Skip to content

fix(hip): restrict wavefront guard to device pass #628

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
pepuscz wants to merge 1 commit into
base: main
Choose a base branch
from
+7 −4
Open

fix(hip): restrict wavefront guard to device pass #628

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
11 changes: 7 additions & 4 deletions server/src/flashprefill_kernels.hip.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -29,12 +29,15 @@
// which is a 32-lane instruction. A wave64 target would need a different WMMA
// instruction and fragment layout, not a warp-width tweak, so we fail the build
// loudly rather than emit silently-wrong results. All currently declared AMD
// targets (gfx1100 / gfx1151, RDNA3) are wave32. __AMDGCN_WAVEFRONT_SIZE(__) is
// defined only in the device compile pass.
#if defined(__AMDGCN_WAVEFRONT_SIZE__) && (__AMDGCN_WAVEFRONT_SIZE__ != 32)
// targets (gfx1100 / gfx1151, RDNA3) are wave32. Check the compiler-provided
// wavefront macros only in the device pass: some ROCm versions expose a
// host-side definition that is not an integer constant expression.
#if defined(__HIP_DEVICE_COMPILE__) && __HIP_DEVICE_COMPILE__
# if defined(__AMDGCN_WAVEFRONT_SIZE__) && (__AMDGCN_WAVEFRONT_SIZE__ != 32)
# error "flashprefill_kernels.hip.cu requires a 32-lane wavefront (RDNA v_wmma_f32_16x16x16 fragment layout). Build for a wave32 target (gfx10/gfx11) or provide a wave64 WMMA rewrite."
#elif !defined(__AMDGCN_WAVEFRONT_SIZE__) && defined(__AMDGCN_WAVEFRONT_SIZE) && (__AMDGCN_WAVEFRONT_SIZE != 32)
# elif !defined(__AMDGCN_WAVEFRONT_SIZE__) && defined(__AMDGCN_WAVEFRONT_SIZE) && (__AMDGCN_WAVEFRONT_SIZE != 32)
# error "flashprefill_kernels.hip.cu requires a 32-lane wavefront (RDNA v_wmma_f32_16x16x16 fragment layout). Build for a wave32 target (gfx10/gfx11) or provide a wave64 WMMA rewrite."
# endif
#endif

namespace dflash::common {
Expand Down
Loading