archive/rpcs3
1
0
Fork 0
mirror of https://github.com/RPCS3/rpcs3.git synced 2025-07-12 17:58:37 +12:00
Code Issues Projects Releases Packages Wiki Activity Actions 1

rsx/fp: Fix some decompiler bugs

Browse source
This commit is contained in:
kd-11 2017-03-07 13:40:38 +03:00
parent 1c8cb3b7d3
commit be4bb48476
13 changed files with 174 additions and 126 deletions
Download patch file Download diff file Expand all files Collapse all files

82
rpcs3/Emu/RSX/Common/FragmentProgramDecompiler.cpp
View file

@ -41,6 +41,10 @@ void FragmentProgramDecompiler::SetDst(std::string code, bool append_mask)
{
code = saturate(code);
}
else if (!dst.no_dest)
{
code = NoOverflow(code);
}
code += (append_mask ? "$m" : "");
@ -60,6 +64,13 @@ void FragmentProgramDecompiler::SetDst(std::string code, bool append_mask)
std::string dest = AddReg(dst.dest_reg, dst.fp16) + "$m";
if (dst.exp_tex)
{
//TODO
//If exp_tex really sets _bx2 flag, we may need to perform extra modifications to the src
AddCode("//TODO: exp tex flag is set");
}
AddCodeCond(Format(dest), code);
//AddCode("$ifcond " + dest + code + (append_mask ? "$m;" : ";"));
@ -148,6 +159,45 @@ std::string FragmentProgramDecompiler::AddTex()
return m_parr.AddParam(PF_PARAM_UNIFORM, sampler, std::string("tex") + std::to_string(dst.tex_num));
}
//Both of these were tested with a trace SoulCalibur IV title screen
//Failure to catch causes infinite values since theres alot of rcp(0)
std::string FragmentProgramDecompiler::NotZero(const std::string& code)
{
return "(max(abs(" + code + "), 1.E-10) * sign(" + code + "))";
}
std::string FragmentProgramDecompiler::NotZeroPositive(const std::string& code)
{
return "max(" + code + ", 1.E-10)";
}
std::string FragmentProgramDecompiler::NoOverflow(const std::string& code)
{
//FP16 is expected to overflow alot easier at 0+-65504
//FP32 can still work upto 0+-3.4E38
//See http://http.download.nvidia.com/developer/Papers/2005/FP_Specials/FP_Specials.pdf
if (dst.exp_tex)
{
//If dst.exp_tex really is _bx2 postfix, we need to unpack dynamic range
return "((" + code + "- 0.5) * 2.)";
}
switch (dst.prec)
{
case 0:
break;
case 1:
//Disabled: Causes blue output in some games such as persona V and soul calibur IV
//return "clamp(" + code + ", -65504., 65504.)";
break;
case 2:
return "clamp(" + code + ", -1., 1.)";
}
return code;
}
std::string FragmentProgramDecompiler::Format(const std::string& code)
{
const std::pair<std::string, std::function<std::string()>> repl_list[] =
@ -358,10 +408,10 @@ bool FragmentProgramDecompiler::handle_sct(u32 opcode)
switch (opcode)
{
case RSX_FP_OPCODE_ADD: SetDst("($0 + $1)"); return true;
case RSX_FP_OPCODE_DIV: SetDst("($0 / $1.xxxx)"); return true;
case RSX_FP_OPCODE_DIV: SetDst("($0 / " + NotZero("$1.x") + ")"); return true;
// Note: DIVSQ is not IEEE compliant. divsq(0, 0) is 0 (Super Puzzle Fighter II Turbo HD Remix).
