nir: Add a memcpy optimization pass

This pass attempts to optimize three broad categories of memcpy:

 1. Self-copies: These we can discard out-of-hand.

 2. Vector copies: It doesn't matter what the vector size is or if the
    source and destination have different vector types, it's still easy
    enough to emit a load/store pair.

 3. Tightly packed copies:  In the case where a type is tightly packed
    (no padding bits), we can replace the memcpy with a copy_deref
    instruction which the optimizer is far better at handling.

This has proven capable of getting rid of many of the memcpy instances
in some rather gnarly OpenCL C kernels I've been looking at, even after
coming out of LLVM's optimizer.

Reviewed-by: Jesse Natalie <jenatali@microsoft.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6871>
This commit is contained in:
Jason Ekstrand 2020-09-25 13:02:25 -05:00 committed by Marge Bot
parent e363da3bdd
commit f6667cb0ce
4 changed files with 227 additions and 0 deletions

View File

@ -329,6 +329,7 @@ NIR_FILES = \
nir/nir_opt_loop_unroll.c \
nir/nir_opt_large_constants.c \
nir/nir_opt_load_store_vectorize.c \
nir/nir_opt_memcpy.c \
nir/nir_opt_move.c \
nir/nir_opt_peephole_select.c \
nir/nir_opt_rematerialize_compares.c \

View File

@ -210,6 +210,7 @@ files_libnir = files(
'nir_opt_large_constants.c',
'nir_opt_load_store_vectorize.c',
'nir_opt_loop_unroll.c',
'nir_opt_memcpy.c',
'nir_opt_move.c',
'nir_opt_peephole_select.c',
'nir_opt_rematerialize_compares.c',

View File

@ -4126,6 +4126,7 @@ void nir_lower_deref_copy_instr(struct nir_builder *b,
nir_intrinsic_instr *copy);
bool nir_lower_var_copies(nir_shader *shader);
bool nir_opt_memcpy(nir_shader *shader);
bool nir_lower_memcpy(nir_shader *shader);
void nir_fixup_deref_modes(nir_shader *shader);

View File

@ -0,0 +1,224 @@
/*
* Copyright © 2020 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.
*/
#include "nir_builder.h"
static bool
opt_memcpy_deref_cast(nir_intrinsic_instr *cpy, nir_src *deref_src)
{
assert(cpy->intrinsic == nir_intrinsic_memcpy_deref);
nir_deref_instr *cast = nir_src_as_deref(*deref_src);
if (cast == NULL || cast->deref_type != nir_deref_type_cast)
return false;
/* We always have to replace the source with a deref, not a bare uint
* pointer. If it's the first deref in the chain, bail.
*/
nir_deref_instr *parent = nir_src_as_deref(cast->parent);
if (parent == NULL)
return false;
/* If it has useful alignment information, we want to keep that */
if (cast->cast.align_mul > 0)
return false;
/* Casts to uint8 or int8 never do us any good; get rid of them */
if (cast->type == glsl_int8_t_type() ||
cast->type == glsl_uint8_t_type()) {
nir_instr_rewrite_src(&cpy->instr, deref_src,
nir_src_for_ssa(&parent->dest.ssa));
return true;
}
int64_t parent_type_size = glsl_get_explicit_size(parent->type, false);
if (parent_type_size < 0)
return false;
if (!nir_src_is_const(cpy->src[2]))
return false;
/* We don't want to get rid of the cast if the resulting type would be
* smaller than the amount of data we're copying.
*/
if (nir_src_as_uint(cpy->src[2]) < (uint64_t)parent_type_size)
return false;
nir_instr_rewrite_src(&cpy->instr, deref_src,
nir_src_for_ssa(&parent->dest.ssa));
return true;
}
static bool
type_is_tightly_packed(const struct glsl_type *type, unsigned *size_out)
{
unsigned size = 0;
if (glsl_type_is_struct_or_ifc(type)) {
unsigned num_fields = glsl_get_length(type);
for (unsigned i = 0; i < num_fields; i++) {
const struct glsl_struct_field *field =
glsl_get_struct_field_data(type, i);
if (field->offset < 0 || field->offset != size)
return false;
unsigned field_size;
if (!type_is_tightly_packed(field->type, &field_size))
return false;
size = field->offset + field_size;
}
} else if (glsl_type_is_array_or_matrix(type)) {
if (glsl_type_is_unsized_array(type))
return false;
unsigned stride = glsl_get_explicit_stride(type);
if (stride == 0)
return false;
const struct glsl_type *elem_type = glsl_get_array_element(type);
unsigned elem_size;
if (!type_is_tightly_packed(elem_type, &elem_size))
return false;
if (elem_size != stride)
return false;
size = stride * glsl_get_length(type);
} else {
assert(glsl_type_is_vector_or_scalar(type));
if (glsl_get_explicit_stride(type) > 0)
return false;
if (glsl_type_is_boolean(type))
return false;
size = glsl_get_explicit_size(type, false);
}
if (size_out)
*size_out = size;
return true;
}
static bool
try_lower_memcpy(nir_builder *b, nir_intrinsic_instr *cpy)
{
nir_deref_instr *dst = nir_src_as_deref(cpy->src[0]);
nir_deref_instr *src = nir_src_as_deref(cpy->src[1]);
/* A self-copy can always be eliminated */
if (dst == src) {
nir_instr_remove(&cpy->instr);
return true;
}
if (!nir_src_is_const(cpy->src[2]))
return false;
uint64_t size = nir_src_as_uint(cpy->src[2]);
if (size == 0) {
nir_instr_remove(&cpy->instr);
return true;
}
if (glsl_type_is_vector_or_scalar(src->type) &&
glsl_type_is_vector_or_scalar(dst->type) &&
glsl_get_explicit_size(dst->type, false) == size &&
glsl_get_explicit_size(src->type, false) == size) {
b->cursor = nir_instr_remove(&cpy->instr);
nir_ssa_def *data =
nir_load_deref_with_access(b, src, nir_intrinsic_src_access(cpy));
data = nir_bitcast_vector(b, data, glsl_get_bit_size(dst->type));
assert(data->num_components == glsl_get_vector_elements(dst->type));
nir_store_deref_with_access(b, src, data, ~0 /* write mask */,
nir_intrinsic_dst_access(cpy));
return true;
}
unsigned type_size;
if (dst->type == src->type &&
type_is_tightly_packed(dst->type, &type_size) &&
type_size == size) {
b->cursor = nir_instr_remove(&cpy->instr);
nir_copy_deref_with_access(b, dst, src,
nir_intrinsic_dst_access(cpy),
nir_intrinsic_src_access(cpy));
return true;
}
return false;
}
static bool
opt_memcpy_impl(nir_function_impl *impl)
{
bool progress = false;
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 *cpy = nir_instr_as_intrinsic(instr);
if (cpy->intrinsic != nir_intrinsic_memcpy_deref)
continue;
while (opt_memcpy_deref_cast(cpy, &cpy->src[0]))
progress = true;
while (opt_memcpy_deref_cast(cpy, &cpy->src[1]))
progress = true;
if (try_lower_memcpy(&b, cpy)) {
progress = true;
continue;
}
}
}
if (progress) {
nir_metadata_preserve(impl, nir_metadata_block_index |
nir_metadata_dominance);
} else {
nir_metadata_preserve(impl, nir_metadata_all);
}
return progress;
}
bool
nir_opt_memcpy(nir_shader *shader)
{
bool progress = false;
nir_foreach_function(function, shader) {
if (function->impl && opt_memcpy_impl(function->impl))
progress = true;
}
return progress;
}