nishi/milsko
1
Fork 0
mirror of https://gitea.nishi.boats/pyrite-dev/milsko synced 2025-12-31 06:30:52 +00:00
Code Issues Packages Projects Releases Wiki Activity

mmx: use unsigned ops and signed ops interchangably in places where it doesn't matter

Browse Source 
git-svn-id: http://svn2.nishi.boats/svn/milsko/trunk@566 b9cfdab3-6d41-4d17-bbe4-086880011989
This commit is contained in:
IoIxD
2025-11-01 17:44:09 +00:00
parent 3b26524cad
commit 95185ad0e0
1 changed files with 67 additions and 61 deletions
Download Patch File Download Diff File Expand all files Collapse all files

128
src/math/mmx.c
View File

@@ -10,96 +10,102 @@
struct _tyn* t = (struct _tyn*)&m; \
out->un._rty = *t;
static void mmx_add_u8(MwLLVec* a, MwLLVec* b, MwLLVec* out) {
DO_MMX_INTRINSIC(_m_paddusb, u8, u8, _MwLLVecDataU8x8);
static void mmx_add8(MwLLVec* a, MwLLVec* b, MwLLVec* out) {
DO_MMX_INTRINSIC(_m_paddb, u8, u8, _MwLLVecDataU8x8);
}
static void mmx_sub_u8(MwLLVec* a, MwLLVec* b, MwLLVec* out) {
DO_MMX_INTRINSIC(_m_psubusb, u8, u8, _MwLLVecDataU8x8);
static void mmx_sub8(MwLLVec* a, MwLLVec* b, MwLLVec* out) {
DO_MMX_INTRINSIC(_m_psubb, u8, u8, _MwLLVecDataU8x8);
}
static void mmx_equal_u8(MwLLVec* a, MwLLVec* b, MwLLVec* out) {
static void mmx_equal8(MwLLVec* a, MwLLVec* b, MwLLVec* out) {
DO_MMX_INTRINSIC(_m_pcmpeqb, u8, u8, _MwLLVecDataU8x8);
}
static void mmx_greaterThen_u8(MwLLVec* a, MwLLVec* b, MwLLVec* out) {
DO_MMX_INTRINSIC(_m_pcmpgtb, u8, u8, _MwLLVecDataU8x8);
static void mmx_add16(MwLLVec* a, MwLLVec* b, MwLLVec* out) {
DO_MMX_INTRINSIC(_m_paddw, u16, u16, _MwLLVecDataU16x4);
}
static void mmx_add_u16(MwLLVec* a, MwLLVec* b, MwLLVec* out) {
DO_MMX_INTRINSIC(_m_paddusw, u16, u16, _MwLLVecDataU16x4);
static void mmx_sub16(MwLLVec* a, MwLLVec* b, MwLLVec* out) {
DO_MMX_INTRINSIC(_m_psubw, u16, u16, _MwLLVecDataU16x4);
}
static void mmx_sub_u16(MwLLVec* a, MwLLVec* b, MwLLVec* out) {
DO_MMX_INTRINSIC(_m_psubusw, u16, u16, _MwLLVecDataU16x4);
}
static void mmx_shiftRight_u16(MwLLVec* a, MwLLVec* b, MwLLVec* out) {
static void mmx_shiftRight16(MwLLVec* a, MwLLVec* b, MwLLVec* out) {
DO_MMX_INTRINSIC(_m_psrlw, u16, u16, _MwLLVecDataU16x4);
}
static void mmx_shiftLeft_u16(MwLLVec* a, MwLLVec* b, MwLLVec* out) {
static void mmx_shiftLeft16(MwLLVec* a, MwLLVec* b, MwLLVec* out) {
DO_MMX_INTRINSIC(_m_psllw, u16, u16, _MwLLVecDataU16x4);
}
static void mmx_equal_u16(MwLLVec* a, MwLLVec* b, MwLLVec* out) {
static void mmx_equal16(MwLLVec* a, MwLLVec* b, MwLLVec* out) {
DO_MMX_INTRINSIC(_m_pcmpeqw, u16, u16, _MwLLVecDataU16x4);
}
static void mmx_greaterThen_u16(MwLLVec* a, MwLLVec* b, MwLLVec* out) {
