Mesa (master): glsl/nir: Add optimization pass for access flags
GitLab Mirror
gitlab-mirror at kemper.freedesktop.org
Wed Jun 19 12:19:10 UTC 2019
Module: Mesa
Branch: master
Commit: f2d0e48ddc74d01be1d3e259c5687ff37e4c8d69
URL: http://cgit.freedesktop.org/mesa/mesa/commit/?id=f2d0e48ddc74d01be1d3e259c5687ff37e4c8d69
Author: Connor Abbott <cwabbott0 at gmail.com>
Date: Tue Jun 4 14:13:13 2019 +0200
glsl/nir: Add optimization pass for access flags
Right now, this just deduces when we can arbitrarily reorder SSBO and
image loads, matching the existing logic in radeonsi's TGSI->LLVM pass.
This approach can't handle some things that nir_opt_copy_prop_vars can,
but it can handle images, and with GCM it lets us hoist reads outside of
loops. We can also pass this information to LLVM which lets it do its
own optimizations on it.
This is GLSL only as I haven't tested it on Vulkan yet, and it would
probably need a few changes to work there.
Reviewed-by: Timothy Arceri <tarceri at itsqueeze.com>
---
src/compiler/Makefile.sources | 1 +
src/compiler/glsl/gl_nir.h | 2 +
src/compiler/glsl/gl_nir_opt_access.c | 320 ++++++++++++++++++++++++++++++++++
src/compiler/glsl/meson.build | 1 +
4 files changed, 324 insertions(+)
diff --git a/src/compiler/Makefile.sources b/src/compiler/Makefile.sources
index 324e5a0e939..d6b12fcea1b 100644
--- a/src/compiler/Makefile.sources
+++ b/src/compiler/Makefile.sources
@@ -36,6 +36,7 @@ LIBGLSL_FILES = \
glsl/gl_nir_link_xfb.c \
glsl/gl_nir_linker.c \
glsl/gl_nir_linker.h \
+ glsl/gl_nir_opt_access.c \
glsl/gl_nir.h \
glsl/glsl_parser_extras.cpp \
glsl/glsl_parser_extras.h \
diff --git a/src/compiler/glsl/gl_nir.h b/src/compiler/glsl/gl_nir.h
index 5a5e5d28747..9907f732ed7 100644
--- a/src/compiler/glsl/gl_nir.h
+++ b/src/compiler/glsl/gl_nir.h
@@ -44,6 +44,8 @@ bool gl_nir_lower_samplers_as_deref(nir_shader *shader,
bool gl_nir_lower_buffers(nir_shader *shader,
const struct gl_shader_program *shader_program);
+bool gl_nir_opt_access(nir_shader *shader);
+
#ifdef __cplusplus
}
#endif
diff --git a/src/compiler/glsl/gl_nir_opt_access.c b/src/compiler/glsl/gl_nir_opt_access.c
new file mode 100644
index 00000000000..1affeb3881e
--- /dev/null
+++ b/src/compiler/glsl/gl_nir_opt_access.c
@@ -0,0 +1,320 @@
+/*
+ * Copyright © 2019 Valve 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/nir.h"
+#include "gl_nir.h"
+
+/* This pass optimizes GL access qualifiers. So far it does two things:
+ *
+ * - Infer readonly when it's missing.
+ * - Infer ACCESS_CAN_REORDER when the following are true:
+ * - Either there are no writes, or ACCESS_NON_WRITEABLE and ACCESS_RESTRICT
+ * are both set. In either case there are no writes to the underlying
+ * memory.
+ * - If ACCESS_COHERENT is set, then there must be no memory barriers
+ * involving the access. Coherent accesses may return different results
+ * before and after barriers.
+ * - ACCESS_VOLATILE is not set.
+ *
+ * If these conditions are true, then image and buffer reads may be treated as
+ * if they were uniform buffer reads, i.e. they may be arbitrarily moved,
+ * combined, rematerialized etc.
+ */
+
+struct access_state {
+ struct set *vars_written;
+ bool images_written;
+ bool buffers_written;
+ bool image_barriers;
+ bool buffer_barriers;
+};
+
+static void
+gather_intrinsic(struct access_state *state, nir_intrinsic_instr *instr)
+{
+ nir_variable *var;
+ switch (instr->intrinsic) {
+ case nir_intrinsic_image_deref_store:
+ case nir_intrinsic_image_deref_atomic_add:
+ case nir_intrinsic_image_deref_atomic_min:
+ case nir_intrinsic_image_deref_atomic_max:
+ case nir_intrinsic_image_deref_atomic_and:
+ case nir_intrinsic_image_deref_atomic_or:
+ case nir_intrinsic_image_deref_atomic_xor:
+ case nir_intrinsic_image_deref_atomic_exchange:
+ case nir_intrinsic_image_deref_atomic_comp_swap:
+ case nir_intrinsic_image_deref_atomic_fadd:
+ var = nir_intrinsic_get_var(instr, 0);
+
+ /* In OpenGL, buffer images use normal buffer objects, whereas other
+ * image types use textures which cannot alias with buffer objects.