// sqrt(x, 0) might be equal to some big value (in absolute) whose sign is sign(x) but it has to be proven.
case RSX_FP_OPCODE_DIVSQ: SetDst("divsq_legacy($0, $1)"); return true;
case RSX_FP_OPCODE_DIVSQ: SetDst("($0 / sqrt(" + NotZeroPositive("$1.x") + "))"); return true;
case RSX_FP_OPCODE_DP2: SetDst(getFunction(FUNCTION::FUNCTION_DP2)); return true;
case RSX_FP_OPCODE_DP3: SetDst(getFunction(FUNCTION::FUNCTION_DP3)); return true;
case RSX_FP_OPCODE_DP4: SetDst(getFunction(FUNCTION::FUNCTION_DP4)); return true;
@ -372,10 +422,10 @@ bool FragmentProgramDecompiler::handle_sct(u32 opcode)
case RSX_FP_OPCODE_MOV: SetDst("$0"); return true;
case RSX_FP_OPCODE_MUL: SetDst("($0 * $1)"); return true;
// Note: It's higly likely that RCP is not IEEE compliant but a game that uses rcp(0) has to be found
case RSX_FP_OPCODE_RCP: SetDst("rcp_legacy($0)"); return true;
case RSX_FP_OPCODE_RCP: SetDst("(1. / " + NotZero("$0") + ")"); return true;
// Note: RSQ is not IEEE compliant. rsq(0) is some big number (Silent Hill 3 HD)
// It is not know what happens if 0 is negative.
case RSX_FP_OPCODE_RSQ: SetDst("rsq_legacy($0)"); return true;
case RSX_FP_OPCODE_RSQ: SetDst("(" + getFloatTypeName(4) + "(1.) / sqrt(" + NotZeroPositive("$0.x") + "))"); return true;
case RSX_FP_OPCODE_SEQ: SetDst(getFloatTypeName(4) + "(" + compareFunction(COMPARE::FUNCTION_SEQ, "$0", "$1") + ")"); return true;
case RSX_FP_OPCODE_SFL: SetDst(getFunction(FUNCTION::FUNCTION_SFL)); return true;
case RSX_FP_OPCODE_SGE: SetDst(getFloatTypeName(4) + "(" + compareFunction(COMPARE::FUNCTION_SGE, "$0", "$1") + ")"); return true;
@ -394,10 +444,10 @@ bool FragmentProgramDecompiler::handle_scb(u32 opcode)
{
case RSX_FP_OPCODE_ADD: SetDst("($0 + $1)"); return true;
case RSX_FP_OPCODE_COS: SetDst("cos($0.xxxx)"); return true;
case RSX_FP_OPCODE_DIV: SetDst("($0 / $1.xxxx)"); return true;
case RSX_FP_OPCODE_DIV: SetDst("($0 / " + NotZero("$1.x") + ")"); return true;
// Note: DIVSQ is not IEEE compliant. sqrt(0, 0) is 0 (Super Puzzle Fighter II Turbo HD Remix).
// sqrt(x, 0) might be equal to some big value (in absolute) whose sign is sign(x) but it has to be proven.
