[Beignet] [PATCH 1/2] add 3 simd level built-in functions: shuffle, simdsize and simdid

Yang, Rong R rong.r.yang at intel.com
Fri Apr 10 00:48:33 PDT 2015



> -----Original Message-----
> From: Guo, Yejun
> Sent: Tuesday, April 7, 2015 11:34
> To: Yang, Rong R; beignet at lists.freedesktop.org
> Subject: RE: [PATCH 1/2] add 3 simd level built-in functions: shuffle, simdsize
> and simdid
> 
> To make it clear, I copy your comments and explain as below.
> 
> You can use GenRegister::suboffset directly here.
> Also can use GenRegister::suboffset.
> [Yejun]  GenRegister::suboffset does not do the work that I want. For
> example, suppose src0 is r122<8,8,1>:D and offset is 5, what I want is
> r122.5<0,1,0>:D
[Yang, Rong R] GenRegister::vec1(GenRegister) could be used.

> 
> Why not handle SEL_OP_SIMD_ID in instruction selection, just as SIMD_SIZE?
> If handle SEL_OP_SIMD_ID in gen_insn_selection, need not special handle
> here.
> Furthermore, you could try to handle SIMD_ID and SIMD_SIZE in the
> GenWriter, then can avoid NullaryInstruction totally.
> [Yejun] SIMD_ID needs data from payload, need to invoke function 'ra-
> >genReg'.
[Yang, Rong R] Can use the virtual IR register before register allocation.

> 
> Actually, laneid is same as ocl::stackptr, can you reuse or rename it?
> [Yejun] they hold the same data from payload, but ocl::stackptr is changed in
> function emitStackPointer. We can refine the use of stackptr after this big
> patch is accepted in the code.
[Yang, Rong R] OK.

> 
> -----Original Message-----
> From: Yang, Rong R
> Sent: Friday, April 03, 2015 1:47 PM
> To: Guo, Yejun; beignet at lists.freedesktop.org
> Subject: RE: [PATCH 1/2] add 3 simd level built-in functions: shuffle, simdsize
> and simdid
> 
> Some comments.
> 
> Thanks
> > -----Original Message-----
> > From: Beignet [mailto:beignet-bounces at lists.freedesktop.org] On Behalf
> > Of Guo, Yejun
> > Sent: Friday, March 27, 2015 14:17
> > To: beignet at lists.freedesktop.org
> > Subject: Re: [Beignet] [PATCH 1/2] add 3 simd level built-in
> > functions: shuffle, simdsize and simdid
> >
> > Ask for review, thanks.
> >
> > yejun
> >
> > -----Original Message-----
> > From: Guo, Yejun
> > Sent: Friday, March 20, 2015 1:58 PM
> > To: beignet at lists.freedesktop.org
> > Cc: Guo, Yejun
> > Subject: [PATCH 1/2] add 3 simd level built-in functions: shuffle,
> > simdsize and simdid
> >
> > uint __gen_ocl_get_simd_size();
> > returns 8 if SIMD8, returns 16 if SIMD16
> >
> > uint __gen_ocl_get_simd_id();
> > return value ranges from 0 to simdsize - 1
> >
> > floatN __gen_ocl_simd_shuffle(floatN x, uint c);
> > intN   __gen_ocl_simd_shuffle(intN x, uint c);
> > uintN  __gen_ocl_simd_shuffle(uintN x, uint c); the value of x of the
> > c-th channel of the SIMD is returned, for all SIMD channels, the
> > behavior is undefined if c is larger than simdsize - 1
> >
> > Signed-off-by: Guo Yejun <yejun.guo at intel.com>
> > ---
> >  backend/src/backend/gen8_context.cpp               |  29 ++++-
> >  backend/src/backend/gen_context.cpp                | 127 +++++++++++++++--
> --
> > --
> >  backend/src/backend/gen_context.hpp                |   1 +
> >  .../src/backend/gen_insn_gen7_schedule_info.hxx    |   1 +
> >  backend/src/backend/gen_insn_selection.cpp         |  60 ++++++++++
> >  backend/src/backend/gen_insn_selection.hxx         |   2 +
> >  backend/src/backend/program.h                      |   1 +
> >  backend/src/ir/context.hpp                         |   6 +
> >  backend/src/ir/instruction.cpp                     |  32 ++++++
> >  backend/src/ir/instruction.hpp                     |  17 +++
> >  backend/src/ir/instruction.hxx                     |   3 +
> >  backend/src/ir/liveness.cpp                        |   5 +
> >  backend/src/ir/profile.cpp                         |   2 +
> >  backend/src/ir/profile.hpp                         |   5 +-
> >  backend/src/libocl/CMakeLists.txt                  |   2 +-
> >  backend/src/libocl/include/ocl.h                   |   1 +
> >  backend/src/libocl/include/ocl_misc.h              |   8 --
> >  backend/src/libocl/script/ocl_simd.def             |   4 +
> >  backend/src/libocl/tmpl/ocl_simd.tmpl.cl           |  19 +++
> >  backend/src/libocl/tmpl/ocl_simd.tmpl.h            |  34 ++++++
> >  backend/src/llvm/llvm_gen_backend.cpp              |  27 +++++
> >  backend/src/llvm/llvm_gen_ocl_function.hxx         |   4 +
> >  src/cl_command_queue_gen7.c                        |   8 ++
> >  23 files changed, 351 insertions(+), 47 deletions(-)  create mode
> > 100644 backend/src/libocl/script/ocl_simd.def
> >  create mode 100644 backend/src/libocl/tmpl/ocl_simd.tmpl.cl
> >  create mode 100644 backend/src/libocl/tmpl/ocl_simd.tmpl.h
> >
> > diff --git a/backend/src/backend/gen8_context.cpp
> > b/backend/src/backend/gen8_context.cpp
> > index 3f57cf6..144fd00 100644
> > --- a/backend/src/backend/gen8_context.cpp
> > +++ b/backend/src/backend/gen8_context.cpp
> > @@ -240,6 +240,9 @@ namespace gbe
> >    }
> >
> >    void Gen8Context::emitBinaryInstruction(const SelectionInstruction
> > &insn) {
> > +    const GenRegister dst = ra->genReg(insn.dst(0));
> > +    const GenRegister src0 = ra->genReg(insn.src(0));
> > +    const GenRegister src1 = ra->genReg(insn.src(1));
> >      switch (insn.opcode) {
> >        case SEL_OP_SEL_INT64:
> >        case SEL_OP_I64AND:
> > @@ -250,14 +253,34 @@ namespace gbe
> >          break;
> >        case SEL_OP_UPSAMPLE_LONG:
> >        {
> > -        const GenRegister dst = ra->genReg(insn.dst(0));
> > -        const GenRegister src0 = ra->genReg(insn.src(0));
> > -        const GenRegister src1 = ra->genReg(insn.src(1));
> >          p->MOV(dst, src0);
> >          p->SHL(dst, dst, GenRegister::immud(32));
> >          p->ADD(dst, dst, src1);
> >          break;
> >        }
> > +      case SEL_OP_SIMD_SHUFFLE:
> > +      {
> > +        uint32_t simd = p->curr.execWidth;
> > +        if (src1.file == GEN_IMMEDIATE_VALUE) {
> > +          uint32_t offset = src1.value.ud % simd;
> > +          uint32_t nr = src0.nr;
> > +          uint32_t subnr = src0.subnr;
> > +          subnr = subnr + offset;
> > +          if (subnr > 8) {
> > +            nr = nr + 1;
> > +            subnr = subnr - 8;
> > +          }
> You can use GenRegister::suboffset directly here.
> 
> > +          p->MOV(dst, GenRegister::ud1grf(nr, subnr));
> > +        } else {
> > +          uint32_t base = src0.nr * 32 + src0.subnr * 4;
> > +          GenRegister baseReg = GenRegister::immuw(base);
> > +          const GenRegister a0 = GenRegister::addr8(0);
> > +          p->ADD(a0, GenRegister::unpacked_uw(src1.nr, src1.subnr /
> > typeSize(GEN_TYPE_UW)), baseReg);
> > +          GenRegister indirect = GenRegister::to_indirect1xN(src0, 0, 0);
> > +          p->MOV(dst, indirect);
> > +        }
> > +        break;
> > +      }
> >        default:
> >          GenContext::emitBinaryInstruction(insn);
> >      }
> > diff --git a/backend/src/backend/gen_context.cpp
> > b/backend/src/backend/gen_context.cpp
> > index cdf581c..25c7a5a 100644
> > --- a/backend/src/backend/gen_context.cpp
> > +++ b/backend/src/backend/gen_context.cpp
> > @@ -198,6 +198,22 @@ namespace gbe
> >      this->labelPos.insert(std::make_pair(label, p->store.size()));
> >    }
> >
> > +  void GenContext::emitNullaryInstruction(const SelectionInstruction
> > + &insn)
> > {
> > +    const GenRegister dst = ra->genReg(insn.dst(0));
> > +    switch (insn.opcode) {
> > +      case SEL_OP_SIMD_ID:
> > +        {
> > +          const GenRegister selLaneID = this->simdWidth == 8 ?
> > +                                GenRegister::ud8grf(ir::ocl::laneid) :
> > +                                GenRegister::ud16grf(ir::ocl::laneid);
> > +          const GenRegister laneID = ra->genReg(selLaneID);
> > +          p->MOV(dst, laneID);
> > +        }
> > +        break;
> > +      default: NOT_IMPLEMENTED;
> > +    }
> > +  }
> > +
> Why not handle SEL_OP_SIMD_ID in instruction selection, just as SIMD_SIZE?
> Furthermore, you could try to handle SIMD_ID and SIMD_SIZE in the
> GenWriter, then can avoid NullaryInstruction totally.
> 
> >    void GenContext::emitUnaryInstruction(const SelectionInstruction &insn)
> {
> >      const GenRegister dst = ra->genReg(insn.dst(0));
> >      const GenRegister src = ra->genReg(insn.src(0)); @@ -583,6
> > +599,46 @@ namespace gbe
> >            p->MOV(xdst.bottom_half(), xsrc1.bottom_half());
> >          }
> >          break;
> > +      case SEL_OP_SIMD_SHUFFLE:
> > +        {
> > +          uint32_t simd = p->curr.execWidth;
> > +          if (src1.file == GEN_IMMEDIATE_VALUE) {
> > +            uint32_t offset = src1.value.ud % simd;
> > +            uint32_t nr = src0.nr;
> > +            uint32_t subnr = src0.subnr;
> > +            subnr = subnr + offset;
> > +            if (subnr > 8) {
> > +              nr = nr + 1;
> > +              subnr = subnr - 8;
> > +            }
> Also can use GenRegister::suboffset.
> 
> > +            p->MOV(dst, GenRegister::ud1grf(nr, subnr));
> > +          } else {
> > +            uint32_t base = src0.nr * 32 + src0.subnr * 4;
> > +            GenRegister baseReg = GenRegister::immuw(base);
> > +            const GenRegister a0 = GenRegister::addr8(0);
> > +
> > +            p->push();
> > +              if (simd == 8) {
> > +                p->ADD(a0, GenRegister::unpacked_uw(src1.nr,
> > + src1.subnr /
> > typeSize(GEN_TYPE_UW)), baseReg);
> > +                GenRegister indirect = GenRegister::to_indirect1xN(src0, 0, 0);
> > +                p->MOV(dst, indirect);
> > +              }
> > +              else if (simd == 16) {
> > +                p->curr.execWidth = 8;
> > +                p->ADD(a0, GenRegister::unpacked_uw(src1.nr,
> > + src1.subnr /
> > typeSize(GEN_TYPE_UW)), baseReg);
> > +                GenRegister indirect = GenRegister::to_indirect1xN(src0, 0, 0);
> > +                p->MOV(dst, indirect);
> > +
> > +                p->curr.quarterControl = 1;
> > +                p->ADD(a0, GenRegister::unpacked_uw(src1.nr+1,
> > + src1.subnr /
> > typeSize(GEN_TYPE_UW)), baseReg);
> > +                p->MOV(GenRegister::offset(dst, 1, 0), indirect);
> > +              }
> > +              else
> > +                NOT_IMPLEMENTED;
> > +            p->pop();
> > +          }
> > +        }
> > +        break;
> >        default: NOT_IMPLEMENTED;
> >      }
> >    }
> > @@ -2023,41 +2079,46 @@ namespace gbe
> >      } else
> >
> >      fn.foreachInstruction([&](ir::Instruction &insn) {
> > -      const uint32_t srcNum = insn.getSrcNum();
> > -      for (uint32_t srcID = 0; srcID < srcNum; ++srcID) {
> > -        const ir::Register reg = insn.getSrc(srcID);
> > -        if (insn.getOpcode() == ir::OP_GET_IMAGE_INFO) {
> > -          if (srcID != 0) continue;
> > -          const unsigned char bti =
> > ir::cast<ir::GetImageInfoInstruction>(insn).getImageIndex();
> > -          const unsigned char type =
> > ir::cast<ir::GetImageInfoInstruction>(insn).getInfoType();;
> > -          ir::ImageInfoKey key(bti, type);
> > -          const ir::Register imageInfo = insn.getSrc(0);
> > -          if (curbeRegs.find(imageInfo) == curbeRegs.end()) {
> > -            uint32_t offset = this->getImageInfoCurbeOffset(key, 4);
> > -            insertCurbeReg(imageInfo, offset);
> > +      if (insn.getOpcode() == ir::OP_SIMD_ID) {
> > +        if (curbeRegs.find(laneid) == curbeRegs.end())
> > +          allocCurbeReg(laneid, GBE_CURBE_LANE_ID);
> If handle SEL_OP_SIMD_ID in gen_insn_selection, need not special handle
> here.
