Deep Neural Network Library (DNNL)  1.2.0
Performance library for Deep Learning
 All Classes Namespaces Files Functions Variables Typedefs Enumerations Enumerator Groups Pages
Profiling DNNL performance

DNNL uses JIT (just-in-time) code generation based on primitive parameters and instruction set supported by the system. In order to correctly attribute performance event information, profilers need to be notified about address ranges containing JIT-ed code. DNNL supports two profilers: Intel VTune Amplifier and Linux perf.

At build-time, support for this feature is controlled via cmake option DNNL_ENABLE_JIT_PROFILING:

Option Possible Values (defaults in bold) Description
DNNL_ENABLE_JIT_PROFILING ON, OFF Enables integration with performance profilers

At run-time, this feature can be controlled via the following two functions:

or via the DNNL_JIT_PROFILE environment variable which accepts the same values as the dnnl_set_jit_profiling_flags function. The following individual flags may be OR-ed:

The default setting of the profiling flags is to enable integration with Intel VTune Amplifier, therefore it does not require any additional setup and works out of the box. Code integrating DNNL may override this behavior.

Example: profiling with Intel VTune Amplifier

Assuming that environment is set up already.

Collect profiling data:

$ amplxe-cl -collect hotspots -q -no-summary -knob sampling-mode=hw -r dnnl-vtune ./benchdnn --mode=P mb1ic32ih14oc32oh14kh3ph1n"resnet_50:res4a_branch2b*6"
amplxe: Warning: To enable hardware event-base sampling, VTune Amplifier has disabled the NMI watchdog timer. The watchdog timer will be re-enabled after collection completes.
Output template: perf,%engine%,%name%,%desc%,%Gops%,%Gfreq%,%-time%,%-Gflops%,%0time%,%0Gflops%
perf,cpu,resnet_50:res4a_branch2b*6,--conv mb1ic32ih14oc32oh14kh3ph1nresnet_50:res4a_branch2b*6,0.0032768,0,2.13525,1.53462,4.32546,0.757561
tests:1 passed:0 skipped:0 mistrusted:0 unimplemented:0 failed:0 listed:0
total perf: min(ms):2.13525 avg(ms):4.32546
Note
You don't need to set DNNL_JIT_PROFILE environment variable.

Display top 10 hotspots using command-line interface:

$ amplxe-cl -report hotspots -q -r dnnl-vtune -format csv -csv-delimiter ';' -group-by process,module,function -column 'CPU Time:Self' | head -n 10 | column -t -s';'
Column filter is ON.
Process Module Function CPU Time
benchdnn libgomp.so.1.0.0 do_spin 54.796608
benchdnn libgomp.so.1.0.0 do_spin 52.075321
benchdnn libgomp.so.1.0.0 cpu_relax 3.979194
benchdnn libgomp.so.1.0.0 cpu_relax 3.838870
benchdnn [Dynamic code] jit_avx2_conv_fwd_kernel_f32 2.355442
benchdnn vmlinux __lock_acquire 0.801853
benchdnn vmlinux do_raw_spin_lock 0.290672
benchdnn libdnnl.so.1.1 dnnl::impl::cpu::jit_avx2_convolution_fwd_t::execute_forward(dnnl::impl::exec_ctx_t const&) const::{lambda(intint)#1}::operator() 0.260602
benchdnn vmlinux plist_check_prev_next 0.115266

The JIT-ed function jit_avx2_conv_fwd_kernel_f32 is shown as belonging to the [Dynamic code] module.

See more examples in the Intel VTune Amplifier User Guide

Example: profiling with Linux perf

The following command instructs DNNL to enable both jitdump and perfmap profiling modes and write jitdump files into .debug directory in the current directory by setting environment variable JITDUMPDIR to point to the current directory.

$ JITDUMPDIR=. DNNL_JIT_PROFILE=6 perf record -k1 ./tests/benchdnn/benchdnn --mode=P mb1ic32ih14oc32oh14kh3ph1n"resnet_50:res4a_branch2b*6"
Output template: perf,%engine%,%name%,%desc%,%Gops%,%Gfreq%,%-time%,%-Gflops%,%0time%,%0Gflops%
perf,cpu,resnet_50:res4a_branch2b*6,--conv mb1ic32ih14oc32oh14kh3ph1nresnet_50:res4a_branch2b*6,0.0032768,0,0.0131836,248.551,0.0262988,124.599
tests:1 passed:0 skipped:0 mistrusted:0 unimplemented:0 failed:0 listed:0
total perf: min(ms):0.0131836 avg(ms):0.0262988
[ perf record: Woken up 1 times to write data ]
[ perf record: Captured and wrote 0.884 MB perf.data (23102 samples) ]

The following command injects the information from the jitdump files into the performance data:

$ perf inject -j -i perf.data -o perf.data.j

The following command displays the top hotspots:

$ perf report -i perf.data.j --stdio | head -n20
# To display the perf.data header info, please use --header/--header-only options.
#
#
# Total Lost Samples: 0
#
# Samples: 23K of event 'cpu-clock:uhH'
# Event count (approx.): 5775500000
#
# Overhead Command Shared Object Symbol
#
39.33% benchdnn libgomp.so.1.0.0 [.] 0x000000000001d8ba
29.41% benchdnn jitted-31475-0.so [.] jit_avx2_conv_fwd_kernel_f32
20.49% benchdnn libgomp.so.1.0.0 [.] 0x000000000001d712
3.47% benchdnn libdnnl.so.1.1 [.] dnnl::impl::cpu::jit_avx2_convolution_fwd_t::execute_forward(dnnl::impl::exec_ctx_t const&) const::{lambda(int, int)#1}::operator()
1.52% benchdnn libgomp.so.1.0.0 [.] 0x000000000001d8be
0.93% benchdnn libgomp.so.1.0.0 [.] 0x000000000001d716
0.75% benchdnn libgomp.so.1.0.0 [.] 0x000000000001d8c5
0.55% benchdnn libgomp.so.1.0.0 [.] 0x000000000001d8c3
0.46% benchdnn libgomp.so.1.0.0 [.] 0x000000000001d71d
Note
Not every kernel / distribution support displaying detailed profiling information. Symbol resolution (usually) works as long as the perfmap mode is enabled, but annotating a JIT-ed functions disassembly, which requires jitdump, seems to often fail on kernels before 5.x.

See more on the Brendan Gregg's excellent perf examples page