We tested on Samsung Galaxy S6, which has a Mali-T760 GPU.
Platform: ARM Platform
Device: Mali-T760
OpenCL Driver version : 1.1 (Android)
Compute units : 8
Clock frequency : 772 MHz
workgroup sizes : 256
The current configuration is set to link with libOpenCL.so from this device.
To run a test, use
./intelgemm --kernel tn -s [matrix_size] --global-size [global_work_size] --local-size [local_work_size] --validation
The peak floating-point computation performance is benchmarked through https://github.com/krrishnarraj/clpeak.
Platform: ARM Platform
Device: Mali-T760
Driver version : 1.1 (Android)
Compute units : 8
Clock frequency : 772 MHz
workgroups per compute unit : 2048
workgroup sizes : 256
max alloc size : 536870912
Global memory bandwidth (GBPS)
float : 6.05
float2 : 10.51
float4 : 12.15
float8 : 12.17
float16 : 10.68
Single-precision compute (GFLOPS)
globalWIs = 4194304
float : 12.32
float2 : 32.25
float4 : 31.79
float8 : 45.37
float16 : 154.02
Double-precision compute (GFLOPS)
double : 3.04
double2 : 30.11
double4 : 22.09
double8 : 35.68
double16 : 34.79
Integer compute (GIOPS)
int : 5.13
int2 : 12.43
int4 : 11.99
int8 : 15.83
int16 : 72.55
Transfer bandwidth (GBPS)
enqueueWriteBuffer : 4.69
enqueueReadBuffer : 4.08
enqueueMapBuffer(for read) : 2536.00
memcpy from mapped ptr : 4.36
enqueueUnmap(after write) : 2382.92
memcpy to mapped ptr : 4.41
Kernel launch latency : 1399.80 us
One important thing to notice is the following constraints [1].
| Size | CL Kernel | Throughput (GFlops) |
|---|---|---|
| 1024 x 1024 | noblock-v8 | 11.14 |
| 1024 x 1024 | blocking-2x2-v4 | 17.4 |
| 2048 x 2048 | noblock-v8 | 15.45 |
| 2048 x 2048 | blocking-2x2-v4 | 27.3 |
| 4096 x 4096 | noblock-v8 | 15.6 |
| 4096 x 4096 | blocking-2x2-v4 | 4.4 |
| Size | CL Kernel | Throughput (GFlops) |
|---|---|---|
| 1024 x 1024 | noblock-v16 | 6.46 |
| 1024 x 1024 | noblock-v16-dotprod | 11.50 |
| 2048 x 2048 | noblock-v16 | 6.09 |
| 2048 x 2048 | noblock-v16-dotprod | 15.10 |
| 2048 x 2048 | blocking-4x4-v8 | 17.2 |
| 2048 x 2048 | blocking-4x4-v4 | 10.6 |
Available implementations in literature for reference. None of these achieve good performance on our Android test device.
- [Some GEMM implementation in clBLAS] (https://github.com/clMathLibraries/clBLAS/blob/master/src/library/blas/gens/clTemplates/sgemm_gcn_SmallMatrices.cl)
- [NVidia Implementation Sample] (http://www.nvidia.com/content/cudazone/CUDABrowser/downloads/papers/NVIDIA_OpenCL_BestPracticesGuide.pdf)
- [Intel GEMM Sample] (https://software.intel.com/sites/products/vcsource/files/GEMM.pdf)
- [ViennalCL] (https://www.researchgate.net/publication/229430755_ViennaCL_-_A_High_Level_Linear_Algebra_Library_for_GPUs_and_Multi-Core_CPUs)