Profiling CUDA Utilizing Nsight Programs: A Numba Instance | by Carlos Costa, Ph.D. | Might, 2024

Optimization is an important a part of writing excessive efficiency code, regardless of in case you are writing an online server or computational fluid dynamics simulation software program. Profiling means that you can make knowledgeable choices relating to your code. In a way, optimization with out profiling is like flying blind: principally fantastic for seasoned professionals with professional data and fine-tuned instinct, however a recipe for catastrophe for nearly everybody else.

Following my preliminary collection CUDA by Numba Examples (see components 1, 2, 3, and 4), we’ll examine a comparability between unoptimized, single-stream code and a barely higher model which makes use of stream concurrency and different optimizations. We are going to study, from the ground-up, the best way to use NVIDIA Nsight Systems to profile and analyze CUDA code. All of the code on this tutorial will also be discovered within the repo cako/profiling-cuda-nsight-systems.

NVIDIA recommends as greatest apply to comply with the APOD framework (Assess, Parallelize, Optimize, Deploy). There are a number of proprietary, open-source, free, and industrial software program for several types of assessments and profiling. Veteran Python customers could also be accustomed to fundamental profilers akin to cProfile, line_profiler, memory_profiler (sadly, unmaintaned as of 2024) and extra superior instruments like PyInstrument and Memray. These profilers goal particular elements of the “host” akin to CPU and RAM utilization.

Nonetheless, profiling “machine” (e.g., GPU) code — and its interactions with the host — requires specialised instruments supplied by the machine vendor. For NVIDIA GPUs, Nsight Programs, Nsight Compute, Nsight Graphics can be found for profiling completely different elements of computation. On this tutorial we’ll deal with utilizing Nsight Programs, which is a system-wide profiler. We are going to use it to profile Python code which interacts with the GPU by way of Numba CUDA.

To get began, you will have Nsight Programs CLI and GUI. The CLI may be put in individually and might be used to profile the code in a GPGPU-capable system. The total model contains each CLI and GUI. Notice that each variations might be put in in a system and not using a GPU. Seize the model(s) you want from the NVIDIA website.

To make it simpler to visualise code sections within the GUI, NVIDIA additionally offers the Python pip and conda-installable library nvtx which we’ll use to annotate sections of our code. Extra on this later.

On this part we’ll set our growth and profiling surroundings up. Beneath are two quite simple Python scripts: kernels.py and run_v1.py. The previous will include all CUDA kernels, and the latter will function the entry level to run the instance. On this instance we’re following the “cut back” sample launched in article CUDA by Numba Examples Half 3: Streams and Occasions to compute the sum of an array.

#%%writefile kernels.py
import numba
from numba import cuda
THREADS_PER_BLOCK = 256
BLOCKS_PER_GRID = 32 * 40
@cuda.jit
def partial_reduce(array, partial_reduction):
i_start = cuda.grid(1)
threads_per_grid = cuda.blockDim.x * cuda.gridDim.x
s_thread = numba.float32(0.0)
for i_arr in vary(i_start, array.dimension, threads_per_grid):
s_thread += array[i_arr]
s_block = cuda.shared.array((THREADS_PER_BLOCK,), numba.float32)
tid = cuda.threadIdx.x
s_block[tid] = s_thread
cuda.syncthreads()
i = cuda.blockDim.x // 2
whereas i > 0:
if tid < i:
s_block[tid] += s_block[tid + i]
cuda.syncthreads()
i //= 2
if tid == 0:
partial_reduction[cuda.blockIdx.x] = s_block[0]
@cuda.jit
def single_thread_sum(partial_reduction, sum):
sum[0] = numba.float32(0.0)
for component in partial_reduction:
sum[0] += component
@cuda.jit
def divide_by(array, val_array):
i_start = cuda.grid(1)
threads_per_grid = cuda.gridsize(1)
for i in vary(i_start, array.dimension, threads_per_grid):
array[i] /= val_array[0]

