PyTorch Benchmark

This recipe provides a quick-start guide to using PyTorch benchmark module to measure and compare code performance.

Introduction

Benchmarking is an important step in writing code. It helps us validate that our code meets performance expectations, compare different approaches to solving the same problem and prevent performance regressions.

There are many options when it comes to benchmarking PyTorch code including the Python builtin timeit module. However, benchmarking PyTorch code has many caveats that can be easily overlooked such as managing the number of threads and synchronizing CUDA devices. Moreover, generating Tensor inputs for benchmarking can be quite tedious.

This recipe demonstrates how to use PyTorch benchmark module to avoid common mistakes while making it easier to compare performance of different code, generate input for benchmarking and more.

Setup

Before we begin, install torch if it isn’t already available.

pip install torch

Steps

1. Defining functions to benchmark

2. Benchmarking with timeit.Timer

3. Benchmarking with torch.utils.benchmark.Timer

4. Benchmarking with Blocked Autorange

5. Comparing benchmark results

6. Saving/Loading benchmark results

7. Generating inputs with Fuzzed Parameters

8. Collecting instruction counts with Callgrind

1. Defining functions to benchmark

As of the time of this writing, torch.dot does not support batched mode, so we will compare two approaches to implementing it using existing torch operators: one approach uses a combination of mul and sum while the other reduces the problem to bmm.

import torch


def batched_dot_mul_sum(a, b):
    '''Computes batched dot by multiplying and summing'''
    return a.mul(b).sum(-1)


def batched_dot_bmm(a, b):
    '''Computes batched dot by reducing to bmm'''
    a = a.reshape(-1, 1, a.shape[-1])
    b = b.reshape(-1, b.shape[-1], 1)
    return torch.bmm(a, b).flatten(-3)


# Input for benchmarking
x = torch.randn(10000, 64)

# Ensure that both functions compute the same output
assert batched_dot_mul_sum(x, x).allclose(batched_dot_bmm(x, x))

2. Benchmarking with timeit.Timer

First, let’s benchmark the code using Python’s builtin timeit module. We keep the benchmark code simple here so we can compare the defaults of timeit and torch.utils.benchmark.

import timeit

t0 = timeit.Timer(
    stmt='batched_dot_mul_sum(x, x)',
    setup='from __main__ import batched_dot_mul_sum',
    globals={'x': x})

t1 = timeit.Timer(
    stmt='batched_dot_bmm(x, x)',
    setup='from __main__ import batched_dot_bmm',
    globals={'x': x})

print(f'mul_sum(x, x):  {t0.timeit(100) / 100 * 1e6:>5.1f} us')
print(f'bmm(x, x):      {t1.timeit(100) / 100 * 1e6:>5.1f} us')

Output

 mul_sum(x, x):  111.6 us
 bmm(x, x):       70.0 us

3. Benchmarking with torch.utils.benchmark.Timer

PyTorch benchmark module was designed to be familiar to those who have used the timeit module before. However, its defaults make it easier and safer to use for benchmarking PyTorch code. Let’s first compare the same basic API as above.

import torch.utils.benchmark as benchmark

t0 = benchmark.Timer(
    stmt='batched_dot_mul_sum(x, x)',
    setup='from __main__ import batched_dot_mul_sum',
    globals={'x': x})

t1 = benchmark.Timer(
    stmt='batched_dot_bmm(x, x)',
    setup='from __main__ import batched_dot_bmm',
    globals={'x': x})

print(t0.timeit(100))
print(t1.timeit(100))

Output

 <torch.utils.benchmark.utils.common.Measurement object at 0x7fb10400d0f0>
 batched_dot_mul_sum(x, x)
 setup: from __main__ import batched_dot_mul_sum
   379.29 us
   1 measurement, 100 runs , 1 thread
 <torch.utils.benchmark.utils.common.Measurement object at 0x7fb103d67048>
 batched_dot_bmm(x, x)
 setup: from __main__ import batched_dot_bmm
   716.42 us
   1 measurement, 100 runs , 1 thread

Even though the APIs are the same for the basic functionality, there are some important differences. benchmark.Timer.timeit() returns the time per run as opposed to the total runtime like timeit.Timer.timeit() does. PyTorch benchmark module also provides formatted string representations for printing the results.

