From: inode64 overlay
Subject: [PATCH] HIP: match AMDGPU BF16 WMMA builtin operand types

Clang declares both gfx11 and gfx12 BF16 WMMA builtins with vectors of
short operands.  This is also how rocWMMA calls the builtins.  Reinterpret
the packed BF16 fragments as the declared operand type before calling them.

--- a/ggml/src/ggml-cuda/mma.cuh
+++ b/ggml/src/ggml-cuda/mma.cuh
@@ -1258,7 +1258,7 @@ struct mma_config {
             tile<16, 16, float, dl_d> & D, const tile<16, 8, nv_bfloat162, dl_ab> & A, const tile<16, 8, nv_bfloat162, dl_ab> & B) {
 #if defined(AMD_WMMA_AVAILABLE)
 #if defined(RDNA4)
-        using bf16x8_t = __attribute__((ext_vector_type(8))) __bf16;
+        using bf16x8_t = __attribute__((ext_vector_type(8))) short;
         using floatx8_t = __attribute__((ext_vector_type(8))) float;
         floatx8_t& acc_frag = reinterpret_cast<floatx8_t&>(D.x[0]);
         const bf16x8_t& a_frag = reinterpret_cast<const bf16x8_t&>(A.x[0]);
@@ -1266,7 +1266,7 @@ struct mma_config {
         const bf16x8_t& b_frag = reinterpret_cast<const bf16x8_t&>(B.x[0]);
         acc_frag = __builtin_amdgcn_wmma_f32_16x16x16_bf16_w32_gfx12(a_frag, b_frag, acc_frag);
 #elif defined(RDNA3)
-        using bf16x16_t = __attribute__((ext_vector_type(16))) __bf16;
+        using bf16x16_t = __attribute__((ext_vector_type(16))) short;
         using floatx8_t = __attribute__((ext_vector_type(8))) float;
         floatx8_t& acc_frag = reinterpret_cast<floatx8_t&>(D.x[0]);
         const bf16x16_t& a_frag = reinterpret_cast<const bf16x16_t&>(A.x[0]);