try fix

Author: JohannesGaessler

ggml/src/ggml-cuda/common.cuh

@@ -551,7 +551,7 @@ static __device__ __forceinline__ void ggml_cuda_fma(float & acc, const float v,
 
 static __device__ __forceinline__ void ggml_cuda_fma(float & acc, const half2 v, const half2 u) {
 #if defined(GGML_USE_HIP) && defined(GCN)
-    asm volatile("v_dot2_f32_f16 %0, %1, %2, %0" : "=v"(acc) : "v"(v), "v"(u));
+    asm volatile("v_dot2_f32_f16 %0, %1, %2, %0" : "+v"(acc) : "v"(v), "v"(u));
 #else
 #ifdef FAST_FP16_AVAILABLE
     const float2 tmp = __half22float2(v*u);