[Mesa-dev] [PATCH 2/7] nir/int64: Add some more lowering helpers

Matt Turner mattst88 at gmail.com
Wed Oct 17 00:51:14 UTC 2018


On Sun, Oct 14, 2018 at 3:58 PM Jason Ekstrand <jason at jlekstrand.net> wrote:
>
> On October 14, 2018 17:12:34 Matt Turner <mattst88 at gmail.com> wrote:
>
> > From: Jason Ekstrand <jason.ekstrand at intel.com>
> >
> > [mattst88]: Found in an old branch of Jason's.
> >
> > Jason implemented: inot, iand, ior, iadd, isub, ineg, iabs, compare,
> >                   imin, imax, umin, umax
> > Matt implemented:  ixor, imov, bcsel
> > ---
> > src/compiler/nir/nir_lower_int64.c | 186 +++++++++++++++++++++++++++++++++++++
> > 1 file changed, 186 insertions(+)
> >
> > diff --git a/src/compiler/nir/nir_lower_int64.c
> > b/src/compiler/nir/nir_lower_int64.c
> > index 0d7f165b406..6b269830801 100644
> > --- a/src/compiler/nir/nir_lower_int64.c
> > +++ b/src/compiler/nir/nir_lower_int64.c
> > @@ -24,6 +24,192 @@
> > #include "nir.h"
> > #include "nir_builder.h"
> >
> > +static nir_ssa_def *
> > +lower_imov64(nir_builder *b, nir_ssa_def *x)
> > +{
> > +   nir_ssa_def *x_lo = nir_unpack_64_2x32_split_x(b, x);
> > +   nir_ssa_def *x_hi = nir_unpack_64_2x32_split_y(b, x);
> > +
> > +   return nir_pack_64_2x32_split(b, nir_imov(b, x_lo), nir_imov(b, x_hi));
>
> You don't really need the movs...

Thanks. I think that was really a copy-and-paste-and-replace mistake.