+ * Therefore we have to group buffer samplers together with SSBO's.
+ */
+ if (glsl_get_sampler_dim(var->type) == GLSL_SAMPLER_DIM_BUF)
+ state->buffers_written = true;
+ else
+ state->images_written = true;
+
+ if (var->data.mode == nir_var_uniform)
+ _mesa_set_add(state->vars_written, var);
+ break;
+
+ case nir_intrinsic_bindless_image_store:
+ case nir_intrinsic_bindless_image_atomic_add:
+ case nir_intrinsic_bindless_image_atomic_min:
+ case nir_intrinsic_bindless_image_atomic_max:
+ case nir_intrinsic_bindless_image_atomic_and:
+ case nir_intrinsic_bindless_image_atomic_or:
+ case nir_intrinsic_bindless_image_atomic_xor:
+ case nir_intrinsic_bindless_image_atomic_exchange:
+ case nir_intrinsic_bindless_image_atomic_comp_swap:
+ case nir_intrinsic_bindless_image_atomic_fadd:
+ if (nir_intrinsic_image_dim(instr) == GLSL_SAMPLER_DIM_BUF)
+ state->buffers_written = true;
+ else
+ state->images_written = true;
+ break;
+
+ case nir_intrinsic_store_deref:
+ case nir_intrinsic_deref_atomic_add:
+ case nir_intrinsic_deref_atomic_imin:
+ case nir_intrinsic_deref_atomic_umin:
+ case nir_intrinsic_deref_atomic_imax:
+ case nir_intrinsic_deref_atomic_umax:
+ case nir_intrinsic_deref_atomic_and:
+ case nir_intrinsic_deref_atomic_or:
+ case nir_intrinsic_deref_atomic_xor:
+ case nir_intrinsic_deref_atomic_exchange:
+ case nir_intrinsic_deref_atomic_comp_swap:
+ case nir_intrinsic_deref_atomic_fadd:
+ case nir_intrinsic_deref_atomic_fmin:
+ case nir_intrinsic_deref_atomic_fmax:
+ case nir_intrinsic_deref_atomic_fcomp_swap:
+ var = nir_intrinsic_get_var(instr, 0);
+ if (var->data.mode != nir_var_mem_ssbo)
+ break;
+
+ _mesa_set_add(state->vars_written, var);
+ state->buffers_written = true;
+
+ case nir_intrinsic_memory_barrier:
+ state->buffer_barriers = true;
+ state->image_barriers = true;
+ break;
+
+ case nir_intrinsic_memory_barrier_buffer:
+ state->buffer_barriers = true;
+ break;
+
+ case nir_intrinsic_memory_barrier_image:
+ state->image_barriers = true;
+ break;
+
+ default:
+ break;
+ }
+}
+
+static bool
+process_variable(struct access_state *state, nir_variable *var)
+{
+ if (var->data.mode != nir_var_mem_ssbo &&
+ !(var->data.mode == nir_var_uniform &&
+ glsl_type_is_image(var->type)))
+ return false;
+
+ /* Ignore variables we've already marked */
+ if (var->data.image.access & ACCESS_CAN_REORDER)
+ return false;
+
+ if (!(var->data.image.access & ACCESS_NON_WRITEABLE) &&
+ !_mesa_set_search(state->vars_written, var)) {
+ var->data.image.access |= ACCESS_NON_WRITEABLE;
+ return true;
+ }
+
+ return false;
+}
+
+static bool
+can_reorder(struct access_state *state, enum gl_access_qualifier access,
+ bool is_buffer, bool is_ssbo)
+{
+ bool is_any_written = is_buffer ? state->buffers_written :
+ state->images_written;
+
+ /* Can we guarantee that the underlying memory is never written? */
+ if (!is_any_written ||
+ ((access & ACCESS_NON_WRITEABLE) &&
+ (access & ACCESS_RESTRICT))) {
+ /* Note: memoryBarrierBuffer() is only guaranteed to flush buffer
+ * variables and not imageBuffer's, so we only consider the GL-level
+ * type here.
+ */
+ bool is_any_barrier = is_ssbo ?
+ state->buffer_barriers : state->image_barriers;
+
+ return (!is_any_barrier || !(access & ACCESS_COHERENT)) &&
+ !(access & ACCESS_VOLATILE);
+ }
+
+ return false;
+}
+
+static bool
+process_intrinsic(struct access_state *state, nir_intrinsic_instr *instr)
+{
+ switch (instr->intrinsic) {
+ case nir_intrinsic_bindless_image_load:
+ if (nir_intrinsic_access(instr) & ACCESS_CAN_REORDER)
+ return false;
+
+ /* We have less information about bindless intrinsics, since we can't
+ * always trace uses back to the variable. Don't try and infer if it's
+ * read-only, unless there are no image writes at all.
+ */
+ bool progress = false;
+ bool is_buffer =
+ nir_intrinsic_image_dim(instr) == GLSL_SAMPLER_DIM_BUF;
+
+ bool is_any_written =
+ is_buffer ? state->buffers_written : state->images_written;
+
+ if (!(nir_intrinsic_access(instr) & ACCESS_NON_WRITEABLE) &&
+ !is_any_written) {
+ progress = true;
+ nir_intrinsic_set_access(instr,
+ nir_intrinsic_access(instr) |
+ ACCESS_NON_WRITEABLE);
+ }
+
+ if (can_reorder(state, nir_intrinsic_access(instr), is_buffer, false)) {
+ progress = true;
+ nir_intrinsic_set_access(instr,
+ nir_intrinsic_access(instr) |
+ ACCESS_CAN_REORDER);
+ }
+
+ return progress;
+
+ case nir_intrinsic_load_deref:
+ case nir_intrinsic_image_deref_load: {
+ nir_variable *var = nir_intrinsic_get_var(instr, 0);
+
+ if (instr->intrinsic == nir_intrinsic_load_deref &&
+ var->data.mode != nir_var_mem_ssbo)
+ return false;
+
+ if (nir_intrinsic_access(instr) & ACCESS_CAN_REORDER)
+ return false;
+
+ bool progress = false;
+
+ /* Check if we were able to mark the whole variable non-writeable */
+ if (!(nir_intrinsic_access(instr) & ACCESS_NON_WRITEABLE) &&
+ var->data.image.access & ACCESS_NON_WRITEABLE) {
+ progress = true;
+ nir_intrinsic_set_access(instr,
+ nir_intrinsic_access(instr) |
+ ACCESS_NON_WRITEABLE);
+ }
+
+ bool is_ssbo = var->data.mode == nir_var_mem_ssbo;
+
+ bool is_buffer = is_ssbo ||
+ glsl_get_sampler_dim(glsl_without_array(var->type)) == GLSL_SAMPLER_DIM_BUF;
+
+ if (can_reorder(state, nir_intrinsic_access(instr), is_buffer, is_ssbo)) {
+ progress = true;
+ nir_intrinsic_set_access(instr,
+ nir_intrinsic_access(instr) |
+ ACCESS_CAN_REORDER);
+ }
+
+ return progress;
+ }
+
+ default:
+ return false;
+ }
+}
+
+static bool
+opt_access_impl(struct access_state *state,
+ nir_function_impl *impl)
+{
+ bool progress = false;
+
+ nir_foreach_block(block, impl) {
+ nir_foreach_instr(instr, block) {
+ if (instr->type == nir_instr_type_intrinsic)
+ progress |= process_intrinsic(state,
+ nir_instr_as_intrinsic(instr));
+ }
+ }
+
+ if (progress) {
+ nir_metadata_preserve(impl,
+ nir_metadata_block_index |
+ nir_metadata_dominance |
+ nir_metadata_live_ssa_defs |
+ nir_metadata_loop_analysis);
+ }
+
+
+ return progress;
+}
+
+bool
+gl_nir_opt_access(nir_shader *shader)
+{
+ struct access_state state = {
+ .vars_written = _mesa_pointer_set_create(NULL),
+ };
+
+ bool progress = false;
+
+ nir_foreach_function(func, shader) {
+ if (func->impl) {
+ nir_foreach_block(block, func->impl) {
+ nir_foreach_instr(instr, block) {
+ if (instr->type == nir_instr_type_intrinsic)
+ gather_intrinsic(&state, nir_instr_as_intrinsic(instr));
+ }
+ }
+ }
+ }
+
+ nir_foreach_variable(var, &shader->uniforms)
+ progress |= process_variable(&state, var);
+
+ nir_foreach_function(func, shader) {
+ if (func->impl) {
+ progress |= opt_access_impl(&state, func->impl);
+ }
+ }
+
+ _mesa_set_destroy(state.vars_written, NULL);
+ return progress;
+}
+
diff --git a/src/compiler/glsl/meson.build b/src/compiler/glsl/meson.build
index 600b9f5057e..615c1ad1761 100644
--- a/src/compiler/glsl/meson.build
+++ b/src/compiler/glsl/meson.build
@@ -84,6 +84,7 @@ files_libglsl = files(
'gl_nir_link_xfb.c',
'gl_nir_linker.c',
'gl_nir_linker.h',
+ 'gl_nir_opt_access.c',
'gl_nir.h',
'glsl_parser_extras.cpp',
'glsl_parser_extras.h',
More information about the mesa-commit
mailing list