libretro/ppsspp
1
0
Fork 0
mirror of https://github.com/libretro/ppsspp.git synced 2024-12-02 22:26:25 +00:00
Code Issues Actions Packages Projects Releases Wiki Activity

HLE: Replace starocean framebuf clear func.

Browse Source 
This reduces the performance impact significantly, by skipping the memset
uploads for each line.

Fixes #10466.
This commit is contained in:
Unknown W. Brackets 2017-12-28 19:28:39 -08:00
parent 0151c5e649
commit 28da05fa7a
2 changed files with 70 additions and 17 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

86
Core/HLE/ReplaceTables.cpp
View File

@ -41,6 +41,14 @@
#include <emmintrin.h>
#endif
enum class GPUReplacementSkip {
MEMSET = 1,
MEMCPY = 2,
MEMMOVE = 4,
};
static int skipGPUReplacements = 0;
// I think these have to be pretty accurate as these are libc replacements,
// but we can probably get away with approximating the VFPU vsin/vcos and vrot
// pretty roughly.
@ -119,8 +127,10 @@ static int Replace_memcpy() {
// Some games use memcpy on executable code. We need to flush emuhack ops.
currentMIPS->InvalidateICache(srcPtr, bytes);
if (Memory::IsVRAMAddress(destPtr) || Memory::IsVRAMAddress(srcPtr)) {
skip = gpu->PerformMemoryCopy(destPtr, srcPtr, bytes);
if ((skipGPUReplacements & (int)GPUReplacementSkip::MEMCPY) == 0) {
if (Memory::IsVRAMAddress(destPtr) || Memory::IsVRAMAddress(srcPtr)) {
skip = gpu->PerformMemoryCopy(destPtr, srcPtr, bytes);
}
}
if (!skip && bytes != 0) {
u8 *dst = Memory::GetPointer(destPtr);
@ -159,8 +169,10 @@ static int Replace_memcpy_jak() {
return 5;
}
currentMIPS->InvalidateICache(srcPtr, bytes);
if (Memory::IsVRAMAddress(destPtr) || Memory::IsVRAMAddress(srcPtr)) {
skip = gpu->PerformMemoryCopy(destPtr, srcPtr, bytes);
if ((skipGPUReplacements & (int)GPUReplacementSkip::MEMCPY) == 0) {
if (Memory::IsVRAMAddress(destPtr) || Memory::IsVRAMAddress(srcPtr)) {
skip = gpu->PerformMemoryCopy(destPtr, srcPtr, bytes);
}
}
if (!skip && bytes != 0) {
u8 *dst = Memory::GetPointer(destPtr);
@ -197,8 +209,10 @@ static int Replace_memcpy16() {
// Some games use memcpy on executable code. We need to flush emuhack ops.
currentMIPS->InvalidateICache(srcPtr, bytes);
if (Memory::IsVRAMAddress(destPtr) || Memory::IsVRAMAddress(srcPtr)) {
skip = gpu->PerformMemoryCopy(destPtr, srcPtr, bytes);
if ((skipGPUReplacements & (int)GPUReplacementSkip::MEMCPY) == 0) {
if (Memory::IsVRAMAddress(destPtr) || Memory::IsVRAMAddress(srcPtr)) {
skip = gpu->PerformMemoryCopy(destPtr, srcPtr, bytes);
}
}
if (!skip && bytes != 0) {
u8 *dst = Memory::GetPointer(destPtr);
@ -220,8 +234,10 @@ static int Replace_memcpy_swizzled() {
u32 srcPtr = PARAM(1);
u32 pitch = PARAM(2);
u32 h = PARAM(4);
if (Memory::IsVRAMAddress(srcPtr)) {
gpu->PerformMemoryDownload(srcPtr, pitch * h);
if ((skipGPUReplacements & (int)GPUReplacementSkip::MEMCPY) == 0) {
if (Memory::IsVRAMAddress(srcPtr)) {
gpu->PerformMemoryDownload(srcPtr, pitch * h);
}
}
u8 *dstp = Memory::GetPointer(destPtr);
const u8 *srcp = Memory::GetPointer(srcPtr);
@ -258,9 +274,11 @@ static int Replace_memmove() {
bool skip = false;
// Some games use memcpy on executable code. We need to flush emuhack ops.
currentMIPS->InvalidateICache(srcPtr, bytes);
if (Memory::IsVRAMAddress(destPtr) || Memory::IsVRAMAddress(srcPtr)) {
skip = gpu->PerformMemoryCopy(destPtr, srcPtr, bytes);
if ((skipGPUReplacements & (int)GPUReplacementSkip::MEMMOVE) == 0) {
currentMIPS->InvalidateICache(srcPtr, bytes);
if (Memory::IsVRAMAddress(destPtr) || Memory::IsVRAMAddress(srcPtr)) {
skip = gpu->PerformMemoryCopy(destPtr, srcPtr, bytes);
}
}
if (!skip && bytes != 0) {
u8 *dst = Memory::GetPointer(destPtr);
@ -282,7 +300,7 @@ static int Replace_memset() {
u8 value = PARAM(1);
u32 bytes = PARAM(2);
bool skip = false;
if (Memory::IsVRAMAddress(destPtr)) {
if (Memory::IsVRAMAddress(destPtr) && (skipGPUReplacements & (int)GPUReplacementSkip::MEMSET) == 0) {
