ShuriZma
/
dolphin
mirror of https://github.com/dolphin-emu/dolphin.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

JIT: Various JitAsmCommon optimizations 

Use some SSE4 instructions in on CPUs that support them.
Use float instructions instead of int where appropriate (it's a cycle faster
on CPUs with arithmetic unit forwarding penalties).
This commit is contained in:
Fiora 2014-07-26 18:46:09 -07:00
parent 7b0f559ae1
commit 7ab820c6f8
2 changed files with 163 additions and 111 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

32
Source/Core/Core/PowerPC/Jit64/Jit_LoadStorePaired.cpp
View File

@ -22,9 +22,8 @@ void Jit64::psq_st(UGeckoInstruction inst)
JITDISABLE(bJITLoadStorePairedOff); JITDISABLE(bJITLoadStorePairedOff);
FALLBACK_IF(!inst.RA); FALLBACK_IF(!inst.RA);
bool update = inst.OPCD == 61; s32 offset = inst.SIMM_12;
bool update = inst.OPCD == 61 && offset;
int offset = inst.SIMM_12;
int a = inst.RA; int a = inst.RA;
int s = inst.RS; int s = inst.RS;
@ -32,9 +31,16 @@ void Jit64::psq_st(UGeckoInstruction inst)
if (update) if (update)
gpr.BindToRegister(a, true, true); gpr.BindToRegister(a, true, true);
fpr.BindToRegister(s, true, false); fpr.BindToRegister(s, true, false);
MOV(32, R(RSCRATCH_EXTRA), gpr.R(a)); if (offset && gpr.R(a).IsSimpleReg())
if (offset) {
ADD(32, R(RSCRATCH_EXTRA), Imm32((u32)offset)); LEA(32, RSCRATCH_EXTRA, MDisp(gpr.RX(a), offset));
}
else
{
MOV(32, R(RSCRATCH_EXTRA), gpr.R(a));
if (offset)
ADD(32, R(RSCRATCH_EXTRA), Imm32((u32)offset));
}
// In memcheck mode, don't update the address until the exception check // In memcheck mode, don't update the address until the exception check
if (update && offset && !js.memcheck) if (update && offset && !js.memcheck)
MOV(32, gpr.R(a), R(RSCRATCH_EXTRA)); MOV(32, gpr.R(a), R(RSCRATCH_EXTRA));
@ -46,7 +52,7 @@ void Jit64::psq_st(UGeckoInstruction inst)
AND(32, R(RSCRATCH2), PPCSTATE(spr[SPR_GQR0 + inst.I])); AND(32, R(RSCRATCH2), PPCSTATE(spr[SPR_GQR0 + inst.I]));
MOVZX(32, 8, RSCRATCH, R(RSCRATCH2)); MOVZX(32, 8, RSCRATCH, R(RSCRATCH2));
// FIXME: Fix ModR/M encoding to allow [RSCRATCH2*4+disp32] without a base register! // FIXME: Fix ModR/M encoding to allow [RSCRATCH2*8+disp32] without a base register!
if (inst.W) if (inst.W)
{ {
// One value // One value
@ -77,18 +83,24 @@ void Jit64::psq_l(UGeckoInstruction inst)
JITDISABLE(bJITLoadStorePairedOff); JITDISABLE(bJITLoadStorePairedOff);
FALLBACK_IF(!inst.RA); FALLBACK_IF(!inst.RA);
bool update = inst.OPCD == 57; s32 offset = inst.SIMM_12;
int offset = inst.SIMM_12; bool update = inst.OPCD == 57 && offset;
int a = inst.RA; int a = inst.RA;
int s = inst.RS; int s = inst.RS;
gpr.FlushLockX(RSCRATCH_EXTRA); gpr.FlushLockX(RSCRATCH_EXTRA);
gpr.BindToRegister(a, true, update && offset); gpr.BindToRegister(a, true, update && offset);
fpr.BindToRegister(s, false, true); fpr.BindToRegister(s, false, true);
if (offset) if (offset && gpr.R(a).IsSimpleReg())
{
LEA(32, RSCRATCH_EXTRA, MDisp(gpr.RX(a), offset)); LEA(32, RSCRATCH_EXTRA, MDisp(gpr.RX(a), offset));
}
else else
{
MOV(32, R(RSCRATCH_EXTRA), gpr.R(a)); MOV(32, R(RSCRATCH_EXTRA), gpr.R(a));
if (offset)
ADD(32, R(RSCRATCH_EXTRA), Imm32((u32)offset));
}
// In memcheck mode, don't update the address until the exception check // In memcheck mode, don't update the address until the exception check
if (update && offset && !js.memcheck) if (update && offset && !js.memcheck)
MOV(32, gpr.R(a), R(RSCRATCH_EXTRA)); MOV(32, gpr.R(a), R(RSCRATCH_EXTRA));

242
Source/Core/Core/PowerPC/JitCommon/JitAsmCommon.cpp
View File

@ -184,47 +184,63 @@ static const u8 GC_ALIGNED16(pbswapShuffle2x4[16]) = {3, 2, 1, 0, 7, 6, 5, 4, 8,
static const float GC_ALIGNED16(m_quantizeTableS[]) = static const float GC_ALIGNED16(m_quantizeTableS[]) =
{ {
(1ULL << 0), (1ULL << 1), (1ULL << 2), (1ULL << 3), (1ULL << 0), (1ULL << 0), (1ULL << 1), (1ULL << 1), (1ULL << 2), (1ULL << 2), (1ULL << 3), (1ULL << 3),
(1ULL << 4), (1ULL << 5), (1ULL << 6), (1ULL << 7), (1ULL << 4), (1ULL << 4), (1ULL << 5), (1ULL << 5), (1ULL << 6), (1ULL << 6), (1ULL << 7), (1ULL << 7),
(1ULL << 8), (1ULL << 9), (1ULL << 10), (1ULL << 11), (1ULL << 8), (1ULL << 8), (1ULL << 9), (1ULL << 9), (1ULL << 10), (1ULL << 10), (1ULL << 11), (1ULL << 11),
(1ULL << 12), (1ULL << 13), (1ULL << 14), (1ULL << 15), (1ULL << 12), (1ULL << 12), (1ULL << 13), (1ULL << 13), (1ULL << 14), (1ULL << 14), (1ULL << 15), (1ULL << 15),
(1ULL << 16), (1ULL << 17), (1ULL << 18), (1ULL << 19), (1ULL << 16), (1ULL << 16), (1ULL << 17), (1ULL << 17), (1ULL << 18), (1ULL << 18), (1ULL << 19), (1ULL << 19),
(1ULL << 20), (1ULL << 21), (1ULL << 22), (1ULL << 23), (1ULL << 20), (1ULL << 20), (1ULL << 21), (1ULL << 21), (1ULL << 22), (1ULL << 22), (1ULL << 23), (1ULL << 23),
(1ULL << 24), (1ULL << 25), (1ULL << 26), (1ULL << 27), (1ULL << 24), (1ULL << 24), (1ULL << 25), (1ULL << 25), (1ULL << 26), (1ULL << 26), (1ULL << 27), (1ULL << 27),
(1ULL << 28), (1ULL << 29), (1ULL << 30), (1ULL << 31), (1ULL << 28), (1ULL << 28), (1ULL << 29), (1ULL << 29), (1ULL << 30), (1ULL << 30), (1ULL << 31), (1ULL << 31),
1.0 / (1ULL << 32), 1.0 / (1ULL << 31), 1.0 / (1ULL << 30), 1.0 / (1ULL << 29), 1.0 / (1ULL << 32), 1.0 / (1ULL << 32), 1.0 / (1ULL << 31), 1.0 / (1ULL << 31),
1.0 / (1ULL << 28), 1.0 / (1ULL << 27), 1.0 / (1ULL << 26), 1.0 / (1ULL << 25), 1.0 / (1ULL << 30), 1.0 / (1ULL << 30), 1.0 / (1ULL << 29), 1.0 / (1ULL << 29),
1.0 / (1ULL << 24), 1.0 / (1ULL << 23), 1.0 / (1ULL << 22), 1.0 / (1ULL << 21), 1.0 / (1ULL << 28), 1.0 / (1ULL << 28), 1.0 / (1ULL << 27), 1.0 / (1ULL << 27),
1.0 / (1ULL << 20), 1.0 / (1ULL << 19), 1.0 / (1ULL << 18), 1.0 / (1ULL << 17), 1.0 / (1ULL << 26), 1.0 / (1ULL << 26), 1.0 / (1ULL << 25), 1.0 / (1ULL << 25),