Another important difference, and the reason why the results diverge is that PyTorch benchmark module runs in a single thread by default. We can change the number of threads with the num_threads arg.

torch.utils.benchmark.Timer takes several additional arguments including: label, sub_label, description and env which change the __repr__ of the measurement object returned and are used for grouping the results (more on this later).

num_threads = torch.get_num_threads()
print(f'Benchmarking on {num_threads} threads')

t0 = benchmark.Timer(
    stmt='batched_dot_mul_sum(x, x)',
    setup='from __main__ import batched_dot_mul_sum',
    globals={'x': x},
    num_threads=num_threads,
    label='Multithreaded batch dot',
    sub_label='Implemented using mul and sum')

t1 = benchmark.Timer(
    stmt='batched_dot_bmm(x, x)',
    setup='from __main__ import batched_dot_bmm',
    globals={'x': x},
    num_threads=num_threads,
    label='Multithreaded batch dot',
    sub_label='Implemented using bmm')

print(t0.timeit(100))
print(t1.timeit(100))

Output

 Benchmarking on 40 threads
 <torch.utils.benchmark.utils.common.Measurement object at 0x7fb103d54080>
 Multithreaded batch dot: Implemented using mul and sum
 setup: from __main__ import batched_dot_mul_sum
   118.47 us
   1 measurement, 100 runs , 40 threads
 <torch.utils.benchmark.utils.common.Measurement object at 0x7fb16935d2e8>
 Multithreaded batch dot: Implemented using bmm
 setup: from __main__ import batched_dot_bmm
   68.21 us
   1 measurement, 100 runs , 40 threads

Running benchmark with all threads available gives similar results as the timeit module. More importantly, which version is faster depends on how many threads we run the code with. This is why it’s important to benchmark the code with thread settings that are representative of real use cases. Another important thing to remember is to synchronize CPU and CUDA when benchmarking on the GPU. Let’s run the above benchmarks again on a CUDA tensor and see what happens.

x = torch.randn(10000, 1024, device='cuda')

t0 = timeit.Timer(
    stmt='batched_dot_mul_sum(x, x)',
    setup='from __main__ import batched_dot_mul_sum',
    globals={'x': x})

t1 = timeit.Timer(
    stmt='batched_dot_bmm(x, x)',
    setup='from __main__ import batched_dot_bmm',
    globals={'x': x})

# Ran each twice to show difference before/after warmup
print(f'mul_sum(x, x):  {t0.timeit(100) / 100 * 1e6:>5.1f} us')
print(f'mul_sum(x, x):  {t0.timeit(100) / 100 * 1e6:>5.1f} us')
print(f'bmm(x, x):      {t1.timeit(100) / 100 * 1e6:>5.1f} us')
print(f'bmm(x, x):      {t1.timeit(100) / 100 * 1e6:>5.1f} us')

Output

 mul_sum(x, x):   27.6 us
 mul_sum(x, x):   25.3 us
 bmm(x, x):      2775.5 us
 bmm(x, x):       22.4 us

t0 = benchmark.Timer(
    stmt='batched_dot_mul_sum(x, x)',
    setup='from __main__ import batched_dot_mul_sum',
    globals={'x': x})

t1 = benchmark.Timer(
    stmt='batched_dot_bmm(x, x)',
    setup='from __main__ import batched_dot_bmm',
    globals={'x': x})

# Run only once since benchmark module does warmup for us
print(t0.timeit(100))
print(t1.timeit(100))

Output

 <torch.utils.benchmark.utils.common.Measurement object at 0x7fb10400d080>
 batched_dot_mul_sum(x, x)
 setup: from __main__ import batched_dot_mul_sum
   232.93 us
   1 measurement, 100 runs , 1 thread
 <torch.utils.benchmark.utils.common.Measurement object at 0x7fb10400d0f0>
 batched_dot_bmm(x, x)
 setup: from __main__ import batched_dot_bmm
   181.04 us
   1 measurement, 100 runs , 1 thread

The results reveal something interesting. The first run of the bmm version using the timeit module takes much longer than the second run. This is because bmm calls into cuBLAS which needs to be loaded the first time it’s called which takes some time. This is why it’s important to do a warmup run before benchmarking, luckily for us, PyTorch’s benchmark module takes care of that.

The difference in the results between timeit and benchmark modules is because the timeit module is not synchronizing CUDA and is thus only timing the time to launch the kernel. PyTorch’s benchmark module does the synchronization for us.

