Mesa (main): nir: add nir based version of the lower_const_arrays_to_uniforms pass
GitLab Mirror
gitlab-mirror at kemper.freedesktop.org
Sat Jun 4 03:57:00 UTC 2022
Module: Mesa
Branch: main
Commit: 44d6068c5b7a0e3659c3304dcbf32bd130cc1648
URL: http://cgit.freedesktop.org/mesa/mesa/commit/?id=44d6068c5b7a0e3659c3304dcbf32bd130cc1648
Author: Timothy Arceri <tarceri at itsqueeze.com>
Date: Sun Oct 10 20:33:15 2021 +1100
nir: add nir based version of the lower_const_arrays_to_uniforms pass
Doing this in NIR should give better results, but also allows us to
stop calling more GLSL IR optimisations passes.
v2: Skip 8bit and 16bit type that would require further processing
I believe this is an existing bug in the GLSL IR pass also.
v3: rebuild constant initialisers as we want to call this pass
after nir has already lowered them and performed optimisations.
Reviewed-by: Ian Romanick <ian.d.romanick at intel.com> (v1)
Acked-by: Emma Anholt <emma at anholt.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/16770>
---
src/compiler/nir/meson.build | 1 +
src/compiler/nir/nir.h | 2 +
.../nir/nir_lower_const_arrays_to_uniforms.c | 419 +++++++++++++++++++++
3 files changed, 422 insertions(+)
diff --git a/src/compiler/nir/meson.build b/src/compiler/nir/meson.build
index 09bc0b54ceb..905ea86d751 100644
--- a/src/compiler/nir/meson.build
+++ b/src/compiler/nir/meson.build
@@ -143,6 +143,7 @@ files_libnir = files(
'nir_lower_clip_cull_distance_arrays.c',
'nir_lower_clip_disable.c',
'nir_lower_clip_halfz.c',
+ 'nir_lower_const_arrays_to_uniforms.c',
'nir_lower_convert_alu_types.c',
'nir_lower_variable_initializers.c',
'nir_lower_discard_if.c',
diff --git a/src/compiler/nir/nir.h b/src/compiler/nir/nir.h
index 7f7ed57b366..8fabefcb62e 100644
--- a/src/compiler/nir/nir.h
+++ b/src/compiler/nir/nir.h
@@ -4812,6 +4812,8 @@ bool nir_lower_bool_to_bitsize(nir_shader *shader);
bool nir_lower_bool_to_float(nir_shader *shader);
bool nir_lower_bool_to_int32(nir_shader *shader);
bool nir_opt_simplify_convert_alu_types(nir_shader *shader);
+bool nir_lower_const_arrays_to_uniforms(nir_shader *shader,
+ unsigned max_uniform_components);
bool nir_lower_convert_alu_types(nir_shader *shader,
bool (*should_lower)(nir_intrinsic_instr *));
bool nir_lower_constant_convert_alu_types(nir_shader *shader);
diff --git a/src/compiler/nir/nir_lower_const_arrays_to_uniforms.c b/src/compiler/nir/nir_lower_const_arrays_to_uniforms.c
new file mode 100644
index 00000000000..d7efdf6a4a5
--- /dev/null
+++ b/src/compiler/nir/nir_lower_const_arrays_to_uniforms.c
@@ -0,0 +1,419 @@
+/*
+ * Copyright © 2021 Valve Corporation
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
+ * DEALINGS IN THE SOFTWARE.
+ */
+
+/**
+ * Lower constant arrays to uniform arrays.
+ *
+ * Some driver backends (such as i965 and nouveau) don't handle constant arrays
+ * gracefully, instead treating them as ordinary writable temporary arrays.
+ * Since arrays can be large, this often means spilling them to scratch memory,
+ * which usually involves a large number of instructions.
+ *
+ * This must be called prior to gl_nir_set_uniform_initializers(); we need the
+ * linker to process our new uniform's constant initializer.
+ *
+ * This should be called after optimizations, since those can result in
+ * splitting and removing arrays that are indexed by constant expressions.
+ */
+#include "nir.h"
+#include "nir_builder.h"
+#include "nir_deref.h"
+
+struct var_info {
+ nir_variable *var;
+
+ bool is_constant;
+ bool found_read;
+
+ /* Block that has all the variable stores. All the blocks with reads
+ * should be dominated by this block.
+ */
+ nir_block *block;
+};
+
+static void
+set_const_initialiser(nir_deref_instr **p, nir_constant *top_level_init,
+ nir_src *const_src, unsigned writemask)
+{
+ assert(*p);
+
+ nir_constant *ret = top_level_init;
+ for ( ; *p; p++) {
+ if ((*p)->deref_type == nir_deref_type_array) {
+ assert(nir_src_is_const((*p)->arr.index));
+
+ uint64_t idx = nir_src_as_uint((*p)->arr.index);
+
+ /* Just return if this is an out of bounds write */
+ if (idx >= ret->num_elements)
+ return;
+
+ ret = ret->elements[idx];
+ } else if ((*p)->deref_type == nir_deref_type_struct) {
+ ret = ret->elements[(*p)->strct.index];
+ } else {
+ unreachable("Unsupported deref type");
+ }
+ }
+
+ /* Now that we have selected the corrent nir_constant we copy the constant
+ * values to it.
+ */
+ assert(const_src->is_ssa);
+ nir_instr *src_instr = const_src->ssa->parent_instr;
+ assert(src_instr->type == nir_instr_type_load_const);
+ nir_load_const_instr* load = nir_instr_as_load_const(src_instr);
+
+ for (unsigned i = 0; i < load->def.num_components; i++) {
+ if (!(writemask & (1 << i)))
+ continue;
+
+ memcpy(ret->values + i, load->value + i, sizeof(*load->value));
+ }
+
+ return;
+}
+
+static nir_constant *
+rebuild_const_array_initialiser(const struct glsl_type *type, void *mem_ctx)
+{
+ nir_constant *ret = rzalloc(mem_ctx, nir_constant);
+
+ if (glsl_type_is_matrix(type) && glsl_get_matrix_columns(type) > 1) {
+ ret->num_elements = glsl_get_matrix_columns(type);
+ ret->elements = ralloc_array(mem_ctx, nir_constant *, ret->num_elements);
+
+ for (unsigned i = 0; i < ret->num_elements; i++) {
+ ret->elements[i] = rzalloc(mem_ctx, nir_constant);
+ }
+
+ return ret;
+ }
+
+ if (glsl_type_is_array(type) || glsl_type_is_struct(type)) {
+ ret->num_elements = glsl_get_length(type);
+ ret->elements = ralloc_array(mem_ctx, nir_constant *, ret->num_elements);
+
+ for (unsigned i = 0; i < ret->num_elements; i++) {
+ if (glsl_type_is_array(type)) {
+ ret->elements[i] =
+ rebuild_const_array_initialiser(glsl_get_array_element(type), mem_ctx);
+ } else {
+ ret->elements[i] =
+ rebuild_const_array_initialiser(glsl_get_struct_field(type, i), mem_ctx);
+ }
+ }
+ }
+
+ return ret;
+}
+
+static bool
+lower_const_array_to_uniform(nir_shader *shader, struct var_info *info,
+ struct hash_table *const_array_vars,
+ unsigned *free_uni_components,
+ unsigned *const_count, bool *progress)
+{
+ nir_variable *var = info->var;
+
+ if (!info->is_constant)
+ return true;
+
+ if (!glsl_type_is_array(var->type))
+ return true;
+
+ /* TODO: Add support for 8bit and 16bit types */
+ if (!glsl_type_is_32bit(glsl_without_array(var->type)) &&
+ !glsl_type_is_64bit(glsl_without_array(var->type)))
+ return true;
+
+ /* How many uniform component slots are required? */
+ unsigned component_slots = glsl_get_component_slots(var->type);
+
+ /* We would utilize more than is available, bail out. */
+ if (component_slots > *free_uni_components)
+ return false;
+
+ *free_uni_components -= component_slots;
+
+ /* In the very unlikely event of 4294967295 constant arrays in a single
+ * shader, don't promote this to a uniform.
