Skip to content

HIP: add CDNA4 gfx950 (AMD Instinct MI350) support - #302

Open
wensongc wants to merge 1 commit into
RenderKit:masterfrom
wensongc:hip-gfx950
Open

wensongc wants to merge 1 commit into
RenderKit:masterfrom
wensongc:hip-gfx950

Conversation

@wensongc

@wensongc wensongc commented Nov 6, 2026

Copy link
Copy Markdown

Adds CDNA4 (gfx950, Instinct MI350) support to the HIP device.

  • RDNA WMMA/DL kernels are unchanged and stay in the same library.
  • gfx950 MFMA convolution (ck_conv_mfma.cpp) is compiled as its own translation unit with -mwavefrontsize64, because CDNA is wave64 and cannot share a fat binary with the wave32 WMMA kernels.
  • OIDN_HIP_GFX950 (default ON) enables it; the option is skipped automatically if the compiler does not accept --offload-arch=gfx950.
  • Composable Kernel submodule is not modified; ck_conv_mfma.cpp defines warpSize before including CK so that ck::get_warp_size() is constexpr on ROCm 7.

Tested:

  • Default build (RDNA targets + gfx950) on ROCm 7.2.3, MI350X (gfx950): oidnTest --device hip, all 2403 assertions pass.
  • OIDN_HIP_GFX950=OFF build compiles.
  • 640x480 path-traced frame with albedo/normal: HIP vs official 2.5.1 CPU, MAE 1.8e-4, max 2.9e-3.
  • Not tested: RDNA hardware (only compiled). Will do it very soon.

Keep the RDNA kernels in the same device library. gfx950 MFMA convolutions are a separate translation unit because they are wave64 and cannot share a fat binary with the RDNA WMMA kernels.

Co-authored-by: Cursor <cursoragent@cursor.com>
@wensongc wensongc changed the title HIP: add CDNA4 gfx950 (Instinct MI350) support HIP: add CDNA4 gfx950 (AMD Instinct MI350) support Nov 6, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant