[Mesa-dev] [PATCH 16/24] i965: Use LZD to implement nir_op_ifind_msb on Gen < 7

[Matt Turner]

On Wed, Jun 29, 2016 at 2:04 PM, Ian Romanick <idr at freedesktop.org> wrote:
> From: Ian Romanick <ian.d.romanick at intel.com>
>
> Signed-off-by: Ian Romanick <ian.d.romanick at intel.com>
> ---
>  src/mesa/drivers/dri/i965/brw_fs_nir.cpp   | 50 ++++++++++++++++++++++------
>  src/mesa/drivers/dri/i965/brw_vec4_nir.cpp | 52 +++++++++++++++++++++++-------
>  2 files changed, 81 insertions(+), 21 deletions(-)
>
> diff --git a/src/mesa/drivers/dri/i965/brw_fs_nir.cpp b/src/mesa/drivers/dri/i965/brw_fs_nir.cpp
> index f15bf3e..f8db28a 100644
> --- a/src/mesa/drivers/dri/i965/brw_fs_nir.cpp
> +++ b/src/mesa/drivers/dri/i965/brw_fs_nir.cpp
> @@ -623,8 +623,32 @@ fs_visitor::nir_emit_find_msb_using_lzd(const fs_builder &bld,
>                                          bool is_signed)
>  {
>     fs_inst *inst;
> +   fs_reg temp = src;
>
> -   bld.LZD(retype(result, BRW_REGISTER_TYPE_UD), src);
> +   if (is_signed) {
> +      /* LZD of an absolute value source almost always does the right
> +       * thing.  There are two problem values:
> +       *
> +       * * 0x80000000.  Since abs(0x80000000) == 0x80000000, LZD returns
> +       *   0.  However, findMSB(int(0x80000000)) == 30.
> +       *
> +       * * 0xffffffff.  Since abs(0xffffffff) == 1, LZD returns
> +       *   31.  Section 8.8 (Integer Functions) of the GLSL 4.50 spec says:
> +       *
> +       *    For a value of zero or negative one, -1 will be returned.
> +       *
> +       * For all negative number cases, including 0x80000000 and
> +       * 0xffffffff, the correct value is obtained from LZD if instead of

Interesting, both the G45 and IVB docs (I didn't check others, I
suspect they say the same) say

"If the source is signed, the abs source modifier must be used to
convert any negative source value to a positive value."

I suppose that means you're supposed to retype the src to UD below?

> +       * negating the (already negative) value the logical-not is used.  A
> +       * conditonal logical-not can be achieved in two instructions.
> +       */
> +      temp = vgrf(glsl_type::int_type);
> +
> +      bld.ASR(temp, src, brw_imm_d(31));
> +      bld.XOR(temp, temp, src);
> +   }
> +
> +   bld.LZD(retype(result, BRW_REGISTER_TYPE_UD), temp);