[Beignet] [patch v3 1/3] reimplement the LZD instruction in backend.

Zhigang Gong zhigang.gong at linux.intel.com
Tue Jan 27 00:27:03 PST 2015


On Mon, Jan 26, 2015 at 02:57:45PM +0800, xionghu.luo at intel.com wrote:
> From: Luo Xionghu <xionghu.luo at intel.com>
> 
> handle the byte/word/dword/qword input accordingly.
> 
> v2: fix build issue.
> v3: remove duplicate code and unnessesary code.
> 
> Signed-off-by: Luo Xionghu <xionghu.luo at intel.com>
> ---
>  backend/src/llvm/llvm_gen_backend.cpp | 71 ++++++++++++++++++++++++++++++++++-
>  1 file changed, 69 insertions(+), 2 deletions(-)
> 
> diff --git a/backend/src/llvm/llvm_gen_backend.cpp b/backend/src/llvm/llvm_gen_backend.cpp
> index 86030b9..7922ddb 100644
> --- a/backend/src/llvm/llvm_gen_backend.cpp
> +++ b/backend/src/llvm/llvm_gen_backend.cpp
> @@ -3286,10 +3286,77 @@ error:
>            break;
>            case Intrinsic::ctlz:
>            {
> -            ir::Type srcType = getType(ctx, I.getType());
> +            Type *llvmDstType = I.getType();
> +            ir::Type dstType = getType(ctx, llvmDstType);
> +            Type *llvmSrcType = I.getOperand(0)->getType();
> +            ir::Type srcType = getUnsignedType(ctx, llvmSrcType);
> +
>              const ir::Register dst = this->getRegister(&I);
>              const ir::Register src = this->getRegister(I.getOperand(0));
> -            ctx.ALU1(ir::OP_LZD, srcType, dst, src);
> +            int imm_value = 0;
> +            if(srcType == ir::TYPE_U16) {
> +              imm_value = 16;
> +            }else if(srcType == ir::TYPE_U8) {
> +              imm_value = 24;
> +            }else if(srcType == ir::TYPE_U64) {
> +              imm_value = 32;
> +            }
> +
> +            if(srcType == ir::TYPE_U16 || srcType == ir::TYPE_U8) {
> +              ir::ImmediateIndex imm;
> +              ir::Type tmpType = ir::TYPE_S32;
> +              imm = ctx.newIntegerImmediate(imm_value, tmpType);
> +              const ir::RegisterFamily family = getFamily(tmpType);
> +              const ir::Register immReg = ctx.reg(family);
> +              ctx.LOADI(ir::TYPE_S32, immReg, imm);
> +
> +              ir::Register tmp0 = ctx.reg(getFamily(tmpType));
> +              ir::Register tmp1 = ctx.reg(getFamily(tmpType));
> +              ir::Register tmp2 = ctx.reg(getFamily(tmpType));
> +              ctx.CVT(tmpType, srcType, tmp0, src);
> +              ctx.ALU1(ir::OP_LZD, tmpType, tmp1, tmp0);
> +              ctx.SUB(tmpType, tmp2, tmp1, immReg);
> +              ctx.CVT(dstType, tmpType, dst, tmp2);
> +            }
> +            else if(srcType == ir::TYPE_U64) {
> +              ir::ImmediateIndex imm;
> +              ir::Type tmpType = ir::TYPE_U32;
> +              imm = ctx.newIntegerImmediate(imm_value, srcType);
> +              const ir::RegisterFamily family = getFamily(srcType);
> +              const ir::Register immReg = ctx.reg(family);
> +              ctx.LOADI(ir::TYPE_S64, immReg, imm);
> +
> +              const ir::RegisterFamily tmpFamily = getFamily(tmpType);
> +              const ir::ImmediateIndex imm32 = ctx.newIntegerImmediate(32, tmpType);
> +              const ir::Register imm32Reg = ctx.reg(tmpFamily);
> +              ctx.LOADI(ir::TYPE_S32, imm32Reg, imm32);
> +
> +              ir::Register tmp0 = ctx.reg(getFamily(srcType));
> +              ir::Register tmp1 = ctx.reg(getFamily(tmpType));
> +              ir::Register tmp2 = ctx.reg(getFamily(tmpType));
> +              ir::Register tmp3 = ctx.reg(getFamily(tmpType));
> +              ir::Register tmp4 = ctx.reg(getFamily(tmpType));
> +              ir::Register tmp5 = ctx.reg(getFamily(tmpType));
> +              ir::Register tmp6 = ctx.reg(getFamily(tmpType));
Use too many TYPE_U64 temporary registers is not good.

> +              ir::Register cmp = ctx.reg(ir::FAMILY_BOOL);
> +
> +              ctx.SHR(srcType, tmp0, src, immReg);
The SHR on U64 type is very expensive on many gen backend. Please consider
the following way (in llvm IR format) to get the two integers of the U64 data:
  %1 = bitcast i64 %0 to <2 x i32>
  %2 = extractelement <2 x i32> %1, i32 0
  %3 = extractelement <2 x i32> %1, i32 1

The other part LGTM.

Thanks,
Zhigang Gong.

> +              ctx.CVT(tmpType, srcType, tmp1, tmp0);
> +
> +              ctx.ALU1(ir::OP_LZD, tmpType, tmp2, tmp1);
> +              ctx.LT(tmpType, cmp, tmp2, imm32Reg);
> +
> +              ctx.CVT(tmpType, srcType, tmp3, src);
> +              ctx.ALU1(ir::OP_LZD, tmpType, tmp4, tmp3);
> +              ctx.ADD(tmpType, tmp5, tmp4, imm32Reg);
> +
> +              ctx.SEL(tmpType, tmp6, cmp, tmp2, tmp5);
> +              ctx.CVT(dstType, tmpType, dst, tmp6);
> +            }
> +            else
> +            {
> +              ctx.ALU1(ir::OP_LZD, dstType, dst, src);
> +            }
>            }
>            break;
>            case Intrinsic::fma:
> -- 
> 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