/*========================== begin_copyright_notice ============================ Copyright (C) 2017-2021 Intel Corporation SPDX-License-Identifier: MIT ============================= end_copyright_notice ===========================*/ #ifndef _IGA_BACKEND_GED_ENCODER_H_ #define _IGA_BACKEND_GED_ENCODER_H_ #include "../../ErrorHandler.hpp" #include "../../IR/Block.hpp" #include "../../IR/Instruction.hpp" #include "../../IR/Kernel.hpp" #include "../../Timer/Timer.hpp" #include "../EncoderOpts.hpp" #include "GEDBitProcessor.hpp" #include "IGAToGEDTranslation.hpp" #include #include #include namespace iga { #define GED_ENCODE(FUNC, ARG) GED_ENCODE_TO(FUNC, ARG, &m_gedInst) #if defined(GED_TIMER) || defined(_DEBUG) #define START_GED_TIMER() startIGATimer(TIMER_GED) #define STOP_GED_TIMER() stopIGATimer(TIMER_GED) #else #define START_GED_TIMER() #define STOP_GED_TIMER() #endif #if defined(TOTAL_ENCODE_TIMER) || defined(_DEBUG) #define START_ENCODER_TIMER() startIGATimer(TIMER_TOTAL) #define STOP_ENCODER_TIMER() stopIGATimer(TIMER_TOTAL) #else #define START_ENCODER_TIMER() #define STOP_ENCODER_TIMER() #endif // uncomment to emit all the GED calls // #define TRACE_GED_CALLS #ifdef TRACE_GED_CALLS #define TRACE_GED_SETTER(FIELD, ARG, STATUS) \ std::cerr << "Encoder.cpp:" << __LINE__ << ": GED_Set" << #FIELD << "(..," \ << (ARG) << ") results in " << (STATUS) << "\n" #else #define TRACE_GED_SETTER(FIELD, ARG, STATUS) #endif #define GED_ENCODE_TO(FIELD, ARG, GED) \ do { \ GED_RETURN_VALUE _status; \ START_GED_TIMER(); \ _status = GED_Set##FIELD(GED, ARG); \ STOP_GED_TIMER(); \ TRACE_GED_SETTER(FIELD, ARG, _status); \ if (_status != GED_RETURN_VALUE_SUCCESS) { \ handleGedError(__LINE__, #FIELD, _status); \ } \ } while (0) using BlockList = std::list>; class Encoder : protected GEDBitProcessor { public: Encoder(const Model &model, ErrorHandler &eh, const EncoderOpts &eos); void encodeKernel(Kernel &k, MemManager &m, void *&bits, uint32_t &bitsLen); size_t getNumInstructionsEncoded() const; /////////////////////////////////////////////////////////////////////// // PROFILING API FOR TESTING (must be compiled into) // // compile with MEASURE_COMPILATION_TIME double getElapsedTimeMicros(unsigned int idx); int64_t getElapsedTimeTicks(unsigned int idx); std::string getTimerName(unsigned int idx); protected: // TODO: phase these out void encodeKernelPreProcess(Kernel &k); void doEncodeKernelPreProcess(Kernel &k); void encodeFC(const Instruction &i); void encodeTernaryInstruction(const Instruction &inst, GED_ACCESS_MODE accessMode); void encodeTernaryAlign1Instruction(const Instruction &inst); void encodeTernaryAlign16Instruction(const Instruction &inst); void encodeQuinaryInstruction(const Instruction &inst); void encodeSendDirectDestination(const Operand &dst); void encodeSendDestinationDataType(const Operand &dst); void encodeOptionsThreadControl(const Instruction &inst); protected: void encodeDstReg(RegName regName, uint16_t regNum); // just the regnum void encodeImmVal(const ImmVal &val, Type type); template void encodeSrcRegFile(GED_REG_FILE rf); template void encodeSrcRegionVert(Region::Vert v); template void encodeSrcType(Type t); template void encodeSrcAddrMode(GED_ADDR_MODE x); template void encodeSrcModifier(SrcModifier x); template void encodeSrcSubRegNum(uint32_t subRegInByte); template void encodeSrcMathMacroReg(MathMacroExt a); template void encodeSrcReg(RegName regName, uint16_t regNum); template void encodeSrcAddrImm(int32_t addrImm); template void encodeSrcAddrSubRegNum(uint32_t addrSubReg); template void encodeSrcRegion(const Region &r, bool hasRgnWi = true); template void encodeSrcRegionWidth(Region::Width w); template void encodeSrcReducedRegion(const Region& rgn); template void encodeSrcRepCtrl(GED_REP_CTRL rep); template void encodeSrcChanSel(GED_SWIZZLE chSelX, GED_SWIZZLE chSelY, GED_SWIZZLE chSelZ, GED_SWIZZLE chSelW); template void encodeTernarySourceAlign16(const Instruction &inst); void encodeTernaryDestinationAlign16(const