Impaczus/rpcs3
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

PPU: some instructions replaced

Browse Source
This commit is contained in:
Nekotekina
2015-03-29 14:00:10 +03:00
parent d640aba903
commit 3247152ff7
4 changed files with 120 additions and 293 deletions
Download Patch File Download Diff File Expand all files Collapse all files
Utilities/BEType.h
+19 -2
View File
@@ -203,14 +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;
}
@@ -307,6 +314,16 @@ union _CRT_ALIGN(16) u128
return fromV(_mm_cmpeq_epi8(left.vi, right.vi));
}
static __forceinline u128 eq16(const u128& left, const u128& right)
{
return fromV(_mm_cmpeq_epi16(left.vi, right.vi));
}
static __forceinline u128 eq32(const u128& left, const u128& right)
{
return fromV(_mm_cmpeq_epi32(left.vi, right.vi));
}
bool operator == (const u128& right) const
{
return (_u64[0] == right._u64[0]) && (_u64[1] == right._u64[1]);
Utilities/GNU.h
+28
View File
@@ -360,3 +360,31 @@ 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));
}