For an innermost loop kernel in assembly, this tool allows automatic instruction fetching of assembly code and automatic runtime prediction including throughput analysis and detection for critical path and loop-carried dependencies.
OSACA is as a python module with a command line interface.
OSACA is also integrated into the Compiler Explorer at godbolt.org, which allows using OSACA from a browser without any installation. To analyze an assembly snippet, go to https://godbolt.org change language to "Analysis", insert an AArch64 or AT&T(!) x86 assembly code and make sure OSACA is selected in the corresponding analysis panel, e.g., https://godbolt.org/z/shK4f8. When analyzing a high-level language code, use the "Add tool..." menu in the compiler output panel to add OSACA analysis, e.g. https://godbolt.org/z/hbMoPn. To change the micro architecture model, add --arch
and µarch shortname (e.g., SKX
for Skylake, ZEN2
, N1
for ARM Neoverse) to the "Compiler options..." (when using "Analysis" mode) or "Arguments" (when analyzing compiler output of a high-level code).
On most systems with python pip and setuputils installed, just run:
pip install --user osaca
for the latest release.
To build OSACA from source, clone this repository using git clone https://github.com/RRZE-HPC/OSACA
and run in the root directory:
python ./setup.py install
After installation, OSACA can be started with the command osaca
in the CLI.
Necessary equirements are:
- Python3
- Graphviz for dependency graph creation (minimal dependency is
libgraphviz-dev
on Ubuntu) - Python packages:
Optional requirements are:
- Kerncraft >=v0.8.4 for marker insertion
- ibench or asmbench for throughput/latency measurements
- BeautifulSoup4 for scraping instruction form information for the x86 ISA (experimental)
A schematic design of OSACA's workflow is shown below:
The usage of OSACA can be listed as:
osaca [-h] [-V] [--arch ARCH] [--fixed] [--lines LINES]
[--ignore-unknown] [--lcd-timeout SECONDS]
[--db-check] [--import MICROBENCH] [--insert-marker]
[--export-graph GRAPHNAME] [--consider-flag-deps]
[--out OUT] [--yaml-out YAML_OUT] [--verbose]
FILEPATH
-h, --help | prints out the help message. |
-V, --version | shows the program’s version number. |
--arch ARCH | needs to be replaced with the target architecture abbreviation.
Possible options are SNB , IVB , HSW , BDW , SKX , CSX , ICL (Client), ICX (Server), SPR for the latest Intel micro architectures starting from Intel Sandy Bridge and ZEN[1-4] for AMD Zen architectures.
Furthermore, TX2 for Marvell`s ARM-based ThunderX2 , N1 for ARM's Neoverse, A72 for ARM Cortex-A72, TSV110 for the HiSilicon TaiShan v110, A64FX for Fujitsu's HPC ARM architecture, M1 for the Apple M1-Firestorm performance core, and V2 for the Neoverse V2 (used in NVIDIA's Grace CPU) are available.
If no micro-architecture is given, OSACA assumes a default architecture for x86/AArch64. |
--fixed | Run the throughput analysis with fixed port utilization for all suitable ports per instruction. Otherwise, OSACA will print out the optimal port utilization for the kernel. |
--lines | Define lines that should be included in the analysis. This option overwrites any range defined by markers in the assembly. Add either single lines or ranges defined
by "-" or ":", each entry separated by commas, e.g.: --lines 1,2,8-18,20:24 |
--db-check | Run a sanity check on the by "--arch" specified database. The output depends on the verbosity level. Keep in mind you have to provide an existing (dummy) filename in anyway. |
--import MICROBENCH | |
Import a given microbenchmark output file into the corresponding architecture instruction database. Define the type of microbenchmark either as "ibench" or "asmbench". | |
--insert-marker | |
OSACA calls the Kerncraft module for the interactively insertion of IACA byte markers or OSACA AArch64 byte markers in suggested assembly blocks. | |
--export-graph EXPORT_PATH | |
Output path for .dot file export. If "." is given, the file will be stored as "./osaca_dg.dot". After the file was created, you can convert it to a PDF file using dot. | |
--ignore-unknown | |
Force OSACA to apply a throughput and latency of 0.0 cy for all unknown instruction forms. If not specified, a warning will be printed instead if one ore more isntruction form is unknown to OSACA. | |
--lcd-timeout SECONDS | |
Set timeout in seconds for LCD analysis. After timeout, OSACA will continue its analysis with the dependency paths found up to this point. Defaults to 10. | |
-f, --consider-flag-deps | |
Consider flag dependencies for the critical path and loop-carried dependency analysis. By default, those dependencies are ignored. | |
-v, --verbose | Increases verbosity level |
-o OUT, --out OUT | |
Write analysis to this file (default to stdout) | |
--yaml-out YAML_OUT | |
Write analysis as YAML representation to this file |
The FILEPATH describes the filepath to the file to work with and is always necessary, use "-" to read from stdin.
x86 CPUs
Designer | Model/microarch | OSACA flag |
---|---|---|
Intel | Sandy Bridge | SNB |
Intel | Ivy Bridge | IVB |
Intel | Haswell | HSW |
Intel | Broadwell | BDW |
Intel | Skylake-X | SKX |
Intel | Cascadelake-X | CSX |
Intel | Icelake client | ICL |
Intel | Icelake server | ICX |
Intel | Sapphire Rapids | SPR |
AMD | Naples / Zen 1 | ZEN1 |
AMD | Rome / Zen 2 | ZEN2 |
AMD | Milan / Zen 3 | ZEN3 |
AMD | Genoa / Zen 4 | ZEN4 |
ARM AArch64 CPUs
Designer | Model/microarch | OSACA flag |
---|---|---|
ARM | Cortex-A72 | A72 |
ARM | Neoverse N1 | N1 |
ARM | Neoverse V2 | V2 |
Marvell | ThunderX2 | TX2 |
Fujitsu | FX700/A64FX | A64FX |
HiSilicon | TaiShan v110 | TSV110 |
Apple | M1-Firestorm | M1 |
NVIDIA | Neoverse V2/Grace | V2 |
Hereinafter OSACA's scope of function will be described.
As main functionality of OSACA, the tool starts the analysis on a marked assembly file by running the following command with one or more of the optional parameters:
osaca --arch ARCH [--fixed] [--ignore-unknown]
[--export-graph EXPORT_PATH]
file
The file
parameter specifies the target assembly file and is always mandatory.
The parameter ARCH
is positional for the analysis and must be replaced by the target architecture abbreviation.
OSACA assumes an optimal scheduling for all instructions and assumes the processor to be able to schedule instructions in a way that it achieves a minimal reciprocal throughput.
However, in older versions (<=v0.2.2) of OSACA, a fixed probability for port utilization was assumed.
This means, instructions with N available ports for execution were scheduled with a probability of 1/N to each of the ports.
This behavior can be enforced by using the --fixed
flag.
If one or more instruction forms are unknown to OSACA, it refuses to print an overall throughput, CP and
LCD analysis and marks all unknown instruction forms with X
next to the mnemonic.
This is done so the user does not miss out on this unrecognized instruction and might assume an incorrect runtime prediction.
To force OSACA to apply a throughput and latency of 0.0 cy for all unknown instruction forms, the flag --ignore-unknown
can be specified.
To get a visualization of the analyzed kernel and its dependency chains, OSACA provides the option to additionally produce a graph as DOT file, which represents the kernel and all register dependencies inside of it.
The tool highlights all LCDs and the CP.
The graph generation is done by running OSACA with the --export-graph EXPORT_GRAPH
flag.
OSACA stores the DOT file either at the by EXPORT_GRAPH
specified filepath or uses the default filename "osaca_dg.dot" in the current working directory.
Subsequently, the DOT-graph can be adjusted in its appearance and converted to various output formats such as PDF, SVG, or PNG using the dot command, e.g., dot -Tpdf osaca_dg.dot -o
graph.pdf
to generate a PDF document.
For extracting the right kernel, one can mark it in beforehand. Currently, only the detection of markers in the assembly code and therefore the analysis of assembly files is supported by OSACA. If OSACA cannot find any markers in the given input file, all lines will be evaluated.
Marking a kernel means to insert the byte markers in the assembly file in before and after the loop.
For this, the start marker has to be inserted right in front of the loop label and the end marker directly after the jump instruction.
IACA requires byte markers since it operates on opcode-level.
To provide a trade-off between reusability for such tool and convenient usability, OSACA supports both byte markers and comment line markers.
While the byte markers for x86 are equivalent to IACA byte markers, the comment keywords OSACA-BEGIN
and OSACA-END
are based on LLVM-MCA's markers.
Byte markers
movl $111,%ebx #IACA/OSACA START MARKER
.byte 100,103,144 #IACA/OSACA START MARKER
.loop:
# loop body
jb .loop
movl $222,%ebx #IACA/OSACA END MARKER
.byte 100,103,144 #IACA/OSACA END MARKER
Comment line markers
# OSACA-BEGIN
.loop:
# loop body
jb .loop
# OSACA-END
Byte markers
mov x1, #111 // OSACA START .byte 213,3,32,31 // OSACA START .loop: // loop body b.ne .loop mov x1, #222 // OSACA END .byte 213,3,32,31 // OSACA END
Comment line markers
// OSACA-BEGIN .loop: // loop body b.ne .loop // OSACA-END
OSACA in combination with Kerncraft provides a functionality for the automatic detection of possible loop kernels and inserting markers.
This can be done by using the --insert-marker
flag together with the path to the target assembly file and the target architecture.
OSACA supports the automatic integration of new instruction forms by parsing the output of the micro-
benchmark tools asmbench and ibench.
This can be achieved by running OSACA with the command line option --import MICROBENCH
:
osaca --arch ARCH --import MICROBENCH file
MICROBENCH
specifies one of the currently supported benchmark tools, i.e., "asmbench" or "ibench".
ARCH
defines the abbreviation of the target architecture for which the instructions will be added and file must be the path to the generated output file of the benchmark.
The format of this file has to match either the basic command line output of ibench, e.g.,
[INSTRUCTION FORM]-TP: 0.500 (clock cycles) [DEBUG - result: 1.000000] [INSTRUCTION FORM]-LT: 4.000 (clock cycles) [DEBUG - result: 1.000000]
or the command line output of asmbench including the name of the instruction form in a separate line at the beginning, e.g.:
[INSTRUCTION FORM] Latency: 4.00 cycle Throughput: 0.50 cycle
Note that there must be an empty line after each throughput measurement as part of the output so that one instruction form entry consists of four (4) lines.
To let OSACA import the instruction form with the correct operands, the naming conventions for the instruction form name must be followed:
- The first part of the name is the mnemonic and ends with the character "
-
" (not part of the mnemonic in the DB). - The second part of the name are the operands. Each operand must be separated from another operand by the character "
_
". - For each x86 operand, one of the following symbols must be used:
- "
r
" for general purpose registers (rax, edi, r9, ...) - "
x
", "y
", or "z
" for xmm, ymm, or zmm registers, respectively - "
i
" for immediates - "
m
" for a memory address. Add "b
" if the memory address contains a base register, "o
" if it contains an offset, "i
" if it contains an index register, and "s
" if the index register additionally has a scale factor of more than 1.
- "
- For each AArch64 operand, one of the following symbols must be used:
- "
w
", "x
", "b
", "h
", "s
", "d
", or "q
" for registers with the corresponding prefix. - "
v
" followed by a single character ("b
", "h
", "s
", or "d
") for vector registers with the corresponding lane width of the second character. If no second character is given, OSACA assumes a lane width of 64 bit (d
) as default. - "
i
" for immediates - "
m
" for a memory address. Add "b
" if the memory address contains a base register, "o
" if it contains an offset, "i
" if it contains an index register, and "s
" if the index register additionally has a scale factor of more than 1. Add "r
" if the address format uses pre-indexing and "p
" if it uses post-indexing.
- "
Valid instruction form examples for x86 are vaddpd-x_x_x
, mov-r_mboi
, and vfmadd213pd-mbis_y_y
.
Valid instruction form examples for AArch64 are fadd-vd_vd_v
, ldp-d_d_mo
, and fmov-s_i
.
