sgilissen/llamaplayer
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

System: Refactor main loop

Browse Source 
Reduces JIT exits.
Improves runahead performance.
This commit is contained in:
Stenzek
2023-08-15 23:12:21 +10:00
parent 4ebd34fcb3
commit 5b980dafa5
43 changed files with 1343 additions and 923 deletions
Download Patch File Download Diff File Expand all files Collapse all files

146
src/core/gte.cpp
View File

@ -4,13 +4,13 @@
#include "gte.h"
#include "common/assert.h"
#include "common/bitutils.h"
#include "util/state_wrapper.h"
#include "cpu_core.h"
#include "cpu_core_private.h"
#include "host_display.h"
#include "pgxp.h"
#include "settings.h"
#include "timing_event.h"
#include "util/state_wrapper.h"
#include <algorithm>
#include <array>
#include <numeric>
@ -471,11 +471,12 @@ ALWAYS_INLINE static u32 UNRDivide(u32 lhs, u32 rhs)
return std::min<u32>(0x1FFFF, result);
}
static void MulMatVec(const s16 M[3][3], const s16 Vx, const s16 Vy, const s16 Vz, u8 shift, bool lm)
static void MulMatVec(const s16* M_, const s16 Vx, const s16 Vy, const s16 Vz, u8 shift, bool lm)
{
#define M(i, j) M_[((i)*3) + (j)]
#define dot3(i) \
TruncateAndSetMACAndIR<i + 1>(SignExtendMACResult<i + 1>((s64(M[i][0]) * s64(Vx)) + (s64(M[i][1]) * s64(Vy))) + \
(s64(M[i][2]) * s64(Vz)), \
TruncateAndSetMACAndIR<i + 1>(SignExtendMACResult<i + 1>((s64(M(i, 0)) * s64(Vx)) + (s64(M(i, 1)) * s64(Vy))) + \
(s64(M(i, 2)) * s64(Vz)), \
shift, lm)
dot3(0);
@ -483,15 +484,17 @@ static void MulMatVec(const s16 M[3][3], const s16 Vx, const s16 Vy, const s16 V
dot3(2);
#undef dot3
#undef M
}
static void MulMatVec(const s16 M[3][3], const s32 T[3], const s16 Vx, const s16 Vy, const s16 Vz, u8 shift, bool lm)
static void MulMatVec(const s16* M_, const s32 T[3], const s16 Vx, const s16 Vy, const s16 Vz, u8 shift, bool lm)
{
#define M(i, j) M_[((i)*3) + (j)]
#define dot3(i) \
TruncateAndSetMACAndIR<i + 1>( \
SignExtendMACResult<i + 1>(SignExtendMACResult<i + 1>((s64(T[i]) << 12) + (s64(M[i][0]) * s64(Vx))) + \
(s64(M[i][1]) * s64(Vy))) + \
(s64(M[i][2]) * s64(Vz)), \
SignExtendMACResult<i + 1>(SignExtendMACResult<i + 1>((s64(T[i]) << 12) + (s64(M(i, 0)) * s64(Vx))) + \
(s64(M(i, 1)) * s64(Vy))) + \
(s64(M(i, 2)) * s64(Vz)), \
shift, lm)
dot3(0);
@ -499,19 +502,20 @@ static void MulMatVec(const s16 M[3][3], const s32 T[3], const s16 Vx, const s16
dot3(2);
#undef dot3
#undef M
}
static void MulMatVecBuggy(const s16 M[3][3], const s32 T[3], const s16 Vx, const s16 Vy, const s16 Vz, u8 shift,
bool lm)
static void MulMatVecBuggy(const s16* M_, const s32 T[3], const s16 Vx, const s16 Vy, const s16 Vz, u8 shift, bool lm)
{
#define M(i, j) M_[((i)*3) + (j)]
#define dot3(i) \
do \
{ \
TruncateAndSetIR<i + 1>(static_cast<s32>(SignExtendMACResult<i + 1>(SignExtendMACResult<i + 1>( \
(s64(T[i]) << 12) + (s64(M[i][0]) * s64(Vx)))) >> \
(s64(T[i]) << 12) + (s64(M(i, 0)) * s64(Vx)))) >> \
shift), \
false); \
TruncateAndSetMACAndIR<i + 1>(SignExtendMACResult<i + 1>((s64(M[i][1]) * s64(Vy))) + (s64(M[i][2]) * s64(Vz)), \
TruncateAndSetMACAndIR<i + 1>(SignExtendMACResult<i + 1>((s64(M(i, 1)) * s64(Vy))) + (s64(M(i, 2)) * s64(Vz)), \
shift, lm); \
} while (0)
@ -520,82 +524,50 @@ static void MulMatVecBuggy(const s16 M[3][3], const s32 T[3], const s16 Vx, cons
dot3(2);
#undef dot3
#undef M
}
static void Execute_MVMVA(Instruction inst)
{
REGS.FLAG.Clear();
// TODO: Remove memcpy..