> > +}
> > +
> > +static nir_ssa_def *
> > +lower_bcsel64(nir_builder *b, nir_ssa_def *cond, nir_ssa_def *x,
> > nir_ssa_def *y)
> > +{
> > +   nir_ssa_def *x_lo = nir_unpack_64_2x32_split_x(b, x);
> > +   nir_ssa_def *x_hi = nir_unpack_64_2x32_split_y(b, x);
> > +   nir_ssa_def *y_lo = nir_unpack_64_2x32_split_x(b, y);
> > +   nir_ssa_def *y_hi = nir_unpack_64_2x32_split_y(b, y);
> > +
> > +   return nir_pack_64_2x32_split(b, nir_bcsel(b, cond, x_lo, y_lo),
> > +                                    nir_bcsel(b, cond, x_hi, y_hi));
> > +}
> > +
> > +static nir_ssa_def *
> > +lower_inot64(nir_builder *b, nir_ssa_def *x)
> > +{
> > +   nir_ssa_def *x_lo = nir_unpack_64_2x32_split_x(b, x);
> > +   nir_ssa_def *x_hi = nir_unpack_64_2x32_split_y(b, x);
> > +
> > +   return nir_pack_64_2x32_split(b, nir_inot(b, x_lo), nir_inot(b, x_hi));
> > +}
> > +
> > +static nir_ssa_def *
> > +lower_iand64(nir_builder *b, nir_ssa_def *x, nir_ssa_def *y)
> > +{
> > +   nir_ssa_def *x_lo = nir_unpack_64_2x32_split_x(b, x);
> > +   nir_ssa_def *x_hi = nir_unpack_64_2x32_split_y(b, x);
> > +   nir_ssa_def *y_lo = nir_unpack_64_2x32_split_x(b, y);
> > +   nir_ssa_def *y_hi = nir_unpack_64_2x32_split_y(b, y);
> > +
> > +   return nir_pack_64_2x32_split(b, nir_iand(b, x_lo, y_lo),
> > +                                    nir_iand(b, x_hi, y_hi));
> > +}
> > +
> > +static nir_ssa_def *
> > +lower_ior64(nir_builder *b, nir_ssa_def *x, nir_ssa_def *y)
> > +{
> > +   nir_ssa_def *x_lo = nir_unpack_64_2x32_split_x(b, x);
> > +   nir_ssa_def *x_hi = nir_unpack_64_2x32_split_y(b, x);
> > +   nir_ssa_def *y_lo = nir_unpack_64_2x32_split_x(b, y);
> > +   nir_ssa_def *y_hi = nir_unpack_64_2x32_split_y(b, y);
> > +
> > +   return nir_pack_64_2x32_split(b, nir_ior(b, x_lo, y_lo),
> > +                                    nir_ior(b, x_hi, y_hi));
> > +}
> > +
> > +static nir_ssa_def *
> > +lower_ixor64(nir_builder *b, nir_ssa_def *x, nir_ssa_def *y)
> > +{
> > +   nir_ssa_def *x_lo = nir_unpack_64_2x32_split_x(b, x);
> > +   nir_ssa_def *x_hi = nir_unpack_64_2x32_split_y(b, x);
> > +   nir_ssa_def *y_lo = nir_unpack_64_2x32_split_x(b, y);
> > +   nir_ssa_def *y_hi = nir_unpack_64_2x32_split_y(b, y);
> > +
> > +   return nir_pack_64_2x32_split(b, nir_ixor(b, x_lo, y_lo),
> > +                                    nir_ixor(b, x_hi, y_hi));
> > +}
> > +
> > +static nir_ssa_def *
> > +lower_iadd64(nir_builder *b, nir_ssa_def *x, nir_ssa_def *y)
> > +{
> > +   nir_ssa_def *x_lo = nir_unpack_64_2x32_split_x(b, x);
> > +   nir_ssa_def *x_hi = nir_unpack_64_2x32_split_y(b, x);
> > +   nir_ssa_def *y_lo = nir_unpack_64_2x32_split_x(b, y);
> > +   nir_ssa_def *y_hi = nir_unpack_64_2x32_split_y(b, y);
> > +
> > +   nir_ssa_def *res_lo = nir_iadd(b, x_lo, y_lo);
> > +   nir_ssa_def *carry = nir_b2i(b, nir_ult(b, res_lo, x_lo));
> > +   nir_ssa_def *res_hi = nir_iadd(b, carry, nir_iadd(b, x_hi, y_hi));
> > +
> > +   return nir_pack_64_2x32_split(b, res_lo, res_hi);
> > +}
> > +
> > +static nir_ssa_def *
> > +lower_isub64(nir_builder *b, nir_ssa_def *x, nir_ssa_def *y)
> > +{
> > +   nir_ssa_def *x_lo = nir_unpack_64_2x32_split_x(b, x);
> > +   nir_ssa_def *x_hi = nir_unpack_64_2x32_split_y(b, x);
> > +   nir_ssa_def *y_lo = nir_unpack_64_2x32_split_x(b, y);
> > +   nir_ssa_def *y_hi = nir_unpack_64_2x32_split_y(b, y);
> > +
> > +   nir_ssa_def *res_lo = nir_isub(b, x_lo, y_lo);
> > +   /* In NIR, true is represented by ~0 which is -1 */
>
> We've had discussions (had some at XDC this year) about changing booleans
> to one-bit which would break this.  Doing b2i would be safer but this does
> work for now.
>
> > +   nir_ssa_def *borrow = nir_ult(b, x_lo, y_lo);
> > +   nir_ssa_def *res_hi = nir_iadd(b, nir_isub(b, x_hi, y_hi), borrow);
> > +
> > +   return nir_pack_64_2x32_split(b, res_lo, res_hi);
> > +}
> > +
> > +static nir_ssa_def *
> > +lower_ineg64(nir_builder *b, nir_ssa_def *x)
> > +{
> > +   /* Since isub is the same number of instructions (with better dependencies)
> > +    * as iadd, subtraction is actually more efficient for ineg than the usual
> > +    * 2's complement "flip the bits and add one".
> > +    */
> > +   return lower_isub64(b, nir_imm_int64(b, 0), x);
> > +}
> > +
> > +static nir_ssa_def *
> > +lower_iabs64(nir_builder *b, nir_ssa_def *x)
> > +{
> > +   nir_ssa_def *x_hi = nir_unpack_64_2x32_split_y(b, x);
> > +   nir_ssa_def *x_is_neg = nir_ilt(b, x_hi, nir_imm_int(b, 0));
> > +   return nir_bcsel(b, x_is_neg, lower_ineg64(b, x), x);
>
> lower_bcsel?  Or, since we're depending on this running multiple times,
> just nir_ineg?  I go back and forth on whether a pass like this should run
> in a loop or be smart enough to lower intermediate bits on the fly.  We
> should probably pick one.

