thirdparty/ppsspp
2
0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

samplerjit: Skip processing alpha if unused.

Browse source
This commit is contained in:
Unknown W. Brackets 2022-01-09 11:10:12 -08:00
parent fe0b3dbd01
commit 72aa4be879
2 changed files with 24 additions and 22 deletions
Download patch file Download diff file Expand all files Collapse all files

6
GPU/Software/Sampler.h
View file

@ -88,9 +88,9 @@ private:
bool Jit_GetTexData(const SamplerID &id, int bitsPerTexel); bool Jit_GetTexData(const SamplerID &id, int bitsPerTexel);
bool Jit_GetTexDataSwizzled(const SamplerID &id, int bitsPerTexel); bool Jit_GetTexDataSwizzled(const SamplerID &id, int bitsPerTexel);
bool Jit_GetTexDataSwizzled4(const SamplerID &id); bool Jit_GetTexDataSwizzled4(const SamplerID &id);
bool Jit_Decode5650(); bool Jit_Decode5650(const SamplerID &id);
bool Jit_Decode5551(); bool Jit_Decode5551(const SamplerID &id);
bool Jit_Decode4444(); bool Jit_Decode4444(const SamplerID &id);
bool Jit_TransformClutIndex(const SamplerID &id, int bitsPerIndex); bool Jit_TransformClutIndex(const SamplerID &id, int bitsPerIndex);
bool Jit_ReadClutColor(const SamplerID &id); bool Jit_ReadClutColor(const SamplerID &id);
bool Jit_GetDXT1Color(const SamplerID &id, int blockSize, int alpha); bool Jit_GetDXT1Color(const SamplerID &id, int blockSize, int alpha);