s16 M[3][3];
switch (inst.mvmva_multiply_matrix)
static constexpr const s16* M_lookup[4] = {&REGS.RT[0][0], &REGS.LLM[0][0], &REGS.LCM[0][0], nullptr};
static constexpr const s16* V_lookup[4][3] = {
{&REGS.V0[0], &REGS.V0[1], &REGS.V0[2]},
{&REGS.V1[0], &REGS.V1[1], &REGS.V1[2]},
{&REGS.V2[0], &REGS.V2[1], &REGS.V2[2]},
{&REGS.IR1, &REGS.IR2, &REGS.IR3},
};
static constexpr const s32 zero_T[3] = {};
static constexpr const s32* T_lookup[4] = {REGS.TR, REGS.BK, REGS.FC, zero_T};
const s16* M = M_lookup[inst.mvmva_multiply_matrix];
const s16* const* const V = V_lookup[inst.mvmva_multiply_vector];
const s32* const T = T_lookup[inst.mvmva_translation_vector];
s16 buggy_M[3][3];
if (!M)
{
case 0:
std::memcpy(M, REGS.RT, sizeof(s16) * 3 * 3);
break;
case 1:
std::memcpy(M, REGS.LLM, sizeof(s16) * 3 * 3);
break;
case 2:
std::memcpy(M, REGS.LCM, sizeof(s16) * 3 * 3);
break;
default:
{
// buggy
M[0][0] = -static_cast<s16>(ZeroExtend16(REGS.RGBC[0]) << 4);
M[0][1] = static_cast<s16>(ZeroExtend16(REGS.RGBC[0]) << 4);
M[0][2] = REGS.IR0;
M[1][0] = REGS.RT[0][2];
M[1][1] = REGS.RT[0][2];
M[1][2] = REGS.RT[0][2];
M[2][0] = REGS.RT[1][1];
M[2][1] = REGS.RT[1][1];
M[2][2] = REGS.RT[1][1];
}
break;
// buggy
buggy_M[0][0] = -static_cast<s16>(ZeroExtend16(REGS.RGBC[0]) << 4);
buggy_M[0][1] = static_cast<s16>(ZeroExtend16(REGS.RGBC[0]) << 4);
buggy_M[0][2] = REGS.IR0;
buggy_M[1][0] = REGS.RT[0][2];
buggy_M[1][1] = REGS.RT[0][2];
buggy_M[1][2] = REGS.RT[0][2];
buggy_M[2][0] = REGS.RT[1][1];
buggy_M[2][1] = REGS.RT[1][1];
buggy_M[2][2] = REGS.RT[1][1];
M = &buggy_M[0][0];
}
s16 Vx, Vy, Vz;
switch (inst.mvmva_multiply_vector)
{
case 0:
Vx = REGS.V0[0];
Vy = REGS.V0[1];
Vz = REGS.V0[2];
break;
case 1:
Vx = REGS.V1[0];
Vy = REGS.V1[1];
Vz = REGS.V1[2];
break;
case 2:
Vx = REGS.V2[0];
Vy = REGS.V2[1];
Vz = REGS.V2[2];
break;
default:
Vx = REGS.IR1;
Vy = REGS.IR2;
Vz = REGS.IR3;
break;
}
static const s32 zero_T[3] = {};
switch (inst.mvmva_translation_vector)
{
case 0:
MulMatVec(M, REGS.TR, Vx, Vy, Vz, inst.GetShift(), inst.lm);
break;
case 1:
MulMatVec(M, REGS.BK, Vx, Vy, Vz, inst.GetShift(), inst.lm);
break;
case 2:
MulMatVecBuggy(M, REGS.FC, Vx, Vy, Vz, inst.GetShift(), inst.lm);
break;
default:
MulMatVec(M, zero_T, Vx, Vy, Vz, inst.GetShift(), inst.lm);
break;
}
const s16 Vx = *V[0];
const s16 Vy = *V[1];
const s16 Vz = *V[2];
if (inst.mvmva_translation_vector != 2)
MulMatVec(M, T, Vx, Vy, Vz, inst.GetShift(), inst.lm);
else
MulMatVecBuggy(M, T, Vx, Vy, Vz, inst.GetShift(), inst.lm);
REGS.FLAG.UpdateError();
}
@ -874,10 +846,10 @@ static ALWAYS_INLINE void InterpolateColor(s64 in_MAC1, s64 in_MAC2, s64 in_MAC3
