Analyzing the performance of deep neural networks is hard. Getting kernels out of NVProf or NSight Compute provides some generic kernel name and its execution time, but not detailed information regarding the following:
- Which layer launched it: e.g. the association of
ComputeOffsetsKernelwith a concrete PyTorch layer or API is not obvious. - What the tensor dimensions and precision were: without knowing the tensor dimensions and precision, it's impossible to reason about whether the actual (silicon) kernel time is close to maximum performance of such a kernel on the GPU. Knowing the tensor dimensions and precision, we can figure out the FLOPs and bandwidth required by a layer, and then determine how close to maximum performance the kernel is for that operation.
- Forward-backward correlation: currently it's very hard to determine what the forward pass step was that resulted in the particular weight and data gradients (wgrad, dgrad), which makes it difficult to determine the tensor dimensions required by these backprop steps to assess their performance.
- Did the kernel use Tensor Cores?
- Which line in the user's code resulted in launching this particular kernel (program trace)?
PyProf addresses all of the issues above by:
- Instrumenting PyTorch operations to capture the tensor dimensions and precision using NVTX. This information is recorded at profile capture time, e.g. using NvProf.
- Querying the record produced by the profiler to correlate the kernel name and duration with PyTorch API/layer name, tensor dimensions, tensor precision, as well as calculating FLOPs and bandwidth for common operations. In addition, extra information from the profile is added for use by CUDA professionals, such as CUDA launch parameters (block/grid dimensions).
Regarding FLOP and bandwidth implementations, these are usually quite straightforward. For example, for matrices AMxK and BKxN, the FLOP count for a matrix multiplication is 2 * M * N * K, and bandwidth is M * K + N * K + M * N. Note that these numbers are based on the algorithm, not the actual performance of the specific kernel. For more details, see NVIDIA's Deep Learning Performance Guide.
Armed with such information, the user can determine various issues to help them tune the network. For instance, according to the Tensor Core Performance Guide, the M, N and K dimensions that result in Tensor Core usage need to be divisible by 8. In fact, PyProf comes with a flag that lets the user obtain information regarding whether Tensor Cores were used by the kernel. Other useful information might include knowing that a particular kernel did not exploit much thread parallelism, as determined by the grid/block dimensions. Since many PyTorch kernels are open-source (or even custom written by the user, as in CUDA Extensions), this provides the user with information that helps root cause performance issues and prioritize optimization work.
```sh
pip3 install . --user
pip3 uninstall pyprof2
```
-
Add the following lines to your PyTorch network:
import torch.cuda.profiler as profiler import pyprof2 pyprof2.init()
Run the training/inference loop with the PyTorch's NVTX context manager
with torch.autograd.profiler.emit_nvtx(). Optionally, you can useprofiler.start()andprofiler.stop()to pick an iteration (say after warm-up) for which you would like to capture data. Here's an example:iters = 500 iter_to_capture = 100 # Define network, loss function, optimizer etc. # PyTorch NVTX context manager with torch.autograd.profiler.emit_nvtx(): for iter in range(iters): if iter == iter_to_capture: profiler.start() output = net(images) loss = criterion(output, labels) loss.backward() optimizer.step() if iter == iter_to_capture: profiler.stop()
-
Run NVprof to generate a SQL (NVVP) file. This file can be opened with NVVP, as usual.
# If you used profiler.start() and profiler.stop() in net.py nvprof -f -o net.sql --profile-from-start off -- python net.py # Profile everything nvprof -f -o net.sql -- python net.py
Note: if you're experiencing issues with hardware counters and you get a message such as **_ERR_NVGPUCTRPERM The user running <tool_name/application_name> does not have permission to access NVIDIA GPU Performance Counters on the target device_**, please follow the steps described in Hardware Counters.
-
Run parser on the SQL file. The output is an ASCII file. Each line is a python dictionary which contains information about the kernel name, duration, parameters etc. This file can be used as input to other custom scripts as well.