1.0 / (1ULL << 16), 1.0 / (1ULL << 15), 1.0 / (1ULL << 14), 1.0 / (1ULL << 13), 1.0 / (1ULL << 24), 1.0 / (1ULL << 24), 1.0 / (1ULL << 23), 1.0 / (1ULL << 23),
1.0 / (1ULL << 12), 1.0 / (1ULL << 11), 1.0 / (1ULL << 10), 1.0 / (1ULL << 9), 1.0 / (1ULL << 22), 1.0 / (1ULL << 22), 1.0 / (1ULL << 21), 1.0 / (1ULL << 21),
1.0 / (1ULL << 8), 1.0 / (1ULL << 7), 1.0 / (1ULL << 6), 1.0 / (1ULL << 5), 1.0 / (1ULL << 20), 1.0 / (1ULL << 20), 1.0 / (1ULL << 19), 1.0 / (1ULL << 19),
1.0 / (1ULL << 4), 1.0 / (1ULL << 3), 1.0 / (1ULL << 2), 1.0 / (1ULL << 1), 1.0 / (1ULL << 18), 1.0 / (1ULL << 18), 1.0 / (1ULL << 17), 1.0 / (1ULL << 17),
1.0 / (1ULL << 16), 1.0 / (1ULL << 16), 1.0 / (1ULL << 15), 1.0 / (1ULL << 15),
1.0 / (1ULL << 14), 1.0 / (1ULL << 14), 1.0 / (1ULL << 13), 1.0 / (1ULL << 13),
1.0 / (1ULL << 12), 1.0 / (1ULL << 12), 1.0 / (1ULL << 11), 1.0 / (1ULL << 11),
1.0 / (1ULL << 10), 1.0 / (1ULL << 10), 1.0 / (1ULL << 9), 1.0 / (1ULL << 9),
1.0 / (1ULL << 8), 1.0 / (1ULL << 8), 1.0 / (1ULL << 7), 1.0 / (1ULL << 7),
1.0 / (1ULL << 6), 1.0 / (1ULL << 6), 1.0 / (1ULL << 5), 1.0 / (1ULL << 5),
1.0 / (1ULL << 4), 1.0 / (1ULL << 4), 1.0 / (1ULL << 3), 1.0 / (1ULL << 3),
1.0 / (1ULL << 2), 1.0 / (1ULL << 2), 1.0 / (1ULL << 1), 1.0 / (1ULL << 1),
}; };
static const float GC_ALIGNED16(m_dequantizeTableS[]) = static const float GC_ALIGNED16(m_dequantizeTableS[]) =
{ {
1.0 / (1ULL << 0), 1.0 / (1ULL << 1), 1.0 / (1ULL << 2), 1.0 / (1ULL << 3), 1.0 / (1ULL << 0), 1.0 / (1ULL << 0), 1.0 / (1ULL << 1), 1.0 / (1ULL << 1),
1.0 / (1ULL << 4), 1.0 / (1ULL << 5), 1.0 / (1ULL << 6), 1.0 / (1ULL << 7), 1.0 / (1ULL << 2), 1.0 / (1ULL << 2), 1.0 / (1ULL << 3), 1.0 / (1ULL << 3),
1.0 / (1ULL << 8), 1.0 / (1ULL << 9), 1.0 / (1ULL << 10), 1.0 / (1ULL << 11), 1.0 / (1ULL << 4), 1.0 / (1ULL << 4), 1.0 / (1ULL << 5), 1.0 / (1ULL << 5),
1.0 / (1ULL << 12), 1.0 / (1ULL << 13), 1.0 / (1ULL << 14), 1.0 / (1ULL << 15), 1.0 / (1ULL << 6), 1.0 / (1ULL << 6), 1.0 / (1ULL << 7), 1.0 / (1ULL << 7),
1.0 / (1ULL << 16), 1.0 / (1ULL << 17), 1.0 / (1ULL << 18), 1.0 / (1ULL << 19), 1.0 / (1ULL << 8), 1.0 / (1ULL << 8), 1.0 / (1ULL << 9), 1.0 / (1ULL << 9),
1.0 / (1ULL << 20), 1.0 / (1ULL << 21), 1.0 / (1ULL << 22), 1.0 / (1ULL << 23), 1.0 / (1ULL << 10), 1.0 / (1ULL << 10), 1.0 / (1ULL << 11), 1.0 / (1ULL << 11),
1.0 / (1ULL << 24), 1.0 / (1ULL << 25), 1.0 / (1ULL << 26), 1.0 / (1ULL << 27), 1.0 / (1ULL << 12), 1.0 / (1ULL << 12), 1.0 / (1ULL << 13), 1.0 / (1ULL << 13),
1.0 / (1ULL << 28), 1.0 / (1ULL << 29), 1.0 / (1ULL << 30), 1.0 / (1ULL << 31), 1.0 / (1ULL << 14), 1.0 / (1ULL << 14), 1.0 / (1ULL << 15), 1.0 / (1ULL << 15),
(1ULL << 32), (1ULL << 31), (1ULL << 30), (1ULL << 29), 1.0 / (1ULL << 16), 1.0 / (1ULL << 16), 1.0 / (1ULL << 17), 1.0 / (1ULL << 17),
(1ULL << 28), (1ULL << 27), (1ULL << 26), (1ULL << 25), 1.0 / (1ULL << 18), 1.0 / (1ULL << 18), 1.0 / (1ULL << 19), 1.0 / (1ULL << 19),
