diff options
Diffstat (limited to 'src/video_core/shader')
| -rw-r--r-- | src/video_core/shader/shader_interpreter.cpp | 36 | ||||
| -rw-r--r-- | src/video_core/shader/shader_jit_x64.cpp | 52 | ||||
| -rw-r--r-- | src/video_core/shader/shader_jit_x64.h | 5 |
3 files changed, 91 insertions, 2 deletions
diff --git a/src/video_core/shader/shader_interpreter.cpp b/src/video_core/shader/shader_interpreter.cpp index e14de0768..646171a19 100644 --- a/src/video_core/shader/shader_interpreter.cpp +++ b/src/video_core/shader/shader_interpreter.cpp @@ -334,6 +334,42 @@ void RunInterpreter(UnitState<Debug>& state) { Record<DebugDataRecord::CMP_RESULT>(state.debug, iteration, state.conditional_code); break; + case OpCode::Id::EX2: + { + Record<DebugDataRecord::SRC1>(state.debug, iteration, src1); + Record<DebugDataRecord::DEST_IN>(state.debug, iteration, dest); + + // EX2 only takes first component exp2 and writes it to all dest components + float24 ex2_res = float24::FromFloat32(std::exp2(src1[0].ToFloat32())); + for (int i = 0; i < 4; ++i) { + if (!swizzle.DestComponentEnabled(i)) + continue; + + dest[i] = ex2_res; + } + + Record<DebugDataRecord::DEST_OUT>(state.debug, iteration, dest); + break; + } + + case OpCode::Id::LG2: + { + Record<DebugDataRecord::SRC1>(state.debug, iteration, src1); + Record<DebugDataRecord::DEST_IN>(state.debug, iteration, dest); + + // LG2 only takes the first component log2 and writes it to all dest components + float24 lg2_res = float24::FromFloat32(std::log2(src1[0].ToFloat32())); + for (int i = 0; i < 4; ++i) { + if (!swizzle.DestComponentEnabled(i)) + continue; + + dest[i] = lg2_res; + } + + Record<DebugDataRecord::DEST_OUT>(state.debug, iteration, dest); + break; + } + default: LOG_ERROR(HW_GPU, "Unhandled arithmetic instruction: 0x%02x (%s): 0x%08x", (int)instr.opcode.Value().EffectiveOpCode(), instr.opcode.Value().GetInfo().name, instr.hex); diff --git a/src/video_core/shader/shader_jit_x64.cpp b/src/video_core/shader/shader_jit_x64.cpp index 836942c6b..e4b8295b3 100644 --- a/src/video_core/shader/shader_jit_x64.cpp +++ b/src/video_core/shader/shader_jit_x64.cpp @@ -25,8 +25,8 @@ const JitFunction instr_table[64] = { &JitCompiler::Compile_DP4, // dp4 nullptr, // dph nullptr, // unknown - nullptr, // ex2 - nullptr, // lg2 + &JitCompiler::Compile_EX2, // ex2 + &JitCompiler::Compile_LG2, // lg2 nullptr, // unknown &JitCompiler::Compile_MUL, // mul nullptr, // lge @@ -280,6 +280,22 @@ void JitCompiler::Compile_UniformCondition(Instruction instr) { CMP(sizeof(bool) * 8, MDisp(UNIFORMS, offset), Imm8(0)); } +void JitCompiler::Compile_PushCallerSavedXMM() { +#ifndef _WIN32 + SUB(64, R(RSP), Imm8(2 * 16)); + MOVUPS(MDisp(RSP, 16), ONE); + MOVUPS(MDisp(RSP, 0), NEGBIT); +#endif +} + +void JitCompiler::Compile_PopCallerSavedXMM() { +#ifndef _WIN32 + MOVUPS(NEGBIT, MDisp(RSP, 0)); + MOVUPS(ONE, MDisp(RSP, 16)); + ADD(64, R(RSP), Imm8(2 * 16)); +#endif +} + void JitCompiler::Compile_ADD(Instruction instr) { Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1); Compile_SwizzleSrc(instr, 2, instr.common.src2, SRC2); @@ -331,6 +347,38 @@ void JitCompiler::Compile_DP4(Instruction instr) { Compile_DestEnable(instr, SRC1); } +void JitCompiler::Compile_EX2(Instruction instr) { + Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1); + MOVSS(XMM0, R(SRC1)); + + // The following will actually break the stack alignment + ABI_PushAllCallerSavedRegsAndAdjustStack(); + Compile_PushCallerSavedXMM(); + ABI_CallFunction(reinterpret_cast<const void*>(exp2f)); + Compile_PopCallerSavedXMM(); + ABI_PopAllCallerSavedRegsAndAdjustStack(); + + SHUFPS(XMM0, R(XMM0), _MM_SHUFFLE(0, 0, 0, 0)); + MOVAPS(SRC1, R(XMM0)); + Compile_DestEnable(instr, SRC1); +} + +void JitCompiler::Compile_LG2(Instruction instr) { + Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1); + MOVSS(XMM0, R(SRC1)); + + // The following will actually break the stack alignment + ABI_PushAllCallerSavedRegsAndAdjustStack(); + Compile_PushCallerSavedXMM(); + ABI_CallFunction(reinterpret_cast<const void*>(log2f)); + Compile_PopCallerSavedXMM(); + ABI_PopAllCallerSavedRegsAndAdjustStack(); + + SHUFPS(XMM0, R(XMM0), _MM_SHUFFLE(0, 0, 0, 0)); + MOVAPS(SRC1, R(XMM0)); + Compile_DestEnable(instr, SRC1); +} + void JitCompiler::Compile_MUL(Instruction instr) { Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1); Compile_SwizzleSrc(instr, 2, instr.common.src2, SRC2); diff --git a/src/video_core/shader/shader_jit_x64.h b/src/video_core/shader/shader_jit_x64.h index b88f2a0d2..a6ae7fbf1 100644 --- a/src/video_core/shader/shader_jit_x64.h +++ b/src/video_core/shader/shader_jit_x64.h @@ -37,6 +37,8 @@ public: void Compile_ADD(Instruction instr); void Compile_DP3(Instruction instr); void Compile_DP4(Instruction instr); + void Compile_EX2(Instruction instr); + void Compile_LG2(Instruction instr); void Compile_MUL(Instruction instr); void Compile_FLR(Instruction instr); void Compile_MAX(Instruction instr); @@ -67,6 +69,9 @@ private: void Compile_EvaluateCondition(Instruction instr); void Compile_UniformCondition(Instruction instr); + void Compile_PushCallerSavedXMM(); + void Compile_PopCallerSavedXMM(); + /// Pointer to the variable that stores the current Pica code offset. Used to handle nested code blocks. unsigned* offset_ptr = nullptr; |