[Mesa-dev] [PATCH 09/12] nir: Add a array-of-vector variable narrowing pass
Jason Ekstrand
jason at jlekstrand.net
Thu Jul 26 16:00:05 UTC 2018
This pass looks for variables with vector or array-of-vector types and
narrows the type to only the components used.
---
src/compiler/nir/nir.h | 1 +
src/compiler/nir/nir_split_vars.c | 530 +++++++++++++++++++++++++++++-
2 files changed, 523 insertions(+), 8 deletions(-)
diff --git a/src/compiler/nir/nir.h b/src/compiler/nir/nir.h
index c6ed5bb5358..ca437743ff8 100644
--- a/src/compiler/nir/nir.h
+++ b/src/compiler/nir/nir.h
@@ -2611,6 +2611,7 @@ bool nir_split_array_vars(nir_shader *shader, nir_variable_mode modes);
bool nir_split_var_copies(nir_shader *shader);
bool nir_split_per_member_structs(nir_shader *shader);
bool nir_split_struct_vars(nir_shader *shader, nir_variable_mode modes);
+bool nir_narrow_vec_vars(nir_shader *shader, nir_variable_mode modes);
bool nir_lower_returns_impl(nir_function_impl *impl);
bool nir_lower_returns(nir_shader *shader);
diff --git a/src/compiler/nir/nir_split_vars.c b/src/compiler/nir/nir_split_vars.c
index 394ed2be622..8a981dd341d 100644
--- a/src/compiler/nir/nir_split_vars.c
+++ b/src/compiler/nir/nir_split_vars.c
@@ -25,6 +25,9 @@
#include "nir_builder.h"
#include "nir_deref.h"
+/* Needed for _mesa_bitcount() */
+#include "main/macros.h"
+
struct split_var_state {
void *mem_ctx;
@@ -46,15 +49,28 @@ struct field {
};
static const struct glsl_type *
-wrap_type_in_array(const struct glsl_type *type,
- const struct glsl_type *array_type)
+wrap_type_in_matrix_or_array(const struct glsl_type *type,
+ const struct glsl_type *array_type)
{
- if (!glsl_type_is_array(array_type))
+ if (glsl_type_is_array(array_type)) {
+ const struct glsl_type *elem_type =
+ wrap_type_in_matrix_or_array(type, glsl_get_array_element(array_type));
+ return glsl_array_type(elem_type, glsl_get_length(array_type));
+ } else if (glsl_type_is_matrix(array_type)) {
+ if (glsl_type_is_scalar(type)) {
+ /* This can happen if we reduce the number of rows in a matrix to 1.
+ * Just use an array type in this case.
+ */
+ return glsl_array_type(type, glsl_get_matrix_columns(array_type));
+ } else {
+ assert(glsl_type_is_vector(type));
+ return glsl_matrix_type(glsl_get_base_type(type),
+ glsl_get_components(type),
+ glsl_get_matrix_columns(array_type));
+ }
+ } else {
return type;
-
- const struct glsl_type *elem_type =
- wrap_type_in_array(type, glsl_get_array_element(array_type));
- return glsl_array_type(elem_type, glsl_get_length(array_type));
+ }
}
static void
@@ -86,7 +102,7 @@ init_field_for_type(struct field *field, struct field *parent,
} else {
const struct glsl_type *var_type = type;
for (struct field *f = field->parent; f; f = f->parent)
- var_type = wrap_type_in_array(var_type, f->type);
+ var_type = wrap_type_in_matrix_or_array(var_type, f->type);
nir_variable_mode mode = state->base_var->data.mode;
if (mode == nir_var_local) {
@@ -689,6 +705,8 @@ split_array_derefs_impl(nir_function_impl *impl,
/* This level is split, just advance to the next element */
assert(p->deref_type == nir_deref_type_array);
unsigned idx = nir_src_as_const_value(p->arr.index)->u32[0];
+ if (idx >= glsl_get_length(tail_elem->type))
+ idx = 0; /* Overflow; we can give them anything */
tail_elem = &tail_elem->children[idx];
} else {
/* This level isn't split, build a deref */
@@ -794,3 +812,499 @@ nir_split_array_vars(nir_shader *shader, nir_variable_mode modes)
return progress;
}
+
+struct vec_var_usage {
+ nir_component_mask_t comps_read;
+ nir_component_mask_t comps_written;
+