(1ULL << 24), (1ULL << 23), (1ULL << 22), (1ULL << 21), 1.0 / (1ULL << 20), 1.0 / (1ULL << 20), 1.0 / (1ULL << 21), 1.0 / (1ULL << 21),
(1ULL << 20), (1ULL << 19), (1ULL << 18), (1ULL << 17), 1.0 / (1ULL << 22), 1.0 / (1ULL << 22), 1.0 / (1ULL << 23), 1.0 / (1ULL << 23),
(1ULL << 16), (1ULL << 15), (1ULL << 14), (1ULL << 13), 1.0 / (1ULL << 24), 1.0 / (1ULL << 24), 1.0 / (1ULL << 25), 1.0 / (1ULL << 25),
(1ULL << 12), (1ULL << 11), (1ULL << 10), (1ULL << 9), 1.0 / (1ULL << 26), 1.0 / (1ULL << 26), 1.0 / (1ULL << 27), 1.0 / (1ULL << 27),
(1ULL << 8), (1ULL << 7), (1ULL << 6), (1ULL << 5), 1.0 / (1ULL << 28), 1.0 / (1ULL << 28), 1.0 / (1ULL << 29), 1.0 / (1ULL << 29),
(1ULL << 4), (1ULL << 3), (1ULL << 2), (1ULL << 1), 1.0 / (1ULL << 30), 1.0 / (1ULL << 30), 1.0 / (1ULL << 31), 1.0 / (1ULL << 31),
(1ULL << 32), (1ULL << 32), (1ULL << 31), (1ULL << 31), (1ULL << 30), (1ULL << 30), (1ULL << 29), (1ULL << 29),
(1ULL << 28), (1ULL << 28), (1ULL << 27), (1ULL << 27), (1ULL << 26), (1ULL << 26), (1ULL << 25), (1ULL << 25),
(1ULL << 24), (1ULL << 24), (1ULL << 23), (1ULL << 23), (1ULL << 22), (1ULL << 22), (1ULL << 21), (1ULL << 21),
(1ULL << 20), (1ULL << 20), (1ULL << 19), (1ULL << 19), (1ULL << 18), (1ULL << 18), (1ULL << 17), (1ULL << 17),
(1ULL << 16), (1ULL << 16), (1ULL << 15), (1ULL << 15), (1ULL << 14), (1ULL << 14), (1ULL << 13), (1ULL << 13),
(1ULL << 12), (1ULL << 12), (1ULL << 11), (1ULL << 11), (1ULL << 10), (1ULL << 10), (1ULL << 9), (1ULL << 9),
(1ULL << 8), (1ULL << 8), (1ULL << 7), (1ULL << 7), (1ULL << 6), (1ULL << 6), (1ULL << 5), (1ULL << 5),
(1ULL << 4), (1ULL << 4), (1ULL << 3), (1ULL << 3), (1ULL << 2), (1ULL << 2), (1ULL << 1), (1ULL << 1),
}; };
static float GC_ALIGNED16(psTemp[4]); static float GC_ALIGNED16(psTemp[4]);
static const float GC_ALIGNED16(m_65535) = 65535.0f; static const float GC_ALIGNED16(m_65535[4]) = {65535.0f, 65535.0f, 65535.0f, 65535.0f};
static const float GC_ALIGNED16(m_32767) = 32767.0f; static const float GC_ALIGNED16(m_32767) = 32767.0f;
static const float GC_ALIGNED16(m_m32768) = -32768.0f; static const float GC_ALIGNED16(m_m32768) = -32768.0f;
static const float GC_ALIGNED16(m_255) = 255.0f; static const float GC_ALIGNED16(m_255) = 255.0f;
@ -273,14 +289,11 @@ void CommonAsmRoutines::GenQuantizedStores()
RET(); RET();
const u8* storePairedU8 = AlignCode4(); const u8* storePairedU8 = AlignCode4();
SHR(32, R(RSCRATCH2), Imm8(6)); SHR(32, R(RSCRATCH2), Imm8(5));
MOVSS(XMM1, MDisp(RSCRATCH2, (u32)(u64)m_quantizeTableS)); MOVQ_xmm(XMM1, MDisp(RSCRATCH2, (u32)(u64)m_quantizeTableS));
PUNPCKLDQ(XMM1, R(XMM1));
MULPS(XMM0, R(XMM1)); MULPS(XMM0, R(XMM1));
#ifdef QUANTIZE_OVERFLOW_SAFE #ifdef QUANTIZE_OVERFLOW_SAFE
MOVSS(XMM1, M((void *)&m_65535)); MINPS(XMM0, M((void *)&m_65535));
PUNPCKLDQ(XMM1, R(XMM1));
MINPS(XMM0, R(XMM1));
#endif #endif
CVTTPS2DQ(XMM0, R(XMM0)); CVTTPS2DQ(XMM0, R(XMM0));
PACKSSDW(XMM0, R(XMM0)); PACKSSDW(XMM0, R(XMM0));