+ */
+ unsigned limit = ~0;
+ if (*const_count == limit)
+ return false;
+
+ nir_variable *uni = rzalloc(shader, nir_variable);
+
+ /* Rebuild constant initialiser */
+ nir_constant *const_init = rebuild_const_array_initialiser(var->type, uni);
+
+ /* Set constant initialiser */
+ nir_function_impl *impl = nir_shader_get_entrypoint(shader);
+ nir_foreach_block(block, impl) {
+ nir_foreach_instr(instr, block) {
+ if (instr->type != nir_instr_type_intrinsic)
+ continue;
+
+ nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
+ assert(intrin->intrinsic != nir_intrinsic_copy_deref);
+ if (intrin->intrinsic != nir_intrinsic_store_deref)
+ continue;
+
+ nir_deref_instr *deref = nir_src_as_deref(intrin->src[0]);
+ nir_variable *deref_var = nir_deref_instr_get_variable(deref);
+ if (var != deref_var)
+ continue;
+
+ nir_deref_path path;
+ nir_deref_path_init(&path, deref, NULL);
+ assert(path.path[0]->deref_type == nir_deref_type_var);
+
+ nir_deref_instr **p = &path.path[1];
+ set_const_initialiser(p, const_init, &intrin->src[1],
+ nir_intrinsic_write_mask(intrin));
+
+ nir_deref_path_finish(&path);
+ }
+ }
+
+ uni->constant_initializer = const_init;
+ uni->data.how_declared = nir_var_hidden;
+ uni->data.read_only = true;
+ uni->data.mode = nir_var_uniform;
+ uni->type = info->var->type;
+ uni->name = ralloc_asprintf(uni, "constarray_%x_%u",
+ *const_count, shader->info.stage);
+
+ nir_shader_add_variable(shader, uni);
+
+ *const_count = *const_count + 1;
+
+ _mesa_hash_table_insert(const_array_vars, info->var, uni);
+
+ *progress = true;
+
+ return true;
+}
+
+static unsigned
+count_uniforms(nir_shader *shader)
+{
+ unsigned total = 0;
+
+ nir_foreach_variable_with_modes(var, shader, nir_var_uniform) {
+ total += glsl_get_component_slots(var->type);
+ }
+
+ return total;
+}
+
+bool
+nir_lower_const_arrays_to_uniforms(nir_shader *shader,
+ unsigned max_uniform_components)
+{
+ /* This only works with a single entrypoint */
+ nir_function_impl *impl = nir_shader_get_entrypoint(shader);
+
+ unsigned num_locals = nir_function_impl_index_vars(impl);
+ if (num_locals == 0) {
+ nir_shader_preserve_all_metadata(shader);
+ return false;
+ }
+
+ bool progress = false;
+ unsigned uniform_components = count_uniforms(shader);
+ unsigned free_uni_components = max_uniform_components - uniform_components;
+ unsigned const_count = 0;
+
+ struct var_info *var_infos = ralloc_array(NULL, struct var_info, num_locals);
+ nir_foreach_function_temp_variable(var, impl) {
+ var_infos[var->index] = (struct var_info) {
+ .var = var,
+ .is_constant = true,
+ .found_read = false,
+ };
+ }
+
+ nir_metadata_require(impl, nir_metadata_dominance);
+
+ struct hash_table *const_array_vars =
+ _mesa_hash_table_create(NULL, _mesa_hash_pointer, _mesa_key_pointer_equal);
+
+ /* First, walk through the shader and figure out what variables we can
+ * lower to a uniform.
+ */
+ nir_foreach_block(block, impl) {
+ nir_foreach_instr(instr, block) {
+ if (instr->type == nir_instr_type_deref) {
+ /* If we ever see a complex use of a deref_var, we have to assume
+ * that variable is non-constant because we can't guarantee we
+ * will find all of the writers of that variable.