4. Benchmarking with Blocked Autorange

While timeit.Timer.autorange takes a single continuous measurement of at least 0.2 seconds, torch.utils.benchmark.blocked_autorange takes many measurements whose times total at least 0.2 seconds (which can be changed by the min_run_time parameter) subject to the constraint that timing overhead is a small fraction of the overall measurement. This is accomplished by first running with an increasing number of runs per loop until the runtime is much larger than measurement overhead (which also serves as a warm up), and then taking measurements until the target time is reached. This has the useful properties that it wastes less data and allows us to compute statistics to estimate the reliability of the measurements.

m0 = t0.blocked_autorange()
m1 = t1.blocked_autorange()

print(m0)
print(m1)

Output

 <torch.utils.benchmark.utils.common.Measurement object at 0x7fb10400d0f0>
 batched_dot_mul_sum(x, x)
 setup: from __main__ import batched_dot_mul_sum
   231.79 us
   1 measurement, 1000 runs , 1 thread
 <torch.utils.benchmark.utils.common.Measurement object at 0x7fb10400d080>
 batched_dot_bmm(x, x)
 setup: from __main__ import batched_dot_bmm
   Median: 162.08 us
   2 measurements, 1000 runs per measurement, 1 thread

We can also inspect the individual statistics from the returned measurements object.

print(f"Mean:   {m0.mean * 1e6:6.2f} us")
print(f"Median: {m0.median * 1e6:6.2f} us")

Output

 Mean:   231.79 us
 Median: 231.79 us

5. Comparing benchmark results

So far we’ve been comparing our two versions of batched dot against a single input. In practice, we want to try a combination of inputs as well as different number of threads. The Compare class helps display the results of many measurements in a formatted table. It uses the annotations described above (label, sub_label, num_threads, etc.) as well as description to group and organize the table. Let’s use Compare to see how our functions perform for different input sizes and number of threads.

from itertools import product

# Compare takes a list of measurements which we'll save in results.
results = []

sizes = [1, 64, 1024, 10000]
for b, n in product(sizes, sizes):
    # label and sub_label are the rows
    # description is the column
    label = 'Batched dot'
    sub_label = f'[{b}, {n}]'
    x = torch.ones((b, n))
    for num_threads in [1, 4, 16, 32]:
        results.append(benchmark.Timer(
            stmt='batched_dot_mul_sum(x, x)',
            setup='from __main__ import batched_dot_mul_sum',
            globals={'x': x},
            num_threads=num_threads,
            label=label,
            sub_label=sub_label,
            description='mul/sum',
        ).blocked_autorange(min_run_time=1))
        results.append(benchmark.Timer(
            stmt='batched_dot_bmm(x, x)',
            setup='from __main__ import batched_dot_bmm',
            globals={'x': x},
            num_threads=num_threads,
            label=label,
            sub_label=sub_label,
            description='bmm',
        ).blocked_autorange(min_run_time=1))

compare = benchmark.Compare(results)
compare.print()

