[Mesa-dev] [PATCH 13/28] nir: take into account rounding modes in conversions

Connor Abbott cwabbott0 at gmail.com
Wed Dec 5 18:21:29 UTC 2018


On Wed, Dec 5, 2018 at 4:56 PM Samuel Iglesias Gonsálvez
<siglesias at igalia.com> wrote:
>
> Signed-off-by: Samuel Iglesias Gonsálvez <siglesias at igalia.com>
> ---
>  src/compiler/nir/nir.h                       | 15 +++++++
>  src/compiler/nir/nir_constant_expressions.py | 46 +++++++++++++++++---
>  src/compiler/spirv/vtn_alu.c                 | 16 ++++++-
>  3 files changed, 71 insertions(+), 6 deletions(-)
>
> diff --git a/src/compiler/nir/nir.h b/src/compiler/nir/nir.h
> index 65a1f60c3c6..f22ac13b2ac 100644
> --- a/src/compiler/nir/nir.h
> +++ b/src/compiler/nir/nir.h
> @@ -866,6 +866,21 @@ nir_get_nir_type_for_glsl_type(const struct glsl_type *type)
>  nir_op nir_type_conversion_op(nir_alu_type src, nir_alu_type dst,
>                                nir_rounding_mode rnd);
>
> +static inline nir_rounding_mode
> +nir_get_rounding_mode_from_float_controls(unsigned rounding_mode,
> +                                          nir_alu_type type)
> +{
> +   if (nir_alu_type_get_base_type(type) != nir_type_float)
> +      return nir_rounding_mode_undef;
> +
> +   if (rounding_mode & SHADER_ROUNDING_MODE_RTZ)
> +      return nir_rounding_mode_rtz;
> +   if (rounding_mode & SHADER_ROUNDING_MODE_RTE)
> +      return nir_rounding_mode_rtne;
> +
> +   return nir_rounding_mode_undef;
> +}
> +
>  typedef enum {
>     NIR_OP_IS_COMMUTATIVE = (1 << 0),
>     NIR_OP_IS_ASSOCIATIVE = (1 << 1),
> diff --git a/src/compiler/nir/nir_constant_expressions.py b/src/compiler/nir/nir_constant_expressions.py
> index 118af9f7818..a9af1bd233d 100644
> --- a/src/compiler/nir/nir_constant_expressions.py
> +++ b/src/compiler/nir/nir_constant_expressions.py
> @@ -79,6 +79,7 @@ template = """\
>  #include <math.h>
>  #include "util/rounding.h" /* for _mesa_roundeven */
>  #include "util/half_float.h"
> +#include "util/double.h"
>  #include "nir_constant_expressions.h"
>
>  /**
> @@ -300,7 +301,15 @@ struct bool32_vec {
>           % elif input_types[j] == "float16":
>              _mesa_half_to_float(_src[${j}].u16[${k}]),
>           % else:
> -            _src[${j}].${get_const_field(input_types[j])}[${k}],
> +            % if ("rtne" in op.name) and ("float" in input_types[j]) and ("int" in output_type):

This should be done in the opcode definition, not buried inside
internal nir_constant_expressions code, unless there's a very good
reason. The whole reason the constant-folding code is put next to the
opcode is to document what it does, and you're defeating that here.

