[Mesa-dev] [PATCH 1/8] nir: add varying array splitting pass
Timothy Arceri
tarceri at itsqueeze.com
Wed Nov 29 11:20:15 UTC 2017
On 29/11/17 21:47, Nicolai Hähnle wrote:
> On 21.11.2017 04:28, Timothy Arceri wrote:
>> V2: fix matrix support, non-array matrices were being skipped in v1
>> ---
>> src/compiler/Makefile.sources | 1 +
>> src/compiler/nir/meson.build | 1 +
>> src/compiler/nir/nir.h | 1 +
>> src/compiler/nir/nir_lower_io_arrays_to_elements.c | 373
>> +++++++++++++++++++++
>> 4 files changed, 376 insertions(+)
>> create mode 100644 src/compiler/nir/nir_lower_io_arrays_to_elements.c
>>
>> diff --git a/src/compiler/Makefile.sources
>> b/src/compiler/Makefile.sources
>> index 2ab8e163a2..c5094b7f19 100644
>> --- a/src/compiler/Makefile.sources
>> +++ b/src/compiler/Makefile.sources
>> @@ -219,20 +219,21 @@ NIR_FILES = \
>> nir/nir_lower_double_ops.c \
>> nir/nir_lower_drawpixels.c \
>> nir/nir_lower_global_vars_to_local.c \
>> nir/nir_lower_gs_intrinsics.c \
>> nir/nir_lower_load_const_to_scalar.c \
>> nir/nir_lower_locals_to_regs.c \
>> nir/nir_lower_idiv.c \
>> nir/nir_lower_indirect_derefs.c \
>> nir/nir_lower_int64.c \
>> nir/nir_lower_io.c \
>> + nir/nir_lower_io_arrays_to_elements.c \
>> nir/nir_lower_io_to_temporaries.c \
>> nir/nir_lower_io_to_scalar.c \
>> nir/nir_lower_io_types.c \
>> nir/nir_lower_passthrough_edgeflags.c \
>> nir/nir_lower_patch_vertices.c \
>> nir/nir_lower_phis_to_scalar.c \
>> nir/nir_lower_regs_to_ssa.c \
>> nir/nir_lower_returns.c \
>> nir/nir_lower_samplers.c \
>> nir/nir_lower_samplers_as_deref.c \
>> diff --git a/src/compiler/nir/meson.build b/src/compiler/nir/meson.build
>> index e5c8326aa0..b61a07773d 100644
>> --- a/src/compiler/nir/meson.build
>> +++ b/src/compiler/nir/meson.build
>> @@ -107,20 +107,21 @@ files_libnir = files(
>> 'nir_lower_double_ops.c',
>> 'nir_lower_drawpixels.c',
>> 'nir_lower_global_vars_to_local.c',
>> 'nir_lower_gs_intrinsics.c',
>> 'nir_lower_load_const_to_scalar.c',
>> 'nir_lower_locals_to_regs.c',
>> 'nir_lower_idiv.c',
>> 'nir_lower_indirect_derefs.c',
>> 'nir_lower_int64.c',
>> 'nir_lower_io.c',
>> + 'nir_lower_io_arrays_to_elements.c',
>> 'nir_lower_io_to_temporaries.c',
>> 'nir_lower_io_to_scalar.c',
>> 'nir_lower_io_types.c',
>> 'nir_lower_passthrough_edgeflags.c',
>> 'nir_lower_patch_vertices.c',
>> 'nir_lower_phis_to_scalar.c',
>> 'nir_lower_regs_to_ssa.c',
>> 'nir_lower_returns.c',
>> 'nir_lower_samplers.c',
>> 'nir_lower_samplers_as_deref.c',
>> diff --git a/src/compiler/nir/nir.h b/src/compiler/nir/nir.h
>> index f46f614711..c62af4afb9 100644
>> --- a/src/compiler/nir/nir.h
>> +++ b/src/compiler/nir/nir.h
>> @@ -2475,20 +2475,21 @@ bool
>> nir_lower_constant_initializers(nir_shader *shader,
>> nir_variable_mode modes);
>> bool nir_move_vec_src_uses_to_dest(nir_shader *shader);
>> bool nir_lower_vec_to_movs(nir_shader *shader);
>> void nir_lower_alpha_test(nir_shader *shader, enum compare_func func,
>> bool alpha_to_one);
>> bool nir_lower_alu_to_scalar(nir_shader *shader);
>> bool nir_lower_load_const_to_scalar(nir_shader *shader);
>> bool nir_lower_read_invocation_to_scalar(nir_shader *shader);
