Skip to content

[HIP][CUDA] local size bytes #2634

Open

Description

@jinz2014

HIP_FUNC_ATTRIBUTE_LOCAL_SIZE_BYTES | The local memory usage of each thread by this function in bytes.
HIP_FUNC_ATTRIBUTE_NUM_REGS | The number of registers used by each thread of this function.

When the local size bytes is the memory usage of each thread, should the following case be moved to "urKernelGetInfo" ?

  case UR_KERNEL_GROUP_INFO_PRIVATE_MEM_SIZE: {
    // OpenCL PRIVATE == CUDA LOCAL
    int Bytes = 0;
    UR_CHECK_ERROR(cuFuncGetAttribute(
        &Bytes, CU_FUNC_ATTRIBUTE_LOCAL_SIZE_BYTES, hKernel->get()));
    return ReturnValue(uint64_t(Bytes));

Activity

  1. added
    hipHIP adapter specific issues
    cudaCUDA adapter specific issues
    on Jan 29, 2025
  2. changed the title [-][HIP] local size bytes[/-] [+][HIP][CUDA] local size bytes[/+] on Feb 5, 2025
  3. npmiller commented on Feb 18, 2025

    @npmiller
    Contributor

    This is more of a specification change proposal, essentially to move UR_KERNEL_GROUP_INFO_PRIVATE_MEM_SIZE from ur_kernel_group_info_t to ur_kernel_info_t.

    We couldn't change it just for HIP and CUDA.

    We would need to check if this is possible for all targets, but in theory I think it makes sense this information shouldn't be affected by the work size, so it shouldn't need to be in the group info.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Labels

cudaCUDA adapter specific issueshipHIP adapter specific issuesspecificationChanges or additions to the specification

Type

No type

Projects

No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions