OpenCL on BeagleBoard X15

[gtortone]

Hi,
I'm trying to use OpenCL with AM57xx-EVM (BeagleBoard X15 board) and with a recent kernel (e.g. 4.4.21-ti-r45) it does not work; after loading of cmemk module each OpenCL example fails with:

root@arm:/usr/share/ti/examples/opencl/simple# ./simple
Unable to allocate OCL MSMC memory from 0x40500000

(kernel 4.4.21-ti-r45 U-Boot load am57xx-beagle-x15-revb1 device tree but I don't know if this is correct)

Reading BeagleBoard X15 wiki I realized that OpenCL needs linux kernel: 4.1.10-ti-r23/4.1.10-ti-rt-r23 or greater from the v4.1.x-ti branch then I installed 4.1.10-ti-r23 but U-Boot did not find am57xx-beagle-x15-revb1 device tree and after forced U-Boot to load am57xx-beagle-x15.dtb (I don't know if this is correct)
OpenCL examples work but sometime in dmesg i found this:

[  453.129412] ------------[ cut here ]------------
[  453.129435] WARNING: CPU: 0 PID: 6622 at drivers/bus/omap_l3_noc.c:147 l3_interrupt_handler+0x284/0x39c()
[  453.129444] 44000000.ocp:L3 Custom Error: MASTER MPU TARGET DMM_P1 (Read): Data Access in User mode during Functional access
[  453.129450] Modules linked in: rpmsg_proto cmemk(O) 8021q garp mrp stp llc rpmsg_rpc virtio_rpmsg_bus snd_soc_simple_card ti_vpe videobuf2_dma_contig ti_vpdma v4l2_mem2mem snd_soc_omap_hdmi_audio videobuf2_memops pruss_remoteproc snd_soc_tlv320aic3x videobuf2_core v4l2_common videodev media snd_soc_davinci_mcasp snd_soc_edma omap_rng rng_core omap_remoteproc uio_pdrv_genirq uio bnep bluetooth rfkill
[  453.129554] CPU: 0 PID: 6622 Comm: dsplib_fft Tainted: G        W  O    4.1.10-ti-r23 #1
[  453.129560] Hardware name: Generic DRA74X (Flattened Device Tree)
[  453.129578] [<c001a180>] (unwind_backtrace) from [<c0014b38>] (show_stack+0x20/0x24)
[  453.129590] [<c0014b38>] (show_stack) from [<c09b54ec>] (dump_stack+0x8c/0xcc)
[  453.129603] [<c09b54ec>] (dump_stack) from [<c00479c8>] (warn_slowpath_common+0x98/0xc8)
[  453.129613] [<c00479c8>] (warn_slowpath_common) from [<c0047a38>] (warn_slowpath_fmt+0x40/0x48)
[  453.129623] [<c0047a38>] (warn_slowpath_fmt) from [<c05471f8>] (l3_interrupt_handler+0x284/0x39c)
[  453.129638] [<c05471f8>] (l3_interrupt_handler) from [<c009f1e8>] (handle_irq_event_percpu+0xb0/0x254)
[  453.129649] [<c009f1e8>] (handle_irq_event_percpu) from [<c009f3d8>] (handle_irq_event+0x4c/0x6c)
[  453.129658] [<c009f3d8>] (handle_irq_event) from [<c00a23c8>] (handle_fasteoi_irq+0xf0/0x1ac)
[  453.129668] [<c00a23c8>] (handle_fasteoi_irq) from [<c009e70c>] (generic_handle_irq+0x3c/0x4c)
[  453.129677] [<c009e70c>] (generic_handle_irq) from [<c009ea38>] (__handle_domain_irq+0x8c/0xfc)
[  453.129686] [<c009ea38>] (__handle_domain_irq) from [<c000952c>] (gic_handle_irq+0x34/0x6c)
[  453.129697] [<c000952c>] (gic_handle_irq) from [<c09bae40>] (__irq_svc+0x40/0x74)
[  453.129703] Exception stack(0xedcf5c68 to 0xedcf5cb0)
[  453.129710] 5c60:                   000000ff eb4a265f 00000000 000007ff ffeff000 00000000
[  453.129718] 5c80: c0025144 00080000 eb782140 ecde1b88 00000207 edcf5ccc 000eb4a2 edcf5cb0
[  453.129724] 5ca0: c0004000 c00244d4 20070013 ffffffff
[  453.129736] [<c09bae40>] (__irq_svc) from [<c00244d4>] (kmap_atomic+0x110/0x15c)
[  453.129745] [<c00244d4>] (kmap_atomic) from [<c0022110>] (__flush_dcache_page+0xa0/0x108)
[  453.129755] [<c0022110>] (__flush_dcache_page) from [<c00221f0>] (flush_dcache_page+0x78/0x7c)
[  453.129767] [<c00221f0>] (flush_dcache_page) from [<c0173150>] (__get_user_pages+0x408/0x4d0)
[  453.129779] [<c0173150>] (__get_user_pages) from [<c0173548>] (get_user_pages_unlocked+0x13c/0x1c8)
[  453.129791] [<c0173548>] (get_user_pages_unlocked) from [<c0166ef4>] (get_user_pages_fast+0x50/0x58)
[  453.129802] [<c0166ef4>] (get_user_pages_fast) from [<c00c66e0>] (get_futex_key+0xdc/0x244)
[  453.129811] [<c00c66e0>] (get_futex_key) from [<c00c6928>] (futex_wake+0x40/0x148)
[  453.129819] [<c00c6928>] (futex_wake) from [<c00c8af8>] (do_futex+0x140/0xa74)
[  453.129827] [<c00c8af8>] (do_futex) from [<c00c94b8>] (SyS_futex+0x8c/0x178)
[  453.129836] [<c00c94b8>] (SyS_futex) from [<c0045334>] (mm_release+0xf0/0x110)
[  453.129845] [<c0045334>] (mm_release) from [<c004a3e4>] (do_exit+0x16c/0x9e8)
[  453.129855] [<c004a3e4>] (do_exit) from [<c004acac>] (do_group_exit+0x0/0xd4)
[  453.129861] ---[ end trace baa5f7cd7094a3af ]---