>> bool nir_lower_phis_to_scalar(nir_shader *shader);
>> +void nir_lower_io_arrays_to_elements(nir_shader *producer, nir_shader
>> *consumer);
>> void nir_lower_io_to_scalar(nir_shader *shader, nir_variable_mode
>> mask);
>> void nir_lower_io_to_scalar_early(nir_shader *shader,
>> nir_variable_mode mask);
>> bool nir_lower_samplers(nir_shader *shader,
>> const struct gl_shader_program
>> *shader_program);
>> bool nir_lower_samplers_as_deref(nir_shader *shader,
>> const struct gl_shader_program
>> *shader_program);
>> typedef struct nir_lower_subgroups_options {
>> uint8_t subgroup_size;
>> diff --git a/src/compiler/nir/nir_lower_io_arrays_to_elements.c
>> b/src/compiler/nir/nir_lower_io_arrays_to_elements.c
>> new file mode 100644
>> index 0000000000..c41f300edb
>> --- /dev/null
>> +++ b/src/compiler/nir/nir_lower_io_arrays_to_elements.c
>> @@ -0,0 +1,373 @@
>> +/*
>> + * Copyright © 2017 Timothy Arceri
>> + *
>> + * 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.
>> + */
>> +
>> +#include "nir.h"
>> +#include "nir_builder.h"
>> +
>> +/** @file nir_lower_io_arrays_to_elements.c
>> + *
>> + * Split arrays/matrices with direct indexing into individual
>> elements. This
>> + * will allow optimisation passes to better clean up unused elements.
>> + *
>> + */
>> +
>> +static unsigned
>> +get_io_offset(nir_builder *b, nir_deref_var *deref, nir_variable *var,
>> + unsigned *element_index)
>> +{
>> + nir_deref *tail = &deref->deref;
>> +
>> + /* For per-vertex input arrays (i.e. geometry shader inputs), skip
>> the
>> + * outermost array index. Process the rest normally.
>> + */
>> + if (nir_is_per_vertex_io(var, b->shader->info.stage)) {
>> + tail = tail->child;
>> + }
>> +
>> + unsigned offset = 0;
>> + while (tail->child != NULL) {
>> + tail = tail->child;
>> +
>> + if (tail->deref_type == nir_deref_type_array) {
>> + nir_deref_array *deref_array = nir_deref_as_array(tail);
>> + assert(deref_array->deref_array_type !=
>> nir_deref_array_type_indirect);
>> +
>> + unsigned size = glsl_count_attribute_slots(tail->type, false);
>> + offset += size * deref_array->base_offset;
>> +
>> + unsigned num_elements = glsl_type_is_array(tail->type) ?
>> + glsl_get_aoa_size(tail->type) : 1;
>> +
>> + num_elements *=
>> glsl_type_is_matrix(glsl_without_array(tail->type)) ?
>> + glsl_get_matrix_columns(glsl_without_array(tail->type)) : 1;
>> +
>> + *element_index += num_elements * deref_array->base_offset;
>> + } else if (tail->deref_type == nir_deref_type_struct) {
>> + /* TODO: we could also add struct splitting support to this
>> pass */
>> + break;
>
> Maybe make this an assert to emphasize the point?
Sure.
>
>
>> + }
>> + }
>> +
>> + return offset;
>> +}
>> +
>> +static nir_variable **
>> +get_array_elements(struct hash_table *ht, nir_variable *var,
>> + gl_shader_stage stage)
>> +{
>> + nir_variable **elements;
>> + struct hash_entry *entry = _mesa_hash_table_search(ht, var);
>> + if (!entry) {
>> + const struct glsl_type *type = var->type;
>> + if (nir_is_per_vertex_io(var, stage)) {
>> + assert(glsl_type_is_array(type));
>> + type = glsl_get_array_element(type);
>> + }
>> +
>> + unsigned num_elements = glsl_type_is_array(type) ?
