AlexIndustrial/mesa
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

nir/opt_varyings: fix mesh shader miss promote varying to flat

Browse Source 
We still allow mesh shader promote constant output to flat, but
mesh shader like geometry shader may store multi vertices'
varying in a single thread. So mesh shader may store different
constant values to different vertices in a single thread, we
should not promote this case to flat.

I'm not using shader_info.mesh.ms_cross_invocation_output_access
because OpenGL does not require IO to have explicit location, so
when nir_shader_gather_info is called in OpenGL GLSL compiler to
compute ms_cross_invocation_output_access, some implicit output
has -1 location which causes ms_cross_invocation_output_access
unset for it.

Cc: mesa-stable
Closes: https://gitlab.freedesktop.org/mesa/mesa/-/issues/13134
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/35081>
This commit is contained in:
Qiang Yu
2025-05-16 09:55:01 +08:00
committed by Marge Bot
parent bf24d56862
commit 6f2a1e19da
1 changed files with 40 additions and 21 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/compiler/nir/nir_opt_varyings.c
+40 -21
View File
@@ -43,7 +43,7 @@
*
* When an output stores an SSA that is convergent and all stores of that
* output appear in unconditional blocks or conditional blocks with
* a convergent entry condition and the shader is not GS, it implies that all
* a convergent entry condition and the shader is not GS or MS, it implies that all
* vertices of that output have the same value, therefore the output can be
* promoted to flat because all interpolation modes lead to the same result
* as flat. Such outputs are opportunistically compacted with both flat and
@@ -693,9 +693,9 @@ struct linkage_info {
BITSET_DECLARE(xfb32_only_mask, NUM_SCALAR_SLOTS);
BITSET_DECLARE(xfb16_only_mask, NUM_SCALAR_SLOTS);
/* Mask of all TCS inputs using cross-invocation access. */
BITSET_DECLARE(tcs_cross_invoc32_mask, NUM_SCALAR_SLOTS);
BITSET_DECLARE(tcs_cross_invoc16_mask, NUM_SCALAR_SLOTS);
/* Mask of all TCS inputs or MS outputs using cross-invocation access. */
BITSET_DECLARE(cross_invoc32_mask, NUM_SCALAR_SLOTS);
BITSET_DECLARE(cross_invoc16_mask, NUM_SCALAR_SLOTS);
/* Mask of all TCS->TES slots that are read by TCS, but not TES. */
BITSET_DECLARE(no_varying32_mask, NUM_SCALAR_SLOTS);
@@ -795,8 +795,8 @@ print_linkage(struct linkage_info *linkage)
!BITSET_TEST(linkage->indirect_mask, i) &&
!BITSET_TEST(linkage->xfb32_only_mask, i) &&
!BITSET_TEST(linkage->xfb16_only_mask, i) &&
!BITSET_TEST(linkage->tcs_cross_invoc32_mask, i) &&
!BITSET_TEST(linkage->tcs_cross_invoc16_mask, i) &&
!BITSET_TEST(linkage->cross_invoc32_mask, i) &&
!BITSET_TEST(linkage->cross_invoc16_mask, i) &&
!BITSET_TEST(linkage->no_varying32_mask, i) &&
!BITSET_TEST(linkage->no_varying16_mask, i) &&
!BITSET_TEST(linkage->interp_fp32_mask, i) &&
@@ -829,8 +829,8 @@ print_linkage(struct linkage_info *linkage)
BITSET_TEST(linkage->indirect_mask, i) ? " indirect" : "",
BITSET_TEST(linkage->xfb32_only_mask, i) ? " xfb32_only" : "",
BITSET_TEST(linkage->xfb16_only_mask, i) ? " xfb16_only" : "",
BITSET_TEST(linkage->tcs_cross_invoc32_mask, i) ? " tcs_cross_invoc32" : "",
BITSET_TEST(linkage->tcs_cross_invoc16_mask, i) ? " tcs_cross_invoc16" : "",
BITSET_TEST(linkage->cross_invoc32_mask, i) ? " cross_invoc32" : "",
BITSET_TEST(linkage->cross_invoc16_mask, i) ? " cross_invoc16" : "",
BITSET_TEST(linkage->no_varying32_mask, i) ? " no_varying32" : "",
BITSET_TEST(linkage->no_varying16_mask, i) ? " no_varying16" : "",
BITSET_TEST(linkage->interp_fp32_mask, i) ? " interp_fp32" : "",
@@ -889,8 +889,8 @@ slot_disable_optimizations_and_compaction(struct linkage_info *linkage,
BITSET_CLEAR(linkage->interp_explicit_strict16_mask, i);
BITSET_CLEAR(linkage->per_primitive32_mask, i);
BITSET_CLEAR(linkage->per_primitive16_mask, i);
BITSET_CLEAR(linkage->tcs_cross_invoc32_mask, i);
BITSET_CLEAR(linkage->tcs_cross_invoc16_mask, i);
BITSET_CLEAR(linkage->cross_invoc32_mask, i);
