is a known constant, the instructions take operands with a register class appropriate for that configuration (e.g., 1x1, 4x4). Third, the rest of the framework runs as usual. Tile registers from the appropriate class are allocated by the register allocator. No live range of any virtual tile register can pass through the ldtilecfg (because it defines them all), but that's okay, none of live ranges will by construction (the configuration-placement pass ensures this). Thanks again, Hal From: Hal Finkel <hfinkel at anl.gov><mailto:hfinkel at anl.gov> Sent: Wednesday, August 19, 2020 12:52 PM To: Kaylor, Andrew <andrew.kaylor at intel.com><mailto:andrew.kaylor at intel.com>; Luo, Yuanke <yuanke.luo at intel.com><mailto:yuanke.luo 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>; Topper, Craig <craig.topper at intel.com><mailto:craig.topper at intel.com>; Lu, Hongjiu <hongjiu.lu at intel.com><mailto:hongjiu.lu at intel.com> Subject: Re: [llvm-dev] Intel AMX programming model discussion. On 8/19/20 10:24 AM, Kaylor, Andrew wrote:> When the tile shape is unknown at compile time, how do you plan to do the register allocation of the tiles? My question is: do you do the allocation for this case in the same way as you would if you knew the size was 16x16 (i.e., conservatively assume the largest size)?I think what will happen is that the registers are allocated based on a number of runtime values that are assumed to be different from one another but less than or equal to 16. So, for example, we'll allocate registers for MxN tiles, NxM tiles and MxM tiles without knowing what M and N are. Then at runtime the values of these variables will be used to create the actual tile configuration. The instructions that need to know the shape take these runtime values as operands. So you're going to multiversion the code? In any case, my point is that you probably don't need a custom register allocator. If you just define the tile registers and make sure that the ldtilecfgs implicitly defines them all, then the regular infrastructure likely works. You'll have a bunch of register classes, but that's not necessarily a problem. I recommend trying this, and let us know what you discover, before we go down the road of a new, dedicated allocator just for these registers. -Hal There may be some artifacts coming from the front end that conservatively assume a 16x16 tile, but I think those generally go away in SROA or later specialized passes. Yuanke can confirm or correct my understanding of this. From: Hal Finkel <hfinkel at anl.gov><mailto:hfinkel at anl.gov> Sent: Wednesday, August 19, 2020 5:14 AM To: Luo, Yuanke <yuanke.luo at intel.com><mailto:yuanke.luo 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>; Topper, Craig <craig.topper at intel.com><mailto:craig.topper at intel.com>; Lu, Hongjiu <hongjiu.lu at intel.com><mailto:hongjiu.lu at intel.com> Subject: Re: [llvm-dev] Intel AMX programming model discussion. On 8/19/20 5:34 AM, Luo, Yuanke wrote: There is no problem to have 256 register classes. Just a lot of register classes to me. We don't assume the shape of each physical register be 16x16, it is defined by user. For variable shape, I mean the shape is known in runtime and in compile time the shape is unknown. Take below code as an example, the %row and %col are variable instead of constant. Compiler recognizes llvm.x86.tileloadd64 and deduce the shape of %0 is %row x %col. %0 = tail call <256 x i32> @llvm.x86.tileloadd64(i16 %row, i16 %col, i8* getelementptr inbounds ([1024 x i8], [1024 x i8]* @buf, i64 0, i64 0), i64 32) When the tile shape is unknown at compile time, how do you plan to do the register allocation of the tiles? My question is: do you do the allocation for this case in the same way as you would if you knew the size was 16x16 (i.e., conservatively assume the largest size)? Thanks again, Hal From: Hal Finkel <hfinkel at anl.gov><mailto:hfinkel at anl.gov> Sent: Wednesday, August 19, 2020 4:58 PM To: Luo, Yuanke <yuanke.luo at intel.com><mailto:yuanke.luo 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>; Topper, Craig <craig.topper at intel.com><mailto:craig.topper at intel.com>; Lu, Hongjiu <hongjiu.lu at intel.com><mailto:hongjiu.lu at intel.com> Subject: Re: [llvm-dev] Intel AMX programming model discussion. On 8/19/20 2:21 AM, Luo, Yuanke wrote: Hi Hal, There is 3 aspect to be solved. 1. The HW support max shape 16x16, so there are many register classes from 1x1 to 16x16. We need 256 register classes. 2. We want to support variable shape, so compiler don't know what register class to fit tile shape as it is only known in runtime. 3. The tile configure is to configure physical tile register, so we need to allocate register and then we know the shape of each physical tile register and configure the tile register. I think your suggestion is helpful to reduce the complexity if we only support fixed (constant) tile shape. -Yuanke Thanks, Yuanke. It's not clear to me that having 256 register classes is, in itself, a problem. Is it? What does it mean to support variable-shape tiles in this context? Do you do something other than conservatively assume that they are 16x16 for register-allocation purposes? -Hal From: Hal Finkel <hfinkel at anl.gov><mailto:hfinkel at anl.gov> Sent: Wednesday, August 19, 2020 8:20 AM To: Kaylor, Andrew <andrew.kaylor at intel.com><mailto:andrew.kaylor at intel.com>; Philip Reames <listmail at philipreames.com><mailto:listmail at philipreames.com>; Luo, Yuanke <yuanke.luo at intel.com><mailto:yuanke.luo at intel.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>; Topper, Craig <craig.topper at intel.com><mailto:craig.topper at intel.com>; Lu, Hongjiu <hongjiu.lu at intel.com><mailto:hongjiu.lu at intel.com> Subject: Re: [llvm-dev] Intel AMX programming model discussion. Hi, Andy, I don't quite understand everything that's going on here. Could we model this as: 1. Define a collection of register classes, one for 2x4 tiles, one for 4x2 tiles, etc. each populated with a set of tile registers. Registers can have aliasing relationships (instead of worrying of any kind of subregister/superregister relationships -- these won't be useful anyway). 2. Define the tile-configuration instructions so that they implicitly define all of the registers in all of the classes. Then you would still need to pre-schedule the tile operations as you've described, and collect the configuration information in order to add the ldtilecfgs, but the regular register allocator can handle the allocation itself in the usual way. What do you think? -Hal On 8/18/20 6:58 PM, Kaylor, Andrew via llvm-dev wrote: The AMX registers are complicated. The single configuration register (which is mostly used implicitly, similar to MXCSR for floating point) controls the shape of all the tile registers, and if you change the tile configuration every single tile register is cleared. In practice, if we have to change the the configuration while any of the tile registers are live, performance is going to be terrible. We need to handle this case for correctness, but users of this programming interface will need to have enough awareness of the performance issues and the hardware details to prevent this. We'll also want a diagnostic that lets the user know when this has happened. When the tile configuration is set, the shape of each tile is locked in, so the individual tile registers aren't interchangeable at that point. If a function needs 2x4 tiles, 4x2 tiles, and 4x4 tiles, the configuration needs to be set with this in mind. The shape isn't explicit in every instruction and intrinsic. It must be deduced. And again, we'll need a way to tell the user when efficient allocation can't be done. In practice, I don't expect any function to be using more than three tile shapes. The implication of all this is that I don't think the greedy register allocator is well suited to figure all of this out. We need a special pass to pre-allocate these registers. If the function is written in a way that makes good performance possible, it should be a relatively simple task to allocate everything with minimal spilling. If it isn't possible to get good performance, we don't need to do anything especially clever. We can just do something straightforward that is correct and let the user know that they aren't going to be happy with the results. -Andy From: Philip Reames <listmail at philipreames.com><mailto:listmail at philipreames.com> Sent: Friday, August 14, 2020 8:29 PM To: Luo, Yuanke <yuanke.luo at intel.com><mailto:yuanke.luo at intel.