> 
> > +      } else {
> > +        const uint32_t srcNum = insn.getSrcNum();
> > +        for (uint32_t srcID = 0; srcID < srcNum; ++srcID) {
> > +          const ir::Register reg = insn.getSrc(srcID);
> > +          if (insn.getOpcode() == ir::OP_GET_IMAGE_INFO) {
> > +            if (srcID != 0) continue;
> > +            const unsigned char bti =
> > ir::cast<ir::GetImageInfoInstruction>(insn).getImageIndex();
> > +            const unsigned char type =
> > ir::cast<ir::GetImageInfoInstruction>(insn).getInfoType();;
> > +            ir::ImageInfoKey key(bti, type);
> > +            const ir::Register imageInfo = insn.getSrc(0);
> > +            if (curbeRegs.find(imageInfo) == curbeRegs.end()) {
> > +              uint32_t offset = this->getImageInfoCurbeOffset(key, 4);
> > +              insertCurbeReg(imageInfo, offset);
> > +            }
> > +            continue;
> >            }
> > -          continue;
> > +          if (fn.isSpecialReg(reg) == false) continue;
> > +          if (curbeRegs.find(reg) != curbeRegs.end()) continue;
> > +          if (reg == ir::ocl::stackptr) GBE_ASSERT(stackUse.size() > 0);
> > +          INSERT_REG(lsize0, LOCAL_SIZE_X)
> > +          INSERT_REG(lsize1, LOCAL_SIZE_Y)
> > +          INSERT_REG(lsize2, LOCAL_SIZE_Z)
> > +          INSERT_REG(gsize0, GLOBAL_SIZE_X)
> > +          INSERT_REG(gsize1, GLOBAL_SIZE_Y)
> > +          INSERT_REG(gsize2, GLOBAL_SIZE_Z)
> > +          INSERT_REG(goffset0, GLOBAL_OFFSET_X)
> > +          INSERT_REG(goffset1, GLOBAL_OFFSET_Y)
> > +          INSERT_REG(goffset2, GLOBAL_OFFSET_Z)
> > +          INSERT_REG(workdim, WORK_DIM)
> > +          INSERT_REG(numgroup0, GROUP_NUM_X)
> > +          INSERT_REG(numgroup1, GROUP_NUM_Y)
> > +          INSERT_REG(numgroup2, GROUP_NUM_Z)
> > +          INSERT_REG(stackptr, STACK_POINTER)
> > +          INSERT_REG(printfbptr, PRINTF_BUF_POINTER)
> > +          INSERT_REG(printfiptr, PRINTF_INDEX_POINTER)
> > +          do {} while(0);
> >          }
> > -        if (fn.isSpecialReg(reg) == false) continue;
> > -        if (curbeRegs.find(reg) != curbeRegs.end()) continue;
> > -        if (reg == ir::ocl::stackptr) GBE_ASSERT(stackUse.size() > 0);
> > -        INSERT_REG(lsize0, LOCAL_SIZE_X)
> > -        INSERT_REG(lsize1, LOCAL_SIZE_Y)
> > -        INSERT_REG(lsize2, LOCAL_SIZE_Z)
> > -        INSERT_REG(gsize0, GLOBAL_SIZE_X)
> > -        INSERT_REG(gsize1, GLOBAL_SIZE_Y)
> > -        INSERT_REG(gsize2, GLOBAL_SIZE_Z)
> > -        INSERT_REG(goffset0, GLOBAL_OFFSET_X)
> > -        INSERT_REG(goffset1, GLOBAL_OFFSET_Y)
> > -        INSERT_REG(goffset2, GLOBAL_OFFSET_Z)
> > -        INSERT_REG(workdim, WORK_DIM)
> > -        INSERT_REG(numgroup0, GROUP_NUM_X)
> > -        INSERT_REG(numgroup1, GROUP_NUM_Y)
> > -        INSERT_REG(numgroup2, GROUP_NUM_Z)
> > -        INSERT_REG(stackptr, STACK_POINTER)
> > -        INSERT_REG(printfbptr, PRINTF_BUF_POINTER)
> > -        INSERT_REG(printfiptr, PRINTF_INDEX_POINTER)
> > -        do {} while(0);
> >        }
> >      });
> >  #undef INSERT_REG
> > diff --git a/backend/src/backend/gen_context.hpp
> > b/backend/src/backend/gen_context.hpp
> > index 6ca88db..3ac675e 100644
> > --- a/backend/src/backend/gen_context.hpp
> > +++ b/backend/src/backend/gen_context.hpp
> > @@ -124,6 +124,7 @@ namespace gbe
> >
> >      /*! Final Gen ISA emission helper functions */
> >      void emitLabelInstruction(const SelectionInstruction &insn);
> > +    virtual void emitNullaryInstruction(const SelectionInstruction
> > + &insn);
> >      virtual void emitUnaryInstruction(const SelectionInstruction &insn);
> >      virtual void emitUnaryWithTempInstruction(const
> > SelectionInstruction &insn);
> >      virtual void emitBinaryInstruction(const SelectionInstruction
> > &insn); diff -- git
> > a/backend/src/backend/gen_insn_gen7_schedule_info.hxx
> > b/backend/src/backend/gen_insn_gen7_schedule_info.hxx
> > index d054820..fd7e1a4 100644
> > --- a/backend/src/backend/gen_insn_gen7_schedule_info.hxx
> > +++ b/backend/src/backend/gen_insn_gen7_schedule_info.hxx
> > @@ -1,5 +1,6 @@
> >  //                 Family     Latency     SIMD16     SIMD8
> >  DECL_GEN7_SCHEDULE(Label,           0,         0,        0)
> > +DECL_GEN7_SCHEDULE(Nullary,         20,        4,        2)
> >  DECL_GEN7_SCHEDULE(Unary,           20,        4,        2)
> >  DECL_GEN7_SCHEDULE(UnaryWithTemp,   20,        40,      20)
> >  DECL_GEN7_SCHEDULE(Binary,          20,        4,        2)
> > diff --git a/backend/src/backend/gen_insn_selection.cpp
> > b/backend/src/backend/gen_insn_selection.cpp
> > index c240261..1586098 100644
