[Mesa-dev] [PATCH v4 (part2) 17/59] i965/fs/nir: implement unsized array's length calculation

[Matt Turner]

On Wed, Aug 5, 2015 at 1:30 AM, Iago Toral Quiroga <itoral at igalia.com> wrote:
> From: Samuel Iglesias Gonsalvez <siglesias at igalia.com>
>
> Signed-off-by: Samuel Iglesias Gonsalvez <siglesias at igalia.com>
> ---
>  src/mesa/drivers/dri/i965/brw_fs_nir.cpp | 49 ++++++++++++++++++++++++++++++++
>  1 file changed, 49 insertions(+)
>
> diff --git a/src/mesa/drivers/dri/i965/brw_fs_nir.cpp b/src/mesa/drivers/dri/i965/brw_fs_nir.cpp
> index e922a85..fde65e9 100644
> --- a/src/mesa/drivers/dri/i965/brw_fs_nir.cpp
> +++ b/src/mesa/drivers/dri/i965/brw_fs_nir.cpp
> @@ -1578,6 +1578,55 @@ fs_visitor::nir_emit_intrinsic(const fs_builder &bld, nir_intrinsic_instr *instr
>        emit_barrier();
>        break;
>
> +   case nir_intrinsic_ssbo_unsized_array_length: {
> +      nir_const_value *const_uniform_block = nir_src_as_const_value(instr->src[0]);
> +      unsigned ubo_index = const_uniform_block ? const_uniform_block->u[0] : 0;
> +      nir_const_value *const_offset_ir = nir_src_as_const_value(instr->src[1]);
> +      int const_offset = const_offset_ir ? const_offset_ir->u[0] : 0;
> +      nir_const_value *const_stride_ir = nir_src_as_const_value(instr->src[2]);
> +      unsigned unsized_array_stride = const_stride_ir ? const_stride_ir->u[0] : 1;
> +      int reg_width = dispatch_width / 8;
> +
> +      assert(shader->base.UniformBlocks[ubo_index].IsShaderStorage);
> +
> +      /* Set LOD = 0 */
> +      fs_reg source = fs_reg(0);
> +
> +      int mlen = 1 * reg_width;
> +      fs_reg src_payload = fs_reg(GRF, alloc.allocate(mlen),
> +                                  BRW_REGISTER_TYPE_UD);
> +      bld.LOAD_PAYLOAD(src_payload, &source, 1, 0);
> +
> +      fs_reg surf_index = fs_reg(prog_data->binding_table.ubo_start + ubo_index);
> +      fs_reg buffer_size = vgrf(glsl_type::int_type);
> +
> +      fs_inst *inst = bld.emit(FS_OPCODE_UNSIZED_ARRAY_LENGTH, buffer_size,
> +                               src_payload, surf_index);
> +      inst->header_size = 0;
> +      inst->mlen = mlen;
> +      inst->regs_written = 4 * reg_width;
> +      bld.emit(inst);
> +
> +      /* array.length() =
> +          max((buffer_object_size - offset_of_array) / stride_of_array, 0) */
> +      fs_reg temp = vgrf(glsl_type::float_type);
> +      bld.ADD(buffer_size, buffer_size, fs_reg(-const_offset));
> +      bld.MOV(temp, buffer_size);
> +
> +      assert(unsized_array_stride > 0);
> +
> +      fs_reg stride = fs_reg((float)1/unsized_array_stride);

Spaces around operators, and you can just write "1.0f"

> +      bld.MUL(temp, temp, stride);
> +      bld.MOV(dest, fs_reg(0));
> +      bld.CMP(bld.null_reg_f(), temp, fs_reg(0), BRW_CONDITIONAL_G);

Should be BRW_CONDITIONAL_GE, but as Jordan says you probably just
want emit_minmax:

   bld.emit_minmax(BRW_CONDITIONAL_GE, dest, temp, fs_reg(0.0f));

> +      bld.IF(BRW_PREDICATE_NORMAL);
> +      {
> +         bld.MOV(retype(dest, BRW_REGISTER_TYPE_D), temp);
> +      }
> +      bld.emit(BRW_OPCODE_ENDIF);
> +      break;
> +   }
> +
>     default:
>        unreachable("unknown intrinsic");
>     }
> --
> 1.9.1
>
> _______________________________________________
> mesa-dev mailing list
> mesa-dev at lists.freedesktop.org
> http://lists.freedesktop.org/mailman/listinfo/mesa-dev