case RSX_FP_OPCODE_DIVSQ: SetDst("divsq_legacy($0, sqrt($1).xxxx)"); return true;
case RSX_FP_OPCODE_DIVSQ: SetDst("($0 / sqrt(" + NotZeroPositive("$1.x") + "))"); return true;
case RSX_FP_OPCODE_DP2: SetDst(getFunction(FUNCTION::FUNCTION_DP2)); return true;
case RSX_FP_OPCODE_DP3: SetDst(getFunction(FUNCTION::FUNCTION_DP3)); return true;
case RSX_FP_OPCODE_DP4: SetDst(getFunction(FUNCTION::FUNCTION_DP4)); return true;
@ -416,10 +466,10 @@ bool FragmentProgramDecompiler::handle_scb(u32 opcode)
case RSX_FP_OPCODE_MIN: SetDst("min($0, $1)"); return true;
case RSX_FP_OPCODE_MOV: SetDst("$0"); return true;
case RSX_FP_OPCODE_MUL: SetDst("($0 * $1)"); return true;
case RSX_FP_OPCODE_PK2: SetDst("float(packSnorm2x16($0.xy))"); return true;
case RSX_FP_OPCODE_PK4: SetDst("float(packSnorm4x8($0))"); return true;
case RSX_FP_OPCODE_PK16: SetDst("float(packHalf2x16($0.xy))"); return true;
case RSX_FP_OPCODE_PKB: SetDst("packUnorm4x8($0 / 255.)"); return true;
case RSX_FP_OPCODE_PK2: SetDst(getFloatTypeName(4) + "(packSnorm2x16($0.xy))"); return true;
case RSX_FP_OPCODE_PK4: SetDst(getFloatTypeName(4) + "(packSnorm4x8($0))"); return true;
case RSX_FP_OPCODE_PK16: SetDst(getFloatTypeName(4) + "(packHalf2x16($0.xy))"); return true;
case RSX_FP_OPCODE_PKB: SetDst(getFloatTypeName(4) + "(packUnorm4x8($0 / 255.))"); return true;
case RSX_FP_OPCODE_PKG: LOG_ERROR(RSX, "Unimplemented SCB instruction: PKG"); return true;
case RSX_FP_OPCODE_SEQ: SetDst(getFloatTypeName(4) + "(" + compareFunction(COMPARE::FUNCTION_SEQ, "$0", "$1") + ")"); return true;
case RSX_FP_OPCODE_SFL: SetDst(getFunction(FUNCTION::FUNCTION_SFL)); return true;
@ -440,8 +490,11 @@ bool FragmentProgramDecompiler::handle_tex_srb(u32 opcode)
{
case RSX_FP_OPCODE_DDX: SetDst(getFunction(FUNCTION::FUNCTION_DFDX)); return true;
case RSX_FP_OPCODE_DDY: SetDst(getFunction(FUNCTION::FUNCTION_DFDY)); return true;
case RSX_FP_OPCODE_NRM: SetDst("normalize($0)"); return true;
case RSX_FP_OPCODE_NRM: SetDst("normalize($0.xyz)"); return true;
case RSX_FP_OPCODE_BEM: LOG_ERROR(RSX, "Unimplemented TEX_SRB instruction: BEM"); return true;
case RSX_FP_OPCODE_TEXBEM:
//treat as TEX for now
LOG_ERROR(RSX, "Unimplemented TEX_SRB instruction: TEXBEM");
case RSX_FP_OPCODE_TEX:
switch (m_prog.get_texture_dimension(dst.tex_num))
{
@ -462,7 +515,9 @@ bool FragmentProgramDecompiler::handle_tex_srb(u32 opcode)
return true;
}
return false;
case RSX_FP_OPCODE_TEXBEM: SetDst("texture($t, $0.xy, $1.x)"); return true;
case RSX_FP_OPCODE_TXPBEM:
//Treat as TXP for now
LOG_ERROR(RSX, "Unimplemented TEX_SRB instruction: TXPBEM");
case RSX_FP_OPCODE_TXP:
switch (m_prog.get_texture_dimension(dst.tex_num))
{
@ -480,7 +535,6 @@ bool FragmentProgramDecompiler::handle_tex_srb(u32 opcode)
return true;
}
return false;
case RSX_FP_OPCODE_TXPBEM: SetDst("textureProj($t, $0.xyz, $1.x)"); return true;
case RSX_FP_OPCODE_TXD:
switch (m_prog.get_texture_dimension(dst.tex_num))
{
@ -498,7 +552,7 @@ bool FragmentProgramDecompiler::handle_tex_srb(u32 opcode)
return true;
}
return false;
case RSX_FP_OPCODE_TXB: SetDst("texture($t, $0.xy, $1.x)"); return true;
case RSX_FP_OPCODE_TXB:
case RSX_FP_OPCODE_TXL:
switch (m_prog.get_texture_dimension(dst.tex_num))
{