Skip to content

Latest commit

 

History

History
379 lines (332 loc) · 15.3 KB

tuning.md

File metadata and controls

379 lines (332 loc) · 15.3 KB
Raw

CLBlast: Tuning for better performance

This document describes how to tune CLBlast for better performance and lists for which devices tuned kernels are already available. For other information about CLBlast, see the main README.

Already tuned-for devices

The CLBlast library is already tuned for the most commonly used OpenCL devices and it's gradually being extended to other devices as well. For unseen devices CLBlast will make use of common-best tuning values for similar architectures (e.g. AMD Fiji) or in general similar devices (e.g. AMD GPUs), so performance might still be decent. The current release of CLBlast is tuned for the following devices:

  • NVIDIA GPUs:
    • SM 2.0:
      • GeForce GTX 480
      • GeForce GTX 580
    • SM 3.0:
      • GRID K520
      • GeForce GT 650M
      • GeForce GTX 670
      • GeForce GTX 680
      • GeForce GTX 760 Ti
    • SM 3.5:
      • GeForce GT 730
      • GeForce 920A
      • GeForce GTX TITAN
      • GeForce GTX TITAN Black
      • Tesla K20m
      • Tesla K40m
    • SM 5.0:
      • GeForce GTX 920MX
      • GeForce GTX 750
      • GeForce GTX 750 Ti
      • Quadro M2000M
    • SM 5.2:
      • GeForce GTX 970
      • GeForce GTX 980
      • GeForce GTX TITAN X
    • SM 6.0:
      • Tesla P100 16GB
    • SM 6.1:
      • GeForce MX 150
      • GeForce GTX 1060 6GB
      • GeForce GTX 1070
      • GeForce GTX 1070 Ti
      • GeForce GTX 1080
      • GeForce GTX 1080 Ti
      • TITAN X (Pascal)
      • Tesla P4
    • SM 7.0:
      • Quadro GV100
      • Tesla V100
    • SM 7.5:
      • GeForce MX 450
      • GeForce GTX 1650
      • GeForce GTX 1650 Ti
      • GeForce GTX 1650 Super
      • GeForce GTX 2060
      • GeForce GTX 2070 with Max-Q
      • GeForce GTX 2070 Super
      • GeForce GTX 2080 with Max-Q
      • GeForce GTX 2080 Ti
      • Quadro T2000
      • TITAN RTX
      • Tesla T4
    • SM 8.0:
      • Tesla A100 40GB
    • SM 8.6:
      • GeForce GTX 3050 Ti Laptop
      • GeForce GTX 3060 Laptop
      • GeForce GTX 3070
      • GeForce GTX 3070 Ti Laptop
      • GeForce GTX 3080
      • GeForce GTX 3080 Laptop
      • GeForce GTX 3080 Ti
      • GeForce GTX 3090
      • RTX A6000
    • SM 8.9:
      • GeForce GTX 4050 Laptop
      • GeForce RTX 4060
      • GeForce GTX 4060 Ti
      • GeForce GTX 4070 Laptop
      • GeForce GTX 4070 Ti
      • GeForce GTX 4080
      • GeForce GTX 4090
  • AMD GPUs:
    • Turks:
      • Radeon HD 6770M
    • Vancouver:
      • Radeon HD 6750M
    • Tahiti:
      • Radeon HD 7970
    • Oland:
      • Radeon R7 250
    • Pitcairn:
      • Radeon R9 270X
    • Hawaii:
      • FirePro W8100
      • Radeon R9 290X
    • Tonga:
      • Radeon R9 380
    • Fiji:
      • Radeon 500
      • Radeon R9 Fury X
      • Radeon R9 M370X
    • Ellesmere:
      • Radeon RX 480
      • Radeon RX 580 2048SP
      • Radeon RX 590 GME
    • Vega:
      • Radeon RX Vega
    • gfx902:
      • Radeon RX Vega
      • Radeon RX Vega 10
    • gfx906:
      • Radeon VII
    • gfx90c:
      • Ryzen 5600G APU
      • Ryzen 5700G APU
    • gfx1010:
      • Radeon RX 5700
      • Radeon RX 5700 XT
    • gfx1030:
      • Radeon RX 6800 XT
      • Radeon RX 6900 XT
    • gfx1031:
      • Radeon RX 6700 XT
    • gfx1032:
      • Radeon RX 6600 XT
    • gfx1034:
      • Radeon RX 6500 XT
    • gfx1035:
      • Radeon 680M
      • Ryzen 4600G APU
    • gfx1100:
      • Radeon RX 7900 XTX
    • gfx1101:
      • Radeon RX 7800 XT
    • gfx1102:
      • Radeon RX 7600
    • gfx1103:
      • Radeon 780M
    • Other:
      • Radeon Pro 450
      • Radeon Pro 580
  • Intel GPUs:
    • HD Graphics 530
    • HD Graphics 5500 BroadWell U-Processor GT2
    • HD Graphics 6000 BroadWell U-Processor GT3
    • HD Graphics Haswell Ultrabook GT2 Mobile
    • HD Graphics IvyBridge M GT2
    • HD Graphics Skylake ULT GT2
    • UHD Graphics 620
    • UHD Graphics 630
    • UHD Graphics 770
    • Iris
    • Iris Pro
    • Iris Xe Graphics
    • RaptorLake-S Mobile Graphics
    • Arc A750
    • Arc A770
  • Intel CPUs:
    • Core i5-4570
    • Core i5-4590S
    • Core i5-6200U
    • Core i7-920
    • Core i7-2670QM
    • Core i7-3770K
    • Core i7-4790K
    • Core i7-5930K
    • Core i7-6770HQ
    • Core i7-12700H
    • Core i9-9980HK
    • Xeon E5-2630 v3
    • Xeon E5-2630 v4
  • Other devices:
    • ARM Mali-T628 GPU
    • ARM Mali-T760 GPU
    • ARM Mali-G57 MC2 GPU
    • Qualcomm Adreno 330 GPU
    • Qualcomm Adreno 540 GPU
    • Qualcomm Adreno 640 GPU
    • Qualcomm Adreno 730 GPU
    • Qualcomm Adreno 740 GPU
    • Intel MIC
    • Imagination PowerVR B-Series BXE-4-32
    • Apple M1 GPU
    • Apple M2 Max GPU