pyprof2/parse/parse.py net.sql > net.dict -
Run the profiler. The input is the python dictionary created above. The tool can produce a CSV output, a columnated output (similar to
column -tfor terminal readability) and a space separated output (for post processing by AWK for instance). The tool produces 20 columns of information for every GPU kernel but you can select a subset of columns using the-cflag. Note that a few columns might have the value "na" implying either its a work in progress or the tool was unable to extract that information. Assuming the directory ispyprof2/prof, here are a few examples of how to useprof.py.# Print usage and help. Lists all available output columns. prof.py -h # Columnated output of width 150 with some default columns. prof.py -w 150 net.dict # CSV output. prof.py --csv net.dict # Space seperated output. prof.py net.dict # Columnated output of width 130 with columns index,direction,kernel name,parameters,silicon time. prof.py -w 130 -c idx,dir,kernel,params,sil net.dict # CSV output with columns index,direction,kernel name,parameters,silicon time. prof.py --csv -c idx,dir,kernel,params,sil net.dict # Space separated output with columns index,direction,kernel name,parameters,silicon time. prof.py -c idx,dir,kernel,params,sil net.dict # Input redirection. prof.py < net.dict
-
Profile-guided optimization
If kernels that do matrix multiplication/GEMM or convolution use half precision (fp16) data but do not use Tensor Cores (the TC column in the profile analysis output doesn't show a "1"), one can follow some basic steps to increase the likelihood that a Tensor Core-compatible kernel will be chosen. For example, for GEMMs, M, N and K should be divisible by 8, and for convolutions, the number of input and output channels shuold be divisible by 8. For more information, see detailed Tensor Core guides such as:
- Blog Post: Tips for Optimizing GPU Performance Using Tensor Cores
- GTC Talk: Tensor Core Deep Learning Performance Guide
For both Tensor Core and non-Tensor Core Deep Learning performance optimization tips, see NVIDIA's Deep Learning Performance Guide.
- The support for conv transpose is currently missing.
- PyProf currently works only with NvProf, but Nsight Compute support will be added in the future.
- Run
nvprofon the LeNet model inexamples/lenet.py. This will output a SQL file callednet.sql.
nvprof -f -o net.sql --profile-from-start off -- python examples/lenet.pyNote: DO NOT add --analysis-metrics since that will change which table nvprof writes the kernels to (CUPTI_ACTIVITY_KIND_KERNEL instead of the usual CUPTI_ACTIVITY_KIND_CONCURRENT_KERNEL). Support for running with metrics may be added in the future.
If you don't care about a full correlation analysis and you'd just like to view the timeline with detailed NVTX annotations, you can do so, e.g. in the NVIDIA Visual Profiler (NVVP). For example, you can call nvvp net.sql to view the annotated timeline.
- Run the
parse.pyscript onnet.sqlto extract kernel and runtime information and save it asnet.dict.
pyprof2/parse/parse.py net.sql > net.dictThis will produce a text file, which can be parsed by any external tool, but it can also be directly read one line at a time by Python by calling eval on the line being read.
Note: you do not need to process this output manually. Here the output is just shown as an example of modularity - you can process the raw data yourself, or let the next step enrich the information further and dump a CSV.
The output of this step will look as follows. Note that the dictionary has a lot more keys than the ones shown in the example.
>>> with open('torchvision.resnet50.adam.64.dict') as f:
... for line in f:
... d = eval(line)
... print(d['kShortName'], d['op'], d['kDuration'], d['block'], d['grid'], d['device'], d['stream'], d['trace'])
...
nchwToNhwc3To4Kernel ['conv2d'] 376324 (256, 1, 1) (1568, 1, 64) 0 7 ['imagenet.py:137', 'imagenet.py:129', '/opt/conda/lib/python3.6/site-packages/torchvision/models/resnet.py:195']
generic4Channel_kernel ['conv2d'] 10720 (512, 1, 1) (19, 1, 1) 0 7 ['imagenet.py:137', 'imagenet.py:129', '/opt/conda/lib/python3.6/site-packages/torchvision/models/resnet.py:195']
first_layer_fwd_kernel ['conv2d'] 411204 (128, 1, 1) (2, 7, 64) 0 7 ['imagenet.py:137', 'imagenet.py:129', '/opt/conda/lib/python3.6/site-packages/torchvision/models/resnet.py:195']
nhwcToNchwKernel ['conv2d'] 342371 (256, 1, 1) (392, 2, 64) 0 7 ['imagenet.py:137', 'imagenet.py:129', '/opt/conda/lib/python3.6/site-packages/torchvision/models/resnet.py:195']
