OFFICIAL CGMINER mining software thread for linux/win/osx/mips/arm/r-pi 4.11.0 - page 647.

-ck

legendary

Activity: 4088

Merit: 1631

Ruu \o/

Quote from: Diapolo on February 17, 2012, 02:48:56 AM

Quote from: -ck on February 17, 2012, 02:26:03 AM

Quote from: Diapolo on February 17, 2012, 01:39:08 AM

Hey Con,

The global_work_offset value you pass to clEnqueueNDRangeKernel is used as the starting value for get_global_id(0). If global_work_offset would be 10 and we had 5 work-items, the nonces generated would be 20, 21, 22, 23, 24, 25, 26, 27, 28 and 29 so we loose nonces from 10 to 19 and use ones, that should not be base ... would that create duplicate work? You pass -D GOFFSET to the kernel, right?

I suspect the problem is to do with overflow on 32 bit unsigned integers. Imagine a value close to 2^32. Since the values are doubled, and what happens on overflowing 32 bits is undefined, you may well get repeated ranges of nonces checked.

That would prove true for a base of (2147483647 - global worksize) because doubled it's over 2^32. How big are the nonce bases and what would be the global worksize for -I == 14 (as this is the maximum)?

Dia

global worksize = 2^(15 + intensity) so it's 29 max and it's double that for 2 vectors and so on...

cgminer always tries to test the entire range of nonces up to 2^32 so it will *always* get to a value above 2^31 where it will wrap with a doubling of the global id regardless of what intensity it's at.

Diapolo

hero member

Activity: 772

Merit: 500

Quote from: -ck on February 17, 2012, 02:26:03 AM

Quote from: Diapolo on February 17, 2012, 01:39:08 AM

Hey Con,

The global_work_offset value you pass to clEnqueueNDRangeKernel is used as the starting value for get_global_id(0). If global_work_offset would be 10 and we had 5 work-items, the nonces generated would be 20, 21, 22, 23, 24, 25, 26, 27, 28 and 29 so we loose nonces from 10 to 19 and use ones, that should not be base ... would that create duplicate work? You pass -D GOFFSET to the kernel, right?

I suspect the problem is to do with overflow on 32 bit unsigned integers. Imagine a value close to 2^32. Since the values are doubled, and what happens on overflowing 32 bits is undefined, you may well get repeated ranges of nonces checked.

That would prove true for a base of (2147483647 - global worksize) because doubled it's over 2^32. How big are the nonce bases and what would be the global worksize for -I == 14 (as this is the maximum)?

Dia

Diapolo

hero member

Activity: 772

Merit: 500

Quote from: -ck on February 17, 2012, 02:30:05 AM

Quote from: Diapolo on February 17, 2012, 01:39:08 AM

Quote from: -ck on February 17, 2012, 12:42:37 AM

Edit: Oh and I have to warn you, yesterday I tested a new AMD driver, which has a new OpenCL runtime and I was unable to use CGMINER on Windows, so perhaps AMD "fixed" the binary generation but this breaks your solution from a few days ago, because the old message that no .bin could be generated was back. I switched back to a former runtime :-(. Seems to suck!

I don't think this is the old bug but a simple failure to produce a binary because of compilation of the kernel instead. Try running with -T -verbose and it will spew out any kernel compilation errors.

As the kernel compiles fine with AMDs KernelAnalyzer, I doubt it's an compilation error, I looked through the verbose messages yesterday and only saw clBuildProgram has failed (not the exact message), but got no real hint as to why this happens. Is there a check for a specific runtime version, which could cause that behaviour, because new runtime version is unknown ... we can debug this later, as I don't have access to my PC currently.

Dia

TheHarbinger

sr. member

Activity: 378

Merit: 250

Why is it so damn hot in here?

Quote from: DeathAndTaxes on February 16, 2012, 11:34:51 PM

Quote from: jjshabadoo on February 16, 2012, 11:31:12 PM

Oh well, i guess i was wrong, i read that BAMT didn't work with cgminer, maybe they just mean it's not part of the original download which is fine. it didn't make sense to em either, but i fiugured it had something to do with some of the programming.

So i guess you just download BAMT and then download cgminer and you're all set?

what drivers and sdk does BAMT use?

BAMT already has cgminer installed and it is integrated into BAMT other tools (mgpumon, web monitor, gpumon, etc).

It is only cgminer 2.1.2 I believe but you can install newer copy if you like. SDK is 2.4, I am not sure the driver. It doesn't have 100% bug.

So it is write BAMT to flash drive. Run fixer to grab latest updates, change 2 config files (1 for BAMT, 1 for cgminer) and you are mining. You can then take that flash drive, record the image, put that image on 8 flash drives, put them in 8 rigs make a few changes via SSH and power up 20 GH/s farm in a few minutes.

For sure on the 2.4 Sad

I'm still trying to get it downgraded back to 2.1, Dropping my memclocks from 330 to 150 saves me a good amount of wattage on my 5@5870 rig. If you figure out a way to do it, please share the info.

-ck

legendary

Activity: 4088

Merit: 1631

Ruu \o/

Quote from: Diapolo on February 17, 2012, 01:39:08 AM

Quote from: -ck on February 17, 2012, 12:42:37 AM

Edit: Oh and I have to warn you, yesterday I tested a new AMD driver, which has a new OpenCL runtime and I was unable to use CGMINER on Windows, so perhaps AMD "fixed" the binary generation but this breaks your solution from a few days ago, because the old message that no .bin could be generated was back. I switched back to a former runtime :-(. Seems to suck!

I don't think this is the old bug but a simple failure to produce a binary because of compilation of the kernel instead. Try running with -T -verbose and it will spew out any kernel compilation errors.

-ck

legendary

Activity: 4088

Merit: 1631

Ruu \o/

Quote from: Diapolo on February 17, 2012, 02:25:47 AM

Quote from: -ck on February 17, 2012, 02:15:30 AM

Quote from: blandead on February 17, 2012, 01:49:19 AM

I thought the NDRangeKernel went like this...Edit: Oh and Dia, in your CommandQueue try adding this property, "cl.command_queue_properties.OUT_OF_ORDER_EXEC_MODE_ENABLE"

https://github.com/ckolivas/cgminer/blob/master/ocl.c#L710

Code:

clState->commandQueue = clCreateCommandQueue(clState->context, devices[gpu],
CL_QUEUE_OUT_OF_ORDER_EXEC_MODE_ENABLE, &status);

cgminer has used this for a very long time.

As I wrote, I think OoE mode is not supported on AMD GPUs ... is there a debug or vebose message, if that mode was successfully activated?

It is successfully activated on windows and linux, but osx fails. It does not improve throughput with current GPUs but is harmless to enable for if/when they do.

-ck

legendary

Activity: 4088

Merit: 1631

Ruu \o/

Quote from: Diapolo on February 17, 2012, 01:39:08 AM

Hey Con,

The global_work_offset value you pass to clEnqueueNDRangeKernel is used as the starting value for get_global_id(0). If global_work_offset would be 10 and we had 5 work-items, the nonces generated would be 20, 21, 22, 23, 24, 25, 26, 27, 28 and 29 so we loose nonces from 10 to 19 and use ones, that should not be base ... would that create duplicate work? You pass -D GOFFSET to the kernel, right?

I suspect the problem is to do with overflow on 32 bit unsigned integers. Imagine a value close to 2^32. Since the values are doubled, and what happens on overflowing 32 bits is undefined, you may well get repeated ranges of nonces checked.

Diapolo

hero member

Activity: 772

Merit: 500

Quote from: -ck on February 17, 2012, 02:15:30 AM

Quote from: blandead on February 17, 2012, 01:49:19 AM

I thought the NDRangeKernel went like this...Edit: Oh and Dia, in your CommandQueue try adding this property, "cl.command_queue_properties.OUT_OF_ORDER_EXEC_MODE_ENABLE"

https://github.com/ckolivas/cgminer/blob/master/ocl.c#L710

Code:

clState->commandQueue = clCreateCommandQueue(clState->context, devices[gpu],
CL_QUEUE_OUT_OF_ORDER_EXEC_MODE_ENABLE, &status);

cgminer has used this for a very long time.

As I wrote, I think OoE mode is not supported on AMD GPUs ... is there a debug or vebose message, if that mode was successfully activated?

Dia

Diapolo

hero member

Activity: 772

Merit: 500

Quote from: blandead on February 17, 2012, 01:49:19 AM

I thought the NDRangeKernel went like this...

cl_int clEnqueueNDRangeKernel (
cl_command_queue command_queue,
cl_kernel kernel, cl_uint work_dim,
const size_t *global_work_offset,
const size_t *global_work_size,
const size_t *local_work_size,
cl_uint num_events_in_wait_list,
const cl_event *event_wait_list, cl_event *event)

If you tell it the global_work_size and local_work_size, why would the global_work_offset be the starting value for get_global_id(0)? Or am I completely off the ball?

What runtime are you using? I'm using newest 12.2 preview runtime with no problems after adding opencl 1.2 preview drivers

Edit: Oh and Dia, in your CommandQueue try adding this property, "cl.command_queue_properties.OUT_OF_ORDER_EXEC_MODE_ENABLE"

I think it would be beneficial to start incorporating out of order execution in the kernels, especially if you are moving to the NDRangeKernel. After adding this in and a couple other changes, VECTORS8 is actually running faster than VECTORS4.. Getting very high output on my nonces

OoE is not supported on AMD GPUs afaik. But I'm sure Con has code to try to use it with the command queue, if available.

