this is not official test result, just for reference
Here is the test result from internal team on i.MX8QXP MEK board. (8DX use the same GPU as 8QXP)
root@imx8qxpmek:~# ./clpeak
Platform: Vivante OpenCL Platform
Device: Vivante OpenCL Device GC7000L.6214.0000
Driver version : OpenCL 1.2 V6.2.4.p4.190076 (Linux ARM64)
Compute units : 1
Clock frequency : 850 MHz
Global memory bandwidth (GBPS)
float : 2.75
float2 : 5.07
float4 : 6.90
float8 : 5.77
float16 : 4.07
Single-precision compute (GFLOPS)
float : 6.63
float2 : 13.53
float4 : 26.88
float8 : 26.63
float16 : 26.41
No half precision support! Skipped
No double precision support! Skipped
Integer compute (GIOPS)
int : 6.79
int2 : 6.78
int4 : 6.77
int8 : 6.72
int16 : 6.70
Transfer bandwidth (GBPS)
enqueueWriteBuffer : 1.64
enqueueReadBuffer : 0.09
enqueueMapBuffer(for read) : 273.58
memcpy from mapped ptr : 0.09
enqueueUnmap(after write) : 228.22
memcpy to mapped ptr : 1.65
Kernel launch latency : 96.73 us