static void NCS(const s16 V[3], u8 shift, bool lm)
{
// [IR1,IR2,IR3] = [MAC1,MAC2,MAC3] = (LLM*V0) SAR (sf*12)
MulMatVec(REGS.LLM, V[0], V[1], V[2], shift, lm);
MulMatVec(&REGS.LLM[0][0], V[0], V[1], V[2], shift, lm);
// [IR1,IR2,IR3] = [MAC1,MAC2,MAC3] = (BK*1000h + LCM*IR) SAR (sf*12)
MulMatVec(REGS.LCM, REGS.BK, REGS.IR1, REGS.IR2, REGS.IR3, shift, lm);
MulMatVec(&REGS.LCM[0][0], REGS.BK, REGS.IR1, REGS.IR2, REGS.IR3, shift, lm);
// Color FIFO = [MAC1/16,MAC2/16,MAC3/16,CODE], [IR1,IR2,IR3] = [MAC1,MAC2,MAC3]
PushRGBFromMAC();
@ -909,10 +881,10 @@ static void Execute_NCT(Instruction inst)
static void NCCS(const s16 V[3], u8 shift, bool lm)
{
// [IR1,IR2,IR3] = [MAC1,MAC2,MAC3] = (LLM*V0) SAR (sf*12)
MulMatVec(REGS.LLM, V[0], V[1], V[2], shift, lm);
MulMatVec(&REGS.LLM[0][0], V[0], V[1], V[2], shift, lm);
// [IR1,IR2,IR3] = [MAC1,MAC2,MAC3] = (BK*1000h + LCM*IR) SAR (sf*12)
MulMatVec(REGS.LCM, REGS.BK, REGS.IR1, REGS.IR2, REGS.IR3, shift, lm);
MulMatVec(&REGS.LCM[0][0], REGS.BK, REGS.IR1, REGS.IR2, REGS.IR3, shift, lm);
// [MAC1,MAC2,MAC3] = [R*IR1,G*IR2,B*IR3] SHL 4 ;<--- for NCDx/NCCx
// [MAC1,MAC2,MAC3] = [MAC1,MAC2,MAC3] SAR (sf*12) ;<--- for NCDx/NCCx
@ -950,10 +922,10 @@ static void Execute_NCCT(Instruction inst)
static void NCDS(const s16 V[3], u8 shift, bool lm)
{
// [IR1,IR2,IR3] = [MAC1,MAC2,MAC3] = (LLM*V0) SAR (sf*12)
MulMatVec(REGS.LLM, V[0], V[1], V[2], shift, lm);
MulMatVec(&REGS.LLM[0][0], V[0], V[1], V[2], shift, lm);
// [IR1,IR2,IR3] = [MAC1,MAC2,MAC3] = (BK*1000h + LCM*IR) SAR (sf*12)
MulMatVec(REGS.LCM, REGS.BK, REGS.IR1, REGS.IR2, REGS.IR3, shift, lm);
MulMatVec(&REGS.LCM[0][0], REGS.BK, REGS.IR1, REGS.IR2, REGS.IR3, shift, lm);
// No need to assign these to MAC[1-3], as it'll never overflow.
// [MAC1,MAC2,MAC3] = [R*IR1,G*IR2,B*IR3] SHL 4 ;<--- for NCDx/NCCx
@ -999,7 +971,7 @@ static void Execute_CC(Instruction inst)
const bool lm = inst.lm;
// [IR1,IR2,IR3] = [MAC1,MAC2,MAC3] = (BK*1000h + LCM*IR) SAR (sf*12)
MulMatVec(REGS.LCM, REGS.BK, REGS.IR1, REGS.IR2, REGS.IR3, shift, lm);
MulMatVec(&REGS.LCM[0][0], REGS.BK, REGS.IR1, REGS.IR2, REGS.IR3, shift, lm);
// [MAC1,MAC2,MAC3] = [R*IR1,G*IR2,B*IR3] SHL 4
// [MAC1,MAC2,MAC3] = [MAC1,MAC2,MAC3] SAR (sf*12)
@ -1021,7 +993,7 @@ static void Execute_CDP(Instruction inst)
const bool lm = inst.lm;
// [IR1,IR2,IR3] = [MAC1,MAC2,MAC3] = (BK*1000h + LCM*IR) SAR (sf*12)
MulMatVec(REGS.LCM, REGS.BK, REGS.IR1, REGS.IR2, REGS.IR3, shift, lm);
MulMatVec(&REGS.LCM[0][0], REGS.BK, REGS.IR1, REGS.IR2, REGS.IR3, shift, lm);
// No need to assign these to MAC[1-3], as it'll never overflow.
// [MAC1,MAC2,MAC3] = [R*IR1,G*IR2,B*IR3] SHL 4