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>; Kaylor, Andrew <andrew.kaylor at intel.com><mailto:andrew.kaylor at intel.com>; Topper, Craig <craig.topper at intel.com><mailto:craig.topper at intel.com>; Lu, Hongjiu <hongjiu.lu at intel.com><mailto:hongjiu.lu at intel.com> Subject: Re: [llvm-dev] Intel AMX programming model discussion. I find your answer unconvincing. I'm not going to debate it as I don't wish to take the time to build the appropriate context, but my initial response is skepticism. Philip On 8/14/20 4:49 PM, Luo, Yuanke wrote: [Yuanke] AMX register is special. It needs to be configured before use and the config instruction is expensive. To avoid unnecessary tile configure, we collect the tile shape information as much as possible and combine them into one ldtilecfg instruction. The ldtilecfg instruction should dominate any AMX instruction that access tile register. On the other side, the ldtilecfg should post-dominated the instruction that define the tile shape. For tile register spill, it should avoid re-config due to the different tile shape, the spilled register should be reloaded to the register that share the same tile shape. Since tile register allocation is special and it may allocate general virtual register to configure tile register, we can add a sperate pass to do it before general register allocation pass. After register allocation, the tile shape information is not needed anymore, so we can transform the pseudo AMX instruction to real AMX instruction by removing the row and column operands. [Philip] This seems complicated. Reading through the documentation, there appears to be a single global tile config for all tile registers at any time. Why not simply model this tile config as a designated special register and the tile instructions as having an implicit use of this register? That would seem to ensure that the register allocator has all the constraints needed. You'd need to teach it how to spill the special registers with the appropriate instructions, but that seems a lot more straight forward? [Yuanke] In that case user need to configure the tile register by themselves. Spilling configure register is very expensive, because it clears all the tile data register to zero. In our proposal, compiler is responsible to deduce the shape for virtual of tile data register, allocate physical registers for them and then configure those physical register. We may build the dependency as you proposed and it can be used for machine IR check to ensure tile data register is configured before use. From: Philip Reames <listmail at philipreames.com><mailto:listmail at philipreames.com> Sent: Saturday, August 15, 2020 1:17 AM To: Luo, Yuanke <yuanke.luo at intel.com><mailto:yuanke.luo at intel.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>; Kaylor, Andrew <andrew.kaylor at intel.com><mailto:andrew.kaylor at intel.com>; Topper, Craig <craig.topper at intel.com><mailto:craig.topper at intel.com>; Lu, Hongjiu <hongjiu.lu at intel.com><mailto:hongjiu.lu at intel.com> Subject: Re: [llvm-dev] Intel AMX programming model discussion. On 8/14/20 6:27 AM, Luo, Yuanke via llvm-dev wrote: Hi, Intel Advanced Matrix Extensions (Intel AMX) is a new programming paradigm consisting of two components: a set of 2-dimensional registers (tiles) representing sub-arrays from a larger 2-dimensional memory image, and accelerators able to operate on tiles. Capability of Intel AMX implementation is enumerated by palettes. Two palettes are supported: palette 0 represents the initialized state and palette 1 consists of 8 tile registers of up to 1 KB size, which is controlled by a tile control register. The instruction manual is posted at https://software.intel.com/content/www/us/en/develop/download/intel-architecture-instruction-set-extensions-programming-reference.html. The AMX abi proposal is posted at https://groups.google.com/g/x86-64-abi/c/NRejFm7pwb4. This email is to discuss the programming model for AMX. Florian has introduced the matrix type and intrinsics in LLVM community. We'd like to adopt some ideas from it. Here is what we propose for the AMX programming model. 