> > --- a/backend/src/backend/gen_insn_selection.cpp
> > +++ b/backend/src/backend/gen_insn_selection.cpp
> > @@ -477,6 +477,8 @@ namespace gbe
> >      /*! To make function prototypes more readable */
> >      typedef const GenRegister &Reg;
> >
> > +#define ALU0(OP) \
> > +  INLINE void OP(Reg dst) { ALU0(SEL_OP_##OP, dst); }
> >  #define ALU1(OP) \
> >    INLINE void OP(Reg dst, Reg src) { ALU1(SEL_OP_##OP, dst, src); }
> > #define
> > ALU1WithTemp(OP) \ @@ -530,12 +532,15 @@ namespace gbe
> >      ALU2WithTemp(HADD)
> >      ALU2WithTemp(RHADD)
> >      ALU2(UPSAMPLE_LONG)
> > +    ALU2(SIMD_SHUFFLE)
> > +    ALU0(SIMD_ID)
> >      ALU1WithTemp(CONVI_TO_I64)
> >      ALU1WithTemp(CONVF_TO_I64)
> >      ALU1(CONVI64_TO_I)
> >      I64Shift(I64SHL)
> >      I64Shift(I64SHR)
> >      I64Shift(I64ASR)
> > +#undef ALU0
> >  #undef ALU1
> >  #undef ALU1WithTemp
> >  #undef ALU2
> > @@ -622,6 +627,8 @@ namespace gbe
> >      void MATH(Reg dst, uint32_t function, Reg src0, Reg src1);
> >      /*! Extended math function (1 argument) */
> >      void MATH(Reg dst, uint32_t function, Reg src);
> > +    /*! Encode nullary instructions */
> > +    void ALU0(SelectionOpcode opcode, Reg dst);
> >      /*! Encode unary instructions */
> >      void ALU1(SelectionOpcode opcode, Reg dst, Reg src);
> >      /*! Encode unary with temp reg instructions */ @@ -1435,6
> > +1442,11 @@ namespace gbe
> >        insn->dst(i + 1) = tmp[i];
> >    }
> >
> > +  void Selection::Opaque::ALU0(SelectionOpcode opcode, Reg dst) {
> > +    SelectionInstruction *insn = this->appendInsn(opcode, 1, 0);
> > +    insn->dst(0) = dst;
> > +  }
> > +
> >    void Selection::Opaque::ALU1(SelectionOpcode opcode, Reg dst, Reg src)
> {
> >      SelectionInstruction *insn = this->appendInsn(opcode, 1, 1);
> >      insn->dst(0) = dst;
> > @@ -2054,6 +2066,42 @@ namespace gbe
> >  #define DECL_CTOR(FAMILY, INSN_NUM, COST) \
> >    FAMILY##Pattern(void) : OneToManyPattern<FAMILY##Pattern,
> > ir::FAMILY>(INSN_NUM, COST) {}
> >
> > +  /*! Nullary instruction patterns */  class
> > + NullaryInstructionPattern : public SelectionPattern  {
> > +  public:
> > +    NullaryInstructionPattern(void) : SelectionPattern(1,1) {
> > +      for (uint32_t op = 0; op < ir::OP_INVALID; ++op)
> > +        if (ir::isOpcodeFrom<ir::NullaryInstruction>(ir::Opcode(op)) == true)
> > +          this->opcodes.push_back(ir::Opcode(op));
> > +    }
> > +
> > +    INLINE bool emit(Selection::Opaque &sel, SelectionDAG &dag) const {
> > +      using namespace ir;
> > +      const ir::NullaryInstruction &insn = cast<NullaryInstruction>(dag.insn);
> > +      const Opcode opcode = insn.getOpcode();
> > +      const Type type = insn.getType();
> > +      GenRegister dst = sel.selReg(insn.getDst(0), type);
> > +
> > +      sel.push();
> > +      switch (opcode) {
> > +        case ir::OP_SIMD_SIZE:
> > +          {
> > +            const GenRegister src = GenRegister::immud(sel.curr.execWidth);
> > +            sel.curr.execWidth = 1;
> > +            sel.MOV(dst, src);
> > +          }
> > +          break;
> > +        case ir::OP_SIMD_ID:
> > +          sel.SIMD_ID(dst);
> > +          break;
> > +        default: NOT_SUPPORTED;
> > +      }
> > +      sel.pop();
> > +      return true;
> > +    }
> > +  };
> > +
> >    /*! Unary instruction patterns */
> >    DECL_PATTERN(UnaryInstruction)
> >    {
> > @@ -2563,6 +2611,17 @@ namespace gbe
> >          case OP_UPSAMPLE_LONG:
> >            sel.UPSAMPLE_LONG(dst, src0, src1);
> >            break;
> > +        case OP_SIMD_SHUFFLE:
> > +          {
> > +            if (src1.file == GEN_IMMEDIATE_VALUE) {
> > +              sel.SIMD_SHUFFLE(dst, src0, src1);
> > +            } else {
> > +              GenRegister shiftL =
> > + GenRegister::udxgrf(sel.curr.execWidth,
> > sel.reg(FAMILY_DWORD));
> > +              sel.SHL(shiftL, src1, GenRegister::immud(0x2));
> > +              sel.SIMD_SHUFFLE(dst, src0, shiftL);
> > +            }
> > +          }
> > +          break;
> >          default: NOT_IMPLEMENTED;
> >        }
> >        sel.pop();
> > @@ -4789,6 +4848,7 @@ namespace gbe
> >      this->insert<GetImageInfoInstructionPattern>();
> >      this->insert<ReadARFInstructionPattern>();
> >      this->insert<RegionInstructionPattern>();
> > +    this->insert<NullaryInstructionPattern>();
> >
> >      // Sort all the patterns with the number of instructions they output
> >      for (uint32_t op = 0; op < ir::OP_INVALID; ++op) diff --git
> > a/backend/src/backend/gen_insn_selection.hxx
> > b/backend/src/backend/gen_insn_selection.hxx
> > index 09f5aaf..87ccee3 100644
