You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Hi! Is there a way that I can have a list of integers inside a kernel representing pointers to array data of known sizes?
Curretly I am working with big image arrays of different sizes that barely fit in memory, and thus I can not copy them into some sort of contiguous array due to copy delays and memory space. Moreover, I will be loading and unloading a portion of these images in a dynamic way as the computation progresses. Which images will be loaded or unloaded will depend on other factors, but some will remain while others will be unloaded and new ones will be loaded. Since images can have vastly different sizes, I don't think I can use a sort of rolling buffer. I would like to use warp for processing, but currently I have not found a way to declare inside a warp kernel that an integer is an array pointer. Which image I am going to access will depend on computation, and will vary from iteration to iteration inside the same kernel. What I wanted to achive is something like the following:
image0=wp.array(...) # 10000 x 10000 imageimage1=wp.array(...) # 10000 x 10000 imageimage2=wp.array(...) # 10000 x 10000 imageimage_source=wp.array(...) # 10000 x 10000 image@wp.kerneldefkn(image_src, image_trgts, image_trgts_shape):
x, y=wp.tid()
target_image_index= (...) # target image index will depend on some computation for the current image pixeltarget_x= (...) # target image coordinates will depend on some other computationtarget_y= (...) # target image coordinates will depend on some other computation# get the sape of target image from the shapes arraytgt_h=image_trgts_shape[target_image, 0]
tgt_w=image_trgts_shape[target_image, 1]
# declare the array for the target image that I want to accesstarget=wp.array(pointer=image_trgts[target_image], shape=(tgt_h, tgt_w, 3))
pixel_rgb=wp.vec3f(
target[target_y, target_x, 0],
target[target_y, target_x, 1],
target[target_y, target_x, 2])
# finally work with target image pixel
(...)
The real computation has plenty more steps, however this example outlines the issue.
Thank you in advance
The text was updated successfully, but these errors were encountered:
Hi @BernardoCovas, this looks like an interesting use case, and I can see that having the ability to construct an array at kernel time from ptr would be useful indeed.
Let me discuss with the team and get back to you on how we can support this.
Just want to ask a clarifying question - in your example, would image_trgts be an array of e.g.: uint64 addresses representing the array memory for an image?
Hi! Is there a way that I can have a list of integers inside a kernel representing pointers to array data of known sizes?
Curretly I am working with big image arrays of different sizes that barely fit in memory, and thus I can not copy them into some sort of contiguous array due to copy delays and memory space. Moreover, I will be loading and unloading a portion of these images in a dynamic way as the computation progresses. Which images will be loaded or unloaded will depend on other factors, but some will remain while others will be unloaded and new ones will be loaded. Since images can have vastly different sizes, I don't think I can use a sort of rolling buffer. I would like to use warp for processing, but currently I have not found a way to declare inside a warp kernel that an integer is an array pointer. Which image I am going to access will depend on computation, and will vary from iteration to iteration inside the same kernel. What I wanted to achive is something like the following:
The real computation has plenty more steps, however this example outlines the issue.
Thank you in advance
The text was updated successfully, but these errors were encountered: