fdo-mirrors/mesa
1
0
Fork 0
mirror of https://gitlab.freedesktop.org/mesa/mesa.git synced 2026-01-14 11:50:17 +01:00
Code Issues Projects Releases Packages Wiki Activity Actions 1

anv/brw: stop turning load_push_constants into load_uniform

Browse source 
Those intrinsics have different semantics in particular with regards
to divergence. Turning one into the other without invalidating the
divergence information breaks NIR validation. But also the conversion
means we get artificially less convergent values in the shaders.

So just handle load_push_constants in the backend and stop changing
things in Anv.

Fixes a bunch of tests in
   dEQP-VK.descriptor_indexing.*
   dEQP-VK.pipeline.*.push_constant.graphics_pipeline.dynamic_index_*

Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Cc: mesa-stable
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/34546>
(cherry picked from commit df15968813)
This commit is contained in:
Lionel Landwerlin 2025-04-16 08:54:14 +03:00 • committed by Eric Engestrom
parent 747b170d46
commit a3293eb26c
6 changed files with 29 additions and 21 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

2
.pick_status.json
View file

@ -5274,7 +5274,7 @@
"description": "anv/brw: stop turning load_push_constants into load_uniform",
"nominated": true,
"nomination_type": 1,
"resolution": 0,
"resolution": 1,
"main_sha": null,
"because_sha": null,
"notes": null

4
src/intel/compiler/brw_from_nir.cpp
View file

@ -1987,6 +1987,7 @@ get_nir_def(nir_to_brw_state &ntb, const nir_def &def, bool all_sources_uniform)
break;
case nir_intrinsic_load_uniform:
case nir_intrinsic_load_push_constant:
is_scalar = get_nir_src(ntb, instr->src[0], 0).is_scalar;
break;
@ -6154,7 +6155,8 @@ brw_from_nir_emit_intrinsic(nir_to_brw_state &ntb,
break;
}
case nir_intrinsic_load_uniform: {
case nir_intrinsic_load_uniform:
case nir_intrinsic_load_push_constant: {
/* Offsets are in bytes but they should always aligned to
* the type size
*/

21
src/intel/compiler/brw_nir.c
View file

@ -2326,20 +2326,21 @@ brw_nir_create_passthrough_tcs(void *mem_ctx, const struct brw_compiler *compile
}
nir_def *
brw_nir_load_global_const(nir_builder *b, nir_intrinsic_instr *load_uniform,
brw_nir_load_global_const(nir_builder *b, nir_intrinsic_instr *load,
nir_def *base_addr, unsigned off)
{
assert(load_uniform->intrinsic == nir_intrinsic_load_uniform);
assert(load->intrinsic == nir_intrinsic_load_push_constant ||
load->intrinsic == nir_intrinsic_load_uniform);
unsigned bit_size = load_uniform->def.bit_size;
unsigned bit_size = load->def.bit_size;
assert(bit_size >= 8 && bit_size % 8 == 0);
unsigned byte_size = bit_size / 8;
nir_def *sysval;
if (nir_src_is_const(load_uniform->src[0])) {
if (nir_src_is_const(load->src[0])) {
uint64_t offset = off +
nir_intrinsic_base(load_uniform) +
nir_src_as_uint(load_uniform->src[0]);
nir_intrinsic_base(load) +
nir_src_as_uint(load->src[0]);
/* Things should be component-aligned. */
assert(offset % byte_size == 0);
@ -2359,14 +2360,14 @@ brw_nir_load_global_const(nir_builder *b, nir_intrinsic_instr *load_uniform,
}
sysval = nir_extract_bits(b, data, 2, suboffset * 8,
load_uniform->num_components, bit_size);
load->num_components, bit_size);
} else {
nir_def *offset32 =
nir_iadd_imm(b, load_uniform->src[0].ssa,
off + nir_intrinsic_base(load_uniform));
nir_iadd_imm(b, load->src[0].ssa,
off + nir_intrinsic_base(load));
nir_def *addr = nir_iadd(b, base_addr, nir_u2u64(b, offset32));
sysval = nir_load_global_constant(b, addr, byte_size,
load_uniform->num_components, bit_size);
load->num_components, bit_size);
}
return sysval;

4
src/intel/compiler/brw_nir_lower_rt_intrinsics.c
View file

@ -153,7 +153,8 @@ lower_rt_intrinsics_impl(nir_function_impl *impl,
nir_instr_remove(instr);
break;
case nir_intrinsic_load_uniform: {
case nir_intrinsic_load_uniform:
case nir_intrinsic_load_push_constant:
/* We don't want to lower this in the launch trampoline.
*
* Also if the driver chooses to use an inline push address, we
@ -169,7 +170,6 @@ lower_rt_intrinsics_impl(nir_function_impl *impl,
BRW_RT_PUSH_CONST_OFFSET);
break;
}
case nir_intrinsic_load_ray_launch_id:
sysval = nir_channels(b, hotzone, 0xe);

16
src/intel/vulkan/anv_nir.h
View file

@ -42,13 +42,17 @@ extern "C" {
nir_imm_int(b, 0), \
.base = anv_drv_const_offset(field), \
.range = components * anv_drv_const_size(field))
/* Use load_uniform for indexed values since load_push_constant requires that
* the offset source is dynamically uniform in the subgroup which we cannot
* guarantee.
*/
#define anv_load_driver_uniform_indexed(b, components, field, idx) \
nir_load_push_constant(b, components, \
anv_drv_const_size(field[0]) * 8, \
nir_imul_imm(b, idx, \
anv_drv_const_size(field[0])), \
.base = anv_drv_const_offset(field), \
.range = anv_drv_const_size(field))
nir_load_uniform(b, components, \
anv_drv_const_size(field[0]) * 8, \
nir_imul_imm(b, idx, \
anv_drv_const_size(field[0])), \
.base = anv_drv_const_offset(field), \
.range = anv_drv_const_size(field))

3
src/intel/vulkan/anv_nir_compute_push_layout.c
View file

@ -57,6 +57,7 @@ anv_nir_compute_push_layout(nir_shader *nir,
has_const_ubo = true;
break;
case nir_intrinsic_load_uniform:
case nir_intrinsic_load_push_constant: {
unsigned base = nir_intrinsic_base(intrin);
unsigned range = nir_intrinsic_range(intrin);
@ -175,6 +176,7 @@ anv_nir_compute_push_layout(nir_shader *nir,
nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
switch (intrin->intrinsic) {
case nir_intrinsic_load_uniform:
case nir_intrinsic_load_push_constant: {
/* With bindless shaders we load uniforms with SEND
* messages. All the push constants are located after the
@ -184,7 +186,6 @@ anv_nir_compute_push_layout(nir_shader *nir,
*/
unsigned base_offset =
brw_shader_stage_is_bindless(nir->info.stage) ? 0 : push_start;
intrin->intrinsic = nir_intrinsic_load_uniform;
nir_intrinsic_set_base(intrin,
nir_intrinsic_base(intrin) -
base_offset);