Instruction &inst); template void encodeTernarySourceAlign1(const Instruction &inst); template void encodeQuinarySource(const Instruction &inst); void encodeTernaryDestinationAlign1(const Instruction &inst); void encodeTernarySrcRegionVert(SourceIndex S, Region::Vert v); template void encodeTernaryImmVal(const ImmVal &val, Type type); template void encodeSrcRegionHorz(Region::Horz s); void handleGedError(int line, const char *setter, GED_RETURN_VALUE status); // state that is valid over instance life EncoderOpts m_opts; // state that is valid over encodeInst() ged_ins_t m_gedInst; bool m_encodeAlign16 = false; Op m_opcode = Op::INVALID; size_t m_numberInstructionsEncoded; private: void operator delete(void *, MemManager *){}; void *operator new(size_t sz, MemManager *m) { return m->alloc(sz); }; void encodeBlock(Kernel &k, Block *blk); void encodeInstruction(Instruction &inst); void patchJumpOffsets(); /////////////////////////////////////////////////////////////////////// // BASIC INSTRUCTIONS /////////////////////////////////////////////////////////////////////// void encodeBasicInstruction(const Instruction &inst, GED_ACCESS_MODE accessMode); void encodeBasicDestination(const Instruction &inst, const Operand &dst, GED_ACCESS_MODE accessMode = GED_ACCESS_MODE_Align1); template void encodeBasicSource(const Instruction &inst, const Operand &src, GED_ACCESS_MODE accessMode = GED_ACCESS_MODE_Align1); /////////////////////////////////////////////////////////////////////// // BRANCH INSTRUCTIONS /////////////////////////////////////////////////////////////////////// void encodeBranchingInstruction(const Instruction &inst); void encodeBranchingInstructionSimplified(const Instruction &inst); void encodeBranchDestination(const Operand &dst); void encodeBranchSource(const Operand &src); /////////////////////////////////////////////////////////////////////// // SEND INSTRUCTIONS /////////////////////////////////////////////////////////////////////// void encodeSendInstruction(const Instruction &inst); void encodeSendSource0(const Operand &src); void encodeSendDestination(const Operand &dst); void encodeSendgDestinationXe3p(const Operand &dst); void encodeSendgxDestinationXe3p(const Operand &dst); void encodeSendgxSource0Xe3p(const Operand &src); void encodeSendgxSource1Xe3p(const Operand &src); void encodeSendsSource0(const Operand &src); void encodeSendsSource1(const Operand &src); void encodeSendsDestination(const Operand &dst); void encodeSendDescs(const Instruction &inst); void encodeSendDescsPreXe(const Instruction &inst); void encodeSendDescsXe(const Instruction &inst); void encodeSendDescsXeHP(const Instruction &inst); void encodeSendDescsXeHPG(const Instruction &inst); void encodeSendDescsXe2(const Instruction &inst); void encodeSendgDescsXe3(const Instruction &inst); void encodeInlineBinaryInst(Instruction &inst); /////////////////////////////////////////////////////////////////////// // SYNC INSTRUCTIONS /////////////////////////////////////////////////////////////////////// void encodeSyncInstruction(const Instruction &inst); /////////////////////////////////////////////////////////////////////// // OTHER HELPER FUNCTIONS /////////////////////////////////////////////////////////////////////// void setEncodedPC(Instruction *inst, int32_t encodePC); int32_t getEncodedPC(const Instruction *inst) const; bool getBlockOffset(const Block *b, uint32_t &pc); // handles encoding ARF registers as well as the easy GRF case // caller actually pulls the trigger and encodes the bits, but this // call can raise the encoding error uint32_t translateRegNum(int opIx, RegName reg, uint16_t regNum); uint32_t mathMacroRegToBits(int src, MathMacroExt mme); // ChSel and SubReg GED_DST_CHAN_EN mathMacroRegToChEn(MathMacroExt mme); void applyGedWorkarounds(const Kernel &k, size_t bitsLen); void encodeOptions(const Instruction &inst); ////////////////////////////////////////////////////////////////////// // platform specific queries *but sometimes need