PCCR and PCRM registers
A message that is not in the archive (sent off-list or not yet received): [email protected]
+cc mailing list,
Hi Elisa,
Which toolchain are you using? I guess newlib?
It has functions like or1k_mfspr() and or1k_mtspr() see or1k-support.h and or1k-sprs.h headers for details.
-Stafford
Quoted 329 lines · On Wed, Oct 16, 2019, 8:40 PM <[email protected]> wrote:
Hi Stafford,
I am with PCCR and PCRM registers. I have seen that I can access from asm language, but there is functions to access from C? Have you got any example about their usage?
I have already confirmed my subscription to the mailing list.
Thanks
Elisa-----Mensaje original-----
De: Stafford Horne <[email protected]>
Enviado el: miércoles, 9 de octubre de 2019 13:38
Para: ecalvo at 2se.es; Julius Baxter <[email protected]>
Asunto: Re: other doubtHello Elisa,
If you simulate with Icarus or modelsim you will be able to measure pretty much the same performance characteristics as FPGA. So there is no need to go straight to FPGA.
In terms of my example C code is one option. You can also read timer data directly from the tick timer in assembly and achieve the same thing.
If you are interested we can CC the mailing list and get more opinions.
-Stafford
Quoted 4 lines · On Wed, Oct 9, 2019 at 5:09 PM <[email protected]> wrote:
Hi Stafford,
Nice to meet you and, first of all, thanks a lot for your guidance. I am
new on this, and although there is some documentation, sometimes it is difficult some point which maybe it is basic.
Quoted 2 lines
Ok, to your comments. If.."A simulator like QEMU or or1ksim will not
give and exact representation of the CPUs real time performance"...then...if I simulate directly the processor with modelsim, icarus or a similar tool...neither I get a real performance, don’t I? And values for the counters that you tell me to enable, neither are real, isn't it? should I execute it directly on the FPGA and it will depends on the implementation?
Quoted 15 lines
Ok, to C code. I have understood the dependency with toolchain.
Thanks a lot again.
Best regards,
Elisa-----Mensaje original-----
De: Stafford Horne <[email protected]> Enviado el: martes, 8 de octubre de 2019 16:18
Para: Julius Baxter <[email protected]>
CC: ecalvo at 2se.es
Asunto: Re: other doubtQuoted 3 lines
Hi Elisa,
OpenRISC cpu's can run any algorithm, but how well it will perform
depends on many things:
Quoted 4 lines
- Compiler optimization flags (i.e. -O3)
- Whether or not you are doing FPU instructions and have FPU enabled
- Whether or not you use multiply and divide and have theseinstructions
Quoted 4 lines
enabled
- The frequency you are running
- Cache settings Icache Dcache
- The type of algorithm, does it require lots of data which will causemany
Quoted 3 lines
cache misses?
A simulator like QEMU or or1ksim will not give and exact representation
of the CPUs real time performance. It can tell you which intructions will be executed, but not how fast those will run or how many pipeline stalls of cache misses will happen.
Quoted 2 lines
You can use the performance counters, they are supported in mor1kx if
you enable them with the FEATURE_PERFCOUNTERS='ENABLED' parameter. They can help count how many events happen between certain events. Then you can combine them with a timer and watchpoints to detect how many times a loop can execute in 1000 clock cycles etc. Please read about PCCRn and PCMRn in the architecture manual.
Quoted 2 lines
It might be just as easy to use simple timing in a c program though,
depending on the toolchain you use you can compare times between runs of your algorithm.
Quoted 31 lines
i.e.
#include <time.h> #include <stdio.h> static long to_micro(struct timespec *time) { return (time->tv_sec * 1000000) + (time->tv_nsec / 1000); } int main() { int i, j = 0; struct timespec before, after; clock_gettime(CLOCK_MONOTONIC, &before); /* Super complex algorithm */ for (int i = 0; i < 100; i++) { j = (j+1) * (j+2); } clock_gettime(CLOCK_MONOTONIC, &after); printf("time to run algorithm %ld uSecs\n", to_micro(&after) -
to_micro(&before));return 0; }$ or1k-smh-linux-gnu-gcc timer.c $ ./glibc-build-scripts/qemu-or1k-libc ./a.out time to run algorithm 164 uSecsQuoted 11 lines
I hope it helps.
-Stafford
Quoted 6 lines · Julius Baxter
Hi,
No problem.
There are performance counters in the OpenRISC architecture but whether they're implemented in a particular implementation is another
matter.
Quoted 18 lines
Quoted 18 lines
You can use these registers to measure various things the CPU is doing while it's executing. If you read the ISA document it'll tell you about them.
I'm CCing Stafford because he's the main OpenRISC man these days and probably knows about the state of the performance counter registers in various simulators and RTL implementations.
Cheers,
JuliusQuoted 4 lines · On Tue., 8 Oct. 2019, 10:43 pm , <[email protected]> wrote:
Hi Julius,
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Sorry for bothering you again ☹. Can I do you other fast question related to openrisc? If not, ignore the email please.
Is there any way to characterize the type of application that I can run in openrisc? I mean, could you measure (with numbers) if an algorithm can be executed on it and the speed that it will
achieve?
Quoted 29 lines
Quoted 29 lines
Quoted 29 lines
Is it possible to do it using orksim?
Sorry because maybe it is so basic and general ☹
Thanks in advance
Elisa
*De:* Julius Baxter <[email protected]> *Enviado el:* lunes, 16 de septiembre de 2019 13:11
*Para:* ecalvo at 2se.es
*Asunto:* Re: Starting with OpenRISC - IOBsAlso! To let you know, if you're in Spain, we will soon be having our ORConf conference in Europe, and it's in Bordeaux, France, just across the border. :)
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
There are several people there who can help you get up to speed, one of whom is Stafford Horne who knows most about the OpenRISC IP lately. He will be presenting. If you can attend, it'd be helpful,
I'm sure.
Quoted 107 lines
Quoted 107 lines
Quoted 107 lines
All info at https://orconf.org
Cheers,
Julius
On Mon, 16 Sep 2019 at 21:09, Julius Baxter
<[email protected]>
wrote:Hi Elisa,
Sorry for the delay in this response.
You should be using an SoC toplevel. FPGAs have everything you need on board like memories and IO blocks and lots of other FPGA fabric for you to implement other pieces of hardware.
FuseSoC provides a really nice and easy way to build an mor1kx design for the DE0 nano I believe:
https://github.com/olofk/de0_nano
That github page has a rough guide to getting it going.
If you need help I recommend posting to the OpenRISC mailing list and people will respond probably more promptly than I. (I recommend getting to know how to use mailing lists.
https://openrisc.io/communityThere are more resources here: https://openrisc.io/tutorials
I hope that's helpful.
Cheers,
Julius
On Wed, 11 Sep 2019 at 20:09, <[email protected]> wrote:
Hi Julius,
Thanks a lot for the quick answer.
Yes, this is the problem: I am using as top level the mor1kx module itself. You mean that I need to synthetize also in reconfigurable logic these cores, don’t you? I thought that I could have these elements as external in a development board.
Thanks again,
Cheers
Elisa
*De:* Julius Baxter <[email protected]> *Enviado el:* miércoles, 11 de septiembre de 2019 12:02
*Para:* ecalvo at 2se.es
*Asunto:* Re: Starting with OpenRISC - IOBsHi Elisa,
Thanks for getting in touch, that sounds like a cool project.
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Can you tell me about the toplevel - are you using a system toplevel, or is your toplevel the mor1kx module itself?
If it's the latter, then that's not the best way to do it - you need a system toplevel which instantiates memories and some reset circuitry and likely some IO (UART, GPIO, JTAG debug, etc.) to talk
to the outside world.
Quoted 19 lines
Quoted 19 lines
Quoted 19 lines
Is that helpful?
Cheers,
Julius
On Wed, 11 Sep 2019 at 19:47, <[email protected]> wrote:
Dear Dr. Baxter,
Quoted 31 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
My name is Elisa Calvo Gallego. I am writing you because I have started to work with OpenRISC in the framework of a research project developed in the company where I am working (Space Submicron Electronics, 2SE), and I am having some basic troubles.
Could you help me?
Quoted 24 lines
Quoted 23 lines
Quoted 23 lines
Although the FPGA that we are planning to use is larger, I have synthetized mor1kx for a DE0 nano board as first step (this is the board used in the majority of guides and tutorials). My problem is that the results that I have obtained are similar in area and resources, except for IOBs, which are more than available IOBs in the device. Do you know what I am doing wrong? Should I comment debug lines or something like that? I apologize if the question is immediate. I didn't find the answer and I'm new in this.
Thanks very much in advance.
Best regards,
Elisa
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.librecores.org/pipermail/openrisc/attachments/20191017/445c4da1/attachment.html>
Hi Stafford,
Yes, I am using newlib. I had discovered both files; these registers in sprs.h
/******************************/
/* Performance Counters Group */
/******************************/
#define OR1K_SPR_PERF_GROUP 0x07
/* Performance Counters Count Registers */
#define OR1K_SPR_PERF_PCCR_BASE OR1K_UNSIGNED(0x000)
#define OR1K_SPR_PERF_PCCR_COUNT OR1K_UNSIGNED(0x008)
#define OR1K_SPR_PERF_PCCR_STEP OR1K_UNSIGNED(0x001)
#define OR1K_SPR_PERF_PCCR_INDEX(N) (OR1K_SPR_PERF_PCCR_BASE + ((N) * OR1K_SPR_PERF_PCCR_STEP))
#define OR1K_SPR_PERF_PCCR_ADDR(N) ((OR1K_SPR_PERF_GROUP << OR1K_SPR_GROUP_LSB) | OR1K_SPR_PERF_PCCR_INDEX(N))
/* Performance Counters Mode Registers */
#define OR1K_SPR_PERF_PCMR_BASE OR1K_UNSIGNED(0x008)
#define OR1K_SPR_PERF_PCMR_COUNT OR1K_UNSIGNED(0x008)
#define OR1K_SPR_PERF_PCMR_STEP OR1K_UNSIGNED(0x001)
#define OR1K_SPR_PERF_PCMR_INDEX(N) (OR1K_SPR_PERF_PCMR_BASE + ((N) *OR1K_SPR_PERF_PCMR_STEP))
#define OR1K_SPR_PERF_PCMR_ADDR(N) ((OR1K_SPR_PERF_GROUP << OR1K_SPR_GROUP_LSB) | OR1K_SPR_PERF_PCMR_INDEX(N))
/* Performance Counters Configuration */
#define OR1K_SPR_SYS_PCCFGR_INDEX OR1K_UNSIGNED(0x008)
#define OR1K_SPR_SYS_PCCFGR_ADDR OR1K_UNSIGNED(0x0008)
/* Number of Performance Counters */
#define OR1K_SPR_SYS_PCCFGR_NPC_LSB 0
#define OR1K_SPR_SYS_PCCFGR_NPC_MSB 2
#define OR1K_SPR_SYS_PCCFGR_NPC_BITS 3
#define OR1K_SPR_SYS_PCCFGR_NPC_MASK OR1K_UNSIGNED(0x00000007)
#define OR1K_SPR_SYS_PCCFGR_NPC_GET(X) (((X) >> 0) & OR1K_UNSIGNED(0x00000007))
#define OR1K_SPR_SYS_PCCFGR_NPC_SET(X, Y) (((X) & OR1K_UNSIGNED(0xfffffff8)) | ((Y) << 0))
And these functions in support.h
static inline void or1k_mtspr (uint32_t spr, uint32_t value)
static inline uint32_t or1k_mfspr (uint32_t spr)
Despite this I don’t have clear how to use it.
1. If I do: or1k_mtspr (OR1K_SPR_SYS_PCCFGR_ADDR , 0) -> Does it allow me to configure the PCCFGR to one performance counter?
Is This the same than or1k_mtspr (OR1K_SPR_SYS_PCCFGR_ADDR , OR1K_SPR_SYS_PCCFGR_NPC_LSB ) or Do OR1K_SPR_SYS_PCCFGR_NPC_LSB , OR1K_SPR_SYS_PCCFGR_NPC_MSB, etc. provide different functions on each one performance counter?
2. What is the meaning of PCCR_BASE, PCCR_COUNT, PCCR_STEP, PCMR_INDEX(N), PCMR_ADDR(N) ? (the same for PCMR) (Is BASE the base address of all PCCR and ADDR the position of each one of them? ….why PCMR_BASE and COUNT hasta de same value OR1K_UNSIGNED(0x008)? )
3. Should I define first PCCFGR, second PCMR and last get PCCR?
Thanks and sorry for the inconveniences.
Elisa
De: Stafford Horne <[email protected]>
Enviado el: jueves, 17 de octubre de 2019 7:14
Para: ecalvo at 2se.es; Openrisc <[email protected]>
Asunto: Re: PCCR and PCRM registers
+cc mailing list,
Hi Elisa,
Which toolchain are you using? I guess newlib?
It has functions like or1k_mfspr() and or1k_mtspr() see or1k-support.h and or1k-sprs.h headers for details.
