Skip to content

[Issue]: amdgcn-cfg.py fails to create CFG for the FA kernel - i.e., attn_fwd kernel #671

@ravil-mobile

Description

@ravil-mobile

Problem Description

amdgcn-cfg.py fails to create an edge because there is the following basic block in the middle. The script didn't take into account an early kernel terminator

.LBB0_40:
	.loc	1 626 4                         ; flash-attention.py:626:4
	s_endpgm

The CFG logic needs to be fixed (somewhere around get_block_list function).

The output looks as follows If one deletes LBB0_40 label and a jump to this label

attn_fwd svg kernel-0

Operating System

Ubuntu 22.04.5 LTS

CPU

MI300X

GPU

AMD

ROCm Version

ROCm 6.2.3

ROCm Component

ROCm

Steps to Reproduce

No response

(Optional for Linux users) Output of /opt/rocm/bin/rocminfo --support

No response

Additional Information

No response

Metadata

Metadata

Assignees

Labels

No labels
No labels

Type

No type

Projects

No projects

Milestone

No milestone

Relationships

None yet

Development

No branches or pull requests

Issue actions