A basic kernel benchmark can be created with just a few lines of CUDA C++:
void my_benchmark(nvbench::state& state) {
state.exec([](nvbench::launch& launch) {
my_kernel<<<num_blocks, 256, 0, launch.get_stream()>>>();
});
}
NVBENCH_BENCH(my_benchmark);
There are three main components in the definition of a benchmark:
- A
KernelGenerator
callable (my_benchmark
above) - A
KernelLauncher
callable (the lambda passed tonvbench::exec
), and - A
BenchmarkDeclaration
usingNVBENCH_BENCH
or similar macros.
The KernelGenerator
is called with an nvbench::state
object that provides
configuration information, as shown in later sections. The generator is
responsible for configuring and instantiating a KernelLauncher
, which is
(unsurprisingly) responsible for launching a kernel. The launcher should contain
only the minimum amount of code necessary to start the CUDA kernel,
since nvbench::exec
will execute it repeatedly to gather timing information.
An nvbench::launch
object is provided to the launcher to specify kernel
execution details, such as the CUDA stream to use. NVBENCH_BENCH
registers
the benchmark with NVBench and initializes various attributes, including its
name and parameter axes.
By default, a benchmark is named by converting the first argument
of NVBENCH_BENCH
into a string.
This can be changed to something more descriptive if desired.
The NVBENCH_BENCH
macro produces a customization object that allows such
attributes to be modified.
NVBENCH_BENCH(my_benchmark).set_name("my_kernel<<<num_blocks, 256>>>");
NVBench records GPU execution times on a specific CUDA stream. By default, a new
stream is created and passed to the KernelLauncher
via the
nvbench::launch::get_stream()
method, as shown in
Minimal Benchmark. All benchmarked kernels and other
stream-ordered work must be launched on this stream for NVBench to capture it.
In some instances, it may be inconvenient or impossible to specify an explicit
CUDA stream for the benchmarked operation to use. For example, a library may
manage and use its own streams, or an opaque API may always launch work on the
default stream. In these situations, users may provide NVBench with an explicit
stream via nvbench::state::set_cuda_stream
and nvbench::make_cuda_stream_view
.
It is assumed that all work of interest executes on or synchronizes with this
stream.
void my_benchmark(nvbench::state& state) {
cudaStream_t default_stream = 0;
state.set_cuda_stream(nvbench::make_cuda_stream_view(default_stream));
state.exec([](nvbench::launch&) {
my_func(); // a host API invoking GPU kernels on the default stream
my_kernel<<<num_blocks, 256>>>(); // or a kernel launched with the default stream
});
}
NVBENCH_BENCH(my_benchmark);
A full example can be found in examples/stream.cu.
Some kernels will be used with a variety of options, input data types/sizes, and other factors that impact performance. NVBench explores these different scenarios by sweeping through a set of user-defined parameter axes.
A parameter axis defines a set of interesting values for a single kernel
parameter — for example, the size of the input, or the type of values being
processed. These parameter axes are used to customize a KernelGenerator
with
static and runtime configurations. There are four supported types of parameters:
int64, float64, string, and type.
More examples can found in examples/axes.cu.
A common example of a parameter axis is to vary the number of input values a
kernel should process during a benchmark measurement. An int64_axis
is ideal
for this:
void benchmark(nvbench::state& state)
{
const auto num_inputs = state.get_int64("NumInputs");
thrust::device_vector<int> data = generate_input(num_inputs);
state.exec([&data](nvbench::launch& launch) {
my_kernel<<<blocks, threads, 0, launch.get_stream()>>>(data.begin(), data.end());
});
}
NVBENCH_BENCH(benchmark).add_int64_axis("NumInputs", {16, 64, 256, 1024, 4096});
NVBench will run the benchmark
kernel generator once for each specified value
in the "NumInputs" axis. The state
object provides the current parameter value
to benchmark
.
Using powers-of-two is quite common for these sorts of axes. int64_axis
has a
unique power-of-two mode that simplifies how such axes are defined and helps
provide more readable output. A power-of-two int64 axis is defined using the
integer exponents, but the benchmark will be run with the computed 2^N value.