-Stafford
On Wed, Oct 16, 2019, 8:40 PM < <mailto:[email protected]> ecalvo at 2se.es> wrote:
Hi Stafford,
I am with PCCR and PCRM registers. I have seen that I can access from asm language, but there is functions to access from C? Have you got any example about their usage?
I have already confirmed my subscription to the mailing list.
Thanks
Elisa
-----Mensaje original-----
De: Stafford Horne < <mailto:[email protected]> shorne at gmail.com>
Enviado el: miércoles, 9 de octubre de 2019 13:38
Para: <mailto:[email protected]> ecalvo at 2se.es; Julius Baxter < <mailto:[email protected]> juliusbaxter at gmail.com>
Asunto: Re: other doubt
Hello Elisa,
If you simulate with Icarus or modelsim you will be able to measure pretty much the same performance characteristics as FPGA. So there is no need to go straight to FPGA.
In terms of my example C code is one option. You can also read timer data directly from the tick timer in assembly and achieve the same thing.
If you are interested we can CC the mailing list and get more opinions.
-Stafford
Quoted 70 lines · On Wed, Oct 9, 2019 at 5:09 PM < <mailto:[email protected]> ecalvo at 2se.es> wrote:
Hi Stafford,
Nice to meet you and, first of all, thanks a lot for your guidance. I am new on this, and although there is some documentation, sometimes it is difficult some point which maybe it is basic.
Ok, to your comments. If.."A simulator like QEMU or or1ksim will not give and exact representation of the CPUs real time performance"...then...if I simulate directly the processor with modelsim, icarus or a similar tool...neither I get a real performance, don’t I? And values for the counters that you tell me to enable, neither are real, isn't it? should I execute it directly on the FPGA and it will depends on the implementation?
Ok, to C code. I have understood the dependency with toolchain.
Thanks a lot again.
Best regards,
Elisa-----Mensaje original-----
De: Stafford Horne < <mailto:[email protected]> shorne at gmail.com> Enviado el: martes, 8 de octubre de 2019 16:18
Para: Julius Baxter < <mailto:[email protected]> juliusbaxter at gmail.com>
CC: <mailto:[email protected]> ecalvo at 2se.es
Asunto: Re: other doubtHi Elisa,
OpenRISC cpu's can run any algorithm, but how well it will perform depends on many things:
- Compiler optimization flags (i.e. -O3)
- Whether or not you are doing FPU instructions and have FPU enabled
- Whether or not you use multiply and divide and have these instructions enabled
- The frequency you are running
- Cache settings Icache Dcache
- The type of algorithm, does it require lots of data which will cause many cache misses?A simulator like QEMU or or1ksim will not give and exact representation of the CPUs real time performance. It can tell you which intructions will be executed, but not how fast those will run or how many pipeline stalls of cache misses will happen.
You can use the performance counters, they are supported in mor1kx if you enable them with the FEATURE_PERFCOUNTERS='ENABLED' parameter. They can help count how many events happen between certain events. Then you can combine them with a timer and watchpoints to detect how many times a loop can execute in 1000 clock cycles etc. Please read about PCCRn and PCMRn in the architecture manual.
It might be just as easy to use simple timing in a c program though, depending on the toolchain you use you can compare times between runs of your algorithm. i.e.
#include <time.h> #include <stdio.h> static long to_micro(struct timespec *time) { return (time->tv_sec * 1000000) + (time->tv_nsec / 1000); } int main() { int i, j = 0; struct timespec before, after; clock_gettime(CLOCK_MONOTONIC, &before); /* Super complex algorithm */ for (int i = 0; i < 100; i++) { j = (j+1) * (j+2); } clock_gettime(CLOCK_MONOTONIC, &after); printf("time to run algorithm %ld uSecs\n", to_micro(&after) -
to_micro(&before));return 0; }$ or1k-smh-linux-gnu-gcc timer.c $ ./glibc-build-scripts/qemu-or1k-libc ./a.out time to run algorithm 164 uSecs
Quoted 231 lines
I hope it helps.
-Stafford
Quoted 225 lines · Julius Baxter
Hi,
No problem.
There are performance counters in the OpenRISC architecture but whether they're implemented in a particular implementation is another matter.
You can use these registers to measure various things the CPU is doing while it's executing. If you read the ISA document it'll tell you about them.
I'm CCing Stafford because he's the main OpenRISC man these days and probably knows about the state of the performance counter registers in various simulators and RTL implementations.
Cheers,
JuliusQuoted 205 lines · On Tue., 8 Oct. 2019, 10:43 pm , < <mailto:[email protected]> ecalvo at 2se.es>…
Hi Julius,
Sorry for bothering you again ☹. Can I do you other fast question related to openrisc? If not, ignore the email please.
Is there any way to characterize the type of application that I can run in openrisc? I mean, could you measure (with numbers) if an algorithm can be executed on it and the speed that it will achieve? Is it possible to do it using orksim?
Sorry because maybe it is so basic and general ☹
Thanks in advance
Elisa
*De:* Julius Baxter < <mailto:[email protected]> juliusbaxter at gmail.com> *Enviado el:* lunes, 16 de septiembre de 2019 13:11
*Para:* <mailto:[email protected]> ecalvo at 2se.es
*Asunto:* Re: Starting with OpenRISC - IOBsAlso! To let you know, if you're in Spain, we will soon be having our ORConf conference in Europe, and it's in Bordeaux, France, just across the border. :)
There are several people there who can help you get up to speed, one of whom is Stafford Horne who knows most about the OpenRISC IP lately. He will be presenting. If you can attend, it'd be helpful, I'm sure.
All info at <https://orconf.org> https://orconf.org
Cheers,
Julius
On Mon, 16 Sep 2019 at 21:09, Julius Baxter
< <mailto:[email protected]> juliusbaxter at gmail.com> wrote:Hi Elisa,
Sorry for the delay in this response.
You should be using an SoC toplevel. FPGAs have everything you need on board like memories and IO blocks and lots of other FPGA fabric for you to implement other pieces of hardware.
FuseSoC provides a really nice and easy way to build an mor1kx design for the DE0 nano I believe:
<https://github.com/olofk/de0_nano> https://github.com/olofk/de0_nano
That github page has a rough guide to getting it going.
If you need help I recommend posting to the OpenRISC mailing list and people will respond probably more promptly than I. (I recommend getting to know how to use mailing lists.
<https://openrisc.io/community> https://openrisc.io/communityThere are more resources here: <https://openrisc.io/tutorials> https://openrisc.io/tutorials
I hope that's helpful.
Cheers,
Julius
On Wed, 11 Sep 2019 at 20:09, < <mailto:[email protected]> ecalvo at 2se.es> wrote:
Hi Julius,
Thanks a lot for the quick answer.
Yes, this is the problem: I am using as top level the mor1kx module itself. You mean that I need to synthetize also in reconfigurable logic these cores, don’t you? I thought that I could have these elements as external in a development board.
Thanks again,
Cheers
Elisa
*De:* Julius Baxter < <mailto:[email protected]> juliusbaxter at gmail.com> *Enviado el:* miércoles, 11 de septiembre de 2019 12:02
*Para:* <mailto:[email protected]> ecalvo at 2se.es
*Asunto:* Re: Starting with OpenRISC - IOBsHi Elisa,
Thanks for getting in touch, that sounds like a cool project.
Can you tell me about the toplevel - are you using a system toplevel, or is your toplevel the mor1kx module itself?
If it's the latter, then that's not the best way to do it - you
need a system toplevel which instantiates memories and some reset
circuitry and likely some IO (UART, GPIO, JTAG debug, etc.) to talk to the outside world.Is that helpful?
Cheers,
Julius
On Wed, 11 Sep 2019 at 19:47, < <mailto:[email protected]> ecalvo at 2se.es> wrote:
Dear Dr. Baxter,
My name is Elisa Calvo Gallego. I am writing you because I have started to work with OpenRISC in the framework of a research
project developed in the company where I am working (Space
Submicron Electronics, 2SE), and I am having some basic troubles. Could you help me?Although the FPGA that we are planning to use is larger, I have synthetized mor1kx for a DE0 nano board as first step (this is the board used in the majority of guides and tutorials). My problem is that the results that I have obtained are similar in area and resources, except for IOBs, which are more than available IOBs in the device. Do you know what I am doing wrong? Should I comment debug lines or something like that? I apologize if the question is immediate. I didn't find the answer and I'm new in this.
Thanks very much in advance.
Best regards,
Elisa
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.librecores.org/pipermail/openrisc/attachments/20191017/23f38c86/attachment-0001.html>
Hi Elisa,
Right, these are the functions. You only need to be concerned with:
OR1K_SPR_PERF_PCCR_ADDR(n)
OR1K_SPR_PERF_PCMR_ADDR(n)
OR1K_SPR_SYS_PCCFGR_ADDR
OR1K_SPR_SYS_PCCFGR_NPC_GET(x)
The others are used internally.
The PCCFGR is read only, it specifies how many performance counters your CPU has built in. It may be 0-7.
An example of how to use them:
#include <or1k-support.h>
#include <or1k-sprs.h>
#include <stdio.h>
#define PCMR_CISM 1<<3
#define PCMR_CIUM 1<<3
#define PCMR_IF 1<<6
int main() {
int number_of_pcs;
uint32_t pcmr, pccr, upr, pccfgr;
/* Check if PCs are even available */
upr = or1k_mfspr(OR1K_SPR_SYS_UPR_ADDR);
if (OR1K_SPR_SYS_UPR_PCUP_GET(upr)) { pccfgr = or1k_mfspr(OR1K_SPR_SYS_PCCFGR_ADDR);
number_of_pcs = OR1K_SPR_SYS_PCCFGR_NPC_GET(pccfgr) + 1;
printf ("We have %d program counters.\n", number_of_pcs);
pccr = or1k_mfspr(OR1K_SPR_PERF_PCCR_ADDR(0));
printf ("PCCR before setup %x\n", pccr);
/* Turn on counter and enable instruction fetch counting */
pcmr = or1k_mfspr(OR1K_SPR_PERF_PCMR_ADDR(0));
pcmr |= PCMR_CISM | PCMR_CIUM | PCMR_IF;
or1k_mtspr(OR1K_SPR_PERF_PCMR_ADDR(0), pcmr);
/* Read the PCCR after we are done */
printf ("Run a printf.");
pccr = or1k_mfspr(OR1K_SPR_PERF_PCCR_ADDR(0));
printf ("PCCR after printf %x\n", pccr);
} else {
printf ("No performance counters available.\n");}
return 0;
}
Quoted 454 lines · On Fri, Oct 18, 2019 at 12:01 AM <[email protected]> wrote:
Hi Stafford,
Yes, I am using newlib. I had discovered both files; these registers in
sprs.h/******************************/
/* Performance Counters Group */
/******************************/
#define
OR1K_SPR_PERF_GROUP0x07/* Performance Counters Count Registers */
#define
OR1K_SPR_PERF_PCCR_BASEOR1K_UNSIGNED(0x000)#define
OR1K_SPR_PERF_PCCR_COUNTOR1K_UNSIGNED(0x008)#define
OR1K_SPR_PERF_PCCR_STEPOR1K_UNSIGNED(0x001)#define
OR1K_SPR_PERF_PCCR_INDEX(N)(OR1K_SPR_PERF_PCCR_BASE+ ((N) *OR1K_SPR_PERF_PCCR_STEP))#define
OR1K_SPR_PERF_PCCR_ADDR(N)((OR1K_SPR_PERF_GROUP<<OR1K_SPR_GROUP_LSB) |OR1K_SPR_PERF_PCCR_INDEX(N))/* Performance Counters Mode Registers */
#define
OR1K_SPR_PERF_PCMR_BASEOR1K_UNSIGNED(0x008)#define
OR1K_SPR_PERF_PCMR_COUNTOR1K_UNSIGNED(0x008)#define
OR1K_SPR_PERF_PCMR_STEPOR1K_UNSIGNED(0x001)#define
OR1K_SPR_PERF_PCMR_INDEX(N)(OR1K_SPR_PERF_PCMR_BASE+ ((N)*OR1K_SPR_PERF_PCMR_STEP))#define
OR1K_SPR_PERF_PCMR_ADDR(N)((OR1K_SPR_PERF_GROUP<<OR1K_SPR_GROUP_LSB) |OR1K_SPR_PERF_PCMR_INDEX(N))/* Performance Counters Configuration */
#define
OR1K_SPR_SYS_PCCFGR_INDEXOR1K_UNSIGNED(0x008)#define
OR1K_SPR_SYS_PCCFGR_ADDROR1K_UNSIGNED(0x0008)/* Number of Performance Counters */
#define
OR1K_SPR_SYS_PCCFGR_NPC_LSB0#define
OR1K_SPR_SYS_PCCFGR_NPC_MSB2#define
OR1K_SPR_SYS_PCCFGR_NPC_BITS3#define
OR1K_SPR_SYS_PCCFGR_NPC_MASKOR1K_UNSIGNED(0x00000007)#define
OR1K_SPR_SYS_PCCFGR_NPC_GET(X)(((X) >> 0) &OR1K_UNSIGNED(0x00000007))#define
OR1K_SPR_SYS_PCCFGR_NPC_SET(X, Y)(((X) &OR1K_UNSIGNED(0xfffffff8)) | ((Y) << 0))And these functions in
support.hstatic inline void
or1k_mtspr(uint32_tspr,uint32_tvalue)static inline
uint32_tor1k_mfspr(uint32_tspr)Despite this I don’t have clear how to use it.
