[Mesa-dev] [PATCH 090/133] nir: Add a copy splitting pass
Connor Abbott
cwabbott0 at gmail.com
Wed Dec 17 18:01:38 PST 2014
On Tue, Dec 16, 2014 at 1:11 AM, Jason Ekstrand <jason at jlekstrand.net> wrote:
> ---
> src/glsl/Makefile.sources | 1 +
> src/glsl/nir/nir.h | 2 +
> src/glsl/nir/nir_split_var_copies.c | 225 ++++++++++++++++++++++++++++++++++++
> 3 files changed, 228 insertions(+)
> create mode 100644 src/glsl/nir/nir_split_var_copies.c
>
> diff --git a/src/glsl/Makefile.sources b/src/glsl/Makefile.sources
> index 20496ff..84245bc 100644
> --- a/src/glsl/Makefile.sources
> +++ b/src/glsl/Makefile.sources
> @@ -37,6 +37,7 @@ NIR_FILES = \
> $(GLSL_SRCDIR)/nir/nir_opt_peephole_select.c \
> $(GLSL_SRCDIR)/nir/nir_print.c \
> $(GLSL_SRCDIR)/nir/nir_remove_dead_variables.c \
> + $(GLSL_SRCDIR)/nir/nir_split_var_copies.c \
> $(GLSL_SRCDIR)/nir/nir_to_ssa.c \
> $(GLSL_SRCDIR)/nir/nir_validate.c \
> $(GLSL_SRCDIR)/nir/nir_types.cpp \
> diff --git a/src/glsl/nir/nir.h b/src/glsl/nir/nir.h
> index fc16cb5..86cda07 100644
> --- a/src/glsl/nir/nir.h
> +++ b/src/glsl/nir/nir.h
> @@ -1356,6 +1356,8 @@ void nir_dump_dom_frontier(nir_shader *shader, FILE *fp);
> void nir_dump_cfg_impl(nir_function_impl *impl, FILE *fp);
> void nir_dump_cfg(nir_shader *shader, FILE *fp);
>
> +void nir_split_var_copies(nir_shader *shader);
> +
> void nir_lower_variables_scalar(nir_shader *shader, bool lower_globals,
> bool lower_io, bool add_names,
> bool native_integers);
> diff --git a/src/glsl/nir/nir_split_var_copies.c b/src/glsl/nir/nir_split_var_copies.c
> new file mode 100644
> index 0000000..19724b6
> --- /dev/null
> +++ b/src/glsl/nir/nir_split_var_copies.c
> @@ -0,0 +1,225 @@
> +/*
> + * Copyright © 2014 Intel 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.
> + *
> + * Authors:
> + * Jason Ekstrand (jason at jlekstrand.net)
> + *
> + */
> +
> +#include "nir.h"
> +
> +/*
> + * Implements "copy splitting" which is similar to structure splitting only
> + * it works on copy operations rather than the datatypes themselves. The
> + * GLSL language allows you to copy one variable to another an entire
> + * structure (which may contain arrays or other structures) at a time.
> + * Normally, in a language such as C this would be handled by a "structure
> + * splitting" pass that breaks up the structures. Unfortunately for us,
> + * structures used in inputs or outputs can't be split. Therefore,
> + * regardlesss of what we do, we have to be able to copy to/from
> + * structures.
> + *
> + * The primary purpose of structure splitting is to allow you to better
> + * optimize variable access and lower things to registers where you can.
> + * The primary issue here is that, if you lower the copy to a bunch of
> + * loads and stores, you loose a lot of information about the copy
> + * operation that you would like to keep around. To solve this problem, we
> + * have a "copy splitting" pass that, instead of splitting the structures
> + * or lowering the copy into loads and storres, splits the copy operation
> + * into a bunch of copy operations one for each leaf of the structure tree.
> + * If an intermediate array is encountered, it is referenced with a
> + * wildcard reference to indicate that the entire array is to be copied.
> + *
> + * As things become direct, array copies may be able to be losslessly
> + * lowered to having fewer and fewer wildcards. However, until that
> + * happens we want to keep the information about the arrays intact.
> + *
> + * Prior to the copy splitting pass, there are no wildcard references but
> + * there may be incomplete references where the tail of the deref chain is
> + * an array or a structure and not a specific element. After the copy
> + * splitting pass has completed, every variable deref will be a full-length
> + * dereference pointing to a single leaf in the structure type tree with
> + * possibly a few wildcard array dereferences.
> + */
> +
> +struct split_var_copies_state {
> + void *mem_ctx;
> + void *dead_ctx;
> +};
> +
> +static nir_deref *
> +get_deref_tail(nir_deref *deref)
> +{
> + while (deref->child != NULL)
> + deref = deref->child;
> + return deref;
> +}
I think long-term, we should probably make dereferences exec lists so
we don't have to do things like this. In the meantime, though, I
could've sworn I had to write this exact same function somewhere else
(glsl to nir?)... might be a good idea to move this to nir.c and reuse
it, unless nobody else needs this.
> +
> +static void
> +nir_split_var_copy_instr(nir_intrinsic_instr *old_copy,
> + nir_deref *dest_head, nir_deref *src_head,
> + nir_deref *dest_tail, nir_deref *src_tail,
> + struct split_var_copies_state *state)
It took me a while to see exactly what this is doing, why we need
separate dest_tail and src_tail variables even though they're almost
always the same, and how it creates this temporary deref chain that's
shared between source and destination (a no-no in NIR) but that's ok
because they both get copied at the very end of the recursion when we
actually emit the instructions. I'd like to see a comment explaining
how this function works that explains those things.
