Can't launch 3d grid?

Accelerated Computing CUDA CUDA Programming and Performance
1

I use CUDA8.0 and VS2015, with GTX980.

I tried to launch a kernel using 3D grid and 1D block.

dim3 radiance_gird = (4, 4, 4);
dim3 radiance_block = (FirePresets::TOTAL_SAMPLES); // 5
radianceKernel << <radiance_gird, radiance_block >> > (dev_le, dev_l, dev_le_mean, dev_T, dev_xm, dev_ym, dev_zm);

In NVIDIA NSight CUDA Focus it says my valid grid dimension is (4,0,0). Why?

2

Buildable code or it didn’t happen :-) Now would be a good time to familiarize yourself with the concept of an MCVE (see: [url]https://stackoverflow.com/help/mcve[/url]).

Does the actual code (with proper error checking!) compile and run outside of Nsight? Are you compiling for the correct target architecture (-arch=sm_52)? What is the exact error message returned by Nsight?

3

According to what I believe your intent to be, this will not accomplish what you want.

dim3 radiance_gird = (4, 4, 4);

According to what I believe your intent to be, it should be either:

dim3 radiance_gird = dim3(4, 4, 4);

or:

dim3 radiance_gird(4, 4, 4);

You can look up what the C or C++ rules are for evaluation of this construct:

(4,4,4)

The net is that the comma operator and parenthesis reduce it to:

4

And you can certainly initialize a dim3 item with a scalar quantity of 4, in which case it will result in:

(4,1,1)

cross posting:

https://stackoverflow.com/questions/45995990/cuda-cant-launch-3d-grid

demonstrating that the grid in question is reported as (4,1,1) (which is valid) not (4,0,0) which OP stated in original post of this thread (which is not a valid grid dimension)

4

I would not use the word “invalid” to describe the initialization. It has well-defined semantics, which were explained in detail by txbob. So the initialization is valid, but not what the programmer intended.

5

The intent is clear (in my opinion). For that implied intent, the method chosen is invalid. Wordsmithing applied.