备注
点击此处下载完整示例代码
PyTorch 性能基准测试
创建于：2025 年 4 月 1 日 | 最后更新：2025 年 4 月 1 日 | 最后验证：2024 年 11 月 5 日
本食谱提供使用 PyTorch benchmark 模块进行快速入门指南，以衡量和比较代码性能。
简介
基准测试是编写代码的重要步骤。它帮助我们验证代码是否满足性能预期，比较解决同一问题的不同方法，并防止性能退步。
在基准测试 PyTorch 代码方面有很多选择，包括 Python 内置的 timeit 模块。然而，基准测试 PyTorch 代码有许多容易忽视的注意事项，例如管理线程数量和同步 CUDA 设备。此外，生成用于基准测试的张量输入可能相当繁琐。
本示例演示了如何使用 PyTorch benchmark 模块来避免在编写代码时常见的错误，并使比较不同代码的性能、生成基准测试输入等变得更加容易。
设置
在开始之前，如果尚未安装 torch ，请先安装。
pip install torch
步骤 ¶
定义基准测试函数
基准测试使用
timeit.Timer基准测试使用
torch.utils.benchmark.Timer基准测试使用
Blocked Autorange比较基准测试结果
保存/加载基准测试结果
生成带有
Fuzzed Parameters的输入使用
Callgrind收集指令计数
1. 定义用于基准测试的函数 ¶
到本文写作时为止，torch.dot 不支持批量模式，因此我们将比较两种使用现有 torch 操作符实现它的方法：一种方法使用 mul 和 sum 的组合，另一种将问题简化为 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. 基准测试与 timeit.Timer 模块
首先，让我们使用 Python 内置的 timeit 模块来基准测试代码。在这里，我们保持基准测试代码简单，以便我们可以比较 timeit 和 torch.utils.benchmark 的默认设置。
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')
输出 ¶
mul_sum(x, x): 111.6 us
bmm(x, x): 70.0 us
3. 使用 torch.utils.benchmark.Timer 模块进行基准测试
PyTorch benchmark 模块是为那些之前使用过 timeit 模块的人设计的。然而，它的默认设置使得在基准测试 PyTorch 代码时更容易、更安全。让我们首先比较上面相同的基准 API。
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))
输出 ¶
<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
尽管基本功能的 API 相同，但也有一些重要的区别。 benchmark.Timer.timeit() 返回每次运行的耗时，而 timeit.Timer.timeit() 返回的是总运行时间。PyTorch benchmark 模块还提供了格式化的字符串表示，以便打印结果。
另一个重要的区别，以及结果差异的原因是，PyTorch 基准模块默认在单个线程中运行。我们可以通过 num_threads 参数来更改线程数。
torch.utils.benchmark.Timer 包含多个额外的参数，包括： label 、 sub_label 、 description 和 env ，这些参数会改变返回的测量对象的 __repr__，并用于结果分组（关于这一点稍后详细说明）。
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))
输出 ¶
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
使用所有可用线程运行 benchmark 会得到与 timeit 模块类似的结果。更重要的是，哪个版本更快取决于我们用多少线程运行代码。这就是为什么用代表实际用例的线程设置基准测试代码很重要。还有一点需要记住的是，在 GPU 上基准测试时，要同步 CPU 和 CUDA。让我们再次在 CUDA 张量上运行上述基准测试，看看会发生什么。
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 warm-up
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')
输出 ¶
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 warm-up for us
print(t0.timeit(100))
print(t1.timeit(100))
输出 ¶
<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
结果揭示了一些有趣的事情。使用 bmm 版本和 timeit 模块的第一轮运行比第二轮运行花费的时间要长得多。这是因为 bmm 调用 cuBLAS，它需要首次调用时加载，这需要一些时间。这就是为什么在进行基准测试之前进行预热运行很重要的原因，幸运的是，PyTorch 的 benchmark 模块会处理这一点。
timeit 和 benchmark 模块之间结果差异的原因是 timeit 模块没有同步 CUDA，因此只计时启动内核的时间。PyTorch 的 benchmark 模块会为我们完成同步。
4. 使用阻塞自动范围进行基准测试 ¶
当 timeit.Timer.autorange 进行至少 0.2 秒的单次连续测量时，torch.utils.benchmark.blocked_autorange 会进行多次测量，这些测量的总时间至少为 0.2 秒（可以通过 min_run_time 参数进行更改），同时满足时延开销是整体测量中很小一部分的约束。这是通过首先以增加的每次循环运行次数运行，直到运行时间远大于测量开销（这也作为预热），然后进行测量，直到达到目标时间来实现的。这具有节省数据并允许我们计算统计数据以估计测量可靠性的有用特性。
m0 = t0.blocked_autorange()
m1 = t1.blocked_autorange()
print(m0)
print(m1)
输出 ¶
<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
我们还可以检查返回的测量对象中的单个统计数据。
print(f"Mean: {m0.mean * 1e6:6.2f} us")
print(f"Median: {m0.median * 1e6:6.2f} us")
输出 ¶
Mean: 231.79 us
Median: 231.79 us
5. 比较基准结果 ¶
到目前为止，我们一直在比较我们的批量点积的两个版本与单个输入。在实践中，我们想尝试多种输入组合以及不同的线程数。 Compare 类帮助以格式化的表格显示许多测量的结果。它使用上述描述的注解（标签、子标签、线程数等）以及描述来分组和组织表格。让我们使用 Compare 来看看我们的函数在不同输入大小和线程数下的性能表现。
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()
输出 ¶
[--------------- 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).
上面的结果表明，简化为 bmm 的版本更适合在多线程上运行的大张量，而对于较小和/或单线程的代码，另一个版本更好。
还提供了更改表格格式的函数
compare.trim_significant_figures()
compare.colorize()
compare.print()
6. 保存/加载基准测试结果
测量值（以及在第 8 节中描述的 CallgrindStats 模块）可以通过该模块进行序列化。这使得 A/B 测试变得简单，因为您可以从两个不同的环境中收集测量值，将它们序列化，然后在单个环境中加载它们。计时器甚至接受 env 构造函数参数，以便 A/B 测试可以无缝进行。
假设，而不是两个 Python 函数，add/sum 和 bmm 方法分别位于 PyTorch 的两个不同构建中。下面的示例演示了如何进行 A/B 测试。为了简单起见，我们只使用形状的子集，并通过 pickle 来回传递结果，而不是实际使用多个环境并将结果写入磁盘。
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()
输出 ¶
[------------------------------------- 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. 使用模糊参数生成输入 ¶
如前文所述，输入张量可能会带来明显的性能差异。因此，对多种不同输入进行基准测试是个好主意。然而，创建所有这些输入张量可能会很繁琐，这正是 torch.utils.benchmark.Fuzzer 和相关类派上用场的地方。让我们看看如何使用 Fuzzer 来为基准测试创建一些测试用例。
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()
输出 ¶
[--------------------- 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).
定义自己的 fuzzers 有很多灵活性，这对于创建一组强大的输入以进行基准测试来说非常棒。但为了让事情更加简单，PyTorch 基准模块还提供了一些内置的 fuzzers 以满足常见的基准测试需求。让我们看看如何使用这些内置的 fuzzers 之一。
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()
输出 ¶
[----------------------- 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. 使用 Callgrind 收集指令计数 ¶
优化代码的一个挑战是墙时的变化和透明度。有许多非确定性的来源，从自适应时钟速度到与其他进程的资源竞争。此外，端到端时间无法揭示时间花费在哪里，而这正是我们在优化代码时真正感兴趣的。
补充的方法是同时收集指令计数。这些计数是一个代理指标，并不能涵盖所有性能方面（例如内存或 I/O 密集型任务），然而它们确实具有几个有用的特性。指令计数是可重复的，对环境变化的敏感性低，并且可以提供对程序在哪里消耗周期的细粒度洞察。
为了看到指令计数的实用性，让我们看看我们如何减少 batched_dot_mul_sum 的开销。显然的解决方案是将它移到 C++，这样我们就避免了多次在 Python 和 C++之间转换。
幸运的是，源代码几乎相同。在 C++中我们必须问的一个问题是，我们应该通过值还是引用来传递参数。
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())
输出 ¶
<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)
<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
