Moving from loops to kernel based thinking is important a parallel kernel is not a loop and doesn’t have iteratns instead it works on data parallel way and multiple instances of that operation are executed.
Multidimensional Kernels
Think of it as an abstraction over 2D 3D data which maybe 1D data another consideration is if memory is contigous or not.
-> Contiguous 1 | 2 | 3 | 4 ↓ non Contiguous 5 | 6 | 7 | 8
if it’s more than 3d then user must specifiy mapping between multiple dimensions and linear indices manually using modulo arithmetic or another method.
Overview of Language Features
Different types of kernel what we want, what we need we can express in different ways.
Separating Kernels from Host Code
We use lambda expressions like defined in previous chapters.
Different Types of Parallel Kernels
Three Types.
-
Basic Data Parallel Kernels-> easier to understand and reason we give up more access to low level features like scheduling, increased complexity difficult to reason about their performance.
-
Extends Basic called ND-range N Dimensional useful for low level performance tuning. It enables certain kernel instances to be grouped together.
-
Referred as Hierarchical data experimental syntax of second form syntax similar to nested parallel loops.
Date: 2026-08-06
The Basic Kernel
The Range Class
Templatized defining dimensions of the range can be 1d 2d 3d both execution ranges of parallel constructs and sizes of buffers.
template<int Dmensions=1>
class Range{};The ID Class
Same as range class but argument to a kernel function must be forwarded to any where we want to query the index.
The Item Class
ID class only provides index in the range that it is computing if we want to know about range and id both we need item class.
template<int Dimensions = 1, bool WithOffset = true>
class item{};The ND Range Kernel
Not flat anymore items belong to groups some sort of locality different behaviours for different groups gives more insight on how work is mapped to specific hardware platforms.
Work-Items
Individual instances of kernel functions
Work groups
Work Items are organized in work groups and work items in different work groups cannot communicate.
-> work item in work groups -> work group local memory -> work item can synchronize using work group barriers and work group memory fences -> work item in work groups -> access to group functions
Sub-Groups
subset of work items in work groups sub groups can utilize compiler vectorization SIMD register each work-item in subgroup corresponds to SIMD lane.
The nd_range Class
two instances denoting local and global execution range
The nd_item Class
Again encapsulating the execution range of a kernel
The group Class
maps all functionality to work groups.
The sub_group Class
same as groups for sub groups.
Mapping Computation to Work-Items
Mapping can be improved.
- One-to-One Mapping -> size matching amount of work
- Many-to-One mapping -> not amount of work but number of workers
Choosing a kernel form
Flow Chart describes it in the book basically goes like first parallel language use basic kernel if implementation does not exists -> prototype -> basic kernel if it exists and language is OpenCL or Cuda ND-Range Kernel other language embarrasingly parallel basic development performance nd range productivity basic data locality re-use yes nd range no basic_kernel