fdo-mirrors/mesa
1
0
Fork 0
mirror of https://gitlab.freedesktop.org/mesa/mesa.git synced 2026-05-04 20:38:06 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions 2

nir/lower_io_vars: don't insert output stores for unrelated streams before emits
Some checks are pending
macOS-CI / macOS-CI (dri) (push) Waiting to run
Details
macOS-CI / macOS-CI (xlib) (push) Waiting to run
Details

Browse source 
Before every emit_vertex(stream_id = n), we would insert stores for all
outputs, including outputs that are not meant for that stream.
Those stores would end up having no effect while potentially reducing
performance.

Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/38100>
This commit is contained in:
Marek Olšák 2025-10-13 02:59:52 -04:00 • committed by Marge Bot
parent 0ac9ccb4b4
commit a051d4ee6b
1 changed files with 29 additions and 5 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

34
src/compiler/nir/nir_lower_io_vars_to_temporaries.c
View file

@ -48,7 +48,7 @@ struct lower_io_state {
static void
emit_copies(nir_builder *b, struct exec_list *dest_vars,
struct exec_list *src_vars)
struct exec_list *src_vars, unsigned stream)
{
assert(exec_list_length(dest_vars) == exec_list_length(src_vars));
@ -73,6 +73,29 @@ emit_copies(nir_builder *b, struct exec_list *dest_vars,
dest->data.mode == nir_var_mem_pixel_local_in)
continue;
/* Don't copy those variables before emit_vertex that are not emitted
* by that emit_vertex instruction.
*/
unsigned stream_mask = 0;
if (dest->data.stream & NIR_STREAM_PACKED) {
unsigned slots = glsl_count_dword_slots(src->type, false);
assert(slots && slots <= 4);
for (unsigned i = 0; i < slots; i++) {
unsigned stream = (dest->data.stream >> (i * 2)) & 0x3;
stream_mask |= BITFIELD_BIT(stream);
}
} else {
assert(dest->data.stream < 4);
stream_mask = BITFIELD_BIT(dest->data.stream);
}
if (!(BITFIELD_BIT(stream) & stream_mask))
continue;
/* Can't copy the contents of the temporary back to a read-only
* interface variable. The value of the temporary won't have been
* modified by the shader anyway.
@ -102,13 +125,14 @@ emit_output_copies_impl(struct lower_io_state *state, nir_function_impl *impl)
if (intrin->intrinsic == nir_intrinsic_emit_vertex ||
intrin->intrinsic == nir_intrinsic_emit_vertex_with_counter) {
b.cursor = nir_before_instr(&intrin->instr);
emit_copies(&b, &state->new_outputs, &state->old_outputs);
emit_copies(&b, &state->new_outputs, &state->old_outputs,
nir_intrinsic_stream_id(intrin));
}
}
}
} else if (impl == state->entrypoint) {
b.cursor = nir_before_impl(impl);
emit_copies(&b, &state->old_outputs, &state->new_outputs);
emit_copies(&b, &state->old_outputs, &state->new_outputs, 0);
/* For all other shader types, we need to do the copies right before
* the jumps to the end block.
@ -116,7 +140,7 @@ emit_output_copies_impl(struct lower_io_state *state, nir_function_impl *impl)
set_foreach(&impl->end_block->predecessors, block_entry) {
struct nir_block *block = (void *)block_entry->key;
b.cursor = nir_after_block_before_jump(block);
emit_copies(&b, &state->new_outputs, &state->old_outputs);
emit_copies(&b, &state->new_outputs, &state->old_outputs, 0);
}
}
}
@ -297,7 +321,7 @@ emit_input_copies_impl(struct lower_io_state *state, nir_function_impl *impl)
{
if (impl == state->entrypoint) {
nir_builder b = nir_builder_at(nir_before_impl(impl));
emit_copies(&b, &state->old_inputs, &state->new_inputs);
emit_copies(&b, &state->old_inputs, &state->new_inputs, 0);
if (state->shader->info.stage == MESA_SHADER_FRAGMENT)
fixup_interpolation(state, impl, &b);
}