elementwise_kernel ['__iadd__'] 2816 (128, 1, 1) (1, 1, 1) 0 7 ['imagenet.py:137', 'imagenet.py:129', '/opt/conda/lib/python3.6/site-packages/torchvision/models/resnet.py:196']
batch_norm_collect_statistics_kernel ['batch_norm', 'batch_norm'] 929513 (512, 1, 1) (64, 1, 1) 0 7 ['imagenet.py:137', 'imagenet.py:129', '/opt/conda/lib/python3.6/site-packages/torchvision/models/resnet.py:196']
- Run the
prof.pyscript onnet.dictto summarize the results into a CSV file, or to display the pretty-printed results on the screen. This step processes the raw output from step 2 to generate a nice output, but it also adds a lot of extra useful information inferred from the previous step, such as:
- FLOPs
- bandwidth (bytes in and out of GPU DRAM)
- tensor core usage
pyprof2/prof/prof.py --csv net.dict > results.csvYou can choose which columns you'd like to display. Here's a list from calling pyprof2/prof/prof.py -h:
idx: Index
seq: PyTorch Sequence Id
altseq: PyTorch Alternate Sequence Id
tid: Thread Id
layer: User annotated NVTX string (can be nested)
trace: Function Call Trace
dir: Direction
sub: Sub Sequence Id
mod: Module
op: Operattion
kernel: Kernel Name
params: Parameters
sil: Silicon Time (in ns)
tc: Tensor Core Usage
device: GPU Device Id
stream: Stream Id
grid: Grid Dimensions
block: Block Dimensions
flops: Floating point ops (FMA = 2 FLOPs)
bytes: Number of bytes in and out of DRAM
Let's have a look at the pretty-printed output:
pyprof2/prof/prof.py -w 100 -c kernel,op,sil,tc,flops,bytes,device,stream,block,grid torchvision.resnet50.adam.64.dict
Kernel Op Sil(ns) TC FLOPs Bytes Dev Str Block Grid
elementwise_kernel relu 381028 - 51380224 205520896 0 7 512,1,1 100352,1,1
volta_fp16_s884cudn conv2d 160002 1 1644167168 51388416 0 7 256,1,1 784,1,1
elementwise_kernel relu 96545 - 12845056 51380224 0 7 512,1,1 25088,1,1
volta_fp16_s884cudn conv2d 346083 1 6576668672 128483328 0 7 256,1,1 784,2,1
Not using the pretty-print width (-w) option and adding --csv results in a CSV output instead:
pyprof2/prof/prof.py --csv -c kernel,mod,op,dir,sil,tc,flops,bytes,device,stream,block,grid torchvision.resnet50.adam.64.dict
"Kernel","Module","Op","Direction","Sil(ns)","TC","FLOPs","Bytes","Device","Stream","Block","Grid"
"nchwToNhwc3To4Kernel","torch.nn.functional","conv2d","fprop","376324","-","0","0","0","7","256,1,1","1568,1,64"
"generic4Channel_kernel","torch.nn.functional","conv2d","fprop","10720","-","0","0","0","7","512,1,1","19,1,1"
"first_layer_fwd_kernel","torch.nn.functional","conv2d","fprop","411204","-","0","0","0","7","128,1,1","2,7,64"
"nhwcToNchwKernel","torch.nn.functional","conv2d","fprop","342371","-","0","0","0","7","256,1,1","392,2,64"
"elementwise_kernel","Tensor","__iadd__","fprop","2816","-","1.0","8","0","7","128,1,1","1,1,1"
"batch_norm_collect_statistics_kernel","torch.nn.functional","batch_norm","fprop","929513","-","411041792","411041792","0","7","512,1,1","64,1,1"
"batch_norm_transform_input_kernel","torch.nn.functional","batch_norm","fprop","377539","-","411041792","411041792","0","7","512,1,1","64,64,1"
"elementwise_kernel","torch.nn.functional","relu","fprop","381028","-","51380224","205520896","0","7","512,1,1","100352,1,1"
"MaxPoolForward","torch.nn.functional","max_pool2d","fprop","406531","-","0","0","0","7","256,1,1","50176,1,1"
"cudnn::gemm::computeOffsetsKernel","torch.nn.functional","conv2d","fprop","2464","-","0","0","0","7","128,1,1","25,1,1"
Profiling GPU workloads may require access to hardware performance counters. Due to a fix in recent NVIDIA drivers addressing CVE‑2018‑6260, the hardware counters are disabled by default, and require elevated privileges to be enabled again. If you're using a recent driver, you may see the following message when trying to run nvprof:
**_ERR_NVGPUCTRPERM The user running <tool_name/application_name> does not have permission to access NVIDIA GPU Performance Counters on the target device._**
For details, see here.
Permanent solution
Follow the steps here. The current steps for Linux are:
sudo systemctl isolate multi-user
sudo modprobe -r nvidia_uvm nvidia_drm nvidia_modeset nvidia-vgpu-vfio nvidia
sudo modprobe nvidia NVreg_RestrictProfilingToAdminUsers=0
sudo systemctl isolate graphical
The above steps should result in a permanent change.
Temporary solution
When running on bare metal, you can run nvprof with sudo.
If you're running in a Docker image, you can temporarily elevate your privileges with one of the following (oldest to newest syntax):
nvidia-docker run --privileged docker run --runtime nvidia --privileged docker run --gpus all --privileged