-
Notifications
You must be signed in to change notification settings - Fork 441
uftrace for CUDA
Kang Minchul edited this page Oct 24, 2023
·
2 revisions
This document is written by Kang Minchul
Prerequisites
nvcc compiler
$ nvcc --version
nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2022 NVIDIA Corporation
Built on Wed_Sep_21_10:33:58_PDT_2022
Cuda compilation tools, release 11.8, V11.8.89
Build cuda_11.8.r11.8/compiler.31833905_0
GPU hardware
$ nvidia-smi
Mon Oct 23 14:31:49 2023
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 520.61.05 Driver Version: 520.61.05 CUDA Version: 11.8 |
|-------------------------------+----------------------+----------------------+
| GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. |
| | | MIG M. |
|===============================+======================+======================|
| 0 NVIDIA GeForce ... On | 00000000:01:00.0 Off | N/A |
| 0% 41C P8 29W / 340W | 67MiB / 10240MiB | 0% Default |
| | | N/A |
+-------------------------------+----------------------+----------------------+
| 1 NVIDIA GeForce ... On | 00000000:81:00.0 Off | N/A |
| 0% 45C P8 27W / 340W | 5MiB / 10240MiB | 0% Default |
| | | N/A |
+-------------------------------+----------------------+----------------------+
+-----------------------------------------------------------------------------+
| Processes: |
| GPU GI CI PID Type Process name GPU Memory |
| ID ID Usage |
|=============================================================================|
| 0 N/A N/A 1233 G /usr/lib/xorg/Xorg 56MiB |
| 0 N/A N/A 1681 G /usr/bin/gnome-shell 9MiB |
| 1 N/A N/A 1233 G /usr/lib/xorg/Xorg 4MiB |
+-----------------------------------------------------------------------------+
1. Compile Cuda Code with -pg
option
$ cat hellocu.cu
#include "cuda_runtime.h"
#include "device_launch_parameters.h"
#include <stdio.h>
__global__ void helloCUDA(void)
{
printf("Hello CUDA from GPU!\n");
}
int main(void)
{
printf("Hello GPU from CPU!\n");
helloCUDA << <1, 10 >> > ();
cudaDeviceSynchronize();
return 0;
}
$ nvcc -pg hellocu.cu -o hellocu
2. Record the compilied binary with uftrace
$ uftrace record hellocu
Hello GPU from CPU!
Hello CUDA from GPU!
Hello CUDA from GPU!
Hello CUDA from GPU!
Hello CUDA from GPU!
Hello CUDA from GPU!
Hello CUDA from GPU!
Hello CUDA from GPU!
Hello CUDA from GPU!
Hello CUDA from GPU!
Hello CUDA from GPU!
3. replay the recorded data
$ uftrace replay
# DURATION TID FUNCTION
[ 6293] | pthread_once() {
0.520 us [ 6293] | malloc();
0.280 us [ 6293] | pthread_mutexattr_init();
0.130 us [ 6293] | pthread_mutexattr_settype();
0.140 us [ 6293] | pthread_mutexattr_setpshared();
0.450 us [ 6293] | pthread_mutex_init();
0.150 us [ 6293] | pthread_mutexattr_destroy();
0.260 us [ 6293] | __cxa_atexit();
9.651 us [ 6293] | } /* pthread_once */
0.170 us [ 6293] | __cxa_atexit();
[ 6293] | __sti____cudaRegisterAll() {
0.150 us [ 6293] | pthread_once();
0.220 us [ 6293] | malloc();
[ 6293] | __nv_cudaEntityRegisterCallback() {
0.140 us [ 6293] | __nv_save_fatbinhandle_for_managed_rt();
0.100 us [ 6293] | pthread_once();
0.170 us [ 6293] | malloc();
1.520 us [ 6293] | } /* __nv_cudaEntityRegisterCallback */
0.100 us [ 6293] | pthread_once();
0.150 us [ 6293] | pthread_mutex_lock();
0.300 us [ 6293] | calloc();
0.160 us [ 6293] | free();
0.120 us [ 6293] | malloc();
0.210 us [ 6293] | pthread_mutex_unlock();
0.140 us [ 6293] | __cxa_atexit();
14.750 us [ 6293] | } /* __sti____cudaRegisterAll */
[ 6293] | main() {
9.801 us [ 6293] | puts();
0.090 us [ 6293] | dim3::dim3();
0.130 us [ 6293] | dim3::dim3();
0.110 us [ 6293] | pthread_once();
[ 6293] | helloCUDA() {
[ 6293] | __device_stub__Z9helloCUDAv() {
0.120 us [ 6293] | pthread_once();
[ 6293] | cudaLaunchKernel() {
...