+ nir_component_mask_t comps_kept;
+
+ struct set *vars_copied;
+};
+
+static struct vec_var_usage *
+get_vec_var_usage(nir_variable *var,
+ struct hash_table *var_usage_map,
+ bool add_usage_entry, void *mem_ctx)
+{
+ struct hash_entry *entry = _mesa_hash_table_search(var_usage_map, var);
+ if (entry)
+ return entry->data;
+
+ if (!add_usage_entry)
+ return NULL;
+
+ struct vec_var_usage *usage = rzalloc(mem_ctx, struct vec_var_usage);
+ _mesa_hash_table_insert(var_usage_map, var, usage);
+
+ return usage;
+}
+
+static void
+mark_deref_used(nir_deref_instr *deref,
+ nir_component_mask_t comps_read,
+ nir_component_mask_t comps_written,
+ nir_deref_instr *copy_deref,
+ struct hash_table *var_usage_map,
+ nir_variable_mode modes,
+ void *mem_ctx)
+{
+ if (!(deref->mode & modes))
+ return;
+
+ nir_variable *var = nir_deref_instr_get_variable(deref);
+
+ const struct glsl_type *vec_type = glsl_without_array_or_matrix(var->type);
+ if (!glsl_type_is_vector_or_scalar(vec_type))
+ return;
+
+ unsigned num_components = glsl_get_components(vec_type);
+
+ struct vec_var_usage *usage =
+ get_vec_var_usage(var, var_usage_map, true, mem_ctx);
+
+ usage->comps_read |= comps_read & ((1u << num_components) - 1);
+ usage->comps_written |= comps_written & ((1u << num_components) - 1);
+
+ if (copy_deref) {
+ if (usage->vars_copied == NULL) {
+ usage->vars_copied = _mesa_set_create(mem_ctx, _mesa_hash_pointer,
+ _mesa_key_pointer_equal);
+ }
+ _mesa_set_add(usage->vars_copied,
+ nir_deref_instr_get_variable(copy_deref));
+ }
+}
+
+static bool
+src_is_load_deref(nir_src src, nir_src deref_src)
+{
+ assert(src.is_ssa);
+ assert(deref_src.is_ssa);
+
+ if (src.ssa->parent_instr->type != nir_instr_type_intrinsic)
+ return false;
+
+ nir_intrinsic_instr *load = nir_instr_as_intrinsic(src.ssa->parent_instr);
+ if (load->intrinsic != nir_intrinsic_load_deref)
+ return false;
+
+ assert(load->src[0].is_ssa);
+
+ return load->src[0].ssa == deref_src.ssa;
+}
+
+/* Returns all non-self-referential components of a store instruction. A
+ * component is self-referential if it comes from the same component of a load
+ * instruction on the same deref. If the only data in a particular component
+ * of a variable came directly from that component then it's undefined. The
+ * only way to get defined data into a component of a variable is for it to
+ * get written there by something outside or from a different component.
+ *
+ * This is a fairly common pattern in shaders that come from either GLSL IR or
+ * GLSLang because both glsl_to_nir and GLSLang implement write-masking with
+ * load-vec-store.
+ */
+static nir_component_mask_t
+get_non_self_referential_store_comps(nir_intrinsic_instr *store)
+{
+ nir_component_mask_t comps = nir_intrinsic_write_mask(store);
+
+ assert(store->src[1].is_ssa);
+ nir_instr *src_instr = store->src[1].ssa->parent_instr;
+ if (src_instr->type != nir_instr_type_alu)
+ return comps;
+
+ nir_alu_instr *src_alu = nir_instr_as_alu(src_instr);
+
+ if (src_alu->op == nir_op_imov ||
+ src_alu->op == nir_op_fmov) {
+ /* If it's just a swizzle of a load from the same deref, discount any
+ * channels that don't move in the swizzle.
+ */
+ if (src_is_load_deref(src_alu->src[0].src, store->src[0])) {
+ for (unsigned i = 0; i < NIR_MAX_VEC_COMPONENTS; i++) {
+ if (src_alu->src[0].swizzle[i] == i)
+ comps &= ~(1u << i);
+ }
+ }
+ } else if (src_alu->op == nir_op_vec2 ||
+ src_alu->op == nir_op_vec3 ||
+ src_alu->op == nir_op_vec4) {
+ /* If it's a vec, discount any channels that are just loads from the
+ * same deref put in the same spot.