If I do:
or1k_mtspr(OR1K_SPR_SYS_PCCFGR_ADDR, 0) -> Does it allow me to configure the PCCFGR to one performance counter?Is This the same than
or1k_mtspr(OR1K_SPR_SYS_PCCFGR_ADDR,OR1K_SPR_SYS_PCCFGR_NPC_LSB) or DoOR1K_SPR_SYS_PCCFGR_NPC_LSB,OR1K_SPR_SYS_PCCFGR_NPC_MSB, etc. provide different functions on each one performance counter?What is the meaning of
PCCR_BASE,PCCR_COUNT,PCCR_STEP,PCMR_INDEX(N),PCMR_ADDR(N)? (the same for PCMR) (Is BASE the base address of all PCCR and ADDR the position of each one of them? ….whyPCMR_BASEand COUNT hasta de same valueOR1K_UNSIGNED(0x008)? ) Should I define first PCCFGR, second PCMR and last get PCCR?Thanks and sorry for the inconveniences.
Elisa
De: Stafford Horne <[email protected]>
Enviado el: jueves, 17 de octubre de 2019 7:14
Para: ecalvo at 2se.es; Openrisc <[email protected]>
Asunto: Re: PCCR and PCRM registers+cc mailing list,
Hi Elisa,
Which toolchain are you using? I guess newlib?
It has functions like
or1k_mfspr()andor1k_mtspr()seeor1k-support.handor1k-sprs.hheaders for details.-Stafford
On Wed, Oct 16, 2019, 8:40 PM <[email protected]> wrote:
Hi Stafford,
I am with PCCR and PCRM registers. I have seen that I can access from asm language, but there is functions to access from C? Have you got any example about their usage?
I have already confirmed my subscription to the mailing list.
Thanks
Elisa-----Mensaje original-----
De: Stafford Horne <[email protected]>
Enviado el: miércoles, 9 de octubre de 2019 13:38
Para: ecalvo at 2se.es; Julius Baxter <[email protected]>
Asunto: Re: other doubtHello Elisa,
If you simulate with Icarus or modelsim you will be able to measure pretty much the same performance characteristics as FPGA. So there is no need to go straight to FPGA.
In terms of my example C code is one option. You can also read timer data directly from the tick timer in assembly and achieve the same thing.
If you are interested we can CC the mailing list and get more opinions.
-Stafford
Quoted 301 lines · On Wed, Oct 9, 2019 at 5:09 PM <[email protected]> wrote:
Hi Stafford,
Nice to meet you and, first of all, thanks a lot for your guidance. I am new on this, and although there is some documentation, sometimes it is difficult some point which maybe it is basic.
Ok, to your comments. If.."A simulator like QEMU or or1ksim will not give and exact representation of the CPUs real time performance"...then...if I simulate directly the processor with modelsim, icarus or a similar tool...neither I get a real performance, don’t I? And values for the counters that you tell me to enable, neither are real, isn't it? should I execute it directly on the FPGA and it will depends on the implementation?
Ok, to C code. I have understood the dependency with toolchain.
Thanks a lot again.
Best regards,
Elisa-----Mensaje original-----
De: Stafford Horne <[email protected]> Enviado el: martes, 8 de octubre de 2019 16:18
Para: Julius Baxter <[email protected]>
CC: ecalvo at 2se.es
Asunto: Re: other doubtHi Elisa,
OpenRISC cpu's can run any algorithm, but how well it will perform depends on many things:
- Compiler optimization flags (i.e. -O3)
- Whether or not you are doing FPU instructions and have FPU enabled
- Whether or not you use multiply and divide and have these instructions enabled
- The frequency you are running
- Cache settings Icache Dcache
- The type of algorithm, does it require lots of data which will cause many cache misses?A simulator like QEMU or or1ksim will not give and exact representation of the CPUs real time performance. It can tell you which intructions will be executed, but not how fast those will run or how many pipeline stalls of cache misses will happen.
You can use the performance counters, they are supported in mor1kx if you enable them with the FEATURE_PERFCOUNTERS='ENABLED' parameter. They can help count how many events happen between certain events. Then you can combine them with a timer and watchpoints to detect how many times a loop can execute in 1000 clock cycles etc. Please read about PCCRn and PCMRn in the architecture manual.
It might be just as easy to use simple timing in a c program though, depending on the toolchain you use you can compare times between runs of your algorithm. i.e.
#include <time.h> #include <stdio.h> static long to_micro(struct timespec *time) { return (time->tv_sec * 1000000) + (time->tv_nsec / 1000); } int main() { int i, j = 0; struct timespec before, after; clock_gettime(CLOCK_MONOTONIC, &before); /* Super complex algorithm */ for (int i = 0; i < 100; i++) { j = (j+1) * (j+2); } clock_gettime(CLOCK_MONOTONIC, &after); printf("time to run algorithm %ld uSecs\n", to_micro(&after) -
to_micro(&before));return 0; }$ or1k-smh-linux-gnu-gcc timer.c $ ./glibc-build-scripts/qemu-or1k-libc ./a.out time to run algorithm 164 uSecsI hope it helps.
-Stafford
Quoted 225 lines · Julius Baxter
Hi,
No problem.
There are performance counters in the OpenRISC architecture but whether they're implemented in a particular implementation is another matter.
You can use these registers to measure various things the CPU is doing while it's executing. If you read the ISA document it'll tell you about them.
I'm CCing Stafford because he's the main OpenRISC man these days and probably knows about the state of the performance counter registers in various simulators and RTL implementations.
Cheers,
JuliusQuoted 205 lines · On Tue., 8 Oct. 2019, 10:43 pm , <[email protected]> wrote:
Hi Julius,
Sorry for bothering you again ☹. Can I do you other fast question related to openrisc? If not, ignore the email please.
Is there any way to characterize the type of application that I can run in openrisc? I mean, could you measure (with numbers) if an algorithm can be executed on it and the speed that it will achieve? Is it possible to do it using orksim?
Sorry because maybe it is so basic and general ☹
Thanks in advance
Elisa
*De:* Julius Baxter <[email protected]> *Enviado el:* lunes, 16 de septiembre de 2019 13:11
*Para:* ecalvo at 2se.es
*Asunto:* Re: Starting with OpenRISC - IOBsAlso! To let you know, if you're in Spain, we will soon be having our ORConf conference in Europe, and it's in Bordeaux, France, just across the border. :)
There are several people there who can help you get up to speed, one of whom is Stafford Horne who knows most about the OpenRISC IP lately. He will be presenting. If you can attend, it'd be helpful, I'm sure.
All info at https://orconf.org
Cheers,
Julius
On Mon, 16 Sep 2019 at 21:09, Julius Baxter
<[email protected]>
wrote:Hi Elisa,
Sorry for the delay in this response.
You should be using an SoC toplevel. FPGAs have everything you need on board like memories and IO blocks and lots of other FPGA fabric for you to implement other pieces of hardware.
FuseSoC provides a really nice and easy way to build an mor1kx design for the DE0 nano I believe:
https://github.com/olofk/de0_nano
That github page has a rough guide to getting it going.
If you need help I recommend posting to the OpenRISC mailing list and people will respond probably more promptly than I. (I recommend getting to know how to use mailing lists.
https://openrisc.io/communityThere are more resources here: https://openrisc.io/tutorials
I hope that's helpful.
Cheers,
Julius
On Wed, 11 Sep 2019 at 20:09, <[email protected]> wrote:
Hi Julius,
Thanks a lot for the quick answer.
Yes, this is the problem: I am using as top level the mor1kx module itself. You mean that I need to synthetize also in reconfigurable logic these cores, don’t you? I thought that I could have these elements as external in a development board.
Thanks again,
Cheers
Elisa
*De:* Julius Baxter <[email protected]> *Enviado el:* miércoles, 11 de septiembre de 2019 12:02
*Para:* ecalvo at 2se.es
*Asunto:* Re: Starting with OpenRISC - IOBsHi Elisa,
Thanks for getting in touch, that sounds like a cool project.
Can you tell me about the toplevel - are you using a system toplevel, or is your toplevel the mor1kx module itself?
If it's the latter, then that's not the best way to do it - you
need a system toplevel which instantiates memories and some reset
circuitry and likely some IO (UART, GPIO, JTAG debug, etc.) to talk to the outside world.Is that helpful?
Cheers,
Julius
On Wed, 11 Sep 2019 at 19:47, <[email protected]> wrote:
Dear Dr. Baxter,
My name is Elisa Calvo Gallego. I am writing you because I have started to work with OpenRISC in the framework of a research
project developed in the company where I am working (Space
Submicron Electronics, 2SE), and I am having some basic troubles. Could you help me?Although the FPGA that we are planning to use is larger, I have synthetized mor1kx for a DE0 nano board as first step (this is the board used in the majority of guides and tutorials). My problem is that the results that I have obtained are similar in area and resources, except for IOBs, which are more than available IOBs in the device. Do you know what I am doing wrong? Should I comment debug lines or something like that? I apologize if the question is immediate. I didn't find the answer and I'm new in this.
Thanks very much in advance.
Best regards,
Elisa
Hi Stafford,
Sorry for bothering you again.
The program is hanged accessing to PCMR and PCCR registers. Should I change other features apart from perfcounters?
Elisa
-----Mensaje original-----
De: Stafford Horne <[email protected]> Enviado el: jueves, 17 de octubre de 2019 23:42
Para: ecalvo at 2se.es
CC: Openrisc <[email protected]>
Asunto: Re: PCCR and PCRM registers
Hi Elisa,
Right, these are the functions. You only need to be concerned with:
OR1K_SPR_PERF_PCCR_ADDR(n)
OR1K_SPR_PERF_PCMR_ADDR(n)
OR1K_SPR_SYS_PCCFGR_ADDR
OR1K_SPR_SYS_PCCFGR_NPC_GET(x)
The others are used internally.
The PCCFGR is read only, it specifies how many performance counters your CPU has built in. It may be 0-7.
An example of how to use them:
#include <or1k-support.h>
#include <or1k-sprs.h>
#include <stdio.h>
#define PCMR_CISM 1<<3
#define PCMR_CIUM 1<<3
#define PCMR_IF 1<<6
int main() {
int number_of_pcs;
uint32_t pcmr, pccr, upr, pccfgr;
/* Check if PCs are even available */ upr = or1k_mfspr(OR1K_SPR_SYS_UPR_ADDR);
if (OR1K_SPR_SYS_UPR_PCUP_GET(upr)) { pccfgr = or1k_mfspr(OR1K_SPR_SYS_PCCFGR_ADDR);
number_of_pcs = OR1K_SPR_SYS_PCCFGR_NPC_GET(pccfgr) + 1;
printf ("We have %d program counters.\n", number_of_pcs);
pccr = or1k_mfspr(OR1K_SPR_PERF_PCCR_ADDR(0));
printf ("PCCR before setup %x\n", pccr);
/* Turn on counter and enable instruction fetch counting */
pcmr = or1k_mfspr(OR1K_SPR_PERF_PCMR_ADDR(0));
pcmr |= PCMR_CISM | PCMR_CIUM | PCMR_IF;
or1k_mtspr(OR1K_SPR_PERF_PCMR_ADDR(0), pcmr);
/* Read the PCCR after we are done */
printf ("Run a printf.");
pccr = or1k_mfspr(OR1K_SPR_PERF_PCCR_ADDR(0));
printf ("PCCR after printf %x\n", pccr);
} else {
printf ("No performance counters available.\n");}
return 0;
}
Quoted 462 lines · On Fri, Oct 18, 2019 at 12:01 AM <[email protected]> wrote:
Hi Stafford,
Yes, I am using newlib. I had discovered both files; these registers in
sprs.h/******************************/
/* Performance Counters Group */
/******************************/
#define
OR1K_SPR_PERF_GROUP0x07/* Performance Counters Count Registers */
#define
OR1K_SPR_PERF_PCCR_BASEOR1K_UNSIGNED(0x000)#define
OR1K_SPR_PERF_PCCR_COUNTOR1K_UNSIGNED(0x008)#define
OR1K_SPR_PERF_PCCR_STEPOR1K_UNSIGNED(0x001)#define
OR1K_SPR_PERF_PCCR_INDEX(N)(OR1K_SPR_PERF_PCCR_BASE+ ((N) *OR1K_SPR_PERF_PCCR_STEP))#define
OR1K_SPR_PERF_PCCR_ADDR(N)((OR1K_SPR_PERF_GROUP<<OR1K_SPR_GROUP_LSB) |OR1K_SPR_PERF_PCCR_INDEX(N))/* Performance Counters Mode Registers */
#define
OR1K_SPR_PERF_PCMR_BASEOR1K_UNSIGNED(0x008)#define
OR1K_SPR_PERF_PCMR_COUNTOR1K_UNSIGNED(0x008)#define
OR1K_SPR_PERF_PCMR_STEPOR1K_UNSIGNED(0x001)#define
OR1K_SPR_PERF_PCMR_INDEX(N)(OR1K_SPR_PERF_PCMR_BASE+ ((N)*OR1K_SPR_PERF_PCMR_STEP))#define
OR1K_SPR_PERF_PCMR_ADDR(N)((OR1K_SPR_PERF_GROUP<<OR1K_SPR_GROUP_LSB) |OR1K_SPR_PERF_PCMR_INDEX(N))/* Performance Counters Configuration */
#define
OR1K_SPR_SYS_PCCFGR_INDEXOR1K_UNSIGNED(0x008)#define
OR1K_SPR_SYS_PCCFGR_ADDROR1K_UNSIGNED(0x0008)/* Number of Performance Counters */
#define
OR1K_SPR_SYS_PCCFGR_NPC_LSB0#define
OR1K_SPR_SYS_PCCFGR_NPC_MSB2#define
OR1K_SPR_SYS_PCCFGR_NPC_BITS3#define
OR1K_SPR_SYS_PCCFGR_NPC_MASKOR1K_UNSIGNED(0x00000007)#define
OR1K_SPR_SYS_PCCFGR_NPC_GET(X)(((X) >> 0) &OR1K_UNSIGNED(0x00000007))#define
OR1K_SPR_SYS_PCCFGR_NPC_SET(X, Y)(((X) &OR1K_UNSIGNED(0xfffffff8)) | ((Y) << 0))And these functions in
support.hstatic inline void
or1k_mtspr(uint32_tspr,uint32_tvalue)static inline
uint32_tor1k_mfspr(uint32_tspr)Despite this I don’t have clear how to use it.