40
GPU/Software/SamplerX86.cpp
View file

@ -1701,19 +1701,19 @@ bool SamplerJitCache::Jit_ReadTextureFormat(const SamplerID &id) {
case GE_TFMT_5650: case GE_TFMT_5650:
success = Jit_GetTexData(id, 16); success = Jit_GetTexData(id, 16);
if (success) if (success)
success = Jit_Decode5650(); success = Jit_Decode5650(id);
break; break;
case GE_TFMT_5551: case GE_TFMT_5551:
success = Jit_GetTexData(id, 16); success = Jit_GetTexData(id, 16);
if (success) if (success)
success = Jit_Decode5551(); success = Jit_Decode5551(id);
break; break;
case GE_TFMT_4444: case GE_TFMT_4444:
success = Jit_GetTexData(id, 16); success = Jit_GetTexData(id, 16);
if (success) if (success)
success = Jit_Decode4444(); success = Jit_Decode4444(id);
break; break;
case GE_TFMT_8888: case GE_TFMT_8888:
@ -3102,7 +3102,7 @@ bool SamplerJitCache::Jit_Decode5650Quad(const SamplerID &id, Rasterizer::RegCac
return true; return true;
} }
bool SamplerJitCache::Jit_Decode5650() { bool SamplerJitCache::Jit_Decode5650(const SamplerID &id) {
Describe("5650"); Describe("5650");
X64Reg resultReg = regCache_.Find(RegCache::GEN_RESULT); X64Reg resultReg = regCache_.Find(RegCache::GEN_RESULT);
X64Reg temp1Reg = regCache_.Alloc(RegCache::GEN_TEMP1); X64Reg temp1Reg = regCache_.Alloc(RegCache::GEN_TEMP1);
@ -3125,7 +3125,8 @@ bool SamplerJitCache::Jit_Decode5650() {
AND(32, R(temp2Reg), Imm32(0x00FF00FF)); AND(32, R(temp2Reg), Imm32(0x00FF00FF));
// Now's as good a time to put in A as any. // Now's as good a time to put in A as any.
OR(32, R(temp2Reg), Imm32(0xFF000000)); if (id.useTextureAlpha || id.fetch)
OR(32, R(temp2Reg), Imm32(0xFF000000));
// Last, we need to align, extract, and expand G. // Last, we need to align, extract, and expand G.
// 3 to align to G, and then 2 to expand to 8. // 3 to align to G, and then 2 to expand to 8.
@ -3172,7 +3173,7 @@ bool SamplerJitCache::Jit_Decode5551Quad(const SamplerID &id, Rasterizer::RegCac
// First, extend alpha using an arithmetic shift. // First, extend alpha using an arithmetic shift.
// We use 10 to meanwhile get rid of green too. The extra alpha bits are fine. // We use 10 to meanwhile get rid of green too. The extra alpha bits are fine.
PSRAW(quadReg, 10); PSRAW(quadReg, 10);
// This gets rid of those extra alpha bits. // This gets rid of those extra alpha bits and puts blue in place too.
PSLLD(quadReg, 19); PSLLD(quadReg, 19);
// Combine both together, we still need to swizzle. // Combine both together, we still need to swizzle.
@ -3194,7 +3195,7 @@ bool SamplerJitCache::Jit_Decode5551Quad(const SamplerID &id, Rasterizer::RegCac
return true; return true;
} }
bool SamplerJitCache::Jit_Decode5551() { bool SamplerJitCache::Jit_Decode5551(const SamplerID &id) {
Describe("5551"); Describe("5551");
X64Reg resultReg = regCache_.Find(RegCache::GEN_RESULT); X64Reg resultReg = regCache_.Find(RegCache::GEN_RESULT);
X64Reg temp1Reg = regCache_.Alloc(RegCache::GEN_TEMP1); X64Reg temp1Reg = regCache_.Alloc(RegCache::GEN_TEMP1);
@ -3220,14 +3221,15 @@ bool SamplerJitCache::Jit_Decode5551() {
AND(32, R(temp1Reg), Imm32(0x00070707)); AND(32, R(temp1Reg), Imm32(0x00070707));
OR(32, R(temp2Reg), R(temp1Reg)); OR(32, R(temp2Reg), R(temp1Reg));
// For A, we shift it to a single bit, and then subtract and XOR. if (id.useTextureAlpha || id.fetch) {
// That's probably the simplest way to expand it... // For A, we sign extend to get either 16 1s or 0s of alpha.
SHR(32, R(resultReg), Imm8(15)); SAR(16, R(resultReg), Imm8(15));
// If it was 0, it's now -1, otherwise it's 0. Easy. // Now, shift left by 24 to get the lowest 8 of those at the top.
SUB(32, R(resultReg), Imm8(1)); SHL(32, R(resultReg), Imm8(24));
XOR(32, R(resultReg), Imm32(0xFF000000)); OR(32, R(resultReg), R(temp2Reg));
AND(32, R(resultReg), Imm32(0xFF000000)); } else {
OR(32, R(resultReg), R(temp2Reg)); MOV(32, R(resultReg), R(temp2Reg));
}
regCache_.Release(temp1Reg, RegCache::GEN_TEMP1); regCache_.Release(temp1Reg, RegCache::GEN_TEMP1);
regCache_.Release(temp2Reg, RegCache::GEN_TEMP2); regCache_.Release(temp2Reg, RegCache::GEN_TEMP2);
@ -3301,7 +3303,7 @@ bool SamplerJitCache::Jit_Decode4444Quad(const SamplerID &id, Rasterizer::RegCac
alignas(16) static const u32 color4444mask[4] = { 0xf00ff00f, 0xf00ff00f, 0xf00ff00f, 0xf00ff00f, }; alignas(16) static const u32 color4444mask[4] = { 0xf00ff00f, 0xf00ff00f, 0xf00ff00f, 0xf00ff00f, };
bool SamplerJitCache::Jit_Decode4444() { bool SamplerJitCache::Jit_Decode4444(const SamplerID &id) {
Describe("4444"); Describe("4444");
X64Reg resultReg = regCache_.Find(RegCache::GEN_RESULT); X64Reg resultReg = regCache_.Find(RegCache::GEN_RESULT);
X64Reg vecTemp1Reg = regCache_.Alloc(RegCache::VEC_TEMP1); X64Reg vecTemp1Reg = regCache_.Alloc(RegCache::VEC_TEMP1);
@ -3446,13 +3448,13 @@ bool SamplerJitCache::Jit_ReadClutColor(const SamplerID &id) {
switch (id.ClutFmt()) { switch (id.ClutFmt()) {
case GE_CMODE_16BIT_BGR5650: case GE_CMODE_16BIT_BGR5650:
return Jit_Decode5650(); return Jit_Decode5650(id);
case GE_CMODE_16BIT_ABGR5551: case GE_CMODE_16BIT_ABGR5551:
return Jit_Decode5551(); return Jit_Decode5551(id);
case GE_CMODE_16BIT_ABGR4444: case GE_CMODE_16BIT_ABGR4444:
return Jit_Decode4444(); return Jit_Decode4444(id);
case GE_CMODE_32BIT_ABGR8888: case GE_CMODE_32BIT_ABGR8888:
return true; return true;