apex-master (1).zip

## PyProf - PyTorch Profiling tool ### What does this tool do? Analyzing the performance of deep neural networks is hard. Getting kernels out of [NvProf]([https://developer.nvidia.com/nvidia-visual-profiler](https://developer.nvidia.com/nvidia-visual-profiler)) or [NSight Compute]([https://developer.nvidia.com/nsight-compute](https://developer.nvidia.com/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 `ComputeOffsetsKernel` with 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]([https://www.youtube.com/watch?v=yyR0ZoCeBO8](https://www.youtube.com/watch?v=yyR0ZoCeBO8))? - Which line in the user's code resulted in launching this particular kernel (program trace)? PyProf addresses all of the issues above by: 1. Instrumenting PyTorch operations to capture the tensor dimensions and precision using [NVTX](https://devblogs.nvidia.com/cuda-pro-tip-generate-custom-application-profile-timelines-nvtx). This information is recorded at profile capture time, e.g. using [NvProf](https://developer.nvidia.com/nvidia-visual-profiler). 2. 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 A_MxK and B_KxN, 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](https://docs.nvidia.com/deeplearning/sdk/dl-performance-guide/index.html). 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]([https://docs.nvidia.com/deeplearning/sdk/dl-performance-guide/index.html](https://docs.nvidia.com/deeplearning/sdk/dl-performance-guide/index.html)), 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]([https://pytorch.org/tutorials/advanced/cpp_extension.html](https://pytorch.org/tutorials/advanced/cpp_extension.html))), this provides the user with information that helps root cause performance issues and prioritize optimization work. ### How to get started? 1. Add the following lines to your PyTorch network: ```python import torch.cuda.profiler as profiler from apex import pyprof pyprof.nvtx.init() ``` Run the training/inference loop with the [PyTorch's NVTX context manager](https://pytorch.org/docs/stable/_modules/torch/autograd/profiler.html#emit_nvtx) `with torch.autograd.profiler.emit_nvtx()`. Optionally, you can use `profiler.start()` and `profiler.stop()` to pick an iteration (say after warm-up) for which you would like to capture data. Here's an example: ```python 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() ``` 2. Run NVprof to generate a SQL (NVVP) file. This file can be opened with NVVP, as usual. ```sh # 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](#hardware-counters). 3. 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. ```sh python -m apex.pyprof.parse net.sql > net.dict ``` 4. Run the profiler. The input is the python dictionary created above. The tool can produce a CSV output, a columnated output (similar to `column -t` for 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 `-c` flag. 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 is `prof`, here are a few examples of how to use `prof.py`. ```sh # Print usage and help. Lists all available output columns. python -m apex.pyprof.prof -h # Columnated output of width 150 with some default columns. python -m apex.pyprof.prof -w 150 net.dict # CSV output. python -m apex.pyprof.prof --csv net.dict # Space seperated output. python -m apex.pyprof.prof net.dict # Columnated output of width 130 with columns index,direction,kernel name,parameters,silicon time. python -m apex.pyprof.prof -w 130 -c idx,dir,kernel,params,sil net.dict # CSV output with columns index,direction,kernel name,parameters,silicon time. python -m apex.pyprof.prof --csv -c idx,dir,kernel,params,sil net.dict # Space separated output with columns index,direction,kernel name,parameters,silicon time. python -m apex.pyprof.prof -c idx,dir,kernel,params,sil net.dict # Input redirection. python -m apex.pyprof.prof < net.dict ``` 5. 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 convoluti