Merge pull request #1046 from Nekotekina/new_interpreter

Alternative interpreters
2015-04-04 20:03:18 +03:00
parent abdfcf9d0b 9479bcf77b
commit 952098c00f
26 changed files with 6993 additions and 167 deletions
@@ -88,6 +88,7 @@ union _CRT_ALIGN(16) u128
 	double _d[2];
 	__m128 vf;
 	__m128i vi;
+	__m128d vd;

 	class bit_array_128
 	{
@@ -202,14 +203,21 @@ union _CRT_ALIGN(16) u128
 	static u128 from32p(u32 value)
 	{
 		u128 ret;
-		ret.vi = _mm_set1_epi32((int)value);
+		ret.vi = _mm_set1_epi32(static_cast<s32>(value));
+		return ret;
+	}
+
+	static u128 from16p(u16 value)
+	{
+		u128 ret;
+		ret.vi = _mm_set1_epi16(static_cast<s16>(value));
 		return ret;
 	}

 	static u128 from8p(u8 value)
 	{
 		u128 ret;
-		ret.vi = _mm_set1_epi8((char)value);
+		ret.vi = _mm_set1_epi8(static_cast<s8>(value));
 		return ret;
 	}

@@ -227,16 +235,75 @@ union _CRT_ALIGN(16) u128
 		return ret;
 	}

+	static u128 fromF(__m128 value)
+	{
+		u128 ret;
+		ret.vf = value;
+		return ret;
+	}
+
+	static u128 fromD(__m128d value)
+	{
+		u128 ret;
+		ret.vd = value;
+		return ret;
+	}
+
 	static __forceinline u128 add8(const u128& left, const u128& right)
 	{
 		return fromV(_mm_add_epi8(left.vi, right.vi));
 	}

+	static __forceinline u128 add16(const u128& left, const u128& right)
+	{
+		return fromV(_mm_add_epi16(left.vi, right.vi));
+	}
+
+	static __forceinline u128 add32(const u128& left, const u128& right)
+	{
+		return fromV(_mm_add_epi32(left.vi, right.vi));
+	}
+
+	static __forceinline u128 addfs(const u128& left, const u128& right)
+	{
+		return fromF(_mm_add_ps(left.vf, right.vf));
+	}
+
+	static __forceinline u128 addfd(const u128& left, const u128& right)
+	{
+		return fromD(_mm_add_pd(left.vd, right.vd));
+	}
+
 	static __forceinline u128 sub8(const u128& left, const u128& right)
 	{
 		return fromV(_mm_sub_epi8(left.vi, right.vi));
 	}

+	static __forceinline u128 sub16(const u128& left, const u128& right)
+	{
+		return fromV(_mm_sub_epi16(left.vi, right.vi));
+	}
+
+	static __forceinline u128 sub32(const u128& left, const u128& right)
+	{
+		return fromV(_mm_sub_epi32(left.vi, right.vi));
+	}
+
+	static __forceinline u128 subfs(const u128& left, const u128& right)
+	{
+		return fromF(_mm_sub_ps(left.vf, right.vf));
+	}
+
+	static __forceinline u128 subfd(const u128& left, const u128& right)
+	{
+		return fromD(_mm_sub_pd(left.vd, right.vd));
+	}
+
+	static __forceinline u128 maxu8(const u128& left, const u128& right)
+	{
+		return fromV(_mm_max_epu8(left.vi, right.vi));
+	}
+
 	static __forceinline u128 minu8(const u128& left, const u128& right)
 	{
 		return fromV(_mm_min_epu8(left.vi, right.vi));
@@ -247,14 +314,14 @@ union _CRT_ALIGN(16) u128
 		return fromV(_mm_cmpeq_epi8(left.vi, right.vi));
 	}

-	static __forceinline u128 gtu8(const u128& left, const u128& right)
+	static __forceinline u128 eq16(const u128& left, const u128& right)
 	{
-		return fromV(_mm_cmpgt_epu8(left.vi, right.vi));
+		return fromV(_mm_cmpeq_epi16(left.vi, right.vi));
 	}

-	static __forceinline u128 leu8(const u128& left, const u128& right)
+	static __forceinline u128 eq32(const u128& left, const u128& right)
 	{
-		return fromV(_mm_cmple_epu8(left.vi, right.vi));
+		return fromV(_mm_cmpeq_epi32(left.vi, right.vi));
 	}

 	bool operator == (const u128& right) const
@@ -287,6 +354,16 @@ union _CRT_ALIGN(16) u128
 		return from64(~_u64[0], ~_u64[1]);
 	}

+	__forceinline bool is_any_1() const // check if any bit is 1
+	{
+		return _u64[0] || _u64[1];
+	}
+
+	__forceinline bool is_any_0() const // check if any bit is 0
+	{
+		return ~_u64[0] || ~_u64[1];
+	}
+
 	// result = (~left) & (right)
 	static __forceinline u128 andnot(const u128& left, const u128& right)
 	{
@@ -342,15 +342,49 @@ static __forceinline uint64_t cntlz64(uint64_t arg)
 }