Note that the options to define operands are limited, therefore, one might need to adjust the instruction forms in the architecture DB after importing. OSACA parses the output for an arbitrary number of instruction forms and adds them as entries to the architecture DB. The user must edit the ISA DB in case the instruction form shows irregular source and destination operands for its ISA syntax. OSACA applies the following rules by default:
- If there is only one operand, it is considered as source operand
- In case of multiple operands the target operand (depending on the ISA syntax the last or first one) is considered to be the destination operand, all others are considered as source operands.
Since a manual adjustment of the ISA DB is currently indispensable when adding new instruction forms, OSACA provides a database sanity check using the --db-check flag. It can be executed via:
osaca --arch ARCH --db-check [-v] file
ARCH
defines the abbreviation of the target architecture of the database to check.
The file
argument needs to be specified as it is positional but may be any existing dummy path.
When called, OSACA prints a summary of database information containing the amount of missing throughput values, latency values or μ-ops assignments for an instruction form.
Furthermore, it shows the amount of duplicate instruction forms in both the architecture DB and the ISA DB and checks how many instruction forms in the ISA DB are non-existent in the architecture DB.
Finally, it checks via simple heuristics how many of the instruction forms contained in the architecture DB might miss an ISA DB entry.
Running the database check including the -v
verbosity flag, OSACA prints in addition the specific name of the identified instruction forms so that the user can check the mentioned incidents.
For clarifying the functionality of OSACA a sample kernel is analyzed for an Intel CSX core hereafter:
double a[N], double b[N];
double s;
// loop
for(int i = 0; i < N; ++i)
a[i] = s * b[i];
The code shows a simple scalar multiplication of a vector b
and a floating-point number s
.
The result is written in vector a
.
After including the OSACA byte marker into the assembly, one can start the analysis typing
osaca --arch CSX PATH/TO/FILE
in the command line.
The output is:
Open Source Architecture Code Analyzer (OSACA) - v0.3 Analyzed file: scale.s.csx.O3.s Architecture: csx Timestamp: 2019-10-03 23:36:21 P - Throughput of LOAD operation can be hidden behind a past or future STORE instruction * - Instruction micro-ops not bound to a port X - No throughput/latency information for this instruction in data file Combined Analysis Report ----------------------- Port pressure in cycles | 0 - 0DV | 1 | 2 - 2D | 3 - 3D | 4 | 5 | 6 | 7 || CP | LCD | ------------------------------------------------------------------------------------------------- 170 | | | | | | | | || | | .L22: 171 | 0.50 | 0.50 | 0.50 0.50 | 0.50 0.50 | | | | || 8.0 | | vmulpd (%r12,%rax), %ymm1, %ymm0 172 | | | 0.50 | 0.50 | 1.00 | | | || 5.0 | | vmovapd %ymm0, 0(%r13,%rax) 173 | 0.25 | 0.25 | | | | 0.25 | 0.25 | || | 1.0 | addq $32, %rax 174 | 0.00 | 0.00 | | | | 0.50 | 0.50 | || | | cmpq %rax, %r14 175 | | | | | | | | || | | * jne .L22 0.75 0.75 1.00 0.50 1.00 0.50 1.00 0.75 0.75 13.0 1.0 Loop-Carried Dependencies Analysis Report ----------------------------------------- 173 | 1.0 | addq $32, %rax | [173]
It shows the whole kernel together with the optimized port pressure of each instruction form and the overall port binding. Furthermore, in the two columns on the right, the critical path (CP) and the longest loop-carried dependency (LCD) of the loop kernel. In the bottom, all loop-carried dependencies are shown, each with a list of line numbers being part of this dependency chain on the right.
You can find more (already marked) examples and sample outputs for various architectures in the examples directory.
If you use OSACA for scientific work you can cite us as (for the Bibtex, see the Wiki):
- Automated Instruction Stream Throughput Prediction for Intel and AMD Microarchitectures (Pre-print PMBS18)
- Automatic Throughput and Critical Path Analysis of x86 and ARM Assembly Kernels (Pre-print PMBS19)
Implementation: Jan Laukemann, Julian Hammer