Try to optimize ggml a bit

Author: ocrickard

ggml.c

@@ -2288,17 +2288,31 @@ static void ggml_vec_dot_q4_1(const int n, float * restrict s, const void * rest
         const uint8_t * restrict p0 = x[i].qs;
         const uint8_t * restrict p1 = y[i].qs;
 
-        for (int j = 0; j < QK/2; j++) {
-            const uint8_t v0 = p0[j];
-            const uint8_t v1 = p1[j];
-
-            const float f0 = d0*(v0 & 0xf) + m0;
-            const float f1 = d0*(v0 >> 4)  + m0;
-
-            const float f2 = d1*(v1 & 0xf) + m1;
-            const float f3 = d1*(v1 >> 4)  + m1;
-
-            sumf += f0*f2 + f1*f3;
+        for (int j = 0; j < QK/4; j++) {
+            const uint32_t v0 = ((uint32_t *)p0)[j];
+            const uint32_t v1 = ((uint32_t *)p1)[j];
+
+            const uint8_t v0_0 = (v0 >>  0) & 0xf;
+            const uint8_t v0_1 = (v0 >>  4) & 0xf;
+            const uint8_t v0_2 = (v0 >>  8) & 0xf;
+            const uint8_t v0_3 = (v0 >> 12) & 0xf;
+
+            const uint8_t v1_0 = (v1 >>  0) & 0xf;
+            const uint8_t v1_1 = (v1 >>  4) & 0xf;
+            const uint8_t v1_2 = (v1 >>  8) & 0xf;
+            const uint8_t v1_3 = (v1 >> 12) & 0xf;
+
+            const float f0 = d0 * v0_0 + m0;
+            const float f1 = d0 * v0_1 + m0;
+            const float f2 = d0 * v0_2 + m0;
+            const float f3 = d0 * v0_3 + m0;
+
+            const float f4 = d1 * v1_0 + m1;
+            const float f5 = d1 * v1_1 + m1;
+            const float f6 = d1 * v1_2 + m1;
+            const float f7 = d1 * v1_3 + m1;
+
+            sumf += f0 * f4 + f1 * f5 + f2 * f6 + f3 * f7;
         }
     }
 #endif