oh
oh copied to clipboard
aolofsson
mailbox: Writes dissapear
Can you try the following: -run two scenarios, one with 0 noops inserted and one with 16 nops inserted -for each scenario, try twoo addresses, 1.) write to mailboxlo, 2) write to some address in memory
what do you get?
On Tue, Jan 19, 2016 at 2:27 PM, Ola Jeppsson [email protected] wrote:
I need to insert nops between writes to the mailbox from Epiphany.
— Reply to this email directly or view it on GitHub https://github.com/parallella/oh/issues/37.
Andreas Olofsson, CEO/Founder at Adapteva Cell: +1 781 325 6688 Twitter: @adapteva Web: adapteva.com
Linkedin: linkedin.com/in/andreasolofsson
http://www.adapteva.com/
Do you want:
- Write to mailbox
- Write to DRAM
- nops / no nops
or
- Write to mailbox
- nops / no nops
- Write to DRAM
- nops / no nops
// Ola
I want to see what happens with the exact same Epiphany code with the only difference being the addressing being written to. My theory is that you happened on a corner case that affects all writes from Epiphany to FPGA. The reason we may not have see this is that you wrote a tight assembly loop. Our other tests either use DMA (streaming) or slow C code loops. (in other words our elink test regression still needs major work!) Would ftest, matmul, etc code have the same sequence that you wrote in assembly for mailbox test?
(a fix has tested and checked in)
On Tue, Jan 19, 2016 at 3:57 PM, Ola Jeppsson [email protected] wrote:
Do you want:
- Write to mailbox
- Write to DRAM
- nops / no nops
or
- Write to mailbox
- nops / no nops
- Write to DRAM
- nops / no nops
// Ola
— Reply to this email directly or view it on GitHub https://github.com/parallella/oh/issues/37#issuecomment-172984554.
Andreas Olofsson, CEO/Founder at Adapteva Cell: +1 781 325 6688 Twitter: @adapteva Web: adapteva.com
Linkedin: linkedin.com/in/andreasolofsson
http://www.adapteva.com/
Without nops but with an added str to 0x8f100000 all ten writes to the mailbox arrives. The ARM side reads until E_MAILBOXSTATUS == 0.
00000660 <_main>:
660: 4a0b 0552 mov r2,0x5550
664: 754b 0aa2 mov r3,0xaaaa
668: 060b 4072 mov r16,0x730
66c: 4aab 1552 movt r2,0x5555
670: 754b 1aa2 movt r3,0xaaaa
674: 01eb 5812 movt r16,0x810f
678: 200b 0002 mov r1,0x0
67c: 800b 2502 mov r12,0x5000
680: 407c 0800 strd r2,[r16]
684: 220b 18f2 movt r1,0x8f10
688: 954b 3aa2 movt r12,0xaaaa
68c: 845c 2000 str r12,[r1]
690: 407c 0800 strd r2,[r16]
694: 845c 2000 str r12,[r1]
698: 407c 0800 strd r2,[r16]
69c: 845c 2000 str r12,[r1]
6a0: 407c 0800 strd r2,[r16]
6a4: 845c 2000 str r12,[r1]
6a8: 407c 0800 strd r2,[r16]
6ac: 845c 2000 str r12,[r1]
6b0: 407c 0800 strd r2,[r16]
6b4: 845c 2000 str r12,[r1]
6b8: 407c 0800 strd r2,[r16]
6bc: 845c 2000 str r12,[r1]
6c0: 407c 0800 strd r2,[r16]
6c4: 845c 2000 str r12,[r1]
6c8: 407c 0800 strd r2,[r16]
6cc: 845c 2000 str r12,[r1]
6d0: 407c 0800 strd r2,[r16]
6d4: 845c 2000 str r12,[r1]
6d8: 194f 0402 rts
6dc: 0000 beq 6dc <_main+0x7c>
...
EDIT: With commit https://github.com/parallella/oh/commit/b26255dfb59bd28b65e48cca0a882a47d64f2aaa
Bug is still present in e86567241deabc8f070b61becf07be175394a980 Running through epiphany examples now ...
(examples still work)
The magic number with e865672 is 10 cycles (no stalls AFAIK) between each strd
I.e., this works:
00000660 <_main>:
660: 0a0b 0552 mov r0,0x5550
664: 354b 0aa2 mov r1,0xaaaa
668: 460b 0072 mov r2,0x730
66c: 0aab 1552 movt r0,0x5555
670: 354b 1aa2 movt r1,0xaaaa
674: 41eb 1812 movt r2,0x810f
678: 0874 strd r0,[r2]
67a: 01a2 nop
67c: 01a2 nop
67e: 01a2 nop
680: 01a2 nop
682: 01a2 nop
684: 01a2 nop
686: 0a0b 0552 mov r0,0x5550
68a: 354b 0aa2 mov r1,0xaaaa
68e: 0aab 1552 movt r0,0x5555
692: 354b 1aa2 movt r1,0xaaaa
696: 0874 strd r0,[r2]
698: 01a2 nop
...
Remove one nop and things break (only 6 entries in fifo)
@aolofsson
Which are the right values for ELINK_RXDELAY0 and ELINK_RXDELAY1 ?
The following should work:
idelay0=0xaaaaaaaa idleya1=0x0000000a
On Tue, Jan 19, 2016 at 5:00 PM, Ola Jeppsson [email protected] wrote:
@aolofsson https://github.com/aolofsson
Which are the right values for ELINK_RXDELAY0 and ELINK_RXDELAY1 ?
— Reply to this email directly or view it on GitHub https://github.com/parallella/oh/issues/37#issuecomment-173001846.
Andreas Olofsson, CEO/Founder at Adapteva Cell: +1 781 325 6688 Twitter: @adapteva Web: adapteva.com
Linkedin: linkedin.com/in/andreasolofsson
http://www.adapteva.com/
thanks, as expected it doesn't fix this issue but I had to test
// Ola
On 2016-01-20 00:08, Andreas Olofsson wrote:
The following should work:
idelay0=0xaaaaaaaa idleya1=0x0000000a
On Tue, Jan 19, 2016 at 5:00 PM, Ola Jeppsson [email protected] wrote:
@aolofsson https://github.com/aolofsson
Which are the right values for ELINK_RXDELAY0 and ELINK_RXDELAY1 ?
— Reply to this email directly or view it on GitHub https://github.com/parallella/oh/issues/37#issuecomment-173001846.
Andreas Olofsson, CEO/Founder at Adapteva Cell: +1 781 325 6688 Twitter: @adapteva Web: adapteva.com
Linkedin: linkedin.com/in/andreasolofsson
http://www.adapteva.com/
— Reply to this email directly or view it on GitHub https://github.com/parallella/oh/issues/37#issuecomment-173017631.
Does the following break?
00000660 <_main>: 660: 0a0b 0552 mov r0,0x5550 664: 354b 0aa2 mov r1,0xaaaa
- 668: 460b 0072 mov r2,0x0000* 66c: 0aab 1552 movt r0,0x5555 670: 354b 1aa2 movt r1,0xaaaa
- 674: 41eb 1812 movt r2,0x8f10* 678: 0874 strd r0,[r2] 67a: 01a2 nop 67c: 01a2 nop 67e: 01a2 nop 680: 01a2 nop 682: 01a2 nop 684: 01a2 nop 686: 0a0b 0552 mov r0,0x5550 68a: 354b 0aa2 mov r1,0xaaaa 68e: 0aab 1552 movt r0,0x5555 692: 354b 1aa2 movt r1,0xaaaa 696: 0874 strd r0,[r2] 698: 01a2 nop ...