1. Data type. We'd like to have fixed vector type for AMX. Since the shape to AMX register can be configurable, the vector size is the maximum size of AMX register. That means the vector size is 1024 bytes. The C code may look like this. typedef int _tile_data __attribute__((__vector_size__(1024), __aligned__(64))); _tile_data tile; And the LLVM IR may look like this. @tile = dso_local local_unnamed_addr global <256 x i32> zeroinitializer, align 64 For llvm IR, it is nice to have a new type x86_amxtile that can be mapped to AMX registers. 2. AMX Intrinsics. The internal intrinsics are 1:1 mapped to AMX instructions. The parameter m, n, k identifies the shape of the tile. The shape can be variable, but it cannot exceed the size that AMX HW can support. Compiler can deduce shape of the tile from the AMX intrinsics. _tile_data _tile_loadd_internal(char m, short n, const void *base, int stride); _tile_data _tile_dpbssd_internal(char m, short n, short k, _tile_data dst, _tile_data src1, _tile_data src2); _tile_data _tile_dpbf16ps_internal(char m, short n, short k, _tile_data dst, _tile_data src1, _tile_data src2); void _tile_stored_internal(char m, short n, void *base, int stride, _tile_data tile); 3. User interfaces. The tile shape and tile data are combined into a struct in C language. The shape of the tile is only allowed to be initialized once. The user interface looks as this. 3 #define __DEFAULT_FN_AMX \ 4 __attribute__((__always_inline__, __nodebug__, __target__("amx-int8"))) 9 typedef struct __tile_str { 10 const char row; 11 const short col; 12 _tile_data tile; 13 }__tile; 14 15 __DEFAULT_FN_AMX 16 void __tile_loadd(__tile *dst, const void *base, long stride) { 17 dst->tile = _tile_loadd_internal(dst->row, dst->col, base, stride); 18 } 19 20 __DEFAULT_FN_AMX 21 void __tile_dpbsud(__tile *dst, __tile src1, __tile src2) { 22 dst->tile = _tile_dpbssd_internal(src1.row, src2.col, src1.col, dst->tile, src1.tile, src2.tile); 23 } 24 25 __DEFAULT_FN_AMX 26 void __tile_stored(void *base, long stride, __tile src) { 27 _tile_stored_internal(src.row, src.col, base, stride, src.tile); 28 } 4. Example code The example shows how to use the user interface in a function. 51 void api(int cond, short row, short col) { 52 __tile a = {row, col}; 53 __tile b = {row, col}; 54 __tile c = {row, col}; 55 56 if(cond) { 57 __tile_loadd(&a, buf, STRIDE); 58 __tile_loadd(&b, buf, STRIDE); 59 __tile_loadd(&c, buf, STRIDE); 60 } else { 61 __tile_loadd(&a, buf2, STRIDE); 62 __tile_loadd(&b, buf2, STRIDE); 63 __tile_loadd(&c, buf2, STRIDE); 64 } 65 __tile_dpbsud(&c, a, b); 66 __tile_stored(buf, STRIDE, c); 67 } 5. LLVM IR The LLVM intrinsics IR take the row and column information as the input parameter, so that compiler can deduce the shape of tile data. The remaining parameters are what AMX instructions require. This is the LLVM IR corresponding to the example code. 12 define dso_local void @api(i32 %cond, i16 signext %row, i16 signext %col) local_unnamed_addr #2 { 13 entry: 14 %tobool = icmp eq i32 %cond, 0 15 %sext = shl i16 %col, 8 16 %conv.i31 = ashr exact i16 %sext, 8 17 br i1 %tobool, label %if.else, label %if.then 18 19 if.then: ; preds = %entry 20 %0 = tail call <256 x i32> @llvm.x86.tileloadd64(i16 %row, i16 %conv.i31, i8* getelementptr inbounds ([1024 x i8], [1024 x i8]* @buf, i64 0, i64 0), i64 32) #3 21 %1 = tail call <256 x i32> @llvm.x86.tileloadd64(i16 %row, i16 %conv.i31, i8* getelementptr inbounds ([1024 x i8], [1024 x i8]* @buf, i64 0, i64 0), i64 32) #3 22 %2 = tail call <256 x i32> @llvm.x86.tileloadd64(i16 %row, i16 %conv.i31, i8* getelementptr inbounds ([1024 x i8], [1024 x i8]* @buf, i64 0, i64 0), i64 32) #3 23 br label %if.end 24 25 if.else: ; preds = %entry 26 %3 = tail call <256 x i32> @llvm.x86.tileloadd64(i16 %row, i16 %conv.i31, i8* getelementptr inbounds ([1024 x i8], [1024 x i8]* @buf2, i64 0, i64 0), i64 32) #3 27 %4 = tail call <256 x i32> @llvm.x86.tileloadd64(i16 %row, i16 %conv.i31, i8* getelementptr inbounds ([1024 x i8], [1024 x i8]* @buf2, i64 0, i64 0), i64 32) #3 