[Mesa-dev] [PATCH 7/7] i965/fs: Add empirically-determined instruction latencies for gen7.

[Matt Turner]

On Fri, Dec 7, 2012 at 2:58 PM, Eric Anholt <eric at anholt.net> wrote:
> The limited performance testing I've done on this hasn't shown any
> statistically significant differences yet.
> ---
>  .../dri/i965/brw_fs_schedule_instructions.cpp      |  150 +++++++++++++++++++-
>  1 file changed, 147 insertions(+), 3 deletions(-)
>
> diff --git a/src/mesa/drivers/dri/i965/brw_fs_schedule_instructions.cpp b/src/mesa/drivers/dri/i965/brw_fs_schedule_instructions.cpp
> index 3623c13..f3f0079 100644
> --- a/src/mesa/drivers/dri/i965/brw_fs_schedule_instructions.cpp
> +++ b/src/mesa/drivers/dri/i965/brw_fs_schedule_instructions.cpp
> @@ -57,7 +57,7 @@ static bool debug = false;
>  class schedule_node : public exec_node
>  {
>  public:
> -   schedule_node(fs_inst *inst)
> +   schedule_node(fs_inst *inst, int gen)
>     {
>        this->inst = inst;
>        this->child_array_size = 0;
> @@ -67,10 +67,14 @@ public:
>        this->parent_count = 0;
>        this->unblocked_time = 0;
>
> -      set_latency_gen4();
> +      if (gen >= 7)
> +         set_latency_gen7();
> +      else
> +         set_latency_gen4();
>     }
>
>     void set_latency_gen4();
> +   void set_latency_gen7();
>
>     fs_inst *inst;
>     schedule_node **children;
> @@ -120,6 +124,146 @@ schedule_node::set_latency_gen4()
>     }
>  }
>
> +void
> +schedule_node::set_latency_gen7()
> +{
> +   switch (inst->opcode) {
> +   case BRW_OPCODE_MAD:
> +      /* 3 cycles (this is said to be 4 cycles sometimes depending on the
> +       * register numbers in the sources):
> +       * mad(8) g4<1>F g2.2<4,1,1>F.x  g2<4,1,1>F.x g2.1<4,1,1>F.x { align16 WE_normal 1Q };
> +       *
> +       * 20 cycles:
> +       * mad(8) g4<1>F g2.2<4,1,1>F.x  g2<4,1,1>F.x g2.1<4,1,1>F.x { align16 WE_normal 1Q };
> +       * mov(8) null   g4<4,4,1>F                      { align16 WE_normal 1Q };
> +       */
> +      latency = 17;
> +      break;
> +
> +   case SHADER_OPCODE_RCP:
> +      /* 2 cycles:
> +       * math inv(8) g4<1>F g2<0,1,0>F      null            { align1 WE_normal 1Q };
> +       *
> +       * 18 cycles:
> +       * math inv(8) g4<1>F g2<0,1,0>F      null            { align1 WE_normal 1Q };
> +       * mov(8)      null   g4<8,8,1>F                      { align1 WE_normal 1Q };
> +       *
> +       * Same for exp2, log2, rsq, sqrt, sin, cos.

Should those be in the switch as well?