skip = gpu->PerformMemorySet(destPtr, value, bytes);
}
if (!skip && bytes != 0) {
@ -309,7 +327,7 @@ static int Replace_memset_jak() {
}
bool skip = false;
if (Memory::IsVRAMAddress(destPtr)) {
if (Memory::IsVRAMAddress(destPtr) && (skipGPUReplacements & (int)GPUReplacementSkip::MEMSET) == 0) {
skip = gpu->PerformMemorySet(destPtr, value, bytes);
}
if (!skip && bytes != 0) {
@ -599,6 +617,17 @@ static bool GetMIPSStaticAddress(u32 &addr, s32 lui_offset, s32 lw_offset) {
return true;
}
static bool GetMIPSGPAddress(u32 &addr, s32 offset) {
const MIPSOpcode loadOp = Memory::Read_Instruction(currentMIPS->pc + offset, true);
if (MIPS_GET_RS(loadOp) == MIPS_REG_GP) {
s16 gpoff = (s16)(u16)(loadOp & 0x0000FFFF);
addr = currentMIPS->r[MIPS_REG_GP] + gpoff;
return true;
}
return false;
}
static int Hook_godseaterburst_blit_texture() {
u32 texaddr;
// Only if there's no texture.
@ -1032,7 +1061,6 @@ static int Hook_tonyhawkp8_upload_tutorial_frame() {
const u32 fb_address = currentMIPS->r[MIPS_REG_A0];
if (Memory::IsVRAMAddress(fb_address)) {
gpu->PerformMemoryUpload(fb_address, 0x00088000);
CBreakPoints::ExecMemCheck(fb_address, true, 0x00088000, currentMIPS->pc);
}
return 0;
}
@ -1145,7 +1173,6 @@ static int Hook_mytranwars_upload_frame() {
u32 fb_address = currentMIPS->r[MIPS_REG_S0];
if (Memory::IsVRAMAddress(fb_address)) {
gpu->PerformMemoryUpload(fb_address, 0x00088000);
CBreakPoints::ExecMemCheck(fb_address, true, 0x00088000, currentMIPS->pc);
}
return 0;
}
@ -1183,6 +1210,29 @@ static int Hook_marvelalliance1_copy_after() {
return 0;
}
static int Hook_starocean_clear_framebuf_before() {
skipGPUReplacements |= (int)GPUReplacementSkip::MEMSET;
return 0;
}
static int Hook_starocean_clear_framebuf_after() {
skipGPUReplacements &= ~(int)GPUReplacementSkip::MEMSET;
// This hook runs after the copy, this is the final memcpy destination.
u32 framebuf = currentMIPS->r[MIPS_REG_V0] - 512 * 4 * 271;
u32 y_address, h_address;
if (GetMIPSGPAddress(y_address, -204) && GetMIPSGPAddress(h_address, -200)) {
int y = (s16)Memory::Read_U16(y_address);
int h = (s16)Memory::Read_U16(h_address);
DEBUG_LOG(HLE, "starocean_clear_framebuf() - %08x y=%d-%d", framebuf, y, h);
// TODO: This is always clearing to 0, actually, which could be faster than an upload.
gpu->PerformMemoryUpload(framebuf + 512 * y * 4, 512 * h * 4);
}
return 0;
}
#define JITFUNC(f) (&MIPSComp::MIPSFrontendInterface::f)
// Can either replace with C functions or functions emitted in Asm/ArmAsm.
@ -1292,6 +1342,8 @@ static const ReplacementTableEntry entries[] = {
{ "marvelalliance1_copy", &Hook_marvelalliance1_copy_after, 0, REPFLAG_HOOKENTER, 0x638 },
{ "marvelalliance1_copy", &Hook_marvelalliance1_copy_a1_before, 0, REPFLAG_HOOKENTER, 0x664 },
{ "marvelalliance1_copy", &Hook_marvelalliance1_copy_after, 0, REPFLAG_HOOKENTER, 0x69c },
{ "starocean_clear_framebuf", &Hook_starocean_clear_framebuf_before, 0, REPFLAG_HOOKENTER, 0 },
{ "starocean_clear_framebuf", &Hook_starocean_clear_framebuf_after, 0, REPFLAG_HOOKEXIT, 0 },
{}
};
@ -1306,6 +1358,8 @@ void Replacement_Init() {
continue;
replacementNameLookup[entry->name].push_back(i);
}
skipGPUReplacements = 0;
}
void Replacement_Shutdown() {
@ -1313,8 +1367,6 @@ void Replacement_Shutdown() {
replacementNameLookup.clear();
}
// TODO: Do something on load state?
int GetNumReplacementFuncs() {
return ARRAY_SIZE(entries);
}

1
Core/MIPS/MIPSAnalyst.cpp
View File

@ -382,6 +382,7 @@ static const HardHashTableEntry hardcodedHashes[] = {
{ 0xb0ef265e87899f0a, 32, "vector_divide_t_s", },
{ 0xb183a37baa12607b, 32, "vscl_t", },
{ 0xb1a3e60a89af9857, 20, "fabs", },
{ 0xb25670ff47b4843d, 232, "starocean_clear_framebuf" },
{ 0xb3fef47fb27d57c9, 44, "vector_scale_t", },
{ 0xb43fd5078ae78029, 84, "send_commandi_stall", },
{ 0xb43ffbd4dc446dd2, 324, "atan2f", },