small change

Author: JohannesGaessler

ggml-cuda.cu

@@ -1408,19 +1408,20 @@ static __device__ __forceinline__ float vec_dot_q2_K_q8_1(
     const float d = bq2_K->d;
     const float dmin = bq2_K->dmin;
 
+    const int vi = *((int *) &bq2_K->qs[4*iqs]);
+
     for (int i = 0; i < 4; ++i) {
         const int sc = bq2_K->scales[iqs - iqs%8 + (iqs%8) / 4 + 2*i];
         const float dl = d    * (sc & 0xF);
         const float ml = dmin * (sc >> 4);
 
-        int qs = *((int *) &bq2_K->qs[4*iqs]);
-        qs = (qs >> (2*i)) & 0x03030303;
+        const int vii = (vi >> (2*i)) & 0x03030303;
 
-        const block_q8_1 * bq8i = bq8_1 + (iqs < 8 ? 0 : 4) + i;
+        const block_q8_1 * bq8i = bq8_1 + 4 * (iqs/8) + i;
         const float d8 = bq8i->d;
         const int qs8 = *((int*) &bq8i->qs[4*(iqs%8)]);
 
-        sumf += d8*(dl*__dp4a(qs, qs8, 0) - ml*__dp4a(0x01010101, qs8, 0));
+        sumf += d8*(dl*__dp4a(vii, qs8, 0) - ml*__dp4a(0x01010101, qs8, 0));
     }
 
     return sumf;