Fix iqk_mul_mat on AVX512 systems that are missing BF16 support (#204)

* Fix iqk_mul_mat on AVX512 systems that are missing BF16 support

* One more

---------

Co-authored-by: Iwan Kawrakow <[email protected]>

Author: ikawrakow

ggml/src/iqk/iqk_mul_mat.cpp

@@ -8082,6 +8082,9 @@ struct QFBase {
     using Acc  = __m512;
     static inline Data load(const ggml_half * x) { return _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)x)); }
     static inline Data load(const float * x) { return _mm512_loadu_ps(x); }
+    static inline Data load(const ggml_bf16_t * x) {
+        return _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_cvtepu16_epi32(_mm256_loadu_si256((const __m256i*)x)), 16));
+    }
     static inline Acc acc(Acc prev, const Data& y, const Data& x) {
         return _mm512_fmadd_ps(y, x, prev);
     }
@@ -16079,7 +16082,7 @@ inline void iqk_flash_helper_T(KHelper& kh, ggml_type type_v,
             HelperF16<Dv, k_step> vh(v, stride_v);
             iqk_flash_helper<Dk, Dv, k_step>(kh, vh, nq1, nk1, stride_q, stride_m, stride_qkv, q, mask, scale, softcap, qkv);
         } break;
-#ifdef HAVE_FANCY_SIMD
+#ifdef __AVX512BF16__
         case GGML_TYPE_BF16: {
             HelperBF16<Dv, k_step> vh(v, stride_v);
             iqk_flash_helper<Dk, Dv, k_step>(kh, vh, nq1, nk1, stride_q, stride_m, stride_qkv, q, mask, scale, softcap, qkv);