#%%writefile run_v1.py
import argparse
import warnings
import numpy as np
from numba import cuda
from numba.core.errors import NumbaPerformanceWarning
from kernels import (
BLOCKS_PER_GRID,
THREADS_PER_BLOCK,
divide_by,
partial_reduce,
single_thread_sum,
)
# Ignore NumbaPerformanceWarning
warnings.simplefilter("ignore", class=NumbaPerformanceWarning)
def run(dimension):
# Outline host array
a = np.ones(dimension, dtype=np.float32)
print(f"Outdated sum: {a.sum():.3f}")
# Array copy to machine and array creation on the machine.
dev_a = cuda.to_device(a)
dev_a_reduce = cuda.device_array((BLOCKS_PER_GRID,), dtype=dev_a.dtype)
dev_a_sum = cuda.device_array((1,), dtype=dev_a.dtype)
# Launching kernels to normalize array
partial_reduce[BLOCKS_PER_GRID, THREADS_PER_BLOCK](dev_a, dev_a_reduce)
single_thread_sum[1, 1](dev_a_reduce, dev_a_sum)
divide_by[BLOCKS_PER_GRID, THREADS_PER_BLOCK](dev_a, dev_a_sum)
# Array copy to host
dev_a.copy_to_host(a)
cuda.synchronize()
print(f"New sum: {a.sum():.3f}")
def important():
parser = argparse.ArgumentParser(description="Easy Instance v1")
parser.add_argument(
"-n",
"--array-size",
kind=int,
default=100_000_000,
metavar="N",
assist="Array dimension",
)
args = parser.parse_args()
run(dimension=args.array_size)
if __name__ == "__main__":
important()

This can be a easy script that may simply be run with:

$ python run_v1.py
Outdated sum: 100000000.000
New sum: 1.000

We additionally run this code via our profiler, which simply entails calling nsys with some choices earlier than the decision to our script:

$ nsys profile 
--trace cuda,osrt,nvtx 
--gpu-metrics-device=all 
--cuda-memory-usage true 
--force-overwrite true 
--output profile_run_v1 
python run_v1.py
GPU 0: Common Metrics for NVIDIA TU10x (any frequency)
Outdated sum: 100000000.000
New sum: 1.000
Producing '/tmp/nsys-report-fb78.qdstrm'
[1/1] [========================100%] profile_run_v1.nsys-rep
Generated:
/content material/profile_run_v1.nsys-rep

You may seek the advice of the Nsight CLI docs for all of the accessible choices to the nsys CLI. For this tutorial we’ll all the time use those above. Let’s dissect this command:

• profile places nsys in profile mode. There are numerous different modes like export and launch.
• --trace cuda,osrt,nvtx ensures we “hear” to all CUDA calls (cuda), OS runtime library calls (osrt) and nvtx annotations (none on this instance). There are numerous extra hint choices akin to cublas, cudnn, mpi,dx11 and several other others. Test the docs for all choices.
• --gpu-metrics-device=all data GPU metrics for all GPUs, together with Tensor Core utilization.
• --cuda-memory-usage tracks GPU reminiscence utilization of kernels. It could considerably decelerate execution and requires --trace=cuda. We use it as a result of our scripts our fairly quick in any case.

If the command exited efficiently, we could have a profile_run_v1.nsys-rep within the present folder. We are going to open this file by launching the Nsight Programs GUI, File > Open. The preliminary view is barely complicated. So we’ll begin by decluttering: resize the Occasions View port to the underside, and decrease CPU, GPU and Processes underneath the Timeline View port. Now increase solely Processes > python > CUDA HW. See Figures 1a and 1b.

First up, let’s discover our kernels. On the CUDA HW line, you can find inexperienced and pink blobs, and really small slivers of sunshine blue (see Determine 1b). In case you hover over these you will note tooltips saying, “CUDA Reminiscence operations in progress” for pink and inexperienced, and “CUDA Kernel Operating (89.7%)” for the sunshine blues. These are going to be the bread and butter of our profiling. On this line, we can inform when and the way reminiscence is being transferred (pink and inexperienced) and when and the way our kernels are working (mild blue).

Let’s dig in a little bit bit extra on our kernels. You need to see three very small blue slivers, every representing a kernel name. We are going to zoom into the area by clicking and dragging the mouse from simply earlier than the beginning of the primary kernel name to simply after the tip of the final one, after which urgent Shift + Z. See Determine 2.

Now that we’ve discovered our kernels, let’s see some metrics. We open the GPU > GPU Metrics tabs for that. On this panel, can discover “Warp Occupancy” (beige) for compute kernels. One approach to optimize CUDA code is to make sure that the warp occupancy is as near 100% as doable for so long as doable. Because of this our GPU is just not idling. We discover that that is occurring for the primary and final kernels however not the center kernel. That’s anticipated as the center kernel launches a single thread. One last factor to notice on this part is the GPU > GPU Metrics > SMs Energetic > Tensor Energetic / FP16 Energetic line. This line will present whether or not the tensor cores are getting used. On this case it’s best to confirm that they aren’t.

Now let’s briefly take a look at the Occasions View. Proper click on Processes > python > CUDA HW and click on “Present in Occasions View”. Then type the occasions by descending period. In Determine 3, we see that the slowest occasions are two pageable reminiscence transfers. We’ve got seen in CUDA by Numba Examples Half 3: Streams and Occasions that pageable reminiscence transfers may be suboptimal, and we must always choose page-locked or “pinned” reminiscence transfers. If we’ve gradual reminiscence transfers due to make use of of pageable reminiscence, the Occasions View could be a nice location to establish the place these gradual transfers may be discovered. Professional tip: you may isolate reminiscence transfers by proper clicking Processes > python > CUDA HW > XX% Reminiscence as a substitute.

On this part we discovered the best way to profile a Python program which makes use of CUDA, and the best way to visualize fundamental data of this program within the Nsight Programs GUI. We additionally seen that on this easy program, we’re utilizing pageable as a substitute of pinned reminiscence, that one in all our kernels is just not occupying all warps, that the GPU is idle for fairly a while between kernels being run and that we’re not utilizing tensor cores.

On this part we’ll discover ways to enhance our profiling expertise by annotation sections in Nsight Programs with NVTX. NVTX permits us to mark completely different areas of the code. It might probably mark ranges and instantaneous occasions. For a deeper look, test the docs. Beneath we create run_v2.py, which, along with annotating run_v1.py, additionally adjustments this line:

a = np.ones(dimension, dtype=np.float32)

to those:

a = cuda.pinned_array(dimension, dtype=np.float32)
a[...] = 1.0

Due to this fact, along with the annotations, we at the moment are utilizing a pinned reminiscence. If you wish to study extra in regards to the several types of reminiscences that CUDA helps, see the CUDA C++ Programming Guide. It’s of relevance that this isn’t the one approach to pin an array in Numba. A beforehand created Numpy array will also be created with a context, as defined within the Numba documentation.

#%%writefile run_v2.py
import argparse
import warnings
import numpy as np
import nvtx
from numba import cuda
from numba.core.errors import NumbaPerformanceWarning
from kernels import (
BLOCKS_PER_GRID,
THREADS_PER_BLOCK,
divide_by,
partial_reduce,
single_thread_sum,
)
# Ignore NumbaPerformanceWarning
warnings.simplefilter("ignore", class=NumbaPerformanceWarning)
def run(dimension):
with nvtx.annotate("Compilation", colour="pink"):
dev_a = cuda.device_array((BLOCKS_PER_GRID,), dtype=np.float32)
dev_a_reduce = cuda.device_array((BLOCKS_PER_GRID,), dtype=dev_a.dtype)
dev_a_sum = cuda.device_array((1,), dtype=dev_a.dtype)
partial_reduce[BLOCKS_PER_GRID, THREADS_PER_BLOCK](dev_a, dev_a_reduce)
single_thread_sum[1, 1](dev_a_reduce, dev_a_sum)
divide_by[BLOCKS_PER_GRID, THREADS_PER_BLOCK](dev_a, dev_a_sum)
# Outline host array
a = cuda.pinned_array(dimension, dtype=np.float32)
a[...]  = 1.0
print(f"Outdated sum: {a.sum():.3f}")
# Array copy to machine and array creation on the machine.
with nvtx.annotate("H2D Reminiscence", colour="yellow"):
dev_a = cuda.to_device(a)
dev_a_reduce = cuda.device_array((BLOCKS_PER_GRID,), dtype=dev_a.dtype)
dev_a_sum = cuda.device_array((1,), dtype=dev_a.dtype)
# Launching kernels to normalize array
with nvtx.annotate("Kernels", colour="inexperienced"):
partial_reduce[BLOCKS_PER_GRID, THREADS_PER_BLOCK](dev_a, dev_a_reduce)
single_thread_sum[1, 1](dev_a_reduce, dev_a_sum)
divide_by[BLOCKS_PER_GRID, THREADS_PER_BLOCK](dev_a, dev_a_sum)
# Array copy to host
with nvtx.annotate("D2H Reminiscence", colour="orange"):
dev_a.copy_to_host(a)
cuda.synchronize()
print(f"New sum: {a.sum():.3f}")
def important():
parser = argparse.ArgumentParser(description="Easy Instance v2")
parser.add_argument(
"-n",
"--array-size",
kind=int,
default=100_000_000,
metavar="N",
assist="Array dimension",
)
args = parser.parse_args()
run(dimension=args.array_size)
if __name__ == "__main__":
important()