Dynamic Tracing the CUDA-Samples
1. Get cuda-samples
$ git clone https://github.com/NVIDIA/cuda-samples.git
$ cd cuda-samples
$ git checkout v11.8 # checkout to installed CUDA version
2. make cuda-samples
$ sudo make
3. check the binary in cuda-samples/bin/[arch]/linux/release
cuda-samples/bin/x86_64/linux/release$ ls
alignedTypes conjugateGradientCudaGraphs dwtHaar1D matrixMulDrv radixSortThrust simpleCUFFT_MGPU sortingNetworks
asyncAPI conjugateGradientMultiBlockCG dxtc matrixMulDynlinkJIT reduction simpleDrvRuntime stereoDisparity
bandwidthTest conjugateGradientMultiDeviceCG eigenvalues matrixMul_kernel64.fatbin reductionMultiBlockCG simpleHyperQ streamOrderedAllocation
batchCUBLAS conjugateGradientPrecond fastWalshTransform matrixMul_nvrtc scalarProd simpleIPC streamOrderedAllocationIPC
batchedLabelMarkersAndLabelCompressionNPP conjugateGradientUM FDTD3d MC_EstimatePiInlineP scan simpleLayeredTexture streamOrderedAllocationP2P
bf16TensorCoreGemm convolutionFFT2D FilterBorderControlNPP MC_EstimatePiInlineQ segmentationTreeThrust simpleMultiCopy StreamPriorities
binaryPartitionCG convolutionSeparable fp16ScalarProduct MC_EstimatePiP shfl_scan simpleMultiGPU systemWideAtomics
binomialOptions convolutionTexture freeImageInteropNPP MC_EstimatePiQ simpleAssert simpleOccupancy template
binomialOptions_nvrtc cppIntegration globalToShmemAsyncCopy MC_SingleAsianOptionP simpleAssert_nvrtc simpleP2P tf32TensorCoreGemm
BlackScholes cppOverload graphMemoryFootprint memMapIPCDrv simpleAtomicIntrinsics simplePitchLinearTexture threadFenceReduction
BlackScholes_nvrtc cudaCompressibleMemory graphMemoryNodes memMapIpc_kernel64.ptx simpleAtomicIntrinsics_nvrtc simplePrintf threadMigration
boxFilterNPP cudaOpenMP histEqualizationNPP mergeSort simpleAttributes simpleSeparateCompilation threadMigration_kernel64.fatbin
c++11_cuda cudaTensorCoreGemm histogram MersenneTwisterGP11213 simpleAWBarrier simpleStreams topologyQuery
cannyEdgeDetectorNPP cuHook HSOpticalFlow MonteCarloMultiGPU simpleCallback simpleSurfaceWrite transpose
cdpAdvancedQuicksort cuSolverDn_LinearSolver immaTensorCoreGemm newdelete simpleCooperativeGroups simpleTemplates UnifiedMemoryPerf
cdpBezierTessellation cuSolverRf inlinePTX NV12toBGRandResize simpleCubemapTexture simpleTemplates_nvrtc UnifiedMemoryStreams
cdpQuadtree cuSolverSp_LinearSolver inlinePTX_nvrtc nvJPEG simpleCUBLAS simpleTexture vectorAdd
cdpSimplePrint cuSolverSp_LowlevelCholesky interval nvJPEG_encoder simpleCUBLAS_LU simpleTextureDrv vectorAddDrv
cdpSimpleQuicksort cuSolverSp_LowlevelQR jacobiCudaGraphs p2pBandwidthLatencyTest simpleCUBLASXT simpleTexture_kernel64.fatbin vectorAdd_kernel64.fatbin
clock dct8x8 libcuhook.so.1 ptxjit simpleCudaGraphs simpleVoteIntrinsics vectorAddMMAP
clock_nvrtc deviceQuery lineOfSight ptxjit_kernel64.ptx simpleCUFFT simpleVoteIntrinsics_nvrtc vectorAdd_nvrtc
