Function signature for attribute(kernel) function and usage of "OutAllocation"

Question

I intend to process RGB data through RenderScript.

For this I've created Allocations in Java and passing them to RS Kernel function as below:

uchar3 __attribute__((kernel)) invert(uchar3 v_in, uint32_t v_out) {
v_in.r = ...;
v_in.g = ...;
v_in.b = ...;
}

However ideally I would like to work on v_out in similar way i.e. setting values for v_out.r, v_out.g and v_out.b. Currently I can not do this with uint32_t v_out.

Now if I define the above RS kernel as:

uchar3 __attribute__((kernel)) invert(uchar3 v_in, uchar3 v_out) {
...
}

I get below compile time error: error: Unexpected kernel invert() parameter 'v_out' of type 'uchar3 *'

Please suggest how to resolve this.

Compile time:

error: Unexpected kernel invert() parameter 'v_out' of type 'uchar3 *'

score 1 · Accepted Answer · answered Dec 09 '13 at 18:11

1

you should be defining this as

uchar3 __attribute__((kernel)) invert(uchar3 in);

that function will then be reflected as ScriptC_.forEach_invert(Allocation in, Allocation out). each element in in will be passed to invert, and each value returned from invert will be written to the corresponding location in out.

answered Dec 09 '13 at 18:11

Tim Murray

2,205
13
13

Thanks for your reply. This part works well however I still can't get back this "processed" out data in java. My test code(for 5 elements) looks like this: `mScript.forEach_invert(ain, aout); byte[] output = {0, 0, 0, 0, 0}; aout.copyTo(output);` Am I missing something here? – Gagan Dec 10 '13 at 02:52
Yes, I did get the values back in Java. There was a mismatch in byte[] lengths. – Gagan Dec 10 '13 at 05:30
Does writing a RS `kernel` function always ensure that the code runs on GPU? How can we check whether a particular RS function (kernel or invokable) runs on GPU or CPU? – Gagan Dec 10 '13 at 19:26
No, in fact you can write an RS kernel in such a way that it precludes it from running on the GPU, for example recursive functions. I don't think you can check whether you're running on CPU or GPU beyond obvious performance differences... Or you could attach a profiler and look for any GPU activity? No use in production environment ofc. – chrisvarnz Feb 24 '14 at 13:57
some GPUs are capable of recursion. longer term, we're working on better profiler support and more logcat options that spell out why something is running on a particular processor. however, the GPU is not always going to be faster than the CPU or anything like that; the GPU is usually pretty busy with rendering a high-DPI display, CPU and GPU are power/thermal limited, they have the same memory bandwidth, and so on. – Tim Murray Feb 24 '14 at 18:08

score 1 · Answer 2 · answered Dec 09 '13 at 18:42

1

Define your own local variable of uchar3 type and then populate it before returning it:

uchar3 __attribute__((kernel)) invert(uchar3 in) {
    uchar3 out;
    out.r = ...
    out.g = ...
    out.b = ...
    return out;
}

The compiler is clever enough to notice what you are doing, so there won't be additional copies made of the output items.

answered Dec 09 '13 at 18:42

Stephen Hines

2,612
1
13
12

Thanks for your input. I'm progressing with both these approaches and have updated my finding in comment-section above. – Gagan Dec 10 '13 at 02:53
Thanks again for the answer mate. I did get the values back in Java. Now just have to create my custom 24 bit `Element` for my ain and aout allocations. In current implementation, the returned uchar3 data is 32 bit long and I lose r-component of my 2nd pixel and henceforth. – Gagan Dec 10 '13 at 05:29

Function signature for __attribute__(kernel) function and usage of "OutAllocation"

2 Answers2

Function signature for attribute(kernel) function and usage of "OutAllocation"