+ */
+ for (unsigned i = 0; i < nir_op_infos[src_alu->op].num_inputs; i++) {
+ if (src_is_load_deref(src_alu->src[i].src, store->src[0]) &&
+ src_alu->src[i].swizzle[0] == i)
+ comps &= ~(1u << i);
+ }
+ }
+
+ return comps;
+}
+
+static void
+find_used_components_impl(nir_function_impl *impl,
+ struct hash_table *var_usage_map,
+ nir_variable_mode modes,
+ void *mem_ctx)
+{
+ 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);
+ switch (intrin->intrinsic) {
+ case nir_intrinsic_load_deref:
+ mark_deref_used(nir_src_as_deref(intrin->src[0]),
+ nir_ssa_def_components_read(&intrin->dest.ssa), 0,
+ NULL, var_usage_map, modes, mem_ctx);
+ break;
+
+ case nir_intrinsic_store_deref:
+ mark_deref_used(nir_src_as_deref(intrin->src[0]),
+ 0, get_non_self_referential_store_comps(intrin),
+ NULL, var_usage_map, modes, mem_ctx);
+ break;
+
+ case nir_intrinsic_copy_deref: {
+ /* Just mark everything used for copies. */
+ nir_deref_instr *dst = nir_src_as_deref(intrin->src[0]);
+ nir_deref_instr *src = nir_src_as_deref(intrin->src[1]);
+ mark_deref_used(dst, 0, ~0, src, var_usage_map, modes, mem_ctx);
+ mark_deref_used(src, ~0, 0, dst, var_usage_map, modes, mem_ctx);
+ break;
+ }
+
+ default:
+ break;
+ }
+ }
+ }
+}
+
+static bool
+narrow_vec_var_list(struct exec_list *vars,
+ struct hash_table *var_usage_map)
+{
+ /* Initialize the components kept field of each variable. This is the
+ * AND of the components written and components read. If a component is
+ * written but never read, it's dead. If it is read but never written,
+ * then all values read are undefined garbage and we may as well not read
+ * them.
+ */
+ nir_foreach_variable(var, vars) {
+ struct vec_var_usage *usage =
+ get_vec_var_usage(var, var_usage_map, false, NULL);
+ if (usage) {
+ assert(usage->comps_kept == 0);
+ usage->comps_kept = usage->comps_read & usage->comps_written;
+ }
+ }
+
+ /* In order for variable copies to work, we have to have the same data type
+ * on the source and the destination. In order to satisfy this, we run a
+ * little fixed-point algorithm to transitively ensure that we get enough
+ * components for this to hold for all copies.
+ */
+ bool fp_progress;
+ do {
+ fp_progress = false;
+ nir_foreach_variable(var, vars) {
+ struct vec_var_usage *var_usage =
+ get_vec_var_usage(var, var_usage_map, false, NULL);
+ if (!var_usage || !var_usage->vars_copied)
+ continue;
+
+ const unsigned var_num_components =
+ glsl_get_components(glsl_without_array_or_matrix(var->type));
+
+ struct set_entry *copy_entry;
+ set_foreach(var_usage->vars_copied, copy_entry) {
+ struct vec_var_usage *copy_usage =
+ get_vec_var_usage((void *)copy_entry->key,
+ var_usage_map, false, NULL);
+
+ nir_component_mask_t copy_comps;
+ if (!copy_usage) {
+ /* If the copy doesn't have usage information, assume the worst
+ * case that it uses everything.
+ */
+ copy_comps = (1 << var_num_components) - 1;
+ } else {
+ copy_comps = copy_usage->comps_kept;
+ }
+
+ if ((var_usage->comps_kept | copy_comps) != var_usage->comps_kept) {
+ var_usage->comps_kept |= copy_comps;
+ fp_progress = true;
+ }
+ }
+ }
+ } while (fp_progress);
+
+ bool vars_narrowed = false;
+ nir_foreach_variable_safe(var, vars) {
+ struct vec_var_usage *usage =
+ get_vec_var_usage(var, var_usage_map, false, NULL);
+ if (!usage)
+ continue;
+
+ const struct glsl_type *vec_type =
+ glsl_without_array_or_matrix(var->type);
+ assert(usage->comps_kept < (1u << glsl_get_components(vec_type)));
+
+ if (usage->comps_kept == (1u << glsl_get_components(vec_type)) - 1) {
+ /* This variable doesn't need to be narrowed. Remove it from the
+ * hash table so later passes will ignore it.