If your device is not (yet) among this list or if you want to tune CLBlast for specific parameters (e.g. rectangular matrix sizes), you should run the included tuners.

Compiling and running the tuners

The included CLBlast tuners are compiled with the default CMake options. If they are not compiled, make sure you are specifing -DTUNERS=ON, for example as follows:

cmake -DTUNERS=ON ..

Compiling with -DTUNERS=ON will generate a number of tuners, each named clblast_tuner_xxxxx, in which xxxxx corresponds to a .opencl kernel file as found in src/kernels. These kernels corresponds to routines (e.g. xgemm) or to common pre-processing or post-processing kernels (copy and transpose). Running such a tuner will test a number of parameter-value combinations on your device and report which one gave the best performance. Running make alltuners runs all tuners for all precisions in one go. You can set the default device and platform for alltuners by setting the CLBLAST_DEVICE and CLBLAST_PLATFORM environmental variables. Alternatively, you can also manually run each of the tuners for each of the precisions. Here's an example to tune the axpy kernels for 64-bit precision on device 0 of platform 0:

./clblast_tuner_xaxpy --precision 64 --device 0 --platform 0

The kernels gemm and gemm_direct have too many parameters to explore. Therefore, they will run in two stages: a first stage with a fixed limited number of parameter combinations, and a second stage with a random selection from a much larger search space. The random fraction is determined by the fraction argument on the command-line.

There are also several routine-level tuners. They tune inter-kernel parameters and should only be run after the kernels are tuned. However, they do automatically pick up kernel tuning results from the current folder if there are any. An example is the GEMM routine tuner, which determines when to use the direct or the in-direct GEMM kernel.