It is possible to know a working combination of kernel release and device tree to work with OpenCL without any issues ?

Thanks a lot

[RobertCNelson]

@gtortone sorry, haven't worked on this issue this week. Been working on work-arounds for the bbgw wlan0 fun-ness.

For v4.1.x, yes opencl had a lot of messages dumpped to the kernel, you can ignore them.. For the v4.1.x -> v4.4.x transition some things are still broken, and when i tested the last TI sdk for the x15, the opencl speed up was only 0.8x vs, the 3x->4x we saw in v4.1.x

I've noticed quite a few opencl updates coming across the ti mailing list, so it might be working again. (just need to find a free moment)

REgards,

[RobertCNelson]

@gtortone there's a new build:

http://arago-project.org/git/projects/?p=oe-layersetup.git;a=blob;f=configs/processor-sdk/processor-sdk-03.00.00.04-config.txt;hb=HEAD

Can you dl and test:

http://software-dl.ti.com/processor-sdk-linux/esd/AM57X/latest/index_FDS.html

run the opencl-mpp opencl demo, it'll compare the Dual A15's vs the DSP's

If it's back in the 2x/3x speed up range, i'll start porting the stuff over again..

Regards,

[Shim-Apan]

@RobertCNelson

Hi, i am wondering the same the same thing.
I've downloaded the newest processor sdk and installed it. However i could not find the opencl-mpp demo. I did however ran the examples from the opencl example folder (dgemm and matmpy)

Their results are:

root@am57xx-evm:/usr/share/ti/examples/opencl/dgemm# ./dgemm
Generating random data ... Done. Starting Dgemm.
cMaj C[2048,2048] = A[2048,2048] * B[2048,2048]:
2 DSPs: 3.641 Gflops (4.718682s)
1 CPU : 2.217 Gflops (7.747983s) with ATLAS library
PASS!

and

root@am57xx-evm:/usr/share/ti/examples/opencl/matmpy# ./matmpy
float C[512][512] = float A[512][512] x float B[512][512]
OpenCL dispatching to 1 DSP(S): 0.2874 secs
OpenMP dispatching to 4 CPU(S): 0.2582 secs
Passed!

I'm not sure wether this displays an increase or decrease in speedup compared to the cpu. dgemm supports your findings of a 0.8x speedup but i'm not sure what to make of matmpy's result.

-Shim

edit: formatting

[pdp7]

Please re-open if still an issue with our current Debian images:
https://github.com/beagleboard/Latest-Images