The original change was pushed in main as commit f7a23ecece52.
It was then reverted by commit a04f01bab2 because it caused linker failures
on buildbots that don't build the AMDGPU target.
--
Some instructions are not defined well enough within the target’s scheduling
model for llvm-mca to be able to properly simulate its behaviour. The ideal
solution to this situation is to modify the scheduling model, but that’s not
always a viable strategy. Maybe other parts of the backend depend on that
instruction being modelled the way that it is. Or maybe the instruction is quite
complex and it’s difficult to fully capture its behaviour with tablegen. The
CustomBehaviour class (which I will refer to as CB frequently) is designed to
provide intuitive scaffolding for developers to implement the correct modelling
for these instructions.
More details are available in the original commit log message (f7a23ecece52).
Differential Revision: https://reviews.llvm.org/D104149
Some instructions are not defined well enough within the target’s scheduling
model for llvm-mca to be able to properly simulate its behaviour. The ideal
solution to this situation is to modify the scheduling model, but that’s not
always a viable strategy. Maybe other parts of the backend depend on that
instruction being modelled the way that it is. Or maybe the instruction is quite
complex and it’s difficult to fully capture its behaviour with tablegen. The
CustomBehaviour class (which I will refer to as CB frequently) is designed to
provide intuitive scaffolding for developers to implement the correct modelling
for these instructions.
Implementation details:
llvm-mca does its best to extract relevant register, resource, and memory
information from every MCInst when lowering them to an mca::Instruction. It then
uses this information to detect dependencies and simulate stalls within the
pipeline. For some instructions, the information that gets captured within the
mca::Instruction is not enough for mca to simulate them properly. In these
cases, there are two main possibilities:
1. The instruction has a dependency that isn’t detected by mca.
2. mca is incorrectly enforcing a dependency that shouldn’t exist.
For the rest of this discussion, I will be focusing on (1), but I have put some
thought into (2) and I may revisit it in the future.
So we have an instruction that has dependencies that aren’t picked up by mca.
The basic idea for both pipelines in mca is that when an instruction wants to be
dispatched, we first check for register hazards and then we check for resource
hazards. This is where CB is injected. If no register or resource hazards have
been detected, we make a call to CustomBehaviour::checkCustomHazard() to give
the target specific CB the chance to detect and enforce any custom dependencies.
The return value for checkCustomHazaard() is an unsigned int representing the
(minimum) number of cycles that the instruction needs to stall for. It’s fine to
underestimate this value because when StallCycles gets down to 0, we’ll end up
checking for all the hazards again before the instruction is actually
dispatched. However, it’s important not to overestimate the value and the more
accurate your estimate is, the more efficient mca’s execution can be.
In general, for checkCustomHazard() to be able to detect these custom
dependencies, it needs information about the current instruction and also all of
the instructions that are still executing within the pipeline. The mca pipeline
uses mca::Instruction rather than MCInst and the current information encoded
within each mca::Instruction isn’t sufficient for my use cases. I had to add a
few extra attributes to the mca::Instruction class and have them get set by the
MCInst during instruction building. For example, the current mca::Instruction
doesn’t know its opcode, and it also doesn’t know anything about its immediate
operands (both of which I had to add to the class).
With information about the current instruction, a list of all currently
executing instructions, and some target specific objects (MCSubtargetInfo and
MCInstrInfo which the base CB class has references to), developers should be
able to detect and enforce most custom dependencies within checkCustomHazard. If
you need more information than is present in the mca::Instruction, feel free to
add attributes to that class and have them set during the lowering sequence from
MCInst.
Fortunately, in the in-order pipeline, it’s very convenient for us to pass these
arguments to checkCustomHazard. The hazard checking is taken care of within
InOrderIssueStage::canExecute(). This function takes a const InstRef as a
parameter (representing the instruction that currently wants to be dispatched)
and the InOrderIssueStage class maintains a SmallVector<InstRef, 4> which holds
all of the currently executing instructions. For the out-of-order pipeline, it’s
a bit trickier to get the list of executing instructions and this is why I have
held off on implementing it myself. This is the main topic I will bring up when
I eventually make a post to discuss and ask for feedback.