> > --- a/backend/src/backend/gen_insn_selection.hxx
> > +++ b/backend/src/backend/gen_insn_selection.hxx
> > @@ -77,6 +77,8 @@ DECL_SELECTION_IR(RHADD,
> > BinaryWithTempInstruction)  DECL_SELECTION_IR(I64HADD,
> > I64HADDInstruction)  DECL_SELECTION_IR(I64RHADD, I64RHADDInstruction)
> > DECL_SELECTION_IR(UPSAMPLE_LONG, BinaryInstruction)
> > +DECL_SELECTION_IR(SIMD_SHUFFLE, BinaryInstruction)
> > +DECL_SELECTION_IR(SIMD_ID, NullaryInstruction)
> >  DECL_SELECTION_IR(CONVI_TO_I64, UnaryWithTempInstruction)
> > DECL_SELECTION_IR(CONVI64_TO_I, UnaryInstruction)
> > DECL_SELECTION_IR(CONVI64_TO_F, I64ToFloatInstruction) diff --git
> > a/backend/src/backend/program.h b/backend/src/backend/program.h
> index
> > dc5662f..c4023ec 100644
> > --- a/backend/src/backend/program.h
> > +++ b/backend/src/backend/program.h
> > @@ -99,6 +99,7 @@ enum gbe_curbe_type {
> >    GBE_CURBE_THREAD_NUM,
> >    GBE_CURBE_ZERO,
> >    GBE_CURBE_ONE,
> > +  GBE_CURBE_LANE_ID,
> >    GBE_CURBE_SLM_OFFSET,
> >  };
> >
> > diff --git a/backend/src/ir/context.hpp b/backend/src/ir/context.hpp
> > index
> > cf5109d..af65ff3 100644
> > --- a/backend/src/ir/context.hpp
> > +++ b/backend/src/ir/context.hpp
> > @@ -176,6 +176,12 @@ namespace ir {
> >      DECL_THREE_SRC_INSN(MAD);
> >  #undef DECL_THREE_SRC_INSN
> >
> > +    /*! For all nullary functions */
> > +    void ALU0(Opcode opcode, Type type, Register dst) {
> > +      const Instruction insn = gbe::ir::ALU0(opcode, type, dst);
> > +      this->append(insn);
> > +    }
> > +
> >      /*! For all unary functions */
> >      void ALU1(Opcode opcode, Type type, Register dst, Register src) {
> >        const Instruction insn = gbe::ir::ALU1(opcode, type, dst, src);
> > diff --git a/backend/src/ir/instruction.cpp
> > b/backend/src/ir/instruction.cpp index 797552f..9c3331b 100644
> > --- a/backend/src/ir/instruction.cpp
> > +++ b/backend/src/ir/instruction.cpp
> > @@ -131,6 +131,17 @@ namespace ir {
> >        Register src[srcNum]; //!< Indices of the sources
> >      };
> >
> > +    /*! All 0-source arithmetic instructions */
> > +    class ALIGNED_INSTRUCTION NullaryInstruction : public
> NaryInstruction<0>
> > +    {
> > +    public:
> > +      NullaryInstruction(Opcode opcode, Type type, Register dst) {
> > +        this->opcode = opcode;
> > +        this->type = type;
> > +        this->dst[0] = dst;
> > +      }
> > +    };
> > +
> >      /*! All 1-source arithmetic instructions */
> >      class ALIGNED_INSTRUCTION UnaryInstruction : public NaryInstruction<1>
> >      {
> > @@ -1305,6 +1316,10 @@ namespace ir {
> >      }; \
> >    }
> >
> > +START_INTROSPECTION(NullaryInstruction)
> > +#include "ir/instruction.hxx"
> > +END_INTROSPECTION(NullaryInstruction)
> > +
> >  START_INTROSPECTION(UnaryInstruction)
> >  #include "ir/instruction.hxx"
> >  END_INTROSPECTION(UnaryInstruction)
> > @@ -1532,6 +1547,7 @@ END_FUNCTION(Instruction, Register)
> >      return reinterpret_cast<const internal::CLASS*>(this)->CALL; \
> >    }
> >
> > +DECL_MEM_FN(NullaryInstruction, Type, getType(void), getType())
> >  DECL_MEM_FN(UnaryInstruction, Type, getType(void), getType())
> > DECL_MEM_FN(BinaryInstruction, Type, getType(void), getType())
> > DECL_MEM_FN(BinaryInstruction, bool, commutes(void), commutes())
> @@
> > -1586,6 +1602,21 @@ DECL_MEM_FN(GetImageInfoInstruction, uint8_t,
> > getImageIndex(void), getImageIndex
> >    // Implements the emission functions
> >
> > //////////////////////////////////////////////////////////////////////
> > /////
> >
> > +  // For all nullary functions with given opcode  Instruction
> > + ALU0(Opcode opcode, Type type, Register dst) {
> > +    return internal::NullaryInstruction(opcode, type, dst).convert();
> > + }
> > +
> > +  // All unary functions
> > +#define DECL_EMIT_FUNCTION(NAME) \
> > +  Instruction NAME(Type type, Register dst) { \
> > +    return ALU0(OP_##NAME, type, dst);\
> > +  }
> > +
> > +  DECL_EMIT_FUNCTION(SIMD_SIZE)
> > +
> > +#undef DECL_EMIT_FUNCTION
> > +
> >    // For all unary functions with given opcode
> >    Instruction ALU1(Opcode opcode, Type type, Register dst, Register src) {
> >      return internal::UnaryInstruction(opcode, type, dst,
> > src).convert(); @@ -
> > 1645,6 +1676,7 @@ DECL_MEM_FN(GetImageInfoInstruction, uint8_t,
> > getImageIndex(void), getImageIndex
> >    DECL_EMIT_FUNCTION(RHADD)
> >    DECL_EMIT_FUNCTION(I64HADD)
> >    DECL_EMIT_FUNCTION(I64RHADD)
> > +  DECL_EMIT_FUNCTION(SIMD_SHUFFLE)
> >
> >  #undef DECL_EMIT_FUNCTION
> >
> > diff --git a/backend/src/ir/instruction.hpp
> > b/backend/src/ir/instruction.hpp index 24d27aa..6dd3e81 100644