>> + glsl_get_aoa_size(type) : 1;
>> +
>> + num_elements *= glsl_type_is_matrix(glsl_without_array(type)) ?
>> + glsl_get_matrix_columns(glsl_without_array(type)) : 1;
>> +
>> + elements = (nir_variable **) calloc(num_elements,
>> sizeof(nir_variable *));
>> + _mesa_hash_table_insert(ht, var, elements);
>> + } else {
>> + elements = (nir_variable **) entry->data;
>> + }
>> +
>> + return elements;
>> +}
>> +
>> +static void
>> +create_array_deref(nir_intrinsic_instr *arr_intr,
>> + nir_intrinsic_instr *element_intr)
>> +{
>> + assert(arr_intr->variables[0]->deref.child);
>> +
>> + nir_deref *parent = &element_intr->variables[0]->deref;
>> + nir_deref_array *darr =
>> + nir_deref_as_array(arr_intr->variables[0]->deref.child);
>> + nir_deref_array *ndarr = nir_deref_array_create(parent);
>> +
>> + ndarr->deref.type = glsl_get_array_element(parent->type);
>> + ndarr->deref_array_type = darr->deref_array_type;
>> + ndarr->base_offset = darr->base_offset;
>> +
>> + if (ndarr->deref_array_type == nir_deref_array_type_indirect)
>> + nir_src_copy(&ndarr->indirect, &darr->indirect, parent);
>> +
>> + element_intr->variables[0]->deref.child = &ndarr->deref;
>> +}
>> +
>> +static void
>> +lower_array(nir_builder *b, nir_intrinsic_instr *intr, nir_variable
>> *var,
>> + struct hash_table *varyings)
>> +{
>> + b->cursor = nir_before_instr(&intr->instr);
>> +
>> + nir_variable **elements =
>> + get_array_elements(varyings, var, b->shader->info.stage);
>> +
>> + unsigned elements_index = 0;
>> + unsigned io_offset = get_io_offset(b, intr->variables[0], var,
>> + &elements_index);
>> +
>> + nir_variable *element = elements[elements_index];
>> + if (!element) {
>> + element = nir_variable_clone(var, b->shader);
>> + element->data.location = var->data.location + io_offset;
>> +
>> + const struct glsl_type *type =
>> glsl_without_array(element->type);
>> +
>> + /* This pass also splits matrices so we need give them a new
>> type. */
>> + if (glsl_type_is_matrix(type)) {
>> + type = glsl_vector_type(glsl_get_base_type(type),
>> + glsl_get_vector_elements(type));
>> + }
>> +
>> + if (nir_is_per_vertex_io(var, b->shader->info.stage)) {
>> + type = glsl_get_array_instance(type,
>> glsl_get_length(element->type));
>> + }
>> +
>> + element->type = type;
>> + elements[elements_index] = element;
>> +
>> + nir_shader_add_variable(b->shader, element);
>> + }
>> +
>> + nir_intrinsic_instr *element_intr =
>> + nir_intrinsic_instr_create(b->shader, intr->intrinsic);
>> + element_intr->num_components = intr->num_components;
>> + element_intr->variables[0] = nir_deref_var_create(element_intr,
>> element);
>> +
>> + if (var->data.mode == nir_var_shader_in) {
>
> I believe this needs to be keyed off the intrinsic type rather than the
> variable mode, since at least TCS output variables can also be read from.
I think your right. Surprising this didn't break any tests. Will fix.
>
>
>> + nir_ssa_dest_init(&element_intr->instr, &element_intr->dest,
>> + intr->num_components,
>> intr->dest.ssa.bit_size, NULL);
>> +
>> + if (intr->intrinsic == nir_intrinsic_interp_var_at_offset ||
>> + intr->intrinsic == nir_intrinsic_interp_var_at_sample)
>> + nir_src_copy(element_intr->src, intr->src,
>> &element_intr->instr);
>
> &element_int->src[0], &intr->src[0] for clarity, I think.