If I do:
or1k_mtspr(OR1K_SPR_SYS_PCCFGR_ADDR, 0) -> Does it allow me to configure the PCCFGR to one performance counter?Is This the same than
or1k_mtspr(OR1K_SPR_SYS_PCCFGR_ADDR,OR1K_SPR_SYS_PCCFGR_NPC_LSB) or DoOR1K_SPR_SYS_PCCFGR_NPC_LSB,OR1K_SPR_SYS_PCCFGR_NPC_MSB, etc. provide different functions on each one performance counter?What is the meaning of
PCCR_BASE,PCCR_COUNT,PCCR_STEP,PCMR_INDEX(N),PCMR_ADDR(N)? (the same for PCMR) (Is BASE the base address of all PCCR and ADDR the position of each one of them? ….whyPCMR_BASEand COUNT hasta de same valueOR1K_UNSIGNED(0x008)? ) Should I define first PCCFGR, second PCMR and last get PCCR?Thanks and sorry for the inconveniences.
Elisa
De: Stafford Horne <[email protected]>
Enviado el: jueves, 17 de octubre de 2019 7:14
Para: ecalvo at 2se.es; Openrisc <[email protected]>
Asunto: Re: PCCR and PCRM registers+cc mailing list,
Hi Elisa,
Which toolchain are you using? I guess newlib?
It has functions like
or1k_mfspr()andor1k_mtspr()seeor1k-support.handor1k-sprs.hheaders for details.-Stafford
On Wed, Oct 16, 2019, 8:40 PM <[email protected]> wrote:
Hi Stafford,
I am with PCCR and PCRM registers. I have seen that I can access from asm language, but there is functions to access from C? Have you got any example about their usage?
I have already confirmed my subscription to the mailing list.
Thanks
Elisa-----Mensaje original-----
De: Stafford Horne <[email protected]>
Enviado el: miércoles, 9 de octubre de 2019 13:38
Para: ecalvo at 2se.es; Julius Baxter <[email protected]>
Asunto: Re: other doubtHello Elisa,
If you simulate with Icarus or modelsim you will be able to measure pretty much the same performance characteristics as FPGA. So there is no need to go straight to FPGA.
In terms of my example C code is one option. You can also read timer data directly from the tick timer in assembly and achieve the same thing.
If you are interested we can CC the mailing list and get more opinions.
-Stafford
Quoted 302 lines · On Wed, Oct 9, 2019 at 5:09 PM <[email protected]> wrote:
Hi Stafford,
Nice to meet you and, first of all, thanks a lot for your guidance. I am new on this, and although there is some documentation, sometimes it is difficult some point which maybe it is basic.
Ok, to your comments. If.."A simulator like QEMU or or1ksim will not give and exact representation of the CPUs real time performance"...then...if I simulate directly the processor with modelsim, icarus or a similar tool...neither I get a real performance, don’t I? And values for the counters that you tell me to enable, neither are real, isn't it? should I execute it directly on the FPGA and it will depends on the implementation?
Ok, to C code. I have understood the dependency with toolchain.
Thanks a lot again.
Best regards,
Elisa-----Mensaje original-----
De: Stafford Horne <[email protected]> Enviado el: martes, 8 de octubre de 2019 16:18
Para: Julius Baxter <[email protected]>
CC: ecalvo at 2se.es
Asunto: Re: other doubtHi Elisa,
OpenRISC cpu's can run any algorithm, but how well it will perform depends on many things:
- Compiler optimization flags (i.e. -O3)
- Whether or not you are doing FPU instructions and have FPU enabled
- Whether or not you use multiply and divide and have these instructions enabled
- The frequency you are running
- Cache settings Icache Dcache
- The type of algorithm, does it require lots of data which will cause many cache misses?A simulator like QEMU or or1ksim will not give and exact representation of the CPUs real time performance. It can tell you which intructions will be executed, but not how fast those will run or how many pipeline stalls of cache misses will happen.
You can use the performance counters, they are supported in mor1kx if you enable them with the FEATURE_PERFCOUNTERS='ENABLED' parameter. They can help count how many events happen between certain events. Then you can combine them with a timer and watchpoints to detect how many times a loop can execute in 1000 clock cycles etc. Please read about PCCRn and PCMRn in the architecture manual.
It might be just as easy to use simple timing in a c program though, depending on the toolchain you use you can compare times between runs of your algorithm. i.e.
#include <time.h> #include <stdio.h> static long to_micro(struct timespec *time) { return (time->tv_sec * 1000000) + (time->tv_nsec / 1000); } int main() { int i, j = 0; struct timespec before, after; clock_gettime(CLOCK_MONOTONIC, &before); /* Super complex algorithm */ for (int i = 0; i < 100; i++) { j = (j+1) * (j+2); } clock_gettime(CLOCK_MONOTONIC, &after); printf("time to run algorithm %ld uSecs\n", to_micro(&after) -
to_micro(&before));return 0; }$ or1k-smh-linux-gnu-gcc timer.c $ ./glibc-build-scripts/qemu-or1k-libc ./a.out time to run algorithm 164 uSecsI hope it helps.
-Stafford
Quoted 226 lines · Julius Baxter
Hi,
No problem.
There are performance counters in the OpenRISC architecture but whether they're implemented in a particular implementation is another matter.
You can use these registers to measure various things the CPU is doing while it's executing. If you read the ISA document it'll tell you about them.
I'm CCing Stafford because he's the main OpenRISC man these days and probably knows about the state of the performance counter registers in various simulators and RTL implementations.
Cheers,
JuliusQuoted 206 lines · On Tue., 8 Oct. 2019, 10:43 pm , <[email protected]> wrote:
Hi Julius,
Sorry for bothering you again ☹. Can I do you other fast question related to openrisc? If not, ignore the email please.
Is there any way to characterize the type of application that I can run in openrisc? I mean, could you measure (with numbers) if an algorithm can be executed on it and the speed that it will achieve? Is it possible to do it using orksim?
Sorry because maybe it is so basic and general ☹
Thanks in advance
Elisa
*De:* Julius Baxter <[email protected]> *Enviado el:* lunes,
16 de septiembre de 2019 13:11
*Para:* ecalvo at 2se.es
*Asunto:* Re: Starting with OpenRISC - IOBsAlso! To let you know, if you're in Spain, we will soon be having our ORConf conference in Europe, and it's in Bordeaux, France, just across the border. :)
There are several people there who can help you get up to speed, one of whom is Stafford Horne who knows most about the OpenRISC
IP lately. He will be presenting. If you can attend, it'd be helpful, I'm sure.All info at https://orconf.org
Cheers,
Julius
On Mon, 16 Sep 2019 at 21:09, Julius Baxter
<[email protected]>
wrote:Hi Elisa,
Sorry for the delay in this response.
You should be using an SoC toplevel. FPGAs have everything you need on board like memories and IO blocks and lots of other FPGA fabric for you to implement other pieces of hardware.
FuseSoC provides a really nice and easy way to build an mor1kx design for the DE0 nano I believe:
https://github.com/olofk/de0_nano
That github page has a rough guide to getting it going.
If you need help I recommend posting to the OpenRISC mailing list and people will respond probably more promptly than I. (I recommend getting to know how to use mailing lists.
https://openrisc.io/communityThere are more resources here: https://openrisc.io/tutorials
I hope that's helpful.
Cheers,
Julius
On Wed, 11 Sep 2019 at 20:09, <[email protected]> wrote:
Hi Julius,
Thanks a lot for the quick answer.
Yes, this is the problem: I am using as top level the mor1kx module itself. You mean that I need to synthetize also in reconfigurable logic these cores, don’t you? I thought that I could have these elements as external in a development board.
Thanks again,
Cheers
Elisa
*De:* Julius Baxter <[email protected]> *Enviado el:* miércoles, 11 de septiembre de 2019 12:02
*Para:* ecalvo at 2se.es
*Asunto:* Re: Starting with OpenRISC - IOBsHi Elisa,
Thanks for getting in touch, that sounds like a cool project.
Can you tell me about the toplevel - are you using a system toplevel, or is your toplevel the mor1kx module itself?
If it's the latter, then that's not the best way to do it - you
need a system toplevel which instantiates memories and some
reset circuitry and likely some IO (UART, GPIO, JTAG debug, etc.) to talk to the outside world.Is that helpful?
Cheers,
Julius
On Wed, 11 Sep 2019 at 19:47, <[email protected]> wrote:
Dear Dr. Baxter,
My name is Elisa Calvo Gallego. I am writing you because I have started to work with OpenRISC in the framework of a research
project developed in the company where I am working (Space
Submicron Electronics, 2SE), and I am having some basic troubles. Could you help me?Although the FPGA that we are planning to use is larger, I have synthetized mor1kx for a DE0 nano board as first step (this is the board used in the majority of guides and tutorials). My problem is that the results that I have obtained are similar in area and resources, except for IOBs, which are more than available IOBs in the device. Do you know what I am doing wrong? Should I comment debug lines or something like that? I apologize if the question is immediate. I didn't find the answer and I'm new in this.
Thanks very much in advance.
Best regards,
Elisa
Hello,
Which core are you using? With mor1kx I did not have an issue with hanging. I had to enable perf counters via parameters.
I extended and posted my example code here:
https://gist.github.com/stffrdhrn/6343706cb1d8124bbac6bb579b6913b0
The results looks like:
Compile: or1k-elf-gcc or1k-perfcounters.c
Run: fusesoc run --target mor1kx_tb mor1kx-generic --elf-load ./a.out
Example Output:
We have 4 program counters.
PCCR before setup 0
Run a printf, to generate instructions..
IF 1825
ICS 237
This shows that the printf took 1825 instructions, 237 times the pipeline stalled due to instruction cache misses.
Quoted 535 lines · On Fri, Oct 18, 2019 at 7:51 PM <[email protected]> wrote:
Hi Stafford,
Sorry for bothering you again.
The program is hanged accessing to PCMR and PCCR registers. Should I change other features apart from perfcounters?
Elisa
-----Mensaje original-----
De: Stafford Horne <[email protected]> Enviado el: jueves, 17 de octubre de 2019 23:42
Para: ecalvo at 2se.es
CC: Openrisc <[email protected]>
Asunto: Re: PCCR and PCRM registersHi Elisa,
Right, these are the functions. You only need to be concerned with:
OR1K_SPR_PERF_PCCR_ADDR(n)
OR1K_SPR_PERF_PCMR_ADDR(n)
OR1K_SPR_SYS_PCCFGR_ADDR
OR1K_SPR_SYS_PCCFGR_NPC_GET(x)The others are used internally.
