Diamond75/Panda3DS
1
0
Fork 0
mirror of https://github.com/wheremyfoodat/Panda3DS.git synced 2025-04-17 11:11:30 +12:00
Code Issues Projects Releases Packages Wiki Activity Actions

[Shader JIT] Attempt to implement FLR and CALLU

Browse source
This commit is contained in:
wheremyfoodat 2023-06-11 23:51:45 +03:00
parent 137b652840
commit 3f2b713c4b
1 changed files with 34 additions and 0 deletions
Download patch file Download diff file Expand all files Collapse all files

34
src/core/PICA/dynapica/shader_rec_emitter_x64.cpp
View file

@ -5,6 +5,8 @@
#include <bit> #include <bit>
#include <cassert> #include <cassert>
#include <cstddef> #include <cstddef>
#include <immintrin.h>
#include <smmintrin.h>
using namespace Xbyak; using namespace Xbyak;
using namespace Xbyak::util; using namespace Xbyak::util;
@ -115,12 +117,16 @@ void ShaderEmitter::compileInstruction(const PICAShader& shaderUnit) {
case ShaderOpcodes::CALLC: case ShaderOpcodes::CALLC:
recCALLC(shaderUnit, instruction); recCALLC(shaderUnit, instruction);
break; break;
case ShaderOpcodes::CALLU:
recCALLU(shaderUnit, instruction);
break;
case ShaderOpcodes::CMP1: case ShaderOpcodes::CMP2: case ShaderOpcodes::CMP1: case ShaderOpcodes::CMP2:
recCMP(shaderUnit, instruction); recCMP(shaderUnit, instruction);
break; break;
case ShaderOpcodes::DP3: recDP3(shaderUnit, instruction); break; case ShaderOpcodes::DP3: recDP3(shaderUnit, instruction); break;
case ShaderOpcodes::DP4: recDP4(shaderUnit, instruction); break; case ShaderOpcodes::DP4: recDP4(shaderUnit, instruction); break;
case ShaderOpcodes::END: recEND(shaderUnit, instruction); break; case ShaderOpcodes::END: recEND(shaderUnit, instruction); break;
case ShaderOpcodes::FLR: recFLR(shaderUnit, instruction); break;
case ShaderOpcodes::IFC: recIFC(shaderUnit, instruction); break; case ShaderOpcodes::IFC: recIFC(shaderUnit, instruction); break;
case ShaderOpcodes::IFU: recIFU(shaderUnit, instruction); break; case ShaderOpcodes::IFU: recIFU(shaderUnit, instruction); break;
case ShaderOpcodes::JMPC: recJMPC(shaderUnit, instruction); break; case ShaderOpcodes::JMPC: recJMPC(shaderUnit, instruction); break;
@ -372,6 +378,23 @@ void ShaderEmitter::recMOV(const PICAShader& shader, u32 instruction) {
storeRegister(src1_xmm, shader, dest, operandDescriptor); storeRegister(src1_xmm, shader, dest, operandDescriptor);
} }
void ShaderEmitter::recFLR(const PICAShader& shader, u32 instruction) {
const u32 operandDescriptor = shader.operandDescriptors[instruction & 0x7f];
const u32 src = getBits<12, 7>(instruction);
const u32 idx = getBits<19, 2>(instruction);
const u32 dest = getBits<21, 5>(instruction);
loadRegister<1>(src1_xmm, shader, src, idx, operandDescriptor); // Load source 1 into scratch1
if (haveSSE4_1) {
roundps(src1_xmm, src1_xmm, _MM_FROUND_FLOOR);
} else {
cvttps2dq(src1_xmm, src1_xmm); // Truncate and convert to integer
cvtdq2ps(src1_xmm, src1_xmm); // Convert from integer back to float
}
storeRegister(src1_xmm, shader, dest, operandDescriptor);
}
void ShaderEmitter::recMOVA(const PICAShader& shader, u32 instruction) { void ShaderEmitter::recMOVA(const PICAShader& shader, u32 instruction) {
const u32 operandDescriptor = shader.operandDescriptors[instruction & 0x7f]; const u32 operandDescriptor = shader.operandDescriptors[instruction & 0x7f];
const u32 src = getBits<12, 7>(instruction); const u32 src = getBits<12, 7>(instruction);
@ -702,6 +725,17 @@ void ShaderEmitter::recCALLC(const PICAShader& shader, u32 instruction) {
L(skipCall); L(skipCall);
} }
void ShaderEmitter::recCALLU(const PICAShader& shader, u32 instruction) {
Label skipCall;
// z is 0 if the call should be taken, 1 otherwise
checkBoolUniform(shader, instruction);
jz(skipCall);
recCALL(shader, instruction);
L(skipCall);
}
void ShaderEmitter::recJMPC(const PICAShader& shader, u32 instruction) { void ShaderEmitter::recJMPC(const PICAShader& shader, u32 instruction) {
const u32 dest = getBits<10, 12>(instruction); const u32 dest = getBits<10, 12>(instruction);