 // compare 16 packed unsigned bytes (greater than)
-static __forceinline __m128i _mm_cmpgt_epu8(__m128i A, __m128i B)
+inline __m128i sse_cmpgt_epu8(__m128i A, __m128i B)
 {
 	// (A xor 0x80) > (B xor 0x80)
-	return _mm_cmpgt_epi8(_mm_xor_si128(A, _mm_set1_epi8(-128)), _mm_xor_si128(B, _mm_set1_epi8(-128)));
+	const auto sign = _mm_set1_epi32(0x80808080);
+	return _mm_cmpgt_epi8(_mm_xor_si128(A, sign), _mm_xor_si128(B, sign));
 }

-// compare 16 packed unsigned bytes (less or equal)
-static __forceinline __m128i _mm_cmple_epu8(__m128i A, __m128i B)
+inline __m128i sse_cmpgt_epu16(__m128i A, __m128i B)
 {
-	// ((B xor 0x80) > (A xor 0x80)) || A == B
-	return _mm_or_si128(_mm_cmpgt_epu8(B, A), _mm_cmpeq_epi8(A, B));
+	const auto sign = _mm_set1_epi32(0x80008000);
+	return _mm_cmpgt_epi16(_mm_xor_si128(A, sign), _mm_xor_si128(B, sign));
+}
+
+inline __m128i sse_cmpgt_epu32(__m128i A, __m128i B)
+{
+	const auto sign = _mm_set1_epi32(0x80000000);
+	return _mm_cmpgt_epi32(_mm_xor_si128(A, sign), _mm_xor_si128(B, sign));
+}
+
+inline __m128 sse_exp2_ps(__m128 A)
+{
+	const auto x0 = _mm_max_ps(_mm_min_ps(A, _mm_set1_ps(127.4999961f)), _mm_set1_ps(-127.4999961f));
+	const auto x1 = _mm_add_ps(x0, _mm_set1_ps(0.5f));
+	const auto x2 = _mm_sub_epi32(_mm_cvtps_epi32(x1), _mm_and_si128(_mm_castps_si128(_mm_cmpnlt_ps(_mm_setzero_ps(), x1)), _mm_set1_epi32(1)));
+	const auto x3 = _mm_sub_ps(x0, _mm_cvtepi32_ps(x2));
+	const auto x4 = _mm_mul_ps(x3, x3);
+	const auto x5 = _mm_mul_ps(x3, _mm_add_ps(_mm_mul_ps(_mm_add_ps(_mm_mul_ps(x4, _mm_set1_ps(0.023093347705f)), _mm_set1_ps(20.20206567f)), x4), _mm_set1_ps(1513.906801f)));
+	const auto x6 = _mm_mul_ps(x5, _mm_rcp_ps(_mm_sub_ps(_mm_add_ps(_mm_mul_ps(_mm_set1_ps(233.1842117f), x4), _mm_set1_ps(4368.211667f)), x5)));
+	return _mm_mul_ps(_mm_add_ps(_mm_add_ps(x6, x6), _mm_set1_ps(1.0f)), _mm_castsi128_ps(_mm_slli_epi32(_mm_add_epi32(x2, _mm_set1_epi32(127)), 23)));
+}
+
+inline __m128 sse_log2_ps(__m128 A)
+{
+	const auto _1 = _mm_set1_ps(1.0f);
+	const auto _c = _mm_set1_ps(1.442695040f);
+	const auto x0 = _mm_max_ps(A, _mm_castsi128_ps(_mm_set1_epi32(0x00800000)));
+	const auto x1 = _mm_or_ps(_mm_and_ps(x0, _mm_castsi128_ps(_mm_set1_epi32(0x807fffff))), _1);
+	const auto x2 = _mm_rcp_ps(_mm_add_ps(x1, _1));
+	const auto x3 = _mm_mul_ps(_mm_sub_ps(x1, _1), x2);
+	const auto x4 = _mm_add_ps(x3, x3);
+	const auto x5 = _mm_mul_ps(x4, x4);
+	const auto x6 = _mm_add_ps(_mm_mul_ps(_mm_add_ps(_mm_mul_ps(_mm_set1_ps(-0.7895802789f), x5), _mm_set1_ps(16.38666457f)), x5), _mm_set1_ps(-64.1409953f));
+	const auto x7 = _mm_rcp_ps(_mm_add_ps(_mm_mul_ps(_mm_add_ps(_mm_mul_ps(_mm_set1_ps(-35.67227983f), x5), _mm_set1_ps(312.0937664f)), x5), _mm_set1_ps(-769.6919436f)));
+	const auto x8 = _mm_cvtepi32_ps(_mm_sub_epi32(_mm_srli_epi32(_mm_castps_si128(x0), 23), _mm_set1_epi32(127)));
+	return _mm_add_ps(_mm_mul_ps(_mm_mul_ps(_mm_mul_ps(_mm_mul_ps(x5, x6), x7), x4), _c), _mm_add_ps(_mm_mul_ps(x4, _c), x8));
 }
@@ -128,7 +128,7 @@ bool rRename(const std::string &from, const std::string &to)
 #ifdef _WIN32
 	if (!MoveFile(ConvertUTF8ToWString(from).c_str(), ConvertUTF8ToWString(to).c_str()))
 #else
-	if (rename(from.c_str(), to.c_str()))
+	if (int err = rename(from.c_str(), to.c_str()))
 #endif
 	{
 		LOG_ERROR(GENERAL, "Error renaming '%s' to '%s': 0x%llx", from.c_str(), to.c_str(), (u64)GET_API_ERROR);