On Mon, Jun 17, 2019 at 12:42 PM Boris Brezillon <
boris.brezil...@collabora.com> wrote:
> On Mon, 17 Jun 2019 10:53:47 -0500
> Jason Ekstrand <ja...@jlekstrand.net> wrote:
>
> > On Mon, Jun 17, 2019 at 5:49 AM Boris Brezillon <
> > boris.brezil...@collabora.com> wrote:
> >
> > > The V3D driver has an open-coded solution for this, and we need the
> > > same thing for Panfrost, so let's add a generic way to lower TXS(LOD)
> > > into max(TXS(0) >> LOD, 1).
> > >
> > > Signed-off-by: Boris Brezillon <boris.brezil...@collabora.com>
> > > ---
> > > src/compiler/nir/nir.h | 6 ++++
> > > src/compiler/nir/nir_lower_tex.c | 49 ++++++++++++++++++++++++++++++++
> > > 2 files changed, 55 insertions(+)
> > >
> > > diff --git a/src/compiler/nir/nir.h b/src/compiler/nir/nir.h
> > > index 4270df565111..8972b4af7480 100644
> > > --- a/src/compiler/nir/nir.h
> > > +++ b/src/compiler/nir/nir.h
> > > @@ -3426,6 +3426,12 @@ typedef struct nir_lower_tex_options {
> > > */
> > > bool lower_txd_clamp_if_sampler_index_not_lt_16;
> > >
> > > + /**
> > > + * If true, lower nir_texop_txs with a non-0-lod into nir_texop_txs
> > > with
> > > + * 0-lod followed by a nir_ishr.
> > > + */
> > > + bool lower_txs_lod;
> > > +
> > > /**
> > > * If true, apply a .bagr swizzle on tg4 results to handle
> Broadcom's
> > > * mixed-up tg4 locations.
> > > diff --git a/src/compiler/nir/nir_lower_tex.c
> > > b/src/compiler/nir/nir_lower_tex.c
> > > index 53719017a87f..c29581d9b048 100644
> > > --- a/src/compiler/nir/nir_lower_tex.c
> > > +++ b/src/compiler/nir/nir_lower_tex.c
> > > @@ -978,6 +978,50 @@ lower_tg4_offsets(nir_builder *b, nir_tex_instr
> *tex)
> > > return true;
> > > }
> > >
> > > +static bool
> > > +nir_lower_txs_lod(nir_builder *b, nir_tex_instr *tex)
> > > +{
> > > + int lod_idx = nir_tex_instr_src_index(tex, nir_tex_src_lod);
> > > + if (lod_idx < 0 ||
> > > + (nir_src_is_const(tex->src[lod_idx].src) &&
> > > + !nir_src_as_int(tex->src[lod_idx].src)))
> > >
> >
> > Please use == 0 instead of ! here. We're checking an integer, not a
> > boolean.
>
> Sure, I'll change that.
>
> >
> >
> > > + return false;
> > > +
> > > + nir_ssa_def *lod = nir_ssa_for_src(b, tex->src[lod_idx].src, 1);
> > >
> >
> > In theory, this can emit instructions. We should set the cursor before
> > calling it.
>
> Oops, I'll move the cursor initialization before this line.
>
> >
> >
> > > + unsigned dest_size = nir_tex_instr_dest_size(tex);
> > > + nir_ssa_def *shift, *min, *result;
> > > +
> > > + b->cursor = nir_after_instr(&tex->instr);
> > > +
> > > + switch (dest_size) {
> > > + case 3:
> > > + shift = nir_vec3(b, lod, lod, tex->is_array ? nir_imm_int(b, 0)
> :
> > > lod);
> > > + min = nir_imm_ivec3(b, 1, 1, tex->is_array ? 0 : 1);
> > > + break;
> > > + case 2:
> > > + shift = nir_vec2(b, lod, tex->is_array ? nir_imm_int(b, 0) :
> lod);
> > > + min = nir_imm_ivec2(b, 1, tex->is_array ? 0 : 1);
> > > + break;
> > > + case 1:
> > > + shift = lod;
> > > + min = nir_imm_int(b, 1);
> > > + break;
> > > + default:
> > > + unreachable("Invalid nir_tex_instr_dest_size()\n");
> > > + }
> > > +
> > > + /* TXS(LOD) = max(TXS(0) >> LOD, 1) */
> > > + result = nir_imax(b, nir_ishr(b, &tex->dest.ssa, shift), min);
> > >
> >
> > I think it will actually emit less code and be slightly simpler if you do
> >
> > nir_ssa_def *minified =
> > nir_imax(b, nir_ushr(b, &tex->dest.ssa, lod), nir_imm_int(b, 1));
>
> Are you sure for the ushr()? lod is supposed to be an int, so I'd
> expect to have an ishr() here (note that I don't even handle the
> lod < 0 case as I'm not sure what's supposed to be done in that case).
>
I don't think it's going to matter in practice. Neither the lod nor the
size are ever validly going to be negative. GLSL might require that we
clamp the LOD to a valid range before using it in the shift.
> > if (tex->is_array) {
> > nir_ssa_def *comp[4];
> > for (unsigned i = 0; i < dest_size - 1; i++)
> > comp[i] = nir_component(b, minified, i);
> > comp[dest_size - 1] = nir_component(b, &tex->dest.ssa, dest_size - 1);
> > minified = nir_vec(b, comp, dest_size);
> > }
> >
> > That way, it generates one vec() operation instead of two. Note that you
> > don't need to explcitly expand lod out to three components as nir_builder
> > will do that for you.
>
> Oh, good to know.
>
> >
> >
> > > + nir_ssa_def_rewrite_uses_after(&tex->dest.ssa,
> nir_src_for_ssa(result),
> > > + result->parent_instr);
> > > +
> > > + /* Replace the non-0-LOD in the initial TXS operation by a 0-LOD.
> */
> > > + b->cursor = nir_before_instr(&tex->instr);
> > > + nir_instr_rewrite_src(&tex->instr, &tex->src[lod_idx].src,
> > > + nir_src_for_ssa(nir_imm_int(b, 0)));
> > >
> >
> > Personally, I'd put this right after we grab the LOD from the instruction
> > but this is fine too.
>
> Okay, I'll move it there.
>
> Thanks for the review.
>
> Boris
>
_______________________________________________
mesa-dev mailing list
mesa-dev@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/mesa-dev