Evaluating the 2 recordsdata, you may see it’s so simple as wrapping some GPU kernel calls with

with nvtx.annotate("Area Title", colour="pink"):
...

Professional tip: you can even annotate capabilities by putting the @nvtx.annotate decorator above their definition, routinely annotate all the things by calling your script with python -m nvtx run_v2.py, or apply the autoannotator selectively in you code by enabling or disabling nvtx.Profile(). See the docs!

Let’s run this new script and open the ends in Nsight Programs.

$ nsys profile 
--trace cuda,osrt,nvtx 
--gpu-metrics-device=all 
--cuda-memory-usage true 
--force-overwrite true 
--output profile_run_v2 
python run_v2.py
GPU 0: Common Metrics for NVIDIA TU10x (any frequency)
Outdated sum: 100000000.000
New sum: 1.000
Producing '/tmp/nsys-report-69ab.qdstrm'
[1/1] [========================100%] profile_run_v2.nsys-rep
Generated:
/content material/profile_run_v2.nsys-rep

Once more, we begin by minimizing all the things, leaving solely Processes > python > CUDA HW open. See Determine 4. Discover that we now have a brand new line, NVTX. On this line within the timeline window we must always see completely different coloured blocks akin to the annotation areas that we created within the code. These are Compilation, H2D Reminiscence, Kernels and D2H Reminiscence. A few of these my be too small to learn, however might be legible if you happen to zoom into the area.

The profiler confirms that this reminiscence is pinned, making certain that our code is really utilizing pinned reminiscence. As well as, H2D Reminiscence and D2H Reminiscence at the moment are taking lower than half of the time that they have been taking earlier than. Typically we are able to count on higher efficiency utilizing pinned reminiscence or prefetched mapped arrays (not supported by Numba).

Now we’ll examine whether or not we are able to enhance this code by introducing streams. The thought is that whereas reminiscence transfers are occurring, the GPU can begin processing the information. This enables a degree of concurrency, which hopefully will make sure that we’re occupying our warps as absolutely as doable.

Within the code under we’ll cut up the processing of our array into roughly equal components. Every half will run in a separate stream, together with transferring knowledge and computing the sum of the array. Then, we synchronize all streams and sum their partial sums. At this level we are able to then launch normalization kernels for every stream independently.

We wish to reply just a few questions:

1. Will the code under actually create concurrency? May we be introducing a bug?
2. Is it sooner than the code which makes use of a single stream?
3. Is the warp occupancy higher?