>
>
>> +
>> + nir_ssa_def_rewrite_uses(&intr->dest.ssa,
>> +
>> nir_src_for_ssa(&element_intr->dest.ssa));
>> + } else {
>> + nir_intrinsic_set_write_mask(element_intr,
>> nir_intrinsic_write_mask(intr));
>> + element_intr->src[0] = intr->src[0];
>
> I'm still not 100% up on my NIR, but doesn't this also need to be a
> nir_src_copy? I think it makes a difference if there's an indirect
> access involved.
We don't split indirects so it shouldn't matter here.
>
>
>> + }
>> +
>> + if (nir_is_per_vertex_io(var, b->shader->info.stage)) {
>> + create_array_deref(intr, element_intr);
>> + }
>> +
>> + nir_builder_instr_insert(b, &element_intr->instr);
>> +
>> + /* Remove the old load intrinsic */
>> + nir_instr_remove(&intr->instr);
>> +}
>> +
>> +static bool
>> +deref_has_indirect(nir_builder *b, nir_variable *var, nir_deref_var
>> *deref)
>> +{
>> + nir_deref *tail = &deref->deref;
>> +
>> + if (nir_is_per_vertex_io(var, b->shader->info.stage)) {
>> + tail = tail->child;
>> + }
>> +
>> + for (tail = tail->child; tail; tail = tail->child) {
>> + if (tail->deref_type != nir_deref_type_array)
>> + continue;
>> +
>> + nir_deref_array *arr = nir_deref_as_array(tail);
>> + if (arr->deref_array_type == nir_deref_array_type_indirect)
>> + return true;
>> + }
>> +
>> + return false;
>> +}
>> +
>> +/* Creates a mask of locations that contains arrays that are indexed via
>> + * indirect indexing.
>> + */
>> +static void
>> +create_indirects_mask(nir_shader *shader, uint64_t *indirects,
>> + uint64_t *patch_indirects)
>
> Shouldn't this take a nir_variable_mode to select either inputs or outputs?
Yes it should will fix.
Thanks for the review.
>
> Cheers,
> Nicolai
>
>
>> +{
>> + nir_foreach_function(function, shader) {
>> + if (function->impl) {
>> + nir_builder b;
>> + nir_builder_init(&b, function->impl);
>> +
>> + nir_foreach_block(block, function->impl) {
>> + nir_foreach_instr_safe(instr, block) {
>> +
>> + if (instr->type != nir_instr_type_intrinsic)
>> + continue;
>> +
>> + nir_intrinsic_instr *intr =
>> nir_instr_as_intrinsic(instr);
>> +
>> + if (intr->intrinsic != nir_intrinsic_load_var &&
>> + intr->intrinsic != nir_intrinsic_store_var &&
>> + intr->intrinsic !=
>> nir_intrinsic_interp_var_at_centroid &&
>> + intr->intrinsic !=
>> nir_intrinsic_interp_var_at_sample &&
>> + intr->intrinsic !=
>> nir_intrinsic_interp_var_at_offset)
>> + continue;
>> +
>> + nir_variable *var = intr->variables[0]->var;
>> +
>> + uint64_t loc_mask = 1 << var->data.location;
>> + if (var->data.patch) {
>> + if (deref_has_indirect(&b, var, intr->variables[0]))
>> + patch_indirects[var->data.location_frac] |=
>> loc_mask;
>> + } else {
>> + if (deref_has_indirect(&b, var, intr->variables[0]))
>> + indirects[var->data.location_frac] |= loc_mask;
>> + }
>> + }
>> + }
>> + }
>> + }
>> +}
>> +
>> +static void
>> +lower_io_arrays_to_elements(nir_shader *shader, nir_variable_mode mask,
>> + uint64_t *indirects, uint64_t
>> *patch_indirects,
>> + struct hash_table *varyings)
>> +{
>> + nir_foreach_function(function, shader) {
>> + if (function->impl) {
>> + nir_builder b;
>> + nir_builder_init(&b, function->impl);
>> +
>> + nir_foreach_block(block, function->impl) {
>> + nir_foreach_instr_safe(instr, block) {
>> + if (instr->type != nir_instr_type_intrinsic)
>> + continue;
>> +
>> + nir_intrinsic_instr *intr =
>> nir_instr_as_intrinsic(instr);
>> +
>> + if (intr->intrinsic != nir_intrinsic_load_var &&
>> + intr->intrinsic != nir_intrinsic_store_var &&
>> + intr->intrinsic !=
>> nir_intrinsic_interp_var_at_centroid &&
>> + intr->intrinsic !=
>> nir_intrinsic_interp_var_at_sample &&
>> + intr->intrinsic !=
>> nir_intrinsic_interp_var_at_offset)
>> + continue;
>> +
>> + nir_variable *var = intr->variables[0]->var;
>> +
>> + /* Skip indirects */
>> + uint64_t loc_mask = 1 << var->data.location;
>> + if (var->data.patch) {
>> + if (patch_indirects[var->data.location_frac] &
>> loc_mask)
>> + continue;
>> + } else {
>> + if (indirects[var->data.location_frac] & loc_mask)
>> + continue;
>> + }
>> +
>> + nir_variable_mode mode = var->data.mode;
>> +
>> + const struct glsl_type *type = var->type;
>> + if (nir_is_per_vertex_io(var, b.shader->info.stage)) {
>> + assert(glsl_type_is_array(type));
>> + type = glsl_get_array_element(type);
>> + }
>> +
>> + /* Skip types we cannot split.
