Archives Discussions

[bubu]

Do you respect 100% the data passed from the host pointer in a clCreateImage2D call? or do you perform some kind of data conversion/filtering when you upload the data from the CPU to the GPU?

The format is GL_RGBA and GL_FLOAT.

Sampler uses CLK_NORMALIZED_COORDS_FALSE | CLK_ADDRESS_NONE | CLK_FILTER_NEAREST;

I read a float4 with read_imagef(img,samp,int2). The .xyz are floats and the .w is really an int.

I ask you this because  I have some very sensible data that I need to convert using as_int and seems the data is not correctly preserved when I launch the kernel... Or perhaps your as_int() is bugged?(I can't find a call to as_int() in your examples to test). Or perhaps the endianness is affecting my data ( but I upload the data from an Intel CPU to the GPU via clCreateImage2D-hostPtr ...)

I'll put a code example so you can understand my question better:

Host code:

cl_float4 data;

data.s[0] = 1.0f;

data.s[1] = 2.0f;

data.s[2] = 3.0f;

int i = -4;

data.s[3] = *((float*)&i);

.... then I create a 1x1 image:

cl_image_format l_sImageFmt;
l_sImageFmt.image_channel_data_type = CL_FLOAT;
l_sImageFmt.image_channel_order = CL_RGBA;

img = clCreateImage2D ( ctx, CL_MEM_READ_ONLY|CL_MEM_COPY_HOST_PTR, &l_sImageFmt,
                1, 1, sizeof(cl_float4), &data, NULL );

GPU Kernel:

__constant sampler_t imgSamplerInt2 = CLK_NORMALIZED_COORDS_FALSE | CLK_ADDRESS_NONE | CLK_FILTER_NEAREST;

const float4 data = read_imagef ( img, imgSamplerInt2, (int2)(0,0) );

const int i = as_int(data.w); //i should be -4, right? Well, it is NOT! (it's thrash )

And btw... as_int() is much faster than a normal int cast if I have the data properly formatted, isn't it?

[nou]

you must take other way and read from texture as int and cinvert it into float. because small int will be flushed to zero as it is subnormal floats.

[bubu]

So the float4 values are not respected when the texture is uploaded? why!!!! ????

[nou]

http://forums.amd.com/devforum/messageview.cfm?catid=390&threadid=133069&enterthread=y

[MicahVillmow]

Bubu,
Nou answered your question, 'small int will be flushed to zero as it is subnormal floats'. If you want unmodified values, read them as unsigned integers.

[bubu]

Originally posted by: MicahVillmow If you want unmodified values, read them as unsigned integers.

What "read them as unsigned integers" means, pls?

Can you modify the code I posted reflecting that, pls?

thx

[MicahVillmow]

instead of using the function read_imagef, just use read_imageui and change the image_data_type to CL_UNSIGNED_INT32. Then after you read the data, convert to float with as_float.

[bubu]

Originally posted by: MicahVillmow instead of using the function read_imagef, just use read_imageui and change the image_data_type to CL_UNSIGNED_INT32. Then after you read the data, convert to float with as_float.

Ok I understand now.

Btw, do you think that would be faster than type casting?

[MicahVillmow]

bubu,
Please reference section 6.2.4 of the OpenCL spec on type casting in OpenCL.