+ */
+ nir_deref_instr *deref = nir_instr_as_deref(instr);
+ if (deref->deref_type == nir_deref_type_var &&
+ deref->var->data.mode == nir_var_function_temp &&
+ nir_deref_instr_has_complex_use(deref))
+ var_infos[deref->var->index].is_constant = false;
+ continue;
+ }
+
+ if (instr->type != nir_instr_type_intrinsic)
+ continue;
+
+ nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
+
+ bool src_is_const = false;
+ nir_deref_instr *src_deref = NULL, *dst_deref = NULL;
+ switch (intrin->intrinsic) {
+ case nir_intrinsic_store_deref:
+ dst_deref = nir_src_as_deref(intrin->src[0]);
+ src_is_const = nir_src_is_const(intrin->src[1]);
+ break;
+
+ case nir_intrinsic_load_deref:
+ src_deref = nir_src_as_deref(intrin->src[0]);
+ break;
+
+ case nir_intrinsic_copy_deref:
+ assert(!"Lowering of copy_deref with const arrays to uniform is prohibited");
+ break;
+
+ default:
+ continue;
+ }
+
+ if (dst_deref && nir_deref_mode_must_be(dst_deref, nir_var_function_temp)) {
+ nir_variable *var = nir_deref_instr_get_variable(dst_deref);
+ if (var == NULL)
+ continue;
+
+ assert(var->data.mode == nir_var_function_temp);
+
+ struct var_info *info = &var_infos[var->index];
+ if (!info->is_constant)
+ continue;
+
+ if (!info->block)
+ info->block = block;
+
+ /* We only consider variables constant if they only have constant
+ * stores, all the stores come before any reads, and all stores
+ * come from the same block. We also can't handle indirect stores.
+ */
+ if (!src_is_const || info->found_read || block != info->block ||
+ nir_deref_instr_has_indirect(dst_deref)) {
+ info->is_constant = false;
+ }
+ }
+
+ if (src_deref && nir_deref_mode_must_be(src_deref, nir_var_function_temp)) {
+ nir_variable *var = nir_deref_instr_get_variable(src_deref);
+ if (var == NULL)
+ continue;
+
+ assert(var->data.mode == nir_var_function_temp);
+
+ /* We only consider variables constant if all the reads are
+ * dominated by the block that writes to it.
+ */
+ struct var_info *info = &var_infos[var->index];
+ if (!info->is_constant)
+ continue;
+
+ if (!info->block || !nir_block_dominates(info->block, block))
+ info->is_constant = false;
+
+ info->found_read = true;
+ }
+ }
+ }
+
+ /* Now lower the constants to uniforms */
+ for (int i = 0; i < num_locals; i++) {
+ struct var_info *info = &var_infos[i];
+ if (!lower_const_array_to_uniform(shader, info, const_array_vars,
+ &free_uni_components, &const_count,
+ &progress))
+ break;
+ }
+
+ /* Finally rewrite its uses */
+ nir_builder b;
+ nir_builder_init(&b, impl);
+ nir_foreach_block(block, impl) {
+ nir_foreach_instr_safe(instr, block) {
+
+ if (instr->type != nir_instr_type_intrinsic)
+ continue;
+
+ nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
+ if (intrin->intrinsic != nir_intrinsic_load_deref)
+ continue;
+
+ nir_deref_instr *deref = nir_src_as_deref(intrin->src[0]);
+ nir_variable *var = nir_deref_instr_get_variable(deref);
+
+ struct hash_entry *entry =
+ _mesa_hash_table_search(const_array_vars, var);
+ if (!entry)
+ continue;
+
+ b.cursor = nir_before_instr(instr);
+
+ nir_variable *uni = (nir_variable *) entry->data;
+ nir_deref_instr *new_deref_instr = nir_build_deref_var(&b, uni);
+
+ nir_deref_path path;
+ nir_deref_path_init(&path, deref, NULL);
+ assert(path.path[0]->deref_type == nir_deref_type_var);
+
+ nir_deref_instr **p = &path.path[1];
+ for (; *p; p++) {
+ if ((*p)->deref_type == nir_deref_type_array) {
+ new_deref_instr = nir_build_deref_array(&b, new_deref_instr,
+ (*p)->arr.index.ssa);
+ } else if ((*p)->deref_type == nir_deref_type_struct) {
+ new_deref_instr = nir_build_deref_struct(&b, new_deref_instr,
+ (*p)->strct.index);
+ } else {
+ unreachable("Unsupported deref type");
+ }
+ }
+ nir_deref_path_finish(&path);
+
+ nir_ssa_def *new_def = nir_load_deref(&b, new_deref_instr);
+
+ nir_ssa_def_rewrite_uses(&intrin->dest.ssa, new_def);
+ nir_instr_remove(&intrin->instr);
+ }
+ }
+
+ nir_metadata_preserve(impl, nir_metadata_block_index |
+ nir_metadata_dominance);
+
+ ralloc_free(var_infos);
+ _mesa_hash_table_destroy(const_array_vars, NULL);
+
+ return progress;
+}
More information about the mesa-commit
mailing list