> > --- a/backend/src/ir/instruction.hpp
> > +++ b/backend/src/ir/instruction.hpp
> > @@ -198,6 +198,15 @@ namespace ir {
> >    /*! Output the instruction string in the given stream */
> >    std::ostream &operator<< (std::ostream &out, const Instruction
> > &proxy);
> >
> > +  /*! Nullary instruction instructions are typed. */  class
> > + NullaryInstruction : public Instruction {
> > +  public:
> > +    /*! Get the type manipulated by the instruction */
> > +    Type getType(void) const;
> > +    /*! Return true if the given instruction is an instance of this class */
> > +    static bool isClassOf(const Instruction &insn);  };
> > +
> >    /*! Unary instructions are typed. dst and sources share the same type */
> >    class UnaryInstruction : public Instruction {
> >    public:
> > @@ -558,6 +567,12 @@ namespace ir {
> >    /// All emission functions
> >
> > //////////////////////////////////////////////////////////////////////
> > /////
> >
> > +  /*! alu0.type dst */
> > +  Instruction ALU0(Opcode opcode, Type type, Register dst);  /*!
> > + simd_size.type dst */  Instruction SIMD_SIZE(Type type, Register
> > + dst); /*! simd_id.type dst */  Instruction SIMD_ID(Type type,
> > + Register dst);
> >    /*! alu1.type dst src */
> >    Instruction ALU1(Opcode opcode, Type type, Register dst, Register src);
> >    /*! mov.type dst src */
> > @@ -670,6 +685,8 @@ namespace ir {
> >    Instruction GT(Type type, Register dst, Register src0, Register src1);
> >    /*! ord.type dst src0 src1 */
> >    Instruction ORD(Type type, Register dst, Register src0, Register
> > src1);
> > +  /*! simd_shuffle.type dst src0 src1 */  Instruction
> > + SIMD_SHUFFLE(Type type, Register dst, Register src0, Register src1);
> >    /*! BITCAST.{dstType <- srcType} dst src */
> >    Instruction BITCAST(Type dstType, Type srcType, Tuple dst, Tuple
> > src, uint8_t dstNum, uint8_t srcNum);
> >    /*! cvt.{dstType <- srcType} dst src */ diff --git
> > a/backend/src/ir/instruction.hxx b/backend/src/ir/instruction.hxx
> > index de4abfb..76269bd 100644
> > --- a/backend/src/ir/instruction.hxx
> > +++ b/backend/src/ir/instruction.hxx
> > @@ -25,6 +25,8 @@
> >   * \file instruction.hxx
> >   * \author Benjamin Segovia <benjamin.segovia at intel.com>
> >   */
> > +DECL_INSN(SIMD_SIZE, NullaryInstruction) DECL_INSN(SIMD_ID,
> > +NullaryInstruction)
> >  DECL_INSN(MOV, UnaryInstruction)
> >  DECL_INSN(COS, UnaryInstruction)
> >  DECL_INSN(SIN, UnaryInstruction)
> > @@ -57,6 +59,7 @@ DECL_INSN(BSB, BinaryInstruction)  DECL_INSN(OR,
> > BinaryInstruction)  DECL_INSN(XOR, BinaryInstruction)  DECL_INSN(AND,
> > BinaryInstruction)
> > +DECL_INSN(SIMD_SHUFFLE, BinaryInstruction)
> >  DECL_INSN(SEL, SelectInstruction)
> >  DECL_INSN(EQ, CompareInstruction)
> >  DECL_INSN(NE, CompareInstruction)
> > diff --git a/backend/src/ir/liveness.cpp b/backend/src/ir/liveness.cpp
> > index
> > 2b1ffdb..26c4129 100644
> > --- a/backend/src/ir/liveness.cpp
> > +++ b/backend/src/ir/liveness.cpp
> > @@ -66,6 +66,11 @@ namespace ir {
> >          const uint32_t srcNum = insn.getSrcNum();
> >          const uint32_t dstNum = insn.getDstNum();
> >          bool uniform = true;
> > +
> > +        //have no way to decide the dst uniform if there is no source
> > +        if (srcNum == 0)
> > +          uniform = false;
> > +
> >          for (uint32_t srcID = 0; srcID < srcNum; ++srcID) {
> >            const Register reg = insn.getSrc(srcID);
> >            if (!fn.isUniformRegister(reg)) diff --git
> > a/backend/src/ir/profile.cpp b/backend/src/ir/profile.cpp index
> > 4c272bd..55aedb4 100644
> > --- a/backend/src/ir/profile.cpp
> > +++ b/backend/src/ir/profile.cpp
> > @@ -43,6 +43,7 @@ namespace ir {
> >          "zero", "one",
> >          "retVal", "slm_offset",
> >          "printf_buffer_pointer", "printf_index_buffer_pointer",
> > +        "lane_id",
> >          "invalid"
> >      };
> >
> > @@ -86,6 +87,7 @@ namespace ir {
> >        DECL_NEW_REG(FAMILY_DWORD, slmoffset, 1);
> >        DECL_NEW_REG(FAMILY_DWORD, printfbptr, 1);
> >        DECL_NEW_REG(FAMILY_DWORD, printfiptr, 1);
> > +      DECL_NEW_REG(FAMILY_DWORD, laneid, 0);
> >        DECL_NEW_REG(FAMILY_DWORD, invalid, 1);
> >      }
> >  #undef DECL_NEW_REG
> > diff --git a/backend/src/ir/profile.hpp b/backend/src/ir/profile.hpp
> > index
> > 7259d9f..d310128 100644
> > --- a/backend/src/ir/profile.hpp
> > +++ b/backend/src/ir/profile.hpp
> > @@ -71,8 +71,9 @@ namespace ir {
> >      static const Register slmoffset = Register(27);  // Group's SLM
> > offset in total 64K SLM
> >      static const Register printfbptr = Register(28); // printf buffer address .