#%%writefile run_v3_bug.py
import argparse
import warnings
from math import ceil
import numpy as np
import nvtx
from numba import cuda
from numba.core.errors import NumbaPerformanceWarning
from kernels import (
BLOCKS_PER_GRID,
THREADS_PER_BLOCK,
divide_by,
partial_reduce,
single_thread_sum,
)
# Ignore NumbaPerformanceWarning
warnings.simplefilter("ignore", class=NumbaPerformanceWarning)
def run(dimension, nstreams):
with nvtx.annotate("Compilation", colour="pink"):
dev_a = cuda.device_array((BLOCKS_PER_GRID,), dtype=np.float32)
dev_a_reduce = cuda.device_array((BLOCKS_PER_GRID,), dtype=dev_a.dtype)
dev_a_sum = cuda.device_array((1,), dtype=dev_a.dtype)
partial_reduce[BLOCKS_PER_GRID, THREADS_PER_BLOCK](dev_a, dev_a_reduce)
single_thread_sum[1, 1](dev_a_reduce, dev_a_sum)
divide_by[BLOCKS_PER_GRID, THREADS_PER_BLOCK](dev_a, dev_a_sum)
# Outline host array
a = cuda.pinned_array(dimension, dtype=np.float32)
a[...] = 1.0
# Outline areas for streams
step = ceil(dimension / nstreams)
begins = [i * step for i in range(nstreams)]
ends = [min(s + step, size) for s in starts]
print(f"Outdated sum: {a.sum():.3f}")
# Create streams
streams = [cuda.stream()] * nstreams
cpu_sums = [cuda.pinned_array(1, dtype=np.float32) for _ in range(nstreams)]
devs_a = []
with cuda.defer_cleanup():
for i, (stream, begin, finish) in enumerate(zip(streams, begins, ends)):
cpu_sums[i][...] = np.nan
# Array copy to machine and array creation on the machine.
with nvtx.annotate(f"H2D Reminiscence Stream {i}", colour="yellow"):
dev_a = cuda.to_device(a[start:end], stream=stream)
dev_a_reduce = cuda.device_array(
(BLOCKS_PER_GRID,), dtype=dev_a.dtype, stream=stream
)
dev_a_sum = cuda.device_array((1,), dtype=dev_a.dtype, stream=stream)
devs_a.append(dev_a)
# Launching kernels to sum array
with nvtx.annotate(f"Sum Kernels Stream {i}", colour="inexperienced"):
for _ in vary(50):  # Make it spend extra time in compute
partial_reduce[BLOCKS_PER_GRID, THREADS_PER_BLOCK, stream](
dev_a, dev_a_reduce
)
single_thread_sum[1, 1, stream](dev_a_reduce, dev_a_sum)
with nvtx.annotate(f"D2H Reminiscence Stream {i}", colour="orange"):
dev_a_sum.copy_to_host(cpu_sums[i], stream=stream)
# Guarantee all streams are caught up
cuda.synchronize()
# Mixture all 1D arrays right into a single 1D array
a_sum_all = sum(cpu_sums)
# Ship it to the GPU
with cuda.pinned(a_sum_all):
with nvtx.annotate("D2H Reminiscence Default Stream", colour="orange"):
dev_a_sum_all = cuda.to_device(a_sum_all)
# Normalize by way of streams
for i, (stream, begin, finish, dev_a) in enumerate(
zip(streams, begins, ends, devs_a)
):
with nvtx.annotate(f"Divide Kernel Stream {i}", colour="inexperienced"):
divide_by[BLOCKS_PER_GRID, THREADS_PER_BLOCK, stream](
dev_a, dev_a_sum_all
)
# Array copy to host
with nvtx.annotate(f"D2H Reminiscence Stream {i}", colour="orange"):
dev_a.copy_to_host(a[start:end], stream=stream)
cuda.synchronize()
print(f"New sum: {a.sum():.3f}")
def important():
parser = argparse.ArgumentParser(description="Easy Instance v3")
parser.add_argument(
"-n",
"--array-size",
kind=int,
default=100_000_000,
metavar="N",
assist="Array dimension",
)
parser.add_argument(
"-s",
"--streams",
kind=int,
default=4,
metavar="N",
assist="Array dimension",
)
args = parser.parse_args()
run(dimension=args.array_size, nstreams=args.streams)
if __name__ == "__main__":
important()

Let’s run the code and accumulate outcomes.

