Moved lcd/gcd and strideAlign functions from bgfx.

include/bx/uint32_t.h

@@ -679,6 +679,64 @@ namespace bx
 #endif // BX_COMPILER_
 	}
 
+	/// Greatest common divisor.
+	inline uint32_t uint32_gcd(uint32_t _a, uint32_t _b)
+	{
+		do
+		{
+			uint32_t tmp = _a % _b;
+			_a = _b;
+			_b = tmp;
+		}
+		while (_b);
+
+		return _a;
+	}
+
+	/// Least common multiple.
+	inline uint32_t uint32_lcm(uint32_t _a, uint32_t _b)
+	{
+		return _a * (_b / uint32_gcd(_a, _b) );
+	}
+
+	/// Align to arbitrary stride.
+	inline uint32_t strideAlign(uint32_t _offset, uint32_t _stride)
+	{
+		const uint32_t mod    = uint32_mod(_offset, _stride);
+		const uint32_t add    = uint32_sub(_stride, mod);
+		const uint32_t mask   = uint32_cmpeq(mod, 0);
+		const uint32_t tmp    = uint32_selb(mask, 0, add);
+		const uint32_t result = uint32_add(_offset, tmp);
+
+		return result;
+	}
+
+	/// Align to arbitrary stride and 16-bytes.
+	inline uint32_t strideAlign16(uint32_t _offset, uint32_t _stride)
+	{
+		const uint32_t align  = uint32_lcm(16, _stride);
+		const uint32_t mod    = uint32_mod(_offset, align);
+		const uint32_t mask   = uint32_cmpeq(mod, 0);
+		const uint32_t tmp0   = uint32_selb(mask, 0, align);
+		const uint32_t tmp1   = uint32_add(_offset, tmp0);
+		const uint32_t result = uint32_sub(tmp1, mod);
+
+		return result;
+	}
+
+	/// Align to arbitrary stride and 256-bytes.
+	inline uint32_t strideAlign256(uint32_t _offset, uint32_t _stride)
+	{
+		const uint32_t align  = uint32_lcm(256, _stride);
+		const uint32_t mod    = uint32_mod(_offset, align);
+		const uint32_t mask   = uint32_cmpeq(mod, 0);
+		const uint32_t tmp0   = uint32_selb(mask, 0, align);
+		const uint32_t tmp1   = uint32_add(_offset, tmp0);
+		const uint32_t result = uint32_sub(tmp1, mod);
+
+		return result;
+	}
+
 } // namespace bx
 
 #endif // BX_UINT32_T_H_HEADER_GUARD

tests/uint32_t.cpp

@@ -0,0 +1,22 @@
+/*
+ * Copyright 2010-2015 Branimir Karadzic. All rights reserved.
+ * License: http://www.opensource.org/licenses/BSD-2-Clause
+ */
+
+#include "test.h"
+#include <bx/uint32_t.h>
+
+TEST(StrideAlign)
+{
+	CHECK(0  == bx::strideAlign(0, 12) );
+	for (uint32_t ii = 0; ii < 12; ++ii)
+	{
+		CHECK(12 == bx::strideAlign(ii+1, 12) );
+	}
+
+	CHECK(0  == bx::strideAlign16(0, 12) );
+	for (uint32_t ii = 0; ii < 12; ++ii)
+	{
+		CHECK(48 == bx::strideAlign16(ii+1, 12) );
+	}
+}