[Beignet] [PATCH 5/8] Backend: Add the MADM function to gen8 encoder.

[Matt Turner]

On Tue, Sep 15, 2015 at 4:15 AM,  <junyan.he at inbox.com> wrote:
> From: Junyan He <junyan.he at linux.intel.com>
>
> Signed-off-by: Junyan He <junyan.he at linux.intel.com>
> ---
>  backend/src/backend/gen8_encoder.cpp | 56 ++++++++++++++++++++++++++++++++++++
>  backend/src/backend/gen8_encoder.hpp |  2 ++
>  backend/src/backend/gen_defs.hpp     |  2 ++
>  3 files changed, 60 insertions(+)
>
> diff --git a/backend/src/backend/gen8_encoder.cpp b/backend/src/backend/gen8_encoder.cpp
> index 0af27a3..002a8b5 100644
> --- a/backend/src/backend/gen8_encoder.cpp
> +++ b/backend/src/backend/gen8_encoder.cpp
> @@ -591,4 +591,60 @@ namespace gbe
>       this->setSrc0WithAcc(insn, src0, src0Acc);
>       this->setSrc1WithAcc(insn, src1, src1Acc);
>    }
> +
> +  void Gen8Encoder::MADM(GenRegister dst, GenRegister src0, GenRegister src1, GenRegister src2,
> +      uint32_t dstAcc, uint32_t src0Acc, uint32_t src1Acc, uint32_t src2Acc)
> +  {
> +    GenNativeInstruction *insn = this->next(GEN_OPCODE_MADM);
> +    Gen8NativeInstruction *gen8_insn = &insn->gen8_insn;
> +    assert(dst.file == GEN_GENERAL_REGISTER_FILE);
> +    assert(src0.file == GEN_GENERAL_REGISTER_FILE);
> +    assert(src1.file == GEN_GENERAL_REGISTER_FILE);
> +    assert(src2.file == GEN_GENERAL_REGISTER_FILE);
> +    assert(dst.hstride == GEN_HORIZONTAL_STRIDE_1 || dst.hstride == GEN_HORIZONTAL_STRIDE_0);
> +    assert(src0.type == GEN_TYPE_DF || src0.type == GEN_TYPE_F);
> +    assert(src0.type == dst.type);
> +    assert(src0.type == src1.type);
> +    assert(src0.type == src2.type);
> +    int32_t dataType = src0.type == GEN_TYPE_DF ? 3 : 0;
> +
> +    this->setHeader(insn);
> +    gen8_insn->bits1.da3srcacc.dest_reg_nr = dst.nr;
> +    gen8_insn->bits1.da3srcacc.dest_subreg_nr = dst.subnr / 16;
> +    gen8_insn->bits1.da3srcacc.dst_specal_acc = dstAcc;
> +    gen8_insn->bits1.da3srcacc.src_type = dataType;
> +    gen8_insn->bits1.da3srcacc.dest_type = dataType;
> +    gen8_insn->header.access_mode = GEN_ALIGN_16;
> +
> +    assert(src0.file == GEN_GENERAL_REGISTER_FILE);
> +    assert(src0.address_mode == GEN_ADDRESS_DIRECT);
> +    assert(src0.nr < 128);
> +    gen8_insn->bits2.da3srcacc.src0_specal_acc = src0Acc;
> +    gen8_insn->bits2.da3srcacc.src0_subreg_nr = src0.subnr / 4 ;
> +    gen8_insn->bits2.da3srcacc.src0_reg_nr = src0.nr;
> +    gen8_insn->bits1.da3srcacc.src0_abs = src0.absolute;
> +    gen8_insn->bits1.da3srcacc.src0_negate = src0.negation;
> +    gen8_insn->bits2.da3srcacc.src0_rep_ctrl = src0.vstride == GEN_VERTICAL_STRIDE_0;
> +
> +    assert(src1.file == GEN_GENERAL_REGISTER_FILE);
> +    assert(src1.address_mode == GEN_ADDRESS_DIRECT);
> +    assert(src1.nr < 128);
> +    gen8_insn->bits2.da3srcacc.src1_specal_acc = src1Acc;
> +    gen8_insn->bits2.da3srcacc.src1_subreg_nr_low = (src1.subnr / 4) & 0x3;
> +    gen8_insn->bits3.da3srcacc.src1_subreg_nr_high = (src1.subnr / 4) >> 2;
> +    gen8_insn->bits2.da3srcacc.src1_rep_ctrl = src1.vstride == GEN_VERTICAL_STRIDE_0;
> +    gen8_insn->bits3.da3srcacc.src1_reg_nr = src1.nr;
> +    gen8_insn->bits1.da3srcacc.src1_abs = src1.absolute;
> +    gen8_insn->bits1.da3srcacc.src1_negate = src1.negation;
> +
> +    assert(src2.file == GEN_GENERAL_REGISTER_FILE);
> +    assert(src2.address_mode == GEN_ADDRESS_DIRECT);
> +    assert(src2.nr < 128);
> +    gen8_insn->bits3.da3srcacc.src2_specal_acc = src2Acc;
> +    gen8_insn->bits3.da3srcacc.src2_subreg_nr = src2.subnr / 4;
> +    gen8_insn->bits3.da3srcacc.src2_rep_ctrl = src2.vstride == GEN_VERTICAL_STRIDE_0;
> +    gen8_insn->bits3.da3srcacc.src2_reg_nr = src2.nr;
> +    gen8_insn->bits1.da3srcacc.src2_abs = src2.absolute;
> +    gen8_insn->bits1.da3srcacc.src2_negate = src2.negation;
> +  }
>  } /* End of the name space. */
> diff --git a/backend/src/backend/gen8_encoder.hpp b/backend/src/backend/gen8_encoder.hpp
> index 53ec3d1..8e7939b 100644
> --- a/backend/src/backend/gen8_encoder.hpp
> +++ b/backend/src/backend/gen8_encoder.hpp
> @@ -74,6 +74,8 @@ namespace gbe
>
>      void MATH_WITH_ACC(GenRegister dst, uint32_t function, GenRegister src0, GenRegister src1,
>                         uint32_t dstAcc, uint32_t src0Acc, uint32_t src1Acc);
> +    void MADM(GenRegister dst, GenRegister src0, GenRegister src1, GenRegister src2,
> +              uint32_t dstAcc, uint32_t src0Acc, uint32_t src1Acc, uint32_t src2Acc);
>    };
>  }
>  #endif /* __GBE_GEN8_ENCODER_HPP__ */
> diff --git a/backend/src/backend/gen_defs.hpp b/backend/src/backend/gen_defs.hpp
> index a1bd8dd..1b550ac 100644
> --- a/backend/src/backend/gen_defs.hpp
> +++ b/backend/src/backend/gen_defs.hpp
> @@ -174,6 +174,8 @@ enum opcode {
>    GEN_OPCODE_LINE = 89,
>    GEN_OPCODE_PLN = 90,
>    GEN_OPCODE_MAD = 91,
> +  GEN_OPCODE_LRP = 92,

Unrelated to the main purpose of the patch: Do I understand correctly
that Beignet does not emit the LRP instruction?

If not, I'm curious why not? It maps pretty well to the mix() function
(just reverse the argument order), but it does not handle infinities
properly [0], which may be a more severe limitation with OpenCL.

At least with OpenGL shaders, mix() is a pretty common operation so
LRP is very useful.

[0] https://bugs.freedesktop.org/show_bug.cgi?id=89647