> >      static const Register printfiptr = Register(29); // printf index buffer
> address.
> > -    static const Register invalid = Register(30);  // used for valid comparation.
> > -    static const uint32_t regNum = 31;             // number of special registers
> > +    static const Register laneid = Register(30); // printf index buffer address.
> Actually, laneid is same as ocl::stackptr, can you reuse or rename it?
> 
> > +    static const Register invalid = Register(31);  // used for valid comparation.
> > +    static const uint32_t regNum = 32;             // number of special registers
> >      extern const char *specialRegMean[];           // special register name.
> >    } /* namespace ocl */
> >
> > diff --git a/backend/src/libocl/CMakeLists.txt
> > b/backend/src/libocl/CMakeLists.txt
> > index 16f00ee..623affc 100644
> > --- a/backend/src/libocl/CMakeLists.txt
> > +++ b/backend/src/libocl/CMakeLists.txt
> > @@ -90,7 +90,7 @@ MACRO(GENERATE_SOURCE_PY _mod)
> >  	)
> >  ENDMACRO(GENERATE_SOURCE_PY)
> >
> > -SET (OCL_PY_GENERATED_MODULES ocl_common ocl_relational
> ocl_integer
> > ocl_math)
> > +SET (OCL_PY_GENERATED_MODULES ocl_common ocl_relational
> > ocl_integer
> > +ocl_math ocl_simd)
> >  FOREACH(M ${OCL_PY_GENERATED_MODULES})
> >      GENERATE_HEADER_PY(${M})
> >      GENERATE_SOURCE_PY(${M})
> > diff --git a/backend/src/libocl/include/ocl.h
> > b/backend/src/libocl/include/ocl.h
> > index e886670..a53f4c0 100644
> > --- a/backend/src/libocl/include/ocl.h
> > +++ b/backend/src/libocl/include/ocl.h
> > @@ -30,6 +30,7 @@
> >  #include "ocl_image.h"
> >  #include "ocl_integer.h"
> >  #include "ocl_math.h"
> > +#include "ocl_simd.h"
> >  #include "ocl_misc.h"
> >  #include "ocl_printf.h"
> >  #include "ocl_relational.h"
> > diff --git a/backend/src/libocl/include/ocl_misc.h
> > b/backend/src/libocl/include/ocl_misc.h
> > index aa3f504..359025b 100644
> > --- a/backend/src/libocl/include/ocl_misc.h
> > +++ b/backend/src/libocl/include/ocl_misc.h
> > @@ -128,14 +128,6 @@ DEF(ulong)
> >  #undef DEC16
> >  #undef DEC16X
> >
> > -
> > -/* Temp to add the SIMD functions here. */ -
> > //////////////////////////////////////////////////////////////////////
> > ///////
> > -// SIMD level function
> > -/////////////////////////////////////////////////////////////////////
> > ////////
> > -short __gen_ocl_simd_any(short);
> > -short __gen_ocl_simd_all(short);
> > -
> >  struct time_stamp {
> >    // time tick
> >    ulong tick;
> > diff --git a/backend/src/libocl/script/ocl_simd.def
> > b/backend/src/libocl/script/ocl_simd.def
> > new file mode 100644
> > index 0000000..ccda619
> > --- /dev/null
> > +++ b/backend/src/libocl/script/ocl_simd.def
> > @@ -0,0 +1,4 @@
> > +##simd level functions
> > +floatn __gen_ocl_simd_shuffle(floatn x, uint c) intn
> > +__gen_ocl_simd_shuffle(intn x, uint c) uintn
> > +__gen_ocl_simd_shuffle(uintn x, uint c)
> > diff --git a/backend/src/libocl/tmpl/ocl_simd.tmpl.cl
> > b/backend/src/libocl/tmpl/ocl_simd.tmpl.cl
> > new file mode 100644
> > index 0000000..b9da5e2
> > --- /dev/null
> > +++ b/backend/src/libocl/tmpl/ocl_simd.tmpl.cl
> > @@ -0,0 +1,19 @@
> > +/*
> > + * Copyright @ 2015 Intel Corporation
> > + *
> > + * This library is free software; you can redistribute it and/or
> > + * modify it under the terms of the GNU Lesser General Public
> > + * License as published by the Free Software Foundation; either
> > + * version 2.1 of the License, or (at your option) any later version.
> > + *
> > + * This library is distributed in the hope that it will be useful,
> > + * but WITHOUT ANY WARRANTY; without even the implied warranty of
> > + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
> > GNU
> > + * Lesser General Public License for more details.
> > + *
> > + * You should have received a copy of the GNU Lesser General Public
> > + * License along with this library. If not, see
> > <http://www.gnu.org/licenses/>.
> > + *
> > + */
> > +
> > +#include "ocl_simd.h"
> > diff --git a/backend/src/libocl/tmpl/ocl_simd.tmpl.h
> > b/backend/src/libocl/tmpl/ocl_simd.tmpl.h
> > new file mode 100644
> > index 0000000..42afc7b
> > --- /dev/null
> > +++ b/backend/src/libocl/tmpl/ocl_simd.tmpl.h
> > @@ -0,0 +1,34 @@
> > +/*
> > + * Copyright © 2015 Intel Corporation
> > + *
> > + * This library is free software; you can redistribute it and/or
> > + * modify it under the terms of the GNU Lesser General Public
> > + * License as published by the Free Software Foundation; either
> > + * version 2.1 of the License, or (at your option) any later version.
> > + *
> > + * This library is distributed in the hope that it will be useful,
> > + * but WITHOUT ANY WARRANTY; without even the implied warranty of
> > + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
> > GNU
> > + * Lesser General Public License for more details.
