Mesa (master): nir: Add a memcpy optimization pass
GitLab Mirror
gitlab-mirror at kemper.freedesktop.org
Fri Oct 2 07:46:23 UTC 2020
Module: Mesa
Branch: master
Commit: f6667cb0ce828a8e4bc57e182ab1526fdd7a5e35
URL: http://cgit.freedesktop.org/mesa/mesa/commit/?id=f6667cb0ce828a8e4bc57e182ab1526fdd7a5e35
Author: Jason Ekstrand <jason at jlekstrand.net>
Date: Fri Sep 25 13:02:25 2020 -0500
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 at microsoft.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6871>
---
src/compiler/Makefile.sources | 1 +
src/compiler/nir/meson.build | 1 +
src/compiler/nir/nir.h | 1 +
src/compiler/nir/nir_opt_memcpy.c | 224 ++++++++++++++++++++++++++++++++++++++
4 files changed, 227 insertions(+)
diff --git a/src/compiler/Makefile.sources b/src/compiler/Makefile.sources
index f5975896c66..aaea6a08d6b 100644
--- a/src/compiler/Makefile.sources
+++ b/src/compiler/Makefile.sources
@@ -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 \
diff --git a/src/compiler/nir/meson.build b/src/compiler/nir/meson.build
index f6030b093f9..eff000ee066 100644
--- a/src/compiler/nir/meson.build
+++ b/src/compiler/nir/meson.build
@@ -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',
diff --git a/src/compiler/nir/nir.h b/src/compiler/nir/nir.h
index 0402c1c87d0..99421534880 100644
--- a/src/compiler/nir/nir.h
+++ b/src/compiler/nir/nir.h
@@ -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);
diff --git a/src/compiler/nir/nir_opt_memcpy.c b/src/compiler/nir/nir_opt_memcpy.c
new file mode 100644
index 00000000000..fe2a3039be1
--- /dev/null
+++ b/src/compiler/nir/nir_opt_memcpy.c
@@ -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;
+}
More information about the mesa-commit
mailing list