The PCCFGR is read only, it specifies how many performance counters your CPU has built in. It may be 0-7.An example of how to use them:
#include <
or1k-support.h>
#include <or1k-sprs.h>
#include <stdio.h>#define
PCMR_CISM1<<3
#definePCMR_CIUM1<<3
#definePCMR_IF1<<6int
main(){int
number_of_pcs;
uint32_tpcmr, pccr, upr, pccfgr;/* Check if PCs are even available */ upr =
or1k_mfspr(OR1K_SPR_SYS_UPR_ADDR);if (OR1K_SPR_SYS_UPR_PCUP_GET(upr)) {pccfgr = or1k_mfspr(OR1K_SPR_SYS_PCCFGR_ADDR); number_of_pcs = OR1K_SPR_SYS_PCCFGR_NPC_GET(pccfgr) + 1; printf ("We have %d program counters.\n", number_of_pcs); pccr = or1k_mfspr(OR1K_SPR_PERF_PCCR_ADDR(0)); printf ("PCCR before setup %x\n", pccr); /* Turn on counter and enable instruction fetch counting */ pcmr = or1k_mfspr(OR1K_SPR_PERF_PCMR_ADDR(0)); pcmr |= PCMR_CISM | PCMR_CIUM | PCMR_IF; or1k_mtspr(OR1K_SPR_PERF_PCMR_ADDR(0), pcmr); /* Read the PCCR after we are done */ printf ("Run a printf."); pccr = or1k_mfspr(OR1K_SPR_PERF_PCCR_ADDR(0)); printf ("PCCR after printf %x\n", pccr); } else { printf ("No performance counters available.\n");}
return 0;
}Quoted 462 lines · On Fri, Oct 18, 2019 at 12:01 AM <[email protected]> wrote:
Hi Stafford,
Yes, I am using newlib. I had discovered both files; these registers in
sprs.h/******************************/
/* Performance Counters Group */
/******************************/
#define
OR1K_SPR_PERF_GROUP0x07/* Performance Counters Count Registers */
#define
OR1K_SPR_PERF_PCCR_BASEOR1K_UNSIGNED(0x000)#define
OR1K_SPR_PERF_PCCR_COUNTOR1K_UNSIGNED(0x008)#define
OR1K_SPR_PERF_PCCR_STEPOR1K_UNSIGNED(0x001)#define
OR1K_SPR_PERF_PCCR_INDEX(N)(OR1K_SPR_PERF_PCCR_BASE+ ((N) *OR1K_SPR_PERF_PCCR_STEP))#define
OR1K_SPR_PERF_PCCR_ADDR(N)((OR1K_SPR_PERF_GROUP<<OR1K_SPR_GROUP_LSB) |OR1K_SPR_PERF_PCCR_INDEX(N))/* Performance Counters Mode Registers */
#define
OR1K_SPR_PERF_PCMR_BASEOR1K_UNSIGNED(0x008)#define
OR1K_SPR_PERF_PCMR_COUNTOR1K_UNSIGNED(0x008)#define
OR1K_SPR_PERF_PCMR_STEPOR1K_UNSIGNED(0x001)#define
OR1K_SPR_PERF_PCMR_INDEX(N)(OR1K_SPR_PERF_PCMR_BASE+ ((N)*OR1K_SPR_PERF_PCMR_STEP))#define
OR1K_SPR_PERF_PCMR_ADDR(N)((OR1K_SPR_PERF_GROUP<<OR1K_SPR_GROUP_LSB) |OR1K_SPR_PERF_PCMR_INDEX(N))/* Performance Counters Configuration */
#define
OR1K_SPR_SYS_PCCFGR_INDEXOR1K_UNSIGNED(0x008)#define
OR1K_SPR_SYS_PCCFGR_ADDROR1K_UNSIGNED(0x0008)/* Number of Performance Counters */
#define
OR1K_SPR_SYS_PCCFGR_NPC_LSB0#define
OR1K_SPR_SYS_PCCFGR_NPC_MSB2#define
OR1K_SPR_SYS_PCCFGR_NPC_BITS3#define
OR1K_SPR_SYS_PCCFGR_NPC_MASKOR1K_UNSIGNED(0x00000007)#define
OR1K_SPR_SYS_PCCFGR_NPC_GET(X)(((X) >> 0) &OR1K_UNSIGNED(0x00000007))#define
OR1K_SPR_SYS_PCCFGR_NPC_SET(X, Y)(((X) &OR1K_UNSIGNED(0xfffffff8)) | ((Y) << 0))And these functions in
support.hstatic inline void
or1k_mtspr(uint32_tspr,uint32_tvalue)static inline
uint32_tor1k_mfspr(uint32_tspr)Despite this I don’t have clear how to use it.
If I do:
or1k_mtspr(OR1K_SPR_SYS_PCCFGR_ADDR, 0) -> Does it allow me to configure the PCCFGR to one performance counter?Is This the same than
or1k_mtspr(OR1K_SPR_SYS_PCCFGR_ADDR,OR1K_SPR_SYS_PCCFGR_NPC_LSB) or DoOR1K_SPR_SYS_PCCFGR_NPC_LSB,OR1K_SPR_SYS_PCCFGR_NPC_MSB, etc. provide different functions on each one performance counter?What is the meaning of
PCCR_BASE,PCCR_COUNT,PCCR_STEP,PCMR_INDEX(N),PCMR_ADDR(N)? (the same for PCMR) (Is BASE the base address of all PCCR and ADDR the position of each one of them? ….whyPCMR_BASEand COUNT hasta de same valueOR1K_UNSIGNED(0x008)? ) Should I define first PCCFGR, second PCMR and last get PCCR?Thanks and sorry for the inconveniences.
Elisa
De: Stafford Horne <[email protected]>
Enviado el: jueves, 17 de octubre de 2019 7:14
Para: ecalvo at 2se.es; Openrisc <[email protected]>
Asunto: Re: PCCR and PCRM registers+cc mailing list,
Hi Elisa,
Which toolchain are you using? I guess newlib?
It has functions like
or1k_mfspr()andor1k_mtspr()seeor1k-support.handor1k-sprs.hheaders for details.-Stafford
On Wed, Oct 16, 2019, 8:40 PM <[email protected]> wrote:
Hi Stafford,
I am with PCCR and PCRM registers. I have seen that I can access from asm language, but there is functions to access from C? Have you got any example about their usage?
I have already confirmed my subscription to the mailing list.
Thanks
Elisa-----Mensaje original-----
De: Stafford Horne <[email protected]>
Enviado el: miércoles, 9 de octubre de 2019 13:38
Para: ecalvo at 2se.es; Julius Baxter <[email protected]>
Asunto: Re: other doubtHello Elisa,
If you simulate with Icarus or modelsim you will be able to measure pretty much the same performance characteristics as FPGA. So there is no need to go straight to FPGA.
In terms of my example C code is one option. You can also read timer data directly from the tick timer in assembly and achieve the same thing.
If you are interested we can CC the mailing list and get more opinions.
-Stafford
Quoted 302 lines · On Wed, Oct 9, 2019 at 5:09 PM <[email protected]> wrote:
Hi Stafford,
Nice to meet you and, first of all, thanks a lot for your guidance. I am new on this, and although there is some documentation, sometimes it is difficult some point which maybe it is basic.
Ok, to your comments. If.."A simulator like QEMU or or1ksim will not give and exact representation of the CPUs real time performance"...then...if I simulate directly the processor with modelsim, icarus or a similar tool...neither I get a real performance, don’t I? And values for the counters that you tell me to enable, neither are real, isn't it? should I execute it directly on the FPGA and it will depends on the implementation?
Ok, to C code. I have understood the dependency with toolchain.
Thanks a lot again.
Best regards,
Elisa-----Mensaje original-----
De: Stafford Horne <[email protected]> Enviado el: martes, 8 de octubre de 2019 16:18
Para: Julius Baxter <[email protected]>
CC: ecalvo at 2se.es
Asunto: Re: other doubtHi Elisa,
OpenRISC cpu's can run any algorithm, but how well it will perform depends on many things:
- Compiler optimization flags (i.e. -O3)
- Whether or not you are doing FPU instructions and have FPU enabled
- Whether or not you use multiply and divide and have these instructions enabled
- The frequency you are running
- Cache settings Icache Dcache
- The type of algorithm, does it require lots of data which will cause many cache misses?A simulator like QEMU or or1ksim will not give and exact representation of the CPUs real time performance. It can tell you which intructions will be executed, but not how fast those will run or how many pipeline stalls of cache misses will happen.
You can use the performance counters, they are supported in mor1kx if you enable them with the FEATURE_PERFCOUNTERS='ENABLED' parameter. They can help count how many events happen between certain events. Then you can combine them with a timer and watchpoints to detect how many times a loop can execute in 1000 clock cycles etc. Please read about PCCRn and PCMRn in the architecture manual.
It might be just as easy to use simple timing in a c program though, depending on the toolchain you use you can compare times between runs of your algorithm. i.e.
#include <time.h> #include <stdio.h> static long to_micro(struct timespec *time) { return (time->tv_sec * 1000000) + (time->tv_nsec / 1000); } int main() { int i, j = 0; struct timespec before, after; clock_gettime(CLOCK_MONOTONIC, &before); /* Super complex algorithm */ for (int i = 0; i < 100; i++) { j = (j+1) * (j+2); } clock_gettime(CLOCK_MONOTONIC, &after); printf("time to run algorithm %ld uSecs\n", to_micro(&after) -
to_micro(&before));return 0; }$ or1k-smh-linux-gnu-gcc timer.c $ ./glibc-build-scripts/qemu-or1k-libc ./a.out time to run algorithm 164 uSecsI hope it helps.
-Stafford
Quoted 226 lines · Julius Baxter
Hi,
No problem.
There are performance counters in the OpenRISC architecture but whether they're implemented in a particular implementation is another matter.
You can use these registers to measure various things the CPU is doing while it's executing. If you read the ISA document it'll tell you about them.
I'm CCing Stafford because he's the main OpenRISC man these days and probably knows about the state of the performance counter registers in various simulators and RTL implementations.
Cheers,
JuliusQuoted 206 lines · On Tue., 8 Oct. 2019, 10:43 pm , <[email protected]> wrote:
Hi Julius,
Sorry for bothering you again ☹. Can I do you other fast question related to openrisc? If not, ignore the email please.
Is there any way to characterize the type of application that I can run in openrisc? I mean, could you measure (with numbers) if an algorithm can be executed on it and the speed that it will achieve? Is it possible to do it using orksim?
Sorry because maybe it is so basic and general ☹
Thanks in advance
Elisa
*De:* Julius Baxter <[email protected]> *Enviado el:* lunes,
16 de septiembre de 2019 13:11
*Para:* ecalvo at 2se.es
*Asunto:* Re: Starting with OpenRISC - IOBsAlso! To let you know, if you're in Spain, we will soon be having our ORConf conference in Europe, and it's in Bordeaux, France, just across the border. :)
There are several people there who can help you get up to speed, one of whom is Stafford Horne who knows most about the OpenRISC
IP lately. He will be presenting. If you can attend, it'd be helpful, I'm sure.All info at https://orconf.org
Cheers,
Julius
On Mon, 16 Sep 2019 at 21:09, Julius Baxter
<[email protected]>
wrote:Hi Elisa,
Sorry for the delay in this response.
You should be using an SoC toplevel. FPGAs have everything you need on board like memories and IO blocks and lots of other FPGA fabric for you to implement other pieces of hardware.
FuseSoC provides a really nice and easy way to build an mor1kx design for the DE0 nano I believe:
https://github.com/olofk/de0_nano
That github page has a rough guide to getting it going.
If you need help I recommend posting to the OpenRISC mailing list and people will respond probably more promptly than I. (I recommend getting to know how to use mailing lists.
https://openrisc.io/communityThere are more resources here: https://openrisc.io/tutorials
I hope that's helpful.
Cheers,
Julius
On Wed, 11 Sep 2019 at 20:09, <[email protected]> wrote:
Hi Julius,
Thanks a lot for the quick answer.
Yes, this is the problem: I am using as top level the mor1kx module itself. You mean that I need to synthetize also in reconfigurable logic these cores, don’t you? I thought that I could have these elements as external in a development board.
Thanks again,
Cheers
Elisa
*De:* Julius Baxter <[email protected]> *Enviado el:* miércoles, 11 de septiembre de 2019 12:02
*Para:* ecalvo at 2se.es
*Asunto:* Re: Starting with OpenRISC - IOBsHi Elisa,
Thanks for getting in touch, that sounds like a cool project.
Can you tell me about the toplevel - are you using a system toplevel, or is your toplevel the mor1kx module itself?
If it's the latter, then that's not the best way to do it - you
need a system toplevel which instantiates memories and some
reset circuitry and likely some IO (UART, GPIO, JTAG debug, etc.) to talk to the outside world.Is that helpful?
Cheers,
Julius
On Wed, 11 Sep 2019 at 19:47, <[email protected]> wrote:
Dear Dr. Baxter,
My name is Elisa Calvo Gallego. I am writing you because I have started to work with OpenRISC in the framework of a research
project developed in the company where I am working (Space
Submicron Electronics, 2SE), and I am having some basic troubles. Could you help me?Although the FPGA that we are planning to use is larger, I have synthetized mor1kx for a DE0 nano board as first step (this is the board used in the majority of guides and tutorials). My problem is that the results that I have obtained are similar in area and resources, except for IOBs, which are more than available IOBs in the device. Do you know what I am doing wrong? Should I comment debug lines or something like that? I apologize if the question is immediate. I didn't find the answer and I'm new in this.
Thanks very much in advance.
Best regards,
Elisa
A message that is not in the archive (sent off-list or not yet received): [email protected]
A message that is not in the archive (sent off-list or not yet received): [email protected]
+ cc OpenRISC lost
Quoted 680 lines · Stafford Horne
Hi Elisa,
Quoted 5 lines · On Wed, Oct 23, 2019 at 09:47:24AM +0200, ecalvo at 2se.es wrote:
Hi Stafford,
Sorry for the delay in the answer. I have been out of the office.
I am using mor1kx 4.1. What I have done is change
FEATURE_PERFCOUNTERSto ENABLED in
orpsoc_top.cin the library of FuseSoc.. I have also changedOPTION_PERFCOUNTERS_NUMbut I have no seen changes in the output.The versions I am using are (as reported by busesoc):
INFO: Preparing ::adv_debug_sys:3.1.0-r1
INFO: Preparing ::cdc_utils:0.1
INFO: Preparing ::elf-loader:1.0.2
INFO: Preparing ::intgen:0
INFO: Preparing ::jtag_tap:1.13-r1
INFO: Preparing ::jtag_vpi:0-r4
INFO: Preparing ::mor1kx:5.0-r3
INFO: Preparing ::uart16550:1.5.5-r1
INFO: Preparing ::verilog-arbiter:0-r2
INFO: Preparing ::vlog_tb_utils:1.1
INFO: Preparing ::wb_common:1.0.3
INFO: Preparing ::wb_bfm:1.2.1
INFO: Preparing ::wb_intercon:1.2.2
INFO: Preparing ::wb_ram:1.1
INFO: Preparing ::mor1kx-generic:1.1The changes I made in mor1kx-generic are:
mor1kx-generic.core +1
-2@@ -10,7 +10,6 @@ filesets: marocchino_modules: depend: - or1k_marocchino - soc_files: files: - rtl/verilog/wb_intercon.vh: {is_include_file: true}@@ -31,7 +30,7 @@ filesets: file_type: verilogSource depend: - elf-loader - - "=jtag_vpi-r2" + - ">=jtag_vpi-r2" - ">=vlog_tb_utils-1.0" verilator_tb_files: files:rtl/verilog/orpsoc_top.v +2
-0@@ -266,6 +266,8 @@ mor1kx #( .OPTION_RF_NUM_SHADOW_GPR (option_rf_num_shadow_gpr), .IBUS_WB_TYPE ("B3_REGISTERED_FEEDBACK"), .DBUS_WB_TYPE ("B3_REGISTERED_FEEDBACK"), + .FEATURE_PERFCOUNTERS ("ENABLED"), + .OPTION_PERFCOUNTERS_NUM (3), .OPTION_CPU0 (pipeline), .OPTION_RESET_PC (32'h00000100) ) mor1kx0 (If using this version is not possible then it would be best that you look into
traces and try to get some insight to where things are getting stuck.-Stafford
Quoted 17 lines
Output:
We have 1 program counters.
(hanged)I have checked that features have been modified in the fuse build files.
Elisa
-----Mensaje original-----
De: Stafford Horne <[email protected]>
Enviado el: viernes, 18 de octubre de 2019 16:41
Para: ecalvo at 2se.es
CC: Openrisc <[email protected]>
Asunto: Re: PCCR and PCRM registersQuoted 3 lines
Hello,
Which core are you using? With mor1kx I did not have an issue with
hanging. I had to enable perf counters via parameters.
Quoted 11 lines
I extended and posted my example code here:
https://gist.github.com/stffrdhrn/6343706cb1d8124bbac6bb579b6913b0The results looks like:
Compile: or1k-elf-gcc
or1k-perfcounters.c
Run: fusesoc run --targetmor1kx_tbmor1kx-generic --elf-load ./a.outExample Output:
Quoted 7 lines
We have 4 program counters.
PCCR before setup 0
Run a printf, to generate instructions..
IF 1825
ICS 237This shows that the printf took 1825 instructions, 237 times the
pipeline stalled due to instruction cache misses.
Quoted 8 lines · On Fri, Oct 18, 2019 at 7:51 PM <[email protected]> wrote:
Quoted 6 lines
Hi Stafford,
Sorry for bothering you again.
The program is hanged accessing to PCMR and PCCR registers. Should I
change other features apart from perfcounters?
Quoted 12 lines
Quoted 12 lines
Elisa
-----Mensaje original-----
De: Stafford Horne <[email protected]> Enviado el: jueves, 17 de octubre de 2019 23:42
Para: ecalvo at 2se.es
CC: Openrisc <[email protected]>
Asunto: Re: PCCR and PCRM registersHi Elisa,
Quoted 8 lines
Quoted 8 lines
Right, these are the functions. You only need to be concerned with:
OR1K_SPR_PERF_PCCR_ADDR(n)
OR1K_SPR_PERF_PCMR_ADDR(n)
OR1K_SPR_SYS_PCCFGR_ADDR
OR1K_SPR_SYS_PCCFGR_NPC_GET(x)The others are used internally.
The PCCFGR is read only, it specifies how many performance countersyour CPU has built in. It may be 0-7.
Quoted 135 lines
Quoted 135 lines
An example of how to use them:
#include <
or1k-support.h>
#include <or1k-sprs.h>
#include <stdio.h>#define
PCMR_CISM1<<3
#definePCMR_CIUM1<<3
#definePCMR_IF1<<6int
main(){int
number_of_pcs;
uint32_tpcmr, pccr, upr, pccfgr;/* Check if PCs are even available */ upr =
or1k_mfspr(OR1K_SPR_SYS_UPR_ADDR);if (OR1K_SPR_SYS_UPR_PCUP_GET(upr)) {pccfgr = or1k_mfspr(OR1K_SPR_SYS_PCCFGR_ADDR); number_of_pcs = OR1K_SPR_SYS_PCCFGR_NPC_GET(pccfgr) + 1; printf ("We have %d program counters.\n", number_of_pcs); pccr = or1k_mfspr(OR1K_SPR_PERF_PCCR_ADDR(0)); printf ("PCCR before setup %x\n", pccr); /* Turn on counter and enable instruction fetch counting */ pcmr = or1k_mfspr(OR1K_SPR_PERF_PCMR_ADDR(0)); pcmr |= PCMR_CISM | PCMR_CIUM | PCMR_IF; or1k_mtspr(OR1K_SPR_PERF_PCMR_ADDR(0), pcmr); /* Read the PCCR after we are done */ printf ("Run a printf."); pccr = or1k_mfspr(OR1K_SPR_PERF_PCCR_ADDR(0)); printf ("PCCR after printf %x\n", pccr); } else { printf ("No performance counters available.\n");}
return 0;
}Quoted 89 lines · On Fri, Oct 18, 2019 at 12:01 AM <[email protected]> wrote:
Hi Stafford,
Yes, I am using newlib. I had discovered both files; these registers in
sprs.h/******************************/
/* Performance Counters Group */
/******************************/
#define
OR1K_SPR_PERF_GROUP0x07/* Performance Counters Count Registers */
#define
OR1K_SPR_PERF_PCCR_BASEOR1K_UNSIGNED(0x000)#define
OR1K_SPR_PERF_PCCR_COUNTOR1K_UNSIGNED(0x008)#define
OR1K_SPR_PERF_PCCR_STEPOR1K_UNSIGNED(0x001)#define
OR1K_SPR_PERF_PCCR_INDEX(N)(OR1K_SPR_PERF_PCCR_BASE+ ((N) *OR1K_SPR_PERF_PCCR_STEP))#define
OR1K_SPR_PERF_PCCR_ADDR(N)((OR1K_SPR_PERF_GROUP<<OR1K_SPR_GROUP_LSB) |OR1K_SPR_PERF_PCCR_INDEX(N))/* Performance Counters Mode Registers */
#define
OR1K_SPR_PERF_PCMR_BASEOR1K_UNSIGNED(0x008)#define
OR1K_SPR_PERF_PCMR_COUNTOR1K_UNSIGNED(0x008)#define
OR1K_SPR_PERF_PCMR_STEPOR1K_UNSIGNED(0x001)#define
OR1K_SPR_PERF_PCMR_INDEX(N)(OR1K_SPR_PERF_PCMR_BASE+ ((N)*OR1K_SPR_PERF_PCMR_STEP))#define
OR1K_SPR_PERF_PCMR_ADDR(N)((OR1K_SPR_PERF_GROUP<<OR1K_SPR_GROUP_LSB) |OR1K_SPR_PERF_PCMR_INDEX(N))/* Performance Counters Configuration */
#define
OR1K_SPR_SYS_PCCFGR_INDEXOR1K_UNSIGNED(0x008)#define
OR1K_SPR_SYS_PCCFGR_ADDROR1K_UNSIGNED(0x0008)/* Number of Performance Counters */
#define
OR1K_SPR_SYS_PCCFGR_NPC_LSB0#define
OR1K_SPR_SYS_PCCFGR_NPC_MSB2#define
OR1K_SPR_SYS_PCCFGR_NPC_BITS3#define
OR1K_SPR_SYS_PCCFGR_NPC_MASKOR1K_UNSIGNED(0x00000007)#define
OR1K_SPR_SYS_PCCFGR_NPC_GET(X)(((X) >> 0) &OR1K_UNSIGNED(0x00000007))#define
OR1K_SPR_SYS_PCCFGR_NPC_SET(X, Y)(((X) &OR1K_UNSIGNED(0xfffffff8)) | ((Y) << 0))And these functions in
support.hstatic inline void
or1k_mtspr(uint32_tspr,uint32_tvalue)static inline
uint32_tor1k_mfspr(uint32_tspr)Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Despite this I don’t have clear how to use it.
If I do:
or1k_mtspr(OR1K_SPR_SYS_PCCFGR_ADDR, 0) -> Does it allowme to configure the PCCFGR to one performance counter?
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Is This the same than
or1k_mtspr(OR1K_SPR_SYS_PCCFGR_ADDR,
OR1K_SPR_SYS_PCCFGR_NPC_LSB) or DoOR1K_SPR_SYS_PCCFGR_NPC_LSB,OR1K_SPR_SYS_PCCFGR_NPC_MSB, etc. provide different functions on each one performance counter?Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
What is the meaning of
PCCR_BASE,PCCR_COUNT,PCCR_STEP,PCMR_INDEX(N),PCMR_ADDR(N)? (the same for PCMR) (Is BASE the baseaddress of all PCCR and ADDR the position of each one of them? ….why
PCMR_BASEand COUNT hasta de same valueOR1K_UNSIGNED(0x008)? ) Should I define first PCCFGR, second PCMR and last get PCCR?Quoted 24 lines
Quoted 24 lines
Quoted 24 lines
Thanks and sorry for the inconveniences.
Elisa
De: Stafford Horne <[email protected]> Enviado el: jueves, 17 de octubre de 2019 7:14
Para: ecalvo at 2se.es; Openrisc <[email protected]>
Asunto: Re: PCCR and PCRM registers+cc mailing list,
Hi Elisa,
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Which toolchain are you using? I guess newlib?
It has functions like
or1k_mfspr()andor1k_mtspr()see
or1k-support.handor1k-sprs.hheaders for details.Quoted 10 lines
Quoted 10 lines
Quoted 10 lines
-Stafford
On Wed, Oct 16, 2019, 8:40 PM <[email protected]> wrote:
Hi Stafford,
I am with PCCR and PCRM registers. I have seen that I can access
from asm language, but there is functions to access from C? Have you got any example about their usage?
Quoted 12 lines
Quoted 12 lines
Quoted 12 lines
I have already confirmed my subscription to the mailing list.
Thanks
Elisa-----Mensaje original-----
De: Stafford Horne <[email protected]> Enviado el: miércoles, 9 de octubre de 2019 13:38
Para: ecalvo at 2se.es; Julius Baxter <[email protected]>
Asunto: Re: other doubtQuoted 3 lines
Quoted 3 lines
Quoted 3 lines
Hello Elisa,
If you simulate with Icarus or modelsim you will be able to measure
pretty much the same performance characteristics as FPGA. So there is no need to go straight to FPGA.
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
In terms of my example C code is one option. You can also read
timer data directly from the tick timer in assembly and achieve the same thing.
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
If you are interested we can CC the mailing list and get more
opinions.
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
-Stafford
Quoted 4 lines · On Wed, Oct 9, 2019 at 5:09 PM <[email protected]> wrote:
Hi Stafford,
Nice to meet you and, first of all, thanks a lot for your
guidance. I am new on this, and although there is some documentation, sometimes it is difficult some point which maybe it is basic.
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Ok, to your comments. If.."A simulator like QEMU or or1ksim will
not give and exact representation of the CPUs real time performance"...then...if I simulate directly the processor with modelsim, icarus or a similar tool...neither I get a real performance, don’t I? And values for the counters that you tell me to enable, neither are real, isn't it? should I execute it directly on the FPGA and it will depends on the implementation?
Quoted 15 lines
Quoted 15 lines
Quoted 15 lines
Quoted 15 lines
Ok, to C code. I have understood the dependency with toolchain.
Thanks a lot again.
Best regards,
Elisa-----Mensaje original-----
De: Stafford Horne <[email protected]> Enviado el: martes, 8 de octubre de 2019 16:18
Para: Julius Baxter <[email protected]>
CC: ecalvo at 2se.es
Asunto: Re: other doubtQuoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Hi Elisa,
OpenRISC cpu's can run any algorithm, but how well it will perform
depends on many things:
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
- Compiler optimization flags (i.e. -O3)
- Whether or not you are doing FPU instructions and have FPUenabled
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
- Whether or not you use multiply and divide and have these
instructions
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
enabled
- The frequency you are running
- Cache settings Icache Dcache
- The type of algorithm, does it require lots of data which willcause many
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
cache misses?
A simulator like QEMU or or1ksim will not give and exact
representation of the CPUs real time performance. It can tell you which intructions will be executed, but not how fast those will run or how many pipeline stalls of cache misses will happen.
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
You can use the performance counters, they are supported in mor1kx
if you enable them with the FEATURE_PERFCOUNTERS='ENABLED' parameter. They can help count how many events happen between certain events. Then you can combine them with a timer and watchpoints to detect how many times a loop can execute in 1000 clock cycles etc. Please read about PCCRn and PCMRn in the architecture manual.
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
It might be just as easy to use simple timing in a c program
though, depending on the toolchain you use you can compare times between runs of your algorithm.
Quoted 36 lines
Quoted 36 lines
Quoted 36 lines
Quoted 36 lines
i.e.
#include <time.h> #include <stdio.h> static long to_micro(struct timespec *time) { return (time->tv_sec * 1000000) + (time->tv_nsec / 1000); } int main() { int i, j = 0; struct timespec before, after; clock_gettime(CLOCK_MONOTONIC, &before); /* Super complex algorithm */ for (int i = 0; i < 100; i++) { j = (j+1) * (j+2); } clock_gettime(CLOCK_MONOTONIC, &after); printf("time to run algorithm %ld uSecs\n", to_micro(&after)-
to_micro(&before));return 0; }$ or1k-smh-linux-gnu-gcc timer.c $ ./glibc-build-scripts/qemu-or1k-libc ./a.out time to run algorithm 164 uSecsI hope it helps.
-Stafford
Quoted 7 lines
Quoted 7 lines
Quoted 7 lines
Quoted 7 lines · Julius Baxter
Quoted 6 lines
Hi,
No problem.
There are performance counters in the OpenRISC architecture but whether they're implemented in a particular implementation is
another matter.
Quoted 18 lines
Quoted 18 lines
Quoted 18 lines
Quoted 18 lines
Quoted 18 lines
You can use these registers to measure various things the CPU is doing while it's executing. If you read the ISA document it'll tell you about them.
I'm CCing Stafford because he's the main OpenRISC man these days and probably knows about the state of the performance counter registers in various simulators and RTL implementations.
Cheers,
JuliusQuoted 4 lines · On Tue., 8 Oct. 2019, 10:43 pm , <[email protected]> wrote:
Hi Julius,
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Sorry for bothering you again ☹. Can I do you other fast question related to openrisc? If not, ignore the email please.
Is there any way to characterize the type of application that I can run in openrisc? I mean, could you measure (with numbers) if an algorithm can be executed on it and the speed
that it will achieve?
Quoted 30 lines
Quoted 30 lines
Quoted 30 lines
Quoted 30 lines
Quoted 30 lines
Quoted 30 lines
Is it possible to do it using orksim?
Sorry because maybe it is so basic and general ☹
Thanks in advance
Elisa
*De:* Julius Baxter <[email protected]> *Enviado el:* lunes,
16 de septiembre de 2019 13:11
*Para:* ecalvo at 2se.es
*Asunto:* Re: Starting with OpenRISC - IOBsAlso! To let you know, if you're in Spain, we will soon be having our ORConf conference in Europe, and it's in Bordeaux, France, just across the border. :)
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
There are several people there who can help you get up to speed, one of whom is Stafford Horne who knows most about the OpenRISC IP lately. He will be presenting. If you can attend,
it'd be helpful, I'm sure.
Quoted 107 lines
Quoted 107 lines
Quoted 107 lines
Quoted 107 lines
Quoted 107 lines
Quoted 107 lines
All info at https://orconf.org
Cheers,
Julius
On Mon, 16 Sep 2019 at 21:09, Julius Baxter
<[email protected]>
wrote:Hi Elisa,
Sorry for the delay in this response.
You should be using an SoC toplevel. FPGAs have everything you need on board like memories and IO blocks and lots of other FPGA fabric for you to implement other pieces of hardware.
FuseSoC provides a really nice and easy way to build an mor1kx design for the DE0 nano I believe:
https://github.com/olofk/de0_nano
That github page has a rough guide to getting it going.
If you need help I recommend posting to the OpenRISC mailing list and people will respond probably more promptly than I. (I recommend getting to know how to use mailing lists.
https://openrisc.io/communityThere are more resources here: https://openrisc.io/tutorials
I hope that's helpful.
Cheers,
Julius
On Wed, 11 Sep 2019 at 20:09, <[email protected]> wrote:
Hi Julius,
Thanks a lot for the quick answer.
Yes, this is the problem: I am using as top level the mor1kx module itself. You mean that I need to synthetize also in reconfigurable logic these cores, don’t you? I thought that I could have these elements as external in a development board.
Thanks again,
Cheers
Elisa
*De:* Julius Baxter <[email protected]> *Enviado el:* miércoles, 11 de septiembre de 2019 12:02
*Para:* ecalvo at 2se.es
*Asunto:* Re: Starting with OpenRISC - IOBsHi Elisa,
Thanks for getting in touch, that sounds like a cool project.
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Can you tell me about the toplevel - are you using a system toplevel, or is your toplevel the mor1kx module itself?
If it's the latter, then that's not the best way to do it - you need a system toplevel which instantiates memories and some reset circuitry and likely some IO (UART, GPIO, JTAG
debug, etc.) to talk to the outside world.
Quoted 15 lines
Quoted 15 lines
Quoted 15 lines
Quoted 15 lines
Quoted 15 lines
Quoted 15 lines
Is that helpful?
Cheers,
Julius
On Wed, 11 Sep 2019 at 19:47, <[email protected]> wrote:
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Dear Dr. Baxter,
My name is Elisa Calvo Gallego. I am writing you because I have started to work with OpenRISC in the framework of a research project developed in the company where I am working (Space Submicron Electronics, 2SE), and I am having some basic
troubles. Could you help me?
Quoted 9 lines
Quoted 9 lines
Quoted 9 lines
Quoted 9 lines
Quoted 9 lines
Quoted 9 lines
Although the FPGA that we are planning to use is larger, I have synthetized mor1kx for a DE0 nano board as first step (this is the board used in the majority of guides and tutorials). My problem is that the results that I have obtained are similar in area and resources, except for IOBs, which are more than available IOBs in the device. Do you know
Quoted 20 lines
what I am doing wrong?
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Should I comment debug lines or something like that? I apologize if the question is immediate. I didn't find the
answer and I'm new in this.
Quoted 15 lines
Quoted 14 lines
Quoted 13 lines
Quoted 13 lines
Quoted 12 lines
Quoted 12 lines
Thanks very much in advance.
Best regards,
Elisa
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.librecores.org/pipermail/openrisc/attachments/20191024/e74c375f/attachment-0001.html>
Okkk. I will check version numbers first and, if not, I will look into traces.
I will come back to you if I discover the problem.
Thanks a lot Stafford.
-Elisa
De: Stafford Horne <[email protected]>
Enviado el: jueves, 24 de octubre de 2019 1:06
Para: ecalvo at 2se.es
CC: Openrisc <[email protected]>
Asunto: Re: PCCR and PCRM registers
+ cc OpenRISC lost
On Thu, Oct 24, 2019, 6:05 AM Stafford Horne <shorne at gmail.com <mailto:[email protected]> > wrote:
Hi Elisa,
Quoted 5 lines · On Wed, Oct 23, 2019 at 09:47:24AM +0200, ecalvo at 2se.es…
Hi Stafford,
Sorry for the delay in the answer. I have been out of the office.
I am using mor1kx 4.1. What I have done is change
FEATURE_PERFCOUNTERSto ENABLED inorpsoc_top.cin the library of FuseSoc.. I have also changedOPTION_PERFCOUNTERS_NUMbut I have no seen changes in the output.
The versions I am using are (as reported by busesoc):
INFO: Preparing ::adv_debug_sys:3.1.0-r1
INFO: Preparing ::cdc_utils:0.1
INFO: Preparing ::elf-loader:1.0.2
INFO: Preparing ::intgen:0
INFO: Preparing ::jtag_tap:1.13-r1
INFO: Preparing ::jtag_vpi:0-r4
INFO: Preparing ::mor1kx:5.0-r3
INFO: Preparing ::uart16550:1.5.5-r1
INFO: Preparing ::verilog-arbiter:0-r2
INFO: Preparing ::vlog_tb_utils:1.1
INFO: Preparing ::wb_common:1.0.3
INFO: Preparing ::wb_bfm:1.2.1
INFO: Preparing ::wb_intercon:1.2.2
INFO: Preparing ::wb_ram:1.1
INFO: Preparing ::mor1kx-generic:1.1
The changes I made in mor1kx-generic are:
mor1kx-generic.core +1 -2
@@ -10,7 +10,6 @@ filesets: marocchino_modules: depend: - or1k_marocchino - soc_files: files: - rtl/verilog/wb_intercon.vh: {is_include_file: true}
@@ -31,7 +30,7 @@ filesets: file_type: verilogSource depend: - elf-loader - - "=jtag_vpi-r2" + - ">=jtag_vpi-r2" - ">=vlog_tb_utils-1.0" verilator_tb_files: files:
rtl/verilog/orpsoc_top.v +2 -0
@@ -266,6 +266,8 @@ mor1kx #( .OPTION_RF_NUM_SHADOW_GPR (option_rf_num_shadow_gpr), .IBUS_WB_TYPE ("B3_REGISTERED_FEEDBACK"), .DBUS_WB_TYPE ("B3_REGISTERED_FEEDBACK"), + .FEATURE_PERFCOUNTERS ("ENABLED"), + .OPTION_PERFCOUNTERS_NUM (3), .OPTION_CPU0 (pipeline), .OPTION_RESET_PC (32'h00000100) ) mor1kx0 (
If using this version is not possible then it would be best that you look into traces and try to get some insight to where things are getting stuck.
-Stafford
Quoted 37 lines
Output:
We have 1 program counters.
(hanged)I have checked that features have been modified in the fuse build files.
Elisa
-----Mensaje original-----
De: Stafford Horne <shorne at gmail.com <mailto:[email protected]> > Enviado el: viernes, 18 de octubre de 2019 16:41
Para: ecalvo at 2se.es <mailto:[email protected]>
CC: Openrisc <openrisc at lists.librecores.org <mailto:[email protected]> >
Asunto: Re: PCCR and PCRM registersHello,
Which core are you using? With mor1kx I did not have an issue with hanging. I had to enable perf counters via parameters.
I extended and posted my example code here:
https://gist.github.com/stffrdhrn/6343706cb1d8124bbac6bb579b6913b0The results looks like:
Compile: or1k-elf-gcc
or1k-perfcounters.c
Run: fusesoc run --targetmor1kx_tbmor1kx-generic --elf-load ./a.outExample Output:
We have 4 program counters.
PCCR before setup 0
Run a printf, to generate instructions..
IF 1825
ICS 237
Quoted 541 lines
This shows that the printf took 1825 instructions, 237 times the pipeline stalled due to instruction cache misses.
Quoted 537 lines · On Fri, Oct 18, 2019 at 7:51 PM <ecalvo at 2se.es <mailto:[email protected]> >…
Hi Stafford,
Sorry for bothering you again.
The program is hanged accessing to PCMR and PCCR registers. Should I change other features apart from perfcounters?
Elisa
-----Mensaje original-----
De: Stafford Horne <shorne at gmail.com <mailto:[email protected]> > Enviado el: jueves, 17 de octubre de 2019 23:42
Para: ecalvo at 2se.es <mailto:[email protected]>
CC: Openrisc <openrisc at lists.librecores.org <mailto:[email protected]> >
Asunto: Re: PCCR and PCRM registersHi Elisa,
Right, these are the functions. You only need to be concerned with:
OR1K_SPR_PERF_PCCR_ADDR(n)
OR1K_SPR_PERF_PCMR_ADDR(n)
OR1K_SPR_SYS_PCCFGR_ADDR
OR1K_SPR_SYS_PCCFGR_NPC_GET(x)The others are used internally.
