Merge pull request #2086 from psychocrypt/topic-amdOptimizeCNv8Div

OpenCl: optimize cn-v8 div

xmrstak/backend/amd/amd_gpu/opencl/fast_int_math_v2.cl

@@ -68,61 +68,35 @@ inline uint get_reciprocal(const __local uchar *RCP, uint a)
 
 inline uint2 fast_div_v2(const __local uint *RCP, ulong a, uint b)
 {
-	const uint r = get_reciprocal((const __local uchar *)RCP, b);
-	const ulong k = mul_hi(as_uint2(a).s0, r) + ((ulong)(r) * as_uint2(a).s1) + a;
-
-	ulong q;
-	((uint*)&q)[0] = as_uint2(k).s1;
-
-#if defined(cl_amd_device_attribute_query) && (OPENCL_DRIVER_MAJOR == 14)
-	/* The AMD driver 14.XX is not able to compile `(k < a)`
-	 * https://github.com/fireice-uk/xmr-stak/issues/1922
-	 * This is a workaround for the broken compiler.
-	 */
-	 ulong whyAMDwhy;
-	((uint*)&whyAMDwhy)[0] = as_uint2(k).s0;
-	((uint*)&whyAMDwhy)[1] = as_uint2(k).s1;
-	((uint*)&q)[1] = (whyAMDwhy < a) ? 1U : 0U;
-#else
-	((uint*)&q)[1] = (k < a) ? 1U : 0U;
-#endif
-	
-	const long tmp = a - q * b;
-	const bool overshoot = (tmp < 0);
-	const bool undershoot = (tmp >= b);
-
-	return (uint2)(
-		as_uint2(q).s0 + (undershoot ? 1U : 0U) - (overshoot ? 1U : 0U),
-		as_uint2(tmp).s0 + (overshoot ? b : 0U) - (undershoot ? b : 0U)
-	);
+    const uint r = get_reciprocal((const __local uchar *)RCP, b);
+
+    const ulong k = mul_hi(as_uint2(a).s0, r) + ((ulong)(r) * as_uint2(a).s1) + a;
+    const uint q = as_uint2(k).s1;
+    long tmp = a - ((ulong)(q) * b);
+    ((int*)&tmp)[1] -= (as_uint2(k).s1 < as_uint2(a).s1) ? b : 0;
+    const int overshoot = ((int*)&tmp)[1] >> 31;
+    const int undershoot = as_int2(as_uint(b - 1) - tmp).s1 >> 31;
+    return (uint2)(q + overshoot - undershoot, as_uint2(tmp).s0 + (as_uint(overshoot) & b) - (as_uint(undershoot) & b));
 }
-
 inline uint fast_sqrt_v2(const ulong n1)
 {
-	float x = as_float((as_uint2(n1).s1 >> 9) + ((64U + 127U) << 23));
-
-	float x1 = native_rsqrt(x);
-	x = native_sqrt(x);
-
-	// The following line does x1 *= 4294967296.0f;
-	x1 = as_float(as_uint(x1) + (32U << 23));
-
-	const uint x0 = as_uint(x) - (158U << 23);
-	const long delta0 = n1 - (((long)(x0) * x0) << 18);
-	const float delta = convert_float_rte(as_int2(delta0).s1) * x1;
-
-	uint result = (x0 << 10) + convert_int_rte(delta);
-	const uint s = result >> 1;
-	const uint b = result & 1;
-
-	const ulong x2 = (ulong)(s) * (s + b) + ((ulong)(result) << 32) - n1;
-	if ((long)(x2 + b) > 0) --result;
-	if ((long)(x2 + 0x100000000UL + s) < 0) ++result;
-
-	return result;
+    float x = as_float((as_uint2(n1).s1 >> 9) + ((64U + 127U) << 23));
+    float x1 = native_rsqrt(x);
+    x = native_sqrt(x);
+    // The following line does x1 *= 4294967296.0f;
+    x1 = as_float(as_uint(x1) + (32U << 23));
+    const uint x0 = as_uint(x) - (158U << 23);
+    const long delta0 = n1 - (as_ulong((uint2)(mul24(x0, x0), mul_hi(x0, x0))) << 18);
+    const float delta = convert_float_rte(as_int2(delta0).s1) * x1;
+    uint result = (x0 << 10) + convert_int_rte(delta);
+    const uint s = result >> 1;
+    const uint b = result & 1;
+    const ulong x2 = (ulong)(s) * (s + b) + ((ulong)(result) << 32) - n1;
+    if ((long)(x2 + as_int(b - 1)) >= 0) --result;
+    if ((long)(x2 + 0x100000000UL + s) < 0) ++result;
+    return result;
 }
 
 #endif
 
 )==="
-        
\ No newline at end of file