OpenCL explicit vector alignment for structs

OpenCL explicit vector alignment for structs

Bild des Benutzers dpshamonin

Hi,

After checking OpenCL specification (Section 6) for me it is still unclear
whether you have to provide explicit vector alignment for structs initialization.

OpenCL example on different platforms (Intel, NVidia, AMD):

typedef struct
{
float4 f;
float2 g;
float4 h;
} A;

__constant A a = {0,1,2,3,4,5,6,7,8,9}; // Will compiles fine on Intel and NVidia. FAIL on AMD!!!
__constant A a = {{0,1,2,3},{4,5},{6,7,8,9}}; // Will compiles fine on Intel, NVidia and AMD.

Could you provide some explanation how you would prefer to see it according to the standard?

Thanks,
-Denis

3 Beiträge / 0 neu
Letzter Beitrag
Nähere Informationen zur Compiler-Optimierung finden Sie in unserem Optimierungshinweis.
Bild des Benutzers Guy Benyei (Intel)

Hi Denis,

C99 has its standard way to initialize structures by enclosing the initializing values in braces. OpenCL doesn't override this specification.

As for vector elements, the right way to initialize a float4 variable would be:

float4 f = (float4)(0.f,1.f,2.f,3.f);

as stated in the OpenCL 1.1 spec, section 6.1.6.

So the right way to do this whole initialization is:

__constant A d = {(float4)(0,1,2,3),(float2)(4,5),(float4)(6,7,8,9)};

Other forms of initialization might compile, but the result might not be as intended.

For example, __constant A a = {0,1,2,3,4,5,6,7,8,9}; will initialize the first vector (f) with 0 in all its elements, the second vector (g) will be initialized with the value (float2)(1,1), and the third (h) vector will hold (float4)(2,2,2,2). The rest of the initializer list will be thrown away.

Thanks

Guy

Bild des Benutzers Sergey Kostrov
Quoting Guy Benyei (Intel) ...
Other forms of initialization might compile, but the result might not be as intended.
...

I decided to reproducethe casewith anMS C/C++ compiler ( VS 2005 ). A C structure lookslike:

typedef struct tagA
{
float f[4];
float g[2];
float h[4];
} A;

So,after initialization avariable 'v' of a type Alooked like:

v.f[0] = 0
v.f[1] = 1
v.f[2] = 2
v.f[3] = 3

v.g[0]= 4
v.g[1]= 5

v.h[0]= 6
v.h[1]= 7
v.h[2]= 8
v.h[3]= 9

in both cases ( described in the initial post ).

It doesn't matter if you declare a set of numbers for initialization with brackets or without brackets.

Best regards,
Sergey

Melden Sie sich an, um einen Kommentar zu hinterlassen.