+ */
+ _mesa_hash_table_remove_key(var_usage_map, var);
+ continue;
+ }
+
+ if (usage->comps_kept == 0) {
+ /* This variable is dead, remove it */
+ exec_node_remove(&var->node);
+ } else {
+ const struct glsl_type *new_vec_type =
+ glsl_vector_type(glsl_get_base_type(vec_type),
+ _mesa_bitcount(usage->comps_kept));
+ var->type = wrap_type_in_matrix_or_array(new_vec_type, var->type);
+ }
+ vars_narrowed = true;
+ }
+
+ return vars_narrowed;
+}
+
+static bool
+deref_is_dead(nir_deref_instr *deref,
+ struct hash_table *var_usage_map,
+ nir_variable_mode modes)
+{
+ if (!(deref->mode & modes))
+ return false;
+
+ struct vec_var_usage *usage =
+ get_vec_var_usage(nir_deref_instr_get_variable(deref),
+ var_usage_map, false, NULL);
+ if (!usage)
+ return false;
+
+ return usage->comps_kept == 0;
+}
+
+static void
+narrow_vec_var_access_impl(nir_function_impl *impl,
+ struct hash_table *var_usage_map,
+ nir_variable_mode modes)
+{
+ nir_builder b;
+ nir_builder_init(&b, impl);
+
+ nir_foreach_block(block, impl) {
+ nir_foreach_instr_safe(instr, block) {
+ switch (instr->type) {
+ case nir_instr_type_deref: {
+ nir_deref_instr *deref = nir_instr_as_deref(instr);
+ if (!(deref->mode & modes))
+ break;
+
+ /* Clean up any dead derefs we find lying around. They may refer
+ * to variables we're planning to split.
+ */
+ if (nir_deref_instr_remove_if_unused(deref))
+ break;
+
+ /* We don't need to check if this is one of the derefs we're
+ * narrowing because this is a no-op if it isn't. The worst that
+ * could happen is that we accidentally fix an invalid deref.
+ */
+ if (deref->deref_type == nir_deref_type_var) {
+ deref->type = deref->var->type;
+ } else if (deref->deref_type == nir_deref_type_array ||
+ deref->deref_type == nir_deref_type_array_wildcard) {
+ nir_deref_instr *parent = nir_deref_instr_parent(deref);
+ assert(glsl_type_is_array(parent->type) ||
+ glsl_type_is_matrix(parent->type));
+ deref->type = glsl_get_array_element(parent->type);
+ }
+ break;
+ }
+
+ case nir_instr_type_intrinsic: {
+ nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
+
+ /* If we have a copy whose source or destination has been deleted
+ * because we determined the variable was dead, then we just
+ * delete the copy instruction. If the source variable was dead
+ * then it was writing undefined garbage anyway and if it's the
+ * destination variable that's dead then the write isn't needed.
+ */
+ if (intrin->intrinsic == nir_intrinsic_copy_deref) {
+ nir_deref_instr *dst = nir_src_as_deref(intrin->src[0]);
+ nir_deref_instr *src = nir_src_as_deref(intrin->src[1]);
+ if (deref_is_dead(dst, var_usage_map, modes) ||
+ deref_is_dead(src, var_usage_map, modes)) {
+ nir_instr_remove(&intrin->instr);
+ nir_deref_instr_remove_if_unused(dst);
+ nir_deref_instr_remove_if_unused(src);
+ }
+ continue;
+ }
+
+ if (intrin->intrinsic != nir_intrinsic_load_deref &&
+ intrin->intrinsic != nir_intrinsic_store_deref)
+ continue;
+
+ nir_deref_instr *deref = nir_src_as_deref(intrin->src[0]);
+ if (!(deref->mode & modes))
+ continue;
+
+ struct vec_var_usage *usage =
+ get_vec_var_usage(nir_deref_instr_get_variable(deref),
+ var_usage_map, false, NULL);
+ if (!usage)
+ continue;
+
+ if (usage->comps_kept == 0) {
+ if (intrin->intrinsic == nir_intrinsic_load_deref) {
+ nir_ssa_def *u =
+ nir_ssa_undef(&b, intrin->dest.ssa.num_components,
+ intrin->dest.ssa.bit_size);
+ nir_ssa_def_rewrite_uses(&intrin->dest.ssa,
+ nir_src_for_ssa(u));
+ }
+ nir_instr_remove(&intrin->instr);
+ nir_deref_instr_remove_if_unused(nir_src_as_deref(intrin->src[0]));
+ continue;
+ }
+
+ if (intrin->intrinsic == nir_intrinsic_load_deref) {
+ unsigned swizzle[NIR_MAX_VEC_COMPONENTS] = { 0, };
+ unsigned c = 0;
+ for (unsigned i = 0; i < intrin->num_components; i++) {
+ if (usage->comps_kept & (1u << i))
+ swizzle[i] = c++;
+ }
+
+ b.cursor = nir_after_instr(&intrin->instr);
+
+ nir_ssa_def *swizzled =
+ nir_swizzle(&b, &intrin->dest.ssa, swizzle,
+ intrin->num_components, false);
+
+ nir_ssa_def_rewrite_uses_after(&intrin->dest.ssa,
+ nir_src_for_ssa(swizzled),
+ swizzled->parent_instr);
+
+ /* The SSA def is now only used by the swizzle. It's safe to
+ * shrink the number of components.