28 %5 = tail call <256 x i32> @llvm.x86.tileloadd64(i16 %row, i16 %conv.i31, i8* getelementptr inbounds ([1024 x i8], [1024 x i8]* @buf2, i64 0, i64 0), i64 32) #3 29 br label %if.end 30 31 if.end: ; preds = %if.else, %if.then 32 %a.sroa.1186.0 = phi <256 x i32> [ %3, %if.else ], [ %0, %if.then ] 33 %b.sroa.1068.0 = phi <256 x i32> [ %4, %if.else ], [ %1, %if.then ] 34 %c.sroa.1149.0 = phi <256 x i32> [ %5, %if.else ], [ %2, %if.then ] 35 %6 = tail call <256 x i32> @llvm.x86.tdpbssd(i16 %row, i16 %conv.i31, i16 %conv.i31, <256 x i32> %c.sroa.1149.0, <256 x i32> %a.sroa.1186.0, <256 x i32> %b.sroa.1068.0) #3 36 tail call void @llvm.x86.tilestored64(i16 %row, i16 %conv.i31, i8* getelementptr inbounds ([1024 x i8], [1024 x i8]* @buf, i64 0, i64 0), i64 32, <256 x i32> %6) #3 37 ret void 38 } 6. Shape propagation When in -O0 build, some general load/store for tile vector is generated by front-end. We need to root from AMX intrinsics to propagate the shape information to the virtual tile register. If the an AMX intrinsic use the result of load instruction, the shape is propagated to the load and the load is transformed to tile load intrinsic. If the store instruction uses any result of AMX intrinsic, the shape is propagated to store instruction and the store is transformed to tile store intrinsic 7. Machine IR Since the AMX intrinsics take the row and column as the input parameters, we can create a pseudo instruction corresponding to it. The AMX intrinsics are lowered to the pseudo AMX instruction which has extra row and column operands corresponding to AMX intrinsic. The real AMX instructions don't need the row and column operands. The row and column information should be configured by ldtilecfg before executing any AMX instruction. 8. Register allocation AMX register is special. It needs to be configured before use and the config instruction is expensive. To avoid unnecessary tile configure, we collect the tile shape information as much as possible and combine them into one ldtilecfg instruction. The ldtilecfg instruction should dominate any AMX instruction that access tile register. On the other side, the ldtilecfg should post-dominated the instruction that define the tile shape. For tile register spill, it should avoid re-config due to the different tile shape, the spilled register should be reloaded to the register that share the same tile shape. Since tile register allocation is special and it may allocate general virtual register to configure tile register, we can add a sperate pass to do it before general register allocation pass. After register allocation, the tile shape information is not needed anymore, so we can transform the pseudo AMX instruction to real AMX instruction by removing the row and column operands. This seems complicated. Reading through the documentation, there appears to be a single global tile config for all tile registers at any time. Why not simply model this tile config as a designated special register and the tile instructions as having an implicit use of this register? That would seem to ensure that the register allocator has all the constraints needed. You'd need to teach it how to spill the special registers with the appropriate instructions, but that seems a lot more straight forward? 9. Use recommendation Due to the shape configure issue, we recommend user to define the tile shape at the entry of the function entry and inline function as much as possible. The AMX instructions focus on computation instead of storage, so global variable for tile data is not recommended. Thanks Yuanke _______________________________________________ LLVM Developers mailing list llvm-dev at lists.llvm.org<mailto:llvm-dev at lists.llvm.org> https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev _______________________________________________ LLVM Developers mailing list llvm-dev at lists.llvm.org<mailto:llvm-dev at lists.llvm.org> https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev -- 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 -- 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 -- 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/20200821/37d18c18/attachment-0001.html>