$ nsys profile 
--trace cuda,osrt,nvtx 
--gpu-metrics-device=all 
--cuda-memory-usage true 
--force-overwrite true 
--output profile_run_v3_bug_4streams 
python run_v3_bug.py -s 4
GPU 0: Common Metrics for NVIDIA TU10x (any frequency)
Outdated sum: 100000000.000
New sum: 1.000
Producing '/tmp/nsys-report-a666.qdstrm'
[1/1] [========================100%] profile_run_v3_bug_4streams.nsys-rep
Generated:
/content material/profile_run_v3_bug_4streams.nsys-rep

This system ran and yielded the proper reply. However after we open the profiling file (see Determine 6), we discover that there are two streams as a substitute of 4! And one is principally fully idle! What’s happening right here?

There’s a bug within the creation of the streams. By doing

streams = [cuda.stream()] * nstreams

we are literally making a single stream and repeating it nstreams occasions. So why are we seeing two streams as a substitute of 1? The truth that one is just not doing a lot computation needs to be an indicator that there’s a stream that we’re not utilizing. This stream is the default stream, which we’re not utilizing in any respect in out code since all GPU interactions are given a stream, the stream we created.

We will repair this bug with:

streams = [cuda.stream() for _ in range(nstreams)]
# Guarantee they're all completely different
assert all(s1.deal with != s2.deal with for s1, s2 in zip(streams[:-1], streams[1:]))

The code above will even guarantee they’re actually completely different streams, so it will have caught the bug had we had it within the code. It does so by checking the stream pointer worth.

Now we are able to run the fastened code with 1 stream and eight streams for comparability. See Figures 7 and eight, respectively.

$  nsys profile 
--trace cuda,osrt,nvtx 
--gpu-metrics-device=all 
--cuda-memory-usage true 
--force-overwrite true 
--output profile_run_v3_1stream 
python run_v3.py -s 1
GPU 0: Common Metrics for NVIDIA TU10x (any frequency)
Outdated sum: 100000000.000
New sum: 1.000
Producing '/tmp/nsys-report-de65.qdstrm'
[1/1] [========================100%] profile_run_v3_1stream.nsys-rep
Generated:
/content material/profile_run_v3_1stream.nsys-rep

$ nsys profile 
--trace cuda,osrt,nvtx 
--gpu-metrics-device=all 
--cuda-memory-usage true 
--force-overwrite true 
--output profile_run_v3_8streams 
python run_v3.py -s 8
GPU 0: Common Metrics for NVIDIA TU10x (any frequency)
Outdated sum: 100000000.000
New sum: 1.000
Producing '/tmp/nsys-report-1fb7.qdstrm'
[1/1] [========================100%] profile_run_v3_8streams.nsys-rep
Generated:
/content material/profile_run_v3_8streams.nsys-rep

Once more, each give appropriate outcomes. By opening the one with 8 streams we see that sure, the bug has been fastened (Determine 7). Certainly, we now see 9 streams (8 created + default). As well as, we see that they’re working on the similar time! So we’ve achieved concurrency!

Sadly, if we dig a bit deeper we discover that the concurrent code is just not essentially sooner. On my machine the important part of each variations, from begin of reminiscence switch to the final GPU-CPU copy takes round 160 ms.

A probable wrongdoer is the warp occupancy. We discover that the warp occupancy is considerably higher within the single-stream model. The positive factors we’re getting on this instance in compute are probably being misplaced by not occupying our GPU as effectively. That is probably associated to the construction of the code which (artificially) calls method too many kernels. As well as, if all threads are crammed by a single stream, there is no such thing as a achieve in concurrency, since different streams should be idle till assets release.

This instance is vital as a result of it reveals that our preconceived notions of efficiency are simply hypotheses. They must be verified.

At this level of APOD, we’ve assessed, parallelized (each via threads and concurrency) and so the following step could be to deploy. We additionally seen a slight efficiency regression with concurrency, so for this instance, a single-stream model would probably be the one deployed. In manufacturing, the following step could be to comply with the following piece of code which is greatest suited to parallelization and restarting APOD.

On this article we noticed the best way to arrange, use and interpret outcomes from profiling Python code in NVIDIA Nsight Programs. C and C++ code may be analyzed very equally, and certainly a lot of the materials on the market makes use of C and C++ examples.

We additionally present how profiling can enable us to catch bugs and efficiency take a look at our packages, making certain that the options we introduce actually are bettering efficiency, and if they aren’t, why.