BITSET_CLEAR(linkage->cross_invoc16_mask, i);
BITSET_CLEAR(linkage->no_varying32_mask, i);
BITSET_CLEAR(linkage->no_varying16_mask, i);
BITSET_CLEAR(linkage->color32_mask, i);
@@ -1469,9 +1469,9 @@ gather_inputs(struct nir_builder *builder, nir_intrinsic_instr *intr, void *cb_d
if (!is_sysval(vertex_index_instr, SYSTEM_VALUE_INVOCATION_ID)) {
if (intr->def.bit_size == 32)
BITSET_SET(linkage->tcs_cross_invoc32_mask, slot);
BITSET_SET(linkage->cross_invoc32_mask, slot);
else if (intr->def.bit_size == 16)
BITSET_SET(linkage->tcs_cross_invoc16_mask, slot);
BITSET_SET(linkage->cross_invoc16_mask, slot);
else
unreachable("invalid load_input type");
}
@@ -1643,6 +1643,21 @@ gather_outputs(struct nir_builder *builder, nir_intrinsic_instr *intr, void *cb_
unreachable("invalid store_output type");
}
}
if (linkage->producer_stage == MESA_SHADER_MESH &&
intr->intrinsic == nir_intrinsic_store_per_vertex_output) {
nir_src *vertex_index_src = nir_get_io_arrayed_index_src(intr);
nir_instr *vertex_index_instr = vertex_index_src->ssa->parent_instr;
if (!is_sysval(vertex_index_instr, SYSTEM_VALUE_INVOCATION_ID)) {
if (value->bit_size == 32)
BITSET_SET(linkage->cross_invoc32_mask, slot);
else if (value->bit_size == 16)
BITSET_SET(linkage->cross_invoc16_mask, slot);
else
unreachable("invalid store_output type");
}
}
} else {
/* Only TCS output loads can get here.
*
@@ -1746,7 +1761,7 @@ tidy_up_convergent_varyings(struct linkage_info *linkage)
* bit and keep the convergent bit, which means that it's interpolated,
* but can be promoted to flat.
*
* Since the geometry shader is the only shader that can store values
* Since the geometry shader and mesh shader can store values
* in multiple vertices before FS, it's required that all stores are
* equal to be considered convergent (output_equal_mask), otherwise
* the promotion to flat would be incorrect.
@@ -1761,7 +1776,9 @@ tidy_up_convergent_varyings(struct linkage_info *linkage)
BITSET_CLEAR(linkage->convergent32_mask, i);
} else if ((!linkage->can_mix_convergent_flat_with_interpolated &&
BITSET_TEST(linkage->flat32_mask, i)) ||
(linkage->producer_stage == MESA_SHADER_GEOMETRY &&
((linkage->producer_stage == MESA_SHADER_GEOMETRY ||
(linkage->producer_stage == MESA_SHADER_MESH &&
BITSET_TEST(linkage->cross_invoc32_mask, i))) &&
!BITSET_TEST(linkage->output_equal_mask, i))) {
/* Keep the original qualifier. */
BITSET_CLEAR(linkage->convergent32_mask, i);
@@ -1785,7 +1802,9 @@ tidy_up_convergent_varyings(struct linkage_info *linkage)
BITSET_CLEAR(linkage->convergent16_mask, i);
} else if ((!linkage->can_mix_convergent_flat_with_interpolated &&
BITSET_TEST(linkage->flat16_mask, i)) ||
(linkage->producer_stage == MESA_SHADER_GEOMETRY &&
((linkage->producer_stage == MESA_SHADER_GEOMETRY ||
(linkage->producer_stage == MESA_SHADER_MESH &&
BITSET_TEST(linkage->cross_invoc16_mask, i))) &&
!BITSET_TEST(linkage->output_equal_mask, i))) {
/* Keep the original qualifier. */
BITSET_CLEAR(linkage->convergent16_mask, i);
@@ -4951,18 +4970,18 @@ compact_varyings(struct linkage_info *linkage,
8;
if (linkage->consumer_stage == MESA_SHADER_TESS_CTRL) {
/* Make tcs_cross_invoc*_mask bits disjoint with flat*_mask bits
* because tcs_cross_invoc*_mask is initially a subset of flat*_mask,
/* Make cross_invoc*_mask bits disjoint with flat*_mask bits
* because cross_invoc*_mask is initially a subset of flat*_mask,
* but we must assign each scalar slot only once.
*/
BITSET_ANDNOT(linkage->flat32_mask, linkage->flat32_mask,
linkage->tcs_cross_invoc32_mask);
linkage->cross_invoc32_mask);
BITSET_ANDNOT(linkage->flat16_mask, linkage->flat16_mask,
linkage->tcs_cross_invoc16_mask);
linkage->cross_invoc16_mask);
/* Put cross-invocation-accessed TCS inputs first. */
vs_tcs_tes_gs_assign_slots_2sets(linkage, linkage->tcs_cross_invoc32_mask,
linkage->tcs_cross_invoc16_mask,
vs_tcs_tes_gs_assign_slots_2sets(linkage, linkage->cross_invoc32_mask,
linkage->cross_invoc16_mask,
&slot_index, NULL, progress);
/* Remaining TCS inputs. */
vs_tcs_tes_gs_assign_slots_2sets(linkage, linkage->flat32_mask,