Two float[] outputs in one kernel pass (Sobel -> Magnitude and Direction) - allocation

Question

I wrote the following rs code in order to calculate the magnitude and the direction within the same kernel as the sobel gradients.
#pragma version(1)
#pragma rs java_package_name(com.example.xxx)
#pragma rs_fp_relaxed
rs_allocation bmpAllocIn, direction;
int32_t width;
int32_t height;
// Sobel, Magnitude und Direction
float __attribute__((kernel)) sobel_XY(uint32_t x, uint32_t y) {
float sobX=0, sobY=0, magn=0;
// leave a border of 1 pixel
if (x>0 && y>0 && x<(width-1) && y<(height-1)){
uchar4 c11=rsGetElementAt_uchar4(bmpAllocIn, x-1, y-1); uchar4 c12=rsGetElementAt_uchar4(bmpAllocIn, x-1, y);uchar4 c13=rsGetElementAt_uchar4(bmpAllocIn, x-1, y+1);
uchar4 c21=rsGetElementAt_uchar4(bmpAllocIn, x, y-1);uchar4 c23=rsGetElementAt_uchar4(bmpAllocIn, x, y+1);
uchar4 c31=rsGetElementAt_uchar4(bmpAllocIn, x+1, y-1);uchar4 c32=rsGetElementAt_uchar4(bmpAllocIn, x+1, y);uchar4 c33=rsGetElementAt_uchar4(bmpAllocIn, x+1, y+1);
sobX= (float) c11.r-c31.r + 2*(c12.r-c32.r) + c13.r-c33.r;
sobY= (float) c11.r-c13.r + 2*(c21.r-c23.r) + c31.r-c33.r;
float d = atan2(sobY, sobX);
rsSetElementAt_float(direction, d, x, y);
magn= hypot(sobX, sobY);
}
else{
magn=0;
rsSetElementAt_float(direction, 0, x, y);
}
return magn;
}
And the Java part:
float[] gm = new float[width*height]; // gradient magnitude
float[] gd = new float[width*height]; // gradient direction
ScriptC_sobel script;
script=new ScriptC_sobel(rs);
script.set_bmpAllocIn(Allocation.createFromBitmap(rs, bmpGray));
// dirAllocation: reference to the global variable "direction" in rs script. This
// dirAllocation is actually the second output of the kernel. It will be "filled" by
// the rsSetElementAt_float() method that include a reference to the current
// element (x,y) during the passage of the kernel.
Type.Builder TypeDir = new Type.Builder(rs, Element.F32(rs));
TypeDir.setX(width).setY(height);
Allocation dirAllocation = Allocation.createTyped(rs, TypeDir.create());
script.set_direction(dirAllocation);
// outAllocation: the kernel will slide along this global float Variable, which is
// "formally" the output (in principle the roles of the outAllocation (magnitude) and the
// second global variable direction (dirAllocation)could have been switched, the kernel
// just needs at least one in- or out-Allocation to "slide" along.)
Type.Builder TypeOut = new Type.Builder(rs, Element.F32(rs));
TypeOut.setX(width).setY(height);
Allocation outAllocation = Allocation.createTyped(rs, TypeOut.create());
script.forEach_sobel_XY(outAllocation); //start kernel
// here comes the problem
outAllocation.copyTo(gm) ;
dirAllocation.copyTo(gd);
In a nutshell: this code works for my older Galaxy Tab2 (API17) but it creates a crash (Fatal signal 7 (SIGBUS), code 2, fault addr 0x9e6d4000 in tid 6385) with my Galaxy S5 (API 21). The strange thing is that when I use a simpler Kernel that just calculates SobelX or SobelY gradients in the very same way (except the 2nd allocation, here for the direction), it works also on the S5. Thus, the Problem cannot be some compatibility issue. Also, as I said, the kernel itself passes without problems (I can log the Magnitude and direction values) but it struggles with the above .copyTo Statements. As you can see the gm and gd floats have the same dimensions (width*height) as all other allocations used by the kernel. Any idea what the Problem could be? Or is there an alternative, more robust way to do the whole Story?