1 llvm-exegesis - LLVM Machine Instruction Benchmark
2 ==================================================
4 .. program:: llvm-exegesis
9 :program:`llvm-exegesis` [*options*]
14 :program:`llvm-exegesis` is a benchmarking tool that uses information available
15 in LLVM to measure host machine instruction characteristics like latency,
16 throughput, or port decomposition.
18 Given an LLVM opcode name and a benchmarking mode, :program:`llvm-exegesis`
19 generates a code snippet that makes execution as serial (resp. as parallel) as
20 possible so that we can measure the latency (resp. inverse throughput/uop decomposition)
22 The code snippet is jitted and executed on the host subtarget. The time taken
23 (resp. resource usage) is measured using hardware performance counters. The
24 result is printed out as YAML to the standard output.
26 The main goal of this tool is to automatically (in)validate the LLVM's TableDef
27 scheduling models. To that end, we also provide analysis of the results.
29 :program:`llvm-exegesis` can also benchmark arbitrary user-provided code
32 EXAMPLE 1: benchmarking instructions
33 ------------------------------------
35 Assume you have an X86-64 machine. To measure the latency of a single
40 $ llvm-exegesis -mode=latency -opcode-name=ADD64rr
42 Measuring the uop decomposition or inverse throughput of an instruction works similarly:
46 $ llvm-exegesis -mode=uops -opcode-name=ADD64rr
47 $ llvm-exegesis -mode=inverse_throughput -opcode-name=ADD64rr
50 The output is a YAML document (the default is to write to stdout, but you can
51 redirect the output to a file using `-benchmarks-file`):
61 llvm_triple: x86_64-unknown-linux-gnu
62 num_repetitions: 10000
64 - { key: latency, value: 1.0058, debug_string: '' }
66 info: 'explicit self cycles, selecting one aliasing configuration.
72 To measure the latency of all instructions for the host architecture, run:
76 $ llvm-exegesis -mode=latency -opcode-index=-1
79 EXAMPLE 2: benchmarking a custom code snippet
80 ---------------------------------------------
82 To measure the latency/uops of a custom piece of code, you can specify the
83 `snippets-file` option (`-` reads from standard input).
87 $ echo "vzeroupper" | llvm-exegesis -mode=uops -snippets-file=-
89 Real-life code snippets typically depend on registers or memory.
90 :program:`llvm-exegesis` checks the liveliness of registers (i.e. any register
91 use has a corresponding def or is a "live in"). If your code depends on the
92 value of some registers, you have two options:
94 - Mark the register as requiring a definition. :program:`llvm-exegesis` will
95 automatically assign a value to the register. This can be done using the
96 directive `LLVM-EXEGESIS-DEFREG <reg name> <hex_value>`, where `<hex_value>`
97 is a bit pattern used to fill `<reg_name>`. If `<hex_value>` is smaller than
98 the register width, it will be sign-extended.
99 - Mark the register as a "live in". :program:`llvm-exegesis` will benchmark
100 using whatever value was in this registers on entry. This can be done using
101 the directive `LLVM-EXEGESIS-LIVEIN <reg name>`.
103 For example, the following code snippet depends on the values of XMM1 (which
104 will be set by the tool) and the memory buffer passed in RDI (live in).