// Equivalent to above, {16, 64, 256, 1024, 4096} = {2^4, 2^6, 2^8, 2^10, 2^12}
NVBENCH_BENCH(benchmark).add_int64_power_of_two_axis("NumInputs",
{4, 6, 8, 10, 12});
// Or, as shown in a later section:
NVBENCH_BENCH(benchmark).add_int64_power_of_two_axis("NumInputs",
nvbench::range(4, 12, 2));
For floating point numbers, a float64_axis
is available:
void benchmark(nvbench::state& state)
{
const auto quality = state.get_float64("Quality");
state.exec([&quality](nvbench::launch& launch)
{
my_kernel<<<blocks, threads, 0, launch.get_stream()>>>(quality);
});
}
NVBENCH_BENCH(benchmark).add_float64_axis("Quality", {0.05, 0.1, 0.25, 0.5, 0.75, 1.});
For non-numeric data, an axis of arbitrary strings provides additional flexibility:
void benchmark(nvbench::state& state)
{
const auto rng_dist = state.get_string("RNG Distribution");
thrust::device_vector<int> data = generate_input(rng_dist);
state.exec([&data](nvbench::launch& launch)
{
my_kernel<<<blocks, threads, 0, launch.get_stream()>>>(data.begin(), data.end());
});
}
NVBENCH_BENCH(benchmark).add_string_axis("RNG Distribution", {"Uniform", "Gaussian"});
A common use for string axes is to encode enum values, as shown in examples/enums.cu.
Another common situation involves benchmarking a templated kernel with multiple compile-time configurations. NVBench strives to make such benchmarks as easy to write as possible through the use of type axes.
A type_axis
is a list of types (T1
, T2
, Ts
...) wrapped in
a nvbench::type_list<T1, T2, Ts...>
. The kernel generator becomes a template
function and will be instantiated using types defined by the axis. The current
configuration's type is passed into the kernel generator using
a nvbench::type_list<T>
.
template <typename T>
void my_benchmark(nvbench::state& state, nvbench::type_list<T>)
{
thrust::device_vector<T> data = generate_input<T>();
state.exec([&data](nvbench::launch& launch)
{
my_kernel<<<blocks, threads, 0, launch.get_stream()>>>(data.begin(), data.end());
});
}
using my_types = nvbench::type_list<int, float, double>;
NVBENCH_BENCH_TYPES(my_benchmark, NVBENCH_TYPE_AXES(my_types))
.set_type_axes_names({"ValueType"});
The NVBENCH_TYPE_AXES
macro is unfortunately necessary to prevent commas in
the type_list<...>
from breaking macro parsing.
Type axes can be used to encode compile-time enum and integral constants using
the nvbench::enum_type_list
helper. See
examples/enums.cu for detail.
Since parameter sweeps often explore a range of evenly-spaced numeric values, a
strided range can be generated using the nvbench::range(start, end, stride=1)
helper.
assert(nvbench::range(2, 5) == {2, 3, 4, 5});
assert(nvbench::range(2.0, 5.0) == {2.0, 3.0, 4.0, 5.0});
assert(nvbench::range(2, 12, 2) == {2, 4, 6, 8, 10, 12});
assert(nvbench::range(2, 12, 5) == {2, 7, 12});
assert(nvbench::range(2, 12, 6) == {2, 8});
assert(nvbench::range(0.0, 10.0, 2.5) == { 0.0, 2.5, 5.0, 7.5, 10.0});
Note that start and end are inclusive. This utility can be used to define axis values for all numeric axes.
If more than one axis is defined, the complete cartesian product of all axes will be benchmarked. For example, consider a benchmark with two type axes, one int64 axis, and one float64 axis:
// InputTypes: {char, int, unsigned int}
// OutputTypes: {float, double}
// NumInputs: {2^10, 2^20, 2^30}
// Quality: {0.5, 1.0}
using input_types = nvbench::type_list<char, int, unsigned int>;
using output_types = nvbench::type_list<float, double>;
NVBENCH_BENCH_TYPES(benchmark, NVBENCH_TYPE_AXES(input_types, output_types))
.set_type_axes_names({"InputType", "OutputType"})
.add_int64_power_of_two_axis("NumInputs", nvbench::range(10, 30, 10))
.add_float64_axis("Quality", {0.5, 1.0});
This would generate a total of 36 configurations and instantiate the benchmark 6 times. Keep the rapid growth of these combinations in mind when choosing the number of values in an axis. See the section about combinatorial explosion for more examples and information.
In additional to raw timing information, NVBench can track a kernel's throughput, reporting the amount of data processed as:
- Number of items per second
- Number of bytes per second
- Percentage of device's peak memory bandwidth utilized
To enable throughput measurements, the kernel generator can specify the number
of items and/or bytes handled in a single kernel execution using
the nvbench::state
API.
state.add_element_count(size);
state.add_global_memory_reads<InputType>(size);
state.add_global_memory_writes<OutputType>(size);
In general::
- Add only the input element count (no outputs).
- Add all reads and writes to global memory.
More examples can found in examples/throughput.cu.
Sometimes particular combinations of parameters aren't useful or interesting — or for type axes, some configurations may not even compile.