> +{
> + assert(src_tail->type == dest_tail->type);
> +
> + switch (glsl_get_base_type(src_tail->type)) {
> + case GLSL_TYPE_ARRAY: {
> + nir_deref_array *deref = nir_deref_array_create(state->dead_ctx);
> + deref->deref.type = glsl_get_array_element(src_tail->type);
> + deref->deref_array_type = nir_deref_array_type_wildcard;
> +
> + src_tail->child = &deref->deref;
> + dest_tail->child = &deref->deref;
> + nir_split_var_copy_instr(old_copy, dest_head, src_head,
> + dest_tail->child, src_tail->child, state);
> + src_tail->child = NULL;
> + dest_tail->child = NULL;
> + break;
> + }
> +
> + case GLSL_TYPE_STRUCT:
> + for (unsigned i = 0; i < glsl_get_length(src_tail->type); i++) {
> + nir_deref_struct *deref = nir_deref_struct_create(state->dead_ctx, i);
> + deref->deref.type = glsl_get_struct_field(src_tail->type, i);
> +
> + src_tail->child = &deref->deref;
> + dest_tail->child = &deref->deref;
> +
> + nir_split_var_copy_instr(old_copy, dest_head, src_head,
> + dest_tail->child, src_tail->child, state);
> + }
> + src_tail->child = NULL;
> + dest_tail->child = NULL;
> + break;
> +
> + case GLSL_TYPE_UINT:
> + case GLSL_TYPE_INT:
> + case GLSL_TYPE_FLOAT:
> + case GLSL_TYPE_BOOL:
> + if (glsl_type_is_matrix(src_tail->type)) {
> + nir_deref_array *deref = nir_deref_array_create(state->dead_ctx);
> + deref->deref.type = glsl_get_column_type(src_tail->type);
> + deref->deref_array_type = nir_deref_array_type_wildcard;
> +
> + src_tail->child = &deref->deref;
> + dest_tail->child = &deref->deref;
> + nir_split_var_copy_instr(old_copy, dest_head, src_head,
> + dest_tail->child, src_tail->child, state);
> + src_tail->child = NULL;
> + dest_tail->child = NULL;
> + } else {
> + nir_intrinsic_instr *new_copy =
> + nir_intrinsic_instr_create(state->mem_ctx, nir_intrinsic_copy_var);
> +
> + if (old_copy->has_predicate) {
> + new_copy->has_predicate = true;
> + new_copy->predicate = nir_src_copy(old_copy->predicate,
> + state->mem_ctx);
> + }
> +
> + nir_deref *src = nir_copy_deref(state->mem_ctx, src_head);
> + nir_deref *dest = nir_copy_deref(state->mem_ctx, dest_head);
> +
> + new_copy->variables[0] = nir_deref_as_var(dest);
> + new_copy->variables[1] = nir_deref_as_var(src);
> +
> + nir_instr_insert_after(&old_copy->instr, &new_copy->instr);
> + }
> + break;
> +
> + case GLSL_TYPE_SAMPLER:
> + case GLSL_TYPE_IMAGE:
> + case GLSL_TYPE_ATOMIC_UINT:
> + case GLSL_TYPE_INTERFACE:
> + default:
> + unreachable("Cannot copy these types");
> + }
> +}
> +
> +static bool
> +nir_split_var_copies_block(nir_block *block, void *void_state)
> +{
> + struct split_var_copies_state *state = void_state;
> +
> + nir_foreach_instr_safe(block, instr) {
> + if (instr->type != nir_instr_type_intrinsic)
> + continue;
> +
> + nir_intrinsic_instr *intrinsic = nir_instr_as_intrinsic(instr);
> + if (intrinsic->intrinsic != nir_intrinsic_copy_var)
> + continue;
> +
> + nir_deref *dest_head = &intrinsic->variables[0]->deref;
> + nir_deref *src_head = &intrinsic->variables[1]->deref;
> + nir_deref *dest_tail = get_deref_tail(dest_head);
> + nir_deref *src_tail = get_deref_tail(src_head);
> +
> + switch (glsl_get_base_type(src_tail->type)) {
> + case GLSL_TYPE_ARRAY:
> + case GLSL_TYPE_STRUCT:
> + nir_split_var_copy_instr(intrinsic, dest_head, src_head,
> + dest_tail, src_tail, state);
> + nir_instr_remove(&intrinsic->instr);
> + ralloc_steal(state->dead_ctx, instr);
> + break;
> + case GLSL_TYPE_FLOAT:
> + case GLSL_TYPE_INT:
> + case GLSL_TYPE_UINT:
> + case GLSL_TYPE_BOOL:
> + if (glsl_type_is_matrix(src_tail->type)) {
> + nir_split_var_copy_instr(intrinsic, dest_head, src_head,
> + dest_tail, src_tail, state);
> + nir_instr_remove(&intrinsic->instr);
> + ralloc_steal(state->dead_ctx, instr);
> + }
> + break;
> + default:
> + unreachable("Invalid type");
> + break;
> + }
> + }
> +
> + return true;
> +}
> +
> +static void
> +nir_split_var_copies_impl(nir_function_impl *impl)
> +{
> + struct split_var_copies_state state;
> +
> + state.mem_ctx = ralloc_parent(impl);
> + state.dead_ctx = ralloc_context(NULL);
> +
> + nir_foreach_block(impl, nir_split_var_copies_block, &state);
> +
> + ralloc_free(state.dead_ctx);
> +}
> +
> +void
> +nir_split_var_copies(nir_shader *shader)
> +{
> + nir_foreach_overload(shader, overload) {
> + if (overload->impl)
> + nir_split_var_copies_impl(overload->impl);
> + }
> +}
> --
> 2.2.0
>
> _______________________________________________
> mesa-dev mailing list
> mesa-dev at lists.freedesktop.org
> http://lists.freedesktop.org/mailman/listinfo/mesa-dev
More information about the mesa-dev
mailing list