> > + *
> > + * You should have received a copy of the GNU Lesser General Public
> > + * License along with this library. If not, see
> > <http://www.gnu.org/licenses/>.
> > + *
> > + */
> > +#ifndef __OCL_SIMD_H__
> > +#define __OCL_SIMD_H__
> > +
> > +#include "ocl_types.h"
> > +
> > +/////////////////////////////////////////////////////////////////////
> > +//
> > +//////
> > +// SIMD level function
> > +/////////////////////////////////////////////////////////////////////
> > +//
> > +//////
> > +short __gen_ocl_simd_any(short);
> > +short __gen_ocl_simd_all(short);
> > +
> > +uint __gen_ocl_get_simd_size(void);
> > +uint __gen_ocl_get_simd_id(void);
> > +
> > +OVERLOADABLE float __gen_ocl_simd_shuffle(float x, uint c);
> > +OVERLOADABLE int __gen_ocl_simd_shuffle(int x, uint c);
> OVERLOADABLE
> > +uint __gen_ocl_simd_shuffle(uint x, uint c);
> > diff --git a/backend/src/llvm/llvm_gen_backend.cpp
> > b/backend/src/llvm/llvm_gen_backend.cpp
> > index bf03a13..4fcb8bb 100644
> > --- a/backend/src/llvm/llvm_gen_backend.cpp
> > +++ b/backend/src/llvm/llvm_gen_backend.cpp
> > @@ -2790,10 +2790,17 @@ namespace gbe
> >        case GEN_OCL_CONV_F32_TO_F16:
> >        case GEN_OCL_SIMD_ANY:
> >        case GEN_OCL_SIMD_ALL:
> > +      case GEN_OCL_SIMD_SHUFFLE:
> >        case GEN_OCL_READ_TM:
> >        case GEN_OCL_REGION:
> >          this->newRegister(&I);
> >          break;
> > +      case GEN_OCL_SIMD_SIZE:
> > +        this->newRegister(&I, NULL, true);
> > +        break;
> > +      case GEN_OCL_SIMD_ID:
> > +        this->newRegister(&I, NULL, false);
> > +        break;
> >        case GEN_OCL_PRINTF:
> >          break;
> >        default:
> > @@ -3053,6 +3060,26 @@ namespace gbe
> >              ctx.ALU1(ir::OP_SIMD_ANY, ir::TYPE_S16, dst, src);
> >              break;
> >            }
> > +          case GEN_OCL_SIMD_SIZE:
> > +          {
> > +            const ir::Register dst = this->getRegister(&I);
> > +            ctx.ALU0(ir::OP_SIMD_SIZE, getType(ctx, I.getType()), dst);
> > +            break;
> > +          }
> > +          case GEN_OCL_SIMD_ID:
> > +          {
> > +            const ir::Register dst = this->getRegister(&I);
> > +            ctx.ALU0(ir::OP_SIMD_ID, getType(ctx, I.getType()), dst);
> > +            break;
> > +          }
> > +          case GEN_OCL_SIMD_SHUFFLE:
> > +          {
> > +            const ir::Register src0 = this->getRegister(*AI); ++AI;
> > +            const ir::Register src1 = this->getRegister(*AI); ++AI;
> > +            const ir::Register dst = this->getRegister(&I);
> > +            ctx.SIMD_SHUFFLE(getType(ctx, I.getType()), dst, src0, src1);
> > +            break;
> > +          }
> >            case GEN_OCL_READ_TM:
> >            {
> >              const ir::Register dst = this->getRegister(&I); diff
> > --git a/backend/src/llvm/llvm_gen_ocl_function.hxx
> > b/backend/src/llvm/llvm_gen_ocl_function.hxx
> > index 9536a3c..714a293 100644
> > --- a/backend/src/llvm/llvm_gen_ocl_function.hxx
> > +++ b/backend/src/llvm/llvm_gen_ocl_function.hxx
> > @@ -155,6 +155,10 @@ DECL_LLVM_GEN_FUNCTION(CONV_F32_TO_F16,
> > __gen_ocl_f32to16)  DECL_LLVM_GEN_FUNCTION(SIMD_ANY,
> > __gen_ocl_simd_any)  DECL_LLVM_GEN_FUNCTION(SIMD_ALL,
> > __gen_ocl_simd_all)
> >
> > +DECL_LLVM_GEN_FUNCTION(SIMD_SIZE, __gen_ocl_get_simd_size)
> > +DECL_LLVM_GEN_FUNCTION(SIMD_ID, __gen_ocl_get_simd_id)
> > +DECL_LLVM_GEN_FUNCTION(SIMD_SHUFFLE, __gen_ocl_simd_shuffle)
> > +
> >  DECL_LLVM_GEN_FUNCTION(READ_TM, __gen_ocl_read_tm)
> > DECL_LLVM_GEN_FUNCTION(REGION, __gen_ocl_region)
> >
> > diff --git a/src/cl_command_queue_gen7.c
> b/src/cl_command_queue_gen7.c
> > index 253c4f2..3f73de0 100644
> > --- a/src/cl_command_queue_gen7.c
> > +++ b/src/cl_command_queue_gen7.c
> > @@ -202,6 +202,14 @@ cl_curbe_fill(cl_kernel ker,
> >    UPLOAD(GBE_CURBE_WORK_DIM, work_dim);  #undef UPLOAD
> >
> > +  /* __gen_ocl_get_simd_id needs it */  if ((offset =
> > + interp_kernel_get_curbe_offset(ker->opaque,
> > GBE_CURBE_LANE_ID, 0)) >= 0) {
> > +    const uint32_t simd_sz = interp_kernel_get_simd_width(ker->opaque);
> > +    uint32_t *laneid = (uint32_t *) (ker->curbe + offset);
> > +    int32_t i;
> > +    for (i = 0; i < (int32_t) simd_sz; ++i) laneid[i] = i;  }
> > +
> >    /* Write identity for the stack pointer. This is required by the stack
> pointer
> >     * computation in the kernel
> >     */
> > --
> > 1.9.1
> >
> > _______________________________________________
> > Beignet mailing list
> > Beignet at lists.freedesktop.org
> > http://lists.freedesktop.org/mailman/listinfo/beignet


More information about the Beignet mailing list