The nvbench::state
object provides a skip("Reason")
method that can be used
to avoid running these benchmarks. To skip uncompilable type axis
configurations, create an overload for the kernel generator that selects for the
invalid type combination:
template <typename T, typename U>
void my_benchmark(nvbench::state& state, nvbench::type_list<T, U>)
{
// Skip benchmarks at runtime:
if (should_skip_this_config)
{
state.skip("Reason for skip.");
return;
}
/* ... */
};
// Skip benchmarks at compile time -- for example, always skip when T == U
// (Note that the `type_list` argument defines the same type twice).
template <typename SameType>
void my_benchmark(nvbench::state& state,
nvbench::type_list<SameType, SameType>)
{
state.skip("T must not be the same type as U.");
}
using Ts = nvbench::type_list<...>;
using Us = nvbench::type_list<...>;
NVBENCH_BENCH_TYPES(my_benchmark, NVBENCH_TYPE_AXES(Ts, Us));
More examples can found in examples/skip.cu.
By default, NVBench assumes that the entire execution time of the
KernelLauncher
should be measured, and that no syncs are performed
(e.g. cudaDeviceSynchronize
, cudaStreamSynchronize
, cudaEventSynchronize
,
etc. are not called).
Execution tags may be passed to state.exec
when these assumptions are not
true:
nvbench::exec_tag::sync
tells NVBench that the kernel launcher will synchronize internally.nvbench::exec_tag::timer
requests a timer object that can be used to restrict the timed region.
Multiple execution tags may be combined using operator|
, e.g.
state.exec(nvbench::exec_tag::sync | nvbench::exec_tag::timer,
[](nvbench::launch &launch, auto& timer) { /*...*/ });
The following sections provide more details on these features.
If a KernelLauncher
synchronizes the CUDA device internally without passing
this tag, the benchmark will deadlock at runtime. Passing the sync
tag
will fix this issue. Note that this disables batch measurements.
void sync_example(nvbench::state& state)
{
// Pass the `sync` exec tag to tell NVBench that this benchmark will sync:
state.exec(nvbench::exec_tag::sync, [](nvbench::launch& launch) {
/* Benchmark that implicitly syncs here. */
});
}
NVBENCH_BENCH(sync_example);
See examples/exec_tag_sync.cu for a complete example.
For some kernels, the working data may need to be reset between launches. This is particularly common for kernels that modify their input in-place.
Resetting the input data to prepare for a new trial shouldn't be included in the benchmark's execution time. NVBench provides a manual timer mode that allows the kernel launcher to specify the critical section to be measured and exclude any per-trial reset operations.
To enable the manual timer mode, pass the tag object nvbench::exec_tag::timer
to state.exec
, and declare the kernel launcher with an
additional auto& timer
argument.
Note that using manual timer mode disables batch measurements.
void timer_example(nvbench::state& state)
{
// Pass the `timer` exec tag to request a timer:
state.exec(nvbench::exec_tag::timer,
// Lambda now accepts a timer:
[](nvbench::launch& launch, auto& timer)
{
/* Reset code here, excluded from timing */
/* Timed region is explicitly marked.
* The timer handles any synchronization, flushes, etc when/if
* needed for the current measurement.
*/
timer.start();
/* Launch kernel on `launch.get_stream()` here */
timer.stop();
});
}
NVBENCH_BENCH(timer_example);
See examples/exec_tag_timer.cu for a complete example.
Be very careful of how quickly the configuration space can grow. The following example generates 960 total runtime benchmark configurations, and will compile 192 different static parametrizations of the kernel generator. This is likely excessive, especially for routine regression testing.
using value_types = nvbench::type_list<nvbench::uint8_t,
nvbench::int32_t,
nvbench::float32_t,
nvbench::float64_t>;
using op_types = nvbench::type_list<thrust::plus<>,
thrust::multiplies<>,
thrust::maximum<>>;
NVBENCH_BENCH_TYPES(my_benchmark,
NVBENCH_TYPE_AXES(value_types,
value_types,
value_types,
op_types>))
.set_type_axes_names({"T", "U", "V", "Op"})
.add_int64_power_of_two_axis("NumInputs", nvbench::range(10, 30, 5));
960 total configs
= 4 [T=(U8, I32, F32, F64)]
* 4 [U=(U8, I32, F32, F64)]
* 4 [V=(U8, I32, F32, F64)]
* 3 [Op=(plus, multiplies, max)]
* 5 [NumInputs=(2^10, 2^15, 2^20, 2^25, 2^30)]
For large configuration spaces like this, pruning some of the less useful
combinations (e.g. sizeof(init_type) < sizeof(output)
) using the techniques
described in the "Skip Uninteresting / Invalid Benchmarks" section can help
immensely with keeping compile / run times manageable.
Splitting a single large configuration space into multiple, more focused benchmarks with reduced dimensionality will likely be worth the effort as well.