The PCCFGR is read only, it specifies how many performance counters your CPU has built in. It may be 0-7.An example of how to use them:
#include <
or1k-support.h>
#include <or1k-sprs.h>
#include <stdio.h>#define
PCMR_CISM1<<3
#definePCMR_CIUM1<<3
#definePCMR_IF1<<6int
main(){int
number_of_pcs;
uint32_tpcmr, pccr, upr, pccfgr;/* Check if PCs are even available */
upr =or1k_mfspr(OR1K_SPR_SYS_UPR_ADDR);if (OR1K_SPR_SYS_UPR_PCUP_GET(upr)) {pccfgr = or1k_mfspr(OR1K_SPR_SYS_PCCFGR_ADDR); number_of_pcs = OR1K_SPR_SYS_PCCFGR_NPC_GET(pccfgr) + 1; printf ("We have %d program counters.\n", number_of_pcs); pccr = or1k_mfspr(OR1K_SPR_PERF_PCCR_ADDR(0)); printf ("PCCR before setup %x\n", pccr); /* Turn on counter and enable instruction fetch counting */ pcmr = or1k_mfspr(OR1K_SPR_PERF_PCMR_ADDR(0)); pcmr |= PCMR_CISM | PCMR_CIUM | PCMR_IF; or1k_mtspr(OR1K_SPR_PERF_PCMR_ADDR(0), pcmr); /* Read the PCCR after we are done */ printf ("Run a printf."); pccr = or1k_mfspr(OR1K_SPR_PERF_PCCR_ADDR(0)); printf ("PCCR after printf %x\n", pccr); } else { printf ("No performance counters available.\n");}
return 0;
}Quoted 464 lines · On Fri, Oct 18, 2019 at 12:01 AM <ecalvo at 2se.es <mailto:[email protected]> >…
Hi Stafford,
Yes, I am using newlib. I had discovered both files; these registers in
sprs.h/******************************/
/* Performance Counters Group */
/******************************/
#define
OR1K_SPR_PERF_GROUP0x07/* Performance Counters Count Registers */
#define
OR1K_SPR_PERF_PCCR_BASEOR1K_UNSIGNED(0x000)#define
OR1K_SPR_PERF_PCCR_COUNTOR1K_UNSIGNED(0x008)#define
OR1K_SPR_PERF_PCCR_STEPOR1K_UNSIGNED(0x001)#define
OR1K_SPR_PERF_PCCR_INDEX(N)(OR1K_SPR_PERF_PCCR_BASE+ ((N) *OR1K_SPR_PERF_PCCR_STEP))#define
OR1K_SPR_PERF_PCCR_ADDR(N)((OR1K_SPR_PERF_GROUP<<OR1K_SPR_GROUP_LSB) |OR1K_SPR_PERF_PCCR_INDEX(N))/* Performance Counters Mode Registers */
#define
OR1K_SPR_PERF_PCMR_BASEOR1K_UNSIGNED(0x008)#define
OR1K_SPR_PERF_PCMR_COUNTOR1K_UNSIGNED(0x008)#define
OR1K_SPR_PERF_PCMR_STEPOR1K_UNSIGNED(0x001)#define
OR1K_SPR_PERF_PCMR_INDEX(N)(OR1K_SPR_PERF_PCMR_BASE+ ((N)*OR1K_SPR_PERF_PCMR_STEP))#define
OR1K_SPR_PERF_PCMR_ADDR(N)((OR1K_SPR_PERF_GROUP<<OR1K_SPR_GROUP_LSB) |OR1K_SPR_PERF_PCMR_INDEX(N))/* Performance Counters Configuration */
#define
OR1K_SPR_SYS_PCCFGR_INDEXOR1K_UNSIGNED(0x008)#define
OR1K_SPR_SYS_PCCFGR_ADDROR1K_UNSIGNED(0x0008)/* Number of Performance Counters */
#define
OR1K_SPR_SYS_PCCFGR_NPC_LSB0#define
OR1K_SPR_SYS_PCCFGR_NPC_MSB2#define
OR1K_SPR_SYS_PCCFGR_NPC_BITS3#define
OR1K_SPR_SYS_PCCFGR_NPC_MASKOR1K_UNSIGNED(0x00000007)#define
OR1K_SPR_SYS_PCCFGR_NPC_GET(X)(((X) >> 0) &OR1K_UNSIGNED(0x00000007))#define
OR1K_SPR_SYS_PCCFGR_NPC_SET(X, Y)(((X) &OR1K_UNSIGNED(0xfffffff8)) | ((Y) << 0))And these functions in
support.hstatic inline void
or1k_mtspr(uint32_tspr,uint32_tvalue)static inline
uint32_tor1k_mfspr(uint32_tspr)Despite this I don’t have clear how to use it.
If I do:
or1k_mtspr(OR1K_SPR_SYS_PCCFGR_ADDR, 0) -> Does it allow me to configure the PCCFGR to one performance counter?Is This the same than
or1k_mtspr(OR1K_SPR_SYS_PCCFGR_ADDR,OR1K_SPR_SYS_PCCFGR_NPC_LSB) or DoOR1K_SPR_SYS_PCCFGR_NPC_LSB,OR1K_SPR_SYS_PCCFGR_NPC_MSB, etc. provide different functions on each one performance counter?What is the meaning of
PCCR_BASE,PCCR_COUNT,PCCR_STEP,PCMR_INDEX(N),PCMR_ADDR(N)? (the same for PCMR) (Is BASE the base address of all PCCR and ADDR the position of each one of them? ….whyPCMR_BASEand COUNT hasta de same valueOR1K_UNSIGNED(0x008)? ) Should I define first PCCFGR, second PCMR and last get PCCR?Thanks and sorry for the inconveniences.
Elisa
De: Stafford Horne <shorne at gmail.com <mailto:[email protected]> > Enviado el: jueves, 17 de
octubre de 2019 7:14
Para: ecalvo at 2se.es <mailto:[email protected]> ; Openrisc <openrisc at lists.librecores.org <mailto:[email protected]> >
Asunto: Re: PCCR and PCRM registers+cc mailing list,
Hi Elisa,
Which toolchain are you using? I guess newlib?
It has functions like
or1k_mfspr()andor1k_mtspr()seeor1k-support.handor1k-sprs.hheaders for details.-Stafford
On Wed, Oct 16, 2019, 8:40 PM <ecalvo at 2se.es <mailto:[email protected]> > wrote:
Hi Stafford,
I am with PCCR and PCRM registers. I have seen that I can access from asm language, but there is functions to access from C? Have you got any example about their usage?
I have already confirmed my subscription to the mailing list.
Thanks
Elisa-----Mensaje original-----
De: Stafford Horne <shorne at gmail.com <mailto:[email protected]> > Enviado el: miércoles, 9 de octubre de 2019 13:38
Para: ecalvo at 2se.es <mailto:[email protected]> ; Julius Baxter <juliusbaxter at gmail.com <mailto:[email protected]> >
Asunto: Re: other doubtHello Elisa,
If you simulate with Icarus or modelsim you will be able to measure pretty much the same performance characteristics as FPGA. So there is no need to go straight to FPGA.
In terms of my example C code is one option. You can also read timer data directly from the tick timer in assembly and achieve the same thing.
If you are interested we can CC the mailing list and get more opinions.
-Stafford
Quoted 303 lines · On Wed, Oct 9, 2019 at 5:09 PM <ecalvo at 2se.es <mailto:[email protected]> > wrote:
Hi Stafford,
Nice to meet you and, first of all, thanks a lot for your guidance. I am new on this, and although there is some documentation, sometimes it is difficult some point which maybe it is basic.
Ok, to your comments. If.."A simulator like QEMU or or1ksim will not give and exact representation of the CPUs real time performance"...then...if I simulate directly the processor with modelsim, icarus or a similar tool...neither I get a real performance, don’t I? And values for the counters that you tell me to enable, neither are real, isn't it? should I execute it directly on the FPGA and it will depends on the implementation?
Ok, to C code. I have understood the dependency with toolchain.
Thanks a lot again.
Best regards,
Elisa-----Mensaje original-----
De: Stafford Horne <shorne at gmail.com <mailto:[email protected]> > Enviado el: martes, 8 de octubre de 2019 16:18
Para: Julius Baxter <juliusbaxter at gmail.com <mailto:[email protected]> >
CC: ecalvo at 2se.es <mailto:[email protected]>
Asunto: Re: other doubtHi Elisa,
OpenRISC cpu's can run any algorithm, but how well it will perform depends on many things:
- Compiler optimization flags (i.e. -O3)
- Whether or not you are doing FPU instructions and have FPU enabled
- Whether or not you use multiply and divide and have these instructions enabled
- The frequency you are running
- Cache settings Icache Dcache
- The type of algorithm, does it require lots of data which will cause many cache misses?A simulator like QEMU or or1ksim will not give and exact representation of the CPUs real time performance. It can tell you which intructions will be executed, but not how fast those will run or how many pipeline stalls of cache misses will happen.
You can use the performance counters, they are supported in mor1kx if you enable them with the FEATURE_PERFCOUNTERS='ENABLED' parameter. They can help count how many events happen between certain events. Then you can combine them with a timer and watchpoints to detect how many times a loop can execute in 1000 clock cycles etc. Please read about PCCRn and PCMRn in the architecture manual.
It might be just as easy to use simple timing in a c program though, depending on the toolchain you use you can compare times between runs of your algorithm. i.e.
#include <time.h> #include <stdio.h> static long to_micro(struct timespec *time) { return (time->tv_sec * 1000000) + (time->tv_nsec / 1000); } int main() { int i, j = 0; struct timespec before, after; clock_gettime(CLOCK_MONOTONIC, &before); /* Super complex algorithm */ for (int i = 0; i < 100; i++) { j = (j+1) * (j+2); } clock_gettime(CLOCK_MONOTONIC, &after); printf("time to run algorithm %ld uSecs\n", to_micro(&after)-
to_micro(&before));return 0; }$ or1k-smh-linux-gnu-gcc timer.c $ ./glibc-build-scripts/qemu-or1k-libc ./a.out time to run algorithm 164 uSecsI hope it helps.
-Stafford
Quoted 226 lines · Julius Baxter
Hi,
No problem.
There are performance counters in the OpenRISC architecture but whether they're implemented in a particular implementation is another matter.
You can use these registers to measure various things the CPU is doing while it's executing. If you read the ISA document it'll tell you about them.
I'm CCing Stafford because he's the main OpenRISC man these days and probably knows about the state of the performance counter registers in various simulators and RTL implementations.
Cheers,
JuliusQuoted 206 lines · On Tue., 8 Oct. 2019, 10:43 pm , <ecalvo at 2se.es <mailto:[email protected]> >…
Hi Julius,
Sorry for bothering you again ☹. Can I do you other fast question related to openrisc? If not, ignore the email please.
Is there any way to characterize the type of application that
I can run in openrisc? I mean, could you measure (with
numbers) if an algorithm can be executed on it and the speed that it will achieve? Is it possible to do it using orksim?Sorry because maybe it is so basic and general ☹
Thanks in advance
Elisa
*De:* Julius Baxter <juliusbaxter at gmail.com <mailto:[email protected]> > *Enviado el:* lunes,
16 de septiembre de 2019 13:11
*Para:* ecalvo at 2se.es <mailto:[email protected]>
*Asunto:* Re: Starting with OpenRISC - IOBsAlso! To let you know, if you're in Spain, we will soon be having our ORConf conference in Europe, and it's in Bordeaux, France, just across the border. :)
There are several people there who can help you get up to
speed, one of whom is Stafford Horne who knows most about the
OpenRISC IP lately. He will be presenting. If you can attend, it'd be helpful, I'm sure.All info at https://orconf.org
Cheers,
Julius
On Mon, 16 Sep 2019 at 21:09, Julius Baxter <juliusbaxter at gmail.com <mailto:[email protected]> > wrote:
Hi Elisa,
Sorry for the delay in this response.
You should be using an SoC toplevel. FPGAs have everything you need on board like memories and IO blocks and lots of other FPGA fabric for you to implement other pieces of hardware.
FuseSoC provides a really nice and easy way to build an mor1kx design for the DE0 nano I believe:
https://github.com/olofk/de0_nano
That github page has a rough guide to getting it going.
If you need help I recommend posting to the OpenRISC mailing list and people will respond probably more promptly than I. (I recommend getting to know how to use mailing lists.
https://openrisc.io/communityThere are more resources here: https://openrisc.io/tutorials
I hope that's helpful.
Cheers,
Julius
On Wed, 11 Sep 2019 at 20:09, <ecalvo at 2se.es <mailto:[email protected]> > wrote:
Hi Julius,
Thanks a lot for the quick answer.
Yes, this is the problem: I am using as top level the mor1kx module itself. You mean that I need to synthetize also in reconfigurable logic these cores, don’t you? I thought that I could have these elements as external in a development board.
Thanks again,
Cheers
Elisa
*De:* Julius Baxter <juliusbaxter at gmail.com <mailto:[email protected]> > *Enviado el:* miércoles, 11 de septiembre de 2019 12:02
*Para:* ecalvo at 2se.es <mailto:[email protected]>
*Asunto:* Re: Starting with OpenRISC - IOBsHi Elisa,
Thanks for getting in touch, that sounds like a cool project.
Can you tell me about the toplevel - are you using a system toplevel, or is your toplevel the mor1kx module itself?
If it's the latter, then that's not the best way to do it -
you need a system toplevel which instantiates memories and
some reset circuitry and likely some IO (UART, GPIO, JTAG debug, etc.) to talk to the outside world.Is that helpful?
Cheers,
Julius
On Wed, 11 Sep 2019 at 19:47, <ecalvo at 2se.es <mailto:[email protected]> > wrote:
Dear Dr. Baxter,
My name is Elisa Calvo Gallego. I am writing you because I
have started to work with OpenRISC in the framework of a
research project developed in the company where I am working
(Space Submicron Electronics, 2SE), and I am having some basic troubles. Could you help me?Although the FPGA that we are planning to use is larger, I
have synthetized mor1kx for a DE0 nano board as first step
(this is the board used in the majority of guides and
tutorials). My problem is that the results that I have
obtained are similar in area and resources, except for IOBs,
which are more than available IOBs in the device. Do you know what I am doing wrong? Should I comment debug lines or something like that? I
apologize if the question is immediate. I didn't find the answer and I'm new in this.Thanks very much in advance.
Best regards,
Elisa
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.librecores.org/pipermail/openrisc/attachments/20191024/6dba270a/attachment-0001.html>