1
0
mirror of https://github.com/RPCS3/llvm-mirror.git synced 2024-11-22 10:42:39 +01:00
llvm-mirror/docs/CommandGuide/llvm-exegesis.rst
Roman Lebedev 5d534d8259 [llvm-exegesis] Loop unrolling for loop snippet repetitor mode
I really needed this, like, factually, yesterday,
when verifying dependency breaking idioms for AMD Zen 3 scheduler model.

Consider the following example:
```
$ ./bin/llvm-exegesis --mode=inverse_throughput --snippets-file=/tmp/snippet.s --num-repetitions=1000000 --repetition-mode=duplicate
Check generated assembly with: /usr/bin/objdump -d /tmp/snippet-4a7e50.o
---
mode:            inverse_throughput
key:
  instructions:
    - 'VPXORYrr YMM0 YMM0 YMM0'
  config:          ''
  register_initial_values: []
cpu_name:        znver3
llvm_triple:     x86_64-unknown-linux-gnu
num_repetitions: 1000000
measurements:
  - { key: inverse_throughput, value: 0.31025, per_snippet_value: 0.31025 }
error:           ''
info:            ''
assembled_snippet: C5FDEFC0C5FDEFC0C5FDEFC0C5FDEFC0C5FDEFC0C5FDEFC0C5FDEFC0C5FDEFC0C5FDEFC0C5FDEFC0C5FDEFC0C5FDEFC0C5FDEFC0C5FDEFC0C5FDEFC0C5FDEFC0C3
...

```
What does it tell us?
So wait, it can only execute ~3 x86 AVX YMM PXOR zero-idioms per cycle?
That doesn't seem right. That's even less than there are pipes supporting this type of op.

Now, second example:
```
$ ./bin/llvm-exegesis --mode=inverse_throughput --snippets-file=/tmp/snippet.s --num-repetitions=1000000 --repetition-mode=loop
Check generated assembly with: /usr/bin/objdump -d /tmp/snippet-2418b5.o
---
mode:            inverse_throughput
key:
  instructions:
    - 'VPXORYrr YMM0 YMM0 YMM0'
  config:          ''
  register_initial_values: []
cpu_name:        znver3
llvm_triple:     x86_64-unknown-linux-gnu
num_repetitions: 1000000
measurements:
  - { key: inverse_throughput, value: 1.00011, per_snippet_value: 1.00011 }
error:           ''
info:            ''
assembled_snippet: 49B80800000000000000C5FDEFC0C5FDEFC04983C0FF75F2C3
...
```
Now that's just worse. Due to the looping, the throughput completely plummeted,
and now we can only do a single instruction/cycle!?

That's not great.
And final example:
```
$ ./bin/llvm-exegesis --mode=inverse_throughput --snippets-file=/tmp/snippet.s --num-repetitions=1000000 --repetition-mode=loop --loop-body-size=1000
Check generated assembly with: /usr/bin/objdump -d /tmp/snippet-c402e2.o
---
mode:            inverse_throughput
key:
  instructions:
    - 'VPXORYrr YMM0 YMM0 YMM0'
  config:          ''
  register_initial_values: []
cpu_name:        znver3
llvm_triple:     x86_64-unknown-linux-gnu
num_repetitions: 1000000
measurements:
  - { key: inverse_throughput, value: 0.167087, per_snippet_value: 0.167087 }
error:           ''
info:            ''
assembled_snippet: 49B80800000000000000C5FDEFC0C5FDEFC04983C0FF75F2C3
...
```

So if we merge the previous two approaches, do duplicate this single-instruction snippet 1000x
(loop-body-size/instruction count in snippet), and run a loop with 1000 iterations
over that duplicated/unrolled snippet, the measured throughput goes through the roof,
up to 5.9 instructions/cycle, which finally tells us that this idiom is zero-cycle!

