diff options
author | bunnei <bunneidev@gmail.com> | 2018-10-28 18:06:21 +0100 |
---|---|---|
committer | GitHub <noreply@github.com> | 2018-10-28 18:06:21 +0100 |
commit | aa1cf608ed20984c410fc215d9f73937abe76ddc (patch) | |
tree | 1167c5b03eb2526f2704195a27d50f11430f583c /src/video_core | |
parent | Merge pull request #1606 from FearlessTobi/revert-1581-macosx-target-version (diff) | |
parent | Refactor precise usage and add FMNMX, MUFU, FMUL32 and FADD332 (diff) | |
download | yuzu-aa1cf608ed20984c410fc215d9f73937abe76ddc.tar yuzu-aa1cf608ed20984c410fc215d9f73937abe76ddc.tar.gz yuzu-aa1cf608ed20984c410fc215d9f73937abe76ddc.tar.bz2 yuzu-aa1cf608ed20984c410fc215d9f73937abe76ddc.tar.lz yuzu-aa1cf608ed20984c410fc215d9f73937abe76ddc.tar.xz yuzu-aa1cf608ed20984c410fc215d9f73937abe76ddc.tar.zst yuzu-aa1cf608ed20984c410fc215d9f73937abe76ddc.zip |
Diffstat (limited to 'src/video_core')
-rw-r--r-- | src/video_core/renderer_opengl/gl_shader_decompiler.cpp | 55 |
1 files changed, 35 insertions, 20 deletions
diff --git a/src/video_core/renderer_opengl/gl_shader_decompiler.cpp b/src/video_core/renderer_opengl/gl_shader_decompiler.cpp index dec291a7d..dcf6941b0 100644 --- a/src/video_core/renderer_opengl/gl_shader_decompiler.cpp +++ b/src/video_core/renderer_opengl/gl_shader_decompiler.cpp @@ -341,10 +341,10 @@ public: */ void SetRegisterToFloat(const Register& reg, u64 elem, const std::string& value, u64 dest_num_components, u64 value_num_components, - bool is_saturated = false, u64 dest_elem = 0) { + bool is_saturated = false, u64 dest_elem = 0, bool precise = false) { SetRegister(reg, elem, is_saturated ? "clamp(" + value + ", 0.0, 1.0)" : value, - dest_num_components, value_num_components, dest_elem); + dest_num_components, value_num_components, dest_elem, precise); } /** @@ -368,7 +368,7 @@ public: const std::string func{is_signed ? "intBitsToFloat" : "uintBitsToFloat"}; SetRegister(reg, elem, func + '(' + ConvertIntegerSize(value, size) + ')', - dest_num_components, value_num_components, dest_elem); + dest_num_components, value_num_components, dest_elem, false); if (sets_cc) { const std::string zero_condition = "( " + ConvertIntegerSize(value, size) + " == 0 )"; @@ -416,7 +416,7 @@ public: } }(); - SetRegister(reg, elem, result, dest_num_components, value_num_components, dest_elem); + SetRegister(reg, elem, result, dest_num_components, value_num_components, dest_elem, false); } /** @@ -757,7 +757,8 @@ private: * @param dest_elem Optional, the destination element to use for the operation. */ void SetRegister(const Register& reg, u64 elem, const std::string& value, - u64 dest_num_components, u64 value_num_components, u64 dest_elem) { + u64 dest_num_components, u64 value_num_components, u64 dest_elem, + bool precise) { if (reg == Register::ZeroIndex) { LOG_CRITICAL(HW_GPU, "Cannot set Register::ZeroIndex"); UNREACHABLE(); @@ -774,7 +775,18 @@ private: src += GetSwizzle(elem); } - shader.AddLine(dest + " = " + src + ';'); + if (precise && stage != Maxwell3D::Regs::ShaderStage::Fragment) { + shader.AddLine('{'); + ++shader.scope; + // This avoids optimizations of constant propagation and keeps the code as the