+ */
+ assert(list_is_singular(&intrin->dest.ssa.uses));
+ intrin->num_components = c;
+ intrin->dest.ssa.num_components = c;
+ } else {
+ nir_component_mask_t write_mask =
+ nir_intrinsic_write_mask(intrin);
+
+ unsigned swizzle[NIR_MAX_VEC_COMPONENTS];
+ nir_component_mask_t new_write_mask = 0;
+ unsigned c = 0;
+ for (unsigned i = 0; i < intrin->num_components; i++) {
+ if (usage->comps_kept & (1u << i)) {
+ swizzle[c] = i;
+ if (write_mask & (1u << i))
+ new_write_mask |= 1u << c;
+ c++;
+ }
+ }
+
+ b.cursor = nir_before_instr(&intrin->instr);
+
+ nir_ssa_def *swizzled =
+ nir_swizzle(&b, intrin->src[1].ssa, swizzle, c, false);
+
+ /* Rewrite to use the compacted source */
+ nir_instr_rewrite_src(&intrin->instr, &intrin->src[1],
+ nir_src_for_ssa(swizzled));
+ nir_intrinsic_set_write_mask(intrin, new_write_mask);
+ intrin->num_components = c;
+ }
+ break;
+ }
+
+ default:
+ break;
+ }
+ }
+ }
+}
+
+/** Attempt to narrow (remove components) from (arrays of) vectors
+ *
+ * This pass looks at variables which contain a vector or an array (possibly
+ * multiple dimensions) of vectors and attempts to lower to a smaller vector.
+ */
+bool
+nir_narrow_vec_vars(nir_shader *shader, nir_variable_mode modes)
+{
+ assert((modes & (nir_var_global | nir_var_local)) == modes);
+
+ void *mem_ctx = ralloc_context(NULL);
+
+ struct hash_table *var_usage_map =
+ _mesa_hash_table_create(mem_ctx, _mesa_hash_pointer,
+ _mesa_key_pointer_equal);
+
+ bool has_vars_to_narrow = false;
+ nir_foreach_function(function, shader) {
+ if (!function->impl)
+ continue;
+
+ /* Don't even bother crawling the IR if we don't have any variables.
+ * Given that this pass deletes any unused variables, it's likely that
+ * we will be in this scenario eventually.
+ */
+ if (!exec_list_is_empty(&shader->globals) ||
+ !exec_list_is_empty(&function->impl->locals)) {
+ has_vars_to_narrow = true;
+ find_used_components_impl(function->impl, var_usage_map,
+ modes, mem_ctx);
+ }
+ }
+ if (!has_vars_to_narrow) {
+ ralloc_free(mem_ctx);
+ return false;
+ }
+
+ bool globals_narrowed = false;
+ if (modes & nir_var_global)
+ globals_narrowed = narrow_vec_var_list(&shader->globals, var_usage_map);
+
+ bool progress = false;
+ nir_foreach_function(function, shader) {
+ if (!function->impl)
+ continue;
+
+ bool locals_narrowed = false;
+ if (modes & nir_var_local) {
+ locals_narrowed = narrow_vec_var_list(&function->impl->locals,
+ var_usage_map);
+ }
+
+ if (globals_narrowed || locals_narrowed) {
+ narrow_vec_var_access_impl(function->impl, var_usage_map, modes);
+
+ nir_metadata_preserve(function->impl, nir_metadata_block_index |
+ nir_metadata_dominance);
+ progress = true;
+ }
+ }
+
+ ralloc_free(mem_ctx);
+
+ return progress;
+}
--
2.17.1
More information about the mesa-dev
mailing list