CB is a base class where targets implement their own derived classes. If a
target specific CB does not exist (or we pass in the -disable-cb flag), the base
class is used. This base class trivially returns 0 from its checkCustomHazard()
implementation (meaning that the current instruction needs to stall for 0 cycles
aka no hazard is detected). For this reason, targets or users who choose not to
use CB shouldn’t see any negative impacts to accuracy or performance (in
comparison to pre-patch llvm-mca).
Differential Revision: https://reviews.llvm.org/D104149
This is based on the assumption that most simulated instructions don't define
more than one or two registers. This is true for example on x86, where
most instruction definitions don't declare more than one register write.
The default code region size has been increased from 8 to 16. This is based on
the assumption that, for small microbenchmarks, the typical code snippet size is
often less than 16 instructions.
mca::Instruction now uses bitfields to pack flags.
No functional change intended.
The constructor of InOrderIssueStage no longer takes as input a reference to the
target scheduling model. The stage can always query the subtarget to obtain a
reference to the scheduling model.
The ResourceManager is no longer stored internally as a unique_ptr.
Moved a couple of method definitions to the .cpp file.
This is a follow-up for:
D98604 [MCA] Ensure that writes occur in-order
When instructions are aligned by the order of writes, they retire
in-order naturally. There is no need for an RCU, so it is disabled.
Differential Revision: https://reviews.llvm.org/D98628
This patch adds a pipeline to support in-order CPUs such as ARM
Cortex-A55.
In-order pipeline implements a simplified version of Dispatch,
Scheduler and Execute stages as a single stage. Entry and Retire
stages are common for both in-order and out-of-order pipelines.
Differential Revision: https://reviews.llvm.org/D94928
Before this patch, loads and stores were only tracked by their corresponding
queues in the LSUnit from dispatch until execute stage. In practice we should be
more conservative and assume that memory opcodes leave their queues at
retirement stage.
Basically, loads should leave the load queue only when they have completed and
delivered their data. We conservatively assume that a load is completed when it
is retired. Stores should be tracked by the store queue from dispatch until
retirement. In practice, stores can only leave the store queue if their data can
be written to the data cache.
This is mostly a mechanical change. With this patch, the retire stage notifies
the LSUnit when a memory instruction is retired. That would triggers the release
of LDQ/STQ entries. The only visible change is in memory tests for the bdver2
model. That is because bdver2 is the only model that defines the load/store
queue size.
This patch partially addresses PR39830.
Differential Revision: https://reviews.llvm.org/D68266
llvm-svn: 374034
Now that we've moved to C++14, we no longer need the llvm::make_unique
implementation from STLExtras.h. This patch is a mechanical replacement
of (hopefully) all the llvm::make_unique instances across the monorepo.
llvm-svn: 369013
This patch adds an experimental stage named MicroOpQueueStage.
MicroOpQueueStage can be used to simulate a hardware micro-op queue (basically,
a decoupling queue between 'decode' and 'dispatch'). Users can specify a queue
size, as well as a optional MaxIPC (which - in the absence of a "Decoders" stage
- can be used to simulate a different throughput from the decoders).
This stage is added to the default pipeline between the EntryStage and the
DispatchStage only if PipelineOption::MicroOpQueue is different than zero. By
default, llvm-mca sets PipelineOption::MicroOpQueue to the value of hidden flag
-micro-op-queue-size.
Throughput from the decoder can be simulated via another hidden flag named
-decoder-throughput. That flag allows us to quickly experiment with different
frontend throughputs. For targets that declare a loop buffer, flag
-decoder-throughput allows users to do multiple runs, each time simulating a
different throughput from the decoders.