Here are all the tuners included in the make alltuners target (in the same order) with all their precision arguments:

./clblast_tuner_copy_fast -precision 32
./clblast_tuner_copy_fast -precision 64
./clblast_tuner_copy_fast -precision 3232
./clblast_tuner_copy_fast -precision 6464
./clblast_tuner_copy_fast -precision 16
./clblast_tuner_copy_pad -precision 32
./clblast_tuner_copy_pad -precision 64
./clblast_tuner_copy_pad -precision 3232
./clblast_tuner_copy_pad -precision 6464
./clblast_tuner_copy_pad -precision 16
./clblast_tuner_transpose_fast -precision 32
./clblast_tuner_transpose_fast -precision 64
./clblast_tuner_transpose_fast -precision 3232
./clblast_tuner_transpose_fast -precision 6464
./clblast_tuner_transpose_fast -precision 16
./clblast_tuner_transpose_pad -precision 32
./clblast_tuner_transpose_pad -precision 64
./clblast_tuner_transpose_pad -precision 3232
./clblast_tuner_transpose_pad -precision 6464
./clblast_tuner_transpose_pad -precision 16
./clblast_tuner_xaxpy -precision 32
./clblast_tuner_xaxpy -precision 64
./clblast_tuner_xaxpy -precision 3232
./clblast_tuner_xaxpy -precision 6464
./clblast_tuner_xaxpy -precision 16
./clblast_tuner_xdot -precision 32
./clblast_tuner_xdot -precision 64
./clblast_tuner_xdot -precision 3232
./clblast_tuner_xdot -precision 6464
./clblast_tuner_xdot -precision 16
./clblast_tuner_xger -precision 32
./clblast_tuner_xger -precision 64
./clblast_tuner_xger -precision 3232
./clblast_tuner_xger -precision 6464
./clblast_tuner_xger -precision 16
./clblast_tuner_xgemm -precision 32
./clblast_tuner_xgemm -precision 64
./clblast_tuner_xgemm -precision 3232
./clblast_tuner_xgemm -precision 6464
./clblast_tuner_xgemm -precision 16
./clblast_tuner_xgemm_direct -precision 32
./clblast_tuner_xgemm_direct -precision 64
./clblast_tuner_xgemm_direct -precision 3232
./clblast_tuner_xgemm_direct -precision 6464
./clblast_tuner_xgemm_direct -precision 16
./clblast_tuner_xgemv -precision 32
./clblast_tuner_xgemv -precision 64
./clblast_tuner_xgemv -precision 3232
./clblast_tuner_xgemv -precision 6464
./clblast_tuner_xgemv -precision 16
./clblast_tuner_invert -precision 32
./clblast_tuner_invert -precision 64
./clblast_tuner_invert -precision 3232
./clblast_tuner_invert -precision 6464
./clblast_tuner_invert -precision 16
./clblast_tuner_routine_xgemm -precision 32
./clblast_tuner_routine_xgemm -precision 64
./clblast_tuner_routine_xgemm -precision 3232
./clblast_tuner_routine_xgemm -precision 6464
./clblast_tuner_routine_xgemm -precision 16
./clblast_tuner_routine_xtrsv -precision 32
./clblast_tuner_routine_xtrsv -precision 64
./clblast_tuner_routine_xtrsv -precision 3232
./clblast_tuner_routine_xtrsv -precision 6464
./clblast_tuner_routine_xtrsv -precision 16

Using the tuning results

The tuners output a JSON-file with the results. The best results need to be added to src/database/kernels/xxxxx.hpp in the appropriate section. However, this can be done automatically based on the JSON-data using a Python (2.7 or 3.x) script in scripts/database/database.py. If you want the found parameters to be included in future releases of CLBlast, please attach the JSON files to the corresponding issue on GitHub or email the main author.

In summary, tuning the entire library for your device can be done as follows (starting from the root of the CLBlast folder):

mkdir build
cd build
cmake -DTUNERS=ON ..
make
make alltuners
python ../scripts/database/database.py . ..
make

Tuning using the API (advanced users only)

Apart from running the tuning binaries, it is also possible to run the tuners programmatically through the CLBlast API. This could be useful if you want to tune for non-standard arguments (e.g. a rectangular or very small matrix). The tuning results can then also be set programmatically using OverrideParameters.

The tuning API does not perform any disk or stdout I/O, thus it is not possible to track progress. Running the regular tuner binaries should give an idea of the amount of configurations to explore for a particular device, thus giving an indication of a good value for the fraction argument (see the API documentation for more details).