Agreed, I don't love it. A later patch in the series calls
nir_lower_int64() in a loop so I suppose for now I'll rely on that and
use nir_ineg directly.

> > +}
> > +
> > +static nir_ssa_def *
> > +lower_int64_compare(nir_builder *b, nir_op op, nir_ssa_def *x, nir_ssa_def *y)
> > +{
> > +   nir_ssa_def *x_lo = nir_unpack_64_2x32_split_x(b, x);
> > +   nir_ssa_def *x_hi = nir_unpack_64_2x32_split_y(b, x);
> > +   nir_ssa_def *y_lo = nir_unpack_64_2x32_split_x(b, y);
> > +   nir_ssa_def *y_hi = nir_unpack_64_2x32_split_y(b, y);
> > +
> > +   switch (op) {
> > +   case nir_op_ieq:
> > +      return nir_iand(b, nir_ieq(b, x_hi, y_hi), nir_ieq(b, x_lo, y_lo));
> > +   case nir_op_ine:
> > +      return nir_ior(b, nir_ine(b, x_hi, y_hi), nir_ine(b, x_lo, y_lo));
> > +   case nir_op_ult:
> > +      return nir_ior(b, nir_ult(b, x_hi, y_hi),
> > +                        nir_iand(b, nir_ieq(b, x_hi, y_hi),
> > +                                    nir_ult(b, x_lo, y_lo)));
> > +   case nir_op_ilt:
> > +      /* This one is tricky.  Handled below */
> > +      break;
> > +   case nir_op_uge:
> > +      /* Lower as !(x < y) in the hopes of better CSE */
> > +      return nir_inot(b, lower_int64_compare(b, nir_op_ult, x, y));
> > +   case nir_op_ige:
> > +      /* Lower as !(x < y) in the hopes of better CSE */
> > +      return nir_inot(b, lower_int64_compare(b, nir_op_ilt, x, y));
> > +   default:
> > +      unreachable("Invalid comparison");
> > +   }
> > +
> > +   assert(op == nir_op_ilt);
> > +
> > +   nir_ssa_def *neg_x = lower_ineg64(b, x);
> > +   nir_ssa_def *neg_y = lower_ineg64(b, y);
> > +   nir_ssa_def *x_is_neg = nir_ilt(b, x_hi, nir_imm_int(b, 0));
> > +   nir_ssa_def *y_is_neg = nir_ilt(b, y_hi, nir_imm_int(b, 0));
> > +
> > +   nir_ssa_def *imm_true = nir_imm_int(b, NIR_TRUE);
> > +   nir_ssa_def *imm_false = nir_imm_int(b, NIR_FALSE);
> > +   nir_ssa_def *ult = lower_int64_compare(b, nir_op_ult, x, y);
> > +   nir_ssa_def *not_ult_neg =
> > +      nir_inot(b, lower_int64_compare(b, nir_op_ult, neg_x, neg_y));
>
> If you know both are negative, can't you just do an unsigned comparison in
> the opposite direction?  That would save you all the expensive negation.
> You may even be able to re-use the result above though I haven't thought
> through it all the way.

You wrote this code. You tell me :)

In fact, I recently noticed that some piglit comparison tests are
failing, so I suspect there's a bug in here somewhere...


More information about the mesa-dev mailing list