Merge pull request #2085 from psychocrypt/topic-amdOptimizeDiv

OpenCL: optimize cn-heavy div
fireice-uk · Nov 20, 2018 · 1b2b4d3 · 1b2b4d3
2 parents b7ffd6b + 9813e1c
commit 1b2b4d3
Showing 1 changed file with 4 additions and 8 deletions.
diff --git a/xmrstak/backend/amd/amd_gpu/opencl/fast_div_heavy.cl b/xmrstak/backend/amd/amd_gpu/opencl/fast_div_heavy.cl
@@ -6,23 +6,19 @@ inline long fast_div_heavy(long _a, int _b)
 {
 	long a = abs(_a);
 	int b = abs(_b);
-
 	float rcp = native_recip(convert_float_rte(b));
 	float rcp2 = as_float(as_uint(rcp) + (32U << 23));
-
-	ulong q1 = convert_ulong_rte(convert_float_rte(as_int2(a).s1) * rcp2);
+	ulong q1 = convert_ulong(convert_float_rte(as_int2(a).s1) * rcp2);
 	a -= q1 * as_uint(b);
-
-	long q2 = convert_long_rte(convert_float_rtn(a) * rcp);
+	float q2f = convert_float_rte(as_int2(a >> 12).s0) * rcp;
+	q2f = as_float(as_uint(q2f) + (12U << 23));
+	long q2 = convert_long_rte(q2f);
 	int a2 = as_int2(a).s0 - as_int2(q2).s0 * b;
-
 	int q3 = convert_int_rte(convert_float_rte(a2) * rcp);
 	q3 += (a2 - q3 * b) >> 31;
-
 	const long q = q1 + q2 + q3;
 	return ((as_int2(_a).s1 ^ _b) < 0) ? -q : q;
 }
 
 #endif
 )==="
-