DO_MMX_INTRINSIC(_m_pcmpgtw, u16, u16, _MwLLVecDataU16x4);
}
static void mmx_add_u32(MwLLVec* a, MwLLVec* b, MwLLVec* out) {
static void mmx_add32(MwLLVec* a, MwLLVec* b, MwLLVec* out) {
DO_MMX_INTRINSIC(_m_paddd, u32, u32, _MwLLVecDataU32x2);
}
static void mmx_sub_u32(MwLLVec* a, MwLLVec* b, MwLLVec* out) {
static void mmx_sub32(MwLLVec* a, MwLLVec* b, MwLLVec* out) {
DO_MMX_INTRINSIC(_m_psubd, u32, u32, _MwLLVecDataU32x2);
}
static void mmx_shiftRight_u32(MwLLVec* a, MwLLVec* b, MwLLVec* out) {
static void mmx_shiftRight32(MwLLVec* a, MwLLVec* b, MwLLVec* out) {
DO_MMX_INTRINSIC(_m_psrld, u32, u32, _MwLLVecDataU32x2);
}
static void mmx_shiftLeft_u32(MwLLVec* a, MwLLVec* b, MwLLVec* out) {
static void mmx_shiftLeft32(MwLLVec* a, MwLLVec* b, MwLLVec* out) {
DO_MMX_INTRINSIC(_m_pslld, u32, u32, _MwLLVecDataU32x2);
}
static void mmx_equal_u32(MwLLVec* a, MwLLVec* b, MwLLVec* out) {
static void mmx_equal32(MwLLVec* a, MwLLVec* b, MwLLVec* out) {
DO_MMX_INTRINSIC(_m_pcmpeqw, u32, u32, _MwLLVecDataU32x2);
}
static void mmx_greaterThen_u32(MwLLVec* a, MwLLVec* b, MwLLVec* out) {
DO_MMX_INTRINSIC(_m_pcmpgtw, u32, u32, _MwLLVecDataU32x2);
}
static void mmx_add_i8(MwLLVec* a, MwLLVec* b, MwLLVec* out) {
DO_MMX_INTRINSIC(_m_paddsb, i8, i8, _MwLLVecDataI8x8);
};
static void mmx_sub_i8(MwLLVec* a, MwLLVec* b, MwLLVec* out) {
DO_MMX_INTRINSIC(_m_psubsb, i8, i8, _MwLLVecDataI8x8);
static void mmx_addi8(MwLLVec* a, MwLLVec* b, MwLLVec* out) {
DO_MMX_INTRINSIC(_m_paddsb, u8, u8, _MwLLVecDataU8x8);
}
static void mmx_add_i16(MwLLVec* a, MwLLVec* b, MwLLVec* out) {
DO_MMX_INTRINSIC(_m_paddsw, i16, i16, _MwLLVecDataI16x4);
static void mmx_subi8(MwLLVec* a, MwLLVec* b, MwLLVec* out) {
DO_MMX_INTRINSIC(_m_psubsb, u8, u8, _MwLLVecDataU8x8);
}
static void mmx_sub_i16(MwLLVec* a, MwLLVec* b, MwLLVec* out) {
DO_MMX_INTRINSIC(_m_psubsw, i16, i16, _MwLLVecDataI16x4);
static void mmx_addi16(MwLLVec* a, MwLLVec* b, MwLLVec* out) {
DO_MMX_INTRINSIC(_m_paddsw, u16, u16, _MwLLVecDataU16x4);
}
static void mmx_subi16(MwLLVec* a, MwLLVec* b, MwLLVec* out) {
DO_MMX_INTRINSIC(_m_psubsw, u16, u16, _MwLLVecDataU16x4);
}
static void mmx_greaterTheni8(MwLLVec* a, MwLLVec* b, MwLLVec* out) {
DO_MMX_INTRINSIC(_m_pcmpgtb, u8, u8, _MwLLVecDataU8x8);
}
static void mmx_greaterTheni16(MwLLVec* a, MwLLVec* b, MwLLVec* out) {
DO_MMX_INTRINSIC(_m_pcmpgtw, u16, u16, _MwLLVecDataU16x4);
}
static void mmx_greaterTheni32(MwLLVec* a, MwLLVec* b, MwLLVec* out) {
DO_MMX_INTRINSIC(_m_pcmpgtd, u32, u32, _MwLLVecDataU32x2);
}
void mmx_apply(MwLLMathVTable** t) {
t[_MwLLVecTypeU8x8]->Add = mmx_add_u8;
t[_MwLLVecTypeU8x8]->Sub = mmx_sub_u8;
t[_MwLLVecTypeU8x8]->GreaterThen = mmx_greaterThen_u8;
t[_MwLLVecTypeU8x8]->Equal = mmx_equal_u8;
t[_MwLLVecTypeU16x4]->Add = mmx_add_u16;
t[_MwLLVecTypeU16x4]->Sub = mmx_sub_u16;
t[_MwLLVecTypeU16x4]->ShiftLeft = mmx_shiftLeft_u16;
t[_MwLLVecTypeU16x4]->ShiftRight = mmx_shiftRight_u16;
t[_MwLLVecTypeU16x4]->GreaterThen = mmx_greaterThen_u16;
t[_MwLLVecTypeU16x4]->Equal = mmx_equal_u16;
t[_MwLLVecTypeU32x2]->Add = mmx_add_u32;
t[_MwLLVecTypeU32x2]->Sub = mmx_sub_u32;
t[_MwLLVecTypeU32x2]->ShiftLeft = mmx_shiftLeft_u32;
t[_MwLLVecTypeU32x2]->ShiftRight = mmx_shiftRight_u32;
t[_MwLLVecTypeU32x2]->GreaterThen = mmx_greaterThen_u32;
t[_MwLLVecTypeU32x2]->Equal = mmx_equal_u32;
t[_MwLLVecTypeU8x8]->Add = mmx_add8;
t[_MwLLVecTypeU8x8]->Sub = mmx_sub8;
// t[_MwLLVecTypeU8x8]->GreaterThen = mmx_greaterThen8;
t[_MwLLVecTypeU8x8]->Equal = mmx_equal8;
t[_MwLLVecTypeU16x4]->Add = mmx_add16;
t[_MwLLVecTypeU16x4]->Sub = mmx_sub16;
t[_MwLLVecTypeU16x4]->ShiftLeft = mmx_shiftLeft16;
t[_MwLLVecTypeU16x4]->ShiftRight = mmx_shiftRight16;
// t[_MwLLVecTypeU16x4]->GreaterThen = mmx_greaterThen16;
t[_MwLLVecTypeU16x4]->Equal = mmx_equal16;
t[_MwLLVecTypeU32x2]->Add = mmx_add32;
t[_MwLLVecTypeU32x2]->Sub = mmx_sub32;
t[_MwLLVecTypeU32x2]->ShiftLeft = mmx_shiftLeft32;
t[_MwLLVecTypeU32x2]->ShiftRight = mmx_shiftRight32;
// t[_MwLLVecTypeU32x2]->GreaterThen = mmx_greaterThen32;
t[_MwLLVecTypeU32x2]->Equal = mmx_equal32;
t[_MwLLVecTypeI8x8]->Add = mmx_add_i8;
t[_MwLLVecTypeI8x8]->Sub = mmx_sub_i8;
t[_MwLLVecTypeI16x4]->Add = mmx_add_i16;
t[_MwLLVecTypeI16x4]->Sub = mmx_sub_i16;
t[_MwLLVecTypeI16x4]->ShiftLeft = mmx_shiftLeft_u16;
t[_MwLLVecTypeI16x4]->ShiftRight = mmx_shiftRight_u16;
t[_MwLLVecTypeI32x2]->ShiftLeft = mmx_shiftLeft_u32;
t[_MwLLVecTypeI32x2]->ShiftRight = mmx_shiftRight_u32;
t[_MwLLVecTypeU8x8]->Add = mmx_addi8;
t[_MwLLVecTypeU8x8]->Sub = mmx_subi8;
t[_MwLLVecTypeI8x8]->GreaterThen = mmx_greaterTheni8;
t[_MwLLVecTypeI8x8]->Equal = mmx_equal8;
t[_MwLLVecTypeU16x4]->Add = mmx_addi16;
t[_MwLLVecTypeU16x4]->Sub = mmx_subi16;
t[_MwLLVecTypeI16x4]->ShiftLeft = mmx_shiftLeft16;
t[_MwLLVecTypeI16x4]->ShiftRight = mmx_shiftRight16;
t[_MwLLVecTypeI16x4]->GreaterThen = mmx_greaterTheni16;
t[_MwLLVecTypeI16x4]->Equal = mmx_equal16;
t[_MwLLVecTypeI32x2]->Add = mmx_add32;
t[_MwLLVecTypeI32x2]->Sub = mmx_sub32;
t[_MwLLVecTypeI32x2]->ShiftLeft = mmx_shiftLeft32;
t[_MwLLVecTypeI32x2]->ShiftRight = mmx_shiftRight32;
t[_MwLLVecTypeI32x2]->GreaterThen = mmx_greaterTheni32;
t[_MwLLVecTypeI32x2]->Equal = mmx_equal32;
}
#endif