concurrentKernels deviceQueryDrv matrixMul quasirandomGenerator simpleCUFFT_2d_MGPU simpleZeroCopy warpAggregatedAtomicsCG
conjugateGradient dmmaTensorCoreGemm matrixMulCUBLAS quasirandomGenerator_nvrtc simpleCUFFT_callback SobolQRNG watershedSegmentationNPP
4. replay one of them with full dynamic tracing
$ uftrace -P . ./simpleCallback
Starting simpleCallback
Found 2 CUDA capable GPUs
GPU[0] NVIDIA GeForce RTX 3080 supports SM 8.6, capable GPU Callback Functions
GPU[1] NVIDIA GeForce RTX 3080 supports SM 8.6, capable GPU Callback Functions
2 GPUs available to run Callback Functions
Starting 8 heterogeneous computing workloads
Total of 8 workloads finished:
Success
# DURATION TID FUNCTION
[ 31969] | __cudart426() {
[ 31969] | __cudart1635() {
[ 31969] | pthread_once() {
[ 31969] | __cudart1670() {
0.140 us [ 31969] | malloc();
[ 31969] | __cudart1332() {
0.160 us [ 31969] | pthread_mutexattr_init();
0.060 us [ 31969] | pthread_mutexattr_settype();
0.060 us [ 31969] | pthread_mutexattr_setpshared();
0.100 us [ 31969] | pthread_mutex_init();
0.060 us [ 31969] | pthread_mutexattr_destroy();
1.380 us [ 31969] | } /* __cudart1332 */
[ 31969] | atexit() {
0.110 us [ 31969] | __cxa_atexit();
0.380 us [ 31969] | } /* atexit */
2.890 us [ 31969] | } /* __cudart1670 */
3.530 us [ 31969] | } /* pthread_once */
4.530 us [ 31969] | } /* __cudart1635 */
0.060 us [ 31969] | __cxa_atexit();
5.320 us [ 31969] | } /* __cudart426 */
[ 31969] | __sti____cudaRegisterAll() {
[ 31969] | __cudaRegisterFatBinary() {
[ 31969] | __cudart667() {
[ 31969] | __cudart1635() {
0.040 us [ 31969] | pthread_once();
0.320 us [ 31969] | } /* __cudart1635 */
0.570 us [ 31969] | } /* __cudart667 */
[ 31969] | __cudart524() {
0.090 us [ 31969] | malloc();
0.340 us [ 31969] | } /* __cudart524 */
1.150 us [ 31969] | } /* __cudaRegisterFatBinary */
[ 31969] | __nv_cudaEntityRegisterCallback() {
0.120 us [ 31969] | __nv_save_fatbinhandle_for_managed_rt();
[ 31969] | __cudaRegisterFunction() {
[ 31969] | __cudart667() {
[ 31969] | __cudart1635() {
0.050 us [ 31969] | pthread_once();
0.250 us [ 31969] | } /* __cudart1635 */
0.380 us [ 31969] | } /* __cudart667 */
[ 31969] | __cudart530() {
0.060 us [ 31969] | malloc();
0.340 us [ 31969] | } /* __cudart530 */
0.940 us [ 31969] | } /* __cudaRegisterFunction */
1.330 us [ 31969] | } /* __nv_cudaEntityRegisterCallback */
[ 31969] | __cudaRegisterFatBinaryEnd() {
[ 31969] | __cudart667() {
[ 31969] | __cudart1635() {
...
- Home
- Tutorial
- Development
- Practical Use Cases
- GCC
- Clang/LLVM
- Node.js
- Chromium
- MySQL/InnoDB
- FFmpeg
- CPython
- POCO
- Telegram
- yara
- RustPython
- cURL
- bpftrace
- SpiderMonkey
- Apache HTTP Server
- GStreamer
- Squid
- TCPDUMP
- OpenCV
- Libav
- Wireshark
- LXC
- Git
- Radare2
- uftrace on Android
- deno
- parallel sort algorithm
- LevelDB/RocksDB (YCSB)
- Redis
- libjpeg‐turbo (JPEG)
- JM (H.264/AVC)
- HM (HEVC)
- VTM (VVC)
- CUDA
- Erlang/OTP BEAM
- uftrace on Yocto
- TTCN3