On Tue, Jan 19, 2016 at 6:13 PM, Ola Jeppsson [email protected] wrote:
thanks, as expected it doesn't fix this issue but I had to test
// Ola
On 2016-01-20 00:08, Andreas Olofsson wrote:
The following should work:
idelay0=0xaaaaaaaa idleya1=0x0000000a
On Tue, Jan 19, 2016 at 5:00 PM, Ola Jeppsson [email protected] wrote:
@aolofsson https://github.com/aolofsson
Which are the right values for ELINK_RXDELAY0 and ELINK_RXDELAY1 ?
— Reply to this email directly or view it on GitHub https://github.com/parallella/oh/issues/37#issuecomment-173001846.
Andreas Olofsson, CEO/Founder at Adapteva Cell: +1 781 325 6688 Twitter: @adapteva Web: adapteva.com
Linkedin: linkedin.com/in/andreasolofsson
http://www.adapteva.com/
— Reply to this email directly or view it on GitHub < https://github.com/parallella/oh/issues/37#issuecomment-173017631>.
— Reply to this email directly or view it on GitHub https://github.com/parallella/oh/issues/37#issuecomment-173018561.
Andreas Olofsson, CEO/Founder at Adapteva Cell: +1 781 325 6688 Twitter: @adapteva Web: adapteva.com
Linkedin: linkedin.com/in/andreasolofsson
http://www.adapteva.com/
The reason we may not have see this is that you wrote a tight assembly loop. Our other tests either use DMA (streaming) or slow C code loops. (in other words our elink test regression still needs major work!) Would ftest, matmul, etc code have the same sequence that you wrote in assembly for mailbox test?
This works (will check in complete test tomorrow)
FUNC(etest)
.global SYM(etest)
.align 3
SYM(etest):
mov r20, %low(EXPECTED_0)
movt r20, %high(EXPECTED_0)
mov r21, %low(EXPECTED_1)
movt r21, %high(EXPECTED_1)
mov r22, %low(EXPECTED_2)
movt r22, %high(EXPECTED_2)
mov r23, %low(EXPECTED_3)
movt r23, %high(EXPECTED_3)
mov r24, %low(EXPECTED_4)
movt r24, %high(EXPECTED_4)
mov r25, %low(EXPECTED_5)
movt r25, %high(EXPECTED_5)
mov r26, %low(EXPECTED_6)
movt r26, %high(EXPECTED_6)
mov r27, %low(EXPECTED_7)
movt r27, %high(EXPECTED_7)
mov r30, %low(RESULT_ADDR)
movt r30, %high(RESULT_ADDR)
str r20,[r30],+1
str r21,[r30],+1
str r22,[r30],+1
str r23,[r30],+1
str r24,[r30],+1
str r25,[r30],+1
str r26,[r30],+1
str r27,[r30],+1
;; repeat w/ 64-bit writes
strd r20,[r30],+1
strd r22,[r30],+1
strd r24,[r30],+1
strd r26,[r30],+1
rts
Found the bug!! NASTY!!!!!
Remembered that we have a long forgotten mode in the epiphany chip elink (not impemented in the fpga elink) that creates bursts when you write doubles to the same address. (F**K!) So the writes were likely coming in as bursts. Looks like the mailbox works fine when you write in "int"s (I tested it on the board with consecutive) (see "mailbox_test" in elink/sw0)
Should not be a big deal to implement burst with increment set to 0 instead of 8, for now let's work with ints.
Please verify.
On Tue, Jan 19, 2016 at 6:38 PM, Ola Jeppsson [email protected] wrote:
(result_addr is in DRAM)
— Reply to this email directly or view it on GitHub https://github.com/parallella/oh/issues/37#issuecomment-173024031.
Andreas Olofsson, CEO/Founder at Adapteva Cell: +1 781 325 6688 Twitter: @adapteva Web: adapteva.com
Linkedin: linkedin.com/in/andreasolofsson
http://www.adapteva.com/