the instruction too) // // GEN7p5 implicitly scales PC offsets by QW except for a few instructions bool arePcsInQWords(const OpSpec &os) const; // Call need to have src0 region be set to: // SKL and before: <2;2,1> // ICL: <2;4,1> // Later GENs ignore the region completely bool callNeedsSrc0Region221(const Instruction &inst) const; bool callNeedsSrc0Region241(const Instruction &inst) const; ///////////////////////////////////////////////////////////// // state valid over encodeKernel() MemManager *m_mem; uint8_t *m_instBuf = nullptr; // the output bits struct JumpPatch { // JIP and UIP label patching Instruction *inst; // the instruction ged_ins_t gedInst; // the partially constructed GED instruction uint8_t *bits; // where to encode it in the heap JumpPatch(Instruction *i, const ged_ins_t &gi, uint8_t *bs) : inst(i), gedInst(gi), bits(bs) {} }; std::vector m_needToPatch; std::map m_blockToOffsetMap; public: //////////////////////////////////////////////////////////////// static uint64_t typeConvesionHelper(const ImmVal &val, Type type) { uint64_t value = 0; switch (type) { case Type::UD: case Type::F: case Type::V: case Type::UV: case Type::VF: value = (uint64_t)val.u32; break; case Type::D: value = (uint64_t)val.s32; break; case Type::W: value = (uint64_t)val.s16; break; case Type::UW: case Type::HF: case Type::BF: value = (uint64_t)val.u16; break; case Type::DF: case Type::UQ: case Type::Q: value = val.u64; break; case Type::B: case Type::UB: // technically not reachable since we don't permit byte moves // from immediates value = val.u64; break; case Type::BF8: case Type::HF8: value = (uint64_t)val.u8; break; case Type::TF32: value = (uint64_t)val.u32; break; default: break; } return value; } protected: //////////////////////////////////////////////////////////////// // allowable types for ternary Align1 static bool isTernaryAlign1Floating(Type t) { return lowerExecDataType(t) == GED_EXECUTION_DATA_TYPE::GED_EXECUTION_DATA_TYPE_Float; } // allowable types for ternary Align1 static bool isTernaryAlign1Integral(Type t) { return lowerExecDataType(t) == GED_EXECUTION_DATA_TYPE::GED_EXECUTION_DATA_TYPE_Integer; } }; // end: class definition Encoder template void Encoder::encodeSrcRegFile(GED_REG_FILE rf) { if (S == SourceIndex::SRC0) { GED_ENCODE(Src0RegFile, rf); } else if (S == SourceIndex::SRC1) { GED_ENCODE(Src1RegFile, rf); } else if (S == SourceIndex::SRC2) { GED_ENCODE(Src2RegFile, rf); } else if (S == SourceIndex::SRC3) { GED_ENCODE(Src3RegFile, rf); } else if (S == SourceIndex::SRC4) { GED_ENCODE(Src4RegFile, rf); } } template void Encoder::encodeSrcRegionVert(Region::Vert v) { if (S == SourceIndex::SRC0) { GED_ENCODE(Src0VertStride, lowerRegionVert(v)); } else { // (S == SourceIndex::SRC1) GED_ENCODE(Src1VertStride, lowerRegionVert(v)); } // S != SRC2 since ternary Align1 doesn't have bits for that } template void Encoder::encodeSrcType(Type t) { if (S == SourceIndex::SRC0) { GED_ENCODE(Src0DataType, lowerDataType(t)); } else if (S == SourceIndex::SRC1) { GED_ENCODE(Src1DataType, lowerDataType(t)); } else { GED_ENCODE(Src2DataType, lowerDataType(t)); } } template void Encoder::encodeSrcAddrMode(GED_ADDR_MODE x) { if (S == SourceIndex::SRC0) { GED_ENCODE(Src0AddrMode, x); } else { GED_ENCODE(Src1AddrMode, x); } } template void Encoder::encodeSrcModifier(SrcModifier x) { if (S == SourceIndex::SRC0) { GED_ENCODE(Src0SrcMod, lowerSrcMod(x)); } else if (S == SourceIndex::SRC1) { GED_ENCODE(Src1SrcMod, lowerSrcMod(x)); } else { GED_ENCODE(Src2SrcMod, lowerSrcMod(x)); } } template void Encoder::encodeSrcSubRegNum(uint32_t subRegInByte) { if (S == SourceIndex::SRC0) { GED_ENCODE(Src0SubRegNum, subRegInByte); } else if (S == SourceIndex::SRC1) { GED_ENCODE(Src1SubRegNum, subRegInByte); } else if (S == SourceIndex::SRC2) { GED_ENCODE(Src2SubRegNum, subRegInByte); } else if (S == SourceIndex::SRC3) { GED_ENCODE(Src3SubRegNum, subRegInByte); } else if (S == SourceIndex::SRC4) { GED_ENCODE(Src4SubRegNum, subRegInByte); } } template void