Reviewed By: courbet

Differential Revision: https://reviews.llvm.org/D102522
2021-05-25 12:08:27 +03:00

312 lines
12 KiB
ReStructuredText

llvm-exegesis - LLVM Machine Instruction Benchmark
==================================================
.. program:: llvm-exegesis
SYNOPSIS
--------
:program:`llvm-exegesis` [*options*]
DESCRIPTION
-----------
:program:`llvm-exegesis` is a benchmarking tool that uses information available
in LLVM to measure host machine instruction characteristics like latency,
throughput, or port decomposition.
Given an LLVM opcode name and a benchmarking mode, :program:`llvm-exegesis`
generates a code snippet that makes execution as serial (resp. as parallel) as
possible so that we can measure the latency (resp. inverse throughput/uop decomposition)
of the instruction.
The code snippet is jitted and executed on the host subtarget. The time taken
(resp. resource usage) is measured using hardware performance counters. The
result is printed out as YAML to the standard output.
The main goal of this tool is to automatically (in)validate the LLVM's TableDef
scheduling models. To that end, we also provide analysis of the results.
:program:`llvm-exegesis` can also benchmark arbitrary user-provided code
snippets.
EXAMPLE 1: benchmarking instructions
------------------------------------
Assume you have an X86-64 machine. To measure the latency of a single
instruction, run:
.. code-block:: bash
$ llvm-exegesis -mode=latency -opcode-name=ADD64rr
Measuring the uop decomposition or inverse throughput of an instruction works similarly:
.. code-block:: bash
$ llvm-exegesis -mode=uops -opcode-name=ADD64rr
$ llvm-exegesis -mode=inverse_throughput -opcode-name=ADD64rr
The output is a YAML document (the default is to write to stdout, but you can
redirect the output to a file using `-benchmarks-file`):
.. code-block:: none
---
key:
opcode_name: ADD64rr
mode: latency
config: ''
cpu_name: haswell
llvm_triple: x86_64-unknown-linux-gnu
num_repetitions: 10000
measurements:
- { key: latency, value: 1.0058, debug_string: '' }
error: ''
info: 'explicit self cycles, selecting one aliasing configuration.
Snippet:
ADD64rr R8, R8, R10
'
...
To measure the latency of all instructions for the host architecture, run:
.. code-block:: bash
$ llvm-exegesis -mode=latency -opcode-index=-1
EXAMPLE 2: benchmarking a custom code snippet
---------------------------------------------
To measure the latency/uops of a custom piece of code, you can specify the
`snippets-file` option (`-` reads from standard input).
.. code-block:: bash
$ echo "vzeroupper" | llvm-exegesis -mode=uops -snippets-file=-
Real-life code snippets typically depend on registers or memory.
:program:`llvm-exegesis` checks the liveliness of registers (i.e. any register
use has a corresponding def or is a "live in"). If your code depends on the
value of some registers, you have two options:
- Mark the register as requiring a definition. :program:`llvm-exegesis` will
automatically assign a value to the register. This can be done using the
directive `LLVM-EXEGESIS-DEFREG <reg name> <hex_value>`, where `<hex_value>`
is a bit pattern used to fill `<reg_name>`. If `<hex_value>` is smaller than
the register width, it will be sign-extended.
- Mark the register as a "live in". :program:`llvm-exegesis` will benchmark
using whatever value was in this registers on entry. This can be done using
the directive `LLVM-EXEGESIS-LIVEIN <reg name>`.
For example, the following code snippet depends on the values of XMM1 (which
will be set by the tool) and the memory buffer passed in RDI (live in).
.. code-block:: none
# LLVM-EXEGESIS-LIVEIN RDI
# LLVM-EXEGESIS-DEFREG XMM1 42
vmulps (%rdi), %xmm1, %xmm2
vhaddps %xmm2, %xmm2, %xmm3
addq $0x10, %rdi
EXAMPLE 3: analysis
-------------------
Assuming you have a set of benchmarked instructions (either latency or uops) as
YAML in file `/tmp/benchmarks.yaml`, you can analyze the results using the
following command:
.. code-block:: bash
$ llvm-exegesis -mode=analysis \
-benchmarks-file=/tmp/benchmarks.yaml \
-analysis-clusters-output-file=/tmp/clusters.csv \
-analysis-inconsistencies-output-file=/tmp/inconsistencies.html
This will group the instructions into clusters with the same performance
characteristics. The clusters will be written out to `/tmp/clusters.csv` in the
following format:
.. code-block:: none
cluster_id,opcode_name,config,sched_class
...
2,ADD32ri8_DB,,WriteALU,1.00
2,ADD32ri_DB,,WriteALU,1.01
2,ADD32rr,,WriteALU,1.01
2,ADD32rr_DB,,WriteALU,1.00
2,ADD32rr_REV,,WriteALU,1.00
2,ADD64i32,,WriteALU,1.01
2,ADD64ri32,,WriteALU,1.01
2,MOVSX64rr32,,BSWAP32r_BSWAP64r_MOVSX64rr32,1.00
2,VPADDQYrr,,VPADDBYrr_VPADDDYrr_VPADDQYrr_VPADDWYrr_VPSUBBYrr_VPSUBDYrr_VPSUBQYrr_VPSUBWYrr,1.02
2,VPSUBQYrr,,VPADDBYrr_VPADDDYrr_VPADDQYrr_VPADDWYrr_VPSUBBYrr_VPSUBDYrr_VPSUBQYrr_VPSUBWYrr,1.01
2,ADD64ri8,,WriteALU,1.00
2,SETBr,,WriteSETCC,1.01
...
:program:`llvm-exegesis` will also analyze the clusters to point out
inconsistencies in the scheduling information. The output is an html file. For
example, `/tmp/inconsistencies.html` will contain messages like the following :
.. image:: llvm-exegesis-analysis.png
:align: center
Note that the scheduling class names will be resolved only when
:program:`llvm-exegesis` is compiled in debug mode, else only the class id will
be shown. This does not invalidate any of the analysis results though.
OPTIONS
-------
.. option:: -help
Print a summary of command line options.
.. option:: -opcode-index=<LLVM opcode index>
Specify the opcode to measure, by index. Specifying `-1` will result
in measuring every existing opcode. See example 1 for details.
Either `opcode-index`, `opcode-name` or `snippets-file` must be set.
.. option:: -opcode-name=<opcode name 1>,<opcode name 2>,...
Specify the opcode to measure, by name. Several opcodes can be specified as
a comma-separated list. See example 1 for details.
Either `opcode-index`, `opcode-name` or `snippets-file` must be set.
.. option:: -snippets-file=<filename>
Specify the custom code snippet to measure. See example 2 for details.
Either `opcode-index`, `opcode-name` or `snippets-file` must be set.
.. option:: -mode=[latency|uops|inverse_throughput|analysis]
Specify the run mode. Note that some modes have additional requirements and options.
`latency` mode can be make use of either RDTSC or LBR.
`latency[LBR]` is only available on X86 (at least `Skylake`).
To run in `latency` mode, a positive value must be specified
for `x86-lbr-sample-period` and `--repetition-mode=loop`.
In `analysis` mode, you also need to specify at least one of the
`-analysis-clusters-output-file=` and `-analysis-inconsistencies-output-file=`.
.. option:: -x86-lbr-sample-period=<nBranches/sample>
Specify the LBR sampling period - how many branches before we take a sample.
When a positive value is specified for this option and when the mode is `latency`,
we will use LBRs for measuring.
On choosing the "right" sampling period, a small value is preferred, but throttling
could occur if the sampling is too frequent. A prime number should be used to
avoid consistently skipping certain blocks.
.. option:: -repetition-mode=[duplicate|loop|min]
Specify the repetition mode. `duplicate` will create a large, straight line
basic block with `num-repetitions` instructions (repeating the snippet
`num-repetitions`/`snippet size` times). `loop` will, optionally, duplicate the
snippet until the loop body contains at least `loop-body-size` instructions,
and then wrap the result in a loop which will execute `num-repetitions`
instructions (thus, again, repeating the snippet
`num-repetitions`/`snippet size` times). The `loop` mode, especially with loop
unrolling tends to better hide the effects of the CPU frontend on architectures
that cache decoded instructions, but consumes a register for counting
