[Mesa-dev] [PATCH 09/23] glsl: Distribute multiply over b2f

[Ian Romanick]

On 03/20/2015 01:58 PM, Ian Romanick wrote:
> From: Ian Romanick <ian.d.romanick at intel.com>
> 
> Convert cases like (x * bool(b)) to 'mix(0, x, b)'.
> 
> Note: This may hurt the code generated for GPUs that represent Boolean
> values using floating point.  shader-db doesn't play well with i915, so
> I haven't been able to check it.
> 
> On at least BDW shaders/warsow/85.shader_test is hurt by about 10%, so
> that may be worth investigating.

I think I've found the problem caused here.  There's one pattern in the
shader that used to end up with some MADs (4 instructions total), but
now it gets a small pile of hurt (6 instructions total).  And the shader
contains that same pattern 14 times in a row.  28 instructions is on the
right order of magnitude of the regression that I saw.

I have a couple ideas to fix this in the backend.  Watch for a patch on
Monday or Tuesday.

> v2: Use swizzle_if_required on opeands.  Fixes
> arb_texture_buffer_object-formats in debug builds.  Without this,
> ir_validate fails on that test.
> 
> Shader-db results:
> 
> GM45 (0x2A42):
> total instructions in shared programs: 3548093 -> 3545804 (-0.06%)
> instructions in affected programs:     213889 -> 211600 (-1.07%)
> helped:                                543
> HURT:                                  2
> 
> Iron Lake (0x0046):
> total instructions in shared programs: 4978454 -> 4975982 (-0.05%)
> instructions in affected programs:     226063 -> 223591 (-1.09%)
> helped:                                597
> HURT:                                  5
> 
> Sandy Bridge (0x0116):
> total instructions in shared programs: 6806814 -> 6803487 (-0.05%)
> instructions in affected programs:     447885 -> 444558 (-0.74%)
> helped:                                1612
> HURT:                                  36
> 
> Sandy Bridge (0x0116) NIR:
> total instructions in shared programs: 6813527 -> 6811992 (-0.02%)
> instructions in affected programs:     329020 -> 327485 (-0.47%)
> helped:                                1002
> HURT:                                  196
> 
> Ivy Bridge (0x0166):
> total instructions in shared programs: 6283080 -> 6279862 (-0.05%)
> instructions in affected programs:     421859 -> 418641 (-0.76%)
> helped:                                1592
> HURT:                                  36
> 
> Ivy Bridge (0x0166) NIR:
> total instructions in shared programs: 6319944 -> 6317148 (-0.04%)
> instructions in affected programs:     303221 -> 300425 (-0.92%)
> helped:                                998
> HURT:                                  176
> GAINED:                                4
> 
> Haswell (0x0426):
> total instructions in shared programs: 5766971 -> 5764623 (-0.04%)
> instructions in affected programs:     382796 -> 380448 (-0.61%)
> helped:                                1559
> HURT:                                  63
> 
> Haswell (0x0426) NIR:
> total instructions in shared programs: 5793258 -> 5792647 (-0.01%)
> instructions in affected programs:     276929 -> 276318 (-0.22%)
> helped:                                837
> HURT:                                  343
> GAINED:                                4
> 
> Broadwell (0x162E):
> total instructions in shared programs: 6813995 -> 6811377 (-0.04%)
> instructions in affected programs:     469734 -> 467116 (-0.56%)
> helped:                                1772
> HURT:                                  78
> LOST:                                  1
> 
> Broadwell (0x162E) NIR:
> total instructions in shared programs: 7009761 -> 7009142 (-0.01%)
> instructions in affected programs:     298433 -> 297814 (-0.21%)
> helped:                                866
> HURT:                                  373
> 
> Signed-off-by: Ian Romanick <ian.d.romanick at intel.com>
> ---
>  src/glsl/opt_algebraic.cpp | 22 ++++++++++++++++++++++
>  1 file changed, 22 insertions(+)
> 
> diff --git a/src/glsl/opt_algebraic.cpp b/src/glsl/opt_algebraic.cpp
> index 69c03ea..837b080 100644
> --- a/src/glsl/opt_algebraic.cpp
> +++ b/src/glsl/opt_algebraic.cpp
> @@ -560,6 +560,28 @@ ir_algebraic_visitor::handle_expression(ir_expression *ir)
>              }
>           }
>        }
> +
> +      /* If one of the multiplicands is an ir_unop_b2f, we can convert the
> +       * multiply to a simple csel.
> +       *
> +       *     x * b2f(condition) => mix( 0,  x, condition)
> +       */
> +      for (unsigned i = 0; i < 2; i++) {
> +         if (op_expr[i] == NULL)
> +            continue;
> +
> +         if (op_expr[i]->operation != ir_unop_b2f)
> +            continue;
> +
> +         /* swizzle_if_required is necessary on both operands.  The b2f could
> +          * be a scalar (common) with the other a vector, or the b2f could be
> +          * a vector with the other a scalar (as in piglit's
> +          * arb_texture_buffer_object-formats test).
> +          */
> +         return csel(swizzle_if_required(ir, op_expr[i]->operands[0]),
> +                     swizzle_if_required(ir, ir->operands[i ^ 1]),
> +                     ir_constant::zero(ir, ir->type));
> +      }
>        break;
>  
>     case ir_binop_div:
>