> +               % if "float32" in input_types[j]:
> +                  _mesa_roundevenf(_src[${j}].${get_const_field(input_types[j])}[${k}]),
> +               % else:
> +                  _mesa_roundeven(_src[${j}].${get_const_field(input_types[j])}[${k}]),
> +               % endif
> +            % else:
> +               _src[${j}].${get_const_field(input_types[j])}[${k}],
> +            % endif
>           % endif
>        % endfor
>        % for k in range(op.input_sizes[j], 4):
> @@ -328,8 +337,27 @@ struct bool32_vec {
>                 const float src${j} =
>                    _mesa_half_to_float(_src[${j}].u16[_i]);
>              % else:
> -               const ${input_types[j]}_t src${j} =
> -                  _src[${j}].${get_const_field(input_types[j])}[_i];
> +               % if ("rtne" in op.name) and ("float" in input_types[j]) and ("int" in output_type):
> +                  % if "float32" in input_types[j]:
> +                     const ${input_types[j]}_t src${j} =
> +                        _mesa_roundevenf(_src[${j}].${get_const_field(input_types[j])}[_i]);
> +                  % else:
> +                     const ${input_types[j]}_t src${j} =
> +                        _mesa_roundeven(_src[${j}].${get_const_field(input_types[j])}[_i]);
> +
> +                  % endif
> +               % elif ("float64" in input_types[j]) and ("float32" in output_type):
> +                  % if ("rtz" in op.name):
> +                     const ${input_types[j]}_t src${j} =
> +                        _mesa_double_to_float_rtz(_src[${j}].${get_const_field(input_types[j])}[_i]);
> +                  % else:
> +                     const ${input_types[j]}_t src${j} =
> +                        _mesa_double_to_float_rtne(_src[${j}].${get_const_field(input_types[j])}[_i]);
> +                  % endif
> +               % else:
> +                  const ${input_types[j]}_t src${j} =
> +                     _src[${j}].${get_const_field(input_types[j])}[_i];
> +               % endif
>              % endif
>           % endfor
>
> @@ -350,7 +378,11 @@ struct bool32_vec {
>              ## Sanitize the C value to a proper NIR bool
>              _dst_val.u32[_i] = dst ? NIR_TRUE : NIR_FALSE;
>           % elif output_type == "float16":
> -            _dst_val.u16[_i] = _mesa_float_to_half(dst);
> +            % if "rtz" in op.name:
> +               _dst_val.u16[_i] = _mesa_float_to_float16_rtz(dst);
> +            % else:
> +               _dst_val.u16[_i] = _mesa_float_to_float16_rtne(dst);
> +            % endif
>           % else:
>              _dst_val.${get_const_field(output_type)}[_i] = dst;
>           % endif
> @@ -379,7 +411,11 @@ struct bool32_vec {
>              ## Sanitize the C value to a proper NIR bool
>              _dst_val.u32[${k}] = dst.${"xyzw"[k]} ? NIR_TRUE : NIR_FALSE;
>           % elif output_type == "float16":
> -            _dst_val.u16[${k}] = _mesa_float_to_half(dst.${"xyzw"[k]});
> +            % if "rtz" in op.name:
> +               _dst_val.u16[${k}] = _mesa_float_to_float16_rtz(dst.${"xyzw"[k]});
> +            % else:
> +               _dst_val.u16[${k}] = _mesa_float_to_float16_rtne(dst.${"xyzw"[k]});
> +            % endif
>           % else:
>              _dst_val.${get_const_field(output_type)}[${k}] = dst.${"xyzw"[k]};
>           % endif
> diff --git a/src/compiler/spirv/vtn_alu.c b/src/compiler/spirv/vtn_alu.c
> index 629b57560ca..fba2b031c09 100644
> --- a/src/compiler/spirv/vtn_alu.c
> +++ b/src/compiler/spirv/vtn_alu.c
> @@ -329,7 +329,12 @@ vtn_nir_alu_op_for_spirv_opcode(struct vtn_builder *b,
>        }
>        src_type |= src_bit_size;
>        dst_type |= dst_bit_size;
> -      return nir_type_conversion_op(src_type, dst_type, nir_rounding_mode_undef);
> +      unsigned float_controls =
> +         b->shader->info.shader_float_controls_execution_mode;
> +      nir_rounding_mode rounding_mode =
> +         nir_get_rounding_mode_from_float_controls(float_controls,
> +                                                   src_type);
> +      return nir_type_conversion_op(src_type, dst_type, rounding_mode);
>     }
>     /* Derivatives: */
>     case SpvOpDPdx:         return nir_op_fddx;
> @@ -590,12 +595,21 @@ vtn_handle_alu(struct vtn_builder *b, SpvOp opcode,
>                                           glsl_get_bit_size(type));
>        break;
>
> +   case SpvOpConvertFToS:
> +   case SpvOpConvertFToU:
>     case SpvOpFConvert: {
>        nir_alu_type src_alu_type = nir_get_nir_type_for_glsl_type(vtn_src[0]->type);
>        nir_alu_type dst_alu_type = nir_get_nir_type_for_glsl_type(type);
>        nir_rounding_mode rounding_mode = nir_rounding_mode_undef;
> +      unsigned float_controls = b->shader->info.shader_float_controls_execution_mode;
>
>        vtn_foreach_decoration(b, val, handle_rounding_mode, &rounding_mode);
> +
> +      if (rounding_mode == nir_rounding_mode_undef && float_controls) {
> +         rounding_mode =
> +            nir_get_rounding_mode_from_float_controls(float_controls,
> +                                                      src_alu_type);
> +      }
>        nir_op op = nir_type_conversion_op(src_alu_type, dst_alu_type, rounding_mode);
>
>        val->ssa->def = nir_build_alu(&b->nb, op, src[0], src[1], NULL, NULL);
> --
> 2.19.1
>
> _______________________________________________
> mesa-dev mailing list
> mesa-dev at lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/mesa-dev


More information about the mesa-dev mailing list