AMD Fixes Small Bug That Could Make Your Games Faster
This tiny fix could mean smoother gaming and better AI performance on AMD graphics cards.
A HIP fix corrects the bf16 WMMA operand vector type for gfx11/gfx12. The __builtin_amdgcn_wmma_f32_16x16x16_bf16_w32 and _gfx12 builtins take their A/B operands as vectors of short (_ExtVector<16, short> / _ExtVector<8, short>), not __bf16. Declaring the fragments as ext_vector_type(N) __bf16 fails to compile with an error saying a parameter of type 'short attribute((ext_vector_type(16)))' cannot be initialized with an lvalue of type 'const bf16x16_t' (vector of 16 '__bf16' values).
Because short and __bf16 are both 16-bit and the value is only ever passed to the builtin via reinterpret_cast over raw bytes, changing the element type is bit-identical. The operand type has been short since the builtin was introduced, so the change works across LLVM/ROCm versions without a version gate. The fix addresses HIP builds for RDNA3 (gfx1150/1151) and RDNA4 (gfx1200/1201).
- AMD fixed a bug that prevented developers from building software for new AMD graphics cards.
- The fix affects the latest GPUs like Radeon RX 7000 and 8000 series, used in gaming and AI.
- This means developers can more easily create faster and smarter apps for your AMD-powered device.
Why It Matters
This fix could lead to better gaming and AI performance on AMD graphics cards without any action from you.