[Beignet] silently does nothing on large arrays (was ICD interface (only) broken on Haswell)

[Zhigang Gong]

Reproduced. Thanks for reporting the issue and we will look into it soon.

On Mon, Apr 06, 2015 at 09:55:48PM +0100, Rebecca N. Palmer wrote:
> Control: retitle -1 beignet: silently does nothing on large arrays
> (previous discussion:
> https://bugs.debian.org/cgi-bin/bugreport.cgi?bug=781875 )
> 
> This isn't a Haswell-specific problem (and might not be ICD-specific
> either: did you test that at these sizes, or only with the
> testsuite?), it's a large-array-specific problem: I simply hadn't
> tried arrays that big before.
> 
> The array size required to trigger it decreases as the number of
> arrays (arguments to the kernel being run, not total existing)
> increases, suggesting a total-memory limit: on my system,
> approximately 2x240Mifloat, 3x160Mifloat, or 5x100Mifloat, but these
> vary ~10% from run to run.  (Hence, re-adding the per-array size
> limit probably wouldn't completely avoid the problem, though I
> haven't actually tried that.)
> 
> These sizes do _not_ appear to depend on free system memory, but due
> to their variability and the limited range I can test before running
> into "running out of memory in beignet hangs the entire system"
> (https://bugs.launchpad.net/ubuntu/+source/beignet/+bug/1354086 ), I
> cannot be completely sure of this.

> #!/usr/bin/env python3
> #Depends: python3-pyopencl python3-numpy
> from __future__ import division,print_function
> import pyopencl
> import pyopencl.array
> import numpy as np
> import time
> import pyopencl.clmath
> ctx=pyopencl.create_some_context()
> cq=pyopencl.CommandQueue(ctx)
> asize=100*(2**20)#fails above approx. 235 for 2-array, 162 for 3-array, 100 for 5-array, but the exact number varies
> #Warning: very large sizes will hang your system, https://bugs.launchpad.net/ubuntu/+source/beignet/+bug/1354086
> aCL=pyopencl.array.arange(cq,0,asize,1,dtype='float32')
> bCL=pyopencl.array.arange(cq,0,asize,1,dtype='float32')
> cCL=pyopencl.array.arange(cq,0,asize,1,dtype='float32')
> dCL=pyopencl.array.arange(cq,0,asize,1,dtype='float32')
> eCL=pyopencl.array.arange(cq,0,asize,1,dtype='float32')
> print("CL arrays created")
> ans=aCL[0:1000].get()*4
> f2=pyopencl.elementwise.ElementwiseKernel(ctx,pyopencl.tools.dtype_to_ctype(aCL.dtype)+" *a,"+pyopencl.tools.dtype_to_ctype(aCL.dtype)+" *c","c[i]=3*a[i]+c[i]","twoarray")
> f3=pyopencl.elementwise.ElementwiseKernel(ctx,pyopencl.tools.dtype_to_ctype(aCL.dtype)+" *a,"+pyopencl.tools.dtype_to_ctype(aCL.dtype)+" *b,"+pyopencl.tools.dtype_to_ctype(aCL.dtype)+" *c","c[i]=3*a[i]+b[i]","threearray")
> f5=pyopencl.elementwise.ElementwiseKernel(ctx,pyopencl.tools.dtype_to_ctype(aCL.dtype)+" *a,"+pyopencl.tools.dtype_to_ctype(aCL.dtype)+" *b,"+pyopencl.tools.dtype_to_ctype(aCL.dtype)+" *c,"+pyopencl.tools.dtype_to_ctype(aCL.dtype)+" *d,"+pyopencl.tools.dtype_to_ctype(aCL.dtype)+" *e","c[i]=a[i]+b[i]+d[i]+e[i]","fivearray")
> f5b=pyopencl.elementwise.ElementwiseKernel(ctx,pyopencl.tools.dtype_to_ctype(aCL.dtype)+" *a,"+pyopencl.tools.dtype_to_ctype(aCL.dtype)+" *b,"+pyopencl.tools.dtype_to_ctype(aCL.dtype)+" *c,"+pyopencl.tools.dtype_to_ctype(aCL.dtype)+" *d,"+pyopencl.tools.dtype_to_ctype(aCL.dtype)+" *e","c[i]=4*e[i]","fivearray_usetwo")
> f2(aCL,cCL).wait()
> #f3(aCL,bCL,cCL).wait()
> f5(aCL,bCL,cCL,dCL,eCL).wait()
> #f5b(aCL,bCL,cCL,dCL,eCL).wait()
> print("size",len(aCL)," error ",np.max(np.nan_to_num(np.abs(cCL[0:1000].get()-ans))),"first 10 ",ans[0:10],cCL[0:10].get())
> 

> _______________________________________________
> Beignet mailing list
> Beignet at lists.freedesktop.org
> http://lists.freedesktop.org/mailman/listinfo/beignet