108 # LLVM-EXEGESIS-LIVEIN RDI
109 # LLVM-EXEGESIS-DEFREG XMM1 42
110 vmulps (%rdi), %xmm1, %xmm2
111 vhaddps %xmm2, %xmm2, %xmm3
118 Assuming you have a set of benchmarked instructions (either latency or uops) as
119 YAML in file `/tmp/benchmarks.yaml`, you can analyze the results using the
124 $ llvm-exegesis -mode=analysis \
125 -benchmarks-file=/tmp/benchmarks.yaml \
126 -analysis-clusters-output-file=/tmp/clusters.csv \
127 -analysis-inconsistencies-output-file=/tmp/inconsistencies.html
129 This will group the instructions into clusters with the same performance
130 characteristics. The clusters will be written out to `/tmp/clusters.csv` in the
135 cluster_id,opcode_name,config,sched_class
137 2,ADD32ri8_DB,,WriteALU,1.00
138 2,ADD32ri_DB,,WriteALU,1.01
139 2,ADD32rr,,WriteALU,1.01
140 2,ADD32rr_DB,,WriteALU,1.00
141 2,ADD32rr_REV,,WriteALU,1.00
142 2,ADD64i32,,WriteALU,1.01
143 2,ADD64ri32,,WriteALU,1.01
144 2,MOVSX64rr32,,BSWAP32r_BSWAP64r_MOVSX64rr32,1.00
145 2,VPADDQYrr,,VPADDBYrr_VPADDDYrr_VPADDQYrr_VPADDWYrr_VPSUBBYrr_VPSUBDYrr_VPSUBQYrr_VPSUBWYrr,1.02
146 2,VPSUBQYrr,,VPADDBYrr_VPADDDYrr_VPADDQYrr_VPADDWYrr_VPSUBBYrr_VPSUBDYrr_VPSUBQYrr_VPSUBWYrr,1.01
147 2,ADD64ri8,,WriteALU,1.00
148 2,SETBr,,WriteSETCC,1.01
151 :program:`llvm-exegesis` will also analyze the clusters to point out
152 inconsistencies in the scheduling information. The output is an html file. For
153 example, `/tmp/inconsistencies.html` will contain messages like the following :
155 .. image:: llvm-exegesis-analysis.png
158 Note that the scheduling class names will be resolved only when
159 :program:`llvm-exegesis` is compiled in debug mode, else only the class id will
160 be shown. This does not invalidate any of the analysis results though.
167 Print a summary of command line options.
169 .. option:: -opcode-index=<LLVM opcode index>
171 Specify the opcode to measure, by index. Specifying `-1` will result
172 in measuring every existing opcode. See example 1 for details.
173 Either `opcode-index`, `opcode-name` or `snippets-file` must be set.
175 .. option:: -opcode-name=<opcode name 1>,<opcode name 2>,...
177 Specify the opcode to measure, by name. Several opcodes can be specified as
178 a comma-separated list. See example 1 for details.
179 Either `opcode-index`, `opcode-name` or `snippets-file` must be set.
181 .. option:: -snippets-file=<filename>
183 Specify the custom code snippet to measure. See example 2 for details.
184 Either `opcode-index`, `opcode-name` or `snippets-file` must be set.
186 .. option:: -mode=[latency|uops|inverse_throughput|analysis]
188 Specify the run mode. Note that some modes have additional requirements and options.
190 `latency` mode can be make use of either RDTSC or LBR.
191 `latency[LBR]` is only available on X86 (at least `Skylake`).
192 To run in `latency` mode, a positive value must be specified
193 for `x86-lbr-sample-period` and `--repetition-mode=loop`.
195 In `analysis` mode, you also need to specify at least one of the
196 `-analysis-clusters-output-file=` and `-analysis-inconsistencies-output-file=`.
198 .. option:: -x86-lbr-sample-period=<nBranches/sample>
200 Specify the LBR sampling period - how many branches before we take a sample.
201 When a positive value is specified for this option and when the mode is `latency`,
202 we will use LBRs for measuring.
203 On choosing the "right" sampling period, a small value is preferred, but throttling
204 could occur if the sampling is too frequent. A prime number should be used to
205 avoid consistently skipping certain blocks.
207 .. option:: -repetition-mode=[duplicate|loop|min]
209 Specify the repetition mode. `duplicate` will create a large, straight line
210 basic block with `num-repetitions` instructions (repeating the snippet
211 `num-repetitions`/`snippet size` times). `loop` will, optionally, duplicate the
212 snippet until the loop body contains at least `loop-body-size` instructions,
213 and then wrap the result in a loop which will execute `num-repetitions`
214 instructions (thus, again, repeating the snippet
215 `num-repetitions`/`snippet size` times). The `loop` mode, especially with loop
216 unrolling tends to better hide the effects of the CPU frontend on architectures
217 that cache decoded instructions, but consumes a register for counting
218 iterations. If performing an analysis over many opcodes, it may be best to
219 instead use the `min` mode, which will run each other mode,
220 and produce the minimal measured result.