original + // Sadly using the precise keyword causes "linking" errors on fragment shaders. + shader.AddLine("precise float tmp = " + src + ';'); + shader.AddLine(dest + " = tmp;"); + --shader.scope; + shader.AddLine('}'); + } else { + shader.AddLine(dest + " = " + src + ';'); + } } /// Build the GLSL register list. @@ -1510,8 +1522,9 @@ private: ASSERT_MSG(instr.fmul.cc == 0, "FMUL cc is not implemented"); op_b = GetOperandAbsNeg(op_b, false, instr.fmul.negate_b); + regs.SetRegisterToFloat(instr.gpr0, 0, op_a + " * " + op_b, 1, 1, - instr.alu.saturate_d); + instr.alu.saturate_d, 0, true); break; } case OpCode::Id::FADD_C: @@ -1519,8 +1532,9 @@ private: case OpCode::Id::FADD_IMM: { op_a = GetOperandAbsNeg(op_a, instr.alu.abs_a, instr.alu.negate_a); op_b = GetOperandAbsNeg(op_b, instr.alu.abs_b, instr.alu.negate_b); + regs.SetRegisterToFloat(instr.gpr0, 0, op_a + " + " + op_b, 1, 1, - instr.alu.saturate_d); + instr.alu.saturate_d, 0, true); break; } case OpCode::Id::MUFU: { @@ -1528,31 +1542,31 @@ private: switch (instr.sub_op) { case SubOp::Cos: regs.SetRegisterToFloat(instr.gpr0, 0, "cos(" + op_a + ')', 1, 1, - instr.alu.saturate_d); + instr.alu.saturate_d, 0, true); break; case SubOp::Sin: regs.SetRegisterToFloat(instr.gpr0, 0, "sin(" + op_a + ')', 1, 1, - instr.alu.saturate_d); + instr.alu.saturate_d, 0, true); break; case SubOp::Ex2: regs.SetRegisterToFloat(instr.gpr0, 0, "exp2(" + op_a + ')', 1, 1, - instr.alu.saturate_d); + instr.alu.saturate_d, 0, true); break; case SubOp::Lg2: regs.SetRegisterToFloat(instr.gpr0, 0, "log2(" + op_a + ')', 1, 1, - instr.alu.saturate_d); + instr.alu.saturate_d, 0, true); break; case SubOp::Rcp: regs.SetRegisterToFloat(instr.gpr0, 0, "1.0 / " + op_a, 1, 1, - instr.alu.saturate_d); + instr.alu.saturate_d, 0, true); break; case SubOp::Rsq: regs.SetRegisterToFloat(instr.gpr0, 0, "inversesqrt(" + op_a + ')', 1, 1, - instr.alu.saturate_d); + instr.alu.saturate_d, 0, true); break; case SubOp::Sqrt: regs.SetRegisterToFloat(instr.gpr0, 0, "sqrt(" + op_a + ')', 1, 1, - instr.alu.saturate_d); + instr.alu.saturate_d, 0, true); break; default: LOG_CRITICAL(HW_GPU, "Unhandled MUFU sub op: {0:x}", @@ -1573,7 +1587,7 @@ private: regs.SetRegisterToFloat(instr.gpr0, 0, '(' + condition + ") ? min(" + parameters + ") : max(" + parameters + ')', - 1, 1); + 1, 1, false, 0, true); break; } case OpCode::Id::RRO_C: @@ -1602,7 +1616,7 @@ private: regs.SetRegisterToFloat(instr.gpr0, 0, regs.GetRegisterAsFloat(instr.gpr8) + " * " + GetImmediate32(instr), - 1, 1, instr.fmul32.saturate); + 1, 1, instr.fmul32.saturate, 0, true); break; } case OpCode::Id::FADD32I: { @@ -1625,7 +1639,7 @@ private: op_b = "-(" + op_b + ')'; } - regs.SetRegisterToFloat(instr.gpr0, 0, op_a + " + " + op_b, 1, 1); + regs.SetRegisterToFloat(instr.gpr0, 0, op_a + " + " + op_b, 1, 1, false, 0, true); break; } } @@ -2087,8 +2101,9 @@ private: } } - regs.SetRegisterToFloat(instr.gpr0, 0, op_a + " * " + op_b + " + " + op_c, 1, 1, - instr.alu.saturate_d); + regs.SetRegisterToFloat(instr.gpr0, 0, "fma(" + op_a + ", " + op_b + ", " + op_c + ')', + 1, 1, instr.alu.saturate_d, 0, true); + break; } case OpCode::Type::Hfma2: { |