From dbaebd8582c33807ca25acbf36cbd90587c9cfa9 Mon Sep 17 00:00:00 2001 From: ReinUsesLisp Date: Mon, 20 Apr 2020 03:24:00 -0300 Subject: decode/arithmetic_half: Fix HADD2 and HMUL2 absolute and negation bits The encoding for negation and absolute value was wrong. Extracting is now done manually. Similar instructions having different encodings is the rule, not the exception. To keep sanity and readability I preferred to extract the desired bit manually. This is implemented against nxas: https://github.com/ReinUsesLisp/nxas/blob/8dbc38995711cc12206aa370145a3a02665fd989/table.h#L68 That is itself tested against nvdisasm (Nvidia's official disassembler). --- src/video_core/shader/decode/arithmetic_half.cpp | 51 +++++++++++++++++------- 1 file changed, 37 insertions(+), 14 deletions(-) (limited to 'src/video_core/shader/decode') diff --git a/src/video_core/shader/decode/arithmetic_half.cpp b/src/video_core/shader/decode/arithmetic_half.cpp index ee7d9a29d..a276aee44 100644 --- a/src/video_core/shader/decode/arithmetic_half.cpp +++ b/src/video_core/shader/decode/arithmetic_half.cpp @@ -19,22 +19,46 @@ u32 ShaderIR::DecodeArithmeticHalf(NodeBlock& bb, u32 pc) { const Instruction instr = {program_code[pc]}; const auto opcode = OpCode::Decode(instr); - if (opcode->get().GetId() == OpCode::Id::HADD2_C || - opcode->get().GetId() == OpCode::Id::HADD2_R) { + bool negate_a = false; + bool negate_b = false; + bool absolute_a = false; + bool absolute_b = false; + + switch (opcode->get().GetId()) { + case OpCode::Id::HADD2_R: if (instr.alu_half.ftz == 0) { LOG_DEBUG(HW_GPU, "{} without FTZ is not implemented", opcode->get().GetName()); } + negate_a = ((instr.value >> 43) & 1) != 0; + negate_b = ((instr.value >> 31) & 1) != 0; + absolute_a = ((instr.value >> 44) & 1) != 0; + absolute_b = ((instr.value >> 30) & 1) != 0; + break; + case OpCode::Id::HADD2_C: + if (instr.alu_half.ftz == 0) { + LOG_DEBUG(HW_GPU, "{} without FTZ is not implemented", opcode->get().GetName()); + } + negate_a = ((instr.value >> 43) & 1) != 0; + negate_b = ((instr.value >> 56) & 1) != 0; + absolute_a = ((instr.value >> 44) & 1) != 0; + absolute_b = ((instr.value >> 54) & 1) != 0; + break; + case OpCode::Id::HMUL2_R: + negate_a = ((instr.value >> 43) & 1) != 0; + absolute_a = ((instr.value >> 44) & 1) != 0; + absolute_b = ((instr.value >> 30) & 1) != 0; + break; + case OpCode::Id::HMUL2_C: + negate_b = ((instr.value >> 31) & 1) != 0; + absolute_a = ((instr.value >> 44) & 1) != 0; + absolute_b = ((instr.value >> 54) & 1) != 0; + break; } - const bool negate_a = - opcode->get().GetId() != OpCode::Id::HMUL2_R && instr.alu_half.negate_a != 0; - const bool negate_b = - opcode->get().GetId() != OpCode::Id::HMUL2_C && instr.alu_half.negate_b != 0; - Node op_a = UnpackHalfFloat(GetRegister(instr.gpr8), instr.alu_half.type_a); - op_a = GetOperandAbsNegHalf(op_a, instr.alu_half.abs_a, negate_a); + op_a = GetOperandAbsNegHalf(op_a, absolute_a, negate_a); - auto [type_b, op_b] = [&]() -> std::tuple { + auto [type_b, op_b] = [this, instr, opcode]() -> std::pair { switch (opcode->get().GetId()) { case OpCode::Id::HADD2_C: case OpCode::Id::HMUL2_C: @@ -48,17 +72,16 @@ u32 ShaderIR::DecodeArithmeticHalf(NodeBlock& bb, u32 pc) { } }(); op_b = UnpackHalfFloat(op_b, type_b); - // redeclaration to avoid a bug in clang with reusing local bindings in lambdas - Node op_b_alt = GetOperandAbsNegHalf(op_b, instr.alu_half.abs_b, negate_b); + op_b = GetOperandAbsNegHalf(op_b, absolute_b, negate_b); - Node value = [&]() { + Node value = [this, opcode, op_a, op_b = op_b] { switch (opcode->get().GetId()) { case OpCode::Id::HADD2_C: case OpCode::Id::HADD2_R: - return Operation(OperationCode::HAdd, PRECISE, op_a, op_b_alt); + return Operation(OperationCode::HAdd, PRECISE, op_a, op_b); case OpCode::Id::HMUL2_C: case OpCode::Id::HMUL2_R: - return Operation(OperationCode::HMul, PRECISE, op_a, op_b_alt); + return Operation(OperationCode::HMul, PRECISE, op_a, op_b); default: UNIMPLEMENTED_MSG("Unhandled half float instruction: {}", opcode->get().GetName()); return Immediate(0); -- cgit v1.2.3