ggml : fix Q8_0 and Q8_1 rounding

ggerganov · ggerganov · commit 15fb145fc9d7 · 2023-05-11T20:47:41.000+03:00
diff --git a/ggml.c b/ggml.c
@@ -925,8 +925,8 @@ static void quantize_row_q8_0_reference(const float * restrict x, block_q8_0 * r
             const float v0 = x[i*QK8_0 + 2*j + 0]*id;
             const float v1 = x[i*QK8_0 + 2*j + 1]*id;
 
-            y[i].qs[          j] = v0 + 0.5f;
-            y[i].qs[QK8_0/2 + j] = v1 + 0.5f;
+            y[i].qs[          j] = roundf(v0);
+            y[i].qs[QK8_0/2 + j] = roundf(v1);
         }
     }
 }
@@ -1083,8 +1083,8 @@ static void quantize_row_q8_1_reference(const float * restrict x, block_q8_1 * r
             const float v0 = x[i*QK8_1 + 2*j + 0]*id;
             const float v1 = x[i*QK8_1 + 2*j + 1]*id;
 
-            y[i].qs[          j] = v0 + 0.5f;
-            y[i].qs[QK8_1/2 + j] = v1 + 0.5f;
+            y[i].qs[          j] = roundf(v0);
+            y[i].qs[QK8_1/2 + j] = roundf(v1);
 
             sum += y[i].qs[          j];
             sum += y[i].qs[QK8_1/2 + j];

Original file line number	Diff line number	Diff line change
`@@ -925,8 +925,8 @@ static void quantize_row_q8_0_reference(const float * restrict x, block_q8_0 * r`
`925`	`925`	`const float v0 = x[iQK8_0 + 2j + 0]*id;`
`926`	`926`	`const float v1 = x[iQK8_0 + 2j + 1]*id;`
`927`	`927`
`928`		`- y[i].qs[ j] = v0 + 0.5f;`
`929`		`- y[i].qs[QK8_0/2 + j] = v1 + 0.5f;`
	`928`	`+ y[i].qs[ j] = roundf(v0);`
	`929`	`+ y[i].qs[QK8_0/2 + j] = roundf(v1);`
`930`	`930`	`}`
`931`	`931`	`}`
`932`	`932`	`}`
`@@ -1083,8 +1083,8 @@ static void quantize_row_q8_1_reference(const float * restrict x, block_q8_1 * r`
`1083`	`1083`	`const float v0 = x[iQK8_1 + 2j + 0]*id;`
`1084`	`1084`	`const float v1 = x[iQK8_1 + 2j + 1]*id;`
`1085`	`1085`
`1086`		`- y[i].qs[ j] = v0 + 0.5f;`
`1087`		`- y[i].qs[QK8_1/2 + j] = v1 + 0.5f;`
	`1086`	`+ y[i].qs[ j] = roundf(v0);`
	`1087`	`+ y[i].qs[QK8_1/2 + j] = roundf(v1);`
`1088`	`1088`
`1089`	`1089`	`sum += y[i].qs[ j];`
`1090`	`1090`	`sum += y[i].qs[QK8_1/2 + j];`