Output

 [--------------- Batched dot ----------------]
                       |  mul/sum   |    bmm
 1 threads: -----------------------------------
       [1, 1]          |       5.9  |      11.2
       [1, 64]         |       6.4  |      11.4
       [1, 1024]       |       6.7  |      14.2
       [1, 10000]      |      10.2  |      23.7
       [64, 1]         |       6.3  |      11.5
       [64, 64]        |       8.6  |      15.4
       [64, 1024]      |      39.4  |     204.4
       [64, 10000]     |     274.9  |     748.5
       [1024, 1]       |       7.7  |      17.8
       [1024, 64]      |      40.3  |      76.4
       [1024, 1024]    |     432.4  |    2795.9
       [1024, 10000]   |   22657.3  |   11899.5
       [10000, 1]      |      16.9  |      74.8
       [10000, 64]     |     300.3  |     609.4
       [10000, 1024]   |   23098.6  |   27246.1
       [10000, 10000]  |  267073.7  |  118823.7
 4 threads: -----------------------------------
       [1, 1]          |       6.0  |      11.5
       [1, 64]         |       6.2  |      11.2
       [1, 1024]       |       6.8  |      14.3
       [1, 10000]      |      10.2  |      23.7
       [64, 1]         |       6.3  |      16.2
       [64, 64]        |       8.8  |      18.2
       [64, 1024]      |      41.5  |     189.1
       [64, 10000]     |      91.7  |     849.1
       [1024, 1]       |       7.6  |      17.4
       [1024, 64]      |      43.5  |      33.5
       [1024, 1024]    |     135.4  |    2782.3
       [1024, 10000]   |    7471.1  |   11874.0
       [10000, 1]      |      16.8  |      33.9
       [10000, 64]     |     118.7  |     173.2
       [10000, 1024]   |    7264.6  |   27824.7
       [10000, 10000]  |  100060.9  |  121499.0
 16 threads: ----------------------------------
       [1, 1]          |       6.0  |      11.3
       [1, 64]         |       6.2  |      11.2
       [1, 1024]       |       6.9  |      14.2
       [1, 10000]      |      10.3  |      23.8
       [64, 1]         |       6.4  |      24.1
       [64, 64]        |       9.0  |      23.8
       [64, 1024]      |      54.1  |     188.5
       [64, 10000]     |      49.9  |     748.0
       [1024, 1]       |       7.6  |      23.4
       [1024, 64]      |      55.5  |      28.2
       [1024, 1024]    |      66.9  |    2773.9
       [1024, 10000]   |    6111.5  |   12833.7
       [10000, 1]      |      16.9  |      27.5
       [10000, 64]     |      59.5  |      73.7
       [10000, 1024]   |    6295.9  |   27062.0
       [10000, 10000]  |   71804.5  |  120365.8
 32 threads: ----------------------------------
       [1, 1]          |       5.9  |      11.3
       [1, 64]         |       6.2  |      11.3
       [1, 1024]       |       6.7  |      14.2
       [1, 10000]      |      10.5  |      23.8
       [64, 1]         |       6.3  |      31.7
       [64, 64]        |       9.1  |      30.4
       [64, 1024]      |      72.0  |     190.4
       [64, 10000]     |     103.1  |     746.9
       [1024, 1]       |       7.6  |      28.4
       [1024, 64]      |      70.5  |      31.9
       [1024, 1024]    |      65.6  |    2804.6
       [1024, 10000]   |    6764.0  |   11871.4
       [10000, 1]      |      17.8  |      31.8
       [10000, 64]     |     110.3  |      56.0
       [10000, 1024]   |    6640.2  |   27592.2
       [10000, 10000]  |   73003.4  |  120083.2

 Times are in microseconds (us).

The results above indicate that the version which reduces to bmm is better for larger tensors running on multiple threads, while for smaller and/or single thread code, the other version is better.

Compare also provides functions for changing the table format

compare.trim_significant_figures()
compare.colorize()
compare.print()

6. Saving/Loading benchmark results

Measurements (and CallgrindStats which are described in section 8) are pickleable. This makes A/B testing easy, as you can collect measurements from two separate environments, pickle them, and then load both in a single environment. Timer even takes an env constructor argument so that such A/B testing works seamlessly.

Let’s imagine that rather than two Python functions, the add/sum and bmm approaches were in two different builds of PyTorch. The example below demonstrates how one might A/B test them. For simplicity, we only use a subset of shapes, and simply round trip results through pickle rather than actually using multiple environments and writing results to disk.

import pickle

ab_test_results = []
for env in ('environment A: mul/sum', 'environment B: bmm'):
    for b, n in ((1, 1), (1024, 10000), (10000, 1)):
        x = torch.ones((b, n))
        dot_fn = (batched_dot_mul_sum if env == 'environment A: mul/sum' else batched_dot_bmm)
        m = benchmark.Timer(
            stmt='batched_dot(x, x)',
            globals={'x': x, 'batched_dot': dot_fn},
            num_threads=1,
            label='Batched dot',
            description=f'[{b}, {n}]',
            env=env,
        ).blocked_autorange(min_run_time=1)
        ab_test_results.append(pickle.dumps(m))

ab_results = [pickle.loads(i) for i in ab_test_results]
compare = benchmark.Compare(ab_results)
compare.trim_significant_figures()
compare.colorize()
compare.print()

Output

 [------------------------------------- Batched dot -------------------------------------]
                                                |  [1, 1]  |  [1024, 10000]  |  [10000, 1]
 1 threads: ------------------------------------------------------------------------------
   (environment A: mul/sum)  batched_dot(x, x)  |     7    |      36000      |      21
   (environment B: bmm)      batched_dot(x, x)  |    14    |      40000      |      85

 Times are in microseconds (us).

# And just to show that we can round trip all of the results from earlier:
round_tripped_results = pickle.loads(pickle.dumps(results))
assert(str(benchmark.Compare(results)) == str(benchmark.Compare(round_tripped_results)))

7. Generating inputs with Fuzzed Parameters

As we’ve seen in the previous section, there can be some stark performance differences depending on the input tensors. Hence, it is a good idea to run benchmarks on a number of different inputs. However, creating all these input tensors can be tedious which is where torch.utils.benchmark.Fuzzer and related classes come in. Let’s take a look at how we can use the Fuzzer to create some test cases for the benchmark.

from torch.utils.benchmark import Fuzzer, FuzzedParameter, FuzzedTensor, ParameterAlias

# Generates random tensors with 128 to 10000000 elements and sizes k0 and k1 chosen from a
# loguniform distribution in [1, 10000], 40% of which will be discontiguous on average.
example_fuzzer = Fuzzer(
    parameters = [
        FuzzedParameter('k0', minval=1, maxval=10000, distribution='loguniform'),
        FuzzedParameter('k1', minval=1, maxval=10000, distribution='loguniform'),
    ],
    tensors = [
        FuzzedTensor('x', size=('k0', 'k1'), min_elements=128, max_elements=10000000, probability_contiguous=0.6)
    ],
    seed=0,
)

results = []
for tensors, tensor_params, params in example_fuzzer.take(10):
    # description is the column label
    sub_label=f"{params['k0']:<6} x {params['k1']:<4} {'' if tensor_params['x']['is_contiguous'] else '(discontiguous)'}"
    results.append(benchmark.Timer(
        stmt='batched_dot_mul_sum(x, x)',
        setup='from __main__ import batched_dot_mul_sum',
        globals=tensors,
        label='Batched dot',
        sub_label=sub_label,
        description='mul/sum',
    ).blocked_autorange(min_run_time=1))
    results.append(benchmark.Timer(
        stmt='batched_dot_bmm(x, x)',
        setup='from __main__ import batched_dot_bmm',
        globals=tensors,
        label='Batched dot',
        sub_label=sub_label,
        description='bmm',
    ).blocked_autorange(min_run_time=1))

compare = benchmark.Compare(results)
compare.trim_significant_figures()
compare.print()

Output

 [--------------------- Batched dot ---------------------]
                                      |  mul/sum  |   bmm
 1 threads: ----------------------------------------------
       725    x 257                   |      87   |    180
       49     x 383                   |      15   |     30
       34     x 1468                  |      30   |    118
       187    x 5039                  |     400   |   1200
       2140   x 1296 (discontiguous)  |    2000   |  41000
       78     x 1598                  |      74   |    310
       519    x 763                   |     190   |   1500
       141    x 1082                  |      87   |    500
       78     x 5    (discontiguous)  |       9   |     20
       187    x 1                     |      12   |     10

 Times are in microseconds (us).

There is a lot of flexibility for defining your own Fuzzers which is great for creating a powerful set of inputs to benchmark. But to make things even simpler, PyTorch benchmark module comes with some buitin Fuzzers for common benchmarking needs. Let’s take a look at how we can use one of these builtin fuzzers.

from torch.utils.benchmark.op_fuzzers import binary

results = []
for tensors, tensor_params, params in binary.BinaryOpFuzzer(seed=0).take(10):
    sub_label=f"{params['k0']:<6} x {params['k1']:<4} {'' if tensor_params['x']['is_contiguous'] else '(discontiguous)'}"
    results.append(benchmark.Timer(
        stmt='batched_dot_mul_sum(x, x)',
        setup='from __main__ import batched_dot_mul_sum',
        globals=tensors,
        label='Batched dot',
        sub_label=sub_label,
        description='mul/sum',
    ).blocked_autorange(min_run_time=1))
    results.append(benchmark.Timer(
        stmt='batched_dot_bmm(x, x)',
        setup='from __main__ import batched_dot_bmm',
        globals=tensors,
        label='Batched dot',
        sub_label=sub_label,
        description='bmm',
    ).blocked_autorange(min_run_time=1))

compare = benchmark.Compare(results)
compare.trim_significant_figures()
compare.colorize(rowwise=True)
compare.print()

Output

 [----------------------- Batched dot ------------------------]
                                          |  mul/sum  |   bmm
 1 threads: ---------------------------------------------------
       64     x 473  (discontiguous)      |    10000  |   40000
       16384  x 12642115 (discontiguous)  |       31  |      78
       8192   x 892                       |     4800  |   20400
       512    x 64   (discontiguous)      |   110000  |  400000
       493    x 27   (discontiguous)      |     1100  |    2440
       118    x 32   (discontiguous)      |      870  |    2030
       16     x 495  (discontiguous)      |    23600  |   24000
       488    x 62374                     |    90000  |  100000
       240372 x 69                        |    40000  |   16000
       40156  x 32   (discontiguous)      |     2670  |    5000

 Times are in microseconds (us).

8. Collecting instruction counts with Callgrind

One of the challenges of optimizing code is the variation and opacity of wall time. There are many sources of non-determinism, from adaptive clock speeds to resource contention with other processes. Furthermore, end-to-end time gives no insight into where time is being spent, which is really what we’re interested in when optimizing code.

A complementary approach is to also collect instruction counts. These counts are a proxy metric and do not capture all aspects of performance (e.g. memory or I/O bound tasks), however they do have several useful properties. Instruction counts are reproducible, insensitive to environmental variation, and offer fine grained insight into where a program is spending cycles.

To see the utility of instruction counts, let us look at how we might reduce the overhead of batched_dot_mul_sum. The obvious solution is to move it to C++, so we avoid going between Python and C++ multiple times.

Fortunately, the source is nearly identical. One question that we have to ask in C++ is whether we should take arguments by value or reference.

batched_dot_src = """\
/* ---- Python ---- */
// def batched_dot_mul_sum(a, b):
//     return a.mul(b).sum(-1)

torch::Tensor batched_dot_mul_sum_v0(
    const torch::Tensor a,
    const torch::Tensor b) {
  return a.mul(b).sum(-1);
}

torch::Tensor batched_dot_mul_sum_v1(
    const torch::Tensor& a,
    const torch::Tensor& b) {
  return a.mul(b).sum(-1);
}
"""


# PyTorch makes it easy to test our C++ implementations by providing a utility
# to JIT compile C++ source into Python extensions:
import os
from torch.utils import cpp_extension
cpp_lib = cpp_extension.load_inline(
    name='cpp_lib',
    cpp_sources=batched_dot_src,
    extra_cflags=['-O3'],
    extra_include_paths=[
        # `load_inline` needs to know where to find Pybind11 headers.
        os.path.join(os.getenv('CONDA_PREFIX'), 'include')
    ],
    functions=['batched_dot_mul_sum_v0', 'batched_dot_mul_sum_v1']
)

# `load_inline` will create a shared object that is loaded into Python. When we collect
# instruction counts Timer will create a subprocess, so we need to re-import it. The
# import process is slightly more complicated for C extensions, but that's all we're
# doing here.
module_import_str = f"""\
# https://stackoverflow.com/questions/67631/how-to-import-a-module-given-the-full-path
import importlib.util
spec = importlib.util.spec_from_file_location("cpp_lib", {repr(cpp_lib.__file__)})
cpp_lib = importlib.util.module_from_spec(spec)
spec.loader.exec_module(cpp_lib)"""

import textwrap
def pretty_print(result):
    """Import machinery for cpp_lib.so can get repetitive to look at."""
    print(repr(result).replace(textwrap.indent(module_import_str, "  "), "  import cpp_lib"))


t_baseline = benchmark.Timer(
    stmt='batched_dot_mul_sum(x, x)',
    setup='''\
from __main__ import batched_dot_mul_sum
x = torch.randn(2, 2)''')

t0 = benchmark.Timer(
    stmt='cpp_lib.batched_dot_mul_sum_v0(x, x)',
    setup=f'''\
{module_import_str}
x = torch.randn(2, 2)''')

t1 = benchmark.Timer(
    stmt='cpp_lib.batched_dot_mul_sum_v1(x, x)',
    setup=f'''\
{module_import_str}
x = torch.randn(2, 2)''')

# Moving to C++ did indeed reduce overhead, but it's hard to tell which
# calling convention is more efficient. v1 (call with references) seems to
# be a bit faster, but it's within measurement error.
pretty_print(t_baseline.blocked_autorange())
pretty_print(t0.blocked_autorange())
pretty_print(t1.blocked_autorange())

Output

 <torch.utils.benchmark.utils.common.Measurement object at 0x7fb16935d2e8>
 batched_dot_mul_sum(x, x)
 setup:
   from __main__ import batched_dot_mul_sum
   x = torch.randn(2, 2)

   6.92 us
   1 measurement, 100000 runs , 1 thread
 <torch.utils.benchmark.utils.common.Measurement object at 0x7fb16935d2e8>
 cpp_lib.batched_dot_mul_sum_v0(x, x)
 setup:
   import cpp_lib
   x = torch.randn(2, 2)

   5.29 us
   1 measurement, 100000 runs , 1 thread
 <torch.utils.benchmark.utils.common.Measurement object at 0x7fb16935d2e8>
 cpp_lib.batched_dot_mul_sum_v1(x, x)
 setup:
   import cpp_lib
   x = torch.randn(2, 2)

   5.22 us
   1 measurement, 100000 runs , 1 thread

# Let's use Callgrind to determine which is better.
stats_v0 = t0.collect_callgrind()
stats_v1 = t1.collect_callgrind()

pretty_print(stats_v0)
pretty_print(stats_v1)

# `.as_standardized` removes file names and some path prefixes, and makes
# it easier to read the function symbols.
stats_v0 = stats_v0.as_standardized()
stats_v1 = stats_v1.as_standardized()

# `.delta` diffs the instruction counts, and `.denoise` removes several
# functions in the Python interpreter that are known to have significant
# jitter.
delta = stats_v1.delta(stats_v0).denoise()

# `.transform` is a convenience API for transforming function names. It is
# useful for increasing cancelation when diff-ing instructions, as well as
# just generally improving readability.
replacements = (
    ("???:void pybind11", "pybind11"),
    ("batched_dot_mul_sum_v0", "batched_dot_mul_sum_v1"),
    ("at::Tensor, at::Tensor", "..."),
    ("at::Tensor const&, at::Tensor const&", "..."),
    ("auto torch::detail::wrap_pybind_function_impl_", "wrap_pybind_function_impl_"),
)
for before, after in replacements:
    delta = delta.transform(lambda l: l.replace(before, after))

# We can use print options to control how much of the function to display.
torch.set_printoptions(linewidth=160)

# Once parsed, the instruction counts make clear that passing `a` and `b`
# by reference is more efficient as it skips some c10::TensorImpl bookkeeping
# for the intermediate Tensors, and is also works better with PyBind11. This
# is consistent with our noisy wall time observations.
print(delta)

Output

 <torch.utils.benchmark.utils.valgrind_wrapper.timer_interface.CallgrindStats object at 0x7fb0f06e7630>
 cpp_lib.batched_dot_mul_sum_v0(x, x)
 setup:
   import cpp_lib
   x = torch.randn(2, 2)

                            All          Noisy symbols removed
     Instructions:      2392671                    2392671
     Baseline:             4367                       4367
 100 runs per measurement, 1 thread
 Warning: PyTorch was not built with debug symbols.
          Source information may be limited. Rebuild with
          REL_WITH_DEB_INFO=1 for more detailed results.
 <torch.utils.benchmark.utils.valgrind_wrapper.timer_interface.CallgrindStats object at 0x7fb10400d208>
 cpp_lib.batched_dot_mul_sum_v1(x, x)
 setup:
   import cpp_lib
   x = torch.randn(2, 2)

                            All          Noisy symbols removed
     Instructions:      2378978                    2378978
     Baseline:             4367                       4367
     100 runs per measurement, 1 thread
     Warning: PyTorch was not built with debug symbols.
              Source information may be limited. Rebuild with
              REL_WITH_DEB_INFO=1 for more detailed results.
     <torch.utils.benchmark.utils.valgrind_wrapper.timer_interface.FunctionCounts object at 0x7fb1000ab358>
           86  ???:0x000000000020d9e0
       56  ???:0x000000000020db10
    -1100  pybind11::cpp_function::initialize<wrap_pybind_function_impl_<at::Tensor ... r (&)(...), std::integer_sequence<unsigned long, 0ul, 1ul>)::{lambda(...)
    -1600  ???:wrap_pybind_function_impl_<at::Tensor (&)(...), 0ul, 1ul>(at::Tensor (&)(...), std::integer_sequence<unsigned long, 0ul, 1ul>)::{lambda(...)
    -5200  ???:c10::intrusive_ptr<c10::TensorImpl, c10::UndefinedTensorImpl>::reset_()
    -5935  ???:0x000000000022c0e0

 Total: -13693

Learn More

Take a look at these other recipes to continue your learning:

• PyTorch Profiler