Inspecting and changing tuning parameters at run-time

Alternatively, you can also supply your tuning parameters programmatically through the CLBlast API. This is especially useful if you tune for specific non-standard arguments (e.g. a rectangular or a very small matrix). To do so, you can call the OverrideParameters function which will set new parameters for a specific kernel. At the first next call of the target routine, CLBlast will compile a new binary and use it together with the new parameters from then on. Until OverrideParameters is called again of course. This is the API:

StatusCode PUBLIC_API OverrideParameters(const cl_device_id device, const std::string &kernel_name,
                                         const Precision precision,
                                         const std::unordered_map<std::string,size_t> &parameters)

To inspect current behaviour, you can also retrieve the parameters for a specific device and kernel combination:

StatusCode PUBLIC_API RetrieveParameters(const cl_device_id device, const std::string &kernel_name,
                                         const Precision precision,
                                         std::unordered_map<std::string,size_t> &parameters)

These two functions require/retrieve the parameters as given in src/database/kernels, i.e.:

Kernel name Parameters
Xaxpy VW, WGS, WPT
Xdot WGS1, WGS2
Xgemv WGS1, WPT1
XgemvFast VW2, WGS2, WPT2
XgemvFastRot VW3, WGS3, WPT3
Xger WGS1, WGS2, WPT
Xtrsv TRSV_BLOCK_SIZE
Xgemm GEMMK, KREG, KWG, KWI, MDIMA, MDIMC, MWG, NDIMB, NDIMC, NWG, SA, SB, STRM, STRN, VWM, VWN
XgemmDirect KWID, MDIMAD, MDIMCD, NDIMBD, NDIMCD, PADA, PADB, VWMD, VWND, WGD
Copy COPY_DIMX, COPY_DIMY, COPY_VW, COPY_WPT
Pad PAD_DIMX, PAD_DIMY, PAD_WPTX, PAD_WPTY
Transpose TRA_DIM, TRA_PAD, TRA_SHUFFLE, TRA_WPT
Padtranspose PADTRA_PAD, PADTRA_TILE, PADTRA_WPT
Invert INTERNAL_BLOCK_SIZE
TrsvRoutine TRSV_BLOCK_SIZE

Tuning OpenCL compiler options

For all of CLBlast's APIs, it is possible to optionally set an OS environmental variable CLBLAST_BUILD_OPTIONS to pass specific build options to the OpenCL compiler. Also make sure this is set in the same way when running the tuners.

Which kernels are used for which routines?

To find out which tuners to run for which routines, you can use the table below. The kernel names correspond to the tuner binaries, the tuner API, and to the arguments for OverrideParameters and RetrieveParameters.

Routines Kernel(s) / Tuner(s)
AXPY COPY SCAL SWAP OMATCOPY AXPYBATCHED Xaxpy
AMAX ASUM DOT DOTC DOTU NRM2 SUM MAX MIN AMIN Xdot
GBMV GEMV HBMV HEMV HPMV SBMV SPMV SYMV TMBV TPMV TRMV TRSV Xgemv
GER GERC GERU HER HER2 HPR HPR2 SPR SPR2 SYR SYR2 Xger
GEMM HEMM HER2K HERK SYMM SYR2K SYRK TRMM GEMMBATCHED GEMMSTRIDEDBATCHED Xgemm XgemmDirect Copy Pad Transpose Padtranspose
TRSM Xgemm XgemmDirect Copy Pad Transpose Padtranspose Invert
IM2COL COL2IM Copy

A note on clock frequencies for tuning

You should consider limiting the clock speeds of your processors before performing the tuning. Some examples are given below.

To set the CPU frequency on a Linux machine:

sudo cpupower frequency-set -g performance
sudo cpupower frequency-set -u 3100

To set the NVIDIA GPU frequency on a Linux machine:

sudo nvidia-smi -i <device id> -lgc <clock-speed>

You can get the possible frequencies for your NVIDIA GPU using the following command:

sudo nvidia-smi -i <device id> --query-supported-clocks=gr --format=csv

The suggestion is to pick a clock speed that would be stable. Somewhere in the middle of the range of frequencies listed above.