[Mesa-dev] [PATCH 2/2] i965: Emit smarter code for b2i of a comparison

[Matt Turner]

On Tue, Jun 10, 2014 at 7:13 PM, Ian Romanick <idr at freedesktop.org> wrote:
> From: Ian Romanick <ian.d.romanick at intel.com>
>
> Previously we would emit the comparison, emit an AND to mask off extra
> bits from the comparison result, then convert the result to float.  Now,
> do the comparison, then use a cleverly constructed SEL to pick either
> 0.0f or 1.0f.
>
> No piglit regressions on Ivybridge.
>
> NOTE: I have not yet tested actual application performance.  I'll do
> that tomorrow.
>
> total instructions in shared programs: 1642311 -> 1639449 (-0.17%)
> instructions in affected programs:     136533 -> 133671 (-2.10%)
> GAINED:                                0
> LOST:                                  0
>
> Programs that are affected appear to save between 1 and 5 instuctions
> (just by skimming the output from shader-db report.py.
>
> Signed-off-by: Ian Romanick <ian.d.romanick at intel.com>
> ---
>  src/mesa/drivers/dri/i965/brw_vec4.h           |  1 +
>  src/mesa/drivers/dri/i965/brw_vec4_visitor.cpp | 47 ++++++++++++++++++++++++++
>  2 files changed, 48 insertions(+)
>
> diff --git a/src/mesa/drivers/dri/i965/brw_vec4.h b/src/mesa/drivers/dri/i965/brw_vec4.h
> index c2bbd68..27e1e39 100644
> --- a/src/mesa/drivers/dri/i965/brw_vec4.h
> +++ b/src/mesa/drivers/dri/i965/brw_vec4.h
> @@ -587,6 +587,7 @@ public:
>
>     bool try_emit_sat(ir_expression *ir);
>     bool try_emit_mad(ir_expression *ir);
> +   bool try_emit_b2f_of_compare(ir_expression *ir);
>     void resolve_ud_negate(src_reg *reg);
>
>     src_reg get_timestamp();
> diff --git a/src/mesa/drivers/dri/i965/brw_vec4_visitor.cpp b/src/mesa/drivers/dri/i965/brw_vec4_visitor.cpp
> index 249072c..902f02d 100644
> --- a/src/mesa/drivers/dri/i965/brw_vec4_visitor.cpp
> +++ b/src/mesa/drivers/dri/i965/brw_vec4_visitor.cpp
> @@ -1126,6 +1126,48 @@ vec4_visitor::try_emit_mad(ir_expression *ir)
>     return true;
>  }
>
> +bool
> +vec4_visitor::try_emit_b2f_of_compare(ir_expression *ir)
> +{
> +   ir_expression *const cmp = ir->operands[0]->as_expression();
> +
> +   if (cmp == NULL)
> +      return false;
> +
> +   switch (cmp->operation) {
> +   case ir_binop_less:
> +   case ir_binop_greater:
> +   case ir_binop_lequal:
> +   case ir_binop_gequal:
> +   case ir_binop_equal:
> +   case ir_binop_nequal:
> +      break;
> +
> +   default:
> +      return false;
> +   }
> +
> +   cmp->operands[0]->accept(this);
> +   const src_reg cmp_src0 = this->result;
> +
> +   cmp->operands[1]->accept(this);
> +   const src_reg cmp_src1 = this->result;
> +
> +   this->result = src_reg(this, ir->type);
> +
> +   emit(CMP(dst_reg(this->result), cmp_src0, cmp_src1,
> +            brw_conditional_for_comparison(cmp->operation)));
> +
> +   /* If the comparison is false, this->result will just happen to be zero.
> +    */
> +   vec4_instruction *const inst = emit(BRW_OPCODE_SEL, dst_reg(this->result),
> +                                       fix_3src_operand(src_reg(1.0f)),

No need for fix_3src_operand(). That's only needed for things like MAD and LRP.

Otherwise,

Reviewed-by: Matt Turner <mattst88 at gmail.com>

I'd make the commit summary i965/vec4: ... for this and the other patch.

Welcome to the backend compiler! :)