rsx/fp: Improve rgister component gather detection

- Also avoids clobbering register data by keeping gathered bits in a temp var
This commit is contained in:
kd-11 2018-02-23 22:48:51 +03:00
parent 87741141f1
commit 68b3229756
2 changed files with 48 additions and 34 deletions

View File

@ -111,7 +111,7 @@ void FragmentProgramDecompiler::SetDst(std::string code, bool append_mask)
} }
u32 reg_index = dst.fp16 ? dst.dest_reg >> 1 : dst.dest_reg; u32 reg_index = dst.fp16 ? dst.dest_reg >> 1 : dst.dest_reg;
temp_registers[reg_index].tag(dst.dest_reg, !!dst.fp16); temp_registers[reg_index].tag(dst.dest_reg, !!dst.fp16, dst.mask_x, dst.mask_y, dst.mask_z, dst.mask_w);
} }
void FragmentProgramDecompiler::AddFlowOp(std::string code) void FragmentProgramDecompiler::AddFlowOp(std::string code)
@ -283,14 +283,15 @@ std::string FragmentProgramDecompiler::Format(const std::string& code, bool igno
{ "$_i", [this]() -> std::string {return std::to_string(dst.tex_num);} }, { "$_i", [this]() -> std::string {return std::to_string(dst.tex_num);} },
{ "$m", std::bind(std::mem_fn(&FragmentProgramDecompiler::GetMask), this) }, { "$m", std::bind(std::mem_fn(&FragmentProgramDecompiler::GetMask), this) },
{ "$ifcond ", [this]() -> std::string { "$ifcond ", [this]() -> std::string
{ {
const std::string& cond = GetCond(); const std::string& cond = GetCond();
if (cond == "true") return ""; if (cond == "true") return "";
return "if(" + cond + ") "; return "if(" + cond + ") ";
} }
}, },
{ "$cond", std::bind(std::mem_fn(&FragmentProgramDecompiler::GetCond), this) }, { "$cond", std::bind(std::mem_fn(&FragmentProgramDecompiler::GetCond), this) },
{ "$_c", std::bind(std::mem_fn(&FragmentProgramDecompiler::AddConst), this) } { "$_c", std::bind(std::mem_fn(&FragmentProgramDecompiler::AddConst), this) },
{ "$float4", [this]() -> std::string { return getFloatTypeName(4); } }
}; };
if (!ignore_redirects) if (!ignore_redirects)
@ -407,20 +408,13 @@ template<typename T> std::string FragmentProgramDecompiler::GetSRC(T src)
dst.opcode == RSX_FP_OPCODE_UPB || dst.opcode == RSX_FP_OPCODE_UPB ||
dst.opcode == RSX_FP_OPCODE_UPG) dst.opcode == RSX_FP_OPCODE_UPG)
{ {
//TODO: Implement aliased gather for half floats
bool xy_read = false;
bool zw_read = false;
if (src.swizzle_x < 2 || src.swizzle_y < 2 || src.swizzle_z < 2 || src.swizzle_w < 2)
xy_read = true;
if (src.swizzle_x > 1 || src.swizzle_y > 1 || src.swizzle_z > 1 || src.swizzle_w > 1)
zw_read = true;
auto &reg = temp_registers[src.tmp_reg_index]; auto &reg = temp_registers[src.tmp_reg_index];
if (reg.requires_gather(xy_read, zw_read)) if (reg.requires_gather(src.swizzle_x))
{ {
properties.has_gather_op = true; properties.has_gather_op = true;
AddCode(reg.gather_r()); AddReg(src.tmp_reg_index, src.fp16);
ret = getFloatTypeName(4) + reg.gather_r();
break;
} }
} }
} }

View File

@ -24,25 +24,41 @@ class FragmentProgramDecompiler
bool aliased_r0 = false; bool aliased_r0 = false;
bool aliased_h0 = false; bool aliased_h0 = false;
bool aliased_h1 = false; bool aliased_h1 = false;
bool last_write_half = false; bool last_write_half[4] = { false, false, false, false };
u32 real_index = UINT32_MAX; u32 real_index = UINT32_MAX;
void tag(u32 index, bool half_register) void tag(u32 index, bool half_register, bool x, bool y, bool z, bool w)
{ {
if (half_register) if (half_register)
{ {
last_write_half = true;
if (index & 1) if (index & 1)
{
if (x) last_write_half[2] = true;
if (y) last_write_half[2] = true;
if (z) last_write_half[3] = true;
if (w) last_write_half[3] = true;
aliased_h1 = true; aliased_h1 = true;
}
else else
{
if (x) last_write_half[0] = true;
if (y) last_write_half[0] = true;
if (z) last_write_half[1] = true;
if (w) last_write_half[1] = true;
aliased_h0 = true; aliased_h0 = true;
}
} }
else else
{ {
if (x) last_write_half[0] = false;
if (y) last_write_half[1] = false;
if (z) last_write_half[2] = false;
if (w) last_write_half[3] = false;
aliased_r0 = true; aliased_r0 = true;
last_write_half = false;
} }
if (real_index == UINT32_MAX) if (real_index == UINT32_MAX)
@ -54,12 +70,19 @@ class FragmentProgramDecompiler
} }
} }
bool requires_gather(bool xy, bool zw) const bool requires_gather(u8 channel) const
{ {
//Data fetched from the single precision register requires merging of the two half registers //Data fetched from the single precision register requires merging of the two half registers
//TODO: Check individual swizzle channels verify(HERE), channel < 4;
if ((aliased_h0 && xy) || (aliased_h1 && zw)) if (aliased_h0 && channel < 2)
return last_write_half; {
return last_write_half[channel];
}
if (aliased_h1 && channel > 1)
{
return last_write_half[channel];
}
return false; return false;
} }
@ -67,7 +90,7 @@ class FragmentProgramDecompiler
bool requires_split(u32 /*index*/) const bool requires_split(u32 /*index*/) const
{ {
//Data fetched from any of the two half registers requires sync with the full register //Data fetched from any of the two half registers requires sync with the full register
if (!last_write_half && aliased_r0) if (!(last_write_half[0] || last_write_half[1]) && aliased_r0)
{ {
//r0 has been written to //r0 has been written to
//TODO: Check for specific elements in real32 register //TODO: Check for specific elements in real32 register
@ -85,15 +108,12 @@ class FragmentProgramDecompiler
std::string ret = "//Invalid gather"; std::string ret = "//Invalid gather";
if (aliased_h0 && aliased_h1) if (aliased_h0 && aliased_h1)
ret = reg + " = gather(" + h0 + ", " + h1 + ");"; ret = "(gather(" + h0 + ", " + h1 + "))";
else if (aliased_h0) else if (aliased_h0)
ret = reg + ".xy = gather(" + h0 + ");"; ret = "(gather(" + h0 + "), " + reg + ".zw)";
else if (aliased_h1) else if (aliased_h1)
ret = reg + ".zw = gather(" + h1 + ");"; ret = "(" + reg + ".xy, gather(" + h1 + "))";
last_write_half = false;
aliased_h0 = false;
aliased_h1 = false;
return ret; return ret;
} }
}; };