This stage can/will be extended in future. For example, we could add a "buffer
full" event to notify bottlenecks caused by backpressure. flag
-decoder-throughput would probably go away if in future we delegate to another
stage (DecoderStage?) the simulation of a (potentially variable) throughput from
the decoders. For now, flag -decoder-throughput is "good enough" to run some
simple experiments.
Differential Revision: https://reviews.llvm.org/D59928
llvm-svn: 357248
This patch adds a new flag named -bottleneck-analysis to print out information
about throughput bottlenecks.
MCA knows how to identify and classify dynamic dispatch stalls. However, it
doesn't know how to analyze and highlight kernel bottlenecks. The goal of this
patch is to teach MCA how to correlate increases in backend pressure to backend
stalls (and therefore, the loss of throughput).
From a Scheduler point of view, backend pressure is a function of the scheduler
buffer usage (i.e. how the number of uOps in the scheduler buffers changes over
time). Backend pressure increases (or decreases) when there is a mismatch
between the number of opcodes dispatched, and the number of opcodes issued in
the same cycle. Since buffer resources are limited, continuous increases in
backend pressure would eventually leads to dispatch stalls. So, there is a
strong correlation between dispatch stalls, and how backpressure changed over
time.
This patch teaches how to identify situations where backend pressure increases
due to:
- unavailable pipeline resources.
- data dependencies.
Data dependencies may delay execution of instructions and therefore increase the
time that uOps have to spend in the scheduler buffers. That often translates to
an increase in backend pressure which may eventually lead to a bottleneck.
Contention on pipeline resources may also delay execution of instructions, and
lead to a temporary increase in backend pressure.
Internally, the Scheduler classifies instructions based on whether register /
memory operands are available or not.
An instruction is marked as "ready to execute" only if data dependencies are
fully resolved.
Every cycle, the Scheduler attempts to execute all instructions that are ready
to execute. If an instruction cannot execute because of unavailable pipeline
resources, then the Scheduler internally updates a BusyResourceUnits mask with
the ID of each unavailable resource.
ExecuteStage is responsible for tracking changes in backend pressure. If backend
pressure increases during a cycle because of contention on pipeline resources,
then ExecuteStage sends a "backend pressure" event to the listeners.
That event would contain information about instructions delayed by resource
pressure, as well as the BusyResourceUnits mask.
Note that ExecuteStage also knows how to identify situations where backpressure
increased because of delays introduced by data dependencies.
The SummaryView observes "backend pressure" events and prints out a "bottleneck
report".
Example of bottleneck report:
```
Cycles with backend pressure increase [ 99.89% ]
Throughput Bottlenecks:
Resource Pressure [ 0.00% ]
Data Dependencies: [ 99.89% ]
- Register Dependencies [ 0.00% ]
- Memory Dependencies [ 99.89% ]
```
A bottleneck report is printed out only if increases in backend pressure
eventually caused backend stalls.
About the time complexity:
Time complexity is linear in the number of instructions in the
Scheduler::PendingSet.
The average slowdown tends to be in the range of ~5-6%.
For memory intensive kernels, the slowdown can be significant if flag
-noalias=false is specified. In the worst case scenario I have observed a
slowdown of ~30% when flag -noalias=false was specified.
We can definitely recover part of that slowdown if we optimize class LSUnit (by
doing extra bookkeeping to speedup queries). For now, this new analysis is
disabled by default, and it can be enabled via flag -bottleneck-analysis. Users
of MCA as a library can enable the generation of pressure events through the
constructor of ExecuteStage.
This patch partially addresses https://bugs.llvm.org/show_bug.cgi?id=37494
Differential Revision: https://reviews.llvm.org/D58728
llvm-svn: 355308
to reflect the new license.
We understand that people may be surprised that we're moving the header
entirely to discuss the new license. We checked this carefully with the
Foundation's lawyer and we believe this is the correct approach.
Essentially, all code in the project is now made available by the LLVM
project under our new license, so you will see that the license headers
include that license only. Some of our contributors have contributed
code under our old license, and accordingly, we have retained a copy of
our old license notice in the top-level files in each project and
repository.
llvm-svn: 351636