@ -291,14 +304,11 @@ void CommonAsmRoutines::GenQuantizedStores()
RET(); RET();
const u8* storePairedS8 = AlignCode4(); const u8* storePairedS8 = AlignCode4();
SHR(32, R(RSCRATCH2), Imm8(6)); SHR(32, R(RSCRATCH2), Imm8(5));
MOVSS(XMM1, MDisp(RSCRATCH2, (u32)(u64)m_quantizeTableS)); MOVQ_xmm(XMM1, MDisp(RSCRATCH2, (u32)(u64)m_quantizeTableS));
PUNPCKLDQ(XMM1, R(XMM1));
MULPS(XMM0, R(XMM1)); MULPS(XMM0, R(XMM1));
#ifdef QUANTIZE_OVERFLOW_SAFE #ifdef QUANTIZE_OVERFLOW_SAFE
MOVSS(XMM1, M((void *)&m_65535)); MINPS(XMM0, M((void *)&m_65535));
PUNPCKLDQ(XMM1, R(XMM1));
MINPS(XMM0, R(XMM1));
#endif #endif
CVTTPS2DQ(XMM0, R(XMM0)); CVTTPS2DQ(XMM0, R(XMM0));
PACKSSDW(XMM0, R(XMM0)); PACKSSDW(XMM0, R(XMM0));
@ -310,41 +320,47 @@ void CommonAsmRoutines::GenQuantizedStores()
RET(); RET();
const u8* storePairedU16 = AlignCode4(); const u8* storePairedU16 = AlignCode4();
SHR(32, R(RSCRATCH2), Imm8(6)); SHR(32, R(RSCRATCH2), Imm8(5));
MOVSS(XMM1, MDisp(RSCRATCH2, (u32)(u64)m_quantizeTableS)); MOVQ_xmm(XMM1, MDisp(RSCRATCH2, (u32)(u64)m_quantizeTableS));
PUNPCKLDQ(XMM1, R(XMM1));
MULPS(XMM0, R(XMM1)); MULPS(XMM0, R(XMM1));
// PACKUSDW is available only in SSE4 if (cpu_info.bSSE4_1)
PXOR(XMM1, R(XMM1)); {
MAXPS(XMM0, R(XMM1)); #ifdef QUANTIZE_OVERFLOW_SAFE
MOVSS(XMM1, M((void *)&m_65535)); MINPS(XMM0, M((void *)&m_65535));
PUNPCKLDQ(XMM1, R(XMM1)); #endif
MINPS(XMM0, R(XMM1)); CVTTPS2DQ(XMM0, R(XMM0));
PACKUSDW(XMM0, R(XMM0));
MOVD_xmm(R(RSCRATCH), XMM0);
BSWAP(32, RSCRATCH);
ROL(32, R(RSCRATCH), Imm8(16));
}
else
{
XORPS(XMM1, R(XMM1));
MAXPS(XMM0, R(XMM1));
MINPS(XMM0, M((void *)&m_65535));
CVTTPS2DQ(XMM0, R(XMM0)); CVTTPS2DQ(XMM0, R(XMM0));
MOVQ_xmm(M(psTemp), XMM0); MOVQ_xmm(M(psTemp), XMM0);
// place ps[0] into the higher word, ps[1] into the lower // place ps[0] into the higher word, ps[1] into the lower
// so no need in ROL after BSWAP // so no need in ROL after BSWAP
MOVZX(32, 16, RSCRATCH, M((char*)psTemp + 0)); MOVZX(32, 16, RSCRATCH, M((char*)psTemp + 0));
SHL(32, R(RSCRATCH), Imm8(16)); SHL(32, R(RSCRATCH), Imm8(16));
MOV(16, R(RSCRATCH), M((char*)psTemp + 4)); MOV(16, R(RSCRATCH), M((char*)psTemp + 4));
BSWAP(32, RSCRATCH);
}
BSWAP(32, RSCRATCH);
SafeWriteRegToReg(RSCRATCH, RSCRATCH_EXTRA, 32, 0, QUANTIZED_REGS_TO_SAVE, SAFE_LOADSTORE_NO_SWAP | SAFE_LOADSTORE_NO_PROLOG | SAFE_LOADSTORE_NO_FASTMEM); SafeWriteRegToReg(RSCRATCH, RSCRATCH_EXTRA, 32, 0, QUANTIZED_REGS_TO_SAVE, SAFE_LOADSTORE_NO_SWAP | SAFE_LOADSTORE_NO_PROLOG | SAFE_LOADSTORE_NO_FASTMEM);
RET(); RET();
const u8* storePairedS16 = AlignCode4(); const u8* storePairedS16 = AlignCode4();
SHR(32, R(RSCRATCH2), Imm8(6)); SHR(32, R(RSCRATCH2), Imm8(5));
MOVSS(XMM1, MDisp(RSCRATCH2, (u32)(u64)m_quantizeTableS)); MOVQ_xmm(XMM1, MDisp(RSCRATCH2, (u32)(u64)m_quantizeTableS));
// SHUFPS or UNPCKLPS might be a better choice here. The last one might just be an alias though.