222 .. option:: -num-repetitions=<Number of repetitions>
224 Specify the target number of executed instructions. Note that the actual
225 repetition count of the snippet will be `num-repetitions`/`snippet size`.
226 Higher values lead to more accurate measurements but lengthen the benchmark.
228 .. option:: -loop-body-size=<Preferred loop body size>
230 Only effective for `-repetition-mode=[loop|min]`.
231 Instead of looping over the snippet directly, first duplicate it so that the
232 loop body contains at least this many instructions. This potentially results
233 in loop body being cached in the CPU Op Cache / Loop Cache, which allows to
234 which may have higher throughput than the CPU decoders.
236 .. option:: -max-configs-per-opcode=<value>
238 Specify the maximum configurations that can be generated for each opcode.
239 By default this is `1`, meaning that we assume that a single measurement is
240 enough to characterize an opcode. This might not be true of all instructions:
241 for example, the performance characteristics of the LEA instruction on X86
242 depends on the value of assigned registers and immediates. Setting a value of
243 `-max-configs-per-opcode` larger than `1` allows `llvm-exegesis` to explore
244 more configurations to discover if some register or immediate assignments
245 lead to different performance characteristics.
248 .. option:: -benchmarks-file=</path/to/file>
250 File to read (`analysis` mode) or write (`latency`/`uops`/`inverse_throughput`
251 modes) benchmark results. "-" uses stdin/stdout.
253 .. option:: -analysis-clusters-output-file=</path/to/file>
255 If provided, write the analysis clusters as CSV to this file. "-" prints to
256 stdout. By default, this analysis is not run.
258 .. option:: -analysis-inconsistencies-output-file=</path/to/file>
260 If non-empty, write inconsistencies found during analysis to this file. `-`
261 prints to stdout. By default, this analysis is not run.
263 .. option:: -analysis-clustering=[dbscan,naive]
265 Specify the clustering algorithm to use. By default DBSCAN will be used.
266 Naive clustering algorithm is better for doing further work on the
267 `-analysis-inconsistencies-output-file=` output, it will create one cluster
268 per opcode, and check that the cluster is stable (all points are neighbours).
270 .. option:: -analysis-numpoints=<dbscan numPoints parameter>
272 Specify the numPoints parameters to be used for DBSCAN clustering
273 (`analysis` mode, DBSCAN only).
275 .. option:: -analysis-clustering-epsilon=<dbscan epsilon parameter>
277 Specify the epsilon parameter used for clustering of benchmark points
280 .. option:: -analysis-inconsistency-epsilon=<epsilon>
282 Specify the epsilon parameter used for detection of when the cluster
283 is different from the LLVM schedule profile values (`analysis` mode).
285 .. option:: -analysis-display-unstable-clusters
287 If there is more than one benchmark for an opcode, said benchmarks may end up
288 not being clustered into the same cluster if the measured performance
289 characteristics are different. by default all such opcodes are filtered out.
290 This flag will instead show only such unstable opcodes.
292 .. option:: -ignore-invalid-sched-class=false
294 If set, ignore instructions that do not have a sched class (class idx = 0).
296 .. option:: -mcpu=<cpu name>
298 If set, measure the cpu characteristics using the counters for this CPU. This
299 is useful when creating new sched models (the host CPU is unknown to LLVM).
301 .. option:: --dump-object-to-disk=true
303 By default, llvm-exegesis will dump the generated code to a temporary file to
304 enable code inspection. You may disable it to speed up the execution and save
310 :program:`llvm-exegesis` returns 0 on success. Otherwise, an error message is
311 printed to standard error, and the tool returns a non 0 value.