iterations. If performing an analysis over many opcodes, it may be best to
instead use the `min` mode, which will run each other mode,
and produce the minimal measured result.
.. option:: -num-repetitions=<Number of repetitions>
Specify the target number of executed instructions. Note that the actual
repetition count of the snippet will be `num-repetitions`/`snippet size`.
Higher values lead to more accurate measurements but lengthen the benchmark.
.. option:: -loop-body-size=<Preferred loop body size>
Only effective for `-repetition-mode=[loop|min]`.
Instead of looping over the snippet directly, first duplicate it so that the
loop body contains at least this many instructions. This potentially results
in loop body being cached in the CPU Op Cache / Loop Cache, which allows to
which may have higher throughput than the CPU decoders.
.. option:: -max-configs-per-opcode=<value>
Specify the maximum configurations that can be generated for each opcode.
By default this is `1`, meaning that we assume that a single measurement is
enough to characterize an opcode. This might not be true of all instructions:
for example, the performance characteristics of the LEA instruction on X86
depends on the value of assigned registers and immediates. Setting a value of
`-max-configs-per-opcode` larger than `1` allows `llvm-exegesis` to explore
more configurations to discover if some register or immediate assignments
lead to different performance characteristics.
.. option:: -benchmarks-file=</path/to/file>
File to read (`analysis` mode) or write (`latency`/`uops`/`inverse_throughput`
modes) benchmark results. "-" uses stdin/stdout.
.. option:: -analysis-clusters-output-file=</path/to/file>
If provided, write the analysis clusters as CSV to this file. "-" prints to
stdout. By default, this analysis is not run.
.. option:: -analysis-inconsistencies-output-file=</path/to/file>
If non-empty, write inconsistencies found during analysis to this file. `-`
prints to stdout. By default, this analysis is not run.
.. option:: -analysis-clustering=[dbscan,naive]
Specify the clustering algorithm to use. By default DBSCAN will be used.
Naive clustering algorithm is better for doing further work on the
`-analysis-inconsistencies-output-file=` output, it will create one cluster
per opcode, and check that the cluster is stable (all points are neighbours).
.. option:: -analysis-numpoints=<dbscan numPoints parameter>
Specify the numPoints parameters to be used for DBSCAN clustering
(`analysis` mode, DBSCAN only).
.. option:: -analysis-clustering-epsilon=<dbscan epsilon parameter>
Specify the epsilon parameter used for clustering of benchmark points
(`analysis` mode).
.. option:: -analysis-inconsistency-epsilon=<epsilon>
Specify the epsilon parameter used for detection of when the cluster
is different from the LLVM schedule profile values (`analysis` mode).
.. option:: -analysis-display-unstable-clusters
If there is more than one benchmark for an opcode, said benchmarks may end up
not being clustered into the same cluster if the measured performance
characteristics are different. by default all such opcodes are filtered out.
This flag will instead show only such unstable opcodes.
.. option:: -ignore-invalid-sched-class=false
If set, ignore instructions that do not have a sched class (class idx = 0).
.. option:: -mcpu=<cpu name>
If set, measure the cpu characteristics using the counters for this CPU. This
is useful when creating new sched models (the host CPU is unknown to LLVM).
.. option:: --dump-object-to-disk=true
By default, llvm-exegesis will dump the generated code to a temporary file to
enable code inspection. You may disable it to speed up the execution and save
disk space.
EXIT STATUS
-----------
:program:`llvm-exegesis` returns 0 on success. Otherwise, an error message is
printed to standard error, and the tool returns a non 0 value.