Passing C structures to OpenCL kernel

Question

I have next structure in my host program:

typedef struct  s_figure
{
    cl_float        reflection;
    cl_int          color;
    enum e_figure   type;
    cl_float3       vector1;
    cl_float3       vector2;
    cl_float        param1;
    cl_float        param2;
}                       t_figure;

I have next structure in my kernel:

typedef struct          s_figure
{
    float       reflection;
    int         color;
    enum e_figure   type;
    float3      vector1;
    float3      vector2;
    float       param1;
    float       param2;
}                       t_figure;

Also you can see enum for both of it:

enum            e_figure
{
    BadFigure = -1,
    InfinitePlane = 0,
    Sphere = 1,
    InfiniteCylinder = 2,
    InfiniteCone = 3
};

While passing data to the OpenCL kernel in this way (where figures is an correct parsed structures array):

buf_figures = clCreateBuffer(context, CL_MEM_USE_HOST_PTR, sizeof(t_figure) * figures_count, figures, &err);
clEnqueueWriteBuffer(view->cl->queue, buf_figures, CL_TRUE, 0,sizeof(t_figure) * figures_count, figures, 0, NULL, NULL);

I have a problems with data distortion, for example color after transferring data to OpenCL kernel can be changed to very different (0xFFFFFF->0x007FC2). Also raytrace algorithms works in another way every program execution. How can I fix it? I think gcc compiler make structures in different way than openclc but how to sync it?

cl_int is 32-bit, system int may be either 32 or 64. set color to int32_t — Andreas, Aug 24 '18 at 17:35
The second structure is opencl structure. Does it's int not compatible with cl_int? — Andrew Butok, Aug 24 '18 at 17:48
_[THIS](https://stackoverflow.com/questions/15639197/passing-struct-to-gpu-with-opencl-that-contains-an-array-of-floats)_ may, or may not help. Take a look. — ryyker, Aug 24 '18 at 17:57
This may help, section 6/1 beginning on page 197 as well as section 6.1.5 Alignment of Types on page 203 in https://www.khronos.org/registry/OpenCL/specs/opencl-1.2.pdf — Richard Chambers, Aug 24 '18 at 20:30

Andrew Savonichev · Accepted Answer · 2018-08-25T08:32:43.267

1

You need to specify __attribute__ ((packed)) for a struct declaration for both host and device. Otherwise different compilers (host compiler and device compiler) may create different memory layout for a struct.

EDIT: Given that sizeof(t_figure) is 52 on both host and device, then the root cause is probably not related to the struct.

Since you create a buffer with CL_MEM_USE_HOST_PTR, there are some caveats you need to consider:

figures pointer that you specify with CL_MEM_USE_HOST_PTR must point to a valid memory while a buffer is alive, because it is actually used as an underlying memory storage for cl_mem object.
Buffer created with CL_MEM_USE_HOST_PTR already uses a memory (or a copy of it) pointed by figures, so a subsequent write with clEnqueueWriteBuffer is redundant here. More importantly, I'm not even sure how clEnqueueWriteBuffer behaves in this case, because this operation is essentially a memcpy(figures,figures, size).

edited Aug 25 '18 at 08:32

answered Aug 24 '18 at 19:56

Andrew Savonichev

679
4
8

I have specified it for both struct declarations but after that program do the same. – Andrew Butok Aug 25 '18 at 04:36
Since all kernel execution no any function to change figure data, but it not like in host program. – Andrew Butok Aug 25 '18 at 04:46
@AndrewButok, can you check `sizeof(t_figure)` and dump its content using `printf` on host and device? – Andrew Savonichev Aug 25 '18 at 06:07
Both of them have 52 bytes – Andrew Butok Aug 25 '18 at 06:17
I updated my answer to reflect another possible root cause. – Andrew Savonichev Aug 25 '18 at 08:34
I find it helpful to put structures in largest-to-smallest order. It's not great for keeping things grouped nicely, but it reduces alignment issues greatly. – Dithermaster Aug 25 '18 at 19:26

Passing C structures to OpenCL kernel

1 Answers1