The driver, which contains the problematic Runtime is this one: http://support.amd.com/us/kbarticles/Pages/hd7700series7support.aspx

Global worksize is the number of global work-items that are processed in one kernel execution and the local worksize is the work-group size of work-items, that are executed in parallel and share __local memory.

Dia

-ck

legendary

Activity: 4088

Merit: 1631

Ruu \o/

Quote from: blandead on February 17, 2012, 01:49:19 AM

I thought the NDRangeKernel went like this...Edit: Oh and Dia, in your CommandQueue try adding this property, "cl.command_queue_properties.OUT_OF_ORDER_EXEC_MODE_ENABLE"

https://github.com/ckolivas/cgminer/blob/master/ocl.c#L710

Code:

clState->commandQueue = clCreateCommandQueue(clState->context, devices[gpu],
CL_QUEUE_OUT_OF_ORDER_EXEC_MODE_ENABLE, &status);

cgminer has used this for a very long time.

blandead

newbie

Activity: 46

Merit: 0

I thought the NDRangeKernel went like this...

cl_int clEnqueueNDRangeKernel (
cl_command_queue command_queue,
cl_kernel kernel, cl_uint work_dim,
const size_t *global_work_offset,
const size_t *global_work_size,
const size_t *local_work_size,
cl_uint num_events_in_wait_list,
const cl_event *event_wait_list, cl_event *event)

If you tell it the global_work_size and local_work_size, why would the global_work_offset be the starting value for get_global_id(0)? Or am I completely off the ball?

What runtime are you using? I'm using newest 12.2 preview runtime with no problems after adding opencl 1.2 preview drivers

Edit: Oh and Dia, in your CommandQueue try adding this property, "cl.command_queue_properties.OUT_OF_ORDER_EXEC_MODE_ENABLE"

I think it would be beneficial to start incorporating out of order execution in the kernels, especially if you are moving to the NDRangeKernel. After adding this in and a couple other changes, VECTORS8 is actually running faster than VECTORS4.. Getting very high output on my nonces

Diapolo

hero member

Activity: 772

Merit: 500

Quote from: -ck on February 17, 2012, 12:42:37 AM

Quote from: Diapolo on February 16, 2012, 10:11:16 AM

Con, maybe we can talk about integrating global offset parameter support into CGMINER?

Take a short look at http://www.khronos.org/registry/cl/sdk/1.2/docs/man/xhtml/clEnqueueNDRangeKernel.html and the global_work_offset parameter. All that has to be taken into consideration from the kernel-side is in DiaKGCN. OpenCL 1.1 detection is in your code, too, which is needed, but I can't do the other required changes without a compiler.

In short, the nonce-base is not supplied via the base parameter, if GOFFSET is enabled, but instead via the global_work_offset parameter and used via the global work-item ID in the kernel. This saves a few instructions and can give us a small boost.

Dia

I tried writing code to send nonce as the global offset parameter and your code returned duplicate work on 2 or more vectors. Looking at your code:

Code:

u nonce = ((uint)get_global_id(0) << 1) + (u)(0, 1);

This won't be right as global id will now be the global thread id + the global offset parameter so doubling it will give random results.

Hey Con,

The global_work_offset value you pass to clEnqueueNDRangeKernel is used as the starting value for get_global_id(0). If global_work_offset would be 10 and we had 5 work-items, the nonces generated would be 20, 21, 22, 23, 24, 25, 26, 27, 28 and 29 so we loose nonces from 10 to 19 and use ones, that should not be base ... would that create duplicate work? You pass -D GOFFSET to the kernel, right?

Edit: Oh and I have to warn you, yesterday I tested a new AMD driver, which has a new OpenCL runtime and I was unable to use CGMINER on Windows, so perhaps AMD "fixed" the binary generation but this breaks your solution from a few days ago, because the old message that no .bin could be generated was back. I switched back to a former runtime :-(. Seems to suck!

Dia

-ck

legendary

Activity: 4088

Merit: 1631

Ruu \o/

Quote from: blandead on February 17, 2012, 12:53:43 AM

Nonce of that code makes much sense, but I think Dia is suggesting that he wants to use the natively built-in global_work_offset parameter instead of the one you reference by using clEnqueueNDRangeKernel

Umm... I actually do understand the code Wink

I was explaining what was wrong with it.

blandead

newbie

Activity: 46

Merit: 0

Nonce of that code makes much sense, but I think Dia is suggesting that he wants to use the natively built-in global_work_offset parameter instead of the one you reference by using clEnqueueNDRangeKernel

-ck

legendary

Activity: 4088

Merit: 1631

Ruu \o/

Quote from: Diapolo on February 16, 2012, 10:11:16 AM

Con, maybe we can talk about integrating global offset parameter support into CGMINER?

Take a short look at http://www.khronos.org/registry/cl/sdk/1.2/docs/man/xhtml/clEnqueueNDRangeKernel.html and the global_work_offset parameter. All that has to be taken into consideration from the kernel-side is in DiaKGCN. OpenCL 1.1 detection is in your code, too, which is needed, but I can't do the other required changes without a compiler.

In short, the nonce-base is not supplied via the base parameter, if GOFFSET is enabled, but instead via the global_work_offset parameter and used via the global work-item ID in the kernel. This saves a few instructions and can give us a small boost.

Dia

I tried writing code to send nonce as the global offset parameter and your code returned duplicate work on 2 or more vectors. Looking at your code:

Code:

u nonce = ((uint)get_global_id(0) << 1) + (u)(0, 1);

This won't be right as global id will now be the global thread id + the global offset parameter so doubling it will give random results.

-ck

legendary

Activity: 4088

Merit: 1631

Ruu \o/

Quote from: -ck on February 16, 2012, 11:05:33 PM

Quote from: dlasher on February 16, 2012, 09:31:16 PM

Quote from: -ck on February 15, 2012, 10:34:19 PM

cgminer sets clocks all back to default on exit... if it exits cleanly, and of course on windows it's a miracle when it does.

Sadly this is not the case.

windows machine, pair of 6950's, set to 850/1300 for normal operation, in cgminer they are set to 700-880/300... when cgminer exits it leaves the cards at 880/300.

even updated to 12.1 drivers, both 2.4 and 2.6 SDK.

Instead windows decides to blow your balls off and feed them back to you by collecting up the splatter, mincing it, putting it into a glass and forcing you to drink it through a straw via your left nostril.

I'll try and code a workaround for this windows fail next time by stopping mining and resetting device values and pausing for a bit before letting windows crash cgminer when it tries to exit.

DeathAndTaxes

donator

Activity: 1218

Merit: 1079

Gerald Davis

Quote from: jjshabadoo on February 16, 2012, 11:31:12 PM

Oh well, i guess i was wrong, i read that BAMT didn't work with cgminer, maybe they just mean it's not part of the original download which is fine. it didn't make sense to em either, but i fiugured it had something to do with some of the programming.

So i guess you just download BAMT and then download cgminer and you're all set?

what drivers and sdk does BAMT use?

BAMT already has cgminer installed and it is integrated into BAMT other tools (mgpumon, web monitor, gpumon, etc).

It is only cgminer 2.1.2 I believe but you can install newer copy if you like. SDK is 2.4, I am not sure the driver. It doesn't have 100% bug.

So it is write BAMT to flash drive. Run fixer to grab latest updates, change 2 config files (1 for BAMT, 1 for cgminer) and you are mining. You can then take that flash drive, record the image, put that image on 8 flash drives, put them in 8 rigs make a few changes via SSH and power up 20 GH/s farm in a few minutes.

jjshabadoo

hero member

Activity: 535

Merit: 500

Oh well, i guess i was wrong, i read that BAMT didn't work with cgminer, maybe they just mean it's not part of the original download which is fine. it didn't make sense to em either, but i fiugured it had something to do with some of the programming.

So i guess you just download BAMT and then download cgminer and you're all set?

what drivers and sdk does BAMT use?

gnar1ta$

donator

Activity: 798

Merit: 500

Quote from: jjshabadoo on February 16, 2012, 08:54:24 PM

We need a new stripped down version of linuxcoin or call it whatever, just for mining and maybe built around optimal settings/set-up for cgminer. Seems BAMT doesn't supoort or work with

What am I missing here, cuz my 5970 BAMT rig has been mining with cgminer for a week now - and it took 3 mins to set up Huh

-ck

legendary

Activity: 4088

Merit: 1631

Ruu \o/

Quote from: dlasher on February 16, 2012, 09:31:16 PM

Quote from: -ck on February 15, 2012, 10:34:19 PM

cgminer sets clocks all back to default on exit... if it exits cleanly, and of course on windows it's a miracle when it does.

Sadly this is not the case.

windows machine, pair of 6950's, set to 850/1300 for normal operation, in cgminer they are set to 700-880/300... when cgminer exits it leaves the cards at 880/300.

even updated to 12.1 drivers, both 2.4 and 2.6 SDK.

Instead windows decides to blow your balls off and feed them back to you by collecting up the splatter, mincing it, putting it into a glass and forcing you to drink it through a straw via your left nostril.

Topic: OFFICIAL CGMINER mining software thread for linux/win/osx/mips/arm/r-pi 4.11.0 - page 647. (Read 5805983 times)