PUNPCKLDQ(XMM1, R(XMM1));
MULPS(XMM0, R(XMM1)); MULPS(XMM0, R(XMM1));
#ifdef QUANTIZE_OVERFLOW_SAFE #ifdef QUANTIZE_OVERFLOW_SAFE
MOVSS(XMM1, M((void *)&m_65535)); MINPS(XMM0, M((void *)&m_65535));
PUNPCKLDQ(XMM1, R(XMM1));
MINPS(XMM0, R(XMM1));
#endif #endif
CVTTPS2DQ(XMM0, R(XMM0)); CVTTPS2DQ(XMM0, R(XMM0));
PACKSSDW(XMM0, R(XMM0)); PACKSSDW(XMM0, R(XMM0));
@ -395,10 +411,10 @@ void CommonAsmRoutines::GenQuantizedSingleStores()
}*/ }*/
const u8* storeSingleU8 = AlignCode4(); // Used by MKWii const u8* storeSingleU8 = AlignCode4(); // Used by MKWii
SHR(32, R(RSCRATCH2), Imm8(6)); SHR(32, R(RSCRATCH2), Imm8(5));
MOVSS(XMM1, MDisp(RSCRATCH2, (u32)(u64)m_quantizeTableS)); MOVSS(XMM1, MDisp(RSCRATCH2, (u32)(u64)m_quantizeTableS));
MULSS(XMM0, R(XMM1)); MULSS(XMM0, R(XMM1));
PXOR(XMM1, R(XMM1)); XORPS(XMM1, R(XMM1));
MAXSS(XMM0, R(XMM1)); MAXSS(XMM0, R(XMM1));
MINSS(XMM0, M((void *)&m_255)); MINSS(XMM0, M((void *)&m_255));
CVTTSS2SI(RSCRATCH, R(XMM0)); CVTTSS2SI(RSCRATCH, R(XMM0));
@ -406,7 +422,7 @@ void CommonAsmRoutines::GenQuantizedSingleStores()
RET(); RET();
const u8* storeSingleS8 = AlignCode4(); const u8* storeSingleS8 = AlignCode4();
SHR(32, R(RSCRATCH2), Imm8(6)); SHR(32, R(RSCRATCH2), Imm8(5));
MOVSS(XMM1, MDisp(RSCRATCH2, (u32)(u64)m_quantizeTableS)); MOVSS(XMM1, MDisp(RSCRATCH2, (u32)(u64)m_quantizeTableS));
MULSS(XMM0, R(XMM1)); MULSS(XMM0, R(XMM1));
MAXSS(XMM0, M((void *)&m_m128)); MAXSS(XMM0, M((void *)&m_m128));
@ -416,10 +432,10 @@ void CommonAsmRoutines::GenQuantizedSingleStores()
RET(); RET();
const u8* storeSingleU16 = AlignCode4(); // Used by MKWii const u8* storeSingleU16 = AlignCode4(); // Used by MKWii
SHR(32, R(RSCRATCH2), Imm8(6)); SHR(32, R(RSCRATCH2), Imm8(5));
MOVSS(XMM1, MDisp(RSCRATCH2, (u32)(u64)m_quantizeTableS)); MOVSS(XMM1, MDisp(RSCRATCH2, (u32)(u64)m_quantizeTableS));
MULSS(XMM0, R(XMM1)); MULSS(XMM0, R(XMM1));
PXOR(XMM1, R(XMM1)); XORPS(XMM1, R(XMM1));
MAXSS(XMM0, R(XMM1)); MAXSS(XMM0, R(XMM1));
MINSS(XMM0, M((void *)&m_65535)); MINSS(XMM0, M((void *)&m_65535));
CVTTSS2SI(RSCRATCH, R(XMM0)); CVTTSS2SI(RSCRATCH, R(XMM0));
@ -427,7 +443,7 @@ void CommonAsmRoutines::GenQuantizedSingleStores()
RET(); RET();
const u8* storeSingleS16 = AlignCode4(); const u8* storeSingleS16 = AlignCode4();
SHR(32, R(RSCRATCH2), Imm8(6)); SHR(32, R(RSCRATCH2), Imm8(5));
MOVSS(XMM1, MDisp(RSCRATCH2, (u32)(u64)m_quantizeTableS)); MOVSS(XMM1, MDisp(RSCRATCH2, (u32)(u64)m_quantizeTableS));
MULSS(XMM0, R(XMM1)); MULSS(XMM0, R(XMM1));
MAXSS(XMM0, M((void *)&m_m32768)); MAXSS(XMM0, M((void *)&m_m32768));
@ -507,13 +523,19 @@ void CommonAsmRoutines::GenQuantizedLoads()
UnsafeLoadRegToRegNoSwap(RSCRATCH_EXTRA, RSCRATCH_EXTRA, 16, 0); UnsafeLoadRegToRegNoSwap(RSCRATCH_EXTRA, RSCRATCH_EXTRA, 16, 0);
} }
MOVD_xmm(XMM0, R(RSCRATCH_EXTRA)); MOVD_xmm(XMM0, R(RSCRATCH_EXTRA));
PXOR(XMM1, R(XMM1)); if (cpu_info.bSSE4_1)
PUNPCKLBW(XMM0, R(XMM1)); {
PUNPCKLWD(XMM0, R(XMM1)); PMOVZXBD(XMM0, R(XMM0));
}
else
{
PXOR(XMM1, R(XMM1));
PUNPCKLBW(XMM0, R(XMM1));
PUNPCKLWD(XMM0, R(XMM1));
}
CVTDQ2PS(XMM0, R(XMM0)); CVTDQ2PS(XMM0, R(XMM0));
SHR(32, R(RSCRATCH2), Imm8(6)); SHR(32, R(RSCRATCH2), Imm8(5));
MOVSS(XMM1, MDisp(RSCRATCH2, (u32)(u64)m_dequantizeTableS)); MOVQ_xmm(XMM1, MDisp(RSCRATCH2, (u32)(u64)m_dequantizeTableS));
PUNPCKLDQ(XMM1, R(XMM1));
MULPS(XMM0, R(XMM1)); MULPS(XMM0, R(XMM1));
RET(); RET();
@ -524,7 +546,7 @@ void CommonAsmRoutines::GenQuantizedLoads()
UnsafeLoadRegToRegNoSwap(RSCRATCH_EXTRA, RSCRATCH_EXTRA, 8, 0); // RSCRATCH_EXTRA = 0x000000xx UnsafeLoadRegToRegNoSwap(RSCRATCH_EXTRA, RSCRATCH_EXTRA, 8, 0); // RSCRATCH_EXTRA = 0x000000xx
MOVD_xmm(XMM0, R(RSCRATCH_EXTRA)); MOVD_xmm(XMM0, R(RSCRATCH_EXTRA));
CVTDQ2PS(XMM0, R(XMM0)); // Is CVTSI2SS better? CVTDQ2PS(XMM0, R(XMM0)); // Is CVTSI2SS better?
SHR(32, R(RSCRATCH2), Imm8(6)); SHR(32, R(RSCRATCH2), Imm8(5));
MOVSS(XMM1, MDisp(RSCRATCH2, (u32)(u64)m_dequantizeTableS)); MOVSS(XMM1, MDisp(RSCRATCH2, (u32)(u64)m_dequantizeTableS));
MULSS(XMM0, R(XMM1)); MULSS(XMM0, R(XMM1));
UNPCKLPS(XMM0, M((void*)m_one)); UNPCKLPS(XMM0, M((void*)m_one));
@ -542,13 +564,19 @@ void CommonAsmRoutines::GenQuantizedLoads()
UnsafeLoadRegToRegNoSwap(RSCRATCH_EXTRA, RSCRATCH_EXTRA, 16, 0); UnsafeLoadRegToRegNoSwap(RSCRATCH_EXTRA, RSCRATCH_EXTRA, 16, 0);
} }
MOVD_xmm(XMM0, R(RSCRATCH_EXTRA)); MOVD_xmm(XMM0, R(RSCRATCH_EXTRA));
PUNPCKLBW(XMM0, R(XMM0)); if (cpu_info.bSSE4_1)
PUNPCKLWD(XMM0, R(XMM0)); {
PSRAD(XMM0, 24); PMOVSXBD(XMM0, R(XMM0));
}
else
{
PUNPCKLBW(XMM0, R(XMM0));
PUNPCKLWD(XMM0, R(XMM0));
PSRAD(XMM0, 24);
}
CVTDQ2PS(XMM0, R(XMM0)); CVTDQ2PS(XMM0, R(XMM0));
SHR(32, R(RSCRATCH2), Imm8(6)); SHR(32, R(RSCRATCH2), Imm8(5));
MOVSS(XMM1, MDisp(RSCRATCH2, (u32)(u64)m_dequantizeTableS)); MOVQ_xmm(XMM1, MDisp(RSCRATCH2, (u32)(u64)m_dequantizeTableS));
PUNPCKLDQ(XMM1, R(XMM1));
MULPS(XMM0, R(XMM1)); MULPS(XMM0, R(XMM1));
RET(); RET();
@ -559,7 +587,7 @@ void CommonAsmRoutines::GenQuantizedLoads()
UnsafeLoadRegToRegNoSwap(RSCRATCH_EXTRA, RSCRATCH_EXTRA, 8, 0, true); UnsafeLoadRegToRegNoSwap(RSCRATCH_EXTRA, RSCRATCH_EXTRA, 8, 0, true);
MOVD_xmm(XMM0, R(RSCRATCH_EXTRA)); MOVD_xmm(XMM0, R(RSCRATCH_EXTRA));
CVTDQ2PS(XMM0, R(XMM0)); CVTDQ2PS(XMM0, R(XMM0));
SHR(32, R(RSCRATCH2), Imm8(6)); SHR(32, R(RSCRATCH2), Imm8(5));
MOVSS(XMM1, MDisp(RSCRATCH2, (u32)(u64)m_dequantizeTableS)); MOVSS(XMM1, MDisp(RSCRATCH2, (u32)(u64)m_dequantizeTableS));
MULSS(XMM0, R(XMM1)); MULSS(XMM0, R(XMM1));
UNPCKLPS(XMM0, M((void*)m_one)); UNPCKLPS(XMM0, M((void*)m_one));
@ -573,12 +601,18 @@ void CommonAsmRoutines::GenQuantizedLoads()
UnsafeLoadRegToReg(RSCRATCH_EXTRA, RSCRATCH_EXTRA, 32, 0, false); UnsafeLoadRegToReg(RSCRATCH_EXTRA, RSCRATCH_EXTRA, 32, 0, false);
ROL(32, R(RSCRATCH_EXTRA), Imm8(16)); ROL(32, R(RSCRATCH_EXTRA), Imm8(16));
MOVD_xmm(XMM0, R(RSCRATCH_EXTRA)); MOVD_xmm(XMM0, R(RSCRATCH_EXTRA));
PXOR(XMM1, R(XMM1)); if (cpu_info.bSSE4_1)
PUNPCKLWD(XMM0, R(XMM1)); {
PMOVZXWD(XMM0, R(XMM0));
}
else
{
PXOR(XMM1, R(XMM1));
PUNPCKLWD(XMM0, R(XMM1));
}
CVTDQ2PS(XMM0, R(XMM0)); CVTDQ2PS(XMM0, R(XMM0));
