CUDA* to SYCL* Term Mapping Quick Reference

Intel® DPC++ Compatibility Tool Developer Guide and Reference

Download PDF

ID 768918

Date 6/24/2024

Version

Public

A newer version of this document is available. Customers should click here to go to the newest version.

Visible to Intel only — GUID: GUID-9DB6F500-276F-4B8D-9DA8-13F45A4550FC

View Details

CUDA* to SYCL* Term Mapping Quick Reference

This quick reference maps common CUDA* terms to SYCL* terms, to help start the migration process.

Architecture Terminology Mapping

CUDA Capable GPU	SYCL Capable GPU from Intel
	Xe-LP and prior generations	Xe-HPG and Xe-HPC
CUDA Core	Execution Unit (EU)	Vector Engine & Matrix Engine (XVE & XMX)
Streaming Multiprocessor (SM)	Subslice(SS) or Dual Subslice(DSS)	Xe-Core
Processor Clusters (PC)	Slice	Xe-Slice
N/A	N/A	Xe-Stack

Execution Model Mapping

CUDA	SYCL
Thread	Work-item
Warp	Sub-group
Block	Work-group
Grid	ND-range

Memory Model Mapping

CUDA	SYCL
Register	Private memory
Shared memory	Local memory
Constant memory	N/A
Global memory	Global memory

Memory Specifier Mapping

CUDA	SYCL
`__device__`	N/A
`__shared__`	N/A
`__constant__`	N/A
`__managed__`	N/A

Function Execution Space Specifiers Mapping

CUDA	SYCL
`__global__`	Not required
`__device__`	Not required
`__host__`	Not required

Mapping of Key Host Type Used to Offload Task

CUDA	SYCL
device	`sycl::device`
stream	`sycl::queue`
event	`sycl::event`

Kernel Execution Configures Mapping

CUDA	SYCL
`dim3`	`sycl::range<3>`
`Kernel<<<gridDim, blockDim>>>(...)`	Member function `parallel_for` of class `sycl::queue`: `sycl::queue q; q.parallel_for(sycl::nd_range<3>(girdDim * blockDim, blockDim), [=](sycl::nd_item<3> item){ kernel(...); });` Member function `submit` of class `sycl::queue`: `sycl::queue q; q.submit([&](sycl::handler &cgh){ cgh.parallel_for(sycl::nd_range<3>(gridDim * blockDim, blockDim), [=](sycl::nd_item<3> item){ kernel(...); }); });`

CUDA

SYCL

dim3

sycl::range<3>

Kernel<<<gridDim, blockDim>>>(...)

Member function parallel_for of class sycl::queue:

sycl::queue q;
q.parallel_for(sycl::nd_range<3>(girdDim * blockDim, blockDim), [=](sycl::nd_item<3> item){
  kernel(...);
});

Member function submit of class sycl::queue:

sycl::queue q;
q.submit([&](sycl::handler &cgh){
  cgh.parallel_for(sycl::nd_range<3>(gridDim * blockDim, blockDim), [=](sycl::nd_item<3> item){
    kernel(...);
  });
});

Built-In Execution Space Index Mapping

CUDA	SYCL
`gridDim.x/y/z`	`sycl::nd_item.get_group_range(2/1/0)`
`blockDim.x/y/z`	`sycl::nd_item.get_local _range().get(2/1/0)`
`blockIdx.x/y/z`	`sycl::nd_item.get_group(2/1/0)`
`threadIdx.x/y/z`	`sycl::nd_item.get_local_id(2/1/0)`
`warpSize`	`sycl::nd_item. .get_sub_group().get_local_range().get(0)`

Select Your Language

Using Intel.com Search

Quick Links

Recent Searches

Advanced Search

Only search in

Intel® DPC++ Compatibility Tool Developer Guide and Reference

CUDA* to SYCL* Term Mapping Quick Reference

Architecture Terminology Mapping

Execution Model Mapping

Memory Model Mapping

Memory Specifier Mapping

Function Execution Space Specifiers Mapping

Mapping of Key Host Type Used to Offload Task

Kernel Execution Configures Mapping

Built-In Execution Space Index Mapping