Encoder::encodeSrcMathMacroReg(MathMacroExt a) { if (S == SourceIndex::SRC0) { GED_ENCODE(Src0MathMacroExt, lowerSpecialAcc(a)); } else if (S == SourceIndex::SRC1) { GED_ENCODE(Src1MathMacroExt, lowerSpecialAcc(a)); } else { GED_ENCODE(Src2MathMacroExt, lowerSpecialAcc(a)); } } template void Encoder::encodeSrcReg(RegName regName, uint16_t regNum) { uint32_t regBits = 0; if (regName == RegName::GRF_R) { regBits = regNum; // GRF fast path } else { // ARF slower path const RegInfo *ri = m_model.lookupRegInfoByRegName(regName); if (!ri) { errorT("src", (int)S, ": unexpected register on this platform"); } else { uint16_t reg; ri->encode((int)regNum, reg); regBits = reg; // widen for GED } } if (S == SourceIndex::SRC0) { GED_ENCODE(Src0RegNum, regBits); } else if (S == SourceIndex::SRC1) { GED_ENCODE(Src1RegNum, regBits); } else if (S == SourceIndex::SRC2) { GED_ENCODE(Src2RegNum, regBits); } else if (S == SourceIndex::SRC3) { GED_ENCODE(Src3RegNum, regBits); } else if (S == SourceIndex::SRC4) { GED_ENCODE(Src4RegNum, regBits); } } template void Encoder::encodeSrcAddrImm(int32_t addrImm) { if (S == SourceIndex::SRC0) { GED_ENCODE(Src0AddrImm, addrImm); } else { GED_ENCODE(Src1AddrImm, addrImm); } } template void Encoder::encodeSrcAddrSubRegNum(uint32_t addrSubReg) { if (S == SourceIndex::SRC0) { GED_ENCODE(Src0AddrSubRegNum, addrSubReg); } else { GED_ENCODE(Src1AddrSubRegNum, addrSubReg); } } template void Encoder::encodeSrcRegion(const Region &rgn, bool hasRgnWi) { uint32_t v = 0; if (rgn.getVt() == Region::Vert::VT_VxH) { v = 0x3; } else if (rgn.getVt() != Region::Vert::VT_INVALID) { v = static_cast(rgn.v); } else { errorT(S == SourceIndex::SRC0 ? "invalid region vertical stride on src0" : "invalid region vertical stride on src1"); } uint32_t w = static_cast(rgn.getWi()); if (rgn.getWi() == Region::Width::WI_INVALID) { errorT(S == SourceIndex::SRC0 ? "invalid region width on src0" : "invalid region width on src1"); } uint32_t h = static_cast(rgn.getHz()); if (rgn.getHz() == Region::Horz::HZ_INVALID) { h = 1; errorT(S == SourceIndex::SRC0 ? "invalid region horizontal stride on src0" : "invalid region horizontal stride on src1"); } if (S == SourceIndex::SRC0) { GED_ENCODE(Src0VertStride, v); if (hasRgnWi) { GED_ENCODE(Src0Width, w); } else { // some ops have an implicit width region // (e.g. some specialized instructions poaches Src0.RgnWi) // // Within the IR we use 1 so logic that depends on regioning // gets the correct behavior (hardware assumes w=1). w = 1; } GED_ENCODE(Src0HorzStride, h); } else if (S == SourceIndex::SRC1) { GED_ENCODE(Src1VertStride, v); GED_ENCODE(Src1Width, w); GED_ENCODE(Src1HorzStride, h); } else { IGA_ASSERT_FALSE("Encoder::encodeSrcRegion: only works on src0 and src1"); } } template void Encoder::encodeSrcReducedRegion(const Region& rgn) { IGA_ASSERT(S == SourceIndex::SRC1, "Invalid SourceIndex"); if (rgn.isScalar()) GED_ENCODE(Src1ScalarReg, 1); else GED_ENCODE(Src1ScalarReg, 0); return; IGA_ASSERT_FALSE("Unreachable: Unsupported region"); } template void Encoder::encodeSrcRegionWidth(Region::Width w) { if (S == SourceIndex::SRC0) { GED_ENCODE(Src0Width, lowerRegionWidth(w)); } else { // (S == SourceIndex::SRC1) GED_ENCODE(Src1Width, lowerRegionWidth(w)); } // S != SRC2 since ternary Align1 doesn't have bits for that } template void Encoder::encodeTernaryImmVal(const ImmVal &val, Type type) { if (S == SourceIndex::SRC0) { GED_ENCODE(Src0TernaryImm, typeConvesionHelper(val, type)); } else if (S == SourceIndex::SRC2) { GED_ENCODE(Src2TernaryImm, typeConvesionHelper(val, type)); } else { errorT("immediate operands not supported in src1 of ternary formats"); } } template void Encoder::encodeSrcRegionHorz(Region::Horz s) { if (S == SourceIndex::SRC0) { GED_ENCODE(Src0HorzStride, lowerRegionHorz(s)); } else if (S == SourceIndex::SRC1) { GED_ENCODE(Src1HorzStride, lowerRegionHorz(s)); } else { GED_ENCODE(Src2HorzStride, lowerRegionHorz(s)); } } } // namespace iga namespace iga { typedef Encoder Encoder; } #endif