SHR(32, R(RSCRATCH2), Imm8(6)); SHR(32, R(RSCRATCH2), Imm8(5));
MOVSS(XMM1, MDisp(RSCRATCH2, (u32)(u64)m_dequantizeTableS)); MOVQ_xmm(XMM1, MDisp(RSCRATCH2, (u32)(u64)m_dequantizeTableS));
PUNPCKLDQ(XMM1, R(XMM1));
MULPS(XMM0, R(XMM1)); MULPS(XMM0, R(XMM1));
RET(); RET();
@ -589,7 +623,7 @@ void CommonAsmRoutines::GenQuantizedLoads()
UnsafeLoadRegToReg(RSCRATCH_EXTRA, RSCRATCH_EXTRA, 16, 0, false); UnsafeLoadRegToReg(RSCRATCH_EXTRA, RSCRATCH_EXTRA, 16, 0, false);
MOVD_xmm(XMM0, R(RSCRATCH_EXTRA)); MOVD_xmm(XMM0, R(RSCRATCH_EXTRA));
CVTDQ2PS(XMM0, R(XMM0)); CVTDQ2PS(XMM0, R(XMM0));
SHR(32, R(RSCRATCH2), Imm8(6)); SHR(32, R(RSCRATCH2), Imm8(5));
MOVSS(XMM1, MDisp(RSCRATCH2, (u32)(u64)m_dequantizeTableS)); MOVSS(XMM1, MDisp(RSCRATCH2, (u32)(u64)m_dequantizeTableS));
MULSS(XMM0, R(XMM1)); MULSS(XMM0, R(XMM1));
UNPCKLPS(XMM0, M((void*)m_one)); UNPCKLPS(XMM0, M((void*)m_one));
@ -602,12 +636,18 @@ void CommonAsmRoutines::GenQuantizedLoads()
UnsafeLoadRegToReg(RSCRATCH_EXTRA, RSCRATCH_EXTRA, 32, 0, false); UnsafeLoadRegToReg(RSCRATCH_EXTRA, RSCRATCH_EXTRA, 32, 0, false);
ROL(32, R(RSCRATCH_EXTRA), Imm8(16)); ROL(32, R(RSCRATCH_EXTRA), Imm8(16));
MOVD_xmm(XMM0, R(RSCRATCH_EXTRA)); MOVD_xmm(XMM0, R(RSCRATCH_EXTRA));
PUNPCKLWD(XMM0, R(XMM0)); if (cpu_info.bSSE4_1)
PSRAD(XMM0, 16); {
PMOVSXWD(XMM0, R(XMM0));
}
else
{
PUNPCKLWD(XMM0, R(XMM0));
PSRAD(XMM0, 16);
}
CVTDQ2PS(XMM0, R(XMM0)); CVTDQ2PS(XMM0, R(XMM0));
SHR(32, R(RSCRATCH2), Imm8(6)); SHR(32, R(RSCRATCH2), Imm8(5));
MOVSS(XMM1, MDisp(RSCRATCH2, (u32)(u64)m_dequantizeTableS)); MOVQ_xmm(XMM1, MDisp(RSCRATCH2, (u32)(u64)m_dequantizeTableS));
PUNPCKLDQ(XMM1, R(XMM1));
MULPS(XMM0, R(XMM1)); MULPS(XMM0, R(XMM1));
RET(); RET();
@ -618,7 +658,7 @@ void CommonAsmRoutines::GenQuantizedLoads()
UnsafeLoadRegToReg(RSCRATCH_EXTRA, RSCRATCH_EXTRA, 16, 0, true); UnsafeLoadRegToReg(RSCRATCH_EXTRA, RSCRATCH_EXTRA, 16, 0, true);
MOVD_xmm(XMM0, R(RSCRATCH_EXTRA)); MOVD_xmm(XMM0, R(RSCRATCH_EXTRA));
CVTDQ2PS(XMM0, R(XMM0)); CVTDQ2PS(XMM0, R(XMM0));
SHR(32, R(RSCRATCH2), Imm8(6)); SHR(32, R(RSCRATCH2), Imm8(5));
MOVSS(XMM1, MDisp(RSCRATCH2, (u32)(u64)m_dequantizeTableS)); MOVSS(XMM1, MDisp(RSCRATCH2, (u32)(u64)m_dequantizeTableS));
MULSS(XMM0, R(XMM1)); MULSS(XMM0, R(XMM1));
UNPCKLPS(XMM0, M((void*)m_one)); UNPCKLPS(XMM0, M((void*)m_one));