>> + *
>> + * TODO: Add support for struct splitting.
>> + */
>> + if ((!glsl_type_is_array(type) &&
>> !glsl_type_is_matrix(type))||
>> + glsl_type_is_struct(glsl_without_array(type)))
>> + continue;
>> +
>> + if (var->data.location < VARYING_SLOT_VAR0 &&
>> + var->data.location >= 0)
>> + continue;
>> +
>> + /* Don't bother splitting if we can't opt away any unused
>> + * elements.
>> + */
>> + if (var->data.always_active_io)
>> + continue;
>> +
>> + switch (intr->intrinsic) {
>> + case nir_intrinsic_interp_var_at_centroid:
>> + case nir_intrinsic_interp_var_at_sample:
>> + case nir_intrinsic_interp_var_at_offset:
>> + case nir_intrinsic_load_var:
>> + case nir_intrinsic_store_var:
>> + if ((mask & nir_var_shader_in && mode ==
>> nir_var_shader_in) ||
>> + (mask & nir_var_shader_out && mode ==
>> nir_var_shader_out))
>> + lower_array(&b, intr, var, varyings);
>> + break;
>> + default:
>> + break;
>> + }
>> + }
>> + }
>> + }
>> + }
>> +}
>> +
>> +void
>> +nir_lower_io_arrays_to_elements(nir_shader *producer, nir_shader
>> *consumer)
>> +{
>> + struct hash_table *split_inputs =
>> + _mesa_hash_table_create(NULL, _mesa_hash_pointer,
>> + _mesa_key_pointer_equal);
>> + struct hash_table *split_outputs =
>> + _mesa_hash_table_create(NULL, _mesa_hash_pointer,
>> + _mesa_key_pointer_equal);
>> +
>> + uint64_t indirects[4] = {0}, patch_indirects[4] = {0};
>> + create_indirects_mask(producer, indirects, patch_indirects);
>> + create_indirects_mask(consumer, indirects, patch_indirects);
>> +
>> + lower_io_arrays_to_elements(producer, nir_var_shader_out, indirects,
>> + patch_indirects, split_outputs);
>> +
>> + lower_io_arrays_to_elements(consumer, nir_var_shader_in, indirects,
>> + patch_indirects, split_inputs);
>> +
>> + /* Remove old input from the shaders inputs list */
>> + struct hash_entry *entry;
>> + hash_table_foreach(split_inputs, entry) {
>> + nir_variable *var = (nir_variable *) entry->key;
>> + exec_node_remove(&var->node);
>> +
>> + free(entry->data);
>> + }
>> +
>> + /* Remove old output from the shaders outputs list */
>> + hash_table_foreach(split_outputs, entry) {
>> + nir_variable *var = (nir_variable *) entry->key;
>> + exec_node_remove(&var->node);
>> +
>> + free(entry->data);
>> + }
>> +
>> + _mesa_hash_table_destroy(split_inputs, NULL);
>> + _mesa_hash_table_destroy(split_outputs, NULL);
>> +}
>>
>
>
More information about the mesa-dev
mailing list