Emulation/ppsspp
1
0
Fork 0
mirror of https://github.com/hrydgard/ppsspp.git synced 2025-04-02 11:01:50 -04:00
Code Issues Projects Releases Packages Wiki Activity

VFPU-Int: Add a few fast-paths to ReadMatrix/WriteMatrix. Should gain back part of the speed lost in #12217

Browse source 
(which fixed #5399).
This commit is contained in:
Henrik Rydgård 2019-08-06 14:53:11 +02:00
parent 9b30e5393d
commit 8642134419
1 changed files with 83 additions and 19 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

102
Core/MIPS/MIPSVFPUUtils.cpp
View file

@ -17,7 +17,8 @@
#include <cstdint> #include <cstdint>
#include <limits> #include <limits>
#include <stdio.h> #include <cstdio>
#include <cstring>
#include "Common/BitScan.h" #include "Common/BitScan.h"
#include "Common/CommonFuncs.h" #include "Common/CommonFuncs.h"
@ -245,14 +246,44 @@ void ReadMatrix(float *rd, MatrixSize size, int reg) {
default: _assert_msg_(JIT, 0, "%s: Bad matrix size", __FUNCTION__); default: _assert_msg_(JIT, 0, "%s: Bad matrix size", __FUNCTION__);
} }
for (int i = 0; i < side; i++) { // The voffset ordering is now integrated in these formulas,
for (int j = 0; j < side; j++) { // eliminating a table lookup.
int index = mtx * 4; const float *v = currentMIPS->v + (size_t)mtx * 16;
if (transpose) if (transpose) {
index += ((row+i)&3) + ((col+j)&3)*32; if (side == 4 && col == 0 && row == 0) {
else // Simple 4x4 transpose. TODO: Optimize.
index += ((col+j)&3) + ((row+i)&3)*32; for (int j = 0; j < 4; j++) {
rd[j*4 + i] = V(index); for (int i = 0; i < 4; i++) {
rd[j * 4 + i] = v[i * 4 + j];
}
}
} else {
for (int j = 0; j < side; j++) {
for (int i = 0; i < side; i++) {
int index = ((row + i) & 3) * 4 + ((col + j) & 3);
rd[j * 4 + i] = v[index];
}
}
}
} else {
// Fast path
if (side == 4 && col == 0 && row == 0) {
memcpy(rd, v, sizeof(float) * 16);
/*
for (int j = 0; j < side; j++) {
for (int i = 0; i < side; i++) {
int index = mtx * 16 + j * 4 + i;
rd[j * 4 + i] = v[index];
}
}
*/
} else {
for (int j = 0; j < side; j++) {
for (int i = 0; i < side; i++) {
int index = ((col + j) & 3) * 4 + ((row + i) & 3);
rd[j * 4 + i] = v[index];
}
}
} }
} }
} }
@ -277,16 +308,49 @@ void WriteMatrix(const float *rd, MatrixSize size, int reg) {
ERROR_LOG_REPORT(CPU, "Write mask used with vfpu matrix instruction."); ERROR_LOG_REPORT(CPU, "Write mask used with vfpu matrix instruction.");
} }
for (int i = 0; i < side; i++) { // The voffset ordering is now integrated in these formulas,
for (int j = 0; j < side; j++) { // eliminating a table lookup.
// Hm, I wonder if this should affect matrices at all. float *v = currentMIPS->v + (size_t)mtx * 16;
if (j != side -1 || !currentMIPS->VfpuWriteMask(i)) { if (transpose) {
int index = mtx * 4; // Fast path
if (transpose) if (side == 4 && row == 0 && col == 0 && currentMIPS->VfpuWriteMask() == 0x0) {
index += ((row+i)&3) + ((col+j)&3)*32; for (int j = 0; j < side; j++) {
else for (int i = 0; i < side; i++) {
index += ((col+j)&3) + ((row+i)&3)*32; int index = i * 4 + j;
V(index) = rd[j*4+i]; v[index] = rd[j * 4 + i];
}
}
} else {
for (int j = 0; j < side; j++) {
for (int i = 0; i < side; i++) {
// Hm, I wonder if this should affect matrices at all.
if (j != side - 1 || !currentMIPS->VfpuWriteMask(i)) {
int index = ((row + i) & 3) * 4 + ((col + j) & 3);
v[index] = rd[j * 4 + i];
}
}
}
}
} else {
if (side == 4 && row == 0 && col == 0 && currentMIPS->VfpuWriteMask() == 0x0) {
memcpy(v, rd, sizeof(float) * 16);
/*
for (int j = 0; j < side; j++) {
for (int i = 0; i < side; i++) {
int index = j * 4 + i;
v[index] = rd[j * 4 + i];
}
}
*/
} else {
for (int j = 0; j < side; j++) {
for (int i = 0; i < side; i++) {
// Hm, I wonder if this should affect matrices at all.
if (j != side - 1 || !currentMIPS->VfpuWriteMask(i)) {
int index = ((col + j) & 3) * 4 + ((row + i) & 3);
v[index] = rd[j * 4 + i];
}
}
} }
} }
} }