On the various Joule datasheets (1.0, 1.1) there is a clear reference to its video encoding capabilities, up to 4KP60 for H.264 standard.
I benchmarked the libva under OstroXT, and what I achieve is a maximum of 1080P90 on preallocated buffers, already mapped as NV12. Real-world contexts would be far more stressing the system (Bayer input via MIPI, non-zero copy memory contexts and so on)... how to achieve the maximum claimed performance of 1080P240 or 4KP60?
Thanks in advance.
We need more information to help you, can you answer us the following questions:
1. What tools did you used to perform the benchmark?
2. What Ostro image and BIOS version are you using?
3. What are the specs of the power supply used with Joule?
We will be waiting for your answers.
Have a nice day.
Here below the needed information:
BB_VERSION = "1.31.0"
BUILD_SYS = "x86_64-linux"
NATIVELSBSTRING = "openSUSELeap-42.1"
TARGET_SYS = "x86_64-ostro-linux"
MACHINE = "intel-corei7-64"
DISTRO = "ostro-xt"
DISTRO_VERSION = "1.0+snapshot-20160813"
TUNE_FEATURES = "m64 corei7"
TARGET_FPU = ""
LibVA: included in OstroXT (1.7.0 API 0.39.3)
libva info: VA-API version 0.39.3
The test has been executed by monitoring the encoding time of a pure libva encoding example (the one in \libva-1.7.0\test\encode\h264encode.c, modified to avoid probing the memory loading process - IMHO that example is pretty weird).
The measured performance is the one depicted in the first post, not going any further than 1080P90 on a NV12 preallocated buffer.
Hope this helps,
I forgot to add we are using the 570x Development kit, and therefore using the standard power supplier included. Also we managed to mount the passive heatsink on the board, in addition to a *big* auxiliary fan we drive from another PSU. The Joule module is running cool, no throttling is expected (also, we drug the cpufreqd to select the "performance" governor per each core - using -c0,1,2,3 options in subsequent calls).
We have an update on this case, we are still making some tests with